27 views

Uploaded by karlasdfg123

quantum computing

save

- 03 Quantum Connections (Scientific American Mikhail Lukain,Chris Monroe)
- Albert Einstein 1
- 1505.02286
- Quantum Physics in Islamic Gnosticism and Mysticism With Special Reference to Teachings of Syed Muhammad Noorbakhsh
- [Arxiv]a Quantum Bose-Hubbard Model With Evolving Graph as Toy Model for Emergent Spacetime(2010)(Theory)
- 44
- General Morphological Analysis as a Basic Scientific Modelling Method
- [Daniele_C._Struppa,_Jeffrey_M._Tollaksen_(Editors(BookZZ.pdf
- PR Boxes
- Neutrino Oscillations: The Indefinite Mass Defined
- 73 SIP30 Programacion de Sistemas de Computo MCIC
- 86 SIP30 Seguridad en Redes y Computadoras MCIC
- 51 SIP30 Introduccion Al Modelado de Redes MCIC
- 79 SIP30 Redes de Sensores Inalambricas MCIC
- 95 SIP30 Super Computo Para Aplicaciones Desafiantes MCIC
- 56 SIP30 Microtecnologia y Arquitectura de Procesadores MCIC
- 97_SIP30_Teoria_de_grafos_MCIC
- 20 SIP30 Computo Cuantico MCIC
- 80 SIP30 Redes Inalambricas MCIC
- 82 SIP30 Redes Sociales y Complejidad MCIC
- 79_SIP30_Redes_de_sensores_inalambricas_MCIC.pdf
- DUI0068.pdf
- 2N_SIP30_Matematicas_para_las_Ciencias_de_la_Computacion_MCIC.pdf
- 52 SIP30 Lenguajes de Descripcion de Hardware MCIC
- 48 SIP30 Introduccion a Las Redes Neuronales Artificiales MCIC
- Notas de Curso - Capitulo 2
- Notas de Curso - Capitulo 3
- 94 SIP30 Sistemas Operativos MCIC
- Notas de Curso - Capitulo 1
- 92 SIP30 Sistemas de Tiempo Real MCIC
- b7WuVsE_
- Mecanica Clasica FA Julio 2010
- 11 SIP30 Arquitectura de Supercomputadoras MCIC
- tarea 9-2.pdf
- 16 SIP30 Ciberseguridad MCIC
- 1N_SIP30_Diseno_y_analisis_de_algoritmos_MCIC
- Notas_del_curso-Unidad_1.pdf
- 15 SIP30 Ciberforense MCIC
- 19 SIP30 Compiladores Para Computadoras de Alto Rendimiento MCIC
- Tordorcevic - Walks on Ordinals.pdf

You are on page 1of 356

Logic and Algebraic Structures in Quantum Computing

**Arising from a special session held at the 2010 North American Annual Meeting of
**

the ASL, this volume is an international cross-disciplinary collaboration with

contributions from leading experts exploring connections across their respective fields.

Themes range from philosophical examination of the foundations of physics and

quantum logic, to exploitations of the methods and structures of operator theory,

category theory, and knot theory in an effort to gain insight into the fundamental

questions in quantum theory and logic.

The book will appeal to researchers and students working in related fields,

including logicians, mathematicians, computer scientists, and physicists. A brief

introduction provides essential background on quantum mechanics and category

theory, which, together with a thematic selection of articles, may also serve as the

basic material for a graduate course or seminar.

**Je n n i f e r Ch u b b is Assistant Professor of Mathematics at the University of San
**

Francisco, where she teaches a wide range of courses, including quantum computing,

to students in physics, computer science, and mathematics. She has a background in

physics, dynamical systems, and pure and applied math. Her current research focuses

on computable structure theory and algorithmic mathematics.

**Al i Es k a n d a r i a n holds the positions of Dean and Professor at The George
**

Washington University. He is a theoretical physicist and a founding member of the

groups in astrophysics and quantum computing/information. He serves as co-director

of the Center for Quantum Computing, Information, Logic, and Topology.

**Va l e n t i na Ha r i z a n ov is a Professor of Mathematics at The George Washington
**

University, where she also serves as co-director of the Center for Quantum Computing,

Information, Logic, and Topology. She is internationally recognized for her research in

mathematical logic, particularly in computability theory and computable model theory.

L E C T U R E N OT E S I N L O G I C

A Publication of The Association for Symbolic Logic

**This series serves researchers, teachers, and students in the field of symbolic
**

logic, broadly interpreted. The aim of the series is to bring publications to the

logic community with the least possible delay and to provide rapid

dissemination of the latest research. Scientific quality is the overriding

criterion by which submissions are evaluated.

Editorial Board

Jeremy Avigad

Department of Philosophy, Carnegie Mellon University

Zoe Chatzidakis

DMA, Ecole Normale Supérieure, Paris

Peter Cholak, Managing Editor

Department of Mathematics, University of Notre Dame, Indiana

Volker Halbach

New College, University of Oxford

H. Dugald Macpherson

School of Mathematics, University of Leeds

Slawomir Solecki

Department of Mathematics, University of Illinois at Urbana-Champaign

Thomas Wilke

Institut für Informatik, Christian-Albrechts-Universität zu Kiel

**More information, including a list of the books in the series, can be found at
**

http://www.aslonline.org/books-lnl.html

L E C T U R E N OT E S I N L O G I C 4 5

**Logic and Algebraic Structures in
**

Quantum Computing

Edited by

JENNIFER CHUBB

University of San Francisco

ALI ESKANDARIAN

George Washington University, Washington DC

VALENTINA HARIZANOV

George Washington University, Washington DC

association for symbolic logic

University Printing House, Cambridge CB2 8BS, United Kingdom

**Cambridge University Press is part of the University of Cambridge.
**

It furthers the University’s mission by disseminating knowledge in the pursuit of

education, learning, and research at the highest international levels of excellence.

**www.cambridge.org
**

Information on this title: www.cambridge.org/9781107033399

Association for Symbolic Logic

Richard Shore, Publisher

Department of Mathematics, Cornell University, Ithaca, NY 14853

http://www.aslonline.org

© Association for Symbolic Logic 2016

This publication is in copyright. Subject to statutory exception

and to the provisions of relevant collective licensing agreements,

no reproduction of any part may take place without the written

permission of Cambridge University Press.

First published 2016

A catalogue record for this publication is available from the British Library.

Library of Congress Cataloguing in Publication Data

Names: Chubb, Jennifer. | Eskandarian, Ali. | Harizanov, Valentina S.

Title: Logic and algebraic structures in quantum computing / edited by Jennifer Chubb,

University of San Francisco, Ali Eskandarian, George Washington University, Washington

DC, Valentina Harizanov, George Washington University, Washington DC.

Description: Cambridge : Cambridge University Press, 2016. | Series: Lecture notes in

logic | Includes bibliographical references and index.

Identifiers: LCCN 2015042942 | ISBN 9781107033399 (hardback : alk. paper)

Subjects: LCSH: Quantum computing–Mathematics. | Logic, Symbolic and

mathematical. | Algebra, Abstract.

Classification: LCC QA76.889 .L655 2016 | DDC 006.3/843–dc23 LC record available at

http://lccn.loc.gov/2015042942

ISBN 978-1-107-03339-9 Hardback

Cambridge University Press has no responsibility for the persistence or accuracy of

URLs for external or third-party Internet Web sites referred to in this publication

and does not guarantee that any content on such Web sites is, or will remain,

accurate or appropriate.

CONTENTS

Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii

Jennifer Chubb, Ali Eskandarian, and Valentina Harizanov

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

Jennifer Chubb and Valentina Harizanov

A (very) brief tour of quantum mechanics, computation, and category

theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

Allen Stairs

Could logic be empirical? The Putnam-Kripke debate . . . . . . . . . . . . . . 23

William C. Parke

The essence of quantum theory for computers . . . . . . . . . . . . . . . . . . . . . . 42

Adam Brandenburger and H. Jerome Keisler

Fiber products of measures and quantum foundations . . . . . . . . . . . . . . 71

Samson Abramsky and Chris Heunen

Operational theories and categorical quantum mechanics . . . . . . . . . . . 88

Bart Jacobs and Jorik Mandemaker

Relating operator spaces via adjunctions . . . . . . . . . . . . . . . . . . . . . . . . . . . 123

¨

Andreas Doring

Topos-based logic for quantum systems and bi-Heyting algebras . . . . 151

Bob Coecke

The logic of quantum mechanics – Take II . . . . . . . . . . . . . . . . . . . . . . . . . 174

Dimitri Kartsaklis, Mehrnoosh Sadrzadeh, Stephen Pulman, and Bob

Coecke

Reasoning about meaning in natural language with compact closed

categories and Frobenius algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199

Louis H. Kauﬀman

Knot logic and topological quantum computing with Majorana

fermions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337

v

.

Washington. University of Oxford) and an invited lecture on intuitionistic quantum logic (by Klaas Landsman. We hope that the present collection advances this worthwhile program of scientiﬁc and mathematical progress. mathematical. and the ASL and Cambridge University Press for publishing it.ucla. and even experimental physics relevant to quantum information and quantum computation. The articles in this volume by mathematicians. 1. Jennifer Chubb Ali Eskandarian Valentina Harizanov Summer. vii . 2015. no. philosophers. In addition to this session. attendees heard tutorial lectures on quantum computing (given by Bob Coecke. 17 (2011). We would like to thank the authors that contributed to this volume. D. vol. The talks were so well-received by conference participants that we felt a volume of collected works on this subject would be a valuable addition to the literature. pp. Nijmegen).C. This project was partially supported by the George Washington University Centers & Institutes Facilitating Fund Grant and by the University of San Francisco Faculty Development Fund.htm.edu/∼asl/bsl/1701-toc. 1 The full program is available in the Bulletin of Symbolic Logic. 135–137.math. and scientists address foundational issues and fundamental abstract structures arising in highly active areas of theoretical. PREFACE This project grew out of a Special Session on Logic and the Foundations of Physics at the 2010 North American Annual Meeting of the Association for Symbolic Logic1 . Many of the session’s lecturers investigated the role of algebraic structures in the context of the foundations of quantum physics. Many thanks also to Bryan Fregoso (a University of San Francisco student) for his invaluable assistance in assembling this volume. especially in quantum information and computation. available online at https://www. Radboud University.

.

mathematics. As is evident from the title of this volume. physics. on the marriage of the most unlikely and intriguing ﬁelds of quantum theory and logic and ask: Why quantum logic? By many. unfamiliar in the range of velocities we are normally accustomed to. In the resulting body of work. interesting results. One of the signiﬁcant aspects of these developments has been an integration of several ﬁelds of inquiry that not long ago appeared to be evolving. but not necessarily in its most sublime role as a deep intellectual subject underlying the validity of mathematical structures and worthy of investigation and discovery in its own right. However. logic. Chubb. investigators have revealed a deeper connection among the ideas and techniques of (apparently) disparate ﬁelds. Harizanov Lecture Notes in Logic. 45 c 2016. It is often associated with the rules of correct thinking and decision-making. There exists an undeniable interconnection between the deepest theories of nature and mathematical reasoning. along narrow disciplinary paths without any major overlap with each other. Eskandarian and V. A. perhaps. AND VALENTINA HARIZANOV In the last two decades. and notable progress in our conceptual understand- ing of computing and information based on the laws of quantum theory. The sciences. Logic and Algebraic Structures in Quantum Computing Edited by J. INTRODUCTION JENNIFER CHUBB. famously stated by Eugene Wigner as the unreasonable eﬃcacy of mathematics in physical theories. the scientiﬁc community has witnessed a surge in activity. while analyzing natural phenomena for objects moving close to the speed of light and. One might be tempted to dismiss it as contrary to observation. The inquisitive reader might focus. Indeed. more or less. logical deductions based on the postulates of the special relativity theory lead to the correct predictions of experimental observations. “logic” is deemed to be panacea for faulty intuition. should one hold to the intuition developed through experiencing familiar macroscopic scenarios in our routine impressions of natural phenomena. One such example is a statement within the special theory of relativity that the speed of light is the same in all inertial frames. within the realm of the classical theories of nature. therefore. Association for Symbolic Logic 1 . ALI ESKANDARIAN. one may encounter situations that defy comprehension. computer science and information theory are intricately involved in this fascinating story. It certainly deﬁes the common intuition regarding the observation of velocities of familiar objects in relative motion.

the striking applications of quantum theory in the theory of computation. and causality are the seemingly indispensable factors in our understanding of reality. one is forced to consider and question the validity of the premises on which that logical structure is built. development of new algorithms. and the promising prospects for the building of a computing machine operating on the basis of the laws of quantum theory. ALI ESKANDARIAN. as if considering quantum theory as complete would question one’s logical ﬁtness and one’s understanding of reality! Yet. most prominently brought out in the celebrated work of Einstein. particularly scientiﬁc theories. What happens if the syntax and grammar of such a language become inadequate? This seems to have been the case when some of the more esoteric predictions of the then new theory of quantum mechanics began to challenge the scientiﬁc intuition of the times around the turn of the 20th century. long held understanding of what should be taken for granted as “elements of reality” had to be abandoned. and multiple ways of testing the theory. Otherwise. This violation of intuition was so severe that even the most prominent of scientists were not able to reconcile the dictates of their intuition with the experimentally conﬁrmed predictions of the theory. logical deductions based on primitives that were the very essence of reality and logical consistency forced the conclusion of the incompleteness of quantum theory. as it aﬀords them elegance as well as economy of expression. according to EPR. perhaps. necessitate a deeper investigation of alternative logical structures that encompass the elements of this new quantum . quantum theory has consistently outshined the alternatives. have relied on. by necessity. incomplete. Mathematics has become the de facto language of the quantitative sciences. and Rosen (EPR) in the mid 1930s. and a logical structure that seems to fall short in facilitating correct thinking and correct decision making (at least. Quantum theory has not (as yet) suﬀered any such blow. The discomfort with some of the features and predictions of quantum theory were. EPR fueled several decades of investigations on the foundations of quantum theory that continue to this day. are found to be entirely inconsistent with the actual reality around us. AND VALENTINA HARIZANOV and in particular physics. In particular. Furthermore. which are shown to be correct every time subjected to experimental veriﬁcation. in the decades since. in so far as the behavior of natural phenomena at the quantum level is concerned). the economy of mathematical expressions and the eﬃcacy and rigor of mathematical reasoning with its underlying logical structure to make deﬁnite statements and predictions about nature. Here. The main assertion of the EPR work was that quantum theory had to be. with increasing sophistication in experimentation.2 JENNIFER CHUBB. many predictions relying on the sensibilities of classical theories. Podolsky. where concepts such as separability. and beneﬁted from. or to discover alternative structures. Confronted with the stark inability to reconcile the predictions of a theory. and the major discoveries and predictive statements of these theories (whenever possible) are cast in the language of mathematics. locality.

While their endeavor was revolutionary.. Doring¨ gives a topos-based distributive form of quantum logic as an alternative to the quantum logic of Birkhoﬀ and von Neumann. The papers by Coecke and Kartsaklis et al. One must then give credence to the argument that. the fault is not with the revolutionary quantum theory. In this volume. Birkhoﬀ and von Neumann were among the ﬁrst to propose a generalization of Boolean logic in which propositions about quantum systems could be formulated. as investigated from the perspective of logical and al- gebraic structures. Brandenburger and Keisler use ideas from continuous model theory to explore determinism and locality in quantum mechanical systems. address philosophical and histori- cal issues. we present the work of a select group of scholars with an abid- ing interest in tackling some of the fundamental issues facing quantum comput- ing and information theory. it is only natural to search for those logical and algebraic structures that underlie the scaﬀolding of the quantum rules in computations. In his article in the present volume. by Stairs and Parke. Abramsky and Heunen. the Birkhoﬀ-von Neumann quantum logic was not to be the ﬁnal word on the subject of a logic for quantum mechanics. and Jacobs and Mandemaker describe the relationship between the category-theoretic and operator-theoretic approaches to the foundations of quantum physics. Could logic be empirical? The Putnam-Kripke debate. use a diagrammatic calculus in analyzing quantum mechanical systems and. rather. and indeed the investigation continues with increasing urgency. very recently. no doubt. As obvious as it is that Boolean logic underlies classical computing and much of classical reasoning. the pictorial approach used in knot theory is closely related to the quantum categorical logic presented in other articles in this volume. in computational linguistics. INTRODUCTION 3 reality. i.e. Kauﬀman’s article presents an extensive treatment of the prominent role of algebraic structures arising from topological considerations in quantum information and computing. The utility of logical rules is most pronounced when applied to the building and operation of computing machines. This selection. Stairs outlines Hilary Putnam’s position that quantum mechanics provides an empirical basis for a re-evaluation of our . it is with the inadequacies of logical structures that were insuﬃcient to be expanded and applied to a world that does not comply with the notions embodied in our understanding of the macroscopic classical physical theories of nature. within the reasonable limitations of space and cov- erage of topics for a volume of this size. by Allen Stairs. and for the purpose of generating ideas that would fuel further investigation and research in these and related ﬁelds. The ﬁrst two articles. reﬂects the intellectual proclivities and curiosities of the editors. With the advent of computing that takes advantage of the laws of quantum theory. quantum computing. it is equally obvious that it is not suﬃcient to express the logic underlying quantum mechanics or quantum computing. perhaps.

by Adam Branden- burger and H. which are relevant to a deeper understanding of the mathematical and logical structures underlying (or derived) from such theories. and were used recently by Ita¨ı Ben Yaacov and Jerome Keisler in their work on continuous model theory (2009). and the limitations that our current understanding of space-time imposes on building and utilizing computing machines based on the rules of quantum theory. The essence of quantum theory for computers. Parke provides a thorough yet succinct introduction to the elements of physical theories. and in the end oﬀers the beginnings of a compromise. These objects (ﬁber products) were conceived by Rae Shortt in a 1984 paper. by William C. the authors use ﬁber products of (probability) measures within a framework they construct for empirical and hidden-variable models to prove determinization theorems. In this case. and important in the appreciation of the more subtle quandaries of quantum theory. The article wrestles with the idea of whether and how quantum mechanics should inform our logic and reasoning processes. and can be thought of as probabilistic. Jerome Keisler. and for every hidden-variable model satisfying locality and -independence. The emphasis has been placed on the physical content of information and elements of computation from a physicist’s point of view. together with Bell’s theorem. Techniques in continuous model theory are relevant to the notion of models of quantum structures as in that context the “truth value” of a statement may take on a continuum of values.4 JENNIFER CHUBB. without too much focus on unnecessary details. which includes “disjunctive facts. ALI ESKANDARIAN. AND VALENTINA HARIZANOV idea of logic and Saul Kripke’s response. In this article. precludes the existence of a hidden-variable model in which both determinism and -independence hold. in order to set the stage properly and provide motivation for the work of the others on logical and algebraic structures. The notion of -independence was . in which he takes issue with the very idea of a logic that is based on anything empirical. a technique employed in continuous model theory is used in the construction of models in proofs of theorems that assert that every empirical model can be realized by an extension that is a deterministic hidden-variable model. and the notion of “l-complementarity. there is a realization-equivalent (both models extend a common empirical submodel) hidden-variable model satisfy- ing determinism and -independence. Parke. The treatment of the principles of quantum theory is also developed from an advanced point of view.” which can be true even if their disjuncts are not. Stairs carefully interprets their positions. Fiber products of measures and quantum foundations. This includes a treatment of the role of space-time in the development of physical theories from an advanced point of view. The latter statement.” to describe the relationship between statements having non-commuting associated projectors. In this model-theoretic article. leading to its utilization in computation. classical and quantum. but covering the essential conceptual ingredients.

and density operators. introduced by Birkhoﬀ and von Neumann. one based on operational theories (also called general probabilistic theories) and the other on category-theoretic foundation of quantum theory. self-adjoint. by the topos-based distributive form of quantum logic. such as the complex numbers. They use category- theoretic tools to describe relations between various spaces of operators on a ﬁnite-dimensional Hilbert space. which they call process category. It says that the choices made by an entity as to which observable to measure in a system are not inﬂuenced by the process of the determination of the value of a relevant hidden-variable. by Andreas ¨ D¨oring. They further propose to apply a similar analysis to contextuality. Operational theories and categorical quantum mechanics. which comes with a host of conceptual and interpretational problems. by Bart Jacobs and Jorik Mande- maker. eﬀect. Operational theories focus on empirical and observational content. Relating operator spaces via adjunctions. and shows that these clopen subojects form a bi-Heyting algebra. They describe the algebraic structure of these sets of operators in terms of modules over various semirings. Instead of having a non-distributive orthomodular lattice of projections. and quantum mechan- ics occupies one point in a space of possible theories. The authors deﬁne a symmetric monoidal categorical structure of an operational theory. Topos-based logic for quantum systems and bi-Heyting algebras. he introduces two . The authors give a uniform description of such modules via the notion of an algebra of the multiset monad. which arise in quantum theory. the real numbers. They show how some spaces of operators are related by free constructions between categories of modules. Doring ¨ considers clopen subobjects of the presheaf attaching the Gelfand spectrum to each abelian von Neumann algebra. while the other spaces of operators are related by a dual adjunction between convex sets (conveniently described via a monad) and eﬀect modules. Doring replaces the standard quantum logic. which can be viewed as a broader phenomenon than non-locality. including bounded. and exploit the ideas of categorical quantum mechanics to obtain an operational theory as a certain representation of this process category. projection. INTRODUCTION 5 ﬁrst formulated by W. They lift the notion of non-locality to the general level of operational category. positive. Jacobs and Mandemaker clarify and present in a uniﬁed framework various. by Samson Abram- sky and Chris Heunen. Michael Dickson (2005). the non-negative real numbers. seemingly diﬀerent results in the foundation of quantum theory found in the literature. He gives various physical interpretations of the objects in this algebra and of the operations on them. Samson Abramsky and Chris Heunen establish strong and important connections between these two formalisms. By exploiting techniques of category theory. For example. There are two complementary research programs in the foundations of quantum mechanics. he considers a complete bi-Heyting algebra of propositions. More speciﬁcally.

in particular. Grammatical structure is modeled through the derivations of pregroup grammars. Such functors are structure preserving morphims. Coecke closes with speculation about an important question: Where is the traditional logic hiding in all this? Reasoning about meaning in natural language with compact closed categories and Frobenius algebras. Coecke’s framework has been applied both to logic concerned with natural language interpretations. A pregroup is a partially ordered monoid with left and right adjoints for every element in the partial order. the focus is on the former. and Coecke agrees. but has also provided new insight into non-locality. and ¨ gives physical interpretation of the two kinds of negation.6 JENNIFER CHUBB. Imposing minimal additional structure on these categories (to obtain dagger compact categories) allows for the almost trivial derivation of a number of quantum phenomena. by Bob Coecke. Schrodinger ¨ maintained that composition of systems is the heart of quantum computing. Coecke applies the graphical language of dagger compact categories to natural language processing—“from word meaning to sentence meaning”—implementing Lambek’s theory of grammar and the notion of words as “meaning vectors. Stephen Pulman. and complementarity are left by the wayside. The author puts forth a model of quantum logic that is based on composition rather superposition. He suggests that the Birkhoﬀ-von Neumann formulation of quantum logic fails to adequately and elegantly capture composition of quantum systems.” He argues that sentence meaning amounts to more than the meanings of the constituent words. AND VALENTINA HARIZANOV kinds of negation associated with the Heyting and co-Heyting algebras. ALI ESKANDARIAN. In this article. Doring considers the map called outer daseinisation of projections. The model can be expanded (using spiders!) in such a way that all these are captured. The authors build tensors for linguistic constructs with complex types by using a Frobenius algebra. This (now widely adopted) formalism has been used not only to solve open problems in quantum information theory. The logic of quantum mechanics – Take II. but also the way in which they compose. including quantum teleportation and entanglement swapping. measurement. He axiomatizes composition without reference to underlying systems using strict monoidal categories as the basic structures and explains a graphical language that exactly captures these structures. which provides a link between the usual Hilbert space formalism and his topos-based quantum logic. and Bob Coecke. The authors apply category-theoretic methods to computational lingustics by mapping the derivations of the grammar logic to the distributional interpretation via a strongly monoidal functor. In the end. Mehrnoosh Sadrzadeh. observables. Coecke confesses that dagger compact categories do not capture all we might want them to. and to more formal automated reasoning processes. The Frobenius operations allow them to assign and . by Dimitri Kartsaklis.

which is a particle that interacts with itself and can even annihilate itself.C.edu . D. 20052 E-mail: harizanv@gwu. Kauﬀman.edu DEPARTMENT OF MATHEMATICS GEORGE WASHINGTON UNIVERSITY WASHINGTON. and is particularly interested in models based upon the Temperley-Lieb categories. Lomonaco over the last ten years. Kauﬀman shows how knots are related to braiding and quantum operators. Kauﬀman presents several topics exploring the relation- ship between low-dimensional topology and quantum computing. For example. phrases. Knot logic and topological quantum computing with Majorana fermions. and sentences in a single space. and produces unitary representations of the braid groups that are dense in the unitary groups. Kauﬀman also gives a quantum algorithm for computing the Witten-Reshetikhin-Turaev invariant of three manifolds. the negation can generate the fusion algebra for a Majorana fermion. Kauﬀman works with braid groups and their representations. Kauﬀman uses the diagrammatic approach. by Louis H. as well as to quantum set-theoretic foundations. Kauﬀman calls the negation the mark. These topics have been introduced and developed by Kauﬀman and Samuel J. He discusses from several diﬀerent perspectives the Fibonacci model related to the Temperley-Lieb algebra at ﬁfth roots of unity. The authors present their experimental results for the evaluation of their model in a number of natural languages. VIRGINIA 20147 E-mail: ea1102@gwu. CA 94117 E-mail: jcchubb@usfca. INTRODUCTION 7 compare the meanings of diﬀerent language constructs such as words. He describes the Jones polynomial in terms of his bracket polynomial and applies his approach to design a quantum algorithm for computing the colored Jones polynomials for knots and links.edu DEPARTMENT OF PHYSICS VIRGINIA SCIENCE & TECHNOLOGY CAMPUS GEORGE WASHINGTON UNIVERSITY ASHBURN. He investigates the relationship between knot-theoretic recoupling theory and topological quantum ﬁeld theory. Thus. DEPARTMENT OF MATHEMATICS UNIVERSITY OF SAN FRANCISCO SAN FRANCISCO.

Additional references to works in this volume are included throughout. or sometimes Hermitian conjugate) is written as a bra. a bra-ket denotes an inner product. and we adopt it here. |ϕ|. and general references appear at the end. A vector in a complex Hilbert space representing a quantum state is written as a ket. 1A Hilbert space is a complete. say |α and |. in general. Chubb. then. normed metric space. and its conjugate-transpose (adjoint. Logic and Algebraic Structures in Quantum Computing Edited by J. Postulate of quantum mechanics: Representing states of systems. |. AND CATEGORY THEORY JENNIFER CHUBB AND VALENTINA HARIZANOV This chapter is intended to be a brief treatment of the basic mechanics. Part 1 (sections 1– 4) surveys quantum mechanics and computation. |. COMPUTATION. The state space of a composite system is the tensor product of the state spaces of the subsystems. Quantum systems can exist in a superposition of more than one basic state: If a quantum system has access to two basic states. The Dirac bra-ket notation for states of quantum systems is ubiquitous in the literature. The second part (sections 5–7) provides a survey of category theory. framework. the system’s “current state” can be represented by a linear combination of these states in complex Hilbert space: | = c1 |α + c2 |. ϕ|. that corresponds to that system. In this notation. Part 1: Quantum mechanics & computation §1. Harizanov Lecture Notes in Logic. and a ket-bra denotes an outer product. 45 c 2016. H. where the norm and distance function are induced by an inner product deﬁned on the space. with sections organized according to the com- monly known postulates of quantum theory. Eskandarian and V. A. Qubits & quantum states. A (VERY) BRIEF TOUR OF QUANTUM MECHANICS. Association for Symbolic Logic 8 . and concepts relevant to the study of quantum computing and information for review and reference. where ||| = 1. and a state is usually described as a linear combination in a relevant orthonormal basis. The state of a quantum system is represented by a unit-length vector in a complex Hilbert space1 . Each one-dimensional subspace of H corresponds to a possible state of the system. The basis elements are often thought of as basic states.

More precisely. i. c1 and c2 .e. (More about measurement of quantum systems can be found in Section 3 below. The coeﬃcient itself. and states that diﬀer only by a global phase factor are identiﬁed. a single qubit) exists in a pure state that may be a superposition of basic states. suﬃce to specify a distinct state via the decomposition | = cos |0 + e sin iφ |1. ] and [0. the range of values taken on by and φ may be restricted to the intervals [0. The squared norm of the state vector | is the inner product of | with itself. So- called mixed states are not proper state vectors. the basic vector |0 points up and |1 points down. the value |c1 |2 is the probability that the system would be found to be in state |α upon measurement. Composite quantum systems. The basic states for this space are usually thought of as |0 and |1. A classical bit can be in only one of two states at a given time. | = c1 |0 + c2 |1. and ϕ| is the corresponding probability amplitude. a qubit is a 2-dimensional quantum system. The quantity |ϕ||2 is the probability that upon measurement. {|+. and ϕ the longitudinal angle. Qubits. describes the latitudinal angle. A quantum bit or qubit may exist in a superposition of these basic (orthogonal) states.. As described above. is called the probability amplitude. the bra-ket |. so two real numbers. The state space of a qubit is often visualized as a point on the Bloch sphere. The norm of a state vector is always one. but at times other bases are used (for example. |−} or {| ↑.) 1. QUANTUM MECHANICS & CATEGORY THEORY 9 The complex coeﬃcients. c1 . | ↓}). they are classical probabilistic combinations of pure states and are best represented by density matrices. 2 ) without any loss of generality. In this visualization. A composition of systems may exist either in a . of |α and | give classical probabilistic information about the state. and φ. Any unit vector that is a (complex) linear combination of the basic states is a pure state and non-trivial linear combinations are superpositions. Basic states are typically the eigenstates (eigenvectors) of an observable of interest (see discussion of measurement below). a single quantum system (for example. Note that states that diﬀer by a global phase factor will (by design) coincide in this visualization. For example. and the (real) probabilities described by the coeﬃcients are the same. where c1 and c2 are complex probability amplitudes. | will be found to be in state |ϕ. Two vectors in H represent the same state if they diﬀer only by a global phase factor: If | = e i |ϕ.1. the state of which is a unit-length vector in H = C2 . then | and |ϕ represent the same state. 2 2 Respectively. Orthogonal states are antipodal on the Bloch sphere. 1. |0 or |1. and so the corresponding distinct states may be mapped uniquely onto the unit sphere in R3 .2.

2 For more on entangled states. Postulate of quantum mechanics: Evolution of systems. §2.1. 2 Such a state of the composite system that can be written as a tensor product of pure states is called separable. as shown in the simple circuit diagram below: . In quantum computing. see Parke’s article in this volume. Example 1. they are important and canonical examples of entangled states: |00 + |11 |00 − |11 √ √ 2 2 |01 + |10 |01 − |10 √ √ 2 2 Example 1. and continuous. The Hadamard gate. Example 1.2. Suppose √ we have a system of two qubits.1.10 JENNIFER CHUBB AND VALENTINA HARIZANOV separable or an entangled state. In the case of an entangled state. Transformations and quantum gates. the√ﬁrst in state | = (|0 + |1)/ 2 and the second in state |ϕ = (|0 − |1)/ 2. they are non-trivial (complex) linear combinations of separable states. The state of the combined system is 1 | ⊗ |ϕ = ||ϕ = (|00 − |01 + |10 − |11). the subsystems cannot be thought of as existing in states independent of the composed system. algorithms are often described as circuits in which information (and time) ﬂows from left to right. Such transformations preserve inner products and are reversible. or Section 6 of Kauﬀman’s article. The time evolution of a closed quantum system is described by a unitary transformation. deterministic. Quantum gates represent unitary transformations applied to qubits in such a circuit. A transformation is unitary if its inverse is equal to its adjoint. Separable states are states that can be written as tensor products of pure states of the constituent subsystems. The GHZ states (for Greenberger-Horne-Zeilinger) are ex- amples of entangled states in composite systems that have three or more subsystems. Example 2.3. The 1-qubit Hadamard gate has as input and output one qubit. Entangled states cannot be so written. The Bell states of a 2-qubit system are not separable. The GHZ state for a system with n subsystems is |0⊗n + |1⊗n √ .

the gate does nothing. QUANTUM MECHANICS & CATEGORY THEORY 11 Its matrix representation (with respect to the basis |0 = [1 0]T . distinct eigenvalues yield orthogonal eigenvectors. These matrices are often described in terms of their spectral decompositions. Another important quantum gate is the controlled-not or CNOT gate. the gate acts by “ﬂipping” the non- control (target) input as follows: If the target input is in state | = c0 |0+c1 |1. Postulate of quantum mechanics: Measurement. If the control is in state |1 (as it is in the diagram above). Moreover. (It should be noted that not all such matrices describe physically meaningful measurements. 2 1 −1 1 This transformation applied to the basic state |0 = results in the 0 1 superposition H |0 = √12 (|0 + |1) = √12 . Thematrix representation of CNOT is the following (given basis |00 = [1 0 0 0] . T T T with respect to the |11 = [0 0 0 1] ): T ⎡ ⎤ 1 0 0 0 1⎢ 0 1 0 0 ⎥ CNOT = ⎢ ⎥. The gate requires two inputs. see Parke’s and Kauﬀman’s articles in this volume.2. then ﬂipping transforms the state to | = c0 |1 + c1 |0. The gate does not alter the control bit. The notion of measurement is described in terms of observables represented by Hermitian (self-adjoint) matrices. and these represent the possible values obtained upon measurement of the observable. Measurement. 1 Example 2. Upon measurement. |10 = [0 0 1 0] . |01 = [0 1 0 0] . |1 = [0 1]T ) is: 1 1 1 H =√ . a system’s . one designated as the control input (passing through the solid dot) and the other as the target input: When the control input is in state |0. §3.) A Hermitian matrix has all real eigenvalues. 2⎣ 0 0 0 1 ⎦ 0 0 1 0 For more on quantum gates and unitary transformations of quantum systems. The controlled-not gate.

§4. If we consider the result of such a measurement as a random variable.12 JENNIFER CHUBB AND VALENTINA HARIZANOV state (or wave function) experiences a “collapse” and is not preserved.2. If a is the result of the measurement of A on |ϕ. No-go theorems and teleportation. two mutually exclusive conclusions may be reached regarding quantum mechanics: either quantum mechanics is incomplete. If the matrix A corresponding to an observable A has (real) eigenvalue a and corresponding unit-length eigenvector |va . consider the possibility that there does exist such an operator U . Let | be an arbitrary state in state space H. Podolsky. This is not the case in quantum computations. Ein- stein. we must have the following: ϕ| = e|ϕ||e = e|ϕ|U † U ||e = ϕ|ϕ|| = (ϕ|)2 . and Rosen (EPR) questioned the completeness of quantum mechanics in the form of a thought experiment involving the measurement of one part of a 2-particle entangled system. the state of the system is the eigenvector corresponding to the eigenvalue that was the result of the measurement. The EPR paradox. 4. Subsequently. . and Bell’s Theorem. Example 3. and so such a U preserves inner product only selectively—the states |ϕ and | must be identical or orthogonal. According to EPR. if the matrices representing two diﬀerent observables are non- commuting. We see that ϕ| must be either 0 or 1 in order for this equality to hold. then the proba- bility that measuring A on state |ϕ will yield the value a is given by |va |ϕ|2 . yielding ||. it must preserve inner products. Complementary observables suﬀer from necessarily limited precision when measured simultaneously as a result of the Heisenberg Uncertainty Principle. In 1935. To “clone” the state |. U ||e = ||. Theorem 4. then the observables are often referred to as complementary and measurements of these observables are subject to uncertainty limits. it is possible to implement error correction by simply duplicating the classical data as needed. we would need to have a unitary transformation that when applied to ||e replaces the ancillary state with a copy of |. Very brieﬂy. After measurement. There is no unitary operator U so that for all states | and ancillary states |e. In classical computation. hidden variables.1. As U must be unitary.1. hence for any and ϕ. the expected value (expectation value) of that quantity is given by ϕ|A|ϕ. To see why.1 (No-cloning theorem). No cloning. and |e be an ancillary state (independent of |) in an identical state space. the system is left in state |va . 4. or the physical quantities associated with two non-commuting operators cannot have simultaneous reality.

the measurement operators corresponding to these observables (position and momentum) do not commute. Another related theorem is the Kochen-Specker Theorem. asserting that no locally realistic theory can make the predictions of quantum mechanics. and checked. In 1964. However. QUANTUM MECHANICS & CATEGORY THEORY 13 building on the behavior of a two-component system under the laws of quantum theory. In the 1970s and 1980s. physical experiments (carried out most famously by Alain Aspect) demonstrated in favor of the former. using | as the control. and hence an exact knowledge of position entails some uncertainty in the value of momentum. To do this. Imagine that two particles. The following scenario captures the idea of the quandary they posed. A related question is this: How does particle B “know” to have a precisely deﬁned momentum and an uncertain position when particle A’s momentum is measured? According to the principle of locality. A and B. EPR argue for the incompleteness of quantum theory. If he subsequently measures the momentum of particle B. which leads to the conclusion that quantum mechanics must be incomplete. he may compute the momentum of particle B exactly due to entanglement. the result will be exactly that computed value. One qubit is in the possession of entity A (Alice) and the other is in the possession of entity B (Bob). or some “hidden variable” or “element of reality” encoding the information as yet unaccounted for by quantum mechanics (assuming determinism or realism). by judicious choice of measurements on particle A. interact and then part ways. This scenario seems to entail either superluminal transmission of information between the particles (violating locality). This is the idea underlying the famous EPR paradox. |. that is. followed by an application of the Hadamard transformation to |. the particles’ positions may be observed. a pair of qubits in a (entangled) Bell state. which she would like to send to Bob. A basic illustration of the power of entanglement is in the quantum teleportation protocol: An EPR pair. Quantum teleportation. The EPR argument makes a case for being able to assign two diﬀerent wave functions (or states) to the same reality (particle B). It would be diﬃcult to overstate the importance of entanglement in quantum computing and the diﬃculty in representing and interpreting this phenomenon in possible quantum logics. . and gave a set of inequalities that would provide a test of quantum mechanics against a local hidden variable theory. What is known as Bell’s Theorem is the summary of all this. Alice also has a qubit. Alice applies a CNOT transformation to her two qubits. are prepared. 4.3. which says that a non-contextual hidden variable theory (one in which the value of an observable in a system is independent of the apparatus used to measure it) is unable to make the predictions of quantum mechanics. Similarly. a physical process occur- ring in one place should not be able to aﬀect a physical process in another location (outside the light cone of the ﬁrst process). If one measures the momentum of particle A. John Stewart Bell formalized (mathematically) the notions of locality and realism. computed.

see Coecke’s and Kauﬀman’s (respectively) articles in this volume. 2008. and tensor products: boxes for operators. Yanofsky and Mirco A. Cambridge University Press. Mannucci. nor Bell’s Theorem (classical information must be transmitted subluminally). by N. Bob preforms one of four corresponding transformations. Nielsen and Isaac L. and (classically) communicates to Bob the (classical) information that results of her measurements. . in the 1970s. 2007. resulting in the transformation of his qubit into the state |. 2007. Oxford University Press. T . 2011. by Eleanor Rieﬀel and Wolfgang Polak. by Phillip Kaye. Cambridge University Press. For alternative formulations of the quantum teleportation protocol in a graphical language and another (similar) formulation in quantum topology. Raymond Laﬂamme. • Quantum Computer Science. MIT Press. David Mermin. Note that this protocol does not violate the no-cloning theorem (Alice’s copy is destroyed). and Michele Mosca. which are of importance in physics 2 This entire process is sometimes called a Bell measurement. Cambridge University Press. and outgoing wires for subscripts. 2011. At the graduate or research level • Quantum Computation and Quantum Information. These diagrams represented various categories. • Quantum Computing: A Gentle Introduction. by Noson S. • An Introduction to Quantum Computing. their products. which Alice wished to transmit to him. Part 2: Category theory for quantum computing In physics. Penrose used graphical language to represent linear operators. Upon receiving this information.14 JENNIFER CHUBB AND VALENTINA HARIZANOV She then measures both of her qubits2 (they are destroyed in the process). the following texts may be useful: Textbooks at the undergraduate level • Quantum Computing for Computer Scientists. For more detailed exposition on all these ideas and topics. Chuang. incoming wires for superscripts. by Michael A.

or simply hom(A. that is. A morphism f : A → B is called a monomorphism or monic if f ◦ g1 = f ◦ g2 implies g1 = g2 for all morphisms g1 . A morphism f : A → B is called an epimorphism or epic if g1 ◦ f = g2 ◦ f implies g1 = g2 for all morphisms g1 . also called monoidal categories. but the converse may not be true. If a morphism has both a left inverse and a right inverse. also called a section of f. also called maps or arrows with speciﬁc abstract properties. which have been used by S. hom(C). A) contains the identity morphism idA such that for every f : A → B. also called a retraction of f. That is. and h : C → D. the class hom(A. by interchanging the domain and the codomain of each morphism. Hence we have the following deﬁnition. We will give a brief survey of monoidal categories. A morphism f : A → B has a right inverse. which is an associative operation that respects domain and codomain information. B) is a set. if there is a morphism g : B → A such that g ◦ f = idA . we have (ii) f ◦ idA = f and (iii) idB ◦ f = f. and it is called locally small is for every pair of objects A. For every pair of objects. g2 : B → C . §5. A morphism f has a domain dom(f) (also called source) and a codomain cod(f) (also called target). A category C is called small if both ob(C) and hom(C) are sets. g2 : C → A. (i) (f ◦ g) ◦ h = f ◦ (g ◦ h). For every object A. A morphism with a right inverse is an epimorphism. the set hom(A. For more details see [3] and [1]. B). A and B. If . The equations (i)–(iii) can be viewed as the axioms for the categories. g : D → A. The morphisms are equipped with composition ◦. which we write f : dom(f) → cod(f). QUANTUM MECHANICS & CATEGORY THEORY 15 and quantum computing. A category C consists of a class of objects. then the two inverses are equal. a morphism with a left inverse is a monomorphism. if there is a morphism g : B → A such that f ◦ g = idB . Of particular importance are tensor categories. It is denoted by C op . Clearly. where f : A → B. A morphism f : A → B is called an isomorphism if there exists a morphism g : B → A such that f ◦ g = idB and g ◦ f = idA . The opposite category (also called dual category) of C is formed by reversing the morphisms. ob(C). Coecke as a framework for quantum theory. Basic category theory. Their categorical quantum mechanics can be also viewed as a suitable quantum logic. there is a class of morphisms denoted by homC (A. The converse may not be true. B) when the category is clear from the context. A morphism f : A → B has a left inverse. B. Abramsky and B. and a class of morphisms.

B are surjective. B of objects from C. each morphism f ∈ hom(A. In the graphical representation. A functor Φ between locally small categories C and D is called faithful if it is injective when restricted to each set of morphisms that have a given domain and codomain. Φ and Ψ. B) → homD (Φ(A). A functor Φ is called full if the induced functions ΦA. They preserve identity morphisms and composition of morphisms. as well as every morphism of C to a corresponding morphism of D such that the following is satisﬁed. More precisely. g is unique and is called the inverse of f.B : homC (A. a natural transformation N : Φ → Ψ consists of the family of morphisms for every object A of C. Natural transformations capture the notion of a homomorphism between two functors. and hence f is the inverse of g. and two functors from C to D. B of objects in C. Functors capture the notion of a homomorphism between two categories. Ψ(A)) in D such that Ψ(g ◦ h) = Ψ(h) ◦ Ψ(g) ∧ Ψ(idA ) = idΨ(A) . For every pair A. Examples of well-known categories include the category of sets as objects with functions as morphisms.16 JENNIFER CHUBB AND VALENTINA HARIZANOV it exists. That is. the category of vector spaces as objects with linear maps as morphisms. and the category of Hilbert spaces as objects with unitary transformations as morphisms. a faithful functor may not be injective on objects or morphisms. On the other hand. That is. B) in C is mapped to a morphism Φ(f) ∈ hom(Φ(A). C and D. A contravariant functor Ψ from C to D is a map that associates to each object A in C an object Ψ(A) in D. such that for every f ∈ homC (A. The composition is represented by connecting the outgoing edge of one diagram to the incoming edge of another. which reverses the order of composition. B). . Φ(B)) in D such that Φ(g ◦ h) = Φ(g) ◦ Φ(h) ∧ Φ(idA ) = idΦ(A) . object variables label edges (“wires”) and morphism variables label nodes (“boxes”). A functor from C to D is also called a covariant functor. Φ(B)) is injective. given two categories. the induced function ΦA. A : Φ(A) → Ψ(A). while the identity morphism is represented as a continuing edge. we have Ψ(f) ◦ A = B ◦ Φ(f). a functor Φ from a category C to a category D is a function that maps every object A of C to an object Φ(A) of D. for every pair A. B) in C a morphism Ψ(f) ∈ hom(Ψ(B). and associates to each morphism f ∈ hom(A. in order to distinguish it from a contravariant functor.

The tensor product of objects is associative in the sense that for every triple (A. Monoidal categories. (See Coecke’s article in this volume for a wire diagram representation of this equation. there is an isomorphism αA. B. A monoidal category captures the notion of a tensor product as a binary operation of objects. B: A ⊗ idB = (idA ⊗ B ) ◦ αA. A ⊗ B. For morphisms f : A → A . The domain of f ⊗ g is the tensor product of the domains of f and g. we have (f ⊗ (g ⊗ h)) ◦ αA. which means that it satisﬁes the following equations for morphisms: (f1 ⊗ f2 ) ◦ (f3 ⊗ f4 ) = (f1 ◦ f3 ) ⊗ (f2 ◦ f4 ) and idA⊗B = idA ⊗ idB . and the codomain of f ⊗ g is the tensor product of the codomains of f and g.C = αA . .B.B. the following triangle axiom is satisﬁed for every pair of objects A.C ◦ (f ⊗ g) ⊗ h). QUANTUM MECHANICS & CATEGORY THEORY 17 The content of the equation is captured by the following diagram.B .C : (A ⊗ B) ⊗ C → A ⊗ (B ⊗ C ). For every object A. f ◦ A = A ◦ (idI ⊗ f). h : C → C . f ◦ A = A ◦ (f ⊗ idI ). g : B → B . A Φ(A) > Ψ(A) Φ(f) Ψ(f) ∨ ∨ B Φ(B) > Ψ(B) §6. In addition. and of morphisms.I.B . f ⊗ g. C ) of objects. there is an isomorphism (left) A : I ⊗ A → A and an isomorphism (right) A : A ⊗ I → A.) A monoidal category also has a constant unit object denoted by I . The tensor product is a bifunctor.

as well as sets with direct products or disjoint unions.D > αA. which captures the correspondence between the formal language and the graphical language we described. Joyal and Street [2] established a coherence theorem for planar monoidal categories.D ◦ (αA. B. C.D ∨ ((A ⊗ B) ⊗ C ) ⊗ D A ⊗ (B ⊗ (C ⊗ D)) > αA⊗B. αA. we often call it planar monoidal category.B.C. and operation symbols (such as ◦ and ⊗). This relationship is visualized in the following diagram. The formal language of categories uses object variables and morphism variables. These are used to form terms and equations (formulas). The coherence theorem of Joyal and Street states that an equation in the language of monoidal categories follows from the axioms of monoidal categories if and only if it holds in the graphical language. When no additional properties are assumed for a monoidal category.I. Other coherence theorem for special .B.B⊗C.C ⊗ idD )) = αA. Roughly speaking.C ⊗idD idA ⊗αB. Tensor product of morphisms is represented by stacking their diagrams. or Hilbert spaces.C ⊗D ◦ αA⊗B. with either direct sum or tensor product.B. up to planar equivalence.C.D .18 JENNIFER CHUBB AND VALENTINA HARIZANOV Both sides map (A ⊗ I ) ⊗ B to A ⊗ B. here. the tensor product of objects is represented by parallel wires (input or output) from the bottom to the top.D (A ⊗ (B ⊗ C )) ⊗ D > A ⊗ ((B ⊗ C ) ⊗ D) ∧ αA.B⊗C. αA.C. and object constants (such as I ) and morphism constants (such as idA ).D ) ◦ (αA.B. a diagram D1 is planar equivalent to a diagram D2 if it is possible to transform D1 to D2 by continuously moving the boxes and wires of D1 (without crossing or cutting). This equation is captured in the following diagram. D: (idA ⊗ αB.C ⊗D (A ⊗ B) ⊗ (C ⊗ D) In the graphical language.B (A ⊗ I ) ⊗ B > A ⊗ (I ⊗ B) A ⊗idB > idA ⊗B < A⊗B Also. and the unit object is represented by no wire. Both sides map ((A ⊗ B) ⊗ C ) ⊗ D to A ⊗ (B ⊗ (C ⊗ D)). the following pentagon axiom is satisﬁed for every quadruple of objects A.C. Examples of monoidal categories are vector spaces.

QUANTUM MECHANICS & CATEGORY THEORY 19 categories are of the similar nature. and its converse is called a completness theorem. The part of a coherence theorem that states that an equation following from the axioms holds in the graphical language is called a soundness theorem. . A braided monoidal category is a monoidal category with a family of isomorphisms for every pair of objects A. Soundness is guaranteed by assuring that the axioms hold in the graphical language. B.

−1 Hence .A.B : A ⊗ B → B ⊗ A.

A. where −1 .B exists.

C : (idB ⊗ . Two hexagon axioms are satisﬁed for every triple of objects A.B : B ⊗ A → A ⊗ B.A. B.

C ) ◦ αB.A.A.C ◦ (.

A.C.B ⊗ idC ) = αB.A ◦ .

C and −1 −1 −1 (idB ⊗ .A.B.B⊗C ◦ αA.

A ) ◦ αB.C.A.C ◦ (.

A ◦ .B.A ⊗ idC ) = αB.C.

αB.A ◦ αA. The ﬁrst of these axioms is captured in the diagram below.B⊗C.C (B ⊗ A) ⊗ C > B ⊗ (A ⊗ C ) ∧ .B.A.C .

A.B ⊗idC idB ⊗.

C.A.C ∨ (A ⊗ B) ⊗ C B ⊗ (C ⊗ A) ∧ αA.A ∨ .B.C αB.

B⊗C A ⊗ (B ⊗ C ) > (B ⊗ C ) ⊗ A It follows that −1 .A.

A.B ◦ .

B = idA⊗B . Graphical language is extended to picture braiding .A.

A symmetric monoidal category is a braided monoidal category where the −1 braiding .A.(over-) crossing.B and is represented by an under.

A.B is the inverse .

It is called symmetry and is graphically represented by a crossing.B. a functor Φ : C → D is called a monoidal functor if there are also morphisms φA.B : Φ(A) ⊗ Φ(B) → Φ(A ⊗ B) and . For monoidal categories C and D.A .

B.C ) ◦ αΦ(A).20 JENNIFER CHUBB AND VALENTINA HARIZANOV φ : ID → Φ(IC ). For every pair of objects A.Φ(B).B ⊗ idΦ(C ) ) = φA. the functor is called a strict monoidal functor. and two strong monoidal functors from C to D. if they are the identity maps.B ◦ ( A ⊗ B ). we have A⊗B ◦ φA. the functor is called a strong monoidal functor. Φ with φ Φ and Ψ with φ Ψ.B Φ Ψ = φA. the last equation has the diagram: Φ(A) I ⊗ Φ(A) > Φ(A) ∧ φ⊗idΦ(A) Φ(A ) ∨ φI. a natural transformation N : Φ → Ψ with morphisms A : Φ(A) → Ψ(A) is a monoidal natural transformation if for every pair of objects A. B of C.C ) ◦ φA⊗B. which preserve the tensor structure as follows.A Φ(I ) ⊗ Φ(A) > Φ(I ⊗ A) If the maps φA.A ◦ (φ ⊗ idΦ(A) ). a monoidal functor Φ : C → D is called a braided monoidal functor if it is compatible with braiding as follows. For every triple of objects A. Φ(αA. Φ(A) = Φ(A ) ◦ φI. For braided monoidal categories C and D. B of C.I ◦ (idΦ(A) ⊗ φ). Given two monoidal categories. C of C.Φ(C ) . Φ(.B⊗C ◦ (idΦ(A) ⊗ φB.C ◦ (φA.B and φ are also invertible (isomorphisms). Φ(A) = Φ(A ) ◦ φA. For example.B. C and D.

