Cognitive Technologies Managing Editors: D. M. Gabbay J. Siekmann Editorial Board: A. Bundy J. G. Carbonell M. Pinkal H. Uszkoreit M. Veloso W. Wahlster M. J. Wooldridge
Advisory Board: Luigia Carlucci Aiello Franz Baader Wolfgang Bibel Leonard Bolc Craig Boutilier Ron Brachman Bruce G. Buchanan Anthony Cohn Artur d’Avila Garcez Luis Fari˜nas del Cerro Koichi Furukawa Georg Gottlob Patrick J. Hayes James A. Hendler Anthony Jameson Nick Jennings Aravind K. Joshi Hans Kamp Martin Kay Hiroaki Kitano Robert Kowalski Sarit Kraus Maurizio Lenzerini Hector Levesque John Lloyd
For further volumes: http://www.springer.com/series/5216
Alan Mackworth Mark Maybury Tom Mitchell Johanna D. Moore Stephen H. Muggleton Bernhard Nebel Sharon Oviatt Luis Pereira Lu Ruqian Stuart Russell Erik Sandewall Luc Steels Oliviero Stock Peter Stone Gerhard Strube Katia Sycara Milind Tambe Hidehiko Tanaka Sebastian Thrun Junichi Tsujii Kurt VanLehn Andrei Voronkov Toby Walsh Bonnie Webber
•
Dov M. Gabbay
•
Karl Schlechta
Conditionals and Modularity in General Logics
Prof. Dov M. Gabbay Department of Computer Science King’s College London Strand, London WC2R 2LS, UK and Department of Computer Science Bar Ilan University 52900 Ramat-Gan, Israel
Prof. Karl Schlechta Université de Provence CMI 39, rue Joliot-Curie 13453 Marseille Cedex 13 France
[email protected] and Computer Science and Communications Faculty of Sciences University of Luxembourg 6, rue Coudenhove-Kalergi L-1359 Luxembourg
[email protected] Managing Editors Prof. Dov M. Gabbay Augustus De Morgan Professor of Logic Department of Computer Science King’s College London Strand, London WC2R 2LS, UK
Prof. Dr. Jörg Siekmann Forschungsbereich Deduktions- und Multiagentensysteme, DFKI Stuhlsatzenweg 3, Geb. 43 66123 Saarbrücken, Germany
Cognitive Technologies ISSN 1611-2482 ISBN 978-3-642-19067-4 e-ISBN 978-3-642-19068-1 DOI 10.1007/978-3-642-19068-1 Springer Heidelberg Dordrecht London New York ACM Codes 1.2.3, F.4.1 Library of Congress Control Number: 2011934866 © Springer-Verlag Berlin Heidelberg 2011 This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilm or in any other way, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer. Violations are liable to prosecution under the German Copyright Law. The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. Cover design: deblik Printed on acid-free paper Springer is part of Springer Science+Business Media (www.springer.com)
Contents
1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.1 The Main Subjects of This Book.. . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.1.1 An Example.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.1.2 Connections.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.2 Main Definitions and Results . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.2.1 The Monotonic Case . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.2.2 The Non-monotonic Case . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.3 Overview of This Introduction .. . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.4 Basic Definitions.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.5 Towards a Uniform Picture of Conditionals .. . . . .. . . . . . . . . . . . . . . . . . . . 1.5.1 Discussion and Classification . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.5.2 Additional Structure on Language and Truth Values . . . . . . . . 1.5.3 Representation for General Revision, Update, and Counterfactuals . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.6 Interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.6.1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.6.2 Problem and Method . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.6.3 Monotone and Antitone Semantic and Syntactic Interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.6.4 Laws About Size and Interpolation in Non-monotonic Logics .. . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.6.5 Summary .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.7 Neighbourhood Semantics . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.7.1 Defining Neighbourhoods.. . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.7.2 Additional Requirements.. . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.7.3 Connections Between the Various Properties . . . . . . . . . . . . . . . . 1.7.4 Various Uses of Neighbourhood Semantics .. . . . . . . . . . . . . . . . .
1 1 2 5 6 6 7 7 8 9 9 10 11 12 12 13 14 17 24 24 25 25 25 26
v
vi
Contents
1.8
An Abstract View on Modularity and Independence . . . . . . . . . . . . . . . . 1.8.1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.8.2 Abstract Definition of Independence . . . . .. . . . . . . . . . . . . . . . . . . . 1.8.3 Other Aspects of Independence .. . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.9 Conclusion and Outlook .. . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.10 Previously Published Material, Acknowledgements .. . . . . . . . . . . . . . . .
26 26 27 28 30 30
2 Basic Definitions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 2.1.1 Overview of This Chapter.. . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 2.2 Basic Algebraic and Logical Definitions . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 2.2.1 Countably Many Disjoint Sets . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 2.2.2 Introduction to Many-Valued Logics .. . . .. . . . . . . . . . . . . . . . . . . . 2.3 Preferential Structures .. . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 2.3.1 The Minimal Variant . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 2.3.2 The Limit Variant.. . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 2.3.3 Preferential Structures for Many-Valued Logics .. . . . . . . . . . . . 2.4 IBRS and Higher Preferential Structures . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 2.4.1 General IBRS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 2.4.2 Higher Preferential Structures . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 2.5 Theory Revision . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 2.5.1 Theory Revision for Two-Valued Logics .. . . . . . . . . . . . . . . . . . . . 2.5.2 Theory Revision for Many-Valued Logics . . . . . . . . . . . . . . . . . . .
31 31 31 33 42 43 47 48 55 59 63 63 65 77 78 83
3 Towards a Uniform Picture of Conditionals . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3.1.1 Overview of This Chapter.. . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3.2 An Abstract View on Conditionals.. . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3.2.1 A General Definition as Arbitrary Operator.. . . . . . . . . . . . . . . . . 3.2.2 Properties of Choice Functions . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3.2.3 Evaluation of Systems of Sets . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3.2.4 Conditionals Based on Binary Relations .. . . . . . . . . . . . . . . . . . . . 3.3 Conditionals and Additional Structure on Language and Truth Values .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3.3.1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3.3.2 Operations on Language and Truth Values .. . . . . . . . . . . . . . . . . . 3.3.3 Operations on Language Elements and Truth Values Within One Language .. . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3.3.4 Operations on Several Languages . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3.3.5 Operations on Definable Model Sets . . . . .. . . . . . . . . . . . . . . . . . . . 3.3.6 Softening Concepts .. . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3.3.7 Aspects of Modularity and Independence in Defeasible Inheritance .. . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .
85 85 86 89 89 90 91 91 93 93 93 93 95 96 96 97
Contents
3.4
vii
Representation for General Revision, Update, and Counterfactuals.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3.4.1 Importance of Theory Revision for General Structures, Reactivity, and Its Solution .. .. . . . . . . . . . . . . . . . . . . . 3.4.2 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3.4.3 Semantic Representation for Generalized Distance-Based Theory Revision . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3.4.4 Semantic Representation for Generalized Update and Counterfactuals . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3.4.5 Syntactic Representation for Generalized Revision, Update, Counterfactuals . . . . . . .. . . . . . . . . . . . . . . . . . . .
4 Monotone and Antitone Semantic and Syntactic Interpolation . . . . . . . . 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4.1.1 Overview.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4.1.2 Problem and Method . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4.1.3 Monotone and Antitone Semantic and Syntactic Interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4.2 Monotone and Antitone Semantic Interpolation .. . . . . . . . . . . . . . . . . . . . 4.2.1 The Two-Valued Case . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4.2.2 The Many-Valued Case . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4.3 The Interval of Interpolants in Monotonic or Antitonic Logics . . . . . 4.3.1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4.3.2 Examples and a Simple Fact . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4.3.3 C and (in f C and f ) as New Semantic and Syntactic Operators .. . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4.4 Monotone and Antitone Syntactic Interpolation .. . . . . . . . . . . . . . . . . . . . 4.4.1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4.4.2 The Classical Propositional Case . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4.4.3 Finite (Intuitionistic) Goedel Logics . . . . .. . . . . . . . . . . . . . . . . . . . 5 Laws About Size and Interpolation in Non-monotonic Logics.. . . . . . . . . 5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.1.1 A Succinct Description of Our Main Ideas and Results in This Chapter .. . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.1.2 Various Concepts of Size and Non-monotonic Logics . . . . . . 5.1.3 Additive and Multiplicative Laws About Size . . . . . . . . . . . . . . . 5.1.4 Interpolation and Size . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.1.5 Hamming Relations and Size . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.1.6 Equilibrium Logic .. . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.1.7 Interpolation for Revision and Argumentation .. . . . . . . . . . . . . . 5.1.8 Language Change to Obtain Products .. . .. . . . . . . . . . . . . . . . . . . . 5.2 Laws About Size . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.2.1 Additive Laws About Size . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.2.2 Multiplicative Laws About Size . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.2.3 Hamming Relations and Distances . . . . . . .. . . . . . . . . . . . . . . . . . . .
99 99 100 101 106 111 113 113 113 114 115 119 119 121 124 124 125 126 133 133 134 135 153 153 153 153 155 156 158 159 160 160 160 160 163 172
viii
Contents
5.3
5.4 5.5
5.2.4 Summary of Properties .. . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.2.5 Language Change in Classical and Non-monotonic Logic .. . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . Semantic Interpolation for Non-monotonic Logic . . . . . . . . . . . . . . . . . . . 5.3.1 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.3.2 Interpolation of the Form j ˛ ` . . . .. . . . . . . . . . . . . . . . . . . . 5.3.3 Interpolation of the Form ` ˛ j . . . .. . . . . . . . . . . . . . . . . . . . 5.3.4 Interpolation of the Form j ˛ j . . .. . . . . . . . . . . . . . . . . . . . 5.3.5 Interpolation for Distance-Based Revision .. . . . . . . . . . . . . . . . . . 5.3.6 The Equilibrium Logic EQ. . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . Context and Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . Interpolation for Argumentation . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .
179 179 180 180 181 183 184 188 188 193 194
6 Neighbourhood Semantics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6.1.1 Defining Neighbourhoods.. . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6.1.2 Additional Requirements.. . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6.1.3 Connections Between the Various Properties . . . . . . . . . . . . . . . . 6.1.4 Various Uses of Neighbourhood Semantics .. . . . . . . . . . . . . . . . . 6.2 Detailed Overview .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6.2.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6.2.2 Tools to Define Neighbourhoods . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6.2.3 Additional Requirements.. . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6.2.4 Interpretation of the Neighbourhoods . . . .. . . . . . . . . . . . . . . . . . . . 6.2.5 Overview of the Different Lines of Reasoning.. . . . . . . . . . . . . . 6.2.6 Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6.3 Tools and Requirements for Neighbourhoods and How to Obtain Them . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6.3.1 Tools to Define Neighbourhoods . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6.3.2 Obtaining Such Tools . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6.3.3 Additional Requirements for Neighbourhoods . . . . . . . . . . . . . . 6.3.4 Connections Between the Various Concepts . . . . . . . . . . . . . . . . . 6.4 Neighbourhoods in Deontic and Default Logic . .. . . . . . . . . . . . . . . . . . . . 6.4.1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6.4.2 Two Important Examples for Deontic Logic .. . . . . . . . . . . . . . . . 6.4.3 Neighbourhoods for Deontic Systems . . . .. . . . . . . . . . . . . . . . . . . .
197 197 197 198 198 198 198 199 199 200 202 202 203
7 Conclusion and Outlook.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 7.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 7.1.1 Semantic and Syntactic Interpolation . . . .. . . . . . . . . . . . . . . . . . . . 7.1.2 Independence and Interpolation for Monotonic Logic . . . . . . 7.1.3 Independence and Interpolation for Non-monotonic Logic .. . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 7.1.4 Neighbourhood Semantics . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .
217 217 217 217
203 203 207 207 209 215 215 215 216
218 218
Contents
7.2
ix
Outlook .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 219 7.2.1 The Dynamics of Reasoning .. . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 219 7.2.2 A Revision of Basic Concepts of Logic: Justification . . . . . . . 219
References .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 223 Index . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 227
•
Chapter 1
Introduction
Unless stated otherwise, we work in propositional logic.
1.1 The Main Subjects of This Book This text centers around the following main subjects: (1) The concept of modularity and independence in • classical logic and • non-monotonic and other non-classical logic and the consequences on • (syntactic and semantical) interpolation and • language change. In particular, we will show the connection between interpolation for nonmonotonic logic and manipulation of an abstract notion of size. Modularity is, for us, and essentially, the ability to put partial results achieved independently together for a global result. (2) A uniform picture of conditionals, including • many-valued logics and • structure on the language elements themselves (in contrast to structure on the model set) and on the truth value set. (3) Neighbourhood semantics, their connection to independence, and their common points and differences for various logics, e.g., • for defaults and deontic logic, • for the limit version of preferential logics and • for general approximation. D.M. Gabbay and K. Schlechta, Conditionals and Modularity in General Logics, Cognitive Technologies, DOI 10.1007/978-3-642-19068-1 1, © Springer-Verlag Berlin Heidelberg 2011
1
2
1 Introduction
1.1.1 An Example We present now a very simple example which will illustrate some of our main ideas about modularity and interpolation. Example 1.1.1. Take a propositional language with three variables, p; q; r: Models are sequences such as hp; q; ri and hp; q; :ri: We will write 1 2 for the concatenation of two (sub)sequences, e.g., hp; q; ri D hpihq; ri: We will often identify hpi with p: will stand for the product, e.g., fhp; q; ri; hp; q; :rig D fhp; qigfr; :rg: (1) Consider classical logic. We have for WD p ^ q; WD q _ r; ˆ : The set of models of p ^ q; M.p ^ q/; is fhp; qig fr; :rg D fpg fqg fr; :rg: M.q _ r/ is fp; :pg fhq; ri; hq; :ri; h:q; rigI we can write this for clarity as .fp; :pg fqg fr; :rg/ [ .fp; :pg fh:q; rig/: The model sets are constructed in a modular way. It suffices to know the model sets in the sublanguage of the formula, and then we multiply with a fixed set, the set of all possibilities, fr; :rg in the first case, fp; :pg in the second case. This is all trivial, but important. Consider now interpolation. Let ˛ WD q; so, of course, ˆ ˛ ˆ ; ˛ is defined in the common language, and we are done. Why? Write M.˛/ as fp; :pg fqg fr; :rg: Obviously, M./ M.˛/; as the first component of M./; fpg; is a subset of the first component of M.˛/: This holds by the modular definition of M.˛/I as ˛ does not contain p; the first component is the set of all possibilities. Consider now M.˛/ M. /: Here, as did not contain r; already all relevant possibilities had to be models of I it suffices to consider the first part of the union, fp; :pg fqg fr; :rg; to see that this is true. Closer inspection shows that we only used the definition of model sets, and model set inclusion of the consequence relation, in our argument. As a matter of fact all monotone or antitone logics, even those with more than two truth values, have this semantical interpolation property, as we will show in this book; see Proposition 4.2.3 (page 122). Why is this only a semantical interpolation, and not necessarily also a syntactic interpolation? In general, the model set of will have the form † …J 00 ; where † is a set of sequences, each giving a value to all propositional variables in the sublanguage L0 used by ; and …J 00 is the set of all such sequences for the rest of the language, J 00 D L L0 ; if L is the whole language. Semantic interpolation is now to “cut off” † and replace the cut part by another product: Let J 0 L0 be the common language, J be the part of the language occuring only in ; and for each 2 † J; the restriction of to J; † J be the set of all those restrictions; then the interpolant is …J .† J 0 / …J 00 : In general, it is not at all clear that this set is the model set of any formula or set of formulas of the language. This depends on the expressivity of the language. So semantical interpolation may exist, but the interpolant need not be definable; so syntactic interpolation will fail.
1.1 The Main Subjects of This Book
3
(2) Consider a non-monotonic logic defined by a preference relation on the model set of this language; j iff in all minimal models of ; holds. (A model is minimal iff there is no smaller model in the model set.) Let WD p ^ q; WD q ^ r: We define two orders resulting in j I the first will not have interpolation, the second will have interpolation. (2.1) The first ordering: h:p; q; :ri h:p; q; ri hp; q; ri hp; q; :ri and hp; q; :ri m for all other models m: We then close this relation under transitivity. We then have p ^ q j q ^ r; as hp; q; ri hp; q; :ri: We have four candidates for interpolation, which are definable by q only: FALSE, TRUE, q; :q: They all fail, as we show now. p ^ q j6 FALSE; TRUE j6 q ^ r (in fact, TRUE j q ^ :r; as h:p; q; :ri is the smallest model of all, q j q ^ :r; and :q j6 q ^ r; the latter as the :q-models are not ordered among each other. Consequently, we have no interpolation. The crucial observation here is that the p ^ q-models behave differently from the :p ^ q-models, the first prefer r-models, and the second prefer :r-models. We do not have a relation between r and :r; which can be used as a module to construct the global relation. (2.2) The second ordering: In this ordering, we basically prefer r to :r in a homogenous way: For any two models m and m0 ; set m m0 iff m.p/ D m0 .p/; m.q/ D m0 .q/; but m.r/ D TRUE; m0 .r/ D FALSE: Again, we have p ^ q j q ^ r; but also q j q ^ r; so we have (semantical) interpolation for p ^ q j q ^ rI the interpolant ˛ is q: The important property is that we can define the order componentwise: Let D 0 00 ; D 0 00 I then iff 0 0 and 00 00 : Thus, independent of the particular form of 0 and 0 ; if 0 0 ; then the comparison between 00 and 00 will give the result, so we obtain the global comparison by looking at the components. Here, the additional structure, the relation, is also defined in a modular way, this is why we have semantical interpolation. (The problem of syntactic interpolation is the same in the monotone and the non-monotonic cases.) We turn to a more abstract view of the non-monotonic case. The choice function .X / WD fx 2 X W x is -minimal g defines a principal filter on model sets: A X is big or in the filter iff .X / A: The important property is now that for X D X 0 X 00 .X / D .X 0 / .X 00 /; so we can obtain .X / by simple multiplication of .X 0 / and .X 00 / from the components X 0 and X 00 : We can capture this by saying: “big big D big”, and have an (abstract) multiplication law for an abstract size. This is important for the following reason: Many laws for non-monotonic logics have a systematic and concise description as additive laws about abstract size. Moreoever, the law of Rational Monotony can be seen as a property of independence, and is also best expressed as a property of abstract multiplication. Thus, we have connected interpolation for non-monotonic logic to a general abstract picture of these logics.
4
1 Introduction Non-monotonic logic
Monotonic logic
Structure on language, truth values
Hamming relation, Hamming distance
Laws about size
Semantical interpolation
Independence
Classical semantics
Modularity in neighbourhoods
Semantical interpolation
Syntactical interpolation
Syntactical interpolation
Expressivity of language Connections between main concepts
Diagram 1.1.1
(3) There is a second, equally important, way to consider this example, and independence: language change. (We did not include this in Diagram 1.1.1 (page 4), as the diagram might have become too complicated to read.) (3.1) Consider the classical logic side. As stated above, we obtain the model of, say, by multiplying the models in the sublanguage with the full product of the rest: M.p ^ q/ D fhp; qig fr; :rg: It suffices to know the value in the sublanguage to obtain the value in the full language; language change has no effect. Again, this sounds utterly trivial, but it is not; it is a profound conceptual property, and the whole idea of tables of truth values is based on it. (Unfortunately, this is usually not told to beginners, so they learn a technique without understanding that there i s a problem, and why it is solved.) (3.2) Consider the non-monotonic side. In the first ordering, we cannot say “r is better than :r”, as suggested by hp; q; ri hp; q; :ri; and this carries
1.1 The Main Subjects of This Book
5
over to bigger languages. This is not true in the first ordering, as, e.g., h:p; q; :ri h:p; q; ri shows. It is true in the second ordering; we look at r=:r; and can then construct the order in the bigger language of p; q; r: Thus in the sublanguage consisting only of r; we have TRUE j r r; but we have the same in the bigger language, TRUE j p;q;r r: This may be a welcome property; it may be true, but, again, it is not trivial, and we have to be aware of it. In the non-monotonic case, we have to build the robustness under language change into the relation, if we want it. In the classical case, it was there already. We did not pursue this point in detail in this book, yet it is very important, and should be considered in further research — and borne in mind. For some remarks, see Section 5.2.2.2 (page 168) and Table 5.3 (page 178).
1.1.1.1 An Abstract Description of Both Cases in the Above Example In the classical case, we determine the model set of a formula in a modular way: We calculate the model set in the sublanguage used by the formula, and multiply this with the set of all models of the rest of the language. For the latter, we do not need the formula, only its language. So we do a separate calculation of the two components (the second calculation is trivial), and then combine the results to obtain the global result. In the non-monotonic case, if we can do the same, i.e., calculate the minimal models first in the sublanguage used by the formula, and then in the remainder of the language, and combine the results to obtain the globally minimal models, we will also have interpolation. Of course, the same will also apply to any other model choice function of some other logic, once it can be calculated in a modular way. See Diagram 1.8.1 (page 27).
1.1.2 Connections The subjects of this book are not always isolated from one another, and we will sometimes have to go back and forth between them. For instance, a structure on the language elements can itself be given in a modular way, and this then has influence on the modularity of the structure on the model set. Independence seems to be a core idea of logic, as logic is supposed to give the basics of reasoning, so we will not assume any “hidden” connections — everything which is not made explicite otherwise will be assumed to be independent. The main connections between the concepts investigated in this book are illustrated by Diagram 1.1.1 (page 4). The left-hand side concerns non-monotonic logic, the right-hand side monotonic or classical logic. The upper part concerns mainly semantics, the lower part syntax. Independence is at the core of the diagram. It can be generated by Hamming distances and relations, and can be influenced by stuctures on the language and
6
1 Introduction
the truth values. Independence is expressed by the very definition of semantics in classical logic — a formula depends only on the values of the variables occurring in it — and by suitable multiplication laws for abstract size in the non-monotonic case. Essentially, by these laws, a product of two sets is big iff the components are big. In neighbourhood semantics, independence is expressed by independent satisfiability of “close” or “good” along several criteria. Semantical interpolation is the existence of “simple” model sets X between (in the two-valued case “between” by set inclusion) the left- and the right-hand model sets: ˆ results in M./ X M. /: “Simple” means here that X is restricted only in the parts where both M./ and M. / are restricted, and otherwise is the full product, i.e., all truth values may be assumed. For the authors, a surprising result was that monotonic logic and antitonic logic always have semantical interpolation. This results from the independent definition of validity. The same is not necessarily true for full non-monotonic logic (note that we have j iff M..// M. /; where M..// M./; so we have a combined downward and upward movement: M./ ) M..// ) M. //: The reason is that abstract size (the sets of “normal” or “important” elements) need not be defined in an independent way. Semantical interpolation results also in syntactic interpolation if the language and the operators are sufficiently rich to express these semantical interpolants. This holds both for monotonic and non-monotonic logic. We now give a short introduction to these main subjects. These remarks can be found again in the various chapters, sometimes elaborated, but it seems useful to give an overall picture already in the general introduction to this book.
1.2 Main Definitions and Results For better readability, we simplify here in some places a little.
1.2.1 The Monotonic Case We distinguish between semantical and syntactic interpolation, and show that the former always exists for monotone logics, and also for many-valued semantics, and that the latter depends on the strength of the language to define suitable model sets (or model functions in the many-valued case). For f and g such that f .m/ g.m/ for all models m 2 M; we have f C .m; J / WD supff .m0 / W m0 2 M; m J D m0 J g and g .m; J / WD i nf fg.m0 / W m0 2 M; m J D m0 J g where m J is the restriction of m to a subset J of the variables L of the language and J is the set of “common variables” of f and g; such that for
1.3 Overview of This Introduction
7
all m f C .m; J / h.m/ g .m; J / is a semantical interpolant for any h; and, moreoever, f C and g are universal in the sense that they depend only on f (or g/ and J; and not on the details of the other function. See Section 4.2 (page 119) and in particular Proposition 4.2.3 (page 122). This leads us to define new operators on the semantical and syntactic sides, roughly speaking C.f; J / WD f C .m; J / and .f; J / WD f .m; J /: Their syntactic version assures then also syntactic interpolation. See Section 4.3 (page 124).
1.2.2 The Non-monotonic Case Non-monotonic logics correspond to sets of “big” subsets as follows: j iff M. ^ / (the set of models of ^ / is a big subset of M./: Alternatively, if .M.// is the set of minimal elements in a preferential structure, then .M.// is the smallest big subset of M./: Most rules for non-monotonic logics translate to laws about abstract addition of such “big” subsets, e.g., the intersection of two big subsets should be big again. Such laws were examined in detail in [GS08f] and [GS09a]. Interpolation concerns products of model sets, so it is natural that we see here laws about multiplication of such “big” subsets, e.g., . 1/.†0 †00 / D .†0 / .†00 /: (See Definition 5.2.1 (page 169).) This is a law about modularity; we can calculate .†0 †00 / componentwise. Semantic interpolation itself is an expression of modularity (in the classical case, this modularity is just the inductive definition of validity), so it is not surprising that the two meet. We show in Proposition 5.3.5 (page 184) that . 1/ entails non-monotonic interpolation of the following form: If j ; then there is an ˛ such that j ˛ j ; and ˛ is “simple”, i.e. contains only variables common to and : Property . 1/ is related to natural modularity properties of the preference relation in preferential structures; see Section 5.2.3 (page 172). Finally, the logical property of Rational Monotony did not quite find a natural place among the additive laws about “big” sets; it was felt in the past that it had more to do with independence and modularity; it finds now its natural place among the multiplicative laws.
1.3 Overview of This Introduction In the next sections, we give an introduction to the following chapters of the book. In Section 1.8 (page 26), we try to give an abstract definition of independence and modularity (limited to our purposes). We conclude this chapter with remarks on where we used previously published material (basic definitions, etc.) and with acknowledgements.
8
1 Introduction
1.4 Basic Definitions This chapter is relatively long, as we use a number of more or less involved concepts, which have to be made precise. In addition, we also want to put our work a bit more in perspective, and make it self-contained, for the convenience of the reader. Most of the material of this chapter (unless marked as “new”) was published previously; see [Sch04], [GS08b], [GS08c], [GS09a], and [GS08f]. We begin with basic algebraic and logical definitions, including in particular many laws of non-monotonic logics, in their syntactic and semantical variants, showing also the connections between both sides; see Definition 2.2.6 (page 36) and Tables 2.1 (page 38) and 2.2 (page 40). We would like to emphasize the importance of the definability preservation .dp/ property. In the infinite case, not all model sets X are definable, i.e., there need not necessarily be some formula or theory T such that X D M./; the models of ; or X D M.T /; the models of T: It is by no means evident that a model choice function ; applied to a definable model set, gives us back again a definable model set (i.e., is definability preserving, or in short, dp). If does not have this property, some representation results will not hold, which hold if is dp, and representation results become much more complicated; see [Sch04] for positive and impossibility results. In our present context, definability is again an important concept. Even if we have semantical interpolation, if language and operators are not strong enough, we cannot define the semantical interpolants, so we have semantical, but not syntactic interpolation. Examples are found in finite Goedel logics; see Section 4.4 (page 133). New operators guaranteeing the definability of particularly interesting, “universal” interpolants (see Definition 4.3.1 (page 124)) are discussed in Section 4.3 (page 124). They are intricately related to the existence of conjunctive and disjunctive normal forms, as discussed in Section 4.3.3 (page 126). We conclude this part with, to the best of our knowledge, an unpublished result: that we can define only countably many inconsistent formulas; see Example 2.2.2 (page 42). (The problem is due to D. Makinson.) We then give a detailed introduction to the basic concepts of many-valued logics, again, as readers might not be so familiar with the generalizations from two-valued to many-valued logic. In particular, the nice correspondence between two-valued functions and sets does not hold any more, so we have to work with arbitrary functions, which give values to models. We have to redefine what a definable model “set” is, and what semantical interpolation means for many-valued logic. A formula defines such a model value function f ; and we call a model value function f definable iff there is some formula such that f D f : Table 2.3 (page 47) gives an overview. We then give an introduction to preferential structures and the logic they define. These structures are among the best examined semantics for non-monotonic logics, and Chapter 5 (page 153) is based on the investigation of such structures. We first introduce the minimal variant, and then the limit variant. The first variant is the usual one, the second is needed to deal with cases where there are no minimal models, due to infinite descending chains. The limit variant was further investigated in [Sch04],
1.5 Towards a Uniform Picture of Conditionals
9
and we refer the reader to it for representation and impossibility results. An overview of representation results for the minimal variant is given in Table 2.4 (page 52). We introduce a new concept in this section on preferential structures, “bubble structures”, which, we think, present a useful tool for abstraction, and are a semantical variant of independence in preferential structures. Here, we have a global preferential structure between subsets (“bubbles”) of the model set, and a fine-scale structure inside those subsets. Seen from the outside, all elements of a bubble behave the same way, so the whole set can be treated as one element; on the inside, we see a finer structure. We include new material on many-valued preferential structures. We then go into details in the section on IBRS, introduced by D. Gabbay (see [Gab04]) and further investigated in [GS08b] and [GS08f], as they are not so much common knowledge. We also discuss here if and how the limit version of preferential structures might be applied to reactive structures. We then present theory revision, as introduced by Alchorron, Gardenfors, and Makinson; see [AGM85]. Again, we also build on previous results by the authors, when we discuss distance-based revision, introduced by Lehmann, Magidor, and Schlechta (see [LMS95], [LMS01]) and elaborated in [Sch04]. We also include a short paragraph on new material for theory revision based on many-valued logic.
1.5 Towards a Uniform Picture of Conditionals In large part, this chapter should rather be seen more as a sketch for future work than a fully elaborated theory.
1.5.1 Discussion and Classification It seems difficult to say what is not a conditional. The word “condition” suggests something like “if . . . , then . . . ”, but as the condition might be hidden in an underlying structure, and not expressed in the object language, a conditional might also be a unary operator, e.g., we may read the consequence relation j of a preferential structure as “under the condition of normality”. Moreover, as shown at the beginning of Section 3.1 (page 85), in Example 3.1.1 (page 85), it seems that one can define new conditionals ad libitum, binary, ternary, etc. Thus, it seems best to say that a conditional is just any operator. Negation, conjunction, etc., are then included, but excluded from the discussion, as we know them well. The classical connectives have a semantics in the Boolean set operators, but there are other operators, like the -functions of preferential logic, which do not correspond to any such operator, and might even not preserve definability in the
10
1 Introduction
infinite case (see Definition 2.2.4 (page 34)). It seems more promising to order conditionals by the properties of their model choice functions, e.g., by whether those functions are idempotent; see Section 3.2.2 (page 90). Many conditionals can be based on binary relations, e.g., modal conditionals on accessibility relations, preferential consequence relations on preference relations, and counterfactuals and theory revision on distance relations. Thus, it is promising to look at those relations and their properties to bring more order into the vast field of conditionals. D. Gabbay introduced reactive structures (see, e.g., [Gab04]), and added supplementary expressivity to structures based on binary relations; see [GS08b] and [GS08f]. In particular, it was shown there that we can have Cumulativity without the basic properties of preferential structures (e.g., OR). This is discussed in Section 3.2.4 (page 91).
1.5.2 Additional Structure on Language and Truth Values Normally, the language elements (propositional variables) are not structured. This is somewhat surprising, as, quite often, one variable will be more important than another. Size or weight might often be more important than colour for physical objects, etc. It is probably the mathematical tradition which was followed too closely. One of the authors gave a semantics to theory revision using a measure on language elements in [Sch91-1] and [Sch91-3], but, as far as we know, the subject has not since been treated in a larger context. The present book often works with independence of language elements (see in particular Chapters 4 (page 113) and 5 (page 153)), and Hamming type relations and distances between models, where it need not be the case that all variables have the same weight. It can also be fruitful to discuss sizes of subsets of the set of variables, so we may, e.g., neglect differences with classical logic if they concern only a “small” set of propositional variables. On the other hand, classical truth values have a natural order, FALSE < TRUE; and we will sometimes work with more than two truth values; see in particular Chapter 4 (page 113), and also Section 5.3.6 (page 188). So there is a natural question: do we also have a total order, or a Boolean order, or another order on those sets of truth values? Or: Is there a distance between truth values, so that a change from value a to value b is smaller than a change from a to c‹ There is a natural correspondence between semantical structures and truth values, which is best seen by an example: Take finite (intuitionistic) Goedel logics (see Section 4.4.3 (page 135)), say, for simplicity, with two worlds. Now, may hold nowhere, everywhere, or only in the second world (called “there”, in contrast to “here”, the first world). Thus, we can express the same situation by three truth values: 0 for nowhere, 1 for only “there”, 2 for everywhere. In Section 3.3.6 (page 96), we will make some short remarks on “softening” concepts, like neglecting “small” fragments of a language. This way, we can define, e.g., “soft” interpolation, where we need a small set of variables which are not in both formulas.
1.5 Towards a Uniform Picture of Conditionals
11
Inheritance systems, (see, e.g., [TH89, THT86, THT87, TTH91, Tou86, Sch93, Sch97, GS08e], and [GS08f]) present many aspects of independence (see Section 3.3.7 (page 97)). Thus, if two nodes are not connected by valid paths, they may have very different languages, as language elements have to be inherited; otherwise, they are undefined. In addition, a may inherit from b property c; but not property d; as we have a contradiction with d (or :d / via a different node b 0 : These are among the aspects which make them natural for commonsense reasoning, but also quite different from traditional logics.
1.5.3 Representation for General Revision, Update, and Counterfactuals Revision (see [AGM85], and the discussion in Section 2.5 (page 77)), update (see [KM90]), and counterfactuals (see [Lew73] and [Sta68]) are special forms of conditionals, which have received much interest in the artificial intelligence community. Explicitly or implicitly (see [LMS95], [LMS01]), they are based on a distance-based semantics, working with “closest worlds”. In the case of revision, we look at those worlds which are closest to the present set of worlds; in update and counterfactual, we look from each present world individually to the closest worlds, and then take the union. Obviously, the formal properties may be very different in the two cases. There are two generalizations possible, and sometimes necessary. First, “closest” worlds need not exist; there may be infinite descending chains of distances without minimal elements. Second, a distance or ranked order may force too many comparisons when two distances or elements may just simply not be comparable. We address representation problems for these generalizations: (1) We first generalize the notion of distance for revision semantics in Section 3.4.3 (page 101). We mostly consider symmetrical distances, so d.a; b/ D d.b; a/; and we work with equivalence classes Œa; b: Unfortunately, one of the main tools in [LMS01], a loop condition, does not work any more; it is too close to rankedness. We will have to work more in the spirit of general and smooth preferential structures to obtain representation. Unfortunately, revision does not allow many observations (see [LMS01], and, in particular, the impossibility results for revision (“Hamster Wheels”) discussed in [Sch04]), so all we have (see Section 3.4.3.3 (page 103)) are results which use more conditions than what can be observed from revision observations. This problem is one of principles: we showed in [GS08a] (see also [GS08f]) that Cumulativity suffices only to guarantee smoothness of the structure if the domain is closed under finite unions. But the union of two products need not be a product any more. To solve the problem, we use a technique employed in [Sch96-1], using “witnesses” to testify for the conditions.
12
1 Introduction
(2) We then discuss the limit version (when there are no minimal distances) for theory revision. (3) In Section 3.4.4 (page 106), we turn to generalized update and counterfactuals. To solve this problem, we use a technique invented in [SM94], and adapt it to our situation. The basic idea is very simple: we begin (simplified) with some world x; and arrange the other worlds around x; as x sees them, by their relative distances. Suppose we consider now one those worlds, say y: Now we arrange the worlds around y; as y sees them. If we make all the new distances smaller than the old ones, we “cannot look back”. We continue this construction unboundedly (but finitely) often. If we are a little careful, everyone will only see what he is supposed to see. In a picture, we construct galaxies around a center, then planets around suns, moons around planets, etc. The resulting construction is an A-ranked structure, as discussed in [GS08d]; see also [GS08f]. (4) In Section 3.4.5 (page 111), we discuss the corresponding syntactic conditions, using again ideas from [Sch96-1].
1.6 Interpolation 1.6.1 Introduction Chapters 4 (page 113) and 5 (page 153) are probably the core of the present book. We treat very general interpolation problems for monotone and antitone, twovalued and many-valued logics in Chapter 4 (page 113), splitting the question into two parts, “semantical interpolation” and “syntactic interpolation”, showing that the first problem, existence of semantical interpolation, has a simple and general answer, and reducing the second question, existence of syntactic interpolation, to a definability problem. For the latter, we examine the concrete example of finite Goedel logics. We can also show that the semantical problem has two “universal” solutions, which depend only on one formula and the shared variables. In Chapter 5 (page 153), we investigate three variants of semantical interpolation for non-monotonic logics, in syntactic shorthand of the types j ˛ ` ; ` ˛ j ; and j ˛ j ; where ˛ is the interpolant, and see that two variants are closely related to multiplication laws about abstract size, defining (or originating from) the non-monotonic logics. The syntactic problem is analogous to that of the monotonic case.
1.6.1.1 Background Interpolation for classical logic is well known (see [Cra57]), and there are also nonclassical logics for which interpolation has been shown, e.g., for circumscription; see [Ami02]. An extensive overview of interpolation is found in [GM05]. Chapter 1
1.6 Interpolation
13
of this book [GM05] gives a survey and a discussion and proposes that interpolation can be viewed in many different ways, and indeed 11 points of view of interpolation are discussed. The present text presents semantical interpolation; this is a new, twelfth, point of view.
