MATHEMATICAL

ASPECTS OF LOGIC
PROGRAMMING
SEMANTICS
 Chapman & Hall/CRC

Studies in Informatics Series

�

SERIES EDITOR
G. Q. Zhang
Department of EECS
�
Case Western Reserve University
�
Cleveland, Ohio, U.S.A.
�

PUBLISHED TITLES

Stochastic Relations: Foundations for Markov Transition Systems

Ernst-Erich Doberkat

Conceptual Structures in Practice

Pascal Hitzler and Henrik Schärfe

Context-Aware Computing and Self-Managing Systems

Waltenegus Dargie

Introduction to Mathematics of Satisfiability

Victor W. Marek

Ubiquitous Multimedia Computing

Qing Li and Timothy K. Shih

Mathematical Aspects of Logic Programming Semantics

Pascal Hitzler and Anthony Seda
 Chapman & Hall/CRC
Studies in Informatics Series

MATHEMATICAL
ASPECTS OF LOGIC
PROGRAMMING
SEMANTICS

Pascal Hitzler
Kno.e.sis Center at
Wright State University
Dayton, Ohio, USA

antHony seda
University College Cork
Ireland
CRC Press
Boca
TaylorRaton, FL 33487-2742
& Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 33487-2742
© 2011 by Taylor and Francis Group, LLC
No
CRCclaim
Presstoisoriginal U.S.ofGovernment
an imprint worksGroup, an Informa business
Taylor & Francis

No claim to original U.S. Government works

Printed in the United States of America on acid-free paper

10 9 8 7 6 5 4 3 2 1

International Standard Book Number: 978-1-4398-2961-5 (Hardback)

assume
This bookresponsibility for the validity
contains information of all materials
obtained or the consequences
from authentic of their use.
and highly regarded The authors
sources. and
Reasonable
publishers
efforts havehave
beenattempted to trace reliable
made to publish the copyright holders
data and of all material
information, but thereproduced
author andinpublisher
this publication
cannot
and apologize
assume to copyright
responsibility holders
for the if of
validity permission to publish
all materials in this form has
or the consequences not been
of their obtained.
use. The If and
authors any
copyright
publishersmaterial has notto
have attempted been acknowledged
trace the copyrightplease
holderswrite and
of all let us know
material so we may
reproduced rectify
in this in any
publication
future reprint.to copyright holders if permission to publish in this form has not been obtained. If any
and apologize
copyright material has not been acknowledged please write and let us know so we may rectify in any
Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced,
future reprint.
transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or
hereafter
The Open
Except as invented, including
Access version
permitted photocopying,
underofU.S.
this book, available
Copyright microfilming,
Law, at part
no ofand
thisrecording,
book mayor
www.taylorfrancis.com, inreprinted,
behasany information
been made stor-
available
reproduced,
under
age or aretrieval
Creativesystem,
Commons Attribution-Non
without Commercial-No
written permission Derivatives 4.0 license.
from the publishers.

For permission
age or to photocopy
retrieval system, or written
without use material electronically
permission from thefrom this work, please access www.copy-
publishers.
right.com (http://www.copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC), 222
Rosewood
For Drive,
permission toDanvers, MA
photocopy or 01923, 978-750-8400.
use material CCC from
electronically is a not-for-profit organization
this work, please that pro-
access www.copy-
vides licenses
right.com and registration for a variety
(http://www.copyright.com/) of users.the
or contact ForCopyright
organizations that have
Clearance beenInc.
Center, granted a pho-
(CCC), 222
tocopy license by the CCC, a separate system of payment has been arranged.

Trademark Notice:
tocopy license by theProduct
CCC, a or corporate
separate names
system may be trademarks
of payment or registered trademarks, and are
has been arranged.
used only for identification and explanation without intent to infringe.
Trademark Notice:
Visit the Taylor & Francis Web site at
http://www.taylorandfrancis.com
and the CRC Press Web site at
http://www.crcpress.com
Dedication

To Anne, to Martine, and to the memory of Barbara and Ellen Lucille

Contents

List of Figures xi

List of Tables xiii

Preface xv

Introduction xix

About the Authors xxix

1 Order and Logic 1

1.1 Ordered Sets and Fixed-Point Theorems . . . . . . . . . . . 1

1.2 First-Order Predicate Logic . . . . . . . . . . . . . . . . . . . 7

1.3 Ordered Spaces of Valuations . . . . . . . . . . . . . . . . . . 12

2 The Semantics of Logic Programs 23

2.1 Logic Programs and Their Models . . . . . . . . . . . . . . . 23

2.2 Supported Models . . . . . . . . . . . . . . . . . . . . . . . . 28

2.3 Stable Models . . . . . . . . . . . . . . . . . . . . . . . . . . 32

2.4 Fitting Models . . . . . . . . . . . . . . . . . . . . . . . . . . 37

2.5 Perfect Models . . . . . . . . . . . . . . . . . . . . . . . . . . 43

2.6 Well-Founded Models . . . . . . . . . . . . . . . . . . . . . . 56

3 Topology and Logic Programming 65

3.1 Convergence Spaces and Convergence Classes
. . . . . . . . . 66

3.2 The Scott Topology on Spaces of Valuations

. . . . . . . . . 69

3.3 The Cantor Topology on Spaces of Valuations

. . . . . . . . 76

3.4 Operators on Spaces of Valuations Revisited

. . . . . . . . . 83

4 Fixed-Point Theory for Generalized Metric Spaces 87

4.1 Distance Functions in General . . . . . . . . . . . . . . . . . 88

4.2 Metrics and Their Generalizations . . . . . . . . . . . . . . . 91

4.3 Generalized Ultrametrics . . . . . . . . . . . . . . . . . . . . 97

vii
viii Contents

4.4 Dislocated Metrics . . . . . . . . . . . . . . . . . . . . . . . . 102

4.5 Dislocated Generalized Ultrametrics . . . . . . . . . . . . . . 104

4.6 Quasimetrics . . . . . . . . . . . . . . . . . . . . . . . . . . . 106

4.7 A Hierarchy of Fixed-Point Theorems . . . . . . . . . . . . . 112

4.8 Relationships Between the Various Spaces . . . . . . . . . . . 114

4.9 Fixed-Point Theory for Multivalued Mappings . . . . . . . . 125

4.10 Partial Orders and Multivalued Mappings . . . . . . . . . . . 127

4.11 Metrics and Multivalued Mappings . . . . . . . . . . . . . . 129

4.12 Generalized Ultrametrics and Multivalued Mappings . . . . . 129

4.13 Quasimetrics and Multivalued Mappings . . . . . . . . . . . 132

4.14 An Alternative to Multivalued Mappings . . . . . . . . . . . 136

5 Supported Model Semantics 139

5.1 Two-Valued Supported Models . . . . . . . . . . . . . . . . . 140

5.2 Three-Valued Supported Models . . . . . . . . . . . . . . . . 151

5.3 A Hierarchy of Logic Programs . . . . . . . . . . . . . . . . . 159

5.4 Consequence Operators and Fitting-Style Operators . . . . . 161

5.5 Measurability Considerations . . . . . . . . . . . . . . . . . . 166

6 Stable and Perfect Model Semantics 169

6.1 The Fixpoint Completion . . . . . . . . . . . . . . . . . . . . 169

6.2 Stable Model Semantics . . . . . . . . . . . . . . . . . . . . . 171

6.3 Perfect Model Semantics . . . . . . . . . . . . . . . . . . . . 175

7 Logic Programming and Artificial Neural Networks 185

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 185

7.2 Basics of Artificial Neural Networks . . . . . . . . . . . . . . 188

7.3 The Core Method as a General Approach to Integration . . . 191

7.4 Propositional Programs . . . . . . . . . . . . . . . . . . . . . 192

7.5 First-Order Programs . . . . . . . . . . . . . . . . . . . . . . 196

7.6 Some Extensions – The Propositional Case . . . . . . . . . . 212

7.7 Some Extensions – The First-Order Case . . . . . . . . . . . 218

8 Final Thoughts 221

8.1 Foundations of Programming Semantics . . . . . . . . . . . . 221

8.2 Quantitative Domain Theory . . . . . . . . . . . . . . . . . . 222

8.3 Fixed-Point Theorems for Generalized Metric Spaces . . . . 223

8.4 The Foundations of Knowledge Representation and Reasoning 223

8.5 Clarifying Logic Programming Semantics . . . . . . . . . . . 224

8.6 Symbolic and Subsymbolic Representations . . . . . . . . . . 225

8.7 Neural-Symbolic Integration . . . . . . . . . . . . . . . . . . 225

8.8 Topology, Programming, and Artificial Intelligence . . . . . . 226

 Contents ix

Appendix: Transfinite Induction and General Topology 229

A.1 The Principle of Transfinite Induction . . . . . . . . . . . . . 229

A.2 Basic Concepts from General Topology . . . . . . . . . . . . 234

A.3 Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . 237

A.4 Separation Properties and Compactness . . . . . . . . . . . . 238

A.5 Subspaces and Products . . . . . . . . . . . . . . . . . . . . . 239

A.6 The Scott Topology . . . . . . . . . . . . . . . . . . . . . . . 240

Bibliography 243

Index 265

List of Figures

1.1 Hasse diagrams for T HREE and FOU R. . . . . . . . . . . . 16

2.1 Dependency graph for P1 . . . . . . . . . . . . . . . . . . . . . 48

2.2 Dependency graph for P2 . . . . . . . . . . . . . . . . . . . . . 48

4.1 Dependencies between single-valued fixed-point theorems. . . 113

5.1 The main classes of programs discussed in this book. . . . . . 160

7.1 The neural-symbolic cycle. . . . . . . . . . . . . . . . . . . . . 187

7.2 Unit Nk in a connectionist network. . . . . . . . . . . . . . . 188

7.3 Sketch of a 3-layer recurrent network. . . . . . . . . . . . . . 191

7.4 Two 3-layer feedforward networks of binary threshold units. . 194

7.5 Transforming TP into fP . . . . . . . . . . . . . . . . . . . . . 198

7.6 The embedding of the TP -operator for Program 7.5.1. . . . . 199

7.7 The networks from Example 7.5.7. . . . . . . . . . . . . . . . 202

7.8 The embedding and approximation of a TP -operator. . . . . . 202

7.9 An approximating sigmoidal network for Program 7.5.1. . . . 204

7.10 An approximation using the raised cosine function. . . . . . . 205

7.11 An RBF network approximating a TP -operator. . . . . . . . . 206

7.12 A two-dimensional version of the Cantor set. . . . . . . . . . 207

7.13 A construction of the two-dimensional Cantor set. . . . . . . 207

7.14 A vector-based approximating network. . . . . . . . . . . . . 208

7.15 A conjunction unit for FOU R. . . . . . . . . . . . . . . . . . 216

xi
List of Tables

1.1 Belnap’s four-valued logic. . . . . . . . . . . . . . . . . . . . . 11

4.1 Generalized metrics: Definition 4.2.1. . . . . . . . . . . . . . . 92

4.2 (Dislocated) generalized ultrametrics: Definition 4.3.1. . . . . 98

4.3 Summary of single-valued fixed-point theorems. . . . . . . . . 113

5.1 Several truth tables for three-valued logics. . . . . . . . . . . 153

xiii
Preface

This book presents a rigorous, comprehensive, modern, and detailed account

of the mathematical methods and tools required for the semantic analysis of
logic programs. It is, in part, the outcome of a fruitful research collabora­
tion between the authors over the last decade or so and contains many of
the results we obtained during that period. In addition, it discusses the work
of many other authors and places it within the overall context of the sub­
ject matter. A major feature of the book is that it significantly extends the
tools and methods from the order theory traditionally used in the subject
to include non-traditional methods from mathematical analysis depending on
topology, generalized distance functions, and their associated fixed-point the­
ory. The need for such methods arises for several reasons. One reason is the
non-monotonicity of some important semantic operators, associated with logic
programs, when negation is included in the syntax of the underlying language,
and another arises in the context of neural-symbolic integration, as discussed
briefly in the next paragraph and in more detail in the Introduction. Fur­
thermore, it is our belief that certain of our results, although here focused on
logic programming, have much wider applicability and should prove useful in
other parts of theoretical computer science not immediately related to logic
programming. However, we do not discuss this issue in the book in detail and
instead we give references to the literature at appropriate places in the text
in order to aid readers interested in investigating this point more thoroughly.
All the well-known, important semantics in logic programming are devel­
oped in the book from a unified point of view using both order theory and the
non-traditional methods just alluded to, and this provides an illustration of
the main objectives of the book. In addition, the interrelationships between
the various semantics are closely examined. Moreover, a significant amount of
space is devoted to examining the integration of logic programming and con­
nectionist systems (or neural networks) from the point of view of semantics.
Indeed, in the wide sense of integrating discrete models of computation with
continuous models, one can expect to employ a mix of mathematical tools of
both a discrete and continuous nature, as illustrated by the particular choice
of models we make here. Therefore, there is a need in the study of the seman­
tics of logic programming (and in the study of general models of computation)
for a self-contained and detailed exposition of the development of both con­
ventional and non-conventional methods and techniques, as just explained,
and their interaction. This book sets out to provide such an exposition, at

xv
xvi Preface

least in part, and is, we believe, unique in its content and coverage and fills a
significant gap in the literature on theoretical computer science.
The book is mainly aimed at advanced undergraduate students, gradu­
ate students, and researchers interested in the interface between mathemat­
ics and computer science. It presents material from the early days of logic
programming through to topics which are of current importance. It should
be of special interest to those engaged in the foundations of logic program­
ming, theoretical aspects of knowledge representation and reasoning, artifi­
cial intelligence, the integration of logic-based systems with other models of
computation, logic in computer science, semantics of computation, and re­
lated topics. The book should also prove to be of interest to those engaged
in domain theory and in applications of general topology to computer sci­
ence. Indeed, it carries out for logic programming semantics, in a general
model-building sense, something akin to what the well-known treatments of
Abramsky and Jung [Abramsky and Jung, 1994] and Stoltenberg-Hansen et
al. [Stoltenberg-Hansen et al., 1994] set out to do for the semantics of conven­
tional programming languages.
We have inevitably built up a considerable debt of gratitude to a num­
ber of colleagues, collaborators, post-doctoral researchers, and post-graduate
students during the course of conducting the research presented here. It is
therefore a pleasure to record our thanks for insights, comments, and valu­
able discussions to all of them. They include Sebastian Bader, Federico Banti,
Howard Blair, Eleanor Clifford, Artur S. d’Avila Garcez, Ben Goertzel, Bar­
bara Hammer, Roland Heinze, Steffen Hölldobler, Achim Jung, Matthias
Knorr, Ekaterina Komendantskaya, Vladimir Komendantsky, Ralph Kop­
perman, Markus Krötzsch, Kai-Uwe Kühnberger, Luis Lamb, Máire Lane,
Jens Lehmann, Tobias Matzner, Turlough Neary, John Power, Sibylla Prieß-
Crampe, Paulo Ribenboim, Bill Rounds, Sibylle Schwarz, Pawe�l Waszkiewicz,
Matthias Wendt, Andreas Witzel, Damien Woods, and Guo-Qiang Zhang. In
particular, we are grateful to Sebastian Bader for his contribution to Chap­
ter 7, and indeed this chapter was written jointly with him.
Our acknowledgments and thanks are also due to a number of institutions
and individuals for hosting us on a number of research projects and visits and
to various funding agencies for making the latter possible.
In particular, Pascal Hitzler acknowledges the support of Science Foun­
dation Ireland; the Boole Centre for Research at University College Cork;
University College Cork itself; the Deutscher Akademischer Austauschdienst
(DAAD); and Case Western Reserve University, Cleveland, Ohio. While con­
ducting the research which led to the contents of this book, P. Hitzler changed
affiliation several times, and he is grateful to University College Cork Ire­
land; the International Center on Computational Logic at Technical Univer­
sity Dresden; Case Western Reserve University, Cleveland, Ohio; the Institute
for Applied Informatics and Formal Description Methods (AIFB) at the Uni­
versity of Karlsruhe; and the Kno.e.sis Center at Wright State University,
Dayton, Ohio for providing excellent working environments.
 Preface xvii

It is a pleasure for Anthony Seda to thank Professor Steffen Hölldobler

for providing excellent research facilities and hospitality on two visits to the
Knowledge Representation and Reasoning Group at the Artificial Intelligence
Institute of the Department of Computer Science, Technical University Dres­
den, and to thank Deutscher Akademischer Austauschdienst (DAAD) for sup­
port on those occasions. He also acknowledges research support provided by
the Boole Centre for Research in Informatics, Science Foundation Ireland,
and University College Cork and is also considerably indebted to University
College Cork for the leave of absence needed to finish this work.
We also wish to thank several publishers for permission to reproduce parts
of certain of our papers, as follows.
• Elsevier Science Publishers Ltd. for extracts from [Hitzler and Seda, 1999a],
[Hitzler and Seda, 2002b], [Hitzler and Seda, 2003], [Hitzler et al., 2004], and
[Seda, 2006].
• International Information Institute for extracts from [Lane and Seda, 2006],
[Seda, 2002], and [Seda and Lane, 2005].
• IOS Press for permission to reprint the examples on Pages 111 and 112 from
[Seda, 1997].
• Oxford University Press for extracts from [Seda and Hitzler, 2010].
• The Mathematics Department of Auburn University, USA for extracts from
[Hitzler and Seda, 1999c] and [Seda and Hitzler, 1999b].
Finally, we are grateful to Kari A. Budyk, Iris Fahrer and especially to
Randi Cohen, all of Taylor & Francis Group, for their sustained help and
unfailing assistance at all stages in the preparation of the book for publication.
In addition, we would like to express our thanks to Shashi Kumar at the LATEX
Help Desk for solving a number of problems concerned with typesetting and to
the editorial support staff of Taylor & Francis Group for their careful reading
of the manuscript. We are highly indebted to the referees also for their many
helpful comments and suggestions, which led to a considerable broadening and
extension of the scope of the book, and also to the inclusion of a number of
important topics which otherwise would have been omitted. Last, but very far
from least, we extend our heartfelt thanks to our wives, Anne and Martine,
for their constant support and also for their endurance and forbearance in the
face of countless hours spent by us in the preparation of the manuscript.

Pascal Hitzler
Anthony Seda
Dayton and Cork
Introduction

Logic programming is programming with logic. In essence, the idea is to use

formal logic as a knowledge representation language with which to specify a
problem and to view computation as the (automated) deduction of new knowl­
edge from that given. The foundations of logic programming are usually based
upon the seminal paper of Robert Kowalski [Kowalski, 1974], which built on
John Alan Robinson’s well-known paper [Robinson, 1965] wherein foundations
were laid for the field of automated deduction using the resolution principle.
These ideas gave rise, more or less simultaneously, to the programming lan­
guage Prolog, first realized by Alain Colmerauer et al. in Marseilles in 1973,
see [Colmerauer and Roussel, 1993]. In this computing paradigm, a knowledge
base is given in the form of a logic program, which may be thought of as a
conjunctive normal form of a formula in the first-order language L underlying
the program as defined formally in Chapters 1 and 2. Then the program, or
system, can be queried with conjunctions Q of partially instantiated atomic
formulas, that is, with conjunctions of atomic formulas containing variables.
The resulting answers produced by the system are substitutions θ for these
variables by terms in L such that Qθ is a logical consequence of the knowledge
base. The automated deduction performed by the system is usually based on
a restricted form of resolution called SLD(NF)-resolution, see [Apt, 1997].
Since this early work, logic programming has become a major program­
ming paradigm and has developed in a considerable number of different and
diverse directions, including automated deduction (in the context, for exam­
ple, of model checking), natural language processing, databases, knowledge
representation and reasoning (including applications to the Semantic Web),
cognitive robotics, and machine learning, to mention a few. Furthermore, the
industrial applications using the underlying technologies, Prolog in the main,
but also an increasing number of related systems, are growing steadily more
numerous and more and more varied.1

1 For some examples, the proceedings of the annual International Conference on Logic

Programming (ICLP) provide a current view of the subject. The book [Bramer, 2010] con­
tains an introduction to Prolog programming. A standard reference for the theory under­
lying Prolog programming is [Apt, 1997]. The reference [Apt and Wallace, 2007] contains
much about constraint logic programming. See [De Raedt et al., 2008] for details of cur­
rent work in (probabilistic) inductive logic programming. For information about disjunctive
logic programming systems, see [Leone et al., 2006] and the website for the DLV project at
http://www.dbai.tuwien.ac.at/proj/dlv/, and for information concerning the related system
smodels, see [Simons et al., 2002] and the website http://www.tcs.hut.fi/Software/smodels/.

xix
xx Introduction

This book is concerned with the theory of logic programming languages

or, in other words, with their syntax and their semantics, especially the latter.
Very briefly, syntax in this context deals with formal grammar and automated
deduction, as discussed earlier; semantics, as usual, is occupied with meaning.
We will discuss semantics in more detail next. However, it should be observed
straightaway that the semantics of logic programming languages is compli­
cated in a way which is peculiar to them by the introduction of negation into
their syntax. The manner in which one handles negation is important, and
it is worth remarking that its development in logic programming has been
much influenced by the development of negation in non-monotonic reasoning,
a subject familiar in the field of artificial intelligence. Therefore, it will be
helpful to say a little about negation in these terms before describing in de­
tail the precise objectives of the book and its contents. This is because our
treatment of negation and semantics, see Chapter 2, is partly guided by these
considerations and also because negation and semantics are central themes of
the book.
Non-monotonic reasoning came into existence as a result of the desire to
capture certain aspects of human commonsense reasoning based on the obser­
vation that, in many situations occurring in everyday life, humans can reach
conclusions under incomplete or uncertain knowledge. More formally, it is typ­
ically the case that more facts can be derived from given facts or knowledge
when using commonsense reasoning than is the case when first-order logic is
employed. This has the consequence that some conclusions already made may
have to be withdrawn when more facts become known. By contrast, classical
logics such as propositional or predicate logic are monotonic in that whenever
a formula F is entailed by a theory or set of formulas Γ, then Γ ∪ {G} still
entails F , for any formula G.
The non-monotonic aspect of commonsense reasoning, however, has turned
out to be rather difficult to formalize in a satisfactory way. Early work in
this area was mainly based on three entirely different approaches2 : John
McCarthy’s circumscription, see [McCarthy, 1977, McCarthy, 1980]; Robert
Moore’s autoepistemic logic, see [Moore, 1984, Moore, 1985]; and Ray Re­
iter’s default logic, see [Reiter, 1980]. In fact, Prolog naturally includes some
features which can be viewed as being non-monotonic: if the system can prove
that a certain fact A does not follow from a given knowledge base, or program,
then A is considered to be false and hence ¬A is considered to be true. How­
ever, by adding the fact A to the program, we can now prove A, and thus we
have to retract the earlier conclusion ¬A. (Note that the negation occurring
in ¬A should not necessarily be taken here to be the negation encountered,
say, in first-order logic, but rather it symbolizes negation as (finite) failure to
prove A, as introduced in [Clark, 1978].)

2 See [Gabbay et al., 1994] for an excellent account of some of the main approaches to

non-monotonic reasoning including discussions of their advantages and drawbacks, and of

the validity of the intuitions underlying non-monotonic reasoning. Introductory textbooks
are [Antoniou, 1996, Berzati, 2007, Makinson, 2005].
 Introduction xxi

For reasons of this sort, research into non-monotonic reasoning has influ­
enced research into logic programming, and vice-versa, giving rise to impor­
tant and fruitful ideas and research directions in both areas. In particular,
such cross fertilization has led to the realization that logic programs, possi­
bly augmented with some additional syntactic features, provide an excellent
language for knowledge representation in the presence of non-monotonicity.
In addition, such research has led to a number of implementations of non­
monotonic-reasoning-based logic programming systems commonly known as
answer set programming systems.3
Thus, the interaction between logic programming and non-monotonic rea­
soning is important. It is not, however, the main focus of our work. On the
contrary, our main focal points are, in a nutshell, first, the detailed develop­
ment of the mathematical tools and methods required to study the semantics
of logic programs, and second, in order to illustrate these methods, the detailed
development of the main semantics of logic programs per se. In addition, we
give an application of the methods we present to study semantics in the con­
text of neural-symbolic integration, as described in more detail shortly. Thus,
we do not treat procedural matters and matters concerned with implementa­
tion in any depth, and indeed these issues are only touched on incidentally. We
also do not discuss matters primarily concerned with non-monotonic reason­
ing other than in the context of their role in guiding our thinking in relation
to negation in logic programs, as already noted. It will therefore be of value
to say a little more about our precise objectives, and we do this next.
In common with most programming languages, the syntax of logic pro­
gramming is comparatively easy to specify formally, whereas the semantics is
much harder to deal with. Again, in common with other programming lan­
guages, there are several ways of giving logic programs a formal semantics.
First, logic programs have, of course, a procedural or operational semantics,
which describes and is described by their behaviour when executed on some
(abstract) machine. Second, unlike imperative or functional programs, logic
programs have a natural semantics, called their declarative semantics, which
arises simply because a logic program is a consistent set of well-formed formu­
lae and can be viewed as a theory. This semantics is usually captured by means
of models, in the sense of mathematical logic, and will play a dominant role
in our development. Indeed, a central problem in the theory is the question of
selecting the “right” model for a program, namely, a model which reflects the
intended meaning of the programmer and relates it to what the program can
compute. It is here that ideas from non-monotonic reasoning play a funda­
mental role in determining the right models, including well-known ones such
as the supported, stable, and well-founded models. Third, a standard and very
important way of selecting the appropriate models for a logic program is to as­

3 For a discussion of these matters, see [Lifschitz, 1999, Marek and Truszczyński, 1999,

Baral, 2003]. For current developments in non-monotonic reasoning (versus logic program­
ming), one may consult the proceedings series of the International Conferences on Logic
Programming and Non-Monotonic Reasoning (LPNMR), for example.
xxii Introduction

sociate with the program one or more of a number of operators called semantic
operators 4 defined on spaces of interpretations (or valuations) determined by
the program. One then studies the fixed points of these operators, leading to
the fixed-point semantics of the program in question. This latter semantics
can roughly be equated with the denotational semantics of imperative and
functional programs associated with the names of Dana Scott and Christo­
pher Strachey because some, but not all, of the important semantic operators
which have been introduced are Scott continuous in the sense of domain the­
ory, or at least are monotonic. Moreover, fixed points play a fundamental role
also in denotational semantics. Finally, there is a general requirement that
all the semantics described previously should coincide or at least be closely
related in some sense.5
Taking the observations just made a little further forward, we note that
there are several interconnected strands to the programme of analyzing the
fixed points of semantic operators, but three of the main ones are as follows.
First, we consider a number of operators already well-known in the theory, in
addition to introducing several more. In this step, we focus on ensuring that
the operators we study, and their fixed points, correctly reflect the meaning
of programs and their properties. Second, we investigate the properties of the
operators themselves, especially in relation to whether or not they are Scott
continuous and, if not, what properties they do possess. Scott continuity is a
desirable feature for a semantic operator to have because it implies that the
operator has a least fixed point. Furthermore, this least fixed point is often
taken to be the fixed-point semantics of the program in question, and indeed,
operators which are not Scott continuous may in general fail to have any fixed
points at all. Third, we study the fixed-point theory of semantic operators in
considerable generality. In fact, the failure of certain apparently reasonable
semantic operators (already known to capture declarative semantics) to be
Scott continuous often results from the introduction of negation, because the
introduction of negation may render the operators in question to be non-
monotonic and hence to fail to be Scott continuous, as we will see in Chapter 2.
The point just made is important because it is one of the reasons for
introducing alternatives to order theory in studying fixed-point theory in re­
lation to semantics and in establishing fixed-point theorems applicable to non-
monotonic operators, see Chapter 4. Therefore, it will help to give some in­
sight next into the non-traditional methods we introduce and develop, how
they work in the context of negation, and especially how they work in find­
ing models for logic programs with negation. Our point of view is to regard
programs, and logic programs in particular, as (abstract) dynamical systems
whose states change under program execution and whose state changes can be
modelled by an operator T . Starting with some initial state, s0 , say, it is inter­

4 This is a generic term which we use to cover all of a number of specific operators we

will study, such as the TP -operator, see Definition 2.2.1.

5 See Theorem 2.2.3, for example, and [Lloyd, 1987] for details of how procedural seman­

tics relates to declarative semantics.

 Introduction xxiii

esting to observe the behaviour of the sequence of iterates s0 , T (s0 ), T 2 (s0 ),

T 3 (s0 ), . . . of T on the state s0 . Suppose, for example, that s0 is the nowhere
defined partial function on the natural numbers, and T is the operator on
the partial functions determined in the usual way by some well-defined recur­
sive definition on the natural numbers, see [Stoltenberg-Hansen et al., 1994],
for example. Then, typically, the sequence of iterates will form an ω-chain
as defined in Chapter 1 and will converge in the Scott topology (defined in
Chapter 3) to the supremum s of the chain; thus, we have s = lim T n (s0 ) in
the Scott topology on the partial functions. Furthermore, T will typically be
Scott continuous (see Chapter 3 again for the definition of Scott continuity) in
the sense that T (s) = T (lim sn ) = lim T (sn ) whenever sn is a sequence con­
verging to s in the Scott topology, that is, a sequence satisfying lim sn = s. If
T is indeed Scott continuous, then it is now easy to deduce that T (s) = s so
that s is a fixed point, in fact, the least fixed point, of T . (These observations
are the heart of the proof of Kleene’s theorem, Theorem 1.1.9, which is some­
times viewed as the fundamental theorem in semantics. They are also quite
close in form to the proof of the Banach contraction mapping theorem, The­
orem 4.2.3, except that it is order rather than a contraction property which
determines the convergence.) In such a situation, s is usually taken to be the
meaning or semantics of the original recursive definition. Precisely the same
sort of thing happens in relation to logic programming semantics in the case of
logic programs P which do not contain negation or in other words are definite
programs. Specifically, the iterates of the single-step operator (or immediate
consequence operator)6 T applied to the empty interpretation converge in the
Scott topology to an interpretation M . This interpretation M is the (least)
fixed point of T , captures well the declarative semantics for P , and relates well
to the procedural semantics for P under SLD-resolution, see Theorem 2.2.3
and the discussion following it.
Following on from the comments just made in the previous paragraph is
the interesting observation from our point of view, or the mathematical point
of view, that the discussion just presented can quite easily be generalized: all
that one needs is an abstract notion of convergence and an abstract notion
of continuity. Such a setting is provided by the notion of convergence space,
and in particular by convergence classes or equivalently by topological spaces,
as defined in Chapter 3. These notions provide a general setting in which one
can study semantics and in particular logic programming semantics for logic
programs P which may or may not contain negation. The classical case of
definite programs corresponds to taking the Scott topology, but we consider
quite extensively another topology, called the Cantor topology by us, defined
in Chapter 3, which is closely connected to negation, has connections with the
Scott topology, and underlies important classes of programs which do involve
negation such as acceptable programs and their generalizations, see Chap­
ter 5. Indeed, a quite elementary property we use is given in Proposition 3.3.2

6 See Chapter 2 for the definitions of these terms.

xxiv Introduction

and in a rather general form by Theorem 5.4.2 and simply states that if P
is any logic program, and I is an interpretation such that T n (I) converges
in the Cantor topology to an interpretation M , then M is a model for P ; if,
further, T is continuous in the Cantor topology, then M is a fixed point of T
(here, again, T denotes the single-step operator associated with P ). Further
comments on this result are to be found in Remark 3.3.3 and in the comments
immediately following Remark 3.3.3. In particular, this fact is exploited on a
number of occasions to find models in the presence of negation and in partic­
ular in studying acceptable programs, as just mentioned, and also in studying
the perfect model for locally stratified programs in Chapter 6. Indeed, the
working out of this observation together with some of its implications occu­
pies a significant proportion of our time. In addition, because convergence is a
key notion, our development of topology in Chapter 3 is based on it, although
the main conclusions presented there are also given in other equivalent and
familiar forms.
In practice, detecting whether sequences converge or whether operators
have fixed points is most easily done by means of metrics and more general
distance functions (generalized metrics) together with their associated fixed-
point theorems, the latter perhaps being reminiscent of the Banach contrac­
tion mapping theorem, see Theorem 4.2.3. Furthermore, underlying the use
of generalized metrics are topologies defined on spaces of interpretations, and
we study these in Chapter 3 with a view to developing, in conjunction with
Chapter 4, the mathematical analysis we apply later in Chapters 5 and 6 in
studying acceptable programs and related semantics, as already mentioned,
and in Chapter 7 in the context of artificial neural networks in relation to logic
programming. This latter work concerns the problem of integrating different
models of computation in an attempt to combine the best of each in a single
system and understanding the semantics of the combined system. In our case,
we consider the integration of logic programming, perhaps taken as repre­
sentative of discrete systems, with connectionist systems, or neural networks,
considered as continuous systems inspired by biological models of computa­
tion. A means of doing this is to compute semantic operators by means of
neural networks. However, in the case of first-order (non-propositional) pro­
grams, it is necessary to employ approximation techniques (rather than exact
computation) which depend on viewing spaces of interpretations as compact
Hausdorff spaces, that is, to employ yet again methods from mathematical
analysis. Such applications as these are another important reason for devel­
oping a quite extensive body of mathematics which provides alternative tools
to those based on order theory in studying semantics. In fact, one of the main
highlights, themes and motivating features of this book is the analysis we carry
out of foundational structures of various sorts, with an eye to potential ap­
plications in the field of computational logic in general, as exemplified by our
results in, for example, Chapter 7. Indeed, it seems probable that such meth­
ods and tools will prove useful in developing foundations in other areas where
discrete and continuous models of computation are combined, quite apart from
 Introduction xxv

neural-symbolic integration. Such non-classical models of computation are of

great interest generally in present times and may contain both continuous
and discrete components, especially those inspired by physical phenomena.
As such, their study will almost certainly require techniques appropriate to
both their continuous elements and to their discrete elements and may well
be of the sort developed here.
It should be noted that other authors have, to a greater or lesser extent,
employed mathematical analysis in the context of logic programming seman­
tics. Their work is complementary to what we present here, and we briefly
discuss some of it and its relationship with ours next and in more detail in the
body of the text. For example, some of the recent work of Howard Blair and
several of his colleagues on logic programming semantics is much concerned
with the interaction between the continuous and the discrete, and it makes
use of ideas from dynamical systems, convergence spaces, and automata the­
ory to model hybrid systems. We consider this work further in Chapter 3.
We mention also the work of Sibylla Prieß-Crampe and Paulo Ribenboim on
the role of generalized ultrametrics in fixed-point theory in the context of
logic programmimg semantics. They discuss both single-valued and multival­
ued mappings in this context, and we consider their results in considerable
detail in Chapter 4 and some of their applications in Chapter 5. In addition,
we also include in Chapter 4 a discussion of recent work of Umberto Straccia,
Manuel Ojeda-Aciego, and Carlos Damásio on multivalued mappings in the
context of semantics and the relationship between their work and ours. Fi­
nally, we discuss in Chapter 4 also the extensive work of William Rounds and
Guo-Qiang Zhang on the use of domain theory as a theoretical foundation
for logic programming, both from the point of view of procedural aspects and
from the point of view of semantics.
Summarizing the chapters, Chapter 1 contains, in fairly condensed form,
the preliminaries from order theory, domain theory, and logic which we will
employ throughout this book. In addition, we present two well-known fixed-
point theorems, based on order, which are fundamental in applications to
semantics. The next chapter, Chapter 2, introduces logic programs and the
most important ways of assigning semantic operators and declarative seman­
tics to them. The manner in which the material is presented is rather novel and
employs the syntactic notion of level mapping, defined in Chapter 2. Indeed,
we make several different applications of level mappings in our discussions,
and they play a unifying role in several places in the course of developing our
main themes. For example, their use in Chapter 2 provides a uniform and
comprehensive treatment of all of the important different semantics known in
the subject, including those associated with the supported, stable, and well-
founded models mentioned earlier.7 Sets of interpretations are important in
that they are, among other things, the carrier sets for the various semantic

7 The uniform characterizations by means of level mappings which will be given in Chap­

ter 2 are due mainly to [Hitzler and Wendt, 2002].

xxvi Introduction

operators we discuss. Such sets themselves may be endowed with various, use­
ful structures. In Chapter 3, we illustrate the point just made by studying
various topologies on spaces of interpretations, including the Scott topology
and a topology called the Cantor topology, as already mentioned. The conti­
nuity of semantic operators in the Scott topology is examined in Chapter 3,
but the treatment of their continuity in the Cantor topology is deferred until
we reach Chapter 5, where the results are needed. In fact, as noted earlier, it
is convergence in these topologies which is of main interest because it can be
used to find models for logic programs as we show in Chapters 5 and 6, and
thus, convergence is the dominant theme in our development of topology.
We take the theme of structures defined on spaces of interpretations yet
further in Chapter 4 in presenting a detailed account of both various gen­
eralized distance functions defined on spaces of interpretations and their as­
sociated fixed-point theorems. These tools, some of which depend on level
mappings again, are developed specifically for investigating semantic oper­
ators of logic programs with negation, but we believe that Chapter 4 is a
self-contained account of results which are likely to have applications within
computer science outside those areas considered here. In Chapter 5, we com­
bine the developments of Chapters 2, 3, and 4 by applying the fixed-point
theorems of Chapter 4 to the more important semantic operators introduced
in Chapter 2. More specifically, we focus on classes of programs, which we
call unique supported model classes, each of which has the property that all
programs in that class have a unique supported model. An example of such
a class is the class of acceptable programs well-known in termination analy­
sis, but we examine other important unique supported model classes as well.
These classes are interesting because it turns out that for each of the programs
they contain, many of the main semantics studied in the earlier chapters co­
incide, and hence the meaning of each program in a unique supported model
class is unambiguous relative to the most important semantics. In essence,
we obtain these classes by applying to various semantic operators those fixed-
point theorems of Chapter 4 which guarantee a unique fixed point, if there
is a fixed point at all. The process involves working with successively more
general semantic operators, especially Fitting-style operators, and examining
their properties in relation to single-step operators and convergence of their
iterates in the Cantor topology studied in Chapter 3. Indeed, the process cul­
minates in a very general semantic operator T which subsumes many of those
studied in the earlier chapters, and we estabish many of its important proper­
ties in Chapter 5. In particular, we examine in depth the continuity of T in the
Cantor topology, thereby obtaining the corresponding results for single-step
operators and Fitting-style operators. Finally, we note that the work we do
in this chapter consolidates the uniform approach provided in Chapter 2, em­
ploying level mappings, to encompass the additional semantics we introduce
in Chapter 5.
Turning now to Chapter 6, our objectives here are twofold. First, we revisit
the stable model semantics and establish a close connection between the well­
 Introduction xxvii

known Gelfond–Lifschitz operator GLP and the fixpoint completion fix(P ) for
any normal logic program P by deriving the identity GLP (I) = Tfix(P ) (I), for
any two-valued interpretation I, see Theorem 6.1.4. This will make it a simple
and routine matter to prove many facts about GLP , and hence about the sta­
ble model, from properties of the single-step operator, including the derivation
of continuity properties of GLP . Our second objective in Chapter 6 is to re­
visit stratification and the perfect model and to present an iterative process for
obtaining the perfect model for locally stratified normal logic programs. This
approach involves careful control of negation in order to produce monotonic
increasing sequences by means of non-monotonic operators and is interesting
for the insight it gives into the structure of the perfect model. In Chapter 7,
we apply the topological and analytical tools developed earlier in order to
discuss logic programming in the context of dynamical systems and artificial
neural networks with a view, in particular, to presenting a detailed account of
these methods in the foundations of neural-symbolic integration. Specifically,
in Chapter 7, we consider the computation by artificial neural networks of
various semantic operators associated with normal logic programs. We view
this as a means of integrating these two computing paradigms because both
can be represented by functions: the semantic operator on the one hand and
the I/O function of the neural network on the other. In fact, exact computa­
tion of semantic operators is only possible in the case of propositional normal
logic programs. In the case of first-order programs, approximation methods
are required, and this is where analytical and topological methods make their
entrance. Indeed, it turns out that continuity of a semantic operator in the
Cantor topology is a necessary and sufficient condition for this approximation
process to work, see Theorem 7.5.3. This observation is yet further motiva­
tion for studying the Cantor topology, and hence Chapter 7 represents an
important application of analytical ideas in logic programming semantics. In
Chapter 8, we give a brief discussion of further possible applications of our
results and future directions for research involving the methods and results
of this book. In particular, we discuss possible future work in the context
of the foundations of program semantics, quantitative domain theory, fixed-
point theory, the Semantic Web, and neural-symbolic integration, among other
things. In the Appendix, we bring together a summary of those facts from the
theory of ordinals and general topology which will be needed at various points
in our investigations, but are not developed in the main body of the text; its
inclusion makes our treatment essentially self-contained. In particular, the re­
sults of Chapter 3 together with those of the Appendix give a treatment of
the Scott topology in terms of convergence.
Finally, on a point of convention, we note that the symbol • will be em­
ployed as an end marker in two ways in the body of the text. First, it will
be used to indicate the end of every proof. Second, it will be used on a few
occasions to mark clearly the end of any statement (theorem, proposition,
definition, remark, example, program, etc.), where the end of that statement
might otherwise be unclear.
About the Authors

Pascal Hitzler is an assistant professor at the Kno.e.sis Center for

Knowledge-Enabled Computing, which is an Ohio Center of Excellence at
Wright State University in Dayton, Ohio, U.S.A. From 2004 to 2009, he was
an assistant professor at the Institute for Applied Informatics and Formal De­
scription Methods (AIFB) at the University of Karlsruhe in Germany, and
from 2001 to 2004 he was a post-doctoral researcher at the Artificial Intel­
ligence Institute at TU Dresden in Germany. In 2001 he obtained a PhD
in mathematics from the National University of Ireland, University College
Cork, Cork, Ireland, under the supervision of Anthony Seda, and in 1998 he
obtained a Diplom (Masters degree equivalent) in mathematics from the Uni­
versity of Tübingen in Germany. His research record lists over 150 publications
in such diverse areas as the Semantic Web, neural-symbolic integration, knowl­
edge representation and reasoning, denotational semantics, and set-theoretic
topology. He is editor-in-chief of the IOS Press journal Semantic Web – In­
teroperability, Usability, Applicability and the IOS Press book series Studies
on the Semantic Web. He is vice-chair of the steering committee of the con­
ference series on Web Reasoning and Rule Systems (RR) and the RR2010 PC
co-chair. He is co-chair of the 2010 International Semantic Web Conference
(ISWC) “Semantic Web In Use” and “Industry” tracks and co-chair of the
2011 Extended Semantic Web Conference (ESWC) “Reasoning” track. He is
co-author of the W3C Recommendation OWL 2 Primer, of the first German
introductory textbook to the Semantic Web, published by Springer-Verlag,
and of the book Foundations of Semantic Web Technologies published by
CRC Press, 2009. For more information, see http://www.pascal-hitzler.de.

Anthony Karel Seda is a senior lecturer in the Department of Mathematics,

University College Cork, Cork, Ireland. He holds a BSc and a PhD in pure
mathematics from the University of Wales, UK, and an MSc from the Uni­
versity of Warwick, UK, also in pure mathematics. His research record lists
nearly 100 publications in areas ranging from measure theory and functional
analysis to topology and fixed-point theory in computer science through to
denotational semantics and the semantics of logic programs, including two
books of which he is a co-editor. In addition, he has a wide range of teaching
interests, including the research areas just listed. He is a member of several
editorial boards of journals, including the editorial board, and associate ed­
itor for Europe, of the international journal Information; editorial board of
the International Journal of Advanced Intelligence; and he was for ten years a

xxix
xxx About the Authors

member of the board of editors of the Bulletin of the European Association for
Theoretical Computer Science (EATCS), and also, for a number of years, was
on the editorial board of Asian Information-Science-Life: An International
Journal. He is a co-founder of the “Irish Conference on the Mathematical
Foundations of Computer Science and Information Technology (MFCSIT),”
a conference series the proceedings of which are published by Elsevier Sci­
ence Publishers, and has strong connections with a considerable number of
other conferences. These include the conference series “International Informa­
tion Conference,” usually held in Asia, and the “International Workshop on
Formal Methods,” usually held in Ireland. He is a co-founder of the Boole
Centre for Research in Informatics (BCRI) at University College Cork, is a
member of its executive committee, and leader of the “Theory of Computa­
tion” theme within BCRI. He has led various groups of researchers in the
theory of computation and has supervised many Masters degree and PhD de­
gree students to completion. More information can be found at the Web site
http://euclid.ucc.ie/pages/staff/seda/tseda.htm
Chapter 1

Order and Logic

The study of the semantics of logic programs rests on a certain amount of

order theory and logic, and it will be convenient to collect together here in
this first chapter those basic facts we need throughout the book to accomplish
this study.1 At the same time, we establish some notation and terminology
which is common to all the chapters.

1.1 Ordered Sets and Fixed-Point Theorems

We start by presenting the minimum amount that we need of the theory
of ordered sets. In addition, we discuss certain important and well-known
fixed-point theorems applying to functions defined on ordered sets. In fact,
the first of these theorems has fundamental applications in the semantics of
computation in general, as well as in logic programming semantics.
Let D be a set. Recall that a binary relation [ on D is simply a subset [
of D × D. As usual, the symbol [ will be written infix, and hence we write
x [ y rather than (x, y) ∈ [, where x, y ∈ D. Furthermore, we write x c y
if x [ y and x = y. The relation [ on D is called reflexive if, for all x ∈ D,
we have x [ x; it is called antisymmetric if, for all x, y ∈ D, x [ y and y [ x
imply x = y; and it is called transitive if, for all x, y, z ∈ D, x [ y and y [ z
imply x [ z. We call [ a partial order if [ is reflexive, antisymmetric, and
transitive, and in that case we call the pair (D, [), or simply D when [ is
understood, a partially ordered set, a poset, or sometimes a partial order by
abuse of terminology. We may sometimes simply refer to a partially ordered
set (D, [) as an ordered set and to the relation [ as an ordering (on D).
Two elements x and y of a partially ordered set D are said to be comparable
if either x [ y or y [ x holds; otherwise, x and y are called incomparable.
A non-empty subset A ⊆ D is said to be totally ordered by [ or is called
a chain if any two elements of A are comparable with respect to [, that is,
given a, b ∈ A, we have a [ b or b [ a. A partial order [ on D is called a
total order if D itself is totally ordered by [. We call A an ω-chain if A is an
increasing sequence a0 [ a1 [ a2 . . ., where ω denotes the first limit ordinal.
1 The text [Davey and Priestley, 2002] is a useful reference for the subject of ordered sets.

1
2 Mathematical Aspects of Logic Programming Semantics

(We refer the reader to the Appendix for a brief discussion of the theory of
ordinals.) We note that any ω-chain is, of course, a chain.
A non-empty subset A of a partially ordered set (D, [) is called directed if,
for all a, b ∈ A, there is c ∈ A with a [ c and b [ c. An element b in an ordered
set D is called an upper bound of a subset A of D if we have a [ b for all a ∈ A
and is called a least upper bound or supremum of A if b is an upper bound of A
[ b' for all upper bounds b' of A. Of course, by antisymmetry, the
satisfying b 
supremum, A or sup A, of A is unique if it exists.nSimilarly, one defines lower
bound and the greatest lower bound or infimum, A or inf A, of a subset A
of D. An element x of D is called maximal (minimal ) if we do not have x c y
(y c x) for any element y of D. Given an ordering [ on a set D, we define the
dual ordering [d on D by x [d y if and only if y [ x. Lower bounds, greatest
lower bounds, etc. in [ correspond to upper bounds, least upper bounds, etc.
in [d .

1.1.1 Definition Let (D, [) be a partially ordered set.


(1) We call (D, [) an ω-complete partial order or an ω-cpo if A exists in D
for each ω-chain A in D, and D has an element ⊥, called the least element
or bottom element, satisfying ⊥ [ x for all x ∈ D.

(2) We call (D, [) chain complete if every chain in D has a supremum.


(3) We call (D, [) a complete partial order or a cpo if A exists in D for
each directed subset A of D, and D has a bottom element.

(4) We call (D, [) a complete upp
ner semi-lattice if A exists in D for each
directed subset A of D, and A exists for each subset A of D.
 n
(5) We call (D, [) a complete lattice if A and A exist in D for every
subset A of D.

Later on, we will encounter examples of each of these notions in the context
of spaces of valuations. Notice that on taking A = D in the previous definition,
we see that a complete upper semi-lattice or a complete lattice always has
a bottom element and that a complete lattice always has a top element or
greatest element, that is, an element T satisfying a [ T for all a ∈ D.
There are various implications between the notions formulated in Defini­
tion 1.1.1, some of which are obvious. Indeed, as far as the various notions of
completeness are concerned, each defined concept is apparently less general
than its predecessor. For example, since any chain is a directed set, we see that
any complete partial order is chain complete, and any chain-complete poset
with a bottom element is an ω-complete partial order. However, the following
fact, which we simply state, is less trivial.2
2 For a discussion of chain completeness versus completeness (for directed sets), we refer

the reader to [Markowsky, 1976]; see also [Abramsky and Jung, 1994, Proposition 2.1.15].
 Order and Logic 3

1.1.2 Proposition A partially ordered set (D, [) is a complete partial order

if and only if it has a bottom element and is chain complete.

Many aspects of theoretical computer science depend on the notion of

a partially ordered set. More structure is often required, however, than is
provided simply by a partial order or even by a complete partial order or
complete lattice. For example, one needs extra structure in order to model
standard programming language constructs or to provide an abstract theory of
computability, as well as having a satisfactory fixed-point theorem available. It
is now widely recognized that Scott’s theory of domains provides a satisfactory
setting in which to attain all these objectives, and we will find it useful later on
to view spaces of valuations as Scott domains. It will therefore be convenient
to give next the definition of the term “(Scott) domain” in the form in which
we will always use it. First, however, we need to define the notion of compact
element.

1.1.3 Definition Let (D [) be a partially ordered set. We call an element

a ∈ D compact
 or finite if it satisfies the property that whenever A is directed
and a [ A, we have a [ x for some x ∈ A. We denote the set of compact
elements in D by Dc .

Notice that the bottom element in a complete partial order is always a

compact element, and hence the set Dc is always non-empty in this case. The
compact elements are fundamental in domain theory.

1.1.4 Definition A Scott-Ershov domain, Scott domain, or just domain

(D, [) is a consistently complete algebraic complete partial order. Thus, the
following statements hold.

(1) (D, [) is a complete partial order.

(2) For each x∈ D, the set approx(x) = {a ∈ Dc | a [ x} is directed, and we

have x = approx(x) (the algebraicity of D).

consistent (that is, there exists x ∈ D such that

(3) If the set {a, b} ⊆ Dc is 
a [ x and b [ x), then {a, b} exists in D (the consistent completeness
of D).

We next give some simple examples of the concepts defined above; note
that (1) and (2) are special cases of Theorem 1.3.2.

1.1.5 Example (1) The power set D = P(N) of the set N of natural numbers
is a complete lattice when ordered by set inclusion. In this ordering, D is
also a domain in which the compact elements are the finite subsets of N.
Furthermore, the bottom element of D is the empty set ∅ and ∅ is also
the only minimal element of D; the top element of D is N and N is the
only maximal element of D.
4 Mathematical Aspects of Logic Programming Semantics

(2) Let X be a non-empty set, and let D denote the set of all pairs (I + , I − ),
where I + and I − are disjoint subsets of X. We define an ordering on D
by (I + , I − ) [ (J + , J − ) if and only if I + ⊆ J + and I − ⊆ J − . Then D
is a domain in which the bottom element is the pair (∅, ∅), the compact
elements of D are the pairs (I + , I − ) in D in which I + and I − are finite
sets, and the maximal elements are the pairs (I + , I − ) which satisfy I + ∪
I − = X. Note that D is not a complete lattice.

(3) Let D denote the set of all partial functions f : Nn → N ordered by graph
inclusion, that is, f [ g if and only if graph(f ) ⊆ graph(g), where f
and g are partial functions. (Thus, f [ g if and only if whenever f (x)
is defined, so is g(x) and f (x) = g(x).) Then D is a domain in which a
partial function f is a compact element if and only if graph(f ) is a finite
set and the bottom element is the empty function. Here, the maximal
elements of D are the total functions. Again, D is not a complete lattice.

1.1.6 Remark Mathematically speaking, the denotational semantics, or

mathematical semantics, approach to the theory of procedural and functional
programming languages is highly involved with providing a satisfactory frame­
work within which to model constructs made in conventional programming
languages. Such frameworks must be closed under the formation of products,
sums, and function spaces and therefore are, simply, Cartesian closed cate­
gories. One of the most successful Cartesian closed categories to have arisen
out of these considerations is that of Scott domains,3 as formulated in Defini­
tion 1.1.4. Moreover, most functions and operators encountered within domain
theory are order continuous, see Definition 1.1.7, and therefore the most useful
fixed-point theorem in domain theory is Theorem 1.1.9. On the other hand, as
we shall see in the next chapter and subsequent chapters, a logic program has
a well-defined and mathematically precise meaning inherent in its very nature,
namely, its semantics as a first-order logical theory. In addition, certain impor­
tant operators arising in logic programming are not monotonic in general due
to the presence of negation, resulting in Theorems 1.1.9 and 1.1.10 often be­
ing inapplicable, and this has no direct parallel in conventional programming
language semantics. For these reasons, the semantics of logic programming
languages has developed rather differently from that of procedural program­
ming languages. Nevertheless, we shall study domains in Chapter 4, in the
context of fixed-point theory.4

If D is a set, A is a subset of D, and f : D → D is a function, then we

denote the image set {f (a) | a ∈ A} of A under f by f (A). We also define

3 See
[Scott, 1982b].
4 Our
basic references to domain theory are the book [Stoltenberg-Hansen et al., 1994]
and the book chapter [Abramsky and Jung, 1994], but the reader interested in domain
theory may also care to consult the notes of G.D. Plotkin [Plotkin, 1983] and also the
comprehensive treatment to be found in [Gierz et al., 2003].
 Order and Logic 5

iterates of a function f : D → D inductively as follows: f 0 (x) = x, and

f n+1 (x) = f (f n (x)) for all n ∈ N and x ∈ D.

1.1.7 Definition A function f : D → E between posets D and E is called

monotonic if, for all a, b ∈ D with a [ b, we have f (a) [ f (b). Furthermore,
f is called antitonic if, for all a, b ∈ D with a [ b, we have f (b) [ f (a). If
D and E are ω-complete partial orders,  then a function
 f : D → E is called
ω-continuous if it is monotonic and f (A) = f ( A) for each ω-chain A in
D. Finally, if D and E are complete partial orders, then f is called (order )
continuous
 if, again,
 it is monotonic and, for every directed subset A of D, we
have f (A) = f ( A).
We note that if f is monotonic, then the image of any ω-chain under f
is an ω-chain, and similarly the image of any directed set under f is itself
a directed set. Therefore, the two suprema required in making the previous
definition always exist. Indeed, it is easy to see that, equivalently,5 one may
define f to be continuous  by requiring,for each directed set A, that f (A)
is a directed set and that f (A) = f ( A). In fact, if f  is monotonicand
A is directed, then it is easily checked that the inequality f (A) [ f ( A)
always holds. Therefore,
 itfollows that f is continuous if and only if it is
monotonic and f ( A) [ f (A) whenever A ⊆ D is directed. As a matter
of fact, preservation of suprema of chains is enough in defining continuity as
shown by the next result, which again we simply state.6 We note finally that
if a function f between complete partial orders is continuous, then it is clear
that it is ω-continuous as a function between ω-complete partial orders.

1.1.8 Proposition A function f : D → E between

 complete
 partial orders is
continuous if and only if it is monotonic and f (A) = f ( A) for each chain
A in D.
We define ordinal powers of a monotonic function f on a complete partial
order (D, [) inductivelyas follows: f ↑ 0 = ⊥, f ↑ (α + 1) = f (f ↑ α) for any
ordinal α, and f ↑ α = {f ↑ β | β < α} if α is a limit ordinal. Noting that
(D, [) is chain complete, being a complete partial order, it is straightforward
using transfinite induction to see that f ↑ β [ f ↑ α whenever β ≤ α, and
hence that ordinal powers of f are well-defined. More generally, the same
comments apply to the ordinal powers f α (x) for any x ∈ D which satisfies
x [ f (x): we define f 0 (x) = x, f α+1 (x) = f (f α (x)) for any ordinal α, and
f α (x) = {f β (x) | β < α} if α is a limit ordinal.
A fixed point of a function f : D → D is an element x ∈ D satisfying
f (x) = x. A pre-fixed point of a function f on a poset (D, [) is an element
y ∈ D satisfying f (y) [ y. Finally, a post-fixed point of f is an element y ∈ D
satisfying y [ f (y). The least fixed point, lfp(f ), of f is a fixed point x of f
5 Thisis the definition adopted in [Stoltenberg-Hansen et al., 1994].
6A discussion of the various ways of formulating the notion of continuity is to be found
in [Markowsky, 1976].
6 Mathematical Aspects of Logic Programming Semantics

satisfying the property: if y is a fixed point of f , then x [ y. Least pre-fixed

points and least post-fixed points are defined similarly.
The following two theorems are fundamental in handling the semantics of
logic programs.7 Indeed, the first of them, which is frequently referred to as the
fixed-point theorem, is fundamental in procedural and functional programming
as well.8

1.1.9 Theorem (Kleene) Let (D, [) denote an ω-complete partial order

and let f : D → D be ω-continuous. Then f has a least fixed point x = f ↑ ω
which is also its least pre-fixed point.

Proof: We sketch the proof of this well-known result.

The sequence (f ↑ n)n∈N is an ω-chain. It therefore
 has a supremum f↑ω =
x, say. By ω-continuity, we have x = f ↑ ω = {f ↑ (n + 1) | n ∈ N} = f ( {f ↑
n | n ∈ N}) = f (x), and so x is a fixed point of f . If y is a pre-fixed point of
f , then ⊥ [ y, and, by monotonicity of f , we obtain f ↑ 1 = f (⊥) [ f (y) [ y.
Inductively, it follows that f ↑ n [ y for all n ∈ N, and hence x = f ↑ ω [ y.
So x is the least pre-fixed point of f and hence also its least fixed point. •

By our earlier observation that a continuous function is ω-continuous, this

theorem applies, of course, to continuous functions on complete partial orders.
Moreover, if the function is not ω-continuous, but is monotonic, the existence
of a least fixed point can still be guaranteed, as we see next.9

1.1.10 Theorem (Knaster-Tarski) Let (D, [) denote a complete partial

order, let f : D → D be monotonic, and let x ∈ D be such that x [ f (x).
Then f has a least fixed point a above x, meaning x [ a, which is also the
least pre-fixed point of f above x, and there exists a least ordinal α such that
a = f α (x). In particular, f has a least fixed point a which is also its least
pre-fixed point.

Proof: Again, this theorem is well-known, and we just sketch its proof.
Let γ be an ordinal whose cardinality exceeds that of D, and form the
set {f β (x) | β ≤ γ}. By cardinality considerations, there must be ordinals
α < β ≤ γ with f α (x) = f β (x), and we can assume without loss of generality
that α is least with this property. Since f α (x) [ f (f α (x)) [ f β (x) = f α (x),
7 Fixed points of certain operators associated with logic programs are of extreme impor­

tance in the semantics of logic programs, as we shall see in later chapters.

8 A result similar to Kleene’s theorem, in fact, equivalent to it, is the well-known theorem

due to Tarski and Kantorovitch in which ω-chains are replaced by countable chains, see
[Jachymski, 2001]. Indeed, the collection containing [Jachymski, 2001] is an excellent general
reference to fixed-point theory. As noted in [Lloyd, 1987], the reference [Lassez et al., 1982]
contains an interesting discussion of the history of fixed-point theorems on ordered sets.
9 In attributing Theorem 1.1.10 to Knaster and Tarski, we are noting Proposition 1.1.2

and then following Jachymski in [Jachymski, 2001]. Theorem 1.1.9 is usually attributed to
Kleene, since this theorem is an abstract formulation of the first recursion theorem, and we
are consistent with [Jachymski, 2001] in this respect.
 Order and Logic 7

we obtain that f α (x) = f (f α (x)), and so a = f α (x) is a fixed point of f .

Clearly, we have x [ a. Furthermore, if b is any pre-fixed point of f with
x [ b, then by monotonicity of f and the fact that f (b) [ b we obtain
f β (x) [ b for all ordinals β. Hence, a [ b, and so a is both the least pre-fixed
point and the least fixed point of f above x.
To obtain the final conclusion, we simply set x = ⊥ and note then that
x [ f (x). •

Note that, in particular, the least fixed point of f is equal to f ↑ α for

some ordinal α. We call the smallest ordinal α with this property the closure
ordinal of f .
One other point to make in this context is that Kleene’s theorem shows
that ω-continuity ensures that in finding a fixed point the iteration will not
continue beyond the first infinite ordinal ω. This contrasts with the Knaster-
Tarski theorem, where it may be necessary to iterate beyond ω if one only has
monotonicity of the operators in question. This is a significant point in rela­
tion to computability considerations and explains the importance of Kleene’s
theorem in the theory of computation.

1.2 First-Order Predicate Logic

We assume that the reader has a slight familiarity with first-order predicate
logic, but for convenience we summarize next the elementary concepts of the
subject, beginning by formally describing its syntax.10

1.2.1 Syntax of First-Order Predicate Logic

As usual, an alphabet A consists of the following classes11 of symbols:
a (possibly empty) collection of constant symbols a, b, c, d, . . .; a non-empty
collection of variable symbols u, v, w, x, y, z, . . .; a (possibly empty) collection
of function symbols f, g, h, . . .; and a non-empty collection of predicate sym­

10 Our approach to the syntax and semantics of first-order logic is standard and is to be

found in any of the well-known texts on mathematical logic, see, for example, [Hodel, 1995,
Mendelson, 1987]. For fuller details of logic in relation to logic programming, the reader
may care to consult [Apt, 1997] or [Lloyd, 1987].
11 Similarly, our use of classes in the definition of an alphabet is also standard in developing

first-order logic and, in our case, is not intended to hint at foundational issues. In logic
programming practice, the classes referred to, namely, those of constant, variable, function,
and predicate symbols, will be finite sets. When working with the set groundJ (P ) defined in
Chapter 2, J will usually (although not necessarily) denote the Herbrand preinterpretation,
and then we will in effect be working with a set containing a possibly denumerable collection
of elements (atoms, in fact).
8 Mathematical Aspects of Logic Programming Semantics

bols12 p, q, r, . . .. In addition, we have the connectives ¬, ∧, ∨, →, and ↔; the

quantifiers ∀ and ∃ ; and the punctuation symbols “(”, “)” and “, ”. The arity
of a function symbol f or of a predicate symbol p is commonly denoted by
#(f ) or by #(p).
In the following four definitions, we assume that A denotes some fixed,
but arbitrary, alphabet.

1.2.1 Definition We define a term (over ) A inductively13 as follows.

(1) Each constant symbol in A is a term.
(2) Each variable symbol in A is a term.
(3) If f is any n-ary function symbol in A and t1 , . . . , tn are terms, then
f (t1 , . . . , tn ) is a term.
A term is called ground if it contains no variable symbols.

1.2.2 Definition An atom, atomic formula, or proposition A (over A) is an

expression of the form p(t1 , . . . , tn ), where p is an n-ary predicate symbol in
A and t1 , . . . , tn are terms (over A).

1.2.3 Definition A literal L is an atom A or the negation ¬A of an atom

A. Atoms A are sometimes called positive literals, and negated atoms ¬A are
sometimes called negative literals.

1.2.4 Definition A (well-formed ) formula (over A) is defined inductively as

follows.
(1) Each atom is a well-formed formula.
(2) If F and G are well-formed formulae, then so are ¬F , F ∧ G, F ∨ G,
F → G, and F ↔ G.
(3) If F is a well-formed formula and x is a variable symbol, then ∀xF and
∃xF are well-formed formulae also.
A well-formed formula is called ground if it contains no variable symbols.
Thus, in particular, a ground atom is an atom containing no variable symbols.
Of course, brackets are needed in writing down well-formed formulae to
avoid ambiguity. Their use can be minimized, however, by means of the cus­
tomary precedence hierarchy (in descending order) in which ¬, ∀, ∃ have high­
est precedence, followed by that of ∨, followed next by the precedence of ∧,
and finally followed by → and ↔ with the lowest precedence.
12 Constant symbols, variable symbols, function symbols, and predicate symbols are some­

times referred to as simply constants, variables, functions, and predicates, respectively

13 As usual, in giving inductive definitions of sets, we omit the explicit statement of the

closure step and assume that what is being defined is the smallest set satisfying the basis
and induction steps.
 Order and Logic 9

1.2.5 Definition The first-order language L given by an alphabet A consists

of the set of all well-formed formulae determined by the symbols of A. We refer
to terms over A as terms in or over L.

1.2.6 Example Suppose we are given an alphabet A containing constant

symbols a and b; variable symbols x and y; a unary function symbol f
and a binary function symbol g; and a unary predicate symbol p and a bi­
nary predicate symbol q. Then the following are examples of terms over A:
a, b, x, y, f (a), f (x), g(a, f (b)), g(g(a, b), f (y)), f (g(x, b)), . . .. In particular,
we note that, for example, f (a) and g(a, f (b)) are ground terms, whereas
f (g(x, b)) is not.
Furthermore, the following are examples of well-formed formulae in the
first-order language L determined by A: p(a), q(a, g(b, b)), ¬p(x), q(x, g(a, y)),
q(x, g(a, y))∨(p(y)∧¬p(x)), p(x) ← p(f (a))∧q(f (b), g(x, f (y)))∧q(x, g(y, b)),
p(x) ↔ q(f (x), g(x, x)), ∀x(p(x) ← p(a) ∧ ¬q(f (b), g(x, f (x))) ∧ q(x, g(x, b))).
In particular, the last of these is in a form of great significance in logic pro­
gramming. Moreover, p(a) and q(a, g(b, b)), for example, are ground (atomic)
formulas, whereas ∀x∀y(p(x) ← p(a) ∧ ¬q(f (b), g(x, f (y))) ∧ q(x, g(y, b))) is
not ground.

1.2.2 Semantics of First-Order Predicate Logic

The definition formally describes the syntax of first-order predicate logic.
We want now, briefly, to describe formally the semantics or meaning given to
well-formed formulae. In doing this, we adopt the usual set-based approach
from model theory, but with two caveats which direct us. The first is that we
do need to handle more truth values than just the two conventional ones. The
second is that we do not usually need to handle quantified formulae because,
for purposes of the semantics of logic programs P , we usually consider the set
ground(P ), as defined in Chapter 2, instead of P itself, and elements of the
former contain no variable symbols and no quantifiers. However, in order to
proceed further it is necessary to discuss spaces of truth values, and we do
this next.
In classical two-valued logic and almost always in mathematics it is usual
to employ the set T W O = {f , t} of truth values false f and true t. How­
ever, in many places in logic programming and in other areas of comput­
ing, it has been found advantageous to employ more truth values than these.
Indeed, quite early on, Melvin Fitting argued in several places for the use
in logic programming of Kleene’s strong and weak three-valued logics, see
[Fitting, 1985, Fitting and Ben-Jacob, 1990], for example, in which the truth
set is T HREE = {u, f , t}. Here, f denotes false and t denotes true, again, but
u denotes a third truth value which may be thought of as representing under-
defined, none (neither true nor false) or no information, or, in some contexts,
10 Mathematical Aspects of Logic Programming Semantics

non-termination.14 These and other three-valued logics will be encountered

in Chapter 2 and in many other places in Chapters 3, 5, 6, and 7.
Fitting also considered Belnap’s four-valued logic15 in which the truth set
is FOU R = {u, f , t, b}. Here, b denotes a fourth truth value intended to
represent both true and false, both or overdefined, which, it can be argued,
should be used to handle the conflicting information “both true and false”
returned, perhaps, in a distributed logic programming system. On a point of
notation, we remark that the listing of the elements in T W O corresponds to
the truth ordering ≤t , as defined in Section 1.3.2, and in the case of T HREE
and FOU R the listing is derived from the knowledge ordering ≤k , see again
Section 1.3.2, with incomparable elements listed alphabetically.
A fundamental concept throughout this work is that of valuation, or in­
terpretation, and also that of model. Indeed, spaces of interpretations are one
of the central concepts here when viewed as the carrier sets for various se­
mantic operators determined by programs. We will usually work later on in
the truth sets T W O and T HREE and sometimes in FOU R. Nevertheless, in
formulating the concepts of valuation and interpretation, we will work quite
generally, at no extra cost, and allow arbitrary sets of truth values and cer­
tain connectives defined on them. Thus, let T denote an arbitrary set of truth
values or truth set containing at least two elements, one of which will be the
distinguished value t, denoting true. We assume further that certain binary
connectives, namely, conjunction (∧) and disjunction (∨) are given, together
with a unary connective negation (¬), as functions over T . A third binary
connective implication (←) may also be given or it may be defined in terms
of the other connectives, and the latter is the way we will usually handle im­
plication. However, we will defer giving the definition of implication we want
until we have dealt with orderings on truth sets, see Definition 1.3.3. A set
T together with specified definitions of these connectives will be referred to
as a logic and, when the definitions of the connectives are understood, will
be denoted simply by the underlying truth set T without causing confusion.
Quite often, the definitions of ∧, ∨, and ¬ are given by means of a truth
table, and this is the case for most of the logics we encounter here. For ex­
ample, Table 1.1 specifies Belnap’s logic as employed by Fitting and by us.
It contains classical two-valued logic and Kleene’s strong three-valued logic
as sublogics.16 Moreover, FOU R is a complete lattice, as we see later, and is
therefore technically easy to work with. Indeed, these are some of the reasons
why four-valued logic plays an important unifying role in the theory17 and is

14 The truth value u is sometimes denoted in the literature by n, indicating none.

15 We refer to [Belnap, 1977, Fitting, 1991, Fitting, 2002], but note that Fitting worked
with a minor variant of the logic defined in [Belnap, 1977]; we work with this same variant
of Belnap’s definition.
16 The term a sublogic S of a logic T means that S is a subset of the set T of truth values,

and the connectives in S are restrictions to S of the corresponding connectives in T .

17 Fitting has shown the utility of FOU R, when viewed as a bilattice, in giving a unified

treatment of several aspects of logic programming, and we refer the reader to [Fitting, 2002]
and the works cited therein for more details.
 Order and Logic 11

TABLE 1.1: Belnap’s four-valued logic.

p q ¬p p∧q p∨q
u u u u u
u f u f u
u t u u t
u b u f t
f u t f u
f f t f f
f t t f t
f b t f b
t u f u t
t f f f t
t t f t t
t b f b t
b u b f t
b f b f b
b t b b t
b b b b b

the main reason we work with it despite the fact that most of our applications
are to T W O and T HREE. Notice that Kleene’s weak three-valued logic also
uses the truth set T HREE, but its connectives are defined by Table 5.1.18

The next two definitions are fundamental. In presenting the first of them,
we will use the notation commonly employed in logic programming.

1.2.7 Definition Let L be a first-order language and let D be a non-empty

set. A preinterpretation J for L with domain D (of preinterpretation) is an
assignment ·J which satisfies the following: (1) cJ ∈ D for each constant
symbol c in L, and (2) f J is an n-ary function over D for each n-ary function
symbol f in L. A J-variable assignment is a (total) mapping, θ, say, from
variable symbols to elements of D.
Given a preinterpretation J with domain D and a J-variable assignment
θ, we can assign to each term t in L an element of D, called its denotation or
term assignment, inductively as follows: (tθ)J = θ(t) if tA is a variable symbbol,
(tθ)J = tJ if t is a constant symbol, and (tθ)J = f J (t1 θ)J , . . . , (tn θ)J if
t = f (t1 , . . . , tn ) for some n-ary function symbol f and terms t1 , . . . , tn . For
J
an atom A A = p(t
J
b in the language L, we define (Aθ) to be the
1 , . . . , tn ), say,
J
symbol p (t1 θ) , . . . , (tn θ) and call this a J-ground instance of the atom
p(t1 , . . . , tn ). We denote by BL,J the set of J-ground instances of atoms in L.
18 Indeed, disjunction and conjunction in Kleene’s weak three-valued logic are given by

∨2 and ∧3 , respectively, in Table 5.1, see also [Fitting, 1994a].

12 Mathematical Aspects of Logic Programming Semantics

Thus, BL,J is the set of all symbols p(d1 , . . . , dn ), where p is an n-ary predicate
symbol in L and d1 , . . . , dn ∈ D. •

1.2.8 Definition Let L be a first-order language, let J be a preinterpretation

for L with domain D, and let T be a logic. A valuation or interpretation for
L (based on J) with values in T is a mapping v : BL,J → T . Let v : BL,J → T
be a valuation and let θ be a J-variable assignment. Then v and θ determine,
inductively, a well-defined truth value in T for any quantifier-free, well-formed
formula F in L by means of the construction of F and the definitions of the
connectives in T . We say that v is a model for F , written v |= F , if v gives
truth value t to F . We sometimes refer to valuations, interpretations, and
models based on J as J-valuations, J-interpretations, and J-models.

In fact, if T is ordered as a complete lattice (and this issue will be con­

sidered shortly), then a valuation v gives unique truth value in T , in the
standard way, to any closed well-formed formula F in L: universal quantifi­
cation corresponds to the infimum of a set of truth values, and existential
quantification corresponds to the supremum of a set of truth values. The term
closed here has, of course, its normal meaning in mathematical logic, namely,
that each variable symbol occurring in F falls within the scope of a quantifier.
(By default, we allow the term closed to apply to formulae with no variable
symbols and no quantifiers.) Once this observation is made, one can go on, in
the standard way, to define at our present level of generality the terms model,
(un)satisfiable, valid, and logical consequence when applied to sets of closed
well-formed formulae.

1.3 Ordered Spaces of Valuations

Following Definition 1.2.8, we will generally denote the set of all valua­
tions for L based on J with values in T by I(BL,J , T ), and we will consider
I(BL,J , T ) as an ordered set. The orderings we have in mind are derived from
orderings on T , and the set BL,J plays no role in this. Therefore, to ease
notation we will work with an arbitrary set X for the rest of this chapter.
Thus, we regard a valuation or interpretation for the time being as simply a
mapping X → T and denote the set of all these by I(X, T ); typical elements
of I(X, T ) will be denoted by u, v, etc. Later on, in applying the results of this
section, we will of course take X to be a set of ground atoms or of J-ground
instances of atoms, and no confusion will be caused. There is, however, a con­
vention we need to establish concerning the terminology “valuation” versus
“interpretation”, as follows.
 Order and Logic 13

1.3.1 Remark Much of the theory of logic programming semantics is con­

cerned with sets of valuations. It is important, therefore, to have convenient
notation for valuations and to have ways of representing them, which both
facilitate discussion and also allow easy passage backwards and forwards be­
tween the different representations employed. There are three ways of han­
dling valuations, which are commonly used in the literature on the subject
and which we adopt also. Having these three forms available will, in certain
places, greatly increase readability and reduce technical difficulty.
First, when considering general structures such as orderings or topologies
on I(X, T ), the easiest way is to think of valuations as mappings, and this
we will usually do. Thus, in the main, our future use of the term valuation
will refer to mappings whose domain is a set of atoms (or ground instances of
atoms) and whose codomain is a set T of truth values.
Second, when T is a small set containing two, three, or four elements,
say, it is convenient to identify a valuation with the (ordered) tuple of sets
on which it takes the various truth values in T , as discussed in Section 1.3.2.
This is by far the most frequently used representation, and, in common with
most authors, we will in future usually employ the term interpretation when
thinking in these terms. Thus, as we progress, more and more we employ the
terminology interpretation instead of valuation, use the standard notation I,
K, etc. to denote interpretations, and adopt the notation described at the end
of Section 2.1 for sets of interpretations.
Third, there is yet another representation frequently used for interpreta­
tions when T is the set T HREE, namely, signed sets as discussed in Sec­
tion 1.3.3. This form is particularly expressive, as we shall see in Chapter 2,
when one wants to discuss the truth value of conjunctions of literals in relation
to T HREE.

1.3.1 Ordered Spaces of Valuations in General

Usually, the set T of truth values carries an order, ≤, in which (T , ≤)
is perhaps a complete partial order, complete upper semi-lattice, complete
lattice, or Scott domain, with bottom element ⊥, say, or even a bilattice19
when equipped with two compatible orderings. When T carries an ordering,
≤, we can define the corresponding pointwise ordering on I(X, T ), denoted
by [, in which v1 [ v2 if and only if v1 (x) ≤ v2 (x) for all x ∈ X.
It is routine to check that the ordering [ is in fact a partial order if ≤ is
one. Moreover, if T has a bottom element, ⊥, then the valuation which maps
each x in X to ⊥ serves as a bottom element in I(X, T ), and we may denote
this valuation simply by ⊥ again, without causing confusion. Finally, if (T , ≤)
is a Scott domain, we shall say that a valuation v in I(X, T ) is finite if v(x) is

19 A (complete) bilattice is a set D carrying two partial orders in each of which D is a

(complete) lattice. In addition, the two orderings are required to interact with each other
so as to obtain various distributive laws.
14 Mathematical Aspects of Logic Programming Semantics

a compact element in (T , ≤) for each x ∈ X, and the set {x ∈ X | v(x) = ⊥}

is finite.
The structural properties of I(X, T ) may be summarized in the following
result.20

1.3.2 Theorem Let X be a non-empty set, let (T , ≤) be an ordered set of

truth values with bottom element ⊥, and let I(X, T ) be endowed with the
pointwise ordering and bottom element just defined.

(a) If (T , ≤) is a partially ordered set, then so is I(X, T ).

(b) If (T , ≤) is an ω-complete partial order, then so is I(X, T ).

(c) If (T , ≤) is a complete partial order, then so is I(X, T ).

(d) If (T , ≤) is a complete upper semi-lattice, then so is I(X, T ).

(e) If (T , ≤) is a complete lattice, then so is I(X, T ).

(f) If (T , ≤) is a Scott domain, then so is I(X, T ). In this case, the compact

elements of I(X, T ) are the finite valuations.

Proof: (a) As already noted, it is routine in this case to verify that the ordering
on I(X, T ) is a partial ordering, with bottom element as already specified.
(b) The argument in this case is similar to the next and is omitted.
(c) If M ⊆ I(X, T ) is directed, then it is easy to check that, for each
x ∈ X, the set {v(x) | v ∈ M } is directed and hence has a supremum  in T .
It is now clear that the valuation
 vM defined on X by v M (x) = {v(x) |
v ∈ M } is the suprem um, M , of M in I(X, T ). Indeed, for any directed
subset
 M ⊆ I(X, T ), M satisfies the following relationship: for each x ∈ X,
( M )(x) = (M (x)), where M (x) denotes the set{v(x) | v ∈ M }.
(d) By the argument used in (c), the supremum M exists for any directed n
subset M of I(X, T ). Also, n for any subset
n M of I (X, T ), we have that M
exists and is defined by ( M )(x) = (M (x)) for each x ∈ X, where again
M (x) denotes the set {v(x) | v ∈ M }.
(e) It is clear from the argument in (c) that any subset M of I(X, T ) has
a supremum in I(X, T ), and, from (d), M has an infimum in I(X, T ).
(f) We begin by showing that the finite valuations are compact elements.
Suppose that v is a finite valuation and that {x ∈ X | v(x) = ⊥} =
{x1 , . . . , xn }. Suppose that
 M = {uk | k ∈ K} is a directed set of valuations
in I(X, T ) such that v [ M  arbitrary element of {x1 , . . . , xn }.
. Let xi be an
Then we have that v(xi ) ≤ M (xi ) = (M (xi )), that v(xi ) is a compact
element, and that {uk (xi ); k ∈ K} is directed. Therefore, there is uki ∈ M
such that v(xi ) ≤ uki (xi ), and we obtain such uki for i = 1, . . . , n. Since M is
directed, there is u ∈ M such that uki [ u for i = 1, . . . , n, and it now clearly
follows that v [ u. Hence, v is compact.
20 For further details here and in the next three subsections, see [Seda, 2002].
 Order and Logic 15

In the converse direction, suppose that u is any valuation on X. Let M

denote the set of all finite valuations v such that v [ u. Let v1 , v2 ∈ M
and suppose that x ∈ X is such that not both v1 (x) and v2 (x) are equal to
the bottom element (there are only finitely many such x, of course). Noting
that approx(u(x)) in T is directed, that v1 (x), v2 (x) ∈ approx(u(x)) and by
considering one-point valuations (namely, those valuations w such that w(x)
is not equal to the bottom element at at most one value of x), we see that
there is v3 (x) ∈ approx(u(x)) such that both v1 (x) ≤ v3 (x) and v2 (x) ≤ v3 (x).
It follows that there is an element v3 of M such that v1 [ v3 and v2 [ v3 and,
hence, that M is directed. Moreover, given x ∈ X and any a ∈ approx(u(x)),
let vax denote the one-point valuation
 which satisfies vax (x) = a and vax (y) = ⊥
x x
 all y = x. Then va ∈ M , and {va (x) | a ∈ approx(u(x))} = u(x). Thus,
for
M = u.
It now follows from the observations just made that if u is compact, then
there is v ∈ M such that u [ v, and hence the set {x ∈ X | u(x) = ⊥} is finite.
We claim that u(x) is a compact element in (T , ≤) for each x ∈ X. Suppose
otherwise, that is, that there is x0 ∈ X with u(x0) non-compact in (T , ≤).
Then there is a directed set N in T with u(x0 ) ≤ N for which there is no
n ∈ N with u(x0 ) ≤ n. Define the family Nn consisting of the elements un of
I(X, T ), n ∈ N , by setting un (x)= u(x) for all x = x0 and setting un (x0 ) = n.
Then Nn is directed and u [ {un | n ∈ N }, yet we do not have u [ un
for any n ∈ N . This contradicts the fact that u is a compact element, and
hence, for each x ∈ X, u(x) is a compact element. Thus, the compact elements
are indeed the finite valuations, and, moreover, we now see that approx(u) is
directed and that approx(u) = u for each valuation u ∈ I(X, T ).
Finally, if u1 and u2 are two consistent
 finite elements in I(X, T ), then
the valuation v defined by v(x) = {u1 (x), u2 (x)}, for each x ∈ X, is the
supremum of u1 and u2 (and is, in fact, a finite element). This completes the
proof. •

1.3.2 Valuations in Two-Valued and Other Logics

The most prominent declarative semantics for logic programs employ clas­
sical two-valued logic, three-valued logic, or, to a lesser extent, four-valued
logic. The corresponding truth sets T for these logics are T W O, T HREE,
and FOU R, as already discussed. We examine these cases next in some de­
tail in light of Theorem 1.3.2 and also introduce some convenient notation
for these special cases. We begin by considering the orderings involved on the
three sets of truth values that we are currently discussing.
In the case of classical two-valued logic, the ordering usually taken is the
truth ordering. This is the partial ordering ≤t satisfying f <t t and is often
denoted just by ≤; it turns T W O into a complete lattice with f as the bottom
element.
For three-valued logic, there are two natural orderings usually considered:
16 Mathematical Aspects of Logic Programming Semantics

� �

b
/
///
/
≤k f t //
// ≤k f t
/// //
/ //
// //
u //
u

_
≤t _
≤t

FIGURE 1.1: Hasse diagrams for T HREE (left) and FOU R (right).

the knowledge ordering ≤k and the truth ordering ≤t . The first of these, ≤k ,
is the partial order indicated by the Hasse diagram to the left in Figure 1.1 in
which u is the bottom element. This ordering turns T HREE into a complete
upper semi-lattice, but not a complete lattice. The second ordering, ≤t , is
the partial ordering satisfying f <t u and u <t t; it turns T HREE into a
complete lattice with f as the bottom element.
Finally, on FOU R, there are again the two orderings: ≤k , the knowledge
ordering, and ≤t , the truth ordering. They are indicated by the Hasse diagram
on the right-hand side of Figure 1.1. In each of them, FOU R is a complete
lattice and indeed is a complete bilattice, with bottom elements as indicated
by the Hasse diagram.
At this point, having defined the orderings we want on FOU R, it will be
convenient to record the definition we use of implication before resuming the
study of orderings on valuations. Note that the definition reduces to mate­
rial implication in two-valued logic and gives the definition we want later for
Kleene’s strong three-valued logic.

1.3.3 Definition For all truth values t1 and t2 in FOU R, we define implica­
tion by taking the truth value of t1 ← t2 to be f if and only if t1 <t t2 in the
truth ordering ≤t , and t otherwise.

In each of the three cases we are considering, the truth set T is easily seen
to be a Scott domain in the truth ordering and also in the knowledge ordering
in the latter two cases. Furthermore, each element is compact. Therefore, on
applying Theorem 1.3.2 with the induced pointwise orderings involved, we
obtain the following result, which summarizes the previous discussion.
 Order and Logic 17

1.3.4 Theorem Let X be an arbitrary set. Then the following statements

hold.

(a) In case T is the truth set T W O, the set I(X, T ) is a complete lattice in
the ordering [t .

(b) In case T is the truth set T HREE, the set I(X, T ) is a complete up­
per semi-lattice in the ordering [k , but not a complete lattice, and is a
complete lattice in the ordering [t .

(c) In case T is the truth set FOU R, the set I(X, T ) is a complete lattice in
each of the orderings [k and [t .

Furthermore, in each case and in each ordering, the set I(X, T ) is a Scott
domain whose compact elements are precisely those valuations v for which
the set {x ∈ X | v(x) = ⊥} is finite, where ⊥ denotes the appropriate bottom
element. •

Notice that the order structure here is independent of the actual logic in­
volved, as distinct from the underlying truth set. Thus, for example, Kleene’s
strong and weak three-valued logics give rise to precisely the same order struc­
ture on I(X, T HREE); the difference between them is in the definitions of
the connectives, rather than in their order structure.
We next take up the point made in Remark 1.3.1, concerning the repre­
sentation of a valuation in terms of the sets on which it takes various truth
values in T W O, T HREE, or FOU R.
Let v be a valuation, and let vu = v −1 (u), let vf = v −1 (f ), let vt = v −1 (t),
and let vb = v −1 (b); these sets are pairwise disjoint subsets of X, and some
may be empty. A valuation v taking values in T W O is clearly completely
determined by the subset I = vt of X and therefore can be identified with
I. A valuation taking values in T HREE can be identified either with the
pair I = (vt , vf ) of subsets of X or with the pair I = (vt , vu ). The former
choice will be made when we are concerned with the ordering [k , so that
the bottom element is u and this is also the “default” value in the sense that
vu = X \ (vt ∪ vf ). The latter choice will be made when we are concerned with
the ordering [t , so that the bottom element is f and this is also the default
value in that vf = X \ (vt ∪ vu ). Finally, a valuation v with values in FOU R
can be identified either with the triple I = (vt , vf , vb ) of subsets of X when
u is the bottom and default value or with the triple I = (vt , vu , vb ) when f is
the bottom and default value.
Conversely, a subset I of X determines a valuation v : X → T W O with
the property that v(x) = t if and only if x ∈ I. Given the ordering [k , a pair
I = (It , If ) of disjoint subsets of X determines a valuation v : X → T HREE
which takes value t on It , takes value f on If , and, by default, takes value u
on X \ (It ∪ If ). Similarly, given the ordering [t , a pair I = (It , Iu ) of disjoint
subsets of X determines a valuation v : X → T HREE which takes value t on
18 Mathematical Aspects of Logic Programming Semantics

It , takes value u on Iu , and, by default, takes value f on X \(It ∪Iu ). Precisely

the same remarks apply to triples I = (It , If , Ib ) and to triples I = (It , Iu , Ib )
in relation to valuations v : X → FOU R.
This passage between mappings and tuples of subsets will often be made
without explicit mention. However, as noted in Remark 1.3.1, we will, in the
main, use the term valuation to refer to mappings and the term interpretation
to refer to tuples of sets, and it will be convenient to employ the following
terminology.

1.3.5 Definition A valuation or interpretation taking values in T W O,

T HREE, or FOU R will be called two-valued , three-valued , or four-valued ,
respectively.

The identification above, of valuations with tuples of sets, carries the point-
wise ordering of valuations over to the “pointwise” ordering of interpretations,
and we employ exactly the same notation for the orderings in the correspond­
ing cases. We obtain the following result, whose proof is straightforward and
will be omitted. (There is the possibility of confusion here unless one remem­
bers that coordinate positions in the tuples are labelled with truth values and
that the truth value not present is the default value. Thus, for example, in
the case of three-valued valuations, the two coordinate positions are either
ordered with t and f in that order or ordered with t and u in that order,
and similarly for four-valued valuations. The only way to avoid this minor
irritation is to use pairs of sets to represent two-valued valuations, triples of
sets to represent three-valued valuations, and quadruples of sets to represent
four-valued valuations. However, this is not customarily done.)

1.3.6 Theorem The following statements hold in relation to interpretations

on X.

(a) If I and K are two-valued interpretations, then I [t K if and only if

I ⊆ K as subsets of X. The bottom element for the set of two-valued
interpretations is given by the empty set, ∅.

(b) If I and K are three-valued interpretations, then I [k K if and only

if It ⊆ Kt and If ⊆ Kf . Also, I [t K if and only if It ⊆ Kt and
Kf ⊆ If . In both orderings, the bottom element for the set of three-valued
interpretations is given by the appropriate pair (∅, ∅).