A.A ◦ .B = φB.B ) ◦ φA.

Φ(A). An example of a symmetric monoidal category is the category of sets with functions as morphisms. with Cartesian product.Φ(B) . and symmetry given by .

y) = (y. Another example of a symmetric monoidal category is the category of vector spaces with linear maps as morphisms.A.B (x. with tensor product. x). and symmetry given by .

the unit A : I → A∗ ⊗ A and the counit A : A ⊗ A∗ → I . A∗ . A . A and the ﬁrst triangle equality are graphically represented as follows: . (idA∗ ⊗ A ) ◦ ( A ⊗ idA∗ ) = idA∗ . and there are two morphisms. A monoidal category C is called right autonomous if every object A of C has a right dual. which satisfy the following adjunction triangle equalities: idA = (A ⊗ idA ) ◦ (idA ⊗ A ). denoted by A∗ .B (x ⊗ y) = y ⊗ x.A.

A category of sets with binary relations as morphisms and direct product as tensor product and where A∗ = A is a compact closed category. which is identity on the objects and involutive on the morphisms. (g ◦ f)† = f † ◦ g † . Dagger categories. §7. The category of sets with binary relations as morphisms is a dagger category with relational inverse R† as adjoint of R. In general. the adjoint of a diagram is its mirror image. A morphism f is called Hermitian if it is self-adjoint: f † = f. The adjoint is diagrammatically represented by reversing the location but not the direction of the wires and by marking the upper right corner (in contrast to the upper left corner) in the box. and for every morphism g : B → C . A monoidal category is autonomous if it is both right and left autonomous. so the category is autonomous. The category of ﬁnite dimensional vector spaces (or ﬁnite dimensional Hilbert spaces) with tensor product and with A∗ being the dual space of A is a compact closed category. if we allow inﬁnite dimensional vector spaces. A dagger functor Φ between two dagger categories C and D is a functor that satisﬁes the following additional . In a braided right autonomous category. A compact closed category is a right autonomous symmetric monoidal category. The category of Hilbert spaces with bounded linear maps is a dagger category with the usual adjoints. to each morphism f : A → B a morphism f † : B → A is assigned such that † (f † )† = f ∧ idA = idA . the categories of vector spaces and of Hilbert spaces are not autonomous. A dagger category is a category C equipped with a contravariant functor † : C → C. QUANTUM MECHANICS & CATEGORY THEORY 21 idA ⊗ A A > A ⊗ A∗ ⊗ A εA ⊗idA > ∨ idA A A left autonomous monoidal category is deﬁned dually and a left dual of A is denoted by ∗ A. Morphism f † is called the adjoint of f. On the other hand. More speciﬁcally. A morphism f is called unitary if it is an isomorphism and f −1 = f † . a right dual of A is also a left dual of A.

B. A dagger monoidal category C is a category that is both monoidal and dagger and the two structures are compatible in the sense that the morphisms from the monoidal structure. is a dagger symmetric monoidal category that is also compact closed. A . A dagger symmetric monoidal category is a dagger braided monoidal category such that its symmetry (braiding) is unitary. together with a relation to connect the dagger structure to the compact structure. the dagger is used to connect the unit to the counit so that for all objects A in C. A .22 JENNIFER CHUBB AND VALENTINA HARIZANOV equality for every morphism f in C: Φ(f † ) = (Φ(f))† . are unitary and the following equality is satisﬁed for every pair of morphisms f.C . g: (f ⊗ g)† = f † ⊗ g † . also simply called dagger compact category. αA. we have: A = . A dagger compact closed category C. Speciﬁcally.

20052 E-mail: harizanv@gwu. REFERENCES [1] S. editor). [2] A. pp. Finite dimensional Hilbert spaces are complete for dagger compact closed categories. Selinger. That is. Coecke. IEEE. D. 88 (1991). Logical Methods in Computer Science. Springer. Selinger [4] proved a completeness and hence coherence result for dagger compact closed categories. Dagger compact categories are of great importance for foundations of quantum information and computing. pp. pp. Lecture Notes in Physics. pp. Coecke. Joyal and R. DEPARTMENT OF MATHEMATICS UNIVERSITY OF SAN FRANCISCO SAN FRANCISCO. vol. 813. [4] . 8 (2012). A survey of graphical languages for monoidal categories.edu DEPARTMENT OF MATHEMATICS GEORGE WASHINGTON UNIVERSITY WASHINGTON. he established that an equation follows from the axioms of dagger compact closed categories if and only if it holds in ﬁnite dimensional Hilbert spaces.A⊗A∗ ◦ A† . 55–112. Abramsky and B. Proceedings of the 19th Annual IEEE Symposium on Logic in Computer Science. Street. Thus. Advances in Mathematics. this coherence theorem allows us to use the diagrammatic calculus of dagger compact categories to precisely express and verify some fundamental quantum information notions and protocols. 1–12.C. New Structures for Physics (B. [3] P. vol. 2004. 2011. vol. A categorical semantics of quantum protocols. 415– 425. The geometry of tensor calculus I. 289–355. CA 94117 E-mail: jcchubb@usfca.edu .

Some years ago.) In 1974. 45 c 2016. A. Chubb. they can’t converge elsewhere. In particular. very little has been written on the disagreement between Putnam and Kripke. I adopted the device of writing about Paul Kriske and Prof. considerations drawn from quantum mechanics might provide an example. Harizanov Lecture Notes in Logic. We think of logical truths as a special case of necessary truths. Kripke’s paper still hasn’t appeared in print and apart from my 1978 dissertation and a paper I published 28 years later [9].” oﬀering a detailed rebuttal of Putnam’s case. Tupman out of deference to the fact that there is no published version of Kripke’s talk. Not long after Hilary Putnam published “Is Logic Empirical. This paper reviews the debate between Putnam and Kripke. almost 40 years later. This is unfortunate. We would once have said that if two lines are straight and a constant distance apart over some portion of their span. As for the plan of the paper. Page references will be to the reprinted version. For anyone not familiar with Logic and Algebraic Structures in Quantum Computing Edited by J. Here I’ll simply write directly about Putnam and Kripke. Putnam on quantum logic. As of this writing. Association for Symbolic Logic 23 . It suggests the possibility of a “middle way” between Putnam and Kripke: a way in which logic could be broadly a priori but in which empirical considerations could still bear on our views about the logical structure of the world. the issues are well worth investigating.” Saul Kripke presented a critique of Putnam’s argument in a lecture at the University of Pittsburgh. but Putnam reminds us that we now reject certain claims about geometry that once seemed necessary. If I get Kripke wrong. I hope he’ll let us know. Eskandarian and V. we begin by reviewing Putnam’s arguments. Hilary Putnam published a paper called “Is Logic Empiri- cal?” [7] in which he argued that quantum mechanics provides an empirical case for revising our views about logic. In my 2006 paper [9]. Kripke criticized both the substance of Putnam’s version of quantum logic and the idea that one could “adopt” a logic for empirical reasons. This will lead to a larger discussion of what logic and the empirical might have to do with one another.COULD LOGIC BE EMPIRICAL? THE PUTNAM-KRIPKE DEBATE ALLEN STAIRS Abstract. (The paper was republished in his collected works as “The Logic of Quantum Mechanics”. after that we move to Kripke’s rebuttal. §1. Saul Kripke presented a talk at the University of Pittsburgh called “The Question of Logic.

associated with rays α1 and α2 . and lines that depart from geodesics will not seem straighter. (α1 ∧ 2 ). then one also appreciates that the usual ‘linguistic’ moves only help to distort the nature of the discovery and not to clarify it. Putnam claims that this seems as intuitively clear as saying that there are no married bachelors. gives us cases where both disjunctions are true. (p. however. shortest paths are straightest and conversely. (α2 ∧ 1 ). however. 177) Putnam argues that we’ve made a similar discovery about logic itself. The notion of a geodesic preserves this. We pair statements about quantum quantities with subspaces of Hilbert space and we can extend this map from simple statements to compound ones by associating “or” with subspace span (p ∨ q). Putnam insists that this won’t do. but “geodesic” doesn’t mean “straight line.24 ALLEN STAIRS non-Euclidean geometry. likewise. In the case of the lines. Putnam thinks we miss the signiﬁcance of relativity if we represent its geometrical claims as mere changes of meaning. that B has two values b1 and b2 associated with rays 1 and 2 . (1 ∨ 2 ). Quantum mechanics. (α2 ∧ 2 ) and so the corresponding conjunctions are false. That means the conjunction (A = a1 or A = a2 ) and (B = b1 or B = b2 ) is also true. we’ve come to believe not just that the claim might be false but that in some instances it is false. Suppose. If we take the mapping seriously. On our intuitive conception. or that nothing can be scarlet all over and bright green all over at the same time. He writes: The important point is that [‘straight line’] does not ‘change meaning’ in the trivial way one might at ﬁrst suspect. we’ll have to say that there can be points with no straight line between them. we have a conﬂict with classical logic. One way to put it: if we say that geodesics behaving as Putnam describes aren’t straight lines. However each of the following pick out the null subspace of Hilbert space (α1 ∧ 1 ). Now consider the expressions (A = a1 or A = a2 ). read as Putnam reads it. (B = b1 or B = b2 ) and associate them with the subspaces (α1 ∨ α2 ). Suppose the quantity A has two possible values a1 and a2 . and “not” with orthocomplement (p ⊥ ). Hence (A = a1 ∧B = b1 )∨(A = a1 ∧B = b2 )∨(A = a2 ∧B = b1 )∨(A = a2 ∧B = b2 ) . “and” with subspace intersection (p ∧ q). We might say that what Putnam describes applies to geodesics.” However. Once one appreciates that something that was formerly literally unimaginable has indeed happened.

the account he gives (on p.. a position by virtue of the truth of a disjunction of position statements and it also has a momentum by virtue of the truth of disjunction of momentum statements1 . He also argues that quantum logic accounts for the two-slit experiment.” understood as the failure of quantum mechanics to specify joint values for noncommuting quantities comes down to logical incompatibility in quantum logic. “Complementarity. Putnam oﬀers several illustrations. 185) Quantum states are “logically strongest consistent statements” but they aren’t states “in the sense of statements which imply every true proposition about S” (p. look especially at statement (8) and Putnam’s comment on it) but let that pass. the oversimpliﬁ- cation is merely for illustration. (p. this isn’t Putnam’s view. 183) can’t be right for any ﬁnite population of atoms (exercise for the reader. just read the logic oﬀ from the Hilbert space H (S). and that their failure to provide a complete list of all the truths is a reﬂection of our epistemic situation. 179) and the anomalies go away if we change our logic. However. 182) In fact.’ (p. If we treat ‘R ∧ (A1 ∨ A2 )’ as equivalent to ‘(R ∧ A1 ) ∨ (R ∧ A2 )’. 179) The advantage. the state provides a complete description relative to the terms of the theory. we avoid explaining the eﬀect by appeal to a supposedly mysterious ‘disturbance by the measurement. e. In quantum theory. says Putnam. he tells us that a quantum system has. . 185) This might suggest that quantum states are like statistical states in classical mechanics. we must change our logic. (p. Here is Putnam articulating what we will call the value-deﬁniteness thesis: 1 Putnam knows that strictly. To derive the incorrect classical probabilities.” but Putnam writes that A system has no complete description in quantum mechanics. . then we end up with the wrong probabilities. we have to distribute a proposition R about where the photon hits the screen over a disjunction of propositions A1 and A2 about which hole the photon passes through. In classical physics. Putnam also claims that if we analyze barrier penetration quantum logically. such a thing is a logical impossibility (p. the com- plementary quantities don’t share eigenspaces. there are no position and momentum eigenstates. .g. of the system. Rather. Putnam writes: Conclusion: the mapping is nonsense—or. 179) On the other hand. if we do “adopt the heroic course of changing our logic” there’s a straightforward way to proceed: . COULD LOGIC BE EMPIRICAL? THE PUTNAM-KRIPKE DEBATE 25 is false but this discrepancy between the distributed and undistributed formulas is impossible classically. (p. is that all so-called anomalies in quantum mechanics come down to the non-standardness of the logic. there are “states” or “state descriptions.

184) Putnam’s argument for adopting quantum logic is that if we do. The only laws of classical logic that are given up in quantum logic are distributive laws . according to Putnam. . granting that Kripke is right about Putnam’s particular quantum logical proposal wouldn’t show that logic isn’t empirical. For any such question as ‘what is the value of M (S) now?’ where M is a physical magnitude. In what follows. . and the reader will get a better sense of it if s/he reads Kripke’s own words. One deals with the particulars of Putnam’s argument. nor would it show that quantum mechanics doesn’t give us a reason to change our views about logic.1. In the second part of his critique. §2. However. the interpre- tive puzzles of the theory dissolve. The ﬁrst part of Kripke’s argu- ment is intended to show that if we follow Putnam. We could preserve Euclidean geometry. but we can’t predict them all. I quote at length from my partial transcript of Kripke’s talk. and every single anomaly vanishes once we give these up. there exists a statement Ui which was true of S at t0 such that had I known Ui was true at t0 . Kripke on Putnam. To be sure. if we make an appropriate measurement. There Kripke’s case is strong. Likewise for classical logic: we can preserve it only by paying an unacceptable price in the coin of untoward claims about quantum systems. (p. but 2. It is logically impossible to possess a statement Ui which was true of S at t0 from which one could have predicted the value of every magnitude M now. Quantum logic and simple arithmetic. We can predict any one magnitude. I could have predicted the value of M (S) now. but Kripke’s own words do a better job than my paraphrase would of spelling out his view and therefore. is this: These examples makes the principle clear. it seems fairest to him to use those words. This completes the analogy with geometry. we have to agree to the . Kripke argues that the very idea of changing logic for empirical reasons is confused. The advantage of giving up classical logic. I think it is more likely that classical logic is wrong than that there are either hidden variables or “cuts between the observer and the system”. but only by paying the high intellectual price of admitting gratuitous universal forces. we have to say such supposedly objectionable things as that measurements create the values of the quantities measured or that there is a “cut between the observer and the observed” or that there are undetectable hidden variables. The indirect debate between Kripke and Putnam was an important episode. etc. 2. there is a matter of propriety here. . If we insist on classical logic.26 ALLEN STAIRS 1. . Kripke’s critique of Putnam has two parts. But Putnam says .

we must say ‘the particle has no momentum’. It is as simple as that. each with two possible values 1 and 2. . A=2∧B =2 But Kripke argues: . are momentum states. A = 1 ∨ A = 2 2.) We suppose Sz to be known. Is it really clear that Putnam meant this? Here’s a passage that would be hard to make sense of otherwise. etc. A=2∧B =1 4. (Substitute eigenstates of diﬀerent spin components if you prefer. Kripke draws out an untoward consequence. Sz is a position state and T1 . 186) “Simple” or not. According to Putnam.e. T2 . A=1∧B =1 2. COULD LOGIC BE EMPIRICAL? THE PUTNAM-KRIPKE DEBATE 27 untoward conclusion that 2 × 2 ≥ 5. A=1∧B =2 3. the error was in passing from the falsity of Sz · T1 ∨ Sz · T2 ∨ · · · ∨ Sz · TR to the falsity of Sz · (T1 ∨ T2 ∨ · · · ∨ TR ). Suppose we’re given two quantities. he will also say that each of the following are false: 1. it does not create it. or disturb it in any way. Putnam would say. . if M has possible values m1 . and hence that we must reject (T1 ∨ T2 ∨ · · · ∨ TR )—i. and the momentum measurement merely ﬁnds this momentum (while disturbing the position). Then one measures momentum. we could predict the outcome of an M -measurement. If I know that Sz is true. This latter statement is true (assuming Sz ) and so it is true that ‘the particle has a momentum’ . However. and one gets a momentum—so the measurement must have ‘brought it into being’. and if we knew which. A and B. Thus the set {1. 2} is the set of possible values of A and also of B. mn then there is a true statement ascribing one of these values to M . . The statement M = m1 ∨ M = m2 ∨ · · · ∨ M = mn is true. Putnam will say that 1. . the logically strongest statement about the system may not tell us which disjunct is true. However. then I know that for each Tj the conjunction Sz · Tj is false. m2 . The idea that momentum measurement ‘brings into being’ the value found arises very naturally if one does not appreciate the logic being employed in quantum mechanics.. B = 1 ∨ B = 2 are both true. However. . and the summary of the value-deﬁniteness thesis above makes clear what he means: one of the disjuncts really is true. . It is natural to conclude (smuggling in classical logic) that Sz · (T1 ∨ T2 ∨ · · · ∨ TR ) is false. (p.

Also. 1 if we adopt the usual criterion of identity of ordered pairs because that would mean that A = 1 and B equals 1. {c. 1. Kripke insists: he’d only be begging the question if he had assumed a premise that Putnam rejects. B equals 1. d .. B does not equal 1. 2}— AS] because B is either equal to 1 or to 2.e.e. The “fallacy” is that if x comes from {a. 2}) and {c. ” But we can all see the fallacy in any conclusion that these are the only pairs.. is that this is absurd. So the classical arithmetician says “There are four. then we have the disjunction x =a∨x =b and similarly we have y = c ∨ y = d. c. b} and {c. Kripke’s point.. but that it’s where we end up if we follow Putnam. the distributive law isn’t a premise in his argument. b. So the pair A. b} = {1. of course.e. . b. b}. 2} as well. 2. We want to consider how many ordered pairs there are. 2 because it is false that A equals 1 and B equals 2. {c. where x comes from the ﬁrst set and y comes from the second set. but Kripke insists that one obvious rejoinder won’t do: it won’t do to accuse Kripke of begging the question. 2 gives the joint values for A and B. But certainly we cannot say that A. d } are our two two-element sets. 2}—AS] because A is equal either to 1 or to 2. I might say. B does not equal 2. A. B does not equal 2. d } = {1. a. So there is a ﬁfth and hitherto overlooked. The idea is that A was already one of the two numbers 1 and 2 [and] B was already one of the numbers 1 and 2. . 1. Remember that Putnam does not think these are funny pseudo-numbers. then joint values require that there be another pair in the . . 1 [and] A. . However. 2. And A. b} = {1. d } is the set of values of the quantity B (i. ordered pair in the Cartesian product of these two ﬁnite sets. B where these are the two quantities mentioned by Putnam. Then the cardinality of their product is the cardinality of the Cartesian product of the two sets .e. . d . 1. so where x comes from {a. {a. b} and y comes from {c. So A is certainly in the ﬁrst set [i. b} is the set of possible values of the quantity A above (i. namely a. d } where {a. .28 ALLEN STAIRS The usual mathematical deﬁnition of multiplication is this: suppose we have two sets with two elements. 2 . c. B is certainly in the second set [i. though we may not have measured which. and that contradicts [the falsity of (3)]. 2.. Kripke simply reasons from premises that Putnam accepts to the conclusion that if none of the pairs 1. There may be various ways Putnam could respond.) We’ll let Kripke pick up the story: Now I claim that there is a ﬁfth pair A. B is in our Cartesian product. Now suppose that the set {a. {a. d } = {1.

. are begging the question. The impossibility of “adopting” a logic. . Formal systems are not logic. the untoward (and absurd) conclusion follows. Whether Putnam really holds the view of logic that Kripke attributes to him isn’t clear. the mere fact that an old system struck us as the only intuitively acceptable one should be given little weight. As Kripke sees it. I believe: there’s no convincing quantum-logical defense of the value-deﬁniteness thesis. Putnam remarked on our intuitive sense of contradiction when faced with his geometrical example. then we have no place to stand outside of reasoning from which to do this. Once one has a rival system of axioms. and once alternative logics are under consideration. As Kripke puts it in connection with a closely-related example . Kripke reads Putnam this way: Just as in the case of non-Euclidean geometry we throw intuition to the wind and adopt an axiomatic system as supposedly describing the real physical world .” by which Kripke means reasoning. Kripke is right. we will assume that value-deﬁniteness doesn’t hold. Since Putnam would claim that none of the four ordered pairs gives the joint values. he thinks. But there is no neutral place outside logic from which to decide what “logic” to adopt. Once alternative geometries are under consideration. If changing our formal system is supposed to entail changing the way we reason. while we won’t ignore the question of how well Kripke’s criticisms ﬁt Putnam. . “Logics” qua formal systems aren’t logic. because only if my reasoning was invalid did I need any extra premise which I have begged against Putnam. . we abandon any mere intuitive preference for Euclidean geometry. Kripke’s larger point is that there is a problem at the core of Putnam’s view. Putnam. so on every other domain we cannot rely on intuition. 2. and would also claim that both quantities really do have values. COULD LOGIC BE EMPIRICAL? THE PUTNAM-KRIPKE DEBATE 29 Cartesian product. Formal systems may or may not faithfully reﬂect correct principles of reasoning. Therefore.2. to assess the formal systems. . Putnam’s fundamental error lies in missing this point. I think. (See [8] for more discussion) and in what follows. There is reasoning. Speciﬁc formal systems may or may not adequately capture aspects of correct reasoning. Kripke insists that this is incoherent. Kripke thinks there is a deep confusion here. but we have no alternative to using “intuition. We misunderstand logic if we think there are “logics” among which we could somehow choose. we abandon any mere intuitive preference for a particular system of logic. if you say that I am begging the question then you yourself. believes that we could somehow decide to “adopt” a logic. That said. it’s a useful foil for making Kripke’s own view of logic clearer. the exegetical question won’t be our main concern.

the idea that we could change our logic in response to empirical considerations makes no sense. his main objection seems to be that this is a strange notion of “measurement.e. but it doesn’t count against the possibility that things really work this way. In order to rebut a measured version of Putnam’s view. this is too quick. we won’t be able to argue that the probabilities in the two-slit experiment must ﬁt a crude application of the law of total probability. we’ll be blocked from drawing untoward conclusions. However. as Kripke would be the ﬁrst to insist. for instance) typically have the sort of self-supposing quality that Kripke’s point relies on. Putnam’s larger point would be made if sometimes what we discover empirically can properly enter into our deliberations about how to reason. 4]. What William Alston called doxastic practices (see his [1. on logic qua reasoning) but it’s not clear that he means this or needs to say it. we’ll make use of claims that we’ve accepted on the basis of the implicit and explicit rules/practices we use for assessing other kinds of empirical claims. In considering what weight to give memory. Ch. We leave the vexed issue of measurement (or “measurement”) aside and turn to a diﬀerent part of Kripke’s reply: his case that the very idea of “adopting a . that hardly makes a case for merely opportunistic “revisions” of logic. The idea that the interactions we call measurements bring new states of aﬀairs into being might be a reason to pick or invent a diﬀerent word. The distributive law seems to be a correct way to reason. Even if we grant that there’s no place to stand outside reasoning. Putnam’s main argument seems to be that if we give up the distributive law. he merely tells us that he ﬁnds them unlikely.” This threatens to turn the argument into a mere quibble. We can also reason about how to reason. Putnam hasn’t shown us any deep problems with the idea that there are Bohmian-style hidden variables.. we’ll need to rely on at least some such beliefs and hence on the practice of forming beliefs based on sense evidence. Equally important. We can reconsider how to evaluate beliefs based on sensory input. we can’t avoid relying on at least some memories when we do. We can consider what memory can and can’t teach us. And while Putnam might judge that a failure of the distributive law is “more likely” than hidden variables or “cuts between observer and observed. for example. We might be able to avoid any number of unwelcome conclusions if we simply refused to reason in certain ways. we haven’t been presented with an intelligible alternative. Kripke would have to show that reasoning is the one doxastic practice to which the deliverances of other doxastic practices are irrelevant.” Kripke can reply that without something more than a mere and tendentious cost-beneﬁt analysis. Putnam may seem to be saying that we can evaluate logic without relying on logic broadly conceived (i. and we won’t need to say that measurement creates the values that it records. He objects to the idea that measurement might bring the values it yields into being. when we do. Thus. there’s a more general phenomenon here. Be that as it may.30 ALLEN STAIRS Once we grasp this. However. these practices aren’t insulated from one another.

J . each instance follows. Kripke makes similar points about non-contradiction. Perhaps that doesn’t apply to all logical principles. how. COULD LOGIC BE EMPIRICAL? THE PUTNAM-KRIPKE DEBATE 31 logic” makes no sense.” he will say. Imagine further that our poor reasoner is willing to accept Kripke’s authority that all ravens are black and is also willing to accept Kripke’s authority in more general logical matters. . I’m not entirely sure. can one possibly deduce anything from them? Kripke develops the example of universal instantiation at greatest length. It won’t help because he would already have to be in command of the principle to apply it. . Well let me tell you: from every universal statement. What I mean is this: you can’t undermine intuitive reasoning in the case of logic and try to get everything on a much more rigorous basis.” “Yes. We can embody these principles in formal systems. and one can adopt them as one will. You don’t see that. Kripke thinks this would miss the point. and in any case Putnam’s example had to do with giving up rather than adopting a principle.” “Well. hmm. Yes” He agrees. Kripke tells the tale charmingly: So I say to him “Oh. the idea that he could adopt it is incoherent. So I say to him “All universal statements imply their instances. . One has just . then giving him some super-premise like “Every universal statement implies each instance” won’t help him either. standing outside these principles could adopt them. Kripke takes his cue from Lewis Carroll’s “What the Tortoise Said to Achilles” and from Quine’s “Truth by Convention”. adjunction and Lewis Carroll’s modus ponens example. but there’s no sense to the idea that someone. However. He says: The basic problem is this: if logical truths are mere hypotheses . I believe you. Imagine someone who doesn’t see that from a universal claim. therefore J is black” where J is a particular raven.” He will say “Oh. each instance follows. “I don’t really see that I’ve got to accept that!” The problem is clear. Here’s what he says: . unless one has a logic in advance. So ‘All ravens are black’ is a universal statement and ‘This raven is black’ is an instance. This particular statement that all ravens are black implies this particular instance. These are all cases where we couldn’t adopt a particular principle unless we already grasped it intuitively. so to speak. Yes. but he doesn’t see that believing this and accepting that all ravens are black commits him to accepting that J is black. “Ah. . out of our subject’s sight. As Kripke puts it If he was not able to make the simple inference “All ravens are black.” So now I say to him. There’s a raven. I don’t really mean that we adopt as basic just those things to which we can ﬁgure out that this argument applies.

or do empirical considerations sometimes come into play? 2 Kripke talks brieﬂy about cases where we see that an argument we once accepted is invalid. one has to reason. One can only reason as we always did.. Here he claims that the intuitionists introduced new connectives. deﬁned in terms of provability. And if proof by cases was part of our intuitive apparatus then there is no analogy to geometry which says that this should not be respected. We can also agree that questions about how to reason have a special status among the various kinds of questions we can ask. we’ll be tempted to think subalternation is valid. Intuitive reﬂection makes clear that something has to give. He also oﬀers a cursory discussion of intuitionism. What Kripke says is surely right: if we accept subalternation. However. but we do so simply and straightforwardly by reasoning. any reasoning outside the system of postulates can be thrown to the dogs—must be wrong. we overlook the case where ‘P’ is empty. is this always a matter of ordinary reasoning. When we discover that we’ve overlooked a case and accepted an incorrect logical principle as a result. of excluded middle isn’t really in competition with the classical principle.32 ALLEN STAIRS to think not in terms of some formal set of postulates but intuitively. Rejecting subalternation as a case of change in logic. We can agree further that for at least some logical principles there’s no sense to be made of the idea that we might “adopt” them.3. The most useful place to begin is with what he says about the principle of subalternation for universal categoricals—in particular. however. Whether that’s the best reading of intuitionists such as Brouwer I will leave to others to decide. we can ask if this is always so. We correct ourselves by mere reﬂection—by ordinary reasoning. independently of any special set of rules called “logic. and we can even concede that nothing could count as adopting a logic wholesale. To make progress.”2 Logicians once accepted this principle. That is. One can’t just adopt a formal system independently of any reasoning about it because if one tried to do so one wouldn’t understand the directions for setting up the system itself. e. But if there are no deserters. we need to look at what Kripke concedes about changes in logic and how he accounts for them.” in setting up a formal system or in doing anything else. Here we change our beliefs about logic. Whether this scuttles the idea that empirical considerations could bear on logic is less clear.g. that “All P are Q” implies “Some P are Q. it would be very odd to say that some deserters are shot. ‘All deserters will be shot on sight’ may be true. . intuition can be thrown to the dogs—that is. and that may be exactly the reason why there are no deserters. And so any comparison of logic to geometry which says that in the case of logic as in the supposed case of geometry. If we overlook empty terms. Kripke is surely right: logic isn’t just a matter of formal systems. and yet we no longer do. and so the intuitionist’s apparent rejection. 2.

hence ‘not-P’ is true. . pp.3 Another familiar account of future contingents appeals to branching time and supervaluation (see [12]). What if it can’t? One response is to abandon excluded middle—to claim that when ‘P’ is indeterminate. then even if ‘P’ is indeterminate. bivalence and the empirical. according to which the present and the past but not the future are real. Two sorts of views suggest that the answer might be yes. . Consider a debate that Kripke doesn’t mention but that has a long history: whether propositions about the future provide reasons to give up bivalence.] for useful discussion. using ‘F ’ as a future-tense operator. In that case. On this approach. . this permits true disjunctions with no true disjuncts. suppose that presentism is correct. and false if it is false on each such branch. COULD LOGIC BE EMPIRICAL? THE PUTNAM-KRIPKE DEBATE 33 2. there’s a plausible objection: if ‘P’ is indeterminate. but such that on each branch passing through 3 Scope matters here. even though ‘F (not-X )’ is indeterminate. Suppose {P1 .) However. ‘P ∨ not-P’ will be true by virtue of its second disjunct. Future contingents. “There will be a sea battle tomorrow” or “This atom will decay an hour from now”) are contingent in a more-than-merely-logical sense. not logically exhaustive. However. Another is the view that only the present exists. Contingent statements about the future will therefore be neither true nor false.g. past present and future. Then even though future states of aﬀairs don’t exist. On the other hand. Suppose. though it’s not clear how or why. are ontologically on a par. P2 . deterministic laws plus the facts about the present could suﬃce to settle the truth or falsity of propositions about the future. An event’s being undetermined is a matter of its relationship to other events and to the laws of nature. What. ‘P ∨ not-P’ is likewise indeterminate (a view usually associated with Łukasiewicz. ‘not-F (X )’ is true. 82 ﬀ. then it’s not true that P. See Bourne [3. that determinism is false. . If the so-called “block universe view” is correct. all events. Neither future contingent propositions nor presentism alone make the case against bivalence. we’ll use “presentism” for both. a statement about the future is true at the present moment just in case it holds on each branch or history passing through this moment. if presentism is true and determinism false? Perhaps bivalence about future contingent propositions can still be defended.4. usually called “presentism” and the “growing block” view. the claim is that when ‘F (X )’ is indeterminate. Pn } is a set of future contingent propositions that are mutually exclusive. . then. If so. whether we live in a block universe and whether the laws are deterministic are independent questions. that is. Suppose some propositions about the future are contingent in the sense of not being determined by facts about the past. The diﬀerence between presentism proper and the growing block won’t matter for our purposes. the facts about the events making up the block entire settle the truth or falsity of future contingent propositions even if determinism is false. One is that some propositions about the future (e.

Then The coin will come up heads or the coin will come up tails is true at the present moment even though neither disjunct is. not everyone agrees that Putnam’s arguments are sound. The crucial additional assumption is presentism. the possibility of a true disjunction with no true disjuncts isn’t part of logical business as usual. Putnam himself famously invoked special relativity to argue against presentism (albeit not under that name) in “Time and Physical Geometry. Even with excluded middle intact. even if determinism is false. However. In either case. that whether presentism is true depends on the facts about space-time. 4 Of course. empirical considerations plausibly bear on which is correct. empirical matter. the claim is empirical. and reasoning alone won’t tell us if it holds. there are able defenders of the coherence of presentism and of the block universe. However. the point here is not to take sides in this debate. Future contingent propositions give rise to a dilemma: if bivalence holds for such propositions. then. it’s because of something about the world: the falsity of presentism or the truth of determinism. This suggests that matters of logic depend on the way things are. and it might be argued that this is not an empirical matter. However.” [6] His argument that past.4 The broad issue is whether claims about reality could have consequences for logic. An artiﬁcial example: suppose a coin will be tossed. See. and that. Indeed. . Our belief that true disjunctions require true disjuncts came from overlooking a (complex) possibility: the combination of presentism and future contingents. The novelty seems at ﬁrst to sit comfortably with Kripke’s view. for example. If bivalence fails. is enough to make Putnam’s larger point. present and future are equally “real” don’t rest on general philosophical considerations. as Putnam’s view would maintain. it depends on the structure of Minkowski space-time. but that on each branch the outcome is either Heads or Tails. it’s because presentism is true and determinism false. it seems. that the outcome isn’t determined. The status of determinism is a contingent. The question of determinism is certainly empirical and the question of presentism is at least arguably so. The overlooked possibility is a substantive one. certainly the debate has often proceeded as though it’s not. Stein [11] and Bourne [3]. whether bivalence holds is an empirical matter. This particular case for true disjunctions without true disjuncts depends on assumptions about the world: that the block universe view and determinism are both false.34 ALLEN STAIRS the present. Supervaluation preserves excluded middle and non-contradiction. as we noted above. this wouldn’t be enough to undermine bivalence. It may be. If so. Whether it preserves all classical logical truths might be more of an accounting issue than a substantive one. empirical facts about the world. To repeat. one of them is true. Thus. it suggests that assessing the need for the logical revisions at issue in the debate over future contingents depends on contingent. further thought may seem to favor Putnam. If both views are indeed coherent.

broadly speaking. but suppose we grant them. We would still know that the scenario Putnam describes is possible in a broad sense. The logic of the actual world could still be a contingent matter. there’s another approach: treat . The structure is the lattice of subspaces of Hilbert space. That this is possible remains so even if the possibility isn’t realized. The ﬁrst is whether there’s a case of this sort to be made by appeal to quantum mechanics. but that it’s logically impossible for us to know which. It would just be that it’s never actualized. Likewise. that bivalence could fail is arguably not empirical knowledge. the conclusion that bivalence isn’t a correct principle of logic is not an empirical one. We’ll take that up in the following section. There’s a plausible Kripkean reply. We come to it by reﬂecting on the possibilities. This raises two questions. That it is or isn’t pseudo-Riemannian is an empirical claim. but suggests a possibility for d´etente. a priori. but science sometimes takes strange turns) we became convinced that the world is Euclidean after all. but the beginnings of the disagreement with Kripke come from the interpretive overlay. 2. However. one of the disjuncts in parentheses really is true. Quantum logic reconsidered. The question of what the detailed geometry of a world could be would remain. That it does (or does not) fail in a particular way is arguably empirical. but if so the correct conclusion is that bivalence is not a principle of logic. and we discover that there is a genuine possibility we had overlooked: the possibility that there are no facts to ground the truth or falsity of certain propositions about the future. Let ‘Sz = + 12 ’ say that the electron has spin +1/2 in direction z. That the world could be pseudo-Riemannian is not empirical knowledge. Logic writ large (let’s write bold-face ‘logic’ for that) would remain a matter of reasoning. The analogy with geometry helps here. D´etente? What’s just been said concedes something important to Kripke. Suppose (unlikely. Furthermore. broadly under- stood. would say that when 1 1 1 Sz = + ∧ Sx =+ ∨ S x =− 2 2 2 is true. questions about the status of bivalence plausibly count. is it quite so clear that logic really is something we can know by a priori? §3. as we know. the question of what it is in fact would be empirical. Whether bivalence holds might be an empirical matter. And though we won’t try to give a general account of what counts as a question of logic. and similarly for ‘Sx = + 12 ’ and ‘Sx = − 12 .5. The second question will be raised but no more than raised: in light of what quantum mechanics teaches us. COULD LOGIC BE EMPIRICAL? THE PUTNAM-KRIPKE DEBATE 35 The arguments above are skeletal and open to challenge. Putnam’s quantum logical proposal oﬀered a formal structure and some interpretative principles and rationales.’ Putnam.

This is the heart of what Bohr called complementarity and it has two characteristic features. Note that we aren’t following Putnam’s approach. Quantum Bayesians ¨ such as Carleton Caves. Christopher Fuchs and Rudiger Schack are exceptions. we solve the interpretive problems of quantum mechanics. Putnam argued that if we adopt a strong set of logical claims. Quantum mechanics gives reasons of a diﬀerent sort. that’s because the author is waving his hands. Thus: if there’s a true statement P = pi then there is no true statement Q = Qj . First. Those constraints aren’t beyond challenge. The goal isn’t to defend a view in detail (indeed. The paradigmatically curious quantum example is the case of two quantities— call them P and Q—that share no eigenstates. but our purpose isn’t to make an iron-clad case. but those proofs provide another reason.36 ALLEN STAIRS ‘(Sx = + 12 ∨ Sx = − 12 )’ as a disjunctive fact—as a case of a disjunction that’s true in spite of not having a true disjunct. their [4]. and to do it in a way that doesn’t stray far from what a typical physicist would ﬁnd plausible. Stick with our complementary quantities P and Q. When P has a value. See. then it’s impossible for all quantum quantities to have values at the same time. and if the reader ﬁnds it hand- waving. Few physicists would disagree. It’s to make it plausible that quantum mechanics has consequences for logic. I am by no means certain that the view is correct) but simply to make its outlines clear enough to consider. for example. if we’re certain what outcome a P-measurement would yield. What follows is intended merely as a sketch. we are not certain what outcome a Q-measurement would yield. A system can have an energy or a spin in a particular direction.5 The second point goes beyond ordinary common sense but not beyond the common sense of most physicists. The goal in this section is to see how we might move from here to something more clearly relevant to logic. The ﬁrst point is simple: quantum mechanical quantities can have values. there’s no arrangement that measures P and Q at the same time. all values of Q have at least some positive probability. We’re trying to see what quantum mechanics might teach us about logic if we start from things that many physicists already believe. Second. For some relevant discussion see Stairs [10] . It’s that there are purely disjunctive truths about quantum 5 Though few would disagree. If we accept a handful of plausible constraints. There’s no such goal here. Q does not. We’ve already seen reasons of one sort for taking the idea of disjunctive facts seriously. The third point starts with a piece of physics common sense and then moves a bit beyond. No doubt most physicists believed this before no-hidden-variable proofs became widely known. this isn’t the same as saying none would.

the statement Sx = 0 is deﬁnitely false. however. |z0 and |x0 are orthogonal. Kochen and Specker’s [5] partial Boolean algebra approach is a better way to characterize l-complementarity formally. Few physicists would be shocked to be told that ‘(Sx)2 = +1’ is true when ‘Sz = 0’ is true. It’s also of a piece with our second point to say that the system doesn’t have a deﬁnite x-spin. this quantity has a value: +1. the subspace that goes with E=e for some values e of the energy may not be one-dimensional. consider a simple but instructive example: a spin-one system whose z-spin is 0. COULD LOGIC BE EMPIRICAL? THE PUTNAM-KRIPKE DEBATE 37 mechanical systems.6 ‘Sx = +1’ and ‘Sx = −1’ stand in a diﬀerent relationship to ‘Sz = 0’ than ‘Sx = 0’ does. Neither ‘Sx = +1’ nor ‘Sx = −1’ is true. is unclear to say the least. It’s also too restrictive: in principle l-complementarity is more general than Hilbert space non-commutativity. For the states that go with these statements. In the language of Hilbert space. The relationship between ‘Sz = 0’ on the one hand and ‘Sx = +1’ and ‘Sx = −1. I propose l-complementarity. start with a special case of our ﬁrst point: degenerate quantities can have values. To see why this is plausible. and we have a true disjunction with indeterminate disjuncts. Instead. In spite of this. we can say that for (Sx)2 to take the value +1 and for Sx = +1 ∨ Sx = −1 to be true are one and the same fact: ‘Sx = +1 ∨ Sx = −1’ is true even though neither disjunct is. But if the square of the spin is +1. on any orthodox account. Why anyone would insist on any such thing. But while that picks out the sorts of cases we’re interested in. For present purposes. though for diﬀerent reasons than in the case of future contingents. but there’s no standard word for the relationship between the statements themselves. The state |z0 is a superposition of |x+ and |x− . When X and Y are l-complementary they do not belong to a common Boolean 6 Note that even if someone insisted that each disjunct is false. the term is superposition. . there’s nothing strange about saying that the system really can have energy e—that ‘E = e’ can be true. it doesn’t make a connection with logic. In short.’ is not found in classical physics. With that in mind. bivalence fails. it would be gratuitously peculiar to say that ‘Sx = +1’ and ‘Sx = −1’ are both false. ‘Sz = 0’ excludes ‘Sx = 0’ in an old-fashioned classical way: the two are contraries. propositions are l-complementary when their associated projectors don’t commute. But consider the degenerate quantity (Sx)2 — the square of the spin in the x direction. we’d still have a true disjunction with no false disjunct. On any orthodox account. Again. For example: energy is often degenerate.

We can assert the disjunction without accepting the value-deﬁniteness thesis. The expression (Sz = 0 ∧ Sx = +1) ∨ (Sz = 0 ∧ Sx = −1) either fails to pick out an element of the algebra of propositions (on the partial Boolean algebra approach) or picks out a statement that can’t be true (on a lattice approach. 2. it’s one that we came to by way of quantum mechanics. perhaps not. this leaves the logical point unclear. Bivalence fails. then there are possible states of aﬀairs in which X is true but Y is neither true nor false. What I hope to have made plausible is that they aren’t shocking. That’s why we can’t distribute. This is a case of revision in light of ﬁnding an overlooked possibility: the possi- bility of l-complementary propositions. However. This isn’t what Putnam would say.38 ALLEN STAIRS subalgebra of the partial Boolean algebra. consider distribution. it’s clear that Putnam’s picture goes beyond saying that ‘Sx = +1 ∨ Sx = −1’ is true. In particular. X ∨ Y and X ∧ Y are only deﬁned when X and Y belong to a common member of the family. some statements about quantum systems are neither true nor false. and quantum mechanics was an empirical discovery. The proposal on oﬀer is that l-complementarity goes with a particular kind of failure of bivalence: if propositions X and Y are l- complementary. but not in a way that threatens looming arithmetical catastrophe. Unrestricted distribution of “and” over “or” fails. there’s no question of “standing outside logic” and choosing a logic. The proposal under consideration includes these points: 1. once we recognize the possibility of disjunctive facts. 3. Quantum mechanical quantities sometimes have values.7 However. either +1 or -1 but that it’s logically impossible to state this value along with the z-spin value. A full discussion would call for much more detail (see Stairs [9] for some additional thoughts) but we turn to a diﬀerent question: how well does the proposal meet Kripke’s worries? First. Kripke’s “missing pair” is nowhere in the neighborhood. With this in mind. if l-complementarity is a genuine possibility. grasping the 7 A partial Boolean algebra is a family of Boolean algebras that share a common 0 and 1. . The proposal is that both conjuncts are true. but neither disjunct of the disjunction is true. He would say that the x-spin has a deﬁnite value.) The distributive law fails. though not all quantities have values at once. Disjunctions can be true even though none of their disjuncts are. consider Sz = 0 ∧ (Sx = +1 ∨ Sx = −1). Perhaps (1)–(4) ﬁt quantum systems. On the one hand. 4. However.

is that if quantum mechanics is true. Similarly. nor disjunctive facts. Reasoning won’t tell us the actual geometrical structure of the world. but empirical discoveries won’t tell us what the geometrical possibilities are. nor the possibility of a world where distributivity fails. nor for saying that bivalence fails in every domain. The analogy with geometry is still apt: the possible structures. We can (dimly) imagine discovering that the best theory of space-time is Euclidean after all. but the discovery that something is a non-actual possibility is not an empirical discovery. However. Coda: Some loose ends and some thoughts on logic and the limits of thought. Compare: suppose failures of bivalence are possible because it’s possible that determinism and the block universe picture both fail.8 The point. If so. albeit unrealized possibility. for all we can say for sure. That admission wouldn’t call for treating all propositions as neither true nor false. the world embodies a logical relationship that hadn’t been noticed before: the one we’ve called l-complementarity. Recall the case of geometry. we’ll ﬁnd that the correct account of quantum phenomena is some version of Bohmian mechanics. If we do. go beyond the actual. A question that often comes up in discussions of quantum logic is whether it’s meant to apply universally. How- ever. is that something we might have taken to hold in all cases—as a matter of logic—holds only in some. The proposal. it gives no reason at all to think that mathematical propositions are non-bivalent.’ to use the phrase in Bacciagaluppi’s “Is Logic Empirical?” [2] The point of view of this paper is that this is a misleading question. however. not all propositions are bivalent and distributivity fails in certain special circumstances. there’s still a 8 In particular. §4. What’s been said also doesn’t take issue with the idea that our knowledge of logic is a matter of reasoning. to take one important example. it doesn’t apply to every set of propositions. the structure that the world actually instantiates—logical or geometrical—is something we can only discover empirically. . nor failures of distributivity. Even if l-complementarity is a genuine possibility. We have argued that even if Kripke is right and logic is not empirical. Empirical ﬁndings may prompt us to have thoughts we wouldn’t have had otherwise. rather. logical or geometrical. physics would give us no reason to believe that the world exhibits l-complementarity. COULD LOGIC BE EMPIRICAL? THE PUTNAM-KRIPKE DEBATE 39 implications for logic comes from reasoning about the theory and the con- clusions about logic. so to speak—whether quantum logic is ‘the true logic. However even if non-Euclidean geometry didn’t ﬁt this world. this wouldn’t undermine the possibility of l-complementarity. nor the possibility of disjunctive facts. That’s not because this is beyond dispute. non-Euclidean space-time would be a genuine. and it would survive a change of physics. It’s because a central aim of the paper was to see where things stand if we concede to Kripke that what we’ve labeled logic is a matter of reasoning. rather.

However. However.) The point. (A partial Boolean algebra. if this is correct it has an interesting implication: we might not be capable of grasping all of what logic encompasses. The world. That wasn’t inevitable. after all. logic is empirical even if logic isn’t. And a healthy suspicion about our limitations suggests there’s no guarantee we would be . is this. we needed the capacity to grasp the relevant concepts. but about which ones are realized. Logic would come unpinned from reasoning in a diﬀerent way than the one Kripke argued against. partial Boolean algebras. of course. However. is a family of Boolean algebras. In light of this. as abstract mathematical structures is. A proper modesty suggests that there’s no guarantee that we’ll ﬁnd them even if they exist.40 ALLEN STAIRS place for empirical considerations in thinking about logic. collectively. The quantum logical story sketched here sees what we’ve called l-complementarity as a feature of the world. We are. rather. One might dismiss this as a silly kind of skepticism. One might fairly say that the persistent diﬃculty in understanding quantum mechanics has been understanding what it means for the world to have the structure that the mathematics seems to attribute to it. Similar remarks apply to orthomodular lattices. The empirical is not about what the logical possibilities are. The same goes if the l-complementarity-based account of quantum mechanics gets the character of the world right. That would be fair if the suggestion were that we might be deeply and radically ignorant about logic. in turn could have the consequence that we are incapable of grasping the full logical structure of the actual world. And it might be that whatever that structure is lies beyond our cognitive reach. The diﬃculty is in grasping what it means for states of aﬀairs in the world to mirror that structure. there are individual people who lack that capacity. That leaves a perplexing possibility that we’ll raise but not unravel. a bit too cutely. On the contrary (though we haven’t discussed this) a full explication of l-complementarity assumes that propositions are sometimes related exactly as classical logic says they are. not the issue. collective truth though this may be. so this story goes. it might be that the actual geometrical structure or the actual logical structure of the world isn’t what we think it is. Studying. However. Even if we had all been unable to think the right thoughts. Suppose space-time indeed has the structure of a pseudo-Riemannian manifold. the possibility that there might be yet more esoteric exceptions to business as usual doesn’t seem quite so silly. for example. the world would still be pseudo-Riemannian. it doesn’t apply to everyone and need not have applied to anyone. In order to ﬁgure this out. after all. This. lucky enough to be able to grasp the relevant structures and concepts. Go back to the case of geometry. has logical structure just as surely as it has geometrical structure. that’s not the thought. What quantum mechanics may well represent is a case in which we stumbled on a surprising exception to logical business as usual. a full account of l-complementarity calls for positing relationships among properties that we don’t grasp easily.

Gabbay. The problem of hidden variables in quantum mechanics. 174-197. pp. pp. Studies in History and Philosophy of Modern Physics. 255. [2] G. 49–78. 240–247. Bourne. Elsevier. Kochen and E. vol. pp. Springer. Time and physical geometry. Ithaca New York. D. Indeterminist time and truth value gaps. 1975. Cambridge University Press. Demopoulos and I. editors). Alston. Matter and Method. Is logic empirical?. Subjective probability and quantum certainty. Reidel. Theoria. Fuchs. Stairs. pp. [8] A. 2006. Matter and Method. vol. pp. p. Physical Theory and its Interpretation (W. Cambridge University Press. Handbook of Quantum Logic (D. 5. Perceiving God. D. Bacciagaluppi. pp. [9] . Logic in its fullness just might be beyond our grasp. 578–602. 1991. 65 (1968). Oxford. 2006. Thomason. H. Journal of Mathematics and Mechanics. Engesser. [12] R. Schack. Reprinted in Mathematics. and R. vol. 36 (1970). Is logic empirical?. pp. Clarendon Press. DEPARTMENT OF PHILOSOPHY UNIVERSITY OF MARYLAND COLLEGE PARK. Reprinted as The logic of quantum mechanics in Mathematics. Studies in History and Philosophy of Modern Physics. 64 (1967). Journal of Philosophy. pp. editors). [7] . M. Cornell University Press. vol. [4] C. Dordrecht. A loose and separate certainty: Caves. REFERENCES [1] W. and K. Putnam. realism and value-deﬁniteness. Pitowsky. 198-205. 38 (2007). COULD LOGIC BE EMPIRICAL? THE PUTNAM-KRIPKE DEBATE 41 able to recognize them even if they’re there. Fuchs and Schack on quantum probability one. editors). C. Specker. The Journal of Philosophy. Caves. 216–241. Quantum logic. 1968. Kriske. Boston Studies in the Philosophy of Science (Robert S. pp. A Future for Presentism. 59–87. Amsterdam. [5] S. vol. Philosophy of Science. Lehmann. 158–166.edu . [10] . Cohen and Marx W. 2009. vol. 42 (2011). On Einstein-Minkowski space-time. 17 (1967). [3] C. 1975. 50 (1983). MD 20742 E-mail: stairs@umd. [11] H. 5–23. pp. 264– 281. [6] H. Wartofsky. vol. Stein. Tupman and Quantum Logic: the quantum logician’s conundrum. vol. A.

Developing quantum computers requires new ways of thinking beyond those in the familiar classical world. Getting a ﬁrm hold on quantum theory is not an easy task. our predilection for ﬁnding logic behind the behavior of what we observe. Moreover. information storage and transfer is made central. we give a description of the foundational ideas that hold in all of our successful physical models. To help in this thinking. Logic and Algebraic Structures in Quantum Computing Edited by J. Chubb. 2 Our curiosity is enhanced by genetic selection. a description of nature that is hard for us to conceptualize. Eskandarian and V. has led us to quantum theory. Quantum computers take advantage of interfering quantum alternatives in order to handle problems that might be too time consuming with algorithms based on classical logic. As background to quantum theory and quantum computing. so that foundational ideas are explicit. because our experiences and even our genetic predispositions have been developed in a world in which quantum eﬀects are largely washed out. PARKE Abstract. In our description. and 1 Although these days.2 including that of electrons and atoms. Our tact will be to build on the concept of information. Association for Symbolic Logic 42 . Introduction. so that we can anticipate what might happen next. an attempt is made here to give the primitive notions and essential observations that underlie current physical theories. the concept of energy is several steps removed from more basic ideas. A. and to generate new ideas. accurate. For application to quantum computing. energy transfer is used to characterize interactions in current theories. as there is advantage to being able to make sense of what goes on around us. and explains a wide variety of phenomena with only a few statements and input. §1. Having a grasp on the ideas behind a theory helps to apply it correctly. but the Universe. 45 c 2016. However. Our emphasis will be on the proper interpretation of our theories. THE ESSENCE OF QUANTUM THEORY FOR COMPUTERS WILLIAM C. including quantum theory. and a common language is established. and not just their statements. the essence of quantum theory is given. but also lies underneath all natural processes.1 Remarkably. to understand its limitations. together with special precautions and limitations.3 A short description of quantum theory follows. Harizanov Lecture Notes in Logic. 3 Traditionally. macroscopic quantum eﬀects can be seen in the actions of lasers and of quantum ﬂuids. information processing is not only the purpose of computers. but is logical. which lies central to the operation of not just computers.