1.6.2 Problem and Method In classical logic, the problem of interpolation is to find for two formulas and such that ` a “simple” formula ˛ such that ` ˛ ` : “Simple” is defined as “expressed in the common language of and ”. Working on the semantical level often has advantages: • results are robust under logically equivalent reformulations; • in many cases, the semantical level allows an easy reformulation as an algebraic problem, whose results can then be generalized to other situations; • we can split a problem into two parts: a semantical problem and the problem to find a syntactic counterpart (a definability problem); • the semantics of many non-classical logics are built on relatively few basic notions like size, distance, etc., and we can thus make connections to other problems and logics; • in the case of preferential and similar logics, the very definition of the logic is already semantical (minimal models), so it is very natural to proceed on this level. This strategy — translate to the semantical level, do the main work there, and then translate back — has proved fruitful also in the present case. Looking back at the classical interpolation problem, and translating it to the semantical level, it becomes, given M./ M. / (the models sets of and /; is there a “simple” model set A such that M./ A M. /‹ Or, more generally, given model sets X Y; is there a “simple” A such that X A Y ‹ Of course, we have to define in a natural way what “simple” means in our context. This is discussed below in Section 1.6.3.1 (page 14). Our separation of the semantical from the syntactic question pays immediately: (1) We see that monotonic (and antitonic) logics always have a semantical interpolant. But this interpolant may not be definable syntactically. (2) More precisely, we see that there is a whole interval of interpolants in above situation. (3) We also see that our analysis generalizes immediately to many-valued logics, with the same result (existence of an interval of interpolants). (4) Thus, the question remains: under what conditions does a syntactic interpolant exist? (5) In non-monotonic logics, our analysis reveals a deep connection between semantical interpolation and questions about (abstract) multiplication of (abstract) size.
14
1 Introduction
1.6.3 Monotone and Antitone Semantic and Syntactic Interpolation We consider here the semantical property of monotony or antitony, in the following sense (for the two-valued case, the generalization to the many-valued case being straightforward): Let ` be some logic such that ` implies M./ M. / (the monotone case) or M. / M./ (the antitone case). In the many-valued case, the corresponding property is that ! (or `/ respects ; the order on the truth values.
1.6.3.1 Semantic Interpolation The problem (for simplicity, for the two-valued case) reads now as follows: If M./ M. / (or, symmetrically, M. / M.//; is there a “simple” model set A; such that M./ A M. / (or M. / A M.//: Obviously, the problem is the same in both cases. We will see that such A will always exist, so all such logics have semantical interpolation (but not necessarily also syntactic interpolation). The main conceptual problem is to define “simple model set”. We have to look at the syntactic problem for guidance. Suppose is defined using propositional variables p and q; using q and r: ˛ has to be defined using only q: What are the models of ˛‹ By the very definition of validity in classical logic, neither p nor r have any influence on whether m is a model of ˛ or not. Thus, if m is a model of ˛; we can modify m on p and r; and it will still be a model. Classical models are best seen as functions from the set of propositional variables to fTRUE; FALSEg; ft; f g; etc. In this terminology, any m with m ˆ ˛ is “free” to choose the value for p and r; and we can write the model set A of ˛ as ft; f g Mq ft; f g; where Mq is the set of values for q the ˛-models may have (here: ;; ftg; ff g; ft; f g/: So, the semantical interpolation problem is to find sets which may be restricted on the common variables, but are simply the Cartesian product of the possible values for the other variables. To summarize: Let two model sets X and Y be given, where X itself is restricted on variables fp1 ; : : : ; pm g (i.e., the Cartesian product for the rest), Y is restricted on fr1 ; : : : ; rn gI then we have to find a model set A which is restricted only on fp1 ; : : : ; pm g \ fr1 ; : : : ; rn g; and such that X A Y; of course. Formulated this way, our approach, the problem and its solution, has two trivial generalizations: • for multi-valued logics we take the Cartesian product of more than just ft; f gI • may be the hypothesis and the consequence, but also vice versa; there is no direction in the problem. Thus, any result for classical logic carries over to the core part of, e.g., preferential logics.
1.6 Interpolation
15
The main result for the situation with X Y is that there is always such a semantical interpolant A as described above (see Proposition 4.2.1 (page 119) for a simple case, and Proposition 4.2.3 (page 122) for the full picture). Our proof works also for “parallel interpolation”, a concept introduced by Makinson et al. [KM07]. We explain and quote the result for the many-valued case. Suppose we have f; g W M ! V; where, intuitively, M is the set of all models, and V the set of all truth values. Thus, f and g give to each model a truth value, and, intuitively, f and g each code a model set, assigning to m TRUE iff m is in the model set, and FALSE iff not. We further assume that there is an order on the truth value set V: 8m 2 M.f .m/ g.m// corresponds now to M./ M. /; or ` in classical logic. Each model m is itself a function from L; the set of propositional variables to V: Let now J L: We say that f is insensitive to J iff the values of m on J are irrelevant: If m .L J / D m0 .L J /; i.e., m and m0 agree at least on all p 2 L J; then f .m/ D f .m0 /: This corresponds to the situation where the variable p does not occur in the formula I then M./ is insensitive to p; as the value of any m on p does not depend on whether m is a model of ; or not. We need two more definitions: Let J 0 LI then f C .m; J 0 / WD maxff .m0 / W m0 J 0 D m J 0 g and f .m; J 0 / WD mi nff .m0 / W m0 J 0 D m J 0 g: We quote now Proposition 4.2.3 (page 122): Proposition 1.6.1. Let M be rich, f; g W M ! V; f .m/ g.m/ for all m 2 M: Let L D J [ J 0 [ J 00 I let f be insensitive to J; g be insensitive to J 00 : Then f C .m; J 0 / g .m; J 0 / for all m 2 M; and any h W M J 0 ! V which is insensitive to J [ J 00 is an interpolant iff f C .m; J 0 / h.mJ mJ 0 mJ 00 / D h.m J 0 / g .m; J 0 / for all m 2 M: .h can be extended to the full M in a unique way, as it is insensitive to J [ J 00 :/ See Diagram 4.2.1 (page 123).
1.6.3.2 The Interval of Interpolants Our result has an additional reading: it defines an interval of interpolants, with lower bound f C .m; J 0 / and upper bound g .m; J 0 /: But these interpolants have a particular form. If they exist, i.e., iff f g; then f C .m; J 0 / depends only on f and J 0 (and m/; but not on g; and g .m; J 0 / depends only on g and J 0 ; not on f: Thus, they are universal, as we have to look only at one function and the set of common variables. Moreover, we will see in Section 4.3.3 (page 126) that they correspond to simple operations on the normal forms in classical logic. This is not surprising, as we “simplify” potentially complicated model sets by replacing some coordinates with simple products. The question is whether our logic allows us to express this simplification; classical logic does.
16
1 Introduction
1.6.3.3 Syntactic Interpolation Recall the problem described at the beginning of Section 1.6.3.1 (page 14). We were given M./ M. /; and were looking for a “simple” model set A such that M./ A M. /: We just saw that such an A exists, and were able to describe an interval of such an A. But we have no guarantee that any such A is definable, i.e., that there is some ˛ with A D M.˛/: In classical logic, such an ˛ exists; see, e.g., Proposition 4.4.1 (page 134)), and also Section 4.3.3 (page 126). Basically, in classical logic, f C .m; J 0 / and g .m; J 0 / correspond to simplifications of the formulas expressed in normal form; see Fact 4.3.3 (page 128) (in a different notation, which we will explain in a moment). This is not necessarily true in other logics; see Example 4.4.1 (page 145). (We find here again the importance of definability preservation, a concept introduced by one of us in [Sch92].) If we have projections (simplifications) (see Section 4.3 (page 124)), we also have syntactic interpolation. At present, we do not know whether this is a necessary condition for all natural operators. We can also turn the problem around, and just define suitable operators. This is done in Section 4.3.3 (page 126), in Definitions 4.3.2 (page 127) and 4.3.3 (page 127). There is a slight problem, as one of the operands is a set of propositional variables, and not a formula, as usual. One, but certainly not the only one, possibility is to take a formula (or the corresponding model set) and “extract” the “relevant” variables from it, i.e., those which cannot be replaced by a product. Assume now that f is one of the generalized model “sets”; then given f; define (1) .f " J / WD supff .m0 / W m0 2 M; m J D m0 J g (2) .f # J / WD i nf ff .m0 / W m0 2 M; m J D m0 J g (3) Š by fŠ WD f " .L R. // (4) ‹ by f‹ WD f # .L R. //: (We change here the notation from C to " and from – to # for the binary operators, taking f and J as arguments, to avoid confusion with other uses of C and –, whereas writing f C and f seemed to be very natural notations.) We then obtain for classical logic (see Fact 4.3.3 (page 128)) the following: Fact 1.6.2. Let J WD fp1;1 ; : : : ; p1;m1 ; : : : ; pn;1 ; : : : ; pn;mn g: (1) Let i WD ˙pi;1 ^ : : : ^ ˙pi;mi and i .1 ^ 1 / _ : : : _ .n ^ n /: Then " J (2) Let i WD ˙pi;1 _ : : : _ ˙pi;mi and i .1 _ 1 / ^ : : : ^ .n _ n /: Then # J
WD ˙qi;1 ^ : : : ^ ˙qi;ki ; let WD D 1 _ : : : _ n : WD ˙qi;1 _ : : : _ ˙qi;ki ; let WD D 1 ^ : : : ^ n :
In a way, these operators are natural, as they simplify definable model sets, so they can be used as a criterion of the expressive strength of a language and logic: If X is definable, and Y is in some reasonable sense simpler than X; then Y should
1.6 Interpolation
17
also be definable. If the language is not sufficiently strong, then we can introduce these operators, and have also syntactic interpolation.
1.6.3.4 Finite Goedel Logics The semantics of finite (intuitionistic) Goedel logics is a finite chain of worlds, which can also be expressed by a totally ordered set of truth values 0 : : : n (see Section 4.4.3 (page 135)). Let FALSE and TRUE be the minimal and maximal truth values. has value FALSE iff it holds nowhere, and TRUE iff it holds everywhere; it has value 1 iff it holds from world 2 onward, etc. The operators are classical ^ and _I negation : is defined by :.FALSE/ D TRUE and :.x/ D FALSE otherwise. Implication ! is defined by ! is TRUE iff (as truth values), and by the value of otherwise. More precisely, where f is the model value function of the formula W negation : is defined by f: .m/ WD
TRUE iff f .m/ D FALSE FALSE otherwise
implication ! is defined by f! .m/ WD
TRUE iff f .m/ f .m/ otherwise f .m/
see Definition 4.4.2 (page 136) in Section 4.4.3 (page 135). We show in Section 4.4.3.3 (page 145) the well-known result that such logics for three worlds (and thus four truth values) have no interpolation, whereas the corresponding logic for two worlds has interpolation. For the latter logic, we can still find a kind of normal form, though ! cannot always be reduced. At least we can avoid nested implications, which is not possible in the former logic for three worlds. We also discuss several “handmade” additional operators which allow us to define sufficiently many model sets to have syntactical interpolation; of course, we know that we have semantical interpolation. A more systematic approach was discussed above for the operators Š and ‹ :
1.6.4 Laws About Size and Interpolation in Non-monotonic Logics 1.6.4.1 Various Concepts of Size and Non-monotonic Logics A natural interpretation of the non-monotonic rule j is that the set of exceptional cases, i.e., those where holds, but not ; is a small subset of all the
18
1 Introduction
cases where holds, and the complement, i.e., the set of cases where and hold, is a big subset of all -cases. This interpretation gives an abstract semantics to non-monotonic logic, in the sense that definitions and rules are translated to rules about model sets, without any structural justification of those rules, as they are given, e.g., by preferential structures, which provide structural semantics. Yet, they are extremely useful, as they allow us to concentrate on the essentials, forgetting about syntactical reformulations of semantically equivalent formulas; the laws derived from the standard proof-theoretical rules incite to generalize and modify the ideas, and reveal deep connections but also differences. One of those insights is the connection between laws about size and (semantical) interpolation for non-monotonic logics, discussed in Chapter 5 (page 153). To put this abstract view a little more into perspective, we mention three alternative systems, also working with abstract size as a semantics for non-monotonic logics. They are explained in Section 5.1.2 (page 153): • the system of one of the authors for a first-order setting, published in [Sch90] and elaborated in [Sch95-1], • the system of S. Ben-David and R. Ben-Eliyahu, published in [BB94], • the system of N. Friedman and J. Halpern, published in [FH96]. The equivalence of the systems of [BB94] and [FH96] was shown in [Sch97-4]; see also [Sch04]. Historical remarks: Our own view of abstract size was inspired by the classical filter approach, as used in mathematical measure theory. The first time that abstract size was related to non-monotonic logics was, to our the best of knowledge, in the second author’s [Sch90] and [Sch95-1], and, independently, in [BB94]. The approach to size by partial orders is first discussed, to our the best of knowledge, by N. Friedman and J. Halpern; see [FH96]. More detailed remarks can also be found in [GS08c], [GS09a], and [GS08f]. A somewhat different approach is taken in [HM07]. Before we introduce the connection between interpolation and multiplicative laws about size, we give some comments on the laws about size themselves. 1.6.4.2 Additive and Multiplicative Laws About Size We give here a short introduction to and some examples for additive and multiplicative laws about size. A detailed overview is presented in Tables 5.1 (page 164), 5.2 (page 165), and 5.3 (page 178). (The first two tables have to be read together; they are too big to fit on one page.) They show connections and how to develop a multitude of logical rules known from non-monotonic logics by combining a small number of principles about size. We can use them as building blocks to construct the rules from. More precisely, “size” is to be read as “relative size”, since it is essential to change the base sets. In the first two tables, these principles are some basic and very natural postulates, .Opt/; .iM /; .eM I/; .eM F/; and a continuum of power of the notion of “small”,
1.6 Interpolation
19
or, dually, “big”, from .1 s/ to .< ! s/: From these, we can develop the rest except, essentially, Rational Monotony, and thus an infinity of different rules. The probably easiest way to see a connection between non-monotonic logics and abstract size is by considering preferential structures. Preferential structures define a principal filter, generated by the set of minimal elements, as follows: if j holds in such a structure, then ./ M. /; where ./ is the set of minimal elements of M./: According to our ideas, we define a principal filter F over M./ by X 2 F iff ./ X M./: Thus, M./ \ M.: / will be a “small” subset of M./: (Recall that filters contain the “big” sets, and ideals the “small” sets.) We can now go back and forth between rules on size and logical rules, e.g.: (1) The “AND” rule corresponds to the filter property (finite intersections of big subsets are still big). (2) “Right weakening” corresponds to the rule that supersets of big sets are still big. (3) It is natural, but beyond filter properties themselves, to postulate that if X is a small subset of Y; and Y Y 0 ; then X is also a small subset of Y 0 : We call such properties “coherence properties” between filters. This property corresponds to the logical rule .wOR/. (4) In the rule .CM! /; usually called Cautious Monotony, we change the base set a little when going from M.˛/ to M.˛ ^ ˇ/ (the change is small by the prerequisite ˛ j ˇ/; and still have ˛ ^ ˇ j ˇ0 ; if we had ˛ j ˇ 0 : We see here a conceptually very different use of “small”, as we now change the base set, over which the filter is defined, by a small amount. (5) The rule of Rational Monotony is the last one in the first table, and somewhat isolated there. It is better to see it as a multiplicative law, as described in the third table. It corresponds to the rule that the product of medium (i.e, neither big nor small) sets still has medium size. (For details, see Tables 5.1 (page 164), 5.2 (page 165), and 5.3 (page 178).)
1.6.4.3 Interpolation and Size The connection between non-monotonic logic and the abstract concept of size was investigated in [GS09a] (see also [GS08f]). There, we looked among other things at abstract addition of size. Here, we will show a connection to abstract multiplication of size. Our semantical approach uses decomposition of set-theoretical products. An important step is to write a set of models † as a product of some set †0 (which is a restriction of †/ and some full Cartesian product. So, when we speak about size, we will have (slightly simplified) some big subset †1 of one product …1 and some big subset †2 of another product …2 ; and will now check whether †1 †2 is a big subset of …1 …2 ; in shorthand, whether “big big D big”. Such conditions are called coherence conditions, as they do not concern the notion of size itself, but the way the sizes defined for different base sets are connected. Our main results here are Propositions 5.3.3 (page 183) and 5.3.5 (page 184). They say that if the logic under investigation is defined from a notion of size which satisfies sufficiently many
20
1 Introduction
multiplicative conditions, then this logic will have interpolation of type 3 or even 2; see Paragraph 1.6.4.3 (page 20). The condition is made precise in Definition 5.2.1 (page 169), in particular by the rule . 1/ W .X X 0 / D .X / .X 0 /: (Note that the conditions . i / and .S i / are equivalent, as shown in Proposition 5.2.1 (page 170) (for principal filters).) As an example, we repeat .S 1/: .S 1/ †0 †00 is big iff there is D 0 00 such that 0 †0 and 00 †00 are big. Moreover, we can connect these conditions about multiplication of size to conditions about the relation in preferential structures, e.g., . 1/ is roughly equivalent to .GH1/ ^ 0 0 ^ . _ 0 0 / ) 0 0 and
.GH 2/ 0 0 ) _ 0 0 :
(See Proposition 5.2.4 (page 173).) The main result is that the multiplicative size rule . 1/ entails non-monotonic interpolation of the form j ˛ j I see Proposition 5.3.5 (page 184). We take now a closer look at interpolation for non-monotonic logic.
The three variants of non-monotonic interpolation Consider preferential logic, and a rule like j : This means that ./ M. /; where ./ is the set of minimal models of : So we go from M./ to ./; and then to M. /; and, abstractly, we have M./ ./ M. /; so we have neither necessarily M./ M. / nor M./ M. /I the relation between M./ and M. / may be more complicated. Thus, we have neither the monotone nor the antitone case. For this reason, our general results for monotone or antitone logics do not hold any more. But we also see here that classical logic is used. Suppose that there is 0 ; which describes exactly ./I then we can write j 0 ` : So we can split preferential logic into a core part — going from to its minimal models — and a second part, which is just classical logic. (Similar decompositions are also natural for other non-monotonic logics.) Thus, preferential logic can be seen as a combination of two logics, the non-monotonic core and classical logic. It is thus natural to consider variants of the interpolation problem, where j denotes again preferential logic, and ` denotes as usual classical logic:
1.6 Interpolation
21
Given j ; is there “simple” ˛ such that (1) j ˛ ` , or (2) ` ˛ j , or (3) j ˛ j ? In most cases, we will only consider the semantical version, as the problems of the syntactical version are very similar to those for monotonic logics. We turn to the variants. (1) The first variant, j ˛ ` ; has a complete characterization in Proposition 5.3.2 (page 181), provided we have a suitable normal form (conjunctions of disjunctions). The condition says that the relevant variables of ./ have to be relevant for M./: (2) The second variant, ` ˛ j ; is related to very (and in many cases, too) strong conditions about size. We do not have a complete characterization, only sufficient conditions about size. The size conditions we need are (see Definition 5.2.1 (page 169)) the abovementioned . 1/, and . 2/ W .X / Y ) .X A/ Y A where X need not be a product any more. The result is given in Proposition 5.3.3 (page 183). Example 5.2.1 (page 171) shows that . 2/ seems too strong when compared to probability-defined size. We repeat this example here for the reader’s convenience. Example 1.6.1. Take a language of five propositional variables, with X 0 WD fa; b; cg; X 00 WD fd; eg: Consider the model set † WD f˙a ˙ b ˙ cde; a b c d ˙ eg; i.e. of eight models of de and two models of d: The models of de are 8/10 of all elements of †; so it is reasonable to call them a big subset of †: But its projection on X 00 is only 1/3 of †00 : So we have a potential decrease when going to the coordinates. This shows that weakening the prerequisite about X as done in . 2/ is not innocent. We should, however, note that sufficiently modular preferential relations guarantee these very strong properties of the big sets; see Section 5.2.3 (page 172). (3) We turn to the third variant, j ˛ j : This is probably the most interesting one, as (a) it is more general, as it loosens the connection with classical logic, (b) it seems more natural as a rule, and (c) it is also connected to more natural laws about size. Again, we do not have a complete characterization, only sufficient conditions about size. Here, . 1/ suffices, and we have our main result about non-monotonic semantical interpolation, Proposition 5.3.5 (page 184), that . 1/ entails interpolation of the type j ˛ j :
22
1 Introduction
Recall that Proposition 5.2.4 (page 173) shows that . 1/ is (roughly) equivalent to the relation properties .GH1/ ^ 0 0 ^ . _ 0 0 / ) 0 0 (where iff or D /; .GH 2/ 0 0 ) _ 0 0 of a preferential relation. ..GH 2/ means that some compensation is possible, e.g., might be the case, but 0 0 wins in the end, so 0 0 :/ There need not always be a semantical interpolation for the third variant; this is shown in Example 5.3.1 (page 181). So we see that, roughly, semantical interpolation for non-monotonic logics works when abstract size is defined in a modular way — and we find independence again. In a way, this is not surprising, as we use an independent definition of validity for interpolation in classical logic, and we use an independent definition of additional structure (relations or size) for interpolation in non-monotonic logic. 1.6.4.4 Hamming Relations and Size As preferential relations are determined by a relation, and give rise to abstract notions of size and their manipulation, it is natural to take a close look at the corresponding properties of the relation. We already gave a few examples in the preceding sections, so we can be concise here. Our main definitions and results on this subject are to be found in Section 5.2.3 (page 172), where we also discuss distances with similar properties. It is not surprising that we find various types of Hamming relations and distances in this context, as they are, by definition, modular. Neither is it surprising that we see them again in Chapter 6 (page 197), as we are interested there in independent ways to define neighbourhoods. Basically, these relations and distances come in two flavours, the set and the counting variant. This is perhaps best illustrated by the Hamming distance of two sequences of finite, equal length. We can define the distance by the set of arguments where they differ, or by the cardinality of this set. The first results in possibly incomparable distances, the second allows “compensation”; the difference in one argument can be compensated for by equality in another argument. For definitions and results, including those connecting them to notions of size, see Section 5.2.3 (page 172) in particular Definition 5.2.2 (page 172). We show in Proposition 5.2.4 (page 173) that (smooth) Hamming relations generate our size conditions when size is defined as above from a relation (the set of preferred elements generates the principal filter). Thus, Hamming relations determine logics which have interpolation; see Corollary 5.3.4 (page 184). (We define Hamming relations twice, in Sections 5.2.3 (page 172) and in 6.3.1.3 (page 205); their uses and definitions differ slightly.)
1.6 Interpolation
23
1.6.4.5 Equilibrium Logic Equilibrium logic, due to D. Pearce, A. Valverde (see [PV09] for motivation and further discussion), is based on the three-valued finite Goedel logic, also called HT logic, HT for “here and there”. Our results are presented in Section 5.3.6 (page 188). Equilibrium logic (EQ) is defined by a choice function on the model set. First, the models have to be “total”; no variable of the language may have 1 as value, i.e., no variable holds only “there”. Second, if m m0 ; then m is considered better, and m0 discarded, where m m0 iff m and m0 give value 0 to the same variables, and m gives value 2 to strictly less (as subset) variables than m0 does. We can define equilibrium logic by a preferential relation (taking care also of the first condition), but it is not smooth. Thus, our general results from the beginning of this section will not hold, and we have to work with “hand-knitted” solutions. We first show that equilibrium logic has no interpolation of the form ` ˛ j or j ˛ ` ; then that is has interpolation of the form j ˛ j ; and that the interpolant is also definable, i.e., equilibrium logic has semantical and syntactic interpolation of this form. Essentially, semantical interpolation is due to the fact that the preference relation is defined in a modular way, using individual variables — as always when we have interpolation. 1.6.4.6 Interpolation for Revision and Argumentation We have a short and simple result (Lemma 5.3.6 (page 188)) for interpolation in AGM revision. Unfortunately, we need the variables from both sides of the revision operator, as can easily be seen by revising with TRUE. The reader is referred to Section 5.3.5 (page 188) for details. Somewhat surprisingly, we also have an interpolation result for one form of argumentation, where we consider the set of arguments for a statement as the truth value of that statement. As we have maximum (set union), we have the lower bound used in Proposition 4.2.3 (page 122) for the monotonic case, and can show Fact 5.5.3 (page 195). See Section 5.5 (page 194) for details.
1.6.4.7 Language Change to Obtain Products To achieve interpolation and other results of independence, we often need to write a set of models as a nontrivial product. Sometimes, this is impossible, but an equivalent reformulation of the language can solve the problem. As this might be interesting also for the non-specialists, we repeat Example 5.2.5 (page 179) here: Example 1.6.2. Consider p D 3; and let abc; a:bc; a:b:c; :abc; :a:b:c; :ab:c be the 6 D 2 3 positive cases, ab:c; :a:bc the negative ones. (It is a coincidence that we can factorize positive and negative cases — probably iff one of the factors is the full product, here 2, though it could also be 4, etc.)
24
1 Introduction
We divide the cases by three new variables, grouping them together into positive and negative cases. a0 is indifferent; we want this to be the independent factor, the negative ones will be put into :b 0 :c 0 : The procedure has to be made precise still. In the following, (n) is negative. Let a0 code the set abc; a:bc; a:b:c; ab:c (n). Let :a0 code :a:bc (n), :abc; :a:b:c; :ab:c: Let b 0 code abc; a:bc; :a:b:c; :ab:c: Let :b 0 code a:b:c; ab:c (n), :a:bc (n), :abc: Let c 0 code abc; a:b:c; :abc; :a:b:c: Let :c 0 code a:bc; ab:c (n), :a:bc (n), :ab:c: Then the six positive instances are fa0 ; :a0 g fb 0 c 0 ; b 0 :c 0 ; :b 0 c 0 g; the negative ones fa0 ; :a0 g f:b 0 :c 0 g: As we have three new variables, we code again all possible cases, so expressivity is the same. Crucial here is that 6 D 3 2; so we can just rearrange the six models in a different way; see Fact 5.2.9 (page 179). A similar result holds for the non-monotonic case, where the structure must be possible; we can then redefine the language. All details are to be found in Section 5.2.5 (page 179).
1.6.5 Summary We use our proven strategy of “divide et impera”, and transform the problem first into a semantical question, and then into a purely algebraic one: • Classical and basic non-monotonic logic (looking for the sharpest consequence) have surprisingly the same answer; problems show up with definability when going back to the syntactical question. • Thus, we separate algebraic from logical questions, and we see that there are logics with algebraic interpolation, but without logical interpolation, as the necessary sets of models are not definable in the language. This opens the way to make the language richer in order to obtain interpolation, when so desired. • Full non-monotonic logic is more complicated, and finds a partial answer using the concept of size and a novel manipulation of it, justified by certain modular relations. • Finally, our approach also has the advantage of short and elementary proofs.
1.7 Neighbourhood Semantics Neighbourhood semantics (see Chapter 6 (page 197)), probably first introduced by D. Scott and R. Montague in [Sco70] and [Mon70], and used for deontic logic by O. Pacheco in [Pac07], seem to be useful for many logics:
1.7 Neighbourhood Semantics
25
(1) in preferential logics, they describe the limit variant, where we consider neighbourhoods of an ideal, usually inexistent, situation, (2) in approximative reasoning, they describe the approximations to the final result, (3) in deontic and default logic, they describe the “good” situations, i.e., deontically acceptable, or where defaults have fired. Neighbourhood semantics are used when the “ideal” situation does not exist (e.g., preferential systems without minimal elements), or is too difficult to obtain (e.g., “perfect” deontic states).
1.7.1 Defining Neighbourhoods Neighbourhoods can be defined in various ways: • by algebraic systems, such as unions of intersections of certain sets (but not complements), • quality relations, which say that some points are better than others, carrying over to sets of points, • distance relations, which measure the distance to the perhaps inexistent ideal points. The relations and distances may be given by the underlying structure, e.g., in preferential structures, or they can be defined in a natural way, e.g., from a system of sets, as in deontic logic or default logic. In these cases, we can define a distance between two points by the number or set of deontic requirements or default rules which one point satisfies, but not the other. A quality relation is defined in a similar way: a point is better if it satisfies more requirements or rules.
1.7.2 Additional Requirements With these tools, we can define properties neighbourhoods should have. E.g., we may require them to be downward closed, i.e., if x 2 N; where N is a neighbourhood, y x; y is better than x; then y should also be in N: This is a property we will certainly require in neighbourhood semantics for preferential structures (in the limit version). For these structures, we will also require that for every x 62 N; there should be some y 2 N with y x: We may also require that, if x 2 N; y 62 N; and y is in some aspect better than x; then there must be z 2 N; which is better than both, so we have some kind of “ceteris paribus” improvement.
1.7.3 Connections Between the Various Properties There is a multitude of possible definitions (via distances, relations, set systems), and properties, so it is not surprising that one can investigate a multitude of
26
1 Introduction
connections between the different possible definitions of neighbourhoods. We cannot cover all possible connections, so we compare only a few cases, and the reader is invited to complete the picture for the cases which interest him. The connections we examined are presented in Section 6.3.4 (page 209).
1.7.4 Various Uses of Neighbourhood Semantics We also distinguish the different uses of the systems of sets thus characterized as neighbourhoods: we can look at all formulas which hold in (all or some) such sets (as in neighbourhood semantics for preferential logics), or at the formulas which exactly describe them. The latter avoids the infamous Ross paradox of deontic logic. This distinction is simple, but basic, and probably did not receive the attention it deserves in the literature.
1.8 An Abstract View on Modularity and Independence 1.8.1 Introduction We see independence and modularity in many situations. Roughly, it means that we can use components of logic as if they were building blocks that could be put together. The big picture is not more than the elements. One example, which seems so natural that it is hardly ever mentioned, is validity in classical propositional logic, as already discussed. The validity of p in a model does not depend on the values a model assigns to other propositional variables. By induction, this property carries over to more complicated formulas. Consequently, the validity of a formula does not depend on the language: it suffices to know the values for the fragment in which is formulated to decide if holds or not. This is evident, but very important; it justifies what we call “semantical interpolation”: Semantic interpolation will always hold for monotone or antitone logics. It does not follow that the language is sufficiently rich to describe such an interpolant. If this is the case we will also have “syntactic interpolation”. Syntactic interpolation can be guaranteed by the existence of suitable normal forms, which allow us to treat model subsets independently. For preferential non-monotonic logic, we see conditions for the resulting abstract notion of size and its multiplication, which guarantee semantical interpolation also for those logics. Natural conditions for the preference relation result in such properties of abstract size. Independence is also at the basis of an approach to theory revision due to Parikh and his co-authors; see [CP00]. Again, natural conditions on a distance relation result in such independent ways of revision. The rule of Rational Monotony (see Table 2.2 (page 40)) can also be seen as independence: we can “cut up” the domain, and the same rules will still hold in the fragments.
1.8 An Abstract View on Modularity and Independence Σ1
27
f(Σ1)
f(Σ1 ◦ Σ2)= f(Σ1) ◦ f(Σ2)
Σ1 ◦ Σ2
Σ2
f(Σ2)
Note that ◦ and ◦ might be different Independence
Diagram 1.8.1
1.8.2 Abstract Definition of Independence The right notion of independence in our context seems to be as follows: We have compositions ı and ı0 ; and an operation f: We can calculate f .†1 ı †2 / from f .†1 / and f .†2 /; but also conversely, given f .†1 ı †2 / we can calculate f .†1 / and f .†2 /: Of course, in other contexts, other notions of independence might be adequate. More precisely: Definition 1.8.1. Let f W D ! C be any function from domain D to codomain C: Let ı be a “composition function” ı W D D ! DI likewise for ı0 W C C ! C: We say that hf; ı; ı0 i are independent iff for any †i 2 D (1) f .†1 ı †2 / D f .†1 / ı0 f .†2 /; (2) we can recover f .†i / from f .†1 ı †2 /; provided we know how †1 ı †2 splits into the †i ; without using f again.
1.8.2.1 Discussion (1) Ranked structures satisfy it: Let ı D ı0 D [: Let f be the minimal model operator of preferential logic. Let X; Y X [ Y have (at least) medium size, i.e. X \ .X [ Y / ¤ ;; Y \ .X [ Y / ¤ ;; (see Section 5.2.1.1 (page 161)). Then .X [ Y / D .X / [ .Y /; and .X / D .X [ Y / \ X; .Y / D .X [ Y / \ Y: (2) Consistent classical formulas and their interpretations satisfy it: Let ı be a conjunction in the composed language, ı0 be a model set intersection, f ./ D M./: Let ; be classical formulas, defined on disjoint language
28
1 Introduction
fragments L; L0 of some language L00 : Then f . ^ / D M./ \ M. /; and M./ is the projection of M./ \ M. / onto the (models of) language LI likewise for M. /: This is due to the way validity is defined, using only variables which occur in the formula. As a consequence, monotonic logic has semantical interpolation; see [GS09c] and Section 5.3.6.1 (page 188). The definition of being insensitive is justified by this modularity. (3) It does not hold for inconsistent classical formulas: We cannot recover M.a ^ :a/ and M.b/ from M.a ^:a ^b/; as we do not know where the inconsistency came from. The basic reason is trivial: One empty factor suffices to make the whole product empty, and we do not know which factor was the culprit. See Section 1.8.3.3 (page 29) for the discussion of a remedy. (4) Preferential logic satisfies it under certain conditions: If .X Y / D .X / .Y / holds for model products and j; then it holds by definition. An important consequence is that such a logic has interpolation of the form j ı jI see Section 5.3.4 (page 184). (5) Modular revision a la Parikh (see [CP00]) is based on a similar idea.
1.8.3 Other Aspects of Independence 1.8.3.1 Existence of Normal Forms We may see the existence of conjunctive and disjunctive normal forms as a form of independence: A formula may be split into elementary parts, which are then put together by the standard operations of inf .^/ and sup ._/; resulting immediately in the existence of syntactic interpolation, as both the upper and lower limits of interpolation are definable. Note that higher finite Goedel logics do not allow these operations, basically as we cannot always decompose nested intuitionistic implication.
1.8.3.2 Language Change Independence of language fragments gives us the following perspectives: (1) It makes independent and parallel treatment of fragments possible, and offers thus efficient treatment in applications (descriptive logics, etc.): Consider X D X 0 [X 00 ; where X 0 ; X 00 are disjoint. Suppose size is calculated independently, in the following sense: Let Y X I then Z Y is big iff Z \ X 0 Y \ X 0 and Z \ X 00 Y \ X 00 both are big. We can then calculate size independently.
1.8 An Abstract View on Modularity and Independence
29
(2) It results in new rules similar to the classical ones like AND, OR, Cumulativity, etc. We can thus obtain postulates about reasonable behaviour, but also classification by those rules; see Table 5.3 (page 178), Scenario 2, Logical property. (3) It sheds light on notions like “ceteris paribus”, which we saw in the context of obligations; see [GS08g] and Definition 6.3.11 (page 208). (4) It clarifies notions like “normal with respect to ; but not ”; see [GS08e] and [GS08f]. (5) It helps us understand inheritance diagrams where arrows make other information accessible, and we need an underlying mechanism to combine bits of information, given in different languages; see again [GS08e] and [GS08f].
1.8.3.3 A Relevance Problem Consider the formula WD a ^ :a ^ b: Then M./ D ;: But we cannot recover where the problem came from (it might have come from b ^ :b/; and this results in the EFQ rule. We now discuss one, purely algebraic, approach to remedy. Consider three valued models, with a new value b for both, in addition to t and f: The above formula would then have the model m.a/ D b; m.b/ D t: So there is a model, EFQ fails, and we can recover the culprit. To have the usual behaviour of ^ as intersection, it might be good to change the definition so that m.x/ D b is always a model. Then M.b/ D fm.b/ D t; m0 .b/ D bg; M.:b/ D fm.b/ D f; m0 .b/ D bg; and M.b ^ :b/ D fm0 .b/ D bg: It is not yet clear which version to choose, and we have no syntactic characterization.
1.8.3.4 Small Subspaces When considering small subsets in non-monotonic logic, we neglect small subsets of models. What is the analogue when considering small subspaces, i.e., when J D J 0 [ J 00 ; with J 00 small in J in the sense of non-monotonic logic? It is perhaps easiest to consider the relation-based approach first. So we have an order on …J 0 and one on …J 00 ; J 00 is small, and we want to know how to construct a corresponding order on …J: Two solutions come to mind: • a less radical one: we make a lexicographic ordering, where the one on …J 0 has precedence over the one on …J 00 ; • a more radical one: we totally forget about the ordering of …J 00 ; i.e., we do as if the ordering on …J 00 were the empty set, i.e. 0 00 0 00 iff 0 0 and 00 D 00 I we call this condition forget.J 00 /: The less radical one is already covered by our relation conditions .GH /; see Definition 5.2.2 (page 172). The more radical one is probably more interesting. Suppose 0 is written in language J 0 ; 00 in language J 00 I we then have
30
1 Introduction
0 ^ 00 j
0
^
00
iff 0 j
0
and 00 `
00
:
This approach is of course the same as considering on the small coordinate only ALL as a big subset (see the lines x 1=1 x in Table 5.3 (page 178)).
1.9 Conclusion and Outlook In Section 7.2 (page 219), we argue that logics which diverge from classical logic in the sense that they allow us to conclude more or less than classical logic concludes need an additional fundamental concept, a justification. Classical logic has language and truth values, proof theory, and semantics. Here, we need more, justification, why we are allowed to conclude more or less. We have to show that the price we pay (divergence from truth) is justified, e.g., by more efficient reasoning conjectures which “pay”. We think that we need a new fundamental concept, which is on the same level as proof theory and semantics. This is an open research problem, but it seems that our tools such as abstract manipulation of abstract size are sufficient to attack it.
1.10 Previously Published Material, Acknowledgements This text builds upon previous research by the authors. To make the text selfcontained, it is therefore necessary to repeat some previously published material. We give now the parts concerned and their sources. All parts of Chapter 2 (page 31) which are not marked as new material were published in some of [Sch04], [GS08b], [GS08c], [GS09a], [GS08f]. The additive laws on abstract size (see Section 5.2.1 (page 160)) were published in [GS09a] and [GS08f]. The formal material of Chapter 6 (page 197) was already published in [GS08f]; it is put here in a wider perspective. Finally, we would like to thank D. Makinson and D. Pearce for useful comments and very interesting questions.