(c) If I and K are four-valued interpretations, then I [k K if and only if

It ⊆ Kt ∪ Kb , If ⊆ Kf ∪ Kb , and Ib ⊆ Kb . Also, I [t K if and only
if It ⊆ Kt , Iu ⊆ Ku ∪ Kt , and Ib ⊆ Kb ∪ Kt . In both orderings, the
bottom element for the set of four-valued interpretations is given by the
appropriate triple (∅, ∅, ∅). •
 Order and Logic 19

Notice that the difference in the form of the statements in (b) and (c) in
Theorem 1.3.6 concerning the truth ordering [t results from the fact that [t
is a total order in (b), but it is not a total order in (c).
In all cases we are currently considering, except one, we are working in a
complete lattice. Hence, the valuation mapping each element of X to the ap­
propriate top element is itself a top element. The one exception is the case of
three-valued interpretations in the order [k . In this case, it is clear that those
interpretations I = (It , If ) for which It ∪ If = X are maximal elements for
the ordering [k . Moreover, each maximal element I = (It , If ) gives rise to the
two-valued interpretation It , and, conversely, each two-valued interpretation
I gives rise to a maximal three-valued interpretation (I, X \ I). Moreover, this
correspondence is evidently one-to-one. Thus, the two-valued interpretations
can be thought of as maximal three-valued interpretations. Indeed, the max­
imal elements are called total interpretations, while the remaining elements
are called partial interpretations.

1.3.3 Signed Sets and Three-Valued Interpretations

As mentioned in Remark 1.3.1, there is an alternative and useful way of
thinking of three-valued interpretations relative to the ordering [k (so that u
is the current default value in the representation of interpretations as pairs of
sets), and we consider it next.
Let X denote an arbitrary set, and form the set ¬X of symbols ¬ x for
x ∈ X. If X happens to be a set of atoms or of literals, then ¬x is meaningful;
otherwise, we are working formally. In any case, we assume that x and ¬x
are never equal. Given a subset I of X, we let ¬I denote the subset of ¬X
consisting of those ¬ x for x ∈ I. A subset of X ∪ ¬X is called a signed subset
of X and is called consistent if it does not contain both x and ¬x for any x.
Clearly, any signed subset of X has the form I + ∪ ¬I − , where I + and I − are
subsets of X, and is consistent if and only if I + and I − are disjoint.
Every consistent signed subset I = I + ∪ ¬I − of X gives rise to the three-
valued interpretation (I + , I − ). Then, thinking of I as this three-valued inter­
pretation, we have It = I + = {x ∈ X | x ∈ I} and If = I − = {x ∈ X |
¬x ∈ I}. Conversely, every three-valued interpretation I = (It , If ) = (I + , I − )
gives rise to the consistent signed subset I + ∪ ¬I − of X. Moreover, this cor­
respondence is evidently one-to-one, and so I(X, T HREE) can be identified
with the set of all consistent signed subsets of X, and we will quite frequently
use this fact later on without further notice. Indeed, in this representation,
we have I [k K if and only if I + ∪ ¬I − ⊆ K + ∪ ¬K − , and so [k corresponds
to subset inclusion of signed subsets, and, furthermore, the bottom element is
the empty set thought of as a consistent signed subset of X.
Now let X denote a set of atoms in a first-order language L, and let I be
a three-valued interpretation viewed as a consistent signed subset of X. For
a literal L = A, where A is an atom, we write L ∈ I if A ∈ I, and we write
¬L ∈ I if ¬A ∈ I. Similarly, if L = ¬A, we write L ∈ I if ¬A ∈ I, and we
20 Mathematical Aspects of Logic Programming Semantics

write ¬L ∈ I if A ∈ I. Using these observations, we now say that a literal L

is true in I if L ∈ I, that L is false in I if ¬L ∈ I, and that L is undefined
in I otherwise. Notice that these facts depend on, and indeed are equivalent
to, defining the negation operator ¬ from T HREE into itself by means of
Table 1.1, so that ¬(t) = f , ¬(f ) = t, and ¬(u) = u.
Finally, we note that four-valued interpretations can be treated in the same
sort of way as we have just handled three-valued interpretations by including
inconsistent signed sets in the discussion, but we omit the details of this as
we have no need of them.

1.3.4 Operators on Spaces of Valuations

As we have seen, an ordering on a space T of truth values induces an
ordering on the corresponding spaces I(X, T ). Similarly, various connectives
defined on T induce operators defined on I(X, T ), and we close this chapter
by briefly discussing these next. They will be considered further in Chapter 3.
In fact, we concentrate on Belnap’s four-valued logic, in which the truth
set is FOU R and the connectives are determined by the truth table, Ta­
ble 1.1. Since classical two-valued logic and Kleene’s strong three-valued logic
are sublogics of FOU R, they are subsumed in our discussion of FOU R and
therefore need not be considered separately.
The first of these operators arises through negation, and is the opera­
tor mapping I(X, T ) into itself, and still denoted by ¬, in which (¬v)(x) =
¬(v(x)) for each x ∈ X, where v is an arbitrary element of I(X, T ).
Likewise, the connectives ∨ and ∧ determine (binary) operators mapping
I(X, T ) × I(X, T ) into I(X, T ) defined by (u ∨ v)(x) = u(x) ∨ v(x) and
(u ∧ v)(x) = u(x) ∧ v(x), for each x ∈ X, where u and v are arbitrary elements
of I(X, T ). We note that the overloading of the symbols ∨ and ∧ should not
cause any difficulties. Of course, one can similarly deal with other connectives
such as → and ↔.
If v1 , v2 ∈ I(X, T ) satisfy the conditions v1 [t v2 , v1 (x) = f and v2 (x) = t
for some x, then it is clear that ¬v1 [t ¬v2 . Hence, ¬ is not monotonic in this
case. Thus, ¬ is not order continuous in the truth orderings [t . It is, however,
order continuous in the orderings [k , as we shall see in Chapter 3, where we
also consider the continuity of the other operators ∨ and ∧.
The following observation is just one of the many interesting properties
possessed by I(X, T ) when we take T to be the logic FOU R, as we are
currently doing.

1.3.7 Proposition The operators ∨ and ∧ are monotonic in each argument.

Proof: Given v ∈ I(X, T ), it must be shown that the mappings u �→ u ∨ v
and u �→ v ∨ u are both monotonic, and, since ∨ is commutative, it suffices
to show that either is monotonic. It is straightforward to check this from the
truth table, Table 1.1, and the Hasse diagram for FOU R, Figure 1.1, and we
omit details. Precisely the same comments apply also to the operator ∧. •
 Order and Logic 21

Another interesting fact about FOU R, which emerges from its truth table
and its Hasse diagram, is the following result.

1.3.8 Proposition
 n truth ordering ≤t on FOU R, we have
Relative to the
t1 ∨ t2 = {t1 , t2 } and t1 ∧ t2 = {t1 , t2 } for all truth values t1 and t2 . In
particular, in classical two-valued logic and Kleene’s strong three-valued logic
relative to ≤t , we have t1 ∨ t2 = max{t1 , t2 } and t1 ∧ t2 = min{t1 , t2 } for all
truth values t1 and t2 .
Chapter 2

The Semantics of Logic Programs

The objective of this chapter is to introduce the central topic of study in this
work, namely, logic programs, together with several of the main issues and
questions which will be addressed in later chapters. In order to ensure that
our treatment is as self-contained as possible, we will take care to formally
define all concepts which we consider in detail here and later on. In addition, to
assist the reader, we give ample references to those topics which we encounter,
but do not treat in detail.
For the course of this and subsequent chapters, our main focus will be on
declarative semantics, and, as already noted in the Introduction, issues con­
cerning procedural aspects will play only a minor role. In particular, in this
chapter and later in Chapter 5, we will introduce some of the best known
declarative semantics for logic programs, and we will develop a uniform treat­
ment of them applicable not only to resolution-based logic programming, but
also to non-monotonic reasoning as well.
Frequently, a declarative semantics is given by assigning intended models to
logic programs. This is done by selecting from the set of all models for a logic
program, a subset which contains those models with some properties deemed
to be desirable depending on one’s objectives and intended applications. All
the semantics which we will discuss can be described in terms of fixed points
of operators associated with logic programs, and they are all well-established.
Our new and novel contribution in this chapter is the development of a uniform
and operator-free characterization of them.
Our first task, however, is to introduce formally some of the basic concepts
and notation which will be needed throughout the sequel.1

2.1 Logic Programs and Their Models

2.1.1 Definition Given a first-order language L, a clause, program clause,
or rule in L is a formula of the form
(∀x1 ) . . . (∀xl )(A ← L1 ∧ . . . ∧ Ln ),
1 We follow the presentation of semantics from [Hitzler and Wendt, 2002, Hitzler, 2003b,

Hitzler and Wendt, 2005, Hitzler, 2005, Knorr and Hitzler, 2007].

23
24 Mathematical Aspects of Logic Programming Semantics

where l, n ∈ N; A is an atom in L; L1 , . . . , Ln are literals in L; and x1 , . . . , xl

are all the variable symbols occurring in the formula. We will follow common
practice and abbreviate such a clause by writing simply

A ← L1 , . . . , Ln ,

so that the universal quantifiers are understood, and the conjunction symbol
∧ is replaced by a comma. The atom A is called the head of the clause, and
the conjunction L1 , . . . , Ln is called the body of the clause; the literals Li ,
i = 1, . . . , n, in the body L1 , . . . , Ln are called body literals. If a body literal
L is an atom B, say, then we say that B occurs positively in the body of the
clause. If L is a negated atom ¬B, then we say that B occurs negatively in
the body of the clause. By an abuse of notation, we allow n = 0, by which we
mean that the body can be empty, and in this case the clause A ←, or simply
A, is also called a unit clause or a fact. It will sometimes be convenient to
further abbreviate a clause by writing

A ← body,

wherein body denotes the body of the clause. Furthermore, we will use body
not only to denote a conjunction of literals, but also to denote the correspond­
ing set containing these literals. This further abuse of notation will substan­
tially ease matters in some places and will not cause confusion. Note that in
doing this, we are ignoring the ordering of the literals in clause bodies. This will
not matter most of the time, since we are not much concerned with procedural
matters, as already noted, and for this reason we often denote a typical clause
by A ← A1 , . . . , An , ¬B1 , . . . , ¬Bk , say, where all the Ai , i = 1, . . . , n, and all
the Bj , j = 1, . . . , k, are atoms in L. Notice that we allow ourselves a bit of
latitude in the subscripts we employ in writing down clauses, and, for example,
the roles of n in the clause just considered and in the clause A ← L1 , . . . , Ln
above are not identical in general, unless there are no negated atoms present,
of course.
A normal logic program is a finite set of clauses. A definite logic program
is a normal logic program in which no negation symbols occur. The term
program will subsequently always mean a normal program. Definite programs
are sometimes called positive programs, and obviously every definite program
is a normal program. A propositional logic program is a program in which all
predicate symbols are of arity zero. •

In most cases, the underlying first-order language, or simply the underlying

language, LP of a program P will not be given explicitly, but will be under­
stood to be the (first-order) language generated by the constant, variable,
function, and predicate symbols occurring in P . However, when P does not
contain any constant symbols, we add one to LP , so that the underlying lan­
guage always contains at least one constant symbol. Propositional programs
 The Semantics of Logic Programs 25

will be treated slightly differently, and we will return to this point later, see
the examples following Definition 2.1.6.
We illustrate Definition 2.1.1 with a number of example programs, to which
we will return frequently later. At all times, unless stated to the contrary,
we will adhere to the following notational conventions concerning programs:
constant, function, and predicate symbols start with a lowercase letter and
are set in typewriter font unless they consist of a single letter only; variable
symbols start with an uppercase letter.

2.1.2 Program (Tweety1) Let Tweety12 be the program consisting of the

following clauses.

penguin(tweety) ←
bird(bob) ←
bird(X) ← penguin(X)
flies(X) ← bird(X), ¬penguin(X)

Tweety1 is intended to represent the following knowledge: tweety is a penguin,

bob is a bird, all penguins are birds, and every bird which is not a penguin
can fly.

2.1.3 Program (Even) Let Even be the program consisting of the following
clauses.

even(a) ←
even(s(X)) ← ¬even(X)

The intended meaning of this program is as follows: a is the natural number

0, and s is the successor function on natural numbers. Thus, the program
represents the knowledge that 0 is even, and if some number is not even, then
its successor is even.

Many of our later examples will be variations of the Even program theme
and will employ the successor notation for natural numbers. Consider, for
example, the following program.

2.1.4 Program (Length) Let Length be the program consisting of the fol­
lowing clauses.

length([], a) ←
length([H|T ], s(X)) ← length(T, X)

Following Prolog conventions, [] denotes the empty list, and [ · | · ] denotes

2 We borrow Tweety programs, in which a penguin usually called Tweety appears, from

the literature discussing the semantics of non-monotonic reasoning.

26 Mathematical Aspects of Logic Programming Semantics

a binary function whose intended meaning is the list constructor whose first
argument is the head of the list and whose second argument is its tail. Thus,
the program Length is intended to be a recursive definition of the “length of
lists” using the successor notation for natural numbers as in Program 2.1.3.
Length is an example of a definite program.

Thus far, we have specified the syntax of logic programs. We now turn our
attention to dealing with their semantics, and this is based on Definitions 1.2.7
and 1.2.8 of Chapter 1 with some notation peculiar to logic programming.

2.1.5 Definition Let P be a program with underlying language LP , and

let D be a non-empty set. A preinterpretation J for P with domain D is a
preinterpretation J for LP with domain D.
Let J be a preinterpretation for the program P , with domain D, and
let θ be a J-variable assignment. For a typical clause C in P of the form
A ← A1 , . . . , An , ¬B1 , . . . , ¬Bk , we let (Cθ)J denote

(Aθ)J ← (A1 θ)J , . . . , (An θ)J , ¬(B1 θ)J , . . . , ¬(Bk θ)J .

We call (Cθ)J a J-ground instance 3 of C. By groundJ (P ), we denote the set

of all J-ground instances of clauses in P . We denote by BP,J the set BLP ,J of
all J-ground instances of atoms in LP , that is, the collection of all elements
of the form p(d1 , . . . , dn ), where p is an n-ary predicate symbol in LP and
d1 , . . . , dn ∈ D. Usually, we will be working over a fixed, but arbitrary, prein­
terpretation J. In order to ease notation, we will often omit mention of J if it
causes no confusion, and instead of writing BP,J , groundJ (P ), J-ground in­
stance, etc., we will simply write BP , ground(P ), ground instance, etc. We will
frequently abuse notation even further by referring to elements of groundJ (P )
as (ground ) clauses and by applying to ground clauses terminology, such as
“definite”, already defined for program clauses.

Of particular interest is the so-called Herbrand preinterpretation of a pro­

gram. Its importance rests on the fact that, for many purposes, restricting
to Herbrand preinterpretations causes no loss of generality.4 For example, in
classical first-order logic, a set of clauses has a model if and only if it has a Her-
brand model. Indeed, in many cases in the literature on the subject, discussions
of logic programming semantics refer only to Herbrand (pre)interpretations
and Herbrand models.

2.1.6 Definition Given a program P with underlying language LP , the Her-

brand universe UP of P is the set of all ground terms in LP . The Herbrand
3 This extends the notion of a J-ground instance of an atom to a J-ground instance of a

clause, see [Lloyd, 1987, Page 12].

4 Nevertheless, we prefer to formulate the basic definitions in complete generality. For one

thing this is at no extra cost, and for another we require quite general preinterpretations in
our treatment of acceptable programs in Chapter 5.
 The Semantics of Logic Programs 27

preinterpretation J, say, for P , has domain UP and assigns constant and func­
tion symbols as follows, where we use the notation of Definition 1.2.7.

(1) For each constant symbol c ∈ LP , cJ is equal to c.

(2) For each n-ary function symbol f ∈ LP , f J : UPn → UP is the mapping

defined by f J (t1 , . . . , tn ) = f (t1 , . . . , tn ).

We illustrate these definitions by discussing some of the previous examples

in relation to them. For this purpose, and indeed for all example programs
unless otherwise noted, we consider the Herbrand preinterpretation.
For the program Tweety1 (Program 2.1.2), we obtain

UTweety1 = {bob, tweety},

BTweety1 = {penguin(bob), penguin(tweety),
bird(bob), bird(tweety),
flies(bob), flies(tweety)},

and ground(Tweety1) consists of the following clauses.

penguin(tweety) ←
bird(bob) ←
bird(tweety) ← penguin(tweety)
bird(bob) ← penguin(bob)
flies(tweety) ← bird(tweety), ¬penguin(tweety)
flies(bob) ← bird(bob), ¬penguin(bob)

For the successor notation used in the program Even (Program 2.1.3),
the following convention will be convenient: for n ∈ N, we denote the term
s(s(. . . s(x) . . .)), with n occurrences of s, by sn (x). We then obtain for the
Even program

UEven = {sn (a) | n ∈ N} ,

BEven = {even (sn (a)) | n ∈ N} ,

and ground(Even) consists of the following clauses.

even(a) ←
even sn+1 (a) ← ¬even (sn (a))
A b
for all n ∈ N

We note that the set ground(Even) is infinite. In fact, ground(Even) can

be thought of as an infinite propositional program consisting of clauses p0 ←
and pn+1 ← ¬pn , where, for each n ∈ N, pn is a propositional variable re­
placing even (sn (a)). Often, it is conceptually easier to think of ground(P ) as
a (countably) infinite propositional program and to study it rather than P .
28 Mathematical Aspects of Logic Programming Semantics

Indeed, many authors even define a logic program to be a set of propositional

clauses, with the advantage that notation can be considerably eased in some
places. For many of the example programs which we will discuss later, we will
also take advantage of this simpler notation, as in the following.

2.1.7 Program Let P be the following program.

p ← ¬q
q ← ¬p
Then BP = {p, q} and ground(P ) = P . Preinterpretations play no role in this
case.
We now come to the fundamental notions of interpretation and model
for programs. Interpretations and models as defined next are the particular
forms of Definitions 1.2.7 and 1.2.8 that we will use henceforth in studying
the semantics of programs.

2.1.8 Definition Let P be a program, let J be a preinterpretation for P

with domain D, and let T be a logic. An interpretation or valuation for P
(based on J) with values in T is an interpretation or valuation defined on
BP,J with values in T . An interpretation I for P is a model for P if I(C) = t
for each clause C ∈ groundJ (P ). As in Definition 1.2.8, we sometimes refer
to valuations, interpretations, and models for P based on J as J-valuations,
J-interpretations, and J-models, respectively.
We will in future use the notation IP,J,2 for the set of all two-valued in­
terpretations for P based on J. As usual, reference to the preinterpretation
J will often be omitted if it is fixed and understood. Similarly, the number
2 will be omitted if it is understood, and hence the set of all two-valued in­
terpretations for P based on a given, fixed preinterpretation J will often be
denoted by IP,2 or just by IP . Similar comments apply to the set IP,J,3 of
all three-valued interpretations for P based on J and to the set IP,J,4 of all
four-valued interpretations for P based on J.
The three sets just defined have the order-theoretic structure described in
Theorem 1.3.4 relative to the orders we discussed in Chapter 1. In particular,
IP,2 can be identified with the power set of BP .
With these structures in place, we are now ready to begin the main subject
of our study in this chapter, namely, the semantics of logic programs.

2.2 Supported Models

As already noted, a declarative semantics for logic programs is usually
given by selecting models for the programs which satisfy certain desirable
 The Semantics of Logic Programs 29

conditions. This selection is often most conveniently described by an operator,

mapping interpretations to interpretations, whose fixed points are exactly the
models being sought. In this section, we will introduce the first of a number of
operators we study in the context of declarative semantics, namely, the single-
step or immediate consequence operator due to Kowalski and van Emden,
see [van Emden and Kowalski, 1976]. The single-step operator was historically
the first to be studied in relation to logic programming semantics and in
many ways is the most natural. Indeed, it turns out that for definite programs
the single-step operator is order continuous and that its least fixed point, as
given by Kleene’s theorem, Theorem 1.1.9, accords well with a programmer’s
expectations of what a declarative semantics should be and how it should
relate to the procedural semantics.5
For the remainder of this section and for the next, we will work in classical
two-valued logic. Hence, IP or IP,J means IP,J,2 here and in the subsequent
section, where J is a given preinterpretation, and we will on occasions remind
the reader of this notational convenience.
The following is an important definition.

2.2.1 Definition Let P be a normal logic program, and let J be a preinter­

pretation for LP . The single-step operator or immediate consequence operator
TP,J : IP,J → IP,J is defined, for I ∈ IP,J , by setting TP,J (I) to be the set
of all A ∈ BP,J for which there is a clause A ← L1 , . . . , Ln in groundJ (P )
satisfying I |= L1 ∧ . . . ∧ Ln , that is, satisfying I(L1 ∧ . . . ∧ Ln ) = t.
Consistent with our earlier remarks concerning notation, we will usually
denote TP,J simply by TP when J is understood. Furthermore, we will some­
times find it convenient to refer to TP as the TP -operator .
The importance of the immediate consequence operator is clear from the
following proposition.

2.2.2 Proposition The models for P are exactly the pre-fixed points of TP .
Proof: Let I ∈ IP be a model for P , and let A ∈ TP (I). Then there is a
clause A ← L1 , . . . , Ln in ground(P ) with I(L1 ∧ . . . ∧ Ln ) = t; let us denote
this clause by C. Since I is a model for P , we have I(C) = t. Hence, I(A) = t,
and so A ∈ I, giving TP (I) ⊆ I, as required.
Conversely, suppose TP (I) ⊆ I, and let A ← L1 , . . . , Ln be a clause C in
ground(P ) with I(L1 ∧ . . . ∧ Ln ) = t. Then A ∈ TP (I) ⊆ I. Hence, I(A) = t,
and in consequence I(C) = t, as required. •

The notion of model is far too general to capture the declarative semantics
of logic programs without some restrictions being imposed upon it. Indeed, BP
itself is always a model for P , but in general BP fails by far to give a reason­
able “intended meaning” for a program. Standard approaches to declarative
5 As already noted, we will not consider procedural aspects in depth and instead refer

the reader to [Apt, 1997, Lloyd, 1987] for details of procedural semantics.
30 Mathematical Aspects of Logic Programming Semantics

semantics therefore involve the imposition of certain additional conditions

which models must satisfy in order to qualify as intended models. However,
just what conditions it is reasonable to choose in this context depends on one’s
particular understanding of what “intended” could mean, and the remainder
of this chapter will be devoted, in the main, to the presentation and study of
different conditions which have been proposed in the literature to solve this
problem.
The observation that BP is too large, as a two-valued model, suggests the
selection of minimal models. Of particular interest are the cases when there
exists a least model.

2.2.3 Theorem Let P be a definite program, and let J denote a fixed prein­
terpretation for LP . Then the following statements hold.

(a) TP is order continuous on IP .

(b) P has a least (J-)model, which coincides with the least fixed point of TP
and is equal to TP ↑ ω.

(c) The intersection of any non-empty collection of (J-)models for P is itself a

model for P . Therefore, a definite program cannot have two distinct min­
imal models. Furthermore, the intersection of the collection of all models
for P coincides with the least model for P .

Proof: (a) We first show that TP is monotonic. Let I, K ∈ IP with I ⊆ K,

and suppose A ∈ TP (I). Then there is a clause A ← body in ground(P ) with
body ⊆ I. Hence, body ⊆ K, and so A ∈ TP (K), as required.
Now let I = {Iλ |  λ ∈ Λ}be a directed family of two-valued inter­
pretations, and let I = I = I. Since the order under consideration is
set-inclusion and TP is monotonic, we immediately have that TP (I) is di­
rected. Bythe remarks following Definition 1.1.7, it remains to show that
TP (I) ⊆ TP (I). So suppose that A belongs to TP (I). Then there is a
(definite) clause C of the form A ← A1 , . . . , An in ground(P ) satisfying
A1 , . . . , An ∈ I. Therefore, there exist Iλ1 , . . . , Iλn in I with Ai ∈ Iλi for
i = 1, . . . , n. Since I is directed, there is Iλ ∈ I with Iλi ⊆ Iλ for i = 1, . . . , n.
Hence, the body of C  is true in Iλ , and we obtain that A ∈ TP (Iλ ) and,
consequently, that A ∈ TP (I), as required.
(b) By (a), we can apply Kleene’s theorem, Theorem 1.1.9, to see that
TP has a least pre-fixed point, that this least pre-fixed point is in fact the
least fixed point of TP , and that it coincides with TP ↑ ω. Hence, by Proposi­
tion 2.2.2, TP ↑ ω is the least model for P .
(c) The details of the proof of this claim are straightforward and therefore
are omitted. •

It can be shown, furthermore, that the least model for definite programs
corresponds rather well with the procedural behaviour of logic programming
 The Semantics of Logic Programs 31

systems based on resolution.6 Thus, in summary, the least model semantics is

very satisfactory for definite logic programs from all points of view.
Attempts to generalize Theorem 2.2.3 to normal programs, however, fail
in several ways, as we show next.

2.2.4 Program Let P be the normal logic program consisting of the following
clauses.

p ← ¬q
q ← ¬p
r ← ¬r

Then {p, r} and {q, r} are minimal, but incomparable, models so that P has
no least model, TP has no fixed points at all (and hence P has no supported
models, see Proposition 2.2.6), and, since TP (∅) = {p, q, r} and TP ({p, q, r}) =
∅, we see that TP is not monotonic.
It is not entirely clear how to cope with the negative results presented by
Program 2.2.4. Various different methods have been discussed in the literature,
leading to different declarative semantics with varying degrees of success. We
will discuss the more prominent of these approaches in the remainder of this
chapter.
A rather straightforward attack is to study minimal models instead of least
models. However, consider the program Even (Program 2.1.3) with models

K1 = even s2n (a) | n ∈ N

A b
and
A 2n+1 b
K2 = even s (a) | n ∈ N .

Both models are minimal, but it seems to be rather obvious that K1 captures
the intended meaning of Even, while K2 does not. Essentially, this arises from
the fact that even(s(a)) is true with respect to K2 , although the program
itself gives no justification for this. Thus, it would seem intuitively reasonable
that whenever an atom is true in an intended model for a program P , then
it should be true for a reason provided by the program itself. This idea is
captured by the following definition, see [Apt et al., 1988].

2.2.5 Definition An interpretation I for a program P is called supported if

for each A ∈ I there is a clause A ← body in ground(P ) with I(body) = t.
Continuing the Even program discussion above, note that K1 is supported,
whereas K2 is not. Indeed, K1 is the only supported model for Even, as we
will see later. So, for some programs, supportedness is an appropriate require­
ment of models. Supportedness is also captured by the immediate consequence
operator, as follows.
6 A detailed account of resolution-based logic programming can be found in [Apt, 1997,

Lloyd, 1987].
32 Mathematical Aspects of Logic Programming Semantics

2.2.6 Proposition The supported interpretations for a program P are ex­

actly the post-fixed points of TP . The supported models for P are exactly the
fixed points of TP .

Proof: Let I be a supported interpretation for P , and suppose that A ∈ I.

Then there is a clause A ← body in ground(P ) with I(body) = t. But then
A ∈ TP (I), showing that I ⊆ TP (I), as required to see that I is a post-fixed
point of TP .
Conversely, assume that I ⊆ TP (I) is a post-fixed point of TP , and let A ∈
I. Then A ∈ TP (I). Therefore, there exists a clause A ← body in ground(P )
with I(body) = t, showing that I is a supported model for P .
Finally, using Proposition 2.2.2, we obtain that an interpretation for P
is a supported model for P if and only if it is both a pre-fixed point and a
post-fixed point of TP , that is, if and only if it is a fixed point of TP . •

2.2.7 Example Tweety1 from Program 2.1.2 has supported model M , where
M = {penguin(tweety), bird(bob), bird(tweety), flies(bob)}, as is easily
verified. Careful inspection will also convince the reader that M is the unique
supported model for Tweety1, and we give a formal proof of this in Exam­
ple 5.1.7.

From a procedural point of view in the context of resolution-based logic

programming, supported models are better than minimal ones. They capture
the probable intention of a programmer who may think of a clause as a form
of equivalence7 rather than as an implication.
Since the least model for a definite program is a fixed point, by Theo­
rem 2.2.3, we obtain as a corollary that the least model is always supported.
In proving Theorem 2.2.3, we applied Kleene’s theorem. For normal programs,
this theorem is not applicable, nor is the Knaster-Tarski theorem, due to the
non-monotonicity of the immediate consequence operator in general. In order
to study the supported model semantics, that is, in order to obtain fixed points
of non-monotonic immediate consequence operators, it seems natural to em­
ploy fixed-point theorems for mappings which are not necessarily monotonic.
This is the main theme of Chapter 4.

2.3 Stable Models

One of the drawbacks of the supported model semantics is that definite
programs may have more than one supported model.
7 One formal approach to understanding clauses as equivalences is via the notion of the

Clark completion of a program and is related to SLDNF-resolution, see [Clark, 1978].

 The Semantics of Logic Programs 33

2.3.1 Program Let P be the program consisting of the single clause p ← p.

Then both ∅ and {p} are supported models for P .

This unsatisfactory situation is resolved by the introduction of stable mod­

els. Before we give the definition, let us make the following observation.

2.3.2 Proposition The least model TP ↑ ω for a definite program P is the

unique model M for P satisfying the following condition: there exists a map­
ping l : BP → α, for some ordinal α, such that for each A ∈ M there is a
clause A ← body in ground(P ) with M (body) = t and l(B) < l(A) for each
B ∈ body.

Proof: To start with, take M to be the least model TP ↑ ω, choose α = ω,

and define l : BP → α by setting l(A) = min{n | A ∈ TP ↑ (n + 1)}, if A ∈ M ,
and by setting
 l(A) = 0, if A ∈ M . Since ∅ ⊆ TP ↑ 1 ⊆ . . . ⊆ TP ↑ n ⊆ . . . ⊆
TP ↑ ω = m<ω
TP ↑ m, for each n, we see that l is well-defined and that the
least model TP ↑ ω for P has the desired properties.

Conversely, if M is a model for P which satisfies the given condition for

some mapping l : BP → α, then it is easy to show, by induction on l(A), that
A ∈ M implies A ∈ TP ↑ (l(A) + 1). This yields that M ⊆ TP ↑ ω and hence
that M = TP ↑ ω by minimality of the model TP ↑ ω. •

Mappings l from BP into an ordinal are commonly called level mappings.

They will play an important role in several places in the book. On occasions,
we will need to extend such mapping to literals, and unless stated to the
contrary, we will always assume that the extension satisfies l(¬A) = l(A) for
all atoms A.
The following definition of stable model merges the property of MP just
established with that of supportedness.8

2.3.3 Definition An interpretation I for a program P is called a well-

supported interpretation if there exists a level mapping l : BP → α, for some
ordinal α, with the property that, for each A ∈ I, there is a clause C in
ground(P ) of the form A ← A1 , . . . , An , ¬B1 , . . . , ¬Bk such that the body of
C is true in I and l(Ai ) < l(A) for i = 1, . . . , n. A well-supported model for P
is called a stable model for P .

2.3.4 Theorem The following statements hold.

(a) Every stable model is supported, but not vice-versa.

(b) Every stable model is a minimal model, but not vice-versa.

(c) Every definite program has a unique stable model, which is its least model.
8 It is shown in [Fages, 1994] that stable models can be introduced as in Definition 2.3.3.

The original formulation used the Gelfond–Lifschitz operator from Definition 2.3.6.
34 Mathematical Aspects of Logic Programming Semantics

Proof: (a) Supportedness of stable models follows immediately from the def­
inition. The supported model {p} for Program 2.3.1 is not stable.
(b) Let P be a program, let M be a stable model for P , and let l be
a level mapping with respect to which M is well-supported. Assume that
K is a model for P with K ⊂ M . Then there exists A ∈ M \ K, and
we can assume without loss of generality that A is also such that l(A) is
minimal. By the well-supportedness of M , there is a clause C of the form
A ← A1 , . . . , An , ¬B1 , . . . , ¬Bk in ground(P ) such that for i = 1, . . . , n and
j = 1, . . . , k we have Ai ∈ M , l(A) > l(Ai ) and Bj ∈ M . Since K ⊂ M , we
obtain, for j = 1, . . . k, that Bj ∈ K, and by minimality of l(A) we obtain
Ai ∈ K for i = 1, . . . , n. Since K is a model for P and the body of C is true
with respect to K, we conclude that A ∈ K, which contradicts the assumption
that A ∈ M \ K. Hence, M must be a minimal model.
In the opposite direction, Program 2.3.5 below has {p} as its only model,
and hence, this is a minimal model. It is clearly not a stable model, however.
(c) By Proposition 2.3.2, we see that the least model is indeed stable.
Uniqueness follows from (b) and Theorem 2.2.3 (c). •

There are programs with unique supported models which are not stable.

2.3.5 Program The program P consisting of the two clauses

p←p
p ← ¬p

has unique supported model {p}, and this model is not stable.
A unique stable model is always a least model by Theorem 2.3.4 (b). If a
program has a least model, however, this model is not guaranteed to be stable,
as Program 2.3.5 shows in having {p} as its only model.
A characterization of stable models as fixed points of an operator can be
given, and we proceed with this next.

2.3.6 Definition Let P be a normal logic program, and let I ∈ IP . The

Gelfond–Lifschitz transform P/I of P is the set of all clauses A ← A1 , . . . , An
for which there exists a clause A ← A1 , . . . , An , ¬B1 , . . . , ¬Bk in ground(P )
with B1 , . . . , Bk ∈ I.
We note that the Gelfond–Lifschitz transform P/I of a program P is always
definite (as a set of ground clauses) and therefore has a least model TP/I ↑ ω
by Theorem 2.2.3. The operator GLP : I �→ TP/I ↑ ω is called the Gelfond–
Lifschitz operator 9 associated with P .

2.3.7 Theorem The following hold.

9 The Gelfond–Lifschitz operator is named after the authors of the well-known paper

[Gelfond and Lifschitz, 1988] and was introduced by them in defining the stable model se­
mantics.
 The Semantics of Logic Programs 35

(a) The Gelfond–Lifschitz operator is antitonic and, in general, is not mono­

tonic.

(b) An interpretation I is a stable model for a program P if and only if it is

a fixed point of GLP , that is, if and only if it satisfies GLP (I) = I.

Proof: (a) Let P be a program, and let I, K be interpretations for P with

I ⊆ K. Then P/K ⊆ P/I, and it is a straightforward proof by induction to
show that TP/K ↑ n ⊆ TP/I ↑ n for all n ∈ N. Hence, GLP (K) = TP/K ↑ ω ⊆
TP/I ↑ ω = GLP (I), which shows that GLP is antitonic. To see that it is not
generally monotonic, take P to be Program 2.3.5. On setting I = ∅, we obtain
that P/I is the definite program consisting of the clauses p ← p and p ←, and
GLP (I) = {p}; on setting I = {p}, we obtain that P/I consists of the single
clause p ← p, and GLP (I) = ∅. This establishes (a).
For (b), we start by supposing that GLP (I) = TP/I ↑ ω = I. Then I is
the least model for P/I, and hence, is also a model for P , and, by Propo­
sition 2.3.2, is well-supported with respect to any level mapping l satisfying
l(A) = min{n | A ∈ TP/I ↑ (n + 1)} for each A ∈ I. Conversely, let I be
a stable model for P . Then I is well-supported relative to some level map­
ping l, say. Thus, for every A ∈ I, there is a clause C in ground(P ) of the
form A ← A1 , . . . , An , ¬B1 , . . . , ¬Bk such that the body of C is true in I and
l(Ai ) < l(A) for i = 1, . . . , n. But then, for every A ∈ I, there is a clause
A ← A1 , . . . , An in P/I whose body is true in I and such that l(Ai ) < l(A)
for i = 1, . . . , n. By Proposition 2.3.2, this means that I is the least model for
P/I, that is, I = TP/I ↑ ω = GLP (I). •

The Gelfond–Lifschitz transform can be considered as a two-step process:

first, delete each ground clause which has a negative literal ¬B in its body
with B ∈ I; second, delete all negative literals in the bodies of the remaining
clauses. Indeed, the intuition behind it is as follows. We can think of P as a
set of premises and of I as a set of beliefs that a rational agent might hold and
wants to test, given the premises P . Any ground clause that contains ¬B in its
body, where B ∈ I, is useless to the agent and can be discarded. Among the
remaining ground clauses, an occurrence of ¬B with B ∈ I is trivial. Thus,
we can simplify the premises to P/I. If I happens to be the set of atoms that
logically follow from P/I, then the agent is rational.
We will now give some examples.

2.3.8 Example Consider again Tweety1 from Program 2.1.2 and its sup­
ported model M as given in Example 2.2.7. We show that M is stable. The
36 Mathematical Aspects of Logic Programming Semantics

program Tweety1/M is as follows.

penguin(tweety) ←
bird(bob) ←
bird(tweety) ← penguin(tweety)
bird(bob) ← penguin(bob)
flies(bob) ← bird(bob)

The least model for this program turns out to be M , which shows that M is
stable.

A strange feature of the supported model semantics is that the addition

of clauses of the form p ← p may change the semantics.

2.3.9 Program (Tweety2) Consider the following program Tweety2.

penguin(tweety) ←
bird(bob) ←
bird(X) ← penguin(X)
flies(X) ← bird(X), ¬penguin(X)
penguin(bob) ← penguin(bob)

Tweety2 results from Tweety1 by adding the clause penguin(bob) ←

penguin(bob). Intuitively, this addition should not change the semantics of
the program. However, in addition to the supported model M from Example
2.2.7, Tweety2 also has

M ' = {penguin(tweety), penguin(bob), bird(tweety), bird(bob)}

as a supported model. While M is also a stable model for Tweety2, M ' is not.
This can be seen by inspecting the program Tweety2/M ' , as follows, which
has {penguin(tweety), bird(bob), bird(tweety)} = M ' as its least model.

penguin(tweety) ←
bird(bob) ←
bird(tweety) ← penguin(tweety)
bird(bob) ← penguin(bob)
penguin(bob) ← penguin(bob)

We can also use the stable model semantics for modelling choice.
 The Semantics of Logic Programs 37

2.3.10 Program (Tweety3) Consider the program Tweety3, as follows.

eagle(tweety) ← ¬penguin(tweety)
penguin(tweety) ← ¬eagle(tweety)
bird(X) ← eagle(X)
bird(X) ← penguin(X)
flies(X) ← bird(X), ¬penguin(X)

This program has the two stable models

{eagle(tweety), bird(tweety), flies(tweety)}
and
{penguin(tweety), bird(tweety)}.

2.4 Fitting Models

The stable model semantics is more satisfactory than the supported model
semantics in that each definite program has a unique stable model which
coincides with its least model. However, for normal logic programs in general,
uniqueness cannot be guaranteed, as can be seen from Program 2.1.7, which
has two stable models {p} and {q}. It is desirable to be able to associate with
each program a unique model in some natural way. One way of doing this is
by means of three-valued logic, and we discuss this next.10
In fact, we will work with Kleene’s strong three-valued logic as discussed
in Chapter 1 and, in particular, with the knowledge ordering, ≤k , on the truth
values. We find it convenient here to represent three-valued interpretations as
signed sets, see Section 1.3.3, so that the corresponding ordering [k is subset
inclusion of signed sets.
Given a normal logic program P , we define the following operators TP' and
FP on IP = IP,3 = IP,J,3 . First, TP' (I) is the set of all A ∈ BP for which
there is a clause A ← body in ground(P ) with body true in I with respect
to Kleene’s strong three-valued logic. Second, FP (I) is the set of all A ∈ BP
such that for all clauses A ← body in ground(P ) we have that body is false in
I with respect to Kleene’s strong three-valued logic. Finally, we define
ΦP (I) = TP' (I) ∪ ¬FP (I)
for all I ∈ IP . We will call the operator ΦP the Fitting operator for P or the
ΦP -operator .
10 The resulting Kripke-Kleene semantics, herein called the Fitting semantics, is due to

Fitting [Fitting, 1985].

38 Mathematical Aspects of Logic Programming Semantics

Notice that, for any three-valued interpretation I, we have A ∈ ΦP (I)

whenever A is the head of a ground clause and ¬A ∈ ΦP (I) whenever there
is no ground clause whose head is A.

2.4.1 Example We illustrate the calculation of ΦP (I), taking P to be the

program Tweety1 and starting with the three-valued interpretation I = ∅
thought of as a signed subset; of course, ∅ gives truth value u to all ground
atoms in our present context.
We have
TP' (∅) = {penguin(tweety), bird(bob)}
and
¬FP (∅) = ¬{penguin(bob)}.
Therefore,

ΦP (∅) = {penguin(tweety), bird(bob), ¬penguin(bob)}.

Continuing, we have

TP' (ΦP (∅)) = {penguin(tweety), bird(bob), bird(tweety), flies(bob)},

and
¬FP (ΦP (∅)) = ¬{penguin(bob), flies(tweety)}.
Thus, ΦP (ΦP (∅)) = TP' (ΦP (∅)) ∪ ¬FP (ΦP (∅)) is a total three-valued interpre-
tation. It follows from this fact and Proposition 2.4.4 below that ΦP (ΦP (∅))
is, in fact, the least fixed point of ΦP , as can readily be checked in any case
by iterating ΦP once more.

The development of the operator ΦP somewhat parallels that of TP except

that there are two orderings involved, and the following result is analogous to
Proposition 2.2.2.

2.4.2 Proposition Let P be a normal logic program. Then the three-valued

models for P are exactly the pre-fixed points of ΦP in the truth ordering [t .

Proof: Suppose that M is a three-valued interpretation for P satisfying

ΦP (M ) [t M , and let A ∈ BP be arbitrary. Suppose that ΦP (M )(A) = u.
Then we must have M (A) equal to u or to t. Since no clause A ← body
in ground(P ) can have M (body) = t, otherwise ΦP (M )(A) would be equal
to t, we must have M (body) equal to u or to f for each clause A ← body
in ground(P ). But then, on recalling the truth value given to ← in Defini­
tion 1.3.3, we see that A ← body is true in M . The other possible values for
ΦP (M )(A) are handled similarly, and so M is a model for P .
The converse is also handled similarly, and we omit the details. •
 The Semantics of Logic Programs 39

2.4.3 Program Consider the following program P .

p ← ¬q
p ← ¬r
q←q
r←r

Define M as follows: M (p) = f , M (q) = u, and M (r) = t. Then M is a

three-valued interpretation for P satisfying ΦP (M ) [k M , and yet M is not
a model for P .
On the other hand, take P to be Program 2.2.4. Define M as follows:
M (p) = t, M (q) = u, and M (r) = t. Then M is a three-valued model for P ,
but it does not satisfy the inequality ΦP (M ) [k M .
Therefore, neither implication of Proposition 2.4.2 holds in the case of the
knowledge ordering.

The following fact about ΦP is fundamental.

2.4.4 Proposition Let P be a program. Then ΦP is monotonic on IP,3 in

the knowledge ordering [k .
Proof: Let I, K ∈ IP,3 with I ⊆ K. We show ΦP (I) ⊆ ΦP (K). Let A ∈ ΦP (I)
be an atom. Then A ∈ TP' (I). Therefore, there is a ground clause A ← body
such that body is true in I. From Table 1.1, each literal in body must be true
and therefore, noting the results of Section 1.3.3, must belong to I. Hence,
each literal in body belongs to K since I ⊆ K and is therefore true in K.
Hence, body is also true in K, and we obtain that A ∈ TP' (K) ⊆ ΦP (K). Now
let ¬A ∈ ΦP (I) be a negated atom. Then A ∈ FP (I), and so, for all ground
clauses A ← body, we have that body is false in I. So, given such a clause,
from Table 1.1 we see that at least one literal Lj , say, in body, is false. Hence,
by the results of Section 1.3.3 again, we have ¬Lj ∈ I. But I ⊆ K and hence
¬Lj ∈ K. Therefore, Lj is also false in K, and consequently body is false in
K. Thus, we obtain A ∈ FP (K), and hence ¬A ∈ ΦP (K), as required. •

2.4.5 Example Take P to be Program 2.2.4 again. Define three-valued in­

terpretations I and K for P as follows: I(p) = I(q) = I(r) = f , and
K(p) = K(q) = K(r) = t. Then I ct K. Yet ΦP (K) is constant with value f ,
and ΦP (I) is constant with value t. Hence, Φ(K) ct Φ(I), and so ΦP is not
monotonic relative to the truth ordering.

Since the operator ΦP is monotonic relative to the ordering [k , it has

a least fixed point by the Knaster-Tarski theorem, Theorem 1.1.10, and this
least fixed point is an ordinal power ΦP ↑ α, as defined in Section 1.1, for
some ordinal α. The least fixed point of ΦP is called the Kripke-Kleene model
or Fitting model for P . It turns out, as we show later, that ΦP is not order
40 Mathematical Aspects of Logic Programming Semantics

continuous, indeed not even ω-continuous, relative to [k , and so Kleene’s

theorem, Theorem 1.1.9, is not generally applicable to ΦP .

2.4.6 Proposition Let P be a program. Then every fixed point M of ΦP is a

model for P with the following properties. (a) If A ∈ BP is such that M (A) =
t, then there exists a clause A ← body in ground(P ) with M (body) = t.
(b) If A ∈ BP is such that for all clauses A ← body in ground(P ) we have
M (body) = f , then M (A) = f .

Proof: Let A ← body be a clause in ground(P ). If M (body) = t, then M (A) =

ΦP (M )(A) = M (body) = t. If M (A) = f , then ΦP (M )(A) = M (A) = f , and
hence M (body) = f . Finally, if M (A) = u, then ΦP (M )(A) = M (A) = u,
and therefore M (body) = f or M (body) = u. By definition of the truth value
given to ←, we see that this suffices to show that M is a model for P .
In order to show (a), let A ∈ BP , and suppose that M (A) = t. Then
ΦP (M )(A) = M (A) = t, and there is a clause A ← body in ground(P ) with
M (body) = t by definition of ΦP .
To show (b), let A ∈ BP , and assume that for all clauses A ← body in
ground(P ) we have M (body) = f . Then M (A) = ΦP (M )(A) = f , again by
definition of ΦP . •

Proposition 2.4.6 shows that fixed points of ΦP are three-valued supported

models for P , meaning that they satisfy (a) and (b) of Proposition 2.4.6. Note
that a total three-valued supported model is a supported model in the sense
of Definition 2.2.5.

2.4.7 Proposition Let P be a program. Then the fixed points of ΦP are

exactly the three-valued supported models for P .

Proof: Certainly, every fixed point of ΦP is a three-valued supported model

for P by Proposition 2.4.6. Conversely, let M be a three-valued supported
model for P , and let A ∈ BP . If M (A) = t, then, by definition of a three-valued
supported model, there exists a clause A ← body in ground(P ) such that
M (body) = t, and hence ΦP (M )(A) = M (body) = t = M (A). If M (A) = f ,
then for all clauses A ← body in ground(P ) we have that M (body) = f , since
M is a model for P . Hence, ΦP (M )(A) = M (body) = f = M (A). It follows
that M is a fixed point of ΦP , as required. •

Before discussing further properties of the Fitting model, we give an alter­

native characterization of it.
For a program P and a three-valued interpretation I ∈ IP,3 , an I-partial
level mapping for P is a partial mapping l : BP → α with domain dom(l) =
{A | A ∈ I or ¬A ∈ I}, where α is some ordinal. Again, we extend every such
mapping to literals by setting l(¬A) = l(A) for all A ∈ dom(l).
 The Semantics of Logic Programs 41

2.4.8 Definition Let P be a normal logic program, let I be a three-valued

model for P , and let l be an I-partial level mapping for P . We say that P
satisfies (F) with respect to I and l if each A ∈ dom(l) satisfies one of the
following conditions.
(Fi) A ∈ I, and there is a clause A ← L1 , . . . , Ln in ground(P ) such that
Li ∈ I and l(A) > l(Li ) for i = 1, . . . , n.
(Fii) ¬A ∈ I, and for each clause A ← L1 , . . . , Ln in ground(P ) there exists
i ∈ {1, . . . , n} with ¬Li ∈ I and l(A) > l(Li ).
If A ∈ dom(l) satisfies (Fi), then we say that A satisfies (Fi) with respect to I
and l, with similar terminology if A ∈ dom(l) satisfies (Fii).

2.4.9 Theorem Let P be a normal logic program with Fitting model MP .

Then, in the knowledge ordering [k , MP is the greatest model among all
three-valued models I for which there exists an I-partial level mapping l for
P such that P satisfies (F) with respect to I and l.
Proof: We have MP = ΦP ↑ α for some ordinal α, and indeed α may be
taken to be the closure ordinal for MP . Define the MP -partial level mapping
lP : BP → α as follows: lP (A) = β, where β is the least ordinal such that
A is not undefined in ΦP ↑ (β + 1). The proof will be established by showing
the following facts. (1) P satisfies (F) with respect to MP and lP . (2) If I is
a three-valued model for P and l is an I-partial level mapping such that P
satisfies (F) with respect to I and l, then I ⊆ MP .
(1) Let A ∈ dom(lP ), and suppose that lP (A) = β. We consider the two
cases corresponding to (Fi) and (Fii).
Case (Fi). If A ∈ MP , then A ∈ TP' (ΦP ↑ β). Hence, there exists a clause
A ← body in ground(P ) such that body is true in ΦP ↑ β. Therefore, for all
Li ∈ body, we have that Li ∈ ΦP ↑ β, and hence lP (Li ) < β, and also that
Li ∈ MP for all i. Consequently, A satisfies (Fi) with respect to MP and lP .
Case (Fii). If ¬A ∈ MP , then A ∈ FP (ΦP ↑ β). Hence, for each clause
A ← body in ground(P ), there is a literal L ∈ body with ¬L ∈ ΦP ↑ β. But
then lP (L) < β and ¬L ∈ MP . Consequently, A satisfies (Fii) with respect to
MP and lP , and we have established that fact (1) holds.
(2) We show via transfinite induction on β = l(A) that, whenever A ∈ I,
or ¬A ∈ I, we have A ∈ ΦP ↑ (β + 1), or ¬A ∈ ΦP ↑ (β + 1)), respectively. For
the base case, note that if l(A) = 0, then A ∈ I implies that A occurs as the
head of a fact in ground(P ), hence A ∈ ΦP ↑ 1, and ¬A ∈ I implies that there
is no clause with head A in ground(P ), hence ¬A ∈ ΦP ↑ 1. So assume now
that the induction hypothesis holds for all B ∈ BP with l(B) < β and that
l(A) = β. We consider two cases.
Case i. If A ∈ I, then it satisfies (Fi) with respect to I and l. Hence, there
is a clause A ← body in ground(P ) such that body ⊆ I and l(K) < β for all
K ∈ body. Hence, body ⊆ MP by the induction hypothesis, and since MP is
a model for P , we obtain A ∈ MP .
42 Mathematical Aspects of Logic Programming Semantics

Case ii. If ¬A ∈ I, then A satisfies (Fii) with respect to I and l. Hence,

for each clause A ← body in ground(P ), there is K ∈ body with ¬K ∈ I and
l(K) < β. But then, by the induction hypothesis, we have ¬K ∈ MP , and
consequently for each clause A ← body in ground(P ) we obtain that body is
false in MP . Since MP = ΦP (MP ) is a fixed point of the ΦP -operator, we
obtain ¬A ∈ MP . This establishes Fact (2) and concludes the proof. •

The following corollary follows immediately as a special case of the previous

result.

2.4.10 Corollary A normal logic program P has a total Fitting model if and
only if there is a total model I for P and a (total) level mapping l for P such
that P satisfies (F) with respect to I and l.

2.4.11 Example Example 2.4.1 shows that Tweety1 (Program 2.1.2) has
total Fitting model M ∪ ¬(BTweety1 \ M ), where M is as in Example 2.2.7.
Tweety2 (Program 2.3.9) has Fitting model

{penguin(tweety), bird(bob), bird(tweety), ¬flies(tweety)}.

Thus, we cannot decide whether or not bob is a penguin. Hence, the Fitting
semantics suffers from the same deficiency as the supported model semantics,
see our discussion of Program 2.3.9.
Tweety3 (Program 2.3.10) has ∅ as its Fitting model.

The Fitting operator is not ω-continuous in general, not even for definite
programs, as shown by the next example.

2.4.12 Program Consider the program P consisting of the following clauses.

p(s(X)) ← p(X)
q ← p(X)

Then ΦP ↑ n = ¬p sk (0) | k < n for all n ∈ N and ΦP ↑ ω = {¬p(sn (0)) |

A b

n ∈ N}. However, ΦP ↑ (ω + 1) = {¬q, ¬p(sn (0)) | n ∈ N} is the least fixed

point of the operator.

The Fitting operator can be thought of as an approximation to the imme­

diate consequence operator, in the sense of the following proposition.

2.4.13 Proposition Let P be a program. Then for all I ∈ IP,3 , we have that
ΦP (I)+ ⊆ TP (I + ) ⊆ BP \ ΦP (I)− . Furthermore, the Fitting operator maps
total interpretations to total interpretations and coincides with the immediate
consequence operator on these.
 The Semantics of Logic Programs 43

Proof: Let I = I + ∪ ¬I − be a three-valued interpretation, and let A ∈

ΦP (I)+ . Then there is a clause A ← body in ground(P ), where body equals
A1 , . . . , An , ¬B1 , . . . , ¬Bk , say, and is true in the three-valued interpretation
I. Therefore, for all i and j, we have Ai ∈ I + and Bj ∈ I − so that Ai ∈ I +
and Bj ∈ I + . Therefore, body is true in the two-valued interpretation I + , and
so A ∈ TP (I + ). Conversely, if I is total, then Bj ∈ I + means that Bj ∈ I − ,
and hence whenever A ∈ TP (I + ) we have A ∈ ΦP (I)+ . This deals with the
first inclusion.
For the second inclusion, A ∈ ΦP (I)− if and only if for all clauses A ←
body in ground(P ) we have body false in the three-valued interpretation I.
But then one of the literals in body is false, and so, using the notation already
established for body, either some Ai ∈ I − or some Bj ∈ I + , that is, either
some Ai ∈ I + or some Bj ∈ I + . Therefore, body is also false in the two-
valued interpretation I + leading to A ∈ TP (I + ). We thus obtain ΦP (I)− ⊆
BP \TP (I + ) so that TP (I + ) ⊆ BP \ΦP (I)− . If I is total, then BP \ΦP (I)− =
ΦP (I)+ = TP' (I) = TP (I + ). •

From Proposition 2.4.13, we immediately obtain that total Fitting models

are always supported. They are, in fact, also stable in general, as we will see
later in Section 2.6. However, if a program has a unique stable model, it does
not necessarily have a total Fitting model.

2.4.14 Program The program consisting of the three clauses

p ← ¬q
q ← ¬p
p ← ¬p

has unique (two-valued) supported model {p}, which is also stable. However,
its (three-valued) Fitting model is everywhere equal to u.

2.5 Perfect Models

The approach using three-valued models, which was presented in Section
2.4, has the advantage that a unique model, namely, the least fixed point of
the Fitting operator, or the Fitting model, can be associated with each given
program. This avoids the ambiguity present in semantics based on classical
logic, such as the stable model semantics, where a program may have many
associated models.
An alternative way of avoiding this problem is to restrict syntax of pro­
grams in such a way that only programs are allowed whose semantics is unam­
biguous. The restriction is usually put in place by conditions which prevent
44 Mathematical Aspects of Logic Programming Semantics

recursion in certain situations, and the most convenient way of expressing

these conditions is again by the use of level mappings. For example, the alter­
native characterization of the Fitting model in Definition 2.4.8 and Theorem
2.4.9 can be viewed from this standpoint, and we will return to this point later
on in this section and in Section 2.6.
The approach which we present in this section is based on the following
idea: the introduction of negation, and in particular the possibility of allow­
ing recursive dependencies between negated atoms, causes ambiguity from a
declarative point of view. However, if recursion is only allowed through posi­
tive atoms, a standard model, namely, the least model, can be obtained. So it
seems natural to disallow recursion through negative dependencies, while at
the same time allowing recursion through positive ones. This idea is captured
in the following definition.

2.5.1 Definition A program P is called locally stratified 11 if there exists a

level mapping l : BP → α such that for each clause

A ← A1 , . . . , An , ¬B1 , . . . , ¬Bm

in ground(P ) the following hold.

(S1) l(A) ≥ l(Ai ) for i = 1, . . . , n.
(S2) l(A) > l(Bj ) for j = 1, . . . , m.
Furthermore, P is called stratified if it is locally stratified, and for all atoms
A, B ∈ BP with the same predicate symbol, we have l(A) = l(B).
Note that for stratified programs the image of the level mapping involved
is finite, in contrast to locally stratified programs. Stratified programs are
particularly interesting from the procedural point of view. Nevertheless, we
will concentrate here on the more general locally stratified programs.
Along with the introduction of locally stratified programs, a semantics was
developed called the perfect model semantics. We will discuss this semantics
only in passing in this chapter. Indeed, we will focus here on the more general
weakly perfect semantics, which is introduced later in this section and is also
defined for locally stratified programs. However, we will consider the perfect
model semantics in some detail in Section 6.3.

2.5.2 Definition Let P be a locally stratified program, and let l denote the
associated level mapping. Given two distinct models M and N for P , we say
that N is preferable to M if, for every ground atom A in N \ M , there is a
ground atom B in M \ N such that l(A) > l(B). A model M for P is called
perfect if there are no models for P preferable to M .
11 The notion of local stratification and the perfect model semantics were introduced in

the paper [Przymusinski, 1988]. Stratified programs and certain procedural apects of them
were studied in [Apt et al., 1988].
 The Semantics of Logic Programs 45

2.5.3 Example Tweety2 (Program 2.3.9) is locally stratified, indeed strat­

ified, since flies depends both on penguin and on bird, where “depends
on” is defined below, bird depends only on penguin, and penguin does not
depend on any predicate symbol other than itself. We will see in Example
6.3.12 that it has M from Example 2.2.7 as its perfect model.
Tweety3 (Program 2.3.10) is obviously not locally stratified.

We will see later in Section 6.3 that every locally stratified program has
a unique perfect model and that this model is independent of the choice of
the level mapping with respect to which the program is locally stratified. In
fact, we are more interested here in a generalization of the perfect model
semantics to three-valued logic, and of course the objective underlying this
generalization is the usual one, namely, to provide a single intended model for
each given program.
We will proceed next with presenting the rather involved definition of the
weakly perfect model due to Przymusinska and Przymusinski.12 For ease of
notation, it will be convenient to consider (countably infinite) propositional
programs instead of programs over a first-order language, and we recall that
we have already observed in Section 2.1 that this results in no loss of generality
for our purposes.
Let P be a (countably infinite propositional) normal logic program. We
say that an atom A ∈ BP refers to an atom B ∈ BP if either B or ¬B occurs
as a body literal in a clause A ← body in P with head A. We say that A refers
negatively to B if ¬B occurs as a body literal in such a clause. We say that A
depends on B, written B ≤ A, if the pair (A, B) is in the transitive closure of
the relation refers to. We say that A depends negatively on B, written B < A,
if there are C, D ∈ BP such that C refers negatively to D and the following
conditions hold: (1) C ≤ A or C = A (the latter meaning identity), and (2)
B ≤ D or B = D. For A, B ∈ BP , we write A ∼ B if either A = B or A and B
depend negatively on each other, so that A < B and B < A both hold in this
latter case.13 The relation ∼ is an equivalence relation, and its equivalence
classes are called components of P . A component is trivial if it consists of a
single element A with A < A.
Notice that the definitions above can be viewed in a rather intuitive way
by means of the dependency graph GP of a program P , defined as follows. The
vertices of GP are the ground atoms appearing in P ; for each clause A ← body
in ground(P ) there is a positive directed edge in GP from B to A if B occurs
in body, and there is a negative directed edge from B to A in GP if ¬B occurs
in body. Then, in these terms, we have B ≤ A if and only if there is a directed
path in GP from B to A, and we have B < A if and only if there is a directed
path in GP from B to A passing through a negative edge.
12 The notions of weak stratification and the weakly perfect model were introduced in the

paper [Przymusinska and Przymusinski, 1990].

13 It is noted in [Przymusinska and Przymusinski, 1990] that such mutual recursion is the

primary cause of difficulties in defining declarative semantics for logic programs.

46 Mathematical Aspects of Logic Programming Semantics

Let C1 and C2 be two components of a program P . We write C1 � C2 if

and only if C1 = C2 and for each A1 ∈ C1 there is A2 ∈ C2 with A1 < A2 . A
component C1 is called minimal if there is no component C2 with C2 � C1 .
Given a normal logic program P , the bottom stratum S(P ) of P is the union
of all minimal components of P . The bottom layer of P is the subprogram L(P )
of P which consists of all clauses from P with heads belonging to S(P ).
Given a three-valued interpretation I for P , thought of as a signed subset,
we define the reduct of P with respect to I to be the program P/I obtained
from P by performing the following reductions. (1) Remove from P all clauses
which contain a body literal L such that ¬L ∈ I or whose head belongs to
I. (2) Remove from all remaining clauses all body literals L with L ∈ I. (3)
Remove from the resulting program all non-unit clauses whose heads appear
also as heads of unit clauses in the program.
Note that the definition of P/I used here differs from that given in Defi­
nition 2.3.6 in the context of stable models. The new definition just given will
only be used in the present section.

2.5.4 Definition The weakly perfect model MP for a program P is defined by

transfinite induction as follows. Let P0 = P , and let M0 = ∅. For each (count­
able) ordinal α > 0 such that programs Pδ and three-valued interpretations
Mδ have already been defined for all δ < α, let

Nα = Mδ ,
δ<α
Pα = P/Nα ,

Rα is the set of all atoms which are undefined in Nα and were eliminated from
P by reducing it with respect to Nα ,

Sα = S (Pα ) , and
Lα = L (Pα ) .

The construction then proceeds with one of the following three cases. (1) If
Pα is empty, then the construction stops, and MP = Nα ∪ ¬Rα is the (total )
weakly perfect model for P . (2) If the bottom stratum Sα is empty or if the
bottom layer Lα contains a negative literal, then the construction also stops,
and MP = Nα ∪ ¬Rα is the (partial ) weakly perfect model for P . (3) In the
remaining case, Lα is a definite program, and we define Mα = H ∪¬Rα , where
H is the total three-valued model corresponding to the least two-valued model
for Lα , and the construction continues.
For every α, the set Sα ∪ Rα is called the α-th stratum of P , and the
program Lα is called the α-th layer of P .

We now present a detailed example of the calculation of the weakly perfect

model; see also Program 2.6.12 for further discussion of this example.
 The Semantics of Logic Programs 47

2.5.5 Example Consider the program Tweety4, as follows; it is a modifica­

tion of Tweety2 (Program 2.3.9), where the last clause has been changed.

penguin(tweety) ←
bird(bob) ←
bird(X) ← penguin(X)
flies(X) ← bird(X), ¬penguin(X)
penguin(bob) ← penguin(bob), ¬flies(bob)

This program has the weakly perfect model

{bird(bob), bird(tweety), penguin(tweety), ¬flies(tweety)},

and we show here how this model is calculated. We begin by setting P = P0 =

ground(Tweety4), as follows.

penguin(tweety) ←
bird(bob) ←
bird(tweety) ← penguin(tweety)
bird(bob) ← penguin(bob)
flies(tweety) ← bird(tweety), ¬penguin(tweety)
flies(bob) ← bird(bob), ¬penguin(bob)
penguin(bob) ← penguin(bob), ¬flies(bob)

Next, we set M0 = ∅ and carry out reduction of P0 with respect to M0

to obtain P1 = P0 /M0 , which turns out to be equal to P0 with the fourth
clause removed. The dependency graph GP1 of P1 is shown in Figure 2.1,
where we use the obvious abbreviations for the ground atoms in P1 such
as p(t) for penguin(tweety) and so on. Using GP1 , it is simple to check that
the components of P1 are {bird(bob)}, {bird(tweety)}, {penguin(tweety)},
{flies(tweety)}, and {flies(bob), penguin(bob)} and that the minimal
components are the first three of these. Therefore, the bottom stratum
S1 = S(P1 ) of P1 is {penguin(tweety), bird(bob), bird(tweety)}. Hence,
the bottom layer L1 = L(P1 ) of P1 is the definite program

penguin(tweety) ←
bird(bob) ←
bird(tweety) ← penguin(tweety)

whose
 least two-valued model is clearly equal to S1 . Note that N1 =
δ<1 M δ = M0 = ∅. Reduction of P0 with respect to M0 removed one clause,
but did not eliminate any atoms from P ; hence, R1 = ∅. Since L1 is definite,
we put M1 = H ∪¬R1 , where H is the total three-valued model corresponding
to S1 ; thus, M1 = S1 , and the process continues.
48 Mathematical Aspects of Logic Programming Semantics

p(t) p(b) +

+
1
− b(t) − b(b) −

+ +
1 1
f(t) f(b)

FIGURE 2.1: Dependency graph for P1 .

p(b) +

− −

f(b)

FIGURE 2.2: Dependency graph for P2 .

The program P2 = P1 /M1 is

flies(bob) ← ¬penguin(bob)
penguin(bob) ← penguin(bob), ¬flies(bob)

The dependency graph GP2 of P2 , shown in Figure 2.2, has only one com­
ponent {penguin(bob), flies(bob)}, which is therefore equal to the bot­
tom stratum S2 = S(P2 ) of P2 . Furthermore, N2 = M0 ∪ M1 = M1 and
R2 = {flies(tweety)}. Since the bottom layer L2 = L(P2 ) is equal to P2 ,
it is not definite. Therefore, the construction stops, and the weakly perfect
model is N2 ∪ ¬R2 = M1 ∪ ¬R2 , as claimed.

2.5.6 Proposition Let P be a program, and let M be its (partial) weakly

perfect model. Then M is a model with respect to Kleene’s strong three-valued
logic.

Proof: It is straightforward to show that ΦP (M ) = M , and we leave the

details to the reader. •

A weakly stratified program is a program with a total weakly perfect model.

The set of all its strata is then called its weak stratification.

2.5.7 Remark We remark that our definition of weakly perfect model, as

 The Semantics of Logic Programs 49

given in Definition 2.5.4, differs slightly from the version introduced in

[Przymusinska and Przymusinski, 1990]. In order to obtain the original def­
inition, points (2) and (3) of Definition 2.5.4 have to be replaced with the fol­
'
lowing: (2) If the bottom stratum Sα is empty or if the bottom layer Lα has
no least two-valued model, then the construction stops, and MP = Nα ∪ ¬Rα
'
is the (partial) weakly perfect model for P . (3) In the remaining case, Lα
has a least two-valued model, and we define Mα = H ∪ ¬Rα , where H is
the three-valued model for Lα corresponding to its least two-valued model,
and the construction continues. The original definition is more general due to
the fact that every definite program has a least two-valued model. However,
while the least two-valued model for a definite program can be obtained as
the least fixed point of the monotonic (and even continuous) operator TP ,
we know of no similar result, nor of a general operator, for obtaining the
least two-valued model, if it exists, for programs which are not definite. The
original definition therefore seems to be rather awkward, and indeed, even in
[Przymusinska and Przymusinski, 1990], when defining weakly stratified pro­
grams, the more general version was dropped in favour of requiring definite
layers. So Definition 2.5.4 is an adaptation taking the original notion of weakly
stratified program into account and appears to be more natural. Our use,
therefore, of the term weakly perfect model will refer to Definition 2.5.4 unless
stated to the contrary.

Again, an alternative characterization of the weakly perfect model can be

provided using level mappings.

2.5.8 Definition Let P be a normal logic program, let I be a three-valued

model for P , and let l be an I-partial level mapping for P . We say that P
satisfies (WS) with respect to I and l if each A ∈ dom(l) satisfies one of the
following conditions.

(WSi) A ∈ I, and there is a clause A ← L1 , . . . , Ln in ground(P ) such that

Li ∈ I and l(A) > l(Li ) for all i.

(WSii) ¬A ∈ I, and for each clause A ← A1 , . . . , An , ¬B1 , . . . , ¬Bm in

ground(P ) one (at least) of the following conditions holds.

(WSiia) There exists i with ¬Ai ∈ I and l(A) > l(Ai ).

(WSiib) For all k we have l(A) ≥ l(Ak ), for all j we have l(A) >
l(Bj ), and there exists i with ¬Ai ∈ I.
(WSiic) There exists j with Bj ∈ I and l(A) > l(Bj ).

Noting that the condition (Fii) in Definition 2.4.8 implies that either
(WSiia) or (WSiic) holds, we see that the condition (WSii) above is more
general than (Fii); conditions (WSi) and (Fi) are identical.

2.5.9 Theorem Let P be a normal logic program with weakly perfect model
50 Mathematical Aspects of Logic Programming Semantics

MP . Then, in the knowledge ordering [k , MP is the greatest model among all

models I for which there exists an I-partial level mapping l for P such that
P satisfies (WS) with respect to I and l.

We prepare for the proof of Theorem 2.5.9 by introducing some notation

which will help make the presentation transparent.
It will be convenient to consider level mappings which map into pairs (β, n)
of ordinals, where n ≤ ω. So let α be a (countable) ordinal, and consider the
set A of all pairs (β, n), where β < α and n ≤ ω. Of course, A endowed with
the lexicographic ordering is isomorphic to an ordinal. So any mapping from
BP to A can be considered to be a level mapping.
Let P be a program with (partial) weakly perfect model MP . We define
the MP -partial level mapping lP as follows: lP (A) = (β, n), where A ∈ Sβ ∪Rβ
and n is least with A ∈ TLβ ↑ (n + 1), if such an n exists, and n = ω otherwise.
We observe that if lP (A) = lP (B), then there exists α with A, B ∈ Sα ∪ Rα ,
and if A ∈ Sα ∪ Rα and B ∈ Sβ ∪ Rβ with α < β, then l(A) < l(B).
The following notion will help to ease later notation.

2.5.10 Definition Let P and Q be two programs, and let I be an interpre­

tation.

(1) Suppose that C1 = (A ← L1 , . . . , Ln ) and C2 = (B ← K1 , . . . , Km ) are

two clauses. Then we say that C1 subsumes C2 , written C1 � C2 , if A = B
and {L1 , . . . , Ln } ⊆ {K1 , . . . , Km }.

(2) We say that P subsumes Q, written P � Q, if for each clause C1 in P

there exists a clause C2 in Q with C1 � C2 .

(3) We say that P subsumes Q model-consistently (with respect to I), written

P �I Q, if the following conditions hold.

(i) For each clause C1 = (A ← L1 , . . . , Ln ) in P , there exists a clause

C2 = (B ← K1 , . . . , Km ) in Q with C1 � C2 and {K1 , . . . , Km } \
{L1 , . . . , Ln } ⊆ I.
(ii) For each clause C2 = (B ← K1 , . . . , Km ) in Q which satisfies
{K1 , . . . , Km } ⊆ I and B ∈ I, there exists a clause C1 in P such
that C1 � C2 .

Definition 2.5.10 will facilitate the proof of Theorem 2.5.9 by employing

the following lemma.

2.5.11 Lemma With the notation established in Definition 2.5.4, we have

P/Nα �Nα P for all α.

Proof: Condition 3(i) of Definition 2.5.10 holds because every clause C1 =

(A ← L1 , . . . , Ln ) in P/Nα is obtained from a clause C2 = (A ← K1 , . . . , Km )
in P by deleting body literals which are contained in Nα . Clearly, C1 � C2 ,
 The Semantics of Logic Programs 51

and the set difference {K1 , . . . , Km } \ {L1 , . . . , Ln } contains only elements of

Nα . Condition 3(ii) holds because for each clause C2 = (A ← K1 , . . . , Km ) in
P with head A ∈ Nα whose body is true under Nα , Step 2 in the reduction
of P with respect to Nα removes all the body literals Ki . Therefore, we have
that C1 = (A ←) is a fact in P/Nα , and clearly, C1 � C2 . •

The next lemma establishes the induction step in Part (2) of the proof of
Theorem 2.5.9.

2.5.12 Lemma If I is a non-empty three-valued model for a (infinite propo­

sitional normal) logic program P ' and l is an I-partial level mapping such
that P ' satisfies (WS) with respect to I and l, then the following hold for
P = P ' /∅.

(a) The bottom stratum S(P ) of P is non-empty and consists of trivial com­
ponents only.

(b) The bottom layer L(P ) of P is definite.

(c) The three-valued model N corresponding to the least two-valued model

for L(P ) is consistent with I in the following sense: we have I ' ⊆ N , where
I ' is the restriction of I to all atoms which are not undefined in N .

(d) P/N satisfies (WS) with respect to I \ N and l|N , where l|N is the restric­
tion of l to the atoms in I \ N .

Proof: (a) Assume that there exists some component C ⊆ S(P ) which is
not trivial. Then there must exist atoms A, B ∈ C with A < B, B < A,
and A = B. Without loss of generality, we can assume that A is chosen such
that l(A) is minimal. Now let A' be any atom occurring in the body of a
clause with head A. If A' occurs positively, then A > B > A ≥ A' , and so
A > A' ; if A' occurs negatively, then A > A' also. Therefore, by minimality
of the component, we must also have A' > A. Thus, we obtain that all atoms
occurring positively or negatively in the bodies of clauses with head A must
be contained in C. We consider two cases.
Case i. If A ∈ I, then there must be a fact A ← in P ; otherwise, by (WSi)
we have a clause A ← L1 , . . . , Ln (for some n ≥ 1) with L1 , . . . , Ln ∈ I and
l(A) > l(Li ) for all i, contradicting the minimality of l(A). Since P = P ' /∅,
we obtain that A ← is the only clause in P with head A, contradicting the
existence of B = A with B < A.
Case ii. If ¬A ∈ I, then since A was chosen to be minimal with re­
spect to l, we obtain that condition (WSiib) must hold for each clause
A ← A1 , . . . , An , ¬B1 , . . . , ¬Bm with respect to I and l and that m = 0.
Furthermore, all Ai must be contained in C, as already noted above, and
l(A) ≥ l(Ai ) for all i by (WSiib). Also, from Case i, we obtain that no Ai can
be contained in I. We have now established that, for all Ai in the body of any
clause with head A, we have l(A) = l(Ai ) and ¬Ai ∈ I. The same argument
52 Mathematical Aspects of Logic Programming Semantics

holds for all clauses with head Ai , for all i, and the argument repeats itself.
Now, from A > B, we obtain D, E ∈ C with A ≥ E (or A = E), D ≥ B (or
D = B), and E refers negatively to D. As we have just seen, we obtain ¬E ∈ I
and l(E) = l(A). Since E refers negatively to D, there is a clause containing E
in its head and ¬D in its body. Since (WSii) holds for this clause, there must
be a literal L in its body with level less than l(E), so that l(L) < l(A) and
L ∈ C, which is a contradiction. We thus have established that all components
are trivial.
We show next that the bottom stratum is non-empty. Indeed, let A be
an atom such that l(A) is minimal. We will show that {A} is a component.
Assume that this is not the case, that is, assume that there is B with B < A.
Then there exist D1 , . . . , Dk , for some k ∈ N, such that D1 = A, Dj refers
to Dj+1 for all j = 1, . . . , k − 1, and Dk refers negatively to some B ' with
B ' ≥ B (or B ' = B).
We show by induction that, for all j = 1, . . . , k, the following statements
hold: ¬Dj ∈ I, B < Dj , and l(Dj ) = l(A). Indeed, note that for j = 1, that is,
when Dj = A, we have that B < Dj = A and l(Dj ) = l(A). Assuming A ∈ I,
we obtain, by minimality of l(A), that A ← is the only clause in P = P ' /∅ with
head A, contradicting the existence of B < A. So, ¬A ∈ I, and the assertion
holds for j = 1. Now assume that the assertion holds for some j < k. Then
obviously Dj+1 > B since A ≥ D2 ≥ . . . ≥ Dk−1 ≥ Dk > B ' ≥ B. Since
¬Dj ∈ I and l(Dj ) = l(A), we obtain that (WSii) must hold, and, by the
minimality of l(A), we infer that (WSiib) must hold and that no clause with
head Dj contains negated atoms. So, l(Dj+1 ) = l(Dj ) = l(A) holds by (WSiib)
and the minimality of l(A). Furthermore, the assumption Dj+1 ∈ I can be
rejected by the same argument as for A above; otherwise, Dj+1 ← would be
the only clause with head Dj+1 by minimality of l(Dj+1 ) = l(A), contradicting
B < Dj+1 . This concludes the inductive proof.
Summarizing, we obtain that Dk refers negatively to B ' and that ¬Dk ∈ I.
But then there is a clause satisfying (WSii) with head Dk and ¬B ' in its body,
and this contradicts the minimality of l(Dk ) = l(A). This concludes the proof
of statement (a).
(b) Assume that L(P ) is not definite. Then there exists a clause A ← body
in L(P ) with a negated literal ¬B occurring in body. But then B < A, and
since the bottom stratum consists of minimal components only, we also have
A < B, that is, A and B are in the same component, contradicting (a).
(c) First, note that in forming the reduct P of P ' with respect to ∅, the
third step is the only one in the process which has any effect in that it removes
all non-unit clauses whose heads appear also as heads of unit clauses. Now
let A ∈ I ' be an atom with A ∈ N , and assume without loss of generality
that A is chosen such that l(A) is minimal with these properties. By the first
observation and the hypothesis that P ' satisfies (WS) with respect to I and
l, there must be a clause A ← L1 , . . . , Ln in P such that, for all i, Li is true
with respect to I, and hence true with respect to I ' , and l(A) > l(Li ). Hence,
all the literals Li are true with respect to N by minimality of l(A). Thus,
 The Semantics of Logic Programs 53

L1 , . . . , Ln is true in N , and, since N is a model for L(P ), we obtain A ∈ N ,

which contradicts our assumption.
Now let A ∈ N be an atom with A ∈ I ' , and assume without loss of
generality that A is chosen such that n is minimal with A ∈ TL(P ) ↑ (n + 1).
Then there is a definite clause A ← body in L(P ) such that all atoms in
body are true with respect to TL(P ) ↑ n. Hence, these atoms are also true with
respect to I ' , and, since I ' is a model for L(P ), we obtain A ∈ I ' , which
contradicts our assumption.
Finally, let ¬A ∈ I ' . Then we cannot have A ∈ N ; otherwise, A ∈ I ' . So,
¬A ∈ N since N is a total model for L(P ).
(d) From Lemma 2.5.11, we know that P/N �N P . We distinguish two
cases.
Case i. If A ∈ I \ N , then there must be a clause A ← L1 , . . . , Lk in P such
that Li ∈ I and l(A) > l(Li ) for all i. Since it is not possible for A to belong
to N , there must also be a clause in P/N which subsumes A ← L1 , . . . , Lk
and which therefore satisfies (WSi). So, A satisfies (WSi).
Case ii. If ¬A ∈ I \ N , then, for each clause A ← body1 in P/N , there
must be a clause A ← body in P which is subsumed by A ← body1, and, since
¬A ∈ I, we obtain that condition (WSii) must be satisfied by A and also by
the clause A ← body. Since reduction with respect to N removes only body
literals which are true in N , condition (WSii) is still fulfilled. •

We can now proceed with the proof of Theorem 2.5.9.

Proof of Theorem 2.5.9: The proof will proceed by establishing the fol­
lowing facts: (1) P satisfies (WS) with respect to MP and lP . (2) If I is a
model for P and l is an I-partial level mapping such that P satisfies (WS)
with respect to I and l, then I ⊆ MP .
(1) Let A ∈ dom(lP ), and suppose that lP (A) = (α, n). We consider two
cases.
Case i. A ∈ MP . Then A ∈ TLα ↑ (n + 1). Hence, there is a definite clause
A ← A1 , . . . , Ak in Lα with A1 , . . . , Ak ∈ TLα ↑ n. Thus, A1 , . . . , Ak ∈ MP and
lP (A) > lP (Ai ) for all i. By Lemma 2.5.11, P/Nα �Nα P . So there must be a
clause A ← A1 , . . . , Ak , L1 , . . . , Lm in P with literals L1 , . . . , Lm ∈ Nα ⊆ MP ,
and we obtain lP (Lj ) < lP (A) for all j = 1, . . . , m. So, (WSi) holds in this
case.
Case ii. ¬A ∈ MP . Let A ← A1 , . . . , Ak , ¬B1 , . . . , ¬Bm be a clause in
P , noting that (WSii) is trivially satisfied in case no such clause exists. We
consider the following two subcases.
Subcase ii.a. Assume A is undefined in Nα and was eliminated from P by
reducing it with respect to Nα , that is, A ∈ Rα . Then, in particular, there
must be some ¬Ai ∈ Nα , or some Bj ∈ Nα , which yields lP (Ai ) < lP (A), or
lP (Bj ) < lP (A), respectively, and hence one of (WSiia), (WSiic) holds.
Subcase ii.b. Assume ¬A ∈ H, where H is the three-valued model cor­
responding to the least two-valued model for Lα . Since P/Nα subsumes P
54 Mathematical Aspects of Logic Programming Semantics

model consistently with respect to Nα , we obtain that there must be some Ai

with ¬Ai ∈ H, and, by definition of lP , we obtain lP (A) = lP (Ai ) = (α, ω)
and, hence, also lP (Ai/ ) ≤ lP (Ai ) for all i' = i. Furthermore, since P/Nα is
definite, we obtain that ¬Bj ∈ Nα for all j, and hence lP (Bj ) < lP (A) for all
j. So, condition (WSiib) is satisfied.
(2) Suppose that I is a non-empty three-valued model for P and that l
is an I-partial level mapping such that P satisfies (WS) with respect to I
and l. First, note that for all models M , N of P with M ⊆ N , we have
(P/M )/N = P/(M ∪ N ) = P/N and (P/N )/∅ = P/N .
Let Iα denote I restricted to the atoms which are not undefined in Nα ∪Rα .
It suffices to show the following: for all α > 0, we have Iα ⊆ Nα ∪ Rα , and
I \ MP = ∅.
We show next by induction that if α > 0 is an ordinal, then the following
statements hold. (a) The bottom stratum of P/Nα is non-empty and consists
of trivial components only. (b) The bottom layer of P/Nα is definite. (c)
Iα ⊆ Nα ∪ Rα . (d) P/Nα+1 satisfies (WS) with respect to I \ Nα+1 and
l|Nα+1 .
Note that P satisfies the hypothesis of Lemma 2.5.12 and, hence, also its
conclusions. So, on taking α = 1, we have that P/N1 = P/∅ satisfies (WS)
with respect to I \N1 and l|N1 , and by application of Lemma 2.5.12, we obtain
that statements (a) and (b) hold. For (c), note that no atom in R1 can be true
in I, because no atom in R1 can appear as head of a clause in P , and now apply
Lemma 2.5.12 (c). For (d), apply Lemma 2.5.12, noting that P/N2 �N2 P .
For α a limit ordinal, we can show, exactly as in the proof of Lemma 2.5.12
(d), that P satisfies (WS) with respect to I \ Nα and l|Nα . So, Lemma 2.5.12
is applicable, and statements (a) and (b) follow. For (c), let A ∈ Rα . Then
every clause in P with head A contains a body literal which is false in Nα . By
the induction hypothesis, this implies that no clause with head A in P can
have a body which is true in I. So, A ∈ I. Together with Lemma 2.5.12 (c),
this proves statement (c). For (d), apply again Lemma 2.5.12 (d), noting that
P/Nα+1 �Nα+1 P .
For α = β + 1 a successor ordinal, we obtain by the induction hypothesis
that P/Nβ satisfies the hypothesis of Lemma 2.5.12. So, again statements (a)
and (b) follow immediately from this lemma, and (c) and (d) follow as in the
case when α is a limit ordinal.
It remains to show that I \ MP = ∅. Indeed, by the transfinite induction
argument just given, we obtain that P/MP satisfies (WS) with respect to
I \ MP and l|MP . If I \ MP is non-empty, then by Lemma 2.5.12 the bottom
stratum S(P/MP ) is non-empty, and the bottom layer L(P/MP ) is definite
and has model M corresponding to the least two-valued model for L(P/MP ).
Hence, by definition of the weakly perfect model MP for P , we must have that
M ⊆ MP , which contradicts the fact that M is the least model for L(P/MP ).
Hence, I \ MP must be empty, and this concludes the proof. •

The following corollary follows immediately as a special case.

 The Semantics of Logic Programs 55

2.5.13 Corollary A normal logic program P is weakly stratified, that is, has
a total weakly perfect model if and only if there is a total model I for P and
a (total) level mapping l for P such that P satisfies (WS) with respect to I
and l.

The weakly perfect model is in general different from the Fitting model.

2.5.14 Proposition Let P be a program, let M1 be its Fitting model, and

let M2 be its (partial) weakly perfect model. Then M1 ⊆ M2 .

Proof: Let l1 be an M1 -partial level mapping such that P satisfies (F) with
respect to M1 and l1 . Then, trivially, P satisfies (WS) with respect to M1 and
l1 . Since M2 is the largest model among all models I for which there exists
an I-partial level mapping l for P such that P satisfies (WS) with respect to
I and l, by Theorem 2.5.9, we have that M1 ⊆ M2 . •

The Fitting model does not in general coincide with the (partial) weakly
perfect model, nor does it coincide in general with the perfect model for locally
stratified programs.

2.5.15 Program Let P be the program consisting of the single clause p ← p.

Then the Fitting model for P is ∅, but the (partial) weakly perfect model for
P is {¬p}. Note that P is locally stratified with perfect (two-valued) model
in which p is false.

We will see later in Section 6.3 that if P is a locally stratified program,

then P is weakly stratified, and its (total) weakly perfect model is also its
perfect model. So, the weakly perfect model semantics unifies two separate
approaches. On the one hand, it is a generalization of the Fitting semantics
and allows one to assign a single intended model to each program; on the
other hand, it generalizes the perfect model semantics for locally stratified
programs.

2.5.16 Theorem Definite programs are locally stratified and have a total
weakly perfect model.

Proof: The first statement is trivial. For the second statement, let P be a
definite program with least model I. Assign levels l(A) to all A ∈ I according
to Proposition 2.3.2, and set l(B) = 0 for all B ∈ I. Considering the charac­
terization of the weakly perfect model from Theorem 2.5.9, we observe that
all A ∈ I satisfy (WSi), while all other atoms satisfy (WSiib), and this suffices
to establish the result. •
56 Mathematical Aspects of Logic Programming Semantics

2.6 Well-Founded Models

If we compare Definitions 2.4.8 and 2.5.8 and keep in mind that the main
idea underlying stratification is to restrict recursion through negation, one
may be led to ask whether Definition 2.5.8 is the most natural way to achieve
this in a three-valued setting. Indeed, one may be led to propose the following
definition.

2.6.1 Definition Let P be a normal logic program, let I be a model for

P , and let l be an I-partial level mapping for P . We say that P satisfies
(WF) with respect to I and l if each A ∈ dom(l) satisfies one of the following
conditions.
(WFi) A ∈ I, and there is a clause A ← L1 , . . . , Ln in ground(P ) such that
Li ∈ I and l(A) > l(Li ) for all i.
(WFii) ¬A ∈ I, and for each clause A ← A1 , . . . , An , ¬B1 , . . . , ¬Bm in
ground(P ) one (at least) of the following conditions holds.
(WFiia) There exists i with ¬Ai ∈ I and l(A) ≥ l(Ai ).
(WFiib) There exists j with Bj ∈ I and l(A) > l(Bj ).
If A ∈ dom(l) satisfies (WFi), then we say that A satisfies (WFi) with respect
to I and l, and similarly if A ∈ dom(l) satisfies (WFii).
We note that conditions (Fi), (WSi), and (WFi) are identical, and, fur­
thermore, if P satisfies (WS) with respect to I and l, then it satisfies (WF)
with respect to I and l. However, replacing (WFi) by a “stratified version”
such as the following is not satisfactory.
(SFi) A ∈ I, and there is a clause A ← A1 , . . . , An , ¬B1 , . . . , ¬Bm in
ground(P ) such that Ai , ¬Bj ∈ I, l(A) ≥ l(Ai ), and l(A) > l(Bj )
for all i and j.
Indeed, if we do replace condition (WFi) by condition (SFi), then it is not
guaranteed that, for a given program, there is a greatest model satisfying the
desired properties. Consider the program consisting of the two clauses p ← p
and q ← ¬p, the two (total) models {p, ¬q} and {¬p, q}, and the level mapping
l with l(p) = 0 and l(q) = 1. These models are incomparable, yet in both cases
the conditions obtained by replacing (WFi) by (SFi) in (WF) are satisfied.
So, in the light of Theorem 2.4.9, Definition 2.6.1 should provide a natural
stratified version of the Fitting semantics, and indeed it does, see Program
2.6.12 for an instructive example. Furthermore, the resulting semantics coin­
cides with another well-known semantics, called the well-founded semantics,
which is a very satisfactory result. To establish this claim, we need to introduce
well-founded models, and this we do next.
 The Semantics of Logic Programs 57

Given a normal logic program P and I ∈ IP,4 , we say that U ⊆ BP is

an unfounded set (of P ) with respect to I if each atom A ∈ U satisfies the
following condition. For each clause A ← body in ground(P ) at least one of
the following holds.
(US1) Some (positive or negative) literal in body is false in I.
(US2) Some (non-negated) atom in body occurs in U .