7 For example. M. and by its economy. They can also be transformed away in alternate but equivalent theories. Classical electrodynamics in terms of direct interparticle action. Madelung in Quantentheorie in hydrodynamischer Form. Physical theory and reality. Theories which take space as continuous implicitly do so only down to the scale permitted by our instruments. Reviews of Modern Physics. 425–433. however. wave functions are clearly not physical. invented by Wheeler and Feynman. ¨ Dirac to be completely equivalent to Erwin Schrodinger’s. since our measuring instruments are ﬁnite. Feynman. 5 J. There should be no implication that even continuity exists at ﬁner scales. 4 In information theory terms. Neither do various so-called hydrodynamic formulations. 21 (1949). The transformed theory. but predictions of continuous values will never be proved to match nature exactly. A good example is the transformation of Maxwellian electrodynamics into an action-at-a-distance form. pp. and particularly does not say. vol. . whether the proposed theory has only a few relationships and input data needed for its ability to explain observations over a wide realm. 40 (1927). Here is another caution: Predictions of pure counts are testable as either true or false. §2. Electric ﬁelds do not ‘exist’ in nature. it makes the same predictions as Maxwell’s theory. But Maxwell’s theory does make deﬁnite statements about observations using the electric ﬁeld concept.7 Rather. one should think of the symbols and relationships in a theory as tools for making predictions. 6 We generally use Maxwell’s theory to solve electrodynamics problems because the Wheeler- Feynman theory is a more complicated mathematical system. Just as it is possible to transform. In quantum theory. Even so. pp. become too enamored with the auxiliary structures within a successful theory. THE ESSENCE OF QUANTUM THEORY FOR COMPUTERS 43 then applied to quantum computing. they are complex numbers. Zeitschrift f¨ur Physik. in general. it is also possible to so transform a physical theory. They exist as symbols on paper and in our minds. Wheeler and R.6 A lesson from this example and others is that one should not endow physical meaning to all the symbols and relationships in a theory. isomorphically. Only those points in the theory that are stated as predictions can be connected to nature. in areas where conceptual diﬃculties have arisen. A. the information contained in the independent data explained by a theory should be much larger than the information needed to express the theory. vol. Predictions are the touchstones in the theory.e. It is judged by its accuracy in matching measurements. Werner Heisenberg’s formulation of quantum theory. uses no wave functions. shown by P. All else is ancillary. 322–326. a logical structure into an equivalent one involving distinctly diﬀerent rela- tionships and symbols. P. A physical theory is a logical model capable of making predictions of what we observe. such as that of E. focusing on what the theory says.4 We should not. i.5 no longer contains electric or magnetic ﬁelds. A.

a set of tentative propositions and observations underlying all physical theories is proposed.44 WILLIAM C.10 With I = ln2 W . then that system has no information. Each observable subsystem will be referred to as a physical system. 10 If there were two independent systems of multiplicity W and W .8 which describes. of their interactions. vol. the theory is not tight. all of the interactions yet detected. where k is Boltzmann’s constant. In the next section.11 If the multiplicity W of a system decreases. see S. E. The Standard Model has made remarkable and now veriﬁed predictions and agrees with the most precise of measurements made to one part in a trillion. then the multiplicity 1 2 of both together would be W1 W2 . including information transfer in the presence of noise. Szil´ard. Leo´ Szil´ard showed that each bit of information we gather from a system and discard necessarily requires an increase in entropy of at least k ln 2. (C. 53 (1929). except for gravity. The number of ways is called the system’s ‘multiplicity’. and of the yet unexplored regions in nature. 379–423 & 623–656. The condition f(W1 W2 ) = f(W1 ) + f(W2 ) makes f(W ) proportional to ln(W ). On the decrease of entropy in a thermodynamic system by the intervention of intelligent beings. A mathematical theory of communication.) Claude Shannon developed the formalism of information theory. taken to be an additive quantity which grows with the number of distinct ways that the system may be conﬁgured under given physical constraints. S) of a system with S = k ln W . 34 (2004).9 To be additive across independent systems. 12 In the late nineteenth century. 5–13. its reading transmits one bit of information. vol. 9 One of the many remarkable implications of quantum theory is that the count W can be performed over a denumerable number of quantum states of a system. The number of divisions may reach a limit. pp. European Physical Journal C. Information storage and transfer will be seen to be fundamental to natural processes. and so forth. having many unexplained interaction strengths and masses. pp. the equivalent of a yes or a no. Even so. Zeitschrift f¨ur Physik.12 8 For a personal perspective in the development of the Standard Model. Weinberg’s article. 11 If a given system subject to physical constraints cannot be re-conﬁgured. Basic properties of physical systems. If the system has two possible conﬁgurations. we say the system has become more ‘ordered’. the parts may be called ‘components’ of the system. The making of the standard model. PARKE Our best physical theory so far is the so-called ‘Standard Model’. Ludwig Boltzmann introduced the number W (‘Wahrschein- lichkeit’). 840–856. connecting it to the disorder (Clausius’ ‘entropy’. If a physical system can be further divided. 1948). Bell System Technical Journal. W .) . The natural world is divisible into a collection of observable subsystems. A physical system can store information. building toward the foundations of quantum theory and application to quantum computing. the information I in a system must be proportional to ln W . the information is given in ‘bits’. Shannon. pp. (L. with quantum ﬁeld theory. We expect new theories will give a deeper and simpler explanation of particles. but no more. §3. 27 (July & October. vol.

In an N dimensional space. one of which is prepared as a measuring instrument. with only one such state realized by observation. by deﬁnition. . . 59 (1999). each smallest repeating cycle called a period of the clock. . 13 This grounding is particularly poignant in quantum theory. then space has at least N − 1 spatial dimensions. wherein a quantum system is described by a set of interfering possible states for each observable.15 If N localizable systems can be spatially separated from each other by the same distance. 2. such as the photon. Observation is made by allowing two physical systems to interact. speciﬁes when and where an observation has occurred. no. exchanges information between them. and references back to Wolfgang Pauli. up to a selected constant factor. An open physical system can interact with other systems. x N }.13 A statement about a physical system is predictive if it relates a number of observations of that system. with time deﬁned and measured by the number of repeats. A measuring instrument is a physical system whose information gathered from an observed system is capable of being copied with a relatively high assurance. For a deﬁnition. pp. Hawton. . THE ESSENCE OF QUANTUM THEORY FOR COMPUTERS 45 An interaction between two physical systems. x 2 . are found to consistently have the same number N of periods. Two systems with a ﬁnite distance between them are said to be spatially separated. {x 0 .14 An isolated physical system is said to be ‘closed’ when external interactions which might inﬂuence the results of intended measurements of that system are negligible. x 2 . Physical Review A. a complete set of such coordinates for one location is denoted {x 1 . it is localizable. prepared in the same way. x 1 . These coordinates are measured by one observer relative to an ‘origin’. to be the minimum time needed for an observable change in one of those two systems to cause an observable eﬀect on the other. . there would be no signiﬁcant eﬀect. that system can act as a clock. . a location used by that observer to coordinate a set of bodies. these clocks are said to be ‘good’ to a precision of at least one part in 1/N . x N }. The copy will act as a record of the observation. vol. A physical system is isolatable if the measurable eﬀect of all interactions with other external systems can be made arbitrarily small. The distance between two interacting physical systems is deﬁned. . Photon position operator with commuting component. 14 We will use the term ‘small’ for a quantity which has the property that if made smaller. Space is deﬁned to be the set of available distances between all systems. see M. The spatial coordinates of a body are the minimal set of numbers that uniquely determine a deﬁnable location within the body. . . 954–959. is tricky. If a set of observations of a system is found to repeat. 15 Deﬁning the localizability of zero-mass particles with spin greater than 1/2 (in units of Planck’s constant over 2 ). If a large set of independent periodic systems. A system localizable in each spatial dimension can be referred to as a body. Statements about a physical system are veriﬁed only by observations. An event. If one isolatable system can be spatially separated from all others.

i. . each measurement made in a single frame of reference. Systems with many interacting components. Bodies with no macroscopic motion relative to the observer are said to be stationary. and then left alone.e. A ‘free particle’ is a particle whose interactions with other systems can be neglected. A life system is one which is capable of self-replication by interactions with external systems. Some systems. i. have been successfully described when those components can be tracked. The ‘dynamics’ of a physical system. A fundamental particle is a particle that suﬀers no measurable change in its intrinsic information even after engaging in all available interactions.e. With a suﬃcient variety of particles and interactions. quantities that are independent of how the observer measures them. 16 A ‘biosystem’ is a physical system whose activities support life. These schemes are most easily tested using isolatable ‘simple systems’. The velocity of a body is its spatial change per unit observer’s time along each of the independent spatial directions. Recording a complete set of observables in a system determines (to the degree possible) the information present in that system at the time of measurement and before any further interaction with the system. or when statistical likelihood arguments become meaningful. Sys- tematics in the behavior of complex systems make global properties referred to as emergent relationships. A particle is a localizable physical system with some identiﬁable intrinsic characteristics. i. a description of how interacting subsystems change over time. will have a non-zero lower bound on the probability of being found in the initial volume later in time. through the mutual interactions of their particles.46 WILLIAM C. those with only a few discernible component subsystems and low information content. when initially localized in a certain volume with zero average velocity. using information stored within the life system. The selection of observables is made such that the measurement of any one does not change the result that would be found for the measurement of any other in the selected set. Rules for optimal dynamics in biosystems16 form others. and the acceleration is the change of velocity per unit time. called ‘complex systems’ or ‘macrosystems’. Those of thermodynamics and statistical mechanics are examples. Those observables that are time independent are called conserved. So far. The ability to create bound systems gives preference to the evolution of diﬀerentiated systems and to condensation into locally ordered subsystems. follows logical predictive schemes which reveal cause and eﬀect.e. If the spatial separation between two bodies changes with the observer’s time. we say they have relative motion.e. will form bound bodies. i. PARKE A frame of reference characterizes how one observer records events. A conﬁned system is one which. systems that retain their localized character provided external interactions are suﬃciently weak. all physical systems can be described by the interactions of fundamental particles in space-time.

The dimension of our space is at least three. Introduction to Thermodynamics of Irreversible Processes.20 Examination shows that our Universe has a ﬁnite limiting speed. The strong experimental support of our conservation laws in three dimensions suggests that if higher dimensions of space existed.19 The Principle of Relativity allows for the existence of a ﬁnite universal limiting speed for all bodies. frame moves at a speed v away along the positive x-axis if the second inertial of the ﬁrst. pp. vol. Observation shows that. We will use the term ‘inertial observer’ for an observer in an inertial frame. That relationship is the Lorentz transformation. 2265–2279. New York. This makes t2 ∼ = t1 . containing a ﬁxed universal speed called c. so that time becomes universal in this limit. pp. the constancy of the speed of light. under reasonable assumptions about how events are measured in close by inertial frames with relative motion. 405–426. At present. To the precision of current measurement. 20 As demonstrated by H. initially aligned. Reciprocal relations in irreversible processes. Bulletin des sciences math´ematiques. z2 = z1 . or motion relative to the events. there is no evidence for higher extended dimensions than three. then x2 = (x1 − vt1 )/ 1 − (v/c)2 . orientation. and the volume of our Universe apparently is also ﬁnite. Relativity alone. This is the grand ‘Principle of Relativity’.17 including the evolution of life. the interactions due to electromagnetism and gravity carry information between bodies at the universal speed c. 37 & 38 (1931). Interscience. The Galilean transformation is approached when the universal speed in the Lorentz transformation is taken much larger than the relative speeds of the observed bodies. is not needed. an interaction experienced by one body can always be associated with the eﬀect of other local bodies. St. 302–324. I & II. y2 = y1 . t2 = (t1 − (v/c 2 )x1 )/ 1 − (v/c)2 . At small scales. 18 Three dimensions is also the minimum dimension needed to build a computer or brain having more than four devices with mutual connections.18 The distance between widely separated bodies has been growing relative to the size of the smallest bodies since time started. vol. In inertial frames. Space-time as background to quantum theory. Physical Review. Poincar´ e in L’´etat actuel et l’avenir de la physique math´ematique. THE ESSENCE OF QUANTUM THEORY FOR COMPUTERS 47 the evolution of complexity in open subsystems is natural. Onsager. 17 L. There is a limit to the greatest separation between bodies. allows only one relationship between their space-time coordinates. to a good approximation. Louis Conference. The Universe is deﬁned as the collection of everything that can be observed. there exist inertial frames in which an isolated body nearby and initially stationary relative to the observer will continue to be nearly stationary. Explicitly. the relationships between local events can be expressed in a form that is independent of the observer’s position. 1955. Einstein’s second postulate. In our Universe. Prigogine. 28 (1904). but these are taken as part of the initial conditions in dynamics and so do not vitiate the relativity principle. the residual eﬀects of these initial conditions on present observations of local events are often small. §4. The Universe appears to have existed in a ﬁnite number of current clock periods. 19 Radiation from distant galaxies and radiation left over from the hot big bang do establish a unique frame of reference. I. matter and energy would have had extreme diﬃculty passing into or out of it. .

An important example of a four-vector is a particle’s four-momentum. The set of quantities {g } form what is called the metric tensor. The length of {p }/c is the mass of 22 the particle. with cp 0 being the energy of the particle and p its spatial momentum. then the acceleration of the pair will be half the rate of the single one. repeated indices. The inertial mass of a particle is an intrinsic property. {p }. b κ } forms the so-called ‘Poincar´e group’. dx 2 . where x 0 ≡ ct. but acted on by one other body some distance away. b } = {a a . and the length of A is A A . −1. so that dx dx is a ‘scalar’. A general coordinate transformation between frames of reference. x = f(x). 23 Rotations.48 WILLIAM C. PARKE In Relativity. We say the pair has twice the ‘inertial mass’ of the single body. This is Einstein’s Equivalence Principle. in the form {dx } = {dx 0 . so that the metric tensor is well approximated by {g } ≈ diag{1. dx 1 . 22 The energy and momentum of a system are best deﬁned. dx 3 }. This makes the concept of space and time inseparable. Lorentz transformations. ds = g dx dx invariant. (a b + b )}. A vector ‘dual’ to {dx } can be deﬁned by dx ≡ g dx . The set {a . one lower. 24 Reﬂections are excluded by imposing det |a| = 1. i. The sum A B deﬁnes the scalar product of the two vectors.e. If a duplicate of the ﬁrst body is weakly bound to the ﬁrst. should be summed from 0 to 3.24 κ κ A body initially stationary in an inertial frame. and the experiment repeated. with the product rule {a . in our successful theories. through the generators of time and space translations. −1}. Any other ordered set of four quantities forms a four-vector if they transform by coordinate transformations just like {dx } does. The observation of the eﬀects on the motion of bodies due to the acceleration of the observer’s frame with respect to an inertial frame is locally indistin- guishable from the eﬀects of gravity. one observer’s measure of spatial separation between two bodies is related to a combination of space and time coordinates of another observer moving relative to the ﬁrst. −1. . and displacements are included. a number who value is independent of the frame of reference of the observer. These ideas will be presented shortly in the context of Noether’s Theorem and Einstein’s General Relativity Theory.e. with a scale determined by gravity. Each inﬁnitesimal space-time region within any inertial frame can be covered by an orthogonal coordinate grid. independent of the observer’s frame of reference. will accelerate. and gives utility to the idea of a four-vector using the coordinates in space and time for a pair of close by events. Einstein’s Equivalence Principle makes inertial mass the same as ‘gravitational 21 By convention.21 i. becomes a ‘Poincar´e transformation’ when x = a x +b and the coeﬃcients {a } satisfy g a κ a = gκ . Then the transformations are called ‘proper’. one upper. 23 Note that this relation makes the metric components an ‘invariant tensor’. Relativity makes the small interval between two events. in that the components take the same values after a coordinate transformation.

even in the presence of mass. In Riemannian 26 geometry. and the metric tensor {g } can no longer be transformed by a coordinate choice to the form {g } = diag{1. The size of the extra perihelion precession of Mercury’s orbit. THE ESSENCE OF QUANTUM THEORY FOR COMPUTERS 49 mass’. Inﬁnitesimal changes in any vector that are observed while transporting that vector along a path deﬁne the ‘covariant derivative’: Dκ A = ∂κ A − Γ κ A . must be proportional to the area of the loop and the size of the original vector ﬁeld.28 Einstein discovered that in empty space. The changes due to the underlying geometry come from the ‘connections’ Γ κ in the space. where the loop is given orthogonal sides dx and dy . The amount of the gravitational 25 The Equivalence Principle also means that mass m can be measured in distance units by giving Gm/c 2 . κ + ∂ g − ∂ gκ ). where G is Newton’s gravitational constant that determines the strength of gravity. Annalen der Physik. A (x). Die Grundlage der allgemeinen Relativit¨ atstheorie. Curvature can be characterized by the behavior of vectors as they are moved from one point to another across space. requires that the distance measure of space-time in the presence of a gravitating body be non-Euclidean. the condition on the metric curvature tensor29 given by R κ = 0 numerically predicts: Newtonian gravitational ﬁelds when the eﬀects of gravity diﬀer little from ﬂat space. Einstein. The change A in the components of any vector ﬁeld. 28 More general theories can be constructed using higher derivatives of the metric tensor in the ﬁeld equations than the second. vol.25 The mass of any localized system (including the equivalent mass of any associated localized ﬁeld energy) can be measured by using the gravitational pull that system creates on a distant mass. The proportionality constants in each small patch of space-time deﬁnes the curvature tensor {R κ } in that patch. The vector A (x0 ) − x0 Γκ A dx κ is said to be the components of the ‘parallel transport’ of the original vector at x0 along a particular path to x. −1. −1} in any ﬁnite region of the space near the body. pp. 50–205. the connections x are determined by gradients of the metric tensor. 26 In the form g Γ = (1/2)(∂ g κ 27 A. However. 49 (1916).e. there will be intrinsic curvature to the space-time around a body with mass. Einstein showed that the eﬀects of gravity due to masses could be found from conditions on the Riemannian curvature of space-time. i. together with the Principle of Relativity. to wit: A = R κ A dx κ dy . The Equivalence Principle. by carrying the vector in parallel transport around an inﬁnitesimal closed loop. Einstein’s General Theory of Relativity27 is the simplest of a class of theories that incorporate the Equivalence Principle and the Principle of Relativity. inertial observers will still ﬁnd an approximate ﬂat metric in their inﬁnitesimal neighborhood. which is the intrinsic property of a body that determines the strength of its gravitational inﬂuence on nearby systems. . −1. 29 The metric curvature tensor {R } is that part of the local curvature tensor {R } due κ solely to changes in the metric across space-time.

Quantum theory. Without these. They are described in: R. i. and. the very rarely detected ultra-high energy cosmic rays may be scattered by this quantum granularity of space. bodies acted on by gravity follow a ‘geodesic’.6 × 10−35 m.30 In both the Special and the General Theory of Relativity. time is not universal. The diﬀerence in the time elapsed by the clocks will be the diﬀerence between the values of |g dx dx |/c. 1965. Weinberg and E. and cannot spontaneously decay. no. Sitzungsberichte der K¨oniglich Preussischen Akademie der Wissenschaften zu Berlin (1899). Although this is far smaller than the regions we can explore with current accelerators. Quantentheorie schwacher Gravitationsfelder. vol. carry no charge. pp. errors in positions would be unacceptable! 31 This leads to the ‘Twin Paradox’. PARKE deﬂection of light. §5.31 In General Relativity. All ‘dynamical ﬁelds’. P. R. See S. Planck. vol. The essence of quantum theory.50 WILLIAM C. Hibbs. integrated along the path of each clock from the common starting point to the common endpoint. a path that makes the invariant four-dimensional distance ds along the path between ﬁxed initial and ﬁnal points of the motion extreme. Limits on massless particles.32 Einstein’s General Relativity Theory describes how the classical ﬁeld {g } should vary over space-time. 34 M. The resolution came from Einstein using his General Theory of Relativity. 140–157. yet each sees the other move away and then come back. Bronstein. Quantum Mechanics and Path Integrals. and one is set in motion relative to the other. 33 See. 1–2. P.e. a dynamical ﬁeld varies both over space and in time. that one twin can end up younger than the other. pp. Formally. no localizable charge can be carried by a massless particle with spin greater than 1/2. Boiled down to its essence. Feynman and A.33 We expect that the quantum aspects of gravity will be important near the ‘Planck scale’ 34 G/c 3 ≈ 1. for example. All these and more have been conﬁrmed to the precision of current instruments. Physics Letters. Free particles that travel at the ultimate speed c also follow geodesics. 35 Feynman began thinking of these ideas in 1942.1. Uber ¨ irreversible Strahlungsvorg¨ange. Generally. quantum theory follows from a prescription due to Feynman:35 30 Calculations of position on Earth using Global Positioning Satellites at height h and speed v over an Earth of mass ME and radius RE . 59–62. F¨unfte Mitteilung. . B 96 (1980). 32 In relativistic quantum theory. The interval for the slowing of clocks in a gravitational ﬁeld. are necessarily massless. 5. pp. Witten. 9 (1936). McGraw Hill. M. nor can there be a localizable ﬂow of energy and momentum for massless particles with spin greater than 1. If two good clocks are synchronized in one frame of reference. they may diﬀer in the number of periods each had when they are brought back together. ﬁelds which have a kinetic energy term in the Lagrangian for the system are dynamic. must have corresponding quanta. 440–480. Physikalische Zeitschrift der Sowjetunion. have Special Relativity corrections included to order v 2 /c 2 for the relativistic Doppler shift and General Relativity corrections included to order GME h/(c 2 RE 2 ) for clock slowing in a gravitational ﬁeld. to be consistent with quantum theory.

a system can be in an interfering combination of possible realizable events before 36 For an excellent description on how Feynman paths are constructed.36 The function L. called the transition amplitude. 5th edition. depends on the particle coordinates and time changes of coordinates for the possible paths between A and B. In contrast to the determinism of Newtonian theory. Path Integrals in Quantum Mechanics. as a sum of unimodular complex numbers according to: B|A = N exp (2 iS/h) . Doing so creates quantum theory. Singapore. The quantity S is called the action. World Scientiﬁc. (2) A The Feynman sum Eq. the phases of the Feynman amplitudes change sign. All of quantum mechanics follows. so that time reversal of a transition amplitude is equivalent to taking its complex conjugate: B|A = ∗ A|B . The number h is called Planck’s constant.37 quantum theory gives probabilities for the result of each measurement of a system. (1) is carried over all distinct paths between A and B. construct a complex number. 37 The assumption that systems have a deﬁnite state of existence between interactions would follow from having only a ‘single’ path dominate the Feynman sum over paths. (1) paths The factor N will be ﬁxed by a ‘normalization’ condition. We can take this amplitude to be unity and thereby ﬁx the magnitude of the normalization constant N . see H. 2009. and Financial Markets. (3) B B This relation makes it possible to interpret the magnitude square of the Feynman amplitude as a probability for a given transition. such as the Poincar´e symmetry of Relativity. and often can be expressed as the particle’s kinetic energy minus its potential energy. deﬁned by a time-integration from A to B of a function L: B S= L dt . In quantum theory. These probabilities are not simply the result of statistics applied to events. Kleinert. (2). Then B A|B B|A gives the amplitude for the state A to explore all possible alternatives but then return to itself. Polymer Physics. That’s it. The Lagrangian is presumed known. . called the Lagrangian. Let B range over all possible states into which A may evolve. Helping to strongly limit the possible Lagrangians is the imposition of the symmetries we observe. By reversing the order of the time limits in the action integral Eq. We will then have A|B B|A = | B|A |2 = 1. Statistics. THE ESSENCE OF QUANTUM THEORY FOR COMPUTERS 51 For each particle that was initially observed at A and later observed at B. Each exponential term in the sum has a phase given by 2 S/h. introduced shortly.

So far. But John von Neumann showed that quantum theory cannot be trivially subsumed into a bigger deterministic theory. after the appropriate choice of L. Nature Communications. say C1 and C2 . there will be interference of the amplitudes constructed to pass through C1 with those constructed to pass through C2 . Note that the summation of unit complex numbers with wildly diﬀerent phases will tend to cancel (think of adding unit vectors in a plane with arbitrary angles between them). This is the famous ‘Principle of Least Action’. This observation applied to the Feynman path sum shows how to take the classical limit. See J. The classical limit. (4) C Quantum amplitudes contain a linear superposition of possible intermediate states. 5. then quantum ﬁeld theory follows. with the ﬁeld amplitudes as the particle displacements. 411–416. 5. and spacecraft. 2 (2011). PARKE one of these events is determined by interactions with another system such as by measurement.3. from which Newton’s laws and Maxwell’s electrodynamics can be derived. . greatly disturbed Einstein. see R.39 When compared with quantum theory. Colbeck and R.52 WILLIAM C. Mathematical Foundations of Quantum Mechanics. all careful observations are consistent with quantum theory. pp. even ones that Einstein called ‘spooky action at a distance’. After all.2.38 If one takes ﬁeld quantities as a set of equivalent particle oscillators in each inﬁnitesimal volume of space. No extension of quantum theory can have improved predictive power. like baseballs. From the observation that the action satisﬁes SBA = SAC + SCB . Princeton University Press. (1) that B|A = B|C C |A . Newton’s theory predicts natural processes quite well for massive slowly moving bodies. in that those paths causing the least change in the action S relative to the size of h contribute the most to the probability. Classical physics includes only those paths between two events that minimize S. Maxwell’s electrodynamics. Chapter 4. 39 That non-relativistic quantum mechanics has Newtonian theory as a limit is an example of the ‘correspondence limit’ which we impose on any new theory in order to sustain the veriﬁed predictions of earlier observations. moons. A ‘classical computer’ is a dedicated physical system which transforms a prepared initial state into a desired output state by applying the equivalent of Boolean logic in one or more steps between input and output. Superposition. it follows from Eq. Renner. while a collection of such complex numbers with almost the same phase tend to add coherently. This interference can be completely 38 The fact that certain predictions of quantum theory have intrinsic probabilistic character and that the possible realizable states of a system retain strange correlations over arbitrarily long distances between particles. and statistics applied to Newtonian systems with a large number of particles are together in a realm called ‘Classical Physics’. If the allowed Feynman paths from A to B are restricted to only those that pass through two small intermediate regions. Newtonian theory for particles. 1955. vol. von Neumann. For more recent work.

We see that ∗ (x. (The ‘dagger’ here performs a transpose-complex-conjugate operation. This makes the resultant probabilities an intrinsic property of the theory.) From the Feynman path sums. (x. and has no explanation in classical particle theory. Quantum theory lets us calculate these new kinds of probabilities. (3). (5) ∗ From Eq. Individual particles are always found localized. t)|A(x0 . we never observe particles as waves. Dirac recognized that wave functions may be considered a projection of the ‘state of the system’ described by a vector denoted | onto a speciﬁc state (‘eigenstate’) of position: (x. . THE ESSENCE OF QUANTUM THEORY FOR COMPUTERS 53 destructive. the operation U must be unitary: U † U = 1. Rather. t) dx = 1. even for a single particle.) The Feynman path summation divided into small time steps means we can write U = exp(−i H dt/). t)dx is the probability of ﬁnding the particle within the volume dx. t) = B(x. to keep the total probability of ﬁnding the particle anywhere unity. but possibly inﬁnite dimensional. rather than just complex-conjugation. We never ﬁnd a particle ‘spread out’. The symbol dx in the integral is to be interpreted as the volume element in space. New. t) = x| (t) . (6) where. a (The sum over ‘a’ may be given continuous regions as an approximation to discrete sums which are dense in those regions. Each ‘quantum state’ | can be considered a vector in a Hilbert space. and not just due to ignorance of states in a more deterministic theory. t0 ) |(t0 ) .40 Superposition allows us to expand the quantum state into a complete set of basis states: | = |a a| . to include cases in which is taken to have components. The constant = h/(2 ).4. t0 ) . so that repeated searches for a particle at B that was launched from A come up practically empty. Wave functions and quantum states. 5. Addition of amplitudes allows for interference eﬀects. you might say. but isn’t the particle a wave? No. This eﬀect is observed. t)(x. The Feynman transition amplitude for a particle to leave any earlier location A with coordinates x0 at time t0 and arrive at B having the location x at time t is called the wave function for that particle over the spatial coordinates x at the time t: (x. a formulation for probabilities unheard of before the second decade of the 1900s. the state of a system evolves in time according to a linear transformation |(t) = U (t. called the Hamiltonian. Yes. the probability of ﬁnding a particular particle somewhere can be spread out over space. (The sign in the exponent is conventional. because these probabilities are found by ﬁrst adding complex amplitudes.) The operator H . t)(x. satisﬁes the ‘Hermiticity 40 Essentially a vector space with lengths and angles deﬁned.

pp. Particles in relativistic quantum theory. vol. and the ’s must satisfy + = 2g I . and the combined pair corresponding to the electron carrying positive or negative energy. Dirac. so that the wave equation took the form42 3 (i∂ − (e/c)A ) = mc . which formed the basis of the dynamics of quantum ¨ theory originated by Schrodinger. 610–624 & pp. Einstein. The Quantum Theory of the Electron. and the ﬁelds A as the electromagnetic vector potentials due to other charges. the Dirac equation very accurately describes electrons in the ﬁeld of other charges. (7) This is a wave equation. Our present quantum theory incorporates Einstein’s Special Theory of Relativity. in principle. A. The possibility that fundamental particles can be created and destroyed is included into quantum theory by taking the particle wave functions and interacting ﬁelds as quantum ﬁelds. M. The components of electron wave function can be decomposed into two pairs.54 WILLIAM C. t) = H(x. vol. all of chemistry and molecular biology. so that g i∂ i∂ = p p = m 2 c 2 . A. 5. 42 P. 17 (1905). If we add the assumption of reﬂection symmetry.41 P. (8) =0 When the ﬁelds A vanish. This becomes the 41 A. pp. t) . Annalen der Physik. H is a generator of time translations. Taking the ’s to be dimension four. The time-and-space-reversed wave function for a particle describes the forward progression of a corresponding antiparticle. As an indication of the profound reach gained by merging quantum theory and Relativity. a prediction before anyone dreamed of the concept. and that antimatter must exist. For small shifts in time. each pair corresponding to the two possible intrinsic spin directions measurable. In the language of Lie groups. PARKE condition’ H = H † . and therefore atomic structure and. the ’s are square matrices with even dimension at least four. there are plane wave solutions ∝ exp (−ip x )/). Dirac was able to show that the electron spin and its magnetic moment followed from relativistic quantum theory. . 891–921. M. We ﬁnd that if disturbed. wrote the Hamiltonian as a linear operator in the generators of space translation. particle pairs can even ‘bubble’ out of empty space. 351–361. entering into the action S with their own dynamics. recog- nizing that Relativity requires that physical laws be expressible with space and time on an equal footing. A 117 & A 118 (1928). Dirac. satisﬁes a linear equation: i∂t (x. Zur Elektrodynamik bewegter K¨ orper.5. Part I & II. Proceedings of the Royal Society of London.

The characteristic properties of particles following from relativistic quantum theory were ﬁrst described by Eugene Wigner in On unitary representations of the inhomogeneous Lorentz group. ‘green’. there is one or more corresponding charges. It is conventional to use the label s to characterize the particle spin.45 The numerical strength of a particle’s interactions with other particles is always associated with an intrinsic property called its ‘charge’. called their ‘helicities’. as in “The electron has spin 1/2”. 44 Our observations of the sky together with General Relativistic cosmology seem not to allow more than three generations. photons in lasers). leptons have no direct strong interactions. 5. and Veltman. from those with integer spin. ) the phase change is −1. 1.43 The quantum ﬁeld in a three-dimensional space and associated with a pair of identical particles will undergo a phase change when those two particles are exchanged: |(1. Quantum ﬁeld theory distinguishes particles with half-odd integer spin. 2. The fundamental particles making up the structure of materials currently appear to be three generations of the doublet electron-neutrino. Glashow. and gravitational forces. If the total charge of a closed physical system is preserved during a sequence of 43 Particles must have quantized spin with length s(s + 1) and projection along some measurement axis of . 40 (1939). Each generation of quark comes in one of three distinct varieties according to their ‘color charge’. The bound state of a ‘red’. All the observed interactions of one particle with another can be categorized by the so-called strong. parity transformation. For each category of interaction. Weinberg. 3/2. . and ‘blue’ quark and any other ‘color-neutral’ combination of an odd number of quarks generates a ‘baryon’. while no phase change occurs if the particles are bosons (s = 0. 2) = (−1)2s |(2. weak. . including ‘mesons’ coming from bound color-neutral quark-anitiquark systems. In the Standard Model. Higgs. called ‘bosons’. . ’t Hooft. called ‘fermions’. 45 The electromagnetic and weak interactions were linked. ). either along their momentum. The family of electrons and neutrinos are called ‘leptons’. 149–204. This is the ‘Pauli Exclusion Principle’. A zoo of more ﬂeeting particles exist. . all strongly interacting particles. or in the opposite direction. Interactions in quantum theory. 1. and −s ≤ ≤ s. .g. referring to the operations of charge conjugation. Annals of Mathematics. principally by the work of Salam.44 and three generations of a doublet of quarks. such as the familiar proton and neutron. . are called ‘hadrons’. from 1964 to 1975. and time reversal. pp.6. Any number of bosons of the same type can be in the same quantum state (e. where s is either a half or whole integer. This means no two fermions of the same type (such as electrons in atoms) can occupy the same quantum state. If the particles are fermions (s = 1/2. no. THE ESSENCE OF QUANTUM THEORY FOR COMPUTERS 55 ‘CPT Theorem’ in quantum ﬁeld theory. all fermions. vol. . electromagnetic. The large family of baryons and mesons. Particles that move at the speed of c have only two projections of their spin. 1).

In quantum theory.46 The existence of conserved and localizable charges means one can always deﬁne an interaction ﬁeld that has those charges as its source. while the gauge ﬁeld {A } is called the electromagnetic vector potential. The choice of the ‘gauge function’ Λ(x) is open. vol. See P. In the case of electromagnetic interactions. we say the charge has been ‘conserved’. all charges are quantized.e. Dirac. We say each particle with a charge of some kind ‘creates’ an interaction ﬁeld in the space around it. Sitzungsberichte der K¨oniglich Preussischen Akademie der Wissenschaften zu Berlin (1918). with 0123 = 1. v } represents the charge density and currentdensity for a set of charges. 60–71. ∂κ F κ = (4 /c)j and ∂κ F ∗κ = 0. its components are invariant under a proper Poincar´e transformation. A constant shift has no observable eﬀect. the interaction ﬁeld is {F } with components that are the electric and magnetic ﬁeld. A. Gauge ﬁelds are not uniquely determined. using the following argument: If {j } = {c. they come from a countable set. then the local conservation of the total charge. a second kind of gauge transformation occurs when the phase of particle wave functions are shifted. and that ﬁeld acts on other particles having the same kind of charge. Like g . satisfying ∂κ F κ ∝ j . In nature. their interference is observable in the associated particle probability. How particles react to other charges requires knowledge of the dynamics for those particles. Theories whose predictions are independent of the choice of gauge have ‘gauge symmetry’. A 133 (1931).48 Conventional theory describes particle interactions by introducing inter- action ﬁelds which ‘mediate’ the eﬀect of one charge on another.47 If no such dual charge exists in a region of space. F ∗ ≡ (1/2) κ F κ deﬁnes a ‘dual’ conserved charge with current j ∗ ∝ ∂ F ∗ . But making a shift in phase which depends on location will introduce a relative phase between wave components. PARKE interactions within that system. Maxwell’s equations. i. If those component waves converge. can be read from ∂ j = 0. The ﬁeld {A } is called the ‘gauge ﬁeld’ going with the corresponding charge. Proceedings of the Royal Society of London. 46 Dirac showed that if magnetic monopoles exist. 47 The { κ } is the completely antisymmetric tensor in four dimensions. 48 The use of ‘gauge symmetry’ was introduced by Herman Weyl in his consideration of theories with invariance in the scale of length.. But this implies the existence of an ‘interaction ﬁeld’ {F }. and there is no observable local magnetic charge. M. provided ∂ Λdx vanishes for all closed loops in regions where the gauge ﬁeld acts. called the Levi-Civita symbol. (H. An associated ﬁeld. Weyl. then the ﬁeld {F } can be expressed in terms of a vector ﬁeld {A } by F = ∂ A − ∂ A . 465–480. antisymmetric in its indices.) . Dynamics is incorporated into quantum theory. but may be transformed into new ﬁelds {A } which have the same interaction ﬁeld {F } by adding a gradient: A = A + ∂ Λ. then express two conditions: Electric charge is conserved locally. pp. Gravitation und Elektrizit¨at. Q ≡ d 3 x.56 WILLIAM C. Now. pp. Quantised singularities in the electromagnetic ﬁeld. then electric charge must be quantized.

THE ESSENCE OF QUANTUM THEORY FOR COMPUTERS 57 if. Invariante Variationsprobleme. acting on a quantum state for a particle. As the gauge ﬁelds have their own dynamics. and the velocity of the center- of-energy. one can make the combined shifts cancel. along with the phase shift. These quanta are necessarily bosons. Wiss. total angular momentum. lead to conservation laws. These kind of ‘non-abelian’ gauge ﬁelds were introduced by C. This technique for introducing interactions is referred to as the ‘minimal coupling principle’. In this interpretation. the quark color charge follows from an SU3 gauge symmetry. a shift in the derivatives of the wave function occurs. pp. 96 (1954). . In fact. K¨ onig. Mills (Conservation of isotopic spin and isotopic gauge invariance. That means that the interactions between material particles occur only by the exchange of quanta.49 A marvelous theorem was derived by Emmy Noether. Consider the free-electron Dirac equation i∂ = mc. In the case of the Poincar´e symmetry. Yang and R. Noether. These ﬁelds are the gauge ﬁelds described above. Klasse (1918). d. Math. For electromagnetic interactions. d. no. When the gauge ﬁelds themselves are taken to be operators on the internal components of a quantum state. zu G¨ottingen. N. 235–257. Gauge symmetry makes the corresponding charges conserved. We arrive at the full Dirac equation (8). Nachr. Physical Review. the gauge-ﬁeld quantum 49 Gauge symmetry in quantum theory can be re-expressed in terms of the action of ‘covariant’ derivatives D = ∂ + i(e/(c))A . This is the property built into ‘gauge symmetric quantum theories’. castable into a form that requires only knowledge of the ﬁelds of other particles in a small local space- time neighborhood of the aﬀected particle. such as the Poincar´e group and the gauge transformations. vol. pp. 191–195) and are used in the Standard Model to describe interactions between fundamental particles grouped into families. The free-particle wave equation becomes (i∂ − (e/c)∂ Λ) = mc . Symmetry under constant phase shifts of a lepton or baryon wave function makes lepton and baryon number conservation.50 who showed that symmetries of our theories based on continuous groups of transformations. An important example is the symmetry under time translation: If experiments done now with a given system have the same set of results as those done at any time later. all the interactions among fundamental particles have been found to follow from theories which satisfy gauge symmetry! Another property of our current dynamical theories can be called the principle of quasi-local interactions: The known interactions of one particle with another can be described by ‘quasi-local’ eﬀects. the gauge group elements may not commute. For example.-phys. Gesellsch. the conserved quantities are total energy-momentum. 50 E. the interactions arise from the behavior of quantum states by parallel transport across space. quantum theory requires that gauge ﬁelds be quantized. 1. Gauge symmetry can be enforced by adding to the derivative term a gauge ﬁeld A which undergoes a gauge transformation of the ﬁrst kind: A = A + ∂ Λ. A gauge transformation of the second kind on the wave function can be expressed as (x) = exp −i(e/(c))Λ(x) (x). then the system’s energy is ﬁxed.

are all stable at suﬃciently low energy arriving from the outside. since. the quanta of the ﬁeld are called gluons. 5. These stable states are changed only by the input of energies larger than typically available. through their self-interaction.e. can form ﬂux tubes between quarks. PARKE is the photon. i. Some interactions. (9) i If the phases between two or more components of the quantum state are related. Quantum interference between various possible outcomes of a measurement requires some coherence in a quantum state. 52 Such stability is a pure quantum eﬀect. Prepared states and measurement. Gluons also have unit spin but they carry various ‘color’ charges. if the energy states in a bound subsystem were not quantized. Helping the eﬀort are quantum states with unusual stability. and topologically constrained subsystems. For example. By having charge. the gluon ﬁelds. . or even a trapped atom. for which we may only know probabilities for the system to be in given quantum states. making their dynamics more complicated than for photons. The states |φ1 and |φ2 might be two possible interfering states of a single electron. nuclei. and must move slower than the universal speed c. and then stimulating or allowing the system to approach a desired initial state. Isolating a system and determining its initial conﬁguration are often daunting tasks. one with no change in probability densities for its particles. isolating the system from unwanted interactions. must have mass. as contrasted to a ‘mixed quantum state’. if these were observed. so they are eﬀectively isolated until such energies enter the system. Subsystems such as bound electrons. gluons can directly interact with themselves. A quantum system is prepared by ﬁrst selecting a physical system. and therefore.51 For the strong interactions between quarks. such as those from stray ﬁelds or background radiation. any small energy could excite the subsystem. bound atoms. The two states of the atom 51 A particle with charge will carry energy associated with the ﬁeld of that charge. if it can be separated from other particles. has unit spin.58 WILLIAM C.52 After isolation. Each possible quantum state of a system is referred to as a ‘pure quantum state’. and carries no electric charge. may be diﬃcult or impossible to eliminate. The photon at present appears to travel at the maximum speed in Relativity. A pure quantum state made from a superposition of component states is called a ‘coherent quantum state’ when all the phases between its various component states are known to be ﬁxed. a system will evolve by quantum dynamics following a unitary transformation. determining its initial conﬁguration.7. Consider the expansion of a pure quantum state into component states which together span the system’s Hilbert space: | = αi |φi . and may eventually become a ‘steady state’. then these components are said to be in ‘coherence’. The state of most macrosystems will be practically impossible to completely specify.

Browne’s article Physicists put atom in 2 places at once. however. t)A(x. This is a process of ‘decoherence’. For one particle. Interference will arise from terms for which φi | A |φj are not zero for i = j. then within some bounded region its wave function becomes a ‘plane’ wave φ ∝ exp ((ipx x − iEt)/). Boulder. (9). See C. Then the probability distribution of ﬁnding the atom at a speciﬁc location within the trap shows an interference pattern. and then hitting the phosphor. so that the wave function for each state can oscillate back and forth across the trap. Science. the wave function determines the probability density for position across space. Colorado. The average position over all space is predicted to be | x |. those spots produced by electrons passing through two slits in a screen. Those states |a satisfying A |a = a |a 53 This game was played using a Beryllium atom by Dr. . and some may be lost to the environment. published in the New York Times. t) dx . Starting with a set of identically prepared systems in a coherent state rep- resented by Eq. A ‘Schr¨odinger cat’ superposition of an atom. During the measurement.. 1996).53 This quantum eﬀect. However. vol.54 The operators A may also act on the spin and other components of the wave function. some or all of the components of residual quantum states for that system may be left with no well-deﬁned phase relation- ships. is no diﬀerent in principle than that seen as an interference pattern made by the bright spots of light on the surface of a phosphor plate. Some members of the press mis-represented the observation as indicating that one atom can be found in two places at once. THE ESSENCE OF QUANTUM THEORY FOR COMPUTERS 59 might have opposite motions. W. see M. 1996. For example. Monroe et al. pp. One of the important measurements of a system locates the position of particles. pˆ x )(x. Christopher Monroe and colleagues at the National Institute of Standards and Technology. if a quantum system interacts with another system or with a measuring device. so the distribution of measured positions is predicted to be an approximation of (x)∗ (x)dx. 54 The proportionality constant is ﬁxed by noting that if a free particle is left unobserved. May 28. we ﬁnd a distribution of positions. one electron at a time. More generally. (10) wherein pˆ x is taken proportional to the space translation operator in accord with Noether’s Theorem. 1131–1136. 272 (May 24. After a number of such measurements in each small region dx of space. each distinct measurement of a property of the system can be associated with a Hermitian operator A that acts on wave functions for the system as follows: The average value of A will be A = † (x. for which pˆ x φ = −i∂x φ. information is transferred between the system and the measuring device. measurements of the observable A will have an average ∗ ij αi αj φi | A |φj . The order of non-commuting operators in A must be determined by physical arguments.

i. The measurement of A has ‘collapsed’ the quantum state to |a ∝ Pa |. say a. The eﬀect of measurement can be represented by a ‘projection operation’: Pa ≡ |a a|. Then a subsequent measurement of the observable A will have an average value of A = k pk k | A |k . .e. This means that after measurement of A for a number of identically prepared systems. then one of the possible entangled two-electron states can be written √ |0 = (1/ 2)(|01 |12 − |11 |02 ). ˆ In this way. i. For Hermitian operators A. The collapse evidently does not preserve unitarity for the system. The measurement process has involved another interacting system which reduced the system’s available states in a subsequent measurement. but rather has an uncertainty which increases as the phase of the electromagnetic wave becomes more deﬁnite.60 WILLIAM C. which happens to have a total spin of zero. expressed by |(t) = U (t. Entangled states. if we let the quantum state | κ represent the electron labeled by κ and having a spin projection along the z-axis of ( − 1/2) . by deﬁnition. deﬁned by ˆ ≡ k pk |k k |. After a single measurement of the observable A for a system in a pure quantum state | that has A as one of its observables.e. i. ˆ 5. so that A = Tr(A). one of the eigenvalues of A.e. especially likely if the measuring device is macroscopic. and the system will be left in the state |a.8. the observed values will be distributed around the average with a ‘width’ of ΔA. the eigenvalues a will be real numbers. and ‘complete’. 55 There is a special caution for quantum states describing photons. PARKE are called the ‘eigenstates’ of A and a an ‘eigenvalue’. The measured values of A will have an uncertainty deﬁned by ΔA ≡ (A − A)2 . Quantum unitarity applies to isolated systems. the system may be left in a ‘mixed’ state. in that the number of photons is not ﬁxed. and therefore each is a value which may result from a measurement. The interaction (called the ‘coupling’) between two systems during a measure- ment may cause one or more of the phase diﬀerences between the components in the resulting quantum states to become indeterminate.55 For example. This expression can be usefully re-written in terms of a ‘density operator’. they span thespace of possible states. expressible as a |a a| = I by reading from | = a |a a |. for which only the probabilities pk for any particular pure quantum state |k are known. has two or more particles in a quantum state which cannot be factorized into states for each particle. After the measurement. the choice of the mixed state is left implicit. will be found. A† = A. a |a = a a . unless the system was already in an eigenstate of A. t0 ) |(t0 ). A pure state can then be simply characterized by ˆ2 = . The elements of the set {|a} for distinct values a will be ‘orthogonal’. An ‘entangled quantum system’.

is always less than or equal to the universal limiting speed. then this becomes a delayed-choice experiment as introduced by John Wheeler in Mathematical Foundations of Quantum Theory. Next. while the electrons are in ﬂight. 57 Information transmitted by a wave disturbance that started at a certain time cannot be transferred faster than the outgoing wavefront from that disturbance. should be surprised! Even so. then your other friend must ﬁnd a queen even before hearing from your ﬁrst friend. Now send one of the envelopes to one friend. The diﬀerence in phase of that wave when followed around a closed loop is given by e A dx /(c). There are interactions predicted by quantum theory without classical explanation. Take the case when a pair of electrons is prepared in a zero total spin state along a z-axis expressed above. THE ESSENCE OF QUANTUM THEORY FOR COMPUTERS 61 The outcome of a measurement of one of the electrons in an entangled pair will be correlated with the outcome of measurement of the other. R. This is the 56 If the decision on how a component of a system is measured comes after that system has had suﬃcient time to cause interference between quantum alternatives for that component. vol. If one friend opens your envelope and ﬁnds a jack.56 If this ﬁrst observer ﬁnds an electron aligned along her new axis. Yakir Aharonov and David Bohm58 showed that a single electron wave which never enters a region of electric or magnetic ﬁeld could never-the-less have a measurable shift in the probability of ﬁnding that electron after an electric or magnetic ﬁeld changes in the excluded region. when the electron passes on either side but does not enter a tube where a magnetic ﬁeld is conﬁned. those who say so are likely to have been tripped up by picturing each unobserved electron as being localized between observations! 5. pp. the ﬁrst distant observer makes a ‘delayed-choice experiment’. Physical Review. edited by A. i. This kind of correlation also occurs under classical conditions. will ﬁnd the other electron aligned along the negative direction of the new axis constructed by the ﬁrst observer. c. In Special Relativity. The eﬀect occurs. 58 Y. there is a twist in the quantum world. even when they are far apart. . However.e. 115 (1959).57 This suggestion of faster-than-light signaling is a misinterpretation of quantum theory. The result does not mean that the pair interacted after traveling apart. Bohm. then the second observer. far away. where {A } is the electromagnetic potential. also called the signal speed. and then the electrons are allowed to move far apart. Signiﬁcance of electromagnetic potentials in the quantum theory. Academic Press (1978). Marlow. the second envelope to a second friend. have one of the distant observers rotate her electron-spin measuring apparatus away from the z-axis direction to an angle of her choosing.9. this is the way nature acts. Rather. on ﬁrst hearing and with our classical thinking. and such information transfer has not been seen. Non-classical interactions. 485–491. nor was there ‘superluminal’ transmission of information. There is no such restriction on the group velocity or the phase velocity of the wave. Suppose you put a jack in one envelope and a queen in another. Aharonov and D. for example. Now we. the speed of the wavefront.