Chapter 2
Basic Definitions
2.1 Introduction 2.1.1 Overview of This Chapter This chapter contains basic definitions and results, sometimes slightly beyond the immediate need of this book, as we want to put our work a bit more in perspective, and make it self-contained, for the convenience of the reader. Most of the material of this chapter (unless marked as “new”) was published previously; see [Sch04,GS08b, GS08c, GS09a], and [GS08f]. We begin with basic algebraic and logical definitions, including in particular many laws of non-monotonic logics, in their syntactic and semantical variants, showing the connections between both sides; see Definition 2.2.6 (page 36) and Tables 2.1 (page 38) and 2.2 (page 40). It seems to be a little known result that even the classical operators permit an unusual interpretation in the infinite case, but we claim no originality; see Example 2.2.1 (page 36). We first recall some standard concepts, such as filters, and introduce our notations for various algebraic and logical notions. This part ends with Example 2.2.1 (page 36). We then introduce notation and rules particular to non-monotonic logics. This might be the place to recall that such logics and their propeties were constructed from two sides, the syntactical side, e.g., by D. Gabbay in [Gab85], and the semantical side, e.g., for preferential structures by Y. Shoham in [Sho87b]. Both approaches were first brought together by S. Kraus, D. Lehmann, M. Magidor in [KLM90]. The semantical versions of these rules are partly due to one of us (see, e.g., [Sch04]), and described systematically in [GS08c]; see also [GS08f]. We would like to emphasize the importance of the definability preservation (dp) property. In the infinite case, not all model sets X are definable, i.e., there is some formula or theory T such that X D M./; the models of ; or X D M.T /; the models of T: It is by no means evident that a model choice function ; applied D.M. Gabbay and K. Schlechta, Conditionals and Modularity in General Logics, Cognitive Technologies, DOI 10.1007/978-3-642-19068-1 2, © Springer-Verlag Berlin Heidelberg 2011
31
32
2 Basic Definitions
to a definable model set, gives us back again a definable model set (is definability preserving, or dp). If does not have this property, some representation results will not hold, which hold if is dp, and representation results become much more complicated; see [Sch04] for positive and for impossibility results. In our present context, definability is again an important concept. Even if we have semantical interpolation, if language and operators are not strong enough, we cannot define the semantical interpolants, so we have semantical, but not syntactic interpolation. Examples are found in finite Goedel logics; see Section 4.4 (page 133). New operators guaranteeing the definability of particularly interesting, “universal” interpolants (see Definition 4.3.1 (page 124)) are discussed in Section 4.3 (page 124). They are intricately related to the existence of conjunctive and disjunctive normal forms, as discussed in Section 4.3.3 (page 126). We conclude this part with, to our the best of knowledge, the unpublished result that we can define only countably many inconsistent formulas; see Example 2.2.2 (page 42). (The question is due to D. Makinson.) We then give a detailed introduction to the basic concepts of many-valued logics, again, as readers might not be so familiar with the generalizations from two-valued to many-valued logic. In particular, the nice correspondence between two-valued functions and sets does not hold any more, so we have to work with arbitrary functions. We have to redefine what a definable model “set” is, and what semantical interpolation means for many-valued logic. Table 2.3 (page 47) gives an overview. We then give an introduction to preferential structures and the logic they define. These structures are among the best examined semantics for non-monotonic logics, and Chapter 5 (page 153) is based on the investigation of such structures. We first introduce the minimal variant, and then the limit variant. The first variant is the ususal one, the second is needed to deal with cases where there are no minimal models, due to infinite descending chains. (The first variant was introduced by Y. Shoham in [Sho87b], the second variant by P. Siegel et al. in [BS85]. It should, however, be emphasized, that preferential models were introduced as a semantics for deontic logic long before they were investigated as a semantics for non-monotonic logic; see [Han69]). The limit variant was further investigated in [Sch04], and we refer the reader there for representation and impossibility results. An overview of representation results for the minimal variant is given in Table 2.4 (page 52). Again, a systematic view on these correspondences is presented in [GS08c] and [GS08f]. We introduce a new concept in this section on preferential structures, “bubble structures”, which, we think, present a useful tool for abstraction, and are a semantical variant of independence in preferential structures. Here, we have a global preferential structure between subsets (“bubbles”) of the model set, and a fine scale structure inside those subsets. Seen from the outside, all elements of a bubble behave the same way, so the whole set can be treated as one element; on the inside, we see a finer structure. Moreover, new material on many-valued preferential structures is included. We then go into details in the section on IBRS, introduced by D. Gabbay (see [Gab04]) and further investigated in [GS08b] and [GS08f], as they are not so
2.2 Basic Algebraic and Logical Definitions
33
much common knowledge. We also discuss here if and how the limit version of preferential structures might be applied to reactive structures. From a very abstract point of view, traditional preferential structures contain an algebraic “AND” and “OR”: we need to destroy all copies of a given element (this is the AND), but we can do this in several ways, i.e., by various sets of other elements (this is the OR). Attacks on attacks in higher preferential structures provide a “NOT”; thus, we can represent all choice functions for subsets. The necessary definitions for reactive structures and resulting higher preferential structures are quite involved, and the interested reader is referred to [GS08f] for further discussion. We included all main definitions and results for completeness’ sake. We then present theory revision, as introduced by Alchorron, Gardenfors, and Makinson; see [AGM85]. (Revising a theory T by a formula is the problem of “incorporating” into T in a way that the result will be consistent, even if contradicts T:) Again, we also build on previous results by one of the authors, when we discuss distance-based revision, introduced by Lehmann, Magidor, and Schlechta (see [LMS95, LMS01], and elaborated in [Sch04]). We also include a short paragraph on new material for theory revision based on many-valued logic.
2.2 Basic Algebraic and Logical Definitions Notation 2.2.1. We use sometimes FOL as an abbreviation for first-order logic, and NML for non-monotonic logic. To avoid LaTeX complications in bigger expressions, ‚ …„ ƒ we replace x xxxx by xxxxx.
B
Definition 2.2.1. (1) We use WD and W, to define the left-hand side by the righthand side, as in the following two examples: X WD fxg defines X as the singleton with element x: X < Y W, 8x 2 X 8y 2 Y .x < y/ extends the relation < from elements to sets. (2) We use P to denote the power set operator. S …fXi W i 2 I g WD fg W g W I ! fXi W i 2 I g; 8i 2 I:g.i / 2 Xi g is the general Cartesian product, X X 0 is the binary Cartesian product. card.X / shall denote the cardinality of X; and V the set-theoretic universe we work in — the class of all sets. Given a set of pairs X ; and a set X; we let X X WD fhx; i i 2 X W x 2 X g: (When the context is clear, we will sometimes simply write X for X X:) We will use the same notation to denote the restriction of functions and in particular of sequences to a subset of the domain. If † is a set of sequences over an index set X; and X 0 X; we will abuse notation and write † X 0 for f X 0 W 2 †g: Concatenation of sequences, e.g., of and 0 ; will be denoted by juxtaposition: 0 :
34
2 Basic Definitions
(3) A B will denote that A is a subset of B or equal to B; and A B that A is a proper subset of BI likewise for A B and A B: Given some fixed set U we work in, and X U; C .X / WD U X . (4) If Y P.X / for some X; we say that Y satisfies .\/ T iff it is closed under finite intersections, . / iff it is closed under arbitrary intersections, .[/ S iff it is closed under finite unions, . / iff it is closed under arbitrary unions, .C / iff it is closed under complementation, ./ iff it is closed under set difference. (5) We will sometimes write A D B k C for: A D B; or A D C; or A D B [ C: We make ample and tacit use of the Axiom of Choice. Definition 2.2.2. will denote the transitive closure of the relation : If FALSE holds. As such, this assumption is unsatisfactory. The limit version avoids such assumptions. It will still work in the above situation, i.e., when there are not always optimal (closest) elements, it defines what happens when we get “better and better”, i.e., approach the limit (the “best” case). We will have to define what a suitable “neighbourhood” of the best cases is; in our context, this will roughly be a set of elements which minimizes all other elements and is downward closed, i.e., contains all elements better than some x already in the set. We call such sets MISE for minimizing initial segment. We will see (Example 2.3.1 (page 56)) that this definition will not always do what we want it to do, and we will have to impose additional properties. Essentially, we want MISE sets to reflect the properties of the sets of minimal elements, if they exist. Thus, the set of minimal elements should be a special case of a MISE. But we also want MISE sets to be closed under finite intersection, to have the logical (AND) property; see again Example 2.3.1 (page 56). If our definition is such that its properties are sufficiently close to those of the ideal (the minimal elements), then we will also have the desired algebraic and logical properties, but avoid pathologies originating from the empty set (when there are no best elements) — and this is what we wanted. Of course, our definition still has to correspond to the intuition of what an approximation to the ideal case should be. Note that we will also consider limit versions and other generalizations for theory revision and update in Section 3.4 (page 99). We give now the basic definitions for the limit version of preferential and ranked preferential structures. Definition 2.3.5. (1) General preferential structures. (1.1) The version without copies: Let M WD hU; i: Define Y X U is a minimizing initial segment, or MISE, of X iff:
56
2 Basic Definitions
(a) 8x 2 X 9x 2 Y:y x; where y x stands for x y or x D y (i.e., Y is minimizing), and (b) 8y 2 Y; 8x 2 X.x y ) x 2 Y / (i.e., Y is downward closed or an initial part). (1.2) The version with copies: Let M WD hU; i be as above. Define for Y X U Y is a minimizing initial segment or MISE of X iff: (a) 8hx; i i 2 X 9hy; j i 2 Y:hy; j i hx; i i and (b) 8hy; j i 2 Y; 8hx; i i 2 X .hx; i i hy; j i ) hx; i i 2 Y /: (1.3) For X U; let ƒ.X / be the set of a MISE of X: (1.4) We say that a set X of a MISE is cofinal in another set of a MISE X 0 (for the same base set X / iff for all Y 0 2 X 0 ; there is Y 2 X ; Y Y 0 : (1.5) A MISE X is called definable iff fx W 9i:hx; i i 2 X g 2 D L : (1.6) T ˆM iff there is Y 2 ƒ.U M.T // such that Y ˆ : .U M.T / WD fhx; i i 2 U W x 2 M.T /gI if there are no copies, we simplify in the obvious way.) (2) Ranked preferential structures. In the case of ranked structures, we may assume without loss of generality that the MISE sets have a particularly simple form: For X U; A X is a MISE iff X ¤ ; and 8a 2 A8x 2 X.x a _ x?a ) x 2 A/: (A is downward and horizontally closed.) (3) Theory revision. Recall that we have a distance d on the model set, and are interested in y 2 Y which are close to X: Thus, given X; Y; we define analogously: B Y is a MISE iff (1) B ¤ ; and (2) there is a d 0 such that B WD fy 2 Y W 9x 2 X:d.x; y/ d 0 g (we could also have chosen d.x; y/ < d 0 ; but this is not important). And we define 2 T T 0 iff there is a B 2 ƒ.M.T /; M.T 0 // B ˆ : There are basic problems with the limit in general preferential structures, as we shall see now: Example 2.3.1. Let a b; a c; b d; c d (but not transitive!); then fa; bg and fa; cg are such S and S 0 ; but there is no S 00 S \ S 0 which is an initial segment. If, for instance, in a and b holds, in a and c; 0 ; then “in the limit” and 0 will hold, but not ^ 0 : This does not seem right. We should not be obliged to give up to obtain 0 : t u We will therefore require it to be closed under finite intersections, or at least that if S; S 0 are such segments, then there must be S 00 S \ S 0 which is also such a segment. We make this official. Let ƒ.X / be the set of initial segments of X I then we require:
2.3 Preferential Structures
57
.ƒ\/ If A; B 2 ƒ.X / then there is a C A \ B; C 2 ƒ.X /: To familiarize the reader with the limit version, we show two easy but important results. Fact 2.3.2. (Taken from [Sch04], Fact 3.4.3, Proposition 3.10.16, (2a) is new, but only a summary of other properties.) Let the relation be transitive. The following hold in the limit variant of general preferential structures: (1) If A 2 ƒ.Y / and A X Y; then A 2 ƒ.X /: (2) If A 2 ƒ.Y /; A X Y; and B 2 ƒ.X /; then A \ B 2 ƒ.Y /: (2a) We summarize to make finitary semantical Cumulativity evident: Let A 2 ƒ.Y /; A X Y: Then, if B 2 ƒ.Y /; A \ B 2 ƒ.X /: Conversely, if B 2 ƒ.X /; then A \ B 2 ƒ.Y /: (3) If A 2 ƒ.Y /; B 2 ƒ.X /; then there is Z A [ B Z 2 ƒ.Y [ X /: The following hold in the limit variant of ranked structures without copies, where the domain is closed under finite unions and contains all finite sets. (4) A; B 2 ƒ.X / ) A B or B A; (5) A 2 ƒ.X /; YT X; Y \ A T ¤ ; ) Y \ A 2 ƒ.Y /; (6) ƒ0 ƒ.X /; ƒ0 ¤ ; ) ƒ0 2 ƒ.X /; (7) X Y; A 2 ƒ.X / ) 9B 2 ƒ.Y /:B \ X D A: Proof. (1) trivial. (2) (2.1) A \ B is closed in Y W Let hx; i i 2 A \ B; hy; j i hx; i i; then hy; j i 2 A: If hy; j i 62 X I then hy; j i 62 A; a contradiction. So hy; j i 2 X I but then hy; j i 2 B: (2.2) A \ B minimizes Y W Let ha; i i 2 Y: (a) If ha; i i 2 A B X; then there is a hy; j i ha; i i; hy; j i 2 B: By the closure of A, hy; j i 2 A: (b) If ha; i i 62 A; then there is ha0 ; i 0 i 2 A X; ha0 ; i 0 i ha; i iI continue by (a). (2a) For the first part, by (2), A \ B 2 ƒ.Y /; so by (1), A \ B 2 ƒ.X /: The second part is just (2). (3) Let Z WD fhx; i i 2 A: :9hb; j i hx; i i:hb; j i 2 X Bg [ fhy; j i 2 B: :9ha; i i hy; j i:ha; i i 2 Y Ag; where stands for or D : (3.1) Z minimizes Y [ X W We consider Y; the consideration of X is symmetrical. (a) We first show that if ha; ki 2 AZ, then there is hy; i i 2 Z:ha; ki hy; i i: Proof: If ha; ki 2 A-Z, then there is hb; j i ha; ki; hb; j i 2 X -B. Then there is hy; i i hb; j i; hy; i i 2 B: But hy; i i 2 Z; too: If not, there would be ha 0 ; k 0 i hy; i i; ha0 ; k 0 i 2 Y AI but ha0 ; k 0 i ha; ki; contradicting closure of A. (b) If ha00 ; k 00 i 2 Y A; there is ha; ki 2 A; ha; ki ha00 ; k 00 i: If ha; ki 62 Z; continue with (a).
58
2 Basic Definitions
(3.2) Z is closed in Y [X W Let then hz; i i 2 Z; hu; ki hz; i i; hu; ki 2 Y [X: Suppose hz; i i 2 A; the case hz; i i 2 B is symmetrical. (a) hu; ki 2 Y A cannot be, by closure of A: (b) hu; ki 2 X B cannot be, as hz; i i 2 Z; and by definition of Z: (c) If hu; ki 2 A Z; then there is hv; li hu; ki; hv; li 2 X B; so hv; li hz; i i; contradicting (b). (d) If hu; ki 2 B Z, then there is hv; li hu; ki; hv; li 2 Y A; contradicting (a). (4) Suppose not, so there are a 2 A B; b 2 B A: But if a?b; a 2 B and b 2 AI similarly if a b or b a: (5) As A 2 ƒ.X / and Y X; Y \ A is downward and horizontally closed. As Y T T \ A ¤ ;; Y \ A minimizes Y: (6) T ƒ0 is downward and horizontally closed, as all A 2 ƒ0 are. As ƒ0 ¤ ;; ƒ0 minimizes X: (7) Set B WD fb 2 Y W 9a 2 A:a?b or b ag u t We have as immediate logical consequence: Fact 2.3.3. (Fact 3.4.4 of [Sch04].) If is transitive, then in the limit variant hold: (1) .AND/, (2) .OR/. Proof. Let Z be the structure. (1) Immediate by Fact 2.3.2 (page 57), (2) — set A D B: (2) Immediate by Fact 2.3.2 (page 57), (3).
t u
We also have the following: Fact 2.3.4. (Fact 3.4.5 in [Sch04].) Finite Cumulativity holds in transitive limit structures: If j ; then D ^ : See [Sch04] for a direct proof, or above Fact 2.3.2 (page 57), (2a). We repeat now (without proof) our main logical trivialization results on the limit variant of general preferential structures, Proposition 3.4.7 and Proposition 3.10.19 from [Sch04]: Proposition 2.3.5. (1) Let the relation be transitive. If we consider only formulas on the left of j ; the resulting logic of the limit version can also be generated by the minimal version of a (perhaps different) preferential structure. Moreover, this structure can be chosen smooth. (2) Let a logic j be given by the limit variant of a ranked structure without copies. Then there is a ranked structure which gives exactly the same logic, but interpreted in the minimal variant. t u
2.3 Preferential Structures
59
(The negative results for the general, not definability preserving minimal case apply also to the general limit case; see Section 5.2.3 in [Sch04] for details.)
2.3.2.1 New Material on the Limit Variant of Preferential Structures This short section contains new material on the limit variant of preferential structures; a discussion of the limit variant of higher preferential structures will be presented below; see Section 2.4.2.6 (page 74). Consider the following analogon to .PR/ .A B ) .B/ \ A .A//: Fact 2.3.6. Let be transitive, ƒ.X / the MISE systems over X: Let A B; A0 2 ƒ.A/ ) 9B 0 2 ƒ.B/:B 0 \ A A0 : Proof. Consider B 0 WD fb 2 B W b 62 A A0 and :9b 0 2 A A0 :b 0 bg: Thus, B B 0 D fb 2 B W b 2 A A0 or 9b 0 2 A A0 :b 0 bg: (1) B 0 \ A A0 W Trivial. (2) B 0 is closed in B W Let b 2 B 0 ; suppose there is b 0 2 B B 0 ; b 0 b: b 0 2 AA0 is excluded by definition, b 0 is such that 9b 00 2 A A0 :b 00 b 0 by transitivity. (3) B 0 is minimizing: Let b 2 B B 0 : If b 2 A A0 ; then there is an a 2 A0 :a b by minimization of A by A0 : We have to show that a 2 B 0 : If not, there must be b 0 2 A A0 :b 0 a; contradicting closure of A0 in A: If b is such that there is a b 0 2 A A0 :b 0 b; then there has to be an a 2 A0 such that a b 0 b; so a b by transitivity, and we continue as above. t u We have immediately the following: Corollary 2.3.7. ^ 0 [ f 0 g: Proof. Let 2 ^ 0 ; A WD M. ^ 0 /; B WD M./: So there is A0 2 ƒ.A/:A0 ˆ I so there is B 0 2 ƒ.B/:B 0 \ M. 0 / ˆ by Fact 2.3.6 (page 59), so B 0 ˆ 0 ! ; so 0 ! 2 ; so [ 0 ` : t u
2.3.3 Preferential Structures for Many-Valued Logics This section is important; it contains new material, which will later be applied to equilibrium logic; see Section 5.3.6 (page 188). We can, of course, consider for a given the set of models where has maximal truth value TRUE, and then take the minimal ones as usual. The resulting logic j
then makes j true iff the minimal models with value TRUE assign TRUE also to : See Section 5.3.6 (page 188). But this does not seem to be the adequate way. So we adapt the definition of preferential structures to the many-valued situation.
60
2 Basic Definitions
Definition 2.3.6. Let L be given with model set M: Let a binary relation be given on X ; where X is a set of pairs hm; i i; m 2 M; i some index as usual. (We use here the assumption that the truth value is independent of indices.) Let f W M ! V be given; we define .f /; the minimal models of f W
.f /.m/ WD
8 < FALSE iff 8hm; i i 2 X 9hm0 ; i 0 i hm; i i:f .m0 / f .m/ :
f .m/
otherwi se
This generalizes the idea that only models of can destroy models of : Obviously, for all v 2 V; v ¤ FALSE; fm W .f /.m/ D vg fm W f .m/ D vg: A structure is called smooth iff for all f and for all hm; i i such that there is a hm0 ; i 0 i hm; i i with f .m0 / f .m/; there is hm00 ; i 00 i hm; i i with f .m00 / f .m/; and no hn; j i hm00 ; i 00 i with f .n/ f .m00 /: A structure will be called definablity preserving iff for all f ; .f / is again the f for some : Definition 2.3.7. With these ideas, we can also define minimizing initial segments for many-valued structures in a straightforward way: F is a MISE with respect to G iff (0) F G; (1) if F .x/ ¤ 0; y x; G.x/ G.y/; then F .y/ D G.y/ (downward closure), and (2) if G.x/ ¤ 0; F .x/ D 0; then there is a y with y x; G.x/ G.y/; F .y/ D G.y/: We turn to representation questions. Example 2.3.2. This example shows that a suitable choice of truth values can destroy coherence, as it is present in two-valued preferential structures. We want essentially y x in A; A B; but y 6 x in B: The solution will be to make FB .y/ < FB .x/; but FA .y/ FA .x/; and FA FB I e.g., we set FB .x/ D 3; FB .y/ D 2; FA .x/ D FA .y/ D 2: This example leads to the following small representation result: Fact 2.3.8. Let U be the universe we work in; let W P.U / ! P.U / be a function such that (1) .X / X; (2) there is no singleton X D fxg with .X / D ;: Then there is a many-valued preferential structure X which represents : Note that no coherence conditions are necessary.
2.3 Preferential Structures
61
Proof. Let 0 be the smallest truth value, and 8x 2 U x also be a truth value, where for all x ¤ y; x; y 2 U; x?y .x; y as truth values are incomparable). Take as preference relation x y for all x; y 2 U; x ¤ y: Choose X U: Define FX .x/ WD 0 iff x 62 .X /; and FX .x/ WD x iff x 2 .X /: Then all relations x y are effective for y 62 .X /; as then FX .y/ FX .x/; so y will not be minimal. If y 2 .X /; then there is no x ¤ y; FX .y/ FX .x/: t u The above Fact 2.3.8 (page 60) largely solves the problem of finding a preferential representation for arbitrary choice functions by many-valued structures. But one might ask different questions in this context; e.g.: Suppose we have a family of pairs hF; F i of functions giving truth values to all x 2 U: Suppose 8x 2 U:F .x/ F .x/I in short, F F: Suppose for simplicity that we have a minimal element 0 of truth values, with the meaning F .x/ D 0 iff “x 62 F ” (read as a set), so we will not consider x with F .x/ D 0: Suppose further that F .x/ D 0 or F .x/ D F .x/: Then, what are the conditions on the family of hF; F i such that we can represent them by a many-valued preferential structure? The answer is not as trivial as the one to the choice function representation problem above. Consider the following: Example 2.3.3. Consider F; G with F G; F .x/ ¤ 0; F .x/ D 0; F .y/ ¤ 0; F .x/ F .y/: In this case, a relation y x is effective for F: Suppose now that also G.x/ G.y/I then y x is also effective for G: We may say roughly, not only if F G; but for x; y such that F .x/; F .y/ ¤ 0; and for the “derivatives” F 0 and G 0 ; if F 0 G 0 holds in the sense that F .x/ F .y/ ) G.x/ G.y/; then F and G must have the same coherence properties as the two-valued choice functions in order to be preferentially representable — as any relation effective for F will also be effective for G: This may lead us to consider the following brute force solution: We have a global truth value relation in the sense that for all F; G F .x/ F .y/ iff G.x/ G.y/ — apart from cases where, e.g., F .x/ D 0; as “x 62 F ”. In this case, they behave just like normal two-valued structures, but we could now just as well simply omit any relations x y when F .y/ 6 F .x/ (equivalently, for any other G/: So this leads us nowhere interesting. We will leave the problem for further research, and only add a few rudimentary remarks: (1) We may introduce new operators in order to be able to speak about the situation: (a) m ˆ OF iff for all m0 such that m0 ˆ ; F .m/ F .m0 /; (b) ˆ OF .; / iff for all m such that m ˆ and all m0 such that m0 ˆ F .m/ F .m0 /:
;
These expressions are, of course, still semi-classical, and we can replace ˆ by a certain threshold, or consider only m0 such that F .m0 / F .m/:
62
2 Basic Definitions
Then, given sufficient definability power, we can express that all models “in” F have truth values at least as good as those in F F WD F \ C F; ˆ OF .F F; F /; and use this to formulate a coherence condition: OF .F F; F / ) OG .F F; F /: (2) To be able to do some set theory, we will assume that V the set of truth values is a complete Boolean algebra, with symbols ^ (or for many arguments) for infimum, binary a b for a^-b, 0 and 1. For functions F; G; etc.Vwith values in the set V of truth values, V we define ^; _; etc. argumentwise, e.g., Fi is defined by . Fi /.x/ WD .Fi .x//: (3) As an illustration, and for no other purposes, we look at some cases of the crucial Fact 3.3.1 in [Sch04] for representation by smooth structures, which we repeat now here for easier reference, together with its proof: Fact 2.3.9. Let A, U; U 0 ; Y and all Ai be in Y: . / and .PR/ entail the following: S S (1) A D fAi W i 2 I g ! .A/ f.Ai / W i 2 I g; (2) U H.U /; and U U 0 ! H.U / H.U 0 /; (3) .U [ Y / H.U / .Y /: . /; .PR/; .C UM / entail the following: (4) U A; .A/ H.U / ! .A/ U; (5) .Y / H.U / ! Y H.U / and .U [ Y / D .U /; (6) x 2 .U /; x 2 Y .Y / ! Y 6 H.U /; (7) Y 6 H.U / ! .U [ Y / 6 H.U /: S S Proof. S (1) .A/ \ Aj .Aj / .Ai /; so by .A/ A D Ai ; .A/ .Ai /: (2) Trivial. (3) .U [ YS / H.U / .2/ .U [ Y / U ./ .U S [ Y / \ Y .PR/ .Y /: (4) .A/ D f.A/ \ X W .X / U g .PR0 / f.A \ X / W .X / U g: But if .X / U A; then by .X / X; .X / A \ X X !.C UM / .A \ X / D .X / U; so .A/ U: (5) Let .Y / H.U /I then by .U / H.U / and (1), .U [ Y / .U / [ .Y / H.U /; so by (4), .U [ Y / U and U [ Y H.U /: Moreover, .U [ Y / U U [ Y !.C UM / .U [ Y / D .U /: (6) If not, Y H.U /; so .Y / H.U /; so .U [ Y / D .U / by (5), but x 2 Y .Y / !.PR/ x 62 .U [ Y / D .U /; a contradiction. (7) .U [ Y / H.U / !.5/ U [ Y H.U /: t u We translate some properties and arguments: • (1) (a) Fi F ) F ^ Fi Fi for all i by .PR/ (but recall that .PR/ will not always hold; W see Example 2.3.2 (page 60)). (b) F F i Fi :
2.4 IBRS and Higher Preferential Structures
63
W W W Thus F D (by b/ F ^ i Fi D (distributivity) i .F ^ Fi / (by a) i Fi : • (3) We first need an analogue to X Y [ Z ) X Y Z W (a) FX FY _ FZ ) FX FY FZ : Proof: FX FY D FX ^ C FY (prerequisite) .FY ^ C FY / _ .FZ ^ C FY / D 0 _ .FZ ^ C FY / FZ : We then need (b) FX FX 0 ) FY FX 0 FY FX : Proof: FX FX 0 ) C FX 0 C FX ; so FY \ C FX 0 FY \ C FX : Thus, fU [Y fH.U / (by (2) and (b)) fU [Y fU (by . /; (a)) fU [Y ^ fY fY ; by .PR/: • (6) FY FH.U / ) FY FH.U / ) FU [Y D .FU _ FY / D FU by (5). FU [Y ^ FY D .FU _ FV / ^ FY FY by .PR/; so FU ^ FY D FU [Y ^ FY FY : Thus FU ^ FY ^ C FY D ;; contradicting the prerequisite.
2.4 IBRS and Higher Preferential Structures 2.4.1 General IBRS We first define IBRS: Definition 2.4.1. (1) An information bearing binary relation frame IBR, has the form .S; 2: This is the reason we treat them here. We will connect the interpolation problem to the existence of normal forms. The connection
136
4 Monotone and Antitone Semantic and Syntactic Interpolation
is incomplete, as we will show that suitable normal forms entail interpolation, but we do not know if this condition is necessary. We discuss these logics here, as they present a nice and simple example which shows that there need not be syntactic interpolation, though semantical interpolation exists, as we saw above. A second reason is that the finite Goedel logic with two worlds is the basis for the non-monotonic logic EQ, discussed in Section 5.3.6 (page 188). 4.4.3.1 The Definitions Definition 4.4.2. Finite intuitionistic Goedel logics with n C 1 truth values FALSE D 0 1 : : : n D TRUE are defined as follows: (1) f^ .m/ WD i nf ff .m/; f .m/g; (2) f_ .m/ WD supff .m/; f .m/g; (3) negation : is defined by
f: .m/ WD
8 < TRUE iff f .m/ D FALSE :
FALSE
otherwise
(4) implication ! is defined by 8 < TRUE iff f .m/ f .m/
f! .m/ WD
:
f .m/
otherwise
Thus, for n C 1 D 2; this is classical logic. So we assume now n 2: Definition 4.4.3. We will also consider the following additional operators: (1) J is defined by
fJ .m/ WD
8 < f .m/ :
iff f .m/ D FALSE or f .m/ D TRUE
f .m/ C 1
otherwise
The intuitive meaning is “it holds in the next moment”. (2) A is defined by
fA./ .m/ WD
8 < TRUE iff f .m/ D TRUE :
FALSE
otherwise
Thus, A is the dual of negation; we might call it affirmation.
4.4 Monotone and Antitone Syntactic Interpolation
137
(3) F is defined by
fF .m/ WD
8 < FALSE :
iff f .m/ D FALSE or f .m/ D TRUE
f .m/ C 1
otherwise
The intuitive meaning is “it begins to hold in the next moment”. (4) Z (cyclic addition of 1) is defined by
fZ .m/ WD
8 < FALSE :
iff f .m/ D TRUE
f .m/ C 1
otherwise
Note that Z is slightly different from J: We do not know if there is an intuitive meaning. To help the intuition, we give the truth tables of the basic operators for n D 3; and of :; !; J; A; F; Z for n D 4 and n D 6: b a 0 1 2
:a 2 0 0
0 1 2
0 1 2
a!b
a^b 2 2 2 0 2 2 0 1 2
0 1 2 a_b
0 0 0 0 1 1 0 1 2
:a 3 0 0 0
Ja 0 2 3 3
Aa 0 0 0 3
Fa 0 2 3 0
Za 1 2 3 0
:a 5 0 0 0 0 0
Ja 0 2 3 4 5 5
Aa 0 0 0 0 0 5
Fa 0 2 3 4 5 0
Za 1 2 3 4 5 0
2 0 0 0 2 1 0 1 2
0
1 2 3
3 0 0 0
3 3 1 1
a!b
b a 0 1 2 3 4 5
a$b 0 1 2 1 1 2 2 2 2
b a 0 1 2 3
0 1 2
3 3 3 2
3 3 3 3
0 1
2 3 4 5
5 0 0 0 0 0
5 5 5 2 2 2
a!b 5 5 1 1 1 1
5 5 5 5 3 3
5 5 5 5 5 4
5 5 5 5 5 5
4.4.3.2 Normal Forms and f C We work now towards a suitable normal form, using only ^; _; :; !; even though we cannot obtain it for n > 3: This will also indicate a way to repair those logics by introducing suitable additional operators, which allow us to obtain such normal
138
4 Monotone and Antitone Semantic and Syntactic Interpolation
Table 4.1 Neglecting a variable — Part 1 b!a a T T
:a
ab
:b F
.b ! a/ ^ :a F :b F
ac
a!c T T c
a a c a
.a ! c/ ^ a a c c
ac
a!c T T c
::a ::a ::c ::c ::a . ::c; c/
.a ! c/ ^ ::a ::a ::c c . ::c/
1;3
2;2
2;4
forms. In addition, we will see how to obtain f C .m; J / WD maxff .m0 / W m0 J D m J g from normal forms. ! cannot be eliminated by reducing it, e.g., to ^ and : as in classical logic. The main problem in obtaining normal forms is flattening nested !I this is possible for two worlds, but not for more. We then investigate “neglecting” one variable, say x; i.e., considering f C .m; L fxg/: The way to do this is described in Tables 4.1 (page 138) and 4.2 (page 139). We have the following fact: Fact 4.4.5. (1) With one variable a we can define up to semantical equivalence exactly the following six different formulas: a; :a; ::a; TRUE D a ! a; FALSE D :.a ! a/; ::a ! a: The following semantical equivalences hold: (Note: all except (14) hold also for four and six truth values, and so probably for arbitrarily many truth values, but this is not checked so far.) Triple negation can be simplified: (2) :::a $ :a: Disjunction and conjunction combine classically: (3) :.a _ b/ $ :a ^ :b (4) :.a ^ b/ $ :a _ :b (5) a ^ .b _ c/ $ .a ^ b/ _ .a ^ c/ (6) a _ .b ^ c/ $ .a _ b/ ^ .a _ c/: Implication can be eliminated from combined negation and implication: (7) :.a ! b/ $ ::a ^ :b (8) .a ! :b/ $ .:a _ :b/ (9) .:a ! b/ $ .::a _ b/: Implication can be put inside when combined with ^ and _ W (10) .a _ b ! c/ $ ..a ! c/ ^ .b ! c// (11) .a ^ b ! c/ $ ..a ! c/ _ .b ! c// (12) .a ! b ^ c/ $ ..a ! b/ ^ .a ! c//
3;2
3;1
Case 1: b c 1.1: a < b 1.2: a D b 1.3: b a c 1.4: c < a Case 2: c < b 2.1: a c 2.2: c < a < b 2.3: b a
Case 1: b c 1.1: a < b 1.2: a D b 1.3: b a c 1.4: c < a Case 2: c < b 2.1: a c 2.2: c < a < b 2.3: b a
Table 4.2 Neglecting a variable — Part 2
T c c
a a T
a!c
b!a T T T c
T c c
a a T
a T T T
T T T c
a!c
a T T T
b!a
a
a . c/ c c
a .< c/ a . c/ a . c/ c
.b ! a/ ^ .a ! c/ ^ a
a . c/ c c
a T T c
.b ! a/ ^ .a ! c/
c
c
b!c T
(continued)
4.4 Monotone and Antitone Syntactic Interpolation 139
3;4
1.4: c < a Case 2: c < b 2.1: a c 2.2: c < a < b 2.3: b a
Case 1: b c 1.1: a < b 1.2: a D b 1.3: b a c
Table 4.2 (continued) 3;3 Case 1: b c 1.1: a < b 1.2: a D b 1.3: b a c 1.4: c < a Case 2: c < b 2.1: a c 2.2: c < a < b 2.3: b a
T c c
a!c
a a T
b!a
c
T c c
T
a a T
T T T
T T T c
a T T T
a T T
a!c
b!a
F F F
:a :a :a . :c/
ac c c
::a T T
T
a ::c ::a ::c ::a ::c D ::c if a D c c ::c
::a ::a ::a
.b ! a/ ^ .a ! c/ ^ ::a
F :a D :b :a :a ^ c D F
:a :a :a . :b/ :a . :b; :c/
::a
.b ! a/ ^ .a ! c/ ^ :a
:a
c
b!c T
c
b!c T
::c
::c ::c
:b D F
:b :b
c
.b ! c/ ^ ::c ::c
F
.b ! c/ ^ :b :b
140 4 Monotone and Antitone Semantic and Syntactic Interpolation
4.4 Monotone and Antitone Syntactic Interpolation
141
(13) .a ! b _ c/ $ ..a ! b/ _ .a ! c//: Nested implication can be flattened for nesting on the right: (14) .a ! .b ! c// $ ..a ^ b ! c/ ^ .a ^ :c ! :b//: Proof. We use T for TRUE, F for FALSE. (1) The truth table for the six formulas is given by the following: a 0 1 2 ... n
a 0 1 2 ... n
:a n 0 0 ... 0
::a 0 n n ... n
T .D a ! a/ n n n ... n
F .D :.a ! a// 0 0 0 ... 0
::a ! a n 1 2 .. n
We see that the first line takes the values 0 and n; and the n 1 other lines take the vectors of values .1; : : : n/; .0; : : : ; 0/; .n; : : : ; n/; and that all combinations of the first line values and those vectors occur. Thus, we can check closure separately for the first line and the other lines, which is now trivial. (2) Trivial. .3/ C .4/ Both sides can only be T or F: (3): Suppose a b; then :.a _ b/ D T iff b D F; and :a ^ :b D T iff b D F: The case b a is symmetrical. (4): similar: :.a ^ b/ D T iff a D F and :a _ :b D T iff a D F: (5) Suppose b c: Thus .a ^ b/ _ .a ^ c/ D a ^ c: If a c; then a ^ .b _ c/ D a; else a ^ .b _ c/ D c: The case c b is symmetrical. (6) Suppose b c: Then .a _ b/ ^ .a _ c/ D a _ b: If a b; then a _ .b ^ c/ D b; else a _ .b ^ c/ D a: (7) Both sides are T or F: :.a ! b/ D T iff a > b and b D F: ::a ^ :b D T iff b D F and ::a D F: ::a D F iff a > F: (8) Again, both sides are T or F: a ! :b D T iff a :b iff b D F or a D F: (9) :a ! b D T iff a > F or b D T: If a > F; then :a ! b D T and ::a _ b D T: If a D F; then :a ! b D b; and ::a _ b D b: (10) a _ b ! c is a ! c or b ! c: If a b; then b ! c; and a ! c b ! c: The case b a is symmetrical. .11/ .13/ are similar to (10), e.g., (12): If b c; then .a ! b/ ^ c D a ! b; and .a ! b/ ^ .a ! c/ D a ! b: (14) Case 1. b c W Then :c :b; and a ! .b ! c/ D T; a ^ b ! c D T; a ^ :c ! :b D T: Case 2. b > c W So .a ! .b ! c// D .a ! c/; and :b D F: Case 2.1, a b W So a ^ b ! c D a ! c: Case 2.1.1. :c D F W So a ^ :c ! :b D T; and we are done. Case 2.1.2. :c D T W So c D F; a ! c D a ! F; and a ^ :c ! :b D a ! :b D a ! F; and we are done.
142
4 Monotone and Antitone Semantic and Syntactic Interpolation
Case 2.2. a > b W So a > b > c; a ! c D c; a ^ b ! c D b ! c D c; and :b D F: Case 2.2.1. :c D F W So a ^ :c ! :b D T; and we are done. Case 2.2.2. :c D T W So c D F: Thus a ^ :c ! :b D a ! :b D a ! F: But also a ! c D a ! F; and we are done. t u We assume now the fellowing (Assumption) Any formula of the type . ! 0 / ! ˆ containing only flat !’s. We will later show that this is true for n D 2:
is equivalent to a formula
Fact 4.4.6. Let above assumption be true. Then: Every formula can be transformed into a semantically equivalent formula the following form:
of
(1) has the form 1 _ : : : _ n (2) every i has the form i;1 ^ : : : ^ i;m (3) every i;m has one of the following forms: p; or :p; or ::p; or p ! q; where p and q are propositional variables. Note that also ! D TRUE can be replaced by :a _ ::a: Proof. The numbers refer to Fact 4.4.5 (page 138). We first push : downward, towards the interior: • :. ^ / is transformed to : _ : by (4). • :. _ / is transformed to : ^ : by (3). • :. ! / is transformed to :: ^ : by (7). We next eliminate any !
where and
are not propositional variables:
• • • •
: ! is transformed to :: _ by (9). ^ 0 ! is transformed to . ! / _ . 0 ! / by (11). _ 0 ! is transformed to . ! / ^ . 0 ! / by (10). . ! 0 / ! is transformed to flat ˆ by the assumption.
• • • •
! : is transformed to : _ : by (8). ! ^ 0 is transformed to . ! / ^ . ! 0 / by (12). ! _ 0 is transformed to . ! / _ . ! 0 / by (13). ! . ! 0 / is transformed to . ^ ! 0 / ^ . ^ : 0 ! : / by (14).
Finally, we push ^ inside: ^ . _ 0 / is transformed to . ^ / _ . ^ The exact proof is, of course, by induction.
0
/ by (5). t u
This normal form allows us to use the following facts: Fact 4.4.7. We will now work for syntactic interpolation. For this purpose, we show that if f is definable in Proposition 4.2.3 (page 122), i.e. there is with f D f ; then f C in the same Proposition is also definable. Recall that f C .m/ was defined
4.4 Monotone and Antitone Syntactic Interpolation
143
as the maximal f .m0 / for m0 J 0 D m J 0 : We use the normal form just shown to show that conjuncts and disjuncts can be treated separately. Our aim is to find a formula which characterizes the maximum. More precisely, if f D f for some ; we look for 0 such that f 0 .m/ D maxff .m0 / W m0 2 M; m J D m0 J g: First, a trivial fact, which shows that we can treat the elements of J (or L J / one after the other: maxfg.x; y/ W x 2 X; y 2 Y g D maxfmaxfg.x; y/ W x 2 X g W y 2 Y g: (Proof: The interior max on the right-hand side range over subsets of X Y; so they are all than the left-hand side. Conversely, the left-hand side max is assumed for some hx; yi; which also figures on the right-hand side. A full proof would be an induction.) Next, we show that we can treat disjunctions separately for one x 2 L; and also conjunctions, as long as x occurs only in one of the conjuncts. Again, a full proof would be by induction; we only show the crucial arguments. First, some notation: Notation 4.4.1. (1) We write m D.x/ m0 as shorthand for m .L fxg/ D m0 .L fxg/ (2) Let f W M ! V; x 2 LI then f.x/ .m/ WD maxff .m0 / W m0 2 M; m D.x/ m0 g: (3) Let f W M ! V; and .f /.x/ D f 0 for some 0 I then we write .x/ for (some such) 0 : 0 00 ; .x/ both exist, then so does .x/ ; and Fact 4.4.8. (1) If D 0 _ 00 ; and .x/ 0 00 .x/ D .x/ _ .x/ : 0 (2) If D 0 ^ 00 ; .x/ exists, and 00 does not contain x; then .x/ exists, and 0 00 .x/ D .x/ ^ : 0 00 : Proof. (1) We have to show f.x/ D f..x/ _.x/ / 0 _ 00 / .m/ D maxff 0 .m/; f 00 .m/g: By definition of validity of _; we have f..x/ .x/ .x/ .x/ 0 0 _ 00 / .m/ D maxfmaxff 0 .m / W f.x/ .m/ WD maxff .m0 / W m0 D.x/ mg; so f..x/ .x/ m0 D.x/ mg; maxff 00 .m0 / W m0 D.x/ mgg D maxfmaxff 0 .m0 /; f 00 .m0 /g W m0 D.x/ mg D (again by definition of validity of _/ maxff 0 _ 00 .m0 / W m0 D.x/ mg D maxff .m0 / W m0 D.x/ mg D f..x/ / .m/: 0 ^ 00 / : By definition of validity of ^; we have (2) We have to show f.x/ D f..x/ .x/ 0 0 00 0 00 0 00 .m/ D i nf fmaxff 0 .m / W f..x/ ^.x/ / .m/ D i nf ff.x/ .m/; f.x/ .m/g: So f..x/ ^.x/ / 0 0 0 00 m D.x/ mg; maxff 00 .m / W m D.x/ mgg D (as does not contain x/ i nf fmaxff 0 .m0 / W m0 D.x/ mg; f 00 .m/g D maxfi nf ff 0 .m0 /; f 00 .m/g W m0 D.x/ mg D (again by definition of validity of ^; and by the fact that 00 does .m/: not contain x/ maxff 0 ^ 00 .m0 / W m0 D.x/ mg D maxff .m0 / W m0 D.x/ mg D f./ t u
Thus, we can calculate disjunctions separately, and also conjunctions, as long as the latter have no variables in common. In classical logic, we are finished, as we can break down conjunctions into parts which have no variables in common. The problem here are formulas of the type a ! b; as they may have variables in
144
4 Monotone and Antitone Semantic and Syntactic Interpolation
common with other conjuncts, and, as we will see in Fact 4.4.15 (page 150), (2) and (3), they cannot be eliminated. Thus, we have to consider situations like .a ! b/ ^ .b ! c/; a ^ .a ! b/; etc., where without loss of generality none is of the form a ! a; as this can be replaced by TRUE. To consider as many cases together as possible, it is useful to use Fact 4.4.5 (page 138) (10) and (12) backwards, to obtain general formulas. We then see that the cases to examine are of the form: D ..b1 _ : : : _ bn / ! a/ ^ .a ! .c1 ^ : : : ^ cm // ^ a ^ a ^ a; where none of the bi or ci are a; and where n; m may be 0, and ; ; are absent .;; no a/; a; :a; or ::a: We have the following equalities (for F D FALSE; T D TRUE/ W a ^ :a D F; a ^ ::a D a; :a ^ ::a D F; a ^ :a ^ ::a D F: Thus, it suffices to consider as empty, a; :a; or ::a; which leaves us with four cases. Moreover, we see that we always treat b1 _ : : : _ bn and c1 ^ : : : ^ cm as one block, so we can without loss of generality restrict the consideration to the following 12 cases: 1;1 WD .b ! a/; 1;2 WD .b ! a/ ^ a; 1;3 WD .b ! a/ ^ :a; 1;4 WD .b ! a/ ^ ::a; 2;1 WD .a ! c/; 2;2 WD .a ! c/ ^ a; 2;3 WD .a ! c/ ^ :a; 2;4 WD .a ! c/ ^ ::a; 3;1 WD .b ! a/ ^ .a ! c/; 3;2 WD .b ! a/ ^ .a ! c/ ^ a; 3;3 WD .b ! a/ ^ .a ! c/ ^ :a; 3;4 WD .b ! a/ ^ .a ! c/ ^ ::a: We consider now the maximum when we let a float, i.e., consider all m0 such 0 that m L fag D m0 L fag: Let i;j be this maximum. For 1;1 ; 1;2 ; 1;4 ; 0 i;j D T (take a D T /: 0 For 2;1 ; 2;3 ; i;j D T (take a D F /: 0 Next, we consider the remaining simple cases 1;i and 2;i : We show 1;3 D :b; 0 0 2;2 D c; 2;4 D ::cI see Table 4.1 (page 138). (We abbreviate m.a/ < m.b/ by a < b:/. 0 0 0 0 We show now 3;1 D b ! c; 3;2 D c; 3;3 D .b ! c/ ^ :b; 3;4 D.b!c/ ^ ::cI see Table 4.2 (page 139). Remark 4.4.9. We cannot improve the value of ! by taking a detour ! ˛1 ! : : : ! ˛n ! because the destination determines the value: in any column of !; there is only max and a constant value. And if we go further down than needed, we get only worse; going from right to left deteriorates the values in the lines.
4.4 Monotone and Antitone Syntactic Interpolation
145
We can achieve the same result by first closing under the following rules, and then erasing all formulas containing aW (1) ! under transitivity, i.e., ..b1 _ : : : _ bn / ! a/ ^ .a ! .c1 ^ : : : ^ cm // ) ..b1 _ : : : _ bn / ! .c1 ^ : : : ^ cm //: (2) 0 a and ! as follows: ..b1 _ : : : _ bn / ! a/ ^ .a ! .c1 ^ : : : ^ cm //; a ) ..b1 _ : : : _ bn / ! .c1 ^ : : : ^ cm // ^ c1 ^ : : : ^ cm ..b1 _ : : : _ bn / ! a/ ^ .a ! .c1 ^ : : : ^ cm //; ::a ) ..b1 _ : : : _ bn / ! .c1 ^ : : : ^ cm // ^ ::c1 ^ : : : ^ ::cm ..b1 _ : : : _ bn / ! a/ ^ .a ! .c1 ^ : : : ^ cm //; :a ) ..b1 _ : : : _ bn / ! .c1 ^ : : : ^ cm // ^ :b1 ^ : : : ^ :bn : In summary: the semantical interpolant constructed in Section 2.2.2.3 (page 46) is definable if the assumption holds, so the HT logic (see Section 4.4.3.5 (page 150)) has also syntactic interpolation. This result is well known, but we need the techniques for the next section. In some cases, introducing new constants analogous to TRUE or FALSE — in the cited case ONE, TWO when truth starts at world 1 or 2 — might help, but we did not investigate this question. This question is also examined in [ABM03].
4.4.3.3 An Important Example for Non-existence of Interpolation We turn now to an important example. It shows that the logic with three worlds, and thus four truth values, has no interpolation. But first, we show as much as possible for the general case (arbitrarily many truth values). Example 4.4.1. Let ˛.p; q; r/ WD p ! ...q ! r/ ! q/ ! q/ ! p; ˇ.p; s/ WD ..s ! p/ ! s/ ! s: We will show that ˛.p; q; r/ ! ˇ.p; s/ holds in the case of three worlds, but that there is no syntactic interpolant (which could use only p/: Introducing a new operator Jp meaning “from next moment onwards p holds and if now is the last moment then p holds now” gives enough definability to have also syntactic interpolation for ˛ and ˇ above. This will be shown in Section 4.4.3.5 (page 150). First, we give some general results for above example. Fact 4.4.10. Let T; TRUE, be the maximal truth value. (1) WD ..: : : ...a ! b/ ! b/ ! b/ : : :/ ! b/ has the following truth value v./ in a model mW (1.1) if the number n of b on the right of the first ! is odd: if v.a/ v.b/; then v./ D T; otherwise v./ D v.b/; (1.2) if the number m of b on the right of the first ! is even: if v.a/ v.b/; then v./ D v.b/; otherwise v./ D T:
146
4 Monotone and Antitone Semantic and Syntactic Interpolation
(2) WD ..: : : ...a ! b/ ! a/ ! a/ : : :/ ! a/ has the following truth value v./ in a model mW (2.1) if the number n of a on the right of the first ! is odd: if v.b/ < v.a/; then v./ D T; otherwise v./ D v.a/; (2.2) if the number m of a on the right of the first ! is even: if v.b/ < v.a/; then v./ D v.a/; otherwise v./ D T: Proof. (1) We proceed by induction. (1.1) For n D 1; it is the definition of ! : (1.2) Case n D 2 W If v.a/ v.b/; then v.a ! b/ D T; so v..a ! b/ ! b/ D v.b/: If v.a/ > v.b/; then v.a ! b/ D v.b/; so v..a ! b/ ! b/ D T: The general induction works as for the step from n D 1 to n D 2: (2) n D 1W D .a ! b/ ! a: If v.b/ < v.a/; then v.a ! b/ D v.b/; so v./ D T: If v.b/ v.a/; then v.a ! b/ D T; so ./ D v.a/: n ! n C 1W D ! a: If v.b/ < v.a/; then, if n is odd, v. / D T; so v./ D v.a/I if n is even, v. / D v.a/; so v./ D T: If v.b/ v.a/; then, if n is odd, v. / D v.a/; so v./ D T I if n is even, v. / D T; so v./ D v.a/: t u Corollary 4.4.11. Let T be the maximal truth value TRUE. Consider again the formulas of Example 4.4.1 (page 145), ˛.p; q; r/ WD .p ! ...q ! r/ ! q/ ! q// ! p; ˇ.p; s/ WD ..s ! p/ ! s/ ! s: We use Fact 4.4.10 (page 145); the numbers refer to this fact. Let f WD fˇ : Let f 0 .m/ WD mi nff .m0 / W m p D m0 pg: Fix m: By (2.2), if m.p/ D T; then f 0 .m/ D T I if m.p/ < T; then f 0 .m/ D m.p/ C 1: Let g WD f˛ : Let g0 .m/ WD maxfg.m0 / W m p D m0 pg: Fix m: ˛ is of the form .p ! / ! p; so by (2.1), if m0 ./ < m0 .p/; then m0 .˛/ D T I if m0 ./ m0 .p/; then m0 .˛/ D m0 .p/: By (2.2), we have if m0 .r/ < m0 .q/; then m0 ./ D m0 .q/I if m0 .r/ m0 .q/; then m0 ./ D T: Note that m0 ./ > 0: Table 4.3 (page 147) shows that for T D 3; ˛ ` ˇ; for T D 4; ˛ 6` ˇ: Thus, an interpolant h must have h.0/ D 0 or 1, h.1/ D 1 or 2, h.2/ D h.3/ D 3 in the case T D 3: This is impossible by Fact 4.4.5 (page 138). t u 4.4.3.4 The Additional Operators J , A, F , Z The additional operators J and A The following was checked with a small computer program: (1) A alone will generate 12 semantically different formulas with one variable, but it does not suffice to obtain interpolation.
4.4 Monotone and Antitone Syntactic Interpolation
147
Table 4.3 ˛ ` ˇ m.p/ 0 1 2 3
TD3 f 0 .m/ 1 2 3 3
g 0 .m/ 0 1 3 3
TD4 f 0 .m/ 1 2 3 4 4
m.p/ 0 1 2 3 4
g 0 .m/ 0 1 4 4 4
(2) J alone will generate eight semantically different formulas with one variable, and it will solve the interpolation problem for ˛.p; q; r/ and ˇ.p; s/ of Example 4.4.1 (page 145) (3) A and J will generate 48 semantically different formulas with one variable.
The additional operator F For n C 1 truth values, let for k < n k0 WD :.F k .a/ ! F kC1 .a//; k WD a ^ k0 : (For k D n; we take :a:/ Then 8 < n k iff m D n k fk .m/ WD : FALSE otherwise Applying F again, we can increase the value from n k to TRUE. We give the table of k0 .˛/ for six truth values. ˛ 50 D :˛ 40 D :.F 4 .˛/ ! F 5 .˛// 30 D :.F 3 .˛/ ! F 4 .˛// 20 D :.F 2 .˛/ ! F 3 .˛// 10 D :.F .˛/ ! F 2 .˛// 00 D :.˛ ! F .˛// if
0 5 0 0 0 0 0
1 0 5 0 0 0 0
2 0 0 5 0 0 0
3 0 0 0 5 0 0
4 0 0 0 0 5 0
5 0 0 0 0 0 5
This allows definition by cases. Suppose we want ˛ to have the same result as has value p; and the same result as 0 otherwise; more precisely:
F˛ .m/ WD
8 < F .m/ iff F .m/ D p :
F 0 .m/
otherwise
148
4 Monotone and Antitone Semantic and Syntactic Interpolation
then we define 0 . /^ ˛ WD .np
0 / _ ..:np . // ^
0
/:
As k contains !; but only one variable, there is no problem for projections here. Note the following important fact: More generally, we can construct in the same way new functions from old ones by cases, 8 ˆ F .m/ iff F˛ .m/ D s ˆ ˆ ˆ ˆ < F .m/ WD F .m/ iff F˛ .m/ D t ˆ ˆ ˆ ˆ ˆ : F .m/ otherwise But we cannot attribute arbitrary values as in if condition1 holds, then x1; if condition2 holds, then x2; etc. This is also reflected by the fact that by the above, for four truth values, we can (using _/ obtain f0; 3g f0; 1; 2; 3g f0; 2; 3g f0; 3g D 2 4 3 2 D 48 semantically different formulas with one variable, and no more (checked with a computer program). Therefore, F is weaker than Z; which can generate arbitrary functions. Thus, we can also define ! by cases, in a uniform way for all n: Consider, e.g., the case v./ D 3; v. / D 2: v. ! / should be equal to v. /I we take n3 ./ ^ n2 . / ^ : We conclude by the following Fact 4.4.12. F and J C A are interdefinable: Aa $ :.a ! F a/; Ja $ F a _ Aa; F a $ Ja ^ :Aa:
The additional operator Z Definition 4.4.4. We introduce the following, derived, auxiliary operators: Si ./ WD n iff v./ D i; and 0 otherwise, for i D 0; : : : ; n; Ki ./ WD i for any ; for i D 0; : : : ; n: Example 4.4.2. We give here the example for n D 5:
t u
4.4 Monotone and Antitone Syntactic Interpolation
a
0 1
1 2
2 3
3 4
4 5
5 0
a
0 0 1 2 3 4 5
1 0 1 2 3 4 5
2 0 1 2 3 4 5
3 0 1 2 3 4 5
4 0 1 2 3 4 5
5 0 1 2 3 4 5
a
0 5 0 0 0 0 0
1 0 5 0 0 0 0
2 0 0 5 0 0 0
3 0 0 0 5 0 0
4 0 0 0 0 5 0
5 0 0 0 0 0 5
Z K0 K1 K2 K3 K4 K5 S0 S1 S2 S3 S4 S5
149
Fact 4.4.13. (1) We can define Si ./ and Ki ./ for 0 i n from :; ^; _; Z: (2) We can define any m-ary truth function from :; ^; _; Z: Proof. (1)
Si ./ D :Z ni ./; Ki ./ D Z i C1 .:. ^ ://:
(2) Suppose hi1 ; : : : ; im i should have value i; hi1 ; : : : ; im i 7! i I we can express this by Si1 .x1 /^; : : : ; ^Sim .xm / ^ Ki : We then take the disjunction of all such expressions: _
fSi1 .x1 /^; : : : ; ^Sim .xm / ^ Ki W hi1 ; : : : ; im i 7! i g:
Corollary 4.4.14. Any model function is definable from Z; so any semantical interpolant is also a syntactic one. t u
4.4.3.5 Special Finite Goedel Logics We take now a short look at two special finite Goedel logics, first the one based on two worlds, here and there, which will be used again in Section 5.3.6 (page 188) as basis for equilibrium logic, and then the one based on three worlds, which does not have syntactic interpolation.
150
4 Monotone and Antitone Semantic and Syntactic Interpolation
The three-valued intuitionistic logic Here/There HT We now give an introduction to the well-known three-valued intuitionistic logic HT (Here/There), with some results also for similar logics with more than three values. Many of these properties were found and checked with a small computer program. In particular, we show the existence of a normal form, similar to classical propositional logic, but ! cannot be eliminated. Consequently, we cannot always separate propositional variables easily. Our main result here (which is probably well known, and we claim no priority) is that “forgetting” a variable preserves definability in the following sense: Let, e.g., D a ^ b; and M./ be the set of models where has maximal truth value (2 here); then there is 0 such that the set of models where 0 has value 2 is the set of all models which agree with a model of on, e.g., b: We “forget” about a: Our 0 is here, of course, b: Here, the problem is trivial, it is a bit less so when ! is involved, as we cannot always separate the two parts. For example, the result of “forgetting” about a in the formula a ^ .a ! b/ is bI in the formula a ! b it is TRUE. Thus, forgetting about a variable preserves definability, and the above-mentioned semantical interpolation property carries over to the syntactic side, similarly to the result on classical logic; see Section 4.3.3.3 (page 127). Fact 4.4.15. These results were checked with a small computer program: (1) With two variables a; b are definable, using the operators :; !; ^; _; 174 semantically different formulas. _ is not needed, i.e., with or without _ we have the same set of definable formulas. We have, e.g., a _ b $ ..b ! .::a ! a// ! ..:a ! b/ ^ .::a ! a///: (2) With the operators :; ^; _ only 120 semantically different formulas are definable. Thus, ! cannot be expressed by the other operators. Fact 4.4.16. ..a ! b/ ! c/ $ ..:a ! c/ ^ .b ! c/ ^ .a _ :b _ c// holds in the three-valued case. (Thanks to D. Pearce for telling us this.) This Fact is well known, we have verified it by computer, but not by hand. Corollary 4.4.17. The three-valued case has interpolation. Proof. By Fact 4.4.16 (page 150), we can flatten nested ! s; so we have projection. t u Finite Goedel logics with four truth values Interpolation fails for .d ! ...a ! b/ ! a/ ! a/ ! d ` ...c ! d / ! c/ ! c:
4.4 Monotone and Antitone Syntactic Interpolation
151
Consider the table in Corollary 4.4.11 (page 146), and the comment about possible interpolants after the table. By the proof of Fact 4.4.5 (page 138), (1), we see that above formulas have no interpolant. But it is trivial to see that Jp will be an interpolant; see Definition 4.4.3 (page 136). Note that the implication is not true for more than four truth values, as we saw in Corollary 4.4.11 (page 146), so this example will not be a counterexample to interpolation any more. We have checked the following with a computer program, but not by hand: Introducing a new constant, 1, which has always truth value 1, and is thus simpler than the above operator J; gives exactly two different interpolants for the formulas of Example 4.4.1 (page 145): .p ! 1/ ! p and .p ! 1/ ! 1: Introducing an additional constant 2 will give still other interpolants. But if one is permitted in classical logic to use the constants TRUE and FALSE for interpolation, why not 1 and 2 here?
•
Chapter 5
Laws About Size and Interpolation in Non-monotonic Logics
5.1 Introduction 5.1.1 A Succinct Description of Our Main Ideas and Results in This Chapter We connect here properties of abstract multiplication of abstract size to interpolation problems in non-monotonic logic. The size properties are, themselves, connected to properties of the preference relation in preferential logics. Our main result is that interpolation of the type j ˛ j ; i.e., if j ; then there is suitable ˛; defined in the common language of and such that j ˛ j ; exists if the following multiplicative property holds: . 1/ .X Y / D .X / .Y /: (See Proposition 5.3.5 (page 184).) Thus, we can calculate the minimal models (the biggest small subset) of X Y componentwise. Property . 1/ is equivalent to the following two properties of a defining relation between models: .GH1/ 0 0 ^ 00 00 ^ ( 0 0 _ 00 00 ) ) 0 00 0 00 ; .GH 2/ 0 00 0 00 ) 0 0 _ 00 00 : (See Proposition 5.2.4 (page 173).) We should, however, emphasize that the multiplicative laws also have other, equally important, consequences, e.g., for considering consequences ˛ j ˇ in different languages. They are mentioned, but perhaps a bit hidden, too, in Table 5.3 (page 178).
5.1.2 Various Concepts of Size and Non-monotonic Logics A natural interpretation of the non-monotonic rule j is that the set of exceptional cases, i.e., those where holds, but not ; is a small subset of all the D.M. Gabbay and K. Schlechta, Conditionals and Modularity in General Logics, Cognitive Technologies, DOI 10.1007/978-3-642-19068-1 5, © Springer-Verlag Berlin Heidelberg 2011
153
154
5 Laws About Size and Interpolation in Non-monotonic Logics
cases where holds, and the complement, i.e., the set of cases where and hold, is a big subset of all -cases. This interpretation gives an abstract semantics to non-monotonic logic, in the sense that definitions and rules are translated to rules about model sets without any structural justification of those rules, as they are given, e.g., by preferential structures, which provide structural semantics. Yet, they are extremely useful, as they allow us to concentrate on the essentials, forgetting about syntactical reformulations of semantically equivalent formulas; the laws derived from the standard proof-theoretical rules incite to generalizations and modifications, and reveal deep connections but also differences. One of those insights is the connection between laws about size and (semantical) interpolation for non-monotonic logics. To put this abstract view a little more into perspective, we present three alternative systems, also working with abstract size as a semantics for non-monotonic logics: • the system of one of the authors for a first-order setting, published in [Sch90] and elaborated in [Sch95-1], • the system of S. Ben-David and R. Ben-Eliyahu, published in [BB94], • the system of N. Friedman and J. Halpern, published in [FH96]. (1) Defaults as generalized quantifiers: We first recall the definition of a “weak filter”, made official in Definition 2.2.3 (page 34): Fix a base set X: A weak filter on or over X is a set F P.X / such that the following conditions hold: (F1) X 2 F (F2) A B X; A 2 F imply B 2 F .F 30 / A; B 2 F imply A \ B ¤ ;: We use weak filters on the semantical side, and add the following axioms on the syntactical side to a FOL axiomatisation: 1. rx.x/ ^ 8x..x/ ! .x// ! rx .x/; 2. rx.x/ ! :rx:.x/; 3. 8x.x/ ! rx.x/ and rx.x/ ! 9x.x/: A model is now a pair, consisting of a classical FOL model M; and a weak filter over its universe. Both sides are connected by the following definition, where N .M / is the weak filter on the universe of the classical model M W hM; N .M /i ˆ rx.x/ iff there is an A 2 N .M / such that 8a 2 A .hM; N .M /i ˆ Œa/: Soundness and completeness are shown in [Sch95-1]; see also [Sch04]. The extension to defaults with prerequisites by restricted quantifiers is straightforward. (2) The system of S. Ben-David and R. Ben-Eliyahu: Let N 0 WD fN 0 .A/ W A U g be a system of filters for P.U /; i.e. each 0 N .A/ is a filter over A. The conditions are (after slight modification):
5.1 Introduction
155
UC0 W B 2 N 0 .A/ ! N 0 .B/ N 0 .A/; DC0 W B 2 N 0 .A/ ! N 0 .A/ \ P.B/ N 0 .B/; RBC0 W X 2 N 0 .A/; Y 2 N 0 .B/ ! X [ Y 2 N 0 .A [ B/; SRM 0 W X 2 N 0 .A/; Y A ! A Y 2 N 0 .A/ _ X \ Y 2 N 0 .Y /; GTS0 W C 2 N 0 .A/; B A ! C \ B 2 N 0 .B/: (3) The system of N. Friedman and J. Halpern: Let U be a set, < a strict partial order on P.U / (i.e., < is transitive, and contains no cycles). Consider the following conditions for <W (B1) A0 A < B B 0 ! A0 < B 0 ; (B2) if A; B; C are pairwise disjoint, then C < A [ B; B < A [ C ! B [ C < A; (B3) ; < X for all X ¤ ;; (B4) A < B ! A < B-A, (B5) Let X; Y A: If A X < X; then Y < A Y or Y X < X \ Y: The equivalence of the systems of [BB94] and [FH96] was shown in [Sch97-4]; see also [Sch04]. Historical remarks: Our own view as abstract size was inspired by the classical filter approach, as used in mathematical measure theory. The first time that abstract size was related to non-monotonic logics was, to the best of our knowledge, in the second author’s [Sch90] and [Sch95-1], and, independently, in [BB94]. The approach to size by partial orders is first discussed, to the best of our knowledge, by N. Friedman and J. Halpern; see [FH96]. More detailed remarks can also be found in [GS08c, GS09a, GS08f]. A somewhat different approach is taken in [HM07]. Before we introduce the connection between interpolation and multiplicative laws about size, we give some comments on the laws about size.
5.1.3 Additive and Multiplicative Laws About Size We give here a short introduction to and some examples for additive and multiplicative laws about size. A detailed overview is presented in Tables 5.1 (page 164), 5.2 (page 165), and 5.3 (page 178). (The first two tables have to be read together; they are too big to fit on one page.) It seems reasonable to present the information on these additive laws, too, as they show better how the multiplicative laws discussed in this book fit in. They show connections and how to develop a multitude of logical rules known from non-monotonic logics by combining a small number of principles about size. We can use them as building blocks to construct the rules from. More precisely, “size” is to be read as “relative size”, since it is essential to change the base sets. In the first two tables, these principles are some basic and very natural postulates, .Opt/, .iM/, .eM I/, .eM F /, and a continuum of power of the notion of “small”, or, dually, “big”, from .1 s/ to .< ! s/. From these, we can develop the rest, except, essentially, Rational Monotony, and thus an infinity of different rules.
156
5 Laws About Size and Interpolation in Non-monotonic Logics
The probably easiest way to see a connection between non-monotonic logics and abstract size is by considering preferential structures. Preferential structures define principal filters, generated by the set of minimal elements, as follows: if j holds in such a structure, then ./ M. /; where ./ is the set of minimal elements of M./: According to our ideas, we define a principal filter F over M./ by X 2 F iff ./ X M./: Thus, M./ \ M.: / will be a “small” subset of M./: (Recall that filters contain the “big” sets, and ideals the “small” sets.) We can now go back and forth between rules on size and logical rules, e.g.: (For details, see Tables 5.1 (page 164), 5.2 (page 165), and 5.3 (page 178).) (1) The “AND” rule corresponds to the filter property (finite intersections of big subsets are still big). (2) “Right weakening” corresponds to the rule that supersets of big sets are still big. (3) It is natural, but beyond filter properties themselves, to postulate that if X is a small subset of Y; and Y Y 0 ; then X is also a small subset of Y 0 : We call such properties “coherence properties” between filters. This property corresponds to the logical rule .wOR/. (4) In the rule .CM! / usually called Cautious Monotony, we change the base set a little when going from M.˛/ to M.˛ ^ ˇ/ (the change is small by the prerequisite ˛ j ˇ/; and still have ˛ ^ ˇ j ˇ 0 if we had ˛ j ˇ 0 : We see here a conceptually very different use of “small”, as we now change the base set, over which the filter is defined, by a small amount. (5) The rule of Rational Monotony is the last one in the first table, and somewhat isolated there. It is better seen as a multiplicative law, as described in the third table. It corresponds to the rule that the product of medium (i.e, neither big nor small) sets, has still medium size.
5.1.4 Interpolation and Size The connection between non-monotonic logic and the abstract concept of size was investigated in [GS09a]; see also [GS08f]. There, we looked among other things at abstract addition of size. Here, we will show a connection to abstract multiplication of size. Our semantical approach used decomposition of set-theoretical products. An important step was to write a set of models † as a product of some set †0 (which was a restriction of †/ and some full Cartesian product. So, when we speak about size, we will have (slightly simplified) some big subset †1 of one product …1 and some big subset †2 of another product …2 ; and will now check whether †1 †2 is a big subset of …1 …2 ; in shorthand, whether “big big D big”. (See Definition 5.2.1 (page 169) for precise definitions.) Such conditions are called coherence conditions, as they do not concern the notion of size itself, but the way the sizes defined for different base sets are connected. Our main results here are Propositions 5.3.3 (page 183) and 5.3.5 (page 184). They say that if the logic under investigation is defined from a notion of size which satisfies sufficiently many conditions, then this logic will have interpolation of type 3 or even 2 (type 3 is the weakest form discussed here).
5.1 Introduction
157
Consider now some set product X X 0 : (Intuitively, X and X 0 are model sets on sublanguages J and J 0 of the whole language L:/ When we have now a rule such as, if Y is a big subset of X; and Y 0 a big subset of X 0 ; then Y Y 0 is a big subset of X X 0 ; and conversely, we can calculate consequences separately in the sublanguages and put them together to have the overall consequences. But this is the principle behind interpolation: we can work with independent parts. This is made precise in Definition 5.2.1 (page 169), in particular by the rule . 1/: .X X 0 / D .X / .X 0 /: (Note that the conditions . i / and .† i / are equivalent, as shown in Proposition 5.2.1 (page 170) (for principal filters).) The main result is that the multiplicative size rule . 1/ entails non-monotonic interpolation of the form j ˛ j I see Proposition 5.3.5 (page 184). We take now a closer look at interpolation for non-monotonic logic.
The three variants of interpolation Consider preferential logic, a rule like j : This means that ./ M. /: So we go from M./ to ./; the minimal models of ; and then to M. /; and, abstractly, we have M./ ./ M. /; so we have neither necessarily M./ M. /; nor M./ M. /I the relation between M./ and M. / may be more complicated. Thus, we have neither the monotone, nor the antitone case. For this reason, our general results for monotone or antitone logics do not hold any more. But we also see here that classical logic is used, too. Suppose that there is a 0 which describes exactly ./I then we can write j 0 ` : So we can split preferential logic into a core part — going from to its minimal models — and a second part, which is just classical logic. (Similar decompositions are also natural for other non-monotonic logics.) Thus, preferential logic can be seen as a combination of two logics, the non-monotonic core, and classical logic. It is thus natural to consider variants of the interpolation problem, where j denotes again preferential logic, and ` as usual classical logic: Given j ; is there “simple” ˛ such that (1) j ˛ ` , or (2) ` ˛ j , or (3) j ˛ j ? In most cases, we will only consider the semantical version, as the problems of the syntactical version are very similar to those for monotonic logics. We turn to the variants. (1) The first variant, j ˛ ` ; has a complete characterization in Proposition 5.3.2 (page 181), provided we have a suitable normal form (conjunctions of disjunctions). The condition says that the relevant variables of ./ have to be relevant for M./; too.
158
5 Laws About Size and Interpolation in Non-monotonic Logics
(2) The second variant, ` ˛ j ; is related to very (and in many cases, too) strong conditions about size. We do not have a complete characterization, only sufficient conditions about size. The size conditions we need are (see Definition 5.2.1 (page 169)) the abovementioned . 1/; and . 2/: .X / Y ) .X A/ Y A where X need not be a product any more. The result is given in Proposition 5.3.3 (page 183). Example 5.2.1 (page 171) shows that . 2/ seems too strong when compared to probability-defined size. We should, however, note that sufficiently modular preferential relations guarantee these very strong properties of the big sets; see Section 5.2.3 (page 172). (3) We turn to the third variant, j ˛ j : This is probably the most interesting one, as it is more general, loosens the connection with classical logic, seems more natural as a rule, and is also connected to more natural laws about size. Again, we do not have a complete characterization, only sufficient conditions about size. Here, . 1/ suffices, and we have our main result about nonmonotonic semantical interpolation, Proposition 5.3.5 (page 184), that . 1/ entails interpolation of the type j ˛ j : Proposition 5.2.4 (page 173) shows that . 1/ is (roughly) equivalent to the properties .GH1/ ^ 0 0 ^ . _ 0 0 / ) 0 0 (where iff or D /; .GH 2/ 0 0 ) _ 0 0 of a preferential relation. ..GH 2/ means that some compensation is possible, e.g., might be the case, but 0 0 wins in the end, so 0 0 :/ There need not always be a semantical interpolation for the third variant; this is shown in Example 5.3.1 (page 181). So we see that, roughly, semantical interpolation for non-monotonic logics works when abstract size is defined in a modular way — and we see independence again. In a way, this is not surprising, as we use an independent definition of validity for interpolation in classical logic, and we use an independent definition of additional structure (relations or size) for interpolation in non-monotonic logic.
5.1.5 Hamming Relations and Size As preferential relations are determined by a relation, and give rise to abstract notions of size and their manipulation, it is natural to take a close look at the corresponding properties of the relation. We already gave a few examples in the
5.1 Introduction
159
preceding sections, so we can be concise here. Our main definitions and results on this subject are to be found in Section 5.2.3 (page 172), where we also discuss distances with similar properties. It is not surprising that we find various types of Hamming relations and distances in this context, as they are, by definition, modular. Neither is it surprising that we see them again in Chapter 6 (page 197), as we are interested there in independent ways to define neighbourhoods. Basically, these relations and distances come in two flavours, the set and the counting variant. This is perhaps best illustrated by the Hamming distance of two sequences of finite, equal length. We can define the distance by the set of arguments where they differ, or by the cardinality of this set. The first results in possibly incomparable distances, the second allows “compensation”; difference in one argument can be compensated by equality in another argument. For definitions and results, also those connecting them to notions of size, see Section 5.2.3 (page 172), in particular Definition 5.2.2 (page 172). We then show in Proposition 5.2.4 (page 173) that (smooth) Hamming relations generate our size conditions when size is defined as above from a relation (the set of preferred elements generates the principal filter). Thus, Hamming relations determine logics which have interpolation; see Corollary 5.3.4 (page 184).
5.1.6 Equilibrium Logic Equilibrium logic, due to D. Pearce, A. Valverde (see [PV09] for motivation and further discussion) is based on the three-valued finite Goedel logic, also called HT logic, HT for “here and there”. Our results are presented in Section 5.3.6 (page 188). Equilibrium logic (EQ) is defined by a choice function on the model set. First models have to be “total”, no variable of the language may have 1 as value. Second, if m m0 ; then m is considered better, and m0 discarded, where m m0 iff m and m0 give value 0 to the same variables, and m gives value 2 to strictly less (as subset) variables than m0 does. We can define equilibrium logic by a preferential relation (taking care also of the first condition), but it is not smooth. Thus, our general results from the beginning of this section will not hold, and we have to work with “hand-knitted” solutions. We first show that equilibrium logic has no interpolation of the form ` ˛ j or j ˛ ` ; then that is has interpolation of the form j ˛ j ; and that the interpolant is also definable, i.e., equilibrium logic has semantical and syntactic interpolation of this form. Essentially, semantical interpolation is due to the fact that the preference relation is defined in a modular way, using individual variables — as always, when we have interpolation. Thus, equilibrium logic is a nice example to apply our ideas, though not all techniques developed before will work. Seen this way, our results here are an extension to the general approach.
160
5 Laws About Size and Interpolation in Non-monotonic Logics
5.1.7 Interpolation for Revision and Argumentation We have a short and simple result (Lemma 5.3.6 (page 188)) for interpolation in AGM revision. Unfortunately, we need the variables from both sides of the revision operator, as can easily be seen by revising with TRUE. The reader is referred to Section 5.3.5 (page 188) for details. Somewhat surprisingly, we also have an interpolation result for one form of argumentation, where we consider the set of arguments for a statement as the truth value of that statement. As we have maximum (set union), we have the lower bound used in Proposition 4.2.3 (page 122) for the monotonic case, and can show Fact 5.5.3 (page 195). See Section 5.5 (page 194) for details.
5.1.8 Language Change to Obtain Products To achieve interpolation and other results of independence, we often need to write a set of models as a nontrivial product. Sometimes, this is impossible, but an equivalent reformulation of the language can solve the problem; see Example 5.2.5 (page 179). Crucial there is that 6 D 3 2; so we can just rearrange the six models in a different way; see Fact 5.2.9 (page 179). A similar result holds for the nonmonotonic case, where the structure must be possible; we can then redefine the language. Our results show that factorization is a structural problem, and simple redefinitions of the language can achieve factorization if the structural requirements are satisfied. As such, these small results seem to be important. All details are to be found in Section 5.2.5 (page 179).
5.2 Laws About Size 5.2.1 Additive Laws About Size This section presents additive laws for abstract size, whereas our emphasis in this book is on multiplicative laws. We present them mainly to put our work in perspective, to connect this book to more familiar laws about size, and to make the text more self-contained, i.e., as a service to the reader. We now give the main additive rules for manipulation of abstract size from [GS09a]; see Tables 5.1 (page 164) and 5.2 (page 165), “Rules on size”. These laws allow us to give very concise abstract descriptions of laws about nonmonotonic logics, and, by their modularity, to construct new logics by changing some parameters. Moreover, they reveal deep connections between laws like AND and Cumulativity, which are not so obvious in their usual formulations.
5.2 Laws About Size
161
The notation is explained with some redundancy, so the reader will not have to leaf back and forth to Chapter 2 (page 31).
5.2.1.1 Notation (1) P.X / is the power set of X I is the subset relation, the strict part of ; i.e. A B iff A B and A ¤ B: The operators ^; :; _; ! and ` have their usual, classical interpretation. (2) I.X / P.X / and F .X / P.X / are dual abstract notions of size; I.X / is the set of “small” subsets of X; F.X / the set of “big” subsets of X: They are dual in the sense that A 2 I.X / , X A 2 F .X /: “I” evokes “ideal”, “F ” evokes “filter” though the full strength of both is reached only in .< ! s/: “s” evokes “small”, and “.x s/” stands for “x small sets together are still not everything”. (3) If A X is neither in I.X /; nor in F.X /; we say it has medium size, and we define M.X / WD P.X / .I.X / [ F .X //: MC .X / WD P.X / I.X / is the set of subsets which are not small. (4) rx is a generalized first-order quantifier; it is read “almost all x have property ”. rx. W / is the relativized version, read “almost all x with property have also property ”. To keep the table “Rules on size” simple, we write mostly only the non-relativized versions. Formally, we have rx W, fx W .x/g 2 F.U / where U is the universe, and rx. W / W, fx W . ^ /.x/g 2 F.fx W .x/g/: Soundness and completeness results on r can be found in [Sch95-1]. (5) Analogously, for propositional logic, we define ˛ j ˇ W, M.˛ ^ ˇ/ 2 F.M.˛//; where M./ is the set of models of : (6) In preferential structures, .X / X is the set of minimal elements of X: This generates a principal filter by F.X / WD fA X W .X / Ag: Corresponding properties about are not listed systematically. (7) The usual rules .AND/ etc. are named here .AND! /; as they are in a natural ascending line of similar rules, based on strengthening of the filter/ideal properties. (8) For any set of formulas T; and any consequence relation j; we will use T WD f W T ` g; the set of classical consequences of T; and T WD f W T j g; the set of consequences of T under the relation j : (9) We say that a set X of models is definable by a formula (or a theory) iff there is a formula (or a theory T / such that X D M./ (or X D M.T //; the set of models of (or T /: (10) Most rules are explained in the table “Logical rules”, and “RW” stands for Right Weakening.
162
5 Laws About Size and Interpolation in Non-monotonic Logics
5.2.1.2 The Groups of Rules The rules concern properties of I.X / or F.X /; or dependencies between such properties for different X and Y: All X; Y; etc. will be subsets of some universe, say V: Intuitively, V is the set of all models of some fixed propositional language. It is not necessary to consider all subsets of V I the intention is to consider subsets of V; which are definable by a formula or a theory. So we assume all X; Y; etc. taken from some Y P.V /; which we call the domain. In the former case, Y is closed under set difference, in the latter case not necessarily so. (We will mention it when we need some particular closure property.) The rules are divided into five groups: (1) .Opt/; which says that “All” is optimal, i.e., when there are no exceptions, then a soft rule j holds. (2) three monotony rules: (2.1) .iM / is inner monotony; a subset of a small set is small; (2.2) .eM I/ is external monotony for ideals: enlarging the base set keeps small sets small; (2.3) .eM F/ is external monotony for filters: a big subset stays big when the base set shrinks. These three rules are very natural if “size” is anything coherent over a change of base sets. In particular, they can be seen as weakening. (3) . / keeps proportions; it is present here mainly to point the possibility out. (4) a group of rules x s; which say how many small sets will not yet add to the base set. The notation “.< ! s/” is an allusion to the full filter property, that filters are closed under finite intersections. (5) Rational Monotony, which can best be understood as robustness of MC I see .MCC /; (3). We will assume all base sets to be nonempty in order to avoid pathologies and in particular clashes between .Opt/ and .1 s/: Note that the full strength of the usual definitions of a filter and an ideal are reached only in line .< ! s/:
Regularities (1) The group of rules .x s/ use ascending strength of I=F: (2) The column .MC / contains interesting algebraic properties. In particular, they show a strengthening from .3 s/ up to Rationality. They are not necessarily equivalent to the corresponding .Ix / rules, not even in the presence of the basic rules. The examples show that care has to be taken when considering the different variants. (3) Adding the somewhat superflous .CM2 /; we have increasing Cautious Monotony from .wCM/ to full .CM! /:
5.2 Laws About Size
163
(4) We have increasing “or” from .wOR/ to full .OR! /: (5) The line .2 s/ is only there because there seems to be no .MC 2 /I otherwise we could begin .n s/ at n D 2:
Summary We can obtain all rules except .RatM/ and . / from .Opt/, the monotony rules — .iM/, .eM I/, .eM F/ — and .x s/ with increasing x:
5.2.1.3 Table The following table is split into two, as it is too big for printing on one page. (See Tables 5.1 (page 164), “Rules on size — Part I” and 5.2 (page 165), “Rules on size — Part II”.
5.2.2 Multiplicative Laws About Size We are mainly interested in non-monotonic logic. In this domain, independence is strongly connected to multiplication of abstract size, and an important part of the present text treats this connection and its repercussions. We have at least two scenarios for multiplication; one is decribed in Diagram 5.2.1 (page 166), the second in Diagram 5.2.2 (page 168). In the first scenario, we have nested sets, in the second, we have set products. In the first scenario, we consider subsets which behave as the big set does; in the second scenario we consider subspaces, and decompose the behaviour of the big space into behaviour of the subspaces. In both cases, this results naturally in multiplication of abstract sizes. When we look at the corresponding relation properties, they are quite different (rankedness vs. some kind of modularity). But this is perhaps to be expected, as the two scenarios are quite different. Other scenarios which might be interesting to consider in our framework are as follows: • When we have more than two truth values, say three, and two is considered a big subset, and we have n propositional variables, and m of them are considered many, then 2m might give a “big” subset of the total of 3n situations. • Similarly, when we fix one variable, consider two cases of the possible three, and multiply this with a “big” set of models. • We may also consider the utility or cost of a situation, and work with a “big” utility, and “many” situations, etc. • Note that, in the case of distances, subspaces add distances, and do not multiply them: d.xy; x 0 y 0 / D d.x; x 0 / C d.y; y 0 /: These questions are left for further research; see also Section 3.3 (page 93).
164
5 Laws About Size and Interpolation in Non-monotonic Logics
Table 5.1 Rules on size — Part I “Ideal” .Opt/
.iM /
.eM I/
.eM F /
./
.1 s/ .2 s/
.n s/ .n 3/
; 2 I.X /
. . .
Rules on size — Part I “Filter”
MC
Optimal proportion X 2 F .X /
r 8x˛ ! rx˛
Monotony (Improving proportions). .iM /: internal monotony, .eM I/: external monotony for ideals, .eM F /: external monotony for filters A B 2 I.X / . A 2 F .X /, rx˛ ^ 8x.˛ ! ˛0 / ) . AB X ! rx˛0 A 2 I.X / . ) B 2 F .X / X Y ) . rx.˛ W ˇ/^ I.X / I.Y / . 8x.˛ 0 ! ˇ/ ! . rx.˛ _ ˛0 W ˇ/ . . . . . X Y ) rx.˛ W ˇ/^ . F .Y / \ P.X / 8x.ˇ ^ ˛ ! ˛ 0 / ! . F .X / rx.˛ ^ ˛0 W ˇ/ . Keeping proportions .I [ d isj / . .F [ d isj / .MC [ d isj / A 2 I.X /; . A 2 F .X /; A 2 MC .X /; B 2 I.Y /; . B 2 F .Y /; B 2 MC .Y /; X \Y D;) . X \Y D;) X \Y D;) A [ B 2 I.X [ Y / .A [ B 2 F .X [ Y / A [ B 2 MC .X [ Y / . . .I1 / X 62 I.X / .I2 / A; B 2 I.X / ) A[B ¤X .In / A1 ; :; An 2 I.X / ) A1 [ : [ An ¤ X
Robustness of proportions: n small ¤ Al l . .F1 / . ; 62 F .X / . .F2 / . A; B 2 F .X / ) . A\B ¤; . .Fn / . A1 ; :; An 2 I.X / . ) . A1 \ : \ An ¤ ; .
.< ! s/
.I! / . .F! / A; B 2 I.X / ) . A; B 2 F .X / ) A [ B 2 I.X / . A \ B 2 F .X / . . . . . . . . . .
.MCC /
. . . . . . . . . . .
C
.Mn / X1 2 F .X2 /; :; Xn1 2 F .Xn / ) X1 2 MC .Xn / C
.M! / (1) A 2 F .X /; X 2 MC .Y / ) A 2 MC .Y / (2) A 2 MC .X /; X 2 F .Y / ) A 2 MC .Y / (3) A 2 F .X /; X 2 F .Y / ) A 2 F .Y / (4) A; B 2 I.X / ) A B 2 I.X -B)
Robustness of MC
.MCC / (1) A 2 I.X /; B 62 F .X / ) A B 2 I.X B/ (2) A 2 F .X /; B 62 F .X / ) A B 2 F .X B/ (3) A 2 MC .X /; X 2 MC .Y / ) A 2 MC .Y /
rx.˛ W ˇ/^ rx.˛ 0 W ˇ/^ :9x.˛ ^ ˛0 / ! rx.˛ _ ˛0 W ˇ/
.r1 / rx˛ ! 9x˛ .r2 / rx˛ ^ rxˇ ! 9x.˛ ^ ˇ/ .rn / rx˛1 ^ : ^ rx˛n ! 9x.˛1 ^ : ^ ˛n / .r! / rx˛ ^ rxˇ ! rx.˛ ^ ˇ/
5.2 Laws About Size
165
Table 5.2 Rules on size — Part II various rules
Rules on size — Part II AND
OR
Caut./Rat. Mon.
Optimal proportion .Opt/
.S C / ˛`ˇ)˛ j ˇ Monotony (Improving proportions)
.iM /
.eM I/
.RW / ˛ j ˇ; ˇ ` ˇ 0 ) ˛ j ˇ0 .PR0 / ˛ j ˇ; ˛ ` ˛ 0 ; ˛ 0 ^ :˛ ` ˇ ) ˛0 j ˇ .PR/ X Y ) .Y / \ X .X /
.wOR/ ˛ j ˇ; ˛ 0 ` ˇ ) ˛ _ ˛0 j ˇ .wOR/ .X [ Y / .X / [ Y
.eM F /
.wCM / ˛ j ˇ; ˛ 0 ` ˛; ˛ ^ ˇ ` ˛0 ) ˛0 j ˇ Keeping proportions
./
.NR/ ˛ j ˇ) ˛^ j ˇ or ˛ ^ : j ˇ
.1 s/
.CP / ˛ j ?)˛`?
.2 s/
.n s/ .n 3/
.< ! s/
.MCC /
.d isjOR/ ˛ j ˇ; ˛0 j ˇ0 ˛ ` :˛ 0 ; ) ˛ _ ˛0 j ˇ _ ˇ0 .d isjOR/ X \Y D;) .X [ Y / .X / [ .Y / Robustness of proportions: n small ¤ Al l .AND1 / ˛ j ˇ ) ˛ 6` :ˇ .AND2 / .OR2 / ˛ j ˇ; ˛ j ˇ0 ) ˛ j ˇ ) ˛ 6 j :ˇ 0 ˛ 6` :ˇ _ :ˇ .ANDn / .ORn / ˛ j ˇ1 ; :; ˛ j ˇn ˛1 j ˇ; :; ˛n1 j ˇ ) ) ˛ 6` :ˇ1 _ : _ :ˇn ˛1 _ : _ ˛n1 6 j :ˇ .AND! / ˛ j ˇ; ˛ j ˇ0 ) ˛ j ˇ ^ ˇ0
.OR! / ˛ j ˇ; ˛0 j ˇ) ˛ _ ˛0 j ˇ .OR/ .X [ Y / .X / [ .Y /
.CM2 / ˛ j ˇ ) ˛ 6 j :ˇ .CMn / ˛ j ˇ1 ; :; ˛ j ˇn1 ) ˛ ^ ˇ1 ^ : ^ ˇn2 6 j :ˇn1 .CM! / ˛ j ˇ; ˛ j ˇ0 ) ˛ ^ˇ j ˇ0 .CM / .X / Y X ) .Y / .X /
Robustness of MC .RatM / ˛ j ˇ; ˛ 6 j :ˇ 0 ) ˛ ^ ˇ0 j ˇ .RatM / X Y; X \ .Y / ¤ ; ) .X / .Y / \ X
166
5 Laws About Size and Interpolation in Non-monotonic Logics
5.2.2.1 Multiplication of Size for Subsets Here we have nested sets, A X Y; A is a certain proportion of X; and X of Y; resulting in a multiplication of relative size or proportions. This is a classical subject of non-monotonic logic; see the last section, taken from [GS09a]; it is partly repeated here to stress the common points with the other scenario.
Properties Diagram 5.2.1 (page 166) is to be read as follows: The whole set Y is split into X and Y X; X is split into A and X A: X is a small/medium/big part of Y; and A is a small/medium/big part of X: The question is, is A a small/medium/big part of Y ‹ Note that the relation of A to X is conceptually different from that of X to Y; as we change the base set by going from X to Y; but not when going from A to X: Thus, in particular, when we read the diagram as expressing multiplication, commutativity is not necessarily true. We looked at this scenario in [GS09a], but from an additive point of view, using various basic properties like .iM /, .eM I/, .eM F/; see Section 5.2.1 (page 160). Here, we use just multiplication — except sometimes for motivation.
A
X Y Scenario 1
Diagram 5.2.1
5.2 Laws About Size
167
We examine different rules: If Y D X or X D A; there is nothing to show, so 1 is the neutral element of multiplication. If X 2 I.Y / or A 2 I.X /; then we should have A 2 I.Y /: (Use for motivation .iM / or .eM I/ respectively.) So it remains to look at the following cases, with the “natural” answers given already: (1) (2) (3) (4)
X X X X
2 F .Y /; A 2 F .X / ) A 2 F.Y /; 2 MC .Y /; A 2 F.X / ) A 2 MC .Y /; 2 F .Y /; A 2 MC .X / ) A 2 MC .Y /; 2 MC .Y /; A 2 MC .X / ) A 2 MC .Y /:
But (1) is case (3) of .MC ! / in [GS09a]; see Table “Rules on size” in Section 5.2.1 (page 160). (2) is case (1) of .MC ! / there, (3) is case (2) of .MC ! / there, and finally, (4) is .MCC / there. So the first three correspond to various expressions of .AND! /, .OR! /, .CM! /, the last one to .RatM/. But we can read them also the other way round, e.g.: (1) corresponds to ˛ j ˇ; ˛ ^ ˇ j ) ˛ j ; (2) corresponds to ˛ j6 :ˇ; ˛ ^ ˇ j ) ˛ j6 :.ˇ ^ /; (3) corresponds to ˛ j ˇ; ˛ ^ ˇ j6 : ) ˛ j6 :.ˇ ^ /: All these rules might be seen as too idealistic, so as we did in [GS09a], we can consider milder versions: We might for instance consider a rule which says that big : : : big; n times, is not small. Consider for instance the case n D 2: So we would conclude that A is not small in Y: In terms of logic, we then have ˛ j ˇ; ˛ ^ ˇ j ) ˛ j6 .:ˇ _ : /: We can obtain the same logical property from 3 small ¤ all:
5.2.2.2 Multiplication of Size for Subspaces Our main interest here is multiplication for subspaces, which we discuss now. The reason for this interest is that it is an abstract view on model sets defined for sublanguages and their composition. Consequently, it offers an abstract view on properties, assuring interpolation for non-monotonic logics. Here we will, however, only look at these laws in general, which are interesting also beyond interpolation. In particular, they also describe what may happen when we go from one language L to a bigger language L0 ; and to the consequence relations in both languages. A priori, they may be totally different, but it often seems reasonable to postulate certain coherence properties between those consequence relations. For instance, if ˛ j ˇ in L; then we will often implicitly assume that also ˛ j ˇ in L0 — but this is not trivial. This corresponds to the following coherence property for sizes: If
168
5 Laws About Size and Interpolation in Non-monotonic Logics
Γ1
Γ2
Σ1
Σ2 Scenario 2
Diagram 5.2.2
† …L is a big subset, then † ….L0 L/ …L0 is also a big subset. This is the kind of property we investigate here.
Properties In this scenario, †i are sets of sequences, (see Diagram 5.2.2 (page 168)) corresponding, intuitively, to a set of models in language Li I †i will be the set of ˛i -models and the subsets i are to be seen as the “best” models where ˇi will hold. The languages are supposed to be disjoint sublanguages of a common language L: As the †i have symmetrical roles, there is no intuitive reason for multiplication not to be commutative. We can interpret the situation twofold: First, we work separately in sublanguages L1 and L2 ; and, say, ˛i and ˇi are both defined in Li ; we look at ˛i j ˇi in the sublanguage Li or we consider both ˛i and ˇi in the big language L and look at ˛i j ˇi in L: These two ways are a priori completely different. Speaking in preferential terms, it is not at all clear why the orderings on the submodels should have anything to do with the orderings on the whole models. It seems a very desirable property, but we have to postulate it, which we do now (an overview is given in Table 5.3 (page 178)). We first give informally a list of such rules, mainly to show the connection with the first scenario. Later (see Definition 5.2.1 (page 169)), we will introduce formally some rules for which we show a connection with interpolation. Here “.big big ) big/” stands for “if both factors are big, so will be the product”; this will be abbreviated by “b b ) b” in Table 5.3 (page 178). We have .big 1 ) big/ Let 1 †1 I if 1 2 F .†1 /; then 1 †2 2 F.†1 †2 / (and we have the dual rule for †2 and 2 /:
5.2 Laws About Size
169
This property preserves proportions, so it seems intuitively quite uncontested whenever we admit coherence over products. (Recall that there was nothing to show in the first scenario.) When we reconsider the above case, and suppose ˛ j ˇ is in the sublanguage, so M.ˇ/ 2 F.M.˛// holds in the sublanguage, so by .big 1 ) big/; M.ˇ/ 2 F.M.˛// in the big language L: We obtain the dual rule for small (and likewise, medium size) sets: .small1 ) small/ Let 1 †1 I if 1 2 I.†1 /; then 1 †2 2 I.†1 †2 / (and we have the dual rule for †2 and 2 /; establishing All D 1 as the neutral element for multiplication. We look now at other plausible rules: .small x ) small/ 1 2 I.†1 /; 2 †2 ) 1 2 2 I.†1 †2 /: .big big ) big/ 1 2 F.†1 /; 2 2 F .†2 / ) 1 2 2 F .†1 †2 /: .big medium ) medium/ 1 2 F .†1 /; 2 2 MC .†2 / ) 1 2 2 MC .†1 †2 /: .medium medium ) medium/ 1 2 MC .†1 /; 2 2 MC .†2 / ) 1 2 2 MC .†1 †2 /: When we accept all the above rules, we can invert .big big ) big/; as a big product must be composed of big components. Likewise, at least one component of a small product has to be small; see Proposition 5.2.1 (page 170). We see that these properties give a lot of modularity. We can calculate the consequences of ˛ and ˛0 separately — provided ˛; ˛0 use disjoint alphabets — and put the results together afterwards. Such properties are particularly interesting for classification purposes, where subclasses are defined with disjoint alphabets. Recall that we work here with a notion of “big” and “small” subsets, which may be thought of as defined by a filter (ideal), though we usually will not need the full strength of a filter (ideal). But assume as usual that A B C and A C is big together imply B C is big, and that C C is big, and define A B as small iff .B A/ B is big; call all subsets which are neither big nor small medium size. For an extensive discussion, see [GS09a]. Let X 0 [ X 00 D X be a disjoint cover, so …X D …X 0 …X 00 : We consider subsets †; etc. of …X: If not stated otherwise, †; etc. need not be a product †0 †00 : We will sometimes write …0 WD …X 0 ; †00 WD † X 00 : The roles of X 0 and X 00 are interchangeable, e.g., instead of X 0 † X 0 ; we may also write X 00 † X 00 : We consider here the following two sets of three finite product rules about size and : Both sets will be shown to be equivalent in Proposition 5.2.1 (page 170). Definition 5.2.1. .S 1/ †0 †00 is big iff there is D 0 00 such that 0 †0 and 00 †00 are big. .S 2/ † is big ) X 0 † X 0 is big, where † is not necessarily a product. .S 3/ A † is big ) there is a B …0 †00 big such that B X 00 A X 00 I again, † is not necessarily a product. . 1/ .†0 †00 / D .†0 / .†00 /:
170
5 Laws About Size and Interpolation in Non-monotonic Logics
. 2/ .†/ ) .† X 0 / X 0 : . 3/ .…X 0 †00 / X 00 .†/ X 00 : .s s/ Let i †i I then 1 2 †1 †2 is small iff 1 †1 is small or 1 †1 is small. A generalization to more than two factors is obvious. One can also consider weakenings, e.g., .S 10 / 0 †00 †0 †00 is big iff 0 †0 is big. Proposition 5.2.1. (1) Let .S 1/ hold. Then 0 00 †0 †00 is small iff 0 †0 or 00 †00 is small. (2) If the filters over A are principal filters, generated by .A/; i.e. B A is big iff .A/ B A for some .A/ A; then .S i / is equivalent to . i /, i D 1; 2; 3: (3) Let the notion of size satisfy .Opt/, .iM/, and .< ! s/; see the tables “Rules on size” in Section 5.2.1 (page 160). Then . 1/ and .s s/ are equivalent. Proof. (1) “(”: Suppose 0 †0 is small. Then †0 0 †0 is big and .†0 0 /†00 †0 †00 is big by .S 1/: But . 0 00 / \ ..†0 0 / †00 / D ;; so 0 00 †0 †00 is small. “)”: For the converse, suppose that neither 0 †0 nor 00 †00 is small. Let A †0 †00 be big; we show that A \ . 0 00 / ¤ ;: By .S 1/ there are B 0 †0 and B 00 †00 big, and B 0 B 00 A: Then B 0 \ 0 ¤ ;; B 00 \ 00 ¤ ;; so there is hx 0 ; x 00 i 2 .B 0 B 00 / \ . 0 00 / A \ . 0 00 /: (2.1) “)” “”: .†0 / †0 and .†00 / †00 are big, so by .S 1/ .†0 / .†00 / 0 † †00 is big, so .†0 †00 / .†0 / .†00 /: “ ”: .†0 †00 / †0 †00 is big ) by .S 1/ there is 0 00 .†0 †00 / and 0 †0 ; 00 †00 big ) .†0 / 0 ; .†00 / 00 ) .†0 / .†00 / .†0 †00 /: “(” Let 0 †0 be big, 00 †00 be big, 0 00 I then .†0 / 0 ; .†00 / 00 ; so by . 1/ .†/ D .†0 / .†00 / 0 00 ; so is big. Let † be big; then by . 1/ .†0 / .†00 / D .†/ : (2.2) “)” .†/ ) † big ) by .S 2/ X 0 † X 0 big ) .† X 0 / X 0: “(” † big ) .†/ ) by . 2/ .† X 0 / X 0 ) X 0 † X 0 big. (2.3) “)”
5.2 Laws About Size
171
.†/ † big ) 9B …X 0 †00 big such that B X 00 .†/ X 00 by .S 3/I thus in particular .…X 0 †00 / X 00 .†/ X 00 : “(” A † big ) .†/ A: .…X 0 †00 / …X 0 †00 is big, and by . 3/ .…X 0 †00 / X 00 .†/ X 00 A X 00 : (3) “)”: (1) Let 0 †0 be small; we show that 0 00 †0 †00 is small. So 0 † 0 †0 is big, so by .Opt/ and . 1/ .†0 0 / †00 †0 †00 is big, so 0 †00 D .†0 †00 / ..†0 0 / †00 / †0 †00 is small, so by .iM/ 0 00 †0 †00 is small. (2) Suppose 0 †0 and 00 †00 are not small; we show that 0 00 †0 †00 is not small. So †0 0 †0 and †00 00 †00 are not big. We show that Z WD ..†0 †00 /. 0 00 // †0 †00 is not big. Z D .†0 .†00 00 //[..†0 0 /†00 /: Suppose X 0 X 00 ZI then X 0 †0 0 or X 00 †00 00 : Proof: Let 0 X 6 †0 0 and X 00 6 †00 00 ; but X 0 X 00 Z: Let 0 2 X 0 .†0 0 /; 00 2 X 00 .†00 00 /I consider 0 00 : 0 00 62 .†0 0 / †00 ; as 0 62 †0 0 ; 0 00 62 †0 .†00 00 /; as 00 62 †00 00 ; so 0 00 62 Z: By the prerequisite, †0 0 †0 is not big, †00 00 †00 is not big, so by .iM/ no X 0 with X 0 †0 0 is big, no X 00 with X 00 †00 00 is big, so by . 1/ or .S 1/ Z †0 †00 is not big, so 0 00 †0 †00 is not small. “(”: (1) Suppose 0 †0 is big; 00 †00 is big, we have to show 0 00 0 † †00 is big. †0 0 †0 is small, †00 00 †00 is small, so by .s s/ .†0 0 / †00 †0 †00 is small and †0 .†00 00 / †0 †00 is small, so by .< ! s/ .†0 †00 / . 0 00 / D ..†0 0 / †00 / [ .†0 .†00 00 // †0 †00 is small, so 0 00 †0 †00 is big. (2) Suppose 0 00 †0 †00 is big; we have to show 0 †0 is big, and 00 †00 is big. By the prerequisite, .†0 †00 / . 0 00 / D ..†0 0 / †00 / [ .†0 .†00 00 // †0 †00 is small, so by .iM/ †0 .†00 00 / †0 †00 is small, so by .Opt/ and .s s/; †00 00 †00 is small, so 00 †00 is big, and likewise 0 †0 is big. t u
Discussion We compare these rules to probability defined size. Let “big” be defined by “more than 50%”. If …X 0 and …X 00 have three elements each, then subsets of …X 0 or …X 00 of card 2 are big. But taking the product may give 4=9 < 1=2: So the product rule “big big D big” will not hold there. One direction will hold, of course. Next, we discuss the prerequisite † D †0 †00 : Consider the following example: Example 5.2.1. Take a language of five propositional variables, with X 0 WD fa; b; cg; X 00 WD fd; eg: Consider the model set † WD f˙a ˙ b ˙ cde; a
172
5 Laws About Size and Interpolation in Non-monotonic Logics
b c d ˙ eg; i.e., of eight models of de and two models of d: The models of de are 8/10 of all elements of †; so it is reasonable to call them a big subset of †: But its projection on X 00 is only 1/3 of †00 : So we have a potential decrease when going to the coordinates. This shows that weakening the prerequisite about † as done in .S 2/ is not innocent. Remark 5.2.2. When we set small sets to 0, big sets to 1, we have the following Boolean rules for filters: (1) (2) (3) (4) (5)
0C0D0 1Cx D1 0 D 1; 1 D 0 0x D0 1 1 D 1:
There are no such rules for medium size sets, as the union of two medium size sets may be big, but also stay medium. Such multiplication rules capture the behaviour of Reiter defaults and of defeasible inheritance.
5.2.3 Hamming Relations and Distances 5.2.3.1 Hamming Relations and Multiplication of Size We now define Hamming relations in various flavours, and then (see Proposition 5.2.4 (page 173)) show that (smooth) Hamming relations generate a notion of size which satisfies our conditions, defined in Definition 5.2.1 (page 169). Corollary 5.3.4 (page 184) will put our results together, and show that (smooth) Hamming relations generate preferential logics with interpolation. We will conclude this section by showing that our conditions . 1/ and . 2/ essentially characterise Hamming relations. Note that we redefine Hamming relations in Section 6.3.1.3 (page 205), as already announced in Section 1.6.4.4 (page 22). Definition 5.2.2. We abuse notation, and define a relation on …X 0 ; …X 00 ; and …X 0 …X 00 : (1) Define x y W, x y and x ¤ yI thus iff or D : (2) We say that a relation satisfies .GH 3/ iff .GH 3/ 0 00 0 00 , 0 0 and 00 00 : (Thus, 0 00 0 00 iff 0 00 0 00 and . 0 0 or 00 00 /:/ (3) Call a relation a GH (D general Hamming) relation iff the following two conditions hold:
5.2 Laws About Size
173
.GH1/ 0 0 ^ 00 00 ^ . 0 0 _ 00 00 / ) 0 00 0 00 (where 0 0 iff 0 0 or 0 D 0 /; 0 00 .GH 2/ 0 00 ) 0 0 _ 00 00 : .GH 2/ means that some compensation is possible, e.g., 0 0 might be the case, but 00 00 wins in the end, so 0 00 0 00 : We use .GH / for .GH1/ C .GH 2/: Example 5.2.2. The circumscription relation satisfies .GH 3/ with :p p and V V ˙qi ˙qi0 iff 8i.˙qi ˙qi0 /: Remark 5.2.3. (1) The independence makes sense because the concept of models, and thus the usual interpolation for classical logic, rely on the independence of the assignments. (2) This corresponds to social choice for many independent dimensions. (3) We can also consider such factorisation as an approximation: we can do part of the reasoning independently. Definition 5.2.3. Given a relation ; define as usual a principal filter F .X / generated by the -minimal elements: .X / WD fx 2 X W :9x 0 x:x 0 2 X g; F.X / WD fA X W .X / Ag: The following proposition summarizes various properties for the different Hamming relations: Proposition 5.2.4. Let …X D …X 0 …X 00 ; † …X: (1) Let be a smooth relation satisfying .GH 3/. Then . 2/ holds, and thus .S 2/ by Proposition 5.2.1 (page 170), (2). (2) Let again †0 WD † X 0 ; †00 WD † X 00 : Let be a smooth relation satisfying .GH 3/. Then .†0 / .†00 / † ) .†/ D .†0 / .†00 /: (Here † D †0 †00 will not necessarily hold.) (3) Let again †0 WD † X 0 ; †00 WD † X 00 : Let be a relation satisfying .GH 3/, and † D †0 †00 : Then . 1/ holds, and thus, by Proposition 5.2.1 (page 170), (2), .S 1/ holds. (4) Let be a smooth relation satisfying .GH 3/I then . 3/ holds, and thus by Proposition 5.2.1 (page 170), (2), .S 3/ holds. (5) . 1/ and . 2/ and the usual axioms for smooth relations characterize smooth relations satisfying .GH 3/: (6) Let , 62 .f; g/ and be smooth. Then satisfies . 1/ (or, by Proposition 5.2.1 (page 170), equivalently .s s// iff is a GH relation. (7) Let 0 †0 ; 00 †00 ; 0 00 †0 †00 be small; let .GH 2/ hold; then 0 †0 is small or 00 †00 is small. (8) Let 0 †0 be small, 00 †00 I let .GH1/ hold; then 0 00 †0 †00 is small.
174
5 Laws About Size and Interpolation in Non-monotonic Logics
Proof. (1) Suppose .†/ and 0 2 † X 0 X 0 I we show 0 62 .† X 0 /: Let D 0 00 2 †I then 62 ; so 62 .†/: So here is ; 2 .†/ by smoothness. Let D 0 00 : We have 0 0 by .GH 3/: 0 D 0 cannot be, as
0 2 X 0 ; and 0 62 X 0 : So 0 0 ; and 0 62 .† X 0 /: (2) “ ”: Let 0 2 .†0 /; 00 2 .†00 /: By the prerequisite, 0 00 2 †: Suppose 0 00 I then 0 0 or 00 00 ; a contradiction. “”: Let 2 .†/I suppose 0 62 .†0 / or 00 62 .†00 /: So there are 0 0 ; 00 00 with 0 2 .†0 /; 00 2 .†00 / by smoothness. Moreoever, 0 0 or 00 00 : By the prerequisite, 0 00 2 † and 0 00 ; so 62 .†/: (3) “ ”: As in (2), the prerequisite holds trivially. “”: As in (2), but we do not need 0 2 .†0 /; 00 2 .†00 /; as 0 00 will be in † trivially. So smoothness is not needed. (4) Let again †00 D † X 00 : Let WD …X 0 †00 ; D 0 00 2 . /: Suppose 00 62 .†/ X 00 : There cannot be any ; 2 †; by † : So 62 †; but 00 2 †00 ; so there is 2 † 00 D 00 : As is not minimal, there must be minimal D 0 00 ;
2 † by smoothness. As is minimal, 00 ¤ 00 ; and as ; 00 00 by .GH 3/: By the prerequisite, 0 00 2 ; and 0 00 ; a contradiction. Note that smoothness is essential. Otherwise, there might be an infinite descending chain i below ; all with i00 D 00 ; but none below : (5) If is smooth and satisfies .GH 3/; then . 1/ and . 2/ hold by (1) and (3). For the converse define as usual W, 62 .f; g/: Let D 0 00 ; D 0 00 : We have to show iff 0 0 and 00 00 and . 0 0 or 00 00 /: “(”: Suppose 0 0 and 00 00 : Then .f 0 ; 0 g/ D f 0 g; and .f 00 ; 00 g/ D f 00 g (either 00 00 or 00 D 00 ; so in both cases .f 00 ; 00 g/ D f 00 g/: As 0 62 .f 0 ; 0 g/; 62 .f 0 ; 0 g f 00 ; 00 g/ D (by . 1// .f 0 ; 0 g/ .f 00 ; 00 g/ D f 0 g f 00 g D fg; so by smoothness : “)”: Conversely, if ; so WD fg D .†/ for † WD f; g; so by . 2/ .† X 0 / D .f 0 ; 0 g/ X 0 D f 0 g; so 0 0 I analogously .† X 00 / D .f 00 ; 00 g/ X 00 D f 00 g; so 00 00 ; but both cannot be equal. (6) (6.1) . 1/ entails the GH relation conditions. .GH1/ W Suppose 0 0 and 00 00 : Then 0 62 .f 0 ; 0 g/ D f 0 g; and .f 00 ; 00 g/ D f 00 g (either 00 00 or 00 D 00 ; so in both cases .f 00 ; 00 g/ D f 00 g/: As 0 62 .f 0 ; 0 g/; 0 00 62 .f 0 ; 0 g f 00 ; 00 g/ D.1/ .f 0 ; 0 g/ .f 00 ; 00 g/ D f 0 g f 00 g D f 0 00 g; so by smoothness 0 00 0 00 : .GH 2/ W Let X WD f 0 ; 0 g; Y WD f 00 ; 00 g; so X Y D f 0 00 ; 0 00 ; 0 00 ; 0 00 g: Suppose 0 00 0 00 ; so 0 00 62 .X Y / D.1/ .X / .Y /: If 0 6 0 ;
5.2 Laws About Size
175
then 0 2 .X /I likewise if 00 6 00 ; then 00 2 .Y /; so 0 00 2 .X Y /; a contradiction. (6.2) The GH relation conditions generate . 1/: .X Y / .X / .Y / W Let 0 2 X; 00 2 Y; 0 00 62 .X / .Y /I then 0 62 .X / or 00 62 .Y /: Suppose 0 62 .X /I let 0 2 X; 0 0 I so by condition .GH1/ 0 00 0 00 ; so 0 00 62 .X Y /: .X / .Y / .X Y / W Let 0 2 X; 00 2 Y; 0 00 62 .X Y /I so there is 0 00 0 00 ; 0 2 X; 00 2 Y; so by .GH 2/ either 0 0 or 00 00 ; so 0 62 .X / or 00 62 .Y /; so 0 00 62 .X / .Y /: (7) Suppose 0 †0 is not small; so there is 0 2 0 and no 0 2 †0 with 0 0 : Fix this 0 : Consider f 0 g 00 : As 0 00 †0 †00 is small, there is for each 0 00 ; 00 2 00 some 0 00 2 †0 †00 ; 0 00 0 00 : By .GH 2/ 0 0 or 00 00 ; but 0 0 was excluded, so for all 00 2 00 there is 00 2 †00 with 00 00 ; so 00 †00 is small. (8) Let 0 2 0 I so there is 0 2 †0 and 0 0 : By .GH1/; for any 00 2 00 ; 0 00 0 00 ; so no 0 00 2 0 00 is minimal. t u Example 5.2.3. Even for smooth relations satisfying .GH 3/, the converse of .2/ is not necessarily true: Let 0 0 ; 00 00 ; † WD f; gI then .†/ D †; but .†0 / D f 0 g; .†00 / D f 00 g; so .†/ ¤ .†0 / .†00 /: We need the additional assumption that .†0 / .†00 / †I see Proposition 5.2.4 (page 173) (2). Example 5.2.4. The following are examples of GH relations: Define on all components Xi a relation i : (1) The set variant Hamming relation: Let the relation be defined on …fXi W i 2 I g by iff for all j j j j ; and there is at least one i such that i i i : (2) The counting variant Hamming relation: Let the relation be defined on …fXi W i 2 I g by iff the number of i such that i i i is bigger than the number of i such that i i i : (3) The weighed counting Hamming relation: Like the counting relation, but we give different (numerical) importance to different i: E.g., 1 1 may count 1, 2 2 may count 2, etc. t u
Note Note that . 1/ results in a strong independence result in the second scenario: Let 0 0 I then 00 00 for all 00 : Thus, whether f 00 g is small or medium size (i.e., 00 2 .†0 //; the behaviour of † f 00 g is the same. This we do not have in the first scenario, as small sets may behave very differently from medium size sets. (But, still, their internal structure is the same; only the minimal elements change.)
176
5 Laws About Size and Interpolation in Non-monotonic Logics
When . 2/ holds, if 0 0 and ¤ ; then ; i.e., we need not have 0 D 0:
5.2.3.2 Hamming Distances and Revision This short section is mainly intended to put our work in a broader perspective, by showing a connection of Hamming distances to modular revision as introduced by Parikh and his co-authors. The main result here is Corollary 5.2.7 (page 177). We will not go into details of motivation here, and refer the reader to [Par96] for further discussion. Thus, we have modular distances and relations, i.e., Hamming distances and relations, we have modular revision as described below, and we have modular logic, which has the (semantical) interpolation property. We want to point out here in particular this cross-reference from modular revision to modular logic, i.e., logic with interpolation. We recall the following: Definition 5.2.4. Given a distance d; define for two sets X; Y X j Y WD fy 2 Y W 9x 2 X.:9x 0 2 X; y 0 2 Y:d.x 0 ; y 0 / < d.x; y//g: We assume that X j Y ¤ ; if X; Y ¤ ;: Note that this is related to the consistency axiom of AGM theory revision: revising by a consistent formula gives a consistent result. The assumption may be wrong due to infinite descending chains of distances. Definition 5.2.5. Given j; we can define an AGM revision operator as follows: T WD T h.M.T / j M.// where T is a theory, and T h.X / is the set of formulas which hold in all x 2 X: It was shown in [LMS01] that a revision operator thus defined satisfies the AGM revision postulates. Definition 5.2.6. Let d be an abstract distance on some product space X Y and its components. (We require of distances only that they be comparable, that d.x; y/ D 0 iff x D y; and that d.x; y/ 0:/ d is called a generalized Hamming distance .GHD/ iff it satisfies the following two properties: .GHD1/ d.; / d.˛; ˇ/ and d. 0 ; 0 / d.˛0 ; ˇ 0 / and .d.; / < d.˛; ˇ/ or d. 0 ; 0 / < d.˛ 0 ; ˇ 0 // ) d. 0 ; 0 / < d.˛˛0 ; ˇˇ0 /; .GHD2/ d. 0 ; 0 / < d.˛˛0 ; ˇˇ0 / ) d.; / < d.˛; ˇ/ or d. 0 ; 0 / < d.˛ 0 ; ˇ 0 /: (Compare this definition to Definition 5.2.2 (page 172).) We have a result analogous to the relation case: Fact 5.2.5. Let j be defined by a generalized Hamming distance; then j satisfies (1) .j / .†1 †01 / j .†2 †02 / D .†1 j †2 / .†01 j †02 /;
5.2 Laws About Size
177
(2) .†01 j †02 / .†001 j †002 / †2 and .†02 j †01 / .†002 j †001 / †1 ) .†1 / j .†2 / D .†01 j †02 / .†001 j †002 / if the distance is symmetric (where †i is not necessarily †0i †00i ; etc.). Proof. (1) and (2). “”: Suppose d. 0 ; 0 / is minimal. If there is an ˛ 2 †1 ; ˇ 2 †2 such that d.˛; ˇ/ < d.; /; then d.˛ 0 ; ˇ 0 / < d. 0 ; 0 / by .GHD1/; so d.; / and d. 0 ; 0 / have to be minimal. “ ”: For the converse, suppose d.; / and d. 0 ; 0 / are minimal, but d. 0 ; 0 / is not, so d.˛˛ 0 ; ˇˇ 0 / < d. 0 ; 0 / for some ˛˛ 0 ; ˇˇ 0 I then d.˛; ˇ/ < d.; / or d.˛ 0 ; ˇ 0 / < d. 0 ; 0 / by .GHD2/; a contradiction. t u These properties translate to logic as follows: Corollary 5.2.6. If and are defined on a separate language from that of 0 and 0 ; and the distance satisfies .GHD1/ and .GHD2/; then for revision the following holds: . ^ 0 / . ^ 0 / D . / ^ . 0 0 /: Corollary 5.2.7. By Corollary 5.2.6 (page 177), Hamming distances generate decomposable revision operators a la Parikh (see [Par96]) also in the generalized form of variable K and : We conclude with a small result on partial (semantical) revision: Fact 5.2.8. Let j be defined by a Hamming distance, then …X j † ) …X 0 j .† X 0 / X 0 : (Recall that …0 is the restriction of … to X 0 :/ Proof. Let t 2 † X 0 X 0 I we show t 62 …X 0 j .† X 0 /: Let 2 † be such that 0 D tI then 62 (otherwise t 2 X 0 /; so 62 …X j †; so there is an ˛ D ˛ 0 ˛ 00 2 …X; ˇ D ˇ 0 ˇ 00 2 †; with d.˛; ˇ/ minimal, so d.˛; ˇ/ < d.; / for all 2 …X: If d. 0 ; 0 / were minimal for some ; then we would consider 0 ˛ 00 ; 0 ˇ 00 I then d.˛0 ˛ 00 ; ˇ 0 ˇ 00 / < d. 0 ˛ 00 ; 0 ˇ 00 / is impossible by (GHD2), so 0 ˇ 00 2 …X j †; so 0 ˇ 00 2 ; and t 2 X 0 ; a contradiction. t u
5.2.3.3 Discussion of Representation It would be nice to have a representation result like the one for Hamming relations; see Proposition 5.2.4 (page 173), (5). But this is impossible, for the following reason: In constructing the representing distance from revision results, we made arbitrary choices (see the proofs in [LMS01] or [Sch04]). I.e., we choose sometimes arbitrarily d.x; y/ d.x 0 ; y 0 / when we do not have enough information to decide. (This is an example of the fact that the problem of “losing ignorance” should not be
C
.iM / A B 2 I.X / ) A 2 I.X / .eM I/ X Y ) I.X / I.Y /; X Y ) F .Y / \ P.X / F .X /
J 0 small
- (Filter)
-
-
Relation property
.
1 2 F .†1 /; 2 2 F .†2 / ) 1 2 2 F .†1 †2 /
Non-monotonic logic 1 2 F .†1 / ) 1 †2 2 F .†1 †2 / 1 2 I.†1 / ) 1 2 2 I.†1 †2 /
Algebraic property .i †i /
Multiplication laws
Rational Monotony
˛ j6 L1 :ˇ; ˛ 0 j L2 ˇ 0 ) ˛ ^ ˛ 0 j6 L :.ˇ ^ ˇ 0 /
˛ j L1 ˇ; ˛ 0 j L2 ˇ 0 ) ˛ ^ ˛ 0 j L ˇ ^ ˇ 0
˛ j L1 ˇ; ˇ 0 `L2 ˛ 0 ) ˛ ^ ˛0 j .ˇ ^ ˛ 0 / _ .˛ ^ ˇ 0 /
˛ j Li ˇ ) ˛ j L ˇ
bb ,b . 1/
.GH /
j ı j
Scenario 2 . symmetrical, only 1 side shown) (see Diagram 5.2.2 (page 168)) Logical property Interpolation .˛; ˇ in L1 ; ˛0 ; ˇ 0 in L2 Multiplic. Relation InterL D L1 [ L2 (disjoint)) law property polation
ranked 1 2 MC .†1 /; 2 2 MC .†2 / ˛ j6 L1 :ˇ; ˛ 0 j6 L2 :ˇ0 ) ) ˛ ^ ˛ 0 j6 L :.ˇ ^ ˇ 0 / 1 2 2 MC .†1 †2 / ˛ j ˇ ) ˛ L1 j ˇ L1 . 1/ C (GH3) ` ı j and . 2/ 0 0 ˛ j L1 ˇ; ˛ j L2 ˇ ) ˛ ^ ˛ 0 j L ˇ ^ ˇ 0 ˛ ^ ˛ 0 j ˇ ^ ˇ0 , f orget.J 0 / ˛ j ˇ; ˛ 0 ` ˇ 0 Theory revision .j / W .GHD/ . ^ 0 / . ^ 0 / `
.†1 †01 / j .†2 †02 / D . ^ 0 / . ^ 0 / D ) 0 0 `
.†1 j †2 / .†01 j †02 / . / ^ . 0 0 / ; in J; 0 ; 0 ; in L J
˛ j6 :ˇ; ˛ ^ ˇ j - (Filter) ) ˛ j6 :ˇ _ :
˛ j ˇ; ˛ ^ ˇ j6 : - (Filter) 1 2 F .†1 /; 2 2 MC .†2 / ) ) ˛ j6 :ˇ _ : 1 2 2 MC .†1 †2 /
˛ j ˇ; ˛ ^ ˇ j ) ˛ j
˛ j :ˇ ) ˛ j :ˇ _ ˛ ^ ˇ j : ) ˛ j :ˇ _ :
trivial trivial
Scenario 1 (see Diagram 5.2.1 (page 166)) Corresponding algebraic Logical property addition property
.< ! s/; .M! / (3) A 2 F .X /; X 2 F .Y / ) A 2 F .Y / C b m ) m .< ! s/; .M! / (2) A 2 MC .X /; X 2 F .Y / ) A 2 MC .Y / C mb )m .< ! s/; .M! / (1) A 2 F .X /; X 2 MC .Y / ) A 2 MC .Y / mm)m .MCC / A 2 MC .X /; X 2 MC .Y / ) A 2 MC .Y / b b , b; pr.b/ D b . 2/
bb )b . 1/
sx )s
x 1 )x 1x )x x s ) s
Multiplication law
Table 5.3 Multiplication laws
178 5 Laws About Size and Interpolation in Non-monotonic Logics
5.2 Laws About Size
179
underestimated; see e.g. [GS08f].) As we do not follow the same procedure for all cases, there is no guarantee that the different representations will fit together. Of course, it might be possible to come to a uniform choice, and one could then attempt a representation result. This is left as an open problem.
5.2.4 Summary of Properties We summarize in this section properties related to multiplicative laws. They are collected in Table 5.3 (page 178). pr.b/ D b means the projection of a big set on one of its coordinates is big again. Note that A B X Y big ) A X big is intuitively better justified than the other direction, as the proportion might increase in the latter, and decrease in the former. See the table “Rules on size”, Section 5.2.1 (page 160), “increasing proportions”.
5.2.5 Language Change in Classical and Non-monotonic Logic This section is very short, but it seems important. It shows that, by redefining the language, we can make problems modular iff they have the right structure, irrespective of the way they are formulated. Fact 5.2.9. We can obtain factorization by language change, provided cardinalities permit this. Proof. Consider k variables; suppose we have p D m n positive instances, and 0 00 that we can divide k into k 0 and k 00 such that 2k m; 2k n; then we can factorize: Choose m sequences of 0/1 of length k 0 ; n sequences of length k 00 : They will code the positive instances: there are p D m n pairs of the chosen sequences. Take any bijection between these pairs and the positive instances, and continue the bijection arbitrarily between other pairs and negative instances. We can do the same also for two sets corresponding to K; to have a common factorization; they both have to admit common factors like m; n above. We then choose first the pairs for K; then for ; then the rest. t u Example 5.2.5. Consider p D 3; and let abc, a:bc; a:b:c; :abc; :a:b:c; :ab:c be the 6 D 2 3 positive cases, ab:c; :a:bc be the negative ones. (It is a coincidence that we can factorize positive and negative cases, probably iff one of the factors is the full product, here 2, it could also be 4, etc.)
180
5 Laws About Size and Interpolation in Non-monotonic Logics
We divide the cases by three new variables, grouping them together into positive and negative cases. a0 is indifferent; we want this to be the independent factor; the negative ones will be put into :b 0 :c 0 : The procedure has to be made precise still. “(n)” will mark the negative cases. Let a0 code the set abc, a:bc; a:b:c; ab:c (n). Let :a0 code :a:bc (n), :abc; :a:b:c; :ab:c: Let b 0 code abc; a:bc; :a:b:c; :ab:c: Let :b 0 code a:b:c; ab:c (n), :a:bc (n), :abc: Let c 0 code abc; a:b:c; :abc; :a:b:c: Let :c 0 code a:bc; ab:c (n), :a:bc (n), :ab:c: Then the six positive instances are fa0 ; :a0 g fb 0 c 0 ; b 0 :c 0 ; :b 0 c 0 gI the negative ones are fa0 ; :a0 g f:b 0 :c 0 g: As we have three new variables, we code again all possible cases, so expressivity is the same. t u The same holds for non-monotonic logic. We give an example: Example 5.2.6. Suppose we have the rule that “positive is better than negative” (all other things being equal). Then, for two variables, a and b; we have the comparisons ab a:b :a:b; and ab :ab :a:b: Suppose now we are given the situation :cd c:d cd and :cd :c:d cd; which has the same order structure, but with negations not fitting. We put c:d and :cd into a new variable a 0 ; cd and :c:d into :a0 ; :cd and :c:d into b 0 ; and c:d and cd into :b 0 : Then a0 b 0 corresponds to :cd; a0 :b 0 to c:d; :a0 b 0 to :c:d; :a0 :b 0 to cd — and we have the desired structure. Thus, if the geometric structure is possible, then we can change the language and obtain the desired pattern. But we cannot obtain by language change a pattern of the type ab a:b without any other comparison if it is supposed to be based on a componentwise comparison. We summarize: We can cut the model set as we like: Choose half to go into p0 ; half into :p0; again half of p0 into p0 ^ p1 ; half into p0 ^ :p1 ; etc.
5.3 Semantic Interpolation for Non-monotonic Logic 5.3.1 Discussion This section is perhaps the central section of the book. We discuss here the full non-monotonic case, i.e., downward and upward. We consider here a non-monotonic logic j : We look at the interpolation problem in three ways.
5.3 Semantic Interpolation for Non-monotonic Logic
181
Given j ; there is an interpolant ˛ such that (1) j ˛ ` ; see Section 5.3.2 (page 181), (2) ` ˛ j ; see Section 5.3.3 (page 183), (3) j ˛ j ; see Section 5.3.4 (page 184). The first variant will be fully characterized below; interpolation in this sense exists iff “does not make things more complicated”; see Proposition 5.3.2 (page 181). The second and third variants have no full characterization at the time of this writing (to the authors’ best knowledge), but are connected to very interesting properties about multiplication of size and componentwise independent relations. As can be expected, the condition for the second variant is stronger than the one for the third variant, and its size version shows that this condition will often be too strong; see Example 5.2.1 (page 171). But the condition will be more natural in its relation version. All results and proofs are not difficult technically. We begin with the following negative result: Example 5.3.1. Full non-monotonic logics, i.e., down and up, does not necessarily have interpolation. Consider the model order pq p:q :p:q :pq: Then :p j :q; there are no common variables, and TRUE j q (and, of course, :p j6 FALSE/: (The full consequence of :p is :p:q; so this has trivial interpolation.)
5.3.2 Interpolation of the Form j ˛ ` We show here that interpolation of the type j ˛ ` exists iff, roughly, .X / is not more complicated than X; more precisely, iff I.†/ I..†//: Fact 5.3.1. Let var./ be the set of relevant variables of : Let † …X; var.˛/ \ var.ˇ/ D ;; var.ˇ/ \ R.†/ D ;; ˇ not a tautology; then † M.˛ _ ˇ/ ) † M.˛/: Proof. Suppose not, so there a is 2 † such that ˆ ˛ _ ˇ; 6ˆ ˛: As ˇ is not a tautology, there is an assignment to var.ˇ/ which makes ˇ wrong. Consider such that D except on var.ˇ/; where makes ˇ wrong using this assignment. By var.˛/ \ var.ˇ/ D ;; ˆ :˛: By var.ˇ/ \ R.†/ D ;; 2 †: So 6ˆ ˛ _ ˇ for some 2 †; a contradiction. t u Proposition 5.3.2. We use here the normal form (conjunctions of disjunctions). Consider a finite language. Let a semantical choice function be given, as discussed in Section 1.6.1 (page 12), defined for sets of sequences (read models). j has interpolation iff for all †; I.†/ I..†// holds.
182
5 Laws About Size and Interpolation in Non-monotonic Logics
In the infinite case, we need as an additional prerequisite that .†/ is definable if † is. Proof. Work with reformulations of †; etc. which use only essential .D relevant) variables. “)”: Suppose the condition is wrong. Then X WD I.†/ I..†// D I.†/ \ R..†// ¤ ;: Thus there is some 0 2 .†/ R.†/ which cannot be continued by some choice in X [ .I.†/ \ I..†/// in .†/; i.e., 0 62 .†/: We first address the finite case: We identify models with their formulas. Consider the formula WD 0 ! : D : 0 _ : : We have T h.†/ j ; as .†/ M./: Suppose †00 is a semantical interpolant for † and : So .†/ †00 M./; and †00 does not contain any variables in as essential variables. By Fact 5.3.1 (page 181), .†/ †00 M.: 0 /; but 0 2 .†/ R.†/; a contradiction. We turn to the infinite case. Consider again 0 : As 0 62 .†/; and .†/ is definable, there is some formula which holds in .†/; but fails in 0 : Thus, T h.†/ j : Write as a disjunction of conjunctions. Let †00 be an interpolant for † and M./: Thus .†/ †00 M./; and 0 62 M./; so .†/ †00 M./ M.: 0 _ : /; so †00 M.: 0 / by Fact 5.3.1 (page 181). So .†/ ˆ : 0 ; a contradiction, as 0 2 .†/ R.†/: (More precisely, we have to argue here with not necessarily definable model sets.) “(”: Let I.†/ I..†//: Let † j †0 ; i.e., .†/ †0 : Write .†/ as a (possibly infinite) conjunction of disjunctions, using only relevant variables. Form †00 from .†/ by omitting all variables in this description which are not in R.†0 /: Note that all remaining variables are in R..†// R.†/; so †00 is a candidate for interpolation. See Diagram 5.3.1 (page 182). (1) .†/ †00 W Trivial. (2) †00 †0 W Let 2 †00 : Then there is a 2 .†/ †0 such that R.†0 / D R.†0 /; so 2 †0 :
Σ = M(φ) Σ = M(ψ)
Σ = M(α) μ(Σ) = μ(φ)
Non-monotonic interpolation, φ ∼ α ψ
Diagram 5.3.1
5.3 Semantic Interpolation for Non-monotonic Logic
183
A shorter argument is as follows: .†/ ˆ M.†0 / has a semantical interpolant by Section 4.2 (page 119), which is by the prerequisite also an interpolant for † and †0 . It remains to show in the infinite case that †00 is definable. This can be shown as in Proposition 4.4.1 (page 134). t u
5.3.3 Interpolation of the Form ` ˛ j This situation is much more interesting than the last one, discussed in Section 5.3.2 (page 181). In this section, and the next one, Section 5.3.4 (page 184), we connect abstract multiplication laws for size to interpolation. To the best of our knowledge, such multiplication laws are considered here for the first time, and therefore so is their connection to interpolation problems. We introduced two sets of three conditions about abstract size (see Definition 5.2.1 (page 169)), and then showed in Proposition 5.2.1 (page 170) that both sets are equivalent. We show now that the first two conditions, or the last condition entails interpolation; see Proposition 5.3.3 (page 183). Recall that in preferential structures, size is generated by a relation. A B is a big subset iff A contains all minimal elements of B (with respect to this relation). Hamming relations (see Definition 5.2.2 (page 172)) generate a notion of size which satisfies our multiplicative conditions (if they are smooth). Thus, if a preferential logic is defined by a smooth Hamming relation, it has semantical interpolation (in our sense here). This is summarized in Corollary 5.3.4 (page 184). Proposition 5.3.3. We assume definability as shown in Proposition 4.4.1 (page 134). Interpolation of the form ` ˛ j exists if (1) both .S 1/ and .S 2/ hold or (2) .S 3/ holds when ˇ j is defined by ˇ j W, .ˇ/ D .M.ˇ// M./; and .X / is the generator of the principal filter over X: (We saw in Example 5.2.1 (page 171) that .S 2/; and thus also .S 3/; will often be too strong.) Proof. Let † WD M./; WD M. /; and X 0 be the set of variables only in I so D …X 0 X 00 ; where …X 0 D …X 0 : Set ˛ WD T h.…X 0 †00 /; where †00 D † X 00 : Note that variables only in are automatically taken care of, as †00 can be written as a product without mentioning them. See Diagram 5.3.2 (page 184).
184
5 Laws About Size and Interpolation in Non-monotonic Logics
Σ =M(φ)
M(α) = ΠX × Σ X
Γ = M(ψ)
X X Non-monotonic interpolation, φ α
ψ
Diagram 5.3.2
By the prerequisite, .†/ I we have to show .…X 0 †00 / : (1) .…X 0 †00 / D .…X 0 /.†00 / by .S 1/ and Proposition 5.2.1 (page 170), (2). By .†/ ; .S 2/; and Proposition 5.2.1 (page 170), (2), .†00 / D .† X 00 / X 00 ; so .…X 0 †00 / D .…X 0 /.†00 / .…X 0 / X 00 : (2) .…X 0 †00 / X 00 .†/ X 00 X 00 by .S 3/ and Proposition 5.2.1 (page 170), (2). So .…X 0 †00 / …X 0 ..…X 0 †00 / X 00 / …X 0 . X 00 / D : t u The following corollary puts our results together. Corollary 5.3.4. Interpolation in the form ` ˛ j exists when j is defined by a smooth Hamming relation, more precisely, a smooth relation satisfying .GH 3/. Proof. We give two proofs: (1) By Proposition 5.2.4 (page 173), .S 1/ and .S 2/ hold. Thus, by Proposition 5.3.3 (page 183) (1), interpolation exists. (2) By Proposition 5.2.4 (page 173), .S 3/ holds, so by Proposition 5.3.3 (page 183), (2), interpolation exists. t u
5.3.4 Interpolation of the Form j ˛ j The following result, together with Proposition 4.2.3 (page 122), is perhaps the main result of the book. The conditions are natural, and not too strong, and the connection between those multiplicative properties and interpolation gives quite deep insights into the basics of non-monotonic logics.
5.3 Semantic Interpolation for Non-monotonic Logic
185
φ ψ
μ(φ)
J
J
J”
Non-monotonic interpolation, φ α ψ Double lines: interpolant ΠJ × (μ(φ) J ) × ΠJ ” Alternative interpolants (in center part): φ J or (φ ∧ ψ ) J
Diagram 5.3.3
Proposition 5.3.5. .1/ entails semantical interpolation of the form j ˛ j in two-valued non-monotonic logic generated by minimal model sets. (As the model sets might not be definable, syntactic interpolation does not follow automatically.) Proof. Let the product be defined on J [ J 0 [ J 00 (i.e., J [ J 0 [ J 00 is the set of propositional variables in the intended application). Let be defined on J [ J 0 ; on J 0 [ J 00 : See Diagram 5.3.3 (page 185). We abuse notation and write j † if ./ †: As usual, ./ abbreviates .M.//: For clarity, even if it clutters up notation, we will be precise about where is formed. Thus, we write J [J 0 [J 00 .X / when we take the minimal elements in the full product, J .X / when we consider only the product on J; etc. XJ will be shorthand for …fXj W j 2 J g: Let j ; i.e., J [J 0 [J 00 ./ M. /: We show that XJ .J [J 0 [J 00 ./ J 0 /XJ 00 is a semantical interpolant, i.e., that J [J 0 [J 00 ./ XJ .J [J 0 [J 00 ./ J 0 / XJ 00 ; and that J [J 0 [J 00 .XJ .J [J 0 [J 00 ./ J 0 / XJ 00 / M. /: The first property is trivial; we turn to the second. (1) As M./ D M./ .J [ J 0 / XJ 00 ; J [J 0 [J 00 ./ D J [J 0 .M./ .J [ J 0 // J 00 .XJ 00 / by . 1/: (2) By . 1/ again, J [J 0 [J 00 .XJ .J [J 0 [J 00 ./ J 0 / XJ 00 / D J .XJ / J 0 .J [J 0 [J 00 ./ J 0 / J 00 .XJ 00 /: So it suffices to show J .XJ / J 0 .J [J 0 [J 00 ./ J 0 / J 00 .XJ 00 / ˆ :
186
5 Laws About Size and Interpolation in Non-monotonic Logics
Proof: Let D J J 0 J 00 2 J .XJ / J 0 .J [J 0 [J 00 ./ J 0 / J 00 .XJ 00 /; so J 2 J .XJ /: By J 0 .J [J 0 [J 00 ./ J 0 / J [J 0 [J 00 ./ J 0 ; there is a 0 D J0 J0 0 J0 00 2 J [J 0 [J 00 ./ such that J0 0 D J 0 ; i.e., 0 D J0 J 0 J0 00 : As 0 2 J [J 0 [J 00 ./; 0 ˆ : By (1) and J 00 2 J 00 .XJ 00 /; also J0 J 0 J 00 2 J [J 0 [J 00 ./; so also 0 J J 0 J 00 ˆ : But does not depend on J; so also D J J 0 J 00 ˆ : t u 5.3.4.1 Remarks for the Converse: from Interpolation to . 1/ Example 5.3.2. We show here in (1.1) and (1.2) that half of the condition . 1/ is not sufficient for interpolation, and in (2) that interpolation may hold, even if . 1/ fails. When looking closer, the latter is not surprising: of subproducts may be defined in a funny way, which has nothing to do with the way on the big product is defined. Consider the language based on p; q; r: We define two orders: (a) on sequences of length 3 by :p:q:r p:q:r; and leave all other sequences of length 3 -incomparable. (b) < on sequences of the same length by < iff there is a :x in ; an x in ; but no y in ; :y in : E.g., :p < p; :pq < pq; :p:q < pq; but :pq 6< p:q: Work now with : Let D :q ^ :r; D :p ^ :q; so ./ D :p ^ :q ^ :r; and j : Suppose there is an ˛; j ˛ j ; ˛ written with q only, so ˛ is equivalent to FALSE, TRUE, q; or :q: j6 FALSE; j6 q: TRUE j6 ; :q j6 : Thus, there is no such ˛; and j has no interpolation. We show in (1.1) and (1.2) that we can make both directions of . 1/ true separately, so they do not suffice to obtain interpolation. (1.1) We make .X Y / .X / .Y / true, but not the converse. Take the order on sequences of length 3 as described above. Do not order any sequences of length 2 or 1, i.e., is there always identity. Thus, .X Y / X Y D .X / .Y / holds trivially. (1.2) We make .X Y / .X / .Y / true, but not the converse. We order all sequences of length 1 or 2 by 0 and m.p/ D 1 for some p; and define m for this m: As the language is finite, the set of such m is finite, and if fX D f ; then f^Vfm W9p:m.p/D1;fX .m/>0g will define f.X / : A model is eliminated iff it contains m.p/ D 1 for some p; or there is some m0 m which contains some such p: So, if we can show this for m with m.p/ D 1 for some p m .m/ D 0; and m .m0 / D 0 if m m0 ; and perhaps m .m0 / if m0 contains some other m0 .p/ D 1; but m .m0 / D 2 for all other m0 ; then we are done. Consider now m with m.p/ D 1 for some p: Let p1 ; : : : ; pk be such that m.pi / D 0; and q1 ; : : : ; qn be such that m.qi / > 0: k may be 0, but not n: Define m WD ::p1 _ : : : _ ::pk _ :q1 _ : : : _ :qn : First, consider m: m.::pi / D 0 for all i; and also m.:qi / D 0 for all i; so m .m/ D 0; as desired. Second, suppose m m0 : As m.x/ D 0 iff m0 .x/ D 0; the values for pi did not change, so still m0 .::pi / D 0; and the values for qi may have changed, but not to 0, so still m0 .:qi / D 0; and m .m0 / D 0: Suppose now m 6 m0 : Case 1: m0 .pi / ¤ 0 for some i I then m0 .::pi / D 2; and we are done. Case 2: m0 .qi / D 0 for some i I then m0 .:qi / D 2; and we are done. Thus, if m .m0 / D 0; then m.x/ D 0 iff m0 .x/ D 0: So m.x/ can only differ from m0 .x/ on the qi ; and m0 .qi / ¤ 0: If for all qi ; 0 m .qi / D 2; then m m0 ; and we are done. If not, then m0 .qi / D 1 for some i; and it should be eliminated anyway. t u
5.3.6.3 The Approach with Models of Value 2 We first define formally what we want to do: Definition 5.3.2. (1) Set M2 ./ WD fm 2 M W m./ D 2 D TRUEg: (2) Set 2 ./ WD fm 2 M W m./ D 2; and m is an equilibrium model g: (3) Set j iff 2 ./ M2 . /: (4) Set ` iff 8m 2 M:m./ m. /: We will show that interpolation of type (a) ` ˛ j and (b) j ˛ ` may fail, but interpolation of type (c) j ˛ j will exist. Definability of the interpolant will be shown using the definability results for HT. We can use the techniques and results developed there (“neglecting” some variables), and see that the semantical interpolant is definable, so we have also syntactical interpolation. Example 5.3.3. (EQ has no interpolation of the form ` ˛ j We work with three variables, a; b; c: Consider † WD fh0; 2; 2i; h2; 1; 0i; h2; 2; 0ig:
.)
5.3 Semantic Interpolation for Non-monotonic Logic
191
By the above, and classical behaviour of “or” and “and”, † is definable by WD .:a ^ b ^ c/ _ .a ^ ::b ^ :c/; i.e., † D fm W m./ D 2g: Note that h2; 2; 0i is total, but h2; 1; 0i h2; 2; 0i; thus .†/ D fh0; 2; 2ig: So † j c D 2 (or † j c/: Let X 0 WD fa; bg; X 00 WD fcg: All possible interpolants must not contain a or b as essential variables, and they must contain †: The smallest candidate is …X 0 f0; 2g: But WD h0; 0; 0i 2 ; is total, and there cannot be any ; so 2 ./; so j6 c D 2: For completeness’ sake, we write all elements of W h0; 0; 0i h0; 0; 2i h0; 1; 0i h0; 1; 2i h0; 2; 0i h0; 2; 2i h1; 0; 0i h1; 0; 2i h1; 1; 0i h1; 1; 2i h1; 2; 0i h1; 2; 2i h2; 0; 0i h2; 0; 2i h2; 1; 0i h2; 1; 2i h2; 2; 0i h2; 2; 2i: Recall that no sequence containing 1 is total, and when we go from 2 to 1, we have a smaller model. Thus, ./ D fh0; 0; 0i; h0; 0; 2ig: Example 5.3.4. (EQ has no interpolation of the form j ˛ ` . ) Consider two variables, a; b; and † WD f0; 2g f0; 1; 2g: No containing 1 can be in .†/I as a matter of fact, .†/ D fh0; 0i; h2; 0ig: † is defined by a _ :a; .†/ is defined by .a _ :a/ ^ :b: So we have a _ :a j b _ :b; and even a _ :a j :b: The only possible interpolants are TRUE or FALSE. a _ :a j6 FALSE; and TRUE 6` :b: Fact 5.3.9. EQ has interpolation of the form j ˛ j
.
Proof. Let j ; i.e., 2 ./ M2 . /: We have to find ˛ such that 2 ./ M2 .˛/; and 2 .˛/ M2 . /: Let J D I./; J 00 D I. /: Consider X WD …J .2 ./ J 0 / …J 00 : By the same arguments (“neglecting” J and J 00 /; X is definable as M2 .˛/ for some ˛: Obviously, 2 ./ M2 .˛/: Consider now 2 .˛/I we have to show 2 .˛/ M2 . /: If 2 .˛/ D ;; we are done, so suppose there is m 2 2 .˛/: Suppose m 62 M2 . /: There is m0 2 2 ./; m0 J 0 D m J 0 : We use now C for concatenation. Consider m00 D .m J / C m0 .J 0 [ J 00 /: As m0 2 2 ./ M2 ./; and M2 ./ D …J M2 ./ .J 0 [ J 00 /; m00 2 M2 ./: m .J [ J 0 / D m00 .J [ J 0 /I thus by J 00 I. /; m00 62 M2 . /: Thus, m00 62 2 ./: So either there is n 2 M2 ./ such that n.y/ D 0 iff m00 .y/ D 0 and fy W n.y/ D 2g fy W m00 .y/ D 2g or m00 .y/ D 1 for some y 2 L: Suppose m00 .y/ D 1 for some y: y cannot be in J 0 [ J 00 ; as m00 .J 0 [ J 00 / D m0 .J 0 [ J 00 /; and m0 2 2 ./: y cannot be in J; as m00 J D m J; and m 2 2 .X /:
192
5 Laws About Size and Interpolation in Non-monotonic Logics
So there must be n 2 M2 ./ as above. Case 1: fy 2 J 0 [ J 00 W n.y/ D 2g
fy 2 J 0 [ J 00 W m00 .y/ D 2g: Then n0 D m0 J C n .J 0 [ J 00 / would eliminate m0 from 2 ./; so this cannot be. Thus, n .J 0 [ J 00 / D m00 .J 0 [ J 00 /: So fy 2 J W n.y/ D 2g fy 2 J W m00 .y/ D 2g D fy 2 J W m.y/ D 2g by m00 J D m J: Consider now n0 D n J C m .J 0 [ J 00 /: n0 2 …J 2 ./ J 0 …J 00 : n0 .y/ D 0 iff m.y/ D 0 by construction of n0 and n: So n0 m; and m 62 2 .…J .2 ./ J 0 / …J 00 /; a contradiction. t u
5.3.6.4 The Refined Approach We consider now more truth values, in the sense that j iff f.f / f ; and not only restricted to value 2, as in Definition 5.3.2 (page 190). The arguments and examples will be the same, they are given for completess’ sake only. Again, we show that there need not be interpolation of the forms ` ˛ j or j ˛ ` ; but there will be interpolation of the type j ˛ j : Example 5.3.5. (EQ has no interpolation of the form ` ˛ j . ) We work with three variables, a; b; c: Models will be written as h0; 0; 0i; etc., with the obvious meaning. Consider WD .:a ^ b ^ c/ _ .a ^ ::b ^ :c/: For h0; 1; 1i; h0; 1; 2i; h0; 2; 1i; h1; 1; 0i; h1; 2; 0i; f has value 1; for h0; 2; 2i; h2; 1; 0i; h2; 2; 0i f has value 2; all other values are 0. The only chosen model is h0; 2; 2iI all others contain 1, or are minimized. So f./ has value 2 for h0; 2; 2iI all other values are 0. Obviously, f./ fc ; so j c: As shown in Fact 4.4.5 (page 138), (1), we can define with c only c; :c; ::c; c ! c; :.c ! c/; and ::c ! c (up to semantical equivalence). But none is an interpolant of the type ` ˛ j W The left-hand condition fails for c; :c; ::c; :.c ! c/I the right-hand condition fails for c ! c and ::c ! c; as f.c!c/ .h0; 0; 0i/ D f.::c!c/ .h0; 0; 0i/ D 2: Example 5.3.6. (EQ has no interpolation of the form j ˛ ` . ) Consider two variables, a; b; WD a _ :a: f .m/ D 1 iff m.a/ D 1; and 2 otherwise. Note that (the model) h1; 0i h2; 0i; but f .h1; 0i/ D 1; f .h2; 0i/ D 2; so this minimization does not “count”. Consequently, f./ .m/ D 2 iff m D h0; 0i or m D h2; 0i; and f./ .m/ D 0 otherwise. Thus, j :b: But j6 FALSE; and TRUE 6` :b: t u Fact 5.3.10. EQ has interpolation of the form j ˛ j
.
Proof. See Diagram 5.3.3 (page 185) for illustration. Let L D J [ J 0 [ J 00 ; J 00 D I./; J D I. /: As does not contain any variables in J 00 in an essential way, f .m/ D f .m0 / if m J [ J 0 D m0 J [ J 0 : Thus, if a 2 J 00 ; and m Lfag D m0 Lfag D m00 Lfag; and m.a/ D 0; m0 .a/ D 1; m00 .a/ D 2; then by f .m/ D f .m0 / D f .m00 /; neither m0 nor m00 survives minimization, i.e., f./ .m0 / D f./ .m00 / D 0: Thus, if m.a/ ¤ 0 for
5.4 Context and Structure
193
some a 2 J 00 ; then f./ .m/ D 0: On the other hand, if m J 0 [ J 00 D m0 J 0 [ J 00 ; then f .m/ D f .m0 /: Define now the semantical interpolant by h.m/ WD supff./ .m0 / W m0 J 0 D m J 0 g: Obviously, f./ h; so, if h D f˛ for some ˛; then j ˛: It remains to show that f.h/ f I then ˛ j ; and we are done. For the same reasons as discussed above, f.h/ .m/ D 0 if m.a/ ¤ 0 for some a 2 J [ J 00 : Take now an arbitrary mI we have to show f.h/ .m/ f .m/: If m.a/ ¤ 0 for some a 2 J [ J 00 ; there is nothing to show. So suppose m.a/ D 0 for all a 2 J [ J 00 : By the above, h.m/ D supff./ .m0 /: m0 J 0 D m J 0 ^ 8a 2 J 00 :m0 .a/ D 0g; so, as m.a/ D 0 for a 2 J 00 ; h.m/ D supff./ .m0 /: m0 J 0 [ J 00 D m J 0 [ J 00 g: By the prerequisite, f./ .m0 / f .m0 / for all m0 ; but does not contain essential variables in J; so if m0 J 0 [ J 00 D m J 0 [ J 00 ; then f .m/ D f .m0 /I thus h.m/ f .m/; but f.h/ h; so f.h/ h.m/ f .m/: t u
5.4 Context and Structure The discussion in this section is intended to expand the perspective and separate support from attack, and, even more broadly, separate logic from manipulation of model sets. But this is not pursued here, and intended to be looked at in future research. We take the importance of condition . 3/ (or .S 3// as occasion for a broader remark. (1) This condition points to a weakening of the Hamming condition: Adding new “branches” in X 0 will not give new minimal elements in X 00 ; but may destroy other minimal elements in X 00 : This can be achieved by a sort of semi-rankedness: If and are different only in the X 0 -part, then iff ; but not necessarily iff : (2) In more abstract terms: When we separate support from attack (support: a branch 0 in X 0 supports a continuation 00 in X 00 iff ı 00 is minimal, i.e., not attacked; attack: a branch in X 0 attacks a continuation 00 in X 00 iff it prevents all ı 00 from being minimal), we see that new branches will not support any new continuations, but may well attack continuations. More radically, we can consider paths 00 as positive information, 0 as potentially negative information. Thus, …0 gives maximal negative information, and thus the smallest set of accepted models. (3) We can interpret this as follows: X 00 determines the base set. X 0 is the context. This determines the choice (subset of the base set). We compare this to preferential structures: In preferential structures, is not part of the language either; it is context. And we have the same behaviour as the one shown in the
194
5 Laws About Size and Interpolation in Non-monotonic Logics
fundamental property of preferential structures: the bigger the set, the bigger the number of possible attacks. (4) The concept of size looks only at the result of support and attack, so it is necessarily somewhat coarse. Future research should investigate both concepts separately. We broaden this. Following a tradition begun by Kripke, one has added structure to the set of classical models, reachability, preference, etc. Perhaps one should emphasize a more abstract approach, as done by one the authors, e.g., in [Sch92], and elaborated in [Sch04]; see in particular the distinction between structural and algebraic semantics in the latter. Our suggestion is to separate structure from logic in the semantics, and to treat what we called context above by a separate “machinery”. So, given a set X of models, we have some abstract function f; which chooses the models where the consequences hold, f .X /: Now, we can put into this “machinery” whatever we want. The abstract properties of preferential or modal structures are well known. But we can also investigate non-static f , where f changes in function of the past — “reacting” to the past. We can look at the usual properties of f , complexity, generation by some simple structure such as a special machine, etc. So we advocate the separation of usual, classical semantics from the additional properties, which are treated “outside”. It might be interesting to forget altogether about logic, classify those functions or more complicated devices which correspond to some logical property, and investigate them and their properties.
5.5 Interpolation for Argumentation Arguments (e.g., in inheritance) are sometimes arranged by a partial order only. We may define: follows from in argumentation iff for every argument for there is a better or equal argument for : It is not sufficient to give just one argument; there might not be a best one. We have to consider the set of all arguments. Consequently, if V D truth value set D set of arguments with a partial order ; we have to look at functions f W M ! P.V /; where to each model m (M the model set) is assigned a set of arguments (which support “m belongs to f ”.) Example: is m a weevil? Yes, it has a long nose. Yes, it has articulate antennae, etc. Thus, fweevi l .m/ D f long nose, articulate antennae g: We have to define on P.V /: We think a good definition is as follows: Definition 5.5.1. For A; B V (V with partial order / we define A B iff 8a 2 A9b 2 B:a b: This seems to be a decent definition of comparison of argument sets. Why not conversely? Suppose we have a very shaky argument b for I then to say that arguments for are better than arguments for ; we would need an even worse argument for : This does not seem right.
5.5 Interpolation for Argumentation
195
Thus, we define for arbitrary model functions f and g the following: Definition 5.5.2. Let f; g W M ! P.V /: We say f entails g iff f g iff 8a 2 f .m/9b 2 g.m/:a b: In total orders, sup and inf were defined. We want for sup A; B sup.A; B/; and A; B C ) sup.A; B/ C: So we define: S Definition 5.5.3. For a set A of argument sets, define sup.A/ WD A: Fact 5.5.1. We have (1) For all A 2 A A sup.A/: (2) If for all A 2 A A B; then sup.A/ B: Proof. Trivial by definition of :
t u
What is the inf? A definition should also work if the order is empty. Then, i nf .A; B/ D A \ B; which may be empty. This is probably not what we want. It is probably best to leave inf undefined. But we can replace A i nf .B; C / by A B and A C; so we can work with inf on the right of without a definition, replacing it by the universal quantifier (or, equivalently, by AND). For interpolation, for L D J [ J 0 [ J 00 ; f insensitive to J; g insensitive to J 00 ; f .m/ g.m/ for all m 2 M I we looked at f C .m; J 0 / WD supff .m0 / W m J 0 D m0 J 0 g and g .m; J 0 / WD i nf fg.m0 / W m J 0 D m0 J 0 g: We showed that f C .m; J 0 / g .m; J 0 /: We have to modify and show the following: S Fact 5.5.2. supff .m0 / W m J 0 D m0 J 0 g WD ff .m0 / W m J 0 D m0 J 0 g
g.m00 / for all m00 such that m J 0 D m00 J 0 : Proof. By definition of ; it suffices to show that 8m0 8m00 .m J 0 D m0 J 0 ^ m J 0 D m00 J 0 ) f .m0 / g.m00 //: Take m0 and m00 as above, so m0 J 0 D m J 0 D m00 J 0 : Define m0 by m0 J D m00 J; m0 J 0 [ J 00 D m0 J 0 [ J 00 : As f is insensitive to J; f .m0 / D f .m0 / g.m0 / by the prerequisite. Note that m0 J [ J 0 D m00 J [ J 0 ; as m0 J 0 D m0 J 0 D m00 J 0 : As g is insensitive to J 00 ; g.m0 / D g.m00 /: So we have f .m0 / D f .m0 / g.m0 / D g.m00 /: t u Fact 5.5.3. f C .m; J 0 / is an interpolant for f and g under the above prerequisites. Proof. Define h.m/ WD f C .m; J 0 /: We have to show that h is an interpolant. f .m/ h.m/ is trivial by definition. It remains to show that h.m/ g.m/ for all m: h.m/ WD supff .m0 / W m J 0 D m0 J 0 g g.m/ iff 8m0 .m0 J 0 D m J 0 ) f .m0 / g.m//; but this is a special case of the proof of Fact 5.5.2 (page 195). t u Note that the same approach may also be used in other contexts, e.g. considering worlds in Kripke structures as truth values, w 2 M./ iff w 2 f : All we really need is some kind of sup and inf.
•
Chapter 6
Neighbourhood Semantics
6.1 Introduction Neighbourhood semantics, probably first introduced by D. Scott and R. Montague in [Sco70] and [Mon70], and used for deontic logic by O. Pacheco in [Pac07] to avoid unwanted weakening of obligations, seem to be useful for many logics: (1) in preferential logics, they describe the limit variant, where we consider neighbourhoods of an ideal, usually inexistent, situation, (2) in approximative reasoning, they describe the approximations to the final result, (3) in deontic and default logic, they describe the “good” situations, i.e., deontically acceptable, or where defaults have fired. Neighbourhood semantics are used when the “ideal” situation does not exist (e.g., preferential systems without minimal elements), or are too difficult to obtain (e.g., “perfect” deontic states).
6.1.1 Defining Neighbourhoods Neighbourhoods can be defined in various ways: • by algebraic systems, such as unions of intersections of certain sets (but not complements), • quality relations, which say that some points are better than others, carrying over to sets of points, • distance relations, which measure the distance to the perhaps inexistent ideal points. The relations and distances may be given already by the underlying structure, e.g., in preferential structures, or they can be defined in a natural way, e.g., from a system of sets, as in deontic logic or default logic. In these cases, we can define a D.M. Gabbay and K. Schlechta, Conditionals and Modularity in General Logics, Cognitive Technologies, DOI 10.1007/978-3-642-19068-1 6, © Springer-Verlag Berlin Heidelberg 2011
197
198
6 Neighbourhood Semantics
distance between two points by the number or set of deontic requirements or default rules which one satisfies, but not the other. A quality relation is defined in a similar way: a point is better if it satisfies more requirements or rules.
6.1.2 Additional Requirements With these tools, we can define properties neighbourhoods should have. E.g., we may require them to be downward closed, i.e., if x 2 N; where N is a neighbourhood, y x; and y is better than x; then y should also be in N: This is a property we will certainly require in neighbourhood semantics for preferential structures (in the limit version). For these structures, we will also require that for every x 62 N; there should be some y 2 N with y x: We may also require that if x 2 N; y 62 N; and y is in some aspect better than x; then there must be z 2 N; which is better than both, so we have some kind of “ceteris paribus” improvement.
6.1.3 Connections Between the Various Properties There is a multitude of possible definitions (via distances, relations, set systems) and properties, so it is not surprising that one can investigate a multitude of connections between the different possible definitions of neighbourhoods. We cannot cover all possible connections, so we compare only a few cases, and the reader is invited to complete the picture for the cases which interest him. The connections we examined are presented in Section 6.3.4 (page 209).
6.1.4 Various Uses of Neighbourhood Semantics We also distinguish the different uses of the systems of sets thus characterized as neighbourhoods: we can look at all formulas which hold in (all or some) such sets (as in neighbourhood semantics for preferential logics), or at the formulas which exactly describe them. The latter reading avoids the infamous Ross paradox of deontic logic. This distinction is simple, but basic, and probably did not receive the attention it deserves in the literature.
6.2 Detailed Overview Our starting point was to give the “derivation” in deontic systems a precise semantical meaning. We extend this now to encompass the following situations: (1) Deontic systems, including contrary-to-duty obligations. (2) Default systems a la Reiter.
6.2 Detailed Overview
199
(3) The limit version of preferential structures. (4) Approximative logic. We borrow the word “neighbourhood” from analysis and topology, but should be aware that our use will be, at least partly, different. Common to topology and our domain is that — for reasons to be discussed — we are interested not only in one ideal point or one ideal set, but also in sets which are in some sense bigger, and whose elements are in some sense close to the “ideal”.
6.2.1 Motivation What are the reasons to consider some kind of “approximation”? (1) First, the “ideal” might not exist, e.g.: (1.1) In preferential structures, minimal models are the ideal, but there might be none, due to infinite descending chains. So the usual approach via minimal models leads to inconsistency; we have to take the limit approach; see Definition 2.3.5 (page 55). The same holds, e.g., for theory revision or counterfactual conditionals without closest worlds. (1.2) Default rules might be contradictory, so the ideal (all defaults are satisfied) is impossible to obtain. (2) Second, the ideal might exist, but be too difficult to obtain, e.g.: (2.1) In deontic logic, the requirements to lead a perfectly moral life might just be beyond human power. The same may hold for other imperative systems. E.g., we might be obliged to post the letter and to water the plants, but not have time for both, so doing one or the other is certainly better than nothing (so the “or” in the Ross paradox is not the problem). (2.2) It might be too costly to obtain perfect cleanliness, so we have to settle with sufficiently clean. (2.3) Approximate reasoning will try to find better and better answers, perhaps without hope of finding an ideal answer. (3) Things might be even more complicated by a (partial or total) hierarchy between aims. E.g., it is a “stronger” law not to kill than not to steal.
6.2.2 Tools to Define Neighbourhoods To define a suitable notion of neighbourhood, we may have various tools: (1) We may have a quality relation between points, where a b says that b is in some sense “better” than a.
200
6 Neighbourhood Semantics
Such a relation is, e.g., given in (1.1) preferential structures, where it is the preference relation; (1.2) defaults, where a (normal) default gives a quality relation: the situations which satisfy it are “better” than those which do not; here, a default gives a quality relation not only to two situations, but usually between two sets of situations; the same holds for deontic logics and other imperative systems; (1.3) borders separating subsets, with a direction, where it is “better” inside (or outside) the border, as in X -logic; see [BS85], and [Sch04]; (1.4) approximation, where one situation might be closer than the other to the ideal. (2) We may have several such relations, which may also partly contradict each other, and we may have a relation of importance between different and 0 (as in the example of not to kill or not to steal). The better a situation is, the closer it should be to our ideal. (3) We may have a distance relation between points, and these distances may be partially or totally ordered. With the distance relation, we can measure the distance of a point to the ideal points (all of them, the closest one, the most distant ideal point, etc.). Even if the ideal points do not exist, we can perhaps find a reasonable measure of distance to them. This can be found in distance semantics for theory revision and counterfactuals. There, it is the “closeness” relation; we are interested only in the closest models, and if they do not exist, in the sufficiently close ones (the limit approach),
6.2.3 Additional Requirements But there might still be other requirements: (1) We might postulate that neighbourhoods not only contain all sufficiently good points, but also that they not contain any points which are too bad. (2) We may require that they be closed under certain operations, e.g.: (2.1) If x is in a neighbourhood X; and y better than x; then it should also be in X (closure under improvement; see Definition 6.3.10 (page 207)). (2.2) For all y and any neighbourhood X; there should be some x 2 X which is better than y: (This and the preceeding requirement are those of MISEs; see Definition 2.3.5 (page 55).) (2.3) If we have a notion of distance, and x; x 0 are in a neighbourhood X; then anything “between” x and x 0 should be in X (X is convex). Thus, when we move in X; we do not risk leaving X: (2.4) Similarly, if x 2 X; and y is an ideal point, then everything between x and y is in X: Or, if x 2 X; and y is an ideal point closest to x; then
6.2 Detailed Overview
(2.5)
(2.6)
(2.7)
(2.8)
(2.9)
201
everything between x and y should be in X: So, when we improve our situation, we will not leave the neighbourhood. (A star-shaped set around an ideal point may satisfy this requirement, without being convex.) (See Definition 6.3.12 (page 208).) If we have to satisfy several requirements, we can ask whether this is possible independently for those requirements, or if we have to sacrifice one requirement in order to satisfy another. If the latter is the case, is there a hierarchy of requirements? Elements in a “good” neighbourhood should be better than the others: If x 2 X and the closest (to x) y 2 C .X /; then x y should hold, and, conversely, if y 2 C .X /; and the closest (to y) x 2 X; then x y should hold; see Definition 6.3.11 (page 208), It might be desirable to improve quality by moving into a good neighbourhood, without sacrificing anything achieved already; this is supposed to capture the “ceteris paribus” idea: If x 2 X and y 62 X satisfy a set R of rules, then there is an x 0 2 X which also satisfies R; and which is better than yI see Definition 6.3.5 (page 204). Given a set of “good” sets, we might be able to construct all good neighbourhoods by simple algebraic operations: Any good neighbourhood X is a union of intersections of the “good” sets; see Definition 6.3.2 (page 204), Finally, if this exists, the set of ideal points should probably satisfy our criteria; the set of ideal points should be a “good” neighbourhood.
Of particular interest are requirements which are in some sense independent: (1) We should try to satisfy an obligation without violating another obligation, which was not violated before. (2) The idea behind the Stalnaker-Lewis semantics for counterfactuals (see [Sta68], [Lew73]) is to look at the closest, i.e., minimally changed, situations. “If it were to rain, I would use an umbrella” means something like “If it were to rain, and there were not a very strong wind” (there is no such wind now), “if I had an umbrella” (I have one now), etc., i.e., if things were mostly as they are now, with the exception that now it does not rain, and in the situation I speak about it rains, and then I will use an umbrella. (3) The distance semantics for theory revision looks also (though with a slightly different formal approach) at the closest, minimally changed, situations. (4) This idea of “ceteris paribus” is the attempt to isolate a necessary change from the rest of the situation, and is thus intimately related to the concept of independence. Of course, a minimal change might not be possible, but small enough changes might do. We consider, e.g., the constant function f W Œ0; 1 ! Œ0; 1; f .x/ WD 0; and look for a minimally changed continous function with f .0:5/ WD 1: This does not exist. So we have to do with approximation, and look at functions “sufficiently” close to the first function. This is one
202
6 Neighbourhood Semantics
of the reasons we have to look at the limit variant of theory revision and counterfactuals. Remark: We do not look here at paths which lead (efficiently?) to better and better situations.
6.2.4 Interpretation of the Neighbourhoods Once we have identified our “good” neighbourhoods, we can interpret the result in several ways: (1) We can ask what holds in all good neighbourhoods. (2) We can ask what holds (finally) in some good neighbourhood; this is the approach for limit preferential structures and similar situations. This version is useful when the quality relation between neighbourhoods corresponds to set inclusion: better neighbourhoods are also smaller. Thus, if holds in some neighbourhood, it will finally hold. (3) We can also consider a more liberal version: we have an arbitrary quality relation between neighbourhoods, and from some neighbourhood N on, will hold in all N 0 at least as good as N: (4) We may not be so much interested in what holds in all or some good neighbourhoods, but in describing them: This is the problem of the semantics of a system of obligations. In short: what distinguishes a good from a bad set of situations. Such a characterization of “good” situations will give us a new semantics not only for deontic logics, and thus a precise semantical meaning for the “derivation” in deontic systems (see below for a justification) but also for defaults, preferential structures, etc. In particular, such descriptions will not necessarily be closed under arbitrary classical weakening; see the infamous Ross paradox, Example 6.4.1 (page 215).
6.2.5 Overview of the Different Lines of Reasoning This chapter is conceptually somewhat complicated; therefore we give now an overview of the different aspects: (1) We look at different tools and ways to define neighbourhoods, using distances and quality relations, and perhaps combining them, or using purely algebraic constructs such as unions and intersections. (2) We look at additional requirements for neighbourhoods, using such tools as closure principles. (3) We investigate how to obtain such natural relations, distances, etc. from different structures, e.g., from obligations, defaults, preferential models, etc.
6.3 Tools and Requirements for Neighbourhoods and How to Obtain Them
203
(4) We look at various ways to interpret the neighbourhood systems we have constructed, e.g., we can ask what holds in all or some neighbourhoods, what finally holds in neighbourhoods (when we have a grading of the neighbourhoods), or what characterizes the neighbourhoods, e.g., in the case of deontic logic. (5) We conclude (unsystematically) with connections between the different concepts.
6.2.6 Extensions This might be the place to make a remark on extensions. An extension is, roughly, a maximal consistent set of information or a smallest nonempty set of models. In default logic, we can follow contradictory default to the end of reasoning, and obtain perhaps contradictory information; likewise in inheritance nets, etc. Usually, one then takes the intersection of extensions, what is true in all extensions, which is — provided the language is adequate — the “OR” of the extensions. But we can also see preferential structures as resulting in extensions, where every minimal model is an extension. Consider a preferential structure with four models, say pq, p:q; :pq; :p:q; ordered by pq p:q; :pq p:q: Then we can see the relation roughly as two defaults: p:q W pq and p:q W :pq; with two extensions; pq and :pq: So, we can see a preferential structure as having usually many extensions (unless there is a single best model, of course), and we take as result the intersection of extensions, i.e., the theory which holds in all minimal models. In preferential structures, the construction of the set of minimal models is a onestep process: a model is in or out. In defaults, for instance, the construction is more complicated, we branch in the process. This is what may make the construction problematic, and gives rise to different approaches like taking immediately the intersection of extensions in inheritance networks. But this difference with preferential structures is in the process of construction; it is not in the outcome. These questions are intimately related to our neighbourhood semantics, as the constructions can be seen as an approximation to the ideal, the final outcome.
6.3 Tools and Requirements for Neighbourhoods and How to Obtain Them 6.3.1 Tools to Define Neighbourhoods Background We often work with an additional structure, some O P.U /; where U is the universe (intuitively, U is a set of propositional models), which allows us to define
204
6 Neighbourhood Semantics
distances and quality relations in a natural way. Intuitively, O is a base set of “good” sets, from which we will construct other “good” sets. Basically, x is better than y iff x is in more (as a set or by counting) O 2 O than y is, and the distance between x and y is the set (or cardinality) of O 2 O where x 2 O; y 62 O; or vice versa. Sometimes, it is more appropriate to work with sequences of 0/1, where 1 stands for O; 0 for C .O/ for O 2 O: Thus, we work with sets † of sequences. Note that † need not contain all possible sequences, corresponding to the possibility that, e.g., O \ O 0 D ; for O; O 0 2 O: Moreover, we may have a difference in quality between O and C .O/: if O is an obligation, then x 2 O is — at least for this obligation — better than x 0 62 O: The same holds for defaults of the type W =; with O D M./: We will follow the tradition of preferential structures, and “smaller” will mean “better”.
6.3.1.1 Algebraic Tools Let here again O P.U /: Definition 6.3.1. Given a finite propositional laguage L defined by the set v.L/ of propositional variables, let L^ be the set of all consistent conjunctions of elements from v.L/ or their negations. Thus, p ^ :q 2 L^ if p; q 2 v.L/; but p _ q; :.p ^ q/ 62 L^ : Finally, let L_^ be the set of all (finite) disjunctions of formulas from L^ : (As we will later not consider all formulas from L^ ; this will be a real restriction.) V Given a set of modelsVM for a finite language L; define M WD fp 2 v.L/ W 8m 2 M:m.p/ D vg ^ f:p W p 2 v.L/; 8m 2 M:m.p/ D f g 2 L^ : (If there are no such p; set M WD TRUE:) This is the strongest 2 L^ which holds in M: (for union of intersections) iff there is a family Definition 6.3.2. X U 0 is .ui S/ T Oi O; i 2 I such that X D . f Oi W i 2 I g/ \ U 0 : Unfortunately, as we will see later, this definition is not very useful for simple relativization. Definition 6.3.3. Let O0 O: Define for m 2 U and ı W O0 ! 2 D f0; 1g m ˆ ı W, 8O 2 O0 .m 2 O , ı.O/ D 1/ Definition 6.3.4. O is independent iff 8ı W O ! 2:9m 2 U:m ˆ ı: Obviously, independence does not inherit downward to subsets of U: Definition 6.3.5. D.O/ WD fX U 0 : 8O0 O 8ı W O0 ! 2 ..9m; m0 2 U; m; m0 ˆ ı; m 2 X; m0 62 X / ) .9m00 2 X:m00 ˆ ı ^m00 s 0 m //g
6.3 Tools and Requirements for Neighbourhoods and How to Obtain Them
205
This property expresses that we can satisfy obligations independently: If we respect O; we can, in addition, respect O 0 ; and if we are hopeless kleptomaniacs, we may still not be murderers. If X 2 D.O/; we can go from U X into X by improving on all O 2 O; which we have not fixed by ı; if ı is not too rigid.
6.3.1.2 Relations We may have an abstract relation of quality on the domain, but we may also define it from the structure of the sequences, as we will do now. Definition 6.3.6. Consider the case of sequences. Given a relation (of quality) on the codomain, we extend this to sequences in †: x y W, 8i 2 I.x.i / y.i //; x y W, 8i 2 I.x.i / y.i //; x y W, 8i 2 I.x.i / y.i // and 9i 2 I.x.i / y.i //: In the 2-case, we will consider x 2 i better than x 62 i: As we have only two values, TRUE and FALSE, it is easy to count the positive and negative cases (in more complicated situations, we might be able to multiply), so we have an analogue of the two Hamming distances, which we might call the Hamming quality relations. Let O P.U / be given now. (Recall that we follow the preferential tradition; “smaller” will mean “better”.) x s y W, O.x/ D O.y/; x s y W, O.y/ O.x/; x s y W, O.y/ O.x/; x c y W, card.O.x// D card.O.y//; x c y W, card.O.y// card.O.x//; x c y W, card.O.y// < card.O.x//:
6.3.1.3 Distances Note that we defined Hamming relations in Section 5.2.3 (page 172), as stated in Section 1.6.4.4 (page 22). Definition 6.3.7. Given x; y 2 †; a set of sequences over an index set I; the Hamming distance comes in two flavours: ds .x; y/ WD fi 2 I W x.i / ¤ y.i /g; the set variant, dc .x; y/ WD card.ds .x; y//; the counting variant. We define ds .x; y/ ds .x 0 ; y 0 / iff ds .x; y/ ds .x 0 ; y 0 /I thus, s-distances are not always comparable. Consequently, readers should be aware that ds -values are not always comparable, even though < and may suggest a linear order. We use these symbols to be in line with other distances. There are straightforward generalizations of the counting variant:
206
6 Neighbourhood Semantics
We can also give different importance to different i in the counting variant, so, e.g., dc .hx; x 0 i; hy; y 0 i/ might be 1 if x ¤ y and x 0 D y 0 ; but 2 if x D y and x0 ¤ y0: If the x 2 † may have more than two different values, then a varying individual distance may also reflect in the distances in †: So, (for any distance d ) if d.x.i /; x 0 .i // < d.x.i /; x 00 .i //; then (the rest being equal) we may have d.x; x 0 / < d.x; x 00 /: Fact 6.3.1. (1) If the x 2 † have only two values, say TRUE and FALSE, then ds .x; y/ D fi 2 I W x.i / D TRUEg4fi 2 I W y.i / D TRUEg; where 4 is the symmetric set difference. (2) dc has the normal addition, set union takes the role of addition for ds ; ; takes the role of 0 for ds I both are distances in the following sense: (2.1) d.x; y/ D 0 iff x D y; (2.2) d.x; y/ D d.y; x/; (2.3) the triangle inequality holds for the set variant in the form ds .x; z/ ds .x; y/ [ ds .y; z/: Proof. (2.3) If i 62 ds .x; y/ [ ds .y; z/; then x.i / D y.i / D z.i /; so x.i / D z.i / and i 62 ds .x; z/: The others are trivial. t u Recall that the 2 † will often stand for a sequence of possibilities O=C .O/ with O 2 O: Thus, the distance between two such sequences and 0 is the number or set of O; where codes being in O and 0 codes being in C .O/; or vice versa. Remark 6.3.1. If the x.i / are equivalence classes, one has to be careful not to confuse the distance between the classes and the resulting distance between elements of the classes, as two different elements in the same class have distance 0. So in Fact 6.3.1 (page 206), (2.1), only one direction holds. Definition 6.3.8. (1) We can define for any distance d with some minimal requirements a notion of “between”. If the codomain of d has an ordering ; but no addition, we define: hx; y; zid W, d.x; y/ d.x; z/ and d.y; z/ d.x; z/: If the codomain has commutative addition, we define hx; y; zid W, d.x; z/ D d.x; y/ C d.y; z/ — in ds C will be replaced by [; i.e., hx; y; zis W, d.x; z/ D d.x; y/ [ d.y; z/: For the above two Hamming distances, we will write hx; y; zis and hx; y; zic . (2) We further define Œx; zd WD fy 2 X W hx; y; xid g; where X is the set we work in. We will write Œx; zs and Œx; zc when appropriate. (3) For x 2 U; X U; set x kd X WD fx 0 2 X W :9x 00 ¤ x 0 2 X:d.x; x 0 / d.x; x 00 /g: Note that if X ¤ ;; then x k X ¤ ;:
6.3 Tools and Requirements for Neighbourhoods and How to Obtain Them
207
We omit the index when this does not cause confusion. Again, when adequate, we write ks and kc . For problems with characterizing “between”, see [Sch04]. Fact 6.3.3. (0) hx; y; zid , hz; y; xid : Consider the situation of a set of sequences †; with W I ! S for 2 † Let A WD A; 00 WD f 0 W 8i 2 I..i / D 00 .i / ! 0 .i / D .i / D 00 .i //g: Then (1) if 0 2 A; then ds .; 00 / D ds .; 0 / [ ds . 0 ; 00 /; so h; 0 ; 00 is ; (2) if 0 2 A and S consists of two elements (as in classical two-valued logic), then ds .; 0 / and ds . 0 ; 00 / are disjoint, (3) Œ; 00 s D A; (4) if, in addition, S consists of two elements, then Œ; 00 c D A. Proof. (0) Trivial. (1) “” follows from Fact 6.3.1 (page 206), (2.3). Conversely, if, e.g., i 2 ds .; 0 /; then by the prerequisite, i 2 ds .; 00 /: (2) Let i 2 ds .; 0 / \ ds . 0 ; 00 /I then .i / ¤ 0 .i / and 0 .i / ¤ 00 .i /; but then by card.S / D 2 .i / D 00 .i /; but 0 2 A; a contradiction. We turn to (3) and (4): If 0 62 A; then there is i 0 such that .i 0 / D 00 .i 0 / ¤ 0 .i 0 /: On the other hand, for all i such that .i / ¤ 00 .i /; i 2 ds .; 0 / [ ds . 0 ; 00 /: Thus: (3) By (1) 0 2 A ) h; 0 ; 00 is : Suppose 0 62 AI so there is i 0 such that i 0 2 ds .; 0 / ds .; 00 /; so h; 0 ; 00 is cannot be. (4) By (1) and (2) 0 2 A ) h; 0 ; 00 ic : Conversely, if 0 62 A; then card.ds .; 0 // C card.ds . 0 ; 00 // card.ds .; 00 // C 2. t u
6.3.2 Obtaining Such Tools We consider a set of sequences † for x 2 † x W I ! S; I a finite index set, S some set. Often, S will be f0; 1gI x.i / D 1 will mean that x 2 i when I P.U / and x 2 U: To abbreviate, we will call this (unsystematically, often context will tell) the 2-case. Often, I will be written OI intuitively, O 2 O is then an obligation, and x.O/ D 1 means x 2 O; or x “satisfies” the obligation O: Definition 6.3.9. In the 2-case, set O.x/ WD fO 2 O W x 2 Og:
6.3.3 Additional Requirements for Neighbourhoods Definition 6.3.10. Given any relation (of quality), we say that X U is (downward) closed (with respect to ) iff 8x 2 X 8y 2 U.y x ) y 2 X /.
208
6 Neighbourhood Semantics
(Warning: we follow the preferential tradition; “smaller” will mean “better”.) Fact 6.3.4. Let be given. (1) Let D U 0 U 00 ; D closed in U 00 I then D is also closed in U 0 : (2) Let D U 0 U 00 ; D closed in U 0 ; U 0 closed in US00 I then D is closed T in U 00 : 0 (3) Let Di U be closed for all i 2 I I then so are fDi W i 2 I g and fDi W i 2 I g: Proof. (1) Trivial. (2) Let x 2 D U 0 ; x 0 x; x 0 2 U 00 I then x 0 2 U 0 by closure of U 00 ; so x 0 2 D by closure of U 0 : (3) Trivial. u t Definition 6.3.11. Given a quality relation between elements, and a distance d; we extend the quality relation to sets and define: (1) x Y W, 8y 2 .x k Y /:x y: (The closest elements — i.e., there are no closer ones — of Y; seen from x; are less good than x:) Analogously X y W, 8x 2 .y k X /:x y: (2) X l Y W, 8x 2 X:x Y and 8y 2 Y:X y (X is locally better than Y ). When necessary, we will write l;s or l;c to distinguish the set from the counting variant. For the next definition, we use the notion of size: r iff for almost all holds, i.e., the set of exceptions is small. (3) X l Y W, rx 2 X:x Y and ry 2 Y:X y: We will likewise write l;s etc. This definition is supposed to capture quality difference under minimal change, the “ceteris paribus ” idea: X l C X should hold for an obligation X: Minimal change is coded by k; and “ceteris paribus” by minimal change. Fact 6.3.5. If X l C X; and x 2 U an optimal point (there is no better one), then x 2 X: Proof. If not, take x 0 2 X closest to xI this must be better than x; a contradiction. t u Definition 6.3.12. Given a distance d; we define: (1) Let X Y U 0 I then Y is a neighbourhood of X in U 0 iff 8y 2 Y 8x 2 X (x is closest to y among all x 0 with x 0 2 X ) Œx; y \ U 0 Y ). (Closest means that there are no closer ones.) When we also have a quality relation ; we define: (2) Let X Y U 0 I then Y is an improving neighbourhood of X in U 0 iff 8y 2 Y 8x((x is closest to y among all x 0 with x 0 2 X and x 0 y) ) Œx; y \ U 0 Y /: When necessary, we will have to say for (3) and (4) which variant, i.e., set or counting, we mean.
6.3 Tools and Requirements for Neighbourhoods and How to Obtain Them
209
Definition 6.3.13. Given a Hamming distance and a Hamming relation, X is called a Hamming neighbourhood of the best cases iff for any x 2 X and y a best case with minimal distance from x; all elements between x and y are in X: Fact 6.3.6. (1) If X X 0 †; and d.x; y/ D 0 ) x D y; then X and X 0 are Hamming neighbourhoods of X in X 0 : (2) If X Yj X 0 † Yj are Hamming neighbourhoods of S for j 2 J; and allT X in X 0 ; then so are fYj W j 2 J g and fYj W j 2 J g: Proof. (1) Trivial (we need here that d.x; y/ D 0 ) x D y). (2) Trivial.
t u
6.3.4 Connections Between the Various Concepts Fact 6.3.7. If x; y are models, then Œx; y D M.fx;yg /: (See Definitions 6.3.8 (page 206) and 6.3.1 (page 204).) Proof. m 2 Œx; y , 8p.x and x ˆ 6 p; y 6ˆ p ) m 6ˆ V ˆ p; y ˆ p ) m ˆ pV p/; m ˆ fx;yg , m ˆ fp W x.p/ D y.p/ D vg ^ f:p W x.p/ D y.p/ D f g: t u The requirement of closure causes a problem for the counting approach: Given two obligations O; O 0 ; any two elements in just one obligation have the same quality, so if one is in, the other should be, too. But this prevents now any of the original obligations from having the desirable property of closure. In the counting case, we will obtain a ranked structure, where elements satisfy 0, 1, 2, etc. obligations, and we are unable to differentiate inside those layers. Moreover, the set variant seems to be closer to logic, where we do not count the propositional variables which hold in a model, but consider them individually. For these reasons, we will not pursue the counting approach as systematically as the set approach. One should, however, keep in mind that the counting variant gives a ranking relation of quality, as all qualities are comparable, and the set variant does not. A ranking seems to be appreciated sometimes in the literature, though we are not really sure why. Of particular interest is the combination of ds and s .dc and c / respectively, where by s we also mean s and s ; etc. We turn to this now. Fact 6.3.8. We work in the 2-case. (1) x s y ) ds .x; y/ D O.x/ O.y/: Let a s b s c: Then (2) ds .a; b/ and ds .b; c/ are not comparable, (3) ds .a; c/ D ds .a; b/ [ ds .b; c/; and thus b 2 Œa; cs : This does not hold in the counting variant, as Example 6.3.1 (page 210) shows. (4) Let x s y and x 0 s y with x; x 0 s -incomparable. Then ds .x; y/ and ds .x 0 ; y/ are incomparable.
210
6 Neighbourhood Semantics
(This does not hold in the counting variant, as then all distances are comparable.) (5) If x s z; then for all y 2 Œx; zs x s y s z: Proof. (1) Trivial. (2) We have O.c/ O.b/ O.a/; so the results follows from (1). (3) By definition of ds and (1). (4) x and x 0 are s -incomparable, so there are O 2 O.x/ O.x 0 /; O 0 2 O.x 0 / O.x/: As x; x 0 s y; O; O 0 62 O.y/; so O 2 ds .x; y/ds .x 0 ; y/; O 0 2 ds .x 0 ; y/ ds .x; y/: (5) x s z ) O.z/ O.x/; ds .x; z/ D O.x/ O.z/: By the prerequisite, ds .x; z/ D ds .x; y/ [ ds .y; z/: Suppose x 6s y: Then there is an i 2 O.y/ O.x/ ds .x; y/; so i 62 O.x/ O.z/ D ds .x; z/; a contradiction. Suppose y 6s z: Then there is an i 2 O.z/ O.y/ ds .y; z/; so i 62 O.x/ O.z/ D ds .x; z/; a contradiction. t u Example 6.3.1. In this and similar examples, we will use the model notation. Some propositional variables p; q; etc. are given, and models are described by p:qr; etc. Moreover, the propositional variables are the obligations, so in this example we have the obligations p; q; r: Consider x WD :p:qr; y WD pq:r; z WD :p:q:r: Then y c x c z; dc .x; y/ D 3; dc .x; z/ D 1; dc .z; y/ D 2; so x 62 Œy; zc : t u Fact 6.3.9. Take the set version. If X l;s C X; then X is downward s -closed. Proof. Suppose X l;s C X; but X is not downward closed. Case 1: There are x 2 X; y 62 X; y s x: Then y 2 x ks C X; but x 6 y; a contradiction. Case 2: There are x 2 X; y 62 X; y s x: By X l;s CX; the elements in X closest to y must be better than y: Thus, there is an x 0 s y; x 0 2 X; with minimal distance from y: But then x 0 s y s x; so ds .x 0 ; y/ and ds .y; x/ are incomparable by Fact 6.3.8 (page 209), so x is among those with minimal distance from y; so X l;s C X does not hold. t u Example 6.3.2. We work with the set variant. This example shows that s -closed does not imply X l;s C X; even if X contains the best elements. Let O WD fp; q; r; sg; U 0 WD fx WD p:q:r:s; y WD :pq:r:s; x 0 WD pqrsg; X WD fx; x 0 g: x 0 is the best element of U 0 ; so X contains the best elements, and X is downward closed in U 0 ; as x and y are not comparable. ds .x; y/ D fp; qg; ds .x 0 ; y/ D fp; r; sg; so the distances from y are not comparable, so x is among the closest elements in X; seen from y; but x 6s y: The lack of comparability is essential here, as the following fact shows. t u
6.3 Tools and Requirements for Neighbourhoods and How to Obtain Them
211
We have, however, for the counting variant the following: Fact 6.3.10. Consider the counting variant. Then if X is downward closed, then X l;c C X: Proof. Take any x 2 X; y 62 X: We have y c x or x c y; as any two elements are c -comparable. y c x contradicts closure, so x c y; and X l;c C X holds trivially. t u Example 6.3.3. Let U 0 WD fx; x 0 ; y; y 0 g with x 0 WD pqrs; y 0 WD pqr:s; x WD :p:qr:s; y WD :p:q:r:s: Consider X WD fx; x 0 g: The counting version: Then x 0 has quality 4 (the best), y 0 has quality 3, x has 1, y has 0. dc .x 0 ; y 0 / D 1; dc .x; y/ D 1; dc .x; y 0 / D 2: Then above “ceteris paribus” criterion is satisfied, as y 0 and x do not “see” each other, so X l;c C X: But X is not downward closed; below x 2 X is a better element y 0 62 X: This seems an argument against X being an obligation. The set version: We still have x 0 s y 0 s x s y: As shown in Fact 6.3.8 (page 209), ds .x; y/ (and also ds .x 0 ; y 0 // and ds .x; y 0 / are not comparable, so our argument collapses. As a matter of fact, we have the result that the “ceteris paribus” criterion entails downward closure in the set variant; see Fact 6.3.9 (page 210). t u In the following Sections 6.3.4.1 (page 211) and 6.3.4.2 (page 213), we will assume a set O of obligations to be given. We define the relation WD O as described in Definition 6.3.6 (page 205), and the distance d is the Hamming distance based on OI see Definition 6.3.7 (page 205). We work here mostly in the set version, the 2-case, only in the final Section 6.3.4.3 (page 214); will we look at the counting case.
6.3.4.1 The Not Necessarily Independent Case Example 6.3.4. We work in the set variant. We show that X s -closed does not necessarily imply that X contains all the s -best elements. Let O WD fp; qg; U 0 WD fp:q; :pqgI then all elements of U 0 have the best quality in U 0 I X WD fp:qg is closed, but does not contain all the best elements. u t Example 6.3.5. We work in the set variant. We show that X s -closed does not necessarily imply that X is a neighbourhood of the best elements, even if X contains them. Consider x WD pq:rstu; x 0 WD :pqrs:t:u; x 00 WD p:qr:s:t:u; y WD p:q:r:s:t:u; z WD pq:r:s:t:u: U WD fx; x0 ; x 00 ; y; zg; the s -best elements are x; x 0 ; x 00 I they are contained in X WD fx; x 0 ; x 00 ; zg: ds .z; x/ D fs; t; ug; ds .z; x 0 / D fp; r; sg; ds .z; x 00 / D fq; rg; so x 00 is one of the best elements closest
212
6 Neighbourhood Semantics
to z: d.z; y/ D fqg; d.y; x 00 / D frg; so Œz; x 00 D fz; y; x 00 g; y 62 X; but X is downward closed. u t Fact 6.3.11. We work in the set variant. Let X ¤ ;; X s -closed. Then (1) X does not necessarily contain all the best elements. Assume now that X contains, in addition, all the best elements. Then (2) X l;s C X does not necessarily hold, (3) X is .ui /; (4) X 2 D.O/ does not necessarily hold, (5) X is not necessarily a neighbourhood of the best elements, (6) X is an improving neighbourhood of the best elements. Proof. (1) See Example 6.3.4 (page 211). (2) See Example 6.3.2 (page 210). (3) If there is an m 2 X; m 62 O for all O 2 O; then by closure X D U; take Oi WD ;: S T For m 2 X; let Om WD fO 2 O W m T 2 Og: Let X 0 WD f Om W m 2 X g: X X 0 : trivial, as T m 2 X ! m 2 Om X 0 : X 0T X : LetT m0 2 Om for some m 2 X: It suffices to show that m0 s m: m0 2 Om D fO 2 O W m 2 Og; so for all O 2 O .m 2 O ! m0 2 O/: (4) Consider Example 6.3.2 (page 210); let dom.ı/ D fr; sg; ı.r/ D ı.s/ D 0: Then x; y ˆ ı; but x 0 6ˆ ı and x 2 X; y 62 X; but there is no z 2 X; z ˆ ı and z y; so X 62 D.O/: (5) See Example 6.3.5 (page 211). (6) By Fact 6.3.8 (page 209), (5). t u Fact 6.3.12. We work in the set variant. (1.1) (1.2) (2.1) (2.2) (3.1) (3.2) (4.1) (4.2)
X l;s C X implies that X is s -closed. X l;s C X ) X contains all the best elements. X is .ui / ) X is s -closed. X is .ui / does not necessarily imply that X contains all the s -best elements. X 2 D.O/ ) X is s -closed. X 2 D.O/ implies that X contains all the s -best elements. X is an improving neighbourhood of the s -best elements ) X is s -closed. X is an improving neighbourhood of the best elements ) X contains all best elements.
Proof. (1.1) By Fact 6.3.9 (page 210). (1.2) By Fact 6.3.5 (page 208). (2.1) Let O 2 O; then O is downward closed (no y 62 O can be better than x 2 O/: The rest follows from Fact 6.3.4 (page 208), (3). (2.2) Consider Example 6.3.4 (page 211); p is .ui / (formed in U Š/; but p \ X does not contain :pq: (3.1) Let X 2 D.O/; but let X not be closed. Thus, there are m 2 X; m0 s m; m0 62 X:
6.3 Tools and Requirements for Neighbourhoods and How to Obtain Them
213
Case 1: Suppose m0 m: Let ım W O ! 2; ım .O/ D 1 iff m 2 O: Then m; m0 ˆ ım ; and there cannot be any m00 ˆ ım ; m00 s m0 ; so X 62 D.O/: Case 2: m0 s m: Let O 0 WD fO 2 O W m 2 O , m0 2 Og; dom.ı/ D 0 O ; ı.O/ WD 1 iff m 2 O for O 2 O0 : Then m; m0 ˆ ı: If there is an O 2 O such that m0 62 O; then by m0 s m; m 62 O; so O 2 O0 : Thus for all O 62 dom.ı/:m0 2 O: But then there is no m00 ˆ ı; m00 s m0 ; as m0 is already optimal among the n with n ˆ ı: (3.2) Suppose X 2 D.O/; x 0 2 U X is a best element; take ı WD ;; x 2 X: Then there must be x 00 x 0 ; x 00 2 X; but this is impossible as x 0 was best. (4.1) By Fact 6.3.8 (page 209), (4), all minimal elements have incomparable distance. But if z y; y 2 X; then either z is minimal or it is above a minimal element, with minimal distance from y; so z 2 X by Fact 6.3.8 (page 209), (3). (4.2) Trivial. t u
6.3.4.2 The Independent Case Assume now the system to be independent, i.e., all combinations of O are present. Note that there is now only one minimal element, and the notions of Hamming neighbourhood of the best elements and improving Hamming neighbourhood of the best elements coincide. Fact 6.3.13. We work in the set variant. Let X ¤ ;; X s -closed. Then (1) (2) (3) (4) (5)
X X X X X
contains the best element, l;s C X; is .ui /; 2 D.O/; is an (improving) Hamming neighbourhood of the best elements.
Proof. (1) Trivial. (2) Fix x 2 X I let y be closest to x; y 62 X: Suppose x 6 yI then there must be an O 2 O such that y 2 O; x 62 O: Choose y 0 such that y 0 is like y; only y 0 62 O: If y 0 2 X; then by closure y 2 X; so y 0 62 X: But y 0 is closer to x than y is, a contradiction. Fix y 2 U -X. Let x be closest to y; x 2 X: Suppose x 6 yI then there is an O 2 O such that y 2 O; x 62 O: Choose x 0 such that x 0 is like x; only x 0 2 O: By closure of X; x 0 2 X; but x 0 is closer to y than x is, a contradiction. (3) By Fact 6.3.11 (page 212), (3). (4) Let X be closed, and O 0 O; ı W O0 ! 2; m; m0 ˆ ı; m 2 X; m0 62 X: Let m00 be such that m00 ˆ ı; and for all O 2 O dom.ı/ m00 2 O: This exists by independence. Then m00 s m0 ; but also m00 s m; so m00 2 X: Suppose m00 m0 I then m0 s m00 ; so m0 2 X; a contradiction, so m00 s m0 : (5) Trivial by (1), the remark preceding this fact, and Fact 6.3.11 (page 212), (6).
214
6 Neighbourhood Semantics
Fact 6.3.14. We work in the set variant. (1) (2) (3) (4)
X X X X
l;s C x ) X is s -closed, is .ui / ) X is s -closed, 2 D.O/ ) X is s -closed, is a (improving) neighbourhood of the best elements ) X is s -closed.
Proof. (1) Suppose there are x 2 X; y 2 U X; y x: Choose them with minimal distance. If card.ds .x; y// > 1; then there is z; y s z s x; z 2 X or z 2 U -X, contradicting minimality. So card.ds .x; y// D 1: So y is among the closest elements of U X seen from x; but then by the prerequisite x y; a contradiction. (2) By Fact 6.3.12 (page 212), (2.1). (3) By Fact 6.3.12 (page 212), (3.1). (4) There is just one best element z; so if x 2 X; then Œx; z contains all y y x by Fact 6.3.8 (page 209), (3). t u The D.O/ condition seems to be adequate only for the independent situation, so we stop considering it now. Fact 6.3.15. Let Xi U; i 2 I; a family of sets; we note the following about closure under unions and intersections: (1) If the Xi are downward closed, then so are their unions and intersections. (2) If the Xi are .ui /; then so are their unions and intersections. Proof. Trivial.
t u
We do not know whether l;s is preserved under unions and intersections; it does not seem an easy problem. Fact 6.3.16. (1) Being downward closed is preserved while going to subsets. (2) Containing the best elements is not preserved (and thus neither is the neighbourhood property). (3) The D.O/ property is not preserved. (4) l;s is not preserved. Proof. (4) Consider Example 6.3.3 (page 211), and eliminate y from U 0 I then the closest to x not in X is y 0 ; which is better. t u
6.3.4.3 Remarks on the Counting Case Remark 6.3.17. In the counting variant all qualities are comparable. So if X is closed, it will contain all minimal elements. Example 6.3.6. We measure distance by counting. Consider a WD :p:q:r:s; b WD :p:q:rs; c WD :p:qr:s; d WD pqr:sI let U WD fa; b; c; d g; X WD fa; c; d g: d is the best element, Œa; d D fa; d; cg; so X is an improving Hamming neighbourhood, but b a; so X 6l;c C X:
6.4 Neighbourhoods in Deontic and Default Logic
215
Fact 6.3.18. We measure distances by counting. X l;c C X does not necessarily imply that X is an improving Hamming neighbourhood of the best elements. Proof. Consider Example 6.3.3 (page 211). There X l;c C X: x 0 is the best element, and y 0 2 Œx 0 ; x; but y 0 62 X: t u
6.4 Neighbourhoods in Deontic and Default Logic 6.4.1 Introduction Deontic and default logic have very much in common. Both have a built-in quality relation, where situations which satisfy the deontic rules are better than those which do not, or closer to the normal case in the default situation. They differ in the interpretation of the result. In default logic, we want to know what holds in the “best” or most normal situations, in deontic logic, we want to characterize the “good” situations, and avoid paradoxes like the Ross paradox. Note that our treatment concerns only obligations and defaults without prerequisites, but this suffices for our purposes: to construct neighbourhood semantics for both. When we work with prerequisites, we have to consider the possibilities of branching into different “extensions”, which is an independent problem. We discussed MISE extensively in Section 2.3.2 (page 55), so it will not be necessary to repeat the presentation.
6.4.2 Two Important Examples for Deontic Logic Example 6.4.1. The original version of the Ross paradox reads: If we have the obligation to post a letter, then we have the obligation to post or burn the letter. Implicit here is the background knowledge that burning the letter implies not posting it, and is even worse than not posting it. We prefer a modified version, which works with two independent obligations: We have the obligation to post the letter, and we have the obligation to water the plants. We conclude by unrestricted weakening that we have the obligation to post the letter or not to water the plants. This is obvious nonsense. Example 6.4.2. Normally, one should not offer a cigarette to someone, out of respect for his health. But the considerate assassin might do so nonetheless, on the cynical reasoning that the victim’s health is going to suffer anyway: (1) One should not kill, :k: (2) One should not offer cigarettes, :o: (3) The assassin should offer his victim a cigarette before killing him; if k; then o:
216
6 Neighbourhood Semantics
Here, globally, :k and :o are best, but among k-worlds, o is better than :o: The model ranking is :k ^ :o :k ^ o k ^ o k ^ :o:
6.4.3 Neighbourhoods for Deontic Systems A set R of deontic or default rules defines naturally quality and distance relations: (1) A situation (model) m is better than a model m0 iff m satisfies more rules than m0 does. “More” can be defined by counting, or by the superset relation. In both cases, we will note the relation here by : (See Definition 6.3.6 (page 205).) (2) The distance between two models m; m0 is the number or the set of rules satisfied by one, but not by the other. In both cases, we will note the distance here by d: Given a distance, we can define “between”: a is between b and c iff d.b; c/ D d.b; a/ C d.a; c/ (in the case of sets, C will be [/: See Definitions 6.3.7 (page 205) and 6.3.8 (page 206). We have here in each case two variants of Hamming relations or distances. With these ideas, we can define “good” sets X in a number of ways. If R is a family of rules, and if x and x 0 are in the same subset R0 R of rules, then a rule derived from R should not separate them. More precisely, if x 2 O 2 R , x 0 2 O 2 R; and D is a derived rule, then x 2 D , x 0 2 D: We think that being closed is a desirable property for obligations: what is at least as good as one element in the obligation should be “in”, too. But it may be a matter of debate which of the different possible notions of neighbourhood should be chosen for a given deontic system. It seems, however, that we should use the characterization of neighbourhoods to describe acceptable situations. Thus, we determine the “best” situations, and all neighbourhoods of the best situations are reasonable approximations to the best situations, and can thus be considered “derived” from the original system of obligations.
Chapter 7
Conclusion and Outlook
7.1 Conclusion An important part of this book concerns the concept of independence. One of the authors described it as “homogenousness” in his first book, [Sch97]. But it took quite some time and several detours to find a reasonable, and in hindsight obvious, answer.
7.1.1 Semantic and Syntactic Interpolation We usually try to decompose a logical problem — often formulated in the language of this logic — into a semantical part, and then translate it to syntax. This has proven fruitful in the past, and does so here, too. The reason is that the semantical problems are often very different from the translation problems. The latter concern usually definability questions, which tend to be similar for various logical problems. Here, we were able to see that semantical interpolation will always exist for monotonic or antitonic logics, but that the language and the operators may be too weak to define the interpolants syntactically. In contrast, semantical interpolants for non-monotonic logics need not always exist. We detail this now briefly.
7.1.2 Independence and Interpolation for Monotonic Logic Independence is closely related to (semantical) interpolation; as a matter of fact, in monotonic logic, the very definition of validity is based on independence, and guarantees semantical interpolation even for many-valued logics, provided the order on the truth values is sufficiently strong, as we saw in Chapter 4 (page 113), whereas the expressive strength of the language determines whether we can define D.M. Gabbay and K. Schlechta, Conditionals and Modularity in General Logics, Cognitive Technologies, DOI 10.1007/978-3-642-19068-1 7, © Springer-Verlag Berlin Heidelberg 2011
217
218
7 Conclusion and Outlook
the semantical interpolant (or interpolants), (see the same chapter). We also saw that we often have an interval of interpolants, and the upper and lower limits are universal interpolants in the following sense: they depend only on one formula, plus the set of propositional variables common to both formulas, but not on the second formula itself.
7.1.3 Independence and Interpolation for Non-monotonic Logic Perhaps the central chapter of the book is Chapter 5 (page 153), where we connect interpolation of non-monotonic logics to multiplicative laws about abstract size. The key to this is seeing that logics like preferential logics define an abstract notion of size by considering the exceptional cases as forming a small subset of all cases, and, dually, the normal cases as a big subset of all cases. Then, laws for non-monotonic logics can be seen as laws about addition of small and big subsets, and about general good behaviour of the notions of big and small. E.g., if X is a small subset of Y; and Y Y 0 ; then X should also be a small subset of Y 0 : These connections were investigated systematically first (to the best of our knowledge) in [GS09a]; see also [GS08f]. It was pointed out there that the rule of Rational Monotony does not fit well into laws about addition, and that it has to be seen rather as a rule about independence. It fits now well into our laws about multiplication, which express independence for non-monotonic logics. It is natural that the laws we need now for semantical interpolation are laws about multiplication, as we speak about products of model sets. Interpolation for non-monotonic logic has (at least) three different forms, where we may mix the non-monotonic consequence relation j with the classical relation ` : We saw that two variants are connected to such multiplicative laws, and especially, the weakest form has a translation to a very natural law about size. We can go on and relate these laws to natural laws about preferential relations when the logic is preferential. The problem of syntactic interpolation is the same as that for the monotonic case. These multiplicative laws about size have repercussions beyond interpolation, as they also say what should happen when we change the language, e.g., if we have a rule j in language L; and now change to a bigger language L0 ; whether we can still expect j to hold. This seems a trivial problem; it is not, and somehow seems to have escaped attention so far.
7.1.4 Neighbourhood Semantics The concluding chapter of the book concerns neighbourhood semantics; see Chapter 6 (page 197). Such semantics are ubiquitous in non-classical logic; they can be found as systems of ever better sets in the limit version of preferential logics, as a semantics for deontic and default logics, for approximative logics, etc. We looked
7.2 Outlook
219
at the common points, how to define them, and what properties to require for them. This chapter should be seen as a toolbox, where one finds the tools to construct the semantics one needs for the particular case at hand.
7.2 Outlook We think that further research should concern the dynamic aspects of reasoning, such as iterated revision, revising one non-monotonic logic with another nonmonotonic logic. Moreover, it seems to us that any non-classical logic (which is not an extension of the former, like modal logic, but diverges in its results from classical logic) needs a justification; so such logics do not only consist of language, proof theory, and semantics, but of language, proof theory, semantics, andjustification.
7.2.1 The Dynamics of Reasoning So far, most work on non-monotonic and related logics concern one step in a reasoning process only. Notable exceptions are, [Spo88, DP94], and [DP97]. It seems quite certain that there is no universal formalism if, e.g., in a theory revision task, we are given ; and then :; as information; we can imagine situations where we should come back to the original state, and others where this should not be the case. So, the dynamics of reasoning need further investigation. We took some steps here when we investigated generalized revision (see Chapter 3 (page 85), Section 3.4 (page 99)), as the results can be applied to revising one preferential logic with another. (Usually, such logics will not have a natural ranked order, so traditional revision will not work.) But we need more than tools; we need satisfactory systems.
7.2.2 A Revision of Basic Concepts of Logic: Justification Some logics like inductive logics (“proving” a theory from a limited number of cases), non-monotonic logics, and revision and update logics go beyond classical logic; they allow to derive formulas which cannot be derived in classical logic. Some might also be more modest, allowing less derivations, and some might be a mixture, e.g. approximative logics, allowing us to derive some formulas which cannot be derived in classical logic, and not allowing to derive other formulas which can be derived in classical logic. Let us call all those logics “bold logics”.
220
7 Conclusion and Outlook
Suppose that we agree that classical logic corresponds to “truth”. But then we need a justification to reason in other logic than in classical logic, as we know or suspect — or someone else knows or suspects — that our reasoning is in some cases false. (Let us suppose for simplicity that we know this erroneousness ourselves.) Whatever this justification may be, we have now a fundamentally new situation. Classical logic has language, proof theory, and semantics. Representation theorems say that proof theory and semantics correspond. Non-monotonic logic also has (language and) proof theory, and semantics. But something is missing: the justification — which we do not need for classical logic, as we do not have any false reasoning to justify. Thus, • classical logic consists of (1) language (variables and operators), (2) proof theory, (3) semantics; • bold logic consists of (1) (2) (3) (4)
language (variables and operators), proof theory, semantics, justification.
If a bold logic has no justification — whatever that may be — it is just foolishness, and the bolder it is (the more it diverges from classical logic), the more foolish it is. So let us consider justifications, in a far from exhaustive list. (1) First, on the negative side, costs: (1.1) A false result has a cost. This cost depends on the problem we try to solve. Suppose we have a case “man, blond”. Classifying this case falsely as “man, black hair” has a different cost when we try to determine the amount of hair dyes to buy, and when we are on the lookout for a blond serial killer on the run. (1.2) Calculating our bold logic has a cost, too (time and space). Usually, this will also depend on the case; the cost is not the same for all cases. E.g., let T D p _ .:p _ q/I then the cost to determine whether m ˆ T is smaller for p-models than for :p-models, as we have to check now in addition q. In addition, there may be a global cost of calculation. (2) Second, on the positive side, benefits: (2.1) Classical logic also has its costs of calculation, similar to the above. (2.2) In some cases, classical logic may not be strong enough to decide the case at hand. Hearing a vague noise in the jungle may not be enough to decide
7.2 Outlook
221
whether it is a lion or not, but we climb up the tree nonetheless. Bold logic allows us to avoid disaster, by precaution. (2.3) Parsimony, elegance, promises of future elaboration may also be considered benefits. We can then say that a bold logic is justified iff the benefits (summarized over all cases to consider) are at least as big as the costs (summarized over all cases to consider). Diverging more from classical logic incurs a bigger cost, so the bolder a logic is, the stronger its benefits must be; otherwise it is not rational to choose it for reasoning. When we look at preferential logic and its abstract semantics of “big” and “small” sets (see, e.g., [GS09a, GS08f]), we can consider this semantics as being an implicit justification: The cases wrongly treated are together just a “small” or “unimportant” subset of all cases. (Note that this says nothing about the benefits.) But we want to make concepts clear, and explicit, however they may be treated in the end. See also [GW08] for related reflections.
•
References
[ABM03] C. Areces, P. Blackburn, M. Marx, “Repairing the interpolation theorem in quantified modal logic”, Annals of Pure and Applied Logics, 123 (1–3), pp. 287–299, 2003 [AGM85] C. Alchourron, P. Gardenfors, D. Makinson, “On the logic of theory change: partial meet contraction and revision functions”, Journal of Symbolic Logic, Vol. 50, pp. 510– 530, 1985 [Ami02] E. Amir, “Interpolation systems for nonmonotonic reasoning systems”, JELIA 2002, pp. 233–244 [BB94] S. Ben-David, R. Ben-Eliyahu, “A modal logic for subjective default reasoning”, Proceedings LICS-94, July 1994, Paris, France, S. Abramsky (ed.), pp. 477–486, IEEE Computer Science Press, 1994 [BLS99] S. Berger, D. Lehmann, K. Schlechta, “Preferred history semantics for iterated updates”, Journal of Logic and Computation, Vol. 9, No. 6, pp. 817–833, 1999 [BS85] G. Bossu, P. Siegel, “Saturation, nonmonotonic reasoning and the closed-world assumption”, Artificial Intelligence 25 (1985), pp. 13–63 [Bou90a] C. Boutilier, “Conditional logics of normality as modal systems”, AAAI 1990, Boston, p. 594 [CP00] S. Chopra, R. Parikh, “Relevance sensitive belief structures”, Annals of Mathematics and Artificial Intelligence, vol. 28, No. 1–4, pp. 259–285, 2000 [Cra57] W. Craig, “Three uses of the Herbrand-Gentzen theorem in relating model theory and proof theory”, Journal of Symbolic Logic 22 (1957), No. 3, pp. 269–285 [DP94] A. Darwiche, J. Pearl, “On the logic of iterated belief revision”, in: “Proceedings of the Fifth Conference on Theoretical Aspects of Reasoning about Knowledge”, R. Fagin (ed.), pp. 5–23, Morgan Kaufmann, Pacific Grove, 1994 [DP97] A. Darwiche, J. Pearl, “On the logic of iterated belief revision”, Journal of Artificial Intelligence, Vol. 89, No. 1–2, pp. 1–29, 1997 [FH96] N. Friedman, J. Halpern, “Plausibility measures and default reasoning”, Journal of the ACM, Vol. 48, pp. 1297–1304, 1996 [GM05] D. Gabbay, L. Maksimova, “Interpolation and definability”, Clarendon Press, Oxford, 2005 [GS08a] D. Gabbay, K. Schlechta, “Cumulativity without closure of the domain under finite unions”, Review of Symbolic Logic 1 (3): pp. 372–392, 2008 [GS08b] D. Gabbay, K. Schlechta, “Reactive preferential structures and nonmonotonic consequence”, Review of Symbolic Logic, Vol. 2, No. 2, pp. 414–450 [GS08c] D. Gabbay, K. Schlechta, “Roadmap for preferential logics”, Journal of Applied Nonclassical Logic, Vol. 19/1, pp. 43–95, Hermes, Paris, 2009, see also hal-00311941, arXiv 0808.3073
D.M. Gabbay and K. Schlechta, Conditionals and Modularity in General Logics, Cognitive Technologies, DOI 10.1007/978-3-642-19068-1, © Springer-Verlag Berlin Heidelberg 2011
223
224
References
[GS08d] D. Gabbay, K. Schlechta, “A theory of hierarchical consequence and conditionals”, Journal of Logic, Language and Information, 19:1, pp. 3–32, Jan. 2010 [GS08e] D. Gabbay, K. Schlechta, “Defeasible inheritance systems and reactive diagrams”, Logic Journal of the IGPL, 17:1–54, 2009 [GS08f] D. Gabbay, K. Schlechta, “Logical tools for handling change in agent-based systems” Springer, 2009, ISBN 978-3-642-04406-9 [GS08g] D. Gabbay, K. Schlechta, “A semantics for obligations”, hal-00339393, arXiv 0811.2754 [GS09a] D. Gabbay, K. Schlechta, “Size and logic”, Review of Symbolic Logic, Vol. 2, No. 2, pp. 396–413 [GS09c] D. Gabbay, K. Schlechta, “Semantic interpolation”, submitted, preliminary version: arXiv.org 0906.4082 [GW08] D. Gabbay, J. Woods, “Resource-origins of nonmonotonicity”, Studia Logica 88:85– 112, 2008 [Gab04] D. M. Gabbay, “Reactive Kripke semantics and arc accessibility”, Proceedings CombLog04, July 28–30, 2004, W. Carnielli, F. M. Dionesio, P. Mateus (eds.), Centre of Logic and Computation, University of Lisbon, 2004, pp. 7–20 [Gab08] D. M. Gabbay, “Reactive Kripke models and contrary to duty obligations”, DEON 2008, Deontic Logic in Computer Science, July 15–18, 2008, Luxembourg, R.v.d. Meyden, L.v.d. Torre (eds.), Lecture Notes in Artificial Intelligence, LNAI 5076, pp. 155–173, Springer, 2008 [Gab85] D. M. Gabbay, “Theoretical foundations for non-monotonic reasoning in expert systems”. K. R. Apt (ed.), “Logics and Models of Concurrent Systems”, Springer, pp. 439–457, 1989 [HM07] J. Hawthorne, D. Makinson, “The quantitative/qualitative watershed for rules of uncertain inference”, Studia Logica 86.2/; pp. 247–297, 2007 [Han69] B. Hansson, “An analysis of some deontic logics”, Nous 3, 373–398. Reprinted in R. Hilpinen (ed.) “Deontic Logic: Introductory and Systematic Readings”, pp. 121–147, Reidel, Dordrecht, 1971 [KLM90] S. Kraus, D. Lehmann, M. Magidor, “Nonmonotonic reasoning, preferential models and cumulative logics”, Artificial Intelligence, 44 (1–2), pp. 167–207, July 1990 [KM07] G. Kourousias, D. Makinson, “Parallel interpolation, splitting, and relevance in belief change”, Journal of Symbolic Logic, 72, pp. 994–1002, Sept. 2007 [KM90] H. Katsuno, A. O. Mendelzon, “On the difference between updating a knowledge base and revising it”, Univ. of Toronto Tech. Rept., KRR-TR-90-6 [LMS01] D. Lehmann, M. Magidor, K. Schlechta, “Distance semantics for belief revision”, Journal of Symbolic Logic, Vol. 66, No. 1, pp. 295–317, March 2001 [LMS95] D. Lehmann, M. Magidor, K. Schlechta, “A Semantics for Theory Revision”, LIM Research Report 1995 - 126 [Lew73] D. Lewis, “Counterfactuals”, Blackwell, Oxford, 1973 [Mon70] R. Montague, “Universal grammar”, Theoria 36, pp. 373–98, 1970 [PV09] D. Pearce, A. Valverde, “Interpolation in equilibrium logic and answer set programming”, submitted [Pac07] O. Pacheco, “Neighbourhood semantics for deontic and agency logics” CIC 0 07; October 2007 [Par96] R. Parikh, “Belief, belief revision, and splitting languages”, L. S. Moss, J. Ginzburg and M. de Rijke (eds.) Proceed. Logic, Language and Computation, CSLI 1999, pp. 266–278 [SM94] K. Schlechta, D. Makinson, “Local and global metrics for the semantics of counterfactual conditionals”, Journal of Applied Nonclassical Logics, Vol. 4, No. 2, pp. 129–140, Hermes, Paris, 1994 [Sch04] K. Schlechta, “Coherent systems”, Elsevier, 2004 [Sch90] K. Schlechta, “Semantics for defeasible inheritance”, L. G. Aiello (ed.), “Proceedings ECAI 90”, Stockholm, Sweden, 1990, Pitman, London/Boston, Springer, pp. 594–597, 1990
References
225
[Sch91-1] K. Schlechta, “Theory revision and probability”, Notre Dame Journal of Formal Logic 32, No.2, pp. 307–319, 1991 [Sch91-3] K. Schlechta, “Some results on theory revision”, A. Fuhrmann, M. Morreau (eds.), “The Logic of Theory Change”, Lecture Notes in Artificial Intelligence, LNAI 465, pp. 72–92, Springer, 1991 [Sch92] K. Schlechta, “Some results on classical preferential models”, Journal of Logic and Computation, Oxford, Vol.2, No.6, pp. 675–686, 1992 [Sch93] K. Schlechta, “Directly sceptical inheritance cannot capture the intersection of extensions”, Journal of Logic and Computation, Oxford, Vol. 3, No. 5, pp. 455–467, 1995 [Sch95-1] K. Schlechta, “Defaults as generalized quantifiers”, Journal of Logic and Computation, Oxford, Vol. 5, No. 4, pp. 473–494, 1995 [Sch95-2] K. Schlechta, “Logic, topology, and integration”, Journal of Automated Reasoning, 14:353–381, 1995 [Sch96-1] K. Schlechta, “Some completeness results for stoppered and ranked classical preferential models”, Journal of Logic and Computation, Oxford, Vol. 6, No. 4, pp. 599–622, 1996 [Sch96-3] K. Schlechta, “Completeness and incompleteness for plausibility logic”, Journal of Logic, Language and Information, 5:2, pp. 177–192, 1996 [Sch97-4] K. Schlechta, “Filters and partial orders”, Journal of the Interest Group in Pure and Applied Logics, Vol. 5, No. 5, pp. 753–772, 1997 [Sch97] K. Schlechta, “Nonmonotonic logics: basic concepts, results, and techniques”, Lecture Notes in Artificial Intelligence, LNAI 1187, Springer, 1997 [Sco70] D. Scott, “Advice in modal logic”, “Philosophical problems in logic”, K. Lambert (ed.), Reidel, 1970 [Sho87b] Yoav Shoham, “A semantical approach to nonmonotonic logics”. Proc. Logics in Computer Science, pp. 275–279, Ithaca, NY, 1987, IEEE Computer Society, and Proceedings IJCAI 87, Milan, Italy, August 1987, J. P. McDermott (ed.), pp. 388–392, Morgan Kaufmann, San Francisco, 1987 [Spo88] W. Spohn, “Ordinal conditional functions: a dynamic theory of epistemic states”. W. L. Harper, B. Skyrms (eds.), “Causation in decision, belief change, and statistics”, vol. 2, pp. 105–134, Reidel, Dordrecht, 1988 [Sta68] Stalnaker, “A theory of conditionals”, N. Rescher (ed.), “Studies in logical theory”, Blackwell, Oxford, pp. 98–112 [TH89] R. H. Thomason, J. F. Horty, “Logics for inheritance theory”, M. Reinfrank, J. de Kleer, M. L. Ginsberg, E. Sandewall (eds.), “Non-monotonic reasoning”, Proceedings 2nd Int. Workshop, Grassau, 1988, Lecture Notes in Artificial Intelligence, LNAI 346, pp. 220– 237, Springer, 1989 [THT86] R. H. Thomason, J. F. Horty, D. S. Touretzky, “A calculus for inheritance in monotonic semantic nets”, Tech. Rept. CMU-CS-86-138, Comp. Sci. Dept., Carnegie Mellon University, 1986 [THT87] D. S. Touretzky, J. F. Horty, R. H. Thomason, “A clash of intuitions: the current state of nonmonotonic multiple inheritance systems”, Proceed. IJCAI 1987, pp. 476–482 [TTH91] D. S. Touretzky, R. H. Thomason, J. F. Horty, “A skeptic’s menagerie: conflictors, preemptors, reinstaters, and zombies in nonmonotonic inheritance”, Proceed. IJCAI 1991, pp. 478–483 [Tou86] D. S. Touretzky, “The mathematics of inheritance systems”, Pitman, London/Morgan Kaufmann, Los Altos, 1986
•
Index
.CM /, 40 .C U T /, 38 . /, 40 ./, 40 T , 35 T , 35 h˛; ki W x ! y, 48 hM; N .M /i, 154 hS; T ii , 112 h˛; ki W hx; i i ! hy; i i, 48 hx; y; zic , 206 hx; y; zid , 206 hx; y; zis , 206 ./, 34 .AND/, 58 .C CL/, 73 .GH /, 29, 173, 178 .GH1/, 20, 22, 153, 158, 172, 173 .GH 2/, 20, 22, 153, 158, 173 .GH 3/, 172, 173, 175, 178, 184 .GHD/, 176, 178 .GHD1/, 176 .GHD2/, 176 .LLE/, 73 .OR/, 58 .Opt /, 155, 163, 170 .RatM /, 163, 167 .SC /, 73 .T /, 104 .UpC um/, 108 .eMI /, 178 .iM /, 155, 163, 166, 170, 178 .ui /, 204 .wOR/, 156 A, 136 GH , 172, 173, 175
l;s , 208 k, 34 kc , 207 ks , 207 @, 146 ƒ.X/, 56 …, 33 ….D; ˛/, 73 ….O; ˛/, 73 ˛.p; q; r/, 145 ˛ W hx; i i ! hy; i i, 48 ˛ W x ! y, 48 ˇ.p; s/, 145 2-case, 207 ` ˛ j , 21, 157, 183, 190, 192 j ˛ ` , 21, 157, 181, 191, 192 j ˛ j , 21, 157, 184, 191, 192 M , 204 .˛/, 65 .X/, 68 M , 48 J 0 0 , 35 , 34 , 34 , 34 , 34 , 34, 44 `, 35 ` ı j, 178 ˆ, 35 j ı j, 178 C , 34 D.˛/, 73 D L , 35 D L -smooth, 50 O.˛/, 73
D.M. Gabbay and K. Schlechta, Conditionals and Modularity in General Logics, Cognitive Technologies, DOI 10.1007/978-3-642-19068-1, © Springer-Verlag Berlin Heidelberg 2011
227
228 A-ranked structure, 12, 107 D .O /, 204 F , 34 I , 34 MC , 164 P , 33 Y -essentially smooth, 71 Y -smooth, 49
r, 164
, 34
l;c , 208
l;s , 208 , 2, 33, 47, 183, 185 . C CL/, 81 . C on/, 81 . Equi v/, 81 . Loop/, 81 . Succ/, 81 .1 s/, 155, 164, 165 .2 s/, 164, 165 .< ! s/, 155, 164, 165, 170, 178 .AND/, 38 .AND1 /, 38, 165 .AND! /, 165, 167 .AND2 /, 165 .ANDn /, 38, 165 .B1/, 155 .B2/, 155 .B3/, 155 .B4/, 155 .B5/, 155 .C CL/, 38, 52 .CM /, 40 .CM2 /, 40, 165 .CMn /, 40, 165 .CM! /, 156, 165, 167 .CP /, 38, 165 .C UM /, 40, 52 .C U T /, 38 .DR/, 40 .EE1/, 79 .EE2/, 79 .EE3/, 79 .EE4/, 79 .EE5/, 79 .F 30 /, 34 .I 30 /, 34 .I [ d i sj /, 38 .I1 /, 38 .I2 /, 38, 40 .I! /, 38, 40 .In /, 38, 40 .K 1/, 79 .K 2/, 79
Index .K 3/, 79 .K 4/, 79 .K 5/, 79 .K 6/, 79 .K 7/, 79 .K 8/, 79 .K 1/, 79 .K 2/, 79 .K 3/, 79 .K 4/, 79 .K 5/, 79 .K 6/, 79 .K 7/, 79 .K 8/, 79 .LLE/, 38, 52 .Log k/, 40 .Log D0 /, 40 .Log[0 /, 40, 52 .Log[/, 40, 52 .NR/, 165 .OR/, 38 .OR! /, 165, 167 .OR2 /, 165 .ORn /, 165 .Opt /, 38, 164, 165 .PR0 /, 165 .PR/, 38, 52 .REF /, 38 .RW /, 38, 165 .RW /C, 52 .RatM /, 40, 52, 165 .RatM D/, 40, 52 .ResM /, 40 .S 10 /, 170 .S 1/, 20, 169, 173, 183 .S 2/, 169, 172, 173, 183 .S 3/, 169, 183 .S i /, 170 .SC /, 38, 52, 165 . /, 163–165 .ƒ\/, 57 .Š /, 129, 130 .‹ /, 131, 133 . 1/, 7, 20, 21, 153, 157, 169, 170, 173, 175, 178, 185 . 2/, 21, 158, 170, 173, 175, 176, 178 . 3/, 170 . i /, 170 . D0 /, 40 . D/, 40 .CM /, 165 .C UM /, 40, 52, 70, 73 .Id /, 104 .OR/, 38, 165
Index .PR0 /, 38 .PR/, 38, 52, 165 .RatM /, 40, 52, 165 .ResM /, 40 .S1/, 102 .S2/, 102 .;f i n/, 38, 52 .;/, 38, 52 . k/, 40 . 2/, 40 .9/, 106 . /, 52, 70, 73 . /, 38, 52, 73 .[0 /, 40 .[/, 40 .d i sjOR/, 38, 165 .dp/, 35, 52 .wOR/, 38, 165 . D0 /, 52 . D/, 52 . k/, 52 . 2/, 52 .[0 /, 52 .[/, T 52 .S/, 34 . /, 34 ./, 52, 73 .\/, 34, 73 .[/, 34, 52 .C /, 34 .F [ d i sj /, 164 .F1 /, 164 .F2 /, 164 .F! /, 164 .Fn /, 164 .I [ d i sj /, 164 .I1 /, 164 .I2 /, 164 .I! /, 164 .In /, 164 .MC [ d i sj /, 164 .MCC /, 40, 164, 165, 178 .MC ! /, 40, 164, 167, 178 .MC n /, 164 .r1 /, 164 .r2 /, 164 .rn /, 164 .r! /, 164 .j /, 178 .j C on/, 81 .j Loop/, 81 .j Succ/, 81 .big 1 ) big/, 168 .big big ) big/, 168, 169
229 .d i sjOR/, 38, 165 .eM F /, 40, 155, 163–166 .eM I /, 38, 40, 155, 163–166 .f " J /, 16, 127 .f " J /.m/, 117 .f # J /, 16, 127 .f # J /.m/, 117 .iM /, 38, 164, 165 .n s/, 164, 165 .s s/, 170 .smal l 1 ) smal l/, 169 .smal l x ) smal l/, 169 .wCM /, 40, 165 .wOR/, 38, 165 .x s/, 163 .big medium ) medium/, 169 .medium medium ) medium/, 169 C.f; J /, 126 C.f; m; J /, 121 .f; J /, 126 .f; m; J /, 121 1 x ) x, 178