2.6.2 Proposition Let P be a program, and let I ∈ IP,4 . Then there exists
a greatest unfounded set of P with respect to I.
Proof: If (Ui )i∈I is a family of sets, each ofwhich is an unfounded set of P
with respect to I, then it is easy to see that i∈I Ui is also an unfounded set
of P with respect to I. •

Let P be a program, and recall the definition of the operator TP' from
Section 2.4. It is straightforward to lift TP' to an operator on IP,4 , namely, by
defining TP' (I), for I ∈ IP,4 , to be the set of all A ∈ BP for which there is a
clause A ← body in ground(P ) with body true in I with respect to Kleene’s
strong three-valued logic. For all I ∈ IP,4 , define UP (I) to be the greatest
unfounded set (of P ) with respect to I. Finally, define14

WP (I) = TP' (I) ∪ ¬UP (I)

for all I ∈ IP,4 . We call WP the WP -operator.

We note that WP does not restrict to a function on IP,3 , which necessitates
using IP,4 instead.

2.6.3 Example Consider Program 2.3.1 and I = {p} ∈ IP,3 . Then TP' (I) =
{p} and UP (I) = {p}, so WP (I) = {p, ¬p} ∈ IP,3 .

2.6.4 Proposition Let P be a program. Then WP is monotonic on IP,4 .

Proof: Let I, K ∈ IP,4 with I ⊆ K. Then we obtain TP' (I) ⊆ TP' (K) as in
the proof of Proposition 2.4.4. So it suffices to show that every unfounded set
of P with respect to I is also an unfounded set of P with respect to K, and
this fact follows immediately from the definition. •

Since WP is monotonic, it has a least fixed point by the Knaster-Tarski

theorem, Theorem 1.1.10. The least fixed point of WP is called the well-
founded model for P , giving the well-founded semantics of P . We will show
shortly that the well-founded model is always in IP,3 , but let us remark first
that the operator WP is not order continuous in general nor even ω-continuous,
as the following example shows.
14 The operator W
P and the well-founded semantics are due to Van Gelder, Ross, and
Schlipf, see [Van Gelder et al., 1991]. However, in the original definition, the operator WP
was not introduced using FOU R.
58 Mathematical Aspects of Logic Programming Semantics

2.6.5 Program Let P be the following program.

p(0) ←
p(s(X)) ← p(X)
q(s(X)) ← ¬p(X)
r ← ¬q(s(X))

Then WP ↑ n = {p(sk (0)) | k < n} ∪ {¬q(sk (0)) | 0 < k < n}, and

WP ↑ ω = {p(sn (0)) | n ∈ N} ∪ {¬q(sn (0)) | n ∈ N, n > 0}

= {p(sn (0)) | n ∈ N} ∪ {¬q(sn (0)) | n ∈ N, n > 0} ∪ {¬r}
= WP ↑ (ω + 1).

2.6.6 Theorem Let P be a program. Then WP ↑ α ∈ IP,3 for all ordinals α.

In particular, the well-founded model for P is in IP,3 .

Proof: We first need some notation. Let M denote the least fixed point of
WP , and for each atom A ∈ M + let l(A) be the least ordinal β such that
A ∈ WP ↑ (β + 1).
Now assume that there is an ordinal γ which is least under the condition
that WP ↑ γ ∈ IP,3 . Then γ must be a successor ordinal, since IP,3 is a
complete partial order; so let I = WP ↑ (γ − 1) ∈ IP,3 . Now consider the set
U = TP' (I) ∩ UP (I). Then for each A ∈ U and each clause A ← body in
ground(P ) such that body is true in I, we have that some (non-negated) atom
B in body occurs in UP (I). We obtain B ∈ UP (I) ∩ I, and since I ⊆ TP' (I)
we get B ∈ U . Now let A ∈ U be chosen such that it is minimal with respect
to l(A) = β, and notice that necessarily β < γ. Then there exists a clause
A ← body in ground(P ) with body true in WP ↑ β ⊆ I, and in particular
B ∈ I and l(B) < l(A) for all (non-negated) atoms B which occur in body.
But now we have just shown that B ∈ U , contradicting minimality of l(A). •

2.6.7 Proposition Let P be a program, and let I ∈ IP,3 . Then ΦP (I) ⊆

WP (I). Furthermore, the three-valued fixed points of WP are three-valued
supported models for P with respect to Kleene’s strong three-valued logic.

Proof: Let A ∈ FP (I). Then for each clause A ← body in ground(P ), we have
that I(body) = f , and so there is a literal L ∈ body with I(L) = f . But then
A is in the greatest unfounded set of P with respect to I, and so A ∈ UP (I).
This shows that ΦP (I) ⊆ WP (I).
Now let M = WP (M ) = TP' (M ) ∪ ¬UP (M ). We show that M = ΦP (M ) =
'
TP (M ) ∪ ¬FP (M ). For this it suffices to show that UP (M ) ⊆ FP (M ). Let
A ∈ UP (M ), and let A ← body be an arbitrary clause in ground(P ) with
head A. Noting that UP (M ) is an unfounded set of P with respect to M , if
condition (US1) in the definition of an unfounded set holds, then body is false
 The Semantics of Logic Programs 59

in M in Kleene’s strong three-valued logic. If (US2) holds, then some atom

in body occurs in UP (M ) and, therefore, is false in M . Consequently, body is
again false in M in Kleene’s strong three-valued logic. Hence, A ∈ FP (M ), as
required. •

We will now show formally that the well-founded model can be character­
ized using Definition 2.6.1.15

2.6.8 Theorem Let P be a normal logic program with well-founded model

M . Then, in the knowledge ordering, M is the greatest model among all
models I for which there exists an I-partial level mapping l for P such that
P satisfies (WF) with respect to I and l.

Proof: Let MP be the well-founded model for P , and define the MP -partial
level mapping lP as follows: lP (A) = α, where α is the least ordinal such that
A is not undefined in WP ↑ (α + 1). The proof will proceed by establishing the
following facts. (1) P satisfies (WF) with respect to MP and lP . (2) If I is a
model for P and l is an I-partial level mapping such that P satisfies (WF)
with respect to I and l, then I ⊆ MP .
(1) Let A ∈ dom(lP ), and suppose that lP (A) = α. We consider the two
cases corresponding to (WFi) and (WFii).
Case i. A ∈ MP . Then A ∈ TP' (WP ↑ α). Hence, there exists a clause
A ← body in ground(P ) such that body is true in WP ↑ α. Thus, for all
Li ∈ body, we have that Li ∈ WP ↑ α. Hence, lP (Li ) < α = lP (A) and
Li ∈ MP for all i. Consequently, A satisfies (WFi) with respect to MP and
lP .
Case ii. ¬A ∈ MP . Then A ∈ UP (WP ↑ α), and so A is contained in the
greatest unfounded set of P with respect to WP ↑ α. Hence, for each clause
A ← body in ground(P ), either (US1) or (US2) holds for this clause with
respect to WP ↑ α and the unfounded set UP (WP ↑ α). If (US1) holds, then
there exists some literal L ∈ body with ¬L ∈ WP ↑ α. Hence, lP (L) < α and
condition (WFiia) holds relative to MP and lP if L is an atom, or condition
(WFiib) holds relative to MP and lP if L is a negated atom. On the other
hand, if (US2) holds, then some (non-negated) atom B in body occurs in
UP (WP ↑ α). Hence, lP (B) ≤ lP (A), and A satisfies (WFiia) with respect to
MP and lP . Thus, we have established that the statement (1) holds.
(2) We show via transfinite induction on α = l(A) that if A ∈ I, or ¬A ∈ I,
then A ∈ WP ↑ (α + 1), or ¬A ∈ WP ↑ (α + 1)), respectively. For the base case,
note that if l(A) = 0, then A ∈ I implies that A occurs as the head of a fact
in ground(P ). Hence, A ∈ WP ↑ 1. If ¬A ∈ I, then consider the set U of all
atoms B with l(B) = 0 and ¬B ∈ I. We show that U is an unfounded set of
P with respect to WP ↑ 0, and this suffices since it implies ¬A ∈ WP ↑ 1 by the
fact that A ∈ U . So let C ∈ U , and let C ← body be a clause in ground(P ).
15 A different characterization using level mappings, which is nevertheless in the same

spirit, can be found in [Lifschitz et al., 1995].

60 Mathematical Aspects of Logic Programming Semantics

Since ¬C ∈ I, and l(C) = 0, we have that C satisfies (WFiia) with respect to

I and l, and so condition (US2) is satisfied showing that U is an unfounded
set of P with respect to I. Assume now that the induction hypothesis holds
for all B ∈ BP with l(B) < α. We consider two cases.
Case i. A ∈ I. Then A satisfies (WFi) with respect to I and l. Hence,
there is a clause A ← body in ground(P ) such that body ⊆ I and l(K) < α
for all K ∈ body. Hence, body ⊆ WP ↑ α, and we obtain A ∈ TP' (WP ↑ α), as
required.
Case ii. ¬A ∈ I. Consider the set U of all atoms B with l(B) = α and
¬B ∈ I. We show that U is an unfounded set of P with respect to WP ↑ α,
and this suffices since it implies ¬A ∈ WP ↑ (α + 1) by the fact that A ∈ U .
So let C ∈ U , and let C ← body be a clause in ground(P ). Since ¬C ∈ I,
we have that C satisfies (WFii) with respect to I and l. If there is a literal
L ∈ body with ¬L ∈ I and l(L) < l(C), then by the induction hypothesis
we obtain ¬L ∈ WP ↑ α, and therefore condition (US1) is satisfied for the
clause C ← body with respect to WP ↑ α and U . In the remaining case, we
have that C satisfies condition (WFiia), and there exists an atom B ∈ body
with ¬B ∈ I and l(B) = l(C). Hence, B ∈ U showing that condition (US2) is
satisfied for the clause C ← body with respect to WP ↑ α and U . Hence, U is
an unfounded set of P with respect to WP ↑ α. •

As a special case, we immediately obtain the following corollary.

2.6.9 Corollary A normal logic program P has a total well-founded model

if and only if there is a total model I for P and a (total) level mapping l such
that P satisfies (WF) with respect to I and l.

The well-founded model is in general different from the weakly perfect

model, but always contains it.

2.6.10 Proposition Let P be a program, let M1 be its (partial) weakly per­

fect model, and let M2 be its well-founded model. Then M1 ⊆ M2 .

Proof: Let l1 be an M1 -partial level mapping such that P satisfies (WS) with
respect to M1 and l1 . Then P satisfies (WF) with respect to M1 and l1 , as
noted earlier. By Theorem 2.6.8, M2 is largest among all models I for which
there exists an I-partial level mapping l for P such that P satisfies (WF) with
respect to I and l, and hence M1 ⊆ M2 . •

2.6.11 Program Let P be the program consisting of the two clauses

p ← q, ¬p
q←p

Then the reduct P1 of P with respect to the empty set is P itself, that is,
 The Semantics of Logic Programs 61

P1 = P/∅ = P . The only minimal component of P1 is the set {p, q}, and hence
the bottom layer of P1 is P ; it follows that the (partial) weakly perfect model
for P is ∅. However, by applying Theorem 2.6.8, it is easy to see that {¬p, ¬q}
is the well-founded model for P . Indeed, more directly, we have TP' (∅) = ∅,
and UP (∅) = {p, q}. Therefore, WP ↑ 2 = WP (WP ↑ 1) = WP (∅ ∪ {¬p, ¬q}) =
{¬p, ¬q} = WP ↑ 1, and it follows that the well-founded model for P is indeed
{¬p, ¬q}.
An irregular property of the weakly perfect model semantics is that certain
changes in the program affect the semantics, although inutitively they should
not.

2.6.12 Program (Tweety4) Consider again the program Tweety4 of Ex­

ample 2.5.5. As noted earlier, this program is a variation of Tweety2 (Pro­
gram 2.3.9), with the last clause changed; it is intuitively clear that this change
should not alter the semantics of the program.
While the program Tweety2, which is locally stratified, has the expected
weakly perfect model as discussed in Example 2.5.3, the program Tweety4 has
weakly perfect model

{penguin(tweety), bird(bob), bird(tweety), ¬flies(tweety)},

as shown in Example 2.5.5. So again we are unable to determine whether or

not bob is a penguin.
The well-founded semantics, however, does not suffer from the same defi­
ciency. Indeed, it turns out to be M ∪ ¬(BP \ M ), where M is as in Example
2.2.7. So in this semantics bob is not a penguin and flies.
An alternative way of characterizing the well-founded semantics is via the
Gelfond–Lifschitz operator from Section 2.3. Recall from Theorem 2.3.7 that
the Gelfond–Lifschitz operator is antitonic. In particular, this means that
for any program P , the operator GL2P , obtained by applying GLP twice, is
monotonic. Therefore, by the Knaster-Tarski theorem, GL2P has a least fixed
point, LP . Note further that IP,2 is a complete lattice in the dual of the truth
ordering on IP,2 . So, on applying the Knaster-Tarski theorem again, we also
obtain that GL2P has a greatest fixed point, GP . Since LP ⊆ GP , we obtain
that LP ∪ ¬(BP \ GP ) is a three-valued interpretation for P and is, in fact, a
model for P , as we show next, called the alternating fixed point model for P .
We are going to show that the alternating fixed point model coincides
with the well-founded model. Let us first introduce some temporary notation,
where P is an arbitrary program.

L0 = ∅ G0 = BP
Lα+1 = GLP (Gα ) Gα+1 = GLP (Lα ) for any ordinal α

�
Lα = Lβ Gα = Gβ for a limit ordinal α.
β<α β<α
62 Mathematical Aspects of Logic Programming Semantics

Since ∅ ⊆ BP , we obtain L0 ⊆ L1 ⊆ G1 ⊆ G0 , and, by transfinite induc­

tion, it can easily be shown that Lα ⊆ Lβ ⊆ Gβ ⊆ Gα whenever α ≤ β.

2.6.13 Theorem Let P be a program. Then the following hold.

(a) LP = GLP (GP ) and GP = GLP (LP ).

(b) For every stable model S for P , we have LP ⊆ S ⊆ GP .

(c) M = LP ∪ ¬(BP \ GP ) is the well-founded model for P .

2
Proof: (a) We obtain GLP (GLP (LP )) = GLP (GL2P (LP )) = GLP (LP ), so
GLP (LP ) is a fixed point of GL2P , and hence LP ⊆ GLP (LP ) ⊆ GP . Similarly,
LP ⊆ GLP (GP ) ⊆ GP . Since LP ⊆ GP , we get from the antitonicity of GLP
that LP ⊆ GLP (GP ) ⊆ GLP (LP ) ⊆ GP . Similarly, since GLP (LP ) ⊆ GP , we
obtain GLP (GP ) ⊆ GL2P (LP ) = LP ⊆ GLP (GP ), so GLP (GP ) = LP , and
hence GP = GL2P (GP ) = GLP (LP ).
(b) It suffices to note that S is a fixed point of GLP , by Theorem 2.3.7,
and, hence, is a fixed point of GL2P .
(c) We prove this statement by applying Theorem 2.6.8. First, we define
an M -partial level mapping l. For convenience, we will take as image set of l,
pairs (α, n) of ordinals, where n ≤ ω, with the lexicographic ordering. This can
be done without loss of generality because any set of pairs of ordinals, lexico­
graphically ordered, is certainly well-ordered and therefore order-isomorphic
to an ordinal, as noted earlier. For A ∈ LP , let l(A) be the pair (α, n), where
α is the least ordinal such that A ∈ Lα+1 , and n is the least ordinal such that
A ∈ TP/Gα ↑ (n + 1). For B ∈ GP , let l(B) be the pair (β, ω), where β is the
least ordinal such that B ∈ Gβ+1 . We show next by transfinite induction that
P satisfies (WF) with respect to M and l.
Let A ∈ L1 = TP/BP ↑ ω. Since P/BP consists of exactly all clauses from
ground(P ) which contain no negation, we have that A is contained in the least
two-valued model for a definite subprogram of P , namely, P/BP , and (WFi)
is satisfied, by Proposition 2.3.2. Now let ¬B ∈ ¬(BP \ GP ) be such that
B ∈ (BP \ G1 ) = BP \ TP/∅ ↑ ω. Since P/∅ contains all clauses from ground(P )
with all negative literals removed, we obtain that each clause in ground(P )
with head B must contain a positive body literal C ∈ G1 , which, by definition
of l, must have the same level as B; hence, (WFiia) is satisfied.
Assume now that, for some ordinal α, we have shown that A satisfies (WF)
with respect to M and l for all n ≤ ω and all A ∈ BP with l(A) ≤ (α, n).
Let A ∈ Lα+1 \ Lα = TP/Gα ↑ ω \ Lα . Then A ∈ TP/Gα ↑ n \ Lα for some
n ∈ N; note that all (negative) literals which were removed by the Gelfond–
Lifschitz transformation from clauses with head A have level less than (α, 0).
Then the assertion that A satisfies (WF) with respect to M and l follows
again by Proposition 2.3.2.
Let A ∈ (BP \ Gα+1 ) ∩ Gα . Then we have A ∈ TP/Lα ↑ ω. Let A ←
A1 , . . . , Ak , ¬B1 , . . . , ¬Bm be a clause in ground(P ). If Bj ∈ Lα for some j,
 The Semantics of Logic Programs 63

then l(A) > l(Bj ). Otherwise, since A ∈ TP/Lα ↑ ω, we have that there exists
Ai with Ai ∈ TP/Lα ↑ ω, and hence l(A) ≥ l(Ai ), and this suffices.
This finishes the proof that P satisfies (WF) with respect to M and l. It
therefore only remains to show that M is greatest with this property.
So assume that M1 = M is the greatest model such that P satisfies (WF)
with respect to M1 and some M1 -partial level mapping l1 .
Assume L ∈ M1 \ M , and, without loss of generality, let the literal L be
chosen such that l1 (L) is minimal. We consider the following two cases.
Case i. If L = A is an atom, then there exists a clause A ← body in
ground(P ) such that body is true in M1 and l1 (L) < l1 (A) for all literals
L in body. Hence, body is true in M , and A ← body transforms to a clause
A ← A1 , . . . , An in P/GP with A1 , . . . , An ∈ LP = TP/GP ↑ ω. But this implies
A ∈ M , contradicting A ∈ M1 \ M .
Case ii. If L = ¬A ∈ M1 \ M is a negated atom, then ¬A ∈ M1 and
A ∈ GP = TP/LP ↑ ω, so A ∈ TP/LP ↑ n for some n ∈ N. We show by induction
on n that this leads to a contradiction to finish the proof.
If A ∈ TP/LP ↑ 1, then there is a unit clause A ← in P/LP , and any
corresponding clause A ← ¬B1 , . . . , ¬Bk in ground(P ) satisfies B1 , . . . , Bk ∈
LP . Since ¬A ∈ M1 , we also obtain by Theorem 2.6.8 that there is i ∈
{1, . . . , k} such that Bi ∈ M1 and l1 (Bi ) < l1 (A). By minimality of l1 (A), we
obtain Bi ∈ M , and hence Bi ∈ LP , which contradicts Bi ∈ LP .
Now assume that there is no ¬B ∈ M1 \ M with B ∈ TP/LP ↑ k for any
k < n + 1, and let ¬A ∈ M1 \ M with A ∈ TP/LP ↑ (n + 1). Then there is a
clause A ← A1 , . . . , Am in P/LP with A1 , . . . , Am ∈ TP/LP ↑ n ⊆ GP , and we
note that we cannot have ¬Ai ∈ M1 \ M for any i ∈ {1, . . . , m} by our current
induction hypothesis. Furthermore, it is also impossible for ¬Ai to belong to
M for any i; otherwise, we would have Ai ∈ BP \ GP . Thus, we conclude
that we cannot have ¬Ai ∈ M1 for any i. Moreover, there is a corresponding
clause A ← A1 , . . . , Am , ¬B1 , . . . , ¬Bm1 in ground(P ) with B1 , . . . , Bm1 ∈
LP . Hence, by Theorem 2.6.8, we know that there is i ∈ {1, . . . , m1 } such
that Bi ∈ M1 and l1 (Bi ) < l1 (A). By minimality of l1 (A), we conclude that
Bi ∈ M , so that Bi ∈ LP , and this contradicts Bi ∈ LP . •

It follows from Theorem 2.6.13 (b) that total well-founded models are
unique stable models. The converse, however, does not hold. Indeed, Program
2.4.14 has well-founded model ∅, as can easily be seen by noting that GLP (∅) =
BP and GLP (BP ) = ∅.

2.6.14 Theorem Let P be a program with a total Fitting model. Then P

has a total well-founded model and a total weakly perfect model. Moreover,
P also has a unique stable and a unique supported model. Furthermore, all
these models coincide.
Proof: By Propositions 2.5.14 and 2.6.10, P has a total well-founded and a
total weakly perfect model, both of which coincide with the Fitting model.
By Theorem 2.6.13 (b), P has a unique stable model, and this coincides with
64 Mathematical Aspects of Logic Programming Semantics

the well-founded model by Theorem 2.6.13 (c). Finally, by Proposition 2.4.13,

P has a unique supported model, and this model coincides with its Fitting
model. •
Chapter 3

Topology and Logic Programming

In this chapter, we consider the role of topology in logic programming se­

mantics. There is a considerable history of topology being used in computer
science in general, much of it stemming from the role of the Scott topology in
domain theory and in conventional programming language semantics. How­
ever, topological methods have been employed in a number of other areas
of importance in computing, including digital topology in image processing,
software engineering, and the use of metric spaces in concurrency, for exam­
ple. In addition, topological methods and ideas have been used in founda­
tional investigations via the topology of observable properties of M.B. Smyth,
see [Smyth, 1992]. Again, Blair et al. have made considerable use of con­
vergence spaces in unifying discrete and continuous models of computation
and, hence, in providing models for hybrid systems. Indeed, these authors,
see [Blair et al., 1999] and [Blair and Remmel, 2001], for example, view any
model of computation in which there is a notion of evolving state as a dy­
namical system. Such models of computation include, of course, Turing ma­
chines, finite state machines, logic programs, neural networks, etc. On the
other hand, convergence spaces, as already noted earlier, provide a very gen­
eral framework in which to study convergence and continuity, either by means
of nets or by filters, and include topologies as a special case. It is shown in
[Blair et al., 1999] and [Blair and Remmel, 2001] that the execution traces of
a dynamical system can be realized as those solutions of a certain type of
constraint on a convergence space that yield continuous instances of the con­
straint. This work provides a foundation for hybrid systems. Furthermore, the
papers [Blair et al., 1997a, Blair et al., 1997b, Blair, 2007, Blair et al., 2007]
give many other interesting applications of ideas of a dynamical systems and
analytical nature to the theory of computation, including logic programming
in particular.
Here, we want to explore the role of topology in finding models for logic
programs and its role as a foundational framework for logic programming
semantics.1 Thus, our focus is the study of topologies and their properties
on spaces I(X, T ) of interpretations, and we work with general truth sets
T wherever possible, only imposing conditions as appropriate and necessary.
There are two main topologies which we discuss in this chapter and which have
1 The thesis [Ferry, 1994] and the paper [Heinze, 2003] contain results concerning the

characterization in topological terms of the various standard models for logic programs
discussed in Chapter 2.

65
66 Mathematical Aspects of Logic Programming Semantics

important properties in relation to logic programming semantics, namely, the

well-known Scott topology and a topology, called the Cantor topology by us,2
which has connections with the Scott topology. Our goal is to establish the
basic facts about these two topologies and to consider continuity of semantic
operators in them. In fact, we deal with continuity in the Scott topology in this
chapter, but postpone our discussion of continuity in the Cantor topology until
Chapter 5. Later on, we will see how the results we establish can be employed
in studying acceptable programs and termination issues, and we will also see
that the topologies we discuss underlie the fixed-point structures we introduce
in later chapters.
In fact, in many ways it is the convergence properties of these topologies
which are most important, as already noted in the Introduction, and therefore
we take convergence as a fundamental notion and base our discussion upon
it. Nevertheless, we quite easily obtain descriptions of the topologies we study
in terms of more familiar notions such as basic open sets. Actually, conver­
gence per se is formalized completely generally via the concept of convergence
spaces, and therefore we take convergence spaces as our starting point. In fact,
we focus mainly on the so-called convergence classes, which form a subclass
of the convergence spaces, because convergence classes correspond to conven­
tional topologies, whereas convergence spaces give more general theories of
convergence than are needed here.
As can be seen from the results of Chapter 2, the notion of order is not
entirely satisfactory as a foundation for logic programming semantics due to
the failure in general of the immediate consequence operator to be monotonic
in the natural order present. However, order can be expressed through conver­
gence, as we show here. Indeed, convergence spaces and convergence classes
are to a considerable extent appropriate structures with which to investigate
semantical questions in computer science in general and in logic programming
in particular.

3.1 Convergence Spaces and Convergence Classes

The theory of convergence can be based either on nets or on filters,3 and
these two approaches are equivalent in that any result which can be estab­
lished by the one can equally well be established by the other. We will work
exclusively with nets since they give rather intuitive descriptions of the sort
of conditions we want to consider in logic programming. The facts we need
2 The Cantor topology was introduced in [Batarekh and Subrahmanian, 1989a] and in

[Batarekh and Subrahmanian, 1989b], see also [Batarekh, 1989], under a restriction called
the matching condition and was treated in complete generality in [Seda, 1995].
3 Our basic references to the theory of nets and filters are the books [Kelley, 1975] and

[Willard, 1970].
 Topology and Logic Programming 67

concerning nets, and our notation in this respect, can be found in the Ap­
pendix.4 Indeed, all the basic facts we need concerning general topology have
been collected together in the Appendix.
We begin with some basic definitions.

3.1.1 Definition Let X be a non-empty set. We call the pair (X, S) =

(X, (Ss )s∈X ) a convergence space if, for each s ∈ X, Ss is a non-empty collec­
tion of nets in X with the following properties.
(1) If (si ) is a constant net, that is, si = s ∈ X for all i, then (si ) ∈ Ss .
(2) If (si )i∈I ∈ Ss and (tj )j∈J is a subnet of (si ), then (tj )j∈J ∈ Ss .
If (si ) ∈ Ss , we say si converges to s and sometimes write si → s to indicate
this.

3.1.2 Definition Let X be a non-empty set, and suppose that C is a class of

pairs ((si ), s), where (si )i∈I is a net in X and s is an element of X. We call C
a convergence class if it satisfies the conditions below, in which we will write
that si converges (C) to s or that limi si ≡ s (C) if and only if ((si ), s) ∈ C.
(1) (Constant nets) If (si ) is a net such that si = s for all i, then ((si ), s) ∈ C.
(2) (Convergence of subnets) If (si ) converges (C) to s, then so does every
subnet of (si ).
(3) (Non-convergence)5 If (si ) does not converge (C) to s, then there is a
subnet of (si ), of which no subnet converges (C) to s.
(4) (Iterated limits) Suppose that I is a directed set and that Jm is
 a directed
set for each m ∈ I. Form the fibred product F ' = I ×I m∈I Jm =
{(m, n) | m ∈ I, n ∈ Jm }, andssuppose that x : F ' → X. Let F denote
the product directed set6 I × m∈I Jm , and let r : F → F ' be defined
by r(m, f ) = (m, f (m)). If limm limn x(m, n) ≡ s (C), then the net x ◦ r
converges (C) to s.
The principal result concerning convergence classes, see [Kelley, 1975,
Chapter 2] or [Seda et al., 2003], is that each convergence class C on X in­
duces a closure operator on X which in turn induces a topology on X, in
accordance with Theorem A.2.9, in which the convergent nets and their limits
are precisely those given in C. More precisely, we have the following result
which shows that the notion of convergence may be taken as fundamental.
4 We refer the reader again to [Kelley, 1975] for more details.
5 This formulation is as given in [Kelley, 1975]. An equivalent form, given in positive
terms, is as follows: if every subnet of a net (si ) has a subnet converging to s, then (si )
converges to s.
6 By a product directed set
Q
Q m∈I Im , we understand, of course, the pointwiseQordering
on the product m∈I Im of the directed sets Im ; thus, for elements f and g of m∈I Im ,
we have f ≤ g if and only if f (m) ≤ g(m) for each m ∈ I.
68 Mathematical Aspects of Logic Programming Semantics

3.1.3 Theorem Let C be a convergence class in a non-empty set X. For each

A ⊆ X, let Ac = {s ∈ X | there is a net (si ) in A with ((si ), s) ∈ C}. Then
·c is a closure operator on X and, hence, defines a topology τ on X, called
the topology associated with C. Moreover, we have ((si ), s) ∈ C if and only if
si → s with respect to τ .
Conversely, suppose that τ is a topology on a non-empty set X. Let C
denote the set of all pairs ((si ), s), where s ∈ X and (si )i∈I is a net in X
which converges to s in the topology τ . Then C is a convergence class in X
whose associated topology coincides with τ .
Proof: The proof of the first part of the theorem is well-known and will be
omitted, and we refer the reader to [Kelley, 1975] or [Seda et al., 2003] for
details.
For the converse, we note that properties (1), (2), and (3) in the definition
of a convergence class are immediate for the class C by elementary properties
of nets converging in a topology (see Definition A.3.3). Property (4) of the
definition follows from the Theorem on Iterated Limits, see [Kelley, 1975, Page
69], and, hence, the class C is a convergence class. Finally, let A ⊆ X be an
arbitrary subset of X. By the definition of the closure operator determined by
C as given in the first statement in the theorem, we have s ∈ Ac if and only
if there is a net (si ) in A converging to s. But this is equivalent to s ∈ A by
statement (a) of Theorem A.3.5, and it follows that the associated topology
of C coincides with τ . •

Another basic definition is that of continuous function, as follows.

3.1.4 Definition Let (X, S) and (Y, T ) be convergence spaces. Then a func­
tion f : X → Y is said to be continuous at s ∈ X if (f (si )) ∈ Tf (s) whenever
(si ) ∈ Ss , that is, if f (si ) converges to f (s) whenever si converges to s.
There are a few points to be made about these definitions. First, suppose
that C is a convergence class on X. For each s ∈ X, let Ss denote the collection
of nets (si ) such that ((si ), s) ∈ C. Then conditions (1) and (2) in the definition
of C show that (X, (Ss )s∈X ) is, in fact, a convergence space. Second, since a
function f : X → Y between topological spaces is continuous at s ∈ X
if and only if f (si ) converges to f (s) whenever the net si converges to s,
see (d) of Theorem A.3.5, we note that the notion of continuity just defined
coincides with topological continuity when the convergence spaces in question
are actually convergence classes. Finally, definitions equivalent to these can
be given entirely in terms of filters, but we omit the details.7
It is known that the full generality of convergence spaces is needed in mod­
elling hybrid systems, as observed earlier. Here, in fact, all the convergence
conditions we consider give rise to convergence classes and, hence, to topolo­
gies, rather than to strict convergence spaces, and therefore our focus is on
convergence classes as already noted.
7 We refer the reader to [Seda et al., 2003] for a treatment in terms of filters.
 Topology and Logic Programming 69

3.2 The Scott Topology on Spaces of Valuations

The Scott topology is normally encountered in domain theory in the con­
text of solving recursive domain equations and in understanding self reference.
However, it also has a role in logic programming, which we discuss in this sec­
tion, and indeed, in a certain sense, it naturally underpins definite programs.
We begin with the following basic definition and refer the reader to the
Appendix, both for proofs of the results we simply state here and also for a
development of the elements of the Scott topology.

3.2.1 Definition Let (D, [) be a complete partial order. A set O ⊆ D is

called Scott open 8 if it satisfies the following two conditions: (1) O is upwards
closed in the sense that wheneverx ∈ O and x [ y, we have y ∈ O, and (2)
whenever A ⊆ D is directed and A ∈ O, then A ∩ O = ∅.

In the case of a domain D, this topology has a rather simple description

in that the collection {↑ a | a ∈ Dc } is a base for the Scott topology on D,
where ↑ x = {y ∈ D | x [ y} for any x ∈ D, as we see in the next proposition.

3.2.2 Proposition Let (D, [) be a domain. Then the following statements

hold.

(a) The Scott-open sets form a topology on D called the Scott topology.

(b) For each compact element a ∈ Dc , the set ↑ a is a Scott-open set.

(c) The collection {↑ a | a ∈ Dc } is a base for the Scott topology on D.

Proof: (a) That ∅ and D are Scott open is easy to see. If O1 and O2 are
Scott open, if x ∈ O1 ∩ O2 , and if x [ y, then it is clear that y ∈ O1 ∩ O2 .
Suppose that A is directed and A ∈ O1 ∩ O2 . Then there are a1 , a2 ∈ A
such that a1 ∈ O1 and a2 ∈ O2 . Therefore, by directedness of A, there is
a3 ∈ A such that a1 [ a3 and a2 [ a3 . But then a3 ∈ O1 ∩ O2 , and hence
a3 ∈ A ∩ (O1 ∩ O2 ), as required
 to see that O1 ∩ O2 is Scott open. Finally, it is
easy to check that a union i∈I Oi of Scott-open sets Oi , i ∈ I, is itself Scott
open.
(b) If x ∈ ↑ a and x[ y, then it is immediate that y ∈ ↑ a.
 Now suppose
that A is directed and A ∈ ↑ a. Then a is compact and a [ A. Therefore,
there is a' ∈ A such that a [ a' . Hence, a' ∈ ↑ a by definition of ↑ a, that is,
a' ∈ A ∩ ↑ a showing that A ∩ ↑ a = ∅, as required.
(c) First we show that this collection is a base for some topology on D.
Let x ∈ D be arbitrary. Then approx(x) is directed and is non-empty;
 let a ∈
approx(x). Then, a ∈ Dc and a [ x, so that x ∈ ↑ a, and hence a∈D ↑ a = D.
8 See [Abramsky and Jung, 1994, Gierz et al., 2003, Stoltenberg-Hansen et al., 1994].
70 Mathematical Aspects of Logic Programming Semantics

Now suppose that a1 and a2 are compact elements and that z ∈ ↑ a1 ∩ ↑ a2 .

Then a1 , a2 ∈ approx(z), and by directedness there is a3 ∈ approx(z) such
that a1 [ a3 and a2 [ a3 . Hence, we have a3 ∈ ↑ a1 ∩ ↑ a2 . But ↑ a1 ∩ ↑ a2 is
clearly upwards closed, and so we obtain z ∈ ↑ a3 ⊆ ↑ a1 ∩ ↑ a2 and a3 ∈ Dc , as
required.
Finally, we show that the collection {↑ a | a ∈ Dc } is a base for the Scott
topology on D. Let O be any Scott-open set, and let x ∈ O. Then approx(x)
is directed, and we have that approx(x) = x ∈ O. Therefore, there is some
a ∈ approx(x) such that a ∈ O. But then a ∈ Dc and a [ x. Therefore,
x ∈ ↑ a ⊆ O, where a is a compact element, as required. •

We refer to the elements of the Scott topology as Scott-open sets. Likewise,

we refer to neighbourhoods in the Scott topology as Scott neighbourhoods,
and so on.
We next give a simple example of the Scott topology in the context of IP,2 .

3.2.3 Example Consider the definite program P as follows.

p(a) ←
p(s(X)) ← p(X)

This program is intended to compute the natural numbers, where a is the

natural number 0, and s is the successor function on the natural numbers.
In accordance with Theorem 1.3.4, the set IP = IP,2 of all two-valued
interpretations for P is a domain, and, furthermore, its compact elements are
the finite subsets I of BP , where as usual we are identifying a two-valued
interpretation with the set of ground atoms which are true in I. Therefore, a
typical basic open set in the Scott topology on IP is the set ↑ I = {I ' ⊆ BP |
I ⊆ I ' } of all supersets of the finite set I.
One of our main aims here is to present the Scott topology in terms of
convergence, and we proceed to do this next.9

3.2.4 Theorem Let (D, [) denote a domain, let (si ) be a net in D, and let
s denote an element of D. Define limi si ≡ s (C) to mean that
for each a ∈ approx(s), there is an index i0 such that a [ si whenever i0 ≤ i.
Then the condition just given determines a convergence class C whose associ­
ated topology is the Scott topology on D. Therefore, a net si converges to s
in the Scott topology on D if and only if it satisfies the condition just stated.
Proof: We first verify that the conditions (1), (2), (3), and (4) in the definition
of a convergence class, see Definition 3.1.2, hold with the given meaning of
limi si ≡ s (C).
9 For further details of this result and of several more in this chapter, see [Seda, 2002].
 Topology and Logic Programming 71

(1) Suppose that si = s for all i ∈ I is a constant net, and let a ∈

approx(s). Thus, a is a compact element satisfying a [ s. Therefore, we have
a [ si for all i. So, ((si ), s) ∈ C.
(2) Suppose that ((si ), s) ∈ C and that (tj )j∈J is a subnet of (si )i∈I . Thus,
there is a function φ : J → I such that (i) tj = sφ(j) for all j ∈ J , and (ii)
for each i0 ∈ I, there is j0 ∈ J such that i0 ≤ φ(j) whenever j0 ≤ j. Let
a ∈ approx(s) be arbitrary. Then because ((si ), s) ∈ C, there is an i0 ∈ I such
that a [ si whenever i0 ≤ i. Since tj is a subnet of si , there is j0 ∈ J such
that i0 ≤ φ(j) whenever j0 ≤ j. But then we have a [ sφ(j) whenever j0 ≤ j,
that is, a [ tj whenever j0 ≤ j. Therefore, ((tj ), s) ∈ C.
(3) Suppose that ((si ), s) ∈ C. Then there exists a ∈ approx(s) such that
for each index i0 there is an index j0 ≥ i0 with a [ sj0 . Let J denote the
collection of all these j0 . Then clearly J is cofinal in I, and hence (tj )j ∈J is
a subnet of (si ), where tj = sj for each j ∈ J . It is clear that if (rk ) is any
subnet of (tj ), then we have ((rk ), s) ∈ C.
(4) Suppose that the conditions stated in (4) of Definition 3.1.2 all hold
and that limm limn x(m, n) ≡ s (C), where x : F ' → D. Let a ∈ approx(s)
be arbitrary. Because limm limn x(m, n) ≡ s (C), there is an index m0 ∈ I
such that a [ limn x(m, n) whenever m ≥ m0 . But now we see that a ∈
approx(limn x(m, n)). Therefore, for each fixed m ≥ m0 , there s is an index
nm ∈ Jm such that a [ x(m, n) whenever n ≥ nm . Define f ∈ m∈I Jm by
setting f (m) = nm ∈ Jm whenever m ≥ m0 , and otherwise letting f (m) ∈ Jm
be arbitrary. Suppose that (m' , g) ≥ (m0 , f ). Then m' ≥ m0 and g ≥ f , so that
g(m' ) ≥ f (m' ) = nm/ , that is, g(m' ) ≥ nm/ . Thus, a [ x(m' , g(m' )) whenever
(m' , g) ≥ (m0 , f ). Hence, a [ x ◦ r(m' , g) whenever (m' , g) ≥ (m0 , f ), and it
follows that (x ◦ r, s) ∈ C, as required.
Next, we verify that the topology induced on D by the convergence con­
dition coincides with the Scott topology on D. Let O be open in the topology
associated with the convergence class C, let x ∈ O, and suppose that x [ y;
suppose further that y ∈ O, that is, suppose that y is in the closed set D \ O.
Then there is a net si → y with si ∈ D \ O for all i. Let a ∈ approx(x) be ar­
bitrary. Then a ∈ approx(y) and, hence, a [ si eventually. It follows from this
that si → x. Therefore, by (b) of Theorem A.3.5, we see that si is eventually
in O. This contradiction sho  ws that y is, in fact, in O. Next, suppose that A
is a directed set with x = A ∈ O. Then by Proposition A.6.1 we have that,
as a net, A → x. Therefore, A is eventually in O, and so A ∩ O = ∅. Hence, O
is a Scott-open set.
Conversely, suppose that O is a Scott-open set, and let x ∈ O. We show
that O is open in the topology associated with the convergence class C by
establishing that, whenever si → x, we have si eventually in O, and then the
result follows 
from (b) of Theorem A.3.5 again. Now, approx(x) is a directed
set, and x = approx(x) ∈ O. Therefore, there is an element a ∈ approx(x)
such that a ∈ O. Since si → x, it now follows that there is i0 such that for
i0 ≤ i we have a [ si . But then, since a ∈ O and O is Scott open, we have
si ∈ O whenever i0 ≤ i, as required to finish the proof. •
72 Mathematical Aspects of Logic Programming Semantics

Of course, a function f : D → E from a domain D to a domain E is called

Scott continuous if it is continuous in the Scott topologies on D and E. How­
ever, it is well-known that a function f between domains is Scott continuous
if and only if it is continuous in the sense of Definition 1.1.7, see Proposi­
tion A.6.4. Moreover, by virtue of Theorem 1.3.2 and Proposition A.6.5, we
have the following result.

3.2.5 Proposition Suppose that the truth set T is a domain. Then in the
Scott topology I(X, T ) is a compact T0 topological space, but is not T1 in
general.
Nets (and convergence classes), like sequences, are normally simple to han­
dle, and their use makes checking continuity relatively straightforward, as we
will see later on in several places. However, we move next to consider the sig­
nificance of Theorem 3.2.4 in the case of spaces I(X, T ) of valuations, where
the set (T , ≤) of truth values is a domain. Indeed, suppose that (T , ≤) is a do­
main and that the net (vi ) converges to v in the Scott topology on the domain
I(X, T ). According to Theorem 3.2.4, this holds if and only if for each finite
valuation u with u [ v, there is an index i0 such that u [ vi whenever i0 ≤ i.
In fact, when applied to the particular truth sets discussed in Section 1.3.2,
Theorem 3.2.4 gives the following result.

3.2.6 Theorem Suppose that (Ii ) is a net of interpretations and that I is an

interpretation.
(a) Let T denote the truth set T W O. Then, in the ordering [t on I(X, T ), we
have that (Ii ) converges to I in the Scott topology if and only if whenever
x ∈ I, eventually x ∈ Ii .
(b) Let T denote the truth set T HREE. Then the following statements hold.
(i) In the ordering [k on I(X, T ), we have that (Ii ) converges to I in
the Scott topology if and only if whenever x ∈ It , eventually x ∈ Iit ,
and whenever x ∈ If , eventually x ∈ Iif .
(ii) In the ordering [t on I(X, T ), we have that (Ii ) converges to I in
the Scott topology if and only if whenever x ∈ It , eventually x ∈ Iit ,
and whenever x ∈ Iu , eventually x ∈ Iiu ∪ Iit .
(c) Let T denote the truth set FOU R. Then the following statements hold.
(i) In the ordering [k on I(X, T ), we have that (Ii ) converges to I
in the Scott topology if and only if whenever x ∈ It , eventually
x ∈ Iit ∪ Iib , whenever x ∈ If , eventually x ∈ Iif ∪ Iib , and whenever
x ∈ Ib , eventually x ∈ Iib .
(ii) In the ordering [t on I(X, T ), we have that (Ii ) converges to I
in the Scott topology if and only if whenever x ∈ Iu , eventually
x ∈ Iiu ∪ Iit , whenever x ∈ Ib , eventually x ∈ Iib ∪ Iit , and whenever
x ∈ It , eventually x ∈ Iit .
 Topology and Logic Programming 73

Proof: We prove the first of the claims in (c), with the others being proved
similarly. Let v denote the valuation corresponding to the interpretation I,
and, for each index i, let vi denote the valuation corresponding to the in­
terpretation Ii . Suppose first that (vi ) converges to v in the Scott topology
on I(X, T ), and let x ∈ X. Suppose further that x ∈ vt , so that v(x) = t.
Define u ∈ I(X, T ) by u(x) = t, and, for y = x, set u(y) = u. Then u is a
finite element satisfying u [k v. Therefore, by Theorem 3.2.4, there exists i0
such that u [k vi whenever i ≥ i0 , and hence eventually either vi (x) = t or
vi (x) = b. Thus, eventually x ∈ vit ∪ vib . A similar argument holds in case
x ∈ vf or x ∈ vb , and hence we obtain the stated condition.
Conversely, suppose that the given condition holds. Let u be a finite valua­
tion such that u [k v, and suppose further that u takes value u at all points of
X except possibly at one point x, say. Let us first suppose that u(x) = t. Then
either x ∈ vt or x ∈ vb . But then, by the given condition, either eventually
x ∈ vit ∪ vib or eventually x ∈ vib , and in either case, eventually u [k vi . A
similar argument holds in case u(x) = f or u(x) = b. By a standard argument
using the directedness of the index set of the net vi , it follows that, for any
finite valuation u [k v, we have eventually u [k vi . Hence, (vi ) converges to
v in the Scott topology on I(X, T ), as required. •

Thus, we obtain a uniform description of net convergence in the Scott

topology on (I(X, T ), [), where (T , ≤) is any one of the main sets of truth
values which are important in logic programming. Indeed, the convergence
conditions involved are simple, natural, and intuitive, and this is one of the
advantages of approaching this topic via convergence.
In fact, it is Part (a) of Theorem 3.2.6 which we will use most often, and
we illustrate its use next with an example.

3.2.7 Example The following statements concerning convergence in the

Scott topology hold in two-valued logic.10

(1) Any net (Iλ ) of interpretations converges to the empty interpretation ∅.

(2) If (Iλ ) is a net of interpretations which is monotonic

 in the sense that
Iλ ⊆ Iγ whenever λ ≤ γ, then (Iλ ) converges to λ Iλ .

(3) If a net (Iλ ) of interpretations converges to an interpretation I, and J ⊆ I,

then (Iλ ) converges to J. Thus, in general, a net (Iλ ) of interpretations
has many limits. A specific example of this can be given as follows. Sup­
pose that L is a first-order language containing a unary predicate sym­
bol p, a unary function symbol s, and a constant symbol a, such as the
language underlying Example 3.2.3, say. Consider the sequence (In ) of
interpretations defined as follows: In is the set {p(a), p(s(a))} if n is even
and is the set {p(a), p(s(a)), p(s2 (a))} if n is odd. Then (In ) converges to
10 For further results in this direction, see [Seda, 1995].
74 Mathematical Aspects of Logic Programming Semantics

each of the interpretations ∅, {p(a)}, {p(s(a))}, {p(a), p(s(a))}, but not

to {p(a), p(s(a)), p(s2 (a))}.
Again, if In is the interpretation defined by taking it to be the set
{p(a), p(s(a)), . . . , p(sn (a))} if n is even and taking it to be the set
{p(a), p(s(a)), . . . , p(s2n (a))} if n is odd, then the sequence (In ) converges
to the interpretation {p(a), p(s(a)), p(s2 (a)), . . .}; note that (In ) is not
monotonic in the sense of Part (2).

Although we have taken convergence as the basic concept, it is easy to ex­

hibit properties of the Scott topology in other familiar terms, as the following
example shows.

3.2.8 Example In the context of spaces of interpretations, Proposition 3.2.2

gives a simple description of the basic open sets in the Scott topology, and
we briefly consider this point here. In the case of T W O, for example, let
A1 , . . . , An ∈ X and let G(A1 , . . . , An ) = {I ∈ I(X, T W O) | A1 , . . . , An ∈
I}. By means of (f) of Theorem 1.3.2 and (c) of Proposition 3.2.2, it is
clear that the sets G(A1 , . . . , An ) form a base for the Scott topology on
I(X, T WO). Indeed, the sets G(A) = {I ∈ I(X, T W O) � | A ∈ I} form a
subbase for the Scott topology, since G(A1 , . . . , An ) = i∈{1,...,n} G(Ai ). As
another example, consider this time the knowledge ordering [k in the case of
T HREE. Take elements A1 , . . . , An , B1 , . . . , Bm ∈ X, where n, m ≥ 0, and
let G(A1 , . . . , An ; B1 , . . . , Bm ) be the set {I ∈ I(X, T HREE) | A1 , . . . , An ∈
It and B1 , . . . , Bm ∈ If }. Then these sets form a base for the Scott topology
on I(X, T HREE). Indeed, the sets G(A; B) clearly form a subbase for this
topology, where G(A; B) = {I ∈ I(X, T HREE) | A ∈ It and B ∈ If }.
The other cases dealt with in Section 1.3.2 can be treated similarly.
We turn next to consider the continuity of the immediate consequence
operator in the Scott topology. By virtue of (a) of Theorem 2.2.3 and Propo­
sition A.6.4, we have immediately that TP is Scott continuous whenever P is a
definite program. However, we will take the trouble to include a self-contained
proof of this fact next.

3.2.9 Theorem Let P be a definite program. Then TP is continuous in the

Scott topology on IP,2 .
Proof: Let I ∈ IP,2 , and let Ii → I be a net converging to I in the Scott
topology; we show that TP (Ii ) → TP (I) in the Scott topology. If TP (I) = ∅,
then the required conclusion is immediate since, by Theorem 3.2.4, every net
in a domain converges in the Scott topology to the bottom element. So suppose
that TP (I) = ∅, and let A belong to TP (I). Then there is a ground instance
A ← A1 , . . . , An of a clause in P such that I(A1 ∧ . . . ∧ An ) = t, where
n ≥ 0. Since Ii → I, we have, by (a) of Theorem 3.2.6, that eventually
Ii (A1 ∧ . . . ∧ An ) = t. Therefore, A ∈ TP (Ii ) eventually. It now follows from
Theorem 3.2.6 that TP (Ii ) → TP (I) in the Scott topology, as required. •
 Topology and Logic Programming 75

It is not difficult to see that the converse of the previous result fails. For
example, the program P1 with clauses p(a) ← p(a), p(a) ← ¬p(a), and p(b) ←
p(a) and the program P2 with clauses p(a) ← and p(b) ← p(a) have the same
(Scott continuous) immediate consequence operator.
By contrast, recall that Program 2.4.12 showed that the Fitting operator
is not order continuous, and hence not Scott continuous, for definite pro­
grams. Nevertheless, Theorem 3.2.9 justifies our earlier statement that the
Scott topology naturally underpins definite programs.
A theme which is important in this chapter and in later ones concerns the
convergence to some interpretation I of sequences TPn (M ) of iterates of TP
on an interpretation M , and under what conditions I is a model for P . We
discuss this briefly now for definite programs and take it up in more detail in
the next section for normal programs.
In general, if (vi ) is a net converging to v in the Scott topology on I(X, T ),
then it is clear from Theorem 3.2.4, see also Example 3.2.7, that (vi ) converges
to u whenever u [ v and, hence, that the set of limits of (vi ) is downwards
closed.11 Indeed, since (vi ) always converges to ⊥, this latter set is always
non-empty also. Furthermore, when T denotes the complete lattice T W O, we
have by Theorem 1.3.4 that I(X, T ) is itself a complete lattice. Thus, in this
case, the supremum of the set of all limits, in the Scott topology, of a net (vi )
exists and is easily seen to be a limit of (vi ) also, by Theorem 3.2.6. We refer
to this limit as the greatest limit of (vi ) and denote it by gl(vi ). In fact, it
is readily checked that gl(vi ) takes value t precisely on the set of all x ∈ X
at which eventually vi takes value t, and this property completely determines
gl(vi ), see [Seda, 1995] for more details.
Of course, a sequence TPn (M ) always converges to the empty interpretation
∅, as already noted, but the interpretation ∅ need not be a model for P .
However, we do have the following result.

3.2.10 Proposition Let P be a definite logic program, and let M be an inter­

pretation for P . Then the greatest limit gl(TPn (M )) of the sequence (TPn (M ))
is a model for P .

Proof: Let I denote gl(TPn (M )). Then the sequence (TPn (M )) converges to
I in the Scott topology. Hence, by the Scott continuity of TP , the sequence
(TP (TPn (M ))) converges to TP (I). Thus, (TPn (M )) converges to TP (I), and we
obtain, by definition of the greatest limit, that TP (I) ⊆ I, as required. •

Finally, we note that if we take M to be the bottom element in I(X, T ),

then gl(TPn (M )) coincides with the least fixed point of TP and, hence, is the
least model for the definite logic program P , see [Seda, 1995]. Thus, the usual
two-valued semantics for definite programs can be expressed entirely in terms
of convergence in the Scott topology.
11 A subset O of a partially ordered set (D, r) is called downwards closed if, whenever

x ∈ O and y r x, we have y ∈ O.
76 Mathematical Aspects of Logic Programming Semantics

We close this section with an example which, despite its simplicity, illus­
trates the main points discussed previously.

3.2.11 Example Consider again the program P of Example 3.2.3.

p(a) ←
p(s(X)) ← p(X)

Let M = ∅, thought of as a two-valued interpretation, and let In denote

the n-th iterate of TP on M . Then In = {p(a), p(s(a)), . . . , p(sn−1 (a))} for
any n ≥ 1. By Part (2) of Example 3.2.7, the sequence (In ) converges in the
Scott topology to the set I = {p(a), p(s(a)), . . . , p(sn (a)), . . .} of all natural
numbers. Moreover, I is clearly the greatest limit of the sequence (In ) and,
hence, by Theorem 3.2.10, is a model for P . Indeed, by the comments im­
mediately prior to this example, I is the least model for P by the results of
[Seda, 1995].

3.3 The Cantor Topology on Spaces of Valuations

As just noted in the previous section, one of the sources of motivation for
studying topology in relation to logic programming is the role of convergence
of sequences of iterates of the immediate consequence operator in relation
to semantics and also, in fact, in relation to termination. We take this dis­
cussion further now, but this time in the context of normal programs and the
construction of certain standard models for them, and in more detail in Chap­
ter 5. We also refer the reader to Chapter 5 for details of how convergence
enters into questions concerned with the so-called acceptable programs and
problems concerned with termination, see Corollary 5.2.5, Proposition 5.2.7,
Theorem 5.2.8 and Theorem 5.4.14, for example.
We begin with a result concerning product topologies.
Let X and Y be arbitrary sets, and let [X → Y ] denote the set of all total
functions mapping X into Y . When Y is ordered, perhaps as a set of truth
values T , then so is [X → Y ], and, as we have just seen, important topologies
can be defined on [X → Y ] by quite natural convergence conditions which
make use of the order. However, important topologies can also be defined on
[X → Y ] using natural convergence conditions which do not depend on any
order, as we show next.12

3.3.1 Theorem Let (si ) be a net in [X → Y ], and let s ∈ [X → Y ]. Then

the condition
12 Again, see [Seda, 2002].
 Topology and Logic Programming 77

limi si ≡ s (C) if and only if for each x ∈ X eventually si (x) = s(x)

determines a convergence class on [X → Y ] whose associated topology Q is
the product of X copies of the discrete topology on Y .

Proof: We must verify that the conditions (1), (2), (3), and (4) in the defini­
tion of a convergence class, see Definition 3.1.2, hold with the given meaning
of limi si ≡ s (C).
(1) Suppose that si = s for all i ∈ I is a constant net. Then si (x) = s(x)
for all x and all i. Hence, for all x, eventually si (x) = s(x), and so ((si ), s) ∈ C.
(2) Suppose that ((si ), s) ∈ C and that (tj )j ∈J is a subnet of (si )i∈I . Let
x ∈ X be arbitrary, and let i0 be such that si (x) = s(x) for all i ≥ i0 . Since
(tj ) is a subnet of (si ), there is φ : J → I and j0 ∈ J such that i0 ≤ φ(j)
whenever j0 ≤ j. But then, if j0 ≤ j, we have tj (x) = sφ(j) (x) = s(x), and
hence ((tj ), s) ∈ C.
(3) Suppose that (si )i∈I does not converge (C) to s. Then there is x ∈ X
and a cofinal subset J of I such that, whenever j ∈ J , we have sj (x) = s(x).
Let tj = sj for each j ∈ J . Then (tj ) is a subnet of (si ), and clearly no subnet
of (tj ) converges (C) to s.
(4) Suppose that the conditions stated in (4) of Definition 3.1.2 all hold
and that limm limn x(m, n) ≡ s (C), where x : F ' → [X → Y ]. Consider the
net x ◦ r : F → [X → Y ]. Let y ∈ X be arbitrary. Since limm limn x(m, n) ≡
s (C), there is m0 ∈ I such that, for all m ≥ m0 , limn x(m, n) ≡ sm (C)
for some sm ∈ [X → Y ], and limm sm ≡ s (C). Therefore, for m ≥ m0 ,
there is nm ∈ Jm such that x(m, n)( sy) = sm (y) for all n ≥ nm . But for
m ≥ m0 , sm (y) = s(y). Define f ∈ m∈I Jm by setting f (m) = nm ∈ Jm
whenever m ≥ m0 and otherwise letting f (m) ∈ Jm be arbitrary. Suppose
(m, g) ≥ (m0 , f ). Then m ≥ m0 and g ≥ f so that g(m) ≥ f (m) = nm . But
then we have x(m, g(m))(y) = sm (y) = s(y). In other words, (x◦r)(m, g)(y) =
x(m, g(m))(y) = s(y) whenever (m, g) ≥ (m0 , f ). Thus, (x◦r)(y) is eventually
equal to s(y), and hence x ◦ r converges (C) s to s.
Finally, viewing [X → Y ] as the product x∈X Yx , where Yx = Y for each
x ∈ X, then, as is well-known, a net (si ) converges in such a product to s if and
only if si (x) → s(x) in Y for each x, see Theorem A.5.2 (e). But, given that
Y is endowed with the discrete topology, this latter condition si (x) → s(x)
holds if and only if si (x) is eventually equal to s(x), as required. •

Theorem 3.3.1 holds with X taken as BP (= BP,J ), where P is a normal

logic program, and Y taken as any set T of truth values, and in particular it
holds with T taken as T WO. With these choices, we obtain the following re­
sult, which is analogous to Proposition 3.2.10, but applies to normal programs
in general.

3.3.2 Proposition Let P be a normal logic program. Suppose that C is any

convergence class on IP,2 whose elements satisfy the condition stated in The­
orem 3.3.1:
78 Mathematical Aspects of Logic Programming Semantics

if ((Ii ), I) ∈ C, then, for each A ∈ BP , eventually Ii (A) = I(A).

Then, whenever M is an interpretation for P such that ((TPn (M )), I) ∈ C, we

have that I is a model for P .

Proof: By Proposition 2.2.2, it suffices to show that TP (I) [ I. So, sup­

pose therefore that TP (I)(A) = t. Then there is a ground instance A ←
A1 , . . . , An , ¬B1 , . . . , ¬Bm of a clause in P such that I(A1 ∧ . . . ∧ An ∧ ¬B1 ∧
. . . ∧ ¬Bm ) = t. Taking the sequence TPn (M ), we have, by the property
stated in the hypothesis (applied to each literal in the conjunction under
consideration), that eventually TPn (M )(A1 ∧ . . . ∧ An ∧ ¬B1 ∧ . . . ∧ ¬Bm ) =
I(A1 ∧ . . . ∧ An ∧ ¬B1 ∧ . . . ∧ ¬Bm ) = t. Therefore, eventually TPn (M )(A) = t,
and, by the property stated in the hypothesis again, we obtain I(A) = t.
Hence, whenever TP (I)(A) = t, we have I(A) = t. Thus, TP (I) [ I, as
required. •

3.3.3 Remark (1) Theorem 3.3.1 shows that the largest convergence class
C to which Proposition 3.3.2 applies is the convergence class C(Q) deter­
mined by the topology Q. Therefore, Q is the coarsest topology among
the topologies determined by those convergence classes to which Proposi­
tion 3.3.2 can be applied.

(2) In topological terms, Proposition 3.3.2 says that if M is an interpretation

for a normal logic program P such that the sequence (TPn (M )) of iterates
converges in the topology Q to some interpretation I for P , then I is a
model for P .

In fact, we note that the construction of the perfect model semantics for
locally stratified programs P , which we give in Chapter 6, rests on the second
of the facts stated in the previous remark.
Notice that Proposition 3.3.2 holds in any convergence class contained in
C(Q). In other words, it holds for any convergence class determined by a topol­
ogy finer than Q. Furthermore, Q is not the only naturally definable topology
determined by a convergence class for which Proposition 3.3.2 holds. For ex­
ample, if we define limi vi ≡ v (C) to mean that eventually vi = v, we obtain
another natural convergence class which trivially satisfies Proposition 3.3.2,
and this convergence class generates the discrete topology on I(X, T ).
Next, we want to investigate the properties of I(X, T ) when endowed with
the topology Q, and indeed the representation of Q given in Theorem 3.3.1 as
a product space makes this relatively easy.

3.3.4 Theorem Let P be a normal logic program, let J be a preinterpreta­

tion for P with domain D, let X = BP,J , and let T be a truth set endowed
with the discrete topology. Then in the topology Q on I(X, T ) we have the
following results.
 Topology and Logic Programming 79

(a) A net (Ii ) of interpretations converges to an interpretation I if and only

if, for each ground atom A, we have that Ii (A) is eventually equal to I(A).

(b) I(X, T ) is a totally disconnected Hausdorff space.

(c) I(X, T ) is compact if and only if T is a finite set.

(d) I(X, T ) is metrizable13 if and only if D is countable.

(e) I(X, T ) is second countable if and only if D and T are both countable.

(f) Suppose that D is denumerable and that T is finite. Then I(X, T ) is

homeomorphic to the Cantor set in the closed unit interval within the
real line.

Proof: Statement (a) follows immediately from Theorem 3.3.1, and all the
remaining statements follow from general and well-known results concern­
ing product spaces, see the Appendix. Specifically, they can be found in
[Willard, 1970], where unfamiliar terms are also defined, as follows: for (b),
see Page 72, Theorem 13.8, and Page 210, Theorem 29.3; (c) follows from Ty­
chonoff’s theorem (Page 120, Theorem 17.8) and the fact that a discrete space
is compact if and only if it is finite; for (d), see Page 161, Theorem 22.3; for
(e), see Page 108, Theorem 16.2; and finally, for (f), see Page 217, Corollary
30.6. •

Because of Part (f) of Theorem 3.3.4, we refer to the topology Q as the

Cantor topology.
Notice that I(X, T ) is a Hausdorff space in the topology Q, and hence
the limit of any net convergent in Q is unique, see Theorem A.4.2, unlike the
situation in the Scott topology where a convergent net has many limits in
general, as shown by Example 3.2.7.
In the case of two-valued interpretations, we have the following result. It
follows immediately from Part (a) of Theorem 3.3.4 and will be used quite
often later on.

3.3.5 Proposition A net (Ii ) of interpretations in IP,2 converges to I in the

topology Q if and only if whenever A ∈ I, eventually A ∈ Ii , and whenever
A ∈ I, eventually A ∈ Ii . Moreover, the unique limit I coincides with the set
{A ∈ BP | A eventually belongs to Ii }.

The following example illustrates Proposition 3.3.5.

13 Metrics are defined in Section 4.2 and studied extensively in Chapter 4. A topological

space is said to be metrizable if its open sets can be defined in terms of some metric as
discussed in Section 4.1. The representation of Q as a product topology makes it easy to
determine metrics for Q, see [Seda, 1995].
80 Mathematical Aspects of Logic Programming Semantics

3.3.6 Example Consider again the program Even, see Program 2.1.3. To
ease notation here, it will be convenient to denote this program by P and also
to replace the predicate symbol even by p. Thus P denotes the program

p(a) ←
p(s(X)) ← ¬p(X)

We consider the iterates of TP on the interpretation ∅, as follows.

TP0 (∅) = ∅
TP1 (∅) = {p(a), p(s(a)), p(s2 (a)), p(s3 (a)), p(s4 (a)), . . .}
TP2 (∅) = {p(a)}
TP3 (∅) = {p(a), p(s2 (a)), p(s3 (a)), p(s4 (a)), p(s5 (a)), . . .}
TP4 (∅) = {p(a), p(s2 (a))}
TP5 (∅) = {p(a), p(s2 (a)), p(s4 (a)), p(s5 (a)), p(s6 (a)), . . .}
TP6 (∅) = {p(a), p(s2 (a)), p(s4 (a))}
TP7 (∅) = {p(a), p(s2 (a)), p(s4 (a)), p(s6 (a)), p(s7 (a)), . . .}
TP8 (∅) = {p(a), p(s2 (a)), p(s4 (a)), p(s6 (a))}

and so on. On letting In denote TPn (∅) and also letting I denote the set
{p(a), p(s2 (a)), p(s4 (a)), . . .} of “even” natural numbers, we note that the se­
quence (In ) oscillates quite wildly about I. Nevertheless, it is easy to see by
means of Proposition 3.3.5 that (In ) converges in Q to I. Therefore, by Re­
mark 3.3.3, I is a model for P . Indeed, I is a fixed point of TP and is the
unique supported model for P .
In fact, the oscillatory behaviour exhibited in this example in relation to
the single-step operator is typical of programs containing negation. Indeed,
for this example, TP is not Scott continuous, and therefore Theorem 1.1.9 is
not applicable to TP . Hence, the theory developed for the semantics of definite
programs in Chapter 2 is not applicable here either.

3.3.7 Example It is immediate from (a) of Theorem 3.2.6 and Proposi­

tion 3.3.5 that whenever a net (Ii ) converges to I in Q, then it converges
to I in the Scott topology, and this is borne out by Example 3.2.8 and Corol­
lary 3.3.10, just below, which show that the topology Q is finer than the Scott
topology in the case of two-valued interpretations.
On the other hand, the sequence (In ) defined in the first paragraph of (3)
of Example 3.2.7 converges in the Scott topology (to several interpretations),
but does not converge (to anything) in Q.

The point of view that the topology Q is appropriate for studying the
semantics of logic programs with negation is given strong support by examples
such as Example 3.3.6. It is given further support in the most usual case, where
 Topology and Logic Programming 81

the domain of interpretation is countable, as shown in the following example.

In fact, in this next example, we show that a sequence (In ) of two-valued
interpretations converges in Q to a two-valued interpretation I if and only
if the symmetric difference14 In D I of the sets representing In and I can be
made arbitrarily small (in the sense described in Example 3.3.8), and this fact
appears to be in accord with one’s intuition regarding negation. Indeed, the
symmetric difference provides a simple metric for the topology Q, as we see
next.

3.3.8 Example Let P denote a normal logic program, and, to make the
discussion non-trivial, suppose that the underlying first-order language L of
P contains at least one function symbol. Thus, BP is denumerable, and we
can suppose that the elements of BP are given some fixed listing, so that
BP = (A1 , A2 , A3 , . . .), say. (In fact, the exact nature of BP plays no role
here, and we could work equally well over any preinterpretation J for L whose
domain is denumerable and can therefore be listed.) Now o∞ let di be a real
number satisfying 0 < di < 1, for each i, and such that i=1 di = 1; each di
is a weight to be attached to the element Ai of BP . Now define the metric d
on IP by
d(I, I ' ) = /

di ,
Ai ∈IfI
for I, I '
∈ IP . Note that it is routine to check that d does indeed define a
metric on IP , and we show that d generates the topology Q on IP . To do this,
it suffices to show that an arbitrary sequence (In ) converges to I, say, in Q if
and only if it converges to I in the metric d.
Suppose that (In ) is a sequence of interpretations in IP , and In → I in
the metric d. Thus, d(In , I) → 0 as n → ∞. So, given E > 0, there
o is a natural
number n0 such that whenever n ≥ n0 we have d(In , I) = Ai ∈In fI di <
E. Suppose that Aj ∈ I. o Choose E so small that E < dj , and obtain the
corresponding n0 such that Ai ∈In fI di < E whenever n ≥ n0 . Then obviously
dj does not occur in this sum for any n ≥ n0 . In other words, Aj ∈ In ∩ I for
all n ≥ n0 , and so Aj is eventually in In . On the oother hand, suppose that
Aj ∈ I. If Aj belongs to infinitely many In , then Aj ∈In fI di ≥ dj infinitely
often, contradicting d(In , I) → 0. Thus, Aj belongs to only finitely many In ,
and so Aj is eventually not in In . Therefore, by Proposition 3.3.5, convergence
in d implies convergence in Q.
Conversely, suppose In → I in Q. Given E > 0, choose integers n0 so large
that i≥n0 di < E and n'0 ≥ n0 so large that whenever n ≥ n0' , In DI only
o
contains elements Aj with j ≥ n0 or is empty (this situation can be achieved
by finitely many applications of Proposition 3.3.5 since the set {Aj ; j < n0 }
is finite and, in fact, contains n0 − 1 elements). Then, whenever n ≥ n'0 , we
have
d(In , I) = dj ≤ di < E
Aj ∈In fI i≥n0

14 We remind the reader that the symmetric difference of sets A and B is defined by

A6B = (A \ B) ∪ (B \ A).
82 Mathematical Aspects of Logic Programming Semantics

and so In → I in the metric d. Thus, d generates Q, as claimed.

Furthermore, we note that, in particular, the weights di can be taken to
be 21i for each i, in which case the metric d takes the natural form

1
d(I, I ' ) = ,
Ai ∈IfI / 2i
for I, I ' ∈ IP . In any case, if In → I in Q, then
o In → I in d, and hence, given
any E > 0, there is n0 such that d(In , I) = Ai ∈In fI di < E whenever n ≥ n0 ,
and conversely. It is in this sense that the symmetric difference In D I can be
made arbitrarily small if In → I in Q. •

Because Q is a product topology, it is easy to describe the basic open sets

of I(X, T ) in Q as follows (the nature of X is actually irrelevant, although it is
being taken here to be BP,J ). First, given any truth value t ∈ T , the singleton
set {t} is open in T , since T is endowed with the discrete topology. Therefore,
see Section A.5, the basic open sets here are of the form πi−1 1
(ti1 )∩. . .∩πi−1
n
(tin ).
They therefore can be written in the form G(Ai1 , . . . , Ain ; ti1 , . . . , tin ) = {I ∈
I(X, T ) | I(Aij ) = tij for j = 1, . . . , n}, where Ai1 , . . . , Ain are arbitrary, but
fixed, elements of X.
Thus, we have the following result, which describes Q in the familiar terms
of basic open sets.

3.3.9 Proposition With the notation above, the basic open sets in the topol­
ogy Q on the set I(X, T ) take the form G(Ai1 , . . . , Ain ; ti1 , . . . , tin ) = {I ∈
I(X, T ) | I(Aij ) = tij for j = 1, . . . , n}, where Ai1 , . . . , Ain are arbitrary, but
fixed, elements of X and ti1 , . . . , tin are arbitrary, but fixed, elements of T for
j = 1, . . . , n. Furthermore, the subbasic open sets in Q are those basic open
sets G(A; t) determined by taking n = 1 in the set G(Ai1 , . . . , Ain ; ti1 , . . . , tin ).
We denote by G the subbase for Q consisting of the sets G(A; t), where
A ∈ X and t ∈ T .
In particular, the previous proposition has the following corollary when T
is the truth set T W O.

3.3.10 Corollary When T is the truth set T W O, the basic open sets in Q
take the form G(A1 , . . . , An ; B1 , . . . , Bm ) = {I ∈ I(X, T ) | Ai ∈ I, for i =
1, . . . , n, and, for j = 1, . . . , m, Bj ∈ I}, where the Ai and the Bj are fixed,
but arbitrary, elements of X, and n, m ≥ 0. Furthermore, the subbasic open
sets can be described similarly on taking n and m to be at most 1 in the set
G(A1 , . . . , An ; B1 , . . . , Bm ).
Finally, we close this section by noting that a natural question to consider is
that of the continuity of the TP operator relative to the topology Q. However,
as already noted, we defer a discussion of this matter until Chapter 5, see
Theorem 5.4.11, since we treat this question in more generality there in a
 Topology and Logic Programming 83

context within which it naturally arises; in particular, we provide necessary

and sufficient conditions for the continuity of TP in Q to hold. Some results are
also known which ensure discontinuity of TP , see [Seda, 1995], for example,
and we pause briefly to consider an interesting example of this.

3.3.11 Example Consider the program P consisting of the single clause

p ← ¬q(X), whose underlying first-order language L is assumed to contain a
constant symbol o, a function symbol s, and predicate symbols r and t in addi­
tion to the symbols present in P . For each binary sequence a = (an )n∈N (of 0s
and 1s), we form the set Aa = {A1 , A2 , A3 , ...}, where Ai = r(si (o)) if ai = 0
and Ai = t(si (o)) if ai = 1. Finally, let Kn = {q(0), q(s(0)), ..., q(sn (0))} for
each n ∈ N.
Then for each binary sequence a, the sequence of interpretations In =
Aa ∪ Kn converges in Q to the interpretation Ia = Aa ∪ {q(sn (o)) | n ∈ N}
by Theorem 3.3.4. On the other hand, TP (In ) = {p}, whereas TP (Ia ) = ∅.
Hence, TP (In ) does not converge to TP (Ia ) in Q, and so TP is discontinuous
at Ia .
Since we have uncountably many binary sequences a, TP has uncountably
many points of discontinuity in Q.

3.4 Operators on Spaces of Valuations Revisited

Finally, we want to briefly return to the operators defined on I(X, T ),
which were discussed in Section 1.3.4, namely, the operators ¬, ∨, and ∧.
We have already noted in Section 1.3.4 that ¬ is not order continuous and,
hence, not Scott continuous relative to the orderings ≤t , in which f ≤t t. It
is, however, Scott continuous in the orderings ≤k , as we now see. Of course,
one can similarly deal with other connectives such as → and ↔ in the same
way. However, as we have seen earlier, these are usually made to depend on
the three connectives we have already considered and therefore need not be
pursued further.
Our objective here is to examine the continuity of the operators ¬, ∨, and
∧ relative to the Scott and Cantor topologies, and we first deal with the Scott
topology. Again, we concentrate on the truth set FOU R for precisely the same
reasons as stated in Section 1.3.4.

3.4.1 Theorem Let T denote Belnap’s logic FOU R. Then the following
statements hold.
(a) The negation operator ¬ : I(X, T ) → I(X, T ) is continuous in the Scott
topology relative to the knowledge ordering [k , but not relative to the
truth ordering [t . The same statement is true in the case of Kleene’s
strong three-valued logic.
84 Mathematical Aspects of Logic Programming Semantics

(b) Take ≤ to be either ≤k or ≤t on the logic FOU R. Form the domain

I(X, T ) with the corresponding pointwise order [ and the corresponding
product domain I(X, T ) × I(X, T ). Then both ∨ and ∧ are Scott continu­
ous as mappings from I(X, T )×I(X, T ) to I(X, T ). The same statements
are true in the case of classical two-valued logic (where the ordering has
to be ≤t ) and Kleene’s strong three-valued logic.
Proof: For (a), the statements concerning the truth ordering [t have already
been established. To deal with [k , we use the criteria for convergence pre­
sented in Theorem 3.2.6. Let (vi ) be a net converging in the Scott topology to
v in I(X, T ). Suppose that x ∈ (¬v)t . Then (¬v)(x) = t, and hence x ∈ vf .
Since vi → v, we have that eventually x ∈ vif ∪vib , that is, eventually vi (x) = f
or vi (x) = b. But then eventually ¬vi (x) = t or ¬vi (x) = b, and so eventually
x ∈ (¬vi )t ∪ (¬vi )b . The other cases are handled similarly. Thus, the net (¬vi )
converges to ¬v in the Scott topology, as required.
For (b), we establish the result stated concerning ∨, noting that the proof
for ∧ is entirely similar. Now, as is well-known, it suffices to show continuity
in each argument15 of ∨, and, by commutativity, it in fact suffices to show
continuity in one argument, the first, say. So, fix v ∈ I(X, T ), and suppose
that ui → u in the Scott topology on I(X, T ). Let x ∈ X be arbitrary. Then
(u ∨ v)(x) = u(x) ∨ v(x). Since ui → u, we have eventually that u(x) ≤ ui (x)
by Theorem 3.2.6. Therefore, by Proposition 1.3.7, we have eventually that
u(x) ∨ v(x) ≤ ui (x) ∨ v(x), and this suffices, by Theorem 3.2.6, to show that
ui ∨ v → u ∨ v, as required. •

We now turn our attention to these same operators in relation to the

topology Q. Indeed, we close this chapter with the following result.

3.4.2 Theorem Let T denote Belnap’s logic FOU R. Then the following
statements hold.
(a) The negation operator ¬ : I(X, T ) → I(X, T ) is continuous in the topol­
ogy Q. Hence, it is continuous in Q when T denotes either classical two-
valued logic or Kleene’s strong three-valued logic.
(b) Both ∨ and ∧ are continuous as mappings from I(X, T ) × I(X, T ) to
I(X, T ), where I(X, T ) × I(X, T ) is endowed with the product topology
of Q with itself. Hence, the same result holds relative to either classical
two-valued logic or Kleene’s strong three-valued logic.
Proof: For (a), let (vi ) be a net converging to v in I(X, T ) relative to the
topology Q, and let x ∈ X be arbitrary. Then eventually vi (x) = v(x). There­
fore, eventually (¬vi )(x) = (¬v)(x). Therefore, ¬vi → ¬v in Q, and the result
follows.
For (b), let (ui , vi ) → (u, v) in the product topology. Then ui → u in Q
15 See Proposition 2.4 of [Stoltenberg-Hansen et al., 1994].
 Topology and Logic Programming 85

and vi → v in Q. Let x ∈ X be arbitrary. Then there exist i1 and i2 such

that ui (x) = u(x) whenever i ≥ i1 and vi (x) = v(x) whenever i ≥ i2 . By
directedness, there is i3 such that, for i ≥ i3 , we have both ui (x) = u(x) and
vi (x) = v(x). Therefore, whenever i ≥ i3 , we have ui (x) ∨ vi (x) = u(x) ∨ v(x)
and ui (x) ∧ vi (x) = u(x) ∧ v(x). Therefore, ui ∨ vi → u ∨ v and ui ∧ vi → u ∧ v,
as required. •

There are several interesting topics relating to topology and logic pro­
gramming semantics which are examined in the literature on the subject,
but are not pursued here. These include, among other things, the consis­
tency of program completions and of the union of program completions, see
[Batarekh and Subrahmanian, 1989b]; compactness of spaces of models for a
program; and continuity in Q of TP for a normal program P at the point
TP ↓ ω and the coincidence of TP ↓ ω with the greatest fixed point of TP . For
further discussion of all these points and others, see [Seda, 1995].
In conclusion, we note that order is a very satisfactory foundation for the
semantics of procedural and imperative programming languages as exempli­
fied through the denotational semantics approach to programming language
theory. On the other hand, order is not an entirely satisfactory foundation
for the semantics of logic programming languages in the presence of negation,
and yet negation is a natural part of most logics. However, our treatment here
and in later chapters shows that one can consider convergence instead as a
foundation for a unified approach by which one can recover conventional order-
theoretic semantics and at the same time display some important standard
models in logic programming languages as limits of a sequence of iterates.
In addition, convergence conditions involving nets arise very naturally in a
number of areas within theoretical computer science and are simple to state
and to comprehend. Moreover, nets usually give short and technically simple
proofs, as demonstrated in several places in this chapter.
Chapter 4

Fixed-Point Theory for Generalized

Metric Spaces

In Chapters 1 and 2, we gave ample evidence of the fundamental role played

by the Kleene and Knaster-Tarski fixed-point theorems, Theorems 1.1.9 and
1.1.10, in logic programming semantics. Moreover, we have also seen that the
operator TP need not be monotonic for normal programs and, hence, that
the theorems just cited are not generally applicable to TP in this case. It is,
therefore, of interest to consider possible alternatives to Theorems 1.1.9 and
1.1.10, and in this chapter we discuss a number of such fixed-point theorems
and some related results which will be put to use later on.
Almost always, alternatives to the theorems of Kleene and Knaster-Tarski
employ distance functions in their formulations and in their applications.1
Logic programming is no exception to this rule, and we will consider a num­
ber of ways in which distance functions can be naturally introduced into this
subject along with appropriate fixed-point theorems. Part of this process con­
sists of working with quite general distance functions, relaxing in one way
or another the standard axioms for a metric, and establishing corresponding
fixed-point theorems analogous to the Banach contraction mapping theorem.
Nevertheless, the applications we make later and the examples we discuss
show that these general distance functions do quite easily and naturally arise
in logic programming, although applications will be deferred until Chapter 5.
Indeed, Sections 4.1 to 4.7 in this chapter deal with the different generalized
metrics and corresponding fixed-point theorems we develop for single-valued
mappings, while in Section 4.8 we examine the interconnections between the
spaces underlying the various distance functions we study and also discuss a
number of relevant examples. In Sections 4.9 to 4.14, we consider the corre­
sponding results for multivalued mappings. Hence, in summary, this chapter
is a self-contained account of the pure metric fixed-point theory appropriate
to logic programming and also provides the tools needed for the application of
distance functions in developing a unified approach to the fixed-point theory
of very general and significant classes of logic programs in Chapters 5 and
6. In addition, the methods and results discussed in this chapter have poten­
tial applications to a wider spectrum of topics in computer science than just
simply logic programming, but none of these will be pursued here.
1 We refer again to [Kirk and Sims, 2001] as an excellent source of information on fixed-

point theory in general.

87
88 Mathematical Aspects of Logic Programming Semantics

Remark We refer the reader to the paper [Seda and Hitzler, 2010] for a
discussion of many recent and fairly recent applications of distance func­
tions to various parts of computer science. The areas in question range from
conventional semantics ([Arnold and Nivat, 1980b, Arnold and Nivat, 1980a,
Bukatin and Scott, 1997, O’Neill, 1996, Smyth, 1992]) and the study of con­
currency ([de Bakker and de Vink, 1996, Reed et al., 1991]) to domain theory
([Künzi et al., 2006, Krötzsch, 2006, Martin, 2000, Waszkiewicz, 2003]) to in­
formation theory, cognitive processes and unique fingerprinting of time series
([Albeverio et al., 1999, Khrennikov, 1998, Khrennikov, 2004, Murtagh, 2004,
Murtagh, 2005]) to abstract interpretation ([Crazzolara, 1997]) to complex­
ity and and its connections with semantics ([Castro-Company et al., 2007,
Romaguera and Schellekens, 2003, Rodrı́guez-López et al., 2008]), to neural-
symbolic integration ([Bader et al., 2006, Hitzler et al., 2004, Seda, 2006]),
to measuring the distance between programs in software engineering
([Bukatin, 2002, Seda and Lane, 2003]), through to bioinformatics and the
properties of p-adic numbers of DNA sequences and degeneracy of genetic
codes ([Dragovich and Dragovich, 2006, Khrennikov and Kozyrev, 2007]), and
beyond.

4.1 Distance Functions in General

At a completely general level, a distance function d defined on a set X is
simply a mapping d : X × X → A, where A is some suitable set of values
(a distance set or value set), and the distance between x and y is taken to
be the element d(x, y) of A. Second, and again at a completely general level,
the related notion of closeness can be defined by assigning to each element
x of a set X a family Ux of subsets U of X; then y can be thought of as
close to x if y belongs to some element U of Ux . These notions are somewhat
dual to each other, even synonymous, as we shall see shortly. However, the
present level of generality is too high to be useful, and therefore we will impose
a variety of restrictions as we proceed.2 In fact, it is our intention to begin
by briefly considering a uniform, conceptual framework, namely, continuity
spaces,3 within which all the particular distance functions we encounter can be
described. Indeed, this framework is such that the notions of distance function
and closeness are actually dual to each other when the set Ux is taken, for
each x ∈ X, to be the neighbourhood base of x, as defined in the Appendix,

2 [Waszkiewicz,
2002] contains a very general study of spaces based on the notion of
distance function.
3 Our treatment of continuity spaces follows [Kopperman, 1988] closely. We refer also

to [Flagg and Kopperman, 1997] and related papers, where the notion of continuity space
has been developed further in a number of directions, and to [Künzi, 2001] for further
background.
 Fixed-Point Theory for Generalized Metric Spaces 89

see Theorem A.2.5 in particular. This last observation connects topology and
distance in full generality, and this setting, while not the most general to
have been found to be of interest in computer science, as already noted in
Chapter 3, is sufficient for our purposes here. In fact, we shall make no actual
use of continuity spaces and present them purely as a framework within which
to work. However, continuity spaces do provide a smooth transition from the
topology presented in Chapter 3 to the work of this chapter, and indeed they
bridge the two chapters.
Before turning to the details of continuity spaces in general, it will be worth
considering first the familiar case of distance functions d which are metrics,
see Definition 4.2.1 and Remark 4.2.2. In this case, the usual value set A of d is
the interval [0, ∞). Given some real number ε > 0, one defines the (open) ball
Nε (x) of radius ε about a point x ∈ X by setting Nε (x) = {y ∈ X | d(x, y) <
ε}. A subset O of X is then declared to be open if, for each x ∈ X, there
is some ε > 0 such that Nε (x) ⊆ O. It is easy to see that the collection of
such open sets O forms a topology on X. Notice that in defining “open” sets
O here, one can equivalently require Bε/ (x) ⊆ O for suitable ε' > 0, where
Bε (x) = {y ∈ X | d(x, y) ≤ ε} denotes the (closed) ball of radius ε about a
point x ∈ X.
However, it is not true that every topology on X arises thus via a met­
ric d, and, for example, this statement applies to the Scott topology since
this topology in not even T1 in general, see Proposition A.6.5, whereas every
metrizable topology is Hausdorff. Nevertheless, every topology can be gener­
ated by means of a suitable distance function, as already noted, and we next
consider briefly the details of one way of establishing this claim, beginning
with several definitions.

4.1.1 Definition A semigroup is a set A together with an (additive) asso­

ciative binary operation + : A × A → A. If + is also commutative, then the
semigroup is called commutative or Abelian. A semigroup A is called a semi-
group with identity if there exists an element 0 ∈ A, called the identity, such
that 0 + a = a + 0 = a for all a ∈ A. We note that an (additive) Abelian semi-
group with identity is also called a commutative monoid or Abelian monoid.
By an ordered semigroup with identity we mean a semigroup A with 0, say,
on which there is defined an ordering ≤ satisfying: 0 ≤ a for all a ∈ A, and if
a1 ≤ a2 and a'1 ≤ a2' , then a1 + a1' ≤ a2 + a2' for all a1 , a1' , a2 , a2' ∈ A.

4.1.2 Definition A value semigroup A is an additive Abelian semigroup with

identity 0 and absorbing element ∞,4 where ∞ = 0, satisfying the following
axioms.

(1) For all a, b ∈ A, if a + x = b and b + y = a for some x, y ∈ A, then a = b.

(Note that, using this property, we can define a partial order ≤ on A by
setting a ≤ b if and only if b = a + x for some x ∈ A; we call ≤ the partial
4 An element satisfying a + ∞ = ∞ + a = ∞ for all a ∈ A.
90 Mathematical Aspects of Logic Programming Semantics

order induced on A by the operation +. It is immediate that A equipped

with this partial order is an ordered semigroup, as just defined.)

(2) For each a ∈ A, there is a unique b (= a2 ) ∈ A such that b + b = a.

(3) For all a, b ∈ A, the infimum a ∧ b of a and b exists in A relative to the

partial order ≤ defined in (1).

(4) For all a, b, c ∈ A, (a ∧ b) + c = (a + c) ∧ (b + c).

Note that if {(Ai , +i , 0i , ∞i ) | i ∈ I} is a family of value semigroups, then

so is their product (A, +, 0, ∞), where +, 0, and ∞ are defined coordinatewise.

4.1.3 Definition A set P of positives in a value semigroup A is a subset P

of A satisfying the following axioms.

(1) If r, s ∈ P , then r ∧ s ∈ P .

(2) If r ∈ P and r ≤ a, then a ∈ P .

r
(3) If r ∈ P , then 2 ∈ P.

(4) If a ≤ b + r for all r ∈ P , then a ≤ b.

4.1.4 Example The set R of extended real numbers [0, ∞] together with
addition forms a value semigroup, the set (0, ∞] is a set of positives for this
example, and the induced partial order ≤ is the usual one on R.

4.1.5 Definition A continuity space is a quadruple X = (X, d, A, P ), where

X is a non-empty set, A is a value semigroup, P is a set of positives in A,
and d : X × X → A is a function, called a continuity function, satisfying the
following axioms.

(1) For all x ∈ X, d(x, x) = 0.

(2) For all x, y, z ∈ X, d(x, z) ≤ d(x, y) + d(y, z).

Finally, we define the topology generated by a continuity space.

4.1.6 Definition Suppose that X = (X, d, A, P ) is a continuity space. Let

x ∈ X, and let b ∈ P . Then Bb (x) = {y ∈ X | d(x, y) ≤ b} is called the ball
of radius b about x. The topology T (X ) generated by X consists of all those
subsets O of X satisfying the property: if x ∈ O, then Bb (x) ⊆ O for some
b ∈ P.

The main result concerning continuity spaces is the following theorem due
to R. Kopperman [Kopperman, 1988].
 Fixed-Point Theory for Generalized Metric Spaces 91

4.1.7 Theorem Given a continuity space X = (X, d, A, P ), the collection

T (X ) of subsets of X is a topology on X. Conversely, given a topology T on
a set X, there is a continuity space X = (X, d, A, P ) with the property that
T = T (X ).

Given a topology T on X, it is worth noting that the continuity space

X = (X, d, A, P ) with the property that T = T (X ) used in the proof of
Theorem 4.1.7 is obtained, see [Kopperman, 1988], by taking A to be the
product of T copies of R and P to be the product of T copies of (0, ∞]. The
continuity function d is defined coordinatewise by d(x, y)(S) = dS (x, y) for
each S ∈ T , where dS (x, y) = 0 if (x ∈ S implies y ∈ S), and dS (x, y) = q
otherwise, where q is an element of (0, ∞] fixed once and for all.

4.2 Metrics and Their Generalizations

As already noted, it is our intention, with applications in mind, to choose
suitable value sets for distance functions and to impose various useful condi­
tions on the distance functions themselves. We begin by considering the most
familiar of these, where the value set is taken to be the set of non-negative
real numbers.

4.2.1 Definition Let X be a set, and let � : X × X → R+ 0 be a distance

function, where R+
0 denotes the set of non-negative real numbers. We consider
the following conditions on �.

(M1) For all x ∈ X, �(x, x) = 0.

(M2) For all x, y ∈ X, if �(x, y) = �(y, x) = 0, then x = y.

(M3) For all x, y ∈ X, �(x, y) = �(y, x).

(M4) For all x, y, z ∈ X, �(x, y) ≤ �(x, z) + �(z, y).

(M5) For all x, y, z ∈ X, �(x, y) ≤ max{�(x, z), �(z, y)}.

If � satisfies conditions (M1) to (M4), it is called a metric and is called an ultra-

metric if it also satisfies (M5).5 If it satisfies conditions (M1), (M3), and (M4),
it is called a pseudometric. If it satisfies (M2), (M3), and (M4), we will call
it a dislocated metric (or simply a d-metric). Finally, if it satisfies conditions
(M1), (M2), and (M4), it is called a quasimetric. Condition (M4) is usually
5 For elementary properties and notions relating to conventional metrics, such as Cauchy

sequences and completeness, we refer to [Willard, 1970]; these notions will, in any case, be
defined later in this chapter in greater generality.
92 Mathematical Aspects of Logic Programming Semantics

TABLE 4.1: Generalized metrics: Definition 4.2.1.

notion satisfies (M1) (M2) (M3) (M4) (M5)

metric × × × ×
ultrametric × × × (×) ×
pseudometric × × ×
pseudo-ultrametric × × (×) ×
quasimetric × × ×
quasi-ultrametric × × (×) ×
dislocated metric × × ×
dislocated ultrametric × × (×) ×
dislocated quasimetric × ×
dislocated quasi-ultrametric × (×) ×
quasi-pseudometric × ×
quasi-pseudo-ultrametric × (×) ×

called the triangle inequality. Furthermore, if a (pseudo, quasi, d-)metric sat­

isfies the strong triangle inequality (M5), then it is called a (pseudo-, quasi-,
d-)ultrametric. These notions are displayed in Table 4.1, where the symbol
× indicates that the respective condition is satisfied and the symbol (×) in­
dicates that the respective condition is automatically satisfied; for example,
since the condition (M5) implies (M4), any distance function satisfying (M5)
automatically satisfies (M4).

Note that one can take the codomain of � to be [0, ∞] in Definition 4.2.1
rather than R+ 0 . We note then that all the distance functions just considered
in Definition 4.2.1, apart from dislocated metrics, are continuity functions,
as is easily checked. However, even dislocated metrics give rise to topologies,
and essentially the same correspondence between them and topologies holds
between continuity spaces and topologies, as we see later.6 Indeed, each d-
metric gives rise to its associated metric, see Definition 4.8.9, and each d-
generalized ultrametric gives rise to its associated generalized ultrametric, see
Definition 4.8.19.

4.2.2 Remark As far as notation for distance functions is concerned, we will,

generally, although not rigidly, use d and occasionally λ to denote metrics,
ultrametrics, pseudometrics, and quasimetrics, all as just defined; we will use
� and occasionally ρ to denote d-metrics, to denote generalized ultrametrics
as introduced in Section 4.3, and to denote the extensions of these notions
studied in Section 4.4 and beyond. This convention will be employed both
in the context of single-valued mappings and in the context of multivalued
6 See [Hitzler and Seda, 2000] for full details of the topology determined by a d-metric.
 Fixed-Point Theory for Generalized Metric Spaces 93

mappings and is intended to help the reader to remember the nature of the
distance function under consideration at any given time. The one exception to
this occurs in Section 4.8.4, where we encounter two generalized ultrametrics
the second of which is derived from the first. In this instance, we retain the
notation � for the first of these generalized ultrametrics and d for the second;
essentially, the same comment applies to Section 5.1, where the results of
Section 4.8.4 are applied.
The most widely used of the distance functions just defined is that of
metric, and to that extent we regard metric distance functions as basic and
think of departures from them as variants.
The following well-known theorem, usually referred to as the Banach con­
traction mapping theorem, is fundamental in many areas of mathematics. It
is prototypical of a large number of extensions and refinements, including all
those we discuss in this chapter. We give the well-known proof in detail for
later reference.

4.2.3 Theorem (Banach) Let (X, d) be a complete metric space, let 0 ≤

λ < 1, and let f : X → X be a contraction with contractivity factor λ, that
is, f is a (single-valued) function satisfying d(f (x), f (y)) ≤ λd(x, y) for all
x, y ∈ X with x = y. Then f has a unique fixed point, which can be obtained
as the limit of the sequence (f n (y)) for any y ∈ X.
Proof: The proof consists of the following three steps. It is shown that (1)
(f n (y))n≥0 is a Cauchy sequence for all y ∈ X, (2) the limit of this Cauchy
sequence is a fixed point of f , and (3) this fixed point is unique.
(1) Let m, n ∈ N, suppose that m > n, and put k = m − n. Then we obtain

d (f n (y), f m (y)) = d f n (y), f n f k (y) ≤ λn d y, f k (y)

A A bb A b

k−1 k−1
≤ λn d f i (y), f i+1 (y) ≤ λn λi d(y, f (y))
A b
i=0 i=0
k−1 ∞
= λn d(y, f (y)) λi ≤ λn d(y, f (y)) λi
i=0 i=0
λn
= d(y, f (y)).
1−λ
The latter term converges to 0 as n → ∞, and this establishes (1).
(2) Now X is complete, and so (f n (y))n≥0 has a limit x. Thus, we obtain

f (x) = f (lim f n (y)) = lim f n+1 (y) = x

by continuity of f . Therefore, x is a fixed point of f .

(3) Assume now that z is also a fixed point of f . Then d(x, z) =
d(f (x), f (z)) ≤ λd(x, z). Since λ < 1, we obtain d(x, z) = 0, and hence,
by (M2), we have x = z, as required. •
94 Mathematical Aspects of Logic Programming Semantics

Notice that the condition x = y is not actually needed in the statement of

the previous result, but is included for the sake of consistency with what we
want to say next, namely, that it is well-known7 that the requirement λ < 1
cannot be relaxed in general. This can be seen by considering the function
f : R → R defined by
1
x+ x for x ≥ 1,
f (x) =
2 otherwise.

This function satisfies the condition d(f (x), f (y)) < d(x, y) for all x, y ∈ R
with x = y, where d is the usual metric on R, but has no fixed point since
f (x) > x for all x ∈ R. If X is compact, however, the requirement on λ can
be relaxed.

4.2.4 Theorem Let (X, d) be a compact metric space, and let f : X → X

be a function which is strictly contracting, that is, f satisfies d(f (x), f (y)) <
d(x, y) for all x, y ∈ X with x = y. Then f has a unique fixed point.
Proof: The function d(x) = d(x, f (x)) is continuous since f is continuous.
It therefore achieves a minimum m on X. Assume d(x0 ) = m > 0. Then
d(f (x0 )) = d(f (x0 ), f (f (x0 ))) < d(x0 , f (x0 )) = d(x0 ) = m, which is a contra­
diction. Hence, m = 0, and so f has a fixed point.
Assume x and y are fixed points of f and x = y. Then d(x, y) =
d(f (x), f (y)) < d(x, y), which is a contradiction. Therefore, the fixed point
of f is unique. •

There is quite a lot of interest in establishing results which can be viewed

in one way or another as converses of the Banach theorem.8 The following
is such a result. It was originally inspired by certain applications to logic
programming, to be given in Chapters 5 and 6, of the results presented in this
chapter.

4.2.5 Theorem Let (X, τ ) be a T1 topological space, and let f : X → X

be a function which has a unique fixed point a and is such that, for each
x ∈ X, the sequence (f n (x)) converges to a in τ . Then there exists a function
d : X × X → R such that (X, d) is a complete ultrametric space and such that
for all x, y ∈ X we have d(f (x), f (y)) ≤ 12 d(x, y).
Proof: The proof is divided into several steps, numbered consecutively.
(1) Given x ∈ X, we define the set T (x) ⊆ X to be the smallest subset of
X which is closed under the following rules.
(1.1) x ∈ T (x).
7 The results of Section 4.2 can be found in many places including [Kirk and Sims, 2001,

Dugundji and Granas, 1982], for example.

8 A discussion of this question can be found in [Kirk and Sims, 2001] and its references

and in [Istrăţescu, 1981].

 Fixed-Point Theory for Generalized Metric Spaces 95

(1.2) If y ∈ T (x) and f (y) = a, then f (y) ∈ T (x).

(1.3) If y ∈ T (x) and y = a, then f −1 (y) ⊆ T (x).
It is clear that the intersection of the family of all sets closed under these rules
is itself closed under these rules, and hence T (x) exists. Moreover, it is also
clear that each of the sets T (x) is non-empty. Now let T = {T (x) | x ∈ X},
and observe the following facts.
(i) T (a) = {a}. To see this, we note that (1.1), (1.2), and (1.3) are all true
relative to the set {a}. Therefore, by minimality, we have T (a) = {a}.
(ii) If x = a, then a ∈ T (x), and so T (a)∩T (x) = ∅. Hence, either T (a) and
T (x) are equal or they are disjoint. To see this, suppose x = a, and consider
rule (1.3). Clearly, we cannot have a ∈ f −1 (x); otherwise, f (a) = x, and hence
a = x, which is a contradiction. Thus, rules (1.2) and (1.3) applied repeatedly
and starting with x never place a in T (x), and, by minimality, the process
just described generates T (x).
(iii) If T (x) = T (a) and T (y) = T (a), then either T (x) and T (y) are equal
or they are disjoint. To see this, suppose z ∈ T (x)∩T (y). Then the rules (1.1),
(1.2), and (1.3) under repeated application starting with z force T (x) = T (y).
Thus, the collection T is a partition of X.
(2) We next inductively define a mapping l : T → Z ∪ {∞} on each T ∈ T .
(2.1) We set l(a) = ∞, and this defines l on T = T (a). If T = T (a), we
choose an arbitrary x ∈ T and set l(x) = 0 (of course, x = a) and proceed as
follows.
(2.2) For each y ∈ T with f (y) = a and l(y) = k, let l(f (y)) = k + 1.
(2.3) For each y ∈ T with l(y) = k, let l(z) = k − 1 for all z ∈ f −1 (y).
We will henceforth assume that all this is done for every T ∈ T so that l is a
function defined on all of X. It is clear that the mapping l is well-defined since
(X, τ ) is a T1 space.9 For, if there is a cycle in the sequence f n (x) of iterates
for some x ∈ X, then we can arrange for some element y in this sequence to
be frequently not in some neighbourhood of a, using the fact that X is T1 ,
which contradicts the convergence of the sequence f n (x) to a.
(3) Define a mapping ι : Z ∪ {∞} → R by

0 if k = ∞,
ι(k) =
2−k otherwise.

Furthermore, define a mapping δ : X × X → R by

δ(x, y) = max{ι(l(x)), ι(l(y))}

and a mapping d : X × X → R by

δ(x, y) if x = y,
d(x, y) =
0 if x = y.

9 We can weaken the requirement of τ being T by replacing it with the following condi­
1
tion: for every y ∈ X there exists an open neighbourhood U of a with y �∈ U .
96 Mathematical Aspects of Logic Programming Semantics

(4) We show that (X, d) is an ultrametric space. (M1) Let d(x, y) = 0, and
assume that x = y. Then we have δ(x, y) = d(x, y) = 0. Therefore, we obtain
max{ι(l(x)), ι(l(y))} = 0, so ι(l(x)) = ι(l(y)) = 0. Hence, l(x) = l(y) = ∞
and x = y = a by construction of l, which is a contradiction.
(M2) This is true by definition of d.

(M3) This is true by symmetry of δ and, hence, of d.

(M5) Let x, y, z ∈ X. Assume without loss of generality that ι(l(x)) <

ι(l(z)) so that d(x, z) = ι(l(z)). If ι(l(y)) ≤ ι(l(z)), then d(y, z) = ι(l(z)).

If ι(l(y)) > ι(l(z)), then d(y, z) = ι(l(y)) > ι(l(z)). In both cases we get
d(y, z) ≥ d(x, z), as required.
(5) (X, d) is complete as a metric space. In order to show this, let (xn ) be
a Cauchy sequence in X. If (xn ) is eventually constant, then it converges triv­
ially. So now assume that (xn ) is not eventually constant. We proceed to show
that xn converges to a in d, for which it suffices to show that (ι(l(xn )))n∈N
converges to 0. Let ε > 0. Then there exists n0 ∈ N such that for all m, n ≥ n0
we have d(xm , xn ) < ε. In particular, we have d(xm , xn0 ) < ε for all m ≥ n0 ,
and, since (xn ) is not eventually constant, we thus obtain ι(l(xn0 )) < ε and
also ι(l(xm )) < ε for all m ≥ n0 . Since ε was chosen arbitrarily, we see that
(ι(l(xn )))n∈N converges to 0.
(6) We note that for f (x) = a, we have l(f (x)) = l(x) + 1 by definition of
l, and hence ι(l(f (x))) = 12 ι(l(x)).
(7) For all x, y ∈ X, we have that d(f (x), f (y)) ≤ 12 d(x, y). In order
to establish this claim, let x, y ∈ X, and assume without loss of generality
that x = y. Now let d(x, y) = 2−k , say, so that max{ι(l(x)), ι(l(y))} = 2−k .
Then d(f (x), f (y)) = max{ι(l(f (x))), ι(l(f (y)))} = 12 max{ι(l(x)), ι(l(y))} =
1
2 d(x, y), as required. •

It should be noted that Theorem 4.2.5 is not a true converse of the Banach
theorem in that we do not start out with a metrizable space and attempt to
obtain a metric for it relative to which f is a contraction. Thus, Theorem 4.2.5
is quite different from those discussed, for example, in Section 3.6 of the text
[Istrăţescu, 1981], in which a number of converses of the Banach theorem are
considered. Even the result of Bessaga discussed there, which applies to an
abstract set, is very different from ours in that we do not require all iterations
of f to have a unique fixed point, but we do require topological convergence of
the iterates of any point. Indeed, we can only make the following observations
on the relationship between the original topology and the one created by the
metric constructed in the proof of Theorem 4.2.5.

4.2.6 Proposition With the notation of the proof of Theorem 4.2.5, the
following hold.
(a) Any x = a is an isolated point with respect to d, that is, {x} is open and
closed in the topology generated by d.
(b) If (xn ) is a sequence in X which converges in d to some x = a, then the
sequence (xn ) is eventually constant.
 Fixed-Point Theory for Generalized Metric Spaces 97

(c) The metric d does not in general generate τ , but the iterates (f n (x)) of
f converge to a both with respect to τ and with respect to d.
Proof: (a) Let x = a, and let ι(l(x)) = 2−k , say. Then, for any y ∈ X, we have
δ(x, y) ≥ 2−k , and hence, for each y = x, we have d(x, y) ≥ 2−k . Therefore,
{y ∈ X | d(x, y) < 2−k } = {x}, which is consequently open in d. Closedness
is trivial.
(b) In order to see this, it suffices to show that {x} is open with respect to
d for any x = a, which is true by (i) in Step (1) of the proof of Theorem 4.2.5.
(c) Indeed, the topology τ is not in general metrizable. By the proof of
the Banach contraction mapping theorem, (f n (x)) converges to a with respect
to d. Convergence with respect to τ follows from the hypothesis of Theorem
4.2.5. •

4.3 Generalized Ultrametrics

The first generalization of the standard notion of metric which we consider
is actually obtained from Definition 4.2.1 by replacing the codomain of �
(the value set of �), namely, the set R+ 0 of non-negative real numbers, by
an arbitrary partially ordered set rather than by relaxing any axioms. This
leads to the notion of “generalized ultrametric” found in parts of algebra
such as valuation theory and first applied to logic programming semantics
by Prieß-Crampe and Ribenboim. Indeed, the main theorem of this section,
Theorem 4.3.6, is due to Prieß-Crampe and Ribenboim.10

4.3.1 Definition Let X be a set, and let Γ be a partially ordered set with
least element 0. We call (X, � , Γ), or simply (X, �), a generalized ultrametric
space (gum) if � : X × X → Γ is a function such that the following statements
hold for all x, y, z ∈ X and all γ ∈ Γ.
(U1) �(x, x) = 0.
(U2) If �(x, y) = 0, then x = y.
(U3) �(x, y) = �(y, x).
(U4) If �(x, z) ≤ γ and �(z, y) ≤ γ, then �(x, y) ≤ γ.
If � satisfies conditions (U2), (U3), and (U4), but not necessarily (U1), we
call (X, �) a dislocated generalized ultrametric space or simply a d-gum space,
10 The material contained in Section 4.3 up to Theorem 4.3.6 can be found in the following

three papers: [Prieß-Crampe and Ribenboim, 1993, Prieß-Crampe and Ribenboim, 2000a,
Prieß-Crampe and Ribenboim, 2000c].
98 Mathematical Aspects of Logic Programming Semantics

TABLE 4.2: (Dislocated) generalized ultrametrics: Definition 4.3.1.

notion satisfies (U1) (U2) (U3) (U4)

generalized ultrametric (gum) × × × ×
dislocated generalized ultrametric (d-gum) × × ×

see Table 4.2. Condition (U4) will be called the strong triangle inequality for
gums. We note that any gum is a d-gum.

4.3.2 Remark It is clear that every ultrametric space is also a generalized

ultrametric space. However, at the level of generality of the previous definition,
the function � this time is not a continuity function, that is, Γ need not be
a value semigroup. However, in the applications we will actually consider, Γ
will be a value semigroup, and � will indeed be a continuity function, and we
consider this point next.
Let γ > 0 denote an arbitrary ordinal, and denote by Γγ the set {2−α |
α < γ} of symbols 2−α . Then Γγ is totally ordered by 2−α < 2−β if and only
if β < α. Notice that Γγ is really nothing other than γ endowed with the dual
of the usual ordering on ordinals, but it is convenient to use the symbols 2−α
rather than the symbols α to denote typical elements, as will be seen later in
Section 4.8.2 and beyond. Notice also, as is commonly done, that we view an
ordinal γ as the set of all ordinals n such that n ∈ γ, that is, as the set of
ordinals n such that n < γ. Finally, we define the binary operation + on Γγ
by
2−α + 2−β = max{2−α , 2−β }
noting that 2−0 is an absorbing element for this operation. In particular,
applying this construction to the ordinal γ + 1, we note that 2−γ is both the
bottom element of Γγ+1 and the identity element for the operation + defined
on Γγ+1 . Furthermore, 2−γ = 2−0 since γ > 0, where 0 denotes the finite
limit ordinal zero, and we note that we will sometimes also use 0 to denote
2−γ where this does not cause confusion. Then Γγ+1 is a value semigroup in
which a2 = a, where a = 2−α denotes a typical element of Γγ+1 , and moreover,
the partial order induced on Γγ+1 by + coincides with that already defined.
Furthermore, the set {2−α | α < γ} is a set of positives in Γγ+1 . It is the case
Γ = Γγ+1 which is of most interest to us. Therefore, in these cases of most
interest, (X, � , Γ) is a continuity space. In fact, we shall take these points
further later on in this chapter by turning a domain (D, [) into a generalized
ultrametric space, see Sections 4.8.2 and 4.8.3 (and also Section 5.1.1).

The following definitions prepare the way for the main result of this sec­
tion, namely, Theorem 4.3.6, which provides the main fixed-point theorem
applicable to gums. We note that the requisite form of completeness here is
 Fixed-Point Theory for Generalized Metric Spaces 99

that of spherical completeness, defined next, and that the next two definitions
and the following lemma apply to gums as a special case of d-gums.

4.3.3 Definition Let (X, � , Γ) be a d-gum space. For 0 = γ ∈ Γ and x ∈ X,

the set Bγ (x) = {y ∈ X | �(x, y) ≤ γ} is called a (γ-)ball in X with centre or
midpoint x. A d-gum space is called spherically complete if, for�any chain C,
with respect to set-inclusion, of non-empty balls in X we have C = ∅.

The stipulation in the definition of spherical completeness that all balls be

non-empty can be dropped when working in a gum rather than in a d-gum,
since in the former case all balls are clearly non-empty.

4.3.4 Definition Let (X, � , Γ) be a d-gum space, and let f : X → X be a

function.

(1) f is called non-expanding if �(f (x), f (y)) ≤ �(x, y) for all x, y ∈ X.

(2) f is called strictly contracting on orbits 11 if �(f 2 (x), f (x)) < �(f (x), x)
for every x ∈ X with x = f (x).

(3) f is called strictly contracting (on X) if �(f (x), f (y)) < �(x, y) for all
x, y ∈ X with x = y.

We will need the following observations, which are well-known for ordinary
ultrametric spaces.

4.3.5 Lemma Let (X, � , Γ) be a d-gum space. For α, β ∈ Γ and x, y ∈ X,

the following statements hold.

(a) If α ≤ β and Bα (x) ∩ Bβ (y) = ∅, then Bα (x) ⊆ Bβ (y).

(b) If Bα (x) ∩ Bα (y) = ∅, then Bα (x) = Bα (y). In particular, each element

of a ball is also its centre.

(c) B1(x,y) (x) = B1(x,y) (y).

Proof: Let a ∈ Bα (x), and let b ∈ Bα (x) ∩ Bβ (y). Then �(a, x) ≤ α and
�(b, x) ≤ α; hence, �(a, b) ≤ α ≤ β. Since �(b, y) ≤ β, we have �(a, y) ≤ β
and, hence, a ∈ Bβ (y), and this proves the first statement. The second follows
by symmetry and the third by replacing �(x, y) by α and applying (b). •

The following theorem is the analogue of the Banach contraction mapping

theorem applicable to generalized ultrametrics.12 It will be proved later by
virtue of proving the more general Theorem 4.5.1.
11 An orbit of f is a subset of X of the form {f n (x) | n ∈ N} for some x ∈ X.
12 Theorem 4.3.6 can be found in [Prieß-Crampe and Ribenboim, 2000c]. An earlier and
less general version appeared in [Prieß-Crampe, 1990].
100 Mathematical Aspects of Logic Programming Semantics

4.3.6 Theorem (Prieß-Crampe and Ribenboim) Let (X, �, Γ) be a

spherically complete generalized ultrametric space, and let f : X → X be
non-expanding and strictly contracting on orbits. Then f has a fixed point.
Moreover, if f is strictly contracting on X, then f has a unique fixed point.

Note that every compact ultrametric space is spherically complete by the

finite intersection property. The converse is not true: let X be an infinite set,
and let d be the ultrametric defined by setting d(x, y) = 1 if x = y and
taking d(x, x) = 0 for all x ∈ X. Then (X, d) is not compact but is spherically
complete.
The relationship between spherical completeness and completeness is given
by the next proposition.13

4.3.7 Proposition Let (X, d) be an ultrametric space. If X is spherically

complete, then it is complete. The converse does not hold in general.

Proof: Assume that (X, d) is spherically complete and that (xn ) is a Cauchy
sequence in (X, d). Then, for every k ∈ N, there exists a least nk ∈ N such
1
that for all n, m ≥ nk we have d(xn , xm ) ≤
: k . We note that nk increases
with k. Now consider the set of balls B = B k1 (xnk ) | k ∈ N . By (U4), B
is a decreasing chain of balls and has non-empty intersection B by spherical
completeness of (X, d). Let a ∈ B. Then it is easy to see that (xn ) converges
to a. Hence, B = {a} is a one-point set since limits in (X, d) are unique.
Therefore, (X, d) is complete.
In order to show that the converse does not hold in general, define an
ultrametric d on N as follows. For n, m ∈ N, let d(n, m) = 1 + 2− min{m,n}
if n = m, and set d(n, n) = 0 for all n ∈ N. The topology induced by d is
the discrete topology on N, and the Cauchy sequences with respect to d are
exactly the sequences which are eventually constant; hence, (N, d) is complete.
−n
Now consider the chain of balls Bn of the form {m ∈ N | d(m,
� n) ≤ 1 + 2 }.
Then we obtain Bn = {m | m ≥ n} for all n ∈ N. Hence, Bn = ∅. •

Note also that, with the notation from the second part of the proof, the
successor function n �→ n + 1 is strictly contracting, but does not have a fixed
point. By Proposition 4.3.7 and the remarks preceding it, we see that the
notion of spherical completeness is strictly less general than completeness and
is strictly more general than compactness.
Spherical completeness can also be characterized by means of transfinite
sequences, and we consider this next.14
13 Similarstudies of this issue have been undertaken in [Prieß-Crampe, 1990] in the case of
totally ordered distance sets. The topology of generalized ultrametric spaces is investigated
in [Heckmanns, 1996].
14 Here, we follow a line of thought developed in [Prieß-Crampe, 1990], only slightly

changed (the original version was established under the assumption that the distance sets
in question were linearly ordered) and with the proofs adapted to the more general setting.
 Fixed-Point Theory for Generalized Metric Spaces 101

4.3.8 Definition Let (xδ )δ<η be a (possibly transfinite) sequence of elements

of a gum (X, �, Γ). Then (xδ ) is said to be pseudo-convergent if, for all α < β <
γ < η, we have �(xβ , xγ ) < �(xα , xβ ). The transfinite sequence (πδ )δ+1<η with
πδ = �(xδ , xδ+1 ) is then strictly monotonic decreasing. If η is a limit ordinal,
then any x ∈ X with �(x, xδ ) ≤ πδ for all δ < η is called a pseudo-limit of the
transfinite sequence (xδ )δ<η .
The space (X, �, Γ) is called trans-complete if every pseudo-convergent
transfinite sequence (xδ )δ<η , where η is a limit ordinal, has a pseudo-limit in
X.

4.3.9 Proposition Suppose that x is a pseudo-limit of (xδ )δ<η , where η is a

limit ordinal. Then the set of all pseudo-limits of (xδ ) is given by Lim(xδ ) =
{z ∈ X | �(x, z) < πδ for all δ < η}.

Proof: Let z ∈ Lim(xδ ). Since �(z, x) < πδ and �(x, xδ ) ≤ πδ , we obtain

�(z, xδ ) ≤ πδ for all δ, and hence z is a pseudo-limit. Conversely, let z be
a pseudo-limit of (xδ ). Since �(x, xδ+1 ), �(z, xδ+1 ) ≤ πδ+1 for all δ < η, we
obtain �(x, z) ≤ πδ+1 < πδ for all δ < η, as required. •

4.3.10 Proposition A generalized ultrametric space is spherically complete

if and only if it is trans-complete.

Proof: Let X be trans-complete, and let B be a decreasing chain of balls in X.

Without loss of generality, assume that B does not have a minimal element and
is, in fact, strictly decreasing. Then we can select a coinitial subchain (Bδ )δ<η
of B, where η is a limit ordinal, so that (Bδ )δ<η is a transfinite sequence of
balls. Since this transfinite sequence is strictly decreasing, we know that for
every δ there exists xδ ∈ Bδ \ Bδ+1 , and the transfinite sequence (xδ )δ<η is
pseudo-convergent; hence, it has a pseudo-limit x. Since �(x, xδ ) ≤� �(xδ , xδ+1 )
and xδ , xδ+1 ∈ Bδ , we obtain x ∈ Bδ for all δ, and therefore, x ∈ B.
Conversely, let X be spherically complete, and let (xδ ) be pseudo-
convergent. Let πδ = �(xδ , xδ+1 ), and let Bδ = Bπδ (xδ ). For α < β, we
have that xβ ∈ Bα ∩ Bβ , and therefore (Bδ ) is a decreasing chain � of balls by
Lemma 4.3.5. By spherical completeness, there is some x ∈ Bδ , and it is
immediate that x is a pseudo-limit of (xδ ). •

We close this section by considering briefly how pseudo-convergent se­

quences may be generated when the set Γ is linearly ordered. Thus, in what
follows, let (X, �, Γ) be a generalized ultrametric space in which Γ is a linearly
ordered set.

4.3.11 Lemma Let x, y, z ∈ X with �(x, y) < �(y, z). Then �(x, z) = �(y, z).

Proof: We have �(x, z) ≤ max{�(x, y), �(y, z)} ≤ �(y, z) on using the strong
triangle inequality. Now assume �(y, z) ≤ �(x, z). Then, because Γ is linearly
102 Mathematical Aspects of Logic Programming Semantics

ordered, we have �(x, z) < �(y, z), and by the strong triangle inequality again
we obtain �(y, z) ≤ max{�(x, y), �(x, z)} < �(y, z), which is impossible. •

4.3.12 Lemma Let n ≥ 2, and suppose that (x1 , x2 , . . . , xn ) is an n-tuple of

elements of X satisfying �(xi+1 , xi+2 ) < �(xi , xi+1 ) for i = 1, . . . , n − 2. Then
�(x1 , xn ) = �(x1 , x2 ).

Proof: We show by induction on n that the identity �(x1 , x2 ) = �(x1 , xn )

holds. This is trivial for n = 2. So assume n > 2 and that the assertion
holds for n − 1. Then �(x1 , x2 ) = �(x1 , xn−1 ), and consequently �(xn−1 , xn ) <
�(x1 , x2 ) = �(x1 , xn−1 ). So Lemma 4.3.11 applies to the points x1 , xn−1 and
xn and gives �(x1 , xn ) = �(x1 , xn−1 ) = �(x1 , x2 ), as required. •

We can now establish the following result.

4.3.13 Proposition Let (X, �, Γ) be a generalized ultrametric space in which

Γ is a linearly ordered set. Furthermore, let f : X → X be strictly contracting,
let x0 ∈ X, and let xi = f i (x0 ) for all i < ω. Then the sequence (xi )i<ω is
pseudo-convergent.

Proof: Let α < β < γ < ω, and note then that (xα , xα+1 , . . . , xβ , . . . , xγ )
satisfies the hypothesis of Lemma 4.3.12 because f is strictly contracting.
So we obtain �(xα , xβ ) = �(xα , xα+1 ) and �(xβ , xγ ) = �(xβ , xβ+1 ). Thus,
�(xβ , xγ ) = �(xβ , xβ+1 ) < �(xα , xα+1 ) = �(xα , xβ ), as desired. •

4.4 Dislocated Metrics

Dislocated metrics were first studied by S.G. Matthews under the name of
metric domains in the context of Kahn’s dataflow model.15 We proceed now
with the definitions needed for stating the main theorem of Matthews, which,
in fact, is the form of the Banach contraction mapping theorem applicable to
these spaces. Thus, we will define the notions of convergence, Cauchy sequence,
and completeness for dislocated metrics. As it turns out, these notions can be
carried over directly from the corresponding conventional ones.
15 The contents of Section 4.4, including Theorem 4.4.6, can be found in [Matthews, 1986].

Matthews and other authors have argued that the slightly less general notion of (weak )
partial metric is more appropriate than that of dislocated metric from a domain-theoretic
point of view. We refer the reader to [Matthews, 1994, Heckmann, 1999, Waszkiewicz, 2002]
for an account of this, since we have no direct need of it, and indeed dislocated metrics are
well-suited to our purposes.
 Fixed-Point Theory for Generalized Metric Spaces 103

4.4.1 Definition A sequence (xn ) in a d-metric space (X, �) converges with

respect to � or in � if there exists x ∈ X such that �(xn , x) converges to 0 as
n → ∞. In this case, x is called a limit of (xn ) in �.

4.4.2 Proposition Limits in d-metric spaces are unique.

Proof: Let x and y be limits of the sequence (xn ) in a d-metric space (X, �).
By properties (M3) and (M4) of Definition 4.2.1, it follows that �(x, y) ≤
�(xn , x) + �(xn , y) → 0 as n → ∞. Hence, �(x, y) = 0, and by property (M2)
of Definition 4.2.1, we obtain x = y. •

4.4.3 Definition A sequence (xn ) in a d-metric space (X, �) is called a

Cauchy sequence if, for each ε > 0, there exists n0 ∈ N such that for all
m, n ≥ n0 we have �(xm , xn ) < ε.

4.4.4 Proposition Every convergent sequence in a d-metric space is a

Cauchy sequence.

Proof: Let (xn ) be a sequence which converges to some x in a d-metric space

(X, �), and let ε > 0 be chosen arbitrarily. Then there exists n0 ∈ N with
�(xn , x) < 2ε for all n ≥ n0 . For m, n ≥ n0 , we then obtain �(xm , xn ) ≤
�(xm , x) + �(x, xn ) < 2 · 2ε = ε. Hence, (xn ) is a Cauchy sequence. •

4.4.5 Definition A d-metric space (X, �) is called complete if every Cauchy

sequence in X converges with respect to �. Furthermore, a function f : X → X
is called a contraction if there exists 0 ≤ λ < 1 such that �(f (x), f (y)) ≤
λ�(x, y) for all x, y ∈ X.

4.4.6 Theorem (Matthews’ theorem) Let (X, �) be a complete d-metric

space, and let f : X → X be a contraction. Then f has a unique fixed point.

Proof: The proof follows the pattern of the proof of Theorem 4.2.3. Indeed,
Parts (1) and (3) of that proof do not make use of condition (M1) and there­
fore can be carried over literally. Part (2), however, needs to be modified since
we do not have a suitable notion of topological convergence available for dis­
located metric spaces.16 With the notation from the proof of Theorem 4.2.3,
so that x denotes the limit of the Cauchy sequence (f n (y)), we make the

16 It is possible to carry over the complete proof of Theorem 4.2.3, but the constructions

needed are rather involved. Details can be found in [Hitzler and Seda, 2000, Hitzler, 2001];
see also [Hitzler and Seda, 2003].
104 Mathematical Aspects of Logic Programming Semantics

following calculations for all n ∈ N:

�(f (x), x) ≤ �(f (x), f n (x)) + �(f n (x), x)

< � x, f n−1 (x) + �(f n (x), x)
A b

≤ � x, f n−1 (y) + � f n−1 (y), f n−1 (x) + �(f n (x), f n (y))

A b A b

+ �(f n (y), x)
≤ � x, f n−1 (y) + λn−1 �(y, x) + λn �(x, y) + �(f n (y), x).
A b

Since all four terms in the last line converge to 0 as n → ∞, we obtain

�(f (x), x) = 0, and therefore f (x) = x by (M3) and (M2). •

4.5 Dislocated Generalized Ultrametrics

The following theorem gives a partial unification of Matthews’ theo­
rem, Theorem 4.4.6, and the Prieß-Crampe and Ribenboim theorem, The­
orem 4.3.6.17

4.5.1 Theorem Let (X, � , Γ) be a spherically complete d-gum, and let f :

X → X be non-expanding and strictly contracting on orbits. Then f has a
fixed point. If f is strictly contracting on X, then the fixed point is unique.

Proof: Assume that f has no fixed point. Then for all x ∈ X, we have
�(x, f (x)) = 0. We now define the set B by B = {B1(x,f (x)) (x) | x ∈ X}, and
note that each ball in this set is non-empty. We also note that B1(x,f (x)) (x) =
B1(x,f (x)) (f (x)) by Lemma 4.3.5. Now let � C be a maximal chain in B. Since
X is spherically complete, there exists z ∈ C. We show that B1(z,f (z)) (z) ⊆
B1(x,f (x)) for all x ∈ X and, hence, by maximality, that B1(z,f (z)) (z) is the
smallest ball in the chain. Let B1(x,f (x)) (x) ∈ C. Since z ∈ B1(x,f (x)) (x), and
noting our earlier observation that B1(x,f (x)) (x) = B1(x,f (x)) (f (x)) for all x,
we get �(z, x) ≤ �(x, f (x)) and �(z, f (x)) ≤ �(x, f (x)). By non-expansiveness
of f , we get �(f (z), f (x)) ≤ �(z, x) ≤ �(x, f (x)). It follows by (U4) that
�(z, f (z)) ≤ �(x, f (x)) and therefore by Lemma 4.3.5 that B1(z,f (z)) (z) ⊆
B1(x,f (x)) (x) for all x ∈ X, since x was chosen arbitrarily. Now, since f is
strictly contracting on orbits, �(f (z), f 2 (z)) < �(z, f (z)), and therefore z ∈
B1(f (z),f 2 (z)) (f (z)) ⊂ B1(z,f (z)) (f (z)). By Lemma 4.3.5, this is equivalent to
B1(f (z),f 2 (z)) (f (z)) ⊂ B1(z,f (z)) (z), which is a contradiction to the maximality
of C. So f has a fixed point.
17 The proof of Theorem 4.4.6 given here is, in fact, identical to that of Theorem 4.3.6

from [Prieß-Crampe and Ribenboim, 1993].

 Fixed-Point Theory for Generalized Metric Spaces 105

Now let f be strictly contracting on X, and assume that x and y are two
distinct fixed points of f . Then we get �(x, y) = �(f (x), f (y)) < �(x, y), which
is impossible. So the fixed point of f is unique in this case. •

We next give an iterative proof of a special case of Theorem 4.5.1.

4.5.2 Theorem Let (X, �, Γ) be a spherically complete, dislocated general­

ized ultrametric space with Γ = {2−α | α ≤ γ} for some ordinal γ. We order
Γ by 2−α < 2−β if and only if β < α, and denote 2−γ by 0. Thus, Γ is the set
Γγ+1 of Remark 4.3.2. If f : X → X is any strictly contracting function on
X, then f has a unique fixed point.

Proof: Let x ∈ X. Then we have f (x) ∈ f (X) and �(f (x), x) ≤ 2−0 ,
since 2−0 is the maximum possible distance between any two points in X.
Now, �(f (f (x)), f (x)) ≤ 2−1 ≤ 2−0 since f is strictly contracting, and by
(U4), it follows that �(f 2 (x), x) ≤ 2−0 . By the same argument, we obtain
�(f 3 (x), f 2 (x)) ≤ 2−2 ≤ 2−1 , and therefore �(f 3 (x), f (x)) ≤ 2−1 . In fact, an
easy induction argument along these lines shows that �(f n+1 (x), f m (x)) ≤
2−m for m ≤ n. Again by (U4), we obtain that the sequence of balls of the
form B2−n (f n (x)) is a descending chain (with respect to set-inclusion) if n is
increasing and, therefore, has non-zero intersection Bω since X is assumed to
be spherically complete. We therefore conclude that there is xω ∈ Bω with
�(xω , f n (x)) ≤ 2−n for each n ∈ N.
Next, for each n ∈ N, we now argue as follows. Since �(f (xω ), f n+1 (x)) <
�(xω , f n (x)) ≤ 2−n and also �(xω , f n+1 (x)) ≤ 2−(n+1) ≤ 2−n , we therefore
obtain �(f (xω ), xω ) ≤ 2−n . Since this is the case for all n ∈ N, it follows that
�(f (xω ), xω ) ≤ 2−ω .
It is straightforward to cast the above observations into a transfinite in­
duction argument, and we obtain the following construction. Choose x ∈ X
arbitrarily. For each ordinal α ≤ γ, we define f α (x) as follows. If α is a succes­
sor ordinal, then f α (x) = f (f α−1 (x)), as usual. If α is a limit ordinal, then we
choose f α (x) as some xα which has the property that �(xα , f β (x)) ≤ 2−β , not­
ing that the existence of such an xα is guaranteed by spherical completeness
of X.
The resulting transfinite sequence f α (x) has the property that, for all
α ≤ γ, �(f α+1 (x), f α (x)) ≤ 2−α . Consequently, �(f γ+1 (x), f γ (x)) = 2−γ = 0,
and therefore f γ (x) must be a fixed point of f .
Finally, xγ = f γ (x) can be the only fixed point of f . To see this, suppose
y = xγ is another fixed point of f . Then we obtain �(y, xγ ) = �(f (y), f (xγ )) <
�(y, xγ ), from the fact that f is strictly contracting, and this is impossible. •
106 Mathematical Aspects of Logic Programming Semantics

4.6 Quasimetrics
Quasimetrics are a convenient way of reconciling metric and order struc­
tures, see Example 4.6.4. We give the relevant definitions in order to state and
prove the Rutten-Smyth theorem,18 which is the appropriate analogue of the
Banach theorem for quasimetric spaces.

4.6.1 Definition A sequence (xn ) in a quasimetric space (X, d) is a (forward )

Cauchy sequence if, for all ε > 0, there exists n0 ∈ N such that for all n ≥
m ≥ n0 we have d(xm , xn ) < ε. A Cauchy sequence (xn ) converges to x ∈ X
if, for all y ∈ X, d(x, y) = limn→∞ d(xn , y). Finally, X is called CS-complete
if every Cauchy sequence in X converges.
Note that limits of Cauchy sequences in quasimetric spaces are unique.
Given a quasimetric space (X, d), d induces a partial order ≤d on X, called
the partial order induced by d, by setting x ≤d y if and only if d(x, y) = 0.
Furthermore, if (X, d) is a quasimetric space, then (X, d∗ ) is a metric space,
where d∗ (x, y) = max{d(x, y), d(y, x)}, and d∗ is called the metric induced by
d. We call a quasimetric space (X, d) totally bounded if for every ε > 0 there
exists a finite set E ⊆ X such that for every y ∈ X there is an e ∈ E with
d∗ (e, y) < ε.

4.6.2 Definition Let X be a quasimetric space, and let f : X → X be a

function.
(1) f is called CS-continuous if, for all Cauchy sequences (xn ) in X which
converge to x, (f (xn )) is a Cauchy sequence which converges to f (x).
(2) f is called non-expanding if d(f (x), f (y)) ≤ d(x, y) for all x, y ∈ X.
(3) f is called contractive if there exists some c with 0 ≤ c < 1 such that
d(f (x), f (y)) ≤ c · d(x, y) for all x, y ∈ X.
Contractive mappings are not necessarily CS-continuous: consider the set
N ∪ {∞} with the natural order and the distance function
⎧
⎨0 if x ≤ y,
⎪
d(x, y) = 12 if x = 1 and y = 0,
⎪
1 otherwise.
⎩

Then the function f which maps any n ∈ N to 0 and ∞ to 1 is contractive,

but not continuous since limn∈N n = ∞, whereas lim f (n) = 0 = 1 = f (∞).
18 We give Theorem 4.6.3 in the form in which it appears in [Rutten, 1996]; see also the

paper [Rutten, 1995]. A more general version of this result was given in [Smyth, 1987] in
the context of quasi-uniformities.
 Fixed-Point Theory for Generalized Metric Spaces 107

4.6.3 Theorem (Rutten-Smyth) Let (X, d) be a CS-complete quasimetric

space, and let f : X → X be non-expanding.
(a) If f is CS-continuous and there exists x ∈ X with x ≤d f (x), then f has
a fixed point, and this fixed point is least above x with respect to ≤d .
(b) If f is CS-continuous and contractive, then f has a unique fixed point.
Moreover, in both cases the fixed point can be obtained as the limit of the
Cauchy sequence (f n (x)), where in (a) x is the given point, and in (b) x can
be chosen arbitrarily.
Proof: (a) For all n, k ∈ N and k ≥ 1, we have d(f n (x), f n+1 (x)) ≤
ok−1
d(x, f (x)) = 0 and d(f n (x), f n+k (x)) ≤ i=0 d(f
n+i
(x), f n+i+1 (x)) = 0.
n
Hence, (f (x)) is a Cauchy sequence and has a unique limit y, say. Since
f (y) = f (lim f n (x)) = lim f (f n (x)) = lim f n (x) = y, y is a fixed point of f .
Now let z be a fixed point of f with x ≤d z. Then d(y, z) = lim d(f n (x), z) = 0,
since d(f n (x), f n (z)) ≤ d(x, z) = 0. Hence, y ≤d z.
(b) The proof given for Theorem 4.2.3 does not depend on condition (M3)
other than implicitly for deriving continuity of f from the fact that it is a
contraction. Since CS-continuity is a hypothesis in statement (b), the proof
of Theorem 4.2.3 can be carried over by simply replacing “Cauchy sequence”
by “forward Cauchy sequence” and “continuous” by “CS-continuous”, etc. •

4.6.4 Example Let (X, ≤) be a partially ordered set. Define a function d≤

on X × X by
0 if x ≤ y,
d≤ (x, y) =
1 otherwise.
Then it is easily checked that (X, d≤ ) is a quasi-ultrametric space; we call d≤
the discrete quasimetric on X. Note that ≤d≤ and ≤ coincide for a given par­
tial order ≤, and moreover (X, d) is totally bounded if and only if X is finite.
By virtue of this definition and the definition of ≤d for a given quasimetric
d, Part (a) of Theorem 4.6.3 generalizes Kleene’s theorem, Theorem 1.1.9,
and Part (b) of Theorem 4.6.3 generalizes the Banach contraction mapping
theorem, Theorem 4.2.3.19

4.6.5 Example Note that it is easy to see that a sequence (In ) in IP,2 is for­
ward Cauchy relative to the discrete quasimetric d if and only if it is eventually
increasing in the sense that there is a natural number k with the property that
In ⊆ In+1 whenever k ≤ n, see [Seda, 1997, Proposition 1].
Consider the sequence (In ) in the power set P(N) of the natural numbers
determined by setting In = N if n is even and setting In = {0} otherwise.
Then {0} is the greatest limit, gl(In ), of (In ), yet (In ) is not forward Cauchy
19 For further observations on this point, see [Smyth, 1987, Rutten, 1996].
108 Mathematical Aspects of Logic Programming Semantics

in the discrete quasimetric simply because it is not eventually increasing.

Thus, it appears not to be possible to directly characterize the property of
being forward Cauchy relative to the discrete quasimetric in terms of con­
vergence in the Scott topology. This contrasts with the situation where the
(forward) Cauchy sequences relative to the quasimetric determined by a level
mapping, see Definition 4.6.9, can be described in terms of convergence in Q,
see Proposition 4.6.8 and Corollary 4.6.12. •

Using the observations made thus far, it is straightforward to recover the

usual fixed-point semantics of definite logic programs, namely, to recover The­
orem 2.2.3 Part (b) in terms of quasimetrics, by employing Theorem 4.6.3 Part
(a) and the discrete quasimetric on (IP , ⊆). We briefly sketch this next and
refer the reader to [Seda, 1997] for full details.

4.6.6 Example Let P denote an arbitrary definite logic program, and let d
denote the discrete quasimetric defined on the partially ordered set (IP,2 , ⊆).
Then it is shown in [Seda, 1997] that (IP,2 , d) is a CS-complete quasimetric
space and that TP is CS-continuous. We show here that, in fact, TP is non-
expansive and hence that Theorem 4.6.3 is applicable.
Suppose first that d(I1 , I2 ) = 0. Then I1 ⊆ I2 so that TP (I1 ) ⊆ TP (I2 ),
and hence d(TP (I1 ), TP (I2 )) = 0, as required. Next suppose that d(I1 , I2 )
takes value 1. Then immediately d(I1 , I2 ) ≥ d(TP (I1 ), TP (I2 )), as required.
Thus, TP is indeed non-expansive relative to d. We note that, in contrast, TP
is not usually a contraction relative to any metric or quasimetric, since fixed
points of TP are not usually unique. In any event, we are now in a position to
apply Theorem 4.6.3 since we have the following facts.

(1) (IP , d) is a CS-complete quasimetric space.

(2) TP : IP,2 → IP,2 is non-expansive and CS-continuous.

(3) The empty set ∅ is a point in IP,2 such that d(∅, TP (∅)) = 0.

Thus, on applying Theorem 4.6.3 and examining its proof, we conclude that
n
TP has a fixed npoint equal to the greatest limit gl(TP (∅)), and this, in turn,
is equal to TP (∅) = TP ↑ ω, as shown in Chapter 3. Thus, we recover the
classical least fixed point of TP , as required.

We will now use quasimetrics to characterize continuity in the Cantor

topology of the immediate consequence operator for normal logic programs.20

4.6.7 Definition Let (D, [) be a domain, and let r : Dc → N be a function,

20 For more details of the results presented in this section, see [Seda, 1997].
 Fixed-Point Theory for Generalized Metric Spaces 109

called a rank function,21 such that r−1 (n) is a finite set for each n ∈ N. Define
dr : D × D → R by22

dr (x, y) := inf{2−n | (c [ x =⇒ c [ y) for all c ∈ Dc with r(c) < n}.

Then dr is called the quasi-ultrametric induced by r.

It is straightforward to see that (D, dr ) is a quasi-ultrametric space. Fur­
thermore, dr induces the Scott topology on D, and (D, dr ) is totally bounded,
see Proposition 4.6.10.
In order to discuss the relationships between quasimetrics and the Cantor
topology on spaces of interpretations, we need the following proposition.

4.6.8 Proposition Let (X, d) be a totally bounded quasimetric space, and

let (xn ) be a Cauchy sequence in X. Then, for all ε > 0, there exists k ∈ N
such that for all l, m ≥ k, d∗ (xl , xm ) < ε. (A sequence with this property is
usually called a bi-Cauchy sequence.)
Proof: Choose ε > 0 and a finite subset E ⊆ X together with a map h : N →
E such that d∗ (xn , h(n)) < 3ε , using total boundedness. Since (xn ) is a Cauchy
sequence, there exists k0 ∈ N such that for all m ≥ l ≥ k0 , d(xl , xm ) < 3ε .
Now choose k1 ≥ k0 such that for every e ∈ E, the set h−1 (e) ∩ {n | n ≥ k1 }
is either infinite or empty. Choose now l, m ≥ k1 , and let p ≥ l be minimal
such that h(p) = h(m). Then
ε
d(xl , xm ) ≤ d(xl , xp ) + d(xp , h(p)) + d(h(p), xm ) < 3 · = ε,
3
and by symmetry d∗ (xl , xm ) < ε. •

We next define totally bounded quasi-ultrametrics on IP , for a given pro­

gram P , by using level mappings and show that these are closely related to
the Cantor topology Q.

4.6.9 Definition Let P be a normal logic program, and let l : BP → N be a

level mapping for P such that l−1 (n) is finite for every n ∈ N. The mapping
l induces a rank function r : Ic → N defined by

r(I) = max{l(A)},
A∈I

where we take Ic = (IP )c to be the set of all finite subsets of BP . By Definition

4.6.7, r induces a quasi-ultrametric dr on IP .
For a given normal logic program P , we will denote IP,2 by IP for the rest
of this section.
21 The notion of rank function will be given in more generality in Definition 4.8.12.
22 The definition of dr is similar to one made by M.B. Smyth in Example 5 of the paper
[Smyth, 1991].
110 Mathematical Aspects of Logic Programming Semantics

4.6.10 Proposition With the notation established above, (IP , dr ) is a totally

bounded quasi-ultrametric space.
Proof: Choose ε = 2−n , where n ∈ N, and let E be the set of all subsets
of BP , the atoms of which are all of level less than or equal to n. Then E is
finite by our assumption on l. For every I ∈ IP , let e be the restriction of I to
atoms of level less than or equal to n. Then d∗r (e, I) < ε, as is easily verified.
•

We have the following characterization of Cauchy sequences in IP .

4.6.11 Proposition A sequence (In ) in (IP , dr ) is a Cauchy sequence if and

only if for every n ∈ N there exists kn ∈ N such that for all l, m ≥ kn we have
that Il and Im agree on all atoms of level less than n.
Proof: Let (In ) be a Cauchy sequence in IP . Choose n ∈ N, and let ε = 2−n .
Since IP is totally bounded, there exists kn ∈ N such that for all l, m ≥ kn ,
d∗r (Il , Im ) ≤ 2−n . By definition of dr , we obtain that Il and Im agree on all
atoms of level less than n. The converse follows since the argument above
clearly reverses. •

4.6.12 Corollary Let (In ) be a sequence in (IP , dr ). Then (In ) is a Cauchy

sequence if and only if (In ) converges in Q to some I. Moreover, lim In = I,
so (IP , dr ) is complete.
Proof: By Proposition 3.3.5 and the previous proposition, (In ) is a Cauchy
sequence if and only if (In ) converges in Q to some I. It is easily verified that
lim In = I by noting that I = {A ∈ BP | A ∈ In eventually}. It follows that
(IP , dr ) is complete. •

The previous result allows us to characterize CS-continuity in terms of Q.

4.6.13 Proposition Suppose that l : BP → N is a level mapping such that

l−1 (n) is finite for all n. Then the immediate consequence operator TP is
CS-continuous if and only if it is continuous in Q.
Proof: Suppose that TP is CS-continuous and that (In ) is an arbitrary se­
quence in IP which converges in Q to some I ∈ IP . Then (In ) is a Cauchy
sequence, and by Corollary 4.6.12, lim In = I. By CS-continuity of TP , we have
lim TP (In ) = TP (I), and again by Corollary 4.6.12, we have TP (In ) → TP (I)
in Q, as required.
Conversely, suppose TP is continuous in Q and that (In ) is a Cauchy
sequence with lim In = I, say. By Corollary 4.6.12, In → I in Q, and, by
continuity of TP in Q, we get TP (In ) → TP (I), which yields lim TP (In ) =
TP (I), again by Corollary 4.6.12. •

Our next observation shows that non-expansiveness implies CS-continuity.

 Fixed-Point Theory for Generalized Metric Spaces 111

4.6.14 Proposition Let l : BP → N be an arbitrary level mapping satisfying

the condition that l−1 (n) is finite for each n ∈ N. If TP is non-expanding, then
TP is continuous in Q and hence is CS-continuous.
Proof: Let TP be non-expanding, and let (In ) be a Cauchy sequence with
lim In = I. Since TP is non-expansive, we obtain
0 ≤ dr (TP (In ), TP (I)) ≤ dr (In , I) → 0
and
0 ≤ dr (TP (I), TP (In )) ≤ dr (I, In ) → 0
by total boundedness of IP . By definition of dr and Proposition 4.6.11, it
follows that TP (In ) is a Cauchy sequence and, by Proposition 3.3.5 and the
previous inequalities, TP (In ) converges in Q to TP (I). Hence, lim TP (In ) =
TP (I), again by Corollary 4.6.12. •

We close with a brief discussion of several simple examples illustrating the

methods and results of this section as applied to normal logic programs P
relative to TP defined on IP . For full details the reader is again referred to
[Seda, 1997]. Thus, suppose that P is a normal logic program, that dr is the
quasimetric determined by a level mapping l defined on BP and satisfying the
property that l−1 (n) is finite for all n, and that TP is CS-continuous relative
to dr or equivalently that TP is continuous in the topology Q.

4.6.15 Example Consider again the program P of Example 3.2.3

p(a) ←
p(s(X)) ← p(X)
and define l on BP by l(p(sn (a))) = n. Then we see that dr (TP (I1 ), TP (I2 )) ≤
1
2 dr (I1 , I2 ) for all I1 , I2 ∈ IP . Therefore, TP is a contraction and is continuous
in Q and, hence, is CS-continuous. Thus, Theorem 4.6.3 applies and produces
a unique fixed point of TP . Of course, this fixed point coincides with the usual
one produced by considering powers TPn (∅) of ∅.

4.6.16 Example Consider the program P

p(s(X), a) ← p(s(X), a)
with the level mapping l defined on BP by l(p(sn (a), sm (a))) = n + m. Then
it is readily checked that TP is non-expansive (and therefore continuous in
Q), but not contractive, relative to the quasimetric dr determined by l, since
it is easy to find distinct I1 and I2 such that dr (TP (I1 ), TP (I2 )) = dr (I1 , I2 ).
Thus, Theorem 4.6.3 is applicable and, needless to say, produces numerous
fixed points of TP . For this reason, it follows that TP cannot be a contraction
relative to any metric. Thus, the approach to finding fixed points based on
metrics and the Banach contraction mapping theorem fails even for the rather
simple program P .
112 Mathematical Aspects of Logic Programming Semantics

4.6.17 Example Consider again the program P of Example 3.3.6

p(a) ←
p(s(X)) ← ¬p(X)
and note that P is not stratified nor even locally stratified. Define the level
mapping l on BP by l(p(sn (a))) = n for each n. We note that in this
case TP is not non-expansive, for if we take I1 = {p(a), p(s(a))} and I2 =
{p(a), p(s(a)), p(s2 (a))}, then TP (I1 ) = {p(a), p(s3 (a)), p(s4 (a)), p(s5 (a)), . . .}
and TP (I2 ) = {p(a), p(s4 (a)), p(s5 (a)), . . .}. Thus, we have dr (I1 , I2 ) = 0 yet
dr (TP (I1 ), TP (I2 )) = 2−2 , and therefore TP is not non-expansive. Next, con­
sider powers In = TPn (∅), the first few of which, as we have already seen, are as
follows: I1 = BP , I2 = {p(a)}, I3 = BP \ {p(s(a))}, I4 = {p(a), p(s2 (a))}, I5 =
BP \ {p(s(a)), p(s3 (a))}, etc. Then we obtain that dr (In , In+1 ) takes value 0
if n is even and takes value 2−n+1 if n is odd. Therefore, the sequence (In )
is Cauchy and converges to I, say, in Q. By Proposition 3.3.5, we have that
(In ) converges in Q to the set {p(a), p(s2 (a)), p(s4 (a)), . . .}, which therefore
coincides with I. It follows that I is a fixed point of TP , since TP is contin­
uous in Q, and indeed I is the only fixed point of TP , as already noted in
Example 3.3.6.

4.6.18 Example Let P be the program

p(X) ← ¬q(X)
r(s(X)) ← r(X)
q(X) ← q(a), ¬r(X)
which is a slight modification of an example in [Apt et al., 1988, Page 97] and
is stratified. Again, TP is continuous relative to Q, but in this case TP is not
non-expansive for any choice of level mapping l and corresponding quasimetric
dr . To see this, put I1 = {q(a)} and I2 = TP (I1 ) = {p(s(a)), p(s2 (a)), . . .} ∪
{q(a), q(s(a)), . . .}. Then dr (I1 , I2 ) = 0 for any dr simply because I1 ⊆ I2 .
Since TP (I2 ) = {q(a), q(s(a)), . . .}, we must have dr (TP (I1 ), TP (I2 )) > 0 for
any dr or in other words for any choice of l and corresponding dr , so that TP
is never non-expansive. Taking I = {r(a)} and setting In = TPn (I), we have
In = {r(sn (a))}∪{p(a), p(s(a)), p(s2 (a)), . . .}. Clearly, (In ) is Cauchy (for any
choice of level mapping and corresponding dr ), and In converges in Q to the
fixed point {p(a), p(s(a)), p(s2 (a)), . . .}.

4.7 A Hierarchy of Fixed-Point Theorems

For the reader’s convenience, we have collected together in Table 4.3 the
main fixed-point theorems presented in this chapter, at least for single-valued
 Fixed-Point Theory for Generalized Metric Spaces 113

TABLE 4.3: Summary of single-valued fixed-point theorems.

space name of theorem reference number symbol

ω-cpo Kleene 1.1.9 K
cpo Knaster-Tarski 1.1.10 KT
complete metric Banach 4.2.3 B
compact metric — 4.2.4 cp
gum Prieß-Crampe and 4.3.6 PCR
Ribenboim
d-metric Matthews 4.4.6 M
d-gum — 4.5.1 dPCR
quasimetric Rutten-Smyth 4.6.3 RS

cpu\
/// \\
/ \\
// \\
// \\
cp \\
\\
\\
1
K B PCR
//
///
/
1 // 1 1
KT RS M dPCR

FIGURE 4.1: Dependencies between fixed-point theorems from Chapters 1

and 4. The lower a theorem is placed in the diagram, the more general it is.
See Table 4.3 for the abbreviations.

mappings. In fact, we will consider generalizations of several of them to mul­

tivalued mappings as well in the later sections of this chapter. Furthermore,
the dependencies between these theorems are depicted in Figure 4.1, where
the letters abbreviate the theorems as listed in Table 4.3. (The abbreviation
“cpu” represents the statement that strictly contracting functions on compact
ultrametric spaces have unique fixed points, which follows immediately from
Theorem 4.2.4.)
114 Mathematical Aspects of Logic Programming Semantics

4.8 Relationships Between the Various Spaces

We move on next to study the relationships which exist between the var­
ious different spaces we have introduced in this chapter. In particular, we
focus on the representation of certain relationships in terms of others. This
will in some cases lead to alternative proofs of fixed-point theorems we have
already considered. The one exception to this comment is the interplay be­
tween quasimetrics and partial orders. It is clear from the results of Section 4.6
that this interplay is strong. But we will not consider it again other than in
the context of multivalued mappings, see Sections 4.10 and 4.13; see also
[Smyth, 1987, Smyth, 1991, Bonsangue et al., 1996, Rutten, 1996] for further
details.

4.8.1 Metrics and Dislocated Metrics

Our intention here is to establish relationships between metrics and dis­
located metrics. Furthermore, we will examine several methods of obtaining
dislocated metrics from metrics, some of which will be applied later, and we
will show how Matthews’ theorem can be derived from the Banach contraction
mapping theorem.
We begin by noting that if f is a contraction with contractivity factor λ on
a d-metric space (X, �), then we have �(f (x), f (x)) ≤ λ�(x, x) for all x ∈ X.
Furthermore, the property �(x, x) = 0 for all x ∈ X, if � happens to satisfy
this, simply means that the d-metric � is actually a metric. It follows, therefore,
that we are interested in studying the function u1 : X → R associated with
any d-metric �.

4.8.1 Definition Let (X, �) be a d-metric space. We define the function u1 :

X → R by u1 (x) = �(x, x), for all x ∈ X, and call it the dislocation function
of �.

Depending on the context, dislocation functions are sometimes also called

weight functions, see, for example, [Matthews, 1994, Waszkiewicz, 2002].
The following result gives a rather general method by which d-metrics can
be obtained from metrics.

4.8.2 Proposition Let (X, d) be a metric space, let u : X → R+ 0 be a func­

tion, and let T : R+0 × R+
0 → R +
0 be a symmetric function which satisfies the
triangle inequality. Then (X, �), where

�(x, y) = d(x, y) + T (u(x), u(y))

for all x, y ∈ X is a d-metric space, and u1 (x) = T (u(x), u(x)) for all x ∈ X.
In particular, if T (x, x) = x for all x ∈ R+
0 , then u1 = u.
 Fixed-Point Theory for Generalized Metric Spaces 115

Proof: We check the axioms for a d-metric. (M2) If �(x, y) = 0, then

d(x, y) + T (u(x), u(y)) = 0. Hence, d(x, y) = 0, and so x = y. (M3) Obvi­
ous by symmetry of d and T . (M4) Obvious since d and T satisfy the triangle
inequality. •

Completeness also carries over if some continuity conditions are imposed.

4.8.3 Proposition Using the notation of Proposition 4.8.2, let u be contin­

uous as a function from (X, d) to R+0 (where X is endowed with the topology
determined by d, and R+ 0 is endowed with its usual topology), and let T be
continuous as a function from the topological product space (R+ 2 +
0 ) to R0 ,
satisfying the additional property T (x, x) = x for all x. If (X, d) is a complete
metric space, then (X, �) is a complete d-metric space.

Proof: Let (xn ) be a Cauchy sequence in (X, �). Thus, for each ε > 0, there
exists n0 ∈ N such that for all m, n ≥ n0 we have d(xm , xn ) ≤ d(xm , xn ) +
T (u(xm ), u(xn )) = �(xm , xn ) < ε. So (xn ) is also a Cauchy sequence in (X, d)
and therefore has a unique limit x in (X, d). In particular, we have xn → x
in (X, d), and also u(xn ) → u(x) and T (u(xn ), u(x)) → T (u(x), u(x)) = u(x).
We have to show that �(xn , x) converges to 0 as n → ∞. For all n ∈ N, we
obtain �(xn , x) = d(xn , x) + T (u(xn ), u(x)) → u(x) = u1 (x), and it remains
to show that �(x, x) = 0. But this follows from the fact that (xn ) is a Cauchy
sequence, since it implies that u(xn ) = u1 (xn ) = �(xn , xn ) → 0 as n → ∞,
and hence by continuity of u we obtain u(x) = 0. •

An example of a natural function T which satisfies the requirements of

Propositions 4.8.2 and 4.8.3 is
1
T : R+ + +
0 × R0 → R0 : (x, y) �→ (x + y).
2
We discuss a few more examples of d-metrics; they are partly taken from
[Matthews, 1992].

4.8.4 Example Let d be the metric d(x, y) = 12 |x−y| on R+ + +

0 , let u : R0 → R0
1
be the identity function, and define T (x, y) = 2 (x + y). Then � as defined in
Proposition 4.8.2 is a d-metric, and �(x, y) = 12 |x − y| + 12 (x + y) = max{x, y}
for all x, y ∈ R+
0.

4.8.5 Example Let I be the set of all closed intervals in R. Then d : I ×I →

R+
0 defined by
1
d([a, b], [c, d]) = (|a − c| + |b − d|)
2
is a metric on I. Let u : I → R+
0 be defined by

u([a, b]) = b − a
116 Mathematical Aspects of Logic Programming Semantics

and let T be defined as in Example 4.8.4. Then the construction in Proposition

4.8.2 yields a d-metric � such that

�([a, b], [c, d]) = max{b, d} − min{a, c}

for all [a, b], [c, d] ∈ I.

Indeed, we obtain
1 1 1 1
�([a, b], [c, d]) = d([a, b], [c, d]) + b − a + d − c
2 2 2 2
1
= (|b − d| + b + d + |a − c| − a − c)
2
1 1
= (|b − d| + (b + d)) + (|a − c| − (a + c))
2 2
= max{b, d} − min{a, c}.

4.8.6 Example (R+0 , �) is a dislocated metric space, where � is defined by

�(x, y) = x + y.

The following proposition gives an alternative way of obtaining d­

ultrametrics from ultrametrics. We will apply this later in Section 5.1.2.

4.8.7 Proposition Let (X, d) be an ultrametric space, and let u : X → R+

0
be a function. Then (X, �), where

�(x, y) = max{d(x, y), u(x), u(y)}

for all x, y ∈ X, is a d-ultrametric, and �(x, x) = u(x) for all x ∈ X. If

u is continuous as a function on (X, d), then completeness of (X, d) implies
completeness of (X, �).

Proof: (M2) and (M3) are obvious.

(M5) We obtain for all x, y, z ∈ X

�(x, y) = max{d(x, y), u(x), u(y)}

≤ max{d(x, z), d(z, y), u(x), u(y)}
≤ max{d(x, z), u(x), u(z), d(z, y), u(y)}
= max{�(x, z), �(z, y)}.

For completeness, let (xn ) be a Cauchy sequence in (X, �). Then (xn ) is a
Cauchy sequence in (X, d) and converges to some x ∈ X. We then obtain
�(xn , x) = max{d(xn , x), u(xn ), u(x)} → u(x) as n → ∞. As in the proof of
Proposition 4.8.3, we obtain u(x) = 0, and this completes the proof. •

We want to investigate next the relationship between Matthews’ theorem,

Theorem 4.4.6, and the Banach contraction mapping theorem, Theorem 4.2.3.
 Fixed-Point Theory for Generalized Metric Spaces 117

4.8.8 Proposition Let (X, �) be a d-metric space, and define d : X ×X → R

by setting d(x, y) = �(x, y) for x = y and by setting d(x, x) = 0 for all x ∈ X.
Then d is a metric on X.

Proof: We obviously have d(x, x) = 0 for all x ∈ X. If d(x, y) = 0, then either

x = y or �(x, y) = 0, and from the latter we also obtain x = y. Symmetry is
clear. We want to show that d(x, y) ≤ d(x, z) + d(z, y) for all x, y, z ∈ X. If
d(x, z) = �(x, z) and d(z, y) = �(z, y), then the inequality is clear. If d(x, z) =
0, then x = z, and the inequality reduces to d(x, y) ≤ d(x, y), which holds. If
d(z, y) = 0, then z = y, and the inequality reduces to d(x, y) ≤ d(x, y), which
also holds. •

4.8.9 Definition The metric d just defined from the d-metric � is called the
metric associated with �.

Considering Step (3) of the proof of Theorem 4.2.5, we easily verify that δ
is a dislocated ultrametric and note also that d is the metric associated with
δ.
The following proposition allows one to derive from completeness of d, in
general, that � itself is complete.

4.8.10 Proposition Let (X, �) be a d-metric space, and let d denote the
metric associated with �. If the metric d is complete, then so is �. If f is a
contraction relative to �, then f is a contraction relative to d with the same
contractivity factor.

Proof: Suppose that (xn ) is a Cauchy sequence in �. Then for all ε > 0,
there exists n0 such that �(xk , xm ) < ε for all k, m ≥ n0 . Consequently, we
also obtain d(xk , xm ) < ε for all k, m ≥ n0 . Since d is complete, the sequence
(xn ) converges in d to some x, and d(xn , x) → 0 as n → ∞. We show that
�(xn , x) → 0 as n → ∞, and to do this we consider two cases.
Case i. Assume that the sequence (xn ) is such that there exists n0 satisfying
the property that for all m ≥ n0 , we have xm = x. Then �(xm , x) = d(xm , x)
for all m ≥ n0 so that �(xm , x) → 0, and hence �(xn , x) → 0.
Case ii. Assume that there exist infinitely many nk ∈ N such that xnk = x.
Since (xn ) is a Cauchy sequence with respect to �, we obtain �(xnk , x) < ε
for all ε > 0, and so �(x, x) = 0. Hence, �(xn , x) = d(xn , x) for all n ∈ N, and
we obtain that �(xn , x) → 0 as n → ∞, as required.
Let λ ∈ [0, 1) be such that �(f (x), f (y)) ≤ λ�(x, y) for all x, y ∈ X,
and let x, y ∈ X. If f (x) = f (y), then we have d(f (x), f (y)) = 0, hence
d(f (x), f (y)) ≤ λd(x, y). If f (x) = f (y), then x = y, and so d(f (x), f (y)) =
�(f (x), f (y)) ≤ λ�(x, y) = λd(x, y), as required. •
118 Mathematical Aspects of Logic Programming Semantics

4.8.11 Proposition Let (X, �) be a complete d-metric space, and let d de­
note the metric associated with �. Then the metric d is complete. However,
if f is a contraction relative to d, it does not follow that f is necessarily a
contraction relative to �.
Proof: Let (xn ) be a Cauchy sequence in d. If (xn ) eventually becomes con­
stant, then it obviously converges in d. So, assume that this is not the case.
Then the sequence (xn ) must contain infinitely many distinct points; other­
wise, it would not be a Cauchy sequence. We define a subsequence (yn ) of (xn )
which is obtained by removing multiple occurrences of points in (xn ). For each
n ∈ N, let yn = xk , where k is minimal with the property that, for all m < n,
we have ym = xk . Since (yn ) is a subsequence of the Cauchy sequence (xn ),
we see that (yn ) is also a Cauchy sequence relative to d. But, for any two
elements y, z in the sequence (yn ), we have that d(y, z) = �(y, z) by definition
of d. Therefore, (yn ) is a Cauchy sequence in � and, hence, converges in � to
some yω ∈ X. So, (yn ) also converges in d to yω . We show that (xn ) converges
to yω in d. Let ε > 0 be chosen arbitrarily. Since (xn ) is a Cauchy sequence
with respect to d, there exists an index n1 such that d(xk , xm ) < 2ε for all
k, m ≥ n1 . Since (yn ) converges to yω in �, we also know that there is an index
n2 with yn2 = xn3 for some index n3 such that n3 ≥ n1 and d(yn2 , yω ) < 2ε .
For all xn with n ≥ n3 , we then obtain d(xn , yω ) ≤ d(xn , xn3 )+d(xn3 , yω ) < ε,
as required.
Let X = {0, 1}, and define the mapping f : X → X by setting f (x) = 0
for all x ∈ X. Let � be constant and equal to 1. Then � is a complete d-metric,
and f is a contraction relative to d. However, �(f (0), f (1)) = �(0, 0) = �(0, 1),
and so f is not a contraction relative to �. •

The results we have just established put us in a position to prove Matthews’

theorem, Theorem 4.4.6, by using the Banach contraction mapping theorem,
Theorem 4.2.3, and this we do next.

Proof of Theorem 4.4.6 Let (X, �) be a complete d-metric space, and let f
be a contraction relative to �. Let d be the metric associated with �. Then d is
a complete metric, and f is a contraction relative to d. Hence, f has a unique
fixed point by the Banach contraction mapping theorem, Theorem 4.2.3. •

4.8.2 Domains as GUMS

It is our intention here to cast Scott domains into ultrametric spaces, a
construction we will use later in Chapter 5. Usually, domains are endowed with
the Scott topology, see Section A.6. However, as we will see next, domains can
be endowed with the structure of a spherically complete ultrametric space.
This is not something normally considered in domain theory. However, as
already noted at the beginning of the chapter, one of the objectives of the
chapter is to discuss a variety of distance functions, including (generalized)
 Fixed-Point Theory for Generalized Metric Spaces 119

ultrametrics, which have applications both in logic programming and more

generally in theoretical computer science.23

As in Remark 4.3.2, let γ denote an arbitrary ordinal, and let Γγ denote

the set {2−α | α < γ} of symbols 2−α ordered by 2−α < 2−β if and only
if β < α. As already noted, this ordering is, in effect, the dual of the usual
ordering on γ. However, we find it convenient to work with the set Γγ and the
ordering just defined, rather than with the dual ordering on γ, especially in
the context of contraction mappings whose contractivity factor is 12 , see, for
example, Proposition 4.8.17 and particularly Theorem 5.1.6.
We recall that the set of compact elements in a domain D is denoted by
Dc , see Definition 1.1.4.

4.8.12 Definition Let r : Dc → γ be a function, called a rank function, form

Γγ+1 , and denote 2−γ by 0. Define �r : D × D → Γγ+1 by �r (x, y) = inf{2−α |
c [ x if and only if c [ y for every c ∈ Dc with r(c) < α}.

It is readily checked that (D, �r ) is a generalized ultrametric space. We

call �r the generalized ultrametric induced by the rank function r. Indeed, the
intuition behind �r is that two elements x and y of the domain D are close
if they dominate the same compact elements up to a certain rank (and hence
agree in this sense up to this rank); the higher the rank giving agreement, the
closer are x and y. Furthermore, (D, �r ) is spherically complete. The proof
of this claim does not make use of the existence of a bottom element of D,
so this requirement can be omitted. The main idea of the proof is captured
in the next lemma, which shows that chains of balls give rise to chains of
elements in the domain. It depends on the following two elementary facts,
which result immediately from Lemma 4.3.5: (1) if γ ≤ δ and x ∈ Bδ (y), then
Bγ (x) ⊆ Bδ (y), and (2) if Bγ (x) ⊂ Bδ (y), then δ ≤ γ (thus, γ < δ, if Γ is
totally ordered).
In order to simplify notation in the following proofs, we will denote the
ball B2−α (x) by B α (x).

4.8.13 Lemma Let B β (y) and B α (x) be arbitrary balls in (D, �r ). Then the
following statements hold.

(a) For any z ∈ B β (y), we have {c ∈ approx(z) | r(c) < β} = {c ∈ approx(y) |

r(c) < β}.
 
(b) Bβ = {c ∈ approx(y) | r(c) < β} and Bα = {c ∈ approx(x) | r(c) <
α} both exist.
23 This point of view is further developed in a number of papers including the following:

[Kuhlmann, 1999], [Ribenboim, 1996], [Bouamama et al., 2000], [Prieß-Crampe, 1990]; also
the papers [Prieß-Crampe and Ribenboim, 1993], [Prieß-Crampe and Ribenboim, 2000c],
[Prieß-Crampe and Ribenboim, 2000b], [Prieß-Crampe and Ribenboim, 2000a] should be
consulted.
120 Mathematical Aspects of Logic Programming Semantics

(c) Bβ ∈ B β (y) and Bα ∈ B α (x).

(d) Whenever B α (x) ⊆ B β (y), we have Bβ [ Bα .
Proof: (a) Since �r (z, y) ≤ 2−β , the first statement follows immediately from
the definition of �r .
(b) Since the set {c ∈ approx(z) | r(c) < β} is bounded by z, for any z and
β, the second statement follows immediately from the consistent completeness
of D.
(c) By definition, we obtain Bβ [ y. Since Bβ and y agree on all c ∈ Dc
with r(c) < β, the first statement in (c) holds, and the second similarly.
(d) First note that x ∈ B β (y), so that B β (y) = B β (x), and the hypothesis
can be written as B α (x) ⊆ B β (x). We consider two cases.
β
 i. If β ≤ α, then using (a) and
Case  noting again that x ∈ B (y), we
 get
Bβ = {c ∈ approx(y) | r(c) < β} = {c ∈ approx(x) | r(c) < β} [ {c ∈
approx(x) | r(c) < α} = Bα , as required.
Case ii. If α < β, then we cannot have B α (x) ⊂ B β (x), and we therefore
obtain B α (x) = B β (x) and consequently B α (Bβ ) = B β (Bβ ) = B β (Bα ) using
(c). With the argument of Case i and noting that y ∈ B α (x), it follows that
Bα [ Bβ . We want to show that Bα = Bβ . Assume, in fact, that Bα c Bβ .
Since any poin t of a ball is its centre, we can take z = Bβ in (b), twice, to
obtain Bβ = {c ∈ approx(Bβ ) | r(c) < β} and Bα = {c ∈ approx(Bβ ) |
r(c) < α}. Thus, the supposition Bα c Bβ means that {c ∈ approx(Bβ ) |
r(c) < α} c {c ∈ approx(Bβ ) | r(c) < β }. Since {c ∈ approx(Bβ ) | r(c) <
α} ⊆ {c ∈ approx(Bβ ) | r(c) < β }, there must be some d ∈ {c ∈ approx(Bβ ) |
r(c) < β} with d [ {c ∈ approx(Bβ ) | r(c) < α} = Bα . Thus, there is an
element d ∈ Dc with r(d) < β satisfying d [ Bα and d [ Bβ . This contradicts
the fact that �r (Bα , Bβ ) ≤ 2−β . Hence, Bα c Bβ . Since Bα [ Bβ , it follows
that Bα = Bβ and therefore that Bβ [ Bα , as required. •

4.8.14 Theorem The ultrametric space (D, �r ) is spherically complete.

Proof: By the previous lemma, every chain (B α (x α )) of balls in D gives rise
to a chain (Bα ) in D in reverse order. Let B = Bα . Now let B α (xα ) be
an arbitrary ball in the chain. It suffices to show that B ∈ B α (xα ). Since
Bα ∈ B α (xα ), we have �r (Bα , xα ) ≤ 2−α . But �r is a generalized ultrametric,
and so it suffices to show that �r (B, Bα ) ≤ 2−α . For every compact element
c [ Bα , we have c [ B by construction of B. Now let c [ B with c ∈ Dc and
r(c) < α. We have to show that c [ Bα . Since c is compact and c [ B, there
exists Bβ in the chain with c [ Bβ . If B α (xα ) ⊆ B β (xβ ), then Bβ [ Bα by
Lemma 4.8.13, and therefore c [ Bα . If B β (xβ ) ⊂ B α (xα ), then α < β, and,
since c [ Bβ , we see that c is an element of the set {c ∈ approx(xβ ) | r(c) <
α} = {c ∈ approx(xα ) | r(c) < α}. Since Bα is the supremum of the latter
set, we have c [ Bα , as required. •

We will apply this result in Section 5.1.1.

 Fixed-Point Theory for Generalized Metric Spaces 121

4.8.3 GUMS and Chain Complete Posets

In this section, we will invert the point of view of the previous one by
associating a chain-complete partial order with any generalized ultrametric
space (X, �, Γ) whose distance set Γ is an ordinal endowed with, essentially,
the dual ordering as considered in the previous section. Thus, for the duration
of this section, Γ is the set Γγ+1 for some ordinal γ with the ordering described
in Remark 4.3.2. For convenience, we will henceforth call such a generalized
ultrametric space a gum with ordinal distances; recall that we denote 2−γ by
0.
The motivation for adopting our current point of view is to provide a
domain-theoretic proof of the Prieß-Crampe and Ribenboim theorem.24 In
fact, we will prove the Prieß-Crampe and Ribenboim theorem using the
Knaster-Tarski theorem in this special case of gums with ordinal distances.
As a matter of fact, this special case will suffice for all our purposes since, in
applications, all the gums we encounter have ordinal distances, simply because
they arise from level mappings.
Our main technical tool is the space of formal balls associated with a given
metric space, see [Edalat and Heckmann, 1998]. Our first task is to extend this
notion to generalized ultrametrics. 25
Let (X, �, Γ) be a generalized ultrametric space with ordinal distances,
and let B ' X be the set of all pairs (x, α) with x ∈ X and α ∈ Γ. We define
an equivalence relation ∼ on B ' X by setting (x1 , α1 ) ∼ (x2 , α2 ) if and only
if α1 = α2 and �(x1 , x2 ) ≤ α1 . The quotient space BX = B ' X/ ∼ will be
called the space of formal balls associated with (X, �, Γ), and it carries an
ordering [ which is well-defined (on representatives of equivalence classes) by
(x, α) [ (y, β) if and only if �(x, y) ≤ α and β ≤ α. We denote the equivalence
class of (x, α) by [(x, α)], and note of course that the use of the same symbol
[ between equivalence classes and their representatives should not cause any
confusion.

4.8.15 Proposition The set BX is partially ordered by [. Moreover, X is

spherically complete if and only if BX is chain complete.
Proof: That BX is partially ordered by [ is clear.
Let X be spherically complete, and let [(xβ , β)] be an ascending chain in
BX.�Then Bβ (xβ ) is a chain of balls in X with non-empty intersection; let
x ∈ Bβ (xβ ). Then �(xβ , x) ≤ β for all β. Hence, the chain [(xβ , β)] in BX
has [(x, 0)] as an upper bound. Now consider the set A of all α ∈ Γ such
that [(x, α)] is an upper bound of [(xβ , β)]. Since we are working with ordinal
distances only, the set A has a supremum γ, and hence [(x, γ)] is the least
upper bound of the chain [(xβ , β)].
Now suppose BX is chain complete, and let (Bβ (xβ ))β∈Λ be a chain of
24 This approach is inspired by [Edalat and Heckmann, 1998], where the Banach contrac­

tion mapping theorem is derived from Kleene’s theorem.

25 For more details, see [Hitzler and Seda, 2003].
122 Mathematical Aspects of Logic Programming Semantics

ascending chain in BX and has

balls in X, where Λ ⊆ Γ. Then [(xβ , β)] is an �
least upper bound (x, γ), and hence Bγ (x) ⊆ β∈Λ Bβ (xβ ). •

4.8.16 Proposition The function ι : X → BX, where ι(x) = [(x, 0)] for each
x ∈ X, is injective, and ι(X) is the set of all maximal elements of BX.
Proof: Injectivity of ι follows from (U2). The observation that the maximal
elements of BX are exactly the elements of the form [(x, 0)] completes the
proof. •

Now suppose that f is a strictly contracting mapping on a generalized ul­

trametric space (X, �, Γ) with ordinal distances. We use f to induce a mapping
Bf : BX → BX defined by

f (x), 2−(α+1) if 2−α = 0,

A b
Bf (x, 2−α ) =
(f (x), 0) if 2−α = 0.

4.8.17 Proposition If f is strictly contracting, then Bf is monotonic.

Proof: Let (x, 2−α ) [ (y, 2−β ), so that �(x, y) ≤ 2−α and α ≤ β. If 2−α = 0,
there is nothing to show, so assume 2−α = 0. It then remains to show that
�(f (x), f (y)) ≤ 2−(α+1) , and this holds since f is strictly contracting and
because the following Statements (i) and (ii) hold, as is easily verified, namely,
(i) α + 1 ≤ β + 1 if 2−β = 0, and (ii) α + 1 ≤ β if 2−β = 0 and α = β. •

Alternative Proof of Theorem 4.3.6 Let (X, �, Γ) be a spherically com­

plete generalized ultrametric space with ordinal distances, and let f : X → X
be strictly contracting. Then B X is a chain-complete partially ordered set,
and Bf is a monotonic mapping on BX. For B0 ∈ BX, we denote by ↑ B0 the
upper cone of B0 , that is, the set of all B ∈ BX with B0 [ B, as defined in
Section 3.2.
Let x ∈ X be arbitrarily chosen, assume without loss of generality that
x = f (x), and let αbbe an ordinal such that �(x, f (x)) = 2−α . Then
A also−(α+1)
−α
(x, 2 ) [ f (x), 2 , and by monotonicity of Bf we obtain that Bf
maps ↑ [(x, 2−α )] into itself. Since ↑ [(x, 2−α )] is a chain-complete partial order
with bottom element [(x, 2−α )], we obtain by the Knaster-Tarski theorem,
Theorem 1.1.10, that Bf has a least fixed point in ↑ [(x, 2−α )], which we will
denote by B0 .
It is clear by definition of Bf that B0 must be maximal in BX and, hence,
is of the form [(x0 , 0)]. From Bf [(x0 , 0)] = [(x0 , 0)], we obtain f (x0 ) = x0 , so
that x0 is a fixed point of f .
Now assume that y = x0 is another fixed point of f . Then �(x0 , y) =
�(f (x0 ), f (y)) < �(x0 , y) since f is strictly contracting. This contradiction
establishes that f has no fixed point other than x0 . •
 Fixed-Point Theory for Generalized Metric Spaces 123

We note finally that the constructions used for casting domains into gener­
alized ultrametrics as in Section 4.8.2 and for casting generalized ultrametrics
into chain-complete partial orders as in Section 4.8.3 are not inverses of each
other, and the exact relationship between these processes remains to be de­
termined.

4.8.4 GUMS and d-GUMS

We move next to study relationships between gums and d-gums and pro­
vide results somewhat parallel to those of Section 4.8.1, where we contrasted
metrics and d-metrics. Indeed, our main objective here is to investigate the
relationship between the Prieß-Crampe and Ribenboim theorem, Theorem
4.3.6, and its dislocated version, Theorem 4.5.2.

4.8.18 Proposition Let (X, �, Γ) be a dislocated generalized ultrametric

space, and define d : X × X → Γ by setting d(x, y) = �(x, y) for x = y
and setting d(x, x) = 0 for all x ∈ X. Then d is a generalized ultrametric.

Proof: The proof is straightforward following Proposition 4.8.8. •

4.8.19 Definition The generalized ultrametric d just defined from the d-

generalized ultrametric � is called the generalized ultrametric associated with
�.

4.8.20 Proposition Let (X, �, Γ) be a dislocated generalized ultrametric

space, and let d denote the generalized ultrametric associated with �. If d is
spherically complete, then � is spherically complete. If f is strictly contracting
relative to �, then f is strictly contracting relative to d.

Proof: We first show that non-empty balls in � contain all their midpoints.
So let {y | �(x, y) ≤ α} be some non-empty ball in � with midpoint x. Then
there is some z ∈ {y | �(x, y) ≤ α}, and we obtain �(x, x) ≤ �(x, z) by (U4).
Since �(x, z) ≤ α, we have x ∈ {y | �(x, y) ≤ α}. Hence, every non-empty ball
in � is also a ball with respect to d.
Now let B be a chain of non-empty balls in �. Then B is also a chain of
balls in d and has non-empty intersection by spherical completeness of d, as
required.
Let x, y ∈ X with x = y, and assume �(f (x), f (y)) < �(x, y). If f (x) =
f (y), then d(f (x), f (y)) = 0, and hence d(f (x), f (y)) < d(x, y). If f (x) = f (y),
then x = y, and so d(f (x), f (y)) = �(f (x), f (y)) < �(x, y) = d(x, y), as
required. •
124 Mathematical Aspects of Logic Programming Semantics

4.8.21 Proposition Let (X, �, Γ) be a spherically complete dislocated gen­

eralized ultrametric space, and let d denote the generalized ultrametric associ­
ated with �. Then d is spherically complete. However, if f is strictly contract­
ing relative to d, it does not follow that f is necessarily strictly contracting
relative to �.

Proof: Let B be a chain of balls in d. If B contains a ball B = {x} for some

x ∈ X, then x is in the intersection of the chain. So assume that all balls in
B contain more than one point.
Now let Bγ (xm ) = {x | d(x, xm ) ≤ γ} be a ball in B, and let z ∈ Bγ (xm )
with z = xm . Then �(xm , xm ) ≤ �(z, xm ) = d(z, xm ) ≤ γ; hence Bγ (xm ) =
{x | �(x, xm ) ≤ γ}. It follows that B is also a chain of balls in � and, hence,
has non-empty intersection by spherical completeness of �, as required.
Let X = {0, 1}, and define a mapping f : X → X by f (x) = 0 for all
x ∈ X. Let � be constant and equal to 1. Then (X, �, {0, 1}), where 0 < 1,
is a spherically complete d-gum and f is strictly contracting relative to d.
However, �(f (0), f (1)) = �(0, 0) = �(0, 1), and so f is not strictly contracting
relative to �. •

We can now use Theorem 4.3.6 to give an easy proof of Theorem 4.5.2,
as follows. With the notation used in Theorems 4.3.6 and 4.5.2 and using
Proposition 4.8.18, we obtain a generalized ultrametric space (X, d , Γ) which is
spherically complete by Proposition 4.8.21. By Proposition 4.8.20, the function
f is strictly contracting relative to d. Hence, by Theorem 4.3.6, f has a unique
fixed point.
We close this section by giving two constructions of d-gums from gums.

4.8.22 Proposition Let (X, d , Γ) be a generalized ultrametric space with

ordinal distances, and let u : X → Γ be a function. Then the distance function
� defined by
�(x, y) = max{d(x, y), u(x), u(y)}
is a dislocated generalized ultrametric on X.

Proof: (U2) and (U3) are trivial. For (U4), see the proof of Proposition 4.8.7.
•

This result will be applied in Section 5.1.3.

4.8.23 Proposition Let (X, d , Γ) be a generalized ultrametric space with

ordinal distances, let z ∈ X, and define the distance function � by

�(x, y) = max{d(x, z), d(y, z)}.

Then (X, � , Γ) is a spherically complete, dislocated generalized ultrametric

space.
 Fixed-Point Theory for Generalized Metric Spaces 125

Proof: Clearly, � is a d-gum. For spherical completeness, note that every

non-empty ball in (X, �, Γ) contains z, and this suffices. •

This result will be applied in Section 5.1.4.

4.9 Fixed-Point Theory for Multivalued Mappings

We close this chapter with a discussion of multivalued mappings and some
of the fixed-point theorems which are applicable to them.
Let X be a set. Then a multivalued mapping T defined on X is simply a
mapping T : X → P(X) from X to the power set P(X) of X; thus, for each
x ∈ X, T (x) is a subset of X. Furthermore, a fixed point of a multivalued
mapping T is an element x of X such that x ∈ T (x). Such mappings are
important in studying semantics in the presence of non-determinism because
at any step in the execution of a non-deterministic program, there will in
general be many possible successive states, and therefore the informal meaning
of such a program may be taken to be a multivalued mapping defined on the
set X of states the program may assume. These comments apply in particular
to disjunctive logic programs in which the head of a typical program clause
contains a disjunction of several atoms, rather than a single atom, and in
executing such a program a non-deterministic choice has to be made of an
atom in the head of any clause involved in the execution.
Not surprisingly, given their informal meaning, the formal meaning of dis­
junctive programs involves fixed points of multivalued mappings. Therefore, it
is of interest to consider fixed-point theorems in this context and the methods
used to establish them. Again, not surprisingly, the methods normally used
to establish such theorems depend either on order theory or on generalized
metrics of one type or another, and we consider both approaches.
We begin by considering an interesting recent paper by Straccia, Ojeda-
Aciego, and Damásio, see [Straccia et al., 2009], and relating their work to
ours. In this paper, the authors use methods depending on order theory to
establish a number of results guaranteeing the existence of least and greatest
fixed points of a multivalued mapping T : L → P(L), where L is a complete
lattice. In contrast, the methods we will employ mainly depend on the methods
of analysis. Furthermore, as noted below, the results of [Straccia et al., 2009]
are broadly representative of those obtained by order theory. Therefore, it
will help to state a result of [Straccia et al., 2009], which gives a flavour of
its contents and is typical of results obtained in the field by order theory.
However, to do this requires the statement of some preliminary definitions,
but they will be needed in any case as we proceed.
Given the complete lattice (L, ≤) and its power set P(L), we define three
orderings on P(L) familiar in semantics and domain theory, as follows, see
126 Mathematical Aspects of Logic Programming Semantics

[Abramsky and Jung, 1994]. First, the Smyth ordering �S defined by X �S Y

if and only if for each y ∈ Y there exists x ∈ X such that x ≤ y. Second, we
define the Hoare ordering �H by X �H Y if and only if for each x ∈ X there
exists y ∈ Y such that x ≤ y. Finally, we define the Egli–Milner ordering �EM
by X �EM Y if and only if X �S Y and X �H Y . Next, we say that T is
Smyth monotonic or simply S-monotonic if, for all x, y ∈ X satisfying x ≤ y,
we have T (x) �S T (y). The notions of Hoare monotonicity and Egli–Milner
monotonicity are defined similarly.
We are now in a position to present the following result of Straccia, Ojeda-
Aciego, and Damásio, see [Straccia et al., 2009, Prosposition 3.10].

4.9.1 Proposition Let T : L → P(L) be a multivalued mapping, where L is

a complete lattice.

(a) If T is S-monotonic and for all x ∈ L, T (x) has a least element, then T
has a least fixed point.

(b) If T is H-monotonic and for all x ∈ L, T (x) has a greatest element, then
T has a greatest fixed point.

Straccia et al. also introduce a very general class of logic programs P, a

class much more general than conventional disjunctive logic programs, and
proceed to define a multivalued semantic operator TP associated with each
program P in the class in question. On applying their fixed-point theorems,
they establish a one-to-one correspondence between the models of any program
P and the fixed points of TP . All these results are order-theoretic in nature,
although, in summarizing their conclusions, the question of deriving fixed-
point theorems for multivalued mappings using methods from analysis is raised
by the authors, but not taken up in detail.
Thus, we will focus here mainly on those fixed-point theorems for multi­
valued mappings which employ analytical methods and results in their for­
mulation or in their proofs, rather than on results which depend primarily on
order theory. This is partly for the reason stated at the end of the previous
paragraph and partly because the results of [Straccia et al., 2009] largely sub­
sume the order-theoretic results derived by several other contributors to this
subject anyway, except that the latter are usually presented in the context
of complete partial orders rather than in the less general context of complete
lattices employed by Straccia and his co-authors. On the other hand, most
other authors require the condition that the multivalued mapping T is non-
empty in the sense that, for all x ∈ X, we have T (x) = ∅, a condition that
Straccia et al. do not impose. However, despite the opening sentence of this
paragraph, we do wish to consider a result of our own which gives a form, for
multivalued mappings, of the Rutten-Smyth theorem discussed earlier, Theo­
rem 4.6.3, and its role in unifying the order-theoretic and metric approaches to
the fixed-point theory of multivalued mappings, and this of course necessitates
some discussion of order theory.
 Fixed-Point Theory for Generalized Metric Spaces 127

In fact, it turns out that the majority of the fixed-point theorems we have
already considered earlier in this chapter can be directly carried over to the
multivalued setting, and indeed our main task now is to carry out this exten­
sion. Thus, we present multivalued versions of the Knaster-Tarski theorem, the
Banach contraction mapping theorem, the Rutten-Smyth theorem referred to
in the previous paragraph, and Kleene’s theorem. We do not, however, include
any applications of these results here, although they do indeed have a number
of applications to the semantics of (conventional) disjunctive logic programs,
see [Khamsi et al., 1993, Khamsi and Misane, 1998, Hitzler and Seda, 1999c,
Hitzler and Seda, 2002a].

4.10 Partial Orders and Multivalued Mappings

Throughout, T : X → P(X) will denote a multivalued mapping defined
on X. Furthermore, unless stated to the contrary, T will be assumed to be
non-empty.
We begin by discussing a fixed-point theorem first established by M.A.
Khamsi and D. Misane, see [Khamsi and Misane, 1998]. It can be viewed as a
multivalued version of the Knaster-Tarski theorem, Theorem 1.1.10; a multi­
valued version of Kleene’s theorem, Theorem 1.1.9 will be presented in Section
4.13.

4.10.1 Definition Let T : X → P(X) be a multivalued mapping defined on

X. An orbit of T is a net (xi )i∈I in X, where I denotes an ordinal, such that
xi+1 ∈ T (xi ) for all i ∈ I. An orbit (xi )i∈I of T is called an ω-orbit if I is
the first limit ordinal, ω. An orbit (xi )i∈I of T will be said to be eventually
constant if there is a tail (xi )i0 ≤i of (xi )i∈I which is constant in that xi = xj
for all i, j ∈ I satisfying i0 ≤ i, j.

If T : X → P(X) is a multivalued mapping and x is a fixed point of T ,

then we obtain an orbit of T which is eventually constant by setting x =
x0 = x1 = x2 . . .. Conversely, suppose that (xi )i∈I is an orbit of T with the
property that xi+1 = xi for all i ∈ I satisfying i0 ≤ i, for some ordinal i0 ∈ I.
Then xi0 = xi0 +1 ∈ T (xi0 ), and we have a fixed point xi0 of T . Thus, having
a fixed point and having an orbit which is eventually constant are essentially
equivalent conditions on T .

4.10.2 Definition Suppose that T is a multivalued mapping defined on a

partially ordered set X. An orbit (xi )i∈I of T is said to be increasing if we have
xi ≤ xj for all i, j ∈ I satisfying i ≤ j and is said to be eventually increasing
 (xi )i∈I of T
if some tail of the orbit is increasing. Finally, an increasing orbit
is said to be tight if, for all limit ordinals j ∈ I, we have xj = {xi | i < j}.
128 Mathematical Aspects of Logic Programming Semantics

Suppose that (xi )i∈I is an increasing orbit of T and that j ∈ I is a limit

 xj+1 is an element of T (xj ) such that xi ≤ xj+1 for all i < j, and
ordinal. Then
of course {xi | i < j} ≤ xj ≤ xj+1 if the supremum exists. In particular,
any increasing orbit (xi )i∈I which is tight (if such exists) must satisfy the
following condition: for any limit ordinal j, there exists x = xj+1 such that

x ∈ T( {xi | i < j}) and {xi | i < j} ≤ x. (4.1)

This condition is a slight variant of a condition which was identified by Khamsi

and Misane as a sufficient condition for the existence of fixed points of Hoare
monotonic multivalued mappings. In fact, the following result was established
by them, see [Khamsi and Misane, 1998], except that it was formulated for
decreasing orbits and infima, and we have chosen to work with the dual notions
instead to be consistent with the form of Kleene’s theorem we give later.

4.10.3 Theorem (Knaster-Tarski multivalued) Suppose that X is a

complete partial order and that T : X → P(X) is a multivalued mapping
which is non-empty, Hoare monotonic, and satisfies condition (4.1). Then T
has a fixed point.

We omit details of the proof of this result except to observe that, start­
ing with the bottom element x0 = ⊥ of X, the condition (4.1) permits the
construction, transfinitely, of a tight orbit (xi ) of T . Since this can be carried
out for ordinals whose underlying cardinal is greater than that of X, we are
forced to conclude that (xi ) is eventually constant and therefore that T has a
fixed point.
 
Noting that {xi | i < j} = {xi+1 | i < j}, one can  view condition (4.1)
schematically as the statement “ {T (xi ) | i < j} ≤ T ( {xi | i < j})”, and it
can therefore be thought of as a rather natural, weak continuity condition on
T which is automatically satisfied by any monotonic single-valued mapping T
on a complete partial order. The question of when the orbit constructed in the
previous paragraph becomes constant in not more than ω steps is a question
of continuity, as in the single-valued version, and will be taken up in Section
4.13.
Theorem 4.10.3 was established by Khamsi and Misane in order to
show the existence of (consistent) answer sets for a class of disjunctive
logic programs called signed programs. We have shown elsewhere, see
[Hitzler and Seda, 1999c], that it sometimes is necessary to work transfinitely
in practice, a point which justifies the name “Knaster-Tarski theorem” applied
to Theorem 4.10.3.
Thus, in summary, Hoare monotonicity of T together with (4.1) gives,
for multivalued mappings, an exact analogue of the fixed-point theory for
monotonic single-valued mappings due to Knaster-Tarski. Moreover, there are
applications of it to the semantics of disjunctive logic programs which parallel
those made in the standard, non-disjunctive case.
 Fixed-Point Theory for Generalized Metric Spaces 129

4.11 Metrics and Multivalued Mappings

We discuss here a result established by M.A. Khamsi, V. Kreinovich, and
D. Misane, see [Khamsi et al., 1993], which is a multivalued version of the
Banach contraction mapping theorem, Theorem 4.2.3.

4.11.1 Definition Let (X, d) be a metric space. A multivalued mapping T :

X → P(X) is called a contraction if there exists a non-negative real number
λ < 1 such that for every x ∈ X, for every y ∈ X, and for all a ∈ T (x) there
exists b ∈ T (y) such that d(a, b) ≤ λd(x, y).

The result we wish to state is as follows; a proof of it will be given in

Section 4.13.

4.11.2 Theorem (Banach multivalued) Let X be a complete metric

space, and suppose that T is a multivalued contraction on X such that, for
every x ∈ X, the set T (x) is closed and non-empty. Then T has a fixed point.

This theorem was also established with a specific objective in view, namely,
to show the existence of answer sets for disjunctive logic programs which are
countably stratified, again see [Khamsi et al., 1993].

4.12 Generalized Ultrametrics and Multivalued

Mappings
We next turn our attention to multivalued versions of the Prieß-Crampe
and Ribenboim theorem, Theorem 4.3.6.

4.12.1 Definition Let (X, �, Γ) be a generalized ultrametric space. A mul­

tivalued mapping T defined on X is called strictly contracting (on X) (re­
spectively, non-expanding (on X)) if, for all x, y ∈ X with x = y and for
every a ∈ T (x), there exists an element b ∈ T (y) such that �(a, b) < �(x, y)
(�(a, b) ≤ �(x, y)). Furthermore, the mapping T is called strictly contracting
on orbits if, for every x ∈ X and for every a ∈ T (x) with a = x, there exists
an element b ∈ T (a) such that �(a, b) < �(a, x).

For T : X → P(X), let Πx = {�(x, y) | y ∈ T (x)}, and, for a subset Δ ⊆ Γ,

denote by Min Δ the set of all minimal elements of Δ.
Note that these definitions collapse to those already considered for single-
valued mappings if, in fact, T is single valued, meaning that T (x) is a singleton
set for each x ∈ X.
130 Mathematical Aspects of Logic Programming Semantics

The following theorem was proved by Prieß-Crampe and Ribenboim, see

[Prieß-Crampe and Ribenboim, 2000c], and is a multivalued version of Theo­
rem 4.3.6.

4.12.2 Theorem (Prieß-Crampe and Ribenboim) Let (X, �, Γ) be a

spherically complete, generalized ultrametric space, and let T : X → P(X)
be non-empty, non-expanding, and strictly contracting on orbits. In addition,
assume that for every x ∈ X, Min Πx is finite and that every element of Πx
has a lower bound in Min Πx . Then T has a fixed point.

This result has several corollaries, due to Prieß-Crampe and Ribenboim,

see [Prieß-Crampe and Ribenboim, 2000c], both for multivalued mappings
and for single-valued mappings, and we state two of these next for complete­
ness. Theorem 4.12.2 has been applied to establish the stable model semantics
for disjunctive logic programs, see [Seda and Hitzler, 2010]. Note that Theo­
rem 4.12.4 is a slight extension of Theorem 4.3.6.

4.12.3 Theorem Let (X, �, Γ) be spherically complete, and let Γ be narrow,

that is, such that every trivially ordered subset of Γ is finite. Let f : X → P(X)
be non-empty, strictly contracting on orbits and such that f (x) is spherically
complete for every x ∈ X. Then f has a fixed point.

4.12.4 Theorem Let (X, �, Γ) be a spherically complete, generalized ultra-

metric space, and let f : X → X be non-expanding on X. Then either f
has a fixed point or there exists a ball Bπ (z) such that �(y, f (y)) = π for all
y ∈ Bπ (z). If, in addition, f is strictly contracting on orbits, then f has a
fixed point. Finally, this fixed point is unique if f is strictly contracting on X.

The following ideas are closely related to the notion of value semigroup
given in Definition 4.1.2 and were considered by Khamsi, Kreinovich, and
Misane in the context of the stable model semantics for disjunctive logic pro­
grams, see [Khamsi et al., 1993]. We show that, in fact, these notions basically
coincide with those from generalized ultrametric theory.

4.12.5 Definition Let V be an ordered Abelian semigroup with 0, and let

X be an arbitrary set. A g-metric on X is a mapping ρ : X × X → V which
satisfies the following conditions for all x, y, z ∈ X.

(1) ρ(x, y) = 0 if and only if x = y.

(2) ρ(x, y) = ρ(y, x).

(3) ρ(x, y) ≤ ρ(x, z) + ρ(z, y).

A pair (X, ρ) consisting of a set X and a g-metric ρ on X is called a g-metric

space.
 Fixed-Point Theory for Generalized Metric Spaces 131

In fact, g-metrics were called generalized metrics by Khamsi, Kreinovich,

and Misane, but we have changed the terminology since the term “generalized
metric” is, of course, already used differently by us. Actually, we will not work
with g-metrics in general since the closely related generalized ultrametrics will
suffice for our purposes. Indeed, we consider this relationship next, and we
begin by recalling the observations we made in Remark 4.3.2. Thus, let V
denote the set of all expressions of the type 0 or 2−α , where α > 0 is an
ordinal. An order is defined on V by: 0 ≤ v for every v ∈ V , and 2−α ≤ 2−β if
and only if β ≤ α. As a semigroup operation u + v, we will use the maximum
max(u, v). It will be convenient to write 12 2−α = 2−(α+1) .
The following definition is due to Khamsi, Kreinovich, and Misane, see
[Khamsi et al., 1993].

4.12.6 Definition Assume that α is either a countable ordinal or ω1 , the first

uncountable ordinal, and that v = (vβ )β<α is a decreasing family of elements
of V . Let X be a g-metric space relative to V , and let (xβ )β<α be a family of
elements of X.
(1) (xβ ) is said to v-cluster to x ∈ X if, for all β, we have ρ(xβ , x) < vβ
whenever β < α.
(2) (xβ ) is said to be v-Cauchy if, for all β and γ, we have ρ(xβ , xγ ) < vβ
whenever β < γ < α.
(3) X is said to be v-complete or just complete if, for every v, every v-Cauchy
family v-clusters to some element in X.
(4) A set Y ⊆ X will be called v-complete or just complete if, for every v,
whenever a v-Cauchy family consists of elements of Y , it v-clusters to
some element of Y .
A close relationship exists between the notion of completeness for g-metrics
and the notion of trans-completeness, Definition 4.3.8, for generalized ultra­
metrics. Indeed, we show that these notions coincide by showing equivalence
between completeness for g-metrics and spherical completeness for generalized
ultrametrics, see Proposition 4.3.10.

4.12.7 Definition A multivalued mapping T : X → P(X) is called a 12 ­

A b

contraction if, for every x ∈ X, for every y ∈ X, and for every a ∈ T (x),
there exists b ∈ T (y) such that ρ(a, b) ≤ 12 ρ(x, y).
The following theorem was proved by Khamsi, Kreinovich, and Misane in
[Khamsi et al., 1993].

4.12.8
A1b Theorem Let X be a complete g-metric space, let T be a multivalued
2 -contraction defined on X such that T (x) is not empty for some x ∈ X
(so that T is not identically empty), and suppose that for every x ∈ X the set
T (x) is complete. Then T has a fixed point.
132 Mathematical Aspects of Logic Programming Semantics

We next present some results relating those just given to the notion of
spherical completeness we discussed earlier. Indeed, we show that if (X, ρ) is
a g-metric space with respect to V as given in Definition 4.3.2, then ρ is a
generalized ultrametric space, and vice-versa.

4.12.9 Proposition Let (X, ρ) be a complete g-metric space with respect to

V . Then X is spherically complete as an ultrametric space.
A b
Proof: Let B = Bvβ (xβ ) β<α be a decreasing chain of balls in X, and without
loss of generality assume that � it is strictly decreasing and that α is a limit
ordinal. We have to show that B = ∅. Let v = (vβ )β . Since B is a chain, it
is easy to see that (xβ+1 )β is v-Cauchy and therefore, by completeness of X,
(xβ+1 ) v-clusters to some x ∈ X. By definition, this means that ρ(xβ+1 � , x) <
vβ and therefore that x ∈ Bvβ (xβ+1 ) = Bvβ (xβ ) for all β. Thus, x ∈ B. •

In the opposite direction, we have the following result.

4.12.10 Proposition Let (X, ρ, V ) be a spherically complete, generalized

ultrametric space. Then X is complete as a g-metric space.
Proof: Let v= (vβ ) be a decreasing family of elements of V which is, without
loss of generality, strictly decreasing, and let (xβ ) be v-Cauchy.
u For v ∈ v,
for example, v = 2−α , let v ' denote 2−(α+1) . Then B = Bvβ/ (xβ ) is a
β
decreasing chain of balls �in X. By spherical completeness, it has non-empty
intersection. Choose x ∈ B. Then for all β we obtain ρ (xβ , x) ≤ vβ' < vβ ,
and so (xβ ) v-clusters to x. •

This means, by virtue of Theorem 4.12.2, that we can reformulate the

assumptions in Theorem 4.12.8 and thereby obtain the following result, which,
in fact, is a special case of a theorem of Prieß-Crampe and Ribenboim, see
[Prieß-Crampe and Ribenboim, 2000c, (3.4)].

4.12.11 Theorem Let X be a spherically complete, generalized ultrametric

space (with respect to V ), and let T be a multivalued, non-empty, and strictly
contracting mapping defined on X such that T (x) is spherically complete for
all x ∈ X. Then T has a fixed point.

4.13 Quasimetrics and Multivalued Mappings

We move next to study a multivalued version of the Rutten-Smyth the­
orem, Theorem 4.6.3. As a consequence, we obtain a multivalued version of
Kleene’s theorem, Theorem 1.1.9.26
26 For further details, see [Hitzler and Seda, 1999c].
 Fixed-Point Theory for Generalized Metric Spaces 133

4.13.1 Definition Let (X, d) be a quasimetric space. A multivalued mapping

T : X → P(X) is called a contraction if there is a λ ∈ [0, 1) such that, for
all x, y ∈ X and for all a ∈ T (x), there exists b ∈ T (y) satisfying d(a, b) ≤
λd(x, y). We say that T is non-expanding if, for all x, y ∈ X and for all
a ∈ T (x), there exists b ∈ T (y) satisfying d(a, b) ≤ d(x, y).

Again, these definitions are clearly extensions of the corresponding defi­

nitions made for single-valued mappings and indeed collapse to them in the
case where T is single valued. An obvious and natural definition of continuity
of T is the following: for every Cauchy sequence (xn ) in X with limit x and
for every choice of yn ∈ T (xn ), we have that (yn ) is a Cauchy sequence and
lim yn ∈ T (x). In fact, the weaker definition following, which is implied by the
one just given, suffices for our purposes and will be used throughout.

4.13.2 Definition Let T : X → P(X) be a multivalued mapping defined on

a quasimetric space (X, d). We say that T is continuous if we have lim xn ∈
T (lim xn ) for every ω-orbit (xn ) of T which is a Cauchy sequence.

Once more, this definition collapses to a natural one if T is single valued.

In fact, if T is single valued, it simply states the condition that lim T (xn ) =
lim xn+1 = lim xn = T (lim xn ) for every ω-orbit which is a Cauchy se­
quence, which is a weaker condition than that of CS-continuity as in Definition
4.6.2(1).
Finally, if (X, d) is a quasimetric space, we define the associated partial
order ≤d on X by x ≤d y if and only if d(x, y) = 0, see Section 4.6.
The main result of this section is the following theorem, generalizing the
Rutten-Smyth theorem we gave earlier, Theorem 4.6.3.

4.13.3 Theorem (Rutten-Smyth multivalued) Let (X, d) be a CS-

complete quasimetric space, and let T : X → P(X) denote a non-empty
and continuous multivalued mapping on X. Then T has a fixed point if either
of the following two conditions holds.

(a) T is a contraction.

(b) T is non-expanding, and there is x0 ∈ X and x1 ∈ T (x0 ) such that

d(x0 , x1 ) = 0, that is, x0 ≤d x1 .

Proof: (a) Let x0 ∈ X. Since T (x0 ) = ∅, we can choose x1 ∈ T (x0 ). Since T is

a contraction, there is x2 ∈ T (x1 ) such that d(x1 , x2 ) ≤ λd(x0 , x1 ). Applying
this argument repeatedly, we obtain a sequence (xn ) such that for all n ≥ 0
we have xn+1 ∈ T (xn ) and d(xn+1 , xn+2 ) ≤ λd(xn , xn+1 ). Thus, (xn ) is an
ω-orbit. Using the triangle inequality, we obtain
m−1 m−1

d(xn , xn+m ) ≤ d(xn+i , xn+i+1 ) ≤ λn+i d(x0 , x1 ).

i=0 i=0
134 Mathematical Aspects of Logic Programming Semantics
n
λ
Since the last summation here is dominated by 1−λ d(x0 , x1 ), we see that (xn )
is a (forward) Cauchy sequence in X and therefore is an ω-orbit of T which
is Cauchy. Since X is complete, (xn ) has a limit xω . Now, by continuity of T ,
we obtain xω ∈ T (xω ), and xω is a fixed point of T , as required.
(b) Let x0 ∈ X and x1 ∈ T (x0 ) satisfy d(x0 , x1 ) = 0. Since T is non-
expanding, there is x2 ∈ T (x1 ) with d(x1 , x2 ) ≤ d(x0 , x1 ) = 0. Inductively,
we obtain a sequence (xn ) such that xn+1 ∈ T (xn ) and d(xn , xn+k ) ≤
ok−1
i=0 d(xn+i , xn+i+1 ) = 0. Hence, (xn ) is an orbit of T which is forward
Cauchy and therefore has a limit xω . By continuity of T again, we see that
xω is a fixed point of T . •

The proof given here of Part (a) of Theorem 4.13.3 is, up to the last
step, exactly the same as the first half of the proof of the multivalued Ba­
nach contraction mapping theorem, Theorem 4.11.2, established by Khamsi,
Kreinovich, and Misane, except that we are working with a quasimetric rather
than with a metric and therefore care needs to be taken that no use is made
of symmetry. On the other hand, the proof we give next of Theorem 4.11.2,
which roughly corresponds to the second half of the proof given by Khamsi,
Kreinovich, and Misane, is shorter and technically somewhat simpler than the
proof given by them.

Proof of Theorem 4.11.2 We show that the condition that T (x) is closed
for every x together with that of T being a contraction implies that T is
continuous, and the result then follows from Part (a) of Theorem 4.13.3.
First note that (X, d) being a complete metric space means that (X, d) is
complete as a quasimetric space, and obviously T satisfies Part (a) of Theorem
4.13.3. Now suppose that (xn ) is an orbit of T which is a forward Cauchy
sequence and, hence, a Cauchy sequence; we want to show that xω ∈ T (xω ),
where xω is the limit of (xn ).
Since T is a contraction, for every n there exists yn ∈ T (xω ) such that
d(xn+1 , yn ) ≤ λd(xn , xω ). Therefore, d(yn , xω ) ≤ d(yn , xn+1 ) + d(xn+1 , xω ) ≤
λd(xn , xω ) + d(xn+1 , xω ). Hence, we have yn → xω . But each yn ∈ T (xω ),
and T (x) is closed for every x. Consequently, the limit xω of the sequence yn
also belongs to T (xω ). So, xω ∈ T (xω ), and it follows that T is continuous, as
required. •

Thus, Theorem 4.13.3 contains, as a consequence, the multivalued Banach

contraction mapping theorem, Theorem 4.11.2, discussed earlier. It also con­
tains a natural extension of Kleene’s theorem to multivalued mappings, The­
orem 4.13.6 below, as we show next. Thus, Theorem 4.13.3 gives a unification
of metric and order-theoretic notions in direct analogy with the corresponding
unification given, in the single-valued case, by Theorem 4.6.3.
In order to proceed, we make some preliminary and elementary observa­
tions, as follows, concerning partially ordered sets and the quasimetrics they
carry, see Section 4.6. The proofs are straightforward and are omitted.
 Fixed-Point Theory for Generalized Metric Spaces 135

4.13.4 Proposition Let (X, ≤) be a partial order, and let (X, d) denote the
associated quasimetric space, so that d = d≤ as in Section 4.6. Then the
following hold.
(a) A non-empty multivalued mapping T : X → P(X) is Hoare monotonic if
and only if it is non-expanding.
(b) A sequence (xn ) in X is eventually increasing in (X, ≤) if and only if it
is a Cauchy sequence in (X, d).
(c) The partially ordered set (X, ≤) is ω-complete if and only if (X, d) is
complete as a quasimetric space. Furthermore, in the presence of either
form of completeness, the limit of any Cauchy sequence is the least upper
bound of any increasing tail of the sequence.
Notice that neither Part (c) of this result nor the next definition assumes
the presence of a bottom element.

4.13.5 Definition Let the partial order (X, ≤) be ω-complete, and let T :
X → P(X) be a non-empty multivalued mapping on X. We say that T is
ω-continuous if T is Hoare monotonic,
 and for
 any ω-orbit (xn ) of T which
is eventually increasing, we have (xn ) ∈ T ( (xn )), where the supremum is
taken over any increasing tail of (xn ).
We obtain finally the following form of Kleene’s theorem for multivalued
mappings as an easy corollary of our Theorem 4.13.3. This theorem has been
applied by the present authors to find answer sets for certain classes of dis­
junctive logic programs, see [Hitzler and Seda, 1999c].

4.13.6 Theorem (Kleene multivalued) Let (X, ≤) be an ω-complete par­

tial order (with bottom element), and let T : X → P(X) be a non-empty,
ω-continuous multivalued mapping on X. Then T has a fixed point.
Proof: Since (X, ≤) is ω-complete, the associated quasimetric space (X, d)
(with d = d≤ as in Section 4.6) is complete by Proposition 4.13.4. Furthermore,
T is Hoare monotonic, since it is ω-continuous and is therefore non-expanding
by Proposition 4.13.4 again. On taking x0 = ⊥ and x1 ∈ T (x0 ) arbitrarily,
we have x0 and x1 satisfying d(x0 , x1 ) = 0. The result will therefore follow
from Part (b) of Theorem 4.13.3 as soon as we have established that T is
continuous in the sense of Definition 4.13.2.
Let (xn ) be any ω-orbit of T which is a Cauchy sequence.
 Then(xn ) is
eventually increasing, and, by ω-continuity of T , we have (xn ) ∈ T ( (xn )),
where the supremum is taken over any increasing tail of (xn ). In other words,
we have lim xn ∈ T (lim xn ), and hence we have the continuity of T that we
require. •

Kleene’s theorem for single-valued mappings T asserts that the fixed point
produced by the usual proof is the least fixed point of T . This assertion does
136 Mathematical Aspects of Logic Programming Semantics

not immediately carry over to the case of multivalued mappings T without

additional assumptions. One such simple, though rather strong, condition is
the following: for each x ∈ X, assume that T (x) has a least element Mx and
that Mx ≤ My whenever x ≤ y. To see that this suffices, suppose that x is
any fixed point of T , and construct the orbit (xn ) of T by setting x0 = ⊥ and
xn+1 = Mxn for each n. Then (xn ) converges to a fixed point x. Noting that
⊥ ≤ x and that Mx ≤ x, we see that xn ≤ x for all n. Hence, x ≤ x.

4.14 An Alternative to Multivalued Mappings

As already noted earlier, multivalued mappings arise naturally as semantic
operators in relation to disjunctive logic programs. However, William Rounds
and Guo-Qiang Zhang have shown that the use of multivalued mappings in this
context can be avoided by employing single-valued mappings defined on power
domains instead (we refer the reader to [Stoltenberg-Hansen et al., 1994] for
details of power domains). In fact, this observation is part of a considerable
programme of research undertaken by the authors just mentioned in the appli­
cation of domain theory to logic programming. Since their work complements
that presented here, we intend to make a few remarks about a couple of aspects
of it, and it is convenient to do this next.
The starting point of this programme of work is the observation that do­
mains and logic are strongly related [Zhang, 1991] and that this relationship
may be used as a foundation for a theory of logic programming based on do­
main theory. In [Zhang and Rounds, 1997a, Zhang and Rounds, 1997b] and
[Rounds and Zhang, 2001], Rounds and Zhang use power domains to develop
a domain-theoretic view of default logic, which they call power defaults. In­
deed, in this framework logic programs can be viewed in a rather simple way
as default theories in the sense of [Reiter, 1980]. Default theories constitute
an important formalism in the area of non-monotonic reasoning, and we re­
fer the reader to [Bidoit and Froideveaux, 1991, Gelfond and Lifschitz, 1991,
Bochman, 1995, Lifschitz, 2001] and to the references contained in these pa­
pers for an interesting discussion of the relationship between default logic and
logic programs. Indeed, from this point of view, the standard models of a dis­
junctive program, such as the stable model, correspond to extensions in default
logic: in short, truth in a model corresponds to default theorem. Furthermore,
Rounds and Zhang [Rounds and Zhang, 2001, Zhang and Rounds, 1997a,
Zhang and Rounds, 1997b, Zhang and Rounds, 2001] study a version of de­
fault reasoning from the domain-theoretic point of view. In particular, they
focus on the Smyth powerdomain by making the observation that the Smyth
powerdomain can be used to model non-monotonicity. This results, for ex­
ample, in the implementation of a non-monotonic reasoning system, see
[Klavins et al., 1998], which bears a significant relationship to other answer
 Fixed-Point Theory for Generalized Metric Spaces 137

set programming systems which have been investigated with implementation

in mind, see [Lifschitz, 1999, Marek and Truszczyński, 1999]. In addition, in
[Rounds and Zhang, 2001, Zhang and Rounds, 2001], Rounds and Zhang in­
troduced a domain-theoretic framework for the study of the semantics of logic
programming, both procedural and non-procedural, including an abstract res­
olution rule, together with a treatment of negation, which is not negation as (fi­
nite) failure, however. [Hitzler, 2003a, Hitzler and Wendt, 2003, Hitzler, 2004,
Hitzler and Krötzsch, 2006] further expand on some aspects of the work of
Rounds and Zhang and in particular relate it to Formal Concept Analysis
[Ganter and Wille, 1999] and to answer set programming.27
Of course, the monotonicity notions for multivalued mappings used mainly
in this chapter correspond to orderings encountered in power domains. In par­
ticular, this applies to Hoare montonicity and to Smyth monotonicity. With
this and the comments of the previous paragraph in mind, we note finally that
in Chapter 6 of [Zhang and Rounds, 2001], a treatment is given of the seman­
tics of disjunctive logic programs (as considered here) with the same overall
objective as our own. The treatment is based on the Smyth powerdomain
again. One important feature of this power-domain approach is that by using
the right domain, the concept of multivalued function is avoided and continu­
ity can always be taken to be Scott continuity. Thus, in conclusion, we note
that overall the developments just described appear to hold out, in particular,
the possibility of a domain-theoretic treatment of the declarative semantics
of negation in logic programming and therefore to bring logic programming
semantics more fully into the realm of domain theory, and vice-versa.

27 See Footnote 3 in the Introduction.

Chapter 5

Supported Model Semantics

Among the various semantics for normal logic programs discussed in Chapter
2, the supported model semantics, whether in two-valued or in three-valued
form, is most fundamental: stable and perfect models are two-valued sup­
ported models; and well-founded and weakly perfect models are three-valued
supported models. Furthermore, as shown in Theorem 2.6.14, if the Fitting
model for a program P is total, then P has a unique two-valued supported
model which coincides with the unique model assigned to P by the Fitting,
the well-founded, the weakly perfect, and the stable semantics: the semantics
in this case is unambiguous.
Programs which have unique supported models together with those which
have total Fitting models can therefore be considered to be of fundamental
importance for understanding logic programming semantics as presented in
Chapter 2. The former, namely, programs with unique supported models, are
called by us uniquely determined , while we call the latter Φ-accessible pro­
grams. We know from Theorem 2.6.14, as just noted, that every Φ-accessible
program has a unique supported model. The converse, however, is not true in
general, as the following example shows.

5.0.1 Program The program

p←p
p ← ¬p

has a unique supported model {p} and Fitting model ∅.

In this chapter, we study supported models in two-valued and three-valued

logic, with particular emphasis on uniquely determined and Φ-accessible pro­
grams. In particular, in Section 5.1 we consider two-valued supported models
and apply generalized metric fixed-point theorems from Chapter 4 in order
to show that certain classes of programs are uniquely determined. As is to
be expected, more general fixed-point theorems allow the treatment of more
general classes of programs, so that the hierarchy of fixed-point theorems from
Section 4.7 gives rise to a hierarchy of program classes, each of which has the
property that all programs in the class have unique supported models. Such
program classes are consequently called unique supported model classes.
The same hierarchy of unique supported model classes will be considered

139
140 Mathematical Aspects of Logic Programming Semantics

again in Section 5.2, but this time from the point of view of three-valued sup­
ported models (more precisely by studying variants of Fitting’s Φ-operator).
By analogy with Chapter 2, we will establish a correspondence between se­
mantics defined, on the one hand, by means of monotonic operators, and
characterizations given by means of level mappings, on the other hand. As a
result, we obtain a hierarchy of program classes which extends observations
from Chapter 2. All this will be carried out in this chapter in Section 5.3.
Finally, in Section 5.4, we make some brief observations concerning how
one may approach the results of this chapter from a much more general point
of view.

5.1 Two-Valued Supported Models

We know from Proposition 2.2.6 that the (two-valued) supported mod­
els for a given program P are exactly the fixed points of the corresponding
single-step operator TP . From Program 2.2.4, we know that TP is in general
not monotonic. This fact has the particular consequence that the fixed-point
theorems from Section 1.1 for monotonic operators are not applicable to TP
in this case. The alternative suggested by our development in Chapter 4 is to
apply, to non-monotonic single-step operators, fixed-point theorems utilizing
generalized metrics. In particular, it suggests in our current context the ap­
plication of those theorems which directly generalize the Banach contraction
mapping theorem to the extent that they ensure uniqueness of the resulting
fixed points, if any. Of course, if we successfully apply any of these particu­
lar theorems to a single-step operator, the corresponding program will clearly
be uniquely determined. It follows, therefore, that any approach of this type
employing fixed-point theorems which guarantee uniqueness of the resulting
fixed points cannot, when applied to single-step operators, encompass all (def­
inite) programs. Program 2.3.1, for example, is definite, but has two supported
models and, hence, cannot be uniquely determined.
Throughout the present section, it will be convenient to let IP denote IP,2 .

5.1.1 Acyclic and Locally Hierarchical Programs

Let us first recall the program Even (Program 2.1.3). Iterates of the corre­
sponding immediate consequence operator TEven are easily computed and are
as follows, for all n ∈ N, see Example 3.3.6.

2n
= even s2k (0) | 0 ≤ k < n ,
A b
TEven
2n+1
= BEven \ even s2k+1 (0) | 0 ≤ k < n
A b
TEven
 Supported Model Semantics 141

We notice that the sequence of iterates is alternating in a certain sense. The

iterates with even num
A bers bsuccessively generate the atoms in the supported
model M = even s2n (0) | n ∈ N , while the iterates with odd numbers
successively delete those atoms which are not in M . The order in which the
atoms are generated or deleted is such that atoms with more occurrences of
the function symbol s are generated or deleted later. This corresponds to the
structure of the Even program, whose rules reflect this in the sense that the
atom in the head of a ground instance of the second program clause always
contains one more function symbol than the corresponding body atom.
The following definition abstracts from this and draws on the observation
made in the previous paragraph that iterates of the immediate consequence
operator can in some sense be controlled if there is a strong dependency be­
tween heads of clauses and their corresponding body atoms. This is a theme
which will dominate the discussion of this chapter, and the reader may have
already noticed that it is related to the characterizations of semantics using
level mappings given in Chapter 2. The precise relationship between these two
themes will be made more explicit in Section 5.2.

5.1.1 Definition A normal logic program P is called locally hierarchical 1 if

there exists a level mapping l : BP → α, for some ordinal α, such that for
each clause A ← L1 , . . . , Ln in ground(P ) and for all i = 1, . . . , n we have
l(A) > l(Li ). If α can be chosen here to be ω, then P is called acyclic.2

program is acyclic, as can be seen by defining l : BP → α by

A TheAEven bb
l even sk (0) = k for all k ∈ N.

5.1.2 Program (ExistsEven) Consider the following program, which ex­

tends Even. We call it ExistsEven because intuitively, and also when run
under Prolog, it is a generate-and-test program which tests whether or not
there exists an even number.

nat(0) ←
nat(s(X)) ← nat(X)
even(0) ←
even(s(X)) ← ¬even(X)
existsEven ← nat(X), even(X)

1 Locally hierarchical programs were studied in [Cavedon, 1989]. It was shown in

[Seda and Hitzler, 1999a] that it is possible to compute all partial recursive functions with
locally hierarchical programs under SLDNF-resolution if the use of the meta-logical cut is
allowed.
2 Acyclic programs were studied in [Cavedon, 1989, Cavedon, 1991] under the name of

ω-locally hierarchical programs. The notion of acyclicity was introduced in [Bezem, 1989],
and further studies of it concerning termination properties were undertaken in [Bezem, 1989,
Apt and Bezem, 1990].
142 Mathematical Aspects of Logic Programming Semantics

Certainly, ExistsEven is somewhat pointless as a program. However, it ex­

hibits the basic idea underlying the generate-and-test programming scheme.
If Prolog is called with the query

?- existsEven.

then the interpreter successively generates all instantiations of nat(X) and

tests for each instance of X whether or not it falls under the predicate even.
Obviously, the generator nat and the test even could be replaced by something
much more sophisticated.
In ExistsEven, the subprogram consisting
A ofAthe firstbb fourA clauses
A is acyclic
with respect to the level mapping l with l even sk (0) = l nat sk (0) = k
bb

for all k ∈ N, and we notice that any level mapping with respect to which
this subprogram is acyclic must have an infinite codomain. Consequently,
ExistsEven is not acyclic, but it is locally hierarchical, as can be seen by
extending the level mapping by setting l(existsEven) = ω.

We want to apply generalized metric fixed-point theorems from Chapter 4

to acyclic and locally hierarchical programs, that is, we would like to construct
a (generalized) metric on the set of all interpretations of a program such that
the immediate consequence operator of the program satisfies a corresponding
contractivity property. We follow the construction of Section 4.8.2 with a
minor modification to suit our present purposes.

5.1.3 Definition Let P be a normal logic program, and let l : BP → γ be a

level mapping for P . We consider symbols 2−α for ordinals α, and, essentially
as in Section 4.8.2, define Γl = {2−α | α ≤ γ}. The set Γl is again ordered by
2−α < 2−β if and only if β < α, and we denote 2−γ by 0.
In Sections 4.8.2 and 4.8.3, we used this construction for gums with ordinal
distances, and with the notation established there we have Γl = Γγ+1 , where
l : BP → γ.
Finally, define a mapping dl : IP ×IP → Γl by setting dl (I, J) = 0 if I = J,
and, when I = J, by setting dl (I, J) = 2−α , where I and J differ on some
ground atom of level α, but agree on all ground atoms β satisfying β < α.

In case γ = ω, we can identify each 2−n ∈ Γl with the corresponding

negative power of two, that is, 2−n = 21n ∈ R and 2−ω = 0, and then dl takes
values in the set of real numbers.

5.1.4 Proposition Suppose that P is a normal logic program, and that l is

a level mapping. Then the following statements hold.

(a) If P is locally hierarchical with respect to l, then dl is a spherically com­

plete generalized ultrametric.

(b) If P is acyclic with respect to l, then dl is a complete ultrametric.

 Supported Model Semantics 143

Proof: It suffices to prove (a). We will do this by applying Theorem 4.8.14. For
the given level mapping l, define the rank function rl by setting rl (∅) = 0 and
by setting rl (I) = max{l(A) | A ∈ I} for every non-empty I ∈ (IP )c , where
we identify each element of (IP )c with a finite subset of BP , as usual. The
generalized ultrametric drl induced by rl , as in Definition 4.8.12, is spherically
complete by Theorem 4.8.14. The mappings dl and drl coincide since, for each
I ∈ IP , we have I = sup{{A} | A ∈ I}, with the supremum being taken with
respect to subset inclusion. •

Under certain conditions similar to those discussed in Section 4.6, we can

recover the Cantor topology from dl .

5.1.5 Proposition Let P be a normal logic program, and let l : BP → ω be

a level mapping such that l−1 (n) is finite for each n ∈ N. Then dl induces the
Cantor topology Q on IP .

Proof: It is easily shown by using Proposition 3.3.5 that sequences converge

in Q if and only if they converge with respect to dl , and this observation
suffices. •

We show finally that the immediate consequence operator satisfies the re­
quired contractivity conditions for applying the Prieß-Crampe and Ribenboim
theorem or the Banach contraction mapping theorem, as appropriate.

5.1.6 Theorem Suppose that P is a normal logic program, and that l is a

level mapping. Then the following statements hold.

(a) If P is locally hierarchical with respect to l, then TP is a strictly contract­

ing.

(b) If P is acyclic with respect to l, then TP is a contraction.

Furthermore, in both cases, TP has a unique fixed point, and P has a unique
supported model.

Proof: (a) Suppose I1 , I2 ∈ IP , and that dl (I1 , I2 ) = 2−α for some ordinal α.
Suppose α = 0. Let A ∈ TP (I1 ) with l(A) = 0. Since P is locally hierar­
chical, A must be the head of a unit clause in ground(P ). From this it follows
that A ∈ TP (I2 ) also. By the same argument, if A ∈ TP (I2 ) with l(A) = 0,
then A ∈ TP (I1 ). Therefore, TP (I1 ) and TP (I2 ) agree on all atoms of level less
than 1, and hence we have

dl (TP (I1 ), TP (I2 )) ≤ 2−1 < 2−0 = dl (I1 , I2 ),

as required.
Now suppose α > 0, so that I1 and I2 differ on some element of BP
with level α, but agree on all ground atoms of lower level. Let A ∈ TP (I1 )
144 Mathematical Aspects of Logic Programming Semantics

with l(A) ≤ α. Then there is a clause A ← A1 , . . . , Ak1 , ¬B1 , . . . , ¬Bl1 in

ground(P ), where k1 , l1 ≥ 0, such that for all k, j we have Ak ∈ I1 and
Bj ∈ I1 . Since P is locally hierarchical and I1 , I2 agree on all atoms of level
less than α, it follows that for all k, j we have Ak ∈ I2 and Bj ∈ I2 . Therefore,
A ∈ TP (I2 ). By the same argument, if A ∈ TP (I2 ) with l(A) ≤ α, then
A ∈ TP (I1 ). Hence, we have that TP (I1 ) and TP (I2 ) agree on all atoms of
level less than or equal to α, and it follows that

dl (TP (I1 ), TP (I2 )) ≤ 2−(α+1) < 2−α = dl (I1 , I2 ),

as required.
Thus, TP is strictly contracting, and Theorem 4.3.6 yields that TP has a
unique fixed point and therefore that P has a unique supported model.
The proof just given is easily adapted to establish (b). The operator TP
turns out to be contractive with contractivity factor 12 , and then Theorem
4.2.3 is applied instead of Theorem 4.3.6. •

5.1.7 Example Consider the program Tweety1 from Examples 2.1.2 and
2.2.7. Tweety1 is acyclic with level mapping l(penguin(X)) = 0, l(bird(X)) =
1 and l(flies(X)) = 2 for X ∈ {bob, tweety}. For I0 = {bird(tweety)}, we
obtain

I1 = TTweety1 (I0 ) = {penguin(tweety), bird(bob), flies(tweety)},

I2 = TTweety1 (I1 ) = {penguin(tweety), bird(bob), bird(tweety),
flies(bob)}, and
I3 = TTweety1 (I2 ) = I2 .

Another example is given by the program Even (Program 2.1.3), as dis­

cussed at the beginning of Section 5.1.1.

5.1.2 Acceptable Programs

Historically, acyclic programs were introduced in attempts to capture
procedural properties, such as termination, under SLDNF-resolution, see
[Bezem, 1989, Apt and Bezem, 1990, Cavedon, 1991]. The basic idea behind
acyclic programs was extended to take into account the fact that logic pro­
gramming systems, such as Prolog, evaluate clause bodies from left to right,
and this led to the acceptable programs3 studied in this section. We will focus
on declarative aspects of acceptable programs here, generalizing the approach
of Section 5.1.1.
3 Acceptable programs were introduced by Apt and Pedreschi in AP94. For further read­

ing concerning termination in resolution-based logic programming, see [Marchiori, 1996,

Apt, 1997, Pedreschi et al., 2002].
 Supported Model Semantics 145

5.1.8 Definition Let P be a program, and recall from Section 2.5 that an
atom A ∈ BP refers to an atom B ∈ BP if B or ¬B occurs as a body literal in
a clause A ← body in P . We say that A depends on B if the pair (A, B) is in
the transitive closure of the relation refers to. We further denote by NegP the
set of predicate symbols in P which occur in a negative literal in the body of a
clause in P , and we set Neg∗P = NegP ∪ D, where D is the set of all predicate
symbols in P on which the predicate symbols in NegP depend. Finally, by
P − we denote the set of clauses in P whose head contains a predicate symbol
from Neg∗P .
Finally, a program P is called acceptable with respect to some ω-level
mapping l : BP → ω and some interpretation I ∈ IP if I is a model for P whose
restriction to the predicate symbols in Neg∗P is a supported model for P − , and
the following condition holds. For each ground instance A ← L1 , . . . , Ln of a
clause in P and for all i ∈ {1, . . . , n} we have
i−1
�
if I |= Lj , then l(A) > l(Li ). (5.1)
j=1

The following is an example of an acceptable program.

5.1.9 Program Let G be an acyclic finite graph. We define the program

Game to be the program consisting of the following clauses.4
win(X) ← move(X, Y ), ¬win(Y ).
move(a, b) ← for all (a, b) ∈ G
Game is not acyclic. One of the ground instances of the first clause is
win(a) ← move(a, a), ¬win(a), so if Game were acyclic with respect to some
level mapping l, we would have l(win(a)) < l(win(a)), which is impossible.
In order to show that Game is acceptable, we need to find a suitable level
mapping l and a suitable model I for P . Since G is acyclic and finite, there
exists a function f which assigns a natural number to every vertex of G, and
such that for each vertex a the following holds.

0 if there is no (a, b) ∈ G,
f (a) =
1 + max{f (b) | (a, b) ∈ G} otherwise.

We now define l by setting l(move(a, b)) = f (a) and l(win(a)) = f (a) + 1

for all vertices a, b of G. From acyclicity and finiteness of G, we furthermore
obtain that there exists a function g mapping each vertex to {0, 1} satisfying
the following.

0 if there is no (a, b) ∈ G,
g(a) =
1 − min{g(b) | (a, b) ∈ G} otherwise.
4 This example is taken from [Apt and Pedreschi, 1994]. For further discussion of pro­

grams related to Game, see [Hitzler and Seda, 2003].

146 Mathematical Aspects of Logic Programming Semantics

Finally, let
I = {move(a, b) | (a, b) ∈ G} ∪ {win(a) | g(a) = 1}.
It is straightforward to verify that Game is acceptable with respect to I and
l.
We will now show how to construct a complete dislocated metric for any
given acceptable program with respect to which the immediate consequence
operator associated with the program is a contraction. For this purpose, let
P be a program which is acceptable with respect to a level mapping l and
an interpretation I. For any K ∈ IP , we denote by K ' the set K restricted
to the predicate symbols in Neg∗P . Next, we define a function f : IP → R by
setting f (K) = 0 if K \ K ' ⊆ I and, if K \ K ' ⊆ I, by setting f (K) = 2−n ,
where n ≥ 0 is the smallest integer such that there is an atom A ∈ BP with
l(A) = n, A ∈ K \ K ' and A ∈ I. Now define a function u : IP → R by setting
u(K) = max{f (K), dl (K ' , I ' )}, where dl is the generalized ultrametric from
Definition 5.1.3.
Finally, for all J, K ∈ IP , we set5
�(J, K) = max{dl (J \ J ' , K \ K ' ), u(J), u(K)}.
Thus, for all J, K ∈ IP , we have
�(J, K) = max{dl (J \ J ' , K \ K ' ), f (J), dl (J ' , I ' ), f (K), dl (K ' , I ' )}.
We apply Proposition 4.8.7 in order to show that � is a complete dislocated
ultrametric. We will need the following lemma.

5.1.10 Lemma Let u(K) = max{f (K), dl (K ' , I ' )} for K ∈ IP . Then u is
continuous as a function from (IP , dl ) to R.
Proof: Let Km be a sequence in IP which converges in dl to some K ∈ IP .
'
We need to show that dl (Km , I ' ) converges to dl (K ' , I ' ) and that f (Km )
converges to f (K) as m → ∞. Since (Km ) converges to K with respect to
the metric dl , it follows that for each n ∈ N there is mn ∈ N such that, for
all m ≥ mn , K and Km agree on all atoms of level less than or equal to n.
Suppose that f (K) = 2−n0 , say, and that m ≥ mn0 . Then Km and K agree
on all atoms of level less than or equal to n0 , and it follows that K ' and Km '

agree on all atoms of level less than or equal to n0 and, hence, that K \ K '
'
and Km \ Km agree on all atoms of level less than or equal to n0 . Therefore,
we have f (Km ) = 2−n0 = f (K) for all m ≥ mn0 . Also, if dl (K ' , I ' ) = 2−n0 ,
'
say, then dl (Km , I ' ) = 2−n0 = dl (K ' , I ' ) for all m ≥ mn0 .
The result now follows. •

It remains to show that TP is a contraction with respect to �.

5 This approach was inspired by [Fitting, 1994b]. The function u is usually called a weight

function if it is used for constructing dislocated metrics from metrics, see [Matthews, 1992,
Waszkiewicz, 2002]. Here, and in Section 5.1.3, we follow [Seda and Hitzler, 2010].
 Supported Model Semantics 147

5.1.11 Theorem Let P be a program which is acceptable with respect to

some level mapping l and interpretation I. Then � is a complete dislocated
ultrametric, and TP is a contraction with respect to �. In particular, P has a
unique supported model M and M = lim TPn (I0 ) for any I0 ∈ IP .
Proof: The mapping � is a complete dislocated ultrametric by Lemma 5.1.10
and Proposition 4.8.7. By Matthews’ theorem, Theorem 4.4.6, it remains to
show that TP is a contraction with respect to �. The argument for this is
essentially the same as the slightly more general one in the proof of Theorem
5.1.14, to be given in the next section, so we omit it here. •

5.1.3 Φ∗ -Accessible Programs

We have seen in Section 5.1.2 that application of the Banach contraction
mapping theorem can be replaced by application of Matthews’ theorem when
passing from acyclic to acceptable programs. Likewise, the Priess-Crampe and
Ribenboim theorem can be used in place of Banach’s theorem when passing
from acyclic to locally hierarchical programs. Naturally, the question arises
as to whether or not a class of programs can be described which generalizes
both the acceptable and the locally hierarchical programs such that Theorem
4.5.1, which generalizes both Matthews’ theorem and the Priess-Crampe and
Ribenboim theorem, can be applied. We will describe such a class of programs
in this section.

5.1.12 Definition A program P is called Φ∗ -accessible 6 if and only if there

exists a level mapping l for P and a model I for P whose restriction to Neg∗P
is a supported model for P − such that the following condition holds. For each
clause A ← L1 , . . . , Ln in ground(P ), either we have I |= L1 ∧ · · · ∧ Ln and
l(A) > l(Li ) for all i = 1, . . . , n or there exists i ∈ {1, . . . , n} such that I |= Li
and l(A) > l(Li ).
As an example, we refer again to the generate-and-test scheme described
in Program 5.1.2.

5.1.13 Program Assume that the unary predicate symbols generate and
test are defined via acceptable programs P1 and P2 , and consider the program
P which is the union of P1 , P2 and the following clause.

success ← generate(X), test(X).

It is easy to see that P is Φ∗ -accessible: first note that P1 and P2 are Φ∗ ­

accessible with respect to models I1 and I2 and level mappings l1 and l2 , say,
with codomain ω. We can assume without loss of generality that BP1 and BP2
6 It was shown in [Hitzler and Seda, 2003] that it is possible to compute all partial re­

cursive functions with definite Φ∗ -accessible programs using SLD-resolution.

148 Mathematical Aspects of Logic Programming Semantics

are disjoint and do not contain success. Now define I = I1 ∪ I2 ∪ {success}

and define l : BP → ω + 1 by l(A) = li (A), if A ∈ BPi , and l(success) = ω.
Then P is easily seen to be Φ∗ -accessible with respect to I and l.

We continue to carry over the approach of Section 5.1.2; again, we follow

[Seda and Hitzler, 2010]. So let P be a program which is Φ∗ -accessible with
respect to a level mapping l : BP → γ and an interpretation I. For any
K ∈ IP , we again denote by K ' the set K restricted to the predicate symbols
in Neg∗P . Again, we define a function f on IP , this time taking values in Γl ,
by setting f (K) = 0 if K \ K ' ⊆ I and, if K \ K ' ⊆I, by setting f (K) = 2−α ,
where α is the smallest ordinal such that there is an atom A ∈ BP with
l(A) = α, A ∈ K \ K ' and A ∈ I. Now define a function u : IP → Γl by again
setting u(K) = max{f (K), dl (K ' , I ' )}, where dl is the generalized ultrametric
from Definition 5.1.3.
Finally, for all J, K ∈ IP , we set

�(J, K) = max{dl (J \ J ' , K \ K ' ), u(J), u(K)}

as before. Thus, for all J, K ∈ IP , we have

�(J, K) = max{dl (J \ J ' , K \ K ' ), f (J), dl (J ' , I ' ), f (K), dl (K ' , I ' )}.

In fact, the details of the proof of the main result below will be simplified
by introducing the functions d1 and d2 , where, for all J, K ∈ IP , we set
d1 (J, K) = dl (J ' , K ' ) and d2 (J, K) = dl (J \ J ' , K \ K ' ). Indeed, in these terms
we have

�(J, K) = max{d1 (J, I), d1 (K, I), d2 (J, K), f (J), f (K)}

for all J, K ∈ IP .

5.1.14 Theorem Let P be a Φ∗ -accessible normal logic program. Then the

space (IP , �) is a spherically complete, dislocated generalized ultrametric
space, and TP is strictly contracting with respect to �. In particular, P has a
unique supported model.

Proof: It follows from Proposition 4.8.22 that � is a dislocated generalized

ultrametric. For spherical completeness, let (Bα ) be a (decreasing) chain of
balls in IP with centres Iα . Let K be the set of all atoms which are eventually
in Iα , that is, the set of all A ∈ BP such that there exists some ordinal β with
A ∈ Iα for all α ≥ β. We show that for each ball B2−α (Iα ) in the chain, we
have dl (Iα , I) ≤ 2−α , which suffices to show that K is in the intersection of
the chain. Indeed, it is easy to see by the definition of � that all Iβ with β > α
agree on all atoms of level less than α. Hence, by definition of K we obtain
that K and Iα agree on all atoms of level less than α, as required.
It remains to show that TP is strictly contracting with respect to �, for it
will then follow from Theorem 4.5.1 that the operator TP has a unique fixed
 Supported Model Semantics 149

point, yielding a unique supported model for P . In order to show that TP is

strictly contracting with respect to �, we must show that for all J, K ∈ IP
with J = K we have �(TP (J), TP (K)) < �(J, K). In particular, the following
results hold.
(a) d1 (TP (J), I) < d1 (J, I) whenever d1 (J, I) = 0, and d1 (TP (J), I) = 0
whenever d1 (J, I) = 0.
(b) f (TP (J)), f (TP (K)) < �(J, K).
(c) d2 (TP (J), TP (K)) < �(J, K).
Indeed, it suffices to prove properties (a), (b) and (c), and we proceed to do
this next. For convenience, we identify Neg∗P with the subset of BP containing
predicate symbols from Neg∗P .
(a) First note that d1 (TP (J), I) = d1 (TP − (J), I) since d1 only depends
on the predicate symbols in Neg∗P . Let dl (J, I) = 2−α . We show that
dl (TP − (J), I) ≤ 2−(α+1) . We know that J ' and I ' agree on all ground atoms
of level less than α and differ on an atom of level α. It suffices to show now
that TP − (J)' and I ' agree on all ground atoms of level less than or equal to
α.
Let A be a ground atom in Neg∗P with l(A) ≤ α, and suppose that TP − (J)
and I differ on A. Assume first that A ∈ TP − (J) and A ∈ I. Then there
must be a ground instance A ← L1 , . . . , Lm of a clause in P − such that
J |= L1 ∧ · · · ∧ Lm . Since I is a fixed point of TP − , and using Definition 5.1.12,
there must also be a k such that I |= Lk and l(Lk ) < α. Note that the
predicate symbol in Lk is contained in Neg∗P . So we obtain I |= Lk , J |= Lk
and l(Lk ) < α, which is a contradiction to the assumption that J and I
agree on all atoms in Neg∗P of level less than α. Now assume that A ∈ I and
A ∈ TP − (J). It follows that there is a ground instance A ← L1 , . . . , Lm of
a clause in P − such that I |= L1 ∧ · · · ∧ Lm and l(A) > l(L1 ), . . . , l(Lm ) by
Definition 5.1.12. But then J |= L1 ∧ · · · ∧ Lm since J and I agree on all
atoms of level less than α and consequently A ∈ TP − (J). This contradiction
establishes the first statement in (a). The second statement in (a) follows by
a similar argument, noting that in this case J ' = I ' .
(b) It suffices to show this for K. Assume �(J, K) = 2−α . We show that
f (TP (K)) ≤ 2−(α+1) , for which, in turn, we have to show that, for each
A ∈ TP (K) not in Neg∗P with l(A) ≤ α, we have A ∈ I. Assume that A ∈ I
for such an A. Since A ∈ TP (K), there is a ground instance A ← L1 , . . . , Lm
of a clause in P with K |= L1 ∧ · · · ∧ Lm . Since A ∈ I, there must also be a
k with I |= Lk and l(A) > l(Lk ) by Definition 5.1.12. If the predicate symbol
of Lk belongs to Neg∗P , then, since K and I agree on all atoms in NegP ∗
of
level less than α, we obtain K |= Lk , which contradicts K |= L1 ∧ · · · ∧ Lm .
If the predicate symbol in Lk does not belong to Neg∗P , then Lk is an atom,
and since f (K) ≤ 2−α , we obtain I |= Lk , which is again a contradiction.
(c) Let �(J, K) = 2−α , and let A be not in Neg∗P with l(A) ≤ α and
A ∈ TP (J). By symmetry, it suffices to show that A ∈ TP (K). Since A ∈
150 Mathematical Aspects of Logic Programming Semantics

TP (J), we must have a ground instance A ← L1 , . . . , Lm of a clause in P

with J |= L1 ∧ · · · ∧ Lm . If I |= L1 ∧ · · · ∧ Lm , then l(Lk ) < l(A) ≤ α for
all k, and since J and K agree on all atoms of level less than α, we obtain
K |= L1 ∧· · ·∧Lm , and hence A ∈ TP (K). If there is some Lk such that I |= Lk ,
then without loss of generality l(Lk ) < l(A) ≤ α by Definition 5.1.12. Now,
if the predicate symbol of Lk belongs to Neg∗P , then, since d1 (J, I) ≤ 2−α ,
we obtain from J |= Lk that I |= Lk , which is a contradiction. Also, if the
predicate symbol of Lk does not belong to Neg∗P , then Lk is an atom, and
since f (J) ≤ 2−α , we obtain I |= Lk , again a contradiction. This establishes
(c) and completes the proof. •

5.1.4 Φ-Accessible Programs

Definition 5.1.12 of Φ∗ -accessibility is obviously related to the level map­
ping characterization of the Fitting semantics given in Section 2.4. In the
present section, we will carry over the approach from Section 5.1.3 to programs
with a total Fitting model, and we refer the reader to [Hitzler and Seda, 2003]
for further details. The relationships between the different classes of programs
studied so far in this chapter will be further clarified in Section 5.2.

5.1.15 Definition A program is called Φ-accessible if it has a total Fitting

model.
By Corollary 2.4.10, a program P is Φ-accessible if and only if there is a
(two-valued) model I and a (total) level mapping l for P such that P satisfies
(F) with respect to I ∪ ¬(BP \ I) and l. The restriction of I to Neg∗P is
then a supported model for P − , and it follows easily that every Φ∗ -accessible
program is Φ-accessible. However, the development of Section 5.1.3 does not
generalize without modifications, as the following example shows.

5.1.16 Program Let P be the following program.

p s2 (x) ← p(x)
A b

p(0) ←
p s (0) ← p s5 (0)
A 4 b A b

p s2 (0) ← p s3 (0)
A b A b

The program P is Φ-accessible (and even definite) with respect to the model
BP = {p(sn (0)) | n ∈ N} and the level mapping l : BP → N defined by
l(p(sn (0))) = n. Using the dislocated generalized ultrametric � from Section
5.1.3, we obtain for K = {p(s5 (0))} and J = {p(s3 (0))} that �(K, J) = 2−3
and �(TP (K), TP (J)) = 2−2 ; thus, TP is not a contraction relative to �.
We will modify the methods used in Section 5.1.3 by means of Proposition
4.8.23.
 Supported Model Semantics 151

5.1.17 Theorem Let P be a Φ-accessible program with model I and level

mapping l such that P satisfies (F) with respect to I ∪ ¬(BP \ I) and l.
Then TP is strictly contracting on the spherically complete dislocated gener­
alized ultrametric space (IP , �), where for all J, K ∈ IP we have �(J, K) =
max{dl (J, I), dl (I, K)}. In particular, P has a unique supported model.
Proof: By Proposition 4.8.23, we have that (IP , �) is a spherically complete
dislocated generalized ultrametric space.
In order to show that TP is strictly contracting, let J, K ∈ IP , and assume
that �(J, K) = 2−α . Then J, K and I agree on all ground atoms of level less
than α. We show that TP (J) and I agree on all ground atoms of level less
than or equal to α. A similar argument shows that TP (K) and I agree on all
ground atoms of level less than or equal to α, and this suffices.
Let A ∈ TP (J) with l(A) ≤ α. Then there must be a clause A ← L1 , . . . , Ln
in ground(P ) such that J |= L1 ∧ · · · ∧ Ln . Since I and J agree on all ground
atoms of level less than α, (Fii) cannot hold, because if I |= Li with l(A) >
l(Li ), then J |= Li and consequently J |= L1 ∧· · ·∧Ln , which is a contradiction.
Therefore, (Fi) holds, and so A ∈ TP (I) = I. Hence, A ∈ I.
Conversely, suppose that A ∈ I. Since I = TP (I), there must be a clause
A ← L1 , . . . , Ln in ground(P ) such that I |= L1 ∧ · · · ∧ Ln . Thus, (Fi) must
hold, and so we can assume that A ← L1 , . . . , Ln also satisfies l(A) > l(Li )
for i = 1, . . . , n. Since I and J agree on all ground atoms of level less than α,
we have J |= L1 ∧ · · · ∧ Ln , and hence A ∈ TP (J), as required.
Applying Theorem 4.5.1 now yields a unique fixed point M of the operator
TP , that is, a unique supported model for P . •

The proof of Theorem 4.5.1 yields, moreover, that there must be an ordinal
α such that �(M, M ) = 0. Since the only point of X which has non-zero
distance from itself is I, we conclude that I = M is the unique supported
model for P . This is somewhat unfortunate,7 since I was needed in order to
construct �.

5.2 Three-Valued Supported Models

Recall from Section 2.4 that the three-valued supported models for a pro­
gram P are exactly the fixed points of the corresponding Fitting operator,
while the least fixed point of the operator, that is, the least three-valued sup­
ported model for the program, is called its Fitting model. In this section, we
will study variants of the Fitting operator and relate them to the classes of
programs studied in Section 5.1. Thus, in the present section, unless otherwise
7 We have argued in [Hitzler and Seda, 2003] that self-distance can be understood as a

measure of a priori knowledge, but this needs to be substantiated further.

152 Mathematical Aspects of Logic Programming Semantics

noted, interpretations will be three-valued, and therefore IP here means IP,3

ordered using the knowledge ordering introduced in Section 1.3.2.

5.2.1 Fitting Operators Revisited

We begin with an alternative characterization of the Fitting operator,
which is amenable to generalization in various logics. It involves a program
transformation which we will introduce next. Later on in Section 5.4, we will
consider further generalizations of Fitting operators, called Fitting-style op­
erators, see Definition 5.4.9.
Let P be a program and suppose that A ∈ BP is the head of some clause
in ground(P ). Now let {A ← bodyi | i ∈ Λ} be the set of all clausesowith head
A in ground(P ), where Λ is a suitable index set. We call o A ← i∈Λ bodyi
the pseudo-clause associated with A, we call bodyA = i∈Λ bodyi the body
of the pseudo-clause, and we call A its head . As a matter of notation, we
may sometimes
o denote bodyi by Ci , and henceo we may sometimes denote by
A ← i∈Λ Ci or even more simply by A ← Ci the pseudo-clause associated
with A.
Notice
o that the family {bodyi | i ∈ Λ} of bodies may be denumerable and
that i∈Λ bodyi is formal at this stage. Nevertheless, we next assign truth
values to bodies of pseudo-clauses with respect to an interpretation in cer­
tain three-valued
o logics8 and in more generality in Theorem 5.5.1 and in Sec­
tion 7.6. If i∈Λ bodyi is such a body, then bodyi is a (finite) conjunction for
any i and can be evaluated as usual by means of truth tables for conjunction.
We will consider three different conjunctions and two different disjunctions,
all as given by the truth tables in Table 5.1 on Page 153. Note that ∧1 and ∨1
are exactly the conjunction and disjunction from Kleene’s strong three-valued
logic, specified earlier as a sublogic of Belnap’s logic in Table 1.1, and already
employed in Section 2.4 in evaluating truth values of clause bodies.
With respect to ∨1 , a disjunction p ∨1 q is false if and only if both p and
q are false, is true if and only if one of p and q is true, and is undefined
otherwise. We use this as a definition of truth values for bodies of pseudo-
clauses. Therefore, with respect to ∨1 :
o
the body i∈Λ bodyi of a pseudo-clause is false if and only if
all of the bodyi are false, is true if and only if one of the bodyi
is true, and is undefined otherwise.
With respect to ∨2 , a disjunction p ∨2 q is false if and only if both p and q
are false, is undefined if one of p and q is undefined, and is true otherwise.
Therefore, with respect to ∨2 :
o
the body i∈Λ bodyi of a pseudo-clause is false if and only if
8 Strictly speaking, we discuss different truth tables for logical connectives – or rather

different connectives – over three truth values over the same underlying language. It will be
convenient to think in terms of different logics, however.
 Supported Model Semantics 153

TABLE 5.1: Several truth tables for three-valued logics.

p q p ∧1 q p ∧2 q p ∧3 q
u u u u u
u f f u u
u t u u u
f u f f u
f f f f f
f t f f f
t u u u u
t f f f f
t t t t t

p q p ∨1 q p ∨2 q
u u u u
u f u u
u t t u
f u u u
f f f f
f t t t
t u t u
t f t t
t t t t

p ¬p
u u
f t
t f

all of the bodyi are false, is undefined if and only if one of the
bodyi is undefined, and is true otherwise.
Finally, if A is an atom which does not appear as the head
o of a clause in
ground(P ), then we say, by abuse of notation, that
o A ← i∈∅ bodyi is the
pseudo-clause associated with A, and we take i∈∅ bodyi to be false both
with respect to ∨1 and with respect to ∨2 . Notice now that every element
A of BP is the head of the pseudo-clause associated with A and that this
pseudo-clause is uniquely determined by P for a given A.

The following notation will be convenient. Let A ∈ BP , let bodyA be the

body of the pseudo-clause associated with A, and let I be a three-valued inter­
pretation. Then write Ij,k (bodyA ) for the truth value, under I, of bodyA with
respect to ∧j and ∨k , for j = 1, 2, 3 and k = 1, 2. The following proposition
follows easily from the definitions.
154 Mathematical Aspects of Logic Programming Semantics

5.2.1 Proposition Let P be a program, let I be a three-valued interpretation

for P , and let A ∈ BP . Then ΦP (I)(A) = I1,1 (bodyA ), that is, the truth value
of A under ΦP (I) is exactly I1,1 (bodyA ).

The logics from Table 5.1 give rise to different operators.9

5.2.2 Definition Let P be a program. For any j = 1, 2, 3 and any k = 1, 2,

we define an operator ΦP,j,k : IP,3 → IP,3 by ΦP,j,k (I)(A) = Ij,k (bodyA ).

We can now rephrase Proposition 5.2.1 by saying that the operators ΦP,1,1
and ΦP coincide. The following proposition lists properties of the ΦP,j,k ­
operators. We use the notation of three-valued interpretations as signed sets,
see Section 1.3.3, and of two-valued interpretations as subsets of BP .

5.2.3 Proposition Let P be a program, and let I, J, K ∈ IP,3 . Then the

following hold.

(a) ΦP,j,k is monotonic for j = 1, 2, 3 and k = 1, 2.

(b) ΦP,3,k (I) ⊆ ΦP,2,k (J) ⊆ ΦP,1,k (K) for k = 1, 2 if I ⊆ J ⊆ K.

(c) ΦP,j,2 (I) ⊆ ΦP,j,1 (I) for j = 1, 2, 3.

(d) ΦP,j,2 (I)− = ΦP,j,1 (I)− .

Proof: (a) The proof of this statement is very similar to that of Proposition
2.4.4 and is therefore omitted.
(b) From the truth tables, it follows that for all A ∈ BP and each k ∈ {1, 2}
we have I3,k (bodyA ) ⊆ J2,k (bodyA ) ⊆ K1,k (bodyA ), and this suffices.
(c) From the truth tables, we obtain Ij,2 (bodyA ) ⊆ Ij,1 (bodyA ) for all
A ∈ BP .
− −
(d) By (c), it suffices to showAthat
o ΦP,j,2 (I)b ⊇ ΦP,j,1 (I) . So let A ∈ BP
be such that Ij,1 (bodyA ) = Ij,1 i∈Λ bodyi = f . Then Ij,1 (bodyi ) = f for
all i, and hence Ij,2 (bodyi ) = f for all i. Consequently, Ij,2 (bodyA ) = f , as
required. •

Proposition 5.2.3 shows that the operators are “nested” and that ΦP,1,1 =
ΦP is the least sceptical of them. In particular, for each ordinal α and all j
and k, the following hold.

ΦP,3,k ↑ α ⊆ ΦP,2,k ↑ α ⊆ ΦP,1,k ↑ α

ΦP,j,2 ↑ α ⊆ ΦP,j,1 ↑ α

We can also relate ΦP to the two-valued immediate consequence operator, TP ,

thereby extending Proposition 2.4.13.
9 We refer to the papers [Hitzler and Seda, 1999a, Hitzler and Seda, 2002b] for further

details concerning the results of this section.

 Supported Model Semantics 155

5.2.4 Lemma Let P be a normal logic program, let I ∈ IP,2 , and let K ∈ IP,3
be such that K + ⊆ I ⊆ BP \ K − . Then ΦP (K)+ ⊆ TP (I) ⊆ BP \ ΦP (K)− .
Furthermore, if K + = I = BP \ K − , so that K is total, then ΦP (K)+ =
TP (I) = BP \ ΦP (K)− .
Proof: Suppose that A ∈ ΦP (K)+ . Then A must be the head of a clause
A ← A1 , . . . , Ak1 , ¬B1 , . . . , ¬Bk2 in ground(P ) with Ai ∈ K + and Bj ∈ K −
for all i = 1, . . . , k1 and j = 1, . . . , k2 . By assumption, it follows that for these
values of i and j, Ai ∈ I and Bj ∈ I, and hence A ∈ TP (I).
For the second inclusion, it suffices to show that ΦP (K)− ⊆ BP \
TP (I). Let A ∈ ΦP (K)− . Then, for every clause of the form A ←
A1 , . . . , Ak1 , ¬B1 , . . . , ¬Bk2 in ground(P ), we have some Ai ∈ K − or some
Bj ∈ K + . Hence, for every such clause, we have some Ai ∈ I or some Bj ∈ I,
which implies that A ∈ TP (I).
The final statement was established in Proposition 2.4.13. •

The following straightforward corollary provides the essential link between

the Φ-operator, the single-step operator TP , and convergence in Q.

5.2.5 Corollary Let In = TPn (I) for some I ∈ IP,2 , and let Kn = ΦP ↑ n.
Then, for all n ∈ N, we obtain Kn+ ⊆ In ⊆ BP \ Kn− .
The following is a direct consequence of Lemma 5.2.4.

5.2.6 Proposition Let P be a normal logic program, and let (I + , I − ) be a

total three-valued interpretation I for P . Then I is a fixed point of ΦP if and
only if I + is a fixed point of TP . Furthermore, if ΦP has exactly one total
fixed point M , then M + is the unique fixed point of TP .
Proof: Let I be a fixed point of ΦP . Then I + ⊆ I + ⊆ BP \ I − , and by
Lemma 5.2.4 we obtain I + = ΦP (I)+ ⊆ TP (I + ) ⊆ BP \ ΦP (I)− = BP \ I − =
I + . Conversely, let I + be a fixed point of TP . By Lemma 5.2.4, we obtain
ΦP (I)+ = TP (I + ) = I + = BP \ I − = BP \ ΦP (I)− , and therefore ΦP (I)+ =
I + and ΦP (I)− = I − . The last statement now follows immediately. •

Convergence of iterates with respect to the Cantor topology can now be

described, as follows.

5.2.7 Proposition Let P be a normal logic program, and assume that M =

ΦP ↑ ω is total. Then TPn (∅) converges in Q to M + , and M + is the unique
supported model MP for P .
Using the notation from Corollary 5.2.5, we obtain M + = Kn+ and

Proof: 
M − = Kn− . Since M is total, we obtain from Propositions 3.3.5 and 5.2.6
that M + is the limit in Q of the sequence In . Since totality of ΦP ↑ ω implies
that it is the unique fixed point of ΦP , it therefore equals (M + , M − ), so that
M + is the unique fixed point of TP by Proposition 5.2.6. •
156 Mathematical Aspects of Logic Programming Semantics

Proposition 5.2.7 allows us to apply Theorem 4.2.5 in the following way.

Let P be a normal logic program such that ΦP ↑ ω is total. Then TPn (I)
converges in Q to ΦP ↑ ω for every I, and ΦP ↑ ω is the unique fixed point of
TP . By Theorem 4.2.5, we can therefore find a metric with respect to which
TP is a contraction. However, this metric does not in general coincide with
the metric associated with the dislocated ultrametric � from Theorem 5.1.17,
with respect to which TP is also a contraction under the given condition on
P.
The following result is even stronger than Proposition 5.2.7.

5.2.8 Theorem Let P be a normal logic program, let j ∈ {1, 2, 3}, let k ∈
{1, 2}, and assume that M = ΦP,j,k ↑ α is total for some α. Then M + is
the unique two-valued supported model for P . Furthermore, the transfinite
sequence (ΦP,j,k ↑ β)β converges in the Cantor topology to M + .
Proof: By totality of M , Propositions 5.2.3 and 5.2.6, we obtain M + as a
fixed point of TP . The convergence results follow as in Proposition 5.2.7. •

We can extend the treatment of the Fitting operator from Section 2.4 to
the operators ΦP,j,k introduced in Definition 5.2.2. This will, in turn, lead
us back to the program classes from Section 5.1. We begin with the ΦP,3,2 ­
operator in the next section.

5.2.2 Acyclic and Locally Hierarchical Programs

We first present conditions analogous to Definition 2.4.8, which was used to
characterize the Fitting semantics, beginning with condition (F32 ) as defined
next.

5.2.9 Definition Let P be a normal logic program, let I be a model for P ,

and let l be an I-partial level mapping for P . We say that P satisfies (F32 ) with
respect to I and l if for each A ∈ dom(l) and for all clauses A ← L1 , . . . , Ln
in ground(P ) we have Li ∈ dom(l) and l(A) > l(Li ) for all i = 1, . . . , n, and
furthermore each A ∈ dom(l) satisfies one of the following conditions.
(Fi) A ∈ I, and there is a clause A ← L1 , . . . , Ln in ground(P ) such that
Li ∈ I and l(A) > l(Li ) for all i.
(Fii) ¬A ∈ I, and for each clause A ← L1 , . . . , Ln in ground(P ) there exists
i with ¬Li ∈ I and l(A) > l(Li ).
Conditions (Fi) and (Fii) are identical to those in Definition 2.4.8. The dif­
ference between Definitions 2.4.8 and 5.2.9 lies in the additional very strong
condition “for each A ∈ dom(l) and for all clauses A ← L1 , . . . , Ln in
ground(P ) we have Li ∈ dom(l) and l(A) > l(Li ) for all i = 1, . . . , n”. The
proof of the following theorem is very similar to the proof of Theorem 2.4.9
and is therefore only sketched.
 Supported Model Semantics 157

5.2.10 Theorem Let P be a normal logic program, and let M be the least
fixed point of the operator ΦP,3,2 . Then, in the knowledge ordering, M is the
greatest model among all three-valued models I for which there exists an I-
partial level mapping l for P such that P satisfies (F32 ) with respect to I and
l.

Proof: Let MP be the least fixed point of the operator ΦP,3,2 , and define
the MP -partial level mapping lP as follows: lP (A) = α, where α is the least
ordinal such that A is not undefined in ΦP ↑ (α + 1). The proof proceeds by
established the following facts. (1) P satisfies (F32 ) with respect to M and lP .
(2) If I is a three-valued model for P , and l is an I-partial level mapping such
that P satisfies (F32 ) with respect to I and l, then I ⊆ MP .
(1) Let A ∈ dom(lP ), and suppose that lP (A) = α. We consider two cases.
Case i. If A ∈ MP , then Table 5.1 together with the definition of lP yields
that A satisfies (Fi) with respect to MP and lP . It also yields that l(L) < α
for each literal L in the body of any clause from ground(P ) with head A.
Case ii. If ¬A ∈ MP , then again Table 5.1 together with the definition of
lP yields that A satisfies (Fii) with respect to MP and lP . As before, it also
yields l(L) < α for each literal L in the body of any clause from ground(P )
with head A. This completes the proof of (1).
(2) Similarly to the proof of Step (2) in the proof of Theorem 2.4.9, it can
be shown via transfinite induction on α = l(A) that: whenever A ∈ I we have
A ∈ ΦP,3,2 ↑ (α + 1) and whenever ¬A ∈ I we have ¬A ∈ ΦP,3,2 ↑ (α + 1). This
concludes the proof. •

5.2.11 Corollary A logic program P is acyclic if and only if ΦP,3,2 ↑ ω is

total, and is locally hierarchical if and only if ΦP,3,2 ↑ α is total for some
ordinal α.

Proof: Let P be such that ΦP,3,2 ↑ α is total for some α. Then by Theorem
5.2.10 and Definition 5.2.9 it follows that P is locally hierarchical with respect
to the level mapping lP as defined in the proof of Theorem 5.2.10.
Conversely, let P be locally hierarchical with level mapping l. Then, by
Theorem 5.1.6, P has a unique supported model M , that is, M is the unique
fixed point of the operator TP . We show that P satisfies (F32 ) with respect to
I = M ∪ ¬(BP \ M ) and l. For this it suffices to show that for each A ∈ BP ,
conditions (Fi) and (Fii) hold with respect to I. This, however, is an immediate
consequence of the fact that M is a fixed point of TP and that P is locally
hierarchical.
The argument to show that P is acyclic if and only if ΦP,3,2 ↑ ω is total is
similar. •
158 Mathematical Aspects of Logic Programming Semantics

5.2.3 Acceptable Programs

The treatment of Section 5.2.2 carries over to acceptable programs with
only minor modifications. Given a program P , an interpretation I ∈ IP,3 , and
an I-partial level mapping l, we say that a clause A ← L1 , . . . , Ln is k-safe
(with respect to I and l) if either L1 , . . . , Ln ∈ I and l(A) > l(Li ) for all
i = 1, . . . , n or ¬Lk ∈ I, L1 , . . . , Lk−1 ∈ I and l(A) > l(Li ) for all i = 1, . . . , k.
This notion generalizes condition (5.1) in Definition 5.1.8 in the following
sense: a program P is acceptable with respect to some ω-level mapping l and
some interpretation I ∈ IP,2 if and only if I is a model for P whose restriction
to the predicate symbols in Neg∗P is a supported model for P − , and for each
clause in ground(P ) there exists k such that the clause is k-safe (with respect
to I ∪ ¬(BP \ I) and l).

5.2.12 Definition Let P be a normal logic program, let I be a model for P ,

and let l be an I-partial level mapping for P . We say that P satisfies (F22 )
with respect to I and l if, for each A ∈ dom(l) and for all clauses in ground(P )
with head A, there exists k such that the clause is k-safe, and furthermore,
each A ∈ dom(l) satisfies one of the following conditions.

(Fi) A ∈ I, and there is a clause A ← L1 , . . . , Ln in ground(P ) such that

Li ∈ I and l(A) > l(Li ) for all i.

(Fii) ¬A ∈ I, and for each clause A ← L1 , . . . , Ln in ground(P ) there exists

i with ¬Li ∈ I and l(A) > l(Li ).

The proof of the following theorem is very similar to the proof of Theorem
5.2.10 and is therefore omitted.

5.2.13 Theorem Let P be a normal logic program and let M be the least
fixed point of the operator ΦP,2,2 . Then, in the knowledge ordering, M is the
greatest model among all three-valued models I for which there exists an I-
partial level mapping l for P such that P satisfies (F22 ) with respect to I and
l.

5.2.14 Corollary A normal logic program P is acceptable if and only if

ΦP,2,2 ↑ ω is total.

Proof: Let P be such that ΦP,2,2 ↑ ω is total. From Theorem 5.2.8 we know
that P has a unique supported model whose restriction to predicate symbols
in Neg∗P is a supported model for P − . By Theorem 5.2.10 and Definition 5.2.9,
it easily follows that P is acceptable.
The proof of the converse is similar to that of Corollary 5.2.11. •
 Supported Model Semantics 159

5.2.4 Φ∗ -Accessible Programs

We next give the analogue of Definition 5.2.9 for Φ∗ -accessible programs.
In order to make it more concise, we have chosen to rearrange the statements
of the conditions slightly. The reader will easily identify the parts which cor­
respond to conditions (Fi) and (Fii).

5.2.15 Definition Let P be a normal logic program, let I be a model for P ,

and let l be an I-partial level mapping for P . We say that P satisfies (F12 ) with
respect to I and l if for each A ∈ dom(l) and for all clauses A ← L1 , . . . , Ln
one of the following conditions (F12 i), (F12 ii) holds. Furthermore, if A ∈ I,
there must be at least one clause which satisfies (F12 i), and if ¬A ∈ I, there
must be no clauses which satisfy (F12 i).
(F12 i) Li ∈ I and l(A) > l(Li ) for all i.
(F12 ii) There exists i with ¬Li ∈ I and l(A) > l(Li ).
The proof of the following theorem is very similar to the proof of Theorem
5.2.10 and is therefore omitted.

5.2.16 Theorem Let P be a normal logic program, and let M be the least
fixed point of the operator ΦP,1,2 . Then, in the knowledge ordering, M is the
greatest model among all three-valued models I for which there exists an I-
partial level mapping l for P such that P satisfies (F12 ) with respect to I and
l.
The proof of the following corollary is similar to the proof of Corollary
5.2.14 and is therefore omitted.

5.2.17 Corollary A normal logic program P is Φ∗ -accessible if and only if

ΦP,1,2 ↑ α is total for some ordinal α.

5.2.5 Φ-Accessible Programs

Results for Φ-accessible programs corresponding to those for Φ∗ -accessible
programs in Section 5.2.4 have already been obtained, and we refrain from
repeating them here. Theorem 5.2.16 finds its analogue in Theorem 2.4.9, and
the analogue of Corollary 5.2.17 can be found in Definition 5.1.15.

5.3 A Hierarchy of Logic Programs

In Figure 5.1, we present an overview of the relationships between the
main classes of normal logic programs discussed in this book. Note that dif­
ferent branches of the graph shown are not necessarily disjoint. For example,
160 Mathematical Aspects of Logic Programming Semantics

acyclic definite
///
///
/// 1
locally 1 TP Scott-
covered acceptable stratified
hierarchical continuous
//
//
1 /// 1
/
TP 1 // locally
continuous ∗
Φ -accessible
in Q stratified

1
Φ-accessible
1 1
uniquely weakly
determined stratified
1
total
well-founded
model
1
unique
stable
model

FIGURE 5.1: The main classes of programs discussed in this book. The arrows
indicate class inclusion. See the main text of Section 5.3 for further details.

a program can be locally hierarchical without being acyclic, but still have a
Q-continuous immediate consequence operator, meaning that its immediate
consequence operator is continuous in the topology Q.
Covered programs are defined in Definition 7.5.4. Figure 5.1 indicates that
every acyclic program is covered, but note that this is only the case if we
assume that the underlying language contains at least one function symbol.
Indeed, if this is not the case, then the Herbrand base is finite, and, for ex­
ample, the program P with the single clause

q(a) ← p(x)

is acyclic10 but not covered. Q-continuity of the immediate consequence oper­

ator for covered programs follows from Corollary 5.4.8. Scott continuity of the
immediate consequence operator for definite programs follows from Theorem
2.2.3 (it was called order continuity there).
The remaining relationships shown in Figure 5.1 follow from results in
Chapter 2 and Section 5.2.
10 For example, assume a is the only constant symbol. Then ground(P ) is q(a) ← p(a),

and so P is obviously acyclic.

 Supported Model Semantics 161

5.4 Consequence Operators and Fitting-Style Operators

We close this section by discussing some natural extensions of certain ear­
lier results. These are obtained by defining a rather general semantic operator
T modelled on Fitting operators, but defined over abstract finite logics T
rather than over logics containing two, three, or four elements. We call the
resulting operators consequence operators, and an important special case of
them we call Fitting-style operators. Our main result here is a careful anal­
ysis of the continuity of these operators T in the Cantor topology Q, which
yields necessary and sufficient conditions for the continuity in Q of the single-
step operator as a special case, see Theorem 5.4.11. Once these results are
established, the aforementioned extensions we require are straightforward to
present.
Thus, let T denote a finite set {t1 , . . . , tn } of truth values containing at
least the two distinguished values t1 and tn , which are interpreted as being
the truth values for “false” and for “true”, respectively. We assume that we
have truth tables for the usual connectives ∨, ∧, ←, and ¬. Given a normal
logic program P , we denote the set of all (Herbrand) interpretations or valu­
ations in this logic by IP,n ; thus, IP,n is the set of all functions I : BP → T .
If n is clear from the context, we will use the notation IP instead of IP,n ,
and we note that this usage is consistent with that already established for
n = 2, 3, and 4. As usual, any interpretation I can be extended, using the
truth tables, to give a truth value in T to any variable-free formula in the
language L underlying P . We assume throughout this section that our un­
derlying language L contains at least one function symbol, and hence BP is
denumerable. Finally, we endow IP,n with the Cantor topology Q studied in
Chapter 3, see Theorem 3.3.1, and recall that this is the product topology of
BP copies of the discrete topology on T . We refer the reader to Theorem 3.3.4
and Proposition 3.3.9 for a summary of the properties of Q. We note that our
present assumption that BP is denumerable and that T is finite mean that Q
is second countable.
We proceed next with introducing a rather general notion of semantic oper­
ator T which subsumes many of the particular operators we have encountered
in the earlier chapters. As already noted, our main objective here is to study
the continuity of T in the topology Q.11

5.4.1 Definition An operator T on IP is called a consequence operator for

P if for every I ∈ IP the following condition holds: for every clause A ← body
in ground(P ), where T (I)(A) = ti and I(body) = tj , say, we have that the
truth table for ← yields the truth value tn , that is, true for ti ← tj .

11 We refer the reader to [Hitzler et al., 2004] for further details concerning the material

of this section.
162 Mathematical Aspects of Logic Programming Semantics

It turns out that this notion of consequence operator relates nicely to Q,

yielding the following result.

5.4.2 Theorem If T is a consequence operator for P and if for any I ∈ IP

we have that the sequence of iterates T m (I) converges in Q to some M ∈ IP ,
then M is a model for P in the sense that every clause in ground(P ) evaluates
to tn under M . Furthermore, continuity of T yields that M is a fixed point of
T.
Proof: Suppose that A ∈ BP and that M (A) = ti , and let A ← body belong
to ground(P ), where body has the form A1 , . . . , Am , ¬B1 , . . . , ¬Bm/ . Then
eventually T (T k (I))(A) = ti . Suppose M (A1 ∧ . . . ∧ Am ∧ ¬B1 ∧ . . . ∧ ¬Bm/ ) =
tj , say. Taking the sequence T k (I), we have, by the property stated in the
hypothesis (applied to each literal in the conjunction under consideration),
that eventually T k (I)(A1 ∧ . . . ∧ Am ∧ ¬B1 ∧ . . . ∧ ¬Bm/ ) = M (A1 ∧ . . . ∧ Am ∧
¬B1 ∧ . . . ∧ ¬Bm/ ) = tj . Since T (T k (I))(A) ← T k (I)(A1 ∧ . . . ∧ Am ∧ ¬B1 ∧
. . . ∧ ¬Bm/ ) is tn by the fact that T is a consequence operator, we obtain that
M (A ← A1 ∧. . .∧Am ∧¬B1 ∧. . .∧¬Bm/ ) = tn , as required. If T is continuous,
then M = lim T n+1 (I) = T (lim T n (I)) = T (M ). •

Intuitively, consequence operators propagate “truth” along the implication

symbols occurring in the program. From this point of view, we would like the
outcome of the truth value of such a propagation to be dependent only on the
relevant clause bodies. The next definition captures this intuition.

5.4.3 Definition Let A ∈ BP , and denote by BA the set of all body atoms
of clauses with head A which occur in ground(P ). A consequence operator T
is called (P -)local if for every A ∈ BP and any two interpretations I, K ∈ IP
which agree on all atoms in BA , we have T (I)(A) = T (K)(A).
It is our desire to study continuity in Q of local consequence operators.
Since Q is a product topology, it is reasonable to expect that finiteness con­
ditions will play a role in this context, as already observed in Section 3.3.

5.4.4 Definition Let C be a clause in P , and let A ∈ BP be such that A

coincides with the head of C. The clause C is said to be of finite type relative
to A if C has only finitely many different ground instances with head A. The
program P will be said to be of finite type relative to A if each clause in P is
of finite type relative to A, that is, if the set of all clauses in ground(P ) with
head A is finite. Finally, P will be said to be of finite type if P is of finite type
relative to A for every A ∈ BP .
A local variable is a variable which appears in a clause body, but not in
the corresponding head.12 It is easy to see that in the context of Herbrand
12 Local variables appear naturally in implementations, but their occurrence is awkward

from the point of view of semantics, especially if they occur in negated body literals since
this leads to the so-called floundering problem, see [Lloyd, 1987, Apt and Pedreschi, 1994].
 Supported Model Semantics 163

interpretations and if function symbols are present, then the absence of local
variables is equivalent to a program being of finite type.

5.4.5 Proposition Let P be a normal logic program of finite type, and let
T be a local consequence operator for P . Then T is continuous in Q.

Proof: Let I ∈ IP be an interpretation, let G2 = G(A, ti ) be a subbasic

neighbourhood of T (I) in Q, and note that G2 is the set of all K ∈ IP
such that K(A) = ti . We need to find a neighbourhood G1 of I such that
T (G1 ) ⊆ G2 . Since P is of finite type, the set BA is finite. Hence, the set G1 =
�
B∈BA G(B, I(B)) is a finite intersection of open sets and is therefore open.
Since each K ∈ G1 agrees with I on BA , we obtain T (K)(A) = T (I)(A) = ti
for each K ∈ G1 by locality of T . Hence, T (G1 ) ⊆ G2 . •

Now, if P is not of finite type, but we can�ensure by some other property

of P that the, possibly infinite, intersection B∈BA G(B, I(B)) is open, then
the above proof will carry over to programs which are not of finite type, but
satisfy the propert we seek. Alternatively, we would like to be able to disregard
the infinite intersection entirely under conditions which ensure that we have
to consider finite intersections only, as in the case of a program of finite type.
The following definition is, therefore, quite a natural one to make.

5.4.6 Definition Let P be a logic program, and let T be a consequence

operator on IP . We say that T is (P -)locally finite for A ∈ BP and I ∈ IP if
there exists a finite subset S = S(A, I) ⊆ BA such that we have T (J)(A) =
T (I)(A) for all J ∈ IP which agree with I on S. We say that T is (P -)locally
finite if it is locally finite for all A ∈ BP and all I ∈ IP .

Obviously, any locally finite consequence operator is local. Conversely, a

local consequence operator for a program of finite type is locally finite. This
follows from the observation that, for a program of finite type, the sets BA ,
for any A ∈ BP , are finite. But a much stronger result holds.

5.4.7 Theorem A local consequence operator is locally finite for all A ∈ BP

and some I ∈ IP if and only if it is continuous at I in Q.

Proof: Let T be a locally finite consequence operator, let I ∈ IP , let A ∈ BP ,

and let G2 = G(A, T (I)(A)) be a subbasic neighbourhood of T (I) in Q. Since
� finite, there is a finite set S ⊆ BA such that T (J)(A)
T is locally � = T (I)(A) for
all J ∈ B∈S G(B, I(B)). By finiteness of S, the set G1 = B∈S G(B, I(B))
is an open neighbourhood of I, and by the choice of S we have T (G1 ) ⊆ G2 ,
and this suffices for continuity of T at I.
For the converse, assume that T is continuous at I in Q, and let A ∈ BP
be chosen arbitrarily. Then G2 = G(A, T (I)(A)) is a subbasic open neigh­
bourhood of T (I), so that, by continuity of T , there exists a basic open neigh­
bourhood G1 = G(B1 , I(B1 )) ∩ · · · ∩ G(Bk , I(Bk )) of I with T (G1 ) ⊆ G2 . In
164 Mathematical Aspects of Logic Programming Semantics
�
other words, we have T (J)(A) = T (I)(A) for each J ∈ B∈S / G (B, I(B)),
where S ' = {B1 , . . . , Bk } is a finite set. Since T is local, the value of T (J)(A)
depends only on the values J(A) of atoms � A ∈ BA . So if we set S = S ' ∩ BA ,
then T (J)(A) = T (I)(A) for all J ∈ B ∈S G(B, I(B)), which is to say that T
is locally finite for A and I. Since A was chosen arbitrarily, we obtain that T
is locally finite for I and all A ∈ BP . •

The following corollary provides a sufficient condition13 for continuity in

Q.

5.4.8 Corollary Let P be a program, let T be a local consequence operator,

and let l : BP → ω be a level mapping for P with the property that l−1 (n)
is finite for any n ∈ ω and such that the following property holds: for each
A ∈ BP there exists an nA ∈ ω satisfying l(B) < nA for all B ∈ BA . Then T
is continuous in Q.

Proof: It follows easily from the given conditions that BA is finite for all
A ∈ BP , and hence T is locally finite. •

We turn now to the study of a particular type of local consequence op­

erator, which we call a Fitting-style operator, and its continuity. Recall from
Section 5.2.1 that bodies of pseudo-clauses may consist of infinite “disjunc­
tions”, but this will not pose any particular difficulties with respect to the
logics we are going to discuss. We note that a program P is of finite type if
and only if all bodies of all pseudo-clauses in P are finite.
Now, if we are given (suitable) truth tables for negation, conjunction, and
disjunction, then we are able to evaluate the truth values of bodies of pseudo-
clauses relative to given interpretations, as was done in Section 5.2.1.

5.4.9 Definition Let P be a normal logic program. Define the mapping FP :

IP,n → IP,n relative to a given (suitable) logic with n truth
o values by FP (I)
o =
J, where J assigns to eachoA ∈ BP the truth value I( Ci ) of the body Ci
of the pseudo-clause A ← Ci with head A.

We call operators which satisfy Definition 5.4.9 Fitting-style operators or

the FP -operator. If we impose the mild assumption that tj ← tj evaluates
to true for every j with respect to the underlying logic, then we immediately
obtain that every Fitting-style operator is a local consequence operator. We
will impose this condition, namely, that tj ← tj evaluates to true for every j,
for the remainder of this section.
If the chosen logic is classical two-valued logic, then the corresponding
Fitting-style operator is the immediate consequence operator TP (for a given
program P ). Now, if TP (I)(A) = t, then there exists a clause A ← body in
ground(P ) such that I(body) is true, and we obtain TP (J)(A) = t whenever
13 Communicated to us by Howard A. Blair.
 Supported Model Semantics 165

J(body) = t. The observation that bodies of clauses are finite conjunctions

leads us to conclude the following lemma.

5.4.10 Lemma If TP (I)(A) = t, then TP is locally finite for A and I. Fur­

thermore, TP is continuous at I if and only if it is locally finite for all A with
TP (I)(A) = f .
o
A body Ci of a pseudo-clause with head A is false in classical logic if
and only if all the Ci are false. Since TP is a Fitting-style operator, we obtain
TP (I)(A) = f if and only if all the Ci are false. If we require TP to be locally
finite for A and I, then there must be a finite set S ⊆ BA such that any J ∈ IP
which agrees with I on S renders all the Ci false. Conversely, if S ⊆ BA is
a finite set such that any J ∈ IP which agrees with I on S renders all the
Ci false, then T is locally finite for A and I. We have just established the
following theorem.14

5.4.11 Theorem Let P be a normal logic program, and let I ∈ IP . Then

TP is continuous in Q at I if and only if whenever TP (I)(A) = f , then ei­
ther there is no clause with head A or there exists a finite set S(I, A) =
{A1 , . . . , Ak , B1 , . . . , Bm } ⊆ BP with the following properties.
(a) I(Ai ) = t and I(Bj ) = f for all i and j.
(b) For every clause A ← body in ground(P ) at least one ¬Ai or at least one
Bj occurs in body.
In the case of Kleene’s strong three-valued logic, we obtain the following
lemma.

5.4.12 Lemma If ΦP (I)(A) = t, then ΦP is locally finite for A and I. Fur­

thermore, ΦP is continuous if and only if it is locally finite for all A and I
with ΦP (I)(A) ∈ {u, f }.
Similar considerations apply to the Fitting-style operators from Section
5.2.1.15 We mention in passing that the non-monotonic Gelfond–Lifschitz op­
erator is not a consequence operator in the sense discussed here, and attempts
to characterize the continuity of it involve different methods, some of which
will be studied in Chapter 6.
We will finally provide a generalization of Theorem 5.1.6 for acyclic pro­
grams. So let P be acyclic with level mapping l, and let T be a local conse­
quence operator for P . Again, we define the mapping d : IP × IP → R by
d(I, J) = 2−n , where n is least such that I and J differ on some atom A with
l(A) = n, see Definition 5.1.3 and the remarks following it. It follows from
Propositions 4.3.7 and 5.1.4 that d is a complete ultrametric on IP , a fact
which is easily verified directly.
14 A direct proof without using the notion of local finiteness was given in [Seda, 1995].
15 The operator ΨP defined by means of Belnap’s four-valued logic, see [Fitting, 2002,
Clifford and Seda, 2000], for example, is also a Fitting-style operator.
166 Mathematical Aspects of Logic Programming Semantics

5.4.13 Proposition With the hypotheses stated in the previous paragraph,

any local consequence operator T is a contraction with respect to d.

Proof: Suppose d(I, J) = 2−n . Then I and J coincide on all atoms of level
less than n. Now let A ∈ BP with l(A) = n. Then by acyclicity of P we
have that all atoms in the body of the pseudo-clause with head A are of
level less than n, and by locality of T we have that T (I)(A) = T (J)(A). So
d(T (I), T (J)) ≤ 2−(n+1) . •

We finally obtain the following theorem.

5.4.14 Theorem Let P be an acyclic program, and let T be a local conse­

quence operator for P . Then, for any I ∈ IP , we have that T n (I) converges
in Q to the unique fixed point of T .

Proof: Since d is a complete metric, we can apply Proposition 5.4.13 and

the Banach contraction mapping theorem. This yields convergence of T n (I)
in d to a unique fixed point M of T . By definition of d, the convergence of the
sequence of valuations T n (I) to M is pointwise and, hence, is also convergence
in Q. •

Theorem 5.4.14 is remarkable since the existence of a fixed point of the

given semantic operator can be guaranteed without any particular or further
knowledge about the underlying multivalued logic.

5.5 Measurability Considerations

As we shall see in Chapter 7, continuity in Q of Fitting-style operators
FP , and TP in particular, is central in relation to whether or not we can
compute them approximately by neural networks. However, in the context
of approximate computation by neural networks, the weaker notion of mea­
surability has some interest, although rather less than that of continuity, see
[Hornik et al., 1989], for example. Thus, we shall close this chapter by briefly
discussing this topic next.16
In the previous section, we defined Fitting-style operators over finite truth
sets, see Definition 5.4.9. However, unlike the case of the topology Q, finite­
ness of the truth set T is not of much importance here. Therefore, we begin
by noting that we can, in principle, work over any logic T in which the truth
value in T of disjunctions of possibly infinite countable collections of elements
16 We do not formally introduce the notion of measurability and refer to [Bartle, 1966] for

necessary background. For full details of the results we sketch here, we refer the reader to
[Seda and Lane, 2005].
 Supported Model Semantics 167

of T can be evaluated. Given this much and a normal logic program P , one
can then easily define a Fitting-style
o operator as an operator FP : IP,T → IP,T
whichosatisfies FP (I)(A) = I( Ci ) for all I ∈ IP,T and all A ∈ BP . Here,
A ← Ci is the pseudo-clause associated with A, and IP,T denotes the set
of all interpretations defined on BP taking values in T . The question then
arises of providing suitable conditions under which possibly infinite count­
able collections of truth values can be evaluated. This issue is taken up in
Section 7.6, where the notion of finitely determined disjunctions is given in
Definition 7.6.1 and is seen to be adequate for our present purposes. In fact, if
disjunctions are finitely determined, then disjunction is idempotent, commu­
tative, and associative. Furthermore, the converse of this last statement holds
if T is finite.
For a collection M of subsets of a set X, we denote by σ(M ) the smallest
σ-algebra containing M , called the σ-algebra generated by M . Recall that
a function f : X → X is measurable with respect to σ(M ) if and only if
f −1 (A) ∈ σ(M ) for each A ∈ M . If β is the subbase of a topology τ and β is
countable, then σ(β) = σ(τ ).
It turns out that Fitting-style operators are not always measurable with
respect to the σ-algebra σ(Q) generated by Q, at least if the underlying truth
set is unrestricted. However, under quite mild conditions, Fitting-style oper­
ators are always measurable, with no syntactic conditions on the program P
whatsoever, as we see next in the following result. (Note also that we make
no technical use here of the condition that tj ← tj evaluates to true for each
truth value tj ∈ T .)

5.5.1 Theorem Suppose T is a logic in which T is a countable set and

disjunctions are finitely determined. Then for any normal logic program P ,
the Fitting-style operator FP determined by P is measurable with respect to
the σ-algebra σ(Q).
As we shall see in Section 7.6, many logics of interest in logic programming
satisfy the requirement that disjunction is finitely determined. Indeed, it is
satisfied for Belnap’s logic FOU R, and hence TP , ΦP , and ΨP are all always
measurable for any normal logic program P .
Chapter 6

Stable and Perfect Model Semantics

The stable model semantics turns out to be the one which receives the most
attention these days. Some of the most popular implementations of non-
monotonic reasoning systems are based on it.1 In this chapter, we provide
means to lift our results on the supported model semantics to the stable
model semantics. This is done by the so-called fixpoint completion of pro­
grams, which we will introduce in Section 6.1. This construction will enable
us to draw almost effortlessly a number of corollaries on the stable model
semantics, and we will do this in Section 6.2. Finally, in Section 6.3, we will
close our discussion with some additional observations on stratification and
the perfect model semantics.

6.1 The Fixpoint Completion

The fixpoint completion is a program transformation which is based on
the notion of unfolding, meaning the replacement of a body atom A by the
body of a clause which also has head A. In essence, the fixpoint completion of
a given program is obtained by performing (recursively) a complete unfolding
through all positive body atoms and disregarding all clauses which after this
process still contain positive body atoms. We will describe this formally in the
following definition.

6.1.1 Definition A quasi-interpretation 2 is a set of clauses of the form

A ← ¬B1 , . . . , ¬Bm , where A and Bi are ground atoms for all i = 1, . . . , m.
Given a normal logic program P and a quasi-interpretation Q, we define
TP' (Q) to be the quasi-interpretation consisting of the set of all clauses
A ← body1 , . . . , bodyn , ¬B1 , . . . , ¬Bm for which there exists a clause A ←
A1 , . . . , An , ¬B1 , . . . , ¬Bm in ground(P ) and clauses Ai ← bodyi in Q for all
i = 1, . . . , n. We explicitly allow the cases n = 0 or m = 0 in this definition.
1 See [Leone et al., 2006] for details of the DLV system and [Simons et al., 2002] for details

of the smodels system, for example.

2 This notion is due to [Dung and Kanchanasut, 1989]. We stick to the old terminology,

although quasi-interpretations should really be thought of as, and indeed are, programs
with negative body literals only.

169
170 Mathematical Aspects of Logic Programming Semantics

Note that the set of all quasi-interpretations is a complete partial order

with respect to set-inclusion.

6.1.2 Proposition Given a normal logic program P , the operator TP' is Scott
continuous on the set of all quasi-interpretations.
Proof: We show first that TP' is monotonic. So let Q ⊆ R be quasi-
interpretations, and let A ← body be in TP' (Q). If A ← body results from
the unfolding of some clause A ← body0 in P with some clauses Bi ← bodyi
in Q, then Bi ← bodyi is contained in R for all i by assumption, and by the
existence of the clause A ← body0 in P we obtain A ← body in TP' (R) by
unfolding. If A ← body ∈ TP' (Q) does not result from some unfolding, then it
is already contained in P and, hence, in TP' (R). Thus, TP' is monotonic.
Now let Q = {Qλ | λ ∈ Λ}be an indexed directed family of quasi-interpre­
tations, and let Q = Q = Q. Since the order under consideration is set­
inclusion and TP' is monotonic, we immediately have that TP' (Q) is directed.
By the remarks following Definition 1.1.7, it therefore remains to show that
TP' (Q) ⊆ TP' (Q). So suppose that A ← body belongs to TP' (Q). If A ← body

does not result from an unfolding, then it is already contained in P , hence also
in TP' (Q). Otherwise, A ← body results from the unfolding of some A ← body0
in P with some Bi ← bodyi in Q. But then there is λ such that all Bi ← bodyi
are contained in Qλ ; hence, A ← body is contained in TP' (Qλ ) ⊆ TP' (Q), as
required. •

Given a normal logic program P , we define the fixpoint completion fix(P )

of P by fix(P ) = TP' ↑ ω.

6.1.3 Example Consider again the example program Tweety2, see Program
2.3.9. We obtain the following.
'
TTweety2 ↑0 = ∅
'
TTweety2 ↑ 1 = {penguin(tweety) ←, bird(bob) ←}
' '
TTweety2 ↑ 2 = TTweety2 ↑ 1 ∪ {bird(tweety), flies(bob) ← ¬penguin(bob)}
' '
TTweety2 ↑ 3 = TTweety2 ↑ 2 ∪ {flies(tweety) ← ¬penguin(tweety)}
'
fix(Tweety2) = TTweety2 ↑ 3.
The importance of the fixpoint completion lies in the fact that the stable
models of a given program P are exactly the supported models of fix(P ). We
can prove an even stronger result.3
3 The proof of Theorem 6.1.4 is taken directly from [Wendt, 2002a], which appeared in

compressed form as [Wendt, 2002b]. This correspondence can also be carried over to the
Fitting/well-founded semantics. More precisely, it was shown in [Wendt, 2002b] that for any
normal logic program P and any three-valued interpretation I, we have ΨP (I) = Φfix(P ) (I),
where ΨP is the operator due to [Bonnier et al., 1991] used for characterizing three-valued
stable models, but is not treated here. A corollary of the result just mentioned is that the
well-founded model for a given program P coincides with the Fitting model for fix(P ).
 Stable and Perfect Model Semantics 171

6.1.4 Theorem For any normal logic program P and (two-valued) interpre­
tation I, we have
GLP (I) = Tfix(P ) (I).

Proof: We show first that for every A ∈ GLP (I) there exists a clause in fix(P )
with head A whose body is true in I, and hence A ∈ Tfix(P ) (I). We show this
by induction on the powers of TP/I ; recall that GLP (I) = TP/I ↑ ω.
For the base case TP/I ↑ 0 = ∅, there is nothing to show.
So assume now that for all A ∈ TP/I ↑ n there exists a clause in fix(P )
with head A whose body is true in I. For A ∈ TP/I ↑ (n + 1), there exists a
clause A ← A1 , . . . , An in P/I such that A1 , . . . , An ∈ TP/I ↑ n, and hence
by construction of P/I there is a clause A ← A1 , . . . , An , ¬B1 , . . . , ¬Bm in
ground(P ) with B1 , . . . , Bm ∈ I. By our induction hypothesis, we obtain
that for each i = 1, . . . , n there exists a clause Ai ← bodyi in fix(P ) with
I |= bodyi , and hence Ai ∈ Tfix(P ) (I). So by definition of TP' the clause
A ← body1 , . . . , bodyn , ¬B1 , . . . , ¬Bm is contained in fix(P ). From I |= bodyi
and B1 , . . . , Bm ∈ I, we obtain A ∈ Tfix(P ) (I), as desired. This finishes the
induction argument, and hence GLP (I) ⊆ Tfix(P ) (I).
Now conversely, assume that A ∈ Tfix(P ) (I). We show that A ∈ GLP (I)
by proving inductively on k that TTP/ ↑k (I) ⊆ GLP (I) for all k ∈ N.
For the base case, we have TTP/ ↑0 (I) = ∅, so there is nothing to show.
So assume now that TTP/ ↑k (I) ⊆ GLP (I), and let A ∈ TTP/ ↑(k+1) (I)\TTP/ ↑k (I).
Then there is a clause A ← body1 , . . . , bodyn , ¬B1 , . . . , ¬Bm in TP' ↑ (k + 1)
whose body is true in I. Thus, B1 , . . . , Bm ∈ I, and for each i = 1, . . . , n
there is a clause Ai ← bodyi in TP' ↑ k with bodyi true in I. So Ai ∈
TTP/ ↑k (I) ⊆ GLP (I). Furthermore, by definition of TP' , there exists a clause
A ← A1 , . . . , An , ¬B1 , . . . , ¬Bm in ground(P ), and since B1 , . . . , Bm ∈ I, we
obtain A ← A1 , . . . , An ∈ P/I. Since we know that A1 , . . . , An ∈ GLP (I),
we obtain A ∈ GLP (I), and hence TTP/ ↑(k+1) (I) ⊆ GLP (I). This finishes the
induction argument, and we obtain Tfix(P ) (I) ⊆ GLP (I). •

The following corollary is an immediate consequence of Theorem 6.1.4.

6.1.5 Corollary Let P be a normal logic program. Then the stable models
of P are exactly the supported models of fix(P ).

6.2 Stable Model Semantics

Theorem 6.1.4 enables us to carry over results on the single-step operator
and on the supported model semantics to the Gelfond–Lifschitz operator, re­
spectively, the stable model semantics. We will first consider continuity issues.
The following observation is of technical importance.
172 Mathematical Aspects of Logic Programming Semantics

6.2.1 Proposition Let P be a definite logic program, let A ∈ BP , and let

n ∈ N. Then A ∈ TP ↑ n if and only if A ← is a clause in TP' ↑ n.
Proof: Let A ∈ TP ↑ n for some n ∈ N. We proceed by induction on n. If
n = 1, then there is nothing to show. So assume that n > 1. Then there is a
clause A ← body in ground(P ) such that all atoms Bi in body are contained
in TP ↑ (n − 1), and by the induction hypothesis there are clauses Bi ← in
TP' ↑ (n − 1). Unfolding these clauses with A ← body shows that A ← is also
contained in TP' ↑ n.
Conversely, assume there is a clause A ← in TP' ↑ n. We proceed again by
induction. If n = 1, there is nothing to show. So let n > 1. Then there exists
a clause A ← A1 , . . . , Ak in ground(P ) and clauses Ai ← in TP' ↑ (n − 1). By
the induction hypothesis, we obtain Ai ∈ TP ↑ (n − 1) for all i, and hence
A ∈ TP ↑ n. •

Given a program P , we know by Theorem 6.1.4 that GLP is continuous

at some I ∈ IP in Q if and only if Tfix(P ) is continuous at I. This gives rise to
the following theorem.

6.2.2 Theorem Let P be a normal logic program, and let I ∈ IP . Then

GLP is continuous at I in Q if and only if whenever GLP (I)(A) = f , then
either there is no clause with head A in ground(P ) or there exists a finite set
S(I, A) = {A1 , . . . , Ak } ⊆ BP such that I(Ai ) = t for all i and for every clause
A ← body in ground(P ) at least one ¬Ai or some B with GLP (I)(B) = f
occurs in body.
Proof: By Theorem 5.4.11 and Theorem 6.1.4, and by observing that there
are no positive body atoms occuring in fix(P ), we obtain the following.
GLP is continuous at I if and only if whenever GLP (I)(A) = f ,
then either there exists no clause with head A in fix(P ) or
there exists a finite set S(I, A) = {A1 , . . . , Ak } ⊆ BP such
that I(Ai ) = t for all i and for every clause A ← body in
fix(P ) at least one ¬Ai occurs in body.
So let P be such that GLP is continuous at I. If there is no clause with
head A in ground(P ), then there is nothing to show. So assume that there
is a clause with head A in ground(P ). Then we already know that there
exists a finite set S(I, A) = {A1 , . . . , Ak } ⊆ BP such that I(Ai ) = t for
all i and for every clause A ← body in fix(P ) at least one ¬Ai occurs in
body. Now let A ← B1 , . . . , Bk , ¬C1 , . . . , ¬Cm be a clause in ground(P ),
and assume that no ¬Ai occurs in its body. We show that there is some
Bi in body with GLP (I)(Bi ) = f . Assume the contrary, that is, that
GLP (I)(Bi ) = t for all i. Then for each Bi we have Bi ∈ GLP (I) = TP/I ↑ ω.
As in the proof of Proposition 6.2.1, we conclude that there is a clause
A ← ¬D1 , . . . , ¬Dn , ¬C1 , . . . , ¬Cm in fix(P ) with Dj ∈ I for all j = 1, . . . , n.
Since the clause A ← ¬D1 , . . . , ¬Dn , ¬C1 , . . . , ¬Cm is contained in fix(P ), we
 Stable and Perfect Model Semantics 173

know that some atom from the set S(I, A) must occur in its body. It cannot
occur as any Di because I(Dj ) = f for all j. It also cannot occur as any Ci
by assumption. So we obtain a contradiction, which finishes the argument.
Conversely, let P be such that the condition on GLP in the statement of
the theorem holds. We will again make use of the observation made at the
beginning of this proof. So let A ∈ BP with GLP (I)(A) = f . If there is no
clause with head A in fix(P ), then there is nothing to show. So assume there
is a clause with head A in fix(P ). Then there is a clause with head A in P , and
by assumption we know that there exists a finite set S(I, A) = {A1 , . . . , Ak } ⊆
BP such that I(Ai ) = t for all i and for every clause A ← body in ground(P )
at least one ¬Ai or some B with GLP (I)(B) = f occurs in body. Now let
A ← ¬B1 , . . . , ¬Bn be a clause in fix(P ) = TP' ↑ ω. Then there is k ∈ N with
A ← ¬B1 , . . . , ¬Bn contained in TP' ↑ k. Note that n = 0 is impossible since
this would imply GLP (I)(A) = t, contradicting the assumption on A. We
proceed by induction on k. If k = 1, then A ← ¬B1 , . . . , ¬Bn is contained
in ground(P ); hence, one of the Bj is contained in S(I, A), and this suffices.
For k > 1, there is a clause A ← C1 , . . . , Cm , ¬D1 , . . . , ¬Dm/ in ground(P )
and clauses Ci ← bodyi in TP' ↑ (k − 1) which unfold to A ← ¬B1 , . . . , ¬Bn .
By assumption we either have Dj ∈ S(I, A) for some j, in which case there
remains nothing to show, or we have that GLP (I)(Ci ) = f for some i. In the
latter case we obtain that bodyi is non-empty by an argument similar to that
of the proof of Proposition 6.2.1. So by assumption there is a (negated) atom
B in bodyi , and hence B is in {B1 , . . . , Bn }. So again one of the Bj is in
S(I, A), and this observation finishes the proof. •

We also have the following special instance of Theorem 6.2.2.

6.2.3 Corollary Let P be a normal logic program without local variables.

Then GLP is continuous in Q.

Proof: We apply Theorem 6.2.2. Let I ∈ IP and A ∈ BP be such that

GLP (I)(A) = f . Since P has no local variables, it is of finite type. Therefore,
the set B of all negated body atoms in clauses with head A is finite. Let
S(I, A) = {B ∈ B | I(B) = f }; then S(I, A) is also finite. If each clause
with head A contains some negated atom from S(I, A), there is nothing to
prove. So assume that there is a clause A ← A1 , . . . , An , ¬B1 , . . . , ¬Bm in
ground(P ) with Bj ∈ S(I, A) for all j, that is, suppose I(Bj ) = t for all j.
Then A ← A1 , . . . , An is a clause in P/I and A ∈ TP/I ↑ ω. It now follows that
there is some i with Ai ∈ TP/I ↑ ω = GLP (I), and this observation finishes
the argument by Theorem 6.2.2. •

Measurability is much simpler to deal with, as we see next.

6.2.4 Theorem Let P be a normal logic program. Then GLP is measurable

with respect to σ(Q).
174 Mathematical Aspects of Logic Programming Semantics

Proof: By Theorem 5.5.1 we obtain that Tfix(P ) is measurable with respect

to σ(Q), and by Theorem 6.1.4 we know that Tfix(P ) = GLP . •

The following variant of Theorem 5.4.2 can be proven directly.

6.2.5 Theorem Let P be a normal logic program, and let GLP be continuous
and such that the sequence of iterates GLnP (I) converges in Q to some M ∈ IP .
Then M is a stable model for P .

Proof: By continuity we obtain M = lim GLnP (I) = GLP (lim GLnP (I)) =
GLP (M ). •

We can also exploit our knowledge about the relationships between the
single-step operator and the Fitting operator.

6.2.6 Proposition Let P be a normal logic program, and assume that M =

Φfix(P ) ↑ ω is total.4 Then GLnP (∅) converges in Q to M + , and M + is the
unique stable model for P .

Proof: This follows immediately from Proposition 5.2.7 and Theorem 6.1.4.
•

Metric-based approaches also carry over to our present context; we restrict

our discussion to the following corollary of Theorem 5.1.6.

6.2.7 Theorem Let P be a locally stratified normal logic program with cor­
responding level mapping l. Then GLP is strictly contracting with respect to
dl . If the codomain of l is ω, then GLP is a contraction with respect to dl .
Furthermore, in both cases, GLP has a unique fixed point, and therefore P
has a unique stable model.

Proof: If P is locally stratified with respect to l, then fix(P ) is locally hierar­

chical with respect to l. It thus suffices to apply Theorem 5.1.6 in conjunction
with Theorem 6.1.4. •

6.2.8 Remark With the comments already made concerning the fact that
the well-founded model for a given program P coincides with the Fitting
model for fix(P ), for any normal program P , we can also derive the following
result.
4 We mentioned earlier in this chapter that Φ
fix(P ) coincides with the operator ΨP from
[Bonnier et al., 1991] for characterizing three-valued stable models.
 Stable and Perfect Model Semantics 175

Let P be a program with total well-founded model I ∪¬(BP \I),

where I ⊆ BP . Then GLP is strictly contracting on the spher­
ically complete dislocated generalized ultrametric space (IP , �),
where we have �(J, K) = max{dl (J, I), dl (I, K)} for all J, K ∈
IP , and l is defined by taking l(A) to be the minimal α such
that Φfix(P ) ↑ (α + 1)(A) = I(A).

Indeed, the program P has a total well-founded model in this case, and this
implies that fix(P ) has a total Fitting model. So l as just defined is, in fact,
well-defined, and fix(P ) satisfies (F) with respect to I ∪ ¬(BP \ I) and l. Now
apply Theorem 5.1.17.

6.3 Perfect Model Semantics

We return to matters of stratification and the perfect model semantics.
More precisely, we will describe an iterative method for obtaining the perfect
model for locally stratified programs.5

6.3.1 Definition Let P be a normal logic program, and let l : BP → γ be a

level mapping, where γ > 1. For each n satisfying 0 < n ≤ γ, let P[n] denote
the set of all clauses in ground(P ) in which only atoms A with l(A) < n occur,
and denote by Ln the set of all atoms A of level l(A) less than n. We define
T[n] : P(Ln ) → P(Ln ) by T[n] (I) = TP[n] (I). The mapping T[n] is called the
immediate consequence operator restricted at level n.

Thus, the idea formalized by this definition is to “cut-off” at level n.

6.3.2 Definition Let P be a locally stratified normal logic program, and let
l : BP → γ be a level mapping, where γ > 1. We construct the transfinite
sequence (In )n∈γ inductively
∞ as follows. For each m ∈ N, we put I[1,m] =
m
T[1] (∅) and set I1 = m=0 I[1,m] . If n ∈ γ, where n > 1, is a successor ordinal,
m
∞
then for each m ∈ N we put I[n,m] = T[n] (In−1 ) and set In = m=0 I[n,m] . If
 
n ∈ γ is a limit ordinal, we put In = m<n Im . Finally, we put I[P ] = n<γ In .

6.3.3 Example Consider again the example program Tweety2, Pro­

gram 2.3.9, where penguin(X) is assigned level 0, bird(X) is assigned level
1, and flies(X) is assigned level 2, for all X ∈ {tweety, bob}. We obtain the

5 For further details, we refer the reader to the paper [Seda and Hitzler, 1999b].
176 Mathematical Aspects of Logic Programming Semantics

following.

I1 = {penguin(tweety)}
I2 = I1 ∪ {bird(bob), bird(tweety)}
I3 = I2 ∪ {flies(bob)}
I[Tweety2] = I3 .

The main technical lemma we need is as follows. For its proof, which is by
transfinite induction, it will be convenient to put I[n,m] = In for all m ∈ N
whenever n is a limit ordinal; thus, statement (b) in the lemma makes sense
for all ordinals n.

6.3.4 Lemma Let P be a normal logic program which is locally stratified

with respect to the level mapping l : BP → γ, where γ > 1. Then the following
statements hold.

(a) The sequence (In )n∈γ is monotonic increasing in n.

(b) For every n ∈ γ, where n ≥ 1, the sequence (I[n,m] ) is monotonic increasing

in m.

(c) For every n ∈ γ, where n ≥ 1, In is a fixed point of T[n] .

(d) If l(B) < n and B ∈ In , where B ∈ BP , then for every m ∈ γ with n < m
we have B ∈ Im and, hence, B ∈ I[P ] . In particular, if l(B) < n and
B ∈ I[n+1,m] for some m ∈ N, then B ∈ In and, hence, B ∈ I[P ] .

Proof: It is immediate from the construction that the sequence (In )n∈γ is
monotonic increasing in n, and this establishes (a).
The main work is in proving (b) and (c), which we treat simultaneously. To
do this, we need to note the technical fact that, for each n ∈ γ, we can partition
P[n+1] as P[n] ∪ P (n), where P (n) denotes the subset of ground(P ) consisting
of those clauses whose head has level n. Thus, T[n+1] (I) = T[n] (I) ∪ TP (n) (I)
for any I ∈ IP ; note that if A ∈ TP (n) (I), then l(A) = n.
Let P(n) be the proposition, depending on the ordinal n, that (I[n,m] )
is monotonic increasing in m and that In is a fixed point of T[n] . Suppose
that P(n) holds for all n < α, where α ≤ γ is some ordinal. We must show
that P(α) holds. Indeed, P(1) holds since P[1] is a definite program and the
construction of I1 is simply the classical construction of the least fixed point
of T[1] . Therefore, we may assume that α > 2. It will be convenient to break
up the details of the case when α is a successor ordinal into the four steps (1)
to (4) below.
Case i. α = k + 1 is a successor ordinal. Thus, P(k) holds.
 Stable and Perfect Model Semantics 177

(1) We establish the recursion equations6 :

I[k+1,0] = Ik
I[k+1,m+1] = Ik ∪ TP (k) (I[k+1,m] )

and the first is immediate. Putting m = 0, we have I[k+1,1] = T[k+1] (Ik ) =

T[k] (Ik ) ∪ TP (k) (Ik ) = Ik ∪ TP (k) (Ik ) = Ik ∪ TP (k) (I[k+1,0] ), using the fact that
Ik is a fixed point of T[k] . Now suppose that the second of these equations
holds for some m > 0. Then

I[k+1,(m+1)+1] = T[k+1] (I[k+1,m+1] )

= T[k] (I[k+1,m+1] ) ∪ TP (k) (I[k+1,m+1] )
= T[k] (Ik ∪ TP (k) (I[k+1,m] )) ∪ TP (k) (I[k+1,m+1] ),

and it suffices to show that T[k] (Ik ∪ TP (k) (I[k+1,m] )) = Ik . So suppose that
A ∈ T[k] (Ik ∪ TP (k) (I[k+1,m] )). Thus, there is a clause in P[k] of the form
A ← A1 , . . . , Ak1 , ¬B1 , . . . , ¬Bl1 , where A1 , . . . , Ak1 ∈ Ik ∪ TP (k) (I[k+1,m] )
and B1 , . . . , Bl1 ∈ Ik ∪ TP (k) (I[k+1,m] ). But then level considerations and the
hypothesis concerning P imply that A1 , . . . , Ak1 ∈ Ik and B1 , . . . , Bl1 ∈ Ik .
Therefore, A ∈ T[k] (Ik ) = Ik , and the inclusion T[k] (Ik ∪ TP (k) (I[k+1,m] )) ⊆ Ik
holds. The reverse inclusion is demonstrated in like fashion, showing that the
second of the recursion equations holds with m replaced by m + 1 and, hence,
by induction on m, that it holds for all m.
(2) We have the inclusions TP (k) (Ik ) ⊆ TP (k) (Ik ∪ TP (k) (Ik )) ⊆ TP (k) (Ik ∪
TP (k) (Ik ∪ TP (k) (Ik ))) . . .. These inclusions are established by methods similar
to those we have just employed, and we omit the details.
It is now clear from this fact and the recursion equations in Step (1)
that (I[k+1,m] ), or (I[α,m] ), is monotonic increasing in m. Since monotonic
increasing sequences converge to their union in Q, and I[k+1,m] is an iterate
of Ik , it now follows by Theorem 5.4.2 that Ik+1 is a model for P[k+1] .
(3) If B ∈ BP and l(B) < k, then B ∈ Ik+1 if and only if B ∈ Ik .
Indeed, if B ∈ Ik , then it is clear from the recursion equations of Step (1)
that B ∈ Ik+1 . On the other hand, if B ∈ Ik , then it is equally clear from
the recursion equations and level considerations that, for every m ∈ N, B ∈
I[k+1,m] and, hence, that B ∈ Ik+1 , as required.
(4) Ik+1 is a supported model for P[k+1] . ∞
To see that this claim holds, suppose that A ∈ Ik+1 = m=0 I[k+1,m] . Then
m+1
there is m0 ∈ N such that A ∈ I[k+1,m+1] = T[k+1] (Ik ) for all m ≥ m0 .
m0
Thus, A ∈ T[k+1] (T[k+1] (Ik )) = T[k+1] (I[k+1,m0 ] ). Hence, there is a clause
A ← A1 , . . . , Ak1 , ¬B1 , . . . , ¬Bl1 in P[k+1] such that each Ai ∈ I[k+1,m0 ] and
no Bj ∈ I[k+1,m0 ] . But l(Bj ) < k for each j since P is locally stratified. Since
6 As shown here, it results from these equations that the process of constructing

I[k+1,m+1] in terms of I[k+1,m] is inflationary, where, formally, an operator G defined on

a collection of sets is said to be inflationary if X ⊆ G(X) for each set X in the given
collection; see also the corresponding recursion equations in Corollary 6.3.5.
178 Mathematical Aspects of Logic Programming Semantics

Bj ∈ I[k+1,m0 ] , we now see from the recursion equations that Bj ∈ Ik . From

the result in Step (3) we now deduce that, for each j, Bj ∈ Ik+1 . Since it is
obvious that each Ai belongs to Ik+1 , we obtain that A ∈ T[k+1] (Ik+1 ). Thus,
Ik+1 ⊆ T[k+1] (Ik+1 ), and therefore Ik+1 is a supported model for P[k+1] , or a
fixed point of T[k+1] , as required.
Thus, P(α) holds when α is a successor ordinal.
Case ii. α is a limit ordinal.
In this case, it is trivial that (I[α,m] ) is monotonic increasing in m.
Thus, we have only to show that Iα is a fixed point of T[α] , that is, a sup­
ported model for P[α] , and we show first that Iα is a model for P[α] . Let
A ∈ T[α] (Iα ). Then there is a clause A ← A1 , . . . , Ak1 , ¬B1 , . . . , ¬Bl1 in
P[α] such that A1 , . . . , Ak1 ∈ Iα and B1 , . . . , Bl1 ∈ Iα . Indeed, by the defi­
nition of P[α] and the hypothesis concerning P , there is n0 < α such that the
clause A ← A1 , . . . , Ak1 , ¬B1 , . . . , ¬Bl1 belongs  to P[n0 ] . Since the sequence
(In )n∈γ is monotone increasing and Iα = n<α In , there is n1 < α such
that A1 , . . . , Ak1 ∈ In1 and B1 , . . . , Bl1 ∈ In1 . Choosing n2 = max{n0 , n1 },
we have A ← A1 , . . . , Ak1 , ¬B1 , . . . , ¬Bl1 ∈ P[n2 ] and also A1 , . . . , Ak1 ∈ In2
and B1 , . . . , Bl1 ∈ In2 . Therefore, on using the induction hypothesis, we have
A ∈ T[n2 ] (In2 ) = In2 ⊆ Iα . Hence, T[α] (Iα ) ⊆ Iα , as required.
To see that Iα is supported,  let A ∈ Iα . By monotonicity of (In )n∈γ
again and the identity Iα = n<α In , there is a successor ordinal n0 ≥ 1
such that A ∈ In  for all n such that n0 ≤ n < α. In particular, we
∞
have A ∈ In0 = m=0 I[n0 ,m] . Therefore, there is m1 ∈ N such that
m1
A ∈ I[n0 ,m1 +1] = T[n0 ] (T[n 0]
(In0 −1 )). Consequently, there is a clause A ←
m1
A1 , . . . , Ak1 , ¬B1 , . . . , ¬Bl1 in P[n0 ] such that A1 , . . . , Ak1 ∈ T[n 0]
(In0 −1 ) =
I[n0 ,m1 ] ⊆ In0 ⊆ Iα and B1 , . . . , Bk1 ∈ I[n0 ,m1 ] . But l(Bj ) < n0 − 1 for each
j, and so no Bj belongs to In0 −1 by Step (3) of the previous case. Therefore,
by this step, no Bj belongs to In0 , and by iterating this we see that, for ev­
ery m ∈ N, no Bj belongs to In0 +m . Therefore, no Bj belongs to Iα . Hence,
we have A ∈ T[n0 ] (Iα ) ⊆ T[α] (Iα ) or, in other words, that Iα ⊆ T[α] (Iα ), as
required.
It now follows that P(n) holds for all ordinals n, and this completes the
proof of (b) and (c). In particular, we see that the recursion equations obtained
in Step (1) hold for all ordinals k, and we record this fact in the corollary below.
Indeed, all that is needed to establish these equations is the fact that each Ik
is a fixed point of T[k] and to note that the proof just given shows also that
I[P ] is a fixed point of TP . In turn, (d) of the lemma now follows from this
observation by iterating Step (3).
The proof of the lemma is therefore complete. •

It can be seen here, and it will be seen again later, that the importance of
(d) is the control it gives over negation in the manner illustrated in the proof
just given that Ik+1 is a supported model for P[k+1] . It is also worth noting
 Stable and Perfect Model Semantics 179

that the construction produces a monotonic increasing sequence by means of

a non-monotonic operator.7

6.3.5 Corollary Suppose the hypothesis of Lemma 6.3.4 holds. Then the
following statements hold.

(a) For all ordinals n and all m ∈ N, we have the recursion equations

I[n+1,0] = In , and
I[n+1,m+1] = In ∪ TP (n) (I[n+1,m] ).

(b) If P is, in fact, locally hierarchical, then for every ordinal n ≥ 1 we have
I[n+1,m] = In ∪ TP (n) (In ) for all m ∈ N, where P (n) is defined as in the
proof of Lemma 6.3.4, and therefore the iterates stabilize after one step.

Proof: That (a) holds has already been noted in the proof of Lemma 6.3.4.
For (b), it suffices to prove that TP (n) (In ) = TP (n) (In ∪ TP (n) (In )). So
suppose therefore that A ∈ TP (n) (In ∪ TP (n) (In )). Then there is a clause A ←
A1 , . . . , Ak1 , ¬B1 , . . . , ¬Bl1 in P (n) such that A1 , . . . , Ak1 ∈ In ∪ TP (n) (In )
and B1 , . . . , Bk1 ∈ In ∪ TP (n) (In ). From these statements and by level con­
siderations, we have A1 , . . . , Ak1 ∈ In and B1 , . . . , Bk1 ∈ In . Therefore,
A ∈ TP (n) (In ), so that TP (n) (In ∪ TP (n) (In )) ⊆ TP (n) (In ). The reverse in­
clusion is established similarly to complete the proof. •

Statement (b) of this corollary makes the calculation of iterates very easy
to perform in the case of locally hierarchical programs.

6.3.6 Theorem Suppose that P is a normal logic program which is locally

stratified with respect to the level mapping l : BP → γ. Then I[P ] is a minimal
supported model for P .

Proof: That I[P ] is a supported model for P follows from the proof of
Lemma 6.3.4, and so it remains to show that I[P ] is minimal. To do this,
we establish by transfinite induction the following proposition: “if J ⊆ I[P ]
and TP (J) ⊆ J, then In ⊆ J for all n ∈ γ, where n ≥ 1”, and this clearly
suffices. Indeed, T[1] (J) ⊆ TP (J) ⊆ J, and therefore J is a model for P[1] .
But, as already noted in proving Lemma 6.3.4, I1 is the least model for P[1]
by construction, since P[1] is definite. Therefore, I1 ⊆ J, and the proposition
holds with n = 1.
Now assume that the proposition holds for all ordinals n < α for some
ordinal α ∈ γ, where α > 1; we show that it holds with n = α.
Case i. α = k + 1 is a successor ordinal, where k > 0. We have Ik ⊆
J. We show by induction on m that I[k+1,m] ⊆ J for all m. Indeed, with
m = 0, we have I[k+1,0] = Ik ⊆ J. Suppose, therefore, that I[k+1,m0 ] ⊆ J
7 Lemma 6.3.4 plays a role here similar to that played by [Apt et al., 1988, Lemma 10].
180 Mathematical Aspects of Logic Programming Semantics
m0
for some m0 > 0. Let A ∈ I[k+1,m0 +1] = T[k+1] (T[k+1] (Ik )). Then there is
a clause A ← A1 , . . . , Ak1 , ¬B1 , . . . , ¬Bl1 in P[k+1] such that A1 , . . . , Ak1 ∈
m0
T[k+1] (Ik ) = I[k+1,m0 ] and B1 , . . . , Bl1 ∈ I[k+1,m0 ] . But l(Bj ) < k for each j.
Applying Lemma 6.3.4 (d) we see that no Bj belongs to I[P ] , and consequently
no Bj belongs to J because J ⊆ I[P ] . Since I[k+1,m0 ] ⊆ J by assumption, we
have A1 , . . . , Ak1 ∈ J. Therefore, A ∈ T[k+1] (J) ⊆ TP (J) ⊆ J, and from this
we obtain that I[k+1,m0 +1] ⊆ J, as required to complete  the proof in this case.
Case ii. α is a limit ordinal. In this case, Iα = n<α In and In ⊆ J for all
n < α by hypothesis. Therefore, Iα ⊆ J, as required.
Thus, the result follows by transfinite induction. •

We can strengthen Theorem 6.3.6.

6.3.7 Theorem Suppose that P is a normal logic program which is locally

stratified with respect to a level mapping l : BP → γ, where γ is a countable
ordinal. Then I[P ] is a perfect model for P .

Proof: Suppose that there is a model N for P which is preferable to I[P ] (and
therefore distinct from I[P ] ); we will derive a contradiction.
First note that N \ I[P ] must be non-empty; otherwise, we have N ⊆ I[P ] .
But this inclusion forces equality of N and I[P ] since I[P ] is a minimal model
for P , and therefore N and I[P ] are not distinct. This means that there is a
ground atom A in N \ I[P ] , which can be chosen so that l(A) has minimum
value; let B be a ground atom in I[P ] \ N corresponding to A in accordance
with Definition 2.5.2 and satisfying l(A) > l(B).
Next we note that T[1] (N ) ⊆ TP (N ) ⊆ N , since N is a model for P . Hence,
N is a model for P[1] , which implies that I1 ⊆ N since I1 is the least model for
the definite program P[1] . Therefore, B can be chosen so that B ∈ In0 \N , with
minimal  n0 > 1. Now n0 cannot be a limit ordinal; otherwise, we would have
In0 = m<n0 Im , from which we would conclude that B ∈ Im \ N for some
m < n0 contrary to the choice of n0 . Thus, n0 must be a successor ordinal,
and therefore B can be chosen so that B ∈ I[n0 ,m0 ] \ N , where m0 is such that
I[n0 ,m1 ] \ N = ∅ whenever m1 < m0 , ; indeed, since I1 ⊆ N , we must have
n0 > 1 and m0 ≥ 1 also. Consequently, B ∈ T[n0 ] (I[n0 ,m0 −1] )\N , showing that
there is a clause B ← C1 , . . . , Ck1 , ¬D1 , . . . , ¬Dl1 in P[n0 ] with the property
that each Ci ∈ I[n0 ,m0 −1] and no Dj ∈ I[n0 ,m0 −1] . Since l(Dj ) < n0 − 1 for
each j, we see that none of the Dj belong to I[P ] by Lemma 6.3.4 (d). But
all the Ci , if there are any, must belong to N by the choice of the numbers
n0 and m0 . Moreover, there must be at least one Dj and indeed at least one
belonging to N . For if there were no Dj or we had each Dj ∈ N , then we
would have B ∈ TPn0 (N ) ⊆ TP (N ) ⊆ N , using again the fact that N is a
model for P . But this leads to the conclusion that B ∈ N , which is contrary
to B ∈ I[P ] \ N . Thus, there is a D = Dj ∈ N \ I[P ] , for some j, satisfying
l(D) < l(B) < l(A). Since A was chosen in N \ I[P ] to have smallest level, we
have a contradiction.
 Stable and Perfect Model Semantics 181

This contradiction shows that I[P ] must be a perfect model for P , as re­
quired. •

6.3.8 Program Since locally stratified programs are a generalization of lo­

cally hierarchical programs, it is clear that each locally hierarchical program
has a unique perfect model. This does not hold, however, for Φ∗ -accessible
programs. Indeed, the program

p ← ¬q
q ← r, ¬p

is Φ∗ -accessible (even acceptable) with respect to the unique supported model

M = {p}. However, I = {q} is also a model for this program, and while I
is preferable to M , M , in turn, is also preferable to I, so P does not have a
perfect model.

We finally return to the special case of stratified programs. We temporarily

introduce the powers of an operator T mapping a complete lattice to itself:8

T ↑ 0(I) = I
T ↑ (n + 1)(I) = T (T ↑ n(I)) ∪ T ↑ n(I)

∞
T ↑ ω(I) = T ↑ n(I).
n=0

Of course, T ↑ n(I) is not equal to T n (I) unless T happens to be monotonic

and I ⊆ T (I). Indeed, the sequence (T ↑ n(I))n is always monotonic increasing
whether or not T is monotonic. However, this concept can be used to construct
an associated model MP for any stratified program P as follows. We put M0 =
∅, M1 = TP1 ↑ ω(M0 ), . . . , Mm = TPm ↑ ω(Mm−1 ). Finally, let MP = Mm .
We will show that MP is the perfect model for P , for stratified P . To do
this, it will be convenient to introduce the concept T ⇑ n(I) for a mapping
T : IP → IP and I ∈ IP . In fact, T ⇑ n(I) is defined inductively as follows:

T ⇑ 0(I) = I
T ⇑ (n + 1)(I) = T (T ⇑ n(I)) ∪ I

∞
T ⇑ ω(I) = T ⇑ n(I).
n=0

6.3.9 Theorem Let P be a stratified normal logic program. Then I[P ] = MP .

Proof: As usual, we take the stratification to be P = P1 ∪ . . . ∪ Pm , and we

will show by induction that Ik = Mk for k = 1, . . . , m and that Ik = Mm for
k > m. From this we clearly have I[P ] = Mm = MP , as required.

8 This and the following construction of MP was introduced in [Apt et al., 1988].
182 Mathematical Aspects of Logic Programming Semantics

With the definition of the level mapping we are currently using and with
the conventions we have made regarding the stratification, we note first that
the equalities P[k] = ground(P1 ∪P2 ∪. . .∪Pk ) and P (k−1) = ground(Pk ) both
hold for k = 1, . . . , m, where P (k) is as defined in the proof of Lemma 6.3.4.
Now P[1] = ground(P1 ) is definite, even if empty, and so it is immediate
that TP1 ⇑ i(M0 ) = TP1 ↑ i(M0 ) for all i and that I1 = M1 . So suppose
next that TPk+1 ⇑ i(Mk ) = TPk+1 ↑ i(Mk ) for all i and that Ik+1 = Mk+1 for
some k > 0. Then TPk+2 ⇑ 0(Mk+1 ) = Mk+1 = TPk+2 ↑ 0(Mk+1 ) and also
I[k+2,0] = Ik+1 = Mk+1 = TPk+2 ↑ 0(Mk+1 ). So now suppose that TPk+2 ⇑
m(Mk+1 ) = TPk+2 ↑ m(Mk+1 ) and that I[k+2,m] = TPk+2 ↑ m(Mk+1 ) for some
m > 0. Then TPk+2 ⇑ (m + 1)(Mk+1 ) = TPk+2 (TPk+2 ⇑ m(Mk+1 )) ∪ Mk+1
and TPk+2 ↑ (m + 1)(Mk+1 ) = TPk+2 (TPk+2 ↑ m(Mk+1 )) ∪ TPk+2 ↑ m(Mk+1 ),
and it is clear that TPk+2 ⇑ (m + 1)(Mk+1 ) ⊆ TPk+2 ↑ (m + 1)(Mk+1 ). For
the reverse inclusion, we note that under our present hypotheses we have
TPk+2 ↑ (m + 1)(Mk+1 ) = TPk+2 (TPk+2 ⇑ m(Mk+1 )) ∪ TPk+2
⇑ m(Mk+1 ), and so
it suffices to show that TPk+2 ⇑ m(Mk+1 ) ⊆ TPk+2 (TPk+2 ⇑ m(Mk+1 ))∪Mk+1 or,

in other words, that I[k+2,m] ⊆ TP (k+1) (I[k+2,m] ) ∪ Ik+1 . Since this latter set is
equal to I[k+2,m+1] by the recursion equations of Corollary 6.3.5, the inclusion
we want follows from the monotonicity of the sets I[k+2,m] relative to m. We
conclude, therefore, that TPk+2 ⇑ (m + 1)(Mk+1 ) = TPk+2 ↑ (m + 1)(Mk+1 ).
Finally, I[k+2,m+1] = Ik+1 ∪ TP (k+1) (I[k+2,m] ) = Mk+1 ∪ TPk+2 (TPk+2 ↑
m(Mk+1 )) = Mk+1 ∪ TPk+2 (TPk+2 ⇑ m(Mk+1 )) = TPk+2 ⇑ (m + 1)(Mk+1 ) =
TPk+2 ↑ (m + 1)(Mk+1 ), by the conclusions of the previous paragraph. There-
fore, I[k+2,m+1] = TPk+2 ↑ (m + 1)(Mk+1 ). From this we obtain, by induction,
the equality I[k+2,m] = TPk+2 ↑ m(Mk+1 ) for all m and with it the equality
Ik+2 = Mk+2 , as required. •

The details of the induction proof just given also establish the following
proposition.

6.3.10 Proposition Let P = P1 ∪ . . . ∪ Pm be a stratified normal logic

program. Then we have that TPk+1 ⇑ i(Mk ) = TPk+1 ↑ i(Mk ) for all i and
k = 0, . . . , m − 1.
Finally, we show that locally stratified programs have a unique perfect
model, which is also their total weakly perfect model.

6.3.11 Theorem Let P be locally stratified. Then P has a total weakly per­
fect model which is a perfect model for P . Furthermore, this model is in­
dependent of the choice of level mapping with respect to which P is locally
stratified.9
Proof: We will employ Theorem 2.5.9 to establish the claim. Let P be lo­
cally stratified with respect to some level mapping l' . Consider the equations
9 In fact, it is known that every locally stratified program has a unique perfect model,

see [Przymusinski, 1988].

 Stable and Perfect Model Semantics 183

established in Corollary 6.3.5 (a) and define the level mapping l mapping to
pairs of ordinals as follows. For A ∈ I[P ] let l(A) = (l' (A), m), where m is
least such that A ∈ I[l/ (A)+1,m+1] . For A ∈ I[P ] let l(A) = (l' (A) + 1, 0). The
recursion equations from Corollary 6.3.5 (a) together with the fact that P is
locally stratified thus allow us to conclude that (WSi), (WSiib), or (WSiic)
is always satisfied with respect
A to I[Pb] and l. Since I[P ] is total, we obtain by
Theorem 2.5.9 that I[P ] ∪ BP \ I[P ] is the (total) weakly perfect model for
P . Since every program has only one weakly perfect model, and we have just
seen that the weakly perfect model for P coincides with I[P ] , we conclude that
the model I[P ] as constructed by Theorem 6.3.7 is independent of the choice
of level mapping with respect to which P is locally stratified. •

6.3.12 Example Consider Tweety2 from Example 2.5.3 again. It is (locally)

stratified with respect to the level mapping given in Example 6.3.3. We calcu­
late the perfect model for Tweety2 by employing powers of the operator TP
as discussed just prior to the statement of Theorem 6.3.9. Indeed, with the
notation used there, we obtain

M1 = {penguin(tweety)},

M2 = {bird(bob), bird(tweety), penguin(tweety)},

M3 = MTweety2 , and

M4 = M 3.

As discussed in Example 2.5.3, the latter model is the perfect model for
Tweety2.
Chapter 7

Logic Programming and Artificial

Neural Networks

Sebastian Bader,1 Pascal Hitzler,2 and Anthony Seda3

7.1 Introduction
One of the ultimate goals of artificial intelligence is the creation of agents
with human-like intelligence, and many, varied approaches have been made
in attempts to realize this goal. Of course, an agent endowed with human­
like intelligence should be able to represent and reason with well-structured
data and processes, such as those encountered in logic or in mathematics and
related subjects, just as human beings can. On the other hand, that same
agent should also be able to represent and reason with uncertain, noisy, and
incomplete data, again, just as human beings can, at least to a certain extent.
Furthermore, the agent should be able to learn by example and refine the
reasoning process as a result.
These two aspects of the general process of reasoning and intelligence just
considered are complementary and yet are integrated in human intelligence.
Thus, their integration within a single artificial computing system is an im­
portant objective in the search for true artificial intelligence.4 Logic-based
symbolic systems are good implementations of the first, the formal, style of
reasoning, whereas neural networks or connectionist systems are good imple­
mentations of the second, less formal, style. They are therefore good candi­
dates, and indeed are among the most prominent such candidates, for attempt­
ing this integration, with each representing one of the two aspects. Certainly,
there has been a considerable amount of interest in recent years in exactly this

1 MMIS, Department of Computer Science, University of Rostock, Germany.

2 Kno.e.sisCenter for Knowledge-Enabled Computing, Wright State University, Dayton,
Ohio, USA.
3 Department of Mathematics, University College Cork, Cork, Ireland.
4 See [Hitzler and Kühnberger, 2009] for a more detailed discussion of this point.

185
186 Mathematical Aspects of Logic Programming Semantics

integration, known as neural-symbolic integration, with a view to combining

the best of both styles of reasoning within a single system.5
It will be worth contrasting a little further these two, very different, com­
puting paradigms in order to appreciate better the issues involved in their
integration. First, symbolic systems are usually based on a logic of one type
or another. They possess a declarative semantics, and knowledge can be mod­
elled in them in a human-like fashion. Thus, their use makes it easy to process
knowledge and also to handle structured objects. Unfortunately, such sys­
tems are hard to refine from real world data, which usually is noisy, and
they are hard to design if no expert knowledge is available. They are essen­
tially discrete models of computation and have been successfully used in many
applications. On the other hand, artificial neural networks are a powerful ap­
proach to machine learning, inspired by biology and neuroscience. They are
trainable from raw data, even if the data is noisy and inconsistent, and thus
are capable of adapting to new situations. They are, furthermore, robust in
the sense that they degrade gracefully: even if parts of the system fail, the
system still works. Unfortunately, they do not possess a declarative semantics
and have difficulties in handling structured data. Available (symbolic) back­
ground knowledge, which exists in many application domains, is also difficult
to use in such systems. Being modelled on natural phenomena, connectionist
systems are basically continuous models of computation, and they also have
been used successfully in many applications.
Figure 7.1 shows the Neural-Symbolic Cycle which depicts, in general
terms, our approach to the process of integration followed here. Starting from
a symbolic system, which is both readable and writable by humans, we cre­
ate a neural or connectionist system into which the symbolic knowledge is
embedded. The neural system can then be trained using powerful connection­
ist training methods, which allows modification of the rules by generalization
from raw data. If this learned or refined knowledge is later extracted from the
neural system, we obtain a readable version of the acquired knowledge.6 In
fact, it is our intention to show in this chapter how to embed knowledge about
semantic operators into connectionist systems. More specifically, we show how
semantic operators of propositional logic programs P may be computed ex­
actly by neural systems and how these same operators may be approximated
in the case of first-order programs. One consequence of this is that a neural
system acquires a sort of semantics. Another consequence is that this chapter
may be viewed as providing a model of computation for the concepts of the
previous chapters, and it deals to a certain extent with the implementation
aspects of this model. This chapter therefore is a natural continuation of the
earlier ones and gives an example of the use and application of certain of the
methods we have developed. Indeed, the notion of approximation just men­

5 See [Bader and Hitzler, 2005, Hammer and Hitzler, 2007] for overviews of the area.
6 We do not deal with knowledge extraction here, but instead refer the reader to the
papers [Jacobsson, 2005, Bader and Hitzler, 2005, Lehmann et al., 2010] for pointers to the
literature.
 Logic Programming and Artificial Neural Networks 187

writable embedding

trainable
Symbolic Connectionist
System System

readable extraction

FIGURE 7.1: The neural-symbolic cycle.

tioned occurs in the context of a theorem of Funahashi, see Theorem 7.2.2,

and employs the methods of Chapters 3 and 4 in that it casts sets of inter­
pretations into compact metric spaces. This fact permits familiar techniques
from analysis to be employed, and their occurrence is to be expected given the
continuous nature of neural systems, as already noted. Such methods using
approximation are, in fact, forced on us if we wish to employ conventional
neural networks having only finitely many neurons because, for first-order
programs P , both BP and ground(P ) are infinite sets.
Thus, the main objective of this chapter is to give a detailed account of the
foundations of neural-symbolic integration, and the main contents of the chap­
ter are as follows. First, in Section 7.2, we introduce neural networks and the
basic definitions and notation we need throughout, including the statement
of Funahashi’s theorem in the form in which we use it. Next, in Section 7.3,
we discuss in some detail the so-called core method as a general and well-
known approach to neural-symbolic integration. Indeed, it is the method we
adopt here, and it is already summarized in the previous paragraph. In Sec­
tion 7.4, we commence the study of the main topic of the chapter, namely,
the process of embedding semantic operators of logic programs into neural
networks. Thus, in Section 7.4, we start with a basic result, Theorem 7.4.1,
applying to propositional logic programs P and due originally to Hölldobler
and Kalinke [Hölldobler and Kalinke, 1994]. This result provides a procedure
which, when given a normal propositional logic program P , shows how to
construct a neural network which computes the TP -operator for P . The next
section, Section 7.5, is the heart of the chapter and takes up the issue of
the approximate computation of the TP -operator for first-order normal logic
programs P . Starting with the propositional approximation of TP based on
the previous section, we go on to study the approximate computation of TP
by sigmoidal networks, radial-basis-function networks, and vector-based net­
works, in turn, before closing the section with a discussion of the approximate
computation of the least fixed point of the TP -operator for definite normal
logic programs P . It should be noted that, thus far, we have concentrated on
the TP -operator, but we take up the study of the computation and the ap­
proximate computation of other semantic operators, and their fixed points, in
188 Mathematical Aspects of Logic Programming Semantics
v1 (t) w
k1
.
.
.
wkj
vj (t) pk (t) vk (t + ∆t)
. k
. kn
. w
vnk (t) θk

FIGURE 7.2: Unit Nk in a connectionist network.

Sections 7.6 and 7.7. In particular, in Section 7.6, we sketch the construction
of neural networks which extend Theorem 7.4.1 to compute the Fitting-style
operator FP for propositional normal logic programs P . Then, in Section 7.7,
we consider approximate computation for the operators FP and GLP , among
others, for first-order normal logic programs P .
At certain places in this chapter, the material we present is just sketched,
and detail is provided only to the extent to which it serves to outline the
application area under discussion. This is simply because the inclusion of full
detail at the places in question would lead us far astray from the main topic of
the book. We do give ample references to the literature, however, to facilitate
the reader who is interested in studying the relevant matters further.

7.2 Basics of Artificial Neural Networks

We begin by briefly summarizing what we need relating to artificial neural
networks or just neural networks for short.7

7.2.1 Definition A neural network or connectionist network 8 is simply a

weighted directed graph, or weighted digraph, endowed with extra structure,
as follows. A typical unit (or node) Nk in this digraph is shown in Figure 7.2.
We denote by Ik = {1, . . . , nk }, say, the finite set of indices j for which there
is a digraph connection from Nj to Nk , and we let wkj ∈ R denote the weight
of the digraph connection from a unit Nj to a unit Nk , if there is such a
connection, noting that wkj may be 0. Then the unit Nk is characterized,
at time t, by the following data: its input vector (ik1 (t), . . . , iknk (t)), where
ikj (t) = wkj vj (t) is the input received by Nk from Nj at time t; its thresh­
old θk ∈ R; its potential pk (t); and its value vk (t). The units are updated
synchronously; time becomes t + Δt; at each update the potential pk (t) is
calculated by means of an activation function; and the output value for Nk ,
7 Our terminology and notation are fairly standard, and the reader is referred to the pa­

pers [Hitzler et al., 2004, Fu, 1994, Hertz et al., 1991] for further details concerning neural
networks; in particular, we follow [Hitzler et al., 2004] closely here.
8 Also called a connectionist system.
 Logic Programming and Artificial Neural Networks 189

vk (t + Δt), is calculated by means of an output function whose argument is

pk (t). In fact, the activation function we will use most often in our work is the
weighted sum of the inputs uo minus the threshold. In other words, in most of
our discussions pk (t) = j∈Ik wkj vj (t) − θk ∈ R. We say that a unit Nk
becomes active at time t if pk (t) ≥ 0. On the other hand, we consider a number
of different types of units distinguished mainly by their output function, as
follows. A unit is said to be a binary threshold unit if its output function is a
threshold function or Heaviside function H, so that

1 if pk (t) ≥ 0,
vk (t + Δt) = H(pk (t)) =
0 otherwise.

A unit is said to be a linear unit if its output function is the identity as a

function of pk (t) and its threshold θ is 0. A unit is said to be a sigmoidal unit
or a squashing unit if its output function φ is non-decreasing and is such that
limx→∞ φ(x) = 1 and limx→−∞ φ(x) = 0. Such functions are called squashing
functions. •

We will only consider connectionist networks where the units can be orga­
nized in layers, although a variant of this will be encountered in Section 7.6.
A layer is a vector of units. An n-layer feedforward network F consists of the
input layer, n − 2 hidden layers, and the output layer, where n ≥ 2. Each unit
occurring in the i-th layer is connected to each unit occurring in the (i+1)-st
layer, 1 ≤ i < n. Let r and s be the number of units occurring in the input
and output layers, respectively. A connectionist network F is called a mul­
tilayer feedforward network if it is an n-layer feedforward network for some
n. A multilayer feedforward network F computes a function fF : Rr → Rs ,
called the input-output mapping of F or the network function of F, as fol­
lows. The input vector (the argument of fF ) is presented to the input layer
at time t0 and propagated through the hidden layers to the output layer. At
each time point, all units update their potential and value, as noted above.
At time t0 + (n − 1)Δt, the output vector (the image under fF of the input
vector) is read off the output layer.
For a 3-layer feedforward network with r linear units in the input layer,
squashing units in the hidden layer, and a single linear unit in the output
layer, the input-output function of the network as described in the previous
paragraph can thus be obtained as a mapping f : Rr → R with
� �
f (x1 , . . . , xr ) = cj φ wji xi − θj ,
j i

where cj is the weight associated with the connection from the j-th unit of the
hidden layer to the single unit in the output layer, φ is the squashing output
function of the units in the hidden layer, wji is the weight associated with the
190 Mathematical Aspects of Logic Programming Semantics

connection from the i-th unit of the input layer to the j-th unit of the hidden
layer, and θj is the threshold of the j-th unit of the hidden layer.
It is our aim to establish results in the following sections on the represen­
tation and approximation of various semantic operators, the TP -operator in
particular, by input-output functions of 3-layer feedforward networks. Some
of our results rest on the following theorem, which is due to Funahashi, see
[Funahashi, 1989].

7.2.2 Theorem (Funahashi) Suppose that φ : R → R is a non-constant,

bounded, monotone increasing and continuous function. Let K ⊆ Rn be com­
pact, let f : K → R be a continuous function, and let ε > 0. Then there exists a
3-layer feedforward network F with squashing function φ whose input-output
mapping fF : K → R satisfies maxx∈K d(f (x), fF (x)) < ε, where d is a metric
which induces the natural topology9 on R.

In other words, each continuous function f : K → R can be uniformly

approximated by input-output functions of 3-layer (feedforward) networks.
Furthermore, on a point of terminology, suppose given ε > 0. We will write
Y approximates X up to ε if d(Y, X) < ε, where d is some appropriate
metric for the objects X, Y in question.10 There are two cases here where
the definition just given will be applied, as follows. In the first case, X is a
semantic operator and Y is an operator which we are using to approximate X;
d is either the uniform metric used in Theorem 7.2.2 or the metric λ discussed
in Section 7.5.2. In the other case, X is a fixed point of a semantic operator and
Y is an interpretation which we are using to approximate X; d is the metric
dl determined by a level map (taking values in ω) as in Definition 5.1.3, see
again Section 7.5.2 and also Section 7.5.6. We will paraphrase the import of
Theorem 7.2.2, noting that it holds for all ε > 0, by writing that approximating
networks exist for f . Furthermore, for our purposes later, it will suffice to
assume that K is a compact subset of the set of real numbers, so that n can
be taken to be equal to 1 in the statement of the theorem.
An n-layer recurrent network F consists of an n-layer feedforward network
such that the number of units in the input layer is equal to the number of units
in the output layer. Furthermore, each unit in the output layer is connected
with weight 1 to the unit in the corresponding position in the input layer.
Figure 7.3 shows a 3-layer recurrent network. The subnetwork consisting of
the three layers and the connections between the input and the hidden layer
as well as between the hidden and the output layer is a 3-layer feedforward
network called the kernel of F.
Notice that any neural network in which the number of units in the input
layer is equal to the number of units in the output layer can be made recur­
rent just by adding the necessary obvious connections with weight 1. Notice
9 For example, d(x, y) = |x − y|.
10 The fact that d is symmetric will not render this definition ambiguous, because in
practice it will be clear which object is which.
 Logic Programming and Artificial Neural Networks 191

FIGURE 7.3: Sketch of a 3-layer recurrent network containing, from left to

right, 3 input, 4 hidden, and 3 output units and showing also the recurrent
connections from output layer to input layer.

also that a recurrent network can perform iterated computations because the
output values can be returned to the input layer via the connections just de­
scribed; it can thus perform computation of the iterates T k (I), k ∈ N, for
example, where I is an interpretation and T is a semantic operator.

7.3 The Core Method as a General Approach to

Integration
In this section, we outline the idea underlying the approach presented be­
low. Suppose given a normal logic program P and any one of the semantic
operators TP : IP → IP we have thus far associated with P , using TP and
IP as generic symbols for a semantic operator and its underlying set of in­
terpretations. For simplicity, we assume the interpretations in question are
Herbrand interpretations taking values in a truth set T , although the con­
clusions we make here are valid over any preinterpretation J whose domain
D is countable. Can one find, or at least show the existence of, a multilayer
feedforward network FP which computes TP in some sense? Furthermore, can
this network FP , or some other appropriate network, compute the least fixed
point of TP assuming the least fixed point of TP exists?
A few general remarks are in order at this point. To begin with, multi-
layer feedforward networks, even 3-layer feedforward networks, are known to
be extremely powerful computing devices and indeed are known to be univer­
sal approximators in the sense made precise in the statement of Funahashi’s
theorem, Theorem 7.2.2, earlier.11 Therefore, one might expect them to have
the ability to carry out the required computations, and this is so. Indeed, sup­
pose that P is a first-order program and endow IP with the Cantor topology,
11 See [Funahashi, 1989, Hornik et al., 1989] for full details.
192 Mathematical Aspects of Logic Programming Semantics

assuming that the set T of truth values is finite. Then we obtain a compact
Hausdorff space homeomorphic to the Cantor subset of the unit interval in the
real line as shown in Theorem 3.3.4. Thus, whenever TP is continuous in the
Cantor topology on IP (see Theorem 7.5.3), we can apply Theorem 7.2.2, tak­
ing f = fFP , taking K = IP , and given a value of ε > 0, to assert the existence
of a 3-layer feedforward network satisfying the conclusion of Theorem 7.2.2.
Furthermore, by making such a network recurrent, it can also compute iter­
ates of TP provided that conditions prevail under which the error estimate
is uniformly well-behaved relative to ε under iteration. Again, under suitable
conditions and with a suitable choice of initial input I0 ∈ I (perhaps the bot­
tom element of I), the iterates fFnP (I0 ) will converge to a fixed point (perhaps
the least) of TP , and these observations will be examined in Sections 7.5.2 and
7.5.6, see also Corollary 7.4.3. Finally, as one might expect, if P is actually a
propositional program, then the need for approximation disappears, and in­
deed a 3-layer network can be constructed which actually computes TP and,
again under suitable conditions, computes fixed points of TP . In fact, in the
case of propositional programs, networks of binary threshold units suffice for
these purposes, as we shall see. This general method is nowadays known as
the core method, and a number of instances of it are presented in the following
sections.
It is important to note that the proof of Theorem 7.2.2 is non-constructive,
and much of our work in the following sections of this chapter is concerned with
the problem of constructing suitable approximations to semantic operators
in the case of first-order programs.12 However, we will begin by discussing
propositional programs in these terms in the next section.

7.4 Propositional Programs

The previous section delineates the problem we wish to study in this chap­
ter, and we begin by studying the propositional case first relative to the imme­
diate consequence operator. Before doing this however we note that networks
yet simpler than those just described, namely, 2-layer feedforward networks
of binary threshold units, do not in general suffice to compute the immediate
consequence operator for (definite) propositional logic programs, although we
give no details of this claim here.13
We now present the main result of this section.14

12 We know of no constructive proof of Theorem 7.2.2 and refer the reader to the papers

[Cybenko, 1989, Funahashi, 1989, Hornik et al., 1989] for well-known versions of the proof.
13 See [Hitzler et al., 2004] for a discussion of this fact.
14 This result was first established in [Hölldobler and Kalinke, 1994]; here, and in the rest

of this section, we follow [Hitzler et al., 2004].

 Logic Programming and Artificial Neural Networks 193

7.4.1 Theorem For each propositional normal logic program P , a 3-layer

feedforward network can be constructed which computes the immediate con­
sequence operator TP .

Proof: Let m and n be the number of propositional variables and the number
of clauses occurring in P , respectively. Without loss of generality, we may
assume that the variables are ordered. The network associated with P can
now be constructed by the following translation algorithm.

(1) Both the input and output layers are vectors of binary threshold units of
length m, where the i-th unit in either of these layers represents the i-th
variable, 1 ≤ i ≤ m. The threshold of each unit occurring in the input or
output layer is set to 0.5.

(2) For each clause of the form A ← L1 , . . . , Lk , k ≥ 0, occurring in P , do

the following.

(2.1) Add a binary threshold unit c to the hidden layer.

(2.2) Connect c to the unit representing A in the output layer with weight
1.
(2.3) For each literal Lj , 1 ≤ j ≤ k, connect the unit representing Lj in
the input layer to c and, if Lj is an atom, then set the weight to 1;
otherwise, set the weight to −1.
(2.4) Set the threshold θc of c to l − 0.5, where l is the number of positive
literals occurring in L1 , . . . , Lk .

Each interpretation I for P can be represented by a binary vector

(v1 , . . . , vm ). Such an interpretation is given as input to the network by exter­
nally activating corresponding units of the input layer at time t0 . It remains
to show that TP (I)(A) = t if and only if the unit representing A in the output
layer becomes active at time t0 + 2Δt.
If TP (I)(A) = t, then there is a clause A ← L1 , . . . , Lk in P such that
for all 1 ≤ j ≤ k we have I(Lj ) = t. Let c be the unit in the hidden layer
associated with this clause according to (2.1) of the construction. From (2.3)
and (2.4) we conclude that c becomes active at time t0 + Δt. Consequently,
(2.2) and the fact that units occurring in the output layer have a threshold of
0.5 (see Step (1) of the construction) ensure that the unit representing A in
the output layer becomes active at time t0 + 2Δt.
Conversely, suppose that the unit representing the atom A in the output
layer becomes active at time t0 + 2Δt. From the construction of the network,
we find a unit c in the hidden layer which must have become active at time
t0 + Δt. This unit is associated with a clause A ← L1 , . . . , Lk . If k = 0,
that is, if the body of the clause is empty, then, according to (2.4), c has
a threshold of −0.5. Furthermore, according to (2.3), c does not receive any
input, that is, pc = 0+0.5, and consequently c will always be active. Otherwise,
if k ≥ 1, then c becomes active only if each unit in the input layer representing
194 Mathematical Aspects of Logic Programming Semantics
A A A 1 A
0.5 0.5 0.5 −0.5 0.5

1
0.5

1
-1
-1
B B B -1 B

-1
0.5 0.5 0.5 0.5 0.5

1
1
0.5

1
1
1
C C C 1 C
0.5 0.5 0.5 0.5 0.5

FIGURE 7.4: Two 3-layer feedforward networks of binary threshold units

computing TP1 and TP2 , respectively. Only connections with non-zero weight
are shown. The numbers occurring within units denote thresholds.

a positive literal and no unit representing a negative literal in the body of

the clause is active at time t0 (see (2.3) and (2.4)). Hence, we have found a
clause A ← L1 , . . . , Lk such that for all 1 ≤ j ≤ k we have I(Lj ) = t, and
consequently TP (I)(A) = t. •

7.4.2 Example As an example of Theorem 7.4.1, consider the following two

programs P1 (on the left) and P2 (on the right):

C ← A, ¬B A←
C ← ¬A, B C ← A, ¬B
C ← ¬A, B

Their corresponding connectionist networks are shown in Figure 7.4. One

should observe that P2 exemplifies the representation of unit clauses in 3­
layer feedforward networks.15
It is worth noting that the number of units and the number of connections
in a network F corresponding to a program P are bounded by O(m + n) and
O(m × n), respectively, where m is the number of propositional variables and
n is the number of clauses occurring in P . Furthermore, TP (I) is computed in
two steps. As the sequential time to compute TP (I) is bounded by O(n × m)
(assuming that no literal occurs more than once in the conditions of a clause),
the parallel computational model is optimal.16
We mention in passing and in the context of Theorem 7.4.1 that one can
apply the Banach contraction mapping theorem, Theorem 4.2.3, to obtain the
following result.

7.4.3 Corollary Let P be a normal propositional logic program such that

15 We can save the unit in the hidden layer corresponding to the unit clause if we change

the threshold of the unit representing A in the output layer to −0.5.

16 A parallel computational model requiring p(n) processors and t(n) time to solve a

problem of size n is optimal if p(n) × t(n) = O(T (n)), where T (n) is the sequential time to
solve this problem, see, for example, [Karp and Ramachandran, 1990].
 Logic Programming and Artificial Neural Networks 195

TP is a contraction with respect to some (necessarily complete) metric. Then

a 3-layer recurrent network can be constructed such that each computation,
starting with an arbitrary initial input, converges and yields the unique fixed
point of TP or, in other words, yields the unique supported model for P .
Indeed, there is even a kind of converse of Corollary 7.4.3 also, as follows.
Let P be a propositional logic program such that the corresponding network
has the property that each computation starting with an arbitrary initial input
converges, and in all cases converges to the same state. Then it results that
iteration of the TP -operator exhibits the same behaviour, that is, for each
initial interpretation it yields one and the same constant value after a finite
number of iterations. This fact suffices to guarantee the existence of a complete
metric which renders TP a contraction, and the claim therefore follows.17
Returning to the programs P1 and P2 again, we observe that the asso­
ciated TP -operators are contractions.18 Hence, Figure 7.4 shows the kernels
of corresponding recurrent networks which compute the least fixed point of
TP1 (the interpretation represented by the vector (0, 0, 0)) and of TP2 (the
interpretation represented by the vector (1, 0, 1)).
The time needed by the network to settle down into the unique stable state
is equal to the time needed by a sequential machine to compute the least fixed
point of TP in the worst case. As an example, consider the definite program
P3 as follows, where 1 ≤ i < n

A1 ←

Ai+1 ← Ai

The least fixed point of TP3 is the interpretation which evaluates each Ai ,
1 ≤ i ≤ n, to t, and it can be computed in O(n) steps.19 Obviously, the
parallel computational model needs as many steps. More generally, let P be a
propositional definite program containing n clauses. The time needed by the
network to settle down into the unique stable state is 3n in the worst case,
and thus, the time is linear with respect to the number of clauses occurring in
the program. This comes as no surprise as satisfiability of propositional Horn
formulae is P-complete and, thus, is unlikely to be in the class NC.20 On the
other hand, consider the program P4 containing the following clauses
Ai ←
Ai+1 ← Ai
17 See [Hitzler and Seda, 2001, Bessaga, 1959, Jachymski, 2000]; a direct proof of this ob­

servation is given in [Hölldobler and Kalinke, 1994].

18 These programs are actually acceptable, as can be seen by mapping C to 2 and A as

well as B, to 1 and considering the model I(A) = I(C) = t and I(B) = f .

19 Using techniques described in [Dowling and Gallier, 1984] and [Scutellà, 1990]. To be

more precise, the algorithm described in [Dowling and Gallier, 1984] needs O(n) time, where
n denotes the total number of occurrences of propositional variables in the formula.
20 See, for example, [Jones and Laaser, 1977] and [Karp and Ramachandran, 1990].
196 Mathematical Aspects of Logic Programming Semantics

where 1 ≤ i ≤ n and i is even. The least model for P4 maps each atom to t
and is computed in five steps by the recurrent network corresponding to P4 .
We note that the networks constructed by the translation algorithm pre­
sented previously cannot be trained by the usual learning methods applied
to connectionist systems. It was observed in [d’Avila Garcez et al., 1997] (see
also [d’Avila Garcez and Zaverucha, 1999, d’Avila Garcez et al., 2002]) that
results similar to Theorem 7.4.1 and Corollary 7.4.3 can be obtained if the
binary threshold units occurring in the hidden layer of the feedforward kernels
are replaced by sigmoidal units. We omit the technical details here and refer
to the above cited literature. Such a move renders the kernels accessible to
the backpropagation algorithm, a standard technique for training feedforward
networks [Rumelhart et al., 1986].

7.5 First-Order Programs

A central problem for neural-symbolic integration is the determination of
a good representation of first-order rules within a connectionist setting. Such
a representation would result, at least, in the computation or approximation
of the associated semantic operators. That approximating networks exist for
the immediate consequence operators of acyclic logic programs was the first
result obtained in this regard, see [Hölldobler et al., 1999], but it was shown
with the help of Funahashi’s theorem, which is non-constructive as we have
already observed. In this section, we outline the ideas underlying the general
problem and also discuss different constructive approaches to it. But before
going into details, we need to answer the following questions.

• Why do we need to approximate operators such as the TP -operator?

• What does approximation mean in our context?

The first question is easily answered: even a single application of the TP ­

operator can lead to infinite results. For example, assume P is a program
containing the fact p(X). Applying the TP -operator once (to an arbitrary
interpretation) leads to a result containing infinitely many atoms, namely, all
p(X)-atoms for every X. In this simple example, we might be able to represent
this particular result in a finite way, but things might become arbitrarily
complex for other programs using the same or similar representations.21
21 Indeed, the so-called rational models were developed to tackle this representational

problem for certain programs, see [Bornscheuer, 1996]. Unfortunately, there is no way to
compute an upper bound on the size of this rational representation, and hence it does not
give us any immediate advantages. Because we are not aware of any other finite represen­
tation, we will concentrate here on the standard representation using Herbrand interpreta­
tions.
 Logic Programming and Artificial Neural Networks 197

In principle, there are two ways to approximate a given TP -operator. On

the one hand, we can design an approximating function to meet a given level
of accuracy. This leads, as accuracy increases, to increasing numbers of units
in the hidden layer in the resulting networks, and we call this method approx­
imation in space. The approaches presented in this section follow this line of
attack. Alternatively, we can construct a system which approximates a single
application of the TP -operator better and better the longer it runs, and we
call this method approximation in time.22
Our discussion here has concentrated on the operator TP , but all our con­
siderations apply equally well to any of the other semantic operators we have
studied, and we will return to this point in Sections 7.6 and 7.7. However,
unless stated to the contrary, for a given normal logic program P , we will
focus on the operator TP and the space IP of two-valued interpretations in
Section 7.5.1 through to Section 7.5.6.

7.5.1 Feasibility of the First-Order Approach

As mentioned previously, it is well-known that multilayer feedforward net­
works are universal approximators for certain real functions and, in particular,
for all continuous real functions on compact subsets of Rn . Hence, if we can
find a suitable way of representing first-order interpretations by (finite vectors
of) real numbers, say, then feedforward networks may be used to approxi­
mate the meaning function of suitable programs. It is necessary of course that
such representations are compatible with both the logic-programming and the
neural-network paradigms.

7.5.1 Program (Even2) We use the following variant of the program Even,
Program 2.1.3, as a running example. The equations on the right define a level
mapping l assigning odd numbers to even(si (a))-atoms and even numbers to
odd(si (a))-atoms.

even(a) ← l(even(si (a))) := 2i + 1

even(s(X)) ← odd(X) l(odd(si (a))) := 2i + 2
odd(X) ← ¬even(X)

We next define a homeomorphic embedding of the space of interpreta­

tions of a given normal logic program into some (compact) subset of the real
numbers. In doing this, we use level mappings23 to realize this embedding.
For much of this chapter, although not everywhere, we assume that the level
mapping in question is bijective, even though some of the results we discuss
can be extended to the case of non-bijective level mappings.24
22 This method was employed in [Bader and Hitzler, 2004] and [Bader et al., 2005a].
23 We are following [Hölldobler et al., 1999] here.
24 See [Seda, 2006], for example, where the requirement on level mappings l : B → ω is
P
the already familiar one that l−1 (n) be a finite set for each n.
198 Mathematical Aspects of Logic Programming Semantics

IP
TP
_ IP

ι ι

1 1
K _K
fP

FIGURE 7.5: Transforming TP into fP .

7.5.2 Definition Let l : BP → ω be a bijective level mapping defined on the

Herbrand base BP of some normal logic program P , and let b be a natural
number such that b > 2. We define a function ι on IP by setting

ι(I) = b−l(A)
A∈I

for each I ∈ IP .

In fact, ι(I) gives a binary representation in the number system with base b
to each interpretation I, and moreover ι is an embedding of IP into the number
system with base b. It is straightforward to show that ι is a homeomorphism,
and it follows from Theorem 3.3.4 that not only is the set K ⊂ [0, 1] of all
embedded interpretations compact, but that it is also homeomorphic to the
Cantor set whenever IP is endowed with the Cantor topology. Using ι, we can
construct the real-valued version fP = ι(TP ) of the immediate consequence
operator TP by defining fP (x) := ι(TP (ι−1 (x))) or, in other words, by forcing
the diagram in Figure 7.5 to commute.
Furthermore, since ι is a homeomorphism, it follows that fP is contin­
uous if and only if TP is continuous in the Cantor topology on IP . Now,
using Funahashi’s result, Theorem 7.2.2, we can conclude that approximating
networks exist for suitable programs, namely, those for which the immediate
consequence operator TP is continuous in the Cantor topology on IP .
Conversely, suppose that P is a normal logic program and that approx­
imating networks exist for TP . Then TP must be continuous in the Cantor
topology on IP , and we have the following theorem.25

7.5.3 Theorem Suppose that P is a normal logic program. Then approxi­

mating networks exist for TP if and only if TP is continuous in the Cantor
topology on IP .
25 See [Seda, 2006, Theorem 3.24]. In fact, the theorem just cited was established for

Fitting-style operators (over finite truth sets, not just for two truth values).
 Logic Programming and Artificial Neural Networks 199

0.3̄

0.3̄

FIGURE 7.6: The embedding of the TP -operator for Program 7.5.1.

Thus, at this point, we know that approximating networks exist for suitable
normal logic programs, but we do not yet know how to construct them. This
issue will be taken up in the following sections.
Before discussing the constructions in detail, we will take a closer look
at the space of embedded interpretations and at the embedding of the TP ­
operator associated with Program 7.5.1. Using the embedding ι defined above
with b = 3 and taking the level mapping shown in Program 7.5.1, we obtain
the embedding of the TP -operator shown in Figure 7.6. As already mentioned
earlier, the space IP of interpretations is homeomorphic to the Cantor set. This
can also be seen by looking at the domain of the graph shown in Figure 7.6.

7.5.2 First-Order Programs by Propositional Approximation

By completely grounding a first-order program P , that is, by forming the
set ground(P ), we obtain a de facto propositional version of it. In particular,
the associated immediate consequence operators of P and of ground(P ) are
identical. Unfortunately, the ground version of most programs of interest turns
out to be an infinite set. Nevertheless, it is a major point to make that we
can approximate the immediate consequence operator of P by taking the
immediate consequence operator of a subset of ground(P ) instead, and we
consider this process now.
It will be helpful to say first a few words about the metrics which are
useful in the process.26 Suppose l : BP → ω is a level mapping,27 and form
the metric dl induced by l, see Definition 5.1.3. Then we can define a metric
λ on the set of all mappings from IP to IP by28

λ(f, g) = sup dl (f (I), g(I)),

I∈IP

for f, g : IP → IP . Similarly, we write |ι(f ) − ι(g)| to denote the uniform

metric supx∈K |ι(f )(x) − ι(g)(x)| defined on the set of all functions mapping
K into itself. Of course, the definition for λ just given can be made generally
26 We refer the reader to [Seda, 2006, Section 3.1] for more details.

27 Itis enough for l to satisfy the property that l−1 (n) is finite for each n.

28 The supremum can be replaced by maximum if f and g are continuous.

200 Mathematical Aspects of Logic Programming Semantics

and not just for dl , but this suffices for what we want to say here. Now, given
a level n, we form the subset Pn of ground(P ) containing all those clauses
whose heads have level ≤ n. Then, for all A ∈ BP with l(A) ≤ n and for all
I ∈ IP , we have A ∈ TPn (I) if and only if A ∈ TP (I), or equivalently, by
definition of dl , we have dl (TPn (I), TP (I)) ≤ 2−(n+1) for all I ∈ IP . Hence,
λ(TPn , TP )o≤ 2−(n+1) . Now suppose that ε > 0 is given. Choose n ∈ N so
large that i>n b−i < ε, and form Pn . Then for all I ∈ IP , TPn (I) and TP (I)
agree on all atoms A with l(A) ≤ n. Therefore, the expansions ι(TPn (I)) and
ι(TP (I)) agree in their first n terms. Hence, for all I ∈ IP we have, from
Figure 7.5, that

|fPn (ι(I)) − fP (ι(I))| = |ι(TPn (I)) − ι(TP (I))| < ε.

In other words, given any ε > 0, we obtain the approximation |fPn − fP | < ε
provided n is sufficiently large. In addition, approximation can be thought of in
terms of dl and λ at the level of interpretations and of TP itself independently
of the embedding ι chosen. We refer to this process of working with Pn as
approximating TP up to level n, and we will see shortly that it can be used to
show that approximating networks exist for TP for certain programs P . Indeed,
in this terminology the estimates just made show that TPn approximates TP
up to ε provided TPn approximates TP up to level n for large enough n.
Unfortunately, the subsets Pn of ground(P ) which, as we have just seen,
are appropriate for approximation can be infinitely large. For example, there
are infinitely many ground instances of the clause a ← p(X). Therefore, we
consider only so-called covered logic programs in the rest of this section, ex­
cluding Section 7.5.6, and we define the notion of a covered program next.

7.5.4 Definition A logic program is called covered if it has no local variables,

that is, if every variable symbol occurring in the body of a clause also occurs
in the head of the same clause.

7.5.5 Proposition Let P be a covered logic program, let l be a bijective level

mapping from BP to ω, and let n ∈ ω be fixed. Then the program Pn defined
above by

Pn := {C | C ∈ ground(P ) with l(H) ≤ n, where H is the head of C}

is finite.

Proof: The finiteness of Pn follows directly from the fact that, for a given
level m, there is at most one ground clause C whose head has level m. •

Using this finiteness property, we can directly obtain the following theorem
showing the existence of approximating networks for a given covered logic
program.
 Logic Programming and Artificial Neural Networks 201

7.5.6 Theorem Let P be a covered logic program, and let n ∈ N. Then we

can construct a 3-layer feedforward network whose network function approx­
imates TP up to level n.

Proof: We can obtain such an approximating network by

(1) Constructing Pn as defined above.

(2) Using the construction presented in the proof of Theorem 7.4.1 to obtain
a network computing TPn .

Since TPn coincides with TP for all atoms of level ≤ n, we conclude that the
network we have constructed approximates TP up to level n, as required. •

7.5.7 Example Take P to be Program 7.5.1 introduced earlier. We obtain

the corresponding program Pn by means of the level mapping defined in Pro­
gram 7.5.1. The level of the head atom of the clauses is shown below on the
right.

P1 = {even(a) ←} l(even(a)) = 1
P2 = {even(a) ←, l(even(a)) = 1
odd(a) ← ¬even(a)} l(odd(a)) = 2
P3 = {even(a) ←, l(even(a)) = 1
odd(a) ← ¬even(a), l(odd(a)) = 2
even(s(a)) ← odd(a)} l(even(s(a))) = 3

The corresponding networks are shown in Figure 7.7.

7.5.3 Approximation by Sigmoidal Networks

In this section, we take a different approach to the approximation of the
embedded meaning function. We start by presenting the underlying intuitions
and continue with a detailed discussion.29
Using the embedding ι defined earlier for b = 3 and the level mapping
shown in Program 7.5.1, we obtain the embedding of the TP -operator shown
in Figure 7.8 on the left. Under the condition that P is covered and the
level mapping l is bijective, we can approximate this graph using a set of
appropriately chosen constant pieces. These, in turn, can be computed as a
sum of threshold functions, shown in Figure 7.8 in the middle. By replacing
the threshold functions by sigmoidals, we obtain an approximation which can
directly be implemented within a neural network.
29 The interested reader is referred to [Bader et al., 2005b] and [Bader, 2009] for further

details and for implementations.

202 Mathematical Aspects of Logic Programming Semantics

e(a) 1.0 e(a)

P1 : 0.5 −0.5 0.5

e(a) 1.0 e(a)

P2 : 0.5 −0.5 0.5

-1.0
o(a) 1.0 o(a)
0.5 0.5 0.5

e(a) 1.0 e(a)

P3 : 0.5 −0.5 0.5
-1.0

o(a) 1.0 o(a)

0.5 0.5 0.5
1.0

e(s(a)) 1.0 e(s(a))

0.5 0.5 0.5

FIGURE 7.7: The networks corresponding to P1 , P2 , and P3 from Exam­

ple 7.5.7.

0.3̄ 0.3̄ 0.3̄

0.3̄ 0.3̄ 0.3̄

FIGURE 7.8: The embedding of the TP -operator of Program 7.5.1 is shown

on the left. In the middle and on the right, approximations using threshold
and sigmoidal functions are depicted.
 Logic Programming and Artificial Neural Networks 203

1. Approximate the embedded TP -operator using constant pieces. As be­

fore, we start by constructing Pn for a given level n. After embedding
the approximating operator TPn , we find that the resulting function is
a piecewise constant function. Due to the finiteness of the resulting pro­
gram, we obtain the greatest relevant input level by taking the maximal
level of an atom occurring in any of the bodies. Since no atom of a
greater level influences the result of the TP -operator, we see that it is a
piecewise constant function.

2. Approximating the embedded TP -operator using threshold functions. Ob­

viously, every piecewise constant function R → R can be represented as
a sum of (parametrized) threshold functions. To approximate the em­
bedded TP -operator of Program 7.5.1 up to level 3, we need the three
0.016 −0.078 0.016
functions: H0.042 , H0.167 , H0.292 , where Hpy (x) := y · H(x − p) denotes
an h-step at position p.

3. Approximating the embedded TP -operator using sigmoidal functions. To

enable the construction of sigmoidal networks, we need to replace the
threshold functions with sigmoidal functions. This can be done because
(a) we are only interested in the approximation of embedded interpre­
tations, and (b) we can place the threshold functions so that the jumps
are located between two embedded interpretations. First, we construct
the threshold approximation not for the greatest relevant input level
n as introduced earlier, but up to level n + 1. Every approximation of
this function up to ε' := b−(n+1) results in a sufficient approximation
of the embedded TP -operator. Under these conditions, we can replace
the threshold functions by appropriately set up sigmoidal functions. We
just need to make sure that the sigmoidal functions approximate the
threshold functions on all embedded interpretations up to ε' . For the
example of Program 7.5.1, see also Example 7.5.7, we obtain the follow­
0.016 −0.078 0.016
ing sigmoidal functions: S0.042,135.994 , S0.167,53.864 , S0.292,135.994 , where
h h
Sp,s (x) := 1+e−s(x−p) .

4. Approximating the embedded TP -operator using a sigmoidal network.

The approximating sigmoidal functions constructed in Step 3 can easily
be embedded into a standard 3-layer sigmoidal network as follows: the
input and output layer contain exactly one unit computing the identity
function. The hidden layer contains a sigmoidal unit for every sigmoidal
function constructed in Step 3. The weights from input to hidden layer
are set up such that they represent the steepness of the constructed sig­
moidal. The thresholds of the hidden layer correspond to the locations
of the sigmoidal functions, and the weights from hidden to output layer
coincide with the step width of the underlying threshold functions.

Figure 7.9 shows the resulting network for ε = 0.04 corresponding to an

approximation of the TP -operator up to level 3.
204 Mathematical Aspects of Logic Programming Semantics
−5.67

0.01
.99
135

6
−8.98
i 53.864 -0.078 o
0.5 0.32

135

6
0.01
.99
−39.66

4
FIGURE 7.9: An approximating sigmoidal network for Program 7.5.1.

We are now in a position to state the following theorem.30

7.5.8 Theorem Let P be a covered logic program, let b > 2, and let ε > 0.
Then we can construct a 3-layer feedforward sigmoidal network whose network
function approximates TP up to ε.
Both approaches presented in the last two sections are based on a subset of
ground(P ) and embedding the approximated TP -operator. While the approach
presented in Section 7.5.2 creates an input and output unit for every ground
atom, we created just a single unit here. Thus, to increase the accuracy of
the network we simply have to add a unit to the hidden layer, but the input
and output layers can be kept unchanged. Unfortunately, using only a single
unit limits the overall accuracy once the network is implemented on a real
computer.

7.5.4 Approximation by Radial-Basis-Function Networks

Radial-basis-function (RBF) networks are another common neural network
architecture.31 As in the case of sigmoidal networks, they are known to be
universal approximators for continuous functions on compact subsets of Rn .
An RBF network consists of three layers: the input, hidden, and output layers.
The activation of units in the input layer is set from outside. But in contrast to
the networks discussed so far, the hidden units do not compute the weighted
sum, but compute the distance between the vector of input unit activations
and the weight vector of the corresponding connection. That is, the potential of
unit k with nk incoming connections is computed as pk (t) = m(�v , w�k ), with m
denoting a metric over nk -dimensional vectors, �v denoting the vector of input
unit activations, and wk denoting the vector of weights of the connections to
unit kk . Usually, the Euclidean distance between the two vectors is used as
the distance function m.
30 The proof and all details of the construction involved in this result can be found in

[Bader, 2009].
31 Good introductions to them can be found in [Rojas, 1996] and [Bishop, 1995].
 Logic Programming and Artificial Neural Networks 205
y
0.3̄

0.3̄

FIGURE 7.10: The raised cosine activation function and an approximation

of the embedded TP -operator of Program 7.5.1 using raised cosine activation
functions. Each constant piece is represented using two raised cosines.

In the constructions below, we use the raised cosine function (see Fig­
ure 7.10) to compute the activation of the hidden units:

h x−p
A A bb
· 1 + cos π · if |x − p| ≤ w,
rcoshp,w : R → R : x �→ 2 w
0 otherwise.

Note that if two raised cosines rcoshp1 ,w and rcoshp2 ,w with |p1 − p2 | = w are
added, we obtain a function that is constant on the interval [p1 , p2 ]. Therefore,
we can represent each constant piece from above by two raised cosines. Fig­
ure 7.10 shows the approximation of the TP -operator for our running example.
As above, the approximation by raised cosines can easily be implemented
using an RBF network. The resulting network contains a single input and
output unit serving as interface. Every raised cosine necessary for the approx­
imation is computed by a single hidden unit. The weight between the input
and the hidden layer contains the position, and the weight between the hidden
and the output unit represents the height of the function. Figure 7.11 shows
the RBF network for Program 7.5.1. Using these insights, we can state the
following theorem, again without proof.

7.5.9 Theorem Let P be a covered logic program, let b > 2, and let ε > 0.
Then we can construct an RBF network whose network function approximates
TP up to ε.

Unfortunately, the two approaches discussed in Sections 7.5.3 and 7.5.4

only allow for limited accuracy when implemented on a real computer. This is
due to the fact that a single unit is used in the input layer and in the output
layer. Even though we can assume unlimited accuracy of real number opera­
tions in theory, we cannot assume this when using a computer. To overcome
this drawback, we discuss another approach in the following section.
206 Mathematical Aspects of Logic Programming Semantics

0.323
0.0

0.323
0.021

0.323
0.062

0.
3
08

32

i o
0.

0.0 0.0
0
0.

26
25

0.
0
0.271

0.260
0.312

0.260
0.333

0.260

FIGURE 7.11: An RBF network approximating the TP -operator of Pro­

gram 7.5.1.
 Logic Programming and Artificial Neural Networks 207
y
0.3̄

x
0.3̄

FIGURE 7.12: A two-dimensional version of the Cantor set obtained by em­

bedding all interpretations using a two-dimensional bijective level mapping.

y y y y
0.3̄ 0.3̄ 0.3̄ 0.3̄

x x x x
0.3̄ 0.3̄ 0.3̄ 0.3̄

FIGURE 7.13: A construction of the two-dimensional version of the Cantor

set.

7.5.5 Approximation by Vector-Based Networks

The approaches presented above are based on level mappings with co­
domain ω. Here we extend this approach to multi-dimensional level mappings,
which permits the embedding of interpretations into vectors of real numbers.
An n-dimensional level mapping is a function L : BP → ω × {1, . . . , n}, that
is, to each atom A we assign a level Ll (A) ∈ ω and some dimension Ld (A) ∈
{1, . . . , n}. As above, we assume a bijective level mapping. On embedding
interpretations into n-dimensional real vectors, we obtain an n-dimensional
version of the classical Cantor set. A two-dimensional version is shown in
Figure 7.12.
Unfortunately, the results obtained so far cannot be extended to the n-
dimensional case, at least we do not know how to make such an extension.
But nevertheless we can construct approximating networks employing certain
knowledge that we have about the set of embedded interpretations. Figure 7.13
shows a possible way of constructing the two-dimensional Cantor set. Starting
from a square, in every iteration the current version is copied and scaled
down four times. Afterwards, the four copies are placed in the corners. The
squares occurring in the n-th step of the construction are referred to below as
hypercubes of level n.
As for the one-dimensional case, the TPn -operator turns out to be a piece-
wise constant function. Let Pn be as previously defined, and let ñ be the
maximal level of a body atom in Pn . Then the operator TPn is constant on all
208 Mathematical Aspects of Logic Programming Semantics

0.3̄

0.3̄

FIGURE 7.14: A depiction of the approximation of the TP -operator for Pro­

gram 7.5.1 using a vector-based network.

those interpretations which agree on all atoms up to level ñ, and those areas
coincide with the hypercubes of level ñ.32
Vector-based networks33 can be thought of as a generalization of the so-
called self-organizing maps.34 A number of units are distributed over the input
space. For every input given to the network, the closest unit is selected as the
winner unit. The winner’s activation is set to 1, and the activation of all other
units is set to 0. Thus, only the winner influences the output of the network.
By setting up a network such that there is a unit for every hypercube
of level ñ, we can directly embed the TPn -operator into the weights of the
connections from those units to the output units. Figure 7.14 shows what such
a network for the one-dimensional case could look like.35 For every hypercube
(coinciding with intervals in the one-dimensional case) a unit is added to the
network. The weights between the input and hidden layers define (as for RBF
networks) the location of the unit, and the weights between the hidden and
output layers define the output, that is, the value of the embedded TP -operator
for an interpretation within the input area of the unit. As before, we are now
in a position to state a theorem asserting the existence of approximating
vector-based networks, as follows.

7.5.10 Theorem Let P be a covered logic program, let b > 2, and let ε >
0. Then we can construct a vector-based network whose network function
approximates TP up to ε.

By using an m-dimensional level mapping, we fix the network to have

m input and m output units. That is, we can increase the accuracy of the
network by using more units. Unfortunately, the number of hidden units grows
exponentially with the dimension of the input layer. Nonetheless, we are now
in a position to trade accuracy against space, which has not been possible
before.
32 See [Bader, 2009] for details.
33 See [Martinetz and Schulten, 1991, Fritzke, 1998] for further details.
34 [Kohonen, 1981, Haykin, 1994].
35 The n-dimensional case for n > 1 is hard to depict because the graphics need to be

(n + 1)-dimensional.
 Logic Programming and Artificial Neural Networks 209

Just as for the network architectures described previously, we can train

vector-based networks using a set of input-output pairs. The position of the
units, that is, the weights between the input and hidden layer, are modified
such that a unit is located in the centre of all the inputs it is responsible for.
The output weights are trained such that they represent the average output
of all inputs within the unit’s responsibility. If, furthermore, two neighbouring
units have similar output weights, then one of them can be removed because
the other unit will take over in that eventuality. A unit whose accumulated
error is very large can be replaced by two units that can be adapted indepen­
dently, thus allowing the network to refine its input-output function in certain
areas.
The first experiments which reported on this approach36 showed the ap­
plicability of this learning method in the area of neural-symbolic integration.
A randomly initialized network was trained using the embedded versions of
an interpretation I as input values and of TP (I) as output values for a given
program P . The network learned the mapping and could be used iteratively
by adding recurrent connections between the output and input layers.

7.5.6 Approximating the (Least) Fixed Point of TP

Thus far, we have discussed at some length the issue of the approximate
computation of the TP -operator for first-order normal logic programs P . We
turn now to discussing, fairly briefly, the question of the approximate compu­
tation of its fixed points. One approach is to carry forward the work of the
previous sections and employ iterates of (recurrent) neural networks which
approximate TP to approximate iterates of the operator TP , but, as already
noted earlier, the problem then emerges of uniformly controlling the error
estimates under iteration.37
On the other hand, one can approach the problem of computing the least
fixed point of TP for arbitrary definite logic programs P by a modification of
the previous approach employing the subset Pn of ground(P ), except that we
do not assume that P is covered, and instead we ensure that the appropriate
subset of ground(P ) is finite by other means.
Thus, let P denote an arbitrary (first-order) definite logic program, and
denote by I the least fixed point of TP . Let l : BP → ω be a level mapping
with the property that l−1 (n) is a finite set for each n ∈ ω. We proceed to
sketch the details of the construction of a finite subset P n of ground(P ), where
n is a given natural number, which will play the sort of role here that Pn plays
in Proposition 7.5.5 and its companion results.38 We start with the following
claim.
36 See [Bader et al., 2007].
37 This point is discussed in [Hitzler et al., 2004, Section 4.3], but quite strong conditions,
for example, Lipschitz continuity [Hitzler et al., 2004, Theorem 4.19], are required for things
to work satisfactorily.
38 See [Seda, 2006] for full details.
210 Mathematical Aspects of Logic Programming Semantics

Claim. Suppose that A ∈ TP ↑ k. Then there is a clause A ← body in

ground(P ) such that A does not occur in body and TP ↑ (k − 1) |= body.
To establish this claim, we first note that it is clear that k ≥ 1. Suppose
that A ∈ TP ↑ k0 = TP (TP ↑ (k0 − 1)) and that k0 is the smallest natural
number with this property. Then there is a clause A ← body in ground(P )
such that TP ↑ (k0 −1) |= body. By definition of k0 , we have A ∈ TP ↑ (k0 −1),
and hence A does not occur in body. Finally, by monotonicity, we obtain that
TP ↑ (k − 1) |= body, as required.
Since P is definite, we have

∞
TP ↑ 0 ⊆ TP ↑ 1 ⊆ · · · ⊆ TP ↑ n ⊆ · · · ⊆ I = TP ↑ n,
n=1

where TP ↑ n denotes the n-th upward power TPn (∅) of TP , as usual.

Given n ∈ N, there are only finitely many atoms A1 , A2 , . . . , Am ∈ I
with l(Ai ) ≤ n for i = 1, . . . , m, and, by directedness, there is (a smallest)
k = kn ∈ N such that A1 , A2 , . . . , Am ∈ TP ↑ kn .39 Consider the atom Ai ,
where 1 ≤ i ≤ m, and the following three steps.
(1) We have Ai ∈ TP ↑ kn = TP (TP ↑ (kn − 1)). Therefore, there is a
clause
m(i)
Ai ← A1i (1), . . . , Ai (1)

m(i)
in ground(P ) such that A1i (1), . . . , Ai (1) ∈ TP ↑ (kn − 1). Note that this
clause may be a unit clause, that is, m(i) ≥ 0, and there may be many such
clauses with head Ai ; we choose one of them.
m(i)
(2) Because A1i (1), . . . , Ai (1) ∈ TP ↑ (kn − 1) = TP (TP ↑ (kn − 2)),
there are clauses in ground(P ) as follows.

m(i,1)
A1i (1) ← A1i,1 (2), . . . , Ai,1 (2)
m(i,2)
A2i (1) ← A1i,2 (2), . . . , Ai,2 (2)
.. ..
. ← .
m(i) 1 m(i,m(i))
Ai (1) ← Ai,m(i) (2), . . . , Ai,m(i) (2),

where each of the atoms Ari,j (2) in each of the bodies belongs to TP ↑ (kn − 2).
(3) Because each of the Ari,j (2) in Step (2) belongs to TP ↑ (kn − 2) =
TP (TP ↑ (kn − 3)), we have a finite collection of ground clauses (one for each

39 Notice that, depending on l, there may be no atoms A with l(A) ≤ n; this case is

handled by the abuse of notation obtained by allowing m to be 0.

 Logic Programming and Artificial Neural Networks 211

of the Ari,j (2) in Step (2)) as follows.

1 m(i,1,1)
Ai,1 (2) ← A1i,1,1 (3), . . . , Ai,1,1 (3)
m(i,1,2)
A2i,1 (2) ← A1i,1,2 (3), . . . , Ai,1,2 (3)
.. ..
. ← .
m(i,1) 1 m(i,1,m(i,1))
Ai,1 (2) ← Ai,1,m(i,1) (3), . . . , Ai,1,m(i,1)
(3)
m(i,2,1)

A1i,2 (2) ← A1i,2,1 (3), . . . , Ai,2,1 (3)

m(i,2,2)
A2i,2 (2) ← A1i,2,2 (3), . . . , Ai,2,2 (3)
.. ..
. ← .
m(i,2) 1 m(i,2,m(i,2))
Ai,2 (2) ← Ai,2,m(i,2) (3), . . . , Ai,2,m(i,2) (3)
.. ..
. ← .
m(i,m(i),1)
A1i,m(i) (2) ← A1i,m(i),1 (3), . . . , Ai,m(i),1 (3)
m(i,m(i),2)
A2i,m(i) (2) ← A1i,m(i),2 (3), . . . , Ai,m(i),2 (3)
.. ..
. ← .
m(i,m(i)) 1 m(i,m(i),m(i,m(i)))
Ai,m(i) (2) ← Ai,m(i),m(i,m(i)) (3), . . . , Ai,m(i),m(i,m(i)) (3),

where each atom in each body belongs to TP ↑ (kn − 3).

Note that at each stage in this process we select a ground clause in which
the head of the clause does not occur in the body by means of the claim
established earlier.
This process terminates producing unit clauses in its last step. Let Pi,n
denote the (finite) subset of ground(P ) consisting of all the clauses which
result; it is clear that TPi,n ↑ kn consists of the heads of all the clauses in
Pi,n . We carry out this construction for i = 1, . . . , m to obtain programs
P1,n , . . . , Pm,n such that, for i = 1, . . . , m, TPi,n (TPi,n ↑ kn ) = TPi,n ↑ kn
(indeed, TPi,n ↑ kn is the least fixed point of TPi,n by Kleene’s theorem, The­
orem 1.1.9), Ai ∈ TPi,n ↑ kn , and TPi,n ↑ r ⊆ TP ↑ r ⊆ I for all r ∈ N. Let
P n denote the program P1,n ∪ . . . ∪ Pm,n . Then P n is a finite subprogram
of ground(P ), and TPi,n ↑ kn ⊆ TP n ↑ kn ⊆ TP ↑ kn ⊆ I for i = 1, . . . , m.
Furthermore, A1 , . . . , Am ∈ TP n ↑ kn , and TP n ↑ kn is the least fixed point I n
of TP n
.
This completes the construction of the program P n .

7.5.11 Example We illustrate the process just described with k = kn = 3.

Suppose that A1 ∈ TP ↑ 3 = TP (TP ↑ 2). Then there is a ground clause A1 ←
B1 , B2 , say, with B1 , B2 ∈ TP ↑ 2 = TP (TP ↑ 1). Therefore, there exist ground
clauses B1 ← C1 , C2 , C3 and B2 ←, say, with C1 , C2 , C3 ∈ TP ↑ 1 = TP (∅).
212 Mathematical Aspects of Logic Programming Semantics

It follows that there are unit clauses C1 ←, C2 ←, and C3 ← in ground(P ).

Thus, P1 is the program

C1 ←
C2 ←
C3 ←
B2 ←
B1 ← C1 , C2 , C3
A1 ← B1 , B2

Then we have the following calculations: TP1 ↑ 0 = ∅, TP1 ↑ 1 = TP1 (∅) =

{B2 , C1 , C2 , C3 }, TP1 ↑ 2 = TP1 ({B2 , C1 , C2 , C3 }) = {B1 , B2 , C1 , C2 , C3 },
TP1 ↑ 3 = {A1 , B1 , B2 , C1 , C2 , C3 }, and TP1 ↑ 4 = TP1 (TP1 ↑ 3) =
{A1 , B1 , B2 , C1 , C2 , C3 } = TP1 ↑ 3. Thus, TP1 ↑ 3 is a fixed point of TP1
and indeed is the least such fixed point. Moreover, A1 ∈ TP1 ↑ 3.

Further properties of P n can be found in [Seda, 2007].

Now let ε > 0 be given and choose n so large that 2−n < ε. Then
dl (I n , I) ≤ 2−n < ε, where I n is the least fixed point of TP n , I is the least fixed
point of TP , as noted above, and dl is the metric associated with l. Now apply
the algorithm of Theorem 7.4.1 to the propositional program P n and make
the resulting network Fn (which computes TP n ) recurrent. On inputting the
interpretation ∅ to this network and iterating n times, we obtain I n as output.
Thus, Fn approximates I up to ε, and in this sense the family {Fn | n ∈ N}
computes I.

7.5.12 Example Take P to be as in Example 3.2.3, that is, the program

p(a) ←
p(s(X)) ← p(X)

Applying the procedure above to P , we obtain a sequence Fn of 3-layer feed-

forward recurrent neural networks which computes the least fixed point of TP
and hence computes the set of natural numbers.

7.6 Some Extensions – The Propositional Case

So far in this chapter, we have concentrated on the operator TP . However,
in this section and the next we want to briefly consider extensions of our results
to other operators and hence to other semantics. In the present section, we will
focus on propositional normal logic programs P and extensions of the results of
 Logic Programming and Artificial Neural Networks 213

Section 7.4. In particular, we consider extensions of Theorem 7.4.1 to Fitting-

style operators FP , including of course the special cases of ΦP for Kleene’s
strong three-valued logic and the corresponding operator ΨP for Belnap’s
logic FOU R. In the next section, Section 7.7, we will consider extensions
of Section 7.5.2, or in other words we will consider approximations of local
consequence operators, including Fitting-style operators, and the Gelfond–
Lifschitz operator.
In fact, one can adopt an algebraic approach to the material presented in
this section at little extra cost, but with the benefit that the results apply
to constraint logic programs (with constraints belonging to a given semiring)
and to logic programs involving uncertainty expressed via many-valued logics,
as well as to conventional logic programs. We shall not do that, however,
as it would take us too far afield, requiring a definition of logic programs
allowing elements of an abstract set (the set C in the next definition) in clause
bodies and a corresponding new definition of Fitting-style operators. Instead,
we content ourselves with sketching the development for conventional logic
programs.40 Nevertheless, we will present the material in full generality where
it helps, ultimately specializing to logics T . Thus, we next present one of the
main definitions we need in full generality, as follows.

7.6.1 Definition Suppose that C is a set equipped with a binary operation 8.

We say that 8 is finitely determined or that products (relative to 8) are finitely
determined in C if, for each c ∈ C, there exists a countable (possibly infinite)
collection {(Rcn , Ecn ) | n ∈ J } of pairs of sets Rcn ⊆ C and Ecn ⊆
gC, where each
Rcn is finite, such that a countable (possibly infinite) product i∈M ci in C is
equal to c if and only if for some n ∈ J the following statements hold.

(1) Rcn ⊆ {ci | i ∈ M }.

/ Ecn , that is, {ci | i ∈ M } ⊆ (Ecn )co , where (Ecn )co

(2) For all i ∈ M , ci ∈
denotes the complement of the set Ecn .

We call the elements of Ecn excluded values for c, we call the elements of
Anc = (Ecn )co allowable values for c, and in particular we call the elements of
Rcn required values for c; note that, for each n ∈ J , we have Rcn ⊆ Acn , so
that each required value is also an allowable value (but not conversely). More
generally, given
g gc ∈ C, we call s ∈ C an excluded value for cgif no product
i∈M ci with i∈M ci = c contains s, that is, in any product i∈M ci whose
value is equal to c, we have ci = s for no i ∈ M . We let Ec denote the set of
all excluded values for c, and let Ac denote the complement (Ec )co of Ec and
call it the set of all allowable values for c. Note finally that when confusion
might otherwise result, we will superscript each of the sets introduced above
40 For full details of the sketch we present here, the reader should consult the following

papers: [Seda and Lane, 2005], [Lane and Seda, 2006], [Komendantskaya et al., 2007] and
also [Lane and Seda, 2009].
214 Mathematical Aspects of Logic Programming Semantics

with the operation in question. Thus, for example, A�

c denotes the allowable
set for c relative to the operation 8. •

In particular, we can take C as a logic T and 8 as either disjunction or

conjunction defined on it. Indeed, the following example and the paragraph
following it show the thinking behind Definition 7.6.1, and in fact we shall take
FOU R as a running example throughout this section. Note that, throughout
this section, we take FOU R to be the set {u, f , t, b} with this given listing of
its elements, as in Chapter 1.

7.6.2 Example Consider again Belnap’s logic FOU R. Taking 8 to be dis­

junction ∨, the sets E and R are as follows.
(1) For u, we have n = 1, Eu∨ = {t, b}, and Ru∨ = {u}.
(2) For f , we have n = 1, Ef∨ = {u, t, b}, and Rf∨ = {f }.

(3) For t, n takes the values 1 and 2, Et∨ = ∅, Rt∨,1 = {t}, and Rt∨,2 = {u, b}.
(4) For b, we have n = 1, Eb∨ = {u, t}, and Rb∨ = {b}.
o
Thus, for example, a countable disjunction i∈M si takes value t if and only
if either (i) at least one of the si takes value t or (ii) at least one of the si
takes value b and at least one takes value u; no truth value is excluded.
Now taking 8 to be conjunction ∧, the sets E and R are as follows.
(1) For u, we have n = 1, Eu∧ = {f , b}, and Ru∧ = {u}.
(2) For f , n takes the values 1 and 2, Ef∧ = ∅, Rf∧,1 = {f }, and Rf∧,2 = {u, b}.
(3) For t, we have n = 1, Et∧ = {u, f , b}, and Rt∧ = {t}.
(4) For b, we have n = 1, Eb∧ = {u, f }, and Rb∧ = {b}.
In fact, Definition 7.6.1 was motivated by the problem, already mentioned,
of defining truth values of bodies of pseudo-clauses over various three-valued
logics, see [Hitzler and Seda, 1999b] and Sections 5.2.1 and 5.5 herein. The
following facts show how it works, where we take the countable set M to be
N without loss of generality. If 8 is finitely determined, then it is idempotent,
commutative,
g and associative, as already noted in Section 5.5. Furthermore,
if i∈M s i = c, then the sequence s1 , s1 8 s2 , s1 8 s2 8 s3 , . . . is eventually
constant with value c. In the converse direction, suppose C is a countable set
and 8 is idempotent, commutative, and associative. Suppose further that,
for any set {si | i ∈ M } of elements of C where M is countable, the sequence
s1 , s1 8s2 , s1 8s2 8s3 , . . . is eventually constant with value c. Thengall products
in C are (well-defined
g and) finitely determined, where we take i∈M si = c
to define i∈M si .
For a finitely determined binary operation 8 on C, we define the partial
 Logic Programming and Artificial Neural Networks 215

order ≤� on C by s ≤� t if and only if s 8 t = t. (So that s ≤+ t if and only if

s + t = t, and s ≤× t if and only if s × t = t, for finitely determined operations
+ and ×, and similarly for finitely determined operations of disjunction ∨ and
conjunction ∧ in case C is a logic T .)

7.6.3 Example In FOU R, we have t ≤∧ u ≤∧ f , and t ≤∧ b ≤∧ f . Also,

f ≤∨ u ≤∨ t, and f ≤∨ b ≤∨ t.

In fact, the allowable and excluded sets for s ∈ C can easily be characterized
in terms of the partial orders just defined: s ∈ A� t if and only if s ≤� t, see
[Seda and Lane, 2005, Proposition 3.10]. Because of this fact, we have the
following result.

7.6.4 Proposition Suppose that 8 is a finitely determinedg binary operation

on C and that M is a countable set. Then a product i∈M ti evaluates to
the element s ∈ C, where s is the least element in the ordering ≤� such that
ti ∈ A�s for all i ∈ M .

Having now determined how we evaluate the truth values of the bod­
ies of pseudo-clauses in relation to Fitting-style operators FP , we move
next to consider the computation of these operators by neural networks in
the case of propositional normal logic programs P . Indeed, it is shown in
[Lane and Seda, 2006] that one can construct conventional 3-layer feedforward
networks to compute ΦP and ΨP containing only binary threshold units, in
the style of Theorem 7.4.1.41 However, extending this approach to the general
case of FP is not so simple, as the constructions become overly complicated.
Therefore, we will adopt a modular approach in which we construct two types
of 2-layer neural networks of binary threshold units. The first of these (the
multiplication unit) will compute products or conjunctions of elements of C,
and the second of them (the addition unit) will compute sums or disjunc­
tions of elements of C. It then remains to construct 3-layer neural networks
to compute FP in which the hidden layer consists of multiplication units and
the output layer consists of addition units; strictly speaking, these networks
have five layers of course. In this context, it is worth noting that the partial
ordering ≤� , defined previously, and Proposition 7.6.4 play a crucial role in
establishing the results we discuss here.
For the rest of this section, we shall focus on finite sets C with n elements
listed in some fixed order, C = {c1 , c2 , . . . , cn } or C = {t1 , t2 , . . . , tn }, say. In
order to simulate the operations in C by means of neural networks, we need
to represent the elements of C in a form amenable to their manipulation by
neural networks. To do this, we represent elements of C by vectors of n units,
and it is convenient sometimes to view them as column vectors, where the first
unit represents c1 , the second unit represents c2 , and so on. Hence, a vector of
41 See the thesis [Kalinke, 1994], where these results are stated. We thank S. Hölldobler

for drawing this reference to our attention.

216 Mathematical Aspects of Logic Programming Semantics
u f t b
0.5
� 0.5 0.5
� 0.5
/ � / �
/ /
/ /
/ /
/ /
/ /
/ /
u f t b
l − 0.5 l − 0.5 l − 0.5 l − 0.5

FIGURE 7.15: A conjunction unit for FOU R. The full arrows represent con­
nections with weight 1, and the broken arrows represent connections with
weight −1.

units with the first unit activated, or containing 1, represents c1 , a vector with
the second unit activated, or containing 1, represents c2 , etc. Indeed, it will
sometimes be convenient to denote such vectors by binary strings of length n
and to refer to the unit in the i-th position of a string as the i-th unit or the
ci -unit or the unit ci ; as is common, we represent these vectors geometrically
by strings of not-necessarily adjacent rectangles. Note that we do not allow
more than one unit to be activated at any given time in any of the vectors
representing elements of C, and hence all but one of the units in such vectors
contain 0. Furthermore, when the input is consistent with this, it turns out
from the constructions we make that the output of any network we employ is
consistent with it also.

7.6.5 Example Suppose that C = FOU R = {u, f , t, b}. Then u is repre­

sented by 1000, f by 0100, t by 0010, and b by 0001.

In general, the operations in C are not linearly separable, and therefore

we need two layers to compute addition (or disjunction) and two to compute
multiplication (or conjunction). As usual, we take the standard threshold for
binary threshold units to be 0.5. This ensures that the Heaviside function H
outputs 1 if the input is strictly greater than 0, rather than greater than or
equal to 0.

7.6.6 Definition A multiplication (×) unit or a conjunction (∧) unit MU

for a given set C is a 2-layer neural network in which each layer is a vector of
n binary threshold units c1 , c2 , . . . , cn corresponding to the n elements of C.
The units in the input layer have thresholds l − 0.5, where l is the number of
elements being multiplied or conjoined, and all output units have threshold
0.5. We connect input unit ci to the output unit ci with weight 1 and to any
unit cj in the output layer, where ci <× cj , with weight −1.

An input layer representing a product of l elements of C is connected to

 Logic Programming and Artificial Neural Networks 217

a multiplication unit MU in the following way. For each element c of the

product, where c is represented by the n units c1 , c2 , . . . , cn , the unit cj is
connected, with weight 1, to the cj -unit in the input layer of MU and is also
connected, with weight 1, to any unit ck in the input layer of MU for which
cj <× ck . For a negated element d = ¬c in the product, we connect, with
weight 1, cj to the unit representing ¬cj in the input layer of MU and also,
with weight 1, to any unit ck in the input layer of MU for which ¬cj <× ck .

7.6.7 Proposition A multiplication or conjunction unit MU computes the

value of a product of l elements of C when it is connected to an input layer
as just described.

7.6.8 Example Consider again C = FOU R, and input the two elements
u and b to a multiplication unit MU, where l = 2. It is readily checked
that the potentials of the units u, t, and b in the input layer of MU are,
respectively, −0.5, −1.5, and −0.5; their outputs are all equal to 0; and the
outputs of the units u, t, and b in the output layer of MU are also all equal
to 0. On the other hand, the f -unit in the input layer of MU has potential
1 × 1 + 1 × 0 + 1 × 0 + 1 × 0 + 1 × 0 + 1 × 0 + 1 × 0 + 1 × 1 − 1.5 = 0.5, and
therefore the output of this unit is H(0.5) = 1. Furthermore, the input to the
f -unit in the output layer of MU is −1 × 0 + 1 × 1 − 1 × 0 − 1 × 0 = 1. Hence,
the output of this unit is H(1 − 0.5) = 1, and so MU outputs 0100 or f , and
this indeed is the value of u ∧ b, as required.

The ideas behind multiplication units work, with minor changes, for addi­
tion or disjunction, and we obtain addition (+) or disjunction (∨) units AU
which compute the sum or disjunction of k, say, elements of C.
We are now in a position to state the main theorem of this section, where
we take the set C to be a logic T endowed with the operations of disjunction
(∨) and conjunction (∧).

7.6.9 Theorem Suppose that both operations of disjunction and conjunction

in T are finitely determined and that P is a propositional logic program
defined over T . Then we can construct a 3-layer feedforward neural network
F which contains multiplication units in its middle layer and addition units
in its output layer such that F computes FP .

In closing this section, we mention that there is yet another class of logic
programs one can consider in our present context of extending Theorem 7.4.1,
namely, the class of propositional annotated (bi)lattice-based logic programs.
This class is also a very general class of programs capable of handling uncer­
tainty, in this case using lattices and bilattices to model belief estimates for
and against a proposition. However, its study would take us too far from our
current goal, and instead we refer the reader to [Komendantskaya et al., 2007]
again for full details.
218 Mathematical Aspects of Logic Programming Semantics

7.7 Some Extensions – The First-Order Case

So far, in this chapter we have described how certain methods developed
in earlier chapters give rise to approaches to the problem of integrating logic
programs and artificial neural networks. The key insight into this integration
is the observation that the two paradigms can be formally related by means of
functions: on the one hand, semantic operators for logic programs capture the
meaning of logic programs; on the other hand, the input-output function of
an artificial neural network completely characterizes its functional behaviour.
Approaches to neural-symbolic integration thus arise out of methods which
allow us to understand semantic operators as I/O functions of artificial neural
networks, and vice-versa.
Most of this chapter has focused on the single-step operator in logic pro­
gramming which, via its fixed points, determines the supported model seman­
tics of logic programs. However, in Section 7.6, we have just seen that some of
these methods carry over to other semantics, for propositional programs, via
the computation of FP . In this section, we will now consider the first-order
case and extensions of the approximation results we have established for TP to
FP and other semantic operators. At the same time, we briefly discuss further
alternative semantics as treated throughout the book and discuss conclusions
which can be drawn concerning neural-symbolic integration in general.
Our conceptual starting point is Theorem 7.5.3, which tells us that approx­
imating networks exist if and only if the single-step operator is continuous in
the Cantor topology.42 We can now use this result to leverage several new
results on the relationship between the supported model semantics and other
semantics in order to derive similar results for these other semantics.
In Section 5.4, we considered a very general family of semantic operators
and also examined the question of how one may characterize Cantor continuity
for them. The following result is thus an easy corollary of Theorem 5.4.7.

7.7.1 Theorem Let P be a program with a locally finite local consequence

operator T . Then T can be uniformly approximated by 3-layer feedforward
networks in the sense of Theorem 7.2.2.
42 We briefly remark that a result established in [Hornik et al., 1989], which states that

every measurable function can be approximated almost everywhere by a 3-layer sigmoidal

feedforward network, is not necessarily useful for our purposes. This is so despite the fact
that it was shown in Theorem 5.5.1 that many semantic operators, including the single-step
operator, are always measurable, and hence also the Gelfond–Lifschitz operator, see Theo­
rem 6.2.4. However, it should be noted that the Cantor set is a set of (Lebesgue) measure zero
when viewed as a subspace of the reals. Thus, the result just quoted of [Hornik et al., 1989]
need not necessarily lead to useful approximation results for these operators. Indeed, such
approximations arising from the result of [Hornik et al., 1989] may fail to approximate the
operator in question at every point. Nevertheless, it remains to be investigated whether non­
zero measures exist on the Cantor set, which yield useful approximations in conjunction with
the results of [Hornik et al., 1989].
 Logic Programming and Artificial Neural Networks 219

Theorem 7.7.1 covers, among others, all Fitting-style operators from Sec­
tion 5.2.1.
The fixpoint completion, studied in Section 6.1, also turns out to be very
useful, since it allows us to reduce treatments of the Gelfond–Lifschitz oper­
ator to the single-step operator. According to Theorem 7.2.2, we are first of
all interested in carrying over continuity results with respect to the Cantor
topology. From Theorem 6.2.2 we thus obtain the following result.

7.7.2 Theorem Let P be a normal logic program, and let the following con­
dition be satisfied for all I ∈ IP and A ∈ BP : whenever GLP (I)(A) = f ,
then either there is no clause with head A in ground(P ) or there exists a
finite set S(I, A) = {A1 , . . . , Ak } ⊆ BP such that I(Ai ) = t for all i, and
for every clause A ← body in ground(P ) at least one ¬Ai or some B with
GLP (I)(B) = f occurs in body. Then GLP can be uniformly approximated
by 3-layer feedforward networks in the sense of Theorem 7.2.2.
We also obtain the following corollary, taking Corollary 6.2.3 into account.

7.7.3 Corollary Let P be a covered normal logic program. Then GLP can
be uniformly approximated by 3-layer feedforward networks in the sense of
Theorem 7.2.2.
Likewise, the remark given in Footnote 3 on page 170 of Chapter 6 together
with Lemma 5.4.12 allows us to derive similar characterizations of continuity
for the operator characterizing three-valued stable models.
In principle, one can use the results recorded earlier to embark on inves­
tigations similar to those undertaken in Sections 7.5.3 to 7.5.5, for example.
However, a direct application of the results for the Gelfond–Lifschitz oper­
ator is hardly satisfactory since the computation of the fixpoint completion
can only be carried out in an approximate manner. How one deals with this
problem, and what it entails, remains to be investigated.
It should be clear from Section 7.6 how one carries over the approach
using a finite subset of the grounding of a program to other locally finite
local consequence operators. For operators like the Gelfond–Lifschitz operator,
however, a straightforward approach is rather unsatisfactory due to the fact
that one iteration of the Gelfond–Lifschitz operator involves the taking of
a limit of the single-step operator for definite programs. As an alternative
approach, we could again first compute the fixpoint completion of the program
and employ Theorem 6.1.4 in conjunction with the methods from Section 7.6,
but alas, we have noted already that computation of the fixpoint completion
can only be done in an approximate manner. How to deal with this problem in
an appropriate manner again is something which remains to be investigated.

Before closing this chapter, we would like to remark that there is a plethora
of work which has been done on the integration of logic and neural networks.43
43 See, for example, [Bader and Hitzler, 2005, Hammer and Hitzler, 2007] for overviews.
220 Mathematical Aspects of Logic Programming Semantics

In particular, the propositional core method (see Section 7.4) has spawned
a lot of investigations, including extended semantics for propositional logic
programs.44 But alternative methods are also under investigation which do
not connect directly with the investigations into the mathematical foundations
of logic programming we have presented in this book.45

44 Most notable is the body of work done by Artur d’Avila Garcez of City Univer­

sity London on, for example, modal logic, see [d’Avila Garcez et al., 2007], intuitionistic
logic, see [d’Avila Garcez et al., 2006], and epistemic and temporal logic, see, for example,
[d’Avila Garcez and Lamb, 2006]. See also [d’Avila Garcez et al., 2009].
45 For further notable recent work based on methods other than those reported here, the

reader should consult the following papers [Gust et al., 2007, Hölldobler and Ramli, 2009,
Komendantskaya, 2010, Buillame-Bert et al., 2010].
Chapter 8

Final Thoughts

In this book, we have provided a comprehensive treatment of logic program­

ming semantics from the perspective of fixed-point semantics. In doing so, we
have covered a lot of material which also relates to other areas of interest
outside the realm of logic programming as such. In this final chapter, we dis­
cuss contributions to and relationships between the content of this book and a
rather diverse mix of topics, ranging from foundations of computing via arti­
ficial intelligence to cognitive science. We do so with the usual understanding
that the impact of foundational research is more often than not indirect in
nature in providing results, methods, and insights, which can be carried for­
ward by research communities at large until a critical mass is reached, thereby
enabling significant or even major advances to take place.

8.1 Foundations of Programming Semantics

The classical semantic analysis of programs in the sense of denotational
semantics is based on monotonic, order-continuous operators, via their least
fixed points using Theorem 1.1.9 or Theorem 1.1.10. This approach, however,
fails for paradigms where the semantics is expressed by fixed points of op­
erators which are not monotonic in general. In particular, it fails for logic
programming in several of its variants, as studied throughout this book.
By developing methods for the fixed-point semantic analysis of programs
with non-monotonic semantic operators, we therefore widen the scope of ap­
plicability of fixed-point semantics. In particular, we provide sufficiency con­
ditions for the existence of fixed points (Chapter 4) and show how they can
be applied to various semantics based on non-monotonic operators (Sections
5.1 and 5.4 and Chapter 6).
It seems evident that these methods should carry over to other such
paradigms. However, a limitation of some of the work presented in this book
is that certain of the fixed-point theorems provided in Chapter 4 always guar­
antee the existence of a unique fixed point, if there is a fixed point at all,
thus rendering the theorems in question of limited applicability to paradigms
(or programs) where multiple fixed points are the norm. The latter situation

221
222 Mathematical Aspects of Logic Programming Semantics

is encountered in the logic programming paradigm in the case of the stable

model semantics, for example, and indeed our analysis in Section 6.2 is lim­
ited in this respect. Multiple fixed points also arise naturally in the context of
disjunctive logic programs, that is, logic programs where additionally disjunc­
tions of atoms are allowed in rule heads as discussed briefly in Section 4.9. The
application of fixed-point theorems for multivalued mappings as provided, for
example, in Sections 4.9 to 4.13 may provide a remedy when this line of work
has been fully worked out. In particular, an approach to this problem based on
the Rutten-Smyth theorem and a careful analysis and choice of quasimetrics
(perhaps based on level mappings) holds out considerable prospects in this
respect, see [Seda, 1997]. In addition, approaches such as the one mentioned
in Section 4.14 also overcome this problem to a considerable extent.1

8.2 Quantitative Domain Theory

Domain theory,2 based on order continuity of semantic operators, is the
dominant theory underlying the denotational semantics of programming lan­
guages. However, an alternative tradition in the semantics of programming
languages to that using domains is an approach based on the use of metric
spaces, as already mentioned in Chapter 4.3 A reconciliation of these two ap­
proaches is of obvious interest for the theory of programming semantics, and a
considerable body of work has been done on this very topic,4 resulting in the
area of quantitative domain theory. Indeed, the Rutten-Smyth theorem arose
out of precisely these considerations.
In contrast to mainstream work on this reconciliation, which is driven by
a mainly conceptual motivation to unite two theories, the work presented
in this book is driven by a clear application, namely, the semantic analysis
of logic programming. In pursuing this application, we have developed sev­
eral results which provide conceptual insights into the relationships between
domain-theoretic semantics and metric semantics. A key role is played by the
Scott and Cantor topologies (Chapter 3) as the underlying spaces. Another
key role is played by the relationship between ordered spaces and (generalized)
metric spaces (Section 4.8) and by the various fixed-point theorems which can
be provided for these spaces (Chapter 4), some of which have been taken
directly from work on quantitative domain theory.
A theme which has not been taken up in this book in detail and which pro­
vides scope for further work is to investigate more closely how the application­
1 See also [Hitzler and Seda, 1999c, Straccia et al., 2009] for some more investigations

into these matters.

2 See [Scott, 1982a].
3 See [de Bakker, 2002].
4 Initiated by work such as [Smyth, 1987].
 Final Thoughts 223

driven work of quantitative domain theory, as discussed herein, relates to

theory-driven advances in the same topic, which were developed in parallel.5

8.3 Fixed-Point Theorems for Generalized Metric Spaces

Fixed-point theorems have a rightful place in the core arsenal of mathemat­
ical tools applicable to theoretical computer science, with many applications
outside this realm of course. The Banach contraction mapping theorem, The­
orem 4.2.3, which is the starting point for many of the investigations in this
book, is one of the most fundamental of these theorems.
In this book, we have contributed to the study of fixed-point theory by ex­
ploring generalized metrics and thereby providing a compilation of extensions
of the Banach theorem, together with results concerning their relationships
with order-theoretic fixed-point theorems (Chapter 4). We furthermore pro­
vide evidence of the usefulness of these theorems by applying them, throughout
the book, to the study of the semantics of logic programs.
Another theme which has not been taken up here, and again is scope for
further work, is a systematic investigation of the extent to which the Banach
theorem, and its relatives, remains valid with respect to generalized distance
functions under weaker and weaker conditions. Specifically, how weak can
the ambient spaces be and still support a reasonable version of the Banach
theorem?

8.4 The Foundations of Knowledge Representation and

Reasoning
Knowledge Representation and Reasoning (KR) is one of the classical
branches of Artificial Intelligence. Currently, it is experiencing massive re­
newed interest due to the advent of the Semantic Web.6 In a nutshell, the Se­
mantic Web strives to improve the World Wide Web by making Web content
machine-understandable, and it does so by using KR methods, more precisely,
by endowing Web content with additional meta-content in the form of knowl­
edge bases (so-called ontologies), which describe the content in a logic-based
format.
Several KR languages have been developed and standardized by the World
5 The papers [Waszkiewicz, 2002, Waszkiewicz, 2003, Waszkiewicz, 2006, Krötzsch, 2006,

Künzi et al., 2006, Künzi and Kivuvu, 2008], for example, may be consulted.
6 See [Hitzler et al., 2009b] for an introductory textbook.
224 Mathematical Aspects of Logic Programming Semantics

Wide Web Consortium7 for this purpose. One of them, called RIF,8 is essen­
tially a logic programming language, and other ontology languages can also
be understood as logic programming variants.9
In the light of such recent developments, theoretical investigations into
logic programming, as provided in this book, gain further interest. It can be
conjectured that the methods developed in this book may be used for design
and analysis of new KR languages suitable for application purposes.
Conceptually interesting from this point of view is the observation that
the methods of analysis provided herein are close to a denotational seman­
tics approach and thus complement the historically model-theory-driven se­
mantics in KR languages. In particular, there may be scope for the study
of decidability and/or semi-decidability of KR languages based on the level-
mappings approach discussed in Chapter 2,10 a topic which has so far been
largely neglected for logic programming, although it has played a major role in
the development of the currently main ontology language, the Web Ontology
Language OWL.11

8.5 Clarifying Logic Programming Semantics

In this book, we have covered the most important semantics for normal
logic programs. However, many more different semantics for normal logic pro­
grams and generalizations of this paradigm have been defined in the literature.
The rationale behind these various semantics has been manifold, depending
on one’s point of view, which may be that of a programmer or inspired by
commonsense reasoning. Consequently, the constructions which lead to these
semantics are technically very diverse, and the exact relationships between
them have not yet been fully understood.
Our work, and in particular the treatment in Chapter 2, but also Sec­
tion 5.2, provides a uniform perspective of different logic programming seman­
tics, and it should be clear from the proofs that the approach adopted there
can be lifted to other fixed-point semantics, in particular to those involving
monotonic operators. It thus reconciles these semantics within an overarching
framework which can be used for easy comparison of semantics with respect
to syntactic structures that can be employed with them, that is, to deter­
mine the extent to which a semantics is able to break up positive or negative
dependencies or loops between atoms in programs.
7 W3C, http://www.w3.org/
8 See [Boley and Kifer, 2010, Hitzler et al., 2009b].
9 OWL RL [Hitzler et al., 2009a, Reynolds, 2010], ELP [Krötzsch et al., 2008], or F-Logic

[Kifer et al., 1995], for example.

10 For a preliminary investigation into this, see [Cherchago et al., 2007].
11 See [Hitzler et al., 2009a, Hitzler et al., 2009b].
 Final Thoughts 225

It still remains to be seen, however, how far this approach can be carried,12
and whether or not it is possible to establish a meta-theory which goes beyond
mere characterization.13

8.6 Symbolic and Subsymbolic Representations

How to overcome the gap between symbolic and subsymbolic represen­
tations, and how to integrate them in an efficient and effective manner, is
a topic of growing interdisciplinary importance. It is driven by advances in
neuroimaging, which call for the modelling of findings in neuroscience on a
higher and higher level of abstraction, and by the search in Cognitive Science
for suitable cognitive architectures to model complex behaviour. By symbolic
we mean, of course, knowledge representation formalisms based on logic or
similar algebraic structures, while the term subsymbolic refers to paradigms
such as artificial neural networks, where knowledge is not represented in a
crisp, declarative way.
The topology-driven view of logic programming semantics which we pursue
herein indirectly embraces this theme by providing a conceptual bridge be­
tween the discrete (symbolic) world of logic and the continuous (subsymbolic)
world of topology and analysis on the reals.
While, originally, we developed this point of view purely for the purposes of
analyzing logic programs and in order to advance quantitative domain theory,
it bears, at least conceptually, on the symbolic/subsymbolic issue. However, we
have not pursued this in any structured manner, apart from developing neural-
symbolic integration (Chapter 7), albeit with a different initial motivation
(see Section 8.7). The question remains open to what extent our insights can
contribute to the larger quest.

8.7 Neural-Symbolic Integration

Our work on neural-symbolic integration started as a straightforward ap­
plication of our topological approach to logic programming semantics. The
pursuit (Chapter 7) was then driven mainly by an engineering motivation (as

12 Disjunctive well-founded semantics were compared using this approach in the paper

[Knorr and Hitzler, 2007], but only with limited success since the characterizations became
rather complicated.
13 In [Cherchago et al., 2007], for example, level mappings were used to study decidability

properties.
226 Mathematical Aspects of Logic Programming Semantics

opposed to a cognitive science motivation as discussed in Section 8.6), that

is, by the idea of combining logic programming and artificial neural networks
in such a way that the best of both worlds – declarativeness, trainability,
robustness, and reasoning capabilities – is retained.
Indeed, this effort has paid off, and while we provide only the theoretical
underpinnings in Chapter 7, we are indeed able to show that a declarative,
trainable, robust, and reasonable system can be developed on these grounds,14
although it has to be said that the advance remains conceptual in nature
because the system is severely limited in terms of the size of the knowledge
base involved. Nevertheless, it is to date one of the two reported systems with
these capabilities.15
Significant further advances on this front, in particular with respect to
the integration of learning and reasoning, would be highly appreciated in
practice.16

8.8 Topology, Programming, and Artificial Intelligence

It has been argued that there is a strong relationship between topologi­
cal dynamics (chaos theory), logic programming, neural networks, and other
paradigms, and in particular this is so in the context of emergent behaviour
as represented by cellular automata, say.17 Indeed, from a bird’s eye perspec­
tive each seems to be capable of being mapped onto the others. At the same
time, the study in any one of these paradigms seems to pose the same sort of
obstacles found in the others, particularly is this so in relation to the handling
of chaotic dynamics and emergence.
Some of the work in this book contributes to this discussion, especially with
respect to topological dynamics, logic programming, and neural networks, as
discussed in Section 7.5.18 Obviously, this is only a small stepping stone in
the pursuit of these issues which, once fully understood, will provide a major

14 See [Bader et al., 2007, Bader et al., 2008, Bader, 2009] for details.
15 The approach in [Gust et al., 2007] achieves similar results with entirely different meth­
ods.
16 For a discussion of the Semantic Web (see Section 8.4) as a potential test case for neural-

symbolic integration, see [Hitzler et al., 2005]. For a general discussion of the need for the
integration of learning and reasoning for Semantic Web applications, see [Hitzler, 2009,
Hitzler and van Harmelen, 2010].
17 See, for example, [Blair et al., 1997a, Blair et al., 1999]
18 In [Bader and Hitzler, 2004], it was shown that there is indeed a tight relationship

between logic programs and fractals in the sense in which they arise as attractors of iterated
function systems.
 Final Thoughts 227

advance in our overall understanding of complex phenomena. However, we may

not yet have the mathematical tools available to really understand them.19

19 This last sentence is a citation from a keynote talk given by Howard A. Blair, of Syracuse

University, at the MFCSIT2000 conference in Cork, Ireland.

Appendix

Transfinite Induction and General

Topology

In order to help make our discussions relatively self-contained, it will be con­

venient to collect together in this Appendix the basic facts and notation we
need from the theory of ordinals1 and from the subject of general topology.

A.1 The Principle of Transfinite Induction

We begin with a brief discussion of the theory of ordinals and transfinite
induction. In particular, we give a statement of the principle of transfinite
induction in the form in which we make use of it on a number of occasions.

A.1.1 Definition A partially ordered set X is well-ordered or is a well-

ordering if each non-empty subset of X has a first or least element.

A.1.2 Example (1) The set N of natural numbers is well-ordered in the usual
ordering ≤ on N.
(2) The set Z of integers is not well-ordered in the usual ordering ≤ on Z.

A.1.3 Lemma The following statements hold.

(a) Every well-ordered set is linearly ordered.

(b) No well-ordered set contains an infinite strictly descending sequence.

Proof: (a) Let (X, ≤X ) be a well-ordered set, and let x, y ∈ X. Then the set
{x, y} is a non-empty subset of X and hence has a least element, x, say. But
then x ≤X y, which establishes (a).
For (b), suppose that (xn )n∈N is an infinite strictly decreasing sequence in
the well-ordered set (X, ≤X ). Then {xn | n ∈ N} itself is a non-empty subset
1 Our treatment of these matters is informal and non-axiomatic and is in the spirit of the

book [Halmos, 1998] to which we refer the reader for further details.

229
230 Mathematical Aspects of Logic Programming Semantics

of X which has no least element, contradicting the hypothesis that (X, ≤X )

is well-ordered. •

Given two well-ordered sets (X, ≤X ) and (Y, ≤Y ), we call f : X → Y

monotonic if a ≤X b implies f (a) ≤Y f (b) for all a, b ∈ X. If f is also
injective, then f is called an embedding of X into Y . If f is both monotonic
and bijective, then f is called an order isomorphism between X and Y , and
in this case the two well-orderings X and Y are called isomorphic. Note that
all these definitions are consistent with the definitions concerning orderings
made in Chapter 1.

A.1.4 Definition Suppose that (X, ≤X ) is a well-ordered set and that x0 ∈

X. We call the set I = I(x0 ) = {x ∈ X | x ≤X x0 } the initial segment of X
determined by x0 . We call an initial segment I of X a proper initial segment
if I is a proper subset of X.

A.1.5 Definition Suppose that (X, ≤X ) and (Y, ≤Y ) are two well-ordered
sets. Then we write X ≤ Y if X is isomorphic to an initial segment of Y . We
write X < Y if X is isomorphic to a proper initial segment of Y .

A.1.6 Theorem For any two well-ordered sets X and Y , exactly one of the
following statements holds.

(a) X < Y .

(b) X > Y .

(c) X and Y are isomorphic.

Proof: We first prove the following statement.

(1) No well-ordered set (Z, ≤Z ) is isomorphic to a proper initial segment of
itself.
To see this, suppose f : I → Z is an isomorphism, where I is a proper
initial segment of Z. Then we cannot have f (x) = x for all x ∈ I; otherwise, f
would not be surjective. Let x0 be the least element of the set of those elements
x of I such that f (x) = x. Noting in particular that f (x0 ) = x0 , we see that
we cannot have f (x0 ) <Z x0 otherwise f (f (x0 )) = f (x0 ) by minimality of
x0 , and this yields the contradiction that f is not injective. Hence it must
be the case that x0 <Z f (x0 ). Now let x1 ∈ I be such that f (x1 ) = x0 .
Then x1 = x0 because f (x0 ) = x0 . If x1 <Z x0 , then by definition of x0
again, we obtain x0 = f (x1 ) = x1 <Z x0 , which is impossible. If x0 <Z x1 ,
then f (x1 ) = x0 <Z f (x0 ), which contradicts the monotonicity of f . Hence,
Statement (1) holds.
We will also need the following statement.
(2) Suppose the well-ordered sets (W, ≤W ) and (Z, ≤Z ) are isomorphic. Then
there is a unique isomorphism f : (W, ≤W ) → (Z, ≤Z ).
 Transfinite Induction and General Topology 231

In order to see this, suppose f, g : W → Z are isomorphisms. We show

f = g. Assume this is not the case, and let w0 ∈ W be the ≤W -least w
such that f (w) = g(w); suppose in fact that f (w0 ) <Z g(w0 ) for the sake
of argument. Let w1 ∈ W be such that g(w1 ) = f (w0 ). Then w1 = w0 .
If w1 <W w0 , then by minimality of w0 and monotonicity of f , we obtain
g(w1 ) = f (w1 ) <W f (w0 ) = g(w1 ), which is impossible. If w0 <W w1 , then
by monotonicity of g, we have f (w0 ) <Z g(w0 ) <Z g(w1 ) = f (w0 ), which is
also impossible. So Statement (2) holds.
We now turn to the proof of the theorem.
Define the relation R from X to Y by R(x, y) if and only if the initial
segments {w ∈ X | w ≤X x} and {v ∈ Y | v ≤Y y} are isomorphic, where
x ∈ X and y ∈ Y . First note that R(x, y1 ) and R(x, y2 ) implies y1 = y2 by
Statement (1). So R is a partial function. By symmetry, transitivity, and (1)
again, R is also injective.
We next show that dom(R) is an initial segment of X. Suppose x2 ∈
dom(R), say, R(x2 , y2 ), and let x1 <X x2 . Let f be the isomorphism between
the initial segments corresponding to x2 and y2 . Then the initial segments
corresponding to x1 and f (x1 ) are also isomorphic, so R(x1 , f (x1 )), and hence
x1 ∈ dom(R). We have also shown that R is order-preserving.
A similar argument shows that the range of R is an initial segment of Y .
Hence, R is an isomorphism from an initial segment I(x0 ), say, of X to an
initial segment J(y0 ), say, of Y ; thus, R(x0 , y0 ) holds.
Now consider the following cases. If I(x0 ) = X, but J(y0 ) = Y , then case
(a) holds. If I(x0 ) = X, but J(y0 ) = Y , then case (b) holds. If I(x0 ) = X
and J(y0 ) = Y , then case (c) holds. Suppose finally that I(x0 ) = X and
J(y0 ) = Y . Let x1 be the first element of X \ I(x0 ) and y1 be the first element
of Y \ J(y0 ); then x1 is not in the domain of R (and y1 is not in the range
of R). But clearly I(x0 ) ∪ {x1 } is the initial segment I(x1 ) and J(y0 ) ∪ {y1 }
is the initial segment J(y1 ), and, furthermore, I(x1 ) and J(y1 ) are clearly
isomorphic by an isomorphism which, by (2), must be an extension of R. We
therefore obtain the contradiction that x1 is in the domain of R.
Hence, only one of (a), (b), (c) holds, as required. •

Next, we state without proof a well-known theorem usually attributed to

E. Zermelo. This theorem has the consequence that any set is a carrier set for
some ordinal, see [Halmos, 1998] for details.

A.1.7 Theorem (The Well-Ordering Theorem) Every set can be well-

ordered.

A.1.8 Definition An ordinal or ordinal number is an equivalence class of a

well-ordering under the equivalence relation of isomorphism.
The ordinals themselves can be ordered as follows. First, for any well-
ordered set A, let #A denote the equivalence class of A under the equivalence
relation of isomorphism. Suppose that α = #A and β = #B are ordinals. We
232 Mathematical Aspects of Logic Programming Semantics

define the ordering ≤ on the ordinals by α ≤ β if and only if A ≤ B, and we

note that ≤ is easily seen to be well-defined on the ordinals. Furthermore, by
Theorem A.1.6, the ordering ≤ is a partial order and, as we show next, is in
fact a well-order.

A.1.9 Lemma Let X be a linearly ordered partially ordered set which is not
well-ordered. Then X contains an infinite strictly descending sequence.

Proof: If X is not well-ordered, then there exists a subset X0 of X which

does not contain a least element. Choose some x0 ∈ X0 , and note that X1 =
{y ∈ X | y < x0 } does not contain a least element. Now assume that some
xi ∈ X has been chosen such that the set Xi+1 = {y ∈ X | y < xi } does
not contain a least element. Then we can choose xi+1 ∈ Xi+1 arbitrarily and
obtain xi+1 < xi and also that Xi+2 = {y ∈ X | y < xi+1 } does not contain
a least element. By the inductive argument just given, we obtain an infinite
strictly descending sequence (xn ), as required. •

A.1.10 Proposition Every set of ordinals is itself well-ordered by ≤.

Proof: We begin by noting that if α and β are ordinals such that α ≤ β and
α = #A and β = #B, then we can assume without loss of generality that
A ⊆ B; we will make use of this observation in what follows.
Let X be a set of ordinals which is not well-ordered. Then, by Lemma
A.1.9, X contains an infinite descending sequence α0 > α1 > α2 > . . . of
ordinals. For each i ∈ N, suppose that αi = #Ai and that Ai ⊃ Ai+1 . Then
for each i ∈ N there exists ai ∈ Ai \ Ai+1 . Hence, {ai | i ∈ N} ⊆ A0 is a subset
of A0 without a least element, which is impossible. •

It is common practice to identify any ordinal α with the set of all ordinals
β such that β < α; so, in these terms, β < α if and only if β ∈ α. We
will follow this practice in the following. In particular, when we speak of a
mapping f : X → α, where α is an ordinal, we mean, in fact, a mapping
f : X → {β | β < α}.
Ordinals fall into two classes. A successor ordinal is an ordinal α such that
there is a greatest ordinal β with β < α. In this case, α is called the successor
of β and may be denoted by β + 1; we also call β the predecessor of α and
may denote it by α − 1. Any ordinal which is not a successor ordinal is called
a limit ordinal .
Any ordinal has a successor. To see this, let α be an ordinal and identify
it with the set of ordinals {β | β < α}. Then α ∪ {α} is an ordinal above α
and indeed is the least ordinal above α and therefore is the successor α + 1 of
α.
We next give an example containing details of some familiar ordinals.
 Transfinite Induction and General Topology 233

A.1.11 Example It is easy to see that any finite set A = {a1 , . . . , an }, con­
taining n elements, can be well-ordered in essentially one way. Thus, if A
and B are any well-ordered sets containing n elements, then A and B are
isomorphic. Standard notation for the finite ordinals, together with canonical
representatives for them, is as follows: 0 = #∅, 1 = #{∅}, 2 = #{∅, {∅}},
3 = #{∅, {∅}, {∅, {∅}}}, etc. Thus, we are using the same symbols 0, 1, 2, 3, . . .
to denote natural numbers and ordinal numbers (as well as cardinal numbers),
but the context in which they occur will determine their meaning. Often, we
consider an ordinal to be the set of all its predecessors, as already noted, in
which case we view the ordinal n as the set {0, 1, . . . , n − 1} for each n. Fur­
thermore, 0 is the least ordinal, 1 is the successor of 0, 2 is the successor of 1,
etc. Thus, we have 0 < 1 < 2 < 3 < · · · as ordinals.
Turning now to ordinals determined by infinite sets, we note first that in­
finite sets can be well-ordered in more than one way. For example, the set N
of natural numbers can be well-ordered by writing it as {1, 3, 5, . . . ; 2, 4, 6, . . .}
and ordering it from left to right. The resulting well-order is clearly not iso­
morphic to N well-ordered by the usual order on N. Indeed, the first infi­
nite ordinal or least infinite ordinal , denoted by ω, is the ordinal determined
by N in its usual order, that is, ω = #N. Thus, ω is the first limit ordi­
nal. The successor of ω is ω + 1 = {0, 1, 2, . . . , ω}, the successor of which
is ω + 2 = (ω + 1) + 1 = {0, 1, 2, . . . , ω, ω + 1}, etc. The next, or second,
limit ordinal is denoted by ω2 = {0, 1, 2, . . . , ω, ω + 1, ω + 2, . . . , ω + n, . . .}
etc. In this way, the ordinals form a transfinite sequence, and indeed any
non-finite ordinal is sometimes called a transfinite number. Thus, we have
0 < 1 < 2 < 3 < · · · < ω < ω + 1 < ω + 2 < · · · < ω2 < ω2 + 1 < ω2 + 2 <
· · · < ω3 < · · · < ωn < · · · < ωω = ω 2 < ω 2 + 1 < ω 2 + 2 < · · · as ordi­
nal numbers. Note also that all the ordinals we have so far displayed in this
example are determined by countable sets. The first uncountable ordinal is
denoted by ω1 and as a set is the uncountable well-ordered set containing all
the countable ordinals. •

We are now in a position to consider the principle of transfinite induction.

The reader may note that it is an extension, from N to arbitrary well-ordered
sets, of the well-known strong form2 of the principle of mathematical induc­
tion.

A.1.12 Theorem (Principle of Transfinite Induction) Suppose that A

is any well-ordered set and B is a subset of A which satisfies the statement
that a ∈ B whenever x ∈ B for all x < a. Then B = A.

Proof: If B = A, then A \ B = ∅. By well-ordering of A and therefore of any

subset B of A, A \ B has a first element x0 , say. But now we have x ∈ B for

2 Also known as course of values induction.

234 Mathematical Aspects of Logic Programming Semantics

all x < x0 , and the induction hypothesis leads to the conclusion that x0 must
belong to B. This contradiction shows that B = A, as required. •

A.1.13 Corollary Suppose that A is a well-ordered set and {p(a) | a ∈ A} is

a set of statements indexed by A. Suppose further that for all b ∈ A it follows
that p(b) is true if p(x) is true for all x < b. Then p(a) is true for all a ∈ A.
In fact, the form in which we will usually apply the principle of transfinite
induction is as follows.

A.1.14 Corollary Suppose that p(α) is a statement depending on the ordi­

nal α. Suppose further that for all ordinals β, p(β) is true if p(γ) is true for
all γ < β. Then p(α) is true for all ordinals α.
When applying the principle of transfinite induction as a proof principle, as
formulated in Corollary A.1.14, it is usually convenient to split the argument
into two cases. The first of these is when β is assumed to be a successor ordinal,
and the second is when β is assumed to be a limit ordinal.

A.2 Basic Concepts from General Topology

We next turn to giving a brief overview of the general topology we need
at various points in our discussions.3 In addition, we include here the proofs
of the results we stated without proof in our treatment of the Scott topology
in Chapter 3.

A.2.1 Definition A topology on a set X is a collection τ of subsets of X,

called the open sets of τ , satisfying the following properties.
(1) Any union of elements of τ belongs to τ .
(2) Any finite intersection of elements of τ belongs to τ .
(3) ∅ and X belong to τ .
The pair (X, τ ), or simply X by an abuse of notation, is called a topological
space.

A.2.2 Definition Given two topologies τ1 and τ2 on a set X, we say that τ1

is weaker or coarser than τ2 , or that τ2 is stronger or finer than τ1 , if τ1 ⊆ τ2 .
3 Our background references for the material we need from general topology are the books

[Kelley, 1975] and [Willard, 1970] to which we refer the reader for proofs of the results we
simply state.
 Transfinite Induction and General Topology 235

Given a set X, the coarsest topology which can be defined on X is the

indiscrete topology in which the only open sets are ∅ and X. At the other
extreme, the finest topology which can be defined on X is the discrete topology
in which all subsets of X are taken as open sets.

A.2.3 Definition If X is a topological space and x ∈ X, then a neighbour­

hood of x is a set U containing an open set V containing x, that is, x ∈ V ⊆ U ,
where V is open. The neighbourhood system Ux of x is the collection of all
neighbourhoods of x.

A.2.4 Definition A neighbourhood base at x in the topological space X is

a subcollection Bx ⊆ Ux such that, for each U ∈ Ux , there exists V ∈ Bx
satisfying V ⊆ U . Thus, Ux = {U ⊆ X | V ⊆ U for some V ∈ Bx }. The
elements of Bx are called basic neighbourhoods of x.

A.2.5 Theorem Let X be a topological space, and, for each x ∈ X, let Bx

be a neighbourhood base at x. Then the following properties hold.

(a) If V ∈ Bx , then x ∈ V .

(b) If V1 , V2 ∈ Bx , then there is V3 ∈ Bx satisfying V3 ⊆ V1 ∩ V2 .

(c) If V ∈ Bx , there is some V0 ∈ Bx such that if y ∈ V0 , then there is W ∈ By

satisfying W ⊆ V .

(d) G ⊆ X is open if and only if G contains a basic neighbourhood of each of

its points.

Conversely, suppose that X is a set and that a collection Bx of subsets of X,

called basic neighbourhoods of x, is assigned to each element x ∈ X in such a
way that (a), (b), and (c) above are satisfied. If we then define a set G to be
open if and only if it contains a basic neighbourhood of each of its points, as
in (d), we obtain a topology on X in which Bx is a neighbourhood base at x
for each x ∈ X.

A.2.6 Definition In a topological space (X, τ ), a base for τ (or a base for X
by an abuse of terminology) is a collection B ⊆ τ of subsets of X such that
each element of τ is a union of elements of B. Equivalently, B is a base for τ if
and only if whenever V ∈ τ and x ∈ V , there is U ∈ B such that x ∈ U ⊆ V .
Furthermore, a collection C ⊆ τ is called a subbase for τ (or a subbase for X)
if the collection of all finite intersections of elements of C forms a base for τ .

A.2.7 Theorem A collection B of subsets of a set X is a base for a topology

on X if and only if the following conditions are satisfied.

(a) B ∈B B = X.
236 Mathematical Aspects of Logic Programming Semantics

(b) Whenever B1 , B2 ∈ B and x ∈ B1 ∩ B2 , there is B3 ∈ B satisfying

x ∈ B3 ⊆ B1 ∩ B2 .

Furthermore, any collection C of subsets of X is a subbase for some topology

on X, namely, the topology formed by taking all arbitrary unions of finite
intersections of elements of C.

A.2.8 Theorem Suppose that B is a collection of open sets in a topological

space X. Then B is a base for X if and only if, for each x ∈ X, the collection
Bx = {B ∈ B | x ∈ B} is a neighbourhood base at x.

As noted in Definition A.2.1, the elements of τ are called the open sets in
the given topology on X. By definition, we call a subset F of X closed if its
complement, X \ F , is open. It follows immediately that ∅ and X are closed
sets, that any finite union of closed sets is itself closed, and that an arbitrary
intersection of closed sets is closed. Therefore, given an arbitrary subset E of
X, the intersection E of all the closed sets containing E is a closed set, the
smallest closed set containing E, and is called the closure of E. Clearly, a set
F is closed if and only if F = F . Dually, one defines the interior U o of a
subset U of X to be the largest open set contained in U , and it is of course
the union of all the open sets contained in U . Moreover, it is also clear that a
set O is open if and only if O = Oo .
A closure operator (also known as a Kuratowski, or topological, closure
operator) on a set X is a mapping c : P(X) → P(X), from the power set
P(X) of X into itself, subject to the following axioms.

(1) ∅c = ∅.

(2) A ⊆ Ac for all A ⊆ X.

(3) (A ∪ B)c = Ac ∪ B c for all A, B ⊆ X.

(4) Ac = (Ac)c for all A ⊆ X.

Just as the notion of an open set can be taken as basic in defining topolo­
gies, so clearly can the notion of a closed set. More interesting is the fact that
closure can be taken as fundamental, and indeed the characteristic properties
of closure are precisely the four just stated in defining a closure operator, in
the following sense.

A.2.9 Theorem Let X be a non-empty set, and let c : P(X) → P(X) be a

closure operator on X. Then τ = {X \ A | A ⊆ X, A = Ac} is a topology on
X, called the topology associated with c, in which we have A = Ac for each
subset A of X. Thus, Ac is the topological closure in X of each subset A of
X with respect to the topology τ associated with c.
 Transfinite Induction and General Topology 237

A.3 Convergence
It is well known, see [Willard, 1970, Chapter 4], that sequences are not
adequate to describe all basic notions in topological spaces other than in the
class of first countable spaces (a topological space is called first countable if
it has a countable neighbourhood base at each of its points). One therefore
needs notions more general than that of sequence. Such generalizations are
provided by nets and filters, either of which is adequate to describe all topo­
logical concepts. Indeed, convergence itself can be taken as the fundamental
concept in developing topology, see Theorem 3.1.3, and this is the point of
view adopted in Chapter 3. However, we choose to work here only with nets
for reasons already mentioned in Chapter 3.

A.3.1 Definition A net in a set X is a mapping s : I → X, where (I, ≤) or

simply I is a directed set in which the ordering ≤ is reflexive and transitive.
For each i ∈ I, we denote s(i) by si and denote the net s : I → X by (si )i∈I or
simply by (si ) or just by si if no confusion results. Similarly, sequences (sn )n∈N ,
being special cases of nets, may be denoted simply by (sn ) or sn . Given a net
(si )i∈I in X and an element i0 of I, we call the set (si )i0 ≤i = {si | i0 ≤ i}
a tail of (si )i∈I . A property will be said to hold eventually with respect to a
net (si )i∈I if it holds for some tail of the net.

A.3.2 Definition A subnet t of a net s : I → X is a net t : J → X satisfying

(i) t = s ◦ ϕ, where ϕ is a function mapping J into I, and (ii) for each i0 ∈ I,
there exists j0 ∈ J such that ϕ(j) ≥ i0 whenever j ≥ j0 . The point s ◦ ϕ (j)
is often denoted by sij , and we refer to the subnet (sij )j∈J of (si )i∈I .

A.3.3 Definition Let X be a topological space, and let x ∈ X. A net (si )i∈I
in X will be said to converge to x, written si → x or limi si = x, if, for each
neighbourhood U of x, there exists i0 ∈ I such that si ∈ U whenever i0 ≤ i.
If si → x, then we call x a limit of si .

Since the singleton set {x} is a neighbourhood of x if X is endowed with

the discrete topology, it follows that si → x in the discrete topology if and
only if (si ) is eventually constant.
The notion of continuous function between topological spaces is fundamen­
tal in the subject. There are several ways of formulating this concept, but the
following is perhaps the most intuitive.

A.3.4 Definition Let X and Y be topological spaces, and suppose that f :

X → Y is a function. Then f is said to be continuous at x ∈ X if, for each
neighbourhood V of f (x) in Y , there is a neighbourhood U of x in X such
that f (U ) ⊆ V . We say f is continuous if it is continuous at x for each x ∈ X.
238 Mathematical Aspects of Logic Programming Semantics

The sense, mentioned earlier, in which nets can describe all basic topolog­
ical notions can now be clarified.

A.3.5 Theorem Let X and Y be topological spaces. Then the following

statements hold.
(a) Let E ⊆ X. Then x ∈ E if and only if there is a net (si ) in E such that
si → x.
(b) A subset O of X is open if and only if, whenever x ∈ O and (si ) is a net
such that si → x, we have that (si ) is eventually in O.
(c) A subset F of X is closed if and only if, whenever (si ) is a net in F and
si → x, we have x ∈ F .
(d) A function f : X → Y is continuous at x ∈ X if and only if, whenever
si → x in X, we have f (si ) → f (x) in Y .
Proof: We include a proof of (b) here since we have specific need of the result.
Suppose that O is open, that x ∈ O, and that si → x. Then it is clear from
the definition of net convergence that (si ) is eventually in O.
Conversely, assuming the stated condition, we show that O contains a
neighbourhood of each of its points and hence is open. Let x ∈ O, and let
Ux be the neighbourhood system of x. Let I = {(y, U ) | y ∈ U ∈ Ux } or­
dered by (y1 , U1 ) ≤ (y2 , U2 ) if and only if U2 ⊆ U1 . Then it is easy to see
that the ordering ≤ directs I and also that the net s : I → X defined by
s(y, U ) = y converges to x. By our current hypothesis, this net is eventually
in O. Let (y0 , U0 ) be such that s(y,U ) = y ∈ O whenever (y0 , U0 ) ≤ (y, U ).
Since (y0 , U0 ) ≤ (y, U0 ) for all y ∈ U0 , we conclude that x ∈ U0 ⊆ O, as
required. •

A.4 Separation Properties and Compactness

It is important to have sufficiently many open sets to be able to distinguish,
in some way, between points in a topological space by means of the open sets.
This is usually done by means of the following axioms.

A.4.1 Definition Let X be a topological space.

(1) We call X a T0 -space if, whenever x and y are distinct points of X, there
is an open set containing one but not the other.
(2) We call X a T1 -space if, whenever x and y are distinct points of X, there
is a neighbourhood of each not containing the other.
 Transfinite Induction and General Topology 239

(3) We call X a T2 -space or a Hausdorff space if, whenever x and y are

distinct points of X, there are disjoint neighbourhoods of x and y.

One of the important properties of Hausdorff spaces is that stated in the

following result.

A.4.2 Theorem A topological space is Hausdorff if and only if every con­

vergent net in X has a unique limit.

On the other hand, it is important that there not be too many open sets
in a certain sense.

A.4.3 Definition Let X be a topological space. Then  an open cover {Ui |

i ∈ I} of X is a collection of open sets Ui such that i∈I Ui = X. A subcover
of an open cover {Ui | i ∈ I} is a cover {Vj | j ∈ J }, where J ⊆ I. We call a
topological space X compact if every open cover of X has a finite subcover.

A.5 Subspaces and Products

There are several ways in which one can create new topological spaces
from given ones. We discuss here just two of these, namely, the process of
forming subspaces of topological spaces and the process of forming products
of families of topological spaces.

A.5.1 Definition Let (X, τ ) be a topological space, and let S ⊆ X be a

subset of X. Then the collection τS = {S ∩ O | O ∈ τ } gives a topology on S,
called the relative topology or the subspace topology for S. The space (S, τS )
is called a subspace of (X, τ ) or just a subspace of X.

Whenever one has a topological space X and a subset S of X, it will be

assumed that S has been endowed with the subspace topology of X unless
stated to the contrary. Notice that the sets S ∩ O, where O is open in X, need
not be open in X unless S itself is an open set of X.
Now suppose that Xi is a topological space for each i, where i is an element
s set I. As usual, we
of some index  denote the product of the family {Xi | i ∈ I}
of sets by i∈I Xi = {f : I → i∈I Xi | f (i) ∈ Xis }. Associated with any such
product are the mappings πj , j ∈ I, where πj : i∈I Xi → Xj is defined by
πj (f ) = f (j). Indeed, πj is termed the projection ons the j-th factor .
There is a natural topology one can define on i∈I Xi determined by the
projections as follows. Choose any finite set {i1 , . . . , in } of elements of I, and
choose corresponding open sets Uij in Xij , for j = 1, . . . , n. Then we take
−1 −1
the collection of s sets of the form πi1 (Ui1 ) ∩ . . . ∩ πin (Uin ) as a base for
a topology on i∈I Xi called the product topology or the Tychonoff product
240 Mathematical Aspects of Logic Programming Semantics

topology. Indeed, the sets πi−1 (Ui ) form a subbase for this topology, where
i ∈ I and Ui is an open set in Xi . It is immediate that each of the projections
πi is continuous relative to the product topology and the given topology on
the factor Xi . s
Subspaces of X and products i∈I Xi inherit certain properties enjoyed
by X and the Xi , respectively, as one would expect. We summarize next the
ones relevant to our needs in the following theorem.

A.5.2 Theorem The following statements hold.

(a) Subspaces of T0 or Hausdorff spaces are T0 or Hausdorff, respectively.
(b) If X is compact and S is a closed subset of X, then S is compact (as a
topological space in its own right). If X is Hausdorff and S is compact,
then S is a closed subset of X.
s
(c) A non-empty product i∈I Xi is T0 or Hausdorff if and only if each factor
space Xi is T0 or Hausdorff, respectively.
s
(d) (Tychonoff ’s theorem) A non-empty product i∈I Xi is compact if
and only if each factor space is compact.
s
(e) A net (fλ ) in a product space i∈I Xi converges to f if and only if, for
each index i ∈ I, we have πi (fλ ) → πi (f ) in Xi .

A.6 The Scott Topology

We present here the proofs of those results which were simply stated in
Chapter 3 concerning the Scott topology. In fact, our development constitutes
a treatment of the Scott topology from the point of view of convergence.
Unless stated to the contrary, (D, [) will denote throughout some fixed, but
arbitrary, domain with set Dc of compact elements.

A.6.1 Proposition Suppose that A⊆ D is a directed set. Then A is a net in

D, and, as a net, we have that A → A in the Scott topology. In particular,
for each s ∈ D, approx(s) → s in the Scott topology.
Proof: Write A = {ai | i ∈ I} for some index set I, which we identify with A.
Then I is clearly directed by the ordering ≤ obtained by restricting
 [ to A.
Therefore, the inclusion map I → D is a net in D. Let A = A  and suppose
that O is a neighbourhood of A in the Scott topology. Thus, A ∈ O, and
hence there exists some index i0 such that ai0 ∈ O. But O is upwards closed,
and therefore ai ∈ O whenever i0 ≤ i. Thus, A → A, as required. •
 Transfinite Induction and General Topology 241

A.6.2 Proposition Suppose that f : D → E is continuous in the Scott

topologies on domains D and E. Then whenever x ∈ D, a ∈ Dc , and a [ x,
we have f (a) [ f (x).

Proof: Let a ∈ Dc . Since f is continuous at a, given any Scott neighbourhood

V of f (a), there is a Scott neighbourhood U of a such that f (U ) ⊆ V . Let
b ∈ approx(f (a)) be arbitrary. Then V = ↑ b is a Scott neighbourhood of
f (a). Furthermore, ↑ a is a Scott neighbourhood of a contained in any Scott
neighbourhood U of a. Therefore, we have f (↑ a) ⊆ ↑ b. Thus, if a [ x, then
x ∈ ↑ a. Therefore, f (x) ∈ ↑ b, that is, b [ f (x). But b ∈ approx(f (a)) is
arbitrary. Therefore, f (a) [ f (x), as required. •

A.6.3 Proposition Suppose that f : D → E is continuous in the Scott

topologies on domains D and E. Then f is monotonic.

Proof: Suppose that x [ y in D. Note that if a ∈ approx(x) is arbitrary,

then a ∈ Dc and a [ x, so that a [ y. By Proposition A.6.2, we then have
f (a) [ f (y). Now, approx(x) can be thought of as a net approx(x) = {ai | i ∈
I}, as in Proposition A.6.1, and moreover ai → x. Therefore, f (ai ) → f (x).
Hence, by Theorem 3.2.4, for each b ∈ approx(f (x)) there is i0 such that
b [ f (ai ) whenever i0 ≤ i. But ai [ x [ y, for each i, and so ai [ y and
hence f (ai ) [ f (y) whenever i0 ≤ i by our first
 observation. From this we see
that b [ f (y). Finally, we now have f (x) = {b | b ∈ approx(f (x))} [ f (y)
so that f (x) [ f (y), as required. •

A.6.4 Proposition A function f : D → E, where D and E are domains,

is continuous in the Scott topologies on D and E if and only if it is order
continuous in the sense of Definition 1.1.7.

Proof: Suppose that f is continuous in the Scott topologies on D and E.

Then f  is monotonic by Proposition A.6.3. Let A ⊆ D be a directed set, and
let A = A. By Proposition A.6.1, A = {ai | i ∈ I} → A as a net, and hence
f (ai ) → f (A) by our hypothesis concerning f . Therefore, by Theorem 3.2.4,
for each b ∈ approx(f (A)), thereexists i0such that b [ f (ai ) whenever i0 ≤ i.
From this we  obtain f (A) = f
( A) = {b | b ∈ approx(f (A))} [ {f (ai ) |
i ∈ I} = f (A). Thus, f ( A) [ f (A), and it follows that f is order
continuous by the remarks following Definition 1.1.7.
Conversely, suppose that f is order continuous and that si → s in the Scott
topology on D. Now, f is monotonic. Therefore, on noting that approx(s) is
directed and thinking of it as the net {aj| j ∈ J }, we  have that the set
{f
 (aj ) | j ∈ J } is directed and f (s) = f ( approx(s)) = f (approx(s)) =
{f (aj ) | j ∈ J }. Therefore, given any b ∈ approx(f (s)), there is j ∈ J
such that b [ f (aj ), where aj ∈ approx(s). Since si → s, it follows from
Theorem 3.2.4 that there is i0 such that aj [ si whenever i0 ≤ i. Hence,
242 Mathematical Aspects of Logic Programming Semantics

by the monotonicity of f , we have that b [ f (aj ) [ f (si ) whenever i0 ≤ i.

Consequently, we have that f (si ) → f (s) in the Scott topology on E, and so
f is continuous in the Scott topologies, as required. •

Finally, we consider briefly the separation and compactness properties of

the Scott topology.

A.6.5 Proposition When endowed with the Scott topology, any domain
(D, [) is a compact T0 topological space, but is not T1 in general.

Proof: Suppose that {Ui | i ∈ I} is an open cover of D. Then we have ⊥ ∈ Uk ,

where Uk is some element of the given cover and ⊥ denotes the bottom element
of D. But ⊥ [ x for each x ∈ D and Uk is upwards closed, being Scott open.
Therefore, D ⊆ Uk , and so {Uk } is an open subcover of {Ui | i ∈ I}, and
hence D is compact.
We show next that D is T0 . Suppose that x, y ∈ D and x = y. First,
suppose that x and y are comparable, that is, either x [ y or y [ x; suppose
for the sake of argument that x [ y and, hence, that x c y since x = y.
We claim that there is a compact element a [ y such that either x c a [ y
or x and a are incomparable. If not, then for all compact elements a [ y,
we have that x and a are comparable and indeed a [ x. It follows now that
the supremum of such a is less than or equal to x, which is a contradiction
since in fact this supremum is y. But then, given the claim, ↑ a is a Scott
neighbourhood of y which does not contain x. Notice that if a is any compact
element and a [ x, then a [ y. So, any Scott neighbourhood of x contains y,
and we see that the condition in the definition of T0 is not symmetric in this
case.
Now suppose that x and y are incomparable. We claim this time that
there is a compact element a ∈ approx(x) such that a and y are incomparable.
Suppose that this is not the case, that is, suppose that for each a ∈ approx(x),
a and y are comparable. Certainly, it cannot be the case that y [ a; otherwise,
we immediately have y [ x. So  it must be the case that a [ y for each a ∈
approx(x). But then we have {a | a ∈ approx(x)} [ y, that is, x [ y, which
is again a contradiction. Now, given this claim, ↑ a is a Scott neighbourhood
of x not containing y. Notice that, by symmetry, in this case we also have
a Scott neighbourhood of y not containing x; thus, the T1 property actually
applies to some pairs in D (the incomparable pairs), but not to all pairs. In
any case, we now see that D is T0 .
Finally, take the two element domain D = {⊥, a}, where ⊥ c a. The Scott
topology on D contains ∅, ↑ ⊥ = D and ↑ a = {a} as its open sets (the set
{⊥} is not Scott open). This space D is not T1 since any neighbourhood of ⊥
contains a. •
Bibliography

[Abramsky and Jung, 1994] Abramsky, S. and Jung, A. (1994). Domain the­
ory. In Abramsky, S., Gabbay, D. M., and Maibaum, T. S., editors,
Handbook of Logic in Computer Science, Volume 3, pages 1–168. Oxford
University Press, Oxford, UK. An expanded and corrected version is
available at http://www.cs.bham.ac.uk/∼ axj/papers.html.

[Albeverio et al., 1999] Albeverio, S., Khrennikov, A. Y., and Kloeden, P.

(1999). Memory retrieval as a p-adic dynamical system. Biosystems,
49:105–115.

[Antoniou, 1996] Antoniou, G. (1996). Non-Monotonic Reasoning. MIT

Press, Cambridge, MA.

[Apt, 1997] Apt, K. R. (1997). From Logic Programming to Prolog. Inter­

national Series in Computer Science. Prentice Hall, Upper Saddle River,
NJ.

[Apt and Bezem, 1990] Apt, K. R. and Bezem, M. (1990). Acyclic programs.
In Warren, D. H. and Szeredi, P., editors, Proceedings of the Seventh
International Conference on Logic Programming, pages 617–633. MIT
Press, Cambridge, MA.

[Apt and Pedreschi, 1994] Apt, K. R. and Pedreschi, D. (1994). Modular ter­
mination proofs for logic and pure Prolog programs. In Levi, G., editor,
Advances in Logic Programming Theory, pages 183–229. Oxford Univer­
sity Press, Oxford, UK.

[Apt and Wallace, 2007] Apt, K. R. and Wallace, M. (2007). Constraint Logic
Programming Using Eclipse. Cambridge University Press, London; New
York.

[Apt et al., 1988] Apt, K. R., Blair, H. A., and Walker, A. (1988). Towards
a theory of declarative knowledge. In Minker, J., editor, Foundations
of Deductive Databases and Logic Programming, pages 89–148. Morgan
Kaufmann Publishers, Los Altos, CA.

[Arnold and Nivat, 1980a] Arnold, A. and Nivat, M. (1980a). Metric inter­
pretations of infinite trees and semantics of non-deterministic recursive
programs. Theoretical Computer Science, 11:181–205.

243
244 Bibliography

[Arnold and Nivat, 1980b] Arnold, A. and Nivat, M. (1980b). The metric
space of infinite trees: Algebraic and topological properties. Fundamenta
Informaticae, 3(4):445–476.

[Bader, 2009] Bader, S. (2009). Neural-Symbolic Integration. PhD thesis,

Department of Computer Science, TU Dresden, Dresden, Germany.

[Bader and Hitzler, 2004] Bader, S. and Hitzler, P. (2004). Logic programs,
iterated function systems, and recurrent radial basis function networks.
Journal of Applied Logic, 2(3):273–300.

[Bader and Hitzler, 2005] Bader, S. and Hitzler, P. (2005). Dimensions of

neural-symbolic integration — A structured survey. In Artemov, S. N.,
Barringer, H., d’Avila Garcez, A. S., Lamb, L. C., and Woods, J., editors,
We Will Show Them! Essays in Honour of Dov Gabbay, Volume One,
pages 167–194. College Publications, London, UK.

[Bader et al., 2005a] Bader, S., Hitzler, P., and d’Avila Garcez, A. S. (2005a).
Computing first-order logic programs by fibring artificial neural net­
works. In Russell, I. and Markov, Z., editors, Proceedings of the 18th
International Florida Artificial Intelligence Research Society Conference,
FLAIRS05, Clearwater Beach, Florida, May 2005, pages 314–319. AAAI
Press, Menlo Park, CA.

[Bader et al., 2005b] Bader, S., Hitzler, P., and Witzel, A. (2005b). Integrat­
ing first-order logic programs and connectionist systems — A construc­
tive approach. In d’Avila Garcez, A., Elman, J., and Hitzler, P., editors,
Proceedings of the IJCAI-05 Workshop on Neural-Symbolic Learning and
Reasoning, NeSy’05, Edinburgh. AAAI Press, Menlo Park, CA.

[Bader et al., 2006] Bader, S., Hitzler, P., and Hölldobler, S. (2006). The
integration of connectionism and first-order knowledge representation and
reasoning as a challenge for artificial intelligence. Information, 9(1):7–20.

[Bader et al., 2007] Bader, S., Hitzler, P., Hölldobler, S., and Witzel, A.
(2007). A fully connectionist model generator for covered first-order
logic programs. In Veloso, M., editor, Proceedings of the International
Joint Conference on Artificial Intelligence IJCAI07, pages 666–671, Hy­
derabad, India. AAAI Press, Menlo Park, CA.

[Bader et al., 2008] Bader, S., Hitzler, P., and Hölldobler, S. (2008). Con­
nectionist model generation: A first-order approach. Neurocomputing,
71(13–15):2420–2432.

[Baral, 2003] Baral, C. (2003). Knowledge Representation, Reasoning and

Declarative Problem Solving. Cambridge University Press, London; New
York.
 Bibliography 245

[Bartle, 1966] Bartle, R. G. (1966). The Elements of Integration. John Wiley

& Sons, New York.

[Batarekh, 1989] Batarekh, A. (1989). Topological Aspects of Logic Program­

ming. PhD thesis, Syracuse University, Syracuse, New York.

[Batarekh and Subrahmanian, 1989a] Batarekh, A. and Subrahmanian, V. S.

(1989a). The query topology in logic programming. In Monien, B. and
Cori, R., editors, Proceedings of the 1989 Symposium on Theoretical As­
pects of Computer Science, STACS 89, Paderborn, Germany, February,
1989, Lecture Notes in Computer Science, Volume 349, pages 375–387.
Springer, Berlin.

[Batarekh and Subrahmanian, 1989b] Batarekh, A. and Subrahmanian, V. S.

(1989b). Topological model set deformations in logic programming. Fun­
damenta Informaticae, 12(3):357–400.

[Belnap, 1977] Belnap, N. D. (1977). A useful four-valued logic. In Dunn,

J. M. and Epstein, G., editors, Modern Uses of Multiple-Valued Logic,
pages 5–37. Reidel, Dordrecht, The Netherlands.

[Berzati, 2007] Berzati, D. (2007). Non-Monotonic Reasoning: A Unified

Framework. Nova Science Publishers, New York.

[Bessaga, 1959] Bessaga, C. (1959). On the converse of the Banach fixed-point

principle. Colloquium Mathematicum, 7:41–43.

[Bezem, 1989] Bezem, M. (1989). Characterizing termination of logic pro­

grams with level mappings. In Lusk, E. L. and Overbeek, R. A., editors,
Proceedings of the North American Conference on Logic Programming,
pages 69–80. MIT Press, Cambridge, MA.

[Bidoit and Froideveaux, 1991] Bidoit, N. and Froideveaux, C. (1991). Nega­

tion by default and unstratifiable logic programs. Theoretical Computer
Science, 78:85–112.

[Bishop, 1995] Bishop, C. M. (1995). Neural Networks for Pattern Recogni­

tion. Oxford University Press, Oxford.

[Blair, 2007] Blair, H. A. (2007). Elementary differential calculus on dis­

crete, continuous and hybrid spaces. In Kopperman, R., Panan­
gaden, P., Smyth, M. B., and Spreen, D., editors, Computational Struc­
tures for Modelling Space, Time and Causality, August, 2006, Volume
06341 of Dagstuhl Seminar Proceedings. Internationales Begegnungs- und
Forschungszentrum fuer Informatik (IBFI), Schloss Dagstuhl, Germany.

[Blair and Remmel, 2001] Blair, H. A. and Remmel, J. (2001). Hybrid au­
tomata: Convergence spaces and continuity. In Proceedings of the Joint
IIIS & IEEE Meeting of the 5th World Multiconference on Systemics,
246 Bibliography

Cybernetics and Informatics, SCI2001, and the 7th International Confer­

ence on Information Systems Analysis and Synthesis, ISAS2001, Volume
XVII, Orlando, Florida, USA, pages 218–222. International Institute of
Informatics and Systemics: IIIS, Winter Garden, FL.

[Blair et al., 1997a] Blair, H. A., Chidella, J., Dushin, F., Ferry, A., and Hu­
menn, P. (1997a). A continuum of discrete systems. Annals of Mathe­
matics and Artificial Intelligence, 21(2–4):155–185.

[Blair et al., 1997b] Blair, H. A., Dushin, F., and Humenn, P. (1997b). Sim­
ulations between programs as cellular automata. In Dix, J., Furbach,
U., and Nerode, A., editors, Logic Programming and Non-Monotonic
Reasoning, Proceedings of the 4th International Conference LPNMR97,
Dagstuhl, Lecture Notes in Artificial Intelligence, Volume 1265, pages
115–131. Springer, Berlin.

[Blair et al., 1999] Blair, H. A., Dushin, F., Jakel, D. W., Rivera, A. J., and
Sezgin, M. (1999). Continuous models of computation for logic programs.
In Apt, K. R., Marek, V. W., Truszczyński, M., and Warren, D. S., ed­
itors, The Logic Programming Paradigm: A 25-Year Persepective, pages
231–255. Springer, Berlin.

[Blair et al., 2007] Blair, H. A., Jakel, D. W., Irwin, R. J., and Rivera, A.
(2007). Elementary differential calculus on discrete and hybrid struc­
tures. In Artëmov, S. N. and Nerode, A., editors, Logical Foundations
of Computer Science, International Symposium, LFCS 2007, New York,
NY, USA, June 4-7, 2007, Proceedings, Lecture Notes in Computer Sci­
ence, Volume 4514, pages 41–53. Springer, Berlin.

[Bochman, 1995] Bochman, A. (1995). Default consequence relations as a

logical framework for logic programs. In Marek, V. W. and Nerode,
A., editors, Proceedings of the Third International Conference on Logic
Programming and Non-Monotonic Reasoning (LPNMR’95), Lexington,
KY, USA, June 26-28, 1995, Lecture Notes in Computer Science, Volume
928, pages 245–258. Springer-Verlag, Berlin.

[Boley and Kifer, 2010] Boley, H. and Kifer, M., editors (22 June, 2010). RIF
Basic Logic Dialect. W3C Recommendation. Available from the website
http://www.w3.org/TR/rif-bld/.

[Bonnier et al., 1991] Bonnier, S., Nilsson, U., and Näslund, T. (1991). A
simple fixed-point characterization of the three-valued stable model se­
mantics. Information Processing Letters, 40(2):73–78.

[Bonsangue et al., 1996] Bonsangue, M. M., van Breugel, F., and Rutten, J. J.
(1996). Alexandroff and Scott topologies for generalized metric spaces.
In Andima, S., Flagg, R., Itzkowitz, G., et al., editors, Papers on Gen­
eral Topology and Applications: Eleventh Summer Conference on General
 Bibliography 247

Topology and Applications, University of Southern Maine, Maine, USA,

August, 1996, Annals of the New York Academy of Sciences, pages 49–68.
New York Academy of Sciences, New York.
[Bornscheuer, 1996] Bornscheuer, S.-E. (1996). Rational models of normal
logic programs. In KI-96: Advances in Artificial Intelligence, Lecture
Notes in Artificial Intelligence, Volume 1137, pages 1–4. Springer, Berlin.
[Bouamama et al., 2000] Bouamama, S., Misane, D., and Priess-Crampe, S.
(2000). An application of ultrametric spaces in logic programming. In
Proceedings of the Sixth Maghrebian Conference on Computer Sciences.
Maghrebian Information Processing Society, Fes, Moroco.
[Bramer, 2010] Bramer, M. (2010). Logic Programming with Prolog. Springer,
Berlin.
[Buillame-Bert et al., 2010] Buillame-Bert, M., Broda, K., and d’Avila
Garcez, A. (2010). First-order logic learning in artificial neural networks.
In Proceedings IJCNN 2010. IEEE. In press.
[Bukatin, 2002] Bukatin, M. A. (2002). Mathematics of Domains. PhD thesis,
Brandeis University, Waltham, MA.
[Bukatin and Scott, 1997] Bukatin, M. A. and Scott, J. S. (1997). Towards
computing distances between programs via Scott domains. In Adian,
S. and Nerode, A., editors, Logical Foundations of Computer Science,
4th International Symposium, LFCS’97, Yaroslavl, Russia, July, 1997,
Proceedings, Lecture Notes in Computer Science, Volume 1234, pages 33–
43. Springer-Verlag, Berlin.
[Castro-Company et al., 2007] Castro-Company, F., Romaguera, S., Sánchez­
´
Alvarez, J., and Tirado, P. (2007). A quasimetric lattice approach for
access prediction in replicated database protocols. Technical report, In­
stituto de Matemática Pura y Aplicada, Universidad Politécnica de Va­
lencia, Spain.
[Cavedon, 1989] Cavedon, L. (1989). Continuity, consistency, and complete­
ness properties for logic programs. In Levi, G. and Martelli, M., editors,
Proceedings of the 6th International Conference on Logic Programming,
pages 571–584. MIT Press, Cambridge, MA.
[Cavedon, 1991] Cavedon, L. (1991). Acyclic programs and the completeness
of SLDNF-resolution. Theoretical Computer Science, 86:81–92.
[Cherchago et al., 2007] Cherchago, N., Hitzler, P., and Hölldobler, S. (2007).
Decidability under the well-founded semantics. In Marchiori, M., Pan,
J. Z., and de Sainte Marie, C., editors, Web Reasoning and Rule Systems,
First International Conference, RR 2007, Innsbruck , Austria, June 7­
8, 2007, Proceedings, Lecture Notes in Computer Science, Volume 4524,
pages 269–278. Springer, Berlin.
248 Bibliography

[Clark, 1978] Clark, K. L. (1978). Negation as failure. In Gallaire, H. and

Minker, J., editors, Logic and Data Bases, pages 293–322. Plenum Press,
New York.

[Clifford and Seda, 2000] Clifford, E. and Seda, A. K. (2000). Uniqueness of

the fixed points of single-step operators in many-valued logics. Journal
of Electrical Engineering, Slovak Academy of Sciences, 51(12/s):54–58.

[Colmerauer and Roussel, 1993] Colmerauer, A. and Roussel, P. (1993). The

birth of Prolog. In ACM SIGPLAN Notices, Volume 28(3), pages 37–52.
ACM Press, New York.

[Crazzolara, 1997] Crazzolara, F. (1997). Quasimetric spaces as domains for

abstract interpretation. In Falaschi, M., Navarro, M., and Policriti, A.,
editors, Proceedings of the 1997 Joint Conference on Declarative Pro­
gramming, APPIA-GULP-PRODE’97, Grado, Italy, June, 1997, pages
45–56. University of Udine, Italy.

[Cybenko, 1989] Cybenko, G. (1989). Approximation by superpositions of a

sigmoidal function. Mathematics of Control, Signals, and Systems, 2:303–
314.

[Davey and Priestley, 2002] Davey, B. A. and Priestley, H. A. (2002). Intro­

duction to Lattices and Order. Cambridge University Press, Cambridge,
UK, second edition.

[d’Avila Garcez and Lamb, 2006] d’Avila Garcez, A. S. and Lamb, L. C.

(2006). A connectionist computational model for epistemic and temporal
reasoning. Neural Computation, 18(7):1711–1738.

[d’Avila Garcez and Zaverucha, 1999] d’Avila Garcez, A. S. and Zaverucha,

G. (1999). The connectionist inductive learning and logic programming
system. Applied Intelligence, Special Issue on Neural Networks and Struc­
tured Knowledge, 11(1):59–77.

[d’Avila Garcez et al., 1997] d’Avila Garcez, A. S., Zaverucha, G., and
de Carvalho, L. A. (1997). Logical inference and inductive learning in
artificial neural networks. In Hermann, C., Reine, F., and Strohmaier,
A., editors, Knowledge Representation in Neural Networks, pages 33–46.
Logos Verlag, Berlin.

[d’Avila Garcez et al., 2002] d’Avila Garcez, A. S., Broda, K. B., and Gabbay,
D. M. (2002). Neural-Symbolic Learning Systems — Foundations and
Applications. Perspectives in Neural Computing. Springer, Berlin.

[d’Avila Garcez et al., 2006] d’Avila Garcez, A. S., Lamb, L. C., and Gabbay,
D. M. (2006). Connectionist computations of intuitionistic reasoning.
Theoretical Computer Science, 358(1):34–55.
 Bibliography 249

[d’Avila Garcez et al., 2007] d’Avila Garcez, A. S., Lamb, L. C., and Gabbay,
D. M. (2007). Connectionist modal logic: Representing modalities in
neural networks. Theoretical Computer Science, 371(1):34–53.
[d’Avila Garcez et al., 2009] d’Avila Garcez, A. S., Lamb, L. C., and Gabbay,
D. M. (2009). Neural-Symbolic Cognitive Reasoning. Cognitive Technolo­
gies. Springer, Berlin.
[de Bakker, 2002] de Bakker, J. W. (2002). Fixed points in metric semantics.
Electronic Notes in Theoretical Computer Science, 40:70–71.
[de Bakker and de Vink, 1996] de Bakker, J. W. and de Vink, E. P. (1996).
Control Flow Semantics. Foundations of Computing Series. MIT Press,
Cambridge, MA.
[De Raedt et al., 2008] De Raedt, L., Frasconi, P., Kersting, K., and Mug­
gleton, S., editors (2008). Probabilistic Inductive Logic Programming:
Theory and Applications. Springer-Verlag, Berlin.
[Dowling and Gallier, 1984] Dowling, W. F. and Gallier, J. H. (1984). Linear-
time algorithms for testing the satisfiability of propositional Horn formu­
lae. The Journal of Logic Programming, 1(3):267–284.
[Dragovich and Dragovich, 2006] Dragovich, B. and Dragovich, A. (2006). A
p-adic model of DNA sequence and genetic code. Preprint is available at
http://www.arxiv.org/abs/q-bio.GN/0607018.
[Dugundji and Granas, 1982] Dugundji, J. and Granas, A. (1982). Fixed
Point Theory. Monografie Matematyczne. Polish Scientific Publishers,
Warsaw.
[Dung and Kanchanasut, 1989] Dung, P. M. and Kanchanasut, K. (1989). A
fixpoint approach to declarative semantics of logic programs. In Lusk,
E. L. and Overbeek, R. A., editors, Logic Programming, Proceedings of
the North American Conference 1989, NACLP’89, Cleveland, Ohio, pages
604–625. MIT Press, Cambridge, MA.
[Edalat and Heckmann, 1998] Edalat, A. and Heckmann, R. (1998). A com­
putational model for metric spaces. Theoretical Computer Science,
193:53–73.
[Fages, 1994] Fages, F. (1994). Consistency of Clark’s completion and the
existence of stable models. Journal of Methods of Logic in Computer
Science, 1:51–60.
[Ferry, 1994] Ferry, A. P. (1994). Topological Characterizations for Logic Pro­
gramming Semantics. PhD thesis, University of Michigan, MI.
[Fitting, 1985] Fitting, M. C. (1985). A Kripke-Kleene semantics for general
logic programs. The Journal of Logic Programming, 2:295–312.
250 Bibliography

[Fitting, 1991] Fitting, M. C. (1991). Bilattices and the semantics of logic

programming. The Journal of Logic Programming, 11:91–116.
[Fitting, 1994a] Fitting, M. C. (1994a). Kleene’s three-valued logics and their
children. Fundamenta Informaticae, 20:113–131.
[Fitting, 1994b] Fitting, M. C. (1994b). Metric methods: Three examples and
a theorem. The Journal of Logic Programming, 21(3):113–127.
[Fitting, 2002] Fitting, M. C. (2002). Fixpoint semantics for logic program­
ming — A survey. Theoretical Computer Science, 278(1–2):25–51.
[Fitting and Ben-Jacob, 1990] Fitting, M. C. and Ben-Jacob, M. (1990).
Stratified, weak stratified, and three-valued semantics. Fundamenta In­
formaticae, 13:19–33.
[Flagg and Kopperman, 1997] Flagg, B. and Kopperman, R. (1997). Continu­
ity spaces: Reconciling domains and metric spaces. Theoretical Computer
Science, 177(1):111–138.
[Fritzke, 1998] Fritzke, B. (1998). Vektorbasierte Neuronale Netze. Habilita­
tion, Technische Universität Dresden, Germany.
[Fu, 1994] Fu, L. (1994). Neural Networks in Computer Intelligence. McGraw-
Hill, New York.
[Funahashi, 1989] Funahashi, K.-I. (1989). On the approximate realization of
continuous mappings by neural networks. Neural Networks, 2:183–192.
[Gabbay et al., 1994] Gabbay, D. M., Hogger, C. J., and Robinson, J. A.
(1994). Non-Monotonic Reasoning and Uncertain Reasoning, Volume 3
of Handbook of Logic in Artificial Intelligence and Logic Programming.
Clarendon Press, Oxford University Press, Oxford, UK.
[Ganter and Wille, 1999] Ganter, B. and Wille, R. (1999). Formal Concept
Analysis: Mathematical Foundations. Springer, Berlin.
[Gelfond and Lifschitz, 1988] Gelfond, M. and Lifschitz, V. (1988). The sta­
ble model semantics for logic programming. In Kowalski, R. A. and
Bowen, K. A., editors, Logic Programming. Proceedings of the 5th Inter­
national Conference and Symposium on Logic Programming, pages 1070–
1080. MIT Press, Cambridge, MA.
[Gelfond and Lifschitz, 1991] Gelfond, M. and Lifschitz, V. (1991). Classical
negation in logic programs and disjunctive databases. New Generation
Computing, 9:365–385.
[Gierz et al., 2003] Gierz, G., Hofmann, K. H., Keimel, K., Lawson, J. D.,
Mislove, M., and Scott, D. S. (2003). Continuous Lattices and Domains,
Volume 93 of Encyclopedia of Mathematics and its Applications. Cam­
bridge University Press, Cambridge, UK.
 Bibliography 251

[Gust et al., 2007] Gust, H., Kühnberger, K.-U., and Geibel, P. (2007). Learn­
ing models of predicate logical theories with neural networks based on
topos theory. In Hammer, B. and Hitzler, P., editors, Perspectives of
Neural-Symbolic Integration, Volume 77 of Studies in Computational In­
telligence, pages 233–264. Springer, Berlin.
[Halmos, 1998] Halmos, P. R. (1998). Naive Set Theory. Undergraduate Texts
in Mathematics. Springer-Verlag, Berlin.
[Hammer and Hitzler, 2007] Hammer, B. and Hitzler, P., editors (2007). Per­
spectives of Neural-Symbolic Integration, Volume 77 of Studies in Com­
putational Intelligence. Springer, Berlin.
[Haykin, 1994] Haykin, S. (1994). Neural Networks. A Comprehensive Foun­
dation. Macmillan College Publishing Company, New York.
[Heckmann, 1999] Heckmann, R. (1999). Approximation of metric spaces by
partial metric spaces. Applied Categorical Structures, 7:71–83.
[Heckmanns, 1996] Heckmanns, U. (1996). On the topology of ultrametric
spaces. In Simon, P., editor, Proceedings of the 8th Prague Topological
Symposium, August 1996, Prague, pages 149–156. Charles University and
Topology Atlas, Prague.
[Heinze, 2003] Heinze, R. (2003). Topological investigations of the opera­
tors of the well-founded and alternating fixed-point semantics of nor­
mal logic programs. In Hurley, T., MacanAirchinnigh, M., Schellekens,
M., and Seda, A. K., editors, Proceedings of the Second Irish Conference
on the Mathematical Foundations of Computer Science and Information
Technology (MFCSIT2002), July, 2002, Galway, Ireland, Volume 74 of
Electronic Notes in Theoretical Computer Science, pages 51–68. Elsevier
Science Publishers, Amsterdam; New York.
[Hertz et al., 1991] Hertz, J., Krogh, A., and Palmer, R. G. (1991). Intro­
duction to the Theory of Neural Computation. Addison-Wesley, Reading,
MA.
[Hitzler, 2001] Hitzler, P. (2001). Generalized Metrics and Topology in Logic
Programming Semantics. PhD thesis, Department of Mathematics, Na­
tional University of Ireland, University College Cork, Cork, Ireland.
[Hitzler, 2003a] Hitzler, P. (2003a). A resolution theorem for algebraic do­
mains. In Gottlob, G. and Walsh, T., editors, Proceedings of the 18th
International Joint Conference on Artificial Intelligence, Acapulco, Mex­
ico, August, 2003, pages 1339–1340. Morgan Kaufmann Publishers, San
Mateo, CA.
[Hitzler, 2003b] Hitzler, P. (2003b). Towards a systematic account of different
logic programming semantics. In Günter, A., Kruse, R., and Neumann,
252 Bibliography

B., editors, Proceedings of the 26th German Conference on Artificial In­

telligence, KI2003, Hamburg, Germany, September, 2003, Lecture Notes
in Artificial Intelligence, Volume 2821, pages 105–119. Springer, Berlin.

[Hitzler, 2004] Hitzler, P. (2004). Default reasoning over domains and con­
cept hierarchies. In Biundo, S., Frühwirth, T., and Palm, G., editors,
KI2004: Advances in Artificial Intelligence. Proceedings of the 27th An­
nual German Conference on AI, KI2004, Ulm, Germany, September,
2004, Lecture Notes in Artificial Intelligence, Volume 3238, pages 351–
365. Springer-Verlag, Berlin.

[Hitzler, 2005] Hitzler, P. (2005). Towards a systematic account of differ­

ent semantics for logic programs. Journal of Logic and Computation,
15(3):391–404.

[Hitzler, 2009] Hitzler, P. (2009). Towards reasoning pragmatics. In Janowicz,

K., Raubal, M., and Levashkin, S., editors, GeoSpatial Semantics, Third
International Conference, GeoS 2009, Mexico City, Mexico, December 3­
4, 2009. Proceedings, Lecture Notes in Computer Science, Volume 5892,
pages 9–25. Springer, Berlin.

[Hitzler and Krötzsch, 2006] Hitzler, P. and Krötzsch, M. (2006). Querying

formal contexts with answer set programs. In Schärfe, H., Hitzler, P.,
and Øhrstrøm, P., editors, Conceptual Structures: Inspiration and Appli­
cation, Proceedings of the 14th International Conference on Conceptual
Structures, ICCS 2006, Aalborg, Denmark, July, 2006, Lecture Notes in
Artificial Intelligence, Volume 4068, pages 413–426. Springer, Berlin.

[Hitzler and Kühnberger, 2009] Hitzler, P. and Kühnberger, K.-U. (2009).

The importance of being neural-symbolic — A wilde position. In Go­
ertzel, B., Hitzler, P., and Hutter, M., editors, Artificial General Intelli­
gence. Second Conference on Artificial General Intelligence, AGI 2009,
Arlington, Virginia, USA, March 6-9, 2009. Proceedings, pages 208–209.
Atlantis Press, Amsterdam; Paris.

[Hitzler and Seda, 1999a] Hitzler, P. and Seda, A. K. (1999a). Acceptable pro­
grams revisited. In Etalle, S. and Smaus, J.-G., editors, Proceedings of
the Workshop on Verification in Logic Programming, 16th International
Conference on Logic Programming, ICLP’99, Las Cruces, New Mexico,
November, 1999, Electronic Notes in Theoretical Computer Science, Vol­
ume 30, pages 59–76. Elsevier Science Publishers, Amsterdam; New York.

[Hitzler and Seda, 1999b] Hitzler, P. and Seda, A. K. (1999b). Characteri­

zations of classes of programs by three-valued operators. In Gelfond,
M., Leone, N., and Pfeifer, G., editors, Logic Programming and Non-
Monotonic Reasoning, Proceedings of the 5th International Conference
on Logic Programming and Non-Monotonic Reasoning (LPNMR’99), El
 Bibliography 253

Paso, Texas, USA, December, 1999, Lecture Notes in Artificial Intelli­

gence, Volume 1730, pages 357–371. Springer, Berlin.

[Hitzler and Seda, 1999c] Hitzler, P. and Seda, A. K. (1999c). Some issues
concerning fixed points in computational logic: Quasimetrics, multival­
ued mappings and the Knaster-Tarski theorem. In Comfort, W., Heck­
mann, R., Kopperman, R., and Narici, L., editors, Proceedings of the 14th
Summer Conference on Topology and Its Applications: Special Session on
Topology in Computer Science, Long Island, USA, Volume 24 of Topology
Proceedings, pages 223–250.

[Hitzler and Seda, 2000] Hitzler, P. and Seda, A. K. (2000). Dislocated

topologies. Journal of Electrical Engineering, Slovak Academy of Sci­
ences, 51(12/s):3–7.

[Hitzler and Seda, 2001] Hitzler, P. and Seda, A. K. (2001). A “converse” of

the Banach contraction mapping theorem. Journal of Electrical Engi­
neering, Slovak Academy of Sciences, 52(10/s):3–6.

[Hitzler and Seda, 2002a] Hitzler, P. and Seda, A. K. (2002a). The fixed-
point theorems of Priess-Crampe and Ribenboim in logic programming.
In Kuhlmann, F.-V., Kuhlmann, S., and Marshall, M., editors, Valuation
Theory and Its Applications, Proceedings of the International Valuation
Theory Conference, University of Saskatchewan in Saskatoon, Canada,
July, 1999, Volume 32 of Fields Institute Communications Series, pages
219–235. American Mathematical Society, Providence, RI.

[Hitzler and Seda, 2002b] Hitzler, P. and Seda, A. K. (2002b). On the coin­
cidence of semantics for uniquely determined programs. In Hurley, T.,
MacanAirchinnigh, M., Schellekens, M., and Seda, A. K., editors, Pro­
ceedings of the First Irish Conference on the Mathematical Foundations
of Computer Science and Information Technology (MFCSIT2000), Cork,
Ireland, July, 2000, Electronic Notes in Theoretical Computer Science,
Volume 40, pages 189–205. Elsevier Science Publishers, Amsterdam; New
York.

[Hitzler and Seda, 2003] Hitzler, P. and Seda, A. K. (2003). Generalized met­
rics and uniquely determined logic programs. Theoretical Computer Sci­
ence, 305(1–3):187–219.

[Hitzler and van Harmelen, 2010] Hitzler, P. and van Harmelen, F. (2010). A
reasonable Semantic Web. Semantic Web – Interoperability, Usability,
Applicability. In press.

[Hitzler and Wendt, 2002] Hitzler, P. and Wendt, M. (2002). The well-
founded semantics is a stratified Fitting semantics. In Jarke, M., Koehler,
J., and Lakemeyer, G., editors, Proceedings of the 25th Annual Ger­
man Conference on Artificial Intelligence, KI2002, Aachen, Germany,
254 Bibliography

September 2002, Lecture Notes in Artificial Intelligence, Volume 2479,

pages 205–221. Springer, Berlin.
[Hitzler and Wendt, 2003] Hitzler, P. and Wendt, M. (2003). Formal concept
analysis and resolution in algebraic domains. In de Moor, A. and Ganter,
B., editors, Using Conceptual Structures — Contributions to ICCS 2003,
pages 157–170. Shaker Verlag, Aachen.
[Hitzler and Wendt, 2005] Hitzler, P. and Wendt, M. (2005). A uniform ap­
proach to logic programming semantics. Theory and Practice of Logic
Programming, 5(1-2):123–159.
[Hitzler et al., 2004] Hitzler, P., Hölldobler, S., and Seda, A. K. (2004). Logic
programs and connectionist networks. Journal of Applied Logic, 2(3):245–
272.
[Hitzler et al., 2005] Hitzler, P., Bader, S., and d’Avila Garcez, A. (2005).
Ontology learning as a use case for neural-symbolic integration. In
d’Avila Garcez, A. S., Hitzler, P., and Ellman, J., editors, Proceedings
of the IJCAI-05 Workshop on Neural-Symbolic Learning and Reasoning,
NeSy’05.
[Hitzler et al., 2009a] Hitzler, P., Krötzsch, M., Parsia, B., Patel-Schneider,
P., and Rudolph, S., editors (2009a). OWL 2 Web Ontology Language:
Primer. W3C Recommendation, 27 October 2009. Preprint can be ob­
tained from the website http://www.w3.org/TR/owl-primer.
[Hitzler et al., 2009b] Hitzler, P., Krötzsch, M., and Rudolph, S. (2009b).
Foundations of Semantic Web Technologies. Chapman & Hall/CRC,
Boca Raton, FL.
[Hodel, 1995] Hodel, R. E. (1995). Introduction to Mathematical Logic. Thom­
son Learning PWS, New York.
[Hölldobler and Kalinke, 1994] Hölldobler, S. and Kalinke, Y. (1994). To­
wards a massively parallel computational model for logic programming.
In Proceedings ECAI94 Workshop on Combining Symbolic and Connec­
tionist Processing, Amsterdam, July, 1994, pages 68–77. European Co­
ordinating Committee for Artificial Intelligence (ECCAI).
[Hölldobler and Ramli, 2009] Hölldobler, S. and Ramli, C. D. P. K. (2009).
Logics and networks for human reasoning. In Alippi, C., Polycarpou,
M. M., Panayiotou, C. G., and Ellinas, G., editors, Artificial Neural Net­
works (ICANN 2009), 19th International Conference, Limassol, Cyprus,
September 14-17, 2009, Proceedings, Lecture Notes in Computer Science,
Volume 5769, pages 85–94. Springer, Berlin.
[Hölldobler et al., 1999] Hölldobler, S., Kalinke, Y., and Störr, H.-P. (1999).
Approximating the semantics of logic programs by recurrent neural net­
works. Applied Intelligence, 11:45–58.
 Bibliography 255

[Hornik et al., 1989] Hornik, K., Stinchcombe, M., and White, H. (1989).
Multilayer feedforward networks are universal approximators. Neural
Networks, 2:359–366.
[Istrăţescu, 1981] Istrăţescu, V. I. (1981). Fixed Point Theory — An Introduc­
tion, Volume 7, Mathematics and Its Applications. D. Reidel Publishing
Company, Dordrecht, Holland.
[Jachymski, 2000] Jachymski, J. (2000). A short proof of the converse of the
contraction principle and some related results. Topological Methods in
Nonlinear Analysis, 15:179–186.
[Jachymski, 2001] Jachymski, J. (2001). Order-theoretic aspects of metric
fixed-point theory. In Kirk, W. A. and Sims, B., editors, Handbook of
Metric Fixed Point Theory, pages 613–641. Kluwer Academic Publishers,
Dordrecht, The Netherlands.
[Jacobsson, 2005] Jacobsson, H. (2005). Rule extraction from recurrent neural
networks: A taxonomy and review. Neural Computation, 17(6):1223–
1263.
[Jones and Laaser, 1977] Jones, N. D. and Laaser, W. T. (1977). Complete
problems for deterministic sequential time. Theoretical Computer Sci­
ence, 3:105–117.
[Kalinke, 1994] Kalinke, Y. (1994). Ein massiv paralleles Berechnungsmodell
für normale logische Programme. PhD thesis, Department of Computer
Science, Dresden University of Technology, Dresden, Germany.
[Karp and Ramachandran, 1990] Karp, R. M. and Ramachandran, V. (1990).
Parallel algorithms for shared-memory machines. In van Leeuwen, J.,
editor, Handbook of Theoretical Computer Science, Chapter 17, pages
869–941. Elsevier Science Publishers, New York.
[Kelley, 1975] Kelley, J. L. (1975). General Topology. Graduate Texts in
Mathematics, Volume 27. Springer-Verlag, Berlin.
[Khamsi and Misane, 1998] Khamsi, M. A. and Misane, D. (1997/1998). Dis­
junctive signed logic programs. Fundamenta Informaticae, 32(3–4):349–
357.
[Khamsi et al., 1993] Khamsi, M. A., Kreinovich, V., and Misane, D. (1993).
A new method of proving the existence of answer sets for disjunctive logic
programs: A metric fixed-point theorem for multivalued mappings. In
Baral, C. and Gelfond, M., editors, Proceedings of the Workshop on Logic
Programming with Incomplete Information, Vancouver, B.C., Canada,
pages 58–73. University of British Columbia, Vancouver, BC.
[Khrennikov, 1998] Khrennikov, A. Y. (1998). Human subconscious as the
p-adic dynamical system. Journal of Theoretical Biology, 193:179–196.
256 Bibliography

[Khrennikov, 2004] Khrennikov, A. Y. (2004). Information Dynamics in Cog­

nitive, Psychological, Social and Anomalous Phenomena, Volume 138 of
Fundamental Theories of Physics. Kluwer Academic Publishers, Dor­
drecht, Boston, London.
[Khrennikov and Kozyrev, 2007] Khrennikov, A. Y. and Kozyrev, S. V.
(2007). Genetic codenext term on the diadic plane. Physica A: Statistical
Mechanics and Its Applications, 381:265–272.
[Kifer et al., 1995] Kifer, M., Lausen, G., and Wu, J. (1995). Logical founda­
tions of object-oriented and frame-based languages. Journal of the ACM,
42(4):741–843.
[Kirk and Sims, 2001] Kirk, W. A. and Sims, B., editors (2001). Handbook
of Metric Fixed Point Theory. Kluwer Academic Publishers, Dordrecht,
The Netherlands.
[Klavins et al., 1998] Klavins, E., Rounds, W. C., and Zhang, G.-Q. (1998).
Experimenting with power default reasoning. In Proceedings of the Fif­
teenth National Conference on Artificial Intelligence and Tenth Innova­
tive Applications of Artificial Intelligence Conference, AAAI 98, IAAI 98,
July, 1998, Madison, Wisconsin, USA, pages 846–852. AAAI Press/MIT
Press.
[Knorr and Hitzler, 2007] Knorr, M. and Hitzler, P. (2007). A comparison
of disjunctive well-founded semantics. In Hitzler, P., Roth-Berghofer,
T., and Rudolph, S., editors, Foundations of Artificial Intelligence FAInt
2007, Osnabrück, Germany, September 10, 2007, Volume 277, CEUR
Workshop Proceedings.
[Kohonen, 1981] Kohonen, T. (1981). Self-organized formation of topologi­
cally correct feature maps. Biological Cybernetics, 43(1):59–69.
[Komendantskaya, 2010] Komendantskaya, E. (2010). Unification neural net­
works: Unification by error-correction learning. Logic Journal of the
IGPL. In press.
[Komendantskaya et al., 2007] Komendantskaya, E., Lane, M., and Seda,
A. K. (2007). Connectionist representation of multivalued logic programs.
In Hammer, B. and Hitzler, P., editors, Perspectives of Neural-Symbolic
Integration, Volume 77 of Studies in Computational Intelligence, pages
283–313. Springer, Berlin.
[Kopperman, 1988] Kopperman, R. (1988). All topologies come from gener­
alized metrics. American Mathematical Monthly, 95(2):89–97.
[Kowalski, 1974] Kowalski, R. A. (1974). Predicate logic as a programming
language. In Rosenfeld, J. L., editor, Proceedings IFIP’74, Stockholm,
Sweden, August, 1974, pages 569–574. North-Holland, Amsterdam.
 Bibliography 257

[Krötzsch, 2006] Krötzsch, M. (2006). Generalized ultrametric spaces in quan­

titative domain theory. Theoretical Computer Science, 368(1–2):30–49.
[Krötzsch et al., 2008] Krötzsch, M., Rudolph, S., and Hitzler, P. (2008).
ELP: Tractable rules for OWL 2. In Sheth, A., Staab, S., Dean, M.,
Paolucci, M., Maynard, D., Finin, T., and Thirunarayan, K., editors,
Proceedings of the 7th International Semantic Web Conference (ISWC­
08), Lecture Notes in Computer Science, Volume 5318, pages 649–664.
Springer, Berlin.
[Kuhlmann, 1999] Kuhlmann, F.-V. (1999). A theorem about maps on spher­
ically complete ultrametric spaces, and its applications. Preprint, De­
partment of Mathematics and Statistics, University of Saskatchewan in
Saskatoon, 20 pages.
[Künzi, 2001] Künzi, H.-P. A. (2001). Non-symmetric distances and their
associated topologies: About the origins of basic ideas in the area of
asymmetric topology. In Aull, C. E. and Lowen, R., editors, Handbook
of the History of General Topology, Volume 3, pages 853–968. Kluwer
Academic Publishers, Dordrecht.
[Künzi and Kivuvu, 2008] Künzi, H.-P. A. and Kivuvu, C. M. (2008). A dou­
ble completion for an arbitrary T0 -quasimetric space. Journal of Logic
and Algebraic Programming, 76(2):251–269.
[Künzi et al., 2006] Künzi, H.-P. A., Pajoohesh, H., and Schellekens, M.
(2006). Partial quasimetrics. Theoretical Computer Science, 365:237–
246.
[Lane and Seda, 2006] Lane, M. and Seda, A. K. (2006). Some aspects of
the integration of connectionist and logic-based systems. Information,
9(4):551–562.
[Lane and Seda, 2009] Lane, M. and Seda, A. K. (2009). Properties of gen­
eral semantic operators determined by logic-based systems. In Hurley,
T., MacanAirchinnigh, M., Schellekens, M., Seda, A. K., and Strong, G.,
editors, Proceedings of the Fourth Irish Conference on the Mathematical
Foundations of Computer Science and Information Technology (MFC­
SIT2006), University College Cork, August, 2006, Electronic Notes in
Theoretical Computer Science, Volume 225, pages 181–194. Elsevier Sci­
ence Publishers, Amsterdam; New York.
[Lassez et al., 1982] Lassez, J.-L., Nguyen, V., and Sonenberg, L. (1982).
Fixed-point theorems and semantics: A folk tale. Information Processing
Letters, 14(3):112–116.
[Lehmann et al., 2010] Lehmann, J., Bader, S., and Hitzler, P. (2010). Ex­
tracting reduced logic programs from artificial neural networks. Applied
Intelligence, 32(3):249–266.
258 Bibliography

[Leone et al., 2006] Leone, N., Pfeifer, G., Faber, W., Eiter, T., Gottlob, G.,
Perri, S., and Scarcello, F. (2006). The dlv system for knowledge rep­
resentation and reasoning. ACM Transactions on Computational Logic,
7(3):499–562.

[Lifschitz, 1999] Lifschitz, V. (1999). Answer set planning. In Schreye, D. D.,

editor, Logic Programming. Proceedings of the 16th International Confer­
ence on Logic Programming, ICLP’99, Las Cruces, New Mexico, Novem­
ber, 1999, pages 23–37. MIT Press, Cambridge, MA.

[Lifschitz, 2001] Lifschitz, V. (2001). Success of default logic. In Hayes, P.,

Sandewall, E., Amarel, S., et al., editors, Logical Foundations for Cog­
nitive Agents: Contributions in Honor of Ray Reiter, pages 208–212.
Springer-Verlag, New York, Secaucus, NJ.

[Lifschitz et al., 1995] Lifschitz, V., McCain, N., Przymusinski, T. C., and
Stärk, R. F. (1995). Loop checking and the well-founded semantics. In
Marek, V. W. and Nerode, A., editors, Logic Programming and Non-
Monotonic Reasoning, Proceedings of the 3rd International Conference,
LPNMR’95, Lexington, KY, USA, June 1995, Lecture Notes in Computer
Science, Volume 928, pages 127–142. Springer, Berlin.

[Lloyd, 1987] Lloyd, J. W. (1987). Foundations of Logic Programming, Second

Edition. Springer, Berlin.

[Makinson, 2005] Makinson, D. (2005). Bridges from Classical to Non-

Monotonic Logic. King’s College Publications, London.

[Marchiori, 1996] Marchiori, E. (1996). On termination of general logic pro­

grams with respect to constructive negation. The Journal of Logic Pro­
gramming, 26(1):69–89.

[Marek and Truszczyński, 1999] Marek, V. W. and Truszczyński, M. (1999).

Stable models and an alternative logic programming paradigm. In Apt,
K. R., Marek, V. W., Truszczyński, M., and Warren, D. S., editors, The
Logic Programming Paradigm: A 25-Year Perspective, pages 375–398.
Springer, Berlin.

[Markowsky, 1976] Markowsky, G. (1976). Chain-complete posets and di­

rected sets with applications. Algebra Universalis, 6:53–68.

[Martin, 2000] Martin, K. (2000). A Foundation for Computation. PhD the­

sis, Department of Mathematics, Tulane University, New Orleans, LA.

[Martinetz and Schulten, 1991] Martinetz, T. and Schulten, K. (1991). A

“Neural-Gas” network learns topologies. Artificial Neural Networks,
I:397–402.
 Bibliography 259

[Matthews, 1986] Matthews, S. G. (1986). Metric domains for completeness.

Technical Report 76, Department of Computer Science, University of
Warwick, Coventry, UK. Ph.D. Thesis, 1985.

[Matthews, 1992] Matthews, S. G. (1992). The cycle contraction mapping

theorem. Technical Report 228, Department of Computer Science, Uni­
versity of Warwick, Coventry, UK.

[Matthews, 1994] Matthews, S. G. (1994). Partial metric topology. In

Andima, S., Itzkowitz, G., Kong, Y., et al., editors, Proceedings of the
Eighth Summer Conference on General Topology and Its Applications,
Queens College, CUNY, New York, USA, June, 1992, Annals of the
New York Academy of Sciences, Volume 728, pages 183–197. New York
Academy of Sciences, New York.

[McCarthy, 1977] McCarthy, J. (1977). Epistemological problems of artifi­

cial intelligence. In Proceedings of the International Joint Conference on
Artificial Intelligence (IJCAI-77), pages 1038–1044.

[McCarthy, 1980] McCarthy, J. (1980). Circumscription — A form of non-

monotonic reasoning. Artificial Intelligence, 13(1):27–39.

[Mendelson, 1987] Mendelson, E. (1987). Introduction to Mathematical Logic.

Wadsworth & Brooks/Cole Advanced Books and Software, Monterey,
CA.

[Moore, 1984] Moore, R. (1984). Possible-worlds semantics for autoepistemic

logic. In Proceedings of the 1984 Non-Monotonic Reasoning Workshop.
AAAI, Menlo Park, CA.

[Moore, 1985] Moore, R. (1985). Semantical considerations on non-monotonic

logic. Artificial Intelligence, 25(1):75–94.

[Murtagh, 2004] Murtagh, F. (2004). On ultrametricity, data coding, and

computation. Journal of Classification, 21:167–184.

[Murtagh, 2005] Murtagh, F. (2005). Identifying the ultrametricity of time

series. European Physical Journal, 43(4):573–579.

[O’Neill, 1996] O’Neill, S. J. (1996). Partial metrics, valuations, and do­

main theory. In Andima, S., Flagg, R., Itzkowitz, G., et al., editors,
Papers on General Topology and Applications: Eleventh Summer Con­
ference on General Topology and Applications, University of Southern
Maine, Maine, August, 1996, Annals of the New York Academy of Sci­
ences, pages 304–315. New York Academy of Sciences, New York.

[Pedreschi et al., 2002] Pedreschi, D., Ruggieri, S., and Smaus, J.-G. (2002).
Classes of terminating logic programs. Theory and Practice of Logic
Programming, 2(3):369–418.
260 Bibliography

[Plotkin, 1983] Plotkin, G. D. (1983). Domains. Department of Computer

Science, University of Edinburgh, Scotland.
[Prieß-Crampe, 1990] Prieß-Crampe, S. (1990). Der Banachsche Fixpunktsatz
für ultrametrische Räume. Results in Mathematics, 18:178–186.
[Prieß-Crampe and Ribenboim, 1993] Prieß-Crampe, S. and Ribenboim, P.
(1993). Fixed points, combs and generalized power series. Abh. Math.
Sem. Univ. Hamburg, 63:227–244.
[Prieß-Crampe and Ribenboim, 2000a] Prieß-Crampe, S. and Ribenboim, P.
(2000a). Fixed-point and attractor theorems for ultrametric spaces. Fo­
rum Mathematicum., 12:53–64.
[Prieß-Crampe and Ribenboim, 2000b] Prieß-Crampe, S. and Ribenboim, P.
(2000b). Logic programming and ultrametric spaces. Rendiconti di Math­
ematica, VII:1–13.
[Prieß-Crampe and Ribenboim, 2000c] Prieß-Crampe, S. and Ribenboim, P.
(2000c). Ultrametric spaces and logic programming. The Journal of Logic
Programming, 42:59–70.
[Przymusinska and Przymusinski, 1990] Przymusinska, H. and Przymusinski,
T. C. (1990). Weakly stratified logic programs. Fundamenta Informati­
cae, 13:51–65.
[Przymusinski, 1988] Przymusinski, T. C. (1988). On the declarative seman­
tics of deductive databases and logic programs. In Minker, J., editor,
Foundations of Deductive Databases and Logic Programming, pages 193–
216. Morgan Kaufmann Publishers, Los Altos, CA.
[Reed et al., 1991] Reed, G. M., Roscoe, A. W., and Wachter, R. F., editors
(1991). Topology and Category Theory in Computer Science. Oxford
University Press, Oxford, UK.
[Reiter, 1980] Reiter, R. (1980). A logic for default reasoning. Artificial In­
telligence, 13:81–132.
[Reynolds, 2010] Reynolds, D., editor (2010). OWL 2 RL in RIF. A W3C
Working Group Note. See http://www.w3.org/TR/rif-owl-rl/.
[Ribenboim, 1996] Ribenboim, P. (1996). The new theory of ultrametric
spaces. Periodica Mathematica Hungarica, 32(1–2):103–111.
[Robinson, 1965] Robinson, J. A. (1965). A machine-oriented logic based on
the resolution principle. Journal of the ACM, 12(1):23–41.
[Rodrı́guez-López et al., 2008] Rodrı́guez-López, J., Romaguera, S., and
Valero, O. (2008). Denotational semantics for programming languages,
balanced quasimetrics and fixed points. International Journal of Com­
puter Mathematics, 85(3):623–630.
 Bibliography 261

[Rojas, 1996] Rojas, R. (1996). Neural Networks. Springer, Berlin.

[Romaguera and Schellekens, 2003] Romaguera, S. and Schellekens, M.
(2003). Norm-weightable Riesz spaces and the dual complexity space.
In Hurley, T., MacanAirchinnigh, M., Schellekens, M., and Seda, A. K.,
editors, Proceedings of the Second Irish Conference on the Mathematical
Foundations of Computer Science and Information Technology (MFC­
SIT2002), Galway, Ireland, July, 2002, Electronic Notes in Theoretical
Computer Science, Volume 74, pages 1–17. Elsevier Science Publishers,
Amsterdam; New York.
[Rounds and Zhang, 2001] Rounds, W. C. and Zhang, G.-Q. (2001). Clausal
logic and logic programming in algebraic domains. Information and Com­
putation, 171(2):156–182.
[Rumelhart et al., 1986] Rumelhart, D. E., Hinton, G. E., and Williams., R. J.
(1986). Learning internal representations by error propagation. In Mc-
Clelland, J. L. and Rumelhart, D. E., editors, Parallel Distributed Pro­
cessing, pages 318–362. MIT Press, Cambridge, MA.
[Rutten, 1995] Rutten, J. J. (1995). Elements of generalized ultrametric do­
main theory. Technical Report CS-R9507, CWI, Stichting Mathematisch
Centrum, Amsterdam, The Netherlands.
[Rutten, 1996] Rutten, J. J. (1996). Elements of generalized ultrametric do­
main theory. Theoretical Computer Science, 170:349–381. Revised version
of [Rutten, 1995].
[Scott, 1982a] Scott, D. S. (1982a). Domains for denotational semantics. In
Nielsen, M. and Schmidt, E. M., editors, Automata, Languages and Pro­
gramming, 9th Colloquium, Aarhus, Denmark, July 12-16, 1982, Proceed­
ings, Lecture Notes in Computer Science, Volume 140, pages 577–613.
Springer, Berlin.
[Scott, 1982b] Scott, D. S. (1982b). Lecture notes on a mathematical theory
of computation. In Broy, M. and Schmidt, G., editors, Theoretical Foun­
dations of Programming Methodology, pages 145–292. Reidel, Dordrecht.
[Scutellà, 1990] Scutellà, M. G. (1990). A note on Dowling and Gallier’s top-
down algorithm for propositional Horn satisfiability. The Journal of Logic
Programming, 8:265–273.
[Seda, 1995] Seda, A. K. (1995). Topology and the semantics of logic pro­
grams. Fundamenta Informaticae, 24(4):359–386.
[Seda, 1997] Seda, A. K. (1997). Quasimetrics and the semantics of logic
programs. Fundamenta Informaticae, 29(1):97–117.
[Seda, 2002] Seda, A. K. (2002). Some convergence issues in theoretical com­
puter science. Information, 5(4):447–462.
262 Bibliography

[Seda, 2006] Seda, A. K. (2006). On the integration of connectionist and

logic-based systems. In Hurley, T., MacanAirchinnigh, M., Schellekens,
M., Seda, A. K., and Strong, G., editors, Proceedings of the Third Irish
Conference on the Mathematical Foundations of Computer Science and
Information Technology (MFCSIT2004), Trinity College Dublin, Ireland,
July, 2004, Electronic Notes in Theoretical Computer Science, Volume
161, pages 109–130. Elsevier Science Publishers, Amsterdam; New York.

[Seda, 2007] Seda, A. K. (2007). Morphisms of ANN and the computation of

´
least fixed points of semantic operators. In Mira, J. and Alvarez, J. R.,
editors, Proceedings of the Second International Work-Conference on the
Interplay Between Natural and Artificial Computation (IWINAC2007),
Murcia, Spain, June, 2007, Part 1, Lecture Notes in Computer Science,
Volume 4527, pages 224–233. Springer, Berlin.

[Seda and Hitzler, 1999a] Seda, A. K. and Hitzler, P. (1999a). Strictly level-
decreasing logic programs. In Butterfield, A. and Flynn, S., editors, Pro­
ceedings of the Second Irish Workshop on Formal Methods (IWFM’98),
NUI, Cork, Ireland, July, 1998, Electronic Workshops in Computing,
pages 1–18. British Computer Society, Swindon, UK.

[Seda and Hitzler, 1999b] Seda, A. K. and Hitzler, P. (1999b). Topology and
iterates in computational logic. In Flagg, R., Hart, K., Norden, J., Tym­
chatyn, E., and Tuncali, M., editors, Proceedings of the 12th Summer
Conference on Topology and Its Applications: Special Session on Topol­
ogy in Computer Science, Ontario, Canada, August, 1997, Volume 22 of
Topology Proceedings, pages 427–469.

[Seda and Hitzler, 2010] Seda, A. K. and Hitzler, P. (2010). Generalized dis­
tance functions in the theory of computation. The Computer Journal,
53(4):443–464.

[Seda and Lane, 2003] Seda, A. K. and Lane, M. (2003). On continuous mod­
els of computation: Towards computing the distance between (logic) pro­
grams. In Morris, J., Aziz, B., and Oehl, F., editors, Proceedings of the
Sixth International Workshop in Formal Methods (IWFM’03), Dublin
City University, Dublin, Ireland, July, 2003, Electronic Workshops in
Computing, pages 1–15. British Computer Society, Swindon, UK.

[Seda and Lane, 2005] Seda, A. K. and Lane, M. (2005). On the measurability
of the semantic operators determined by logic programs. Information,
8(1):33–52.

[Seda et al., 2003] Seda, A. K., Heinze, R., and Hitzler, P. (2003). Conver­
gence classes and spaces of partial functions. In Domain Theory, Logic
and Computation. Proceedings of the 2nd International Symposium on
 Bibliography 263

Domain Theory, ISDT (2001), Sichuan University, Chengdu, China, Oc­

tober, 2001, Semantic Structures in Computation, Volume 3, pages 75–
115. Kluwer Academic Publishers, Dordrecht, The Netherlands.
[Simons et al., 2002] Simons, P., Niemelä, I., and Soininen, T. (2002). Extend­
ing and implementing the stable model semantics. Artificial Intelligence,
138:181–234.
[Smyth, 1987] Smyth, M. B. (1987). Quasi uniformities: Reconciling domains
with metric spaces. In Main, M. G., Melton, A., Mislove, M. W., and
Schmidt, D. A., editors, Mathematical Foundations of Programming Lan­
guage Semantics, Lecture Notes in Computer Science, Volume 198, pages
236–253. Springer, Berlin.
[Smyth, 1991] Smyth, M. B. (1991). Totally bounded spaces and compact
ordered spaces as domains of computation. In Reed, G. M., Roscoe,
A. W., and Wachter, R. F., editors, Topology and Category Theory in
Computer Science, pages 207–229. Oxford University Press, Oxford, UK.
[Smyth, 1992] Smyth, M. B. (1992). Topology. In Abramsky, S., Gabbay,
D. M., and Maibaum, T. S., editors, Handbook of Logic in Computer
Science Volume 1, pages 641–761. Oxford University Press, Oxford, UK.
[Stoltenberg-Hansen et al., 1994] Stoltenberg-Hansen, V., Lindström, I., and
Griffor, E. R. (1994). Mathematical Theory of Domains. Cambridge
Tracts in Theoretical Computer Science No. 22. Cambridge University
Press, Cambridge, UK.
[Straccia et al., 2009] Straccia, U., Ojeda-Aciego, M., and Damásio, C. V.
(2009). On fixed points of multivalued functions on complete lattices
and their application to generalized logic programs. SIAM Journal of
Computing, 38(5):1881–1911.
[van Emden and Kowalski, 1976] van Emden, M. H. and Kowalski, R. A.
(1976). The semantics of predicate logic as a programming language.
Journal of the ACM, 23(4):733–742.
[Van Gelder et al., 1991] Van Gelder, A., Ross, K. A., and Schlipf, J. S.
(1991). The well-founded semantics for general logic programs. Jour­
nal of the ACM, 38(3):620–650.
[Waszkiewicz, 2002] Waszkiewicz, P. (2002). Quantitative Continuous Do­
mains. PhD thesis, School of Computer Science, The University of Birm­
ingham, Edgbaston, Birmingham, UK.
[Waszkiewicz, 2003] Waszkiewicz, P. (2003). Quantitative continuous do­
mains. Applied Categorical Structures, 11(1):41–67.
[Waszkiewicz, 2006] Waszkiewicz, P. (2006). Partial metrizability of continu­
ous posets. Mathematical Structures in Computer Science, 16(2):359–372.
264 Bibliography

[Wendt, 2002a] Wendt, M. (2002a). Unfolding the well-founded semantics.

Technical Report WV–02–08, Knowledge Representation and Reasoning
Group, Department of Computer Science, Dresden University of Tech­
nology.

[Wendt, 2002b] Wendt, M. (2002b). Unfolding the well-founded seman­

tics. Journal of Electrical Engineering, Slovak Academy of Sciences,
53(12/s):56–59.

[Willard, 1970] Willard, S. (1970). General Topology. Addison-Wesley, Read­

ing, MA.

[Zhang, 1991] Zhang, G.-Q. (1991). Logic of Domains. Birkhauser, Boston.

[Zhang and Rounds, 1997a] Zhang, G.-Q. and Rounds, W. C. (1997a). Com­
plexity of power default reasoning. In Proceedings of the Twelfth An­
nual IEEE Symposium on Logic in Computer Science, LICS’97, Warsaw,
Poland, pages 328–339. IEEE Computer Society Press.

[Zhang and Rounds, 1997b] Zhang, G.-Q. and Rounds, W. C. (1997b). Rea­
soning with power defaults (preliminary report). In Dix, J., Furbach,
U., and Nerode, A., editors, Proceedings of the Fourth International
Conference on Logic Programming and Non-Monotonic Reasoning (LP­
NMR’97), Dagstuhl, Germany, Lecture Notes in Computer Science, Vol­
ume 1265, pages 152–169. Springer, Berlin.

[Zhang and Rounds, 2001] Zhang, Q.-Z. and Rounds, W. C. (2001). Seman­
tics of logic programs and representation of Smyth powerdomains. In
Keimel, K. et al., editors, Domains and Processes, pages 151–179. Kluwer
Academic Publishers, Dordrecht, The Netherlands.
Index

A logic, 10
Banach contraction mapping theo­
A sublogic, 10
rem, xxiii, 93-108, 113-116, 140­
Abelian monoid, 89
147, 166, 194

Abelian semigroup, 89, 130

multivalued, 129, 134

Absorbing element in a semigroup, 89

Base for a topology, 235

Acceptable program, xxiii, 26, 66, 76,

Basic neighbourhood, 235

145, 158
Basic open set, 66, 70, 74, 82

Activation function, 188

Becomes active at time t, 189, 193

Acyclic program, 141, 147-160, 165,

Belnap’s four-valued logic, 10, 20, 83,

196
152, 165-167, 213

Addition unit, 217

Bi-Cauchy sequence, 109

Algebraicity, 3
Binary threshold unit, 189-196, 215,

Allowable values, 213

216

α-th layer, 46
Body literal, 24, 45, 50, 62, 144, 162

α-th stratum, 46
Body of a clause, 24, 51, 145, 169, 200

Alphabet, 7
Body of a pseudo-clause, 152, 164, 214

Alternating fixed-point model, 61

Both, 10

Answer set programming, xxi, 128,

Bottom element, 2, 13, 74, 98, 122,

135, 137
128, 192, 242

Antisymmetric relation, 1
Bottom layer, 46, 51, 61

Antitonic function, 5
Bottom stratum, 46, 51, 54

Approximates up to ε, 190, 200, 204,

208, 212
Cantor topology, xxiii, 66, 79, 108,

Approximating TP up to level n, 200,

143, 155, 161, 198

203
Cauchy sequence, 91-96

Approximating least fixed points, 209

in a d-metric, 103, 116

Approximating networks exist, 190­ in a quasimetric, 106-111, 133-135

208, 218
Centre of a ball, 99

Approximation in space, 197

Chain complete, 2-5, 121-123

Approximation in time, 197

Chain in an ordered set, 1

Atom, 8
Clark completion, xx, 32

ground, 8
Classical logic, xx, 9, 43, 165

Atomic formula, 8
Clause, 23

of finite type, 162

Ball in a continuity space, 90

pseudo-, 152, 153, 164-167, 214

Ball in a d-gum, 99, 104, 123-125, 148

Closed ball about a point, 89

Ball in a gum, 99, 100, 119, 130, 132

Closed set, 71, 236

265
266 Index

Closure of a set, 236

Covered logic program, 200-208

Closure operator, 67, 68, 236

Cpo, 2

Closure ordinal, 7, 41
CS-complete quasimetric space, 106­
Coarser topology, 234
108

Commutative monoid, 89
CS-continuous function, 106-111

Commutative semigroup, 89

Compact element, 3, 14, 69, 119, 240

d-gum space, 97

Compact topological space, 239

d-metric, 91, 92, 102

Comparable elements, 1, 242

Declarative semantics, xxi, 15, 23,

Complete d-metric space, 103, 115

137, 186

Complete lattice, 2, 61, 75, 125, 181

Definite logic program, 24

Complete partial order, 2, 69, 126, 170

Denotation of a term, 11

Complete upper semi-lattice, 2, 13

Denotational semantics, xxii, 4, 221­
Completeness in metric spaces, 91
224

Dependency graph, 45-48

Component, 45

Depends negatively on, 45

minimal, 46

Depends on, 45, 144

trivial, 45

Directed set, 2, 237

Conjunction, xix, 10, 24, 152, 214

Discrete quasimetric, 107, 108

Connectionist network, 188

Discrete topology, 77, 78, 82, 100, 235

Connectionist system, 185, 188, 198

Disjunction, 10, 214

Connectives, 8, 10, 154, 161

Disjunctive logic programs, xix, 125,

Consequence operator, 161-166, 213

130, 135

locally finite, 163

Dislocated generalized ultrametric space,

Consistent completeness, 3, 120

97, 104

Consistent signed subset, 19

Dislocated metric, 91, 92, 102

Continuity function, 90-92, 98

Dislocation function, 114

Continuity of a function, 237

Distance function, 88-91

at a point, 68, 237

d-gum, 97

Continuity space, 88-92, 98

d-metric, 91, 92, 102

Continuous function, 5, 68, 133, 204,

g-metric, 130-132

237

gum, 92, 97

Contraction mapping, 93, 114

metric, 91

multivalued, 129
pseudometric, 91

multivalued on a quasimetric, 133

quasimetric, 91

on a d-metric, 103
ultrametric, xxv, 91

on a quasimetric, 106, 107

Distance set, 88, 100, 121

Contractive, 106, 144

DLV system, xix, 169

Contractivity factor, 93, 114-119, 144

Domain, 3, 4, 69, 118, 136, 222, 240

Convergence class, xxiii, 67-72, 77, 78

Scott, 3

Convergence in a d-metric, 103

Scott-Ershov, 3

Convergence in a quasimetric, 106

Domain of a preinterpretation, 11, 26

Convergence space, xxiii, 67, 68

Downwards closed set, 75

Convergent net, 67, 237

Dual ordering, 2, 119

Converges to, 67, 237

Core method, 187, 192, 220

Egli–Milner monotonicity, 126

 Index 267

Egli–Milner ordering, 126

FOU R, 10, 16, 72, 167, 213-217

Embedding of well-ordered sets, 230

Four-valued interpretation, 18-20, 28

Even, 25, 27, 140

Four-valued logic, 10, 165

Even2, 197
Four-valued valuation, 18

Eventually, 237
FP -operator, 164

Eventually increasing orbit, 127

Funahashi’s theorem, 187, 190

Excluded values, 213

Function, 5

ExistsEven, 141
ω-continuous, 5, 6, 40, 42, 57, 135

antitonic, 5

Fact, 24
continuous, 5, 68, 133, 204, 237

False, 9, 20, 161

continuous at a point, 68, 237

Family of networks computes I, 212

contraction on a d-metric, 103

Finer topology, 234

contraction on a quasimetric, 106,

Finite element, 3
107

Finite valuation, 13
CS-continuous, 106-111

Finitely determined conjunction, 215

monotonic, 5, 230

Finitely determined disjunction, 167,

non-expanding on d-gums, 99

215
non-expanding on a quasimetric,

Finitely determined operation, 213

106

First infinite ordinal, 233

orbit of, 99

First-order language, 9
order continuous, 5

Fitting model, 39-44, 55, 63, 139, 151,

ordinal powers of, 5

174
Scott continuous, xxii, 72, 74, 170

Fitting operator, 37, 42, 43, 75, 151,

single valued, 129, 133, 134

152, 161, 174

strictly contracting on a metric, 94

Fitting semantics, 37, 42, 55, 56, 150,

strictly contracting on a (d-)gum,

156
99

Fitting-style operator, xxvi, 152, 164­ strictly contracting on orbits, 99

167, 213-219

Fixed point, 5
g-metric, 130-132

least, 5, 30, 209

g-metric space, 130-132

least post-, 6
Game, 145, 146

least pre-, 6
Gelfond–Lifschitz operator, xxvii, 34,

post-, 5
35, 61, 165, 171, 218, 219

pre-, 5
Gelfond–Lifschitz transform, 34, 35,

Fixed point of a multivalued mapping,

62

125
Generalized ultrametric, 92, 97

Fixed-point semantics, xxii

Generalized ultrametric space, 97, 98

Fixpoint completion of a program,

Generate-and-test, 141, 147

xxvii, 169, 170

GLP -operator, 34

Formal ball, 121

Greatest element, 2, 126

Formula, 8
Greatest limit of a net, 75

ground, 8
Greatest lower bound, 2

Forward Cauchy sequence, 106-108,

Ground atom, 8

134
Ground formula, 8

268 Index

Ground instance of a clause, 26

well-supported, 33

Ground instance of an atom, 11, 26

Isomorphic well-orderings, 230

Ground term, 8
Iterates of a function, 5

Gum, 92, 97

Gum associated with a d-gum, 123

J-ground instance of a clause, 26

Gum induced by a rank function, 119

J-ground instance of an atom, 11

Gum with ordinal distances, 121, 142

k-safe clause, 158

A 1
b Kernel of a neural network, 190

2
-contraction, 131
Kleene’s strong three-valued logic, 9,

Hausdorff space, 239

37, 48, 57, 152, 165, 213

Head of a clause, 24

Kleene’s theorem, xxiii, 6, 7

Head of a list, 26

multivalued, 135

Head of a pseudo-clause, 152

Kleene’s weak three-valued logic, 9, 11

Herbrand preinterpretation, 26

Knaster-Tarski theorem, 6, 7

Herbrand universe, 26

multivalued, 128

Hidden layer, 189

Knowledge ordering, 10, 16, 37

Hoare monotonicity, 126-128, 135

Knowledge Representation and Rea­

Hoare ordering, 126

soning, xix, 223

Kripke-Kleene model, 39

I-partial level mapping, 40

Kripke-Kleene semantics, 37

Identity of a semigroup, 89

Immediate consequence operator, 29,

Layer in a network, 189

42, 66, 74, 110, 146, 196

hidden, 189

Immediate consequence operator re­ input, 189-191

stricted at level n, 175

output, 189-191

Implication, 10, 16
Least element, 2, 97, 126, 215, 229

Incomparable elements, 1, 10, 242

Least fixed point, 5, 30, 209

Increasing orbit, 127

Least infinite ordinal, 233

Indiscrete topology, 235

Least model, 30-35, 76, 196

Infimum, 2
Least model semantics, 31

Initial segment, 230, 231

Least post-fixed point, 6

proper, 230
Least pre-fixed point, 6

Input layer, 189-191

Least upper bound, 2, 121, 135

Input vector, 188, 189

Length, 25, 26

Input-output mapping, 189, 190

Level mapping, 33, 44, 109-112, 142,

Intended model, 23, 30, 45, 55

174, 198

Interior of a set, 236

Limit of a net, 237

Interpretation, 12, 28, 161, 167, 191,

Limit of a sequence in a d-metric, 103

198
Limit ordinal, 1, 5, 62, 98, 127, 232­
four-valued, 18-20, 28
234

partial, 19
Linear unit, 189

supported, 31-33
List, 26

three-valued, 18, 28
Literal, 8

total, 19, 42
negative, 8

two-valued, xxvii, 18, 19, 28

positive, 8

 Index 269

Local consequence operator, 162-166,

Kleene’s strong three-valued, 9, 37,

213, 218
48, 57, 152, 165, 213

Local variable, 162, 163, 173, 200

three-valued, 9

Locally finite consequence operator,

two-valued, 9, 10

163
Lower bound, 2, 130

Locally hierarchical program, 141­ greatest, 2

147, 156, 174, 179-181

Locally stratified program, 44

Matthews’ theorem, 103, 104, 114,

Logic program, 23, 24

115, 118, 147

ω-locally hierarchical, 141

Maximal element, 2-4, 122

Φ-accessible, 139, 150, 151

Metric, 91

Φ∗ -accessible, 147-150
Metric associated with a d-metric,

acceptable, xxiii, 26, 66, 76, 145,

117, 118

158
Metric domain, 102

acyclic, 141, 147-160, 165, 196

Metric induced by a quasimetric, 106

covered, 200-208
Metrizable space, 79

definite, 24
Midpoint of a ball, 99

Minimal component, 46

disjunctive, xix, 125, 130, 135

Minimal element, 2, 129

locally hierarchical, 141-147, 156,

Minimal model, 30-34

174, 179-181

Model, 12, 28

locally stratified, 44

alternating fixed point, 61

normal, 24

Fitting, 39-44, 55, 63, 139, 151, 174

of finite type, 162

intended, 23, 30, 45, 55

positive, 24

Kripke-Kleene, 39

propositional, 24-27, 187, 192, 217

least, 30-35, 76, 196

satisfies (F) with respect to I and

minimal, 30-34

l, 41, 150, 151

perfect, xxiv, 43-45, 175-183

satisfies (F12 ) with respect to I and

stable, xxi, 33-37, 43-46, 62, 130­
l, 159
136, 169-174, 219

satisfies (F22 ) with respect to I and

supported, xxi, 31-33

l, 158
three-valued supported, 40, 151

satisfies (F32 ) with respect to I and

weakly perfect, 45-63, 182, 183

l, 156
well-founded, xxi, 57-64, 170-175

satisfies (WF) with respect to I

well-supported, 33

and l, 56
Monotonic function, 5, 230

satisfies (WS) with respect to I and

Multilayer feedforward network, 189­
l, 49
191

stratified, 44
Multiplication unit, 215-217

uniquely determined, 139

Multivalued contraction mapping, 129

weakly stratified, 48
Multivalued mapping, 125, 222

Logic programming, xix

ω-continuous, 135

Logic, 9
ω-orbit of, 127, 133-135

classical, xx, 9, 43, 165

continuous on a quasimetric, 133

four-valued, 10, 165

contraction, 129

270 Index

contraction on a quasimetric, 133

Non-empty multivalued mapping, 126

Egli–Milner monotonic, 126

Non-expanding function in a d-gum,

fixed point of, 125

99

Hoare monotonic, 126-128, 135

Non-expanding function in a quasi-

non-empty, 126
metric, 106

non-expanding on a gum, 129

Non-expanding multivalued mapping

non-expanding on a quasimetric,
on a gum, 129

133
Non-expanding multivalued mapping

orbit of, 127

on a quasimetric, 133

Smyth monotonic, 126

Non-termination, 10

strictly contracting on a (d-)gum,

None, 9

129
Normal logic program, 24

strictly
A 1
b contracting on orbits, 129

2
-contraction, 131
Occurs negatively in body, 24

Occurs positively in body, 24

n-dimensional level mapping, 207

ω-chain, xxiii, 1-6

n-layer feedforward network, 189, 190

ω-complete partial order, 2-6, 14, 135

n-layer recurrent network, 190

ω-continuous function, 5, 6, 40, 42, 57,

Negation, xx, 8, 10, 20, 44, 56, 80

135

Negation as (finite) failure, xx

ω-continuous multivalued mapping,

Negative literal, 8
135

NegP , 144
ω-cpo, 2

Neg∗P , 145
ω-locally hierarchical program, 141

Neighbourhood, 235
ω-orbit of a multivalued mapping,

base, 235
127, 133-135

basic, 235
Ontology, 223, 224

system, 235
Open ball about a point, 89

Net, 237
Open cover of a topological space, 239

convergence of, 67, 237

subcover of, 239

converges to, 67, 237

Open set, 66, 79, 82, 89, 163, 234

eventually (satisfies a property),

Operational semantics, xxi

237
Orbit of a function, 99

greatest limit of, 75

Orbit of a multivalued mapping, 127

limit of, 237

eventually increasing, 127

subnet of, 237

increasing, 127

tail of, 237

tight, 127

Network function, 189

Order isomorphism, 230

Network made recurrent, 190

Order-continuous function, 5

Neural network, 185-191

Ordered semigroup with identity, 89

RBF, 204
Ordered set, 1

sigmoidal, 201
Ordering, 1

vector-based, 207
dual, 2, 119

Neural-Symbolic Cycle, 186

Egli–Milner, 126

Neural-symbolic integration, 185-188

Hoare, 126

Node in a neural network, 188

knowledge, 10, 16, 37

 Index 271

Smyth, 126
Prieß-Crampe and Ribenboim theo­
truth, 10, 15, 16
rem multivalued, 130

Ordinal, 2, 231
Principle of Transfinite Induction, 233

closure, 7, 41
Procedural aspects, xxv, 29

first infinite, 233

Procedural semantics, xxi, xxiii, 29

least infinite, 233

Product topology, 80, 82, 162, 239

limit, 1, 5, 62, 98, 127, 232-234

projection on the j-th factor of, 239

predecessor, 232
Program, 24

successor, 232
see also Logic program

Ordinal number, 231

Program clause, 23

Ordinal powers of a function, 5

Prolog, xix, 25, 141, 144

Output function, 189, 190

Proper initial segment, 230

Output layer, 189-191

Proposition, 8

Overdefined, 10
Propositional approximation, 187, 199

Propositional logic program, 24, 25,

− 27, 187, 192, 217

P , 145

P satisfies (F) with respect to I and

P -local consequence operator, 162­ l, 41, 150, 151

166, 213, 218

P satisfies (F12 ) with respect to I and

Partial interpretation, 19
l, 159

Partial metric, 102

P satisfies (F22 ) with respect to I and

Partial order, 1-6

l, 158

Partial order induced by +, 90

P satisfies (F32 ) with respect to I and

Partial order induced by a quasimet­ l, 156

ric, 106
P satisfies (WF) with respect to I and

Partially ordered set, 1-3

l, 56

Perfect model, xxiv, 43-45, 175-183

P satisfies (WS) with respect to I and

Perfect model semantics, 44, 175-183

l, 49

Φ-accessible program, 139, 150, 151

Pseudo-clause, 152, 153, 164-167, 214

Φ∗ -accessible program, 147-150

body of, 152, 164, 214

ΦP -operator, 37
head of, 152

ΦP,j,k -operator, 154

Pseudo-convergent transfinite sequence,

Poset, 1
101

Positive literal, 8
Pseudo-limit of a transfinite sequence,

Positive logic program, 24

101

Post-fixed point, 5
Pseudometric, 91

Potential, 188

Powers of an operator, 181

Quantitative domain theory, xxvii,

Pre-fixed point, 5
222, 225

Predecessor of an ordinal, 232

Quasi-interpretation, 169, 170

Preference between models, 44, 180,

Quasimetric, 91

181
Quasimetric induced by a rank func­
Preinterpretation, 11, 26
tion, 109

Prieß-Crampe and Ribenboim theo­

rem, 100
Rank function, 109, 119

272 Index

RBF network, 204

Sigmoidal network, 201

Recurrent network, 190

Sigmoidal unit, 189

Reduct of P with respect to I, 46

Signed subset, 13, 19, 20, 37, 46, 154

Refers negatively to, 45

Single-step operator, xxiii, 29, 80, 140,

Refers to, 45, 144

155, 171, 218

Reflexive relation, 1
Single-valued function, 129, 133, 134

Relation, 1
Single-valued mapping, 129, 133, 134

antisymmetric, 1
SLD-resolution, xix, 147

reflexive, 1
SLDNF-resolution, xix, 32, 141, 144

transitive, 1
Smodels system, xix, 169

Relative topology, 239

Smyth monotonicity, 126, 137

Required values, 213

Smyth ordering, 126

Resolution, xix, 31
Space of formal balls, 121

Rule, 23
Spherically complete (d-)gum, 99

Rutten-Smyth theorem, 107

Squashing functions, 189

multivalued, 133
Squashing unit, 189

Stable model, xxi, 33-37, 43-46, 62,

S-monotonic, 126
130-136, 169-174, 219

Scott continuous function, xxii, 72,

Stable model semantics, 34, 36, 37, 43,

74, 170
171

Scott domain, 3
Stratified program, 44

Scott open set, 69-71, 242

Strictly contracting function on a

Scott topology, xxiii, 69-76, 234

metric, 94

Scott-Ershov domain, 3
Strictly contracting function on a

Semantic operator, xxii, 10, 66, 126,

(d-)gum, 99

136, 161, 186

Strictly contracting multivalued func­
Semantic Web, xix, 223
tion on a (d-)gum, 129

Semantics, xx
Strictly contracting function on or­
declarative, xxi, 15, 23, 137, 186
bits, 99

denotational, xxii, 4, 221-224

Strictly contracting multivalued func­
Fitting, 37, 42, 55, 56, 150, 156
tion on orbits, 129

fixed-point, xxii
Strong triangle inequality, 92

Kripke-Kleene, 37
Strong triangle inequality for gums,

least model, 31
98

operational, xxi
Stronger topology, 234

perfect model, 44, 175-183

Subbase for a topology, 74, 82, 167,

procedural, xxi, xxiii, 29

235

stable model, 34, 36, 37, 43, 171

Subbasic open set, 82, 163

weakly perfect, 45-63

Subnet of a net, 237

well-founded, 57-61, 170, 225

Subspace topology, 239

Semigroup, 89
Subsumes, 50

Semigroup with identity, 89

Subsumes model-consistently, 50

Set of positives, 90
Successor notation, 25-27

Set of truth values, 10

Successor ordinal, 232

σ-algebra, 167
Supported interpretation, 31-33

 Index 273

Supported model, xxi, 31-33

associated with a closure operator,

Supremum, xxiii, 2, 6, 12, 75, 128

236

associated with a convergence class,

T0 -space, 238
68

T1 -space, 238
base for, 235

T2 -space, 239
coarser, 234

see also Hausdorff space

discrete, 77, 78, 82, 100, 235

finer, 234

Tail of a list, 26

generated by a continuity space, 90

Tail of a net, 237

indiscrete, 235

Term assignment, 11

product, 80, 82, 162, 239

Term over an alphabet, 8

relative, 239

ground, 8

stronger, 234

Theorem,

subbase for, 74, 82, 167, 235

Banach, xxiii, 93-108, 113-116,

subspace, 239

140-147, 166, 194

weaker, 234

Banach multivalued, 129, 134

Total interpretation, 19, 42

Funahashi, 187, 190

Total order on a set, 1, 19

Kleene, xxiii, 6, 7

Totally bounded quasimetric space,

Kleene multivalued, 135

106-110

Knaster-Tarski, 6, 7
Totally ordered set, 1, 98, 100, 119

Knaster-Tarski multivalued, 128

TP -operator, 29

Matthews, 103, 104, 114, 115, 118,

Trans-complete gum, 101

147
Transfinite sequence, 233

Prieß-Crampe and Ribenboim, 100

Transitive relation, 1

Prieß-Crampe and Ribenboim mul­ Translation algorithm, 193, 196

tivalued, 130
Triangle inequality, 92, 98

Rutten-Smyth, 107
Trivial component, 45

Rutten-Smyth multivalued, 133

True, 9, 10, 20, 161

Tychonoff, 240
Truth ordering, 10, 15, 16

Well-Ordering, 231
Truth set, 10

T HREE, 9, 10, 11, 16

Truth value, 9, 10

Three-valued interpretation, 18, 28

both or overdefined, 10

Three-valued logic, 9
false, 9, 20, 161

Three-valued stable model, 170

none or underdefined, 9

Three-valued supported model, 40,

true, 9, 10, 20, 161

151
Tweety, 25

Three-valued valuation, 18
Tweety1, 25, 27, 32, 35, 36, 38, 42

Threshold, 188
Tweety2, 36, 42, 45, 183

Tight orbit, 127

Tweety3, 37, 42, 45

Top element, 2
Tweety4, 47, 61

Topological space, 234

T W O, 9, 10, 11, 15

compact, 239
Two-valued interpretation, xxvii, 18,

subspace of, 239

19, 28

Topology, 234
Two-valued logic, 9, 10

274 Index

Two-valued valuation, 18
Valuation, 12, 13, 28, 161

Tychonoff’s theorem, 240

finite, 13

four-valued, 18

Ultrametric, xxv, 91
three-valued, 18

Ultrametric induced by a rank func­ two-valued, 18

tion, 119
Value of a node, 188

Undefined, 20
Value semigroup, 89

Underdefined, 9
Value set, 88

Underlying first-order language, 24

Variable assignment, 11

Underlying language, 24
Vector-based network, 207

Unfolding, 169-172

Unfounded set, 57-60

Weak partial metric, 102

Unique supported model class, xxvi,

Weak stratification, 45, 48

139
Weaker topology, 234

Uniquely determined program, 139

Weakly perfect model, 45-63, 182, 183

Unit clause, 24
Weakly perfect semantics, 45-63

Unit in a neural network, 188

Weakly stratified program, 48

addition, 217
Web Ontology Language OWL, 224

binary threshold, 189-196, 215, 216

Weight function, 114, 146

linear, 189
Well-formed formula, xxi, 8-12

multiplication, 215-217
Well-founded model, xxi, 57-64, 170­
sigmoidal, 189
175

squashing, 189
Well-founded semantics, 57-61, 170,

Upper bound, 2, 121, 135, 196

225

least, 2, 121, 135

Well-ordered set, 229

Upwards closed set, 69, 240

Well-ordering, 229

Well-Ordering Theorem, 231

v-Cauchy, 131
Well-supported interpretation, 33

v-cluster, 131
Well-supported model, 33

v-complete g-metric, 131

Winner unit, 208

v-complete subset, 131

WP -operator, 57