60 W. This is called the uncertainty principle of Heisenberg. making the probability † dx time independent. B] ≡ (AB − BA) vanishes. 299 (1982). A 92 (1982).61 A localized and isolated physical system will have denumerable (‘quantized’) possible values for its measurable energies and momenta. The wave function for a particle conﬁned to a ﬁxed region of space and initially localized to a much smaller part of that region and then left with no external interaction will diﬀuse outward in space as time progresses. Periodicities of the wave function also enforce quantization if there is a simply-connected closed path over which the corresponding particle can move.59 The shift in the observed interference pattern produced by the electrons when the magnetic ﬂux is turned on has no explanation in classical physics. If the measurement process transfers complete information about a system. See S. Annals of Physics. D. 59 Mandelstam re-expressed the local interaction with {A } as a non-local eﬀect of the electric and magnetic ﬁelds. 1–24. 802–803. Mandelstam. 61 Steady wave functions necessarily have a sinusoidal time dependence through a factor of the form exp (−it). If a copy of a quantum state could be made. For example. 19 (1962). vol. Dieks. Communication by EPR devices.e. If these two measurements are repeated for identically prepared systems. pp. a topological eﬀect of ﬁelds over space-time. If the ‘commutator’ [A. 6. Physics Letters. vol. i. vol. pp. Zurek. then we could defeat the interfering eﬀect of measurement by ﬁrst making a copy. This eﬀect leads to the ‘no-cloning theorem’. no. Suppose two observables A and B for a given system in the state | are measured in a certain time order. Quantum electrodynamics without potentials. . A single quantum cannot be cloned. then the observables A and B may be measured ‘simultaneously’. periodicity in the azimuthal angle in the wave function makes the measured values of the projection of the orbital angular momentum along a measurement axis become denumerable. Nature. 271–272. without the measurement of one aﬀecting the results of measuring the second.62 WILLIAM C. and then measuring the copy. In general. and its quantum wave function will be steady.e.60 the statement that a general quantum system for which we have no prior knowledge cannot be identically copied. Measurement of a system may disturb the system. leaving the original system undisturbed. and eventually the probability for ﬁnding the particle in any small location will have no measurable change in time. i. one can show that the uncertainties satisfy ΔA ΔB ≥ (1/2) |AB − BA|. The wave for an unobserved particle will spread over the entire allowed region. PARKE ﬂux of magnetic ﬁeld somewhere inside the loop. The state of a physical system can be labeled by a set of measured values for a maximal set of mutually commuting observables that are also conserved over time. a change in the order of measurement may change the probability for ﬁnding a given value for the second observable. that system will no longer contain entangled states. pp.

vol. Boltzmann. Quantum computation. The spontaneous ﬂow of information. pp. 76 (1877). Mechanical computers. the multiplicity W . §6. 21 (1982). Modern classical computers use bistable systems to store information. it will spend most of its time in those conﬁgurations which have many ways to be constructed. Quantum theory for complex systems. as there is an intimate connection between entropy. vol. without moving particles between quantum states. i.63 These ideas incorporate the ﬁrst and second law of thermodynamics. Interactions from the outside can change the total energy. International Journal of Theoretical Physics. Feynman65 considered the possibility that we might take advantage of quantum systems to perform computations quicker than so-called classical computers. Simulating physics with computers.10. pp. and/or by changing the energies {i } of the quantum states. where i is the energy of the corresponding quantum state. Sitzungsberichte der Akademie der Wissenschaften zu Wien. 65 R. Uber¨ die Beziehung zwischen dem zweiten Hauptsatze der mechanischen W¨armetheorie und der Wahrscheinlichkeitsrechnung respektive den S¨atzen u¨ ber das W¨armegleich- gewicht. decoherence. E = i ni i . then that system of the two which has the smaller variation in its information content as its total energy changes will tend to spontaneously transfer information into the second system. simply because as the system evolves through various conﬁgurations. then ni ∝ exp (−i /(kT )). putting heat into a system is a ‘disordering process’. results from statistical likelihood. 66 These discrete-level computers are often referred to as ‘digital’. 373– 435. and heat from memory loss. THE ESSENCE OF QUANTUM THEORY FOR COMPUTERS 63 5. wave function collapse. each with ﬁxed volumes. and molecular . and T is the temperature of the system. non-forced ﬂow. In terms of information. 63 If particles remain in their quantum states. 64 If a system near thermal equilibrium is held at ﬁxed volume and a small amount of energy dE is put in. the second law of thermodynamics implies that if two systems interact. under the physical constraints. Work involves changing the particle energies by changing the volume of the system. then the ratio dE/dI turns out to be proportional to the temperature of that system. of a system either by changing the ‘occupation numbers’ {ni }. in contrast to ‘analog’ computers that use internal signals that are assumed to vary smoothly with time. After a relaxation time for a system containing a large number of interacting particles. By increasing the multiplicity of the system.66 For some problems involving numbers with 62 L. information. 467–488. and logical gates to perform Boolean operations on sets of ones and zeros. which work by the movement and interaction of shaped objects.e. The ﬁrst kind of change is heat transfer and the second is work transfer. Feynman. the most likely distribution of particles in the available quantized energy states will be those that tend to maximize. causing an increase in its information content by dI .64 These are important concepts for quantum computers. no heat is transferred. and the process is called ‘adiabatic’. One can then show62 that if the number of particles ni in each quantum state labeled by the index i is also large.

and N is the average noise power. V. no. An alternative parameterization takes α = cos (/2) and = exp (iφ) sin (/2). Communication in the presence of noise. PARKE n digits and that may require solution times that rise exponentially with n when performed on computers using only Boolean logic. N -particle states in a quantum computer can be made by constructing these quantum states from a linear combination of the states for each of the N particles. 37 (January 1949)... Below are some of the special consequences of quantum theory for quantum computers and communications: The simplest system for the storage of information gives only two possible values by a measurement. 10–21. iN with |i | = 1 . where B is the bandwidth (in cycles per second). Proc.. digital computers can be more tolerant of a small amount of noise than analog devices. . 1.1} N 2 2 = i |i1 i2 i3 . (11) i=1 i In the second line of the equation. A quantum system can be constructed that has only these two values for the outcome of a measurement. Two-valued qubit states are easily realized in nature: The electron spin has only two possible projection values ± 1/2. and the photon has only two possible helicity values ± 1. As it is always possible to expand an arbitrary quantum state into a basis set for that state’s Hilbert space. that work through molecular interactions and transformations. the product base state is represented in a shortened form. . i. These days. These values can be taken as 0 or 1. in which the order of the 0’s and 1’s corresponds to computers. Shannon and Hartley showed that the maximum number of bits per second that can be transmitted from one storage location to another is given by B log2 (1 + S/N ).64 WILLIAM C. . a binary coding. . The phrase digital computer. |iN N {ik =0. C. where α and are complex numbers satisfying 2 2 |α| + || = 1. the changes are made in systems which can ﬂip between oﬀ and on in a speciﬁed clock time. can now mean any device which manipulates information by discrete changes. E.e. Transmission of information. Taking these particles to have only two internal quantum states. This state is called a ‘qubit’. S is the average signal power. L. referring to counts base ten. the state of the computer is expressible by |N = i1 i2 i3 . the computation on a quantum computer may take times that rise no faster than a power of n. in which case the states are called |0 and |1. . Hartley. the possible qubit states can be pictured as points on a unit sphere (called the ‘Bloch sphere’) with |0 at the north pole and |1 at the south. Bell System Technical Journal (July 1928). Classically. Institute of Radio Engineers. pp.iN |i1 1 |i2 2 |i3 3 . such a system stores one bit of information. By using such switching to encode information. but whose quantum state is a linear combination of the two possible outcome states |0 and |1: |q = α |0 + |1 . vol. See R. Evidently. Shannon. are a mixed breed.

Quantum computation takes advantage of entanglement within those states. there are ‘universal’ sets of simple gates that can be used to build arbitrarily close representations of a general unitary transformation.e. To be a non-classical computer. i. US might only be approximated by a ﬁnite sequence of quantum gates.) Quantum gates acting on a single qubit can all be represented by a general unitary transformation U which is an arbitrary rotation in Hilbert space: U = exp (i nˆ · . where the set {Ugi } are called ‘quantum gates’. making it also that state which has maximum information content. (Arbitrarily 2 close here means that if VS is the approximation. (11) will have all |i | = 1/N . . In the classical case. a generalization of classical logic gates. with k = 0 labeling the initial state. quantum computers must operate on the input infor- mation stored in quantum states by unitary transformations. where k labels a particular intermediate state. . at least in probabilistic terms. then || (US − VS ) || is a number that can be made arbitrarily small for all | by increasing the number of universal gates used in VS . In the following. such as US . states with equal probability for all possible conﬁgurations of its component particles. all Boolean operations on a set of bits can be performed by a combination of NAND gates. To sustain coherence. Since a general unitary transformation will contain continuous parameters.67 It follows that a useful initial state of a quantum computer has at least a subset of particles prepared in one of the maximally entangled states. In the quantum case. It is possible that US can be decomposedinto a ﬁnite product of simpler or more universal unitary operations: US = i Ugi . at least some the intermediate states must be entangled. The same is true of NOR gates. and i = i1 i2 i3 . leaving (2N − 1) free relative 69 phases between the basis states. it harbors entanglement. iN is a binary number constructed from the i’s. This makes NAND gates universal for classical computing.68 The maximally entangled states made from 2 qubits as in Eq. THE ESSENCE OF QUANTUM THEORY FOR COMPUTERS 65 the labeling of each of the separate qubits. the substage of a quantum computation holding the intermediate state of a calculation will be called |(k). For a given quantum computer. Each term in the product acts on the state |(k) left by the previous operation labeled by k and produces |(k + 1). If the quantum state |N cannot be factorized. a solution to a solvable problem is a unitary transformation US that carries the input quantum state |(0) encoding the required initial data into an output quantum state that carries the information about the solution.

/2) = cos (/2) I + i nˆ · .

Linden. Physical and Engineering Sciences. 2036. Proceedings of the Royal Society A: Mathematical. 2011–2032. On the role of entanglement in quantum computational speed-up. 67 See.R. Jozsa and N. 68 One learns most when the outcomes are least predictable! 69 These maximally entanglement states are also called ‘generalized Bell states’. no. 459 (2002). for example. pp. vol. sin (/2). .

66 WILLIAM C. PARKE where is an angle of rotation around an axis ﬁxed by the direction n. ˆ and the {.

0 1 0 −i 1 0 . i = 1. 3} are the Pauli matrices.i . 2.

x = . .

.y = .

H. Physical Review Letters. vol. and no wave function ‘collapse’ should occur. a 2N dimensional unitary transformation would be implemented to carry out one step of a computation. A collapse of a quantum state from a superposition of substates to one such substate violates unitarity. and there will be both a ‘coherence time’ and a ‘coherence length’ over which the system retains a semblance of its coherence. pp. and that the transfer is cannot be superluminal. so that a number of components of the wave function may have their phase become stochastically indeterminate. Such a projection into a proper subspace is irreversible and non-unitary. measurement of an observable is the equivalent of projecting out a subspace of the initial state. quantum coher- ence may be degraded or lost. should evolve by a unitary transformation. . There is no easy way out of this paradox. 70 Transferring qubits across space was ﬁrst described by C. The observation of the state of a particle in a multi-particle entangled state removes the entanglement of that particle. Formally. because the initial qubit is destroyed in the process. Note that transferring a qubit from one system to another does not violate the no-cloning theorem. . so that the larger system that contains the observed system and the measuring devices. and/or error correcting schemes. If noise or other spurious interactions occur in the system. 70 (1993). the act of measurement causes the wave function for the system to ‘collapse’. Bennett et al. iN . For an initial state 1 0 consisting of the many qubits (perhaps realized by many particles capable of being in two distinct quantum states). In quantum theory. in Teleporting an unknown quantum state via dual classical and Einstein-Podolsky-Rosen channels. This produces a paradox: The measuring instrument is also a physical system. and therefore is outside the formalism of quantum theory. as two classical bits must be sent from the ﬁrst system to the second before reconstruction of the qubit can take place.70 A new measurement acting on a quantum state generally causes some decoherence. all processes preserve the condition that the probability of ﬁnding any of the possible states of the system add to unity. left unobserved. Systems for transferring qubits over long distances require long coherence lengths. A fault-tolerant quantum computer uses states that have long coherence times. quantum states evolve by a unitary transformation. 1895–1899. In the ‘Copenhagen’ view.z = 1 0 i 0 0 −1 0 1 which act on the base states |0 = and |1 = . As we have seen. such as |i1 i2 i3 . . quantum entangled states with long life times. The resulting state of the system no longer holds information about the projected state.

Being unitary for the combined system of the observed and the measuring device. then attempts to intercept that particle will degrade or destroy the entanglement. but rather actually does lose information over time. the group of such exchanges is non-abelian. vol. For example. . Reversal of the weak measurement of a quantum state in a superconducting phase qubit.72 It is also possible that nature does not just scatter information so much that we cannot easily put systems such as broken eggs back together again. Katz et al. If a set of identical particles are restricted to a two-dimensional surface. the quantum state representing two particles may gain a phase factor of exp(2i p) when the two particles are exchanged. 200401. p. making the observed state highly probable. but rather redistribute the amplitudes for various quantum states. 72 This diﬃculty is related to the ergotic hypothesis in classical mechanics. even the measurement of a ﬁnal state after a computer calculation is a reversible process for the computer. the particles are called ‘anyons’. the measurement device and the surrounding interacting systems. Topological structures have been shown to be important in quantum theory.. and the other possible states in the observed system left with very small amplitudes. 73 By contrast. some information is practically lost. This opens the possibility of ‘absolute’ security in transmission lines. Physical Review Letters.71 But for a multitude of interactions. the continuity condition for the wave function describing particles 71 See. But if the information transferred by erasing a quantum memory state produces heat in the environment. there is a space transformation that will ‘untangle’ the pair. N. such a measurement process is. or the space is not simply connected. The same ideas apply to quantum computers. THE ESSENCE OF QUANTUM THEORY FOR COMPUTERS 67 The measuring devices in the larger combined system must introduce inter- actions that do not project out quantum substates in the combined system. The entanglement of an observed system with a measurement instrument and subsequent restoration of the original quantum state has been demonstrated for simple systems and measuring devices with highly restricted interactions. in principle. and therefore will be detectable. restoration after interactions is typically unfeasible with our current resources. For three particles. in a connected region of three dimensional space. In principle. for example.73 If the phase factor p is not n/2 (where n is integer). This consideration may be important in the construction of quantum computers through the storing of information in the topological braiding of non-abelian anyons as they progress in space-time. 101 (2008). and make p an integer multiple of 1/2. In quantum theory. ‘reversible’. where p need not be integer or half-integer. if the order in which the particles are exchanged produces a diﬀerent wave function phase. If one of the particles in an entangled state is sent to a second observer as a form of communication. and the development of entropy concept in statistical thermodynamics. This possibility is outside the realm of quantum theory. particularly since macroscopically long coherence lengths have been realized with laser beams. no information is lost.

For example. §7. Uncontrollable interactions both within and from the outside a quantum computer will tend to collapse coherent states. which are predicted by Einstein’s General Theory of Relativity to occur when the density of an object of mass m exceeds about 3c 6 /(25 G 3 m 2 ). making the system more ‘fault immune’ against the eﬀect of noise and other external interactions. Overall. particle charges come from topological structures. Gates and memory elements smaller than nanoscale will suﬀer quantum ﬂuctuations. Evidently.2.1. Some mixing and degradation in quantum states can be tolerated by using repeated calculations and/or Implement error corrections which can reconstruct. we see no fundamental limitation. In some models. Strong gravitational ﬁelds exist near black holes. Explorations to ﬁnd strategies which minimize the limitations are ongoing. A variety of promising systems for quantum computers take advantage of the diﬃculty of breaking topological structures in order to preserve quantum coherence. Taking systems at the nanoscale and ﬁnding technology that minimizes heat production toward the Szil´ard value of kTln2 per bit lost gives an upper limit to computer density made from materials. with growing uncertainties in bit structures and Boolean transformations as the size of the elements are reduced. Even our DNA code can be mutated by quantum tunneling. each quantum gate must act within the shortest coherence time. the quantum states being used to store information typically have ﬁnite lifetimes through spontaneous decay. Limits to computing. The techniques to control heat buildup also require volume in the ancillary heat sinks or channels for radiative cooling.68 WILLIAM C. Classical com- puters have practical limitations in density. Practical limits to computing and information storage. If the system has a certain level of noise. Cosmological limits. After suﬃcient time. Memory and gates based on information in light beams have corresponding limits due to pulse duration and wave length uncertainties. Coherence can be maintained for some period of time by using quantum states which have some intrinsic stability and suﬀer little debilitating interactions with adjacent systems or with the environment. unless our ambitions reach across the cosmos. classical correction schemes can eliminate errors. 7. resulting in the collapse of the employed coherent states. Quantum computers require coherence within the involved quantum states of the computer during computation. PARKE adds signiﬁcance to global space-time topology. even though we can anticipate severe practical diﬃculties to building a quantum computer which can outperform its classical cousin. a degraded quantum state. with some assurance. Working against us are physical limitations. Such black . at a cost of size. coupling to the environment will cause decoherence and disentanglement within a quantum system. 7.

General Relativity limits the density of a computer. where k is Boltzmann’s constant. 77 J. 1 (1916). D. 189–196. Black hole explosions?. with a mass loss rate inversely proportional to the square of the hole mass m dm/dt = −c 4 /(3 · 5 · 210 G 2 m 2 ) . One resolution of this paradox is to have the object’s information transferred to a region close to the horizon of the black hole. and inversely proportional to the square of Planck’s length. even of any space-time structure. its density eventually forces the computer to collapse into a black hole. . 5443. which is also the information storage capacity. Even before Hawking proposed that black holes evaporate. 76 It is even possible that the volume surrounded by a black hole horizon is completely empty. THE ESSENCE OF QUANTUM THEORY FOR COMPUTERS 69 holes got their name because no form of radiation can escape from the hole if it starts out within a region around the hole bounded by a surface called the ‘horizon’. this surface has the ‘Schwarzschild radius’74 RS ≡ 2Gm/c 2 . 2333–2346. 248 (1976). 75 S. This leads to the idea that the limiting density of information storage may be eﬀectively two dimensional. Uber ¨ das Gravitationsfeld eines Massenpunktes nach der Einsteinschen Theorie. to be consistent with quantum theory. with the object destined to disappear into the black hole. while the negative energy particles fall into the black hole. Thus. For a non-spinning hole without charge. Hawking showed75 that the ﬂuctuations in particle ﬁelds near but outside the horizon of a black hole can produce particle pairs with some of the positive energy particles having suﬃcient kinetic energy to reach large distances away. with any infalling matter ending up just outside the horizon. and we suspect all large galaxies do. and concurrently the density of information storage. Hawking radiation can carry the stored information back out (so the radiation is not perfectly thermal). quantum theory requires that black holes evaporate. 30–31. Bekenstein. Sitzungsberichte der K¨oniglich Preussischen Akademie der Wissenschaften. vol. Astronomers have found stellar-mass black holes in binary systems by analyzing the orbits of companion stars. Jacob Berkenstein77 conjectured that the entropy of a black hole. vol. Black holes and entropy. no. Some (as yet untested) theories even have the information of the whole Universe reﬂected by a kind of holographic image in one less dimension. The ﬂux of photons emitted is close to that of a hot body at a temperature inversely proportional to m T = c 3 /(8 kGm) . Using quantum theory. with no other interaction but gravity. 4 RS2 . a system initially containing an object and a black hole. Nearby large galaxies are known to contain one or more super-massive black holes at their center. W. 74 K. pp. 7 (1973). Hawking. As a computer becomes larger in a given volume. vol. Physical Review D. Nature.76 In this way. Schwarzschild. with each bit stored in a Planck-size area. pp. cannot lose information: The quantum state of the hole and the object evolves unitarily. is proportional to the area of the hole’s horizon. pp. Hawking then calculated the proportionality constant to be k/4. However.

we should recognize that our physi- cal theories are always tentative. Quantum theory describes these operations based on how nature processes informa- tion. D. DEPARTMENT OF PHYSICS COLUMBIAN SCHOOL OF ARTS & SCIENCES THE GEORGE WASHINGTON UNIVERSITY WASHINGTON. it cannot predict both itself and the Universe. there are a variety of precautions to which we should be attentive when applying and interpreting the theory. Correspondingly. Conclusions. Quantum computers take advantage of quantum opera- tions in physical systems in order to solve well-posed problems. In addition. Even though very successful.70 WILLIAM C. with their own language. PARKE A cosmological limitation on computation also comes from the fact that we appear to live in a ﬁnite Universe. which is needed to unambiguously project the Universe’s future. although there are propositions that connect the very small to the very large.edu . quantum theory makes some rather non-intuitive and thought-provoking predictions. Reﬂecting on the underlying ideas central to quantum theory should help us in the exploration of possibilities for future quantum computers. While we formulate how nature handles information. C. since each theory has a variety of equivalent formulations. and active participants in both information transfer and information storage. Our current theories do not incorporate these kinds of limitations. being that the computer is within the Universe. Any smaller computer cannot hold the data of the Universe at one time. our main focus should be on the predictions of a theory. §8. A computer can be no larger than the Universe itself. Also. 20052 E-mail: wparke@gwu. Each covers a limited realm and has a limited accuracy. Space and time are important primitives in quantum theory.

June 2010. we use the concept of the ﬁber product of measures to prove general versions of two determinization results about hidden-variable models. Andrei Savochkin. for example. a hidden-variable model (an empirical model augmented by unobserved variables). and to the Stern School of Business for ﬁnancial support. Logic and Algebraic Structures in Quantum Computing Edited by J. We know that the password is either p2s4w6r8 or 1a3s5o7d. The two outcomes—when Alice types a password and Bob types a password—are perfectly correlated. The correlation is purely informational: It is our state of knowledge that changes. Chubb.v. editors for very important feedback. we deﬁne the concepts of an empirical model (a probabilistic model describing measurements and outcomes). Speciﬁcally. Clearly. Harizanov Lecture Notes in Logic. Dagstuhl. when Alice types a password on her machine. Gus Stuart. quantum experiments that involve spin measurements at arbitrary relative angles. V¨axjo. we immediately know what will happen when Bob types in one or other of the two passwords. Introduction. Hidden variables are extra variables added to the model of an experiment to explain correlations in the outcomes. Barbara Rifkind. §1. Z. and an extra r. we prove that: (i) every empirical model can be realized by a deterministic hidden-variable model. With a view to quantum foundations. our framework is general enough to include. (random variable) X for Alice’s password. there is a realization- equivalent hidden-variable model satisfying determinism and -independence. Within this framework. not Bob’s computer. Eskandarian and V. JEROME KEISLER Abstract. Z takes the value z1 or z2 according as the two machines were prepared with the ﬁrst or the second We are grateful to Samson Abramsky. Y for Bob’s password.v. Axelle Ferri`ere. Amanda Friedenberg. Thus. Tobias Fritz. but we do not know which it is. and Noson Yanofsky for valuable conversations. we can consider an r. Here is a simple example. Bob Coecke.v. for the case of inﬁnite measurement spaces and ﬁnite outcome spaces. 45 c 2016. A. Linnaeus University. Association for Symbolic Logic 71 . to John Asker. and participants at the conference on Advances in Quantum ¨ June 2010. participants at the workshop on Semantics of Information. Elliot Lipnowski. for helpful input. Formally. (ii) for every hidden-variable model satisfying locality and -independence. and various properties of hidden-variable models. this somehow causes Bob’s machine to acquire the same password. it would be wrong to conclude that. to a referee and the volume Theory. FIBER PRODUCTS OF MEASURES AND QUANTUM FOUNDATIONS ADAM BRANDENBURGER AND H. Alice’s and Bob’s computers have been prepared with the same password. If Alice now types in p2s4w6r8 and this unlocks her computer.v. The r. an r.

and by specifying. a probability measure over pairs of outcomes. 1964]. But.v. question one can ask: (i) The existence question. according to what properties we ask of the model. and Rosen [13. Podolosky. where q satisﬁes other stipulated properties and also realizes e? . which says that the choices of measurement by Alice and Bob are independent of the process determining the values of any h. 1984]. 1935]. Then.v.” We will use the term “free variables” below. 1986]).) model is obtained by starting with the empirical model and then appending to it an extra r. 1964]).v. conditional on the value of Z.’s. We imagine that Alice can make one of several measurements on her part of a certain system. even though X and Y will be perfectly correlated. In this sense. for each pair of measurements. with probability measure p that satisﬁes certain properties and realizes e.’s. Suppose we are given an empirical probability measure e on the observable variables of a system. Each pair of measurements (one by Alice and one by Bob) leads to a pair of outcomes (one for Alice and one for Bob). 1985. The correlations that arise in QM—for example. Can we ﬁnd another h. Shimony [20.v. Let us specify a little more what we mean by an experiment. there are much more complicated examples of hidden-variable analysis. and Bob can make one of several measurements on his part of the system. Z explains the correlation. One property is locality (Bell [3. whose watershed no-go theorem gave conditions under which the answer is negative. Can we ﬁnd an extended space that includes h. even in the classical realm.v. they will also be independent (trivially so). with probability measure q. JEROME KEISLER password. where p satisﬁes certain properties (as above) and realizes (via marginalization) the empirical probability measure e? (ii) The equivalence question. p.v. and a probability measure p on this space. model. Having started with von Neumann [23.v. Another property is -independence (the term is due to Dickson [12. We can deﬁne various types of h. the most famous context for hidden-variable analysis is quantum mechanics (QM). Suppose we are given a certain physical system and an empirical probability measure e on the observable variables of the system. model.v. 95] describes this as the condition that “the settings of instruments are in some sense free variables. We can build an empirical model of the experiment by choosing appropriate spaces for the sets of possible measurements and outcomes.72 ADAM BRANDENBURGER AND H. model. 2005]). in spin measurements—cannot be explained as reﬂecting the presence of hidden variables. Here are two basic types of h. Of course. 1932] and Einstein. the question of whether a hidden-variable formulation of QM is possible was re-ignited by Bell [3. which can be decomposed into parameter independence and outcome independence (Jarrett [18. the extra r.v. An associated hidden-variable (henceforth h. and an h. Bell [5.

The goal of this chapter is to prove the two determinization results at a general measure-theoretic level (Theorems 5. these two results tell us a lot about Bell’s Theorem. model satisfying locality and -independence. Bell’s Theorem asks for an h. Every empirical model (whether generated by a classical or quantum or even superquantum system) can be realized by an h.v. It is also true that for every empirical model. model satisfying determinism. Fine [14. 1984]. Given an h. In light of the second determinization result. 1971] mentioned the idea of the ﬁrst determinization result. model with determinism is possible. Fiber products of .’s determine ‘non-probabilistically’ (formally: almost surely) the outcome of any measurement.3). Bell [4. These positive results involve yet another property of h. (This is a trivial construction. FIBER PRODUCTS OF MEASURES AND QUANTUM FOUNDATIONS 73 Bell’s Theorem is the most famous negative answer to (i). The ﬁber product generalizes independence in probability theory. A notable aspect of our formulation is that we allow for inﬁnite measurement spaces. but it is not possible to believe both.2 and 5. Put together. 1999] in graph theory.) As usually stated.. determinism implies locality.g. Bell’s Theorem teaches us that: It is possible to believe that Nature (in the form of QM) is deterministic. and has in turn been generalized in several directions in the literature (e.v. The construction of these objects comes from Shortt [21. an h.1. experiments that involve spin measurements at arbitrary relative angles.v. As we will see in Section 4. an h.v. 1982] produced the ﬁrst version of the second determinization result. for example. Thus. 2009] and Ben Yaacov and Keisler [7. Both results have been (re-)proved for various formulations in the literature. which we note in Remark 5. or it is possible to believe that measurement choices by experimenters are free variables. We assume that outcome sets are ﬁnite (such as spin up or spin down). and Flori and Fritz [15. In this chapter we will focus on positive results for questions of both types (i) and (ii). 2013] in category theory. The ﬁrst determinization result says that for every empirical model. model satisfying locality and -independence. (ii) Second determinization result. model that satisﬁes determinism and -independence.v. Our treatment uses the concept of the ﬁber product of measures. Thus. see Adler [2. obtained when the physical system is quantum and the properties demanded are locality and -independence. model with -independence is possible.v. there is a realization-equivalent h. We consider the following positive results on questions (i) and (ii): (i) First determinization result. Bell’s Theorem can be equivalently stated as asking for determinism and -independence. Dawid and Studen´y [11. our set-up is general enough to include. 2009] in model theory. the h.v. models: The (strong) determinism property says that for each player.v.

space. We ﬁrst introduce notation and recall some well-known facts about product measures.) model is a probability measure p on Ω.v.v. the property we ask for is conditional independence—which we would only expect once the extra r. we will restrict attention to bipartite systems. §3. Billingsley [8. are (realization-)equivalent if they realize the same empirical model. Thus.2. which is a measurable space (Xb . We say that an h. space. An empirical model is a probability measure e on Ψ. and y and the joint outcome x = (xa . X ) ⊗ (Y. Deﬁnition 2. Y) = (Ya .v. The interest in h. Throughout. Write (X. Y) of two measurable spaces (X. Recall that by a product (X. Bob has a space of possible measurements. Yb ). Ω = (X. We see that an empirical model describes an experiment in which the pair of measurements y = (ya . Yb ). X ) and (Y. Ya ) ⊗ (Yb .74 ADAM BRANDENBURGER AND H. We will come to other properties in Section 4. model is an empirical model which has an extra component. possibly with diﬀerent h. spaces. X ) ⊗ (Y.v. We say that two h. models is that we can ask them to satisfy properties that it would be unreasonable to demand of an empirical model. JEROME KEISLER measures turn out to be well suited to the questions in quantum foundations which we study in this chapter. (We will comment later on the extension to more than two parts.v.v. the h. A hidden-variable (h. L).v. L). X ) ⊗ (Y. (Y. which is a measurable space (Ya . X ) = (Xa . Y) is meant the (Cartesian) product space X × Y equipped with the .v. see e. Alice has a space of possible measurements. An h. Y) ⊗ (Λ.g. Deﬁnition 2. and which reproduces a given empirical model when we average over the values of the h. Xa ). Xa ) ⊗ (Xb ..) There is also an h. Xb ).1. Xb ). which is a measurable space (Yb . which is a measurable space (Xa . Deﬁnition 2. 1995]. which is an unspeciﬁed measurable space (Λ. xb ) ∈ X are distributed according to e.v. Z is introduced. yb ) ∈ Y is randomly chosen according to the probability measure margY e. Ya ). Empirical and hidden-variable models. Y). in the example we began with. model p realizes an empirical model e if e = margΨ p. Ψ = (X. and a space of possible outcomes. Products and ﬁber products of measures. Likewise. For background on the relevant measure theory. viz.v. §2. and a space of possible outcomes. models.3.

where J ∈ X and .-algebra generated by the measurable rectangles J × K .

and z ∈ Z. X ) ⊗ (Y. First. we refer to the concept of conditional probability given a sub . then for each J ∈ X we write p(J ) = p(J × Y ) = q(J ). X )⊗(Y. we write p[J ||Z] for the conditional probability of J given z. Second. and for each q-integrable f : X → R we write f(x) dp = f(x) dp = f(x) dq. Y) and q = margX p. We use the following two conventions. Z). when p is a probability measure on a product space (X. Here. a statement holds for p-almost all x ∈ X if and only if it holds for q-almost all x ∈ X . J J ×Y J Thus. J ∈ X . Y)⊗ (Z. in particular. when p is a probability measure on (X. FIBER PRODUCTS OF MEASURES AND QUANTUM FOUNDATIONS 75 K ∈ Y.

∅} is the trivial .-algebra. p[J ||Z] denotes a function from Z into [0. ∅} ⊗ Z](x. 1995.z) . 1] such that p[J ||Z]z = p[J × Y × Z|{X × Y. see Billingsley [8. Section 33] for a presentation.y. Formally. (Note that {X × Y.

-algebra over X × Y .) We use similar notation for (ﬁnite) products with factors to the left of (X. then q[J ||Z] = p[J ||Z]. Z). so that the right-hand side does not depend on (x. We will also need the concept of conditional expectation given a sub . X ) or to the right of (Z. y). Note that if q = margX ×Z p.

given an integrable function f : X → R. Let f and g be two such functions and let L = {z : f(z) < g(z)}. ∅} ⊗ Z](x.1. Suppose p(J ) > 0. Using the deﬁnition of p[J ||Z].y. where we write for the projection from X × Y × Z to X . f(z) dp = p(J × L). Thus. ∅} ⊗ Z] dp = L X ×Y ×L 1J ×Y ×Z dp = p((X × Y × L) ∩ (J × Y × Z)) = p(J × L). Uniqueness: If p(J ) = 0. and we will use an analogous notation. Lemma 3. 1995. Section 34]). . Existence: Let f(z) = p[J ||Z]z . The mapping z → p[J ||Z]z is the p-almost surely unique Z- measurable function f : Z → [0. we see that f(z) dp = E[1J ×Y ×Z |{X × Y. then f(z) = g(z) = 0 p-almost surely. 1] such that for each set L ∈ Z. X ×Y ×L as required.z) .-algebra (Billingsley [8. L Proof. we deﬁne E[f||Z] by: E[f||Z]z = E[f ◦ |{X × Y. and z ∈ Z.

we get.3. Lemma 3. Similarly. Corollary 3. for each J ∈ X . Then.z) dp = p(J × Y × L1 ) = p(J × L1 ) = p(L1 ). We say that two probability measures p and q on (X. 1} p-almost surely. Y). as required. Proof. q) and (Y. p[J ||Y ⊗ Z](y. Y) and r on (Y. by the properties of probability measures. Given probability measures p on (X. X ) ⊗ (Y. L L L a contradiction. X . so p(L) = 0 and hence f(z) ≥ g(z) p-almost surely. Y) agree on Y if margY p = margY q.2. Let L0 = {z ∈ Z : p[J ||Z]z = 0} and L1 = {z ∈ Z : p[J ||Z]z = 1}.76 ADAM BRANDENBURGER AND H. then p[J ||Y ⊗ Z] = p[J ||Z] p-almost surely. and 0< g(z) dp − f(z) dp = (g(z) − f(z)) dp = 0. If p(J × L) > 0. Y) such that q and r are independent with respect to p. Y ×L0 so p[J ||Y ⊗ Z](y. L1 ∈ Z and p(L0 ∪ L1 ) = 1. X ) ⊗ (Y.z) = 1 = p[J ||Z]z ∀ (y. For the particular of ﬁnite X . Therefore p(J × L) = 0. we say that p is an extension of r if r = margY p.z) dp = p(J × Y × L0 ) = p(J × L0 ) = 0.1. X ) ⊗ (Y. r). z) ∈ Y × L1 . Y ×L1 so p[J ||Y ⊗ Z](y.z) = 0 = p[J ||Z]z ∀ (y. Y. z) ∈ Y × L0 .1 again. p(J × K) = q(J ) × r(K) . so f(z) = g(z) p-almost surely. When x ∈ X . that case x∈X p[x||Z]z = 1 p-almost surely. If p[J ||Z] ∈ {0. the product measure p = q ⊗ r is the unique probability measure p on (X. g(z) ≥ f(z) p-almost surely. L1 By Lemma 3. then p(L) > 0. L 0 p[J ||Z]z dp = p(L1 ) = p(J × L1 ). By Lemma 3. we write p[x||Z]z = p[{x}||Z]z . Let q be the marginal of p on X × Z. Similarly. we have p[J ||Z] = q[J ||Z] q-almost surely. Then L0 . p[J ||Y ⊗ Z](y. Given probability spaces (X. that is. JEROME KEISLER Then L ∈ Z. p[J ||Z]z dp = 0 = p(J × L0 ).

X ) ⊗ (Y. Let (X. X . r) be as above and let p be a common extension of q and r on (X.4. Y). (ii) The . Remark 3. FIBER PRODUCTS OF MEASURES AND QUANTUM FOUNDATIONS 77 for all J ∈ X and K ∈ Y. Note that p is a common extension of q and r. Y. The following are equivalent: (i) p = q ⊗ r. q) and (Y.

Then the following are equivalent: (i) p = q ⊗Z r. Z) be measurable spaces. then it is unique. K ∈ Y.6. for all J ∈ X and K ∈ Y. it is enough to show that p[J ||Z] dp = p(J × K × L). Let q and r be probability measures on X ⊗ Z and Y ⊗ Z.5. X ). Consider any J ∈ X . It is clear that (i). that is. the ﬁber product q ⊗Z r is the common extension of q and r with respect to which q and r are as independent as possible given that they have the same marginal on Z. K ∈ Y. (ii) p[J × K||Z]z = q[J ||Z]z × r[K||Z]z p-almost surely. To prove (iv). K×L . (iii) p[J ||Y]y = p(J ) p-almost surely for all J ∈ X . Y = (Y. We next introduce the notion of a ﬁber product of measures.5.-algebras X ⊗ {Y. We say that a probability measure p on X ⊗ Y ⊗ Z is a ﬁber product of q and r over Z. 1999]). Let q and r be as in Deﬁnition 3. if p(J × K × L) = q[J ||Z]z × r[K ||Z]z ds L for all J ∈ X . There are examples where a ﬁber product does not exist (see Swart [22. ∅} and {X. Deﬁnition 3. and L ∈ Z. Proof. for all J ∈ X and K ∈ Y. ∅} ⊗ Y are independent with respect to p. p(J × K ) = p(J ) × p(K ) for all J ∈ X and K ∈ Y. and let p be a common extension of q. (ii). Next is a characterization of the ﬁber product in terms of conditional probabilities and extensions. r on X ⊗ Y ⊗ Z. 1996] and Dawid and Studen´y [11. Z = (Z. Assume (i). Y). respectively. (iii) p[J × K||Z]z = p[J ||Z]z × p[K||Z]z p-almost surely. Lemma 3. and L ∈ Z. Assume that q and r have the same marginal s on Z. and (iii) are equivalent. (iv) p[J ||Y ⊗ Z](y. Intuitively.z) = p[J ||Z]z p-almost surely. for all J ∈ X . But it is easily seen that if a ﬁber product q ⊗Z r does exist. in symbols p = q ⊗Z r. For the remainder of this section we let X = (X.

1996]: Lemma 3. Therefore p[J ||Z] × 1K dp = p[J ||Z] × p[K ||Z] dp Y ×L L = q[J ||Z] × r[K ||Z] dp. Y ×L As in the preceding paragraph.5. A version g(J. If the space X is ﬁnite.7. It is well known that when X and Z are both Polish spaces. then q[J ||Z]z has a regular version. this is equal to p(J × K × L). It is also easily seen that when X is ﬁnite and Z is any measurable space. K ×L Y ×L By the rules of conditional expectations. E[p[J ||Z] × 1K ||Z] = p[J ||Z] × E[1K ||Z] = p[J ||Z] × p[K ||Z]. then the ﬁber product q ⊗Z r exists. We can now formulate the various properties of h. Then p(J × K × L) = p[J ||Y ⊗ Z] dp = p[J ||Z] dp K×L K ×L = p[J ||Z] × 1K dp. The next lemma is from Swart [22. which shows that (i) implies (iv). p[J ||Z] × 1K dp = q[J ||Z] × r[K ||Z] dp.78 ADAM BRANDENBURGER AND H. then the ﬁber product q ⊗Z r exists. L By (i). then q[J ||Z]z has a regular version. models which we listed in the Introduction (we will not repeat their sources) and establish some relationships among them. At this point. Now assume (iv). Properties of hidden-variable models. z0 ) is a probability measure on X for each ﬁxed z0 ∈ Z. Y ×L L and condition (i) is proved. Let q and r be as in Deﬁnition 3. Let q and r be as in Deﬁnition 3. JEROME KEISLER We have p[J ||Z] dp = p[J ||Z] × 1K dp.8. and Xa and Xb are the respective power sets. z) of the conditional probability q[J ||Z]z is regular if g(·.5. .v. §4. If q[J ||Z]z has a regular version. This is the case we will need in this chapter. we adopt: Assumption: The outcome spaces Xa and Xb are ﬁnite. Corollary 3.

model p satisﬁes locality if for every x ∈ X we have p[x||Y ⊗ L] = p[xa ||Ya ⊗ L] × p[xb ||Yb ⊗ L]. The h. pb = qb ⊗Yb ×Λ r if and only if p[xb ||Y ⊗ L] = p[xb ||Yb ⊗ L] for all xb ∈ Xb .” Fix an h. p satisﬁes parameter independence if and only if pa = qa ⊗Ya ×Λ r and pb = qb ⊗Yb ×Λ r. qb = margXb ×Yb ×Λ p. By the term “measure” we will always mean “probability measure.v. By Lemma 3. . Proof. model p satisﬁes parameter independence if for every xa ∈ Xa we have p[xa ||Y ⊗ L] = p[xa ||Ya ⊗ L]. whenever we write an equation involving conditional probabilities.2. pY = margY p. this holds if and only if p[xa ||Y ⊗ L] = p[xa ||Ya ⊗ L] for all xa ∈ Xa .v. Similarly.3. Deﬁnition 4. model p. pΛ = margΛ p. with a and b interchanged. We will often make use of the following notation: pa = margXa ×Y ×Λ p.v. Here is a characterization of parameter independence in terms of ﬁber products. pa = qa ⊗Ya ×Λ r if and only if pa [xa ||Y ⊗ L] = pa [xa ||Ya ⊗ L] for all xa ∈ Xa . pb = margXb ×Y ×Λ p. The result follows. FIBER PRODUCTS OF MEASURES AND QUANTUM FOUNDATIONS 79 Also. Deﬁnition 4.6. r = margY ×Λ p.1. Corollary 4. p pa qa r pY pΛ Xa Xb Xa Xa Ya Yb Ya Yb Ya Ya Yb Ya Yb Λ Λ Λ Λ Λ All expressions below which are given for Alice have counterparts for Bob. qa = margXa ×Ya ×Λ p. it will be understood to mean that the equation holds p-almost surely. The h. Since p is an extension of pa .

Proof. model p satisﬁes outcome independence if for every x = (xa . then p satisﬁes locality.6. p. Proof. p satisﬁes locality if and only if it satisﬁes parameter independence and outcome independence. p[xb ||Y ⊗ L] = p[xb ||Yb ⊗ L].6. It is easily seen from the deﬁnitions that if p satisﬁes parameter independence and outcome independence. The following corollary characterizes outcome independence in terms of ﬁber products. p satisﬁes outcome independence if and only if p = pa ⊗Y ×Λ pb . Suppose that p satisﬁes locality. It follows that p satisﬁes parameter independence. JEROME KEISLER Deﬁnition 4. The next proposition follows Jarrett [18. supposing that p satisﬁes locality. xb ||Y ⊗ L] = p[xa ||Ya ⊗ L] × p[xb ||Yb ⊗ L]. xb ||Y ⊗ L] xb ∈Xb = (p[xa ||Ya ⊗ L] × p[xb ||Yb ⊗ L]) xb ∈Xb = p[xa ||Ya ⊗ L] × p[xb ||Yb ⊗ L] xb ∈Xb = p[xa ||Ya ⊗ L] × 1 = p[xa ||Ya ⊗ L]. xb ∈Xb so p[xa ||Y ⊗ L] = p[{xa } × Xb ||Y ⊗ L] = p[xa . We have {xa } × Xb = {(xa . Again. we have p[xa . We immediately get a characterization of locality in terms of ﬁber products.v. Similarly.4. . Corollary 4. Proposition 4. The h. xb ||Y ⊗ L] = p[xa ||Y ⊗ L] × p[xb ||Y ⊗ L].80 ADAM BRANDENBURGER AND H. so p satisﬁes outcome independence. and hence p[xa . 582]. This follows easily from Lemma 3. xb )}.5. 1984. xb ) ∈ X we have p[x||Y ⊗ L] = p[xa ||Y ⊗ L] × p[xb ||Y ⊗ L].

p satisﬁes locality if and only if p = pa ⊗Y ×Λ pb . FIBER PRODUCTS OF MEASURES AND QUANTUM FOUNDATIONS 81 Corollary 4.v. The following are equivalent: (i) p satisﬁes -independence. model p such that Λ is a singleton satisﬁes -independence. Deﬁnition 4.7. Remark 4.10. pa = qa ⊗Ya ×Λ r. p[L||Y]y = p(L).9. By Remark 3. model p satisﬁes -independence if for every event L ∈ L.v. We observe: (i) The -independence property for p depends only on r.3 and 4. (ii) Any h. we have: Lemma 4.4.5.6 and Corollaries 4. (iii) The . (ii) The measure r is the product pY ⊗ pΛ . pb = qb ⊗Yb ×Λ r.8. By Proposition 4. The h. Proof.

1}..v. L ∈ L. model p satisﬁes weak determinism if for each x ∈ X we have p[x||Y ⊗ L](y. The h. model p satisﬁes strong determinism if for each xa ∈ Xa we have p[xa ||Ya ⊗ L](ya . This says that the set Y × Λ can be partitioned into sets {Ax : x ∈ X } such that p[x||Ax ] = 1 for each x ∈ X .) ∈ {0. i.) ∈ {0.) ∈ {0. 1}.e.v. (ii) For each xa ∈ Xa we have p[xa ||Y ⊗ L](y.13.-algebras Y and L are independent with respect to p. The following are equivalent: (i) p satisﬁes weak determinism. The h. .11. Deﬁnition 4.12. Lemma 4. Strong determinism is the notion discussed in the Introduction. This says that the set Ya × Λ can be partitioned into sets {Axa : xa ∈ Xa } such that p[xa ||Axa ] = 1 for each xa ∈ Xa . 2008]. The distinction between strong and weak determinism in the next two deﬁnitions is from Brandenburger and Yanofsky [10. p(K × L) = p(K) × p(L) for every K ∈ Y. 1}. Deﬁnition 4.

so p satisﬁes parameter independence. Suppose p satisﬁes strong determinism. Proof. Fix xa ∈ Xa . Proposition 4. so p satisﬁes strong determinism. Proposition 4. If p satisﬁes strong determinism then it satisﬁes weak determinism. p[xa ||Y ⊗ L] = p[xa ||Ya ⊗ L]. and hence p[xa ||Y ⊗ L](y. By weak determinism and Lemma 4. It is clear that (ii) implies (i). suppose p satisﬁes weak determinism and parameter independence.3.) = 1 for each xa ∈ Xa . 1}. 1}.) ∈ {0. JEROME KEISLER Proof. . Proof. If p satisﬁes weak determinism then it satisﬁes outcome independence. By Lemma 3.13(ii). For the converse. so p satisﬁes weak determinism by Lemma 4. p satisﬁes strong determinism if and only if it satisﬁes weak determinism and parameter independence.82 ADAM BRANDENBURGER AND H. we have p[xa ||Y ⊗ L] ∈ {0. Suppose p satisﬁes strong determinism. 1}.3. Assume (i).13.14. p[xa ||Y ⊗ L](y. Suppose p satisﬁes weak determinism. 1}. By Lemma 4. we have p[xa ||Ya ⊗ L] = p[xa ||Y ⊗ L] p-almost surely. ) there is an x ∈ X such that p[x||Y ⊗ L](y. Therefore (ii) holds. p[xa ||Ya ⊗ L] = p[xa ||Y ⊗ L].) = 1.13. Proposition 4.) ∈ {0. Then for p-almost all (y. Therefore p[xa ||Ya ⊗ L](y. By Proposition 4. Therefore p[x||Y ⊗ L] = p[xa ||Y ⊗ L] × p[xb ||Y ⊗ L].14. By Lemma 3. By parameter independence. p satisﬁes weak determinism.16. Proof. as required. and therefore p[xa ||Y ⊗ L] ∈ {0.15.

Every empirical model e can be realized by an h.v.v. space of p. Let s = margX e. L. For every probability space (Λ. L. model p.v.1. Given an h. pΛ ). and 4. By Propositions 4. we can take Λ to be a one-element set and take (Λ. space of p is ﬁnite.17. model p where p satisﬁes strong determinism and the h.15.v. Theorem 5. We can summarize the properties we have considered and the relationships among them in the above Venn diagram.v.v.v. Build an h. space (Λ. model p where p satisﬁes -independence and the h. Build a probability measure d on X × Λ so . pΛ ) to be the trivial probability measure. L. Every empirical model e can be realized by an h.6. FIBER PRODUCTS OF MEASURES AND QUANTUM FOUNDATIONS 83 Corollary 4. pΛ ) the h. In particular.2. Proof. space of p has only one element.16. model that realizes e and satisﬁes -independence. p satisﬁes strong determinism if and only if it satisﬁes weak determinism and locality. s) where Λ is a copy of X and L is the power set of X . Determinization theorems. Proof. Remark 5. 4. We now state and prove our determinization results. we call the probability space (Λ. the product measure p = e ⊗pΛ is an h. §5. Proof. L.v.

Given an h. space (Λ. s(x) if x = x . model p satisfying locality and -independence. ).x ) ∈ {0. 1}. For each xa ∈ Xa and x ∈ Λ. and thus r¯ = pY ⊗ p¯ Λ¯ . Since Λ is a copy of the ﬁnite space X . where. . Let r¯ = r ⊗ ua ⊗ ub . ya . 1]a . ya . 1]a is a copy of the real unit interval. 1]a into A consecutive intervals Ia (xa1 . x ) = 0 otherwise. Ia (xa2 . r = pY ⊗ pΛ . ) has length ua (Ia (xa . Ub .v. d (x. Note that d is an extension of s. Λ is ﬁnite. Ia (xa . ub ). Ua is the set of Borel subsets of [0. ¯ p¯ ¯ ) will be the product of (Λ. ya . for each x ∈ X and x ∈ Λ. .) . 1]a . . model p¯ whose h. L. . Since p satisﬁes -independence. Then p is realization-equivalent to e. for each xa ∈ Xa . Let sa be the unique probability measure on (Xa . ). Here. model p¯ that satisﬁes strong determinism and -independence. [0. Theorem 5. ua ) ⊗ ([0. Proof. . Let p be the ﬁber product p = d ⊗X e. For each ya ∈ Ya and ∈ Λ. )) = p[xa ||Ya ⊗ L](ya . Let Xa = {xa1 . 1]b . for each xa we have p[xa ||Ya ⊗ L](ya . .) . Ua ) . JEROME KEISLER that. . there is a realization-equivalent h. 1]a . 1} p-almost surely.v. ). xaA }. partition [0. Suppose p satisﬁes locality and -independence. By Lemma 3. ya . ya . i=1 We carry out the same construction with b in place of a. pΛ ) Λ and the Lebesgue unit square ([0. L) ⊗ ([0. Ua . Xa ) ⊗ (Ya . similarly for b.3.84 ADAM BRANDENBURGER AND H.v. . and ua is Lebesgue measure on Ua . Note that the boundary point between the ith and (i + 1)th intervals is the (Ya ⊗ L)-measurable function n p[xai ||Ya ⊗ L](ya . we have p[xa ||x ] = d [xa ||x ] ∈ {0. new h. Ya ) ⊗ (Λ.3. Ia (xaA . We will construct a ¯ L. This shows that p satisﬁes strong determinism.v.

satisﬁes -independence because r¯ = pY ⊗ p¯ Λ¯ . sa [xa ||Ya ⊗ L ⊗ Ua ] = 1Ia (xa . sb . 1}.) (α) d r. By Lemma 3. Similarly.ya . Now deﬁne p¯ a . ya . 1]a . Since both p and p¯ satisfy locality. It remains to prove that p¯ is an extension of p.) dr Ka ×L = qa ({xa } × Ka × L). ¯ p¯ = p¯ a ⊗Y ×Λ¯ p¯ b . by Fubini’s Theorem we have p({x} ¯ × K × L) = p[x||Y ¯ ⊗ L] ¯ d r¯ K×L×[0. Similarly for sb .1.ya .1]a r. By Fubini’s Theorem. By Lemma 3. .1]a ×[0. and r.) (α) dua dr Ka ×L 0 = ua (Ia (xa . )) dr Ka ×L = qa [xa ||Ya ⊗ L](ya .1]b = sa [xa ||Ya ⊗ L] ¯ × sb [xb ||Yb ⊗ L]¯ dr K×L×[0. 1}.1]b 1 1 = sa [xa ||Ya ⊗ L] ¯ × sb [xb ||Yb ⊗ L] ¯ dua dub dr K×L 0 0 = qa [xa ||Ya ⊗ L] × qb [xb ||Yb ⊗ L] dr K×L = p({x} × K × L). Thus sa is an extension of qa . Ka ∈ Ya . Ka ×L×Ua where we write α for a typical element of [0. we have sa ({xa } × Ka × L × Ua ) = ¯ 1Ia (xa . FIBER PRODUCTS OF MEASURES AND QUANTUM FOUNDATIONS 85 such that for each xa ∈ Xa .) (α) d r¯ Ka ×L×[0.1]a ×[0. model p¯ is a common extension of sa .ya .) ∈ {0. and p¯ as the ﬁber products p¯ a = sa ⊗Ya ×Λ×[0.1]b r. p¯ b . sa ({xa } × Ka × L) = 1Ia (xa . Deﬁne sb in a similar way. ¯ It also We see that the h. sa [xa ||Ya ⊗ L] ¯ ∈ {0. and p¯ extends r¯ = r ⊗ ua ⊗ ub .ya . ¯ p¯ b = sb ⊗Yb ×Λ×[0. L ∈ L. sb is an extension of qb . Therefore p¯ satisﬁes strong determinism.1]a 1 = 1Ia (xa . and Ua ∈ Ua .3.v.

The only adjustment needed is that parameter independence must now be stated in terms of sets of parts instead of individual parts. To keep things simple.86 ADAM BRANDENBURGER AND H.2 and 5. Interestingly. JEROME KEISLER Thus p¯ is an extension of p. outcome independence and locality do not need to be restated. and Theorems 5. This completes the proof. Endnote.2. All the results in Section 4 (“Properties of hidden-variable models”).3 in this section. §6. we assumed in this chapter that the outcome spaces Xa and Xb are ﬁnite.3 holds assuming only that the outcome spaces have countably generated . extend immediately to multipartite systems. Also. 2012] that all of the results in Section 4 hold for arbitrary outcome spaces Xa and Xb . However. the arguments in [9. 2012] can be adapted to show that Theorem 5. the only result in this chapter that requires this assumption is Theorem 5. and hence p¯ is realization-equivalent to p. We show in [9.

and Zeilinger [16. pp. 171–181. Dialectica. Bell. Conﬂuentes Mathematici. vol. available at http://www. 1993] model). 1 (2009). Adler. [8] P. 3rd ed. [4] . [6] . 39 (1985).adambrandenburger. Academic Press.com and http://www. On the Einstein-Podolsky-Rosen Paradox. 113036. vol. [9] A. vol. [2] H. . A canonical hidden-variable space. Speakable and Unspeakable in Quantum Mechanics. The sheaf-theoretic structure of non-locality and contextuality. An exchange on local beables. 29–39].wisc. Keisler. New Journal of Physics.wisc. p. 2011] establish a strict hierarchy of three properties: non-locality (`a la Bell) is strictly weaker than possibilistic non-locality (exhibited by the Hardy [17. Wiley.) Extending these latter properties to the general measure- theoretic setting appears to be an open direction. 1995. also available at http://www.edu/keisler. vol. Horne. For ﬁnite probability spaces. REFERENCES [1] S. Ben Yaacov and H. Keisler. pp. pp. [7] I. which is strictly weaker than strong contextuality (exhibited by the Greenberger. Foundations of Quantum Mechanics. 13 (2011). pp..). 1987. 197–223. It would be of interest to extend the methods in this chapter to formulate other properties that have usually been studied only for the case of ﬁnite sets of measurements.math. 1–21.’ Course IL (New York). J. (2012). Physics. Cambridge University Press. Introduction to the hidden-variable question. 85–96. 1967] is a model-independent proof of strong contextuality. Billingsley. (In this language. Proceedings of the International School of Physics ‘Enrico Fermi. Randomizations of models as metric structures. vol. Probability and measure. [5] . Brandenburger and H. 1987.math. 9 (2009). A Geometric Introduction to Forking and Thorn-Forking. 1 (1964). pp. (Reprinted in [6.-algebras of events Xa and Xb . J. Journal of Mathematical Logic. pp. Brandenburger. the Kochen-Specker Theorem [19. Abramsky and Brandenburger [1. [3] J.edu/keisler. Abramsky and A. 1971. 195–200. 1989] model).

vol. [21] R. [11] A.math. [16] D. Can quantum-mechanical description of physical reality be considered complete?. (Translated as Mathematical Foundations of Quantum Mechanics. 32–40. 48 (1982). Whittaker. Morgan Kaufmann. pp. On the physical signiﬁcance of the locality conditions in the Bell arguments. pp. 169–176. pp. 47 (1935). STERN SCHOOL OF BUSINESS NEW YORK UNIVERSITY NEW YORK. vol.wisc. Quantum Concepts in Space and Time (R. Specker. [20] A. 59–87. Events and processes in the quantum world. Penrose and C. WI 53706 E-mail: keisler@math. Kochen and E. editors). pp. 17 (1967). editor).edu/∼keisler . pp. vol. Physical Review Letters. Studeny. [23] J. Journal of Mathematics and Mechanics. 425302. [13] A. Dawid and M. M. Princeton University Press. available at http://arxiv. von Neumann. pp. FIBER PRODUCTS OF MEASURES AND QUANTUM FOUNDATIONS 87 [10] A. Proceedings of the 7th Workshop (D. 1955. 41 (2008). Shortt. Shimony. M. 28 (1996).edu URL: www. pp. The problem of hidden variables in quantum mechanics. Fritz. M.adambrandenburger. Fundamenta Mathematicae. [19] S. pp. ´ An Alternative Approach to Conditional Independence. Quantum Chance and Non-Locality: Probability and Non-Locality in the Interpretations of Quantum Mechanics. 121 (1984).edu URL: www. Noˆus. Cambridge University Press. 569–589. Quantum Theory and Conceptions of the Universe (M.com DEPARTMENT OF MATHEMATICS UNIVERSITY OF WISCONSIN-MADISON MADISON. vol.nyu. Kluwer. 71 (1993). Zeilinger. 2005. Journal of Physics A: Mathematical and Theoretical. P. Fine. Heckerman and J. Hardy. [22] J. Isham. A classiﬁcation of hidden-variable properties. vol. Nonlocality for two particles without inequalities for almost all entangled states. Podolsky. and A.6548. Yanofsky. Springer-Verlag. Bell’s Theorem.org/ abs/1308. 1665–1668. 182–203. vol. Dickson. 1986. [12] W. Going beyond Bell’s theorem. (2013). Artiﬁcial Intelligence and Statistics 99. [14] A. and N. Greenberger. vol. vol. A. Brandenburger and N. Jarrett. pp.brandenburger@stern. Universally measurable spaces: An invariance theorem and diverse characteriza- tions. Rosen. [15] C. 1989. Physical Review Letters. 777–780.wisc. A conditional product measure theorem. Flori and T. 69–72. Mathematische Grundlagen der Quantenmechanik. Oxford University Press. [17] L. Hidden variables. [18] J. Horne. 131–135. Physical Review. NY 10012 E-mail: adam. Einstein. B. editors).). p. 1999. Statistics & Probability Letters. 291–295. Swart. pp. Compositories and Gleaves. 1932. joint probability and the Bell inequalities. Kafatos. 18 (1984).

52. 47]. no-cloning [1]. The aim of the present paper is to establish strong connections between these two formalisms. Association for Symbolic Logic 88 . 49. and categorical quantum mechanics [6. and to use this perspective to understand the special features and properties which single it out. §1. 41. A central theme in current work in quantum information and quantum foundations is to see quantum mechanics as occupying one point in a space of possible theories. including Hilbert spaces. They also play a ˆ in current work on axiomatizations of quantum mechanics prominent role [36. including quantum protocols [6]. which can be presented in terms of string-diagram representations of structures in monoidal categories [7]. Two formalisms which have been used in this context are operational theories. and the possibilities for alternative theories. entanglement [24]. A. and to use this perspective to understand the special features and properties which single it out. 10]. Two formalisms which have been used in this context are operational theories [48. sets and relations. quantum. 25]. measurement-based quantum computing [29]. Eskandarian and V. The local. and the possibilities for alternative theories. 12. OPERATIONAL THEORIES AND CATEGORICAL QUANTUM MECHANICS SAMSON ABRAMSKY AND CHRIS HEUNEN Abstract. • Operational theories allow general formulations of results in quantum foundations and quantum information [11. We then show how non-locality can be formulated at this level of generality. We show how models of categorical quantum mechanics have representations as operational theories. and study a number of examples from this point of view. and stochastic maps. 19. Introduction. Chubb. This has proved very eﬀective in providing a conceptually illuminating and technically powerful perspective on a range of topics. and non-locality [22]. 45 c 2016. Logic and Algebraic Structures in Quantum Computing Edited by J. Harizanov Lecture Notes in Logic. • Categorical quantum mechanics enables a high-level approach to quantum information and quantum foundations. and no-signalling models are characterized in these terms. and categorical quantum mechanics. 7]. A central theme in current work in quantum information and quantum foundations is to see quantum mechanics as occupying one point in a space of possible theories.

They are seen as linked to an ‘instrumentalist’ or ‘epistemic’ view of physics. We call these operational representations. the operational framework has proved fruitful as a basis for general results. it may be useful to discuss the motivation for studying them.g. quantum. including Hilbert spaces. From our perspective. give rise to operational representations. this is a useful perspective. and stochastic maps. • By focussing on this empirical and observational content. sets and relations. We then review some elements of categorical quantum mechanics. which may prove useful in ﬁnding ‘deeper’ theories. At a stage in the development of physics where the next step is far from clear. operational theories attract criticism on philosophical grounds. Any viable theory must account for this content. We shall assume some familiarity with the linear-algebraic formalism of quantum mechanics. We then show how a proper formulation of compound systems within the operational framework leads to a view of operational theories as representa- tions of monoidal categories of a particular form. and with the ﬁrst notions of category theory. they focus on the empirical content of theories. but would break up the ﬂow of ideas in the main body of the paper. Why operational theories? Before proceeding to a formal description of operational theories. • Indeed. 25]. 19. These are mathematically interesting. 12. As we see it. operational theories have the following attractions: • Firstly. and no-signalling models are characterized in these terms. on the information processing capabilities of theories under various assumptions [11. e. To make the paper reasonably self-contained. 10]. and study a number of examples from this point of view. we include an appendix which reviews the basic deﬁnitions of monoidal categories. 49. the fact that we study operational theories does not indicate any such philosophical . Thus there is a general passage from categorical quantum mechanics to operational theories. We go on to show how non-locality can be formulated at this level of generality. and provides the setting for recent work on axiomatic reconstructions of quantum mechanics [36. equipped with a trace ideal. We shall begin by reviewing operational theories. The local. and show how monoidal dagger categories. OPERATIONAL THEORIES AND CATEGORICAL QUANTUM MECHANICS 89 The aim of the present paper is to establish strong connections between these two formalisms. We also include another appendix which proves a number of technical results on trace ideals. as opposed to a ‘realistic’ approach. and the means by which we can gain knowledge of the microphysical world. §2. functors and natural transformations. On the debit side. operational theories allow meaningful results to be formulated and proved about the ‘space of theories’ as a whole.

An operational theory is formulated in terms of directly accessible ‘operations’. or parts of theories. This means that each measurement has only ﬁnitely many possible outcomes. More generally. We assume there are several diﬀerent types of system. Compound systems. the function vA (p. For each choice of p and m. in a laboratory. 1] which assigns a probability vA (p. For convenience. For each system type A. which will apply to all measurements. where those outcomes outside O have zero probability of occurring. C . This leads to the following additional requirements. putting systems. a compound system type AB. m. these operations should be subject to axioms yielding a coherent mathematical structure on these notions. yields outcome o when measurement m is performed on it. they are pragmatically useful for the reasons already mentioned. • A set TA of transformations which may be performed on systems of type A. A.g. measurements. Moreover. Operational theories formalized. 3. together. An important additional ingredient is to give an account of compound systems. we shall assume a ﬁxed inﬁnite set of outcomes O. B. and can be seen as expressing some irreducible minimum of empirical content. we can consider transformations TA. which is a function vA : PA × MA × O → [0. on A and B to yield corresponding operations on the compound system AB. Each measurement has a set of possible outcomes.e. o) to the event that a system of type A. etc. m. m.B which can be performed on systems of type A to produce systems of type B. We shall use the function dA : PA × MA → D dA (p. The empirical predictions of the theory are given by its evaluation rule. i. which will have to be accounted for by any presumptive ‘deeper’ theory. m) : o → vA (p. • Ways of combining preparations.1. −) deﬁnes a probability distribution on outcomes. etc. §3. Rather. In this paper. possibly space-like separated. we shall only consider ‘ﬁnite-dimensional’ theories. • A set of measurements MA which can be performed on systems of that type. o) where D is the set of probability distributions of ﬁnite support on O. the theory speciﬁes the following: • A set of preparations PA which produce systems of that type. .90 SAMSON ABRAMSKY AND CHRIS HEUNEN commitment. which can be performed e. • For each pair of system types A. Any measurement with a ﬁnite set of outcomes O ⊆ O can be represented using O. B. prepared by p.

Operational representations: concrete description. dA : PA × MA → D). This immediately suggests the notion of Chu space [14. We can regard states operationally as equivalence classes of preparations [51]. and was applied to the modelling of physical systems in [2]. 3. For each system type A. and also indicates why guidance from category theory is helpful in ﬁnding the right structural axioms. There is a natural equivalence relation on preparations: p is equivalent to p . it can be seen as a generalization of the notion of model of a physical system proposed by Mackey in his inﬂuential work on the foundations of quantum mechanics [48]. where m. which shows the naturalness of the ideas.2. 20]. which have been developed extensively as a setting for quantum mechanics and quantum information in the categorical quantum mechanics programme [6. if for all m ∈ MA : dA (p. which is mathematically elegant but a little abstract. That is. we now turn to morphisms. MA .B induces a map f∗ : PA → PB . We shall therefore proceed by giving a precise formulation of operational theories with compound system structure as a certain class of representations of monoidal categories. we shall give a more concrete account. m) = dA (p. We deﬁne m to be equivalent to m . Quotienting an operational system (PA . A transformation in TA. which has received quite extensive development [54]. m ∈ MA . m). we see the essential elements as provided by monoidal categories. Having identiﬁed operational systems with Chu spaces. Indeed. if for all p ∈ PA : dA (p. and then subjecting it to a transformation procedure t resulting in a system of type B. there is an equivalence relation on mea- surements. we can gather the relevant data provided by an operational theory into a single structure (PA . is itself a procedure for preparing a system of type B. to apply it to a state prepared by p ∈ PA . which we call operational representations. In an entirely symmetric fashion. we apply the transformation . Before giving the ‘oﬃcial’ deﬁnition of operational representation. This is exactly the notion of extensional equivalence in Chu spaces [2]. 7]. m ). MA . We can regard observables operationally as equivalence classes of measurements. Such a transformation can also be seen as a procedure for converting measurements of type B into measurements of type A: given a measurement m ∈ MB . dA ) by these equivalences corresponds to the biextensional collapse of a Chu space [2]. preparing a system of type A according to preparation procedure p. OPERATIONAL THEORIES AND CATEGORICAL QUANTUM MECHANICS 91 Rather than trying to develop such ‘meta-operations’ and axioms from ﬁrst principles. where p. p ∈ PA . m) = dA (p .

which is to be speciﬁed by the theory. For all p ∈ PA . m ⊗ m ) = dA (p. MA . there are a number of coherence conditions which are needed to get a mathematically robust notion. which appears in one form or another in the various formulations of operational theories.92 SAMSON ABRAMSKY AND CHRIS HEUNEN t to obtain a preparation of type B. we should see the notion of compound system as an important degree of freedom. (2) This expresses the probabilistic independence of pure tensors. p ). This sub-category will not in general be full. Rather than writing these down in an ad hoc fashion.B : PA × PB → PA⊗B . dB ). Thus we can also associate a map f ∗ : MB → MA to the transformation t. Thus we see that in an entirely natural way. more precisely of Chu(Set. m) · dB (p . The fundamental property which this inclusion must satisfy relates to the evaluation. While Chu spaces have a standard monoidal structure. Conceptually. D) [54]. However. This is given by maps P A. or p with f ∗ (m). The formal relationship that links the two maps f∗ and f ∗ is that. m ∈ MB . pure tensors arise by preparing states or performing measurements indepen- dently on subsystems. (1) ¨ This can be seen as an abstract form of the relationship between the Schrodinger and Heisenberg ‘pictures’ of quantum dynamics. f ∗ (m)). and similarly for measurements. m) = dA (p. we must have: dA⊗B (p ⊗ p . MB . dB ). we shall write p ⊗ p rather than A. Thus given operational systems A = (PA . M A. dA ) and B = (PB . dA ) → (PB . dA⊗B ). f ∗ ) deﬁnes a morphism of Chu spaces (f∗ . whether we measure f∗ (p) with m. p ∈ PB . f ∗ ) : (PA . this does not yet provide an account of compound systems. we shall now turn to a more systematic way of deﬁning the .B : MA × MB → MA⊗B . What general properties should such a notion satisfy? One important requirement. In addition. Rather. we should be able to form a system A ⊗ B = (PA⊗B .B P (p. we can associate an operational theory with a sub-category of Chu spaces. we should not in general expect that operational theories will give rise to monoidal sub-categories of Chu spaces. is to have an inclusion of pure tensors. since not every Chu morphism will arise from a transformation in the theory. m ). and indeed form ∗-autonomous categories [20]. to which m can be applied. MB . we should observe the same probability distribution on outcomes: dB (f∗ (p). m ∈ MA . MA⊗B . For readability. MA . The equation (1) says exactly that the pair of maps (f∗ .

1) in Set.3. The fact that P and M are . an operational representation of C is speciﬁed by the following data: • A symmetric monoidal sub-category Ct of C. This will usually have the same objects as C. in which these conditions arise naturally from standard notions. • Similarly. viewed as a symmetric monoidal category. Let us now unpack this deﬁnition. We take D to be a commutative monoid under pointwise multiplication. where f∗ := P(f). We shall now take a diﬀerent view. ·. which we think of as a process category. which for each type A gives us a set PA . in which the structure of an operational theory arises from a symmetric monoidal category. Note that Ctop is a symmetric monoidal category. • A contravariant symmetric monoidal functor M : Ctop → Set which for each A represents the measurements on A. and only those morphisms which correspond to admissible transformations. injective on objects and faithful. viewed as a type of system. • A dinatural symmetric monoidal transformation ·· d : P × M → KD which gives the evaluation rule of the theory.e. the corresponding set of preparations or states. we have a variable set P. Moreover. i. Note that a constant symmetric monoidal functor valued at a set M is just a commutative monoid (M. Operational representations: functorial formulation. • A symmetric monoidal functor P : Ct → Set which represents. The types of the theory are the objects of Ct . Thus these functions take preparations on A to preparations on B. The operational theory will amount to a certain form of representation of this process category. ×. 1). this acts functorially on the admissible transformations f : A → B in Ct to produce functions f∗ : PA → PB . • The general point of view is that the structure of the operational theory is controlled by the ‘abstract’ category C. 3. The ﬁrst new ingredient which picks up the issue of monoidal structure is that P and M are required to be monoidal functors. • Rather than a single set of preparations. M are embeddings. Given a symmetric monoidal category C. The receiving category for the representation will be (Set. We shall assume that the functors P. the functor M speciﬁes a variable set MA of measurements for each system type A. Here KD is the constant functor valued at D. as already discussed. OPERATIONAL THEORIES AND CATEGORICAL QUANTUM MECHANICS 93 categorical structure of operational theories. The contravariant action of this functor is again as expected from our previous discussion. for each object A of Ct .

The coherence conditions for monoidal natural transformations complete the required properties of pure tensors. Moreover.B : MA × MB → MA⊗B . Thus we obtain a symmetric monoidal category. By functoriality of P and M.B A. P. MB . f∗ ⊗ f ∗ := (f ⊗ f )∗ . If we are given an operational representation (C.B P A × PB / PA⊗B MA ×O MB / MA⊗B O f∗ ×g∗ (f⊗g)∗ f ∗ ×g ∗ (f⊗g)∗ PA × PB / PA ⊗B MA × MB / MA ⊗B AP . The dinatural transformation dA : PA × MA → D represents the evaluation function. dB ). .2. we have the Chu space (PA . dA ). MA . f ∗ ) : (PA . Naturality means that the diagrams P M A.B commute. M. recovering the picture given in Section 3. i.4.94 SAMSON ABRAMSKY AND CHRIS HEUNEN monoidal means that there are natural transformations P A.B : PA × PB → PA⊗B .B AM . d) we can construct from this a single category. Dinaturality says that for each admissible transformation f : A → B: PB8 × MB f∗ ×1B dB $ PA × MB :D 1A ×f ∗ dA & PA × MA Thus we see that dinaturality is exactly the Chu morphism condition (1). whose underlying category is a sub-category of Chu spaces. M A. By dinaturality of d. since P and M are embeddings. f ∗ ⊗ f := (f ⊗ f )∗ . inclusions of pure tensors. Ct . Monoidality of d is the equation (2). dA ) → (PB . we can push the symmetric monoidal structure on C forward to this sub-category: PA ⊗ PB := PA⊗B . each morphism f : A → B gives rise to a Chu morphism (f∗ . Operational categories. ∗ MA ⊗ MB := MA⊗B . For each object A of C. 3. We call this the operational category arising from the operational representation.e. we obtain a sub-category of Chu spaces. MA .

1) is a commutative monoid with a zero element. . We shall brieﬂy review the deﬁnitions. ·. however. Concretely. OPERATIONAL THEORIES AND CATEGORICAL QUANTUM MECHANICS 95 3. The structural properties of operational representations and categories are independent of the particular choice of the monoid D used in specifying the dinatural transformation d. This yields the deﬁnition of operational representation given previously when W = D. and this assignment satisﬁes: 1† = 1. We now have a general scheme for representing symmetric monoidal cate- gories as operational categories. we have no examples. .5. d). Ct . following the ideas of categorical quantum mechanics [7]. §4. Generalized representations. . strictly involutive functor. f †† = f . A dagger category is a category C equipped with an identity-on-objects. where d now has the form ·· d : P × M → KW and KW is the constant symmetric monoidal functor valued at W. Monoidal dagger categories. where (W. Monoidal dagger categories are the basic structures used in categorical quantum mechanics [7]. So far. and give a number of examples. We deﬁne an arrow f : A → B in a dagger category to be a dagger-isomorphism if: f † ◦ f = 1A . ⊗. contravariant. α. for each arrow f : A → B. there is an arrow f † : B → A. I. f ◦ f † = 1B . P. (g ◦ f)† = f † ◦ g † . We shall deﬁne a generalized operational representation with weights W. to be a tuple (C. We shall now show how monoidal dagger categories give rise to operational repre- sentations in a canonical fashion. M. A symmetric monoidal dagger category is a dagger category with a symmetric monoidal structure (C.

) such that (f ⊗ g)† = f † ⊗ g † and moreover the natural isomorphisms . α. . .

and its (full) sub-category FHilb of ﬁnite-dimensional Hilbert spaces. Examples. . More generally. 28]. This includes categories of (right) Hilbert C*- modules. and the tensor product has its standard interpretation for Hilbert spaces. any symmetric monoidal C*-category is an example [33. are componentwise dagger- isos. Here the dagger is the adjoint. • The category Hilb of Hilbert spaces and bounded linear maps. which are Hilbert spaces whose inner product takes values in an arbitrary C*-algebra instead of C.

I )). the family {M (x. The doubling construction. with the action on matrices given by componentwise multiplication. Any topological or conformal quan- tum ﬁeld theory is a sub-category of the case where D = FHilb and C is the category of cobordisms [45. even though they only form a bicategory [16]. then so is the cate- gory [C. All of the above examples are variations on the theme of matrix categories. is obtained by forming the category FMat(S). Morphisms are natural transformations. However. g : B → A. setting D = FHilb and letting C be a group. g)† = (g. Although the construction is formal. This generalizes to relations valued in a commutative quantale [55]. (This corresponds to the ‘Kronecker product’ of matrices). Composition is deﬁned componentwise. Small categories as objects and profunctors as morphisms behave very similarly to Rel. which we think of as ‘X times Y matrices’. D] of functors F : C → D that preserve the dagger. y)}y∈Y is 2 -summable. Indeed. we obtain the category of unitary representations. Given a category C. This accounts for several interesting models. • A common generalization of FHilb and FRel. and for each y ∈ Y . • An inﬁnitary generalization of FMat(C) is given by LMat. For example. we recover FMat(D(I. Letting C be the discrete category N. The tensor product of X and Y is given by X × Y . g) : A → B is a pair of C-morphisms f : A → B. and to the category of relations for any regular category [18]. while if we take S to be the Boolean semiring {0. the category of ﬁnite sets and relations. this yields a category equivalent to FHilb. 56]. while the dagger is conjugate transpose. This is in fact the object part of the right adjoint to the . and letting D be either FHilb or FRel. it is interesting in our context since it can be seen as a form of quantization. 1} (with trivial involution). we deﬁne a dagger category C as follows. Hilb is equivalent to a (non-full) sub-category of LMat. where conjugation of a matrix means elementwise application of the involution on S. where S is a commutative semiring with involution. and maps X × Y → S as morphisms. the family {M (x. while (f. Here the dagger is relational converse. it seems hard to ﬁnd natural examples which are not of this form. This category has arbitrary sets as objects. and as morphisms matrices M : X × Y → C such that for each x ∈ X . while the monoidal structure is given by the cartesian product. FMat(S) has ﬁnite sets as objects. If we take S = C.96 SAMSON ABRAMSKY AND CHRIS HEUNEN • The category Rel of sets and relations. 8. f). y)}x∈X is 2 -summable. it converts classical process categories into a form in which quantum constructions are meaningful. there is a construction which produces a symmetric monoidal dagger category from any symmetric monoidal category. Composition is by matrix multiplication. we get FRel. and a morphism (f. • If C and D are symmetric monoidal dagger categories. The objects are the same as those of C.

see [38. see [39]. 3. The ﬁrst is zero morphisms: for each pair of objects A. A trace ideal is an endomorphism ideal I. for all f : C → A and g : B → D. g) ⊗ (h. h : B → A ⇒ g ◦ f ◦ h ∈ I(B) f ∈ I(A). An endomorphism ideal in a symmetric monoidal category C is speciﬁed by a set I(A) ⊆ End(A) for each object A. This is subject to the following closure conditions: g : A → B. I is a dagger endomorphism ideal when additionally f ∈ I(A) ⇒ f † ∈ I(A).1. For more examples. then C is a symmetric monoidal dagger category.D . together with a function TrA : I(A) → End(I ) 1 Strictly speaking. we are deﬁning the more restricted notion of global trace of an endomorphism.B ◦ f = 0C. g ⊗ k). Examples. f ∈ I(A). A) is the set of endomorphisms on A. The ﬁnal ingredient we shall require is a trace ideal in the sense of [4]. All the examples of symmetric monoidal dagger categories given above have zero morphisms in an evident fashion. i. form a commutative monoid [44]. 4. a morphism 0A. g ∈ I(B) ⇒ f ⊗ g ∈ I(A ⊗ B).B = 0A. g ◦ 0A. OPERATIONAL THEORIES AND CATEGORICAL QUANTUM MECHANICS 97 evident forgetful functor DagCat → Cat. We shall require two further structural ingredients. we recall that in any monoidal category. Note in particular that the structural isos in C turn into dagger isos in C . where End(A) = C(A.B . This has the universal property with respect to dagger functors C → D for categories D. Zero morphisms in C are pairs of zero morphisms in C. 0† = 0. k) := (f ⊗ h. Thus for each dagger category C.e. we further require that f ⊗ 0 = 0 = 0 ⊗ g. the endomorphisms of the tensor unit I .1 Firstly. If C is a dagger category. 0A. there is a dagger functor C : C → C which is the identity on objects. This restricted notion is all we shall need. they are unique. Functor categories have componentwise zero morphisms. In the context of symmetric monoidal dagger categories. Note that if zero morphisms exist. B. with the monoidal structure deﬁned componentwise: thus (f.B : A → B such that. f † ). Additional structure. I(I ) = End(I ) 0 ∈ I(A). but we will also call these endomorphism ideals for short.17]. the scalars. This cofree construction of a dagger category lifts to the level of symmetric monoidal categories. and sends f to (f. .1. rather than a parameterized trace as in [4]. If C is a symmetric monoidal category.

A s = s. where A ranges over the objects of C. Through the GNS-embedding [33. D] inherits endomor- phism ideals and zero morphisms from D componentwise. The doubling construction turns trace ideals into dagger trace ideals. which is always deﬁned. From categorical quantum mechanics to operational categories. g ◦ f ∈ I(A). and ( A sA tA ) = ( A sA )( A tA ). • We show that we really need to restrict to ideals to consider traces: the category of Hilbert spaces does not support a trace on all morphisms. TrI (s) = s. In Appendix B. andhas a trace function Tr(α) = A Tr(αA ) as soon as D(I. In the case of Hilb. the usual matrix trace is a total operation. Let C be a symmetric monoidal dagger category with zero morphisms and a trace . This is the case when C is a ﬁnite group. but is of mathematical interest in its own right. C is a dagger category with the same structure. and similarly for LMat. the details turn out to be quite subtle. we prove a number of results about trace ideals: • We characterize when trace ideals exist. Any symmetric monoidal dagger sub-category of [C. • As a corollary. TrA (g)). This material would have unduly interrupted the main ﬂow of the paper. and to what extent they are unique. I ) has an operation satisfying † † A sA = ( A sA ) . we prove in some detail that the category of Hilbert spaces indeed has a trace ideal. g) = (TrA (f).14].98 SAMSON ABRAMSKY AND CHRIS HEUNEN for each object A. as well as for topological quantum ﬁeld theories. we interpret trace class in the standard sense for Hilbert spaces. we derive that dual objects in the category of Hilbert spaces are necessarily ﬁnite-dimensional. this also provides a trace ideal for any C*-category. g : B → A. • Finally. g) : A → A. Examples. In the case of relations. We call a morphism f ∈ I(A) trace class. Proposition 1. §5. g) ∈ I(A) if and only if f ∈ I(A) and g ∈ I(A). Thus if C is a symmetric monoidal category with zero morphisms and a trace ideal. f ◦ g ∈ I(B)) TrA⊗B (f ⊗ g) = TrA (f)TrB (g). In the case of ﬁnite matrices. subject to the following axioms: TrA (g ◦ f) = TrB (f ◦ g) (f : A → B. but we will also call these trace ideal for short. deﬁne (f. For (f. A dagger trace ideal additionally satisﬁes TrA (f † ) = TrA (f)† . and TrA (f. the summation over the diagonal becomes a supre- mum in a complete semilattice. All of the examples given above have trace ideals.

A morphism f ∈ End(A) in a dagger category is positive if for some g : A → B. i = j. Given a dagger isomorphism f : A → B in C. We take MA to be the set of projective measurements on A. directly inspired by quantum mechanics. this deﬁnition yields exactly the standard notion of density operator as used in quantum mechanics. Inclusion of pure tensors is given by P A.1. We shall show that C gives rise to an operational representation and operational category in a canonical fashion. It is straightforward to check the coherence conditions. Transformations.e. all morphisms are invertible.3. the function f∗ : PA → PB is deﬁned by f∗ : s → f ◦ s ◦ f † . . It is easily seen to be a monoidal dagger sub-category of C. ◦ † ∈ PA . t) → s ⊗ t. A family {fi }i∈I of endomorphisms on A is: • Pairwise disjoint if fi ◦ fj = 0. h : B → A: [ ∀i ∈ I. on A is an arrow P ∈ End(A) such that P 2 = P. i. OPERATIONAL THEORIES AND CATEGORICAL QUANTUM MECHANICS 99 ideal. In Hilb. Indeed. fi ◦ g = fi ◦ h ] ⇒ g = h. and TrA ( ◦ † ) = TrI ( † ◦ ) = TrI (1) = 1 using our assumption on and the axioms for the trace. We take Ct to be the sub-category with the same objects as C.2. This is a groupoid. A projective measurement on A with ﬁnite set of outcomes O ⊆ O is a family of dagger idempotents {Po }o∈O on A which is pairwise disjoint and jointly monic. and such that TrA (f) = 1. A dagger idempotent. • Jointly monic if for all g. An arrow : I → A has unit norm if † ◦ = 1.B : (s. We write PA for the set of states on A. 5. P = P†. f = g † ◦ g. Pure states can also be deﬁned in this setting. and with dagger-isomorphisms as arrows. We deﬁne a state on A to be a positive morphism f ∈ End(A) which is trace class. Given such an arrow. Functoriality holds. or projector. since g∗ ◦ f∗ (s) = g ◦ (f ◦ s ◦ f † ) ◦ g † = (g ◦ f) ◦ s ◦ (g ◦ f)† = (g ◦ f)∗ (s). 5. this arrow is clearly positive. States. Measurements. 5.

Evaluation. is just: TrB (f ◦ s ◦ f † ◦ Po ) = TrA (s ◦ f † ◦ Po ◦ f). and a dinatural transformation d. o ) = TrA⊗B (s ⊗ s ◦ Po ⊗ Po ) = TrA⊗B (s ◦ Po ⊗ s ◦ Po ) = TrA (s ◦ Po )TrB (s ◦ Po ) = dA (s. We say that the canonical representation is distributional if the monoid of scalars End(I ) has an addition making it a commutative semiring. Functoriality is also easily veriﬁed. The transformation d is deﬁned as follows. Ct . It is easily veriﬁed that f ∗ preserves disjointness and joint monicity of fam- ilies of projectors.100 SAMSON ABRAMSKY AND CHRIS HEUNEN The functorial action of the measurement functor on dagger isomorphisms f : A → B in C is deﬁned by f ∗ (Po ) = f † ◦ Po ◦ f. m ⊗ m )(o. The monoidality of d is veriﬁed as follows: dA⊗B (s ⊗ s . 5. The canonical operational representation. Proposition 3. i. Given a symmetric monoidal dagger category C with zero morphisms and a trace ideal. Note that d is valued in the commutative monoid of scalars W := End(I )O . Note that the combined measurement will have a ﬁnite set of outcomes which. 5. We call this the canonical operational representation of C. d) is an operational representation with weights W. and m = {Po }o∈O ∈ MA : TrA (s ◦ Po ). P.5. the Chu morphism condition. Inclusion of tensors is deﬁned pointwise on projectors: P A. o ∈ O dA (s. where s ∈ PA . we have deﬁned a sub-category Ct . m )(o ). otherwise. monoidal functors P and M.4. m)(o) = 1. and hence carries projective measurements to projective measurements. m)(o) · dB (s . this monoid has a zero element.B : (Po . can be regarded as a subset of O. By the assumption of zero morphisms. and for each state s ∈ PA and measurement m ∈ MA : dA (s. Po ) → Po ⊗ Po . M. We collect the constructions described in this section together. m)(o) := 0.e. perhaps with some relabelling. The tuple (C. The dinaturality of this transformation. The corresponding operational category is the canonical operational category for C. (4) o∈O .

OPERATIONAL THEORIES AND CATEGORICAL QUANTUM MECHANICS 101 We say that it is probabilistic if moreover the image of d embeds into the semiring of non-negative reals.3 for dagger categories. we have the following result. We think of the outcomes as labelling the eigenvalues of the observ- able. then i∈I g = 1A ◦ g = Pi ◦ g = Pi ◦ g i∈I i∈I = Pi ◦ h = Pi ◦ h = 1A ◦ h = h. and probabilities of measurement outcomes. (projective) measurements. Hilbert spaces. i∈I i∈I For the converse. suppose that i∈I Pi = 1A . dagger idempotents correspond exactly to projectors in Hilb. This implies that for some non-zero vector . so the family is not jointly monic. transformations. Clearly. i∈I Indeed. Thus it is immediate that the states in the canonical representation for Hilb are the density matrices. Examples of operational categories. Pi = 1A . We shall now examine the opera- tional categories arising from various examples of symmetric monoidal dagger categories. Pi () = 0 for all i. Proposition 5. while the dagger-isomorphisms are the unitary transformations. the deﬁnition of d matches the standard statistical algorithm of quantum mechanics. if Pi = 1A and Pi ◦ g = Pi ◦ h for all i. The operational category arising from Hilb is of course probabilistic. More precisely. . Proof.e. 6. Then for f : C → A given by 1 → . and so does the notion of a pairwise disjoint family of projectors. Thus we obtain the standard interpretations of states. we have Pi ◦ f = Pi ◦ 0 for all i.1. It remains to show that the joint monicity condition captures the fact that a pairwise disjoint family of projectors {Pi }i∈I yields a resolution of the identity. For measurements. §6. The deﬁnitions of states. observables with ﬁnite discrete spectra correspond exactly to the interpretation in Hilb of the abstract notion of measurements as deﬁned in Section 5. then the family {Po }o∈O should correspond to the spectral decomposition of the observable. Measurements in Hilb have exactly their standard meaning. i. measurements and evaluation are directly inspired by those used in the standard Hilbert-space formulation of quantum mechanics. Finally.

Because the tensor unit in such categories is the constant functor KI . Note that the full sub-category FRel(Ω) of ﬁnite sets is identical to FMat(Ω). the bottom element of the lattice. Note that ΔS ◦ ΔT = ⊥X ×X ⇐⇒ S ∧ T = ⊥X . Rel is the special case that Ω is the Boolean semiring {⊥. where we identify ⊥. and for subcategories of [C. If R : X → Y and S : Y → Z. (6) i∈I i∈I .102 SAMSON ABRAMSKY AND CHRIS HEUNEN The same analysis holds for C*-categories through their GNS-construction. then: x(S ◦ R)z := xRy ∧ ySz. we write X for the Ω-subset of X given by x → . states and measurements in such categories are just natural transformations whose components are states or measurements respectively. with 1. and we are simply in the case of matrices over idempotent semirings. ΔSi = 1A ⇐⇒ Si = X . We write xRy = for R(x. }. the top element. We recall that a locale [42] (also known as a frame or complete Heyting algebra) is a complete lattice Ω such that the following distributive law holds: a∧ bi = a ∧ bi . Therefore the induced operational categories are probabilistic. in the ﬁnite case. and an ‘intersection’ i Si given by x → i Si (x). where we regard Ω as a semiring with idempotent addition and multiplication. i∈I Given a Ω-subset S of X .we mean a functionX → Ω. xΔS y = ⊥ if x = y.2. Si = X . y) = . while the morphisms R : X → Y are Ω-valued relations (or matrices) R : X × Y → Ω. Given a set X . y∈Y Clearly. Composition is relational composition (or matrix multiplication) evaluated in Ω. We shall now give a general analysis of the operational representation for locale-valued relations. and . and ⊥X for the Ω-subset of X given by x → ⊥. Relations. Hilb] such as topological quantum ﬁeld theories. they have the same scalars as Hilb. In particular. completeness of Ω need not be assumed. we deﬁne a Ω-relation ΔS : X → X by S(x) if x = y. This level of generality will be useful when we go on look at non-locality in operational categories. Indeed. i∈I i∈I The category Rel(Ω) has sets as objects. we say that a family {Si }i∈I of Ω-subsets of X is a disjoint cover of X if: Si ∧ Sj = ⊥X (i = j). with 0. We shall take the tensor unit in Rel(Ω) to be I = {•}. By an Ω-subset of a set X . 6. Any family {Si } of Ω-subsets of X has a ‘union’ i Si given by x → i Si (x).

We say that states s. and also •(Pi ◦ S)x = .g. OPERATIONAL THEORIES AND CATEGORICAL QUANTUM MECHANICS 103 Proposition 7. Proof. and for some relation R. which implies that xPi x ≥ xPi y.e. S : I → X by •Rx = = •Sy. •TrX (R)• = xRx. Firstly. Indeed. where {Si }i∈I is a disjoint cover of X . since e. by disjointness of the family. ⊥ < xPj z ≤ xPj x implies Pi ◦ Pj = ⊥. Every state in Rel(Ω) is equivalent to a pure state. For the converse. TrX (s ◦ ΔT ) = [x(s ◦ ΔT )x] x . Thus Pk ◦ R = Pk ◦ S for all k ∈ I . We claim that s is equivalent to PS . If R : X → X is an Ω-valued relation. If s is a state on X . which is satisﬁed to the extent that the relation has a ‘ﬁxpoint’. Proposition 8. Sup- pose for a contradiction that xPi y = > ⊥ where x = y. Then •(Pi ◦ R)x = z •Rz ∧ zPi x = ∧ xPi x = since xPi x ≥ yPi x = . deﬁned by xPS y = S(x) ∧ S(y). a reﬂexive element. x The corresponding pure state is PS . suppose we have a projective measurement {Pi }i∈I on X . The fact that Pi is a projector in Rel(Ω) means that xPi y = yPi x and xPz = y xPy ∧ yPz. Hence Pi must have the form Pi = ΔSi for some Si ⊆ X . Hence Pi ◦ R = Pi ◦ S. and •(Pi ◦ R)z = ⊥ = •(Pi ◦ S)z for other z. Similarly •(Pi ◦ R)y = = •(Pi ◦ S)y. Note that Ω-valued relations R : I → X of unit norm correspond to Ω- subsets S of X satisfying S(x) = . xsy = xRz ∧ yRz. z Deﬁne an Ω-subset S = dom(s) of X by x → xsx. The fact that the family {Si }i∈I is a disjoint cover of X now follows from (6). Deﬁne R. we give an explicit description of the trace. Moreover. t on X are equivalent if for all Ω-subsets S of X : TrX (s ◦ ΔS ) = TrX (t ◦ ΔS ). contradicting joint monicity. i. Pj ◦ R = Pj ◦ S = ⊥ for any j = i. then it satisﬁes = x xsx. and •Rz = ⊥ = •Sz for other z. Proof. Projective measurements on X in Rel(Ω) consist of families of relations {ΔSi }i∈I . Clearly any family of relations of this form is a projective measure- ment. x Thus the trace can be viewed as a predicate on endo-relations. for any Ω-subset T of X . Next we analyze states in Rel(Ω).

Thus we have the following result. These results highlight two important diﬀerences between Rel(Ω) and Hilb as operational categories. every projector can appear as part of a projective measurement. Let ΔS be a state. m)(o) = TrX (ΔS ◦ ΔSo ) o o = xΔS y ∧ yΔSo x o.y = TrX (PS ◦ ΔT ). so that every state is equivalent to a pure one. is lost. Finally. The relevance of this will become apparent when we discuss non-locality in Rel(Ω) in Section 9.y = T (x) ∧ xsx x = T (x) ∧ S(x) x = yPS x ∧ xΔT y x. while in Rel(Ω) the collective conditions of disjointness and joint monicity impose the constraint that projectors have to be sub-identities ΔS . Because states correspond to Ω-subsets S satisfying x S(x) = . In Hilb. The scalars in Rel(Ω) can be identiﬁed with the locale Ω. and measurements to disjoint covers.104 SAMSON ABRAMSKY AND CHRIS HEUNEN = xΔT y ∧ ysx x. and convex combinations to form mixed states. Proposition 9.y = S(x) ∧ So (x) o.2. Then dA (ΔS . and m be a measurement given by a disjoint cover {So } of X . in Rel(Ω) the distinction between superpositions of pure states. Proof. . we consider evaluation.x. Moreover. Discussion.x = S(x) ∧ So (x) x o = S(x) x = . we see that equation (4) is satisﬁed. The operational category arising from Rel(Ω) is distribu- tional.

as follows. 1] which are row-stochastic. we deﬁne f∗ (s)(y) = s(x). However. the dagger would not be given by transpose of (bi-stochastic) matrices. equivalent to C op . The objects are ﬁnite sets. The functorial action of states is described as follows. it is in particular self-dual. Thus if f : X → Y is a function. and the morphisms M : X → Y are the X × Y -matrices valued in [0. operational theories should also include classical physics— or its discrete operational residue. Thus for each x ∈ X . Indeed. However. One might consider using the formal doubling construction on Stoch to obtain a dagger symmetric monoidal category with a dagger trace ideal. • The sub-category Stocht is deﬁned by restricting to the functions (deter- ministic transformations). m)(o) = s(x). Our notion of operational representation is indeed broad enough for this. Proposition 10. An alternative description of Stoch is as the Kleisli category for the monad of discrete probability distributions. • A state on X is a morphism I → X in Stoch. The construction of operational repre- sentations on monoidal dagger categories is directly inspired by quantum me- chanics. or equivalently a probability distribution on X . Proof. as we shall now show. it is easy to give a direct deﬁnition of an operational representation. But this would not yield the expected result. The basic classical setting we shall consider is the category Stoch. m(x)=o . The functorial action on f : X → Y is just m → m ◦ f. Note that if a category C has a dagger structure. The monoidal structure is deﬁned as for FMat(S). • The evaluation is deﬁned by: dX (s. This is just a discrete random variable. see [40]. which is thus not self-dual. There is no dagger structure on Stoch. we have the following result. Given f : X → Y . However. This is the classical notion of mixed state. f(x)=y • A measurement on X is a function m : X → O with ﬁnite image O ⊆ O. for example. OPERATIONAL THEORIES AND CATEGORICAL QUANTUM MECHANICS 105 §7. Note that Stoch is not closed under matrix transposition. the one-element set is terminal but not initial in Stoch.e. we have a probability distribution on Y . i. It follows that we cannot directly apply the construction of Section 5. for each x ∈ X the corresponding probability distribution is f(x) . Classical operational categories. represented as matrices by their characteristic maps.

5]. as in [40]. . Ct . x∈X . . We ﬁx a state s ∈ PA . Having set up a general frame- work for operational categories. there are n agents or sites. and provide the empirical yardstick by which it is judged. We can regard these models as observational ‘windows’ on the operational theory. 8. we have a probability distribution on the joint outcomes of the measurements. on ) for m is given by p(o|m) := dA (s. . we obtain the measurement m := m1 ⊗ · · · ⊗ mn by inclusion of pure tensors. formulate non-locality in terms of the (non-)existence of a joint distribution [31. They represent the directly accessible information predicted by the theory. and perhaps more elegantly. . . We shall follow the traditional route of using hidden variables explicitly. . We are assuming a ﬁxed distributional model. . where mi ∈ MAi for i = 1. Now the probability of obtaining a joint outcome o := (o1 . M. . . We now deﬁne what it means for an empirical model of the kind described in the previous sub-section to exhibit non-locality. . although we could equivalently. For each choice of a measurement setting by each of the agents. m)(o). Empirical models. The above data speciﬁes a probabilistic operational repre- sentation of Stoch. . . namely non-locality. • We can generalize to probability measures over general measure spaces. n. 8. d) on a monoidal category C. In these models. . . For each combination of measurements (m1 . we ﬁx a distributional operational representation (C. with a semiring of weights W. P. Non-locality in operational categories. We shall associate objects A1 .106 SAMSON ABRAMSKY AND CHRIS HEUNEN The following result is easily veriﬁed. Non-locality. Proposition 11. This will still yield a distributional operational representa- tion. we shall now investigate an important founda- tional notion in this general setting. such that d (x) = 1. . An with the n sites. §8. Throughout this section. We deﬁne A := A1 ⊗ · · · ⊗ An . Various generalizations of this construction are possible: • We can generalize to ‘distributions’ over an arbitrary commutative semir- ing.2.1. and each measurement has a number of distinct outcomes. A W-distribution on a set X is a function d : X → W of ﬁnite support. each of which has the choice of one of several measurement settings. This amounts to using the Kleisli category of the Giry monad [34]. . We shall begin by showing how probability models of the form commonly studied in quantum information and quantum founda- tions can be interpreted in the corresponding operational category. mn ).

The required condition for the hidden variable model to realize the empirical model p is that. . . . . Ultimately. relative to a given distributional operational representation. Thus we can lift these ideas to the general level of operational categories.1. We say that the hidden-variable model is local if. . oi =oi . . we have a criterion for ascribing non-locality to monoidal process categories themselves. we recover the empirical probabilities by averaging over the hidden variables. . ∈Λ That is. 1) (a. which is essentially the ﬁnite-dimensional part of standard quantum mechanics. As a standard example—essentially the one used by Bell in his original proof of Bell’s theorem—consider the following table. As expected. OPERATIONAL THEORIES AND CATEGORICAL QUANTUM MECHANICS 107 A hidden-variable model for an empirical model is deﬁned using a set Λ of hidden variables. We say that an operational category exhibits non- locality if it gives rise to a non-local empirical model. and o := (o1 . b) 3/8 1/8 1/8 3/8 (a . 1) (1. (0. Examples of non-locality. for all ∈ Λ. §9. . for all m and o: p(o|m) = q (o|m) · d (). the operational category arising from Hilb. and non-local otherwise. We shall now investigate non-locality in a number of examples. i=1 Here q (oi |mi ) is the marginal: q (oi |mi ) = q (o |m ). b ) 3/8 1/8 1/8 3/8 (a . 0) (1. on ): n q (o|m) = q (oi |mi ). Note that the deﬁnition of non-locality makes sense for any distributional operational category. does exhibit non-locality.mi =mi We say that the empirical model p is local if it is realized by some local hidden-variable model. with a ﬁxed distribution d . Hilbert spaces. b) 1/2 0 0 1/2 (a.2 For each ∈ Λ. the model speciﬁes a distribution q (o|m) on outcomes o for each choice of measurements m. . 0) (0. mn ). m = (m1 . 9. b ) 1/8 3/8 3/8 1/8 2 The assumption of a ﬁxed distribution d is technically the condition of ‘-independence’ [26].

The types are sets X1 . 9. 5]) shows that this table cannot be realized by a local hidden-variable model. does not admit non-local behaviour. Relations. in the style of ‘Bell’s theorem without inequalities’ [35]. and measurements mi = {ΔSoi }o∈O . Note that we are working over Ω (the locale of scalars in Rel(Ω)). This comes down to the following calculation for o ∈ O : p(o|m) = S(x) ∧ Soi (x) = S(x) ∧ Soi i (x) = px (o|m) ∧ ds (x). relational models can be used to give ‘logical’ proofs of non-locality and contextuality. at the model described above then still shows that such operational categories are non-local. A standard argument (see e. taking the hidden variables componentwise. We shall now construct a local hidden-variable model which realizes this empirical model. The key point is that these logical proofs are based on showing the non-existence of global sections compatible with a given empirical model. For each combination of measurements m and outcomes o. there is a state s = ΔS for a Ω-subset S of X := i Xi satisfying x S(x) = .g. . which are exactly sets of global elements. .g. The constant functor valued e. .108 SAMSON ABRAMSKY AND CHRIS HEUNEN It lists the probabilities that one of two outcomes (0 or 1) occurs when simultaneously measured with one of two measurements at two sites (a or a at the ﬁrst site. x x i x We conclude from this that Rel(Ω). and b or b at the second). so this is a well-deﬁned distribution. This stands in interesting counter-point to the fact that. . at a relative angle of /3. we have: x S(x) ∧ So (x) i if o ∈ O . . [15. written in the Z basis as |↑↑ + |↓↓ √ . 2 subjected to spin measurements in the XY -plane of the Bloch sphere. We deﬁne the distribution ds on X as x → S(x). as shown extensively in [3]. by a Bell state. We deﬁne p x (o|m) ≡ i Soi i (xi ). p(o|m) = 0 otherwise. This table can be realized in quantum mechanics. Xn . which sums to 1. so this hidden-variable model is local by construction.g. where {Soi } is a disjoint cover of Xi . e. despite being a ‘quantum-like’ monoidal dagger-category. The same reasoning applies to C*-categories and subcategories of [C. while here we are looking at empirical models generated by states in Rel(Ω).2. using the elements of X as the hidden variables. Hilb]. We must verify that this model agrees with the empirical model. Suppose we are given an empirical model in the distributional operational category obtained from Rel(Ω).

ˆ m(x)=o This is the empirical model realized by the state x. We shall now show that Stoch realizes exactly those models which have local hidden-variable realizations. For each m. as expected. the disjoint union of these sets of measurements. so this hidden-variable model is x local.g. We now consider the case of classical sto- chastic maps. as discussed in Section 7. OPERATIONAL THEORIES AND CATEGORICAL QUANTUM MECHANICS 109 The key feature of quantum mechanics. there is a ‘signed probability measure’ . .4. . and measurements mi : Xi → O. . Bearing in mind that ﬁnite-dimensional quantum mechanics corresponds to the operational category arising from FMat(C). all local models are of this form. thus non-locality can only arise in non-idempotent situations. Given an empirical model realized is a probability distribution on X := i Xi . This is in fact quite similar to the case by sets X1 . It is straightforward to verify that the probabilities p(o|m) are recovered by averaging over the deterministic hidden variables. . For each m = (m1 . . Thus elements of X simultaneously assign outcomes to all measurements. viewed as a probability distribution on the hidden variables X . We can deﬁne px (o|m) := o (m(x)). 9. mn ) ∈ i Mi . we again take the hidden variables to be the elements of X . and as shown e. We deﬁne To see M := i Mi . we can say more than this. . This is the category SStoch of signed stochastic maps. . . a state s which for Rel. . this. x∈X Note that is a probability distribution x on X . Classical stochastic maps. 9. Clearly p (o|m) = i p (oi |mi ). We can write s as a convex combination s= x x . we deﬁne a map mˆ : X → O by mˆ : x → (x(m1 ). and X := O M . is that quantum states under suitable measurements are able to realize families of probability distributions which have no global sections. . .3. x(mn )). We can calibrate the expressiveness of an operational theory in terms of which empirical models it realizes. Mn . Thus we conclude. . this shows that idempotence of the scalars implies that only local behaviour can be realized. Xn . by contrast. in [5]. Signed stochastic maps. while FRel(Ω) is FMat(Ω). . In fact. suppose we are given sets of measurements M1 . . Thus for each input. real matrices such that each row sums to 1. We shall now consider a variant of Stoch which has much greater expressive power in terms of the empirical models it realizes. we get the probability distribution on outcomes given by dm : o → s(x). that Stoch does not exhibit non-locality. . .

on the one hand. the category of sets and relations. −1/2. 50. which are non-trivial. 27. The following result can be extracted from [5. which forms a very useful ‘foil’ model for quantum mechanics in many respects [60. 0. Clearly. 1/2. we have a clear deﬁnition of whether a model of categorical quantum mechanics exhibits non-locality or not. each with outcomes {0. 0. while Hilbert-space quantum mechanics does. 0. this can be taken much further. 1}. The state x can be obtained from the PR-box speciﬁcation by solving a system of linear equations. The class of empirical models which are realized by the operational category obtained from SStoch are exactly the no-signalling models. • We have used our framework of operational categories to study non- locality in a general setting. The distributions it generates for the various measurement combinations can be listed in the following table. 0]. as explained at the end of Section 8. b ) 1/2 0 0 1/2 (a . This says that the operational category obtained from SStoch is more expres- sive. Example. Now consider the following state: x := [1/2. As we saw. −1/2. 1/2. 0) (0. than the canonical operational category derived from Hilb. Final remarks. 0. 1) (1. (0. 0. in terms of the empirical models it realizes.9] for the details. which corresponds to quantum mechanics. §10. 0. This paper makes a ﬁrst precise connection between monoidal categories. 22]. using the same encoding of empirical models which we employed in the previous sub-section. and X = {0. The reader should refer to [5. b ) 0 1/2 1/2 0 This can be recognized as the Popescu-Rohrlich box [53]. which may include ‘negative probabilities’ [61. b) 1/2 0 0 1/2 (a. 0. 30]. In particular. Theorem 5. Thus the disjoint union M of the two sets of measurements has four elements. and operational theories on the other. 0. b) 1/2 0 0 1/2 (a .9]. 1}M has 16 elements. An operational representation can be deﬁned for SStoch in the same fashion as for Stoch. it is still distributional. We consider the bipartite system with two measurements at each site. does not. We note a number of directions which it would be interesting to pursue. Theorem 5. 1) (a. 0) (1. An important further direction is to apply a similar analysis to . Proposition 12. 0. see [5] for details.110 SAMSON ABRAMSKY AND CHRIS HEUNEN on outputs. and the categorical quantum mechanics framework. 1/2. thus they properly contain the quantum models. which achieves super-quantum correlations.

the structure of the concrete category of convex operational theories is investigated.g. which can be seen as a broader phenomenon.S. . • We would also like to examine the issue of axiomatization or ‘reconstruc- tion’ of quantum mechanics from the categorical point of view. a general setting is developed allowing a uni- ﬁed treatment of contextuality and non-locality. • There are various constructions for turning a monoidal category of ‘pure’ states into one of ‘mixed’ states [57. and it is split epic when f ◦ g = 1A for some g : B → A. . . Oﬃce of Naval Research Grant Number N000141010357 is gratefully acknowledged. 23]. A category C has a collection of objects A. • It would also be interesting to interpret some of the general results which have been proved for operational theories. and ultimately to obtain such results for classes of monoidal process categories. and there are identity arrows 1A : A → A for each object A. we can form the composition g ◦ f : A → C . The collection of all arrows with domain A and codomain B is denoted as C(A. to no-broadcasting [11]. Regarding related work. Appendix A. . 1A ◦ g = g. In [5]. g. Composition is associative. OPERATIONAL THEORIES AND CATEGORICAL QUANTUM MECHANICS 111 contextuality. It would be interesting to relate these constructions to our canonical operational categories. . which in particular led to an improved formulation of Proposition 8. . Acknowledgements. there is a category embodying Spekkens’ toy theory [60. An arrow f : A → B is split monic when g ◦ f = 1A for some g : B → A. teleportation [12]. h. . Given arrows f : A → B and g : B → C . Each arrow has speciﬁed domain and codomain objects: notation is f : A → B for an arrow f with domain A and codomain B. B. 21]. in place of the somewhat clumsy device used in the present paper. We would like to extend the present account to this setting. For more detailed background. and information causality [10]. with f ◦ 1A = f. Financial support from EPSRC Senior Research Fel- lowship EP/E052819/1 and the U. C. by abuse of notation. we will write g = f −1 in both cases. We thank Shane Mansﬁeld for a number of useful comments. . We shall review some basic notions from category theory. B). First notions from category theory. of which non- locality is a special case. see [9]. . • Such a development would also lead to a more satisfactory treatment of outcomes. in which compatibility of measurements is explicitly represented. we note that in [13]. It would be of interest to study the associated operational category. relating e. and arrows f. in our categorical framework. leading to a natural sheaf-theoretic structure. An arrow f : A → B is called an iso(morphism) when f ◦ f −1 = 1B and f −1 ◦ f = 1A for some arrow f −1 : B → A. Similarly. for every f : A → B and g : C → A.

the following naturality diagram commutes: FA tA / GA Ff Gf FB / GB tB A natural isomorphism is a natural transformation whose components are isomorphisms. . a functor F : C → D assigns an object FA of D to each object A of C. α. I. · Given functors F.112 SAMSON ABRAMSKY AND CHRIS HEUNEN If C is a category. we write C op for the opposite category. These assignments must preserve composition and identities: F (g ◦ f) = F (g) ◦ F (f). for every f : A → B in C. An equivalence of categories is a pair of functors F : C → D and G : D → C such that there are natural isomorphisms F ◦ G ∼ = 1D and G ◦F ∼ = 1C . and arrows A → B corresponding to arrows B → A in C. and F (1A ) = 1FA . If C and D are categories. . such that. ⊗. with the same objects as C. A symmetric monoidal category is a structure (C. a natural transformation t : F → G is a family of arrows {tA : FA → GA} indexed by the objects of C. . G : C → D. and an arrow Ff : FA → FB of D to every arrow f : A → B of C.

• I is a distinguished object of C (unit). • ⊗ : C × C → C is a functor (tensor). . . • α.) where: • C is a category. .

C : A ⊗ (B ⊗ C ) → (A ⊗ B) ⊗ C A : I ⊗ A → A A : A ⊗ I → A .B. are natural isomorphisms (structural isos) with components αA.

A. • an arrow e : ID → FIC . Let C and D be symmetric monoidal categories. The sym- metric monoidal functor is called strong when m is a natural isomorphism. Products are a classical example of symmetric monoidal structure.B : FA ⊗ FB → F (A ⊗ B).B : A ⊗ B → B ⊗ A such that certain coherence diagrams commute. A symmetric monoidal functor (F. m) : C → D comprises • a functor F : C → D. The symmetric monoidal structure can also support entanglement. subject to coherence conditions with the structural isomorphisms. e. the category is then called compact [7]. . • a natural transformation mA. the category is then called Cartesian.

introduced in Section 4. For example. mA. OPERATIONAL THEORIES AND CATEGORICAL QUANTUM MECHANICS 113 Let (F. we prove in some detail that the category of Hilbert spaces indeed has a trace ideal. As a conceptually satisfying corollary. Finally. e . f2 : Y → X .1. such that f1 = i ◦ f1 and f2 = f2 ◦ i −1 . It presents several technical results that are mathematically interesting. we show that we really need to restrict to ideals to consider traces: the category of Hilbert spaces does not support a trace on all morphisms.B Appendix B. Proposition 13. this was claimed in Section 4. and Y and Y are objects of D. that the scalars are commutative. More precisely: if f1 : X → Y . m). we derive that dual objects in the category of Hilbert spaces are necessarily ﬁnite-dimensional. but the details are quite subtle. f1 ◦ f2 ∈ ID (Y )} .1. but would break up the ﬂow of the main text. Any dagger monoidal tracial sub-category D of C with a trace ideal induces a trace ideal I(X ) = {f ∈ C(X. holding in any monoidal category. Notice that traceability generalizes the fact. This appendix further studies the notion of trace ideal.B I e / FI FA ⊗ FB / F (A ⊗ B) tI tA ×tB tA⊗B e GI GA ⊗ GB / G(A ⊗ B) mA. Existence.1. and f2 ◦ f1 = f2 ◦ f1 . and to what extent they are unique. Trace ideals. then there is a morphism i : Y → Y in D that is either split monic or split epic. f1 :Y O f2 $ X i −1 i :X f1 $ f2 Y The category C is called traceable when the full sub-category consisting of the monoidal unit I is tracial. X ) | f = f2 ◦ f1 with f1 : X → Y. A monoidal natural transformation between them is a natural transformation · t : F → G such that the following diagrams commute. we characterize when trace ideals exist. f2 : Y → X are morphisms of C. m ) : C → D be symmetric monoidal functors. Also. e. B. The question whether a category allows a trace ideal at all can be answered as follows. A sub-category D of C is called tracial when endomorphisms in C factoring through D can only do so in a way unique up to isomorphism. f2 : Y → X and Y in D. (G. f1 : X → Y .

i. I ).e. Proof. If s = ◦ † is a pure state on X . Proposition 15. independent of the trace ideal. A dagger monoidal category has a unique minimal trace ideal I(X ) = {f : X → X | f factors through I } Tr(f) = b ◦ a. follows from the ﬁrst and third axioms of endomorphism ideal. in the sense that they are preserved under equivalence. because the full sub-category consisting of just the monoidal unit I is certainly (totally) traced. and (I. Therefore. Tr is well-deﬁned. We now consider uniqueness of trace ideals.114 SAMSON ABRAMSKY AND CHRIS HEUNEN Tr(f) = TrD (f1 ◦ f2 ) on C. One directly checks that I(X ) is an endomorphism ideal. F (X ⊗ Y ) ∼= F (X ) ⊗ F (Y ) and G(X ⊗ Y ) ∼ = G(X ) ⊗ G(Y ) that interact with the coherence isomorphisms in the appropriate way. then for every outcome o: Tr(s ◦ Po ) = Tr( ◦ † ◦ Po ) = Tr( † ◦ Po ◦ ) = † ◦ Po ◦ . Trace ideals are preserved under dagger monoidal equiva- lence: if F : C → D and G : D → C are strong monoidal functors that preserve daggers and form an equivalence of categories. when f = a ◦ b with a : I → X and b : X → I and hence has any trace ideal whatsoever. and there are natural isomorphisms F (I ) ∼= I . B. in partic- ular I(I ) = D(I. TrI ) is a trace ideal in C. such that F (f † ) = F (f)† and G(f † ) = G(f)† . form a trace ideal in D. the evaluation of measurements on pure states is completely determined by the structure of the category. and {Po } a measurement. Theorem 14. The following proposition proves that trace ideals are a categorical invariant. then J (X ) = G −1 (I(G(X ))) = {g ∈ D(X. That the given data form a trace ideal follows from the previous proposition. To see that this trace ideal is minimal. As a consequence of the previous theorem. G(I ) ∼ = I . X ) | G(g) ∈ I(G(X ))}. Uniqueness. . the only possible freedom the choice of a trace ideal brings comes out in behaviour on mixed states. I ) = C(I. Proof. if and only if it is traceable. Because D is tracial. The axioms for the trace function are also readily veriﬁed. A dagger monoidal equivalence is a pair of functors F : C → D and G : D → C that form an equivalence of categories. TrJ I X (g) = F (TrG(X ) (G(g))). that any trace ideal must contain this one.2.

We will now show that the monoidal category (Hilb. There exists a unital C*-algebra A with distinct tracial states = : A → C [46]. For n. see [59] and [37]. Verifying that TrJ satisﬁes the requirements is completely analogous. . . observe that if f ∈ I(X ). the ﬁrst requirement follows from functoriality of G. Make a category C as follows. More precisely. (1) = 1. the action on morphisms is by scalar multiplication. then Tr(f) = Tr(gfh). and composition is addition. and the third from fullness of G together with monoidality of G. the identity is 0. We give this category a monoidal structure by letting the tensor product of objects n and m be n + m. as the following example shows. g : X → X. resulting in a morphism TrU (f) : X → Y . The monoidal unit is the object 0. compact categories. respectively. . the action on morphisms is clear. This . First. In fact. Objects are natural numbers. the identity is (1. It is a dagger endomorphism ideal because G preserves daggers. and (ab) = (ba). B. where traces are unique. . We will now show that in the category Hilb. because all morphisms are endomorphisms. but requires a ‘partial trace’ of morphisms f : X ⊗ U → Y ⊗ U . morphisms n → n are elements of the n-fold direct sum A ⊕ A ⊕ · · · ⊕ A. OPERATIONAL THEORIES AND CATEGORICAL QUANTUM MECHANICS 115 Proof. and composition is pointwise multiplication. except that the last condition additionally uses F (G(s)) ∼ = s. There are then several additional axioms. an ) = i=1 (ai ) for n ≥ 1. trace ideals need not be unique. A tracial state on a C*-algebra A is a linear map : A → C satisfying (a ∗ a) ≥ 0. The previous example is in stark contrast to Cartesian categories or compact categories. to verify that J is an endomor- phism ideal. there may even be more than one trace function making a ﬁxed endomorphism ideal into a trace ideal. for example. . such as the following naturality: TrU (f) ◦ g = TrU (f ◦ (g ⊗ 1U )) for f : X ⊗ U → Y ⊗ U. 1). the second from the fact that G is monoidal. . However. It remains unclear whether trace ideals on. and Trn (a1 . The need for trace ideals. But the very same construction with gives a diﬀerent trace function.3. This satisﬁes all the conditions needed to make I into a trace ideal. Subsequently. Morphisms 0 → 0 are complex numbers. ⊗) cannot be traced monoidal. Deﬁne Tr0 (z) = z. . For n ≥ 1. and g : X → Y is an isomorphism with inverse h. Then. This notion asks not just for traces of all endomorphisms. The counterexample above is somewhat artiﬁcial. Example 16. are unique. m ≥ 1. . we will show that Hilb is not an instance of the established notion of traced monoidal category [43]. There are only endomorphisms. there exists no trace ideal consisting of all morphisms. If one of n or m is 0. we will show that it does have a trace ideal. Taking I(X ) to be all endomorphisms on X certainly n gives an endomorphism ideal. 1.

Hence F ◦ (|0 ⊗ 1H ) = f and F ◦ (|1 ⊗ 1H ) = g. Lemma 17. |1} for C2 . and write |+ = |0 + |1. ⎛ ⎞ ⎛ ⎞ ⎛ ⎞ 1 0 0 1 0 0 0 0 0 f1 = ⎝0 0 0⎠ f2 = ⎝0 1 0⎠ f3 = ⎝0 1 0⎠ 0 0 0 0 0 0 0 0 0 Let g = ϕ ⊕ : (H ⊕ C) ⊕ H → H ⊕ (C ⊕ H ). ⊗) is not traced monoidal. Suppose (Hilb. Choose an orthonormal basis {|0. ⊗) was traced monoidal. Proof. Suppose (Hilb. which is a contradiction. Deﬁne F : C ⊗ H → H via the block 2 2 matrix f g . ⊗) is traced monoidal. Then there exist isomorphisms ϕ : H ⊕ C → H ∼ = ! and : H→ C ⊕ H . Thus Tr(f2 ) = Tr(f2 ) + 1. Then ⎛ ⎞ ⎛ ⎞ ⎛ ⎞ ϕ1 ϕ2 0 1 0 0 ϕ1 ϕ2 0 g ◦ f2 = ⎝ 0 0 1 ⎠ ◦ ⎝0 1 0⎠ = ⎝ 0 0 0⎠ 0 0 2 0 0 0 0 0 0 ⎛ ⎞ ⎛ ⎞ 1 0 0 ϕ1 ϕ2 0 = ⎝0 0 0⎠ ◦ ⎝ 0 0 1 ⎠ = f1 ◦ g. The monoidal category (Hilb.116 SAMSON ABRAMSKY AND CHRIS HEUNEN justiﬁes working with trace ideals in monoidal categories instead of traced monoidal categories. Write them in block matrix form as ϕ = ϕ1 ϕ2 and 1 = . we know that Tr(f3 ) = Tr(1C ) = 1. Proof. Let H be an inﬁnite- ∼= dimensional Hilbert space. f3 : H ⊕ C ⊕ H → H ⊕ C ⊕ H 2 given by the following block matrices. 0 0 0 0 0 2 Hence Tr(f1 ) = Tr(f1 ◦ g ◦ g −1 ) = Tr(g ◦ f2 ◦ g −1 ) = Tr(f2 ◦ g −1 ◦ g) = Tr(f2 ). But Tr(f2 ) = Tr(f1 + f3 ) = Tr(f1 ) + Tr(f3 ) by Lemma 17. And because f3 has ﬁnite rank. We are indebted to Peter Selinger for the following proof. Consider the morphisms f1 . The third equality uses that composition is bilinear. Then Tr(f + g) = Tr(f) + Tr(g) for all endomorphisms f. f2 . Recall ∼ ! that C ⊗ H = H ⊕ H . Now: Tr(f + g) = Tr(F ◦ (|+ ⊗ 1H )) = Tr(F ) ◦ |+ (by naturality) = (Tr(F ) ◦ |0) + (Tr(F ) ◦ |1) = Tr(F ◦ (|0 ⊗ 1H )) + Tr(F ◦ (|1 ⊗ 1H ) (by naturality) = Tr(f) + Tr(g). Theorem 18. g : H → H . .

Trace class maps form a trace ideal in Hilb. Now. An arbitrary continuous linear map f : H → H is trace class when its absolute value |f| : H → H is trace class. see [44. one can show that if A∼ = A ⊕ I . We need to recall some terminology. No rigorous proof that inﬁnite-dimensional Hilbert spaces cannot have duals has been published. then TrA (1A ) = TrA (1A ) + 1. as far as we know. A positive & continuous linear map f : H → H is trace class when n &en | f(en )& < ∞ for an orthonormal basis (en ) of H . 43.5. Other good references are [32. 58]. A linear map f : H → K between Hilbert spaces is Hilbert-Schmidt when n !f(en )!2K < ∞ for an orthonormal basis (en )&of H . 7]. . Dual objects in Hilb are ﬁnite-dimensional. In Hilb. We thank Jamie Vicary for this observation. This recipe does not work when H is inﬁnite-dimensional. C2 ⊗ H ∼ = H ⊕ H. i. Both deﬁnitions are independent of the choice of . Corollary 19. Let H be an inﬁnite-dimensional Hilbert space. for any choice of orthonormal basis {|i}i=1. as virtually all textbooks only consider endomorphisms. the object C always has a dual. To show that the usual trace of continuous linear maps between Hilbert spaces does in fact give a trace ideal requires some work. then so does H ⊕ K. as far as these make sense ‘locally’. we refer to [17]. for the objects C.e. H ⊕ C ⊕ H . in any monoidal category with biproducts.. this does not exclude the possibility that there might be other H ∗ . Hence the contradiction it results in holds here. because i |i does not converge in that case.4. for just one object H . H ⊕ C. ⊗) with duals are precisely ﬁnite-dimensional Hilbert spaces. ε making H into a dual object. H. OPERATIONAL THEORIES AND CATEGORICAL QUANTUM MECHANICS 117 B. B. too.. L∼= L ⊗ I −−→ L ⊗ (R ⊗ L) ∼ = (L ⊗ R) ⊗ L −−→ I ⊗ L ∼ 1⊗ ε⊗1 =L R∼ = I ⊗ R −−→ (R ⊗ L) ⊗ R ∼ = R ⊗ (L ⊗ R) −−→ R ⊗ I ∼ ⊗1 1⊗ε =R ∗ It is well-known that nif H ∈ Hilb is ﬁnite-dimensional. and if H and K have duals. notice that the proof of Theorem 18 only uses the trace properties ‘locally’. → H∗ ⊗ H ∼ = H ⊗ H ∗ −−−− → H ⊗ H∗ − f⊗1H ∗ ε I − →I This satisﬁes all equations for a trace function. Recall that the main characteristic of compact categories is that objects have duals: objects L. The previous theorem al- lows an interesting corollary. for any unexplained terms. Proof. Suppose H has a dual object H ∗ . In fact. then H and H are dual objects by (1) = i=1 |i ⊗ i| and ε(|i) = 1. Objects in (Hilb. For f : H → H . R in a monoidal category are called dual when there are maps : I → R ⊗ L and ε : L ⊗ R → I making the following two composites identities. deﬁne TrH (f) as the following composite. whereas the deﬁning conditions of trace ideals also involve morphisms between diﬀerent objects...n for H . However.

Proposition 21. then en | f(en ) is absolutely summable. It follows that |g ◦ f| = w † ◦ g ◦ f. The category Hilb has a dagger trace ideal consisting of the usual trace class maps and the usual trace function. Let H o g K be morphisms in Hilb. Because w is a partial isometry. Proof. y of a Hilbert space. Proof. Also. Hence. That the trace class maps on a Hilbert space H are closed under adjoint and tensor products is easily seen. By polar decomposition. & & & & &en | |g ◦ f|(en )& = &en | w † ◦ g ◦ f(en )& n n & & = &g † ◦ w(en ) | f(en )& n !1/2 ≤ !g † ◦ w(en )!2 · !f(en )!2 (by Cauchy-Schwarz) n & & = &!g † ◦ w(en )! · !f(en )!& n 1/2 1/2 † ≤ !g ◦ w(en )! 2 · !f(en )! 2 . the latter inequality holds if and only if n !g(en )!2 < ∞. for an orthonormal basis (en ) of H . The H¨older inequality states that 1/2 1/2 |xn · yn | ≤ |xn |2 · |yn |2 n n n for any two sequences (xn ) and (yn ) of complex numbers with |xn |2 < ∞ n and n |yn |2 < ∞. g ◦ f is trace class if and only if f and g are Hilbert-Schmidt.118 SAMSON ABRAMSKY AND CHRIS HEUNEN basis (en ). The Cauchy-Schwarz inequality states that & & ! &x | y& ≤ !x!2 · !y!2 1/2 for any two elements x. and hence the following trace property holds: Tr(f) = en | f(en ) n is a well-deﬁned complex number. there is a unique partial isometry w : H → K satisfying g ◦ f = w ◦ |g ◦ f| and ker(w) = ker(g ◦ f). f / Lemma 20. Then g ◦ f is trace class if and only if f and g are Hilbert-Schmidt. If f is trace class. n n ¨ (by Holder) Therefore gf is trace class if and only if n !f(en )!2 < ∞ and n !g † ◦ w(en )!2 < ∞. That is. any morphism C → C is .

One easily sees from the trace property that trace is the identity on scalars. y) = (y. and preserves daggers. Mackie. 1–28. 2010. editors). Therefore. OPERATIONAL THEORIES AND CATEGORICAL QUANTUM MECHANICS 119 trivially trace class. y) = (x. en ) m. then ( n n (h) is absolutely convergent and) Tr(h) = n n (h). its range is ker(h † )⊥ = ker(h)⊥ . Now suppose that f : H → H is trace class. Then certainly g ◦ f = 0 is trace class. Thus h : H → H restricts to a function h : ker(h)⊥ → ker(h)⊥ on a separable space. it is not true that if f : H → K and g : K → H are morphisms such that g ◦ f is trace class. By the previous lemma.n so that f ◦ g is not trace class. No-Cloning in Categorical Quantum Mechanics. Semantic Techniques in Quantum Computation (S. and hence that g = g † ◦ g = (f ◦ g)† ◦ (f ◦ g) ≥ 0. Lidskii’s formula still holds. then f ◦ g is trace class. and Tr(f ◦ g) = |f ◦ g|(em . For a counterexample. too. Cambridge University Press. is multiplicative on tensor products. trace class functions h : H → H on any (possibly nonseparable) Hilbert space H . x) and g(x. We have written the above example out in more detail than the reader might have thought necessary. let H = K = 2 (N). 0). If g : H → K and h : K → H are arbitrary morphisms. REFERENCES [1] S.n = em | em + 0 | en = dim(H ) = ∞.7]. 0). and deﬁne f(x. en ) | (em . To prove that Tr(g ◦f) = Tr(f ◦g) for f : H → K and g : K → H with both f ◦g and g ◦f trace class. that g = g † = g † ◦ g. we claim that for positive. m. Therefore |f ◦ g| = g. where n (h) are the eigenvalues counted up to algebraic multiplicity [58. because we may then replace g ◦ f and f ◦ g above by their absolute value. Since h is trace √ class. pp. But it is easy to see that f † (x. so that Tr(g ◦ f) = n n (g ◦ f) = n n (f ◦ g) = Tr(f ◦ g). Pick an orthonormal √ basis {ei } for H . we can write f = f2 ◦ f1 for Hilbert-Schmidt maps fi . Because h is positive. then g ◦ f2 and f1 ◦ h are again Hilbert-Schmidt. Hence ker(h)⊥ = ker( h)⊥ can only contain countably many ei . because it is easy to overlook subtleties. Finally. . g ◦ f ◦ h = (g ◦ f2 ) ◦ (f1 ◦ h) is trace class. g ◦ f and f But ◦ g have precisely the same spectrum. Thus trace class maps indeed form an endomorphism ideal. which ﬁnishes the proof that trace class operators form a trace ideal. For example. Gay and I. by the previous lemma again. y) = (0. Theorem 3. we rely on Lidskii’s trace formula for separable H : if h is trace class. i ei | h(ei ) = i ! h(ei )! is summable. Abramsky.

A categorical semantics of quantum protocols. Quantum Physics and Logic 2012. 175–186. pp. Clark. J. M. Lecture Notes in Mathematics. Topological quantum ﬁeld theories. Barrett. 2008. R. 12 (2010). p. vol. Journal of Pure and Applied Algebra. Leifer. [18] C. Awodey. B. New Journal of Physics. Oxford University Press. Barnum. Abramsky. no. 3–47. Synthese. 103–137. 101 (2013). vol. available as arXiv:1007. 143 (1999). Distributors at work. 2. vol. Wilke. [9] S.E. Blute. vol. Barrett. 752. arXiv:1108. S. 186(3) (2012). Publications Mathematiques de l’I. and R. Abramsky and A. P. 84 (2011). 13 (2011). Coecke. Physics. 1 (1964). DOI 10. On the Einstein-Podolsky-Rosen paradox. ˇ Hilbert space operators in quantum physics. Generalized no-broadcasting theorem. Proceedings of the 19th Annual IEEE Symposium on Logic in Computer Science.1007/s10992-013-9280-8. [13] H. Perinotti. pp.H. editors). 24. vol. [21] B. Wilce. D’Ariano. 240501. Nuclear and trace ideals in tensored *-categories. New Journal of Physics. 270 (2011). and A. Wilce. pp. 2014. Symmetry. Leifer. compact closure and dagger compactness for categories of convex operational models. 42 (2013). 3. Abramsky. [10] H. 68 (1988). pp. Coecke and B. Big toy models: Representing physical systems as Chu spaces. [6] S. Chu. Panangaden. 15–36. Arxiv preprint arXiv:0805. O. and A. N. Physical Review A. pp. Bell. 033024. Wilce. Category Theory. ´ [8] M. vol. 297–308. pp. Springer.pdf. pp. BRICS. (2008). second ed. Languages and Programming. 415– 425. [5] S. Coecke and C. . [20] P. [17] J. Electronic Notes in Theoretical Computer Science. Duncan.de/∼streicher/FIBR/DiWo. The sheaf-theoretic structure of non-locality and contextuality. B´enabou. F. IEEE. vol.. Studia Logica. ∗-autonomous categories. G. [7] . M. Mislove. 1. Technical Report LS-98-2. 2008. Teleportation in general probabilistic theories. 411–452. 99 (2007). In ∗-Autonomous Categories? [14]. Constructing ∗-autonomous categories. Brandenburger. Phase groups and the origin of non-locality for qubits. M. 113036. Informational derivation of quantum theory. no. 2000. Electronic Proceedings in Theoretical Computer Science. pp. Blank. [15] J. (2010). American Mathematical Society.1978. and R. 195– 200. Elsevier Science. Spekkens.-H. vol. The compositional structure of multipartite quantum entan- glement. Butz. L. 2. available at http://www. p. Edwards. Journal of Philosophical Logic. Handbook of Quantum Logic and Quantum Struc- tures: Quantum Logic. [11] H. no. [3] . Pictures of complete positivity in arbitrary dimension. [12] . [19] G. Springer. and P. Spekkens’ toy theory as a category of processes. and P. pp. [16] J. vol. R. tu-darmstadt. [22] B. [4] S. no. Abramsky and B. Abramsky and M. 2012. W. 012311. 501– 523. Stepanik. Coecke.mathematik. vol. Barnum. vol. Mathematical Foundations of Information Flow (S. 2004. 697–718. Exner. 261–324.120 SAMSON ABRAMSKY AND CHRIS HEUNEN [2] S. Chiribella. Automata. Havl´ıcek.2754. Barnum.S. [24] B. pp. Physical review letters. Kissinger. p. no.3553. pp. [23] B. Atiyah. Heunen. Barr. A. and M. Entropy and information causality in general probabilistic theories.. 158. Relational Hidden Variables and Non-Locality. vol. p. 1979. vol. 2010. Regular categories and regular logic. 1–14. [14] M. Edwards. pp. R. Spekkens. J. Coecke and A. October 1998. Categorical quantum mechanics.

Order. [35] D. Zeilinger. quantum correlations and commuting observables. [38] C. editors). Routledge and Kegan Paul. vol. [36] L. 57. 99–124. 1968. M. 1. 1131. New Journal of Physics. Doplicher and J. L. Foundations of Quantum Mechanics. Negative probability. Feynman. 45 (1949). Cambridge University Press. M. A new duality theory for compact groups. Mathematical Structures in Computer Science. Giry. vol. Verity. 2011. no. M. Deep Beauty: Understanding the Quantum World through Mathematical Innovation. vol. Roberts. Garling. 59. 177–212. 2010. Calude and Vladimiro Sassone. pp. Arxiv preprint quant-ph/0101012. [37] M. pp. Masanes and M. Proceedings of the Royal Society of London. Longo. 1963. Joyal. Jacobs. 1–19. [31] A. Springer. 1993. 2. Journal of Mathematical Physics. Coherence for compact closed categories. R. Joint distributions. vol. 180 (1942). Quantum Theory and Beyond : Is Entanglement Special?. 27 (2010). Street. 285–296. Quantum Theory: Concepts and Methods. Laplaza. [34] M. 531–533. Radboud University Nijmegen. J. E. P. Quantum theory from ﬁve reasonable axioms. Peres. American Journal of Physics. Cambridge University Press. and J. E. [41] J. Theoretical Computer Science (Cristian S. A remark on crossed product of C*-algebras. Lima. Horne. [30] R. vol. Journal of Pure and Applied Algebra. [50] J. 23 (1981). 79–109. Dirac. 365–392. 1. Journal of the London Mathematical Society (2). Studies in Advanced Mathematics. A categorical approach to probability theory. Mathematical Proceedings of the Cambridge Philosophical Society. pp. 193–213. 19 (2008). Heunen and B. Categorical Quantum Models and Logics. pp. and D. [28] S. Ludwig. [45] J. 2003. Springer-Verlag. Ph. 13 (2011). Fine. 063001. A. Stone Spaces. Frobenius Algebras and 2D Topological Quantum Field Theories. Springer. T. Inventiones mathematicae. Mathematical and Physical Sciences. Heunen. Berlin. 235–248. Quantum Chance and Non-Locality.D. 323. 120 (1985). Foundations of Quantum Mechanics. M. Cambridge University Press. Muller. IFIP Advances in Information and Communication Technology. and A. R. Quantum Implications: Essays in Honour of David Bohm (B. Shimony. Jauch. Johnstone. pp. duality and eﬀects. 157–218. W ∗ -categories. [43] A. vol. E. vol. Brukner. Moyal. vol. [48] G. Rewriting measurement-based quantum computations with generalised ﬂow. vol. 1983. [49] L. . Quantum logic in dagger kernel categories. Bell’s theorem without inequalities. H. Benjamin. Greenberger. Paciﬁc Journal of Mathematics. Hiley and F. [29] R. Categorical Aspects of Topology and Analysis. Inequalities. pp. Kock. OPERATIONAL THEORIES AND CATEGORICAL QUANTUM MECHANICS 121 [25] B. p. The physical interpretation of quantum mechanics. (2001). 1999. pp. 68–85. 1–40. vol. Hasegawa. Series A. 2007. 1306. Traced monoidal categories. Mathematical Foundations of Quantum Mechanics. Duncan and S. vol. D. On traced monoidal closed categories. 3. A. London Math- ematical Society Student Texts. 1982. M. 3 (1996). pp. (2010). ¨ A derivation of quantum theory from physical requirements. [47] G. 19 (1980). pp. 98 (1989). p. Cambridge University Press. Dickson. vol. Heidelberg. Kluwer. Convexity. vol. [46] R. Languages and Programming. Perdrix. Roberts. J. P. pp. pp. [42] P. thesis. vol. Addison-Wesley. [39] C. 58 (1990). 2009. Cambridge University Press. [32] D. 217–244. 1986. [51] A. [40] B. [44] G. 980. editors). Quantum mechanics as a statistical theory. no. [33] P. pp. M. 447–468. [27] P. Ghez. Dakic and C. Jacobs. 23 (1982). vol. Peat. 1987. Mackey. Automata. Mathematical Proceedings of the Cambridge Philosophical Society. p. no. A. vol. pp. Kelly and M. W. pp. Hardy. [26] W. no.

uk . vol. 749. 2000.122 SAMSON ABRAMSKY AND CHRIS HEUNEN [52] C. Piron. 3.ox. Foundations of Physics. Chu spaces from the representational viewpoint. [59] A. [57] P. [60] R. 032110. pp. Pitman Research Notes in Mathematics. 96 (1999). Mathematical surveys and monographs. vol. [54] V. 75 (2007). no. Quantales and their applications. London Mathematical Society Lecture Note Series. 3. Wigner. 139–163. On the quantum correction for thermodynamic equilibrium. Inc. 24 (1994). Simon. Rosenthal.ac. Physical Review. Cambridge University Press. pp. pp. vol. 170. MA. Plotkin. [58] B. 30–41. no. Evidence for the epistemic view of quantum states: A toy theory. 1-3. Topology. The deﬁnition of conformal ﬁeld theory. no. Simpson and G. 1979. 40 (1932). no. Physical Review A. Pratt. Rohrlich. Logic in Computer Science. [55] K. Annals of Pure and Applied Logic. vol. 120. [56] G. Selinger.abramsky@cs. no. Electronic Notes in Theoretical Computer Science. I. [53] S. Complete axioms for categorical ﬁxed-point operators. Longman Scientiﬁc & Technical. 5. [61] E. DEPARTMENT OF COMPUTER SCIENCE UNIVERSITY OF OXFORD WOLFSON BUILDING. Trace Ideals and Their Applications. 379–385. 1976. Geometry and Quantum Field Theory. 1990. 2004. 308. pp. Quantum Programming Languages. W. p.uk E-mail: heunen@cs. Spekkens. 2007. Quantum nonlocality as an axiom. vol. Elsevier. Segal. vol. 421–577. 319–333. Popescu and D. Reading.ox. Foundations of Quantum Physics. PARKS ROAD OXFORD OX1 3QD E-mail: samson. R. p.ac. pp.. Dagger compact closed categories and completely positive maps. WA Benjamin. American Mathematical Society.

A. Since the algebraic structure of these sets of operators is described in terms of modules over various semirings. via this dual adjunction we can put the work [11] on quantum weakest preconditions in perspective (see especially Remark 15). 45 c 2016. basically via dual operation V → V ∗ . eﬀect and projection operators. The article begins by describing the familiar sets of operators (bounded. and over non-negative real numbers R≥0 (for positive operators). like Theorem 14 that relates density operators and eﬀects via a dual adjunction between convex sets and eﬀect modules (extending earlier work [25]). and research in the foundations of quantum physics. Association for Symbolic Logic 123 . There is a recent exciting line of work connecting research in the semantics of programming languages and logic. such as self-adjoint. it is useful to have a uniform description of such modules. It is provided in Section 3. and maps between them. Of particular interest is the connection with quantum structures. It is in line with many other dual adjunctions and dualities that are relevant in programming logics [31. RELATING OPERATOR SPACES VIA ADJUNCTIONS BART JACOBS AND JORIK MANDEMAKER Abstract. via their description as Eilenberg-Moore algebras of a monad. namely over complex numbers C (for bounded operators). are expressed in terms of dual adjunctions. §1. over real numbers R (for self-adjoint operators). self- adjoint. via a dual adjunction between convex sets and eﬀect modules. This chapter uses categorical techniques to describe relations between various sets of operators on a Hilbert space. including quantum computation and logic. including various Hilbert-Schmidt isomorphisms of the form tr(A−). see [9] for an overview. Harizanov Lecture Notes in Logic. 1. positive) on a (ﬁnite-dimensional) Hilbert space in terms of functors to categories of modules. Introduction. These relations. This is to a large extent not more than a systematic presentation of known results and connections in the (modern) language of category theory. via the notion of algebra of Logic and Algebraic Structures in Quantum Computing Edited by J. However. positive. density. This paper ﬁts in that line of work. the approach leads to clarifying results. Indeed. Eskandarian and V. Chubb. The approach systematically uses categories of modules. The dual adjunctions involved are made explicit. 30]. It concentrates on operators (on Hilbert spaces) and organises and relates these operators according to their algebraic structure. see Section 2.

For instance. For a (ﬁnite-dimensional) Hilbert space H we shall study the following sets of operators H → H . E e Pr(H ) B(H ) o ? _ SA(H ) o ? _ Pos(H ) o ? _ Ef (H ) sk (1) 9Y DM(H ) where: Notation Description Structure B(H ) bounded/continuous linear vector space over C SA(H ) self-adjoint A† = A vector space over R Pos(H ) positive: A ≥ 0 module over R≥0 Ef (H ) eﬀect: 0 ≤ A ≤ I eﬀect module over [0.124 BART JACOBS AND JORIK MANDEMAKER a monad (namely the multiset monad). It describes the algebraic structure of the sets of operators that will be relevant here. see Table 1 for an overview. As is well-known. by taking H to be the trivial space C of complex . The main contribution of the paper thus lies in a systematic description. This duality formalises the diﬀerence between the approaches of Heisenberg (focusing on observables/eﬀects) and ¨ Schrodinger (focusing on states). It allows us to reconstruct all sets of operators on a Hilbert space from its projections. We borrow the probabilistic Gelfand duality between (Banach) eﬀect modules and (compact) convex sets from [29] for the ﬁnal steps in our analysis. For instance. [22]. eﬀects.1. This abstract description provides (co)limits and the monoidal closed structure of such algebras (from [33]) for free. Operator overview. We then use that convex sets can also be described as such algebras of a monad (namely the distribution monad). the set B(H ) of endomaps is not only a vector space over the complex numbers. but actually a C ∗ -algebra. It is not meant to capture all the structure that is present. and elaborate the connection with eﬀect modules (also known as convex eﬀect algebras. In this setting we discuss various ‘Gleason-style’ correspondences. see e. 1. see [36]). operators on Hilbert spaces behave in a certain sense as numbers. We should emphasise that the investigations in this paper concentrate on ﬁnite-dimensional Hilbert and vector spaces.g. between projections. and density matrices. 1] † Pr(H ) 2 projection: A = A = A orthomodular lattice DM(H ) density: A ≥ 0 and tr(A) = 1 convex set The emphasis lies on the ‘structure’ column.

We start by recalling that the category VectC of vector spaces over the complex numbers C carries an involution given by conjugation: for a vector space V we write V for the conjugate space. We refer to [6. This isomorphism (2) is a famous example of a non-natural mapping. The material in this section thus serves as preparation. It will focus on isomorphisms V ∼ = V ∗ . It is not new. 1] tj 8X {1} §2. Operators and duality. This section concentrates on the ﬁrst three sets of operators in (1). for V = B(H ). en . but with scalar multiplication given by z •V x = z •V x. because these functions j | form a ‘dual’ basis for V ∗ . A linear map f : V → W is sometimes called conjugate linear. . because it satisﬁes f(z • v) = z • f(v). in subsequent sections. where V is a vector space of operators on a Hilbert space. 1} Co ?_ Ro ?_ R≥0 o ? _ [0. . This yields an ‘involution’ endofunctor (−) : VectC → VectC which is the identity on morphisms. where the complex number z ∈ C has conjugate z ∈ C. If V is ﬁnite-dimensional. There is the standard correspondence between linear functions U → (V W ) and U ⊗ V → W . This serves as motivation for further investigation of the structures involved. capturing ‘states and statements’ in quantum logic. in categories of modules (or vector spaces). Pos(H ). One uses this exponent to form the dual space V ∗ = V C. 27] for more information on involutions in a categorical setting. except possibly for the presentation in terms of maps of adjunctions. where the ‘bra’ j | :V → C sends a vector w = ( k wk | k ) to its j-th coordinate j | w = k wk j | k = wj . It will play a crucial role below. 16. . RELATING OPERATOR SPACES VIA ADJUNCTIONS 125 numbers. say with a basis e1 . We shall write V W for the ‘exponent’ vector space of linear maps V → W between vector spaces V and W . These isomorphisms turn out to be natural in H . Clearly. the diagram (1) becomes: F f {0. namely on B(H ) ← SA(H ) ← Pos(H ). depending on a choice of basis. Complex ∼ conjugation z → z is an example of a conjugate linear (isomorphism) C −→ = C in VectC . Only later will we study the density and eﬀect operators DM(H ) and Ef (H ). written in ‘ket’ notation as | j = ej . . with the same vectors as V . . SA(H ). there is the familiar isomorphism of V with its dual space V ∗ = V C given as: ∼ V = / V∗ ∼ = V C ! ! (2) zj | j / j j zj j | . this yields an isomorphism ∼ V → = V ∗ .

. (3) T The operator C † = C is the conjugate transpose of C . we write B(H ) for the homset of endomaps H → H in FdHilb. it is adjoint to itself. forming an adjunction as on the right. as basis—assuming a basis | 1 . This trace tr satisﬁes the following . an operator A : H → H single 1 in the j-th row of the k-th can be written as matrix A = j. written as B(C ). As is usual. . It makes FdHilb into a dagger category. The mapping H → B(H ) is functorial. it involves the trace operation tr : B(H ) → C of which we ﬁrst recallsome basic facts. of dimension n 2 with the outer products | j k |. and will be used here as functor B : FdHilb → VectC . let FdHilb be the category of ﬁnite-dimensional Hilbert spaces with bounded linear maps between them. as in: f (−)† V −−→ W + ========= FdHilb k ⊥ FdHilbop W −−→ V f† (−)† In the next result we apply the duality isomorphism V ∼ = V ∗ in (2) for V = B(H ). . One can drop the bounded- ness requirement. This deﬁnition is independent of the choice of matrix/basis. V −→ (W C) ============= (−)∗ =(−)C V ⊗ W −→ C ∼=C + =============== VectC k ⊥ (VectC ) op V ⊗W ∼ = V ⊗ W −→ C ============= (−)∗ =(−)C W −→ (V C) In the next step.126 BART JACOBS AND JORIK MANDEMAKER The mapping V → V ∗ = (V C) yields a functor VectC → (VectC ) . k ≤ n. continuous). This functor (−)∗ is adjoint to itself. . for j. satisfying Cv | w = v | C † w. | n for H . because a linear map between ﬁnite-dimensional spaces is automatically bounded (i.e.g. see e. For A ∈ B(H ) the trace tr(A) can be deﬁned as the sum j Ajj of the diagonal matrix values. On a map C : H → K it yields a linear function B(H ) → B(K). where the matrix entries Ajk may be described as j |A| k . In general. Also. This dagger forms an involution on the vector space B(H ). This set B(H ) of “operators on H ” is a vector space over C. for op a map C : H → K we have C ∗ : K ∗ → H ∗ given by f → f ◦ C . in the sense that there is a bijective correspondence (suggested by the double lines) as on the left below.k Ajk | j k |. and given by: ' ( A C† A C B(C ) H −→ H = K −→ H −→ H −→ K . As we shall see. [2]. Such outer product projections | j k | may be understood as the matrix with only 0s except for a column.

RELATING OPERATOR SPACES VIA ADJUNCTIONS 127 basic properties. · · · describes the function B → · · · . We drop the subscript when confusion is unlikely. this hsB is part of a map of adjunctions (see [34. The subscript B is added because we shall encounter analogues of this isomorphism for other operators. tr(AB † ). tr(A + B) = tr(A) + tr(B) tr(zA) = z tr(A) where z ∈ C tr(AB) = tr(BA) the so-called cyclic property T tr(A ) = tr(A) where (−)T is the transpose operation tr(A† ) = tr(A) which results from previous points tr(A) ≥ 0 when A is positive: A ≥ 0. (−)† - FdHilb m ⊥ FdHilbop † (−) B B (−)∗ .e. Vect (−)∗ C Moreover. B) → tr(AB † ) = hsB (A)(B) is commonly named after them. IV. namely: hsB B(H ) / B(H )∗ = B(H ) C is hsB (A)= B.2 Vect op C B . v | Av ≥ 0. For a ﬁnite-dimensional Hilbert space H the duality isomor- phism (2) applied to the vector space B(H ) of endomaps boils down to a trace calculation. ∼ This map B(H ) −→ = B(H )∗ is independent of the choice of basis. ∼ = + op Pictorially. since the inner product (A. . and used for function abstraction: B. it yields a natural isomorphism involving adjoint (−)† and dual (−∗ = (−) C in: hsB B ◦ (−)† +3 (−)∗ ◦ B. i. Proposition 1. ∼ = where the -notation is borrowed from the -calculus. More categorically. this hsB is a natural transformation FdHilb ∼ = 3 VectC between the two functors FdHilb ⇒ Vectop C given in: (−)† 1 FdHilbop B FdHilb . op VectC m ⊥ VectC (−)∗ The letters ‘h’ and ‘s’ in the map hsB stand for Hilbert and Schmidt.7]) in the following situation.

. because = they involve a trace calculation that is base-independent.128 BART JACOBS AND JORIK MANDEMAKER Proof. see e. Deﬁne an operator hs −1 (f) ∈ B(H ) with matrix entries: ! hs −1 (f) jk = f(| j k |). for each map C : H → K in FdHilb. hs H B(H / B(H )∗ = B(H ) C O ) ∼ = O B(C † ) B(C )∗ hs K B(K) / B(K)∗ = B(K ) C ∼ = This diagram commutes because: (B(C )∗ ◦ hs K )(A)(B) = (hs K (A) ◦ B(C ))(B) = hs K (A)(B(C )(B)) = hs K (A)(CBC † ) = tr(A(CBC † )† ) = tr(ACB † C † ) = tr(C † ACB † ) by the cyclic property † = hs H (C AC )(B) = hs H (B(C † )(A))(B) = (hs H ◦ B(C † ))(A)(B). So suppose we have a linear map f : B(H ) → C. | n is a basis for H . . Naturality amounts to commutation of the following diagram.k Ajk Bjk = B.11]. Since the trace of a matrix is basis-independent. tr(AB † ). . If | 1 . because these hs H ’s are natural in H and componentwise isomorphisms. the inverses hs −1 H are also natural in H . tr(A−) : B(H ) −→ = B(H )∗ = (B(H ) C). Lemma 7. so is this isomorphism hs. Finally. j. we use the basic fact that.g. . j. .k Ajk (B † )kj j. | n for H .k = B. . . [4. ∼ The remarkable thing about this result is that whereas the maps V −→ = V∗ ∼ in (2) are not natural. We brieﬂy describe the ∼ inverse of hs = A. j (AB † )jj = B. (4) . the instantiations hs : B(H ) −→ B(H )∗ are. via a choice of basis | 1 . A= Ajk | j k | −→ B. according to (2). The details of the map of adjunctions in the above diagram are left to the interested reader. . then the map hs : B(H ) → B(H )∗ becomes.

by: ' ( ' ( A C† A C SA(C ) H −→ H = K −→ H −→ H −→ K . this mapping f → Af is independent of a choice of basis.k = f(Bjk | j k |). and so is the trace (as sum of these Ajj ).k = f(| j k |)Bjk j. Again. RELATING OPERATOR SPACES VIA ADJUNCTIONS 129 Then we recover f via the trace calculation: ! ! hs hs −1 (f) (B) = tr hs −1 (f)B † −1 † = tr hs (f)jk | j k |)B j. The mapping H → SA(H ) can be extended to a functor SA : Hilb → VectR . (5) . because f is conjugate linear j. We now restrict ourselves to self-adjoint operators SA(H ) → B(H ).k ! = hs −1 (f)jk tr | j k |B † j.k = f(B). Self-adjoint operators. We recall that an operator A : H → H is called self-adjoint (or Hermitian) if A† = A. In particular.k ! = f(| j k |) tr (B † )kj j. like for B in (3).k ! = f(| j k |) tr k |B † | j j. For instance. In terms of matrices this means that Ajk = Akj . for each complex number z ∈ C and B ∈ B(H ) we have self-adjoint operators: zB + zB † and izB − izB † . all entries Ajj on the diagonal are real numbers. since: !† !† SA(C )(A) = CAC † = C †† A† C † = CAC † = SA(C )(A). There are serveral ways to turn a linear operator into a self-adjoint one.k =f Bjk | j k | j. This is well-deﬁned since if A is self-adjoint then so is SA(C )(A). because its inverse A → tr(A−) does not depend on such a choice. The set of self-adjoint operators SA(H ) forms a vector space over R.

If A. If the real part Re(z) is non-zero.130 BART JACOBS AND JORIK MANDEMAKER In this way we obtain mappings B(H ) → SA(H ) in VectR . making it a split mono. the subset SA(H ) → B(H ) of self-adjoint operators on H is a vector space over R. By Proposition 1 there is a unique A ∈ B(H ) with: f = hsB (A) = tr(A(−)† ) : B(H ) −→ C. Conversely. By moving from B to SA we get the following analogue of Proposition 1. We now put hsSA −1 (f) = 12 (A + A† ) ∈ SA(H ). (6) ∼ = It gives rise to a map of adjunctions: (−)† - FdHilb m ⊥ FdHilbop † (−) SA SA (−)R . as described in (5). VectR m ⊥ VectR op (−)R Proof. B : H → H are self adjoint operators. then tr(AB † ) = tr(AB) is a real number. that B + B † and iB − iB † are self-adjoint. and check for B ∈ SA(H ): ! hsSA hsSA −1 (f) (B) ! = tr hsSA −1 (f)B 1 ! = tr(AB) + tr(A† B) 2 . It is not hard to see that f preserves sums of operators and satisﬁes f (zB) = zf (B). Proposition 2. the mapping: 1 ! B −→ zB + zB † 2Re(z) is a left-inverse of the inclusion a SA(H ) → B(H ). since: ! ! tr(AB) = tr (AB)† = tr B † A† = tr (BA) = tr(AB). For H ∈ FdHilb. This f really extends f since ! in the special case when B is self-adjoint we get f (B) = 12 f(2B) + if(0) = f(B) by linearity. for which one obtains a natural isomorphism in VectR : hsSA SA(H ) / SA(H )∗ = SA(H ) R by hsSA (A)(B) = tr(AB). suppose we have a (linear) map f : SA(H ) → R in VectR . It can be extended to a function f : B(H ) → C via 1 ! f (B) = f(B + B † ) + if(iB − iB † ) 2 using.

In that case . for each x ∈ H . RELATING OPERATOR SPACES VIA ADJUNCTIONS 131 1 ! = tr(AB † ) + tr((BA)† ) since B is self-adjoint 2 1 ! = f (B) + tr(BA) 2 1 ! = f(B) + tr(AB) since f(B) = f (B) when B ∈ SA(H ) 2 1 ! = f(B) + f(B) 2 1 ! = f(B) + f(B) because f(B) is real valued 2 = f(B) since f is linear. We prove uniqueness in the self-adjoint case too. ! In the other direction. an arbitrary basis. Positive operators. ! Similarly. the proof gives a formula for the inverse hsSA −1 of the Hilbert- Schmidt map for self-adjoint operators. . We prove the equality Ajk = Cjk in two steps. An operator A : H → H is called positive if the inner product Ax | x is a non-negative real number. by writing Im(Cjk ) = 1 2 −iCjk + iCjk and using the self-adjoint operator −i| j k | + i| k j |. by proving that both their real and imaginary parts are the same. Implicitly. one obtains hsSA −1 hsSA (A) = A by uniqueness. We plan to show Ajk = Cjk wrt. using that | k j | + | j k | is self-adjoint = . (as before) = Re(Ajk ). We need to † prove C = A = 12 (A + A ). and thus A = C . Im(Cjk ) = Im(Ajk ). 1 ! Re(Cjk ) = Cjk + Cjk 2 1 ! = Cjk + (C † )kj 2 1 ! = Cjk + Ckj 2 1 = ( j |C | k + k |C | j ) 2 1 ! = tr( j |C | k ) + tr( k |C | j ) 2 1 ! = tr(C (| k j | + | j k |)) 2 1 ! = tr(A(| k j | + | j k |)) 2 by assumption. Assume a self-adjoint operator C ∈ SA(H ) also satisﬁes f = hsSA (C ) : SA(H ) → R. .

so that indeed tr(A−) has type Pos(H ) → R≥0 . B. Hence the spaces Pos(H ) → SA(H ) → B(H ) are actually ordered (see also [37. then Pos(C )(A) = CAC † ≥ 0 since for each x ∈ K . In a spectral decomposition A = j j | j j | a positive operator A has eigenvalues j ∈ R≥0 for all j.132 BART JACOBS AND JORIK MANDEMAKER one writes A ≥ 0. Mod R≥0 m ⊥ Mod R≥0 op (−)R≥0 Proof. We do so by ﬁrst writing the spectral decomposition as A = j j | j j |. We ﬁrst have to check that tr(AB) ≥ 0. since A† = (BB † )† = B †† B † = BB † = A. Thus there are inclusion maps Pos(H ) → SA(H ) → B(H ). We can describe taking positive operators as a functor Pos : Hilb → Mod R≥0 from Hilbert spaces to modules over the non-negative real numbers. Proposition 3. and also to: all eigenvalues are non-negative reals. for which there is a natural isomorphism in Mod R≥0 : hsPos Pos(H ) / Pos(H )∗ = Pos(H ) R≥0 by hsPos (A)(B) = tr(AB). 15]). since Bj | j ≥ 0. CAC † x | x = AC † x | C † x ≥ 0. The action of Pos on maps is like for SA and B in (3). (7) ∼ = This isomorphism gives rise to a map of adjunctions: (−)† - FdHilb m ⊥ FdHilbop † (−) Pos Pos (−)R≥0 . B ∈ Pos(H ). the subset Pos(H ) → SA(H ) of positive operators is a module over the non-negative reals R≥0 . This is equivalent to: A = BB † . and is well-deﬁned. A + P = B. Then: tr(AB) = j tr(| j j |B) = j tr( j |B| j ) = j tr(Bj | j) j j j = j Bj | j j ≥ 0. It forms a module over the semiring R≥0 of non-negative reals since positive operators are closed under addition and under scalar multiplication with r ∈ R≥0 . for some operator B. with j ≥ 0. Hence the trace tr(A) is a non-negative real number. for A. . As an aside we recall that via positivity one obtains the Lowner ¨ order on arbitrary operators A. since if C : H → K in Hilb and A ≥ 0. A positive operator is clearly self-adjoint. Thus: A ≤ B iﬀ ∃P ∈ Pos(H ). deﬁned as: A ≤ B iﬀ B − A ≥ 0. The set of positive operators on H is written here as Pos(H ). For H ∈ FdHilb.

Projection operators of the form | j j | are positive. But this is impossible. For a positive operator B ∈ Pos(H ) we then get tr(AB) = f (B) = f(B) ≥ 0. and thus: f (B) = f(Bp ) − f(Bn ) = f(C ) − f(D). This concludes our description of the spaces of operators B(H ) ← SA(H ) ← Pos(H ) on a (ﬁnite-dimensional) Hilbert space H . Then take: Bp = j | j j | and Bn = −j | j j |. and A = Ap is a positive operator. But then C = A by the uniqueness from Proposition 2. as required. D ∈ Pos(H ) also satisfy B = C − D. Hence by Proposition 2 there is a unique A = hsSA −1 (f ) ∈ SA(H ) with f = hsSA (A) = tr(A−) : SA(H ) → R. RELATING OPERATOR SPACES VIA ADJUNCTIONS 133 These maps hsPos = tr(A−) clearly preserve the module structure: additions and scalar multiplication (with a non-negative real number). Like before. One way to do it is to write B = j j | j j | as spectral decomposition. assume we have a linear map f : Pos(H ) → R≥0 in Mod R≥0 . we wish to extend it. so that by linearity: f(Bp ) + f(D) = f(Bp + D) = f(C + Bn ) = f(C ) + f(Bn ). and to separate the (real-valued) eigenvalues j into negative and non-negative ones. since B = Bp for such a positive B. Before we proceed to density operators DM(H ) and eﬀects Ef (H ) . This outcome is independent of the choice of Bp . Bn . then Bp + D = C + Bn . as naturally self-dual modules. then for an arbitrary B ∈ SA(H ). We now write A = Ap − An as in (8). this time to a map f : SA(H ) R. where An = j <0 −j | j j |. since we assumed j < 0. Hence An = 0. so that we get for each j with j < 0 0 ≤ tr(A| j j |) = tr(Ap | j j |) − tr(An | j j |) = 0 − (−j ) = j . Thus we have f = tr(A−) : Pos(H ) → R≥0 . Bn ∈ Pos(H ). so that we can take hsPos −1 (f) = hsSA −1 (f ). since if C. f (B) = f(Bp ) − f(Bn ) = tr(CBp ) − tr(CBn ) = tr(C (Bp − Bn )) = tr(CB). If we have an arbitrary self-adjoint operator B ∈ SA(H ) we can write it as diﬀerence B = Bp − Bn of its positive and negative parts Bp . It is not hard to see that the resulting function f : SA(H ) → R is linear (in VectR ). (8) j ≥0 j <0 Now we can deﬁne f (B) = f(Bp )−f(Bn ) ∈ R. We brieﬂy check uniqueness: if C ∈ Pos(H ) also satisﬁes f = tr(C −). Next.

35. §3. As we now proceed more systematically. SA(H ). These formal sum are quotiented by the usual commutativity and associativity relations. where · is multiplication in S. so that s1 | x + s2 | x is considered to be the same as (s1 + s2 )| x . over a semiring S. with scalar multiplication by elements of the semiring. We recall that a semiring [20] is like a ring but without an additive inverse. R. where multiplication distributes over addition.g. when we talked about positive operators forming a module over the non-negative reals R≥0 . 0) and a multiplicative monoid (S. C giving us a uniform description of the categories of modules in which the spaces of operators Pos(H ). We shall be especially interested in the examples S = R≥0 . 7]—gives us certain structure for free. . In this section we thus start with the standard description of categories of modules. B(H ) on a Hilbert space H live. the same element x ∈ X may be counted multiple times. . because it distinguishes x as element of X and as vector in MS (X ). and multiplication : M S (M S (X )) → M S (X ) is ( i si | ϕi )(x) = i si · ϕi (x). whose unit : X → M S (X ) is (x) = 1| x . . B(H ) via free constructions between categories of modules. This multiset functor is a monad. relates the three spaces of operators Pos(H ). Modules are vector spaces except that the scalars need only be a ring. Also. [34. R. and R≥0 ) in terms of algebras of a monad. . +. we shall see that such a module over a semiring consists of a commutative monoid of vectors. For a function f : X → Y one deﬁnes MS (f) : MS (X ) → MS (Y ) by: MS (f)(ϕ)(y)= x∈f −1 (y) ϕ(x). It will be captured as algebra of the multiset monad. With this formal sum notation one can write the application of MS on a map f as MS (f)( i si | xi ) = i si | f(xi ) . Theorem 6. The ket notation | xi is justiﬁed because these elements are vectors. In fact we have already done in the previous section. . where supp(ϕ) = {x ∈ X | ϕ(x) = 0} is the support of ϕ. and useful. namely of the multiset monad MS associated with S. 5. Here we generalise further and will also consider modules over a semiring. One can deﬁne a “multiset” functor MS : Sets → Sets by: MS (X ) = {ϕ : X → S | supp(ϕ) is ﬁnite}. as categories of algebras of a monad. consisting of a commutative additive monoid (S. Categories of modules as algebras. (9) Such a (ﬁnite) multiset ϕ ∈ Ms (X ) may be written as formal sum s1 | x1 + · · · + sk | xk where supp(ϕ) = {x1 . 1). To start. see Theorem 4 below. let S be a semiring. The general theory of monads—see e. ·. SA(H ). The main result in this section.134 BART JACOBS AND JORIK MANDEMAKER on H we wish to explore and exploit the similarities between these modules (over C. and not a ﬁeld. xk } and si = ϕ(xi ) ∈ S describes the “multiplicity” of the element xi .

Cocompleteness always holds for algebras over Sets and follows from a result of Linton’s. This is less natural. which is both complete and cocomplete. with limits as in A. [35. but with a few subsequent pointers. for the semiring S. where the free algebra functor F : A → Alg(T ) preserves the monoidal structure (i. 1} yields the ﬁnite powerset monad Pﬁn = M2 . see e. The monad MN is also known as the ‘bag’ monad. When S happens to be a ﬁeld. containing ordinary (N-valued) multisets. Thus the category of algebras Alg(MS ) is equivalent to the category Mod S of S-modules. is strong monoidal). one may call MS the ﬁnitary multiset monad. The tensor unit . ⊗) in categories of algebras goes back to [33] (see also [23]). § 9. R. or C. such an S-module structure on acommutative monoid M yields an algebra MS (M ) → M by i si | xi → i si • xi .g. (c) symmetric monoidal closed in case these colimits exist and the monad T is symmetric monoidal (commutative).3. namely colimits of reﬂexive pairs. which is stated without proof. namely: Alg(MR≥0 ) = Mod R≥0 Alg(MR ) = Mod R = VectR Alg(MC ) = Mod C = VectC . This makes X a module. Thus we have a uniform description of the three categories of relevance in the previous section. For the semiring S = N one gets the free commutative monoid MN (X ) on a set X . We shall mostly use this result for A = Sets. Here we shall be mostly interested in the cases where S is R≥0 . RELATING OPERATOR SPACES VIA ADJUNCTIONS 135 In order to emphasise that elements of MS (X ) are ﬁnite multisets. see [5. Theorem 4. The Boolean semiring 2 = {0. 4] using the existence of coequalisers of reﬂexive pairs in Sets. A category of algebras is always “as complete” as its underlying category. We continue this section with a basic result in the theory of monads. The category Alg(T ) of algebras is: (a) also complete. 5]. Prop. Monoidal structure (I.e. In order to include non-ﬁnite multisets. It preserves the additive structure (of S and of X ) in each coordinate separately. and let T : A → A be a monad on A. (b) cocomplete as soon as certain special colimits exist in Alg(T ). so that we don’t have to worry about these special colimits. An (Eilenberg-Moore) algebra α : MS (X ) → X for the multiset monad corresponds to a monoid structure on X —given by x + y = α(1| x + 1| y )— together with a scalar multiplication • : S × X → X given by s • x = α(s| x ). If S = Z one obtains the free Abelian group MZ (X ) on X . Let A be a symmetric monoidal category. this category Mod S is the category VectS of vector spaces over S. one has to assume that suitable inﬁnite sums exist in the underlying semiring S. the monoidal structure on the underlying category Sets is thus cartesian. Conversely.

The associated exponent is written as . for the free algebra functor F : Sets → Alg(T ) and the ﬁnal (singleton) set 1.g. The free functor preserves these tensors. In summary. gives! rise to a map of monads MS ⇒ MS . A homomorphism of semirings f : S → S . This follows from the following general result. see e.136 BART JACOBS AND JORIK MANDEMAKER I is simply F (1). 29]. Algebra maps X ⊗ Y → Z then correspond to bi-homomorphisms UX × UY → UZ. [23. N ∈ Mod S there are obvious correspondences: M / (N S) ================ M ⊗N /S ================== N ⊗M /S ================ N / (M S) This means that there are adjunctions: (−)S + Mod S k ⊥ (Mod S ) op (10) (−)S as used in the previous section. Maps M ⊗ N → K in Mod S correspond to bilinear maps M × N → K (linear in each argument separately). In particular. Thus. like before. there is a universal bi-homomorphism ⊗ : UX × UY → U (X ⊗ Y ). For modules M. and thus to a functor Alg(MS ) → Alg(MS ). with S ∼= MS (1) as tensor unit. we have a sequence of categories of algebras of monads: Alg(MR≥0 ) o Alg(MR ) o Alg(MC ) (11) Mod R≥0 VectR VectC where the maps between them can be understood as arising from maps of monads in the other direction: MR≥0 +3 MR +3 MC via semiring inclusions R≥0 /R / C. In that case categories Mod S are monoidal closed. The multiset monad MS is symmetric monoidal if S is a (multiplicatively) commutative semiring. The tensor ⊗ is obtained as a suitable coequaliser of algebras. by (MS (X ) → X ) −→ (MS (X ) → MS (X ) → X ). by j sj | x j → j f(sj )| xj . preserving both the additive and multiplicative monoid ! structures. Proposition 5. This functor always has a left adjoint. . The left adjoint exists because categories of modules Alg(MS ) = Mod S are cocomplete. modules over their semirings have the structure of a biﬁbration [24]. it can be constructed via a coequaliser.

x2 ) ∼ (y1 . Pos(H ) on a Hilbert space H turn out to be related via free constructions. It is surjective since each A ∈ SA(H ) can be written as A = Ap − An for Ap . y2 ) ⇐⇒ ∃z. The proof uses explicit constructions of the left adjoints R and C in (12). and also in [8]. Thus A = ϕ([Ap . x2 ]+[y1 . An ]). x2 ) = (a • x1 − b • x2 . R C Proof. Similarly. The inclusion morphism Pos(H ) → SA(H ) in Mod R≥0 yields as transpose the map ϕ : R(Pos(H )) → SA(H ) in VectR given by ϕ([B1 . and scalar multiplication • : R × R(X ) → R(X ) via: [r • x1 . A module X over R≥0 can be turned into a vector space over R via the same construction that turns a commutative monoid into a commutative group R(X ) = (X × X )/ ∼ where (x1 . x2 +y2 ]. b • x1 + a • x2 ). Thus we have the following situation of triangles commuting up-to-isomorphism. Theorem 6. B2 ) = B1 + iB2 . The additive structure is obtained pointwise. x1 ]. Mod R≥0 (12) For a ﬁnite-dimensional Hilbert space H . the canonical inclusion morphisms Pos(H ) → SA(H ) in Mod R≥0 . given by (B1 . RELATING OPERATOR SPACES VIA ADJUNCTIONS 137 The diﬀerent spaces of operators B(H ).) A vector space X over R can be turned into a vector space over C. x2 ] = [(−r) • x2 . FdHilb Pos B SA y free $ Mod R≥0 / VectR free / VectC . and scalar multiplication • : C × C(X ) → C(X ) is done as follows. Addition is done componentwise: [x1 . and SA(H ) → B(H ) in VectR yield via these adjunctions (transposed ) maps that turn out to be isomorphisms: ! ∼= / ! ∼ = / R Pos(H ) SA(H ) and C SA(H ) B(H ). SA(H ). the inclusion SA(H ) → B(H ) in VectR gives rise to a transpose : C(SA(H )) → B(H ) in VectC . B2 ]) = B1 − B2 . r • x2 ] if r ≥ 0 r • [x1 . This was used implicitly in the proofs of Propositions 2 and 3 in the previous section. (a + ib) • (x1 . x1 + y2 + z = y1 + x2 + z. y2 ] = [x1 +y1 . simply via C(X ) = X × X . An ∈ Pos(H ) as in (8). Also . (−r) • x1 ] if r < 0 (Notice the reversal of the xi in the second case. x2 ] = [x2 . Write the left adjoints to the two forgetful functors in (11) as: R / VectR C / VectC . minus by reversal: −[x1 .

A positive operator A ∈ Pos(H ) is an eﬀect if and only if all of its eigenvalues are in [0. namely convex sets and eﬀect modules.138 BART JACOBS AND JORIK MANDEMAKER this map is surjective since each A ∈ B(H ) can be written as A = 12 (A + A† ) + 12 i(−iA + iA† ). We start with convex sets. where I is the identity map H → H and ≤ is the Lowner ¨ order (described before Proposition 3). Lemma 7. since I − A = A⊥ is positive. The spaces DM(H ) of density operators (states) and Ef (H ) of eﬀects (statements/predicates) require more subtle structures that will be introduced in this section. assume a positive operator A with spectral decomposition A = j | j j | where the | j form an orthonormal basis and j ∈ [0. and that there exists a map of adjunctions from Hilbert spaces. given as: Pr(H ) = {A ∈ B(H ) | A† = A = AA}. This shows A ≤ I . Then: j A = j j | j j | ≤ j | j j | = I . §4.1. We show that they are related by a dual adjunction. A further subset of Ef (H ) is the set of projections. Hence A = ( 12 (A + A† ). elaborated in Remark 15). they give a “quantum expectation value” tr(AB) for a density matrix B (see the map hsEf in Theorem 14 below. Convex sets and eﬀect modules. SA(H ). Conversely. Before we investigate the algebraic structure of these sets of operators we brieﬂy mention the following alternative formulation of eﬀects. In the previous section we have seen how the spaces of operators Pos(H ). where these eﬀects A are called predicates. 1]. 4. and (conveniently) describe them via a monad. Suppose A is an eﬀect with spectral decomposition A = j j |jj|. B(H ) ﬁt in the context of modules. They satisfy: j = j j | j = j |j | j = j |A| j ≤ j |I | j = j | j = 1. 1]. so that we can beneﬁt from general results like in Theo- rem 4. For a projection A ∈ Pr(H ) there is an orthosupplement A⊥ ∈ Pr(H ) with A + A⊥ = I . where we may assume that the eigenvectors | j form an orthonormal basis. Proof. where A + A† and −iA + A† are self-adjoints. The eigenvalues j are necessarily real and positive. DM(H ) = {A ∈ Pos(H ) | tr(A) = 1} Ef (H ) = {A ∈ Pos(H ) | A ≤ I }. First we recall the deﬁnition of the two sets of operators that are relevant in this section. Convex sets. 12 (−iA + iA† )). It is used for instance in [11]. Analogously to the multiset monad one deﬁnes the distribution monad . like in Section 2.

This multiplication is well-deﬁned since: si ϕi (x) = si · ϕi (x) = si · ϕi (x) = si = 1. if A ∈ DM(H ). The three example categories Alg(MS ) of interest here—that arise from multiset monads MS for S = R≥0 . 1] | supp(ϕ) is ﬁnite and x∈X ϕ(x) = 1}. Aﬃne maps preserve such interpretations of convex combinations. 1}). 13. see [26]. Taking density operators yields a functor DM : FdHilbUn → Conv = Alg(D). We recall that in the present context all such convex combinations involve only ﬁnitely many elements xj . for each r ∈ [0. the set DM(H ) of density operators is convex: given ﬁnitely many Aj ∈ DM(H ) and rj ∈ [0. since: tr(A) = tr rj Aj = rj tr(Aj ) = rj = 1. The category Alg(D) of algebras of the monad D is the category Conv of convex sets with aﬃne maps between them. x i x i i x i The distribution monad D is always symmetric monoidal (commutative). 1] ∼ = D({0. Here it is deﬁned for probabilities in the unit interval [0. y ∈ X there is an interpretation of the convex sum rx + (r − 1)y. It thus consists of a set X in which there is an interpretation a( j sj | xj ) ∈ X for each formal convex combination j sj | xj ∈ D(X ). it is a free one since [0. the operator A = j rj Aj is positive and has trace 1. j j j Moreover. since: ! tr DM(U )(A) = tr(UAU † ) = tr(U † UA) = tr(IA) = tr(A) = 1. UU = I and (thus) U † U = I . † so that U † = U −1 —then DM(U )(A) = UAU † : K → K is in DM(K ). but the more general structure of an “eﬀect monoid” may be used instead. see also [32. 1] satisfy i si = 1. As is well-known. C—are diﬀerent from the category Alg(D) of convex sets in at least three aspects: . (13) Elements of D(X ) are convex combinations s1 | x1 + · · · + sk | xk . The following result goes back to [39]. RELATING OPERATOR SPACES VIA ADJUNCTIONS 139 D : Sets → Sets as: D(X )= {ϕ : X → [0.e. In particular. Theorem 8. Proof. 25]. 1] with j rj = 1. 1]. namely as a(r| x + (1 − r)| y ) ∈ X . Here we shall identify such a convex set simply with an algebra a : D(X ) → X of the monad D. 1] and x. Unit and multiplication making D a monad can be deﬁned as for the multiset monad MS . R. if U : H → K is unitary—i. where the probabilities si ∈ [0. The unit interval [0. 1] of real numbers is an obvious example of a convex set. Let FdHilbUn be the category of ﬁnite-dimensional Hilbert spaces with unitary maps between them. Actually. Lemma 9.

A) → rA. • They have biproducts. that gives rise to an inclusion functor Mod R≥0 = Alg(MR≥0 ) → Alg(D) = Conv. y) = (s + t. But ﬁrst we conclude this part on convex sets with an observation like in Theorem 6. saying that modules over non-negative reals are convex sets—in a trivial manner. [0. where the operator tr(B) has trace 1 by construction. transposing the in- ∼ clusion DM(H ) → Pos(H ) in Conv gives an isomorphism S(DM(H )) −→ = Pos(H ) in Mod R≥0 . with addition for u. The mapping X → Conv(X. and thus A = B. This makes S(X ) a module over R≥0 . B ∈ DM(H ). since if rA = sB for A. v ∈ S(X ). since D(1) ∼ = 1. tr(B) . that can be described explicitly in terms of a representation contruction that goes back to [38] (see also [25]). see [10]. There is an obvious map of monads D ⇒ MR≥0 .140 BART JACOBS AND JORIK MANDEMAKER • These categories Alg(MS ) are dually self-adjoint via the functor (−) S as in (10). x) if u = (t. x) and s = 0. for X a convex set. 1]). this functor has a left adjoint. For general reasons. x) + (t. then r = r · tr(A) = tr(rA) = tr(sB) = s · tr(B) = s. but does lead to an interesting dual adjunction with a category of ‘eﬀect modules’. so that tensors ⊗ in Conv have projections (see [23]). This will be the described in the next subsection. For a ﬁnite-dimensional Hilbert space H . does not yield an adjunction as in (10). s +t s +t A scalar multiplication • : R≥0 × S(X ) → S(X ) is deﬁned as: 0 if u = 0 or s = 0 s •u= (s · t. In this way one obtains a triangle commuting up-to- isomorphism: FdHilbUn DM Pos | free $ Alg(D) = Conv / Mod R = Alg(MR ) ≥0 ≥0 S ϕ Proof. in trivial cases given by u + 0 = u = 0 + u and: s t (s. . because the monads MS are ‘additive’. It is injective. This left adjoint S : Conv → Mod R≥0 is given on X ∈ Conv by: S(X ) = {0} + R>0 × X. x+ y). Theorem 10. • The tensor unit in Alg(D) = Conv is the singleton set 1. The induced map S(DM(H )) = {0} + R≥0 × DM(H ) −→ Pos(H ) is given by 0 → 0 and (r. It is also surjective: ! since each!non-zero B ∈ Pos(H ) can be B B B written as B = tr(B) tr(B) = ϕ tr(B).

SA(H ). 1]). which is commutative and associative in a suitable sense. A partial commutative monoid (PCM) consists of a set M with a zero element 0 ∈ M and a partial binary operation : M × M → M satisfying the three requirements below. RELATING OPERATOR SPACES VIA ADJUNCTIONS 141 By combining this result with Theorem 6 we see that each of the spaces of operators B(H ). Before reading the deﬁnition of PCM. [0. x ⊥ ∈ E is the unique element in E with x x ⊥ = 1. 1.2. 1] is a PCM with sum of r. A homomorphism E → D of eﬀect algebras is given by a function f : E → D between the underlying sets satisfying f(1) = 1. This will be introduced ﬁrst. 2. Associativity: y ⊥ z and x ⊥ (y z) implies x ⊥ y and (x y) ⊥ z and also x (y z) = (x y) z. ) with an orthosupple- ment. we need the notion of eﬀect algebra and of monoid in eﬀect algebras. Eﬀect modules are structurally like modules over a semiring. and undeﬁned otherwise. Thus. The latter is a unary operation (−)⊥ : E → E satisfying: 1. This will be formalised next. see [3]. They involve the notation x ⊥ y for: x y is deﬁned. 1]) and DM(H ) ∼ = EMod(Ef (H ). As we will see in Theorem 14 below. An eﬀect algebra is a PCM (E. Pos(H ) can be obtained from the space DM(H ) of density operators via free constructions. see [14. Such an eﬀect monoid is a monoid in the category of eﬀect algebras. we need the notion of partial commutative monoid (PCM). 17] for more . is an example of a PCM. where 1 = 0⊥ . 1]. and in that case r s = r + s. 1] deﬁned if r + s ≤ 1. Eﬀect algebras and their homomorphisms form a category. The unit interval [0. density matrices and eﬀects can be translated back and forth: Ef (H ) ∼= Conv(DM(H ). Commutativity: x ⊥ y implies y ⊥ x and x y = y x. s ∈ [0. called EA. The notion of eﬀect algebra is due to [18]. obtained by adjoining a new element 0. Hence these density operators and eﬀects are in a sense most fundamental among the operators on a Hilbert space. But instead of a semiring of scalars one uses an eﬀect monoid. just like a semiring is a monoid in the category of commutative monoids. y are called orthogonal. 4. [0. and if x ⊥ x in E then both f(x) ⊥ f(x ) in D and f(x x ) = f(x) f(x ). Zero: 0 ⊥ x and 0 x = x. 0. Eﬀect modules. with u 0 = u = 0u. Deﬁnition 11. But in order to deﬁne an eﬀect algebra. in that case x. This + is obviously only a partial operation. Each orthomodular lattice is an eﬀect algebra. see also [14] for an overview. The unit interval is also an eﬀect algebra with r ⊥ = 1 − r. such as the unit interval [0. x ⊥ 1 ⇒ x = 0. think of the unit interval [0. For each set X the lift {0} + X of X . 2. 1] with addition +. Such structures are also studied under the name ‘partially additive monoid’. in order to deﬁne an eﬀect module. 3.

Most importantly in the current setting. 1. Thus MO(0) = {0. This yields a functor Ef : FdHilbUn → EA. then [0. 1} and MO(1) ∼ = P({0. 1] ⊗ MO(2) is an octahedron. The monoid structure on S is thus determined by a bimorphism · : S × S → S that preserves in each variable separately and satisﬁes 1 · x = x = x · 1. as in: S ⊗ S → S ← 2. . namely 0. 1] ⊗ {0. y) and f(x x . In [26] a notion of ‘convex category’ is introduced in which homsets Hom(X. Example 12. . x ⊥ x =⇒ f(x. y ∈ D. satisfying f(1. Using this symmetric monoidal structure (⊗. for 1 ≤ i ≤ n. One can think of elements of the tensor E ⊗ D as ﬁnite sums j xj ⊗ yj .e. when x ⊥ x and y ⊥ y . which are monoids in the category of commutative monoids. so that [0. 1] ⊗ P(X ) ∼ = [0. 2) are eﬀect algebras (where 2 = 1 + 1 and 1 is ﬁnal). In particular. y) ⊥ f(x. Since 2 is initial. say with n elements. the projections Pr(H ) of a Hilbert space form an eﬀect algebra. y ) and f(x. with P ⊥ Q iﬀ P ≤ Q ⊥ . and so an orthomodular lattice. (−)⊥ ) · and a monoid structure. in a standard way. 1] ⊗ P(X ) of eﬀect algebras is then given by the set of step functions f : X → [0. A⊥ = I − A. A monoid S ∈ Mon(EA) consists of a set S carrying eﬀect algebra structure (0. where one identiﬁes: 0⊗y=0 x ⊗ 0=0 (x x ) ⊗ y = (x ⊗ y) (x ⊗ y) x ⊗ (y y ) = (x ⊗ y) (x ⊗ y ). One writes MO(n) for the orthomodular lattice with 2n + 2 elements. 1}). where morphisms E ⊗ D → C in EA correspond to ‘bimorphisms’ f : E × D → C . since {0. The tensor unit is the two-element eﬀect algebra 2 = {0. consisting of positive operators A ≤ I is an eﬀect algebra. When X is a ﬁnite set. with A ⊥ B iﬀ A + B ≤ I . In [28] it is shown that the category EA is symmetric monoidal. y) y ⊥ y =⇒ f(x. 1]2 . 1) = 1. 1]n . the latter map S ← 2 does not add any structure. y) f(x . written multiplicatively.142 BART JACOBS AND JORIK MANDEMAKER information and examples. Since 2 is at the same time initial in EA we have a ‘tensor with coprojections’ (see [23] for ‘tensors with projections’). 1] ⊗ MO(1) ∼ = [0. objects of Mon(CMon). y) f(x. 1]. V ∈ P(X ) one has U ⊥ V iﬀ U ∩ V = ∅ and in that case U V = U ∪ V . y) ⊥ f(x . y) = f(x. 1} is the tensor unit. y y ) = f(x. It can be shown that [0. and thus an eﬀect algebra. and in that case A B = A + B. x ∈ E and y. For U. the set of eﬀects Ef (H ). As special case we have [0. further. 1} ∼ = [0. with only minimal equations. 2) on EA we can consider. i. i ⊥ . see [21]. such functions have only ﬁnitely many output values. y ). For an arbitrary set X the powerset P(X ) is a Boolean algebra. 1]. Such monoids are similar to semirings. and for all x. the category Mon(EA) of monoids in the category EA of eﬀect algebras. The tensor product [0. 1}. i.

The main example of a (commutative) monoid in EA is the unit interval [0. VII. 1] × Ef (H ) → Ef (H ). f(x) + g(x). Conv l ⊥ EMod op EMod(−. corresponding to a bimorphism S × X → X .1]) Proof. Similarly. VII. given by tensoring with S. It is not hard to see that this mapping H → Ef (H ) yields a functor FdHilbUn → EMod. By “homming into [0. with f ⊥ g iﬀ ∀x ∈ X .[0. These eﬀect modules have been studied earlier under the name ‘convex eﬀect algebras’. A homomorphism of eﬀect modules X → Y consists of a map of eﬀect algebras f : X → Y preserving scalar multiplication f(s • x) = s • f(x) for all s ∈ S and x ∈ X . By completely general reasoning the forgetful functor EMod S → EA has a left adjoint. If r1 + r2 ≤ 1. Given a convex set. Proposition 13.1]) . We prefer the name ‘eﬀect module’ to emphasise the similarity with ordinary modules. In this section an eﬀect module X ∈ ActS (EA) thus consists of an eﬀect algebra X together with an action (or scalar multiplication) • : S ⊗ X → X . In particular. as in: EMod S = ActS (EA) A S⊗(−) (14) EA See [34.§4] for details. It is easy to see that this is again an aﬃne function. see [36]. f(x) + g(x) ≤ 1. 1]” one obtains an adjunction: Conv(−. 1] ∈ EA via ordinary multiplication. [0.§4]). then we have the familiar distributivity in each variable.[0. with the usual scalar multiplication [0. as in: s · (r1 r2 ) = s · (r1 + r2 ) = (s · r1 ) + (s · r2 ) = (s · r1 ) (s · r2 ). 1]’. In that case one deﬁnes f g = x ∈ X . Again this is similar to the situation in Section 3 where the category Mod S of modules over a semiring S may be described as the category ActS (CMon) of commutative monoids with S-scalar multiplication. Here it is strengthened to an adjunction between convex sets and eﬀect modules. The eﬀects Ef (H ) of a Hilbert space form an example of an eﬀect module. In the sequel ‘eﬀect module’ will mean ‘eﬀect module over [0. the pointwise scalar product r • f = . or ‘eﬀect modules’ over S (see [34. 1]) of aﬃne maps is an eﬀect module.1] (EA). we shall write EMod for EMod [0.1] . A (dual) adjunction between convex sets and eﬀect algebras is described in [25]. RELATING OPERATOR SPACES VIA ADJUNCTIONS 143 For such a monoid S ∈ Mon(EA) we can consider the category ActS (EA) = EMod S of S-monoid actions (scalar multiplications).1] = Act[0. the homset Conv(X. We shall be most interested in the associated category EMod [0.

1]) gives a contravariant functor since for h : X → X in Conv pre-composition with h yields a map (−) ◦ h : Conv(X . [0. like in (10). With this adjunction in place we can give a clearer picture of density matrices and eﬀects. x ∈ X . the swapped version f) = y ∈ Y . 1]) is a map of eﬀect modules—and similarly for g. For X ∈ Conv and Y ∈ EMod. 1]) of eﬀect modules. [0. labeled hsEf in (15). [0. [0. This f forms a map of eﬀect modules. we can deﬁne an actual sum f : Y → [0. 1] by f(y) = j rj · fj (y). 1]) DM(H ) / EMod Ef (H ). There is a ‘dual adjunction’ between convex sets and eﬀect modules as in the lower part of the diagram below. where fj : Y→ [0. [0. Conv m ⊥ EMod op EMod(−. f(x)(y) : Y → Conv(X. Again.[0.g. 1]) in Conv ==================== Y / Conv(X. but the framing of the relevant structure in terms of maps of adjunctions is new.[0. there are natural isomorphisms: hsEf hsDM Ef (H ) / Conv DM(H ). in the categories EMod and Conv. In the other direction. 1]) → Conv(X. forming a map of adjunctions (like in Section 2. 1]) in EMod g What needs to be checked is that for a map f of convex sets as indicated. 1]) of eﬀect module maps yields a convex set: for a formal convex sum j rj | fj . r · f(x) yields an aﬃne function. [0. The dual adjunction between Conv and EMod involves a bijective correspon- dence that is obtained by swapping arguments. We start with the ﬁrst one. 1] in EMod. This is straightforward. [8]. we have: f X / EMod(Y. [0.1]) Proof. Theorem 14. 1]) ∼ = ∼ = A / tr(A−) B / tr(B−) (15) that give rise to a map of adjunctions given by states DM and statements (eﬀects) Ef in: (−)† - FdHilbUn m ⊥ FdHilbUn op † (−) DM Ef Conv(−. . [0. Further. This mapping X → Conv(X.144 BART JACOBS AND JORIK MANDEMAKER x ∈ X . functoriality is obtained via pre-composition. This map of adjunctions involves natural isomorphisms (15). the homset EMod(Y. [0. see e. given an eﬀect module Y .1]) . The isomorphisms involved are well-known.

for an arbitrary non-zero element x ∈ H there is a density matrix Bx = | x|x| .e. i. RELATING OPERATOR SPACES VIA ADJUNCTIONS 145 and note that it is well-deﬁned: for A ∈ Ef (H ) and B ∈ DM(H ) one has: hsEf (A)(B) = tr(AB) ≤ tr(IB) = tr(B) = 1. Assume A1 . tr(A1 −) = tr(A2 −) : DM(H ) → [0. 1]. A2 ∈ Ef (H ) satisfy hsEf (A1 ) = hsEf (A2 ). Injectivity of hsEf is obtained as follows.

Thus tr(A1 Bx ) = tr(A2 Bx ). Similarly A1 ≤ A2 . . and thus A2 ≤ A1 . or equivalently. We check linearity of h . Since this equation holds for all x ∈ H . It is easy to see that h (rB) = rh (B). C ∈ Pos(H ) we have: h (B) + h (C ) B C = tr(B) · h + tr(C ) · h tr(B) tr(C ) tr(B) B tr(C ) C = tr(B + C ) · ·h + ·h tr(B + C ) tr(B) tr(B + C ) tr(C ) tr(B) B tr(C ) C = tr(B + C ) · h · + · tr(B + C ) tr(B) tr(B + C ) tr(C ) since h preserves convex sums and: tr(B) tr(C ) tr(B) tr(C ) + = + =1 tr(B + C ) tr(B + C ) tr(B) + tr(C ) tr(B) + tr(C ) B +C = tr(B + C ) · h tr(B + C ) = h (B + C ). 1]. we get A1 − A2 ≥ 0. tr(B) = 0 h (B) = tr(B) · h tr(B) B otherwise. This is well-deﬁned since tr( tr(B)B ) = tr(B) tr(B) = 1. For surjectivity of hsEf assume a morphism of convex sets h : DM(H ) → [0. x | 2 :H → H . including x = 0. and for non-zero B. Then: (A1 − A2 )x | x = x | A1 x − x | A2 x = tr( x |A1 | x ) − tr( x |A2 | x ) = tr(A1 | x x |) − tr(A2 | x x |) ! = |x|2 tr(A1 Bx ) − tr(A2 Bx ) = 0. for r ∈ R≥0 . We turn it into a linear map h : Pos(H ) → R≥0 in the category Mod R≥0 of modules over R≥0 via: 0 ' ( if B = 0. and thus A1 = A2 .

For surjectivity assume a map of eﬀect modules g : Ef (H ) → [0. ' ! ( hsDM −1 ◦ (−) ◦ hsEf ◦ (B) = hsDM −1 ◦ (B) ◦ hsEf = hsDM −1 ◦ A. we extend it to a linear map g : Pos(H ) → R≥0 by: 1 1 g (B) = n · g B where n ∈ N is such that B ∈ Ef (H ). where j ≥ 0. One of the equations that hsEf and hsDM should satisfy to ensure that we have a map of adjunctions is the following. Then for A ∈ Ef (H ) we have g(A) = g (A) = tr(BA) ∈ [0. hsEf (A)(B) . n n m n m m It is easy to see that the map g is linear. Injectivity is obtained like for hsEf . Then 1 n B = j n1 j | j j | is an eﬀect by Lemma 7. 1]) (−)◦hsEf EMod(Ef (H ). Take the spectral decomposition Such an B = j j | j j |. We can ﬁnd an n ∈ N with j ≤ n for each j. is in Ef (H ) → Pos(H ). 1]. We claim that A is an eﬀect. because B is positive. n n n can be found in the following way. 1]) −1 hsDM DM(H ) This triangle commutes since for B ∈ DM(H ).e. h. without loss of generality m ≤ n. In particular 1 = g(I ) = tr(BI ) = tr(B). then we use that g is a map of [0. where the | j form an orthonormal basis. 1]. [0. We turn to the second map hsDM in (15). For a density operator B ∈ DM(H ) → Pos(H ) we get tr(AB) = h (B) = h(B) ∈ [0. 1]). and thus j = tr(A| j j |) = h (| j j |) = h(| j j |) ≤ 1. Each operator | j j | is a density matrix. [0. We also have to check that the deﬁnition of g is independent of the choice of n: if also m1 B ∈ Ef (H ). =B. Write A = j j | j j | as spectral decomposition. assume. and the | j form an orthonormal basis. h(B) DM(H ) / EMod Conv(DM(H ).146 BART JACOBS AND JORIK MANDEMAKER By Proposition 3 there is a unique A = hsPos −1 (h ) ∈ Pos(H ) with h = tr(A−) : Pos(H ) → R≥0 . By Lemma 7 we need to prove j ≤ 1. (B)(hsEf (A)) = hsDM −1 ◦ A. [0. i. 1]. so that B ∈ DM(H ). 1]-actions: 1 m 1 m 1 1 n·g B =n·g · B =n· ·g B =m·g B . Hence by Proposition 3 there is a (unique) B = hsPos −1 (g ) ∈ Pos(H ) with g = tr(B−) : Pos(H ) → R≥0 . the other one is similar and left to the reader. using that each operator | x x | is a projection and thus an eﬀect.

The dual adjunction Conv EMod op from Proposition 13 can be restricted to a (dual) equivalence of categories. Here we ignore complete positivity aspects and simply consider these state transformers as aﬃne maps. 1]) −→ Ef (H ) from (15). 1]) ============================== (Prop. RELATING OPERATOR SPACES VIA ADJUNCTIONS 147 = hsDM −1 ◦ A.g. For a state transformer f : DM(H ) → DM(K ) the corresponding predicate transformer wp(f. Remark 15. by a computer algebra tool). [0. see e. −) : Ef (K ) → Ef (H ) is the “weakest precondition operation”. The (dual) correspondence between state transformers and predicate transformers can then be derived using the adjunction Conv EMod op from Proposition 13 and the isomorphisms (15) from Theorem 14: DM(H ) / DM(K) in Conv ============================= (15) DM(H ) / EMod(Ef (K). Corresponding to such programs there are “predicate transformers” Ef (K ) → Ef (H ) going in the opposite direction. i. [0. It is given by: ! wp(f. [0. see [29]. tr(f(B)A) . tr(AB) = hsDM −1 ◦ A. ∼ where we use the isomorphisms hsDM : DM(K ) −→ = EMod(Ef (K ). In [11] a quantum weakest precondition calculus is developed using eﬀects on a ﬁnite-dimensional Hilbert space as predicates and density matrices as states. A) = hsEf −1 B ∈ DM(H ). the weakest precondition can be computed explicitly (for instance. The map of . By elaborating the = formulas for the matrix entries wp(f. The latter are suitably complete with respect to a deﬁnable norm. as maps in the category Conv. [12]. 13) Ef (K) / Conv(DM(H ). The underlying duality can be made explicit in the current setting. hsDM (f(B))(A) ! = hsEf −1 B ∈ DM(H ). [0. giving a probabilistic version of Gelfand duality. tr(BA) = hsDM −1 ◦ tr(B−) = B. A)jk . 1]) ============================ (15) Ef (K) / Ef (H ) in EMod Such correspondences form the basis of Dijkstra’s seminal work on program correctness. Programs act on states and are thus modeled as “state transformer” maps DM(H ) → DM(K ). Naturally we consider them to be maps of eﬀect modules. One obtains an equivalence CCH obs % BEMod op between ‘observable’ convex compact Hausdorﬀ spaces and Banach eﬀect modules.e. 1]) −1 ∼ and hsEf : Conv(DM(H ).

Special thanks. Table 1.[0. Bas Spitters. CCH obs m % BEMod op Hom(−. This says that eﬀects form the free eﬀect module on projections. [0.148 BART JACOBS AND JORIK MANDEMAKER adjunctions from Theorem 14 then restricts to a map of equivalences: (−)† - FdHilbUn m % FdHilbUn op (−)† DM Ef Hom(−. Klaas Landsman. Operators Formula Description eﬀects Ef (H ) ∼ = [0. 1] ⊗ Pr(H ) Gleason’s Theorem density matrices ∼ DM(H ) = EMod(Ef (H ). This equivalence leads to a reformulation of Gleason’s Theorem [19]. 1]) Theorem 14 positive operators Pos(H ) ∼ = S(DM(H )) Theorem 10 self-adjoint operators ∼ SA(H ) = R(Pos(H )) Theorem 6 bounded operators B(H ) ∼ = C(SA(H )) Theorem 6 This concludes our overview of the categorical structure of the various operators on a (ﬁnite dimensional) Hilbert space. 1] ⊗ Pr(H ). Acknowledgements. Rick Dejonghe.[0. Chris Heunen.1]) We refer to [29] for further details. In original form it says that projections on a Hilbert space H (of dimension at least 3) correspond to measures: DM(H ) ∼ = EA Pr(H ). see Table 1. 1]) . [0. . constructed from the projections Pr(H ). We can now summarise how the whole ediﬁce of operators on a Hilbert space H can be obtained from its projections Pr(H ). go to Bob Coecke. In [29] it is shown that Gleason’s theorem is equivalent to: Ef (H ) ∼ = [0. The ﬁrst steps of the research underlying this work was carried out during a sabbatical visit of the ﬁrst author (BJ) to the Quantum Group at Oxford University in April and May 2010. and Dusko Pavlovi´c. for discussion and/or feedback. Various operators on a Hilbert space H .1]) .

[16] J. 120403. [5] M. Physical Review Letters. editors). An extension of Gleason’s theorem for quantum computation. Semantics of weakening and contraction. Reports on Mathematical Physics. Handbook of Quantum Logic and Quantum Structures: Quantum Logic (K. J. 368–393. vol. A. vol. The Mathematical Language of Quantum Theory. pp. Abramsky and B.html. [11] E. 885–893. North Holland. Golan. Jacobs. 16 (2006). 2000.0585. Erratum and addendum in: 206(12):1476–1484. pp. Scholten. Lehmann. Journal of Mathematics and Mechanics. W. Foundations of Physics. 1827–1840. vol. monads and categories. 43 (2004). 1985. Texts and Monographs in Computer Science. 12 (2009). 1986. [8] P. M. Amsterdam. Oxford Logic Guides. Arbib and E. [9] B. [19] A. Borceux. pp. 1994. Scalars. Foulis and M. Toposes. vol. 1990. . A categorical semantics of quantum protocols. [7] F.-E. Information and Com- putation. M. [12] E. 2365–2376. Lecture Notes in Physics. Predicate Calculus and Program Semantics. pp. Triples and Theories. 12. Gabbay. pp. Eﬀect algebras and unsharp quantum logics. states and symmetries in the context of CB-eﬀect algebras. See arxiv. 1331–1352.org/abs/1003. From Uncertainty to Entanglement. Doberkat. G. pp. pp. D. 261–323. Oxford University Press. vol. 51 (1991). K. International Journal of Theoretical Physics. [20] J. Bar categories and star operations. pp. no. J. Pulmannova. S. Coumans and B. Computer Science Press. [18] D. Berlin. Category Theory. 1–77. and D. 50. Kluwer. 24 (1994). pp. Measures on the closed subspaces of a Hilbert space. [14] A. On involutive monoidal categories. Busch. 204 (2006).cwru. Awodey. 60 (2007). [10] D. [6] E. Ziman. vol. 1999. 2012. 69 (1994). North Holland. 2008. Oxford University Press. [23] B. vol. editors). 1756–1781. 91 (2003). 6 (1957). 2009. 429–451. Abramsky. Coecke (editor). vol. J. D’Hondt and P. Jacobs. Sadrzadeh. Springer. Annals of Pure and Applied Logic. Algebraic Approaches to Program Semantics. [13] E. Wells. revised and corrected version available from URL: www. [24] . [15] A. Categorical Logic and Type Theory. Barr and Ch. vol. Dvurecenskij ˇ and S. New Structures for Physics. Heunen and M. 813. Edalat. vol. Springer. 103–152. vol. vol. [2] S. Theory and Applications of Categories. Encyclopedia of Mathematics. 25 (2011). Springer. 73–106. Berlin. Beggs and S. Handbook of Categorical Algebra. Observables. [17] D. pp. 2011. Berlin. pp. Heinosaari and M. Domain theory in logical form. Quantum weakest preconditions. 1999. Kluwer. Gleason. Dordrecht. Algebras and Representation Theory. 2012. ´ New Trends in Quantum Structures. Eilenberg-Moore algebras for stochastic relations. [3] M. International Journal of Theoretical Physics. Majid. p. Springer. [22] T. Cambridge University Press. Quantum states and generalized observables: a simple proof of Gleason’s Theorem. Egger. Berlin. Panangaden. Mathematical Struc- tures in Computer Science. 51 and 52. 35 (1996). RELATING OPERATOR SPACES VIA ADJUNCTIONS 149 REFERENCES [1] S. Cambridge University Press. Manes. Compositional Methods in Physics and Linguistics (C. Dijkstra and C. Bennett. [21] S. Elsevier. Semirings and their Applications. [4] S. 329–346. Annals of Pure and Applied Logic.edu/artsci/math/wells/pub/ttt. Gudder. vol. Coecke. pp. problems and results in eﬀect algebras. vol. Examples. Foulis. 2006. Engesser.

20 (2010). 2012. Probabilities. THE NETHERLANDS E-mail: bart@cs. Closed categories generated by commutative monads. astr. vol. with a GNS-correspondence. vol. [30] B. Jacobs and J. Johnstone. editors). pp. H. 22 (1974). 14 (2004).nl . Jacobs and A. pp. 2010. Jacobs.doi. Springer. Coreﬂections in algebraic quantum logic. pp. IFIP Advances in Information and Communication Technology. [36] S. 25–30.3805. Keimel. pp. pp. pp. Selinger. Foundations of Physics. G. IFIP Theoretical Computer Science 2010 (Boston) (C.. Mac Lane. P. [35] E. Kock. Representation theorem for convex eﬀect algebras. Topology and its Applications. [29] .org/abs/1112. 932–958. 412 (2011). vol. Cambridge University Press. Calude and V. Springer. Berlin. Selinger. Springer. distribution monads and convex categories. vol. Categories for the Working Mathematician. 39–42. vol. 527–586. available from http://dml. Foundations of Physics. [38] M. Quantum Physics and Logic (QPL) 2011 (B. Sokolova. EPTCS. Theoretical Computer Science. 645–659. 82. Monadic functors and convexity. 156 (2008). (10 May 2012). 874–895. [39] T. des sciences math. vol. T. vol. et phys. Berlin. [27] . to appear. [26] . vol. vol.1007/s10701-012-9654-8. pp. 1041–1068. pp. Mathematical Structures in Computer Science.ru. pp. 3. Pulmannova´ and S. The expectation monad in quantum foundations. Exemplaric expressivity of modal logics. [34] S. 3323–3336. see arxiv. Journal of Logic and Computation. 1971. 39 (1998). editors). Postulates for the barycentric calculus. Algebraic Theories. Sassone. 1–19. pp. vol. Gudder. Polonaise des Sciences. Journal of the Australian Mathematical Society. 405–424. 1982.cz/dmlcz/119041. Bulletin de l’Acad. pp. 42 (2012). Cambridge Studies in Advanced Mathematics. 227–239.. S´er. 1974. Commentationes Mathematicae Universitatis Carolinae. Stone. http://dx. S. INSTITUTE FOR COMPUTING AND INFORMATION SCIENCES (ICIS) RADBOUD UNIVERSITY NIJMEGEN. Swirszcz. 29 (1949). Spitters. duality and eﬀects. [32] K. Manes. Annali di Matematica Pura ed Applicata. XII (1971). Mandemaker. [28] B. Towards a quantum programming language. and B. [37] P. The monad of probability measures over compact ordered spaces and its Eilenberg-Moore algebras. Stone Spaces. Convexity. [33] A.150 BART JACOBS AND JORIK MANDEMAKER [25] . Involutive categories and monoids. vol. [31] P.org/doi:10.

non-distributive lattices with an orthocom- plement (and generalisations thereof) have been used as representatives of the algebra of propositions about the quantum system at hand. co-Heyting and bi-Heyting algebras. we relate quantum systems to complete bi-Heyting algebras in a systematic way. 9. The elements of this algebra represent contextualised propositions about the values of the physical quantities of the quantum system. Section 3 recalls the deﬁni- tions and main properties of Heyting. 14. This generalised space takes the roleˆ of a state space for the quantum system. TOPOS-BASED LOGIC FOR QUANTUM SYSTEMS AND BI-HEYTING ALGEBRAS ¨ ANDREAS DORING Abstract. Since then. 13. 10]. This new form of logic for quantum systems is based on a certain Heyting algebra Subcl Σ of clopen. The main new observation is that the complete Heyting algebra Subcl Σ of clopen subobjects representing propositions is also a complete co-Heyting algebra. Eskandarian and V. The plan of the paper is as follows: in section 2. Introduction. Association for Symbolic Logic 151 . For review of standard quantum logic(s). 45 c 2016. we brieﬂy give some background on standard quantum logic and the main ideas behind the new topos-based form of logic for quantum systems. To each quantum system. Logic and Algebraic Structures in Quantum Computing Edited by J.e. There are a number of well-known conceptual and interpretational problems with this kind of ‘logic’. described by a von Neumann algebra of physical quantities. This includes two notions of implication and two kinds of negation. A. we associate a complete bi-Heyting algebra. 16. Chubb. see the article [6]. 15. Quantum logic started with Birkhoﬀ and von Neumann’s seminal article [4]. 11. one obtains a well-behaved intuitionistic form of logic for quantum systems which moreover has a topological underpinning.) In this way. as discussed in the following sections. Harizanov Lecture Notes in Logic.. Hence. In this article. closed and open subobjects of the spectral presheaf Σ. §1. In the last few years. (All technical notions are deﬁned in the main text. we will continue the development of the topos-based form of logic for quantum systems. i. a diﬀerent form of logic for quantum systems based on generalised spaces in the form of presheaves and topos theory has been developed by Chris Isham and this author [12.

complex Hilbert space H as an example of a von Neumann algebra. If the Hilbert space H is ﬁnite- dimensional.152 ¨ ANDREAS DORING section 4 introduces the spectral presheaf Σ and the algebra Subcl Σ of its clopen subobjects. Heyting-regular and co-Heyting regular elements are characterised and a tentative physical interpretation of the two kinds of negation is given. such lattices (and various kinds of generalisations. The fact that each von Neumann algebra has ‘suﬃciently many’ projections makes it attractive for quantum logic. These propositions are pre-mathematical entities that refer to the ‘world out there’. and the spectral theorem holds in a von Neumann alge- bra. which we don’t consider here) have been considered as quantum logics. In this article. co-Heyting structure are consid- ered. §2. In section 5. the link between standard quantum logic and the topos-based form of quantum logic is established and it is shown that Subcl Σ is a complete bi-Heyting algebra. Standard quantum logic. propositions of the form “A ε Δ” are represented by projection operators via the spectral theorem. If. dim H = n. we assume some familiarity with the most basic aspects of the theory of von Neumann algebras and with basics of category and topos theory. Section 7 concludes. This class of algebras is general enough to describe a large variety of quantum mechanical systems. or is aﬃliated with N in the case that Aˆ is unbounded. including systems with symmetries and/or superselection rules. as we always assume. see e. Background. [28]. then the . Throughout. More speciﬁcally. From the perspective of quantum logic. the two kinds of negations associated with the Heyting resp. or more precisely as algebras representing propositions about quantum systems. Starting from Birkhoﬀ and von Neumann [4]. The reader not familiar with von Neumann algebras can always take the algebra B(H) of all bounded operators on a separable. the physical quantity A is described by a self-adjoint operator Aˆ in a given von Neumann algebra N . providing the link between self-adjoint operators (representing physical quantities) and projections (representing propositions). which is written shortly as “A ε Δ”. then B(H) is nothing but the algebra of complex n × n-matrices. The kind of propositions that we are concerned with (at least in standard quantum logic) are of the form “the physical quantity A has a value in the Borel set Δ of real numbers”. each von Neumann algebra is generated by its projections. we will discuss structures associated with von Neumann algebras. In standard quantum logic. The text is interspersed with some physical interpretations of the mathematical constructions. Von Neumann algebras. the key thing is that the projection operators in a von Neumann algebra N form a complete orthomodular lattice P(N ). In section 6.g.

i. For some conceptual discussion.) Following Birkhoﬀ and von Neumann. We partially order the set of 1 From here on. thus identifying the mathematical notion and its physical interpretation. Each context provides one of many ‘classical perspectives’ on a quantum system. Yet. see in particular [11]. i. 24. Unbounded self-adjoint operators aﬃliated with N can be treated in a straightforward manner. 5. 10] is fundamentally diﬀerent from standard quantum logic. Contexts and coarse-graining. In this way. Contextuality has of course been considered widely in foundations of quantum theory.. 20. 16. co-measurable. Such a set determines and is determined by an abelian von Neumann subalgebra V of the non-abelian von Neumann algebra N of (all) physical quantities. the orthogonal complement of a projection. where {S} denotes the double commutant of a set S of operators (see e. we assume that all the physical quantities A correspond to bounded self-adjoint i operators that lie in N . ˆ is interpreted as negation. For recent. P(N ) is a distributive lattice if and only if N is abelian if and only if all physical quantities considered are mutually compatible. Two key ideas are contextuality and coarse-graining of propositions. 2 We will often use the notation V for a subalgebra of V. V is generated by the operators (Aˆi )i∈I and the identity 1. [6].g. 19] and [1. [28]). ∨ in the projection lattice P(N ) as logical connectives between the propositions represented by the projections. The set of all contexts will be denoted V(N ). Physically. The topos-based form of quantum logic that was established in [13] and developed further in [9.g.1 In fact. e. meets and joins do not distribute over each other. ˆ Aˆi | i ∈ I } . Pˆ := 1ˆ − P. see e. 25. 11.2 Each abelian von Neumann subalgebra V of N will be called a context. in particular since Kochen and Specker’s seminal paper [29]. In fact. 26.g. There are many other conceptual diﬃculties with quantum logics based on orthomodular lattices.e. TOPOS-BASED LOGIC FOR QUANTUM SYSTEMS AND BI-HEYTING ALGEBRAS 153 projection corresponding to “A ε Δ” lies in P(N ).ˆ in the sense that V = {1. related work see also [18. one then interprets the lattice operations ∧. 2]. We trust that this will not lead to confusion. so N is never abelian and P(N ) is non-distributive.. It ﬁrst showed up in work by Chris Isham and Jeremy Butterﬁeld [21. the meet ∧ becomes a conjunction and the join ∨ a disjunction.e. 22] and was substantially developed by this author and Isham. This makes the interpretation of P(N ) as an algebra of propositions somewhat dubious. Each physical quantity Ai in the set is represented by some self-adjoint operator Aˆ in V. co-measurable physical quantities (Ai )i∈I . 27. the systematic implementation of contextuality in the language of presheaves is comparatively new. Quantum systems always have some incompatible physical quantities. 23. Moreover. . which does not mean the commutant of V. a context is nothing but a set of compatible. (For details on the spectral theorem see any book on functional analysis. Crucially. [28].

and P(V ) clearly is a sublattice of P(N ). more limited classical perspective containing fewer physical quantities than V. Since by contravariance we go from S V to S V . Since P(V ) is a Boolean algebra. the component at V . Hence. Each context V ∈ V(N ) has a complete Boolean algebra P(V ) of projections. there is a function deﬁned from S V. . to S V . The presheaves we consider are ‘varying sets’ (S V )V ∈V(N ) . In fact. 1})V ∈V(N ) such that if V = V ∩ V˜ is a subcontext of both V and V. if one considers presheaves instead of sets. this means that a certain presheaf has no global elements. An obstacle arises since the Kochen-Specker theorem seems to show that such a quantum state space cannot exist. This is typical. this problem can be overcome. which can be seen as truth-value assignments as usual. but rather the so-called clopen subobjects. there are consistent truth-value assignments for all propositions “A ε Δ” for propositions about physical quantities within a context. the presheaf map S(iV V ) : S V → S V will implement a form of coarse-graining of the information available at V to that available at V . where V |V is the restriction of V to V the subcontext V . In [7]. Set-valued functor. there are no truth-value assignments for all contexts simultaneously in the following sense: there is no family of Boolean algebra homomorphisms (V : P(V ) → {0. Yet. In the topos approach to quantum theory. indexed by contexts. Heyting algebras are the algebraic representatives of (propositional) intuitionistic logics. and analogously V˜ |V . so we can describe more physics from the perspective of V than from V . dim H ≥ 3. The subobjects of the quantum state space. 27]. there are Boolean algebra homomorphisms : P(V ) → {0. Typically. Propositions “A ε Δ” about the values of physical quantities A in a (physical) context correspond to projections in the (mathematical) context V.154 ¨ ANDREAS DORING contexts V(N ) by inclusion. the set associated with the smaller context V . containing more self-adjoint operators and more projections than the smaller context V . but by suitable subobjects of a quantum state space. form a (complete) Heyting algebra. ˜ then V = V |V = ˜ |V . the presheaves deﬁned over contexts will mirror this fact: the component S V at V contains more information (in a suitable sense. Hence. Whenever V ⊂ V. the set associated with the context V. This makes S = (S V )V ∈V(N ) into a contravariant. 1} % {false. since the subobjects of any object in a topos form a Heyting algebra. As Isham and Butterﬁeld realised [23. The latter also form a complete Heyting algebra. The key result by Kochen and Specker [29] shows that for N = B(H). to be made precise in the examples in section 4) than S V . V is the bigger context. true}. A smaller context V ⊂ V represents a ‘poorer’. propositions are represented not by projections. which will be called the spectral presheaf Σ. it is shown that this result generalises to all von Neumann algebras without a type I2 -summand. there is a built-in idea of coarse-graining. we will not consider all subobjects of the spectral presheaf.

[38]. 11. see e. 31]. these constructions provide an intuitionistic form of logic for quantum systems. since the quantum state space Σ is a generalised space associated with the nonabelian algebra N . with Makkai [35] and Zolfaghari [38]. It is straightforward to show that the underlying lattice of a Heyting algebra is distributive. TOPOS-BASED LOGIC FOR QUANTUM SYSTEMS AND BI-HEYTING ALGEBRAS 155 as was ﬁrst shown in [13] and is proven here in a diﬀerent way. As far as we are aware. in section 5. Lawvere emphasised the importance of co-Heyting and bi-Heyting algebras in category and topos theory.2) i∈I i∈I . The diﬀerence between the set of all subobjects of the spectral presheaf and the set of clopen subobjects is analogous to the diﬀerence between all subsets of a classical state space and (equivalence classes modulo null subsets of) measurable subsets. 8. H is a lattice such that for any two elements A. (3. connected bi-Heyting algebras with modal logic. there exists an exponential A ⇒ B. Majid has suggested to use Heyting and co-Heyting algebras within a tentative representation-theoretic approach to the formulation of quantum gravity [33. If the underlying lattice is complete.g. Moreover. 17]). In a recent paper.1) This means that the product (meet) functor A ∧ : H → H has a right adjoint A ⇒ : H → H for all A ∈ H . which is characterised by the adjunction C ≤ (A ⇒ B) if and only if C ∧ A ≤ B. [3] prove (among other things) new duality theorems for bi-Heyting algebras based on bitopological spaces. nobody has connected quantum systems and their logic with bi-Heyting algebras before. B ∈ H . The use of bi-Heyting algebras in superintuition- istic logic was developed by Rauszer [36. A Heyting algebra H is a lattice with bottom element 0 and top element 1 which is a cartesian closed category. In other words. §3. then the adjoint functor theorem for posets shows that for all A ∈ H and all families (Ai )i∈I ⊆ H . Bi-Heyting algebras. The following deﬁnitions are standard and can be found in various places in the literature. using Galois connections. Reyes. in particular in connection with continuum physics [30. Bezhanishvili et al. but see [13. (3. called the Heyting implication (from A to B). Together with the representation of states (which we will not discuss here. 34]. there is a clear topological underpinning. 37]. the following inﬁnite distributivity law holds: A∧ Ai = (A ∧ Ai ). The construction of the presheaf Σ and its algebra of subobjects incorporates the concepts of contextuality and coarse-graining in a direct way. see sections 4 and 5.

8) It is straightforward to show that the underlying lattice of a co-Heyting algebra is distributive. Some standard properties of the Heyting negation are: A ≤ B implies ¬A ≥ ¬B.9) i∈I i∈I The co-Heyting negation is deﬁned as ∼ : J −→ J op (3. ∼A is the smallest element in J such that A ∨ ∼A = 1. called the co-Heyting implication (from A). (3..156 ¨ ANDREAS DORING The Heyting negation is deﬁned as ¬ : H −→ H op (3. the last property on this list means that in a Heyting algebra the law of excluded middle need not hold: in general. (3. ¬A is the largest element in H such that A ∧ ¬A = 0. i.e.11) The deﬁning adjunction shows that ∼A = {B ∈ J | A ∨ B = 1}.4) ¬¬A ≥ A. the disjunction between a proposition represented by A ∈ H and its Heyting negation (also called Heyting complement. A co-Heyting algebra (also called Brouwer algebra) J is a lattice with bottom element 0 and top element 1 such that the coproduct (join) functor A∨ : J → J has a left adjoint A ⇐ : J → J . (3.6) ¬A ∨ A ≤ 1.5) ¬¬¬A = ¬A (3. then the adjoint functor theorem for posets shows that for all A ∈ J and all families (Ai )i∈I ⊆ J .3) A −→ (A ⇒ 0).10) A −→ (1 ⇐ A). i. Heyting algebras are algebraic representatives of (propositional) intuitionistic logics. or pseudo-complement) ¬A can be smaller than 1. which represents the trivially true proposition. It is characterised by the adjunction (A ⇐ B) ≤ C iﬀ A ≤ B ∨ C. (3. A canonical example of a Heyting algebra is the topology T of a topological space (X. Some properties of the ..e. T ). If the underlying lattice is complete.7) Interpreted in logical terms. (3. (3. with unions of open sets as joins and intersections as meets. the following inﬁnite distributivity law holds: A∨ Ai = (A ∨ Ai ). The deﬁning adjunction shows that ¬A = {B ∈ H | A ∧ B = 0}.

Heyting algebras and co-Heyting algebras are dual notions. in a Boolean algebra. 13. (3. the so-called spectral presheaf. TOPOS-BASED LOGIC FOR QUANTUM SYSTEMS AND BI-HEYTING ALGEBRAS 157 co-Heyting negation are: A ≤ B implies ∼A ≥ ∼B. (3. the conjunction between a proposition represented by A ∈ J and its co-Heyting negation ∼A can be larger than 0. Of course. .16) which is the characterising property of the co-Heyting negation.13) ∼∼∼A = ∼A (3. i.. but in the ﬁnal section 6. the so- called clopen subobjects. For each A ∈ K. 16. This gives the connection with the topological examples.12) ∼∼A ≤ A. The opposite H op of a Heyting algebra is a co-Heyting algebra and vice versa. A canonical example of a co-Heyting algebra is given by the closed sets C of a topological space. are deﬁned and their interpretation is given: clopen subobjects can be seen as families of local propositions. closed and open. we will give and interpretation of the co-Heyting negation showing up in the form of quantum logic to be presented in this article.e. A bi-Heyting algebra K is a lattice which is a Heyting algebra and a co- Heyting algebra. In fact. (3. A distinguished family of subobjects. subsets of its Stone space. for all A ∈ B. Co-Heyting algebras are algebraic representatives of (propositional) paraconsistent logics. (3. With each von Neumann algebra N . the last property on this list means that in a co-Heyting algebra the law of noncontradiction does not hold: in general. we associate a particular presheaf. We will not discuss paraconsistent logic in general. and the functor A ∨ : K → K has a left adjoint K ⇐ : K → K. with unions of closed sets as joins and intersections as meets. which represents the trivially false propo- sition. A canonical example of a bi-Heyting algebra is a Boolean algebra B. The spectral presheaf of a von Neumann algebra and clopen subobjects. §4. the functor A ∧ : K → K has a right adjoint A ⇒ : K → K .15) Interpreted in logical terms. compatible with respect to coarse-graining.) In a Boolean algebra. ¬ = ∼. (Note that by Stone’s representation theorem. we have for the Heyting negation that.14) ∼A ∧ A ≥ 0. A ∨ ¬A = 1. each Boolean algebra is isomorphic to the algebra of clopen. 17]. A bi-Heyting algebra K is called complete if it is complete as a Heyting algebra and complete as a co-Heyting algebra. The constructions presented here summarise those discussed in [12.

(This will play an important roleˆ in the discussion of the Heyting negation in section 6. Let N be a von Neumann algebra. The spectral presheaf Σ of N is the presheaf over V(N ) given (a) on objects: for all V ∈ V(N ). then there is a join-preserving map . then Σ(A) is extremely disconnected.3) Pˆ −→ N o ˆ .e. it holds that o V. A % C (Σ(A)).. then V. V ∈ V(N ) such that V ⊂ V.1) Pˆ −→ V. the elements of the Gel’fand spectrum Σ(A) are the pure states of A.158 ¨ ANDREAS DORING Let N be a von Neumann algebra. (4.W ◦ V. which makes it into a compact Hausdorﬀ space. If W ⊂ V ⊂ V.2) i∈I i∈I where the join on the left hand side is taken in P(V ) and the join on the right hand side is in P(V ). partially ordered under inclusion. equipped with the supremum norm. the inclusion iV V : V → V restricts to a morphism iV V |P(V ) : P(V ) → P(V ) of complete Boolean algebras. iV V preserves all meets. Equivalently. ΣV := Σ(V ). By convention.V ˆ (Pi ). given by the identity operator 1ˆ on the Hilbert space on which N is represented. If N is a von Neumann algebra and M is any von Neumann subalgebra such that their unit elements coincide. we exclude the trivial subalgebra V0 = C1ˆ from V(N ).V Pˆ i = o V. ˆ Recall that the Gel’fand spectrum Σ(A) of an abelian C ∗ -algebra A is the set of algebra homomorphisms : A → C. and let V(N ) be the set of its abelian von Neumann subalgebras.V o ˆ (P) = {Qˆ ∈ P(V ) | Qˆ ≥ P} ˆ that preserves all joins.V (4. the Gel’fand spectrum of V. hence it has a left adjoint : P(V ) −→ P(V ) o V.W o = Vo . We note that distributivity of the lattices P(V ) and P(V ) plays no role ˆ here. that is. obviously. We now deﬁne the main object of interest: Deﬁnition 1. for all families (Pˆi )i∈I ⊆ P(V ). By Gel’fand-Naimark duality. i. In particular. We only consider subalgebras V ⊂ N which have the same unit element as N . For V .) The poset V(N ) is called the context category of the von Neumann algebra N . complex-valued functions on Σ(A).V o . The set Σ(A) is given the relative weak*-topology (as a subset of the dual space of A). If A is an abelian von Neumann algebra. .M (P) = {Qˆ ∈ P(M) | Qˆ ≥ P}.M : P(N ) −→ P(M) o N (4. 1ˆ M = 1ˆ N . A is isometrically ∗-isomorphic to the abelian C ∗ -algebra C (Σ(A)) of continuous.

see Def. V ∈ V(N ). together with the lattice operations and bi-Heyting algebra structure deﬁned below. so S V represents a local proposition at the smaller context V ⊂ V that is coarser than (i. The set of all clopen subobjects of Σ is denoted as Subcl Σ. the set S V is a clopen subset of the Gel’fand spectrum ΣV. The set Subcl Σ. The most direct connection with propositions of the form “A ε Δ” is given by the map called daseinisation..5) Pˆ −→ { ∈ ΣV | (P) ˆ = 1}. respectively. If V ⊂ V. then S V ⊇ Σ(iV V )(S V ) (since S is a subobject). since Σ is a presheaf over the context category V(N ). the component S V ⊆ ΣV represents a local proposition about the value of some physical quantity. V The restriction maps Σ(iV V ) are well-known to be continuous. The elements S ∈ Subcl Σ represent propositions about the values of the physical quantities of the quantum system. Cl (ΣV ) denotes the clopen subsets of ΣV. Clearly. there is an isomorphism of complete Boolean algebras αV : P(V ) −→ Cl (ΣV ) (4. We note that for each abelian von Neumann algebra V (and hence for each context V ∈ V(N )). the deﬁnition of clopen subobjects makes use of the Gel’fand topologies on the components ΣV.6) V ∈V(N ) . Moreover. Σ(iV V ) : ΣV −→ ΣV (4.4) −→ | . this means the following: for each context V ∈ V(N ). section 2) is implemented by this construction. TOPOS-BASED LOGIC FOR QUANTUM SYSTEMS AND BI-HEYTING ALGEBRAS 159 (b) on arrows: for all inclusions iV V : V → V. a consequence of) the local proposition represented by S V. A subobject S of Σ is called clopen if for each V ∈ V(N ). coarse-graining is mathematically realised by the fact that we use subobjects of presheaves. is the algebraic implementation of the new topos-based form of quantum logic. In the case of Σ and its clopen subobjects. We note that the concept of contextuality (cf. one for each context. surjective maps with respect to the Gel’fand topologies on ΣV and ΣV . We equip the spectral presheaf with a distinguished family of subobjects (which are subpresheaves): Deﬁnition 2. such that smaller contexts are assigned coarser propositions. A clopen subobject S ∈ Subcl Σ can hence be interpreted as a collection of local propositions. They are also open and closed.e. There is a purely order-theoretic description of Subcl Σ: let P := P(V ) (4. 3 below. see e. [13].g. Here.

Example 1. Note that the lattice operation is not just componentwise set-theoretic intersection. not just open ones. This guarantees that one obtains clopen subsets at each stage V. Pˆ 2 + Pˆ 3 . Pˆ 2 . Pˆ 1 + Pˆ 2 . We deﬁne a partial order on Subcl Σ in the obvious way: ∀S. since each P(V ) is a complete Boolean algebra.V . The fact that meets and joins are not given by set-theoretic intersections and unions also means that Subcl Σ is not a sub-Heyting algebra of the Heyting algebra Sub Σ of all subobjects of the spectral presheaf. see [8. where f(V ) ∈ P(V ) for all V ∈ V(N ).11) Of course. The diﬀerence between Sub Σ and Subcl Σ is analogous to the diﬀerence between the power set PX of a set X and the complete Boolean algebra BX of measurable subsets (with respect to some measure) modulo null sets. ∀V ∈ V(N ) : Si := cl S i. (4. P(N ) is a Boolean algebra. For illustration. S % Subcl Σ. 1}.V . Pˆ 1 + Pˆ 3 . . (4. 17]. + ∀V ∈ V(N ) : Si := int S i.10) where Pˆ 1 . P is a complete Boolean algebra. (4. into which P(V ) can be included). Pˆ 3 . Consider the subset S of P consisting of those functions for which V ⊂ V implies f(V ) ≥ f(V ) (this comparison is taken in P(V ). Analogously. Pˆ 3 are pairwise orthogonal rank-1 projections on a Hilbert space of dimension 3.7) We deﬁne the corresponding (complete) lattice operations in a stagewise manner. The subset S is closed under all meets and joins (in P). not just closed ones. P(N ) = {0. (4.160 ¨ ANDREAS DORING * be the set of choice functions f : V(N ) → V ∈V(N ) P(V ). at each context V ∈ V(N ) separately: for any family (S i )i∈I . Pˆ 2 . i. Equipped with pointwise operations. In section 5..e. ˆ ˆ Pˆ 1 . The projection lattice P(N ) of N has 8 elements.V ⊆ ΣV is the component at V of the clopen subobject S i . but rather the interior (with respect to the Gel’fand topology) of the intersection. T ∈ Subcl Σ : S ≤ T :⇐⇒ (∀V ∈ V(N ) : S V ⊆ T V ). (4.8) i∈I V i∈I where S i. and clearly. we consider a simple example: let N be an abelian von Neumann of diagonal matrices in 3 dimensions. For results on measures and quantum states from the perspective of the topos approach.9) i∈I V i∈I where the closure of the union is necessary in order to obtain clopen sets. we will show that Subcl Σ is a complete bi-Heyting algebra. This is given by N := CPˆ 1 + CPˆ 2 + CPˆ 3 .

Hence. 3. so we have 3 · 2 = 6 subobjects S with two elements in S N . 3.N )(S N ) = {1+2 } must be contained in. . Then S V1 = ΣV1 and S V2 = ΣV2 . Then S V1 can either be {1 } or {1 . S V2 can either be {1+3 } or {1+3 . there are no conditions arising from restriction maps of the spectral presheaf Σ. 2. (c) Let S be such that S N contains one element. 2+3 (1ˆ − Pˆ 1 ) = 1. This completes the description of the spectral presheaf Σ of the algebra N . the spectrum ΣV2 has two elements 1+3 . 2 }. i = 1. Then the restriction maps of Σ dictate that for each Vi .15) The restriction maps from ΣN to ΣV2 resp. 2 . i = 1. S N = {1 }. We distinguish a number of cases: (a) Let S ∈ Subcl Σ be a subobject such that S N = ΣN = {1 . Moreover. Hence. where i (Pˆ j ) = ij . 3. 2 }. e. 2+3 (Pˆ 1 ) = 0. we can pick an arbitrary subset of ΣVi for i = 1. Since the Vi are not contained in one another. so there are two options. but not necessarily equal to S V3 by the deﬁnition of subobjects). so there are 3 · 23 = 24 subobjects S with one element in S N . First. S N = {1 . 3 }.N )(3 ) = 2+3 . 3 } (since Σ(iV3 . 3 }. Consider the restriction map of the spectral presheaf from ΣN to Σ1 : Σ(iV1 . The Gel’fand spectrum ΣN of N has three elements 1 .12) Hence. 1 (1ˆ − Pˆ 1 ) = 0. 3 .N )(S N ) = ΣVi . 2+3 }. (4. 2.) Analogously. We simply have to determine all subobjects of Σ. (4. (d) Finally. 2. and the spectrum ΣV3 has two elements 1+2 . note that the Gel’fand spectra all are discrete sets. TOPOS-BASED LOGIC FOR QUANTUM SYSTEMS AND BI-HEYTING ALGEBRAS 161 The algebra N has three non-trivial abelian (von Neumann) subalgebras. so we have 43 = 64 subobjects S with S N = ∅. 3. (b) Let S be a subobject such that S N contains two elements. 2 .13) The Gel’fand spectrum ΣV1 of V1 has two elements 1 . e. but S V3 can either be {1+2 } or {1+2 . We will now determine all clopen subobjects of Σ. and S V3 can either be {1+2 } or {1+2 . hence there are 4 subsets of ΣVi for i = 1. there are three ways of picking one element from ΣN .g. 2. Each ΣVi has 2 elements.14) (Note that 1ˆ − Pˆ 1 = Pˆ 2 + Pˆ 3 .N )(2 ) = Σ(iV1 . there are three ways of picking two elements from the three-element set ΣN . Moreover. 3 . 2 . (4. Vi := CPˆ i + C(1ˆ − Pˆ i ). consider a subobject S such that S N = ∅. so S must be Σ itself. (4. the context category V(N ) is the 4-element poset with N as top element and Vi ⊂ N for i = 1. Σ(iV1 .N )(1 ) = 1 . we have S Vi ⊃ Σ(iVi . there are 23 options. 2. ΣV3 are deﬁned analogously. 3. 2+3 such that 1 (Pˆ 1 ) = 1.g. so topological questions are trivial here.

V (P) to the corresponding clopen subset of ΣV. 10]. Moreover. if Pˆ .V (P) may represent “A ε Γ” where Γ ⊃ Δ. Outer daseinisation can hence be seen as a map from propositions of the form “A ε Δ” into the bi-Heyting algebra Subcl Σ of clopen subobjects of the spectral presheaf. A projection P.V (P))V ∈V(N ) . encoded by the complete orthomodular lattice P(N ) of projections. Let N be a von Neumann algebra. ˆ representing a o ˆ proposition “A ε Δ”.) . Representation of propositions and bi-Heyting algebra structure. consisting . based on the clopen subobjects of the spectral presheaf Σ. This map was introduced in [13] and discussed in detail in [11. which are joins.162 ¨ ANDREAS DORING In all. which does not aﬀect the interpretation. Here is a direct argument: being left adjoint to the inclusion of P(V ) into P(N ). Apart from the spectral presheaf. For example.. (Each isomorphism of one projection N o ˆ αV. see [16] for details. Deﬁnition 3. represents “A ε Δ”. §5. The map o : P(N ) −→ Subcl Σ (5.1) Pˆ −→ o (P) ˆ := (αV (N o ˆ .e.V (P) for each context V ∈ V(N ). “the physical quantity A has a value in the Borel set Δ of real numbers”.V (P) represents Since we have N o ˆ ˆ o ˆ a coarser (local) proposition than “A ε Δ” in general. αV . Boolean subalgebras). the map N o . to a form of (super)intuitionistic logic for quantum systems.V (P)))V ∈V(N ) is called outer daseinisation of projections. there are a number of other presheaves that play a role in the new mathematical description of quantum system provided by the topos approach to quantum theory. which conceptually plays the role ˆ of a quantum state space. V ∈ V(N ). It can be seen as a ‘translation’ map from standard quantum logic. just maps the projection N o ˆ . We do not discuss these other presheaves here.V (P) ≥ P for all V. but results from considering all subcontexts (i. The connection between propositions and projections is given by the spectral theorem.D of [13] and in [11]. In standard quantum logic. The increased number of elements is not due to non-distributivity. then N o ˆ o The map preserves all joins.V preserves all colimits. This can be compared with the 8-element Boolean algebra P(N ) that we started from. the projection N . is mapped to a collection (N . Subcl Σ has 64 + 24 + 6 + 1 = 95 elements. and let P(N ) be its lattice of projections. as shown in section 2. The topos in which the spectral presheaf op and the other presheaves lie is the topos SetV(N ) of presheaves over the context category V(N ). the projections Pˆ ∈ P(N ) represent propositions of the form “A ε Δ”. that is.

the Heyting negation ¬ is deﬁned for all S ∈ Subcl Σ by ¬S := (S ⇒ 0). and let S ∈ Subcl Σ. o (P)∧ o ˆ ˆ for any projection Rˆ ∈ P(N ). (5. then |V ∈ T V }. so αV ◦ N o . so o preserves all joins. (S ⇒ T )V = { ∈ ΣV | ∀V ⊆ V : if |V ∈ S V . This holds for all V ∈ V(N ). (5. (5. (5. 11] for proof of these statements.2) ˆ In general. and o (1) ˆ Qˆ ∈ P(N ) : o (Pˆ ∧ Q) ∀P. For meets.) This implies that (S ⇒ T ) = {R ∈ Subcl Σ | R ∧ S ≤ T }. the empty subobject. the functor S ∧ : Subcl Σ −→ Subcl Σ (5. ¬S is the largest element of Subcl Σ such that S ∧ ¬S = 0. This was shown before in [13]. the Heyting implication from S. (5. for each S ∈ Subcl Σ. The Heyting implication is given by the adjunction R∧S ≤T if and only if R ≤ (S ⇒ T ). Moreover.5) This map. (5. we have (0) = 0. ˆ ≤ o (P) ˆ ∧ o (Q).V ). o is order-preserving and injective. so by the adjoint functor theorem for posets. (5. Hence. makes Subcl Σ into a complete Heyting algebra.9) That is. Let (S i )i∈I ⊆ Subcl Σ be a family of clopen subobjects of Σ.11) . but not surjective.10) The stagewise expression for ¬S is (¬S)V = { ∈ ΣV | ∀V ⊆ V : |V ∈ / S V }.8) As usual.6) (Note that S ∧ = ∧ S. (Q) is not of the form o (R) See [13.4) preserves all joins. ˆ (5. a complete Boolean algebra) in which ﬁnite meets distribute over arbitrary joins. TOPOS-BASED LOGIC FOR QUANTUM SYSTEMS AND BI-HEYTING ALGEBRAS 163 is an isomorphism of complete Boolean algebras. and joins in Subcl Σ are deﬁned stagewise. Then ∀V ∈ V(N ) : S ∧ Si = (S V ∧ S i.V preserves all joins. Clearly. it has a right adjoint S ⇒ : Subcl Σ −→ Subcl Σ.7) The stagewise deﬁnition is: for all V ∈ V(N ). o ˆ ˆ = Σ. (5.3) i∈I V i∈I since Cl (ΣV ) is a distributive lattice (in fact.

⇐. [38]): S ⇐ T is the smallest clopen subobject R for which T ∨ R is bigger then S. We deﬁne a co-Heyting negation for each S ∈ Subcl Σ by ∼S := (Σ ⇐ S). ¬. Hence. (Subcl Σ. 0.14) which we call co-Heyting implication. for each S the functor S ∨ : Subcl Σ −→ Subcl Σ (5. For all S ∈ Subcl Σ. ∼S is the smallest clopen subobject such that ∼S ∨ S = Σ (5. so it encodes how much is ‘missing’ from T to cover S. Proof.15) so (S ⇐ T ) = {R ∈ Subcl Σ | S ≤ T ∨ R}. (5. ∼) is a complete bi-Heyting al- gebra. We give direct arguments for the following two facts (which also follow from the general theory of bi-Heyting algebras): Lemma 1. ∼S ∧ S ≥ 0. ∨.V ). It is characterised by the adjunction (S ⇐ T ) ≤ R iﬀ S ≤ T ∨ R. (5. (5.12) i∈I V i∈I since ﬁnite joins distribute over arbitrary meets in Cl (ΣV ). ∀V ∈ V(N ) : S ∨ Si = (S V ∨ S i. we also have.) Hence. This map makes Subcl Σ into a complete co-Heyting algebra. so it has a left adjoint S ⇐ : Subcl Σ −→ Subcl Σ (5. we have ¬S ≤ ∼S.18) holds. (5.164 ¨ ANDREAS DORING In Subcl Σ. We have shown: Proposition 1. ⇒. For all V ∈ V(N ). . Σ. it holds that (¬S)V ⊆ ΣV \S V.13) preserves all meets. ∧. The above lemma and the fact that ¬S is the largest subobject such that ¬S ∧ S = 0 imply Corollary 1.17) (Note that Σ is the top element in Subcl Σ. In general.16) One can think of S ⇐ as a kind of ‘subtraction’ (see e.g. while (∼S)V ⊇ ΣV \S V since (∼S ∨ S)V = (∼S)V ∪ S V = ΣV. since (¬S ∧S)V = (¬S)V ∩ S V = ∅. for all families (S i )i∈I ⊆ Subcl Σ and all S ∈ Subcl Σ.

viz. we will make use of the isomorphism αV : P(V ) → Cl (ΣV ) (deﬁned in (4. ˆ −1 Conversely.5)) between the complete Boolean algebras of projections in an abelian von Neumann algebra V and the clopen subsets of its spectrum ΣV. it is useful to think of it as a collection of projections. 1}. Hence. we have ∼S > ¬S and ∼S ∧ S > 0 for all clopen subobjects except 0 and Σ. In fact. Given a projection Pˆ ∈ P(V ). for S ∈ Cl (ΣV ). we write PS := αV (S). then (P) ˆ 2 ∈ {0. This follows easily from the representation of clopen subobjects as families of projections. (5.1) which consists of one projection for each context V. consider (Pˆ S V )V ∈V(N ) = (αV (S V ))V ∈V(N ) . we will use the notation SPˆ := αV (P). Heyting negation and Heyting-regular elements.2) where we used that Pˆ is idempotent and that is multiplicative. then Pˆ S V ≥ Pˆ S V .6) V ⊆V As we saw above. freedom from contradiction. A minimal context is generated by a single . In this section.11)) for the Heyting negation: (¬S)V = { ∈ ΣV | ∀V ⊆ V : |V ∈ / SV } (6. the larger the associated projection Pˆ S V . & & = ∈ ΣV & ˆ PS V = 0 (6. see beginning of next section. a somewhat stronger result holds: for any von Neumann algebra except for C1ˆ = M1 (C) and M2 (C). (This is another instance of coarse-graining.3) = { ∈ ΣV | ∀V ⊆ V : |V (Pˆ S ) = 0}V (6. The fact that S is a subobject then translates to the fact that if V ⊂ V. Negations and regular elements. only the minimal contexts V contained in V are relevant. we will examine the Heyting negation ¬ and the co-Heyting negation ∼ more closely. Throughout.4) = { ∈ ΣV | ∀V ⊆ V : (Pˆ S V ) = 0} (6. the smaller the context V . ˆ = (Pˆ 2 ) = (P) (6. We will determine regular elements with respect to the Heyting and the co-Heyting algebra structure. for the join in the above expression. (6.) If ∈ ΣV and Pˆ ∈ P(V ). holds. TOPOS-BASED LOGIC FOR QUANTUM SYSTEMS AND BI-HEYTING ALGEBRAS 165 This means that the co-Heyting negation does not give a system in which a central axiom of most logical systems. ˆ Given a clopen subobject S ∈ Subcl Σ. §6. We have a glimpse of paraconsistent logic.5) & . We consider the stagewise expression (see eq.

since V is minimal. we have Pˆ (¬¬S)V = Pˆ S V ≥ Pˆ S V (6.166 ¨ ANDREAS DORING projection Qˆ and the identity. it becomes important that we excluded the trivial context V0 = {1} C1.10) V ∈mV = S1− . . Pˆ (¬¬S)V = 1ˆ − (1ˆ − Pˆ S V ) = Pˆ S V . (6.8) We obtain & . We can now consider double negation (¬¬S)V = S1− . we have Pˆ(¬S)V = 1ˆ − W ∈mV PˆS W .e.. (6. but mV = {V }. We have shown: Proposition 3.15) V ∈mV for all V ∈ V(N ). i. so ˆ Pˆ (¬S) V ∈mV V Pˆ (¬¬S) = 1ˆ − V Pˆ (¬S) . it holds that Pˆ S V = Pˆ S V . Let S ∈ Subcl Σ.14) V ∈mV V ∈mV Since Pˆ S V ≥ Pˆ S V for all V ∈ mV (because S is a subobject).12) V ∈mV where mV = {V ⊆ V | V minimal}. if and only if for all V ∈ V(N ). Then Pˆ (¬S)V = 1ˆ − Pˆ S V . & & (¬S)V = ∈ ΣV & ˆ PS V = 0 (6. Thus. Q (6. VQˆ := {Q.11) ˆ PˆS V ∈mV V This shows: Proposition 2. (6. so ¬¬S ≥ S as expected.7) ˆ = Here. ˆ 1} ˆ (6.16) V ∈mV where mV = {V ⊆ V | V minimal}. V (6. ˆ Let mV := {V ⊆ V | V minimal} = {V ˆ | Qˆ ∈ P(V )}. ¬¬S = S.13) V ∈mV For a V ∈ mV. so Pˆ (¬S)V = 1ˆ − Pˆ S V . and let V ∈ V(N ). ˆ = CQˆ + C1. & = ∈ ΣV && 1ˆ − Pˆ S V = 1 (6. An element S of Subcl Σ is Heyting-regular. (6.9) V ∈mV & .

(6. we only have Σ(iV V )(S V ) ⊆ S V . Pˆ (¬¬S)V ≥ PˆS V from equation (6. (6.V Proposition 4.14) that Pˆ (¬¬S)V = V ∈mV Pˆ S V for all V ∈ V(N ). 1} ˆ of V.19) SV V so. we see that o spectral presheaf to the maps V. Consider the minimal subalgebra VPˆS = {PˆS V . Moreover..17) for all V . an element S ∈ Subcl Σ need not be tight. For any S ∈ Subcl Σ. by its deﬁning property ∼S is the smallest element of Subcl Σ such that S ∨ ∼S = Σ. For arbitrary subobjects.22) . 3). (6. For a tight subobject S ∈ Subcl Σ. V ∈ V(N ) such that V ⊂ V. Then. not necessarily maximal.e.V V ˆ Pˆ ∈ P(N ). We saw in equation (6. a maximal abelian subalgebra (masa) of the non-abelian von Neumann algebra N . 3. Outer daseinisation o : P(N ) → Subcl Σ maps projections into the Heyting-regular elements of Subcl Σ. Then Σ(iV V )(S V ) ⊆ S V ⊆ ΣV . Proposition 5. Proof.18) This key formula relates the restriction maps Σ(iV V ) : ΣV → ΣV of the : P(V ) → P(V ). i.VPˆ (Pˆ S V ) = {Qˆ ∈ P(VPˆS ) | Qˆ ≥ Pˆ S V } = Pˆ S V . A clopen subobject S ∈ Subcl Σ is tight if and only if Pˆ S V = (Pˆ S ) for all V . for all V ∈ V(N ).V V V V o ˆ (PS ).1 in [13] shows that Pˆ Σ(i )(S ) = V. we have V o V. tight subobjects are Heyting-regular. (see Def. (6. A clopen subobject S ∈ Subcl Σ is called tight if Σ(iV V )(S V ) = S V (6. V (6. V ∈ V(N ) such that V ⊆ V. Pˆ (¬¬S)V = Pˆ S V = Pˆ S V . Using this. are tight It is clear that all clopen subobjects of the form o (P). Let V be a maximal context. Then clearly (∼S)V = ΣV \S V . Co-Heyting negation and co-Heyting regular elements.20) V ∈m V Corollary 2. TOPOS-BASED LOGIC FOR QUANTUM SYSTEMS AND BI-HEYTING ALGEBRAS 167 Deﬁnition 4. it holds that ¬¬S = S.21) Let V ∈ V(N ).e.15). Thm. V ∈ V(N ) such that V ⊆ V. We remark that in order to be Heyting-regular. and let V. We deﬁne MV := {V˜ ⊇ V | V˜ maximal}. o V. Let S ∈ Subcl Σ be an arbitrary clopen subobject.. since S is tight. i. so Pˆ Σ(iV V )(S V ) ∈ P(V ).

Proof. If V is non-maximal and V˜ is any maximal context ˆ ˆ ˆ containing V.25) V˜ ∈MV = Vo˜ . So it suﬃces to show that for Pˆ(∼S)V = V˜ ∈MV (Vo˜ . we have MV˜ = {V˜ }. then Pˆ (∼S)V = V. (6. On the other hand. so ∼∼S ≤ S as expected. so Pˆ(∼S)V ∨ Pˆ S V ≥ Pˆ (∼S) ∨ Pˆ S = 1. We have shown: .26) V˜ ∈MV W ∈MV˜ Since V˜ is maximal.V V˜ (6. and let V ∈ V(N ). ∼S is a (clopen) subobject.V (1ˆ − Pˆ S V˜ )).V (Pˆ S V˜ ). (6.23) V˜ ∈MV where MV = {V˜ ⊇ V | V˜ maximal}.29) for all V ∈ V(N ).V (1ˆ − (1ˆ − Pˆ S V˜ )) (6.28) V˜ ∈MV Note that the fact that S is a subobject implies that Pˆ (∼∼S)V ≤ Pˆ S V (6. then Pˆ (∼S)V ≥ Pˆ (∼S)V˜ and Pˆ S V ≥ Pˆ S V˜ .24) V˜ ∈MV since (∼S)V.V (1ˆ − Pˆ S V˜ )).V (1ˆ − Pˆ (∼S)V˜ ) (6.168 ¨ ANDREAS DORING Proposition 6. V˜ (1ˆ − ˆ P SW ) . and hence ∼S ∨ S = Σ. must contain all the restrictions of the components (∼S)V˜ for V˜ ∈ MV (and the above inequality expresses this using the corresponding projections). and the above expression simpliﬁes to Pˆ (∼∼S)V = Vo˜ .V o (1ˆ − Pˆ S V ) = 1ˆ − PˆS V and hence P(∼S)V ∨ PS V = 1.V 1ˆ − o W. V˜ V˜ ˆ For the double co-Heyting negation. If V is maximal. V . Then Pˆ (∼S) = V ( o˜ (1ˆ − Pˆ S )). we obtain Pˆ (∼∼S)V = Vo˜ . ∼S is the smallest clopen subobject such that S ∨ ∼S = Σ. we have Pˆ (∼S)V ∨ Pˆ S V = 1ˆ for all V ∈ V(N ). so we must have Pˆ (∼S)V ≥ (Vo˜ . Let S ∈ Subcl Σ. the component at V. (6.27) V˜ ∈MV = Vo˜ .

then the local proposition represented by S V is coarser than the local proposition represented by S V. 7. then ∼∼S = S. By Prop. Pˆ S V ). (6. Outer daseinisation : P(N ) → Subcl Σ maps projections o into the co-Heyting-regular elements of Subcl Σ. it is important to think of an element S ∈ Subcl Σ as a collection of local propositions S V (resp. represented by S V or equivalently by the projection Pˆ S V . if and only if for all V ∈ V(N ) it holds that Pˆ S V = Vo˜ . so V˜ ∈MV V˜ . Proposition 8. if V ⊂ V.31) V ∈mV where mV is the set of all minimal contexts contained in V. i. but the complement of the disjunction of all the coarse-grainings of this local proposition to all smaller contexts. Let S ∈ Subcl Σ be a clopen subobject. . If S is tight. For this interpretation. (6. since Pˆ S V ≥ Pˆ S V for all V ∈ mV. The Heyting complement ¬S is determined at each stage V as the complement of the join of all the coarse-grainings Pˆ S V of Pˆ S V . 2. and let ¬S be its Heyting complement.V (Pˆ S V˜ ). Moreover. one has to consider all the coarse-grainings of this proposition to minimal contexts (which are the ‘max- imal’ coarse-grainings).. For the Heyting negation of the local proposition in the context V. the result follows.e. one for each context V. If S ∈ Subcl Σ is tight. The coarse-grainings of S V are speciﬁed by the clopen subobject S itself.V (PS V˜ ) = PS V . ∼∼S = S. for allo V ∈ V(N ) and for all V ∈ MV.V (PS V˜ ).e.30) V˜ ∈MV where MV = {V˜ ⊇ V | V˜ maximal}. We conclude this section by giving a tentative physical interpretation of the two kinds of negation. TOPOS-BASED LOGIC FOR QUANTUM SYSTEMS AND BI-HEYTING ALGEBRAS 169 Proposition 7. i. An element S of Subcl Σ is co-Heyting-regular. we Proof.32) V˜ ∈MV where MV is the set of maximal contexts containing V. tight subobjects are co-Heyting regular. then ˜ ˆ o ˆ ˆ ˆ have PS V = V˜ . The projection Pˆ(∼S)V is always larger than or equal to 1ˆ − Pˆ S V . as was argued in the proof of Prop. Corollary 3. the local expression for components of ¬S is given by Pˆ (¬S)V = 1ˆ − Pˆ S V . The component of the co-Heyting complement ∼S at a context V is given by Pˆ(∼S)V = (Vo˜ . Physical interpretation. As shown in Prop. the component of the Heyting complement ¬S at V is not simply the complement of S V. 6.V (1ˆ − Pˆ S V˜ )). (6. In other words.. The projection Pˆ (¬S)V is always smaller than or equal to 1ˆ − Pˆ S V .

Pˆ (∼S)V˜ = 1ˆ − Pˆ S V˜ for all maximal contexts V. hence the corresponding local propositions are not mutually exclusive in general. Since S is a subobject. the spectral presheaf Σ. Pˆ(∼S)V is the disjunction of all the coarse-grainings of complements of (ﬁner.170 ¨ ANDREAS DORING This means that the co-Heyting complement ∼S has a component (∼S)V at V that may overlap with the component S V. Summing up. abelian von Neumann subalgebra) V of N ..e. 7). to an element o (P) ˆ of the bi-Heyting algebra Subcl Σ. considered double negation and characterised Heyting regular elements of Subcl Σ (Prop. one can consider . ˜ At smaller contexts V. We characterised the two forms of negation arising from the Heyting and the co-Heyting structure on Subcl Σ by giving concrete stagewise expressions (see Props. Daseinisation maps a propositions of the form “A ε Δ”. The component (∼S)V is deﬁned in such a way that all the stronger local propositions at maximal contexts V˜ ⊃ V are complemented in the usual sense. 3) is a convenient bridge between the usual Hilbert space formalism and the new topos-based form of quantum logic. we have some coarse-grained local proposition. there is a built-in form of coarse-graining which guarantees that if V ⊂ V is a smaller context. that will in general not be disjoint from (i. which comes with a host of well-known conceptual and interpretational problems. The co-Heyting negation hence gives local propositions that for each context V take into account all those contexts V˜ from which one can coarse-grain to V. each component S V of a clopen subobject represents a proposition about the value of a physical quantity in the context (i.e. we have shown that to each quantum system described by a von Neumann algebra N of physical quantities one can associate a (generalised) quantum state space. 2 and 6). mutually exclusive with) the local proposition represented by Pˆ S V .. together with a complete bi-Heyting algebra Subcl Σ of clopen subobjects. The main thrust of this article is to replace the standard algebraic repre- sentation of quantum logic in projection lattices of von Neumann algebras by a better behaved form based on bi-Heyting algebras. Elements S can be interpreted as families of local propositions. §7. stronger) local propositions at contexts V˜ ⊃ V.e.. Conclusion and outlook. i. represented by a projection Pˆ in the complete orthomodular lattice P(N ) of projections in the von Neumann algebra N . The map called outer daseinisation of projections (see Def. represented by Pˆ (∼S)V . Instead. then the local proposition represented by S V is coarser than the proposition represented by S V. 3) as well as co-Heyting regular elements (Prop. It turns out that daseinisation maps projections into Heyting regular and co-Heyting regular elements of the bi-Heyting algebra of clopen subobjects. where ‘local’ refers to contextuality.e. i.. Instead of having a non-distributive orthomodular lattice of projections.

Abramsky. since each topos possesses an internal higher-order intuitionistic logic. Soares Barbosa. Finally.-) ). Abramsky and A. New Journal of Physics. p. there is no direct notion of points available. for giving me the oppor- tunity to organise a Special Session on “Logic and Foundations of Physics” for the 2010 North American Meeting of the ASL. Acknowledgements. Generalised notions of topology such as frames will be useful to study the connections with bitopological spaces. Proceedings 8th International Workshop on Quantum Physics and Logic. Hence. it is a presheaf which has no global elements. But the spectral presheaf Σ is not a topological (or bitopological) space in the usual sense. S. [2] S. see [17]. this provides a distributive form of quantum logic. I am very grateful to the ASL. Mansﬁeld. The anonymous referee also provided some very useful suggestions. Rather. who read the manu- script at an early stage and made some valuable comments and suggestions. Dominique Lambert’s recent talk at Categories and Physics 2011 at Paris 7 served as an eye-opener on paraconsistent logic (and made me lose my fear of contradictions . 2010. I would like to thank Chris Isham and Rui Soares Barbosa for discussions and support. Washington D. 13 (2011). Nijmegen. which makes it impossible to deﬁne a set underlying the topology (or topologies). . This also provides the means to go beyond propositional logic to predicate logic. a non-distributive lattice with an orthocomplement has been traded for a distributive one with two diﬀerent negations. vol. Valentina Harizanov and Jennifer Chubb personally. but more general structures used in the study of quantum logic(s) remain to be considered.C. and to Reed Solomon. In particular. and R.. A generalisation to complete orthomodular lattices is immediate. We conclude by giving some open problems for further study: (a) It will be interesting to see how far the constructions presented in this article can be generalised beyond the case of von Neumann algebras. (c) All the arguments given in this article are topos-external. TOPOS-BASED LOGIC FOR QUANTUM SYSTEMS AND BI-HEYTING ALGEBRAS 171 a complete bi-Heyting algebra of propositions. Roughly speaking. Many thanks to Dan Marsden. March 17–20. There is an internal analogue of the bi-Heyting algebra Subcl Σ in the form of the power object PO of the so-called outer presheaf. The sheaf-theoretic structure of non-locality and contextuality. REFERENCES [1] S. (b) Bi-Heyting algebras are related to bitopological spaces. Brandenburger. The cohomology of non-locality and contextuality. 113036. see [3] and references therein. which I incorporated. so one can op study many aspects internally in the topos SetV(N ) associated with the quantum system.

von Neumann. 823–843. 032101. [13] . 2 (2009). [3] G. vol. 753–937. Quantum states as generalised valuations. Boston. Communications in Mathematical Physics. Heunen. editors). Categories of systems. vol. editors). 2011. vol. 271–313. 053518. Gabbay and F. October 27-29. International Journal of Theoretical Physics. Dordrecht. Competitive Models (B. Topos theory and ‘neo-realist’ quantum theory. pp. Annali di Matematica Pura ed Applicata. Topos theory and consistent histories: The internal logic of the set of all consistent sets. Landsman. N. Spitters. Cambridge University Press. no. 2011. 291–301. (2011). vol. and B. Quantum logics. [9] . [20] . vol. 053517. Handbook of Philosophical Logic (G. A topos perspective on the Kochen-Specker theorem: I. Butterﬁeld. J. Halvorson. Cambridge University Press. Halvorson. or somewhere in between? the logic of quantum theory. A topos foundation for theories of physics: III. Doring. Bezhanishvili et al. Birkh¨auser. vol. Electronic Proceedings in Theoretical Computer Science. [6] M. ¨ Kochen-specker theorem for von Neumann algebras. 2011 (B. vol. 2011. pp.172 ¨ ANDREAS DORING Netherlands. A topos foundation for theories of physics: IV. International Journal of Theoretical Physics. 291 (2009). pp. New York. ‘What is a Thing?’: Topos Theory in the Foundations of Physics. Guenthner. 53 (2012). VI. vol. pp. 39 (2009). Bojowald. pp. Foundations of Physics. 38 (1999). Adv. International Journal of Theoretical Physics. or is it false. [4] G. P. pp. p. Journal of Mathematical Physics. special issue on “Quantum Gravity. New Structures for Physics (B. Conceptual aspects and classical analogues. editors). and B. [23] C. M. Dalla Chiara and R. 63–110. pp. C. 1–14. A topos foundation for theories of physics: I. p. 20 (2010). Open Publishing Association. Zeidler. editor). 49 (2008). Mathematical Structures in Computer Science. Coecke. Lett. 827–859. vol. vol. [22] . Dordrecht. Bitopological duality for distributive lattices and heyting algebras. . pp. 2012. p. 44 (2005). [17] . 139–160.1007/s11229-011-9918-4. Bohriﬁcation. J. 129–228. [8] . 359–393. Quantum Field Theory. J. Fauser. DOI: 10. Online ﬁrst. International Journal of Theoretical Physics. 36 (1997). vol. vol. 49 ˘ A) (2008).. Kluwer. Oxford. ed. Jacobs. 813. Tolksdorf. [11] . pp. p. Berlin. pp. Quantum theory and the representa- tion of physical quantities with arrows ( ˆ : Σ → R↔ . Landsman. 2002. 37 (1998). N. 785–814. Giuntini.3620 [quant-ph]. A topos foundation for theories of physics: II. Journal of Mathