Domain Theory

Corrected and expanded version

Samson Abramsky1 and Achim Jung2

This text is based on the chapter Domain Theory in the Handbook for Logic in Computer Science, volume 3, edited by S. Abramsky, Dov M. Gabbay, and T. S. E. Maibaum, published by Clarendon Press, Oxford in 1994. While the numbering of all theorems and denitions has been kept the same, we have included comments and corrections which we have received over the years. For ease of reading, small typographical errors have simply been corrected. Where we felt the original text gave a misleading impression, we have included additional explanations, clearly marked as such. If you wish to refer to this text, then please cite the published original version where possible, or otherwise this on-line version which we try to keep available from the page http://www.cs.bham.ac.uk/axj/papers.html We will be grateful to receive further comments or suggestions. Please send them to [email protected] So far, we have received comments and/or corrections from Joseph D. Darcy, Mohamed El-Zawawy, Homeira Pajoohesh, and Dominic van der Zypen.

1 Computing Laboratory, University of Oxford, Wolfson Building, Parks Road, Oxford, OX1 3QD, England. 2 School of Computer Science, University of Birmingham, Edgbaston, Birmingham, B15 2TT, England.

Contents
1 Introduction and Overview 1.1 Origins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 Our approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 Domains individually 2.1 Convergence . . . . . . . . . . . . . . . 2.1.1 Posets and preorders . . . . . . . 2.1.2 Notation from order theory . . . . 2.1.3 Monotone functions . . . . . . . 2.1.4 Directed sets . . . . . . . . . . . 2.1.5 Directed-complete partial orders . 2.1.6 Continuous functions . . . . . . . 2.2 Approximation . . . . . . . . . . . . . . 2.2.1 The order of approximation . . . 2.2.2 Bases in dcpos . . . . . . . . . . 2.2.3 Continuous and algebraic domains 2.2.4 Comments on possible variations 2.2.5 Useful properties . . . . . . . . . 2.2.6 Bases as objects . . . . . . . . . . 2.3 Topology . . . . . . . . . . . . . . . . . 2.3.1 The Scott-topology on a dcpo . . 2.3.2 The Scott-topology on domains . 5 5 7 7 10 10 10 11 13 13 14 15 17 18 18 19 22 24 25 29 29 30 34 34 34 35 36 39 40 41 42 43 44 45 45 45 46 46 47 51

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

3 Domains collectively 3.1 Comparing domains . . . . . . . . . . . . . . . 3.1.1 Retractions . . . . . . . . . . . . . . . 3.1.2 Idempotents . . . . . . . . . . . . . . . 3.1.3 Adjunctions . . . . . . . . . . . . . . . 3.1.4 Projections and sub-domains . . . . . . 3.1.5 Closures and quotient domains . . . . . 3.2 Finitary constructions . . . . . . . . . . . . . . 3.2.1 Cartesian product . . . . . . . . . . . . 3.2.2 Function space . . . . . . . . . . . . . 3.2.3 Coalesced sum . . . . . . . . . . . . . 3.2.4 Smash product and strict function space 3.2.5 Lifting . . . . . . . . . . . . . . . . . 3.2.6 Summary . . . . . . . . . . . . . . . . 3.3 Innitary constructions . . . . . . . . . . . . . 3.3.1 Limits and colimits . . . . . . . . . . . 3.3.2 The limit-colimit coincidence . . . . . 3.3.3 Bilimits of domains . . . . . . . . . . .

4 Cartesian closed categories of domains 4.1 Local uniqueness: Lattice-like domains 4.2 Finite choice: Compact domains . . . . 4.2.1 Binite domains . . . . . . . . 4.2.2 FS-domains . . . . . . . . . . . 4.2.3 Coherence . . . . . . . . . . . 4.3 The hierarchy of categories of domains . 4.3.1 Domains with least element . . 4.3.2 Domains without least element .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

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

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

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

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

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

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

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

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

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

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

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

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

54 55 57 57 60 61 63 63 65 68 68 68 68 69 70 70 71 73 74 74 76 77 79 79 80 81 82

5 Recursive domain equations 5.1 Examples . . . . . . . . . . . . . . . . . . . . 5.1.1 Genuine equations . . . . . . . . . . . 5.1.2 Recursive denitions . . . . . . . . . . 5.1.3 Data types . . . . . . . . . . . . . . . . 5.2 Construction of solutions . . . . . . . . . . . . 5.2.1 Continuous functors . . . . . . . . . . 5.2.2 Local continuity . . . . . . . . . . . . 5.2.3 Parameterized equations . . . . . . . . 5.3 Canonicity . . . . . . . . . . . . . . . . . . . . 5.3.1 Invariance and minimality . . . . . . . 5.3.2 Initiality and nality . . . . . . . . . . 5.3.3 Mixed variance . . . . . . . . . . . . . 5.4 Analysis of solutions . . . . . . . . . . . . . . 5.4.1 Structural induction on terms . . . . . . 5.4.2 Admissible relations . . . . . . . . . . 5.4.3 Induction with admissible relations . . 5.4.4 Co-induction with admissible relations

6 Equational theories 6.1 General techniques . . . . . . . . . . . . . . . . . . 6.1.1 Free dcpo-algebras . . . . . . . . . . . . . . 6.1.2 Free continuous domain-algebras . . . . . . 6.1.3 Least elements and strict algebras . . . . . . 6.2 Powerdomains . . . . . . . . . . . . . . . . . . . . . 6.2.1 The convex or Plotkin powerdomain . . . . . 6.2.2 One-sided powerdomains . . . . . . . . . . . 6.2.3 Topological representation theorems . . . . . 6.2.4 Hyperspaces and probabilistic powerdomains 7 Domains and logic 7.1 Stone duality . . . . . . . . . . . . . . . 7.1.1 Approximation and distributivity . 7.1.2 From spaces to lattices . . . . . . 7.1.3 From lattices to topological spaces 7.1.4 The basic adjunction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

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

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

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

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

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

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

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

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

84 . 84 . 84 . 86 . 91 . 92 . 92 . 95 . 96 . 102 . . . . . 105 105 105 108 109 110

7.2

7.3

Some equivalences . . . . . . . . . . . . . . . . . . . 7.2.1 Sober spaces and spatial lattices . . . . . . . . 7.2.2 Properties of sober spaces . . . . . . . . . . . 7.2.3 Locally compact spaces and continuous lattices 7.2.4 Coherence . . . . . . . . . . . . . . . . . . . 7.2.5 Compact-open sets and spectral spaces . . . . 7.2.6 Domains . . . . . . . . . . . . . . . . . . . . 7.2.7 Summary . . . . . . . . . . . . . . . . . . . . The logical viewpoint . . . . . . . . . . . . . . . . . . 7.3.1 Working with lattices of compact-open subsets 7.3.2 Constructions: The general technique . . . . . 7.3.3 The function space construction . . . . . . . . 7.3.4 The Plotkin powerlocale . . . . . . . . . . . . 7.3.5 Recursive domain equations . . . . . . . . . . 7.3.6 Languages for types, properties, and points . .

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

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

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

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

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

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

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

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

111 111 113 115 116 116 118 120 120 120 125 129 131 135 136 144 144 144 144 145 145 146 146 147 148 148 149 150 151 152 164

8 Further directions 8.1 Further topics in Classical Domain Theory . . . . . 8.1.1 Effectively given domains . . . . . . . . . . . 8.1.2 Universal Domains . . . . . . . . . . . . . . . 8.1.3 Domain-theoretic semantics of polymorphism . 8.1.4 Information Systems . . . . . . . . . . . . . . 8.2 Stability and Sequentiality . . . . . . . . . . . . . . . 8.3 Reformulations of Domain Theory . . . . . . . . . . . 8.3.1 Predomains and partial functions . . . . . . . . 8.3.2 Computational Monads . . . . . . . . . . . . . 8.3.3 Linear Types . . . . . . . . . . . . . . . . . . 8.4 Axiomatic Domain Theory . . . . . . . . . . . . . . . 8.5 Synthetic Domain Theory . . . . . . . . . . . . . . . . 9 Guide to the literature References Index

1 Introduction and Overview

1.1 Origins
Let us begin with the problems which gave rise to Domain Theory: 1. Least xpoints as meanings of recursive denitions. Recursive denitions of procedures, data structures and other computational entities abound in programming languages. Indeed, recursion is the basic effective mechanism for describing innite computational behaviour in nite terms. Given a recursive denition:

How can we give a non-circular account of its meaning? Suppose we are working inside some mathematical structure . We want to nd an element such that substituting for in (1) yields a valid equation. The right-hand-side of (1) can be read as a function of , semantically as . We can now see that we are asking for an element such that that is, for a xpoint of . Moreover, we want a uniform canonical method for constructing such xpoints for arbitrary structures and functions within our framework. Elementary considerations show that the usual categories of mathematical structures either fail to meet this requirement at all (sets, topological spaces) or meet it in a trivial fashion (groups, vector spaces). 2. Recursive domain equations. Apart from recursive denitions of computational objects, programming languages also abound, explicitly or implicitly, in recursive denitions of datatypes. The classical example is the type-free calculus [Bar84]. To give a mathematical semantics for the -calculus is to nd a mathematical structure such that terms of the -calculus can be interpreted as elements of in such a way that application in the calculus is interpreted by function application. Now consider the self-application term . By the usual condition for type-compatibility of a function with its argument, we see that if the second occurrence of in has type , and the whole term has type , then the rst occurrence must have, or be construable as having, type . Thus we are led to the requirement that we have

If we view as a functor C C C over a suitable category C of mathematical structures, then we are looking for a xpoint . Thus recursive datatypes again lead to a requirement for xpoints, but now lifted to the functorial level. Again we want such xpoints to exist uniformly and canonically. This second requirement is even further beyond the realms of ordinary mathematical experience than the rst. Collectively, they call for a novel mathematical theory to serve as a foundation for the semantics of programming languages.

 2

P  Q' C A 

    4

  3 32

10#) $ !  &%#"!  $  2

$ IGFE DC H ! A 8  $ 6  B9@#% 5

 (' 

 4 

  

8 #6 5 $

8  $6 97#% 5

A rst step towards Domain Theory is the familiar result that every monotone function on a complete lattice, or more generally on a directed-complete partial order with least element, has a least xpoint. (For an account of the history of this result, see [LNS82].) Some early uses of this result in the context of formal language theory were [Ard60, GR62]. It had also found applications in recursion theory [Kle52, Pla64]. Its application to the semantics of rst-order recursion equations and owcharts was already well-established among Computer Scientists by the end of the 1960s [dBS69, Bek69, Bek71, Par69]. But Domain Theory proper, at least as we understand the term, began in 1969, and was unambiguously the creation of one man, Dana Scott [Sco69, Sco70, Sco71, Sco72, Sco93]. In particular, the following key insights can be identied in his work: 1. Domains as types. The fact that suitable categories of domains are cartesian closed, and hence give rise to models of typed -calculi. More generally, that domains give mathematical meaning to a broad class of data-structuring mechanisms. 2. Recursive types. Scotts key construction was a solution to the domain equation

thus giving the rst mathematical model of the type-free -calculus. This led to a general theory of solutions of recursive domain equations. In conjunction with (1), this showed that domains form a suitable universe for the semantics of programming languages. In this way, Scott provided a mathematical foundation for the work of Christopher Strachey on denotational semantics [MS76, Sto77]. This combination of descriptive richness and a powerful and elegant mathematical theory led to denotational semantics becoming a dominant paradigm in Theoretical Computer Science. 3. Continuity vs. Computability. Continuity is a central pillar of Domain theory. It serves as a qualitative approximation to computability. In other words, for most purposes to detect whether some construction is computationally feasible it is sufcient to check that it is continuous; while continuity is an algebraic condition, which is much easier to handle than computability. In order to give this idea of continuity as a smoothed-out version of computability substance, it is not sufcient to work only with a notion of completeness or convergence; one also needs a notion of approximation, which does justice to the idea that innite objects are given in some coherent way as limits of their nite approximations. This leads to considering, not arbitrary complete partial orders, but the continuous ones. Indeed, Scotts early work on Domain Theory was seminal to the subsequent extensive development of the theory of continuous lattices, which also drew heavily on ideas from topology, analysis, topological algebra and category theory [GHK 80]. 4. Partial information. A natural concomitant of the notion of approximation in domains is that they form the basis of a theory of partial information, which extends the familiar notion of partial function to encompass a whole spectrum of 6

8 $ 9746  5 A 

degrees of denedness. This has important applications to the semantics of programming languages, where such multiple degrees of denition play a key role in the analysis of computational notions such as lazy vs. eager evaluation, and call-by-name vs. call-by-value parameter-passing mechanisms for procedures. General considerations from recursion theory dictate that partial functions are unavoidable in any discussion of computability. Domain Theory provides an appropriately abstract, structural setting in which these notions can be lifted to higher types, recursive types, etc.

1.2 Our approach

It is a striking fact that, although Domain Theory has been around for a quartercentury, no book-length treatment of it has yet been published. Quite a number of books on semantics of programming languages, incorporating substantial introductions to domain theory as a necessary tool for denotational semantics, have appeared [Sto77, Sch86, Gun92b, Win93]; but there has been no text devoted to the underlying mathematical theory of domains. To make an analogy, it is as if many Calculus textbooks were available, offering presentations of some basic analysis interleaved with its applications in modelling physical and geometrical problems; but no textbook of Real Analysis. Although this Handbook Chapter cannot offer the comprehensive coverage of a full-length textbook, it is nevertheless written in the spirit of a presentation of Real Analysis. That is, we attempt to give a crisp, efcient presentation of the mathematical theory of domains without excursions into applications. We hope that such an account will be found useful by readers wishing to acquire some familiarity with Domain Theory, including those who seek to apply it. Indeed, we believe that the chances for exciting new applications of Domain Theory will be enhanced if more people become aware of the full richness of the mathematical theory.

1.3 Overview
Domains individually We begin by developing the basic mathematical language of Domain Theory, and then present the central pillars of the theory: convergence and approximation. We put considerable emphasis on bases of continuous domains, and show how the theory can be developed in terms of these. We also give a rst presentation of the topological view of Domain Theory, which will be a recurring theme. Domains collectively We study special classes of maps which play a key role in domain theory: retractions, adjunctions, embeddings and projections. We also look at construction on domains such as products, function spaces, sums and lifting; and at bilimits of directed systems of domains and embeddings.

Cartesian closed categories of domains A particularly important requirement on categories of domains is that they should be cartesian closed (i.e. closed under function spaces). This creates a tension with the requirement for a good theory of approximation for domains, since neither the category CONT of all continuous domains, nor the category ALG of all algebraic domains is cartesian closed. This leads to a non-trivial analysis of necessary and sufcient conditions on domains to ensure closure under function spaces, and striking results on the classication of the maximal cartesian closed full subcategories of CONT and ALG. This material is based on [Jun89, Jun90]. Recursive domain equations The theory of recursive domain equations is presented. Although this material formed the very starting point of Domain Theory, a full clarication of just what canonicity of solutions means, and how it can be translated into proof principles for reasoning about these canonical solutions, has only emerged over the past two or three years, through the work of Peter Freyd and Andrew Pitts [Fre91, Fre92, Pit93b]. We make extensive use of their insights in our presentation. Equational theories We present a general theory of the construction of free algebras for inequational theories over continuous domains. These results, and the underlying constructions in terms of bases, appear to be new. We then apply this general theory to powerdomains and give a comprehensive treatment of the Plotkin, Hoare and Smyth powerdomains. In addition to characterizing these as free algebras for certain inequational theories, we also prove representation theorems which characterize a powerdomain over as a certain space of subsets of ; these results make considerable use of topological methods. Domains and logic We develop the logical point of view of Domain Theory, in which domains are characterized in terms of their observable properties, and functions in terms of their actions on these properties. The general framework for this is provided by Stone duality; we develop the rudiments of Stone duality in some generality, and then specialize it to domains. Finally, we present Domain Theory in Logical Form [Abr91b], in which a metalanguage of types and terms suitable for denotational semantics is extended with a language of properties, and presented axiomatically as a programming logic in such a way that the lattice of properties over each type is the Stone dual of the domain denoted by that type, and the prime lter of properties which can be proved to hold of a term correspond under Stone duality to the domain element denoted by that term. This yields a systematic way of moving back and forth between the logical and denotational descriptions of some computational situation, each determining the other up to isomorphism.

Acknowledgements
We would like to thank Ji Ad mek, Reinhold Heckmann, Michael Huth, Mathias r a Kegelmann, Philipp S nderhauf, and Paul Taylor for very careful proof reading. Achim u Jung would particularly like to thank the people from the Domain Theory Group at Darmstadt, who provided a stimulating and supportive environment. Our major intellectual debts, inevitably, are to Dana Scott and Gordon Plotkin. The more we learn about Domain Theory, the more we appreciate the depth of their insights.

2 Domains individually
We will begin by introducing the basic language of Domain Theory. Most topics we deal with in this section are treated more thoroughly and at a more leisurely pace in [DP90].

2.1 Convergence
2.1.1 Posets and preorders

Denition 2.1.1. A set with a binary relation : poset if the following holds for all

Small nite partially ordered sets can be drawn as line diagrams (Hasse diagrams). Examples are given in Figure 1. We will also allow ourselves to draw innite posets by showing a nite part which illustrates the building principle. Three examples are given in Figure 2. We prefer the notation to the more common because the order on domains we are studying here often coexists with an otherwise unrelated intrinsic order. The at and lazy natural numbers from Figure 2 illustrate this. If we drop antisymmetry from our list of requirements then we get what is known as preorders. This does not change the theory very much. As is easily seen, the subrelation is in any case an equivalence relation and if two elements from two equivalence classes are related by , then so is any pair of elements from and . We can therefore pass from a preorder to a canonical partially ordered set by taking equivalence classes. Pictorially, the situation then looks as in Figure 3. Many notions from the theory of ordered sets make sense even if reexivity fails. Hence we may sum up these considerations with the slogan: Order theory is the study of transitive relations. A common way to extract the order-theoretic content from a relation is to pass to the transitive closure of , dened as . Ordered sets can be turned upside down:
B  D @ B CA @

Proposition 2.1.2. If

is an ordered set then so is

The at booleans

The four-element lattice

The four-element chain

Figure 1: A few posets drawn as line diagrams. 10

# 7 5 3 1 ) ' % 986420(&$

P GFE

YE a`

YE

" !

 P  

E ` Q X WUS V T



3.



 



2.

  
 

  

E ` aY

"

YE Q I H RP4G

1.

(Reexivity)

(Transitivity) (Antisymmetry)

 P 9#P 

is called a partially ordered set or

 
#

 

ordinal

at

lazy

Figure 2: Three versions of the natural numbers.

Figure 3: A preorder whose canonical quotient is the four-element lattice. One consequence of this observation is that each of the concepts introduced below has a dual counterpart. 2.1.2 Notation from order theory The following concepts form the core language of order theory.

Denition 2.1.3. Let

be an ordered set.

implies for all . We denote 1. A subset of is an upper set if by the set of all elements above some element of . If no confusion is to be feared then we abbreviate as . The dual notions are lower set and . 2. An element is called an upper bound for a subset , if is above every element of . We often write in this situation. We denote by the set of all upper bounds of . Dually, denotes the set of lower bounds of . 3. An element is maximal if there is no other element of above it: . Minimal elements are dened dually. For a subset the minimal elements of are called minimal upper bounds of . The set of all minimal upper bounds of is denoted by . 11


 ! 



 #

 



 





 $







  

0
E

YE

 P

YE

1
E

YE

` `

YE

 $

 

 '

E
E

 "

4. If all elements of are below a single element , then is said to be the largest element. The dually dened least element of a poset is also called bottom and is commonly denoted by . In the presence of a least element we speak of a pointed poset. 5. If for a subset the set of upper bounds has a least element , then is called the supremum or join. We write in this case. In the other direction we speak of inmum or meet and write . 6. A partially ordered set is a -semilattice ( -semilattice) if the supremum (inmum) for each pair of elements exists. If is both a - and a -semilattice then is called a lattice. A lattice is complete if suprema and inma exist for all subsets. The operations of forming suprema, resp. inma, have a few basic properties which we will use throughout this text without mentioning them further. Proposition 2.1.4. Let be a poset such that the suprema and inma occurring in the and all are subsets of .) following formulae exist. (
"

2.

Proof. We illustrate order theoretic reasoning with suprema by showing (3). The element is above each element by (1), so it is an upper bound of the set . Since is the least upper bound of this set, we have . Conversely, each is contained in some and therefore below the corresponding which in turn is below . Hence the right hand side is an upper bound of and as is the least such, we also have . Let us conclude this subsection by looking at an important family of examples of complete lattices. Suppose is a set and is a family of subsets of . We call a closure system if it is closed under the formation of intersections, that is, whenever each member of a family belongs to then so does . Because we have allowed the index set to be empty, this implies that is in . We call the members of hulls or closed sets. Given an arbitrary subset of , one can form . This is the least superset of which belongs to and is called the hull or the closure of . Proposition 2.1.5. Every closure system is a complete lattice with respect to inclusion. Proof. Inma are given by intersections and for the supremum one takes the closure of the union.

12

  

'

'

                    

'

'

'

"

'

3. If

    

and

then

"

'

"

'

1.

implies

and

and similarly for the inmum.

 

 

"

"

2.1.3 Monotone functions

Monotone is really an abbreviation for monotone order-preserving, but since we have no use for monotone order-reversing maps ( ), we have opted for the shorter expression. Alternative terminology is isotone (vs. antitone) or the other half of the full expression: order-preserving mapping. The set of all monotone functions between two posets, when ordered pointwise (i.e. if for all , ), gives rise to another partially ordered set, the monotone function space between and . The category POSET of posets and monotone maps has pleasing properties, see Exercise 2.3.9(9). Proposition 2.1.7. If is a complete lattice then every monotone map from a xpoint. The least of these is given by

the largest by

and by monotonicity.

For lattices, the converse is also true: The existence of xpoints for monotone maps implies completeness. But the proof is much harder and relies on the Axiom of Choice, see [Mar76]. 2.1.4 Directed sets Denition 2.1.8. Let be a poset. A subset of is directed, if it is nonempty and each pair of elements of has an upper bound in . If a directed set has a supremum then this is denoted by . Directed lower sets are called ideals. Ideals of the form are called principal. The dual notions are ltered set and (principal) lter. Simple examples of directed sets are chains. These are non-empty subsets which are totally ordered, i.e. for each pair either or holds. The chain of natural numbers with their natural order is particularly simple; subsets of a poset isomorphic to it are usually called -chains. Another frequent type of directed set is given by the set of nite subsets of an arbitrary set. Using this and Proposition 2.1.4(3), we get the following useful decomposition of general suprema. Proposition 2.1.9. Let be a non-empty subset of a -semilattice for which ists. Then the join of can also be written as

13

nite and non-empty

Proof. Let and . For each and . Taking the inmum we get follows. On the other hand, always implies Applying this to yields and hence .

       (             "                   P       

we have

  




 

   


P

 

      

Denition 2.1.6. Let and called monotone if for all

be partially ordered sets. A function with we also have

! $ )  
in

 P

 

is

8 B 5 $ 6

to

has

 



ex-

General directed sets, on the other hand, may be quite messy and unstructured. Sometimes one can nd a well-behaved conal subset, such as a chain, where we say that is conal in , if for all there is an above it. Such a conal subset will have the same supremum (if it exists). But conal chains do not always exist, as Exercise 2.3.9(6) shows. Still, every directed set may be thought of as being equipped externally with a nice structure as we will now work out. Denition 2.1.10. A monotone net in a poset is a monotone function from a directed set into . The set is called the index set of the net. Let be a monotone net. If we are given a monotone function , where is directed and where for all there is with , then we call a subnet of . A monotone net has a supremum in , if the set has a supremum in . Every directed set can be viewed as a monotone net: let the set itself be the index set. On the other hand, the image of a monotone net is a directed set in . So what are nets good for? The answer is given in the following proposition (which seems to have been stated rst in [Kra39]). be a monotone net. Then has a Lemma 2.1.11. Let be a poset and let subnet , whose index set is a lattice in which every principal ideal is nite.

This lemma allows us to base an induction proof on an arbitrary directed set. This was recently applied to settle a long-standing conjecture in lattice theory, see [TT93]. Proposition 2.1.12. Let be directed and be a monotone net. Under the assumption that the indicated directed suprema exist, the following equalities hold:
' ' ' ' ' '

2.1.5 Directed-complete partial orders

Denition 2.1.13. A poset in which every directed subset has a supremum we call a directed-complete partial order, or dcpo for short. Examples 2.1.14. Every complete lattice is also a dcpo. Instances of this are powersets, topologies, subgroup lattices, congruence lattices, and, more generally, closure systems. As Proposition 2.1.9 shows, a lattice which is also a dcpo is almost complete. Only a least element may be missing.

0 1 ( 0 1 ( ( '&   P    Q )P      Q )P       )P    $ H  !  

14

It is obvious that

is monotone and denes a subnet.

$ %# P  

" #"



"

! 

any element of any upper bound of the set

 $ !

Proof. Let be the set of nite subsets of . Clearly, cipal ideal is nite. We dene the mapping of the elements of :

is a lattice in which every prinby induction on the cardinality

 '! $

         

$ I !



 $ !

 

 

"

 )

$ ' !

 (! $

"

 B 

$ ! $  !





Every nite poset is a dcpo.

More examples will follow in the next subsection. There we will also discuss the question of whether directed sets or -chains should be used to dene dcpos. Arbitrarily long chains have the full power of directed sets (despite Exercise 2.3.9(6)) as the following proposition shows. Proposition 2.1.15. A partially ordered set has a supremum. is a dcpo if and only if each chain in

The proof, which uses the Axiom of Choice, goes back to a lemma of Iwamura [Iwa44] and can be found in [Mar76]. The following, which may also be found in [Mar76], complements Proposition 2.1.7 above. Proposition 2.1.16. A pointed poset on has a xpoint. 2.1.6 Continuous functions Denition 2.1.17. Let and be dcpos. A function is (Scott-) continuous if it is monotone and if for each directed subset of we have . We denote the set of all continuous functions from to , ordered pointwise, by . A function between pointed dcpos, which preserves the bottom element, is called strict. We denote the space of all continuous strict functions by . The identity function on a set is denoted by , the constant function with image by . The preservation of joins of directed sets is actually enough to dene continuous maps. In practice, however, one usually needs to show rst that is directed. This is equivalent to monotonicity. Proposition 2.1.18. Let rected suprema in and be dcpos. Then are calculated pointwise. is again a dcpo. Di

is a dcpo if and only if every monotone map

15

 

8 B$ 6  5  ! $   )

8 46  5 $

$6 8 #% 5

8 $#% 6  5

 " 

2 2 2

The set of natural numbers with the usual order does not form a dcpo; we have to add a top element as done in Figure 2. In general, it is a difcult problem how to add points to a poset so that it becomes a dcpo. Using Proposition 2.1.15 below, Markowsky has dened such a completion via chains in [Mar76]. Luckily, we need not worry about this problem in domain theory because here we are usually interested in algebraic or continuous dcpos where a completion is easily dened, see Section 2.2.6 below. The correct formulation of what constitutes a completion, of course, takes also morphisms into account. A general framework is described in [Poi92], Sections 3.3 to 3.6. The points of a locale form a dcpo in the specialization order, see [Vic89, Joh82].

Proof. Let be a directed collection of functions from function, which is dened by . Let
'

to

. Let be directed.

This shows that

is continuous.

The class of all dcpos together with Scott-continuous functions forms a category, which we denote by DCPO. It has strong closure properties as we shall see shortly. For the moment we concentrate on that property of continuous maps which is one of the main reasons for the success of domain theory, namely, that xpoints can be calculated easily and uniformly. Theorem 2.1.19. Let

be a pointed dcpo. on

1. Every continuous function .

has a least xpoint. ,

F %

Proof. (1) The set is a chain. This follows from and the monotonicity of . Using continuity of we get and the latter is clearly equal to . If is any other xpoint of then from we get and so on by induction. Hence is an upper bound of all and that is why it must be above . (2) Let us rst look at the -fold iteration operator which maps to . We show its continuity by induction. The 0th iteration operator equals so nothing has to be shown there. For the induction step let be a directed family of continuous functions on . We calculate: denition ind. hypothesis Prop. 2.1.18 continuity of Prop. 2.1.12
F
 

The pointwise supremum of all iteration operators (which form a chain as we have seen in (1)) is precisely and so the latter is also continuous.

16

) ' 0(%

) ' 0(%

) ' 02%

2. The assignment
%

  Q   Q   Q  C Q  C     C   C 1)9746  5 ! $8 $        )   )  )     Q   $8  $  $ 7B97B6 

is continuous.

$  !

                             9   
'

'

'

'

'

'

'

'

'

 

) ' 0(%

'

'

'

 

 

 

 

 5 !

 C 

 ) 

 

) ' 0(%

be the

It is given by

The least xpoint operator is the mathematical counterpart of recursive and iterative statements in programming languages. When proving a property of such a statement semantically, one often employs the following proof principle which is known under the name xpoint induction (see [Ten91] or any other book on denotational semantics). Call a predicate on (i.e. a subset of) a dcpo admissible if it contains and is closed under suprema of -chains. The following is then easily established:

Lemma 2.1.20. Let be a dcpo, an admissible predicate, and a Scott-continuous function. If it is true that satises whenever satises satises . then it must be true that

We also note the following invariance property of the least xpoint operator. In fact, it characterizes uniquely among all xpoint operators (Exercise 2.3.9(16)). Lemma 2.1.21. Let

and

be pointed dcpos and let

Proof. Using continuity of , commutativity of the diagram, and strictness of we calculate:

2.2 Approximation
In the last subsection we have explained the kind of limits that domain theory deals with, namely, suprema of directed sets. We could have said much more about these convergence spaces called dcpos. But the topic can easily become esoteric and lose its connection with computing. For example, the cardinality of dcpos has not been restricted yet and indeed, we didnt have the tools to sensibly do so (Exercise 2.3.9(18)). We will in this subsection introduce the idea that elements are composed of (or approximated by) simple pieces. This will enrich our theory immensely and will also give the desired connection to semantics. 17

be a commutative diagram of continuous functions where .

is strict. Then

in turn

 7B)  $ !

  

     

  
% ) ' 0(%

) ' 0(%

) ' 0(%

Q



2.2.1 The order of approximation Denition 2.2.1. Let and be elements of a dcpo . We say that approximates if for all directed subsets of , implies for some . We say that is compact if it approximates itself. We introduce the following notation for and :

approximates

The relation is traditionally called way-below relation. M.B. Smyth introduced the expression order of denite renement in [Smy86]. Throughout this text we will refer to it as the order of approximation, even though the relation is not reexive. Other common terminology for compact is nite or isolated. The analogy to nite sets is indeed very strong; however one covers a nite set by a directed collection of sets, will always be contained in some already. In general, approximation is not an absolute property of single points. Rather, we could phrase as is a lot simpler than , which clearly depends on as much as it depends on . An element which is compact approximates every element above it. More generally, we observe the following basic properties of approximation.
6 6

Proposition 2.2.2. Let

be a dcpo. Then the following is true for all ;

1. 2.

2.2.2 Bases in dcpos

Denition 2.2.3. We say that a subset of a dcpo is a basis for , if for every element of the set contains a directed subset with supremum . We call elements of approximants to relative to . We may think of the rational numbers as a basis for the reals (with a top element added, in order to get a dcpo), but other choices are also possible: dyadic numbers, irrational numbers, etc.

"

3. Every superset of

is also a basis for 18

"

2.

contains

"

"

1. For every

the set

is directed and

 

"

Proposition 2.2.4. Let

be a dcpo with basis

'

  B P QP  P

compact

 ( 

 

                   "           P 

"

'

"

  

"

 &



 

"

   ' 

  

 

"

 

  

Proof. (1) It is clear that the join of equals . The point is directedness. From the denition we know there is some directed subset of with . Let now be elements approximating . There must be elements in above , respectively. These have an upper bound in , which by denition belongs to . (2) We have to show that every element of belongs to . Indeed, since there must be an element above . All of is below , so is actually equal to . (3) is immediate from the denition.
"

Corollary 2.2.5. Let

be a dcpo with basis

1. The largest basis for

"

is

itself.
"

2.

is the smallest basis for

if and only if

The only if part of (2) is not a direct consequence of the preceding proposition. We leave its proof as Exercise 2.3.9(26). 2.2.3 Continuous and algebraic domains Denition 2.2.6. A dcpo is called continuous or a continuous domain if it has a basis. It is called algebraic or an algebraic domain if it has a basis of compact elements. We say is -continuous if there exists a countable basis and we call it -algebraic if is a countable basis. Here we are using the word domain for the rst time. Indeed, for us a structure only qualies as a domain if it embodies both a notion of convergence and a notion of approximation. In the light of Proposition 2.2.4 we can reformulate Denition 2.2.6 as follows, avoiding existential quantication. Proposition 2.2.7. holds. 1. A dcpo is continuous if and only if for all

2. It is algebraic if and only if for all

holds.

The word algebraic points to algebra. Let us make this connection precise. Denition 2.2.8. A closure system closed under directed union.

(cf. Section 2.1.2) is called inductive, if it is

Proposition 2.2.9. Every inductive closure system is an algebraic lattice. The compact elements are precisely the nitely generated hulls. and if is a directed family of hulls such Proof. If is the hull of a nite set that , then is already contained in some . Hence hulls of nite sets are compact elements in the complete lattice . On the other hand, every closed set is the directed union of nitely generated hulls, so these form a basis. By Proposition 2.2.4(2), there cannot be any other compact elements.

19

PB 
,
6

   

"

"

"

 P4 
"

      

"

 

 
'

 

"

" !

"



 

"

'

"

  

"

P 6B
'

 
 

Given a group, (or, more generally, an algebra in the sense of universal algebra), then there are two canonical inductive closure systems associated with it, the lattice of subgroups (subalgebras) and the lattice of normal subgroups (congruence relations). Other standard examples of algebraic domains are: Any set with the discrete order is an algebraic domain. In semantics one usually adds a bottom element (standing for divergence) resulting in so-called at domains. (The at natural numbers are shown in Figure 2.) A basis must in either case contain all elements. The set of partial functions between sets and ordered by graph inclusion. Compact elements are those functions which have a nite carrier. It is naturally isomorphic to and to . Every nite poset.

Continuous domains: Every algebraic dcpo is also continuous. This follows directly from the denition. The order of approximation is characterized by if and only if there exists a compact element between and . The unit interval is a continuous lattice. It plays a central role in the theory of continuous lattices, see [GHK 80], Chapter IV and in particular Theorem 2.19. Another way of modelling the real numbers in domain theory is to take all closed intervals of nite length and to order them by reversed inclusion. Single element intervals are maximal in this domain and provide a faithful representation of the real line. A countable basis is given by the set of intervals with rational endpoints. The lattice of open subsets of a sober space forms a continuous lattice if and only if is locally compact. Compact Hausdorff spaces are a special case. Here holds if and only if there exists a compact set such that . This meeting point of topology and domain theory is discussed in detail in [Smy92, Vic89, Joh82, GHK 80] and will also be addressed in Chapter 7.

At this point it may be helpful to give an example of a non-continuous dcpo. The easiest to explain is depicted in Figure 4 (labelled ). We show that the order of approximation on is empty. Pairs and cannot belong to the order of approximation because they are not related in the order. Two points in the same leg are still not approximating because is a directed set with supremum above but containing no element above . A non-continuous distributive complete lattice is much harder to visualize by a line diagram. From what we have said we know that the topology of a sober space which is not locally compact is such a lattice. Exercise 2.3.9(21) discusses this in detail. If is pointed then the order of approximation is non-empty because a bottom element approximates every other element. A basis not only gives approximations for elements, it also approximates the order relation: 20

 

8 #$ 6 5

 (  P

) ' 0(%

 ( P 

8 #6 5 $

8 5 

2 2 2 2 2 2

Figure 4: A continuous ( ) and a non-continuous ( ) dcpo.

` Y

implies there The form in which we will usually apply this proposition is: exists with . A picture of this situation is given in Figure 5. In the light of Proposition 2.2.10 we can now also give a more intuitive reason why the dcpo in Figure 4 is not continuous. A natural candidate for a basis in is the collection of all s and s (certainly, doesnt approximate anything). Proposition 2.2.10 expresses the idea that in a continuous domain all information about how elements are related is contained in the basis already. And the fact that holds in is precisely what is not visible in the would-be basis. Thus, the dcpo should look rather like in the same gure (which indeed is an algebraic domain). Bases allow us to express the continuity of functions in a form reminiscent of the - denition for real-valued functions. and

21

 

Proposition 2.2.11. A map between continuous domains and , respectively, is continuous if and only if for each there exists with .

     

and

 

"

"

 

  

"

with bases

"

Proposition 2.2.10. Let elements of . Then



be a continuous domain with basis and let and are all equivalent.

Figure 5: Basis element witnesses that

is not below . and be

E E E

E E E E ` `

YY

` YY

` Y

E
`

Y Y

  




Y E %


5

) ' 02%

 

 % !
"

 ) 
" %

) ' 0(%

 

"

Proof. By continuity we have . Since approximates , there exists with . Monotonicity of then implies . For the converse we rst show monotonicity. Suppose holds but is not below . By Proposition 2.2.10 there is and from our assumption we get such that . Since belongs to this is a contradiction. Now let be a directed subset of with as its join. Monotonicity implies . If the converse relation does not hold then we can again choose with and for some we have . Since approximates , some is above and we get contradicting our choice of . Finally, we cite a result which reduces the calculation of least xpoints to a basis. The point here is that a continuous function need not preserve compactness nor the order of approximation and so the sequence need not consist of basis elements.

A proof may be found in [Abr90b].

2.2.4 Comments on possible variations directed sets vs. -chains Let us start with the following observation. Proposition 2.2.13. If a dcpo has a countable basis then every directed subset of contains an -chain with the same supremum.

This raises the question whether one shouldnt build up the whole theory using chains. The basic denitions then read: An -ccpo is a poset in which every -chain has a supremum. A function is -continuous if it preserves joins of -chains. An element is -approximating if implies for some . An -ccpo is continuous if there is a countable subset such that every element is the -approximating it. Similarly for algebraicity. join of an -chain of elements from (This is the approach adopted in [Plo81], for example.) The main point about these denitions is the countability of the basis. It ensures that they are in complete harmony with our set-up, because we can show: Proposition 2.2.14. 1. Every continuous -ccpo is a continuous dcpo.


2. Every algebraic -ccpo is an algebraic dcpo.

22

"

) ' 0(%

) ' 0(%

"

) ' 0(%

3.

 )   

 

2.

1.

, , .

"

Proposition 2.2.12. If is a pointed -continuous domain with basis is a continuous map, then there exists an -chain basis elements such that the following conditions are satised:

   

and if of

 


 

 

               $     D                                 

 

P  P  P

'

" !

 

"

 

"

" 

  

 $ 3) !

3. Every -continuous map between continuous -ccpos is continuous. Proof. (1) Let be an enumeration of a basis for . We rst show that the continuous -ccpo is directed-complete, so let be a directed subset of . Let be the set of basis elements which are below some element of and, for simplicity, assume that . We construct an -chain in as follows: let be an element of which is above . Then let be the rst basis element not below . It must be below some and we set to be an upper bound of and in . We proceed by induction. It does not follow that the resulting chain is conal in but it is true that its supremum is also the supremum of , because both subsets of dominate the same set of basis elements. This construction also shows that -approximation is the same as approximation in a continuous -ccpo. The same basis may then be used to show that is a continuous domain. (The directedness of the sets follows as in Proposition 2.2.4(1).) (2) follows from the proof of (1), so it remains to show (3). Monotonicity of the function is implied in the denition of -continuity. Therefore a directed set is mapped onto a directed set in and also holds. Let be an -chain in with , as constructed in the proof of (1). Then we have . If we drop the crucial assumption about the countability of the basis then the two theories bifurcate and, in our opinion, the theory based on -chains becomes rather bizarre. To give just one illustration, observe that simple objects, such as powersets, may fail to be algebraic domains. There remains the question, however, whether in the realm of a mathematical theory of computation one should start with -chains. Arguments in favor of this approach point to pedagogy and foundations. The pedagogical aspect is somewhat weakened by the fact that even in a continuous -ccpo the sets happen to be directed. Glossing over this fact would tend to mislead the student. In our eyes, the right middle ground for a course on domain theory, then, would be to start with -chains and motivations from semantics and then at some point (probably where the ideal completion of a poset is discussed) to switch to directed sets as the more general concept. This suggestion is hardly original. It is in direct analogy with the way students are introduced to topological concepts. Turning to foundations, we feel that the necessity to choose chains where directed subsets are naturally available (such as in function spaces) and thus to rely on the Axiom of Choice without need, is a serious stain on this approach. To take foundational questions seriously implies a much deeper re-working of the theory: some pointers to the literature will be found in Section 8. We do not feel the need to say much about the use of chains of arbitrary cardinality. This adds nothing in strength (because of Proposition 2.1.15) but has all the disadvantages pointed out for -chains already. bases vs. intrinsic descriptions. The denition of a continuous domain given here differs from, and is in fact more complicated than the standard one (which we presented as Proposition 2.2.7(1)). We nevertheless preferred this approach to the concept of approximation for three reasons. Firstly, the standard denition does not allow the restriction of the size of continuous domains. In this respect not the cardinality of a do-

23

) ' 02%

 

  

"

                   

) ' 0(%

 
%

"

) ' 0(%

"

) ' 0(%

"

) ' 02%

) ' 02% 5

   
" %


"

main but the minimal cardinality of a basis is of interest. Secondly, we wanted to point out the strong analogy between algebraic and continuous domains. And, indeed, the proofs we have given so far for continuous domains specialize directly to the algebraic case if one replaces by throughout. Thus far at least, proofs for algebraic domains alone would not be any shorter. And, thirdly, we wanted to stress the idea of approximation by elements which are (for whatever reason) simpler than others. Such a notion of simplicity does often exist for continuous domains (such as rational vs. real numbers), even though its justication is not purely order-theoretical (see 8.1.1). algebraic vs. continuous. This brings up the question of why one bothers with continuous domains at all. There are two important reasons but they depend on denitions introduced later in this text. The rst is the simplication of the mathematical theory of domains stemming from the possibility of freely using retracts (see Theorem 3.1.4 below). The second is the observation that in algebraic domains two fundamental concepts of domain theory essentially coincide, namely, that of a Scott-open set and that of a compact saturated set. We nd it pedagogically advantageous to be able to distinguish between the two. continuous dcpo vs. continuous domain. It is presently common practice to start a paper in semantics or domain theory by dening the subclass of dcpos of interest and then assigning the name domain to these structures. We fully agree with this custom of using domain as a generic name. In this article, however, we will study a full range of possible denitions, the most general of which is that of a dcpo. We have nevertheless avoided calling these domains. For us, domain refers to both ideas essential to the theory, namely, the idea of convergence and the idea of approximation. 2.2.5 Useful properties Let us start right away with the single most important feature of the order of approximation, the interpolation property.

It is clearly non-empty. It is directed because if and then by the directedness of there is such that and and again by the directedness of there is with and . We calculate the supremum of : let be any element approximating . Since we have that . This holds for all so by continuity . All elements of are less than , so in fact equality holds: . Remember that we started out with a set whose elements approximate . By denition there is with for each . Let be an upper bound of the in . By denition, for some , , and we can take as an interpolating element 24

       

 



  



Proof. Given

in

we dene the set

"

Lemma 2.2.15. Let be a continuous domain and let whose elements approximates . Then there exists holds. If is a basis for then may be chosen from between and .)

be a nite set each of such that . (We say, interpolates

   

           %        ! 

 

"

   

"

 

Corollary 2.2.16. Let be a continuous domain with a basis and let be a directed subset of . If is an element approximating then already approximates some . As a formula:
"

Intersecting with the basis on both sides gives

"

Next we will illustrate how in a domain we can restrict attention to principal ideals.
6

2. The order of approximation on a continuous domain is the union of the orders of approximation on all principal ideals. 3. A dcpo is continuous if and only if each principal ideal is continuous.
' $

4. For a continuous domain

we have

5. A dcpo is algebraic if and only if each principal ideal is algebraic. Proposition 2.2.18. 1. In a continuous domain minimal upper bounds of nite sets of compact elements are again compact. 2. In a complete lattice the sets are -sub-semilattices.

3. In a complete lattice the join of nitely many compact elements is again compact. Corollary 2.2.19. A complete lattice is algebraic if and only if each element is the join of compact elements. The inmum of compact elements need not be compact again, even in an algebraic lattice. An example is given in Figure 6. 2.2.6 Bases as objects In Section 2.2.2 we have seen how we can use bases in order to express properties of the ambient domain. We will now study the question of how far we can reduce domain theory to a theory of (abstract) bases. The resulting techniques will prove useful in later chapters but we hope that they will also deepen the readers understanding of the nature of domains. We start with the question of what additional information is necessary in order to reconstruct a domain from one of its bases. Somewhat surprisingly, it is just the order of approximation. Thus we dene: 25

Proposition 2.2.17. 1. If is a continuous domain and if then approximates if and only if for all directed sets is an such that .

are elements in , with there

 P

     

      

"

'

'



 

 &

between enter .

 I 

and . The proof remains the same if we allow only basis elements to



Abstract bases were introduced in [Smy77] where they are called R-structures. Examples of abstract bases are concrete bases of continuous domains, of course, where the relation is the restriction of the order of approximation. Axiom (INT) is satised because of Lemma 2.2.15 and because we have required bases in domains to have directed sets of approximants for each element. Other examples are partially ordered sets, where (INT) is satised because of reexivity. We will shortly identify posets as being exactly the bases of compact elements of algebraic domains. In what follows we will use the terminology developed at the beginning of this chapter, even though the relation on an abstract basis need neither be reexive nor antisymmetric. This is convenient but in some instances looks more innocent than it is. An ideal in a basis, for example, has the property (following from directedness) that for every there is another element with . In posets this doesnt mean anything but here it becomes an important feature. Sometimes this is stressed by using the expression is a round ideal. Note that a set of the form is always an ideal because of (INT) but that it need not contain itself. We will refrain from calling principal in these circumstances. Denition 2.2.21. For a basis let be the set of all ideals ordered by inclusion. It is called the ideal completion of . Furthermore, let denote the function which maps to . If we want to stress the relation with which is equipped then we write for the ideal completion.
B C

Proposition 2.2.22. Let

be an abstract basis.

26

$  0 ! 

"

" 9



P   I   P  I

"

"

holds for all elements

and nite subsets

of

"

"

V " !

B C

"

" U@

P IU"@

(INT)

"

Denition 2.2.20. An (abstract) basis is given by a set relation on , such that

together with a transitive

Figure 6: The meet of the compact elements

    

and is not compact.

E `

Y `

YE

E E E

` Y `

YE

 



"

"



Proof. (1) holds because clearly the directed union of ideals is an ideal. Roundness can be written as . This union is directed implies that every because is directed. This proves (2) and also (3). The fourth claim follows from the characterization of the order of approximation. The last clause holds because there is only one basis of compact elements for an algebraic domain. Dening the product of two abstract bases as one does for partially ordered sets, we have the following:
" " " "

Proposition 2.2.23.

Our completion has a weak universal property:

B C

be an abstract basis and let be a dcpo. For every Proposition 2.2.24. Let monotone function there is a largest continuous function such that is below . It is given by .

Proof. Let us rst check continuity of . To this end let be a directed collection of ideals. Using general associativity (Proposition 2.1.4(3)) we can calculate: . Since is assumed to be monotone, is an upper bound for . This proves that is below . If, on the other hand, is another continuous function with this property then we have . The claim about the continuity of the assignment is shown by the usual switch of directed suprema. If is a preorder then we can show that : . 27
' ' ' $

"

'

"

"

       Q        !      '              

'

'

'

'

The assignment If the relation

is a Scott-continuous map from is reexive then equals .

      Q    $          $               8 $ 97B6  5 9@#$ 6 5 8

to

 $  ) !

"

 

  

'

 H  A  H

"

"

"

"

  0

"

5. If

is a poset then
V



"

4. If

is reexive then

is algebraic. , and

are all isomorphic.

"

"

3.

is a continuous domain and a basis of

is given by

 

"

  


V "

P IU"@

 $ ) !

"

"

B C P U@ "  "V   

2.

holds in .

if and only if there are

in

such that

   
.

"

1. The ideal completion of

is a dcpo.


V

'

'

A particular instance of this proposition is the case that and are two abstract bases and is monotone. By the extension of to we mean the map . It maps an ideal to the ideal .
B "

Proposition 2.2.25. Let be a continuous domain with basis an abstract basis, we have the following:
" "

. Viewing

for all elements, so . Proof. In a continuous domain we have Composing the maps the other way round we need to see that every which approximates , where is an ideal in , actually belongs to . We interpolate: and using the dening property of the order of approximation, we nd above . Therefore approximates and belongs to . The calculation for (2) is straightforward: . Corollary 2.2.26. A continuous function from a continuous domain completely determined by its behavior on a basis of . to a dcpo

As we now know how to reconstruct a continuous domain from its basis and how to recover a continuous function from its restriction to the basis, we may wonder whether it is possible to work with bases alone. There is one further problem to overcome, namely, the fact that continuous functions do not preserve the order of approximation. The only way out is to switch from functions to relations, where we relate a basis . This can be axiomatized as follows. element to all basis elements approximating
"

Denition 2.2.27. A relation between abstract bases imable if the following conditions are satised:
# # 6 "

and

is called approx-

The following is then proved without difculties.

Theorem 2.2.28. The category of abstract bases and approximable relations is equivalent to CONT, the category of continuous dcpos and continuous maps. The formulations we have chosen in this section allow us immediately to read off the corresponding results in the special case of algebraic domains. In particular: Theorem 2.2.29. The category of preorders and approximable relations is equivalent to ALG, the category of algebraic dcpos and continuous maps. 28

W #

" 

"

4.

 

"

3.

 

"

2.

and

Q 

1.

"

      

"

Q     9 0  

 

2. For every dcpo and continuous function is the restriction of to .

"

we have

where

is

1.

"

is isomorphic to . The isomorphism of the embedding of into . Its inverse maps elements
"

is the extension to .

IU"@ P

! $ 3)   '  &$  !

  
" " V

"

"

"

 

" @

 9 0  4      # I  P    I 3      P   3 # B  

 
#

       

!  $ $  " )  !

"

"

"

 

"

"

as

 7

"



2.3 Topology
By a topology on a space we understand a system of subsets of (called the open sets), which is closed under nite intersections and innite unions. It is an amazing fact that by a suitable choice of a topology we can encode all information about convergence, approximation, continuity of functions, and even points of themselves. To a student of Mathematics this appears to be an immense abstraction from the intuitive beginnings of analysis. In domain theory we are in the lucky situation that we can tie up open sets with the concrete idea of observable properties. This has been done in detail earlier in this handbook, [Smy92], and we may therefore proceed swiftly to the mathematical side of the subject. 2.3.1 The Scott-topology on a dcpo Denition 2.3.1. Let be a dcpo. A subset is called (Scott-)closed if it is a lower set and is closed under suprema of directed subsets. Complements of closed sets are called (Scott-)open; they are the elements of , the Scott-topology on . for the smallest closed set containing . Similarly, We shall use the notation will stand for the open kernel of . A Scott-open set is necessarily an upper set. By contraposition it is characterized by the property that every directed set whose supremum lies in has a non-empty intersection with . Basic examples of closed sets are principal ideals. This knowledge is enough to show the following: Proposition 2.3.2. Let
6

be a dcpo. the following are equivalent:

1. For elements


(a) (c)

,
V

.
5

2. The Scott-topology satises the is trivial.

separation axiom.

Thus we can reconstruct the order between elements of a dcpo from the Scotttopology. The same is true for limits of directed sets. Proposition 2.3.3. Let be a directed set in a dcpo . Then is the supremum of if and only if it is an upper bound for and every Scott-neighborhood of contains an element of .

Proof. Indeed, the closed set bounds of .

separates the supremum from all other upper

29

 ' 

3.

is a Hausdorff (

) topological space if and only if the order on

(b) Every Scott-open set which contains

also contains ,

 I P  

   

Proposition 2.3.4. For dcpos and , a function from to is Scott-continuous if and only if it is topologically continuous with respect to the Scott-topologies on and . Proof. Let be a continuous function from to and let be an open subset of . It is clear that is an upper set because continuous functions are monotone. If maps the element into then we have and by denition there must be some which is mapped into . Hence is open in . For the converse assume that is topologically continuous. We rst show that must be monotone: Let be elements of . The inverse image of the Scottclosed set contains . Hence it also contains . Now let be directed. Look at the inverse image of the Scott-closed set . It contains and is Scott-closed, too. So it must also contain . Since by monotonicity is an upper bound of , it follows that is the supremum of . So much for the theme of convergence. Let us now proceed to see in how far approximation is reected in the Scott-topology. 2.3.2 The Scott-topology on domains In this subsection we work with the second-most primitive form of open sets, namely those which can be written as . We start by characterizing the order of approximation. Proposition 2.3.5. Let for all pairs :

be a continuous domain. Then the following are equivalent

1.

Comment: Of course, (1) is equivalent to (3) in all dcpos.

Proposition 2.3.6. Let be a continuous domain with basis . Then openness of a subset of can be characterized in the following two ways:
' $

1. 2.

This can be read as saying that every open set is supported by its members from the basis. We may therefore ask how the Scott-topology is derived from an abstract basis.

30

"

"

"

"

Proposition 2.3.7. Let the set are of this form.

" V

be an abstract basis and let be any subset of . Then is Scott-open in and all open sets on

"

  P    

      

'

3.

2.

    
   
'

   

'

    

     )      

 

  

'

           P

 

'

This, nally, nicely connects the theory up with the idea of an observable property. If we assume that the elements of an abstract basis are nitely describable and nitely recognisable (and we strongly approve of this intuition) then it is clear how to observe a property in the completion: we have to wait until we see an element from a given set of basis elements. We also have the following sharpening of Proposition 2.3.6: Lemma 2.3.8. Every Scott-open set in a continuous domain is a union of Scott-open lters. Proof. Let be an element in the open set . By Proposition 2.3.6 there is an element which approximates . We repeatedly interpolate between and . This gives us a sequence . The union of all is a Scott-open lter containing and contained in . In this subsection we have laid the groundwork for a formulation of Domain Theory purely in terms of the lattice of Scott-open sets. Since we construe open sets as properties we have also brought logic into the picture. This relationship will be looked at more closely in Chapter 7. There and in Section 4.2.3 we will also exhibit more properties of the Scott-topology on domains. Exercises 2.3.9. 1. Formalize the passage from preorders to their quotient posets.
%

2. Draw line diagrams of the powersets of a one, two, three, and four element set. 3. Show that a poset which has all suprema also has all inma, and vice versa. 4. Rene Proposition 2.1.7 by showing that the xpoints of a monotone function on a complete lattice form a complete lattice. Is it a sublattice? 5. Show that nite directed sets have a largest element. Characterize the class of posets in which this is true for every directed set. 6. Show that the directed set of nite subsets of real numbers does not contain a conal chain.

8. Let be a monotone map between complete lattices subset of . Prove: .

and

and let

9. Show that the category of posets and monotone functions forms a cartesian closed category. 10. Draw the line diagram for the function space of the at booleans (see Figure 1). 11. Show that an ideal in a (binary) product of posets can always be seen as the product of two ideals from the individual posets.

31

 

7. Which of the following are dcpos: integers)?

(unit interval),

(negative be a



 1 8P 5


% 

 

  '

12. Show that a map between two dcpos all directed sets in , need to be required explicitly).

13. Give an example of a monotone map on a pointed dcpo for which is not a xpoint. (Some xpoint must exist by Proposition 2.1.16.) 14. Use xpoint induction to prove the following. Let functions on a pointed dcpo with , and .
F F

15. (Dinaturality of xpoints) Let be pointed dcpos and let be continuous functions. Prove

17. Prove: Given pointed dcpos and and a continuous function there is a continuous function such that holds. (This is the general denition of a category having xpoints.) How does Theorem 2.1.19 follow from this? 18. Show that each version of the natural numbers as shown in Figure 2 is an example of a countable dcpo whose function space is uncountable. 19. Characterize the order of approximation on the unit interval. What are the compact elements? 20. Show that in nite posets every element is compact. 21. Let be the lattice of open sets of , where is equipped with the ordinary metric topology. Show that no two non-empty open sets approximate each other. Conclude that is not continuous.

23. Extend Proposition 2.2.10 in the following way: For every nite subset of a continuous dcpo with basis there exists , such that is an order-isomorphism between and and such that for all , the element belongs to . 24. Prove Proposition 2.2.17. 25. Show that elements of an abstract basis, which approximate no other element, may be deleted without changing the ideal completion. 26. Show that if is a non-compact element of a basis for a continuous domain then is still a basis. (Hint: Use the interpolation property.)
"

32



22. Prove Proposition 2.2.10.

$ % ! I  3) H

 $  

" 

$  !  

16. Show that Lemma 2.1.21 uniquely characterizes

among all xpoint operators.

$ ) !

  #! #P  $ 

be continuous . Then

 

  )  

and is continuous if and only if for holds (i.e., monotonicity does not

P 
"

"

 
  "
B

 ) 
F

 @$ ! P

 P



) ' 0(% "

27. The preceding exercise shows that different bases can generate the same domain. Show that for a xed basis different orders of approximation may also yield the same domain. Show that this will denitely be the case if the two orders and satisfy the equations and .

29. Let be a relation on a set such that holds. Give an example showing that Axiom (INT) (Denition 2.2.20) need not be satised. Nevertheless, is a continuous domain. What is the advantage of our axiomatization over this simpler concept? 30. Spell out the proof of Theorem 2.2.28. 31. Prove that in a dcpo every upper set is the intersection of its Scottneighborhoods. 32. Show that in order to construct the Scott-closure of a lower set of a continuous domain it is sufcient to add all suprema of directed subsets to . Give an example of a non-continuous dcpo where this fails. 33. Given a subset in a dcpo let be the smallest superset of which is closed against the formation of suprema of directed subsets. Show that the cardinality can be no greater than . (Hint: Construct a directed suprema closed of superset of by adding all existing suprema to .)

33

28. What is the ideal completion of

"

 P
"

3 Domains collectively
3.1 Comparing domains
3.1.1 Retractions A reader with some background in universal algebra may already have missed a discussion of sub-dcpos and quotient-dcpos. The reason for this omission is quite simple: there is no fully satisfactory notion of sub-object or quotient in domain theory based on general Scott-continuous functions. And this is because the formation of directed suprema is a partial operation of unbounded arity. We therefore cannot hope to be able to employ the tools of universal algebra. But if we combine the ideas of sub-object and quotient then the picture looks quite nice. Denition 3.1.1. Let and be posets. A pair , of monotone functions is called a monotone section retraction pair if is the identity on . In this situation we will call a monotone retract of . If and are dcpos and if both functions are continuous then we speak of a continuous section retraction pair. We will omit the qualifying adjective monotone, respectively continuous, if the properties of the functions are clear from the context. We will also use s-r-pair as a shorthand. One sees immediately that in an s-r-pair the retraction is surjective and the section is injective, so our intuition about being both a sub-object and a quotient of is justied. In such a situation inherits many properties from : Proposition 3.1.2. Let and tone section retraction pair. 1. Let be any subset of in . It is given by 2. If

be posets and let . If

be a mono-

has a supremum in then . Similarly for the inmum.

has a supremum

is a (pointed) dcpo, a semilattice, a lattice or a complete lattice then so is .

Proof. Because of and the monotonicity of it is clear that is an upper bound for . Let be another such. Then by the monotonicity of we have that is an upper bound of and hence it is above . So we get . The property of being a (pointed) dcpo, semilattice, etc., is dened through the existence of suprema or inma of certain subsets. The shape of these subsets is preserved by monotone functions and so (2) follows from (1). Let us now turn to continuous section retraction pairs. Lemma 3.1.3. Let be a continuous section retraction pair between dcpos and and let be a basis for . Then is a basis for .

34



$ !

$ ! $ !

$ !

"

  

Q   Q   

 P

"

Proof. Let be an approximant to for . We show that approximates . To this end let be a directed subset of with . By the continuity of we have and so for some , must hold. This implies . The continuity of gives us that is the supremum of . Theorem 3.1.4. A retract of a continuous domain via a continuous s-r-pair is continuous. The analogous statement for algebraic domains does not hold in general. Instead of constructing a particular counterexample, we use our knowledge about the ideal completion to get a general, positive result which implies this negative one. Theorem 3.1.5. Every ( -) continuous domain is the retract of an ( -) algebraic domain via a continuous s-r-pair. In more detail, we have:
"

Proof. The continuity of follows from general associativity, Proposition 2.1.4, and are directed unions. For the continuity of we the fact that directed suprema in use the interpolation property in the form of Proposition 2.2.16(2). 3.1.2 Idempotents Often the section part of an s-r-pair is really a subset inclusion. In this case we can hide it and work with the map on alone. It is idempotent, because . Proposition 3.1.7. 1. The image of a continuous idempotent map on a dcpo is a dcpo. The suprema of directed subsets of , calculated in , coincide with those calculated in . The inclusion is Scott-continuous. 2. The set of all continuous idempotent functions on a dcpo is again a dcpo. Proof. (1) The rst part follows from Proposition 3.1.2 because the inclusion is surely monotone. For the second part let be a directed set contained in . We need to see that belongs to again. This holds because is continuous: . (2) Let be a directed family of continuous idempotents. For any we

35

 

 ''

 %

 %

P  $  P 

Proposition 3.1.6. Let

A " V

be a continuous domain with basis and tinuous section retraction pair between and .

. Then the maps constitute a con-

       Q       I           

 

 7$

  %

"

 $  I  !   $ QP  P $ B! P 

"

"

 %

"

      

'

"

 

"

Hence the supremum of continuous idempotents is again an idempotent function. We have proved in Proposition 2.1.18 that it is also continuous. If is a continuous idempotent map on a continuous domain then we know that its image is again continuous. But it is not true that the order of approximation is the restriction of the order of approximation on . For example, every on constant map is continuous and idempotent. Its image is an algebraic domain with one element, which is therefore compact. But surely not every element of a continuous domain is compact. However, we can say something nice about the Scott-topology on the image: Proposition 3.1.8. If is a continuous idempotent function on a dcpo topology on is the restriction of the Scott-topology on to

Proof. This follows immediately because a continuous idempotent function rise to a continuous s-r-pair between and . Useful examples of idempotent self-maps are retractions They are given by if otherwise

Their continuity follows from the fact that is always Scott-closed. Dually, we can dene a retraction onto a principal lter . It is Scott-continuous if (but not only if) its generator is compact. 3.1.3 Adjunctions An easy way to avoid writing this subsection would be to refer to category theory and to translate the general theory of adjoint functors into the poset setting. However, we feel that the right way to get used to the idea of adjointness is to start out with a relatively simple situation such as is presented by domain theory. (In fact, we will use adjoint functors later on, but really in a descriptive fashion only.) Let us start with the example of a surjective map from a poset onto a poset . It is natural to ask whether there is a one-sided inverse for , i.e. a map such that holds. Figure 7 illustrates this situation. Such a map must

36

  % 

onto principal ideals.

$ !

      (   Q  (    (     (    

G QH

'

 

'

'

'

'

'

  %



(    (      PB P  

can calculate

'

G 8Q H

'

 %

 %

then the Scott. gives

Figure 7: The right inverse problem for a surjective function pick out a representative from for each . Set-theoretically this can be done, but the point here is that we want to be monotone. If we succeed then and form a (monotone) section retraction pair. Even nicer would it be if we could pick out a canonical representative from , which in the realm of order theory means that we want to have a least (or largest) element. If this is the case then how can we ensure that the assignment is monotone? The solution is suggested by the observation that if is monotone then is not only the least but also of . This condition is also sufcient. The switch element of from to (and this is a trick to remember) may allow us to construct a partial right inverse even if is not surjective. Thus we arrive at a rst, tentative denition of an adjunction.

This denition is simple and easy to motivate. But it brings out just one aspect of adjoint pairs, namely, that is uniquely determined by . There is much more:

Proposition 3.1.10. Let and be posets and tone functions. Then the following are equivalent:

and

(2) Pick an element . Then because we have from (1) Proof. (1) that holds. So belongs to . Now let be any element of , or, equivalently, . Using (1) again, we see that this can only happen if holds. So is indeed the largest element of . The converse is proved analogously, of course. (1) and (2) together immediately give both (3) and (4). From (3) we get (4) by applying the monotone map to the inequality and using . 37

(For (4)

(1) the monotonicity of



and is not needed.)

 

4.

 



3.

and

 

2.

1.

Denition 3.1.9. (preliminary) Let and be posets and let be monotone functions. We say that is an adjunction between every we have that is the least element of .

and and

be mono-

$ )!

$ !

$ ) !

 


 
 

                            Q  %   Q   %    

      % $   3!    

  

$ !

 

 P

 

 

      

 




if for

Assuming (4) we see immediately that is a lower bound for . But because and hence we have that also belongs to . We get the monotonicity of as follows: If holds in then because we have and by transitivity . Using (4) again, we get . We conclude that despite the lopsided denition, the situation described by an adjunction is completely symmetric. And indeed, adjunctions are usually introduced using either (3) or (4). Denition 3.1.11. (ofcial) Let and be posets and let and be functions. We say that is an adjunction between and if for all and we have . We call the lower and the upper adjoint and write .

2. The image of and the image of are order-isomorphic. The isomorphisms are given by the restrictions of and to and , respectively.

Proof. (1) We use Proposition 3.1.10(3) twice: . (2) The equations from (1) say precisely that on the images of and , and , respectively, act like identity functions. (3) If is surjective then we can cancel on the right in the equation and get . From this it follows that must be injective. (5) Let for . By monotonicity, for each . Conversely, let be any upper bound of . Then is an upper bound for each which in turn is above . So holds and this is equivalent to . The last property in the preceding proposition may be used to dene an adjunction in yet another way, the only prerequisite being that there are enough sets with an inmum (or supremum). This is the Adjoint Functor Theorem for posets. Proposition 3.1.13. Let be a monotone function from a complete lattice to a poset. Then the following are equivalent: 1. 2. preserves all inma, has a lower adjoint. preserves all suprema if and only if 38 has an upper adjoint.

And similarly:

 

 

 

 

  

$ !

   

5.

preserves existing suprema,

preserves existing inma.

4.

is surjective

3.

is surjective

is injective,

is injective,

 %  %

1.

and

Proposition 3.1.12. Let

be an adjunction between posets.

       



) $ ' !

$ !

          Q



   P 

  Q         

!

! 

 &

Proof. We already know how to dene a candidate for a lower adjoint ; we try . All that remains, is to show that belongs to . This follows because preserves meets: . This proposition gives us a way of recognizing an adjoint situation in cases where only one function is explicitly given. It is then useful to have a notation for the missing mapping. We write for the upper and for the lower adjoint of . Now it is high time to come back to domains and see what all this means in our setting.

3. If

is continuous then the converse of (2) is also true.

Proof. Continuity of the lower adjoint follows from Proposition 3.1.12(5). So let be elements in and let be a directed subset of such that holds. This implies and from the continuity of we deduce . Hence some is above which, going back to , means . (3) For the converse let be any directed subset of . Monotonicity of yields . In order to show that the other inequality also holds, we prove that is above every approximant to . Indeed, if we have by assumption. So some is above and for this we have . 3.1.4 Projections and sub-domains Let us now combine the ideas of Section 3.1.1 and 3.1.3. Denition 3.1.15. Let and be dcpos and let and be continuous functions. We say that is a continuous embedding projection pair (or e-p-pair) if and . We note that the section retraction pair between a continuous domain and its ideal completion as constructed in Section 3.1.1 is really an embedding projection pair. From the general theory of adjunctions and retractions we already know quite a bit about e-p-pairs. The embedding is injective, is surjective, preserves existing suprema and the order of approximation, preserves existing inma, is continuous if is continuous, and, nally, embeddings and projections uniquely determine each other. Because of this last property the term embedding has a well-dened meaning; it is an injective function which has a Scott-continuous upper adjoint. An injective lower adjoint also reects the order of approximation:

39

Proposition 3.1.16. Let be elements of . Then

be an e-p-pair between dcpos and let and holds in if and only if approximates in .

                

 

 0 ! $

$ !

 P 

!   3!    

  Q                      

2. If

1.

is Scott-continuous.

is Scott-continuous then preserves the order of approximation.

Proposition 3.1.14. Let

               
be an adjunction between dcpos.

  

! 3! 

  

Let us also look at the associated idempotent on . As it is below the identity, it makes good sense to call such a function a kernel operator, but often such maps are just called projections. We denote the set of kernel operators on a dcpo by . It is important to note that while a kernel operator preserves inma as a map from to its image, it does not have any preservation properties as a map from to besides Scott-continuity. What we can say is summarized in the following proposition. Proposition 3.1.17. Let 1.

be a dcpo.

is a dcpo.

3. The image of a kernel operator is closed under existing suprema.

Proof. (1) is proved as Proposition 3.1.7 and (2) follows because together with the inclusion of into form an adjunction. This also shows (4). Finally, (3) and (5) are direct consequences of (2). In the introduction we explained the idea that the order on a semantic domain models the relation of some elements being better than others, whereat least in semanticsbetter may be replaced more precisely by better termination. Thus we view elements at the bottom of a domain as being less desirable than those higher up; they are proto-elements from which fully developed elements evolve as we go up in places at the botthe order. Now, the embedding part of an e-p-pair tom of . Following the above line of thought, we may think of as being a collection of proto-elements from which the elements of evolve. Because there is the projection part as well, every element of exists in some primitive form in already. Also, contains some information about the order and the order of approximation on . We may therefore think of as a preliminary version of , as an approximation to on the domain level. This thought is made fruitful in Sections 4.2 and 5. Although the word does not convey the whole fabric of ideas, we name a sub-domain of , just in case there is an e-p-pair . 3.1.5 Closures and quotient domains The sub-domain relation is preeminent in domain theory but, of course, we can also combine retractions and adjunctions the other way around. Thus we arrive at continuous insertion closure pairs (i-c-pairs). Because adjunctions are not symmetric as far as the order of approximation is concerned, Proposition 3.1.14, the situation is not just the order dual of that of the previous subsection. We know that the insertion preserves existing inma and so on, but in addition we now have that the surjective part preserves the order of approximation and therefore, is algebraic if is.

40

!

 4! 

5. For kernel operators

on

we have

if and only if

 %   %

! 3!  

4.

. .

2. If

is a kernel operator on .

then for all

we have that

  8   $ 97B6  5   I
T

 



  %  P F Q % H  %     "   %  8  $ 97#6  5

The associated idempotent is called a closure operator. For closure operators the same caveat applies as for kernel operators; they need not preserve suprema. Worse, such functions do no longer automatically have a Scott-continuous (upper) adjoint. This is the price we have to pay for the algebraicity of the image. Let us formulate this precisely. Proposition 3.1.18. Let be an algebraic domain and let be a monotone idempotent function above . Then is again an algebraic domain if and only if it is closed under directed suprema. The reader will no doubt recognize this statement as being a reformulation and generalization of our example of inductive closure systems from Chapter 2, Proposition 2.2.9. It is only consequent to call a quotient domain of the continuous domain if there exists an i-c-pair .

3.2 Finitary constructions

In this section we will present a few basic ways of putting domains together so as to build up complicated structures from simple ones. There are three aspects of these constructions which we are interested in. The rst one is simply the order-theoretic denition and the proof that we stay within dcpos and Scott-continuous functions. The second one is the question how the construction can be described in terms of bases and whether the principle of approximation can be retained. The third one, nally, is the question of what universal property the construction has. This is the categorical viewpoint. Since this Handbook contains a chapter on category theory, [Poi92] (in particular, Chapter 2), we need not repeat here the arguments for why this is a fruitful and enlightening way of looking at these type constructors. There are, however, several categories that we are interested in and a construction may play different roles in different settings. Let us therefore list the categories that, at this point, seem suitable as a universe of discourse. There is, rst of all, DCPO, the category of dcpos and Scott-continuous functions as introduced in Section 2.1. We can restrict the objects by taking only continuous or, more special, algebraic domains. Thus we arrive at the full subcategories CONT and ALG of DCPO. Each of these may be further restricted by requiring the objects to have a bottom element (and Theorem 2.1.19 tells us why one would be interested in doing so) resulting in the categories DCPO , CONT , and ALG . The presence of a distinguished point in each object suggests that morphisms should preserve them. But this is not really appropriate in semantics; strict functions are tied to a particular evaluation strategy. For us this means that there is yet another cascade of categories, DCPO , CONT , and ALG , where objects have bottom elements and morphisms are strict and Scott-continuous. Finally, we may bound the size of (minimal) bases for continuous and algebraic domains to be countable. We indicate this by the prex -.

41

  @$ 3!

 %

 !  ! 

3.2.1 Cartesian product Denition 3.2.1. The cartesian product of two dcpos ing data:
6 B @ (

This is just the usual product of sets, augmented by the coordinatewise order. Through induction, we can dene the cartesian product for nite non-empty collections of dcpos. For the product over the empty index set we dene the result to be a xed one-element dcpo . Proposition 3.2.2. The cartesian product of dcpos is a dcpo. Suprema and inma are calculated coordinatewise. With each product

there are associated two projections:

and

These projections are always surjective but they are upper adjoints only if and are pointed. So there is a slight mismatch with Section 3.1.4 here. Given a dcpo and and , we denote the mediating morphism continuous functions from to by . It maps to . Proposition 3.2.3. Projections and mediating morphisms are continuous. are Scott-continuous, then so is the mediating map . The common notation for it is . Since our construction is completely explicit, we have thus dened a functor in two variables on DCPO. Proposition 3.2.4. Let
B 6 @ B

and

be dcpos. in if and only if approximates

3. 4.

is continuous if and only if .

and

are.

The categorical aspect of the cartesian product is quite pleasing; it is a categorical product in each case. But we can say even more. Lemma 3.2.5. Let C be a full subcategory of DCPO or DCPO products. Then these are isomorphic to the cartesian product.

42

"

"

"

"

2. If

and .

are bases for

and

, respectively, then

1. A tuple approximates a tuple in and approximates in .

@

which has nite

 H

If

and

is a basis for

C 

 $ H 4!

  P   $ 4! '  $ 4#!PC CC 

H  H      H  H    P  H P  

if and only if

and


and

is given by the follow-

 P   P  P  IP    P   H  
@

 7$ H  !

H  $ $ ! H #!  $ 3 )P! 

% H

2 B B


% C H
B

In a restricted setting this was rst observed in [Smy83a]. The general proof may be found in [Jun89]. A useful property which does not follow from general categorical or topological considerations, is the following. Lemma 3.2.6. A function in each variable separately.

Proof. Assume is separately continuous. Then is monotone, because given we can ll in and use coordinatewise monotonicity twice. The same works for continuity: if is directed then
6 6

This proves the interesting direction. 3.2.2 Function space

We have introduced the function space in Section 2.1.6 already. It consists of all continuous functions between two dcpos ordered pointwise. We know that this is again a dcpo. The rst morphism which is connected with this construction is , . It is continuous because it is continuous in each argument separately: in the rst because directed suprema of functions are calculated pointwise, in the second, because contains only continuous functions. The second standard morphism is the operation which rearranges a function of two arguments into a combination of two unary functions. That is, if maps to , then is the mapping which assigns to the function which assigns to the element . is a continuous function because of Lemma 3.2.6. And for completely general reasons we have that itself is a continuous operation from to . Another derived operation is composition which is a continuous operation from to . All this shows that the continuous function space is the exponential in DCPO. Taking cartesian products and function spaces together we have shown that DCPO is cartesian closed. We turn the function space construction into a functor from DCPO DCPO to DCPO by setting , where , and is an element of .

8 7B6  5 C $ 8 C7#6 $ 6 5 $ 5 H8 #88$ %C7#6 5 $#% 5 8 7#6 % 5 6 C $ H P    4 8 7#6 $  4!  C $ 5 

43

 C H  ( 

$ H GF! E  $ 4!

H PH I

         P            P        &   P     P      H%    P  C 7 P H  B) P  $  ! C @$ H  ) !

is continuous if and only if it is continuous

8 46  5 $

'

'

'

H PH I

$6 P   $ # $  H 8 #% 5 !

'

' '

  #P8 #$%64$6  5 5 8

B 9

'

H PH I

Figure 8: The coalesced sum of two pointed dcpos. As for the product we can show that the choice of the exponential object is more or less forced upon us. This again was rst noticed by Smyth in the above mentioned reference. Lemma 3.2.7. Let C be a cartesian closed full subcategory of DCPO. The exponential of two objects and of C is isomorphic to . Let us now turn to the theme of approximation in function spaces. The reader should brace himself for a profound disappointment: Even for algebraic domains it may be the case that the order of approximation on the function space is empty! (Exercise 3.3.12(11) discusses an example.) This fact together with Lemmas 3.2.5 and 3.2.7 implies that neither CONT nor ALG are cartesian closed. The only way to remedy this situation is to move to more restricted classes of domains. This will be the topic of Chapter 4. 3.2.3 Coalesced sum In the category of sets the coproduct is given by disjoint union. This works equally well for dcpos and there isnt really anything interesting to prove about it. We denote it by . Disjoint union, however, destroys the property of having a least element and this in turn is indispensable in proving that every function has a xpoint, Theorem 2.1.19. One therefore looks for substitutes for disjoint union which retain pointedness, but, of course, one cannot expect a clean categorical characterization such as for cartesian product or function space. (See also Exercise 3.3.12(12).) In fact, it has been shown in [HP90] that we cannot have cartesian closure, the xpoint property and coproducts in a non-degenerate category. So let us now restrict attention to pointed dcpos. One way of putting a family of them together is to identify their bottom elements. This is called the coalesced sum and denoted . Figure 8 illustrates this operation. Elements from different from carry a label which indicates where they came from. We write them in the form , . Proposition 3.2.8. The coalesced sum of pointed dcpos is a pointed dcpo.

44

8 4% 5 $6

 P       " !


 

The individual dcpos may be injected into the sum in the obvious way:

A universal property for the sum holds only in the realm of strict functions: Proposition 3.2.9. The coalesced sum of pointed dcpos is the coproduct in DCPO , CONT , and ALG . Once we accept the restriction to bottom preserving functions it is clear how to turn the coalesced sum into a functor. 3.2.4 Smash product and strict function space It is clear that inside DCPO a candidate for the exponential is not the full function space but rather the set of strict continuous functions from to . However, it does not harmonize with the product in DCPO , which, as we have seen, must be the cartesian product. We do get a match if we consider the so-called smash product. It is dened like the cartesian product but all tuples which contain at least one bottom element are identied. Common notation is . We leave it to the reader to check that smash product and strict function space turn DCPO into a monoidal closed category. 3.2.5 Lifting Set-theoretically, lifting is the simple operation of adding a new bottom element to a dcpo. Applied to , the resulting structure is denoted by . Clearly, continuity or algebraicity dont suffer any harm from this. Associated with it is the map which maps each to its namesake in . The categorical signicance of lifting stems from the fact that it is left adjoint to the inclusion functor from DCPO into DCPO. (Where a morphism is lifted by mapping the new bottom element of to the new bottom element of .) 3.2.6 Summary

For quick reference let us compile a table of the constructions looked at so far. A indicates that the category is closed under the respective construction, a says that, in addition, the construction plays the expected categorical role as a product, exponential or coproduct, respectively. Observe that for the constructions considered in this section it makes no difference if we restrict the size of a (minimal) basis.

45

$ #! 

 (

  &$ #!

 

and

 P  $  P "  !   P !   P  " 
$


8 B6  $ 5

3.3 Innitary constructions

The product and sum constructions from the previous section have innitary counterparts. Generally, these work nicely as long as we are only concerned with questions of convergence, but they cause problems with respect to the order of approximation. This is exemplied by the fact that an innite power of a nite poset may fail to be algebraic. In any case, there is not much use of these operations in semantics. Much more interesting is the idea of incrementally building up a domain in a limit process. This is the topic of this section. 3.3.1 Limits and colimits Our limit constructions are to be understood categorically and hence we refer once more to [Poi92] for motivation and general results. Here are the basic denitions. A diagram in a category C is given by a functor from a small category I to C. We can describe, somewhat sloppily but more concretely, a diagram by a pair of a family of objects and a family of connecting morphisms. The shape of the diagram is thus encoded in the index sets (which correspond to the objects of I) and (which correspond to the morphisms of I) and in the maps which corresponds to the and map on I. What is lost is the information about composition in I. In the cases which interest us, this is not a problem. A cone over such a diagram is given by an object and a family of morphisms such that for all we have . A cone is limiting if for every other cone there is exactly one morphism such that for all , . If is a limiting cone, then is called limit and the are called limiting morphisms. The dual notions are cocone, colimit, and colimiting morphism. Theorem 3.3.1. DCPO has limits of arbitrary diagrams. Proof. The proof follows general category theoretic principles. We describe the limit of the diagram as a set of particular elements of the product of all s, the so-called commuting tuples.

46

(

 (

P )    ) $ !  '  P    7$ 3!   (  %  %  $ ! P      7$  (   ! ( P  

(   (  (      !           7$  (   ! !( P  

DCPO

DCPO

DCPO

CONT ALG

CONT ALG

CONT ALG

'

'

' (  (

'

' (  (

'

@ W

'

 8 #6  $  5 8 #% 5 H $ 6
@

'

Figure 9: An expanding sequence of nite domains. The order on the limit object is inherited from the product, that is, tuples are ordered coordinatewise. It is again a dcpo because the coordinatewise supremum of commuting tuples is commuting as all are Scott-continuous. This also proves that the projections restricted to are continuous. They give us the maps needed to complement to a cone. , we dene the mediating morphism Given any other cone by . Again, it is obvious that this is well-dened and continuous, and that it is the only possible choice. We also have the dual:
B

Theorem 3.3.2. DCPO has colimits of arbitrary diagrams. This was rst noted in [Mar77] and, for a somewhat different setting, in [Mes77]. The simplest way to prove it is by reducing it to completeness a la Theorem 23.14 ` of [HS73]. This appears in [LS81]. A more detailed analysis of colimits appears in [Fie92]. There the problem of retaining algebraicity is also addressed. Theorem 3.3.3. DCPO is cartesian closed, complete and cocomplete. Theorem 3.3.4. DCPO

is monoidal closed, complete and cocomplete.

How about DCPO , where objects have least elements but morphisms need not preserve them? The truth is that both completeness and cocompleteness fail for this category. On the other hand, it is the right setting for denotational semantics in most cases. As a result of this mismatch, we quite often must resort to detailed proofs on the element level and cannot simply apply general category theoretic principles. Let us nevertheless write down the one good property of DCPO : Theorem 3.3.5. DCPO is cartesian closed. 3.3.2 The limit-colimit coincidence The theorems of the previous subsection fall apart completely if we pass to domains, that is, to CONT or ALG. To get better results for limits and colimits we must restrict both the shape of the diagrams and the connecting morphisms used.

    
47
E E

E E

'

  !    $ !  @$ !  P   $   ! ( 

YE

(
@

E ' E

For motivation let us look at a chain of domains where each is a sub-domain of in the sense of Section 3.1.4. Taking up again the animated language from that section we may think of the points of as growing out of points of , the latter being the buds which contain the leaves and owers to be seen at later stages. Figure 9 shows a, hopefully inspiring, example. Intuition suggests that in such a situation a well-structured limit can be found by adding limit points to the union of the , and that it will be algebraic if the are. Denition 3.3.6. A diagram in the category DCPO is called an expanding sequence, if the following conditions are satised: .)
% % B 0' )
 

Note that because lower adjoints determine upper adjoints and vice versa, we have whenever . It turns out that, in contrast to the general situation, the colimit of an expanding sequence can be calculated easily via the associated projections.
B 0' )

Then

holds.

to

to

Proof. We have already shown in Theorem 3.3.1 that a limit of the diagram is given by and that the mediating morphism has the (rst) postulated form. For the rest, let us start by showing that the functions are well-dened, i.e. that is a commuting tuple. Assume . Then we have

48

&         "

% %

 % %  

) ' 0(%

&  %

 " 

 P 

) ' 02%

  % &

        P  

by

B 0' )

) ' 02%

B 0(% ) ' B 0(% ) '

3.

is a colimit of the diagram is another cocone, then the mediating map from or .

) ' 0(%

by

B 0' )

) ' 0(%

B 0(% ) ' B 0(% ) '

2.

is a limit of the diagram is another cone, then the mediating map from or .

. If is given . If is given

) 0'

 

1. The maps

, form embedding projection pairs and

)    )    !    PP     P      !      P    P    P     P  &    P  !       $ QP  7$ 3!  P    P  P  $   !   P  7$ 3! !   

 

 

) ' 0(%

) ' 0(%

Theorem 3.3.7. Let in DCPO. Dene

% @ (

 

3.

% %

2.

be an expanding sequence

1. Each denote by
%

is an embedding. (The associated projection

 7$ 4!  P   

 7$ 3! P      I   $  !

P QP 
%

) ' 0(%

 

we

% %

. The assignment is Scott-continuous because of general associativity and because only Scott-continuous maps are involved in the denition. Next, let us now check that and form an e-p-pair.

This proves that the , , add up to the identity, as stated in (1). Putting this to use, we easily get the second representation for the mediating map into viewed as . a limit: It remains to prove the universal property of as a colimit. To this end let be a cocone over the expanding sequence. We have to check that is well-dened in the sense that the supremum is over a directed set. So let . We get . It commutes with the colimiting maps because

We also have to show that there is no other choice for . Again the equation in (1) comes in handy: Let be any mediating morphism. It must satisfy and so . Forming the supremum on both sides gives which is the denition of . This fact, that the colimit of an expanding sequence is canonically isomorphic to the limit of the associated dual diagram, is called the limit-colimit coincidence. It is 49

) 0'

A closer analysis reveals that unchanged for which

will leave all those elements of the tuple

 

 

) 0'

B 0(% ) '

) ' 02%

 

and

 )          )   P      &  &       &    &       Q  !   !       !   &      !   !  &  !   &   !   &   !    

B
 

 

 
% % %


$
%

 

 
 
% % ) 0'



@ @

one of the fundamental tools of domain theory and plays its most prominent role in the solution of recursive domain equations, Chapter 5. Because of this coincidence we will henceforth also speak of the bilimit of an expanding sequence and denote it by . We can generalize Theorem 3.3.7 in two ways; we can replace by an arbitrary directed set (in which case we will speak of an expanding system) and we can use general Scott-continuous adjunctions instead of e-p-pairs. The rst generalization is harmless and does not need any serious adjustments in the proofs. We will freely use it from now on. The second, on the other hand, is quite interesting. By the passage from embeddings to, no longer injective, lower adjoints, we allow domains not only to grow but also to shrink as we move on in the index set. Thus points, which at some stage looked different, may at a later stage be recognised to be the same. The interested reader will nd an outline of the mathematical theory of this in the exercises. For the main text, we must remind ourselves that this generalization has so far not found any application in semantics. Part (1) of the preceding theorem gives a characterization of bilimits:

Proof. Necessity is Part (1) of Theorem 3.3.7. For sufciency we show that the bilimit as constructed there, is isomorphic to . We already have maps and because is the bilimit. These commute with the limiting and the colimiting morphisms, respectively. So let us check that they compose to identities:

We note that in the proof we have used the condition only for the rst calculation. Without it, we still get that and form an e-p-pair. Thus we have:
% % ) ' 02%

)    

50

) 0'

     

and

$ 4!

 P

      !    1 

) ' 0(%

B 0' )

) ' 0(%

) ' 0(%

) ' 0(%

% B 0(% ) '

Lemma 3.3.8. Let , are Scott-continuous functions and, secondly,

% % % @ % ) ' 02% % ) ' 0(% % @

be a cocone for the expanding sequence . It is colimiting if and only if, rstly, there such that each is an e-p-pair holds.

  7  ! $   $ !   "    P$  !   

  P   %
% % @ V

of the sequence is a sub-domain of In other words:

Corollary 3.3.10. The bilimit of an expanding sequence is also the colimit (limit) in the restricted category of dcpos with embeddings (projections) as morphisms. 3.3.3 Bilimits of domains its bilimit.
B 0' )

we rst show that is directed. To Proof. Given an element this end it is sufcient to show that for all and for each there is with . Well, because approximates and because embeddings preserve the order of approximation, we have . Since , some is . This implies . above The set gives back because . It consists solely of approximants to because the are lower adjoints. Exercises 3.3.12. 1. Let be a continuous domain and let be an holds in the idempotent Scott-continuous function. Show that image of if and only if there exists in such that . In the case that is algebraic conclude that an element of is compact if and only if there is with . 2. Let in be a kernel operator with nite image. Show that and that itself is compact in . is called order dense if for each pair .

% % % % % " % ) ' 0(% % " % " % % $

is contained

(a) Let be an order dense chain in an algebraic domain continuous idempotent function on with algebraic (it must be continuous by Theorem 3.1.4).

. Construct a and not

(b) Let, conversely, be a continuous and idempotent function on an algebraic dcpo such that its image is not algebraic. Show that contains an order dense chain.

51

3. [Hut92] A chain such that

there exists

) ' 0(%

) ' 0(%

"

"

) ' 0(%

"

% ) ' 02%

"

    F 

2. If all

are ( -)algebraic then so is

and

"

"

) ' 0(%

 

1. If all each

are ( -)continuous then so is then a basis for is given by

    %    %  9    8  $6 9@#% 5  %          %         $   !     F F  F            F                     F                 F          F                    P

. If we are given bases .

) ' 0(%

B 0(% ) ' @

Theorem 3.3.11. Let quence and

%

 7$ 3! P    P  

B ) 0' B 0(% ) '

Proposition 3.3.9. Let

% % ) ' 0(% % % @

 7$  !  P    P

$ @

be a cocone over the expanding sequence where the are embeddings. Then the bilimit

be an expanding sefor

 

 

) ' 0(%

"

(c) An algebraic domain is called retraction stable if every idempotent on has an algebraic image. Prove that an algebraic domain is retraction stable if and only if does not contain an order dense chain. (d) Formulate a similar result for projection stable domains.

5. Formulate and prove a generalization of Proposition 3.1.13 for arbitrary posets. 6. Formulate an analogue of Proposition 3.2.4 for innite products. Proceed as follows: First restrict to pointed dcpos. Next nd an example of a (non-pointed) nite poset which has a non-algebraic innite power. This should give you enough intuition to try the general case. 7. A dcpo may be seen as a topological space with respect to the Scott-topology. Given two dcpos we can form their product in DCPO. Show that the Scotttopology on the product need not be the product topology but that the two topologies coincide if one of the factors is a continuous domain. 8. Construct an example which shows that Lemma 3.2.6 does not hold for innite products.

9. Derive

and composition as maps in an arbitrary cartesian closed category.

10. Let C be a cartesian closed full subcategory of DCPO. Let R-C be the full subcategory of DCPO whose objects are the retracts of objects of C. Show that R-C is cartesian closed. 11. Let be the negative integers with the usual ordering. Show that the order of approximation on is empty. Find a pointed algebraic dcpo in which a similar effect takes place. 12. Show that DCPO does not have coproducts. 13. Show that CONT does not have equalizers for all pairs of morphisms. (Hint: First convince yourself that limits in CONT, if they exist, have the same underlying dcpo as when they are calculated in DCPO.) 14. Complement the table in Section 3.2.6 with the innitary counterparts of cartesian product, disjoint union, smash product and sum. Observe that for these the cardinality of the basis does play a role, so you have to add columns for -CONT etc. 15. Show that the embeddings into the bilimit of an expanding sequence are given more concretely by with

Find a similar description for expanding systems. 52

P P       !   

 

8 46 5 $

4. Let Show that

be an embedding projection pair between -semilattices. is a lower set in if and only if for all in we have .

 

I      !   %  ! 

H WI

16. Redo Section 3.3.2 for directed index sets and Scott-continuous adjunctions. The following are the interesting points: (a) The limit-colimit coincidence, Theorem 3.3.7, holds verbatim. (b) The characterization of bilimits given in Lemma 3.3.8 does not sufce. It states that must not contain superuous elements. Now we also need to say that does not identify too many elements. (c) Given an expanding system with adjunctions, we can pass to quotient domains by setting . Show that the original adjunctions when restricted and corestricted to the become ep-pairs and that these dene the same bilimit. 17. Let be the space of Scott-continuous idempotents on a dcpo . Apply the previous exercise to show that in implies (where the connecting adjunctions are given by restricting the retractions to the respective image). 18. Prove that the Scott-topology on a bilimit of continuous domains is the restriction of the product topology on the product of the individual domains.
V # '

  A  % %     %  ( P   
53
B

 %

4 Cartesian closed categories of domains

In the last chapter we have seen that our big ambient categories DCPO and DCPO are, among other things, cartesian closed and we have already pointed out that for the natural classes of domains, CONT and ALG, this is no longer true. The problematic construction is the exponential, which as we know by Lemma 3.2.7, must be the set of Scott-continuous functions ordered pointwise. If, on the other hand, we nd a full subcategory of CONT which is closed under terminal object, cartesian product and function space, then it is also cartesian closed, because the necessary universal properties are inherited from DCPO. Let us study more closely why function spaces may fail to be domains. The fact that the order of approximation may be empty tells us that there may be no natural candidates for basis elements in a function space. This we can better somewhat by requiring the image domain to contain a bottom element.

More generally, we will use to and everything else to . Proposition 4.0.2.

F

for the function which maps the Scott-open set

1. Step functions are Scott-continuous. is pointed and let approximates .

2. Let and be dcpos where approximates then 3. If, in addition,

be continuous. If

and

are continuous then

is a supremum of step functions.

Proof. (1) Continuity follows from the openness of , respectively . . Suprema in (2) Let be a directed family of functions with are calculated pointwise so we also have . This implies that for some , holds. A simple case distinction then shows that must be above everywhere. (3) We show that for each and each in there is a step function approximating which maps to . Indeed, from we get and so for some we have . The desired step function is therefore given by . Continuity of implies that we can get arbitrarily close to this way. Note that the supremum in (3) need not be directed, so we have not shown that is again continuous. Was it a mistake to require directedness for the set of approximants? The answer is no, because without it we could not have proved (3) in the rst place. The problem of joining nitely many step functions together, so as to build directed collections of approximants, comes up already in the case of two step functions

54

8 #% $6 5

8 B6  5 $             '                               

! $ 3)

if otherwise.

      P    ' P   

 

'

 

  

Denition 4.0.1. For and we dene the step function

dcpos where

has a least element and by

 P

$  !   

 

Figure 10: Finding an upper bound for two step functions. and which approximate a given continuous function . The situation is illustrated in Figure 10. The problem is where to map the (Scott-open but otherwise unstructured) set . It has to be done in such a way that the resulting function still approximates . As it will turn out, it sufces to make special assumptions about either the image domain the topic of Section 4.1 or about the pre-image domain the topic of Section 4.2. In both cases we restrict our attention to pointed domains, and we work with step functions and joins of these. From these we can pass to more general domains in again two ways. This will be outlined briey in Section 4.3.2. The question then arises whether we have not missed out on some alternative way of building a cartesian closed category. This is not the case as we will see in Section 4.3. The basic tool for this fundamental result, Lemma 4.3.1, will nicely connect up with the dichotomy distinguishing 4.1 and 4.2.

4.1 Local uniqueness: Lattice-like domains

The idea for adjusting the image domain is simple; we assume that and have a least upper bound (if bounded at all). Mapping the intersection to (and to and to ) results in a continuous function which is above and and still approximates . This is seen as follows: Suppose is a directed collection of functions with supremum above . Some must be above and some must be above . Then by construction every upper bound of in is above . In fact, we do not need that the join of and exists globally in . It sufces to form the join for every inside , because we have seen in Proposition 2.2.17 that all considerations about the order of approximation can be performed inside principal ideals. We have the following list of denitions. Denition 4.1.1. Let be a pointed continuous domain. We say that is ;

55

2. a bounded-complete domain (or bc-domain), if each bounded pair has a supremum;

 P

 

1. an L-domain, if each pair

bounded by

has a supremum in

   


``



` `

 
`

`` "


`

`


` `` `

YY

YY

  
Y


YY



!


Y


 P

 

   

`` `

  I  P

``

 


` `

` ` ` `

Y`Y `



        

! 

Y Y

Y


Y Y




Figure 11: Separating examples for the categories of lattice-like domains.



We denote the full subcategories of CONT corresponding to these denitions by L, BC, and LAT. For the algebraic counterparts we use aL, aBC, and aLAT. All this still makes sense if we forget about approximation but, surely, at this point the reader does not suffer from a lack of variety as far as categories are concerned. We would like to point out that continuous lattices are the main objects of study in [GHK 80], a mathematically oriented text, whereas the objects of -aBC are often the domains of choice in semantics, where they appear under the name Scott-domain. Typical examples are depicted in Figure 11. They even characterize the corresponding categories, see Exercise 4.3.11(3). Since domains have directed joins anyway, we see that in L-domains every subset of a principal ideal has a supremum in that ideal. We also know that complete lattices can alternatively be characterized by inma. The same game can be played for the other two denitions: Proposition 4.1.2. Let be a pointed continuous domain. Then is an L-domain, a bc-domain, or a continuous lattice if and only if it has inma for bounded non-empty, non-empty, or arbitrary subsets, respectively. The consideration of inma may seem a side issue in the light of the problem of turning function spaces into domains. Its relevance becomes clear when we remember that upper adjoints preserve inma. The second half of the following is therefore a simple observation. The rst half follows from Proposition 3.1.2 and Theorem 3.1.4. Proposition 4.1.3. Retracts and bilimits of L-domains (bc-domains, continuous lattices) are again L-domains (bc-domains, continuous lattices). We can treat continuous and algebraic lattice-like domains nicely in parallel because the ideal completion respects these denitions: 56

3. (repeated for comparison) a continuous lattice, if each pair supremum.

P 

ALG L

L BC

BC LAT

` `

 
F ` ` Y Y Y

` `

has a

Proposition 4.1.4. Let be an L-domain (bc-domain, continuous lattice). Then is an algebraic L-domain (bc-domain, lattice). Thus L, BC, and LAT contain precisely the retracts of objects of aL, aBC, and aLAT, respectively. We conclude this section by stating the desired closure property of lattice-like domains. Proposition 4.1.5. Let be a continuous domain and an L-domain (bc-domain, continuous lattice). Then is again an L-domain (bc-domain, continuous lattice). Corollary 4.1.6. The categories L, BC, LAT, and their algebraic counterparts are cartesian closed.

4.2 Finite choice: Compact domains

Let us now turn our attention to the rst argument of the function space construction, which means by the general considerations from the beginning of this chapter, the study of open sets and their nite intersections. Step functions are dened using basic open sets of the form , and the fact that there is a single generator was crucial in the proof that approximates whenever approximates . Arbitrary open sets are unions of such basic opens (Proposition 2.3.6) but in general this is an innite union and so the proof of Proposition 4.0.2 will no longer work. For the rst time we have now reached a point in our exposition where the theory of algebraic domains is denitely simpler and better understood than that of continuous domains. Let us therefore treat this case rst. 4.2.1 Binite domains Step functions may in the algebraic case be dened using compact elements only, where the characteristic pre-image is actually equal to . Taking up our line of thought from above, we want for the algebraicity of the function space that the intersection is itself generated by nitely many compact points: . Note that the must be minimal upper bounds of . For each we choose a compact element below and above . New intersections then come up, this time between the different s. Let us therefore further assume that after nitely many iterations this process stops. It is an easy exercise to show that the function constructed in this way is a compact element below and above and . We hope that this provides sufcient motivation for the following list of denitions. Denition 4.2.1. Let

be a poset. (Think of

as the basis of an algebraic domain.)

1. We say that is mub-complete (or: has property m) if for every upper bound of a nite subset of there is a minimal upper bound of below . Written as a formula: . n 2. For a subset of let its mub-closure be the smallest superset of which for every nite also contains . 57
I

QP 

  



  $ %   %   %    ! %   

 

8 #% 5  $6

'



    



    



 P 

3. We say that has the nite mub property if it is mub-complete and if every nite subset has a nite mub-closure. If, in addition, has a least element, then we call it a Plotkin-order. 4. An algebraic domain whose basis of compact elements is a Plotkin-order is called a binite domain. The full subcategory of ALG of binite domains we denote by B. With this terminology we can formulate precisely how nitely many step functions combine to determine a compact element in the function space [Abr91b]. Denition 4.2.2. Let nite subset of

be a binite domain and let is called joinable if

be pointed and algebraic. A

The function which we associate with a joinable family



is

Lemma 4.2.3. If is a binite domain and is pointed and algebraic, then every joinable subset of gives rise to a compact element of . If and are joinable families then the corresponding functions are related if and only if and The expected result, dual to Proposition 4.1.5 above, then is: is a binite domain and Proposition 4.2.4. If is algebraic. All compact elements of lies.

is pointed and algebraic, then arise from joinable fami-

Note that this is strictly weaker

Comment: Unfortunately, it is as yet unproven.

than Proposition 4.1.5 and we do not immediately get that B is cartesian closed. For this we have to nd alternative descriptions. The fact that we can get an algebraic function space by making special assumptions about either the argument domain or the target domain was noted in a very restricted form in [Mar81]. The concept of nite mub closure is best explained by illustrating what can go wrong. In Figure 12 we have the three classical examples of algebraic domains which are not binite; in the rst one the basis is not mub-complete, in the second one there is an innite mub-set for two compact elements, and in the third one, although all mubsets are nite, there occurs an innite mub-closure. On a more positive note, it is clear that every nite and pointed poset is a Plotkin-order and hence binite. This trivial example contains the key to a true understanding of biniteness; we will now prove that binite domains are precisely the bilimits of nite pointed posets.

58

 

   8 #6% 5 $  P C    3 C   P  

8 46  5 $

P  C    P     4  H             $  

Q !

 # C   C  H   C

8 B6  5 $

Figure 12: Typical non-binite domains. Proposition 4.2.5. Let be an algebraic domain with mub-complete basis let be a set of compact elements. Then there is a least kernel operator which keeps xed. It is given by .

Proof. First note that is well-dened because the supremum is indeed over a directed set. This follows from mub-completeness. Continuity follows from Corollary 2.2.16. On the other hand, it is clear that a kernel operator which xes must also x each element of the mub-closure , and so is clearly the least monotone function with the desired property. In a binite domain nite sets of compact elements have nite mub-closures. By the preceding proposition this implies that there are many kernel operators on such a domain which have a nite image. In fact, we get a directed family of them, because the order on kernel operators is completely determined by their images, Proposition 3.1.17. For the sake of brevity, let us call a kernel operator with nite image an idempotent deation. Theorem 4.2.6. Let 1. be a pointed dcpo

. The following are equivalent

is a binite domain.

3. The set of all idempotent deations is directed and yields

Proof. What we have not yet said is how algebraicity of follows from the existence of idempotent deations. For this observe that the inclusion of the image of a kernel operator is a lower adjoint and as such preserves compactness. For the implication we use the fact that idempotent deations are in any case compact elements of the function space.

59

'

2. There exists a directed collection supremum equals .

of idempotent deations of

as its join.

  

and on whose

``

YY

  

  % 

 

  

``

YY

YY

``

YY

 

`` Y

! 

It is now only a little step to the promised categorical characterization. Theorem 4.2.7. A dcpo is binite if and only if it is a bilimit of an expanding system of nite pointed posets. Proof. Let be binite and let be a family of idempotent deations generating the identity. Construct an expanding system by taking as objects the images of the deations and as connecting embeddings the inclusion of images. The associated upper adjoint is given by restricted to . is the bilimit of this system by Lemma 3.3.8. If, conversely, is a bilimit of nite posets then clearly the compositions , where is the upper adjoint of , satisfy the requirements of Theorem 4.2.6. So we have three characterizations of biniteness, the original one, which may be called an internal description, a functional description by Theorem 4.2.6, and a categorical one by Theorem 4.2.7. Often, the functional characterization is the most handy one in proofs. We should also mention that binite domains were rst dened by Gordon Plotkin in [Plo76] using expanding sequences. (In our taxonomy these are precisely the countably based binite domains.) The acronym he used for them, SFP, continues to be quite popular. Theorem 4.2.8. The category B of binite domains is closed under cartesian product, function space, coalesced sum, and bilimits. In particular, B is cartesian closed. Proof. Only function space and bilimit are non-trivial. We leave the latter as an exercise. For the function space let and be binite with families of idempotent deations and . A directed family of idempotent deations on is given by the maps , . 4.2.2 FS-domains

Let us now look at continuous domains. The reasoning about what the structure of should be in order to ensure that is continuous is pretty much the same as for algebraic domains. But at the point where we there introduced the mub-closure of a nite set of compact elements, we must now postulate the existence of some nite and nitely supported partitioning of . This is clearly an increase in the logical complexity of our denition and also of doubtful practical use. It is more satisfactory to generalise the functional characterization. Denition 4.2.9. Let be a dcpo and be a Scott-continuous function. We say that is nitely separated from the identity on , if there exists a nite set such that for any there is with . We speak of strong separation if for each there are elements with . A pointed dcpo is called an FS-domain if there is a directed collection of continuous functions on , each nitely separated from , with the identity map as their supremum.

60

8 #% 5 $6

'

 

P     D  7$   )    !

  !( %

H   P

B (

 8 #% 5 $6


'

 P 
'

$ ! C 

   '  

( 0 1' ( (

'

It is relatively easy to see that FS-domains are indeed continuous. Thus it makes sense to speak of FS as the full subcategory of CONT where the objects are the FSdomains. We have exact parallels to the properties of binite domains, but often the proofs are trickier. Proposition 4.2.10. If is an FS-domain and is continuous.

Theorem 4.2.11. The category FS is closed under the formation of products, function spaces, coalesced sums, and bilimits. It is cartesian closed. What we do not have are a categorical characterization or a description of FSdomains as retracts of binite domains. All we can say is the following. Proposition 4.2.12. 1. Every binite domain is an FS-domain.

2. A retract of an FS-domain is an FS-domain. 3. An algebraic FS-domain is binite. To fully expose our ignorance, we conclude this subsection with an example of a well-structured FS-domain of which we do not know whether it is a retract of a binite domain. be the collection of all closed discs in the plane plus the plane Example. Let itself, ordered by reversed inclusion. One checks that the ltered intersection of discs is again a disc, so is a pointed dcpo. A disc approximates a disc if and only if is a neighborhood of . This proves that is continuous. For every we dene a map on as follows. All discs inside the open disc with radius are mapped to their closed -neighborhood, all other discs are mapped to the plane which is the bottom element of . Because the closed discs contained in some compact set form a compact space under the Hausdorff subspace topology, these functions are nitely separated from the identity map. This proves that is a countably based FS-domain. 4.2.3 Coherence This is a good opportunity to continue our exposition of the topological side of domain theory, which we began in Section 2.3. We need a second tool complementing the lattice of Scott-open sets, namely, the compact saturated sets. Here compact is to be understood in the classical topological sense of the word, i.e. a set of a topological space is compact if every covering of by open sets contains a nite subcovering. Saturated are those sets which are intersections of their neighborhoods. In dcpos equipped with the Scott-topology these are precisely the upper sets, as is easily seen using opens of the form . Theorem 4.2.13. Let be a continuous domain. The sets of open neighborhoods of compact saturated sets are precisely the Scott-open lters in .

61

 

 

   

is pointed and continuous then

8 B6  5 $

By Proposition 7.2.27 this is a special case of the Hofmann-Mislove Theorem 7.2.9. Let us denote the set of compact saturated sets of a dcpo , ordered by reverse inclusion, by . We will refer to families in which are directed with respect to reverse inclusion, more concretely as ltered families. The following, then, is only a re-formulation of Corollary 7.2.11. Proposition 4.2.14. Let

be a continuous domain.

1.

is a dcpo. Directed suprema are given by intersection.

2. If the intersection of a ltered family of compact saturated sets is contained in a Scott-open set then some element of it belongs to already.

Proposition 4.2.15. Let

1. 2. 3.

is a continuous domain.


holds in .

if and only if there is a compact saturated set

Proof. All three claims are shown easily using upper sets generated by nitely many points: If is an open neighborhood of a compact saturated set then there exists a nite set of points of with . The interesting point about FS-domains then is, that their space of compact saturated sets is actually a continuous lattice. We already have directed suprema (in the form of ltered intersections) and continuity, so this boils down to the property that the intersection of two compact saturated sets is again compact. Let us call domains for which this is true, coherent domains. Given the intimate connection between and , it is no surprise that we can read off coherence from the lattice of open sets. is coherent if and only if for all Proposition 4.2.16. A continuous domain with and we also have . (In Figure 6 we gave an example showing that the condition is not true in arbitrary continuous lattices.) This result specializes for algebraic domains as follows: Proposition 4.2.17. An algebraic domain is coherent if and only if complete and nite sets of have nite sets of minimal upper bounds.

is mub-

This proposition was named 2/3-SFP Theorem in [Plo81] because coherence rules out precisely the rst two non-examples of Plotkin-orders, Figure 12, but not the third. The only topological characterization of binite domains we have at the moment, makes use of the continuous function space, see Lemma 4.3.2.

62

 

"

     

 

holds in

if and only if there is a Scott-open set

with

"

 P P

3.

is a dcpo. be a continuous domain.



with

We observe that for algebraic coherent domains, and have a common sublattice, namely that of compact-open sets. These are precisely the sets of the form with the compact elements. This lattice generates both and when we form arbitrary suprema. This pleasant coincidence features prominently in Chapter 7. Theorem 4.2.18. FS-domains (binite domains) are coherent. Let us reformulate the idea of coherence in yet another way.

Proposition 4.2.20. Let

be a continuous domain.

1. The Lawson-topology on is Hausdorff. Every Lawson-open set has the form where is Scott-open and is Scott-compact saturated. 2. The Lawson-topology on is compact if and only if

is coherent.

3. A Scott-continuous retract of a Lawson-compact continuous domain is Lawsoncompact and continuous. So we see that FS-domains and binite domains carry a natural compact Hausdorff topology. We will make use of this in Chapter 6.

4.3 The hierarchy of categories of domains

The purpose of this section is to show that there are no other ways of constructing a cartesian closed full subcategory of CONT or ALG than those exhibited in the previous two sections. The idea that such a result could hold originated with Gordon Plotkin, [Plo81]. For the particular class -ALG it was veried by Mike Smyth in [Smy83a], for the other classes by Achim Jung in [Jun88, Jun89, Jun90]. All these classication results depend on the Axiom of Choice. 4.3.1 Domains with least element Let us start right away with the crucial bifurcation lemma on which everything else in this section is based. Lemma 4.3.1. Let and be continuous domains, where is pointed, such that is continuous. Then is coherent or is an L-domain. Proof. By contradiction. Assume is not coherent and is not an L-domain. By Proposition 4.2.16 there exist open sets , and in such that and hold but not . Therefore there is a directed collection of open sets covering , none of which covers . We shall also need interpolating sets and , that is, and . The assumption about not being an L-domain can be transformed into two special cases. Either contains the algebraic domain from Figure 12 (where the descending 63

'

Denition 4.2.19. The Lawson-topology on a dcpo taining all Scott-open sets and all sets of the form

is the smallest topology con. It is denoted by .

P 

   

8 B6  5 $

chain in may generally be an ordinal) or from Figure 11 as a retract. We have left the proof of this as Exercise 4.3.11(3). Note that if is a retract of then is a retract of and hence the former is continuous if the latter is. Let us now prove for both cases that is not continuous. Case 1: . Consider the step functions and . They clearly approximate , which is dened by if if if otherwise.

 

Since approximating sets are directed we ought to nd an upper bound for and approximating . But this impossible: Given an upper bound of below we have the directed collection dened by

No is above because must contain a non-empty piece of and there is strictly below . The supremum of the , however, equals . Contradiction. Case 2: . We choose open sets in as in the previous case. The various functions, giving the contradiction, are now dened by , , if if if otherwise.

The remaining problem is that coherence does not imply that is an FS-domain (nor, in the algebraic case, that it is binite). It is taken care of by passing to higherorder function spaces: Lemma 4.3.2. Let be a continuous domain with bottom element. Then FS-domain if and only if both and are coherent.

(The proof may be found in [Jun90].) Combining the preceding two lemmas with Lemmas 3.2.5 and 3.2.7 we get the promised classication result. Theorem 4.3.3. Every cartesian closed full subcategory of CONT FS or L. 64

is contained in

if if otherwise.


8 $ 97B6  5

if if otherwise.


and

   8 46% 5 $

  P    '  P P   ' P  '    P  ' P #  ' P     P           P P

 '   '  ' 

  

P  PP P # 


8 #6% $ 6 $ 5 8 4% 5


'

 

is an

Adding Proposition 4.2.12 we get the analogue for algebraic domains: Theorem 4.3.4. Every cartesian closed full subcategory of ALG or aL.

Forming the function space of an L-domain may in general increase the cardinality of the basis (Exercise 4.3.11(17)). If we restrict the cardinality, this case is ruled out: Theorem 4.3.5. Every cartesian closed full subcategory of -CONT ( -ALG ) is contained in -FS ( -B). 4.3.2 Domains without least element The classication of pointed domains, as we have just seen, is governed by the dichotomy between coherent and lattice-like structures. Expressed at the element level, and at least for algebraic domains we have given the necessary information, it is the distinction between nite mub-closures and locally unique suprema of nite sets. It turns out that passing to domains which do not necessarily have bottom elements implies that we also have to study the mub-closure of the empty set. We get again the same dichotomy. Coherence in this case means that itself, that is, the largest element of , is a compact element. This is just the compactness of as a topological space. And the property that is lattice-like boils down to the requirement that each element of is above a unique minimal element, so is really the disjoint union of pointed components. Lemma 4.3.6. Let and ous. Then is compact or be continuous domains such that is a disjoint union of pointed domains.

is continu-

The proof is a cut-down version of that of Lemma 4.3.1 above. The surprising fact is that this choice can be made independently from the choice between coherent domains and L-domains. Before we state the classication, which because of this independence, will now involve cases, we have to rene the notion of compactness, because just like coherence it is not the full condition necessary for cartesian closure. Denition 4.3.7. A dcpo is a nite amalgam if it is the union of nitely many pointed dcpos such that every intersection of s is also a union of s. (Compare the denition of mub-complete.) For categories whose objects are nite amalgams of objects from another category C we use the notation F-C. Similarly, we write U-C if the objects are disjoint unions of objects of C. Proposition 4.3.8. A mub-complete dcpo is a nite amalgam if and only if the mubclosure of the empty set is nite. Lemma 4.3.9. If both amalgam. and are compact and continuous then

is a nite

Theorem 4.3.10. 1. The maximal cartesian closed full subcategories of CONT are F-FS, U-FS, F-L, and U-L. 65

8 B6  5 $

is contained in B

8 $ 9746  5

QP P 

2. The maximal cartesian closed full subcategories of ALG are F-B, U-B, F-aL, and U-aL. At this point we can answer a question that may have occurred to the diligent reader some time ago, namely, why we have dened binite domains in terms of pointed nite posets, where clearly we never needed the bottom element in the characterizations of them. The answer is that we wanted to emphasize the uniform way of passing from pointed to general domains. The fact that the objects of F-B can be represented as bilimits of nite posets is then just a pleasant coincidence.

2. Let be a bounded-complete domain. Show that is a Scott-continuous function from to . 3. Characterize the lattice-like (pointed) domains by forbidden substructures:

4. Find a poset in which all pairs have nite mub-closures but in which a triple of points exists with innite mub-closure. 5. Show that if for an algebraic domain is not necessarily mub-complete. the basis is mub-complete then

6. Show that in a binite domain nite sets of non-compact elements may have innitely many minimal upper bounds and, even if these are all nite, may have innite mub-closures.

8. Prove that bilimits of binite domains are binite. 9. Prove the following statements about retracts of binite domains. (a) A pointed dcpo is a retract of a binite domain if and only if there is a directed family of functions on such that each has a nite image and such that . (You may want to do this for countably based domains rst.) (b) The ideal completion of a retract of a binite domain need not be binite. 66

7. Show that if mub-closed.

is a two-element subset of an L-domain then

 !

 

(d)

is a bounded-complete domain but not a lattice if and only if domain in Figure 11 is a retract of .

itself

(c)

is an L-domain but not bounded-complete if and only if domain Figure 11 is a retract of .

'

(b)

is mub-complete but not an L-domain if and only if domain ure 11 is a retract of .

(a)

is -continuous but not mub-complete if and only if domain ure 12 is a retract of .

'

 H   8 $ 97B6  5

Exercises 4.3.11. space

1. [Jun89] Show that a dcpo is continuous.

is continuous if the function

in Figin Figin

is

(c) If is a countably based retract of a binite domain then it is also the image of a projection from a binite domain. (Without countability this is an open problem.) (d) The category of retracts of binite domains is cartesian closed and closed under bilimits. 10. Prove that FS-domains have inma for downward directed sets. As a consequence, an FS-domain which has binary inma, is a bc-domain. 11. Show that in a continuous domain the Lawson-closed upper sets are precisely the Scott-compact saturated sets. 12. Characterize Lawson-continuous maps between binite domains. 13. We have seen that every binite domain is the bilimit of nite posets. As such, it can be thought of as a subset of the product of all these nite posets. Prove that the Lawson-topology on the binite domain is the restriction of the product topology if each nite poset is equipped with the discrete topology. 14. Prove that a coherent L-domain is an FS-domain. 15. Characterize those domains which are both L-domains and FS-domains. 16. Characterize Scott-topology and Lawson-topology on both L-domains and FSdomains by the ideal of functions approximating the identity. 17. [Jun89] Let be an L-domain such that that is an FS-domain.

67

8 4 5 $6

is countably based. Show

5 Recursive domain equations

The study of recursive domain equations is not easily motivated by reference to other mathematical structure theories. So we shall allow ourselves to deviate from our general philosophy and spend some time on examples. Beyond motivation, our examples represent three different (and almost disjoint) areas in which recursive domain equations arise, in which they serve a particular role, and in which particular aspects about solutions become prominent. It is an astonishing fact that within domain theory all these aspects are dealt with in a unied and indeed very satisfactory manner. This richness and interconnectedness of the theory of recursive domain equations, beautiful as it is, may nevertheless appear quite confusing on a rst encounter. As a general guideline we offer the following: Recursive domain equations and the domain theory for solving them comprise a technique that is worth learning. But in order to understand the meaning of a particular recursive domain equation, you have to know the context in which it came up.

5.1 Examples
5.1.1 Genuine equations . Solving this equation in a cartesian The prime example here is closed category gives a model for the untyped -calculus [Sco80, Bar84], in which, as we know, no type distinction is made between functions and arguments. When setting up an interpretation of -terms with values in , where solves this equation, we need the isomorphisms and explicitly. We conclude that even in the case of a genuine equation we are looking not only for an object but an object plus an isomorphism. This is a rst hint that we shall need to treat recursive domain equations in a categorical setting. However, the function space operator is contravariant in its rst and covariant in its second argument and so there is denitely an obstacle to overcome. A second problem that this example illustrates is that there may be many solutions to choose from. How do we recognize a canonical one? This will be the topic of Section 5.3. Besides this classical example, genuine equations are rare. They come up in semantics when one is confronted with the ability of computers to treat information both as program text and as data. 5.1.2 Recursive denitions In semantics we sometimes need to make recursive denitions, for very much the same reasons that we need recursive function calls, namely, we sometimes do not know how often the body of a denition (resp. function) needs to be repeated. To give an example, take the following denition of a space of so-called resumptions:

We read it as follows: A resumption is a map which assigns to a state either a nal state or an intermediary state together with another resumption representing the remaining 68

 $ 9@#% 5 !  9@#% 5 $ B! 2 8  $ 6 8  $6  8 H
#

82 #6 5 A $

$ B6 5 A

computation. Such a recursive denition is therefore nothing but a shorthand for an innite (but regular) expression. Likewise, a while loop could be replaced by an innite repetition of its body. This analogy suggests that the way to give meaning to a recursive denition is to seek a limit of the repeated unwinding of the body of the denition starting from a trivial domain. No doubt this is in accordance with our intuition, and indeed this is how we shall solve equations in general. But again, before we can do this, we need to be able to turn the right hand side of the specication into a functor. 5.1.3 Data types Data types are algebras, i.e. sets together with operations. The study of this notion is known as Algebraic Specication [EM85] or Initial Algebra Semantics [GTW78]. We choose a formulation which ts nicely into our general framework. Denition 5.1.1. Let be a functor on a category C. An -algebra is given by an object and a map . A homomorphism between algebras and is a map such that the following diagram commutes:

For example, if we let be the functor over Set which assigns to (where is the one-point dcpo as discussed in Section 3.2.1), then -algebras are precisely the algebras with one nullary and one binary operation in the sense of universal algebra. Lehmann and Smyth [LS81] discuss many examples. Many of the data types which programming languages deal with are furthermore totally free algebras, or term algebras on no generators. These are distinguished by the fact that there is precisely one homomorphism from them into any other algebra of the same signature. In our categorical language we express this by initiality. Term algebras (alias initial -algebras) are connected with the topic of this chapter because of the following observation: Lemma 5.1.2. If is an initial -algebra then is an isomorphism.

Proof. Consider the following composition of homomorphisms:

where is the unique homomorphism from guaranteed by initiality. Again by initiality, 69

 C C  0$  C4!   C $  #!    C    C  C  C  C  C  C

to must be

. And from the rst

$  4! C H C

 C  C  C $ ! $  C 3C ) $  C ! !

$  3!  C

So in order to nd an initial -algebra, we need to solve the equation . But once we get a solution, we still have to check initiality, that is, we must validate that the isomorphism from to is the right structure map. In category theory we habitually dualize all denitions. In this case we get (nal) co-algebras. Luckily, this concept is equally meaningful. Where the map describes the way how new objects of type are constructed from old ones, a map stands for the opposite process, the decomposition of an object into its constituents. Naturally, we want the two operations to be inverses of each other. In other words, if is an initial -algebra, then we require to be the nal co-algebra. Peter Freyd [Fre91] makes this reasoning the basis of an axiomatic treatment of domain theory. Beyond and above axiomatizing known results, he treats contravariant and mixed variant functors and offers a universal property encompassing both initiality and nality. This will allow us to judge the solution of general recursive domain equations with respect to canonicity.

5.2 Construction of solutions

Suppose we are given a recursive domain equation where the right hand side denes a functor on a suitable category of domains. As suggested by the example in Section 5.1.2, we want to repeat the trick which gave us xpoints for Scottcontinuous functions, namely, to take a (bi-)limit of the sequence . Remember that bilimits are dened in terms of e-p-pairs. This makes it necessary that we, at least temporarily, switch to a different category. The convention that we adopt for this chapter is to let D stand for any category of pointed domains, closed under bilimits. All the cartesian closed categories of pointed domains mentioned in Chapter 4 qualify. We denote the corresponding subcategory where the morphisms are embeddings by D . Some results will only hold for strict functions. Recall that our notation for these were and D for categories. Despite this unhappy (but unavoidable) proliferation of categories, recall that the central limit-colimit Theorem 3.3.7 and Corollary 3.3.10 state a close connection: Colimits of expanding sequences in D are also colimits in D and, furthermore, if the embeddings dening the sequence are replaced by their upper adjoints, the colimit coincides with the corresponding limit. This will bear fruit when we analyze the solutions we get in D from various angles as suggested by the examples in the last subsection. Let us now start by just assuming that our functor restricts to D . 5.2.1 Continuous functors
) ' 0(% B 0(% ) '

70

Q 

 C P 

B 0' )

B )0' % B 0(% ) '

) ' 02%

Denition 5.2.1. A functor panding sequence we have that

D is called continuous, if for every exwith colimit is a colimit of the sequence , .

C  7$ ! 

P Q C PC  QCP

  C A $  3) C!

  C A

quadrangle we get of each other.

. So

and are inverses

 C  C   C 


Q  7$   C !  C C  7$ $3! D! Q P C C P    C 

 C C

#$ 6 4!

$  3!  C
% @ % @

C  &$ !

This, obviously, is Scott-continuity expressed for functors. Whether we formulate it in terms of expanding sequences or expanding systems is immaterial. The question is not, what is allowed to enter the model, but rather, how much do I have to check before I can apply the theorems in this chapter. And sequences are all that is needed. This, then, is the central lemma on which our domain theoretic technique for solving recursive domain equations is based (recall that is our notation for the upper adjoint of ):

quence the maps

. Then

is isomorphic to and

(clipping off the rst approximation makes no difference), where there is also the cocone . The latter is also colimiting by the continuity of . In this situation Theorem 3.3.7 provides us with unique mediating morphisms which are precisely the stated and . They are inverses of each other because both cocones are colimiting. The equations follow from the explicit description of mediating morphisms in Theorem 3.3.7. Note that since we have restricted attention to pointed domains, we always have the initial embedding . The solution to based on this embedding we call canonical and denote it by . 5.2.2 Local continuity

Continuity of a functor is a hard condition to verify. Luckily there is a property which is stronger but nevertheless much easier to check. It will also prove useful in the next section. Denition 5.2.3. A functor from D to E, where D and E are categories of domains, is called locally continuous, if the maps , , are continuous for all objects and from D. Proposition 5.2.4. A locally continuous functor functor from D to E .

$ (C ! $ Q  QCP   C  #6  P  %  % $

E restricts to a continuous

71

  C A

 C  &" !  C $

B 464202% 7 5 3 1 ) '

B 0(% ) '

Proof. We know that

is a colimit over the diagram

 

 

 

For each

they satisfy the equations

  C

B 02% ) '

Lemma 5.2.2. Let embedding

  C P   C $ #6 Q C 7#6  C  C  $      P       C C      C  P  7$ C7$ !!    C   C      46 Q C C  #6  C #6 $ $ $     P    C C $ ! 

 

be a continuous functor on a category D of domains. For each consider the colimit of the expanding se

via

 

) ' 0(% ) ' 0(%

V V

S S

C

We will soon generalize this, so there is no need for a proof at this point. Typically, recursive domain equations are built from the basic constructions listed in Section 3.2. The strategy is to check local continuity for each of these individually and then rely on the fact that composition of continuous functors yields a continuous functor. However, we must realize that the function space construction is contravariant in its rst and covariant in its second variable, and so the technique from the preceding paragraph does not immediately apply. Luckily, it can be strengthened to cover this case as well. D D E, contravariant in its rst, covariant in Denition 5.2.5. A functor its second variable, is called locally continuous, if for directed sets and (where are objects in D and are objects in D) we have

Proposition 5.2.6. If D D E is a mixed variant, locally continuous functor, then it denes a continuous covariant functor from D D to E as follows:

The upper adjoint to Proof. Let

be e-p-pairs in D and D, respectively. We calculate and , so maps indeed pairs of embeddings to embeddings. For continuity, let and be expanding sequences in D and D with colimits and , respectively. By Lemma 3.3.8 this implies and . By local continuity we have
B 0(% ) ' % % @

and

and so
% %

is a colimit of

While it may seem harmless to restrict a covariant functor to embeddings in order to solve a recursive domain equation, it is nevertheless not clear what the philosophical justication for this step is. For mixed variant functors this question becomes even more pressing since we explicitly change the functor. As already mentioned, a satisfactory answer has only recently been found, [Fre91, Pit93b]. We present Peter Freyds solution in the next section. Let us take stock of what we have achieved so far. Building blocks for recursive domain equations are the constructors of Section 3.2, , etc. , each of which is readily seen to dene a locally continuous functor. Translating them to embeddings

72

) ' 02%

) ' 0(%

P$ P H

) ' 0(%

) ' 0(%

B 0' )

) ' 0(%

) ' 0(%

) ' 0(%

) ' 0(%

in

for objects, and

for embeddings. .

is given by

 QP  %


Q  P  &  C P Q  QP P  C C   P  C P  QP  C      P    P   C C 9  P   C  P    C  9  P  C P  C           P   P        P      P   3P C  1  P   C  P  C  P  C  P  C  P  CC   P3 C     P   C    P  C     P P   C    P   P  C    P  C   P  C   P  C   P  C   P P C  QP  C  Q' C H C $ H EF DC ! %  P  PC  P  C    &  P# C     P  C QP  QP      P  %  $ H GFE DC !



'

'

via the preceding proposition, we get continuous functors of one or two variables. We further need the diagonal D D D to deal with multiple occurrences of in the body of the equation. Then we note that colimits in a nite power of D are calculated coordinatewise and hence the diagonal and the tupling of continuous functors are continuous. Finally, we include constant functors to allow for constants to occur in an equation. Two more operators will be added below: the bilimit in the next section and various powerdomain constructions in Chapter 6. 5.2.3 Parameterized equations Suppose that we are given a locally continuous functor in two variables. Given any domain we can solve the equation using the techniques of the preceding sections. Remember that by default we mean the solution according to Lemma 5.2.2 based on , so there is no ambiguity. Also, we have given a concrete representation for bilimits in Theorem 3.3.7, so is also well-dened in this respect. We want to show that it extends to a functor. Notation is a bit of a problem. Let D E E be a functor in two variables. We set for the functor on E which maps to for objects and to for morphisms. Similarly for . The embeddings into the canonical xpoint of , resp. , we denote by and , and we use and for the unique strict function from into and , respectively. Proposition 5.2.7. Let D E E be a locally continuous functor. Then the following denes a locally continuous functor from D to E :


Proof. Let solution to

and

be objects of D and let is given by the bilimit

Y `

be a strict function. The

and similarly for . Corresponding objects of the two expanding sequences are connected by . They commute with the embeddings of the expanding sequences: For we have because there is only one strict map from to . Higher indices follow by induc5
  

73

 

5 `

) ' 0(%

where the sequence

 C   C        C C C   7 4$ 6   3!  C   C  C      C   P   C A  7B6 3)  $  ! P  # C P      !  P $   C 7 $  $)  !! $ H DC !      C C P  P   P C  P   P C C #$ 6 !  P  C $ H DC ! Q P  C P  ' C $ ! 

On objects

is dened recursively by

) ' 0(%

on morphisms

 DC P  C A

$ !

tion:

So we have a second cocone over the sequence dening and using the fact that colimits in E are also colimits in E , we get a (unique) mediating morphism from to . By Theorem 3.3.7 it has the postulated representation. Functoriality comes for free from the uniqueness of mediating morphisms. It remains to check local continuity. So let be a directed set of maps from to . We easily get by induction and the local continuity of . The supremum can be brought to the very front by the continuity of composition and general associativity. Note that this proof works just as well for mixed variant functors. As an application, suppose we are given a system of simultaneous equations

. . .

. . .

We can solve these one after the other, viewing as parameters for the rst equation, substituting the result for in the second equation and so on. It is more direct to pass from D to D , for which Theorem 3.3.7 and the results of this chapter remain true, and then solve these equations simultaneously with the tupling of the . The fact that these two methods yield isomorphic results is known as Beki s c rule [Bek69].

5.3 Canonicity
We have seen in the rst section of this chapter that recursive domain equations arise in various contexts. After having demonstrated a technique for solving them, we must now check whether the solutions match the particular requirements of these applications. 5.3.1 Invariance and minimality Let us begin with a technique of internalizing the expanding sequence into the canonical solution. This will allow us to do proofs about without (explicit) use of the dening expanding sequence. Lemma 5.3.1. Let be a locally continuous functor on a category of domains D and let be an isomorphism. Then there exists a least homomorphism 74

&

$C $

C

 C   C #P   QCP C  # P Q  C   C # P P    # C    PC 
% % % % % % %
 

P P  P P C A  P P C A

 

 

C C C C
%

  

     C  C

'

   C
  

$  4!  C $ Q  C C C

from to every other functional on

is clearly continuous because is locally continuProof. The functional ous and composition is a continuous operation. Since we have globally assumed least elements, the function space contains as a least element. So the least xpoint of calculated as the supremum of the chain exists. We show by induction that it is below every homomorphism . For this is obvious. For the induction step assume . We calculate: . It follows that holds. On the other hand, every xpoint of is a homomorphism: . The claim about composition of least homomorphisms can also be shown by induction. But it is somewhat more elegant to use the invariance of least xpoints, Lemma 2.1.21. Consider the diagram

. Lemma 2.1.21 then gives us

Specializing the second algebra in this lemma to be itself, we deduce that on every xpoint of a locally continuous functor there exists a least endomorphism . Since the identity is always an endomorphism, the least endomorphism must be below the identity and idempotent, i.e. a kernel operator and in particular strict. This we will use frequently below. Theorem 5.3.2. (Invariance, Part 1) Let be a locally continuous functor on a category of domains D and let be an isomorphism. Then the following are equivalent: 1. 2. 3. is isomorphic to the canonical xpoint

is the least endomorphism of ; where



is dened by

75

 & 

"

8 # 5 $ 8 # 5 ! &  $6 $6  C

where is the strict operation which assigns commutes, because (because is an homomorphism) the desired equality: .

to

. The diagram

$  B!  C &  &  DQ & ) &  &  &  &  )8 C 4$ 6 &    C  &  5 I & 8 $6 ) B 5

 C    C   &  %B    C  C    &  &     8  ) $B6 5 &   C & & &


8 $6 ) # 5

8 $ ) #6 5

& 

"

"

Least homomorphisms compose: If .

$  B! C   C  & 


$  B) ) 4C$ 6 5 &  C! 8

-algebra . It equals the least xpoint of the which is dened by

is also an isomorphism, then

C C $  4! 

8 $6 ) B 5

&

"

"

 C   &  

&

&

&

&

4.

Proof. (1 2) The least endomorphism on is calculated as the least xpoint of . With the usual notation for the embeddings of into we get (by induction): and , where the last equality follows because and are mediating morphisms. Lemma 3.3.8 entails that the supremum of the is the identity. The equivalence of (2) and (3) is a reformulation of Lemma 5.3.1. (3 4) Suppose denes an endomorphism of the algebra . We apply the invariance property of least xpoints, Lemma 2.1.21, to the diagram (where now stands for )


. By Lemma 2.1.21 we have . (4 1) By the preceding lemma we have homomorphisms between and . They compose to the least endomorphisms on , resp. , which we know to be strict. But then they must be equal to the identity as we have just shown for and assumed for . If, in the last third of this proof, we do not assume that is the only strict endomorphism on , then we still get an embedding-projection pair between and . Thus we have: Theorem 5.3.3. (Minimality, Part 1) The canonical xpoint of a locally continuous functor is a sub-domain of every other xpoint. So we have shown that the canonical solution is the least xpoint in a relevant sense. This is clearly a good canonicity result with respect to the rst class of examples. For pedagogical reasons we have restricted attention to the covariant case rst, but, as we will see in section 5.3.3, this characterization is also true for functors of mixed variance. 5.3.2 Initiality and nality By a little renement of the proofs of the preceding subsection we get the desired result that the canonical xpoint together with is an initial -algebra. One of the adjustments is that we have to pass completely to strict functions, because Lemma 5.3.1 76

where maps to be strict. The diagram commutes:



8 B 5 8 46 5 $6 $ &   $ C  #!  #$ 6 !       B  B C C   C               C  3 $ ! &    C   C  C  C  C  C Q#) C   C C D# )  B     B      8  #6 I  $ 5 8 B6 5 8 46 5 $ $

%

. This is a strict operation because is assumed to

  %

 

is the only strict endomorphism of .

does not guarantee the existence of strict homomorphisms and only of these can we prove unicity. Theorem 5.3.4. (Initiality) Let D D be a locally continuous functor on a category of domains with strict functions. Then is an initial algebra where is the canonical solution to . Proof. Let be a strict -algebra. The homomorphism we get from Lemma 5.3.1 is strict as we see by inspecting its denition. That there are no others is shown as in the proof of Theorem 5.3.2, (3 4). The relevant diagram for the application of Lemma 2.1.21 is now:
 

By dualizing Lemma 5.3.1 and the proof of Theorem 5.3.2, (3 4), we get the nal co-algebra theorem. It is slightly stronger than initiality since it holds for all co-algebras, not only the strict ones.
V S

5.3.3 Mixed variance

Let us now tackle the case that we are given an equation in which the variable occurs both positively and negatively in the body, as in our rst example . We assume that by separating the negative occurrences from the positive ones, we have a functor in two variables, contravariant in the rst and covariant in the second. As the reader will remember, solving such an equation required the somewhat magical passage to adjoints in the rst coordinate. We will now see in how far we can extend the results from the previous two subsections to this case. Note that for a mixed variant functor the concept of -algebra or co-algebra is no longer meaningful, as there are no homomorphisms. The idea is to pass to pairs of mappings. Lemma 5.3.1 is replaced by Lemma 5.3.6. Let D D D be a mixed variant, locally continuous functor and let and be isomorphisms. Then there exists a least pair of functions and such that

77

"

and

8 $ #6 5 A

P C  P C  P C  $  $  ! P

"

"

"

"

"

"

P C  B C  P C P C $ !! C $ 3! H EF $ C  P 3! 

"

"

"

Theorem 5.3.5. (Finality) Let ical xpoint . Then

8 B6 $ 8  $6 5 97#% 5 $ 4! C ! 4$ 6   3C) C  $  C A   CD! $ (C !

D


D be a locally continuous functor with canonis a nal co-algebra.

$6 8 4% 5

&

C   7$  ! $ C !

8  $6 97#% 5

&
"

 C 

commute. The composition of two such least pairs gives another one. Proof. Dene a Scott-continuous function on by and let be its least xpoint. Commutativity of the two diagrams is shown as in the proof of Lemma 5.3.1. and in this lemma, we get a least endofunction which satises . Again, it must be below the identity. Let us call such endofunctions mixed endomorphisms. Theorem 5.3.7. (Invariance, Part 2) Let locally continuous functor and let following are equivalent: 1. 2. 3. 4.
"

is isomorphic to the canonical xpoint

is the least mixed endomorphism of ; where ;



is dened by

is the only strict mixed endomorphism of .

Proof. The proof is of course similar to that of Theorem 5.3.2, but let us spell out the parts where mixed variance shows up. Recall from Seclooks like: tion 5.2.2 how the expanding sequence dening . If are the colform the cocone into , imiting maps into , then which, by local continuity, is also colimiting. The equations from Lemma 5.2.2 read: and . We show that the -th approximation to the least mixed endomorphism equals . For we get , and for the induction step:

(Note how contravariance in the rst argument of shufes and in just the right way.) 4) The diagram to which Lemma 2.1.21 is applied is as before, but (3 now maps to . The rest can safely be left to the reader. Theorem 5.3.8. (Minimality, Part 2) The canonical xpoint of a mixed variant and locally continuous functor is a sub-domain of every other xpoint. 78
%

$ ! 8 46 $ 8 $ $ 5 B" #6 5 !   C    P  C  P  C      P    C                   P  C   P  C   P  C P P  P  PC  P  C   Q' C P  P    P PC   P C C  &   P C  C 

 & 

 

8 # 5 $ 8 # 5 ! &  $6 $6  C $ H $ GFE  DC C !  !P

 

 

 

 B C  P  & 

 

 P C
%

By equating

P 8 $6 $ 6 P  #  " # 5 H 8 # 5 
D D ;

D be a mixed variant and be an isomorphism. Then the

"

"

P P  # C  P   # C


Now that we have some experience with mixed variance, it is pretty clear how to deal with initiality and nality. The trick is to pass once more to pairs of (strict) functions. Theorem 5.3.9. (Free mixed variant algebra) Let D D D be a mixed variant, locally continuous functor and let be the canonical solution to . Then for every pair of strict continuous functions
"

commute.

We should mention that the passage from covariant to mixed-variant functors, which we have carried out here concretely, can be done on an abstract, categorical level as was demonstrated by Peter Freyd in [Fre91]. The feature of domain theory which Freyd uses as his sole axiom is the existence and coincidence of initial algebras and nal co-algebras for all endofunctors (all to be interpreted in some suitable enriched sense, in our case as all locally continuous endofunctors). Freyds results are the most striking contribution to date towards Axiomatic Domain Theory, for which see 8.4 below.

5.4 Analysis of solutions

We have worked hard in the last section in order to show that our domain theoretic solutions are canonical in various respects. Besides this being reassuring, the advantage of canonical solutions is that we can establish proof rules for showing properties of them. This is the topic of this section. 5.4.1 Structural induction on terms This technique is in analogy with universal algebra. While one has no control over arbitrary algebras of a certain signature, we feel quite comfortable with the initial or term algebra. There, every element is described by a term and no identications are made. The rst property carries over to our setting quite easily. For each of the nitary constructions of Section 3.2, we have introduced a notation for the basis elements of the , variants , one-element constant , constructed domain, to wit, tuples and step-functions . Since our canonical solutions are built as bilimits, starting from , and since every basis element of a bilimit shows up at a nite iteration already, Theorem 3.3.11, these can be denoted by nite expressions. The proof can then be based on structural induction on the length of these terms. 79

"

 ! 

and

 P C  P C  P  C
"

B 0

P 3

 P  C  B C  IP C P
S

  

"

"

and such that

"

there are unique strict functions

$  #6 3! @46 !  $ C  P @#$ 6 !  $ H F E DC !
and
"

B$ 6  P P C !C A 

Unicity, however, is hard to achieve and this is the fault of the function space. One has to dene normal forms and prove conversion rules. A treatment along these lines, based on [Abr91b], is given in Chapter 7.3. 5.4.2 Admissible relations This is a more domain-theoretic formulation of structural induction, based on certain relations. The subject has recently been expanded and re-organized in an elegant way by Andrew Pitts [Pit93b, Pit94]. We follow his treatment closely but do not seek the same generality. We start with admissible relations, which we have met shortly in Chapter 2 already. Denition 5.4.1. A relation on a pointed domain is called admissible if it contains the constantly-bottom tuple and if it is closed under suprema of -chains. We write for the set of all admissible -ary relations on , ordered by inclusion. Unary relations of this kind are also called admissible predicates. This is tailored to applications of the Fixpoint Theorem 2.1.19, whence we preferred the slightly more inclusive concept of -chain over directed sets. If we are given a strict continuous function , then we can apply it to relations pointwise in the usual way:

Proposition 5.4.2. For dcpos and and admissible -ary relations on on the set is an admissible predicate on .
#


and

We also need to say how admissible relations may be transformed by our locally continuous functors. This is a matter of denition because there are several and equally useful possibilities. Denition 5.4.3. Let D D D be a mixed variant and locally continuous functor on a category of domains and strict functions. An admissible action on ( -ary) relations for is given by a function which assigns to each pair a map from to . These maps have to be compatible

(Admittedly, this is a bit heavy in terms of notation. But in our concrete examples it is simply not the case that the behaviour of on and is the same as or in a simple way related to the result of applying the functor to and viewed as dcpos.) Specializing and to identity mappings in this denition, we get:

 &

Proposition 5.4.4. The maps ond variable.

are antitone in the rst and monotone in the sec-

80

 &

and


with strict morphisms in D etc., such that

as follows: If

and , then

and if

P  P & C   P & C  # C          4$ 6 ! 74$ 6 4! ' C H  & C P  Q C   C P  ! $ H F E C

8 #6  $     5     P P     P P    

B %

#$ 6 4!


@ (

 


Theorem 5.4.5. Let D be a category of domains and let be a mixed variant and locally continuous functor from D D to D together with an admissible action on relations. Abbreviate by . Given two admissible relations such that

Proof. We know from the invariance theorem that the identity on is the least xpoint of , where . Let , which we know is an admissible predicate. We want that the identity on belongs to and for this it sufces to show that maps into itself. So suppose : by denition by assumption because by assumption

Indeed,

belongs again to

In order to understand the power of this theorem, we will study two particular actions in the next subsections. They, too, are taken from [Pit93b]. 5.4.3 Induction with admissible relations Denition 5.4.6. Let be a mixed variant functor as before. We call an admissible action on ( -ary) relations logical, if for all objects and and we have .
# #

action, we get:

D Theorem 5.4.7. (Induction) Let D be a category of domains and let D D be a mixed variant and locally continuous functor together with a logical action on admissible predicates. Let be the canonical xpoint of . If is an admissible predicate, for which implies , then must be equal to . The reader should take the time to recognize in this the principle of structural induction on term algebras. We exhibit a particular logical action on admissible predicates for functors which are admissible predicates on the are built from the constructors of Section 3.2. If pointed domains and , then we set

81

     C H F E (C !

P    P    8  $46   '   P  I P     H   5 P   P     

 P  C   

B 6

@ W

$ 8 H 46 5

Specializing
# V

to be the whole in Theorem 5.4.5 and removing the assumption , which for this choice of is always satised for a logical

  

 P C    P  C  P & C

then

holds.

and

   P
%

   I Q P P C  Q P C  PB C    B C      I    P   897#$ 6  5 '   # C    Q P C  P C  

 C H F E 
S #

and analogously for and . (This is not quite in accordance with our notational convention. For example, the correct expression for is .) The denition of the action for the function space operator should make it clear why we chose the adjective logical for it. We get more complicated functors by composing the basic constructors. The actions also compose in a straightforward way: If , , and are mixed variant functors on a category of domains then we can dene a mixed variant composition by setting for objects and similarly for morphisms. Given admissible actions for each of , , and , we can dene an action for by setting . It is an easy exercise to show that this action is logical if all its constituents are. 5.4.4 Co-induction with admissible relations In this subsection we work with another canonical relation on domains, namely the order relation. We again require that it is dominant if put in the covariant position. Denition 5.4.8. Let be a mixed variant functor. We call an admissible action on binary relations extensional, if for all objects and and we have . Theorem 5.4.9. (Co-induction) Let D be a category of domains and let D D D be a mixed variant and locally continuous functor together with an extensional action on binary relations. Let be the canonical xpoint of . If is an admissible relation such that for all we have , then is contained in . If we call an admissible binary relation on a simulation, if it satises the hypothesis of this theorem, then we can formulate quite concisely: Corollary 5.4.10. Two elements of the canonical xpoint of a mixed variant and locally continuous functor are in the order relation if and only if they are related by a simulation. We still have to show that extensional actions exist. We proceed as in the last subsection and rst give extensional actions for the primitive constructors and then rely on the fact that these compose. So let be admissible binary relations on , resp. . We set:
#

or

82

#  B

and and

or

6 @

and

B 6

or

$ $  P & 8 B6 5 8 #6 5

 P P  C P Q P P  P 

 

  P   P      P   P       $ 6   P     P         BP 8 #% 5   QP #P  H   P  P P   P      P 

 P 

B 6

C  P C  P

 P C     P    ! $ H GFE DC  P & C

B 6B

8 #$ 6 5
B B 6 B 2 B 2 B @

@ @ W

 &

@ W

@ W

@ W

B R

$ 8 46 5 H

P @

C
# #

and similarly for and . We call this family of actions extensional because the denition in the case of the function space is the same as for the extensional order on functions. Exercises 5.4.11. 1. Find recursive domain equations which characterize the three versions of the natural numbers from Figure 2. 2. [Ern85] Find an example which demonstrates that the ideal completion functor is not locally continuous. Characterize the solutions to . 3. [DHR71] Prove that only the one-point poset satises

be pointed dcpos and let 4. Verify Beki s rule in the dcpo case. That is, let c and be continuous functions. We can solve the equations
B 2 @

directly by taking the simultaneous xpoint for one variable at a time by dening
6

and setting Verify that

holds by using xpoint induction.

5. Find an example which shows that the Initiality Theorem 5.3.4 may fail for nonstrict algebras. 6. Why does Theorem 5.3.5 hold for arbitrary (non-strict) co-algebras?

and let be its 8. (G. Plotkin) Let be the functor which maps to canonical xpoint. This gives rise to a model of the (lazy) lambda calculus (see [Bar84, Abr90c, AO93]). Prove that the denotation of the Y combinator in this model is the least xpoint function . Proceed as follows:

9. Given an action on relations for a functor in four variables, contravariant in the rst two, covariant in the last two, dene an action for the functor . Prove that the resulting action is logical (extensional) if the original action was logical (extensional).

P Q P P ' C $  P        8  $    97B6  5    $       

holds.

83

(d) Use Theorem 5.3.7 to show that

. Then

and

is also valid.

(a) Dene a multiplication on by of Y is where (b) The interpretation Check that this is a xpoint of . It follows that (c) Dene a subset of by

7. What are initial algebra and nal co-algebra for the functor category of sets? Show that they are not isomorphic as algebras.

8 4$ 6 5 A  P A
V 6

8 4$6 5 $

P  QP  #       P    P  32  P  P#  #  P   P  

P 

H $ % ! 7$ %3)  H !

8 #$ 6 5

. Or we can solve

on the

6 Equational theories
In the last chapter we saw how we can build initial algebras over domains. It is a natural question to ask whether we can also accommodate equations, i.e. construct free algebras with respect to equational theories. In universal algebra this is done by factoring the initial or term algebra with respect to the congruence generated by the dening equations, and we will see that we can proceed in a similar fashion for domains. Bases will play a prominent role in this approach. The technique of the previous chapter, namely, to generate the desired algebra in an iterative process, is no longer applicable. A formal proof for this statement may be found in [AT89], Section III.3, but the result is quite intuitive: Recall that an algebra encodes the algebraic structure on by giving information about the basic operations on , where is the sum of the input domains for each basic operation. Call an equation at if each of the equated terms contains precisely one operation symbol. For example, commutativity of a binary operation is expressed by a at equation while associativity is not. Flat equations can be incorporated into the concept of -algebras by including the input, on which the two operations agree, . For non-at equations such a trick is not available. What we need only once in instead of just the basic operations is a description of all term operations over . In this case, will have to be the free algebra over , the object we wanted to construct! Thus -algebras are not the appropriate categorical concept to model equational theories. The correct formalization, rather, is that of monads and Eilenberg-Moore algebras. We will show the existence of free algebras for dcpos and continuous domains in the rst section of this chapter. For the former, we use the Adjoint Functor Theorem (see [Poi92], for example), for the latter, we construct the basis of the free algebra as a quotient of the term algebra. Equational theories come up in semantics when non-deterministic languages are studied. They typically contain a commutative, associative, and idempotent binary operation, standing for the union of two possible branches a program may take. The associated algebras are known under the name powerdomains and they have been the subject of detailed studies. We shall present some of their theory in the second section.

6.1 General techniques

6.1.1 Free dcpo-algebras Let us recall the basic concepts of universal algebra so as to x the notation for this chapter. A signature consists of a set of operation symbols and a map , assigning to each operation symbol a (nite) arity. A -algebra is given by a carrier set and an interpretation of the operation symbols, in the sense that for , is a map from to . We also write or even for the interpreted operation symbol and speak of the operation on . A homomorphism between two -algebras and is a map which commutes with the operations:
B W

84

P

     P P   Q    P P  ' $ !

"

 C


"

  

$  C!  C C

C  C

$ !

We denote the term algebra over a set with respect to a signature by . It has the universal property that each map from to , where is a algebra, can be extended uniquely to a homomorphism . Let be a xed countable set whose elements we refer to as variables. Pairs of elements of are used to encode equations. An equation is said to hold in an algebra if for each map we have . The pair is also called an instance of the equation . The class of -algebras in which each equation from a set holds, is denoted by Set . Here we are interested in dcpo-algebras, characterized by the property that the carrier set is equipped with an order relation such that it becomes a dcpo, and such that each operation is Scott-continuous. Naturally, we also require the homomorphisms to be Scott-continuous. Because of the order we also can incorporate inequalities. So from now on we let a pair stand for the inequality . We use the notation DCPO for the class of all dcpo-algebras over the signature which satisfy the inequalities in . For these we have: Proposition 6.1.1. For every signature and set of inequalities, the class with Scott-continuous homomorphisms forms a complete category. DCPO Proof. It is checked without difculties that DCPO is closed under products and equalizers, which both are dened as in the ordinary case. This proves that we have one ingredient for the Adjoint Functor Theorem, namely, a complete category DCPO and a (forgetful) functor DCPO DCPO which preserves all limits. The other ingredient is the so-called solution set condition. For this setup it says that each dcpo can generate only set-many non-isomorphic dcpoalgebras. This is indeed the case: Given a dcpo and a continuous map , where is the carrier set of a dcpo-algebra , we construct the dcpo-subalgebra of generated by in two stages. In the rst we let be the (ordinary) subalgebra of which is generated by . Its cardinality is bounded by an expression depending on the cardinality of and . Then we add to all suprema of directed subsets until we get a sub-dcpo of the dcpo . Because we have required the operations on to be Scott-continuous, remains to be a subalgebra. The crucial step in this argument now is that the cardinality of is bounded by as we asked you to show in Exercise 2.3.9(33). All in all, given , the cardinality of has a bound depending on and so there is only room for a set of different dcpo-algebras. Thus we have shown: Theorem 6.1.2. For every signature and set of inequalities, the forgetful functor DCPO DCPO has a left adjoint. Equivalently: For each dcpo the free dcpo-algebra over with respect to and exists. The technique of this subsection is quite robust and has been used in [Nel81] for proving the existence of free algebras under more general notions of convergence than that of directed-completeness. This, however, is not the direction we are interested in, and instead we shall now turn to continuous domains.

85

P  )  P 

 

$4!  $  ) P

B (


P

 

H    $ !  
) P  H     P

! 

P  )

P  )

      

$  ) P

P  )

B W

P 

 

U@

! 

6.1.2 Free continuous domain-algebras None of the categories of approximated dcpos, or domains, we have met so far is complete. Both innite products and equalizers may fail to exist. Hence we cannot rely on the Adjoint Functor Theorem. While this will result in a more technical proof, there will also be a clear advantage: we will gain explicit information about the basis of the constructed free algebra, which may help us to nd alternative descriptions. In the case of dcpos, such concrete representations are quite complicated, see [Nel81, ANR82]. We denote the category of dcpo-algebras, whose carriers form a continuous doand speak of (continuous) domain-algebras. Again there is the main, by CONT obvious forgetful functor CONT CONT. To keep the notation manageable we shall try to suppress mention of , in particular, we will write for on objects and make no distinction between and on morphisms. Let us write down the condition for adjointness on which we will base our proof:

In words: Suppose a signature and a set of inequalities has been xed. Then given a continuous domain we must construct a dcpo-algebra , whose carrier set is a continuous domain, and a Scott-continuous function such that satises the inequalities in and such that given any such domainalgebra and Scott-continuous map there is a unique Scott-continuous homomorphism for which . (It may be instructive to compare this with Denition 3.1.9.) The idea for solving this problem is to work explicitly with bases (cf. Section 2.2.6). So assume that we have xed a basis for the continuous domain . We will construct an abstract basis for the desired free domain-algebra . The underlying set is given by the set of all terms over . On we have two natural order relations. The rst, which we denote by , is induced by the dening set of inequalities. We can give a precise denition in the form of a deduction scheme: Axioms:
"

Rules: (R1) If (R2) If

and

then

86

is an

%

-ary function symbol and if .

A P P A

A A A A  P P  P  P

(A2)

if this is an instance of an inequality from .

(A1)

for all

  C $  !   C

then

C   C

"

"

 $ #! $   #C! 

 P C

IU"@ P

"

B C

CONT

CONT

P  )

   C
G

$  ) P

   C 

"

!  
@

C 

P  )
G

"

  C   C C

The relation is the least substitutive preorder in the terminology of [Sto88]. It is the obvious generalization of the concept of a congruence relation to the preordered case, and indeed, is the free preordered algebra over . The associated equivalence relation we denote by . The factor set is ordered by and this is the free ordered algebra over . Let us now turn to the second relation on , namely, the one which arises from the order of approximation on . We set if and have the same structure and corresponding constants are related by . Formally, is given through the deduction scheme: Axioms:
"

Rules:

Proof. Since relates only terms of the same structure, it is quite obvious that it holds for all is a transitive relation. For the interpolation axiom assume that elements of a nite set . For each occurrence of a constant in let be the set of constants which occur in the same location in one of the terms . Since is nite and since holds by the denition of , we nd interpolating elements between and . Let be the term in which all constants are replaced by the corresponding interpolating element. This is a term which interpolates between and in the relation . The question now is how to combine and . As a guideline we take Proposition 2.2.2(2). If the inequalities tell us that should be below and should be below and if approximates then it should be the case that approximates . Hence we dene , the order of approximation on , to be the transitive closure of . The following, somewhat technical properties will be instrumental for the free algebra theorem:

Proof. (1) Assume . Let be the set of all constants which appear in the derivation of . For each let be the set of constants which appear in at the same place as appears in . Of course, may not occur in at all; in this case will be empty. If it occurs several times then can contain more than one element. In any case, is nite and holds. Let be an interpolating element between and . We now replace each constant in the derivation of by the corresponding constant and we get a valid derivation of a formula . (The 87

AA

   

"

A A

 

2. For every

we have

 A

 A 

Proposition 6.1.4.

1.

is contained in

. .

   

"

"

 

C   

"

Proposition 6.1.3.

is an abstract basis.

Our rst observation is that

satises the interpolation axiom:

 P P

 %P

(R) If

is an -ary function symbol and if .

%P

 P

(A)

if

in

"

"

"

"

"

AI C P
"

then

 
A

catch is that an instance of an inequality is transformed into an instance of the same inequality.) It is immediate from the construction that holds. (2) Using (1) and the reexivity of we get

The general case follows by induction.

B C " @

Lemma 6.1.5.

is an abstract basis.

Proof. Transitivity has been built in, so it remains to look at the interpolation axiom. Let for a nite set . From the denition of we get for each a sequence of terms . The last two steps as we have shown in the preceding may be replaced by proposition. The collection of all is nite and we nd an interpolating term between it and according to Proposition 6.1.3. Because of the reexivity of we have . So we can take as the carrier set of our free algebra over the ideal completion of and from Proposition 2.2.22 we know that this is a continuous domain. The techniques of Section 2.2.6 also help us to ll in the remaining pieces. The operations on are dened pointwise: If are ideals and if is an -ary function symbol then we let be the ideal which is generated by We need to know that this set is directed. It will follow if the operations on are monotone with respect to . So assume we and pairs . By denition, are given an operation symbol each pair translates into a sequence . Now we use Proposition 6.1.4(2) to extend all these sequences to the same length . Then we can apply step by step, using Rules (R1) and (R) alternately:

Using the remark following Proposition 2.2.24 we infer that the operations dened this way are Scott-continuous functions. Thus is a continuous domainvia the extension of the monoalgebra. The generating domain embeds into tone inclusion of into . Theorem 6.1.6. and .

is the free continuous domain-algebra over

with respect to

Proof. We must show that satises the inequalities in and that it has the universal property. For the inequalities let and let be a map. It assigns to each variable an ideal in . We must show that is a subset of . As we have just seen, the ideal is generated by terms of the form where is a map from to , such that for each variable , . So suppose for such a . Then is an instance of the inequality in the term algebra 88


A     '   C         & $ ! C  )P C   C    C  C  C   C    A  P P A  P P P P P P A P P    A  A A  P C P    P P    P P PP P P    C



 A


A A

 A

A A

"

B A

  

"

 

"

"

% "

B C

P I"

and so we know that also holds. The term belongs to , again because the operations on are dened pointwise. So as desired. For the universal property assume that we are given a continuous map for a dcpo-algebra which satises the inequalities from . The restriction of to the has a unique monotone extension to the preordered algebra . We set want to show that also preserves . For an axiom this is clear because is monotone on . For the rules (R) we use that is a homomorphism and that the operations on are monotone:
B

Together this says that preserves the order of approximation on and therefore it can be extended to a homomorphism on the ideal completion . Uniqueness of is obvious. What we have to show is that , when restricted to , equals , because Proposition 2.2.24 does not give an extension but only a best approximation. We can nevertheless prove it here because arose as the restriction of a continuous map on . An element of is represented in as the ideal containing at least all of because of the axioms of our second deductive system. So we have: . Theorem 6.1.7. For any signature and set of inequalities the forgetful functor CONT CONT has a left adjoint . It is equivalent to the restriction and corestriction of the left adjoint from Theorem 6.1.2 to CONT and CONT , respectively. In other words: Free continuous domain-algebras exist and they are also free with respect to dcpo-algebras. The action of the left adjoint functor on morphisms is obtained by assigning to a continuous function the homomorphism which extends .
"

We want to show that is locally continuous (Denition 5.2.3). To this end let us rst look at the passage from maps to their extension.

Proposition 6.1.8. The assignment is Scott-continuous.

, as a map from

89

P  )

AP C $ 3!    
8 #% 5 $6

  C C    P P P P  P P  Q P P

"

         Q             C

"

Q 

"

 C  C   C

  C 
" Q

$ 3!

 !"

$  ) P

P I 

" U@

8 #6   C 5 $

  C
"

"

"

! 

"

to

Proof. By Proposition 2.2.25 it is sufcient to show this for the restriction of to the basis of . Let be a directed collection of monotone maps from to and let be a term in which the constants occur. We calculate:

where we have written for the term in which each occurrence of is replaced by . Restriction followed by homomorphic extension followed by extension to the ideal completion gives a sequence of continuous functions which equals . Cartesian closed categories can be viewed as categories in which the -functor can be internalized. The preceding proposition formulates a similar closure property of the free construction: if the free construction can be cut down to a cartesian closed category then there the associated monad and the natural transformations that come with it can be internalized. This concept was introduced by Anders Kock [Koc70, Koc72]. It has recently found much interest under the name computational monads through the work of Eugenio Moggi [Mog91]. Theorem 6.1.9. For any signature and set is a locally continuous functor on CONT. Proof. The action of and .

of inequalities the composition

on morphisms is the combination of composition with

If is an embedding then we can describe the action of , respectively , quite concretely. A basis element of is the equivalence class of some term . Its image under is the equivalence class of the term , which we get from by replacing all constants in by their image under . If we start out with an algebraic domain then we can choose as its basis , the set of compact elements. The order of approximation on is the order relation inherited from , in particular, it is reexive. From this it follows that the constructed is also reexive, whence the ideal completion of order of approximation on is an algebraic domain. This gives us: Theorem 6.1.10. For any signature and set from ALG to ALG has a left adjoint.

of inequalities the forgetful functor

Finally, let us look at , which maps the generating domain into the free algebra, and let us study the question of when it is injective. What we can say is that if injectivity fails then it fails completely: Proposition 6.1.11. For any in-equational theory the canonical map from a dcpo into the free algebra over is order-reecting if and only if there exists a dcpoalgebra for this theory for which the carrier dcpo is non-trivially ordered.

90

 

8 $ " #6   C 5 $ 8 #% 5 $6 8  P P P  8    P P    5 8     P P      5    P P  

% 

"

 

  C

" 

  C

'

'

"

 5

1  


$ 8 B$ 6 C 5 $ 8 #$ 6 5

  C

"

$ C  !

P  )

B C

P I"

"

"

C


"

Proof. Assume that there exists a dcpo-algebra which contains two elements . Let be any dcpo and two distinct elements. We can dene a continuous map from to , separating from by setting if otherwise.

Since equals , where is the unique homomorphism from it cannot be that holds. The converse is trivial, because must be monotone. 6.1.3 Least elements and strict algebras

to ,

We have come across strict functions several times already. It therefore seems worthwhile to study the problem of free algebras also in this context. But what should a strict algebra be? There are several possibilities as to what to require of the operations on such an algebra: 1. An operation which is applied to arguments, one of which is bottom, returns bottom. 2. An operation applied to the constantly bottom vector returns bottom. 3. An operation of arity greater than 0 applied to the constantly bottom vector returns bottom. Luckily, we can leave this open as we shall see shortly. All we need is: Denition 6.1.12. A strict dcpo-algebra is a dcpo-algebra for which the carrier set contains a least element. A strict homomorphism between strict algebras is a Scottcontinuous homomorphism which preserves the least element. For pointed dcpos the existence of free strict dcpo-algebras can be established as before through the Adjoint Functor Theorem. For pointed domains the construction of the previous subsection can be adapted by adding a further axiom to the rst deduction scheme:
"

Thus we have:

Theorem 6.1.13. Free strict dcpo- and domain-algebras exist, that is, the forgetful functors

have left adjoints.

and

ALG

ALG

91

DCPO CONT

DCPO CONT

$46 $46 $ 46

P  P)  P)  )

(A3)

for all

  C

     #PP   



C



Let us return to the problem of strict operations. The solution is that we can add a nullary operation to the signature and the inequality to without changing the free algebras. Because of axiom (A3) we have and because of the new inequality we have . Therefore the new operation must be interpreted by the bottom element. The advantage of having bottom explicitly in the signature is that we can now formulate equations about strictness of operations. For example, the rst possibility mentioned at the beginning can be enforced by adding to the inequality

for all operation symbols of positive arity and all . The corresponding free algebras then exist by the general theorem. More problematic is the situation with DCPO (respectively CONT and ALG ). The existence of a least element in the generating dcpo does not imply the existence of a least element in the free algebra (Exercise 6.2.23(2)). Without it, we cannot make use of local continuity in domain equations. Furthermore, even if the free algebra has a least element, it need not be the case that is strict (Exercise 6.2.23(3)). The same phenomena appears if we restrict attention to any of the cartesian closed categories exhibited in Chapter 4. The reason is that we require a special structure of the objects of our category but allow morphisms which do not preserve this structure. It is therefore always an interesting fact if the general construction for a particular algebraic theory can be restricted and corestricted to one of these sub-categories. In the case that the general construction does not yield the right objects it may be that a different construction is needed. This has been tried for the Plotkin powerdomain in several attempts by Karel Hrbacek but a satisfactory solution was obtained only at the cost of changing the morphisms between continuous algebras, see [Hrb87, Hrb89, Hrb88]. On a more positive note, we can say: Proposition 6.1.14. If the free functor maps nite pointed posets to nite pointed posets then it restricts and corestricts to binite domains.

6.2 Powerdomains
6.2.1 The convex or Plotkin powerdomain Denition 6.2.1. The convex or Plotkin powertheory is dened by a signature with one binary operation and the equations

3.

(Idempotence)

The operation is called formal union. A dcpo-algebra with respect to this theory is called a dcpo-semilattice. The free dcpo-semilattice over a dcpo is called the Plotkin powerdomain of and it is denoted by .

92

 

2.

1.

(Commutativity) (Associativity)

      P P QP P QP P 

 

           

Every semilattice can be equipped with an order by setting

"

Formal union then becomes the join in the resulting ordered set. On a dcpo-semilattice this order has little to do with the domain ordering and it is not in the focus of our interest. The free semilattice over a set is given by the set of all non-empty nite subsets of , where formal union is interpreted as actual union of sets. This gives us the rst half of an alternative description of the Plotkin powerdomain over a continuous domain with basis . Its basis , which we constructed as the term algebra over , is partitioned into equivalence classes by , the equivalence relation derived from , that is, from the dening equations. These equivalence classes are in one-toone correspondence with nite subsets of . Indeed, given a term from , we can re-arrange it because of associativity and commutativity, and because of idempotence we can make sure that each constant occurs just once. as the transitive Remember that we have set up the order of approximation on closure of . This way we have ensured that an ideal in contains only full equivalence classes with respect to . We may therefore replace by , the set of nite subsets of , where we associate with a term the set of constants appearing in . Let us now also transfer the order of approximation to the new basis.

Here we are talking about nite subsets of , so we write for the EgliMilner relation between nite subsets of . Let us establish the connection between on and on . Firstly, if then by denition each constant in is matched by a constant in which approximates it and vice versa. These are just the conditions for . Since is transitive, we nd that implies in general. Conversely, if two nite subsets and of are related by then we can build terms and , such that , , and hold. This is done as follows. For each let be an element of such that and for each let be an element of such that . Then we can set

We have proved:

Theorem 6.2.3. The Plotkin powerdomain of a continuous domain is given by the ideal completion of . 93
B

P

and

with basis

)     (   P P  

(

           

IU"@ P

 ( 

Denition 6.2.2. Two subsets and the Egli-Milner relation, written as satised:

of a set equipped with a relation are in , if the following two conditions are

8  5

"

"

"

"

CC  CC

"

"

      '  )     



"

if

   (     (     

"

"

 

 (

"

8 5

"

" "

5 8

  8 5 P 8 8  P  5 8 5 5

"

 (

"

IU"@ P

An immediate consequence of this characterization is that the Plotkin powerdomain of a nite pointed poset is again nite and pointed. By Proposition 6.1.14, the Plotkin powerdomain of a binite domain is again binite. This is almost the best result we can obtain. The Plotkin power construction certainly destroys all properties of being lattice-like, see Exercise 6.2.23(8). It is, on the other hand, not completely haphazard, in the sense that not every nite poset is a sub-domain of a powerdomain of some other poset. This was shown in [N 92]. u The passage from terms to nite sets has reduced the size of the basis for the powerdomain drastically. Yet, it is still possible to get an even leaner representation. We present this for algebraic domains only. For continuous domains a similar treatment is possible but it is less intuitive. Remember that abstract bases for algebraic domains are preordered sets.

A set which coincides with its convex hull is called convex.

Comment: In (5) and (6) we have used the notation as an abbreviation for ; -version of equality as dened in 6.2.2 (which would be nothing more than equality on it is not the the powerset).

is only a preordered set, parts (5) and (6) of the precedWhile ing proposition suggests how to replace it with an ordered set. Writing Cx for the set of nitely generated convex subsets of , we have: Proposition 6.2.6. The Plotkin powerdomain of an algebraic domain to the ideal completion of Cx .
B

is isomorphic

This explains the alternative terminology convex powerdomain. We will sharpen this description in 6.2.3 below. For examples of how the Plotkin powerdomain can be used in semantics, we refer to [HP79, Abr91a].

94

 

&

 

P Q 

&

P  

 

6.

5.

if and only if

4.

 

3.

2.

1.

"

                        

B C

" U@

Proposition 6.2.5. Let

be a preordered set and

The following properties are easily checked:


be subsets of .

B C


" @

P      


Denition 6.2.4. For a subset be dened by

of a preordered set

let the convex hull

 

"

P I

6.2.2 One-sided powerdomains Denition 6.2.7. If the Plotkin powertheory is augmented by the inequality

then we obtain the Hoare or lower powertheory. Algebras for this theory are called inationary semilattices. The free inationary semilattice over a dcpo is called the lower or Hoare powerdomain of , and it is denoted by . Similarly, the terminology concerning the inequality

It is a consequence of the new inequality that the semilattice ordering and the domain ordering coincide in the case of the Hoare powertheory. For the Smyth powertheory the semilattice ordering is the reverse of the domain ordering. This forces these powerdomains to have additional structure. Proposition 6.2.8. 1. The Hoare powerdomain of any dcpo is a lattice which has all non-empty suprema and bounded inma. The sup operation is given by formal union. 2. The Smyth powerdomain of any dcpo has binary inma. They are given by formal union. Unfortunately, the existence of binary inma does not force a domain into one of the cartesian closed categories of Chapter 4. We take up this question again in the next subsection. Let us also study the bases of these powerdomains as derived from a given basis of a continuous domain . The development proceeds along the same lines as for the Plotkin powertheory. The equivalence relation induced by the equations and the new inequality has not changed, so we may again replace by the set of . nite subsets of . The difference is wholly in the associated preorder on
B

if and only if

if and only if

for the Smyth powertheory.

The restricted order of approximation is as before given by the Egli-Milner relation . As prescribed by the general theory we must combine it with inclusion (for the lower theory) and with reversed inclusion (for the upper theory), respectively. Without difculties one obtains the following connection

95

if and only if

 8 5



in the case of the Hoare powertheory and

 

 ) 8 5  

Proposition 6.2.9. For

and

nite subsets of a basis

we have

 

"

"

"

is upper or Smyth powerdomain, deationary semilattice, and

 

 

IU"@ P

 

   


"

IU"@ P

if and only if

So each of the one-sided theories is characterized by one half of the Egli-Milner ordering. Writing and for these we can formulate:
B " U@

Theorem 6.2.10. Let

be a continuous domain with basis

1. The Hoare powerdomain of . 2. The Smyth powerdomain of .

B

is isomorphic to the ideal completion of is isomorphic to the ideal completion of

and

Proposition 6.2.12. Let

be an algebraic domain.

From this description we can infer through Proposition 6.1.14 that the Smyth powerdomain of a binite domain is again binite. Since a deationary semilattice has binary inma anyway, we conclude that the Smyth powerdomain of a binite domain is actually a bc-domain. For a more general statement see Corollary 6.2.15. 6.2.3 Topological representation theorems The objective of this subsection is to describe the powerdomains we have seen so far directly as spaces of certain subsets of the given domain, without recourse to bases and the ideal completion. It will turn out that the characterizations of Proposition 6.2.6 and Proposition 6.2.12 can be extended nicely once we allow ourselves topological methods. 96

2. The Smyth powerdomain .

B

1. The Hoare powerdomain .

B C

of

is isomorphic to the ideal completion of is isomorphic to the ideal completion of

of

"

&

"

Writing for the set of nitely generated lower subsets of for the set of nitely generated upper subsets of , we have:
"

and

     

 P  

P    

"

&

&

&

4.

if and only if



3.

2.

if and only if

 



1.

B C

" U@

Proposition 6.2.11. For subsets

and

of a preordered set

P I

"

For algebraic domains we can replace the preorders on both cases.

  8 5   8 5 " 
P

and

 

P P

    

" "

@ @

by an ordered set in we have

Theorem 6.2.13. The Hoare powerdomain of a continuous domain is isomorphic to the lattice of all non-empty Scott-closed subsets of . Formal union is interpreted by actual union. Proof. Let be a basis for . We establish an isomorphism with the representation of Theorem 6.2.10. Given an ideal of nite sets in we map it to , the Scott-closure of the union of all these sets. Conversely, for a non-empty Scott-closed set we let , the set of nite sets of is inbasis elements approximating some element in . We rst check that deed an ideal with respect to . It is surely non-empty as was assumed to contain elements. Given two nite subsets and of then we can apply the interpolation axiom to get nite subsets and with and . An upper bound for and with respect to is then given by . It is also clear that the Scott closure of gives back again because every element of is the directed supremum of basis elements. Hence . Starting out with an ideal , we must show that we get it back from . So let . By the roundness of (see the discussion before Denition 2.2.21) there is another nite set with . So for each there is with . Since all elements of are contained in , we have that belongs to . Conversely, if is an element of then is not empty and therefore must meet as is closed. The set is then below some element of under the -ordering. Monotonicity of the isomorphisms is trivial and the representation is proved. Formal union applied to two ideals returns the ideal of unions of the constituting sets. Under the isomorphism this operation is transformed into union of closed subsets. This theorem holds not just for continuous domains but also for all dcpos and even all -spaces. See [Sch93] for this. We can also get the full complete lattice of all closed sets if we add to the Hoare powertheory a nullary operation and the equations

Alternatively, we can take the strict free algebra with respect to the Hoare powertheory. If the domain has a least element then these adjustments are not necessary, a least element for the Hoare powerdomain is . Homomorphisms, however, will only preserve non-empty suprema. The characterization of the Smyth powerdomain builds on the material laid out in Section 4.2.3. In particular, recall that a Scott-compact saturated set in a continuous domain has a Scott-open lter of open neighborhoods and that each Scott-open lter in arises in this way. Theorem 6.2.14. The Smyth powerdomain of a continuous domain is isomorphic to the set of non-empty Scott-compact saturated subsets ordered by reversed inclusion. Formal union is interpreted as union.

97

' 

" U@

Proof. Let
"

be a basis for . We show that . Given an ideal we let be

is isomorphic to . This constitutes a

 

 

  


 

  

        

"

"

 

 

" 



 



 



 F

 

             


"

 

 &



"



   
$

 I P  P

" @ V

  

monotone map from to by Proposition 4.2.14. In the other direction, we assign to a compact saturated set the set of all nite sets such that . Why is this an ideal? For every open neighborhood of we nd a nite set of basis elements contained in such that because is compact and (Proposition 2.3.6). Then given two nite sets and in an upper bound for them is any such nite set with . Clearly, is monotone as is equipped with reversed inclusion. Let us show that is the identity on . For let be above in the -ordering. Then and so belongs to . Conversely, every neighborhood of contains some with already as we saw in Proposition 4.2.14. So if is contained in for some nite set then there are and in with and . Hence and belongs to . The composition is clearly the identity as we just saw that every neighborhood of a compact set contains a nitely generated one and as every saturated set is the intersection of its neighborhoods. The claim about formal union follows because on powersets union and intersection completely distribute: . For this theorem continuity is indispensable. A characterization of the free deationary semilattice over an arbitrary dcpo is not known. The interested reader may consult [Hec90, Hec93a] and [Sch93] for a discussion of this open problem. Corollary 6.2.15. The Smyth powerdomain of a coherent domain with bottom is a bcdomain.
"

Let us now turn to the Plotkin powerdomain. An ideal of nite sets ordered by will generate ideals with respect to both coarser orders and . We can therefore associate with a Scott-closed set and a compact saturated set . However, not every such pair arises in this way; the Plotkin powerdomain is not simply the product of the two one-sided powerdomains. We will be able to characterize them in two special cases: for countably based domains and for coherent domains. The general situation is quite hopeless, as is illustrated by Exercise 6.2.23(11). In both special cases we do want to show that is faithfully represented by the intersection . In the rst case we will need the following weakening of the Egli-Milner ordering: Denition 6.2.16. For a dcpo we let be the set of all non-empty subsets of which arise as the intersection of a Scott-closed and a compact saturated subset. The elements of we call lenses. On we dene the topological EgliMilner ordering, , by

if

and

98

   

        

 

"

Proof. That two compact saturated sets ply means . In this case very denition of coherence.
 "  

and are bounded by another one, , simis not empty. It is compact saturated by the



" 

                                              

  10

'

"



&   I

' 

 

  

  10

 
'

&     '



  

 

      

 

 
0 1'
' 

 

  "  



' 

' 6

   

Proposition 6.2.17. Let

1. Every lens is convex and Scott-compact.

Proof. Convexity is clear as every lens is the intersection of a lower and an upper set. An open covering of a lens , where is closed and compact saturated, by adding the complement of to the cover. may be extended to a covering of This proves compactness. Since all Scott-open sets are upwards closed, compactness of a set implies the compactness of . Using convexity, we get and using boolean algebra we calculate and , so . Then if we have and . Equality of and follows. Before we can prove the representation theorem we need yet another description of the lens . Lemma 6.2.18. Let be a continuous domain with basis and let . Then directed and .
 ! " $ B "

. Proof. The elements of the set on the right clearly belong to the Scott-closure of They are also contained in because is above some element in for each . Conversely, let and let . The set is Scott-open and must therefore meet some . From the roundness of we get with . The set also approximates and so it is contained in . Hence . Furthermore, given any , let again be such that . Then is above some element of as and therefore holds for some . Theorem 6.2.19. Let be an -continuous domain. The Plotkin powerdomain is isomorphic to . Formal union is interpreted as union followed by topological convex closure. be a countable basis of . We have already dened the map . In the other direction we take the function which assigns to a lens the set . Before we can prove that these maps constitute a pair of isomorphisms, we need the following information about reconstructing and from . 1. : Since is an upper set which contains , only one inclusion can be in doubt. Let and . Firstly, each set in is non-empty and, secondly, we have whenever . Calculating in the continuous domain gives us a nonempty set which is below and contained in the lens .
X Q B " U@

Proof. Let

99

 

     

                                                   D    I $   !   P P                                        D                     $          P             

be an ideal in for all

3. The topological Egli-Milner ordering is anti-symmetric on

 

 !

 D        

2. A canonical representation for a lens

is given by

" 

be a dcpo.

. .

2. : Again, only one inclusion needs an argument. We show that every element of belongs to . Given a basis element approximating some element of then we already know that it belongs to . Let be some set which contains . Using countability of the basis we may assume that extends to a conal chain in (Proposition 2.2.13): . K nigs Lemma then tells us that we can nd a chain of elements o where . The supremum belongs to and is above . 3. is monotone: Let be two ideals in . The larger ideal results in a bigger lower set and a smaller upper set . Using 1 and 2 we can calculate for the corresponding lenses:


as desired. 4. The monotonicity of follows by construction and one half of the topological Egli-Milner ordering: implies if we assume . 5. : Given a lens we clearly have . Using the continuity of and the compactness of we infer that must equal . Every basis element approximating some element of occurs in some set of , so is clear. Proposition 6.2.17 above then implies that gives back . 6. : Given an ideal we know that each covers the lens in the sense of . So is contained in . By (2), we also have that is contained in . Conversely, if for a nite set of basis elements contained in , then for some we have by the Hofmann-Mislove Theorem 4.2.14. For this we have . On the other hand, each element of approximates some and hence belongs to some . in , therefore, is above in which shows An upper bound for and all that must belong to . 7. In the representation theorems for the one-sided powerdomains we have shown that formal union translates to actual union. We combine this for the convex setting:

Note that we used countability of the basis only for showing that can be recovered from . In general, this is wrong. Exercise 6.2.23(11) discusses an example. The substitution of topological closure for downward closure was also necessary, as the example in Figure 13 shows. There, the set is a lens but its downward closure is not Scott-closed, is missing. The set is also a lens. It is below in the topological Egli-Milner order but not in the plain Egli-Milner order. The convex closure of the union of the two lenses and is not a lens, must be added. A better representation theorem is obtained if we pass to coherent domains (Section 4.2.3). (Note that the example in Figure 13 is not coherent, because the set 100

         Q         ) Q       ) )     



 

 F

 !

             Q   )   Q    

So

                       Q                              Q    Q                            P Q  )              P                                     Q       

% ) ' 0(% B $ 5 "

   
%

" 

Figure 13: An algebraic domain in which topological Egli-Milner ordering and ordinary Egli-Milner ordering do not coincide. has innitely many minimal upper bounds, violating the condition in Proposition 4.2.17.) We rst observe that lenses are always Lawson closed sets. If the domain is coherent then this implies that they are also Lawson-compact. Compactness will allow us to use downward closure instead of topological closure. Lemma 6.2.20. Let is Scott-closed. be a Lawson-compact subset of a continuous domain

there Proof. Let be an element of which does not belong to . For each exists such that . The set is Lawson-open and contains . By compactness, nitely many such sets cover . Let be an upper bound for the associated basis elements approximating . Then is an open neighborhood of which does not intersect . Hence is closed. Corollary 6.2.21. The lenses of a coherent domain are precisely the convex Lawsoncompact subsets. For these, topological Egli-Milner ordering and Egli-Milner ordering coincide.

Proof. The differences to the proof of Theorem 6.2.19, which are not taken care of by the preceding corollary, concern part 2. We must show that contains all of . In the presence of coherence this can be done through the Hofmann-Mislove Theorem 4.2.14. The lower set is a continuous domain in itself. For an element of we look at the ltered collection of upper sets . Each of these is non-empty, because belongs to some , and compact saturated because of coherence. Hence is non-empty. It is also contained in and above .

101

      

 

"

 



 

"

Theorem 6.2.22. Let isomorphic to convex closure.

X

be a coherent domain. The Plotkin powerdomain of is . Formal union is interpreted as union followed by

. Then



Y ` `

` F



`
Y

`Y

` Y

 

Y Y ` ` Y


`

YY

` YY `

$

  

P 

 

  



 

  

P

 

6.2.4 Hyperspaces and probabilistic powerdomains In our presentation of powerdomains we have emphasized the feature that they are free algebras with respect to certain (in-)equational theories. From the general existence theorem for such algebras we derived concrete representations as sets of subsets. This is the approach which in the realm of domain theory was suggested rst by Matthew Hennessy and Gordon Plotkin in [HP79] but it has a rather long tradition in algebraic semantics (see e.g. [NR85]). However, it is not the only viewpoint one can take. One may also study certain sets of subsets of domains in their own right. In topology, this study of hyperspaces, as they are called, is a long-standing tradition, starting with Felix Hausdorff [Hau14] and Leopold Vietoris [Vie21, Vie22]. It is also how the subject started in semantics and, indeed, continues to be developed. A hyperspace can be interesting even if an equational characterization cannot be found or can be found only in restricted settings. Recent examples of this are the set-domains introduced by Peter Buneman [BDW88, Gun92a, Hec90, Puh93, Hec91, Hec93b] in connection with a general theory of relational databases. While these are quite natural from a domaintheoretic point of view, their equational characterizations (which do exist for some of them) are rather bizarre and do not give us much insight. The hyperspace approach is developed in logical form in Section 7.3. We should also mention the various attempts to dene a probabilistic version of the powerdomain construction, see [SD80, Mai85, Gra88, JP89, Jon90]. (As an aside, these cannot be restricted to algebraic domains; the wider concept of continuous domain is forced upon us through the necessary use of the unit interval .) They do have an equational description in some sense but this goes beyond the techniques of this chapter. One can then ask abstractly what constitutes a powerdomain construction and build a theory upon such a denition. This approach was taken in [Hec90, Hec91]. The most notable feature of this work is that under this perspective, too, many of the known powerdomains turn out to be canonical in a precise sense. How this (very natural) formulation of canonicity is connected with concerns in semantics, however, is as yet unclear.
"

Exercises 6.2.23. 1. For the proof of Theorem 6.1.6 we can equip the transitive closure of . Show:

2. Describe the free domain algebra for an arbitrary domain signature in the case that is empty.

3. Set up an algebraic theory such that all its dcpo-algebras have least elements but the embeddings are not strict.

102

(d) What is the advantage of

over

"

B C

"

(c) The ideal completions of Exercise 2.3.9(27).)

and

(b) In general,

(a) This relation

satises the interpolation axiom. . are isomorphic. (Use

is different from

and an arbitrary

8P 5 C

also with

Figure 14: Part of an algebraic domain where Theorem 6.2.19 fails. 4. Let be the usual equational theory of groups (or boolean algebras). Show that any dcpo-algebra with respect to this theory is trivially ordered. Conclude that the free construction collapses each connected component of the generating dcpo into a single point. 5. Given signatures and and sets of inequalities and we call the pair a reduct of if and . In this case there is an obvious forgetful functor from C to C , where C is any of the categories considered in this chapter. Show that the general techniques of Theorem 6.1.2 and 6.1.7 sufce to prove that this functor has a left adjoint. 6. Likewise, show that partial domain algebras can be completed freely. 7. Let be a free domain-algebra over an algebraic domain. Is it true that every operation, if applied to compact elements of , returns a compact element?

9. Is the Plotkin powerdomain closed on F-B, the category whose objects are bilimits of nite (but not necessarily pointed) posets? 10. Dene a natural isomorphism between and where is any continuous domain, is a complete lattice, and stands for the set of functions which preserve all suprema (ordered pointwise). 11. We want to construct an algebraic domain to which Theorem 6.2.19 cannot be extended. The compact elements of are arranged in nite sets already such that they form a directed collection in the Egli-Milner ordering, generating the ideal . We take one nite set for each element of , the nite powerset of the reals (or any other uncountable set), and we will have if . So we can arrange the in layers according to the cardinality of . Each contains one white and many black elements. If then the white element of is below every element of . For the order between black elements look at adjacent layers. There are many subsets of 103

   

    8 B6  5 6 6   $



 

B

P  P # P

 

@ W

8 B

B 2

@ (

8. Let

be the four-element lattice (Figure 1) and let . The sets and are elements of the Plotkin powerdomain of . Show that they have two minimal upper bounds. Since is a top element, is not an L-domain.
B @ (



P)



P P P  P #

 P   "H  

P )

B

P )

with cardinality . The many black elements of we partition into many classes of cardinality . So we can let the black elements of a lower neighbor of be just below the equally many black elements of one of these classes. (The idea being that no two black elements have an upper bound.) Figure 14 shows a tiny fraction of the resulting ordered set . Establish the following facts about this domain: (a) Above a black element there are only black, below a white element there are only white elements. (b) i. An ideal in can contain at most one black element from each set. ii. An ideal can contain at most one black element in each layer. iii. An ideal can contain at most countably many black elements. i. An ideal meeting all sets must contain all white elements. ii. If an ideal contains a black element, then it contains the least black element . iii. If an ideal meeting all sets contains then it must contain upper bounds for and the uncountably many white elements of the rst layer. These upper bounds must form an uncountable set and consist solely of black elements.

(c)

(d) From the contradiction between b-iii and c-iii conclude that only one ideal meets all sets, the ideal of white elements. Therefore, conin tains precisely one element, say . Show that equals and that it is Scott-closed. Hence it is far from containing all elements of . (e) Go a step farther and prove that the lenses of are not even directedcomplete by showing that the ideal we started out with does not have an upper bound. 12. (R. Heckmann) Remove idempotence from the Hoare powertheory and study free domain algebras with respect to this theory. These are no longer nite if the generating domain is nite. Show that the free algebra over the four-element lattice (Figure 1) is neither binite nor an L-domain.

104

   



 




1

 6    

 


  6  


 

7 Domains and logic

There are at least three ways in which the idea of a function can be formalized. The rst is via algorithms, which is the Computer Science viewpoint. The second is via value tables or, in more learned words, via graphs. This is the rather recent invention of Mathematics. The third, nally, is via propositions: We can either take propositions about the function itself or view a function as something which maps arguments which satisfy to values which satisfy . The encoding in the latter case is by the set of all . The beauty of the subject, then, lies in the interplay between these such pairs notions. The passage from algorithms (programs) to the extensional description via graphs is called denotational semantics. It requires sophisticated structures, precisely domains in the sense of this text, because of, for example, recursive denitions in programs. The passage from algorithms to propositions about functions is called program logics. If we take the computer scientists point of view as primary then denotational semantics and program logics are two different ways of describing the behaviour of programs. It is the purpose of this chapter to lay out the connection between these two forms of semantics. As propositions we allow all those formulae whose extensions in the domain under consideration are (compact) Scott-open sets. This choice is well justied because it can be argued that such propositions correspond to properties which can be detected in a nite amount of time [Abr87]. The reader will nd lucid explications of this point in [Smy92] and [Vic89]. Mathematically, then, we have to study the relation between domains and their complete lattices of Scott-open sets. Stated for general topological spaces, this is the famous Stone duality. We treat it in Section 7.1. The restriction to domains introduces several extra features which we discuss in a one by one fashion in Section 7.2. The actual domain logic, as a syntactical theory, is laid out in Section 7.3. The whole open-set lattice, however, is too big to be syntactically represented. We must, on this higher level, once more employ ideas of approximation and bases. There is a wide range of possibilities here, which can be grouped under the heading of information systems. We concentrate on one of these, namely, the logic of compact open subsets. This is well motivated by the general framework of Stone duality and also gives the richest logic.

7.1 Stone duality

7.1.1 Approximation and distributivity We start out with a few observations concerning distributivity. So far, this didnt play a role due to the poor order theoretic properties of domains. Now, in the context of open set lattices, it becomes a central theme, because, as we shall see, it is closely related with the concept of approximation. The earliest account of this connection is probably [Ran53]. A word on notation: We shall try to keep a clear distinction between spaces, which in the end will be our domains, and their open-set lattices. We shall emphasize this by using for the less-than-or-equal-to relation whenever we speak of lattices, even


 P 

105

though these do form a special class of domains, too, as you may remember from Section 4.1. Recall that a lattice is said to be distributive if for all the equality

holds. The dual of this axiom is then satised as well. For the innitary version of distributivity, we introduce the following notation for choice functions: If is a family of sets then we write if takes its value in for every . Complete distributivity can then be expressed by the equation

It, too, implies its order dual, see Exercise 7.3.19(1). There is a lot of room for variations of this and we shall meet a few of them in this section. Here comes the rst: Theorem 7.1.1. A complete lattice

is continuous if and only if

Proof. The reader should check for himself that the supremum on the right hand side is indeed over a directed set. Let now be an element approximating the left hand side of the equation. Then for each we have and so there is with . Let be the choice function which selects these . Then and is below the right hand side as well. Assuming to be continuous, this proves . The reverse inequality holds in every complete lattice. and let be the family of all directed For the converse x an element sets for which . From the equality, which we now assume to hold, we get that . We claim that for each choice function ,
 5   '

the corresponding element set with then



is approximating . Indeed, if for some and so .

is a directed

Let us now look at completely distributive lattices which, by the preceding theorem, are guaranteed to be continuous. We can go further and express this stronger distributivity by an approximation axiom, too.

106

Call

prime-continuous if for every

if

 % " '     '      

Denition 7.1.2. For a complete lattice

dene a relation

on

by

holds.

 

 

   $#          $#6  ) !       "   '    !       

 

'

 

'

 

'

holds for all families

of directed subsets of .

   
'

 P 9 B P 
6



    
& '

      

'

   



&

#$ 6  ) !

& ' ' 



       


% &

'

 

'

Note that the relation is dened in just the same way as the order of approximation, except that directed sets are replaced by arbitrary subsets. All our fundamental results about the order of approximation hold, mutatis mutandis, for as well. In particular, we shall make use of Proposition 2.2.10 and Lemma 2.2.15. Adapting the previous theorem we get George N. Raneys characterization of complete distributivity [Ran53]. Theorem 7.1.3. A complete lattice is prime-continuous if and only if it is completely distributive. Let us now turn our attention to approximation from above. The right concept for this is: Denition 7.1.4. A complete lattice is said to be -generated by a subset every , holds. (Dually, we can speak of -generation.)

We will study -generation by certain elements only, which we now introduce in somewhat greater generality than actually needed for our purposes. Denition 7.1.5. An element of a lattice is called -irreducible if whenever for a nite set then it must be the case that for some . We say is -prime if implies for some , where is again nite. Stating these conditions for arbitrary gives rise to the notions of completely -irreducible and completely -prime element. The dual notions are obtained by exchanging supremum for inmum. Note that neither -irreducible nor -prime elements are ever equal to the top element of the lattice, because that is the inmum of the empty set. Proposition 7.1.6. A -prime element is also -irreducible. The converse holds if the lattice is distributive.

Theorem 7.1.7. A continuous (algebraic) lattice pletely) -irreducible elements.

is -generated by its set of (com-

Proof. If and are elements of such that is not below then there is a Scottopen lter which contains but not , because is closed and the Scott-topology is generated by open lters, Lemma 2.3.8. Employing the Axiom of Choice in the form of Zorns Lemma, we nd a maximal element above in the inductive set . It is clearly -irreducible. In an algebraic lattice we can choose to be a principal lter generated by a compact element. The maximal elements in the complement are then completely -irreducible. Theorem 7.1.8. If is a complete lattice which is -generated by -prime elements, then satises the equations

107



  0    

'

 

    

if for

'

 &

     

C

where the sets and are nite. A dual statement holds for lattices which are -generated by -prime elements. Proof. The right hand side is certainly below the left hand side, so assume that is a prime element above . Surely, is above for every and because it is -prime it is above for some . We claim that the set of all covers at least one . Assume the contrary. Then for each there exists and we can dene a choice function . Then contradicts our construction of . So we know that for some all elements of are below and hence is also above . The proof for the second equation is similar and simpler. Note that the two equations are not derivable from each other because of the side condition on niteness. The rst equation is equivalent to
5

which can be stated without choice functions. In this latter form it is known as the frame distributivity law and complete lattices, which satisfy it, are called frames. The basic operations on a frame are those which appear in this equation, namely, arbitrary join and nite meet. 7.1.2 From spaces to lattices Given a topology on a set then consists of certain subsets of . We may think of as an ordered set where the order relation is set inclusion. This ordered set is a complete lattice because arbitrary joins exist. Let us also look at continuous functions. In connection with open-set lattices it seems right to take the inverse image operation which, for a continuous function, is required to map opens to opens. Set-theoretically, it preserves all unions and intersections of subsets, and hence all joins and nite meets of opens. This motivates the following denition. Denition 7.1.9. A frame-homomorphism between complete lattices which preserves arbitrary suprema and nite inma. and

We let CLat stand for the category of complete lattices and frame-homomorphisms. We want to relate it to Top, the category of topological spaces and continuous functions. The rst half of this relation is given by the contravariant functor , which assigns to a topological space its lattice of open subsets and to a continuous map the inverse image function. For an alternative description let be the two-element chain equipped with the Scott-topology. The open sets of a space are in one-to-one correspondence with continuous functions from to , if for each open set we set to be the map which assigns to an element if and only if . The action of on morphisms can then be expressed by . 108

   

is a map

'

 D 



  

'

 

"

'

      #   $ !     4$6 ) !   !  

'

  

'

"

and

'

"

'

Figure 15: A point in a complete lattice. 7.1.3 From lattices to topological spaces For motivation, let us look at topological spaces rst. An element of a topological space is naturally equipped with the following three pieces of information. We can associate with it its lter of open neighborhoods, the complement of its closure, or a map from , the one-element topological space, to . Taking the lter, for example, we observe that it has the additional property that if a union of open sets belongs to it then so does one of the opens. Also, the closure of a point has the property that it cannot be contained in a union of closed sets without being contained in one of them already. The map , which singles out the point, translates to a frame-homomorphism from to . Let us x this new piece of notation: is called prime if implies Denition 7.1.10. A lter for all nite . Allowing to be an arbitrary subset we arrive at the notion of completely prime lter. Dually, we speak of (completely) prime ideals. Proposition 7.1.11. Let lowing are equivalent: 1. 2. 3.

be a complete lattice and let

be a subset of . The fol-

is a completely prime lter. is a lter and for some

for a -prime element

This proposition shows that all three ways of characterizing points through opens coincide (see also Figure 15). Each of them has its own virtues and we will take advantage of the coincidence. As our ofcial denition we choose the variant which is closest to our treatment of topological spaces.

109

4.

is a frame-homomorphism from

to .

$  C

C  

   


` Y

` `

Y ` `

 C

 C

Y ` `

Y ` Y

 $  

C
Y

 

` Y `

 C

C C

` Y

Denition 7.1.12. Let be a complete lattice. The points of are the completely prime lters of . The collection of all points is turned into a topological space by requiring all those subsets of to be open which are of the form

Proposition 7.1.13. The sets

Observe the perfect symmetry of our setup. In a topological space an element belongs to an open set if ; in a complete lattice a point belongs to an open set if . By assigning to a complete lattice the topological space of all points, and to a frame-homomorphism the map which assigns to a point the point (which is readily seen to be a completely prime lter), we get a contravariant functor, also denoted by , from CLat to Top. Again, we give the alternative description based on characteristic functions. The fact is that we can use the same object for this purpose, because it is a complete lattice as well. One speaks of a schizophrenic object in such a situation. As we saw in Proposition 7.1.11, a completely prime lter gives rise to a frame-homomorphism . The action of the functor on morphisms can then be expressed, as before, by . 7.1.4 The basic adjunction A topological space can be mapped into the space of points of its open set lattice, simply map to the completely prime lter of its open neighborhoods. This assignment, which we denote by , is continuous and open onto its image: Let be an open set in . Then we get by simply unwinding the denitions: . It also commutes with continuous functions : . So the family of all constitutes a natural transformation from the identity functor to . The same holds for complete lattices. We let be the map which assigns to . It is a frame-homomorphism as we have seen in the proof of Proposition 7.1.13. To see that this, too, is a natural transformation, we : check that it commutes with frame-homomorphisms , which is essentially the same calculation as for . We have all the ingredients to formulate the Stone Duality Theorem: Theorem 7.1.14. The functors Top CLat and joints of each other. The units are and .

CLat

Top are dual ad-

110

      $ ! $ !     "    )      Q $ ! 

$

$ !

 (

                 $ I)   ! 

'

Proof. We have
' $

, nite, because points are lters and because they are completely prime.

        'P C '    C 
, , form a topology on
G G



  !    
  $ !

 C C  
 G

Proof. It remains to check the triangle equalities

and

For the left diagram let

be an open set in


While our concrete representation through open sets and completely prime lters, respectively, allowed us a very concise proof of this theorem, it is nevertheless instructive to see how the units behave in terms of characteristic functions. Their type is from to and from to , whereby the right hand sides are revealed to be second duals. The canonical mapping into a second dual is, of course, point evaluation: , where . This is indeed what both and do.

7.2 Some equivalences

7.2.1 Sober spaces and spatial lattices In this subsection we look more closely at the units and . We will need the following concept: Denition 7.2.1. A closed subset of a topological space is called irreducible if it is non-empty and cannot be written as the union of two closed proper subsets.

Clearly, an irreducible closed set corresponds via complementation to a irreducible (and hence -prime) element in the lattice of all open sets.

Proposition 7.2.2. Let be a topological space. Then is injective if and only if satises the -separation axiom. It is surjective if and only if every irreducible closed set is the closure of an element of . Proof. The rst half is just one of the various equivalent denitions of -separation: different elements have different sets of open neighborhoods. For the second statement observe that the -prime elements of are in oneto-one correspondence with completely prime lters of open sets. The condition then simply says that every such lter arises as the neighborhood lter of an element of .

111

The calculation for the right diagram is verbatim the same if we exchange and , and , and and .
G

   
and ,

Q $ ! 

 

$                       Q        D  Q    D  

5

    $  $

 $

$  $

Denition 7.2.3. A topological space

is called sober if

Note that if is bijective then it must be a homeomorphism because we know from Section 7.1.4 that it is always continuous and open onto the image. By the preceding proposition, a space is sober if and only if it is and every irreducible closed set is the closure of a point. The intuitive meaning is, of course, that a space is sober if it can be recovered from its lattice of open sets.

Proposition 7.2.4. For any complete lattice the unit and monotone. Furthermore, the following are equivalent: 1.

is injective.

2. The elements of

are separated by completely prime lters.

3.

is -generated by -prime elements.

5.

is order-reecting.

Proof. We have seen in Proposition 7.1.13 that all open sets on are of the form for some . This proves surjectivity. Monotonicity is clear because lters are upper sets. is equivTurning to the equivalent conditions for injectivity, we note that alent to for all completely prime lters . In other words, is injective if and only if the elements of are separated by completely prime lters. Given let be the inmum of all -primes above . We want to show that . If is strictly above then there exists a completely prime lter containing but not . Using the equivalence of Proposition 7.1.11, we see that this is the same as the existence of a -prime element in , a contradiction. From (3) the last two statements follow easily. They, in turn, imply injectivity (which, in a general order-theoretic setting, is strictly weaker than order-reection).

The intuitive meaning in this case is that a spatial lattice can be thought of as a lattice of open sets for some topological space. A direct consequence of Theorem 7.1.8 is the following: Theorem 7.2.6. A spatial lattice is a frame. In particular, it is distributive. Theorem 7.2.7. For any complete lattice the topological space any topological space the lattice is spatial.
5

Proof. The space of points of a lattice is certainly , because if we are given different completely prime lters then there is which belongs to one of them but not the other. Hence, contains one but not the other. For surjectivity of let be an irreducible closed set of lters. First of all, the union of all lters in is a non-empty upper set in which is unreachable by joins. Hence the complement of 112

Denition 7.2.5. A complete lattice

is called spatial if

is bijective.

is sober. For



  



C 

         C  

( 

$

4. If

then there exists a completely prime lter

such that

and

C $

  C

 $ ! 

is bijective.

is surjective

 

is a principal ideal . Also, the complement of in certainly contains . We claim that must be -prime. Indeed, if then is covered by the complements of and , whence it is covered by one of them, say the complement of , which means nothing else than . It follows that is contained in the closure of the point . On the other hand, belongs to the closed set as each of its open neighborhoods contains an element of . The second statement is rather easier to argue for. If and are different open sets then there is an element of contained in one but not the other. Hence the neighborhood lter of , which is always completely prime, separates and . Corollary 7.2.8. The functors and form a dual equivalence between the category of sober spaces and the category of spatial lattices. This result may suggest that a reasonable universe of topological spaces ought to consist of sober spaces, or, if one prefers the lattice-theoretic side, of spatial lattices. This is indeed true as far as spaces are concerned. For the lattice side, however, it has been argued forcefully that the right choice is the larger category of frames (which are dened to be those complete lattices which satisfy the frame distributivity law, Section 7.1.1). The basis of these arguments is the fact that free frames exist, see [Joh82], Theorem II.1.2, a property which holds neither for complete lattices nor for spatial lattices. (More information on this is in [Isb72, Joh82, Joh83].) The choice of using frames for doing topology has more recently found support from theoretical computer science, because it is precisely the frame distributivity law which can be expected to hold for observable properties of processes. Even though this connection is to a large extent the raison d tre for this chapter, we must refer to [Abr87, Abr91b, e Vic89, Smy92] for an in-depth discussion. 7.2.2 Properties of sober spaces Because application of to a space is an essentially idempotent operation, it is best to think of as a completion of . It is commonly called the soberication of . Completeness of this particular kind is also at the heart of the HofmannMislove Theorem, which we have met in Section 4.2.3 already and which we are now able to state in its full generality. Theorem 7.2.9. Let be a sober space. The sets of open neighborhoods of compact . saturated sets are precisely the Scott-open lters in Proof. It is pretty obvious that the neighborhoods of compact subsets are Scott-open lters in . We are interested in the other direction. Given a Scott-open lter then the candidate for the corresponding compact set is . We must show that each open neighborhood of belongs to already. For the sake of contradiction assume that there exists an open neighborhood . By Zorns Lemma we may further assume that is maximal with this property. Because is a lter, is -prime as an element of and this is tantamount to saying that its complement is irreducible as a closed set. By sobriety it must be the closure of a single point . The open sets which do not contain are precisely those which

113

 1

$

 

 

 D

  

  


G

 

   

are contained in . Hence every open set from the lter contains and so belongs to . This, nally, contradicts our assumption that is a neighborhood of . This appeared rst in [HM81]. Our proof is taken from [KP94]. Note that it relies, like almost everything else in this chapter, on the Axiom of Choice. Saturated sets are uniquely determined by their open neighborhoods, so we can reformulate the preceding theorem as follows: be a sober space. The poset of compact saturated sets orCorollary 7.2.10. Let dered by inclusion is dually isomorphic to the poset of Scott-open lters in (also ordered by inclusion). be a sober space. The ltered intersection of a family of Corollary 7.2.11. Let (non-empty) compact saturated subsets is compact (and non-empty). If such a ltered intersection is contained in an open set then some element of the family belongs to already. Proof. By the Hofmann-Mislove Theorem we can switch freely between compact saturated sets and open lters in . Clearly, the directed union of open lters is another such. This proves the rst statement. For the intersection of a ltered family to be contained in means that belongs to the directed union of the corresponding lters. Then must be contained in one of these already. The claim about the intersection of non-empty sets follows from this directly because we can take . Every -space can be equipped with an order relation, called the specialization order, by setting if for all open sets , implies . We may then compare the given topology with topologies dened on ordered sets. One of these which plays a role in this context, is the weak upper topology. It is dened as the coarsest topology for which all sets of the form are closed. Proposition 7.2.12. For an -space the topology on topology derived from the specialization order.
5

is ner than the weak upper

Proposition 7.2.13. A sober space is a dcpo in its specialization order and its topology is coarser than the Scott-topology derived from this order. Proof. By the equivalence between sober spaces and spatial lattices we may think of as the points of a complete lattice . It is seen without difculties that the specialization order on then translates to the inclusion order of completely prime lters. That a directed union of completely prime lters is again a completely prime lter is immediate. be such a directed union. It belongs to an open set if and only if Let for some . This shows that each is Scott-open. A dcpo equipped with the Scott-topology, on the other hand, is not necessarily sober, see Exercise 7.3.19(7). We also record the following fact although we shall not make use of it. Theorem 7.2.14. The category of sober spaces is complete and cocomplete. It is also closed under retracts formed in the ambient category Top. 114

 D





 

 

C 

'

C  

For the readers convenience we sum up our considerations in a table comparing concepts in topological spaces to concepts in for a complete lattice. space point specialization order open set saturated set compact saturated set

completely prime lter (c. p. lter) inclusion order c. p. lters containing some c. p. lters containing some upper set c. p. lters containing a Scott-open lter

7.2.3 Locally compact spaces and continuous lattices We already know that sober spaces may be seen as dcpos with an order-consistent topology. We move on to more special kinds of spaces with the aim to characterize our various kinds of domains through their open-set lattices. Our rst step in this direction is to introduce local compactness. We have: Lemma 7.2.15. Distributive continuous lattices are spatial. Proof. We have shown in Theorem 7.1.7 that continuous lattices are -generated by -irreducible elements. In a distributive lattice these are also -prime. Now recall that a topological space is called locally compact if every element has a fundamental system of compact neighborhoods. This alone does not imply sobriety, as the ascending chain of natural numbers, equipped with the weak upper topology, shows. But in combination with sobriety we get the following beautiful result: Theorem 7.2.16. The functors and restrict to a dual equivalence between the category of sober locally compact spaces and the category of distributive continuous lattices. Proof. We have seen in Section 4.2.3 already that holds in if there is a compact set between and . This proves that the open-set lattice of a locally compact space is continuous. , that is, . A completely For the converse, let be a point in an open set prime lter is Scott-open, therefore there is a further element with . Lemma 2.3.8 tells us that there is a Scott-open lter contained in which contains . We know by the previous lemma that a distributive continuous lattice can be thought of as the open-set lattice of its space of points, which, furthermore, is guaranteed to be sober. So we can apply the Hofmann-Mislove Theorem 7.2.9 and get that the set of points of , which are supersets of , is compact saturated. In summary, is contained in which is a subset of and this is a subset of . From now on, all our spaces are locally compact and sober. The three properties introduced in the next three subsections, however, are independent of each other.
G

115

 

 C C ( 

 

7.2.4 Coherence We have introduced coherence in Section 4.2.3 for the special case of continuous domains. The general denition reads as follows: Denition 7.2.17. A topological space is called coherent, if it is sober, locally compact, and the intersection of two compact saturated subsets is compact. Denition 7.2.18. The order of approximation on a complete lattice is called multiplicative if and imply . A distributive continuous lattice for which the order of approximation is multiplicative is called arithmetic. As a generalization of Proposition 4.2.16 we have: Theorem 7.2.19. The functors and restrict to a dual equivalence between the category of coherent spaces and the category of arithmetic lattices. Proof. The same arguments as in Proposition 4.2.15 apply, so it is clear that the openset lattice of a coherent space is arithmetic. For the converse we may, just as in the proof of Theorem 7.2.16, invoke the Hofmann-Mislove Theorem. It tells us that compact saturated sets of are in one-to-one correspondence with Scott-open lters. Multiplicativity of the order of approximation is just what we need to prove that the pointwise inmum of two Scott-open lters is again Scott-open. 7.2.5 Compact-open sets and spectral spaces By passing from continuous lattices to algebraic ones we get: restrict to a dual equivalence between the Theorem 7.2.20. The functors and category of sober spaces, in which every element has a fundamental system of compactopen neighborhoods, and the category of distributive algebraic lattices. The proof is the same as for distributive continuous lattices, Theorem 7.2.16. We now combine this with coherence. Denition 7.2.21. A topological space, which is coherent and in which every element has a fundamental system of compact-open neighborhoods, is called a spectral space. Theorem 7.2.22. The functors and restrict to a dual equivalence between the category of spectral spaces and the category of algebraic arithmetic lattices. Having arrived at this level, we can replace the open-set lattice with the sublattice of compact-open subsets. Our next task then is to reformulate Stone-duality with bases of open-set lattices. For objects we have: Proposition 7.2.23. Let be an algebraic arithmetic lattice. The completely prime lters of are in one-to-one correspondence with the prime lters of . The topology on is generated by the set of all , where is compact in .

116

Proof. Given a completely prime lter in , we let be the set of compact elements contained in it. This is clearly an upwards closed set in . It is a lter, because is arithmetic. Primeness, nally, follows from the fact that is Scott-open and hence equal to . Conversely, a lter in generates a lter in . For complete primeness let be a subset of with join in . is algebraic. So we may replace by and will still hold. Because is Scott-open, there is a nite subset of with . Some element of must be below and primeness then gives us that some element of belongs to already. The statement about the topology on follows from the fact that every element of is a join of compact elements. A frame-homomorphism between algebraic arithmetic lattices need not preserve compact elements, so in order to represent it through bases we need to resort to relations, as in Section 2.2.6, Denition 2.2.27. Two additional axioms are needed, however, because frame-homomorphisms are more special than Scott-continuous functions. Denition 7.2.24. A relation between lattices if the following conditions are satised:
# !

and

is called join-approximable

The following is then easily established:

Proposition 7.2.25. The category of algebraic arithmetic lattices and framehomomorphisms is equivalent to the category of distributive lattices and joinapproximable relations. By Proposition 7.2.23 we can replace the compound functor by a direct construction of a topological space out of a distributive lattice. We denote this functor by , standing for the spectrum of a distributive lattice. We also contract to . Then we can say: Theorem 7.2.26. The category of spectral spaces and continuous functions is dually equivalent to the category of distributive lattices and join-approximable relations via the contravariant functors and . We supplement the table in Section 7.2.2 with the following comparison of concepts in a topological space and concepts in the spectrum of a distributive lattice.

117

4.

# !

 

3.

# 

 

!#

2.

!#

1.

; ;

 

C    C

    (  Q     Q4(         0      4 (        I 4(       3(  B  P  P

  


 

 

"

 

C
" G

 

 Q C 
#

"

 !

space point specialization order compact-open set open set saturated set compact saturated set

prime lter inclusion order prime lters containing some union of compact open sets prime lters containing some upper set prime lters containing a lter

It has been argued that the category of spectral spaces is the right setting for denotational semantics, precisely because these have a nitary logical description through their distributive lattices of compact-open subsets, see [Smy92], for example. However, this category is neither cartesian closed, nor does it have xpoints for endofunctions, and hence does not provide an adequate universe for the semantics of computation. An intriguing question arises, of how the kinds of spaces traditionally studied in topology and analysis can best be reconciled with the computational intuitions reected in the very different kinds of spaces which arise in Domain Theory. An interesting recent development is Abbas Edalats use of Domain Theory as the basis for a novel approach to the theory of integration [Eda93a]. 7.2.6 Domains Let us now see how continuous domains come into the picture. First we note that sobriety no longer needs to be assumed: Proposition 7.2.27. Continuous domains eqipped with the Scott-topology are sober spaces. Proof. Let be an irreducible closed set in a continuous domain and let . We show that is directed. Indeed, given and in , then neither nor contain all of . By irreducibility, then, they cant cover . Hence there is . But since is Scott-open, there is also some in this set. This gives us the desired upper bound for and . It is plain from Proposition 2.2.10 that is the closure of . The following result of Jimmie Lawson and Rudolf-Eberhard Hoffmann, [Law79, Hof81], demonstrates once again the central role played by continuous domains. Theorem 7.2.28. The functors and restrict to a dual equivalence between CONT and the category of completely distributive lattices. is a union of sets of the form Proof. A Scott-open set in a continuous domain where . For each of these we have in . This proves complete distributivity, as we have seen in Theorem 7.1.3. For the converse, let be completely distributive. We already know that the points of form a dcpo (where the order is given by inclusion of lters) and that the topology on is contained in the Scott-topology of this dcpo. Now we show that every completely prime lter has enough approximants. Observe that certainly holds in all those cases where is an element of as directed suprema 118
G



  

"

 

"

 %  

  



  C

"

 

"



 
G



 

 

of points are unions of lters. Now given we get from prime-continuity that and so there must be some with . Successively interpolating between and gives us a sequence of elements such that , just as in the proof of Lemma 2.3.8. The set then is a completely prime lter containing with inmum in . The directedness of these approximants is clear because is ltered. As a consequence, we have that holds if and only if belongs to . We are not quite nished, though, because we also need to show that we get the Scott-topology back. To this end let be a Scott-open set of points, that is, implies and implies for some . Let be the supremum of all elements of the form , . We claim that . First of all, for each there is with , which, as we have just seen, is tantamount to , hence belongs to all and is proved. Conversely, if a point contains then it must contain some , , because it is completely prime. Hence belongs to , too, and we have shown . To this we can add coherence and we get a dual equivalence between coherent domains and completely distributive arithmetic lattices. Or we can add algebraicity and get a dual equivalence between algebraic domains and algebraic completely distributive lattices. Adding both properties characterizes what can be called 2/3-binite domains in the light of Proposition 4.2.17. We prefer to speak of coherent algebraic domains. As these are spectral spaces, we may also ask how they can be characterized through the lattice of compact open subsets. The answer is rather simple: A compact open set . These, in an algebraic domain is a nite union of sets of the form for in turn, are characterized by being -irreducible and also -prime. Theorem 7.2.29. The dual equivalence of Theorem 7.2.26 cuts down to a dual equivalence of coherent algebraic domains and lattices in which every element is the join of nitely many -primes. Proof. We only need to show that if a lattice satises the condition stated in the theorem, then its ideal completion is completely distributive. But this is trivial because a principal ideal generated by a -prime is completely -prime in the ideal completion and so the result follows from Theorem 7.1.3. All the combined strength of complete distributivity, algebraicity and multiplicativity of the order of approximation, however, does still not restrict the corresponding spaces far enough so as to bring us into one of our cartesian closed categories of domains. Let us therefore see what we have to add in order to characterize binite domains. The only solution in this setting appears to be a translation of mub-closures into the lattice of compact-open subsets, that is to say, the subset of -primes has the upside-down nite mub property (Denition 4.2.1). Let us sum up these considerations in a theorem: Theorem 7.2.30. A lattice is isomorphic to the lattice of compact-open subsets of an F-B-domain (Denition 4.3.7) if and only if, rstly, has a least element, secondly,

119

  C C      C  C   C  C
  

C 

C  7

 C C C ) C C  C   C  9  C C C    C 9 C  C @ C

C 



'

 " %
$

 
%

% 

) ' 0(%

each element of is the supremum of nitely many -primes and, thirdly, for every nite set of -primes there is a nite superset of -primes such that

The additional requirement that there be a largest element which is also -prime, characterizes the lattices of compact-open subsets of binite domains. The extra condition about nite mub-closures is not a rst-order axiom and cannot be replaced by one as was shown by Carl Gunter in [Gun86]. The smaller class of algebraic bc-domains has a rather nicer description: Theorem 7.2.31. A lattice is isomorphic to the lattice of compact-open subsets of an algebraic bc-domain if and only if it has a least element, each element of is the supremum of nitely many -primes and the set of -primes plus least element is closed under nite inma. 7.2.7 Summary We have summarized the results of this section in Figure 16 and Table 1. As labels we have invented a few mnemonic names for categories. We wont use them outside this subsection. The lled dots correspond to categories for which there is also a characterization in terms of compact-open subsets (spectral spaces). A similar diagram appears in [GHK 80] but there not everything, which appears to be an intersection of categories, really is one.

7.3 The logical viewpoint

This material is based on [Abr91b]. 7.3.1 Working with lattices of compact-open subsets Having established the duality between algebraic domains and their lattices of compact-open subsets we can now ask to what extent we can do domain theory through these lattices. We have already indicated that such an approach offers many new insights but for the moment our motivation could simply be that working with lattices is a lot easier than working with dcpos. Doing domain theory refers to performing the domain constructions of Sections 3.2, 3.3, 5 and 6, at least in a rst approximation. Let us try this out. Suppose you know for some binite domain , how do , the lattice of compact-open subsets of the lifted domain? The you construct answer is simple, just add a new top element: . Coalesced sum also works ne:

We encounter the rst problems when we look at the cartesian product. While it is clear that every compact-open subset of is a nite union of products of compact-open

120

H          H              

"

 

"

TOP SOB L-C COH

Topological spaces. No Stone-dual. Sober spaces vs. spatial lattices. Locally-compact sober spaces vs. continuous distributive lattices. Coherent spaces (= locally compact, sober, and intersection of compact saturated is compact) vs. arithmetic lattices (= distributive, continuous, and order of approximation is multiplicative). Sober spaces with a base of compact-open sets vs. distributive algebraic lattices. Continuous domains with Scott-topology vs. completely distributive lattices. Spectral spaces vs. algebraic arithmetic lattices vs. distributive lattices. Coherent domains vs. arithmetic completely distributive lattices. Algebraic domains vs. algebraic completely distributive lattices. Coherent algebraic domains vs. algebraic arithmetic completely distributive lattices vs. distributive lattices in which every element is the nite join of -primes. F-B-domains (Denition 4.3.7) (= bilimits of nite posets). Stone-dual only described through the basis (or base) of compact-open subsets, which is a distributive lattice with extra properties as stated in Theorem 7.2.30. Binite domains. Stone-dual only described through the basis of compact-open subsets, which is a distributive lattice with extra properties as stated in Theorem 7.2.30. Algebraic bounded-complete domains. Stone-dual only described through the basis of compact-open subsets, which is a distributive lattice with extra properties as stated in Theorem 7.2.31.

C-O CONT SPEC C-CONT ALG C-ALG

F-B

aBC

Table 1: The categories and their Stone-duals.

121

C-ALG F-B

aBC

Figure 16: An overview of Stone-dualities in domain theory. subsets in the factors, there seems to be no simple criterion on such unions which would guarantee unique representation. The moral then is that we must allow for multiple representations of compact-open subsets. Instead of lattices we shall study certain preordered structures. At rst glance this may seem as an unwanted complication but we will soon see that it really makes the whole programme work much more smoothly. Lattices are determined by either their order structure or their algebraic structure but this equivalence no longer holds in the preordered case. Instead we must mention both preorder and lattice operations. We also make -primeness explicit in our axiomatization. The reason for this is that we want to keep all our denitions inductive. This point will become clearer when we discuss the function space construction below. Denition 7.3.1. A coherent algebraic prelocale

is a preordered algebra with two

122

SPEC

C-CONT

COH

C-O

TOP SOB L-C

CONT

ALG

binary operations and , two nullary operations and , and a unary predicate on , such that is a supremum for , is an inmum for , is a least, and is a largest element. The preorder on is denoted by , the corresponding equivalence relation by . The predicate is required to hold if and only if is -prime. Finally, every element of must be equivalent to a nite join of -primes. We will not distinguish between a prelocale and its underlying set. The set is abbreviated as . This is essentially the denition which appears in [Abr91b]. There another predicate is included. We can omit this because we will not look at the coalesced sum construction. The expressions a supremum, an inmum, etc., may seem contradictory but they are exactly appropriate in the preordered universe. It is seen without difculties that every coherent algebraic prelocale gives rise to a lattice which is -generated by -primes and hence distributive. A domain prelocale is gotten by incorporating the two extra conditions from Theorem 7.2.30:

n n

is called Denition 7.3.2. Let and be domain prelocales. A function a pre-isomorphism if it is surjective, order-preserving and order-reecting. If is a domain prelocale and is a binite domain and if further there is a pre-isomorphism then we say that is a localic description of via . A pre-isomorphism must preserve suprema, inma, and least and largest element (up to equivalence). Furthermore, it restricts and corestricts to a surjective map . Let us look more closely at the case of a pre-isomorphism . A diagram may be quite helpful:

Remember that are just those compact-open subsets which are of the form for . The inclusion order between such principal lters is dual to the usual order on . Let us now lift the pre-isomorphism to the domain level. In the previous chapters, the natural approach would have been to apply the ideal completion functor to the preisomorphism between and . Here we use Stone-duality and apply to . This yields an isomorphism between and . Composed

123

Q  

 

 

dual

"

$ !

 

and

 #  P

    #  P

Q 

"

 $ !  $   !  $ !

      A  

GFE 

"

"

  



 

 

$ !

 

with the inverse of the unit

it gives us the isomorphism

It will be good to have a concrete idea of the behaviour of , at least for compact elements of . These are lters in which are generated by -prime elements. So let with . It is easily checked that equals that compact element of which is least in the compact-open subset .

Proof. We have just described how to derive an isomorphism from a pre-isomorphism. is surjective, orderFor the converse observe that the unit preserving and order-reecting (Proposition 7.2.4). For more general functions between domains, we can translate join-approximable relations into the language of domain prelocales. The following is then just a slight extension of Theorem 7.2.30. Theorem 7.3.4. The category of domain prelocales and join-approximable relations is dually equivalent to the category of binite domains and Scott-continuous functions. Our attempt to mimic the cartesian product construction forced us to pass to preordered structures but once we have accepted this we can go one step farther and make the prelocales syntactic objects in which no identications are made at all. More precisely, it is no loss of generality to assume that the underlying algebra is a term algebra with respect to the operations , and . As an example, let us describe the onepoint domain in this fashion. We take the term algebra on no generators, that is, every term is a combination of s and s. The preorder is the smallest relation compatible with the requirements in Denition 7.3.1. The effect of this is that there are exactly two equivalence classes with respect to , the terms equivalent to and the terms equivalent to . The former are precisely the -prime terms. We denote the resulting domain prelocale by 1. The syntactic approach also suggests that we look at the following relation between domain prelocales: Denition 7.3.5. Let and be domain prelocales. We say that of if the following conditions are satised:
D "

is a sub-prelocale

1.

is a subalgebra of

with respect to

and . to .
"

2. The preorder on

is the restriction of the preorder on 124

 

Proposition 7.3.3. There exists a map cale is a localic description of the binite domain are isomorphic.

such that the domain preloif and only if and

 7$   !

Q 

  C

 

  

P P

$ !

$ !

 

 

 P P

Q


 
Q

        C

"

"

Proof. It is clear that both and are continuous because directed joins of elements in , resp. , are just directed unions of prime lters. We have because the preorder on is the restriction of that on . For we dont need any special assumptions. The crucial point is that the two functions are well-dened in the sense that they indeed produce prime lters. The lter part follows again from the fact that both operations and preorder on are the restrictions of those on . For primeness assume that for some nite . This means for some . This element itself is a supremum of -primes of and because is a prime lter in we have some -prime element below in . But we have also required that the -prime elements of are precisely those -prime elements of which lie in and therefore some must be above . Primeness of , on the other hand, follows easily because suprema in are also suprema in . Corollary 7.3.7. Assume that is a localic description of via , that describes via , and that . Then the following denes an embedding of into : If , , , , then .

Of course, it happens more often that is a sub-domain of than that is a sub-prelocale of but the fact is that it will be fully sufcient and even advantageous to work with the stronger relation when it comes to solving recursive domain equations. 7.3.2 Constructions: The general technique Before we demonstrate how function space and Plotkin powerdomain can be constructed through prelocales, let us outline the general technique. The overall picture is in the following diagram. We explain how to get its ingredients step by step below.
" "

dual

125

 

P  Q' C

 P

"

Proof. If we denote by the embedding from preceding proposition, then the embedding .
Q X

into as dened in the is nothing else but

C D

"

  $  B!           

C  C P    C    C  P     


"

"

 P  C

"

 P

"

C 

"

Proposition 7.3.6. If is a sub-prelocale of ding projection pair between and

Q X " Q Q X X

"

" !

   I   

   $  ! $   ! 



A Q P  C

"

" 

"

C 

"

"

We write

if

is a sub-prelocale of

"

 C  

 

3.

equals

. . then the following denes an embed:

1. The set-up. We want to study a construction on (binite) domains. This could be any one from the table in Section 3.2.6 or a bilimit or one of the powerdomain constructions from Section 6.2. The diagram illustrates a binary construction. We can assume that we understand the action of the associated functor on binite domains. In particular, we know what the compact elements of are, how they compare and how acts on embeddings (Proposition 5.2.6). Thus we should have a clear understanding of the bottom row of the diagram, in detail:

are the compact elements of

are the compact-open subsets of and these are precisely those upper sets which are of the form for a nite set of compact elements. are the -prime elements of and these are precisely those subsets of which are of the form for a compact element. The order is inclusion which is dual to the usual order on compact elements.

Furthermore, we assume that we are given domain prelocales and which describe the binite domains and , respectively. These descriptions are encoded in preand . isomorphisms 2.The goal. We want to dene a domain prelocale which is a localic description of . This is achieved in the following series of steps. 3. Denition of . This is the creative part of the enterprise. We search for a description of compact-open subsets of based on our knowledge of and . The point is to do this directly, not via the the compact-open subsets of compact elements of , , and . There will be an immediate payoff, as we will gain an understanding of the construction in terms of properties rather than points. Our treatment of the Plotkin powerdomain below illustrates this most convincingly. The denition of will proceed uniformly in all concrete instances. First a set of generators is dened and then is taken to be the term algebra over with respect to , and . An interpretation function is dened based on the interpretations and . It is extended to all of as a lattice homomorphism: , etc. Finally, axioms and rules are given which govern the preorder and -primeness predicate. Next we have to check that our denitions work. This task is also broken into a series of steps as follows. into valid state4. Soundness. We check that axioms and rules translate via ments about compact-open subsets of . This is usually quite easy. From soundness we infer that is monotone and can be restricted and corestricted to a map . 5. Prime generation. Using the axioms and rules, we prove that every element of can be transformed (effectively) into an equivalent term which is a nite supremum of expressions which are asserted to be -prime. This is the crucial step and usually contains the main technical work. It allows us to prove the remaining 126

 P  C   P  C  QP  C

     P Q' C P  P P  P   P P '  Q  C     QP  C  P  Q ' C P  P $ 4 !     $  !

 

 P  P  C $  P !  QP  C

is the effect of the functor

on objects

and

C QP  C 

 P  C

 P  C

Q P  C Q P  C  P  C

2 2 2 2
5

properties of through and for the latter we can use our knowledge of the basis of . 6. Completeness for -primes. We show that is order reecting. 7. Denability for -primes. We show that is surjective. At this point we can ll in the remaining pieces without reference to the concrete construction under consideration. 8. Completeness. The interpretation function itself is order-reecting. Proof. Let be such that . By 5 we can replace these exand . pressions by formal joins of -primes: Soundness ensures that the value under the interpretation function remains unchanged and that each (resp. ) is of the form (resp. ) for compact elements in . The inclusion order on translates into the formula which by the completeness for -primes can be pulled back into : . In every preordered lattice it must follow that holds. 9. Denability. The surjectivity of is an easy consequence of the surjectivity of because we know that compact-open subsets in an algebraic domain are nite unions of compactly generated principal lters. 10. Well-denedness. Of course, is a domain prelocale and we have just shown that preorder and primeness predicate on are preserved and satises the two extra reected by . This constitutes a semantic proof that conditions for domain prelocales. In other words, is a well-dened operation on domain prelocales. 11. Stone-duality. At this point we have shown that is a pre-isomorphism. As in the previous subsection we lift it to an isomorphism between and via Stone duality:

So much for the correspondence on the object level. We also want to see how the construction harmonizes with the sub-prelocale relation, one the one hand, and the isomorphism , on the other hand. Thus we assume that we are given two more prelocales, and , which are localic descriptions of binite domains and , such that and hold. In Corollary 7.3.7 we have seen how to dene from this embeddings and . In Proposition 5.2.6 we have shown how the functors associated with different constructions act on embeddings, hence we may unambiguously write for the result of this action, which is an embedding from to . Embeddings preserve compact elements so restricts and corestricts to a monotone function . Now for both and 127

P  I

"

 P 

"

 P  P C P $C Q QP  C !  P C P C P  C    P C  $  ! $  !

(    P  C  ( QP     

P  P Q  QP  C

 QP  C

 

P  C       P 

"

 

"

" a

P  P  4

"

 P (      QP  

 

P  Q' C

P  Q' C

 

C

we have a diagram such as depicted at the beginning of this subsection. We connect the lower left corners of these by . This gives rise also to a map from to . Our way of dening will be such that it is immediate that is a subset of and hence there is an inclusion map connecting the upper left corners. Our next technical step then is the following. 12. Naturality. We show that the diagram

commutes. On the element level this reads: If and and then . Now we can again get the remaining missing information in a general manner. 13. Monotonicity. We show that . From the form of our construction it will be clear that is a subset of and the axioms and rules will be such that whatever can be derived in can also be derived in . We must show that in the larger prelocale nothing extra can be proved for elements of . The argument is a semantic one. Proof. Let such that holds in . Let , and similarly for . Correctness says that and hence . By naturality we have . Embeddings are order reecting so follows. Completeness then allows us to conclude that holds in as well. In the same way it is seen that the predicate on is the restriction of that on . 14. Least prelocale. It follows from the correctness of the construction that 1 holds. 15. Naturality of . Having established the relation we can look at the embedding which we dened in Proposition 7.3.6. We claim that the following diagram commutes:
" " " " " " Q Q X X

128

In other words, equals the embedding which can be derived from in the general manner of Corollary 7.3.7.

Q P   Q P  I  $ P P  ! P  I  P  Q  P P  I  P   P  P  C    P  C              &      P   &    P  P   #  P  P P  P  P  P  P  P C    &   P        &   P C  Q Q' C P

 P

 P P Q  
5

  &

"

 Q

"

"

"

  P  C  P C   P  C  P  C

P Q I

"

 5

"

"

"

"

"

 P P C  Q P  C

 &

 P
 5

Q QP  C

"

 P

"

"

"

"

"

Proof. This is a diagram of binite domains and Scott-continuous functions. It therefore sufces to check commutativity for compact elements. A compact element in is a lter generated by a term . Its image under is the compact element which generates the compact-open subset . The lter is generated by the same term . Applying to it gives us a compact element which is least in . Step 12 ensures that maps to . 7.3.3 The function space construction

We start out with two preparatory lemmas. The following notation will be helpful. We for the set of functions which map all of into . write Lemma 7.3.8. The Scott-topology on the function space mains and equals the compact-open topology.

Proof. Let be compact and be open and let be a directed set of continuous functions for which maps into . For every we have and because is open, there is with . The collection of open sets of the form , , covers . By compactness, this is true for nitely many already. If we let be an upper bound in for these , then holds which is equivalent to . Hence is a Scott-open set in . If, on the other hand, belongs to a Scott-open open set then with an idempotent deation this is true also for some approximation on , an idempotent deation on . For each element in the image of we have the set . The intersection of all these belongs to the compact-open topology, contains , and is contained in . Lemma 7.3.9. Let and be binite and let Then is compact-open in .
 

Proof. We know that denes an open set by the previous lemma. From biniteness we get idempotent deations on and on such that and . It follows that for the idempotent deation on which maps to . Now let and be domain prelocales describing binite domains and , as outlined in the general scheme in the previous subsection. The two lemmas justify the following choice of generators and interpretation function for our localic function space construction:
%  % % %  %

Note that the elements are just syntactic expressions. Here are axioms and rules for the preorder and -predicate.

  8 @#%  $6 5     "      $      8 746  5              8 7#% 5  )  $6 8  46  5  $   C            C    D8  7 $B 6  C   C       5  C   8 @#% 5  $6

for binite do-

and

be compact-open.

129

&
%



  P  C    Q P I


"

$   (  $  9     P  I     $    

 &

 5

  C   P  
" 

Rules.

and if

A few comments about these formulae are in place. First a convention: we assume that all index sets are nite, so that the expressions , etc., do indeed belong to the term algebra over . Observe the use of the -predicate in the rule . Without it, it would be very difcult to express this property. Also note that we enforce distributivity. This will be a prerequisite to prove prime generation below. It is clear that the rules are sound for the given interpretation, in particular, is the exact mirror image of our denition of joinable families of step functions, Definition 4.2.2. Let us therefore immediately turn to the crucial step 5. We cannot use Lemma 7.3.9 directly because we have not encoded the idempotent deations. We must nd the minimal elements of a compact-open subset explicitly. We illustrate the general technique in an example. Suppose is of the form and is of the form . We get a minimal element of by choosing a value and a value from . Then we must look at the intersection which again is of the form by coherence. For each we must choose a value from . And so on. Biniteness of the argument domain ensures that this process stops after a nite number of iterations and that the result is a joinable family of pairs . Coherence of the result domain guarantees that all in all only nitely many choices are possible. (Note that it can happen that a set of minimal upper bounds in the image domain is empty. In this case we have just been unlucky with our choices. If is not empty then some minimal function exists.) We can mimic this procedure in the prelocale as follows. For simplicity and to make the analogy apparent, we let stand for terms such that and . Similarly for . We get:

(3 more terms)

130

 '    P 

dist

6 6 $  6  66 $$

    



7$  $   $  79     $9 Q$    7$   $ $ Q  Q  $     $ 

'



 P P    P  ! %     P       Q    1          Q 

then and and

'

'

'

  P 

QP

'

   

'

If If If

then and

(dist)

 6 $ 6  6 0$    $     P D        6 0               0$ 6   $ $9     $    $     6  0$$     6        6   $     $    6 60$   $      $   0$
.
' ' ' '

Axioms.





 

then

We follow up only the rst of these four terms. The trick is to smuggle in the -prime terms whose join equals .

and now induction may do its job. Eventually we will have transformed into a disjunction of joinable families. For these, -primeness may be inferred through rule . Note that distributivity allows us to replace every term by an equivalent term of the form and for each term of the form the transformation works as illustrated. Next we show completeness for -primes. So assume and are terms for which the -predicate holds and for which . It must be the case that and are equivalent to joinable families and as there is no other way of deriving -primeness in . The order relation between joinable families has been characterized in Lemma 4.2.3. Here it says: and . Since we assume completeness for the constituting prelocales and , we may infer and . The relation is now easily derived from . Denability for -primes is immediate because we know that all compact functions arise from joinable families (Lemma 4.2.3 and Proposition 4.2.4). Properties 8 through 11 follow for all constructions uniformly. We are left with proving Naturality, Property 12. To this end, let us rst see how the embedding transforms a step function . We have: and . We get the step function . Now let be an element of for which holds. The interpretation of is the upper set generated by the joinable family , where and for all of step functions plying the embedding to these gives us the step functions as we have just seen. By Corollary 7.3.7 we can rewrite these as and . The supremum of the joinable family is least in . This was to be proved.
" Q X Q

Taking to be and to be we can express the faithfulness of our localic construction quite concisely as follows:

and this isomorphism is natural with respect to the sub-prelocale relation. 7.3.4 The Plotkin powerlocale Next we want to describe the lattice of compact-open subsets of the Plotkin powerdois concretely main of a binite domain . By Theorem 6.2.22 we know that 131

 

$6  8 # 5  A 8  

"

"

$B6   5

"

Theorem 7.3.10. Let

and

be domain prelocales. Then

  $     $    $  6  66 $$   7$   $  
'

    1               8 46 5 $        $         Q    Q 8   B6  5

6 0$    (   (      (  ( 4     8 46 $ ( $ ( 1 (    $  5  0     

. Ap-

, where

'

      

 8  #6 5   $       8  #$ 6  5  $  8  #6 5   $                   Q 8  

'

Q  $     6 $    7$   $ $  7$    $   7$     $ P P 

"

'

"

$ 6 4  5

 

represented as the set of lenses in , ordered by the Egli-Milner ordering (Denition 6.2.2). The compact elements in are those lenses which are convex closures of nite non-empty subsets of (Proposition 6.2.6). Idempotent deations on can be lifted to because is a functor. They map a lens to the convex closure of . The compact-open subsets of , however, are not so readily described. The problem is that one half of the Egli-Milner ordering refers to closed lower sets rather than upper sets. We do not follow this up as there is no logical pathway from the order theory to the axiomatization we are aiming for. It is much more efcient to either consult the mathematical literature on hyperspaces (see [Vie21, Vie22, Smy83b]) or to remind ourselves that powerdomains were introduced to model non-deterministic behaviour. If we think of the compact-open subsets in as observations that can be made about outcomes of a computation, then it is pretty clear that there are two ways of using these to make statements about non-deterministic programs: It could be the case that all runs of the program satisfy the property or it could be that at least one run satises it. Let us check the mathematics: Lemma 7.3.11. If is a binite domain and lowing are compact-open subsets in : is compact-open in , then the fol-

Furthermore, if we let range over all compact-open subsets in of all and forms a base for the Scott-topology on

then the collection .

Proof. Let be compact-open. Then is the upper set of nitely many compact . It is clear that elements and we nd an idempotent deation such that for we have both and . Hence these sets are compact-open, too. Let be a compact lens, that is, of the form for . The upper set n of in can be written as . The following denition then comes as no surprise:

Denition 7.3.12. Let be a domain prelocale which is a localic description of the binite domain . We dene the Plotkin powerlocale over as the term algebra over the generators

on the generators and extended to as a lattice homomorphism. Preorder and -predicate are dened as follows

132

with the interpretation function

dened by

           " Q        

P $  P    

    P     $  !   I         

 

'

      
X X

   

 

 

    

Note that we again require distributivity explicitly. The derivation scheme is almost minimal (in combination with the rest, and are equivalent). The following derived axioms are more useful than and : .

(D1) (D2)

We leave it to the interested reader to check soundness and pass straight on to the central Step 5, which is generation by -prime elements. we rst transform it into a disjunction of conProof. Given an expression in junctions by using the distributivity axiom. Thus it sufces to represent a term of the form

as a disjunction of -primes. But we can simplify further. Using we can pack and by (D2) we can assume that for each all -generators into a single term we have . We represent each as a disjunction of -primes of and applying and distributivity again we arrive at a disjunction of terms of the form

As yet we can not apply the -primeness rule because the two sets and may fail to coincide. Looking at the semantics for a moment, we see that in the compact-open subset thus described the minimal lenses are (the convex closures of) the least elements from each plus some 133

where each . Now we write as a disjunction of -primes is below , it doesnt hurt to add these, too. We get:

. Since each

 I

(

 6

           6   6  6  6

(

 QP P     QP P QP P P    6 (   

0 1' (

 

'

'

'

   (

 6  6

Rules.

If If

then holds for all

and , and is non-empty, then .

(dist)
U

                             "        
, , ,
' ' C

Axioms.

'

'

 6  6 6   6 6   6


, ,

 6

U


of the generators of the . We therefore take our term further apart so as to have a -prime expression for each subset of . For this we use (D1). One application (plus some distributivity) yields

and the picture becomes obvious. Next we check that

'

is order-reecting. , respectively and let . Then we have

5

The denability for -primes was shown in Lemma 7.3.11 already. Hence we are left with checking Naturality, which is Step 12. Proof. Let be a -prime element in and let be a sub-prelocale of . Let be the associated embedding from to . The least element in is the convex closure of the set of minimal elements in . Applying to it gives the convex closure of , as we have argued in the remark following Theorem 6.1.9. Corollary 7.3.7 tells us that this is the least element in . As in the case of the function space construction we summarize: Theorem 7.3.13. Let be a domain prelocale. Then

and this isomorphism is natural with respect to the sub-prelocale relation. The prelocales for Hoare and Smyth powerdomain are much easier to describe. All we have to do is to elide all generators and rules which refer to , respectively .

134

  A Q 

      

'

Since we assume that is order-reecting, we get from the rst equation and from the second .

( 5 ( 1' ( 0

P ((     

'

( 1' ( 0

Proof. Assume and be the least compact elements in , that is,

( (     (   

 ( 

P P



5

   9          

'

'

'

5 5

  

'

"

  (  

( 1' ( 0

7.3.5 Recursive domain equations In this subsection we will treat bilimits in the same fashion as we have studied nitary constructions. We assume that we are given domain prelocales such that each describes some binite domain . Corollary 7.3.7 states how the sub-prelocale relation between and , for , translates into an embedding . It is seen easily that is an expanding system, that is, for , holds. We claim that the directed union is a domain prelocale which describes . The rst claim is fairly obvious as all requirements about prelocales refer to nitely many elements only and hence a property of can be inferred from its validity in some . For the second claim we need to specify the interpretation function. To this end let be the embedding of into the bilimit (as dened in Theorem 3.3.7). Then we can set where is such that is contained in . The exact choice of does not matter; if then by Corollary 7.3.7 we have: and applying to this yields . The interpretation function is well-dened because embeddings preserve the order of approximation (Proposition 3.1.14), hence compact elements and compact-open subsets are also preserved. In order to see that is a pre-isomorphism we proceed as before, checking Steps 4, 5, 6, 7, and 12. It is, actually, rather simple. Soundness holds because the are monotone and map compact elements to compact elements. Prime generation holds because it holds in each . Since the are also order-reecting we get completeness from the completeness of the . Denability follows from Theorem 3.3.11; the only compact elements in are the images (under ) of compact elements in the approximating . If we are given a second sequence of prelocales (describing ) such that for each we have then it is clear that holds, too. For Naturality (Step 12) we must relate this to the embedding from to . The exact form of the latter can be extracted from Theorem 3.3.7: , where is the embedding of into and is the embedding derived from . Now let be -prime in . We have

and our proof is complete.

5

135

) ' 0(%

) ' 0(%

Theorem 7.3.14. If

is a chain of domain prelocales, then

      

% "

! %

"

" % 

 %   P  

          1  $ 3! %   P P  

"

  % A   P 

5 "

) ' 02%

 

) ' 0(%

) ' 02%

" ) ' % C0(%




"

   

@$  ! 

) ' % 0(&$

Observe how simple the limit operation for prelocales is if compared with a bilimit. This comes to full ower if we look at recursive domain equations. If is a construction built from those which can be treated localically (we have seen function space, Plotkin powerdomain, and bilimit, but all the others from Section 3.2 can also be included) then we can nd the initial xpoint of the functor on the localic side by simply taking the union of 1 1 1 . Why does this work and why does the result describe the canonical xpoint of ? First of all, we have 1 1 by Step 14. Successively applying to this relation gives us 1 1 by Monotonicity (Step 13). Hence we do have a chain 1 1 1 as stated and we can form its union . It obviously is a xpoint of the construction and therefore the domain described by it is a xpoint of the functor . But notice that we have rather than merely . This is not so surprising as it may seem at rst sight. Domain prelocales are only representations of domains and what we are exploiting here is the simple idea that we can let represent both and via two different interpretation functions. Let us now address the question about canonicity. It sufces to check that the embedding corresponding to 1 1 is equal to where corresponds to 1 1 . This is precisely the naturality of which we listed as Step 15. It follows that the bilimit is the same as the one constructed in Chapter 5. 7.3.6 Languages for types, properties, and points We dene a formal language of type expressions by the following grammar:

where ranges over a set of type variables. More constructions can be added to this list, of course, such as strict function space, smash product, Hoare powerdomain, and Smyth powerdomain. On the other hand, we do not include expressions for basic types, such as integers and booleans, as these can be encoded in our language by simple formulae. We have seen two ways to interpret type expressions. The rst interpretation takes values directly in B, the category of binite domains, and is based on the constructions in Sections 3.2, 3.3, 5.1, and 6.2. Since a type expression may contain free variables, the interpretation can be dened only relative to an environment B, which assigns to each type variable a binite domain. The semantic clauses corresponding to the individual rules of the grammar are as follows:

etc.

The expression denotes the environment which maps cides with at all other variables .

136

where

$ !

 8 9 $ 5   897$ 5    C P  C   '   8  $  #6   5    $ Q   D    

to

  C      C  C Q     C C 

'

         H   $   ! !

 

QH

A 

$  C 9!

QH

 

and coin-

Our work in the preceding subsections suggests that we can also interpret type expressions in the category DomPreloc of domain prelocales. Call the corresponding mappings and . The semantic clauses for this localic interpretation are:

The preceding subsections were meant to convince the reader of the following: Theorem 7.3.15. If domain prelocale sion it holds that and are environments such that for each the is a localic description of , then for every type expresis a localic description of . As a formula:
Q X

The next step is to dene for each type expression a formal language of (computational or observational) properties. This is done through the following inductive denition:

Here we have used the expression to denote the substitution of for in . The usual caveat about capture of free variables applies but let us not dwell on this. The rules exhibited above will generate for each the carrier set of a (syntactical) domain prelocale in the style of the previous subsections. Note that we dont need special properties for a recursively dened type as these are just the properties of the approximating domains bundled together (Theorem 7.3.14). On each we dene a preorder and predicates and (the latter is needed for the coalesced sum construction) through yet another inductive denition. For example, the following axioms and rules enforce that each is a preordered distributive

137

 8 $ 5   P      8   $   5    $       
etc.

       P     Q  #)       H    H )   ) P $   P $ )    )    P

  A Q         D 

where

Q X WUS V T

Q X WUS V T

Q X WUS V T

Q P4G I H

8 5

 

 

 

 

 

 

 

 

 

 8 '  5                P      P      P 

QH

QH

lattice.

We have seen some type specic axioms and rules in the denition of the function space prelocale and the Plotkin powerlocale. For the full list we refer to [Abr91b], describes the p. 49ff. If is a closed type expression then the domain prelocale intended binite domain: Theorem 7.3.16. If is a closed type expression then

(Note that this is a special case of Theorem 7.3.15.) The whole scheme for deriving , , and is designed carefully so as to have nite positive information in the premise of each rule only. Hence the whole system can be seen as a monotone inductive denition (in the technical sense of e.g. [Acz77]). Furthermore, we have already established close connections between the syntactical rules and properties of the described domains. This is the basis of the following result. Theorem 7.3.17. The language of properties is decidable. Proof. The statement is trivial for the domain prelocale 1 because only combinations of and occur in 1 . For composite types we rely on the general development in Section 7.3.2, which at least for three concrete instances we have veried in Sections 7.3.35. First of all, every expression in can be effectively transformed into a nite disjunction of -primes (i.e. expressions satisfying the -predicate); this is Step 5, prime generation. Soundness and completeness ensure that the expressions . satisfying the -predicate are precisely the -primes in the preordered lattice Hence we can decide the preorder between arbitrary expressions if we can decide the preorder between -primes. For the latter we note that our constructions accomplish more than we have stated so far. All -primes, which are produced by the transformation algorithms, are of the explicit form occuring in the rules for deriving the predicate; rather than merely expressions which happen to be equivalent to -primes. The preorder between these explicit -primes is (for each construction) easily characterized through the semantic interpretation function . The task of establishing the 138

 

 )

 )

 A  

     

 

Q H RI 4G




   

Q X WUS V T

  



                     

P  P )


P )


Q X WUS V T

Q I H RP4G

preorder between these primes is then reduced to establishing some formula dened by structural induction on the type . Since every expression in is derived from and in nitely many steps, we will eventually have reduced our task to checking the preorder between certain expressions in 1 . Finally, we introduce a formal language to speak about points of domains. So far, we have done this in a rather roundabout way, trusting in the readers experience with sets and functions. Doing it formally will allow us to establish a precise relationship between (expressions for) points and (expressions for) properties. We assume that for each (closed) type expression we have a denumerable set of typed variables. The terms are dened as follows (where stands for is a term of type ):
Q H RI 4G

to

In the same fashion as for type expressions we have two alternatives for interpreting a term of type . We can either give a direct denotational semantics in the binite domain or we can specify a prime lter in the corresponding domain prelocale . The denotational semantics suffers from the fact that in order to single out a particular element in a domain we use a mathematical language which looks embarrassingly similar to the formal language we intend to interpret. Some of the semantic clauses to follow will therefore appear to be circular. Again we need environments to deal with free variables. They are maps which we assume to respect the typing. In the following

139

over

extend

QH

QH

up lift

 !  ! 8 '   5  !  !   '   ! '  8 '  5 !  H !    ! P  !  !     ! P  !   !       ! P  !   !       !  !   ! P  !   !     ! !    !  P ! P  !   !    !   !!  P  H !  ! P   H ! P  ! P !   !    ! P  $  !   $ !    32 ! 

cases

of

else

and

let
V

be

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 $  I !

 P

QH

 P P   ! 

QH

 

Q X WUS V T

Now let us give the localic, or, as we are now justied in saying, logical interpretation. We use a sequent calculus style of presenting this domain logic. The problem of free variables is dealt with this time by including a nite list of assumptions on variables. We write them in the form and assume that contains at most one of these for each variable . A sequent then takes the form and should be read

! 

B 0

@ W


lift

 

H V

 8 $  P5       1   P      4  P         8  $    P D5          P   8P $  5   P   !   P  $ 8   !    5  8P $   5     P  P 8 $ 4 $   P  P 5      8 $  5  $   P

clauses we will also suppress the type information.

the least element in

                              P  P    3 2  
where
V S

let

be

where

cases

of

else

to

over

extend

where

140

$

as

A few comments may help in reading these clauses. The rst two rules guarantee

Q X WUS V T

H V

     

'

!    !    !     H  ) !   H ) ! !      !     !   !        !      !        !        !     B !     !      !       B !   ) !     ! P    H ) ! P  !      $ ) ! 3 2  ! I   $    !   $ P  !  $ P  !   $ P

 
'

satises

under the assumptions in .

! 

 

let

be

Q X VWUS T B 2

             

'

P #)! P  ) !   !   !  $ P  $ P ! P ! H    ! !    $ ) !   !   $ QP !   ! $ QP  !  $  P P   P     !  

'

cases

of

else

cases

of

else

lift

to

over

extend

over

extend

141
                           

 

Q X WUS V T

 

! $ P  !    !    !   !    ! !   !   !  !  !      ! ! $ P   ! !  $ QP   !    !   !   !  $ P   ! !  B    ! $ P P #) P B  !  $ QP !



Q X WUS V T

that the set of properties which can be deduced for a term forms a lter in the domain prelocale. The third rule expresses the fact that every particular will satisfy properties from a prime lter. In particular, it entails that is always true. The fourth rule (which is the last of the structural rules) is ordinary weakening. We need it to get started in a derivation. In the two rules for the cases-construct the predicate T shows up. Instead of T we could have written but as we said before, we want to keep the whole logic positive, that is to say, we want to use inductive denitions only. The two rules for and may seem a bit boring, but it is precisely at this point where we take advantage of the fact that in the world of domain prelocales we solve domain equation up to equality. The last rule, nally, has to be applied nitely many times, starting from , in order to yield something interesting. Here we may note with regret that our whole system is based on the logic of observable properties. A standard proof principle such as xpoint induction for admissible predicates, Lemma 2.1.20, does not t into the framework. On the other hand, it is hopefully apparent how canonical the whole approach is. For applications, see [Abr90c, Abr91a, Bou91, Hen93, Ong93, Jen91, Jen92]. Let us now compare denotational and logical semantics. We need to say how environments and assumptions t together. First of all, we assume that maps each variable into . Secondly, we want that belongs to the compactopen subset described by the corresponding entry in . But since environments are functions dened on the whole set of variables while assumptions are nite lists, the if for all entries in we following denition is a bit delicate. We write have . Using this convention, we can formulate validity for assertions about terms: if and only if The nal tie-up between the two interpretations of type expressions and terms then is the following: Theorem 7.3.18. The domain logic is sound and complete. As a formula:

if and only if

Exercises 7.3.19. 1. Prove that a completely distributive lattice also satises the dual distributivity axiom: 2. [Ran60] Prove that a complete lattice the following holds for all : is completely distributive if and only if
' '

(Hint: Use Theorem 7.1.3.)

3. Show that a topological space is sober if and only if every irreducible closed set is the closure of a unique point. 4. Find a complete lattice for which

is empty.

142

Q X WUS V T

!

$

  

              !  P P !   

Q H RI 4G

Q X WUS V T

!  

$ QP

 

 B)

 


  

5. Show that every Hausdorff space is sober. Find a -space which is not sober. The converse, a sober space, which is not , ought to be easy to nd. 6. Find a dcpo which is not sober in the Scott-topology. (Reference: [Joh81]. For an example which is a complete lattice, see [Isb82]. There is no known example which is a distributive lattice.)

 

7. Describe the topological space plete lattice . 8. Let

in terms of -prime elements of the com-

be a continuous domain. Identify with the set of -prime elements in . Prove that the Lawson-topology on is the restriction of the Lawsontopology on to .
#

10. Extend Lemma 7.3.8 to other classes of domains. 11. Try to give a localic description of the coalesced sum construction.

143

9. Suppose


is a lattice homomorphism. Show that dened by if is a join-approximable relation. Characterize the continuous functions between spectral spaces which arise from these particular join-approximable relations.

 

$ )  !   

 

8 Further directions
Our coverage of Domain Theory is by no means comprehensive. Twenty-ve years after its inception, the eld remains extremely active and vital. We shall try in this Section to give a map of the parts of the subject we have not covered.

8.1 Further topics in Classical Domain Theory

We mention four topics which the reader is likely to encounter elsewhere in the literature. 8.1.1 Effectively given domains As we mentioned in the Introduction, domain-theoretic continuity provides a qualitative substitute for explicit computability considerations. In order to evaluate this claim rigorously, one should give an effective version of Domain Theory, and check that the key constructions on domains such as product, function space, least xpoints, and solutions of recursive domain equations, all lift to this effective setting. For this purpose, the use of abstract bases becomes quite crucial; we say (simplifying a little for this thumbnail sketch) that an -continuous domain is effectively given if it has an abstract basis which is numbered as in such a way that is recursive in the indices. Similarly, a continuous function between effectively given domains is effective if the corresponding approximable mapping is recursively enumerable. We refer to [Smy77, Kan79, WD80] and the chapter on Effective Structures in this Handbook for developments of effective domain theory on these lines. There have also been some more sophisticated approaches which aim at making effectivity intrinsic by working inside a constructive universe for set theory based on recursive realizability [McC84, Ros86, Pho91]. We shall return to this idea in subsection 8.5. 8.1.2 Universal Domains Let C be a cartesian closed category of domains, and a domain in C. We say that is . Thus universaluniversal for C if, for every in C, there is an embedding ity means that we can, in effect, replace the category C by the single domain . More precisely, we can regard the domain as represented by the idempotent , where is the projection corresponding to . Since , and is again in C and hence embeddable in , we can ultimately identify with an element , which we can think of as a code for . Moreover, constructions such as product and function space induce continuous functions

which act on these codes, so that e.g.

144

8 $ ) #6 5  $ !    $ 3!  

$ #! 

$ 8 #6  5  P

#6 ! P $

' 2%

 %

"

I S

I S

 P
"

In this way, the whole functorial level of Domain Theory which we developed as a basis for the solution of recursive domain equations in Section 5 can be eliminated, and we can solve domain equations up to equality on the codes by nding xpoints of continuous functions over . This approach was introduced by Scott in [Sco76], and followed in the rst textbook on denotational semantics [Sto77]. However, it must be said that, as regards applications, universal domains have almost fallen into disuse. The main reason is probably that the coding involved in the transition from to is confusing and unappealing; while more attractive ways of simplifying the treatment of domain equations, based on information systems, have been found (see 8.1.4). However, there have been two recent developments of interest. Firstly, a general approach to the construction of universal domains, using tools from Model Theory, has been developed by Gunter and Jung and Droste and G bel, and used to construct universal domains for many categories, and to o prove their non-existence in some cases [GJ88, DG90, DG91, DG93]. Secondly, there is one application where universal domains do play an important r le: to provide models for type theories with a type of all types. Again, the original o idea goes back to [Sco76]. We say that a univeral domain admits a universal type for in C is itself a domain in Cand hence admits if the subdomain of all a representation . We can think of as a code for the type of all types. In [Sco76], Scott studied the powerset as a univeral domain for two categories: the category of -continuous lattices (for which domains are taken to be represented by ), and the category of -algebraic lattices (for which domains are idempotents on represented by closures). Ershov [Ers75] and Hosono and Sato [HS77] independently proved that does not admit a universe for the former category; Hancock and Martin-L f proved that it does for the latter (reported in [Sco76]). For recent examples o of the use of universal domains to model a type of all types see [Tay87, Coq89, Ber91]. 8.1.3 Domain-theoretic semantics of polymorphism We have seen the use of continuity in Domain Theory to circumvent cardinality problems in nding solutions to domain equations such as

A much more recent development makes equally impressive use of continuity to give a nitary semantics for impredicative polymorphism, as in the second-order calculus (Girards System F) [Gir86, CGW87, Coq89]. This semantics makes essential use of the functorial aspects of Domain Theory. There have also been semantics for implicit polymorphism based on ideals [MPS86] and partial equivalence relations [AP90] over domains. We refer to the chapter in this volume of the Handbook on Semantics of Types for comprehensive coverage and references. 8.1.4 Information Systems Scott introduced information systems for bounded-complete -algebraic dcpos (Scott domains) in [Sco82]. The idea is, roughly, to represent a category of domains by a category of abstract bases and approximable mappings as in Theorems 2.2.28 and 145



8  $6 97#% 5 A 

2.2.29. One can then dene constructions on domains in terms of the bases, as in Propositions 3.2.4 and 4.2.4. This gives a natural setting for effective domain theory as in 8.1.1 above. Moreover, bilimits are given by unions of information systems, and domain equations solved up to equality, much as in 7.3.5. More generally, information systems correspond to presenting just the coprime elements from the domain prelocales of 7.3. Information system representations of various categories of domains can be found in [Win88, Zha91, Cur93]. A general theory of information systems applicable to a wide class of topological and metric structures can be found in [ES93].

8.2 Stability and Sequentiality

Recall the - style denition of continuity given in Proposition 2.2.11: given it provides with . However, there is no canonical choice of from . In an order-theoretic setting, it is natural to ask for there to be a least such . This leads to the idea of the modulus of stability: , where , is the least such , if it exists. We say that a continuous function is stable if the modulus always exists, and dene the stable ordering on such functions by

We can think of the modulus as specifying the minimum information actually required of a given input in order that the function yields a given information on the output; the stable ordering renes the usual pointwise order by taking this intensional information into account. It turns out that these denitions are equivalent to elegant algebraic notions in the setting of the lattice-like domains introduced (for completely different purposes!) in is Section 4.1. Let , be domains in L. Then a continuous function stable iff it preserves bounded non-empty inma (which always exist in L; cf. Proposition 4.1.2), and iff for all , . This is the rst step in an extensive development of Stable Domain Theory in which stable functions under the stable ordering take the place which continuous functions play in standard Domain Theory. Stable Domain theory was introduced by Berry [Ber78, Ber79]. Some more recent references are [Gir86, CGW87, Tay90, Ehr93]. Berrys motivation in introducing stable functions was actually to try to capture the notion of sequentially computable function at higher types. For the theory of sequential functions on concrete domains, we refer to [KP93, Cur93].

8.3 Reformulations of Domain Theory

At various points in our development of Domain Theory (see e.g. Section 3.2), we have referred to the need to switch between different versions C, C , C of some category of domains, depending on whether bottom elements are required, and if so whether functions are required to preserve them. In some sense C and C are the mathematically natural categories, since what the morphisms must preserve matches the structure that the objects are required to have; while C is the preferred category for semantics, since endomorphisms need not have xpoints at all in C, while least xpoints in C are necessarily trivial. 146



&#) $ !

P P   P Q#    P #        P      

 

P   P #

 $ #) !

  

 ' 

"

 

 

All this suggests that something is lacking from the mathematical framework in order to get a really satisfactory tie-up with the applications. We shall describe a number of attempts to make good this deciency. While no denitive solution has yet emerged, these proposals have contributed important insights to Domain Theory and its applications. 8.3.1 Predomains and partial functions The rst proposal is due to Gordon Plotkin [Plo85]. The idea is to use the objects of C (predomains, i.e. domains without any requirement of bottom elements), but to change the notion of morphism to partial continuous function: where we say that a partial function is continuous if its domain of denition is a Scott-open subset of , and its restriction to this subset is a (total) continuous function. The resulting category is denoted by C . This switch to partial continuous functions carries with it a change in the type structure we can expect to have in our categories of domains: they should be partial cartesian closed categories, as dened e.g. in [RR88, Ros86]. One advantage of this approach is that it brings the usage of Domain Theory closer to that of recursion theory. For example, the hierarchy of (strict) partial continuous functionals over the natural numbers will be given by

rather than


This avoidance of bottom elements also leads to a simpler presentation of product and sum types. For example, there is just one notion of sum, the disjoint union , which is indeed the coproduct in C . An important point is that there is a good correspondence between the operational behaviour of functions with a call-by-value parameter-passing mechanism and the par. For example, there is a good t between and the tial function type function type constructor in Standard ML [MT91, MTH90]. To balance these advantages, we have the complication of dealing with partially dened expressions and partial cartesian closure; and also a less straightforward treatment of xpoints. It is not the case that an arbitrary partial continuous function has a well-dened least xpoint. However, if itself is a partial function type, e.g. , then does have a well-dened least xpoint. This is in accord with computational intuition for call-by-value programming languages, but not so pleasant mathematically. As a nal remark, note that in fact C is equivalent to C ! Thus, in a sense, this approach brings nothing new. However, there is a distinct conceptual difference, and also C is more amenable to constructive proof and categorical axiomatization [Ros86].

147

 3) !

 

8 5

PG8 4$ 6 8 4$ 6 55 8P 4$ 6 5 P 8P 8 55 8P 5 P

#)  !
8 5


8 5 

8.3.2 Computational Monads Computational monads have been proposed by Eugenio Moggi as a general structuring mechanism for denotational semantics [Mog91]. A computational monad on a cartesian category C is a monad together with a tensorial strength, i.e. a natural transformation

satisfying some equational axioms. The import of the strength is that the monad can be internalised along the lines mentioned after Proposition 6.1.8. Now let C be a category of (pre)domains and total continuous functions. Moggis proposal is to make a distinction between values and (denotations of) computations. An element of is a value, an is a computation. A (call-by-value) procedure will denote a morphism element of which accepts an input value of type and produces a computation over . Composition of such morphisms is by Kleisli extension: if , , then composition is dened by

with identities given by the unit . In particular, partiality can be captured in this way using the lifting monad, for which see 3.2.5. Note that this particular example is really just another way of presenting the category C of the previous subsection; there is a natural isomorphism

The value of the monadic approach lies in its generality and in the type distinction it introduces between values and computations. To illustrate the rst point, note that the various powerdomain constructions presented in Section 7.2 all have a natural structure as strong monads, with the monad unit and multiplication given by suitable versions of the singleton and big union operations. For the second point, we refer to the elegant axiomatization of general recursion in terms of xpoint objects given by Crole and Pitts [CP92], which makes strong use of the monadic approach. This work really belongs to Axiomatic Domain Theory, to which we will return in subsection 4 below. 8.3.3 Linear Types Another proposal by Gordon Plotkin is to use Linear Types (in the sense of Linear Logic [Gir87]) as a metalanguage for Domain Theory [Plo93]. This is based on the following observation. Consider a category C of domains with bottom elements and strict continuous functions. This category has products and coproducts, given by cartesian products and coalesced sums. It also has a monoidal closed structure given by smash product and strict function space, as mentioned in 3.2.4. Now lifting, which is a monad on C by virtue of the adjunction mentioned in 3.2.5, is dually a comonad on C ; and the co-Kelisli category for this comonad is C .

148

$ !

"

"

"

$ ) !

P #6 $

H  $
"

$ 8  5 A 8 B6  5

$ #$ 6

"


H ! &
"

 P P

$ #6

"

Indeed, Linear Logic has broader connections with Domain Theory. A key idea of Linear Logic is the linear decomposition of the function space:

One of the cardinal principles of Domain Theory, as we have seen, is to look for cartesian closed categories of domains as convenient universes for the semantics of computation. Linear Logic leads us to look for linear decompositions of these cartesian closed structures. For example, the cartesian closed category of complete lattices and continuous maps has a linear decomposition via the category of complete lattices and sup-lattice homomorphismsi.e. maps preserving all joins, with , the Hoare powerdomain of . There are many other examples [Hoo92, Ehr93, Hut94].

8.4 Axiomatic Domain Theory

We began our account of Domain Theory with requirements to interpret certain forms of recursive denitions, and to abstract some key structural features of computable partial functions. We then introduced some quite specic structures for convergence and approximation. The elaboration of the resulting theory showed that these structures do indeed work; they meet the requirements with which we began. The question remains whether another class of structures might have served as well or better. To address this question, we should try to axiomatize the key features of a category of domains which make it suitable to serve as a universe for the semantics of computation. Such an exercise may be expected to yield the following benets: By making it clearer what the essential structure is, it should lead to an improved meta-language and logic, a renement of Scotts Logic of Computable Functions [Sco93].

Having a clear axiomatization might lead to the discovery of different models, which might perhaps be more convenient for certain purposes, or suggest new applications. On the other hand, it might lead to a representation theorem, to the effect that every model of our axioms for a category of domains can in fact be embedded in one of the concrete categories we have been studying in this Chapter.

Thus far, only a limited amount of progress has been made on this programme. One step that can be made relatively cheaply is to generalize from concrete categories of domains to categories enriched over some suitable subcategory of . Much of the force of Domain Theory carries over directly to this more general setting [SP82, Fre92]. Moreover, this additional generality is not spurious. A recent development in the semantics of computation has been towards a renement of the traditional denotational paradigm, to reect more intensional aspects of computational behaviour. This has led to considering as semantic universes certain categories in which the morphisms are not functions but sequential algorithms [Cur93], information ows [AJ94b], gametheoretic strategies [AJ94a], or concurrent processes [Abr94]. These are quite different from the concrete categories of domains we have been considering, in which the morphisms are always functions. Nevertheless, they have many of the relevant properties 149

$ 8 5 A 8 46 5
" "

2 2

of categories of domains, notably the existence of xpoints and of canonical solutions of recursive domain equations. The promise of axiomatic domain theory is to allow the rich theory we have developed in this Chapter to be transposed to such settings with a minimum of effort. The most impressive step towards Axiomatic Domain Theory to date has been Peter Freyds work on algebraically compact categories [Fre91, Fre92]. This goes considerably beyond what we covered in Section 5. The work by Crole and Pitts on FIXcategories should also be mentioned [CP92]. In another direction, there are limitative results which show that certain kinds of structures cannot serve as categories of domains. One such result appeared as Exercise 5.4.11(3). For another, see [HM93].

8.5 Synthetic Domain Theory

A more radical conceptual step is to try to absorb all the structure of convergence and approximation, indeed of computability itself, into the ambient universe of sets, by working inside a suitable constructive set theory or topos. The slogan is: Domains are Sets. This leads to a programme of Synthetic Domain Theory, by analogy with Synthetic Differential Geometry [Koc81], in which smoothness rather than effectivity is the structure absorbed into the ambient topos. The programme of Synthetic Domain Theory was rst adumbrated by Dana Scott around 1980. First substantial steps on this programme were taken by Rosolini [Ros86], and subsequently by Phoa [Pho91], and Freyd, Mulry, Rosolini and Scott [FMRS90]. Axioms for Synthetic Domain Theory have been investigated by Hyland [Hyl91] and Taylor [Tay91], and the subject is currently under active development.

150

9 Guide to the literature

As mentioned in the Introduction, there is no book on Domain Theory. For systematic accounts by the two leading contributors to the subject, we refer to the lecture notes of Scott [Sco81] and Plotkin [Plo81]. There is also an introductory exposition by Gunter and Scott in [GS90]. An exhaustive account of the theory of continuous lattices can be found in [GHK 80]; a superb account of Stone duality, with a good chapter on continuous lattices, is given in [Joh82]; while [DP90] is an excellent and quite gentle introduction to the theory of partial orders. Some further reading on the material covered in this Chapter: Section 2: [DP90, Joh82]; Section 3: [Plo81, Gun92b, Win93]; Section 4: [Jun89, Jun90]; Section 5: [SP82, Fre91, Fre92, Pit93b, Pit93a]; Section 6: [Plo76, Smy78, Win83, Hec91, Sch93]; Section 7: [Abr90c, Abr91a, AO93, Ong93, Hen93, Bou94, Jen92, Jen91, Smy83b].

Applications of Domain Theory

There is by now an enormous literature on the semantics of programming languages, much of it using substantial amounts of Domain Theory. We will simply list a number of useful textbooks: [Sch86, Ten91, Gun92b, Win93]. In addition, a number of other applications of Domain Theory have arisen: in Abstract Interpretation and static program analysis [Abr90a, BHA86, AJ91] (see also the article on Abstract Interpretation in this Handbook); databases [BDW88, BJO91]; computational linguistics [PS84, PM90]; articial intelligence [RZ94]; fractal image generation [Eda93b]; and foundations of analysis [Eda93a]. Finally, the central importance of Domain Theory is well indicated by the number of other chapters of this Handbook which make substantial reference to Domaintheoretic ideas: Topology, Algebraic Semantics, Semantics of Types, Correspondence between Operational and Denotational Semantics, Abstract Interpretation, Effective Structures.

151

References
[Abr87] [Abr90a] [Abr90b] S. Abramsky. Domain Theory and the Logic of Observable Properties. PhD thesis, University of London, 1987. S. Abramsky. Abstract interpretation, logical relations and Kan extensions. Journal of Logic and Computation, 1(1):540, 1990. S. Abramsky. A generalized Kahn principle for abstract asynchronous networks. In M. Main, A. Melton, M. Mislove, and D. Schmidt, editors, Mathematical Foundations of Programming Semantics, volume 442 of Lecture Notes in Computer Science, pages 121. Springer Verlag, 1990. S. Abramsky. The lazy lambda calculus. In D. Turner, editor, Research Topics in Functional Programming, pages 65117. Addison Wesley, 1990. S. Abramsky. A domain equation for bisimulation. Information and Computation, 92:161218, 1991. S. Abramsky. Domain theory in logical form. Annals of Pure and Applied Logic, 51:177, 1991. S. Abramsky. Interaction categories and communicating sequential processes. In A. W. Roscoe, editor, A Classical Mind: essays in honour of C. A. R. Hoare, chapter 1, pages 116. Prentice Hall International, 1994. P. Aczel. An introduction to inductive denitions. In J. Barwise, editor, The Handbook of Mathematical Logic, volume 90 of Studies in Logic and Foundations of Mathematics, pages 739782. North Holland, 1977. S. Abramsky and T. Jensen. A relational approach to strictness analysis for polymorphic functions. In Conference Record of the 18th Annual ACM Symposium on Principles of Programming Languages, pages 4954. ACM Press, 1991. S. Abramsky and R. Jagadeesan. Games and full completeness for multiplicative linear logic. Journal of Symbolic Logic, 59(2):543574, 1994. S. Abramsky and R. Jagadeesan. New foundations for the geometry of interaction. Information and Computation, 1994. To appear. J. Ad mek, E. Nelson, and J. Reiterman. Tree constructions of free cona tinuous algebras. Journal of Computer and System Sciences, 24:114146, 1982. S. Abramsky and L. Ong. Full abstraction in the lazy lambda calculus. Information and Computation, 105:159267, 1993. M. Abadi and G. D. Plotkin. A Per model of polymorphism and recursive types. In 5th Annual IEEE Symposium on Logic in Computer Science, pages 355365. IEEE Computer Society Press, 1990. 152

[Abr90c] [Abr91a] [Abr91b] [Abr94]

[Acz77]

[AJ91]

[AJ94a] [AJ94b] [ANR82]

[AO93] [AP90]

[Ard60] [AT89] [Bar84] [BDW88]

D. N. Arden. Theory of Computing Machine Design, chapter Delayed logic and nite state machines. Univ. of Michigan Press, 1960. J. Ad mek and V. Trnkov . Automata and Algebras in Categories. Kluwer a a Academic Publishers, 1989. H. P. Barendregt. The Lambda Calculus: Its Syntax and Semantics. NorthHolland, revised edition, 1984. P. Buneman, S. Davidson, and A. Watters. A semantics for complex objects and approximate queries. In Principles of Database Systems. ACM, 1988. H. Beki . Denable operations in general algebras, and the theory of auc tomata and owcharts. Technical report, IBM Laboratory, Vienna, 1969. Reprinted in H. Beki , Programming Languages and Their Denition, volc ume 177 of Lecture Notes in Computer Science, pages 3055. Springer Verlag, 1984. H. Beki . Towards a mathematical theory of programs. Technical Report c TR 25.125, IBM Laboratory, Vienna, 1971. G. Berry. Stable models of typed -calculi. In Proceedings of the 5th International Colloquium on Automata, Languages and Programming, volume 62 of Lecture Notes in Computer Science, pages 7289. Springer Verlag, 1978. G. Berry. Mod` les Compl` ment Ad quats et Stables des Lambda-calculs e e e typ s, 1979. Th` se de Doctorat dEtat, Universit Paris VII. e e e S. Berardi. Retractions on dI-domains as a model for type:type. Information and Computation, 94:204231, 1991. G. Burn, C. Hankin, and S. Abramsky. Strictness analysis for higher order functions. Science of Computer Programming, 7:249278, 1986. P. Buneman, A. Jung, and A. Ohori. Using powerdomains to generalize relational databases. Theoretical Computer Science, 91:2355, 1991. G. Boudol. Lambda calculi for (strict) parallel functions. Technical Report 1387, INRIA - Sophia Antipolis, 1991. To appear in Information and Computation. G. Boudol. Lambda calculi for (strict) parallel functions. Information and Computation, 108:51127, 1994.

[Bek69]

[Bek71] [Ber78]

[Ber79] [Ber91] [BHA86] [BJO91] [Bou91]

[Bou94]

[CGW87] T. Coquand, C. Gunter, and G. Winskel. dI-domains as a model of polymorphism. In Third Workshop on the Mathematical Foundations of Programming Language Semantics, pages 344363. Springer Verlag, 1987.

153

[Coq89] [CP92]

T. Coquand. Categories of embeddings. Theoretical Computer Science, 68:221238, 1989. R. Crole and A. M. Pitts. New foundations for xpoint computations: FIXhyperdoctrines and FIX-logic. Information and Computation, 98:171 210, 1992. P.-L. Curien. Categorical Combinators, Sequential Algorithms and Functional Programming. Progress in Theoretical Computer Science. Birkh user, second edition, 1993. a J. W. de Bakker and D. S. Scott. A theory of programs. Notes, IBM Seminar, Vienna, 1969. M. Droste and R. G bel. Universal domains in the theory of denotational o semantics of programming languages. In IEEE Symposium on Logic in Computer Science, pages 1934. IEEE Computer Society Press, 1990. M. Droste and R. G bel. Universal information systems. International o Journal of Foundations of Computer Science, 1:413424, 1991. M. Droste and R. G bel. Universal domains and the amalgamation propo erty. Mathematical Structures in Computer Science, 3:137159, 1993. R. O. Davies, A. Hayes, and G. Rousseau. Complete lattices and the generalized Cantor theorem. Proceedings of the AMS, 27:253258, 1971. B. A. Davey and H. A. Priestley. Introduction to Lattices and Order. Cambridge University Press, Cambridge, 1990. A. Edalat. Domain theory and integration. draft paper, 1993. A. Edalat. Dynamical systems, measures and fractals via domain theory: extended abstract. In G. Burn, S. Gay, and M. Ryan, editors, Theory and Formal Methods 1993, Workshops in Computing, pages 8299. Springer, 1993. T. Ehrhard. Hypercoherences: A strongly stable model of linear logic. Mathematical Structures in Computer Science, 3:365386, 1993. H. Ehrig and B. Mahr. Fundamentals of Algebraic Specication 1, volume 6 of EATCS Monographs on Theoretical Computer Science. Springer Verlag, 1985. M. Ern . Posets isomorphic to their extensions. Order, 2:199210, 1985. e Y. L. Ershov. The theory of A-spaces (English translation). Algebra and Logic, 12:209232, 1975. A. Edalat and M. B. Smyth. I-categories as a framework for solving domain equations. Theoretical Computer Science, 115(1):77106, 1993. 154

[Cur93]

[dBS69] [DG90]

[DG91] [DG93] [DHR71] [DP90] [Eda93a] [Eda93b]

[Ehr93] [EM85]

[Ern85] [Ers75] [ES93]

[Fie92]

A. Fiech. Colimits in the category CPO. Technical report, Kansas State University, 1992.

[FMRS90] P. Freyd, P. Mulry, G. Rosolini, and D. Scott. Extensional PERs. In Logic in Computer Science, pages 346354. IEEE Computer Society Press, 1990. [Fre91] P. J. Freyd. Algebraically complete categories. In A. Carboni, M. C. Pedicchio, and G. Rosolini, editors, Como Category Theory Conference, volume 1488 of Lecture Notes in Mathematics, pages 95104. Springer Verlag, 1991. P. J. Freyd. Remarks on algebraically compact categories. In M. P. Fourman, P. T. Johnstone, and A. M. Pitts, editors, Applications of Categories in Computer Science, volume 177 of L.M.S. Lecture Notes, pages 95106. Cambridge University Press, 1992.

[Fre92]

[GHK 80] G. Gierz, K. H. Hofmann, K. Keimel, J. D. Lawson, M. Mislove, and D. S. Scott. A Compendium of Continuous Lattices. Springer Verlag, 1980. [Gir86] [Gir87] [GJ88] J.-Y. Girard. The system of variable types: Fifteen years later. Theoretical Computer Science, 45:159192, 1986. J.-Y. Girard. Linear logic. Theoretical Computer Science, 50:1102, 1987. C. Gunter and A. Jung. Coherence and consistency in domains. In Third Annual Symposium on Logic in Computer Science, pages 309317. IEEE Computer Society Press, 1988. S. Ginsburg and H. G. Rice. Two families of languages related to ALGOL. Journal of the ACM, 9:350371, 1962. S. Graham. Closure properties of a probabilistic powerdomain construction. In M. Main, A. Melton, M. Mislove, and D. Schmidt, editors, Mathematical Foundations of Programming Language Semantics, volume 298 of Lecture Notes in Computer Science, pages 213233. Springer Verlag, 1988. C. Gunter and D. S. Scott. 12: Semantic Domains. In Jan van Leeuwen, editor, Handbook of Theoretical Computer Science, volume B: Formal Models and Semantics, pages 633674. Elsevier Science Publishers, Amsterdam, 1990. J. A. Goguen, J. W. Thatcher, and E. G. Wagner. An initial algebra approach to the specication, correctness and implementation of abstract data types. In R. T. Yeh, editor, Current Trends in Programming Methodology IV: Data Structuring, pages 80144. Prentice Hall, 1978. C. Gunter. The largest rst-order axiomatizable cartesian closed category of domains. In Symposium on Logic in Computer Science, pages 142148. IEEE Computer Society Press, 1986. 155

[GR62] [Gra88]

[GS90]

[GTW78]

[Gun86]

[Gun92a] [Gun92b] [Hau14] [Hec90] [Hec91] [Hec93a]

C. Gunter. The mixed power domain. Theoretical Computer Science, 103:311334, 1992. C. Gunter. Semantics of Programming Languages. Structures and Techniques. Foundations of Computing. MIT Press, 1992. F. Hausdorff. Grundzuge der Mengenlehre. Veit, 1914. R. Heckmann. Power Domain Constructions. PhD thesis, Universit t des a Saarlandes, 1990. R. Heckmann. Power domain constructions. Science of Computer Programming, 17:77117, 1991. R. Heckmann. Observable modules and power domain constructions. In M. Droste and Y. Gurevich, editors, Semantics of Programming Languages and Model Theory, volume 5 of Algebra, Logic, and Applications, pages 159187. Gordon and Breach Science Publishers, 1993. R. Heckmann. Power domains and second order predicates. Theoretical Computer Science, 111:5988, 1993. M. Hennessy. A fully abstract denotational model for higher order processes. In Eighth Annual IEEE Symposium on Logic in Computer Science, pages 397408. IEEE Computer Society Press, 1993. K. H. Hofmann and M. Mislove. Local compactness and continuous lattices. In B. Banaschewski and R.-E. Hoffmann, editors, Continuous Lattices, Proceedings Bremen 1979, volume 871 of Lecture Notes in Mathematics, pages 209248. Springer Verlag, 1981. K. H. Hofmann and M. W. Mislove. All compact Hausdorff lambda models are degenerate. Technical Report 931, Tulane University, 1993. R.-E. Hoffmann. Continuous posets, prime spectra of completely distributive complete lattices, and Hausdorff compactication. In B. Banaschewski and R.-E. Hoffmann, editors, Continuous Lattices, Proceedings Bremen 1979, volume 871 of Lecture Notes in Mathematics, pages 159208. Springer Verlag, 1981. R. Hoofman. Non-stable models of linear logic. PhD thesis, University of Utrecht, 1992. M. C. B. Hennessy and G. D. Plotkin. Full abstraction for a simple parallel programming language. In J. Becvar, editor, Mathematical Foundations of Computer Science, volume 74 of Lecture Notes in Computer Science. Springer Verlag, 1979. H. Huwig and A. Poign . A note on inconsistencies caused by xpoints in e a cartesian closed category. Theoretical Computer Science, 73:101112, 1990. 156

[Hec93b] [Hen93]

[HM81]

[HM93] [Hof81]

[Hoo92] [HP79]

[HP90]

[Hrb87] [Hrb88]

K. Hrbacek. Convex powerdomains I. Information and Computation, 74:198225, 1987. K. Hrbacek. A powerdomain construction. In M. Main, A. Melton, M. Mislove, and D. Schmidt, editors, Mathematical Foundations of Programming Semantics, volume 298 of Lecture Notes in Computer Science, pages 200212. Springer Verlag, 1988. K. Hrbacek. Convex powerdomains II. Information and Computation, 81:290317, 1989. H. Herrlich and G. E. Strecker. Category Theory. Allyn and Bacon, 1973. C. Hosono and M. Sato. The retracts in do not form a continuous lattice a solution to Scotts problem. Theoretical Computer Science, 4:137142, 1977. M. Huth. Cartesian closed categories of domains and the space Proj( ). In S. Brookes, M. Main, A. Melton, M. Mislove, and D. Schmidt, editors, Mathematical Foundations of Programming Semantics, volume 598 of Lecture Notes in Computer Science, pages 259271. Springer Verlag, 1992. M. Huth. Linear Domains and Linear Maps. In S. Brookes, M. Main, A. Melton, M. Mislove, and D. Schmidt, editors, Mathematical Foundations of Programming Semantics, volume 802 of Lecture Notes in Computer Science, pages 438453. Springer Verlag, 1994. J. M. E. Hyland. First steps in synthetic domain theory. In A. Carboni, C. Pedicchio, and G. Rosolini, editors, Conference on Category Theory 1990, volume 1488 of Lecture Notes in Mathematics, pages 131156. Springer Verlag, 1991. J. Isbell. Atomless parts of spaces. Mathematica Scandinavica, 31:532, 1972. J. Isbell. Completion of a construction of Johnstone. Proceedings of the American Mathematical Society, 85:333334, 1982. T. Iwamura. A lemma on directed sets. Zenkoku Shijo Sugaku Danwakai, 262:107111, 1944. (In Japanese). T. Jensen. Strictness analysis in logical form. In J. Hughes, editor, Fifth ACM Conference on Functional Programming Languages and Computer Architecture, volume 523 of LNCS. Springer Verlag, 1991. T. Jensen. Disjunctive strictness analysis. In Seventh Annual IEEE Symposium on Logic in Computer Science, pages 174185. IEEE Computer Society Press, 1992.

[Hrb89] [HS73] [HS77]

[Hut92]

[Hut94]

[Hyl91]

[Isb72] [Isb82] [Iwa44] [Jen91]

[Jen92]

157

[Joh81]

P. T. Johnstone. Scott is not always sober. In B. Banaschewski and R.-E. Hoffmann, editors, Continuous Lattices, Proceedings Bremen 1979, volume 871 of Lecture Notes in Mathematics, pages 282283. Springer Verlag, 1981. P. T. Johnstone. Stone Spaces, volume 3 of Cambridge Studies in Advanced Mathematics. Cambridge University Press, Cambridge, 1982. P. T. Johnstone. The point of pointless topology. Bulletin of the American Mathematical Society, 8:4153, 1983. C. Jones. Probabilistic Non-Determinism. PhD thesis, University of Edinburgh, Edinburgh, 1990. Also published as Technical Report No. CST63-90. C. Jones and G. Plotkin. A probabilistic powerdomain of evaluations. In Proceedings of the 4th Annual Symposium on Logic in Computer Science, pages 186195. IEEE Computer Society Press, 1989. A. Jung. New results on hierarchies of domains. In M. Main, A. Melton, M. Mislove, and D. Schmidt, editors, Mathematical Foundations of Programming Language Semantics, volume 298 of Lecture Notes in Computer Science, pages 303310. Springer Verlag, 1988. A. Jung. Cartesian Closed Categories of Domains, volume 66 of CWI Tracts. Centrum voor Wiskunde en Informatica, Amsterdam, 1989. 107 pp. A. Jung. The classication of continuous domains. In Proceedings, Fifth Annual IEEE Symposium on Logic in Computer Science, pages 3540. IEEE Computer Society Press, 1990. A. Kanda. Fully effective solutions of recursive domain equations. In J. Becvar, editor, Mathematical Foundations of Computer Science, vol ume 74. Springer Verlag, 1979. Lecture Notes in Computer Science. S. C. Kleene. Introduction to Metamathematics. Van Nostrand, 1952. A. Kock. Monads on symmetric monoidal closed categories. Archiv der Mathematik, 21:110, 1970. A. Kock. Strong functors and monoidal monads. Archiv der Mathematik, 23:113120, 1972. A. Kock. Synthetic Differential Geometry, volume 51 of LMS Lecture Notes. Cambridge University Press, 1981. G. Kahn and G. Plotkin. Concrete domains. Theoretical Computer Science, 121:187277, 1993. Translation of a technical report from 1978.

[Joh82] [Joh83] [Jon90]

[JP89]

[Jun88]

[Jun89]

[Jun90]

[Kan79]

[Kle52] [Koc70] [Koc72] [Koc81] [KP93]

158

[KP94] [Kra39]

K. Keimel and J. Paseka. A direct proof of the Hofmann-Mislove theorem. Proceedings of the AMS, 120:301303, 1994. M. Krasner. Un type densembles semi-ordon s et ses rapports avec une e hypoth se de M. A. Weil. Bulletin de la Societ Mathematique de France, e e 67:162176, 1939. J. D. Lawson. The duality of continuous posets. Houston Journal of Mathematics, 5:357394, 1979. J.-L. Lassez, V. L. Nguyen, and E. A. Sonenberg. Fixed point theorems and semantics: A folk tale. Information Processing Letters, 14:112116, 1982. D. J. Lehmann and M. B. Smyth. Algebraic specication of data types: A synthetic approach. Mathematical Systems Theory, 14:97139, 1981. M. Main. Free constructions of powerdomains. In A. Melton, editor, Mathematical Foundations of Programming Semantics, volume 239 of Lecture Notes in Computer Science, pages 162183. Springer Verlag, 1985. G. Markowsky. Chain-complete p.o. sets and directed sets with applications. Algebra Universalis, 6:5368, 1976. G. Markowsky. Categories of chain-complete posets. Theoretical Computer Science, 4:125135, 1977. G. Markowsky. A motivation and generalization of Scotts notion of a continuous lattice. In B. Banaschewski and R.-E. Hoffmann, editors, Continuous Lattices, volume 871 of Lecture Notes in Mathematics, pages 298 307. Springer Verlag, 1981. D. C. McCarty. Realizability and Recursive Mathematics. PhD thesis, Oxford University, 1984. J. Meseguer. On order-complete universal algebra and enriched functorial semantics. In M. Karpi ski, editor, Fundamentals of Computation Then ory, volume 56 of Lecture Notes in Computer Science, pages 294301. Springer Verlag, 1977. E. Moggi. Notions of computations and monads. Information and Computation, 93(1):5592, 1991. D. MacQueen, G. D. Plotkin, and R. Sethi. An ideal model for recursive polymorphic types. Information and Control, 71:95130, 1986. R. E. Milne and C. Strachey. A Theory of Programming Language Semantics. Chapman and Hall, London, 1976. R. Milner and M. Tofte. Commentary on Standard ML. MIT Press, 1991.

[Law79] [LNS82]

[LS81] [Mai85]

[Mar76] [Mar77] [Mar81]

[McC84] [Mes77]

[Mog91] [MPS86] [MS76] [MT91]

159

[MTH90] [Nel81]

R. Milner, M. Tofte, and R. Harper. The Denition of Standard ML. MIT Press, 1990. E. Nelson. Free Z-continuous algebras. In B. Banaschewski and R. E. Hoffmann, editors, Continuous Lattices, volume 871 of Lecture Notes in Mathematics, pages 315334. Springer Verlag, 1981. M. Nivat and J. C. Reynolds, editors. Algebraic Methods in Computer Science. Cambridge University Press, 1985. K.-J. N ler. Ordnungstheoretische Modelle fur nicht-deterministische u Programmiersprachen. PhD thesis, Universit t GH Essen, 1992. a C.-H. L. Ong. Nondeterminism in a functional setting. In Eighth Annual IEEE Symposium on Logic in Computer Science, pages 275286. IEEE Computer Society Press, 1993. D. M. Park. Fixpoint induction and proofs of program properties. In B. Meltzer and D. Michie, editors, Machine Intelligence, volume 5, pages 5978. Edinburgh University Press, 1969. W. Phoa. Domain Theory in Realizability Toposes. PhD thesis, University of Cambridge, 1991. A. M. Pitts. Relational properties of domains. Technical Report 321, Cambridge Univ. Computer Laboratory, December 1993. 37 pages. A. M. Pitts. Relational properties of recursively dened domains. In 8th Annual Symposium on Logic in Computer Science, pages 8697. IEEE Computer Society Press, Washington, 1993. A. M. Pitts. A co-induction principle for recursively dened domains. Theoretical Computer Science, 124:195219, 1994. R. Platek. New foundations for recursion theory. PhD thesis, Stanford University, 1964. G. D. Plotkin. A powerdomain construction. SIAM Journal on Computing, 5:452487, 1976. G. D. Plotkin. Post-graduate lecture notes in advanced domain theory (incorporating the Pisa Notes). Dept. of Computer Science, Univ. of Edinburgh. Available from http://www.dcs.ed.ac.uk/home/gdp/ publications/, 1981. G. D. Plotkin. Lectures on predomains and partial functions. Notes for a course given at the Center for the Study of Language and Information, Stanford 1985, 1985.

[NR85] [N 92] u [Ong93]

[Par69]

[Pho91] [Pit93a] [Pit93b]

[Pit94] [Pla64] [Plo76] [Plo81]

[Plo85]

160

[Plo93]

G. D. Plotkin. Type theory and recursion. In Eighth Annual IEEE Symposium on Logic in Computer Science, page 374. IEEE Computer Society Press, 1993. C. Pollard and D. Moshier. Unifying partial descriptions of sets. In P. Hanson, editor, Information, Language and Cognition, volume 1 of Vancouver Studies in Cognitive Science. University of British Columbia Press, 1990. A. Poign . Basic category theory. In S. Abramsky, D. M. Gabbay, and e T. S. E. Maibaum, editors, Handbook of Logic in Computer Science, vol. 1, pages 413640. Clarendon Press, 1992. F. Pereira and S. Shieber. The semantics of grammar formalisms seen as computer languages. In Proceedings of the 10th International Conference on Computational Linguistics, pages 123129. Association for Computational Linguistics, 1984. H. Puhlmann. The snack powerdomain for database semantics. In A. M. Borzyszkowski and S. Sokoowski, editors, Mathematical Foundations of Computer Science, volume 711 of Lecture Notes in Computer Science, pages 650659, 1993. G. N. Raney. A subdirect-union representation for completely distributive complete lattices. Proceedings of the AMS, 4:518522, 1953. G. N. Raney. Tight Galois connections and complete distributivity. Trans. AMS, 97:418426, 1960. G. Rosolini. Continuity and Effectiveness in Topoi. PhD thesis, Oxford University, 1986. E. Robinson and G. Rosolini. Categories of partial maps. Information and Computation, 79:95130, 1988. W. Rounds and G. Q. Zhang. Domain theory meets default logic. Journal of Logic and Computation, 4:124, 1994. D. A. Schmidt. Denotational Semantics. Allyn and Bacon, 1986. A. Schalk. Algebras for Generalized Power Constructions. Doctoral thesis, Technische Hochschule Darmstadt, 1993. 166 pp. D. S. Scott. A type theoretic alternative to ISWIM, CUCH, OWHY. Manuscript, University of Oxford, 1969. D. S. Scott. Outline of a mathematical theory of computation. In 4th Annual Princeton Conference on Information Sciences and Systems, pages 169176, 1970.

[PM90]

[Poi92]

[PS84]

[Puh93]

[Ran53] [Ran60] [Ros86] [RR88] [RZ94] [Sch86] [Sch93] [Sco69] [Sco70]

161

[Sco71]

D. S. Scott. The lattice of ow diagrams. In E. Engeler, editor, Symposium on Semantics of Algorithmic Languages, volume 188 of Lecture Notes in Mathematics, pages 311366. Springer Verlag, 1971. D. S. Scott. Continuous lattices. In E. Lawvere, editor, Toposes, Algebraic Geometry and Logic, volume 274 of Lecture Notes in Mathematics, pages 97136. Springer Verlag, 1972. D. S. Scott. Data types as lattices. SIAM J. Computing, 5:522587, 1976. D. S. Scott. Relating theories of lambda calculus. In J. R. Hindley and J. P. Seldin, editors, To H. B. Curry: Essays in Combinatory Logic, Lambda Calculus and Formalism, pages 403450. Academic Press, 1980. D. S. Scott. Lectures on a mathematical theory of computation. Monograph PRG-19, Oxford University Computing Laboratory, Oxford, 1981. D. S. Scott. Domains for denotational semantics. In M. Nielson and E. M. Schmidt, editors, International Colloquium on Automata, Languages and Programs, volume 140 of Lecture Notes in Computer Science, pages 577 613. Springer Verlag, 1982. D. S. Scott. A type-theoretical alternative to ISWIM, CUCH, OWHY. Theoretical Computer Science, 121:411440, 1993. Reprint of a manuscript written in 1969. N. Saheb-Djahromi. CPOs of measures for nondeterminism. Theoretical Computer Science, 12:1937, 1980. M. B. Smyth. Effectively given domains. Theoretical Computer Science, 5:257274, 1977. M. B. Smyth. Powerdomains. Journal of Computer and Systems Sciences, 16:2336, 1978. M. B. Smyth. The largest cartesian closed category of domains. Theoretical Computer Science, 27:109119, 1983.

[Sco72]

[Sco76] [Sco80]

[Sco81] [Sco82]

[Sco93]

[SD80] [Smy77] [Smy78] [Smy83a]

[Smy83b] M. B. Smyth. Powerdomains and predicate transformers: a topological view. In J. Diaz, editor, Automata, Languages and Programming, volume 154 of Lecture Notes in Computer Science, pages 662675. Springer Verlag, 1983. [Smy86] M. B. Smyth. Finite approximation of spaces. In D. Pitt, S. Abramsky, A. Poign , and D. Rydeheard, editors, Category Theory and Computer e Programming, volume 240 of Lecture Notes in Computer Science, pages 225241. Springer Verlag, 1986. M. B. Smyth. Topology. In S. Abramsky, D. M. Gabbay, and T. S. E. Maibaum, editors, Handbook of Logic in Computer Science, vol. 1, pages 641761. Clarendon Press, 1992. 162

[Smy92]

[SP82] [Sto77] [Sto88] [Tay87] [Tay90] [Tay91] [Ten91] [TT93]

M. B. Smyth and G. D. Plotkin. The category-theoretic solution of recursive domain equations. SIAM J. Computing, 11:761783, 1982. J. E. Stoy. Denotational Semantics: The Scott-Strachey Approach to Programming Language Theory. The MIT Press, 1977. A. Stoughton. Fully Abstract Models of Programming Languages. Research Notes in Theoretical Computer Science. Pitman/Wiley, 1988. P. Taylor. Recursive Domains, Indexed category Theory and Polymorphism. PhD thesis, Cambridge University, 1987. P. Taylor. An algebraic approach to stable domains. Pure and Applied Algebra, 64, 1990. P. Taylor. The xed point property in synthetic domain theory. In 6th LICS conference, pages 152160. IEEE Computer Society Press, 1991. R. D. Tennent. Semantics of Programming Languages. Prentice Hall, 1991. M. Tischendorf and J. T ma. The characterization of congruence lattices u of lattices. Technical Report 1559, Technische Hochschule Darmstadt, 1993. S. J. Vickers. Topology Via Logic, volume 5 of Cambridge Tracts in Theoretical Computer Science. Cambridge University Press, 1989. L. Vietoris. Stetige Mengen. Monatshefte fur Mathematik und Physik, 31:173204, 1921. L. Vietoris. Bereiche zweiter Ordnung. Monatshefte fur Mathematik und Physik, 32:258280, 1922. K. Weihrauch and T. Deil. Berechenbarkeit auf cpos. Technical Report 63, Rheinisch-Westf lische Technische Hochschule Aachen, 1980. a G. Winskel. Powerdomains and modality. In M. Karpinski, editor, Foundations of Computation Theory, volume 158 of Lecture Notes in Computer Science, pages 505514. Springer Verlag, 1983. G. Winskel. An introduction to event structures. In J. W. de Bakker, editor, Linear Time, Branching Time, and Partial Order in Logics and Models for Concurrency, volume 354 of Lecture Notes in Computer Science, pages 364399. Springer Verlag, 1988. G. Winskel. The Formal Semantics of Programming Languages. An Introduction. MIT Press, 1993. G.-Q. Zhang. Logic of Domains. Progress in Theoretical Computer Science. Birkh user, 1991. a

[Vic89] [Vie21] [Vie22] [WD80] [Win83]

[Win88]

[Win93] [Zha91]

163

Index
, 11 -generated lattice, 107 -irreducible, 107 -prime, 107 F-algebra, 69 abstract basis, 26 adjunction, 38 admissible action on relations, 80 admissible predicate, 17 admissible predicates, 80 admissible relation, 80 algebraic dcpo, 19 algebraic domain, 19 approximable relation, 28 approximants to relative to , 18 arithmetic lattice, 116 basis, 18, 26 bc-domain, 55 binite domain, 58 bilimit, 50 bottom, 12 bounded-complete domain, 55 cartesian product, 42 chain, 13 closure operator, 41 closure system, 12 coalesced sum, 44 conal subset, 14 coherent algebraic prelocale, 122 coherent domains, 62 coherent space, 116 compact element, 18 complete lattice, 12 completely -irreducible, 107 completely -prime, 107 completely distributive lattice, 106 completely prime lter, 109 continuous dcpo, 19
"

 ! % 
, 12
I " V

, 26

continuous domain, 19 continuous embedding projection pair, 39 continuous function, 15 continuous functor, 70 continuous insertion closure pairs, 40 continuous section retraction pair, 34 convex hull, 94 convex powertheory, 92 convex set, 94 dcpo, 14 dcpo-algebras, 85 dcpo-semilattice, 92 deationary semilattice, 95 directed set, 13 directed-complete partial order, 14 distributive lattice, 106 domain logic, 140 domain prelocale, 123 domain-algebras, 86 e-p-pair, 39 effectively given, 144 Egli-Milner relation, 93 expanding sequence, 48 expanding system, 50 extensional action on relations, 82 lter, 13 ltered set, 13 nite amalgam, 65 nite element, 18 nite mub property, 58 nitely separated function, 60 at domains, 20 formal union, 92 frame, 108, 113 frame distributivity law, 108 frame-homomorphism, 108 FS-domain, 60 Hasse diagrams, 10

164

Hoare powerdomain, 95 Hoare powertheory, 95 i-c-pairs, 40 ideal, 13 ideal completion, 26 idempotent deation, 59 index set of a net, 14 inductive closure system, 19 inmum, 12 inationary semilattices, 95 interpolation property, 24 irreducible closed set, 111 isolated element, 18 join, 12 join-approximable relation, 117 joinable family, 58 kernel operator, 40 L-domain, 55 largest element, 12 lattice, 12 Lawson-topology, 63 least element, 12 lenses, 98 lifting, 45 limit-colimit coincidence, 49 line diagrams, 10 localic description, 123 locally continuous functor, 71 logical action on relations, 81 lower adjoint, 38 lower powerdomain, 95 lower powertheory, 95 lower set, 11 maximal element, 11 meet, 12 minimal upper bound, 11 monotone function, 13 monotone function space, 13 monotone net, 14 monotone retract, 34 monotone section retraction pair, 34 mub-closure, 57 165

mub-complete poset, 57 multiplicative, 116 order of approximation, 18 partially ordered set, 10 Plotkin powerdomain, 92 Plotkin powerlocale, 132 Plotkin powertheory, 92 Plotkin-order, 58 point of a complete lattice, 110 pointed poset, 12 pointwise order, 13 poset, 10 pre-isomorphism, 123 preorders, 10 prime lter, 109 prime-continuous lattice, 106 principal lter, 13 principal ideal, 13 projection, 40 property m, 57 quotient domain, 41 reduct, 103 round ideal, 26 s-r-pair, 34 saturated set, 61 schizophrenic object, 110 Scott-closed set, 29 Scott-domain, 56 Scott-open set, 29 Scott-topology, 29 semilattice, 12 smash product, 45 Smyth powerdomain, 95 sober, 112 soberication, 113 spatial, 112 specialization order, 114 spectral space, 116 spectrum, 117 step function, 54 strict dcpo-algebra, 91 strict function, 15

strict homomorphism, 91 strongly separated function, 60 sub-domain, 40 sub-prelocale, 124 subnet, 14 supremum, 12 supremum of a net, 14 topological Egli-Milner ordering, 98 upper adjoint, 38 upper bound, 11 upper powerdomain, 95 upper set, 11 weak upper topology, 114

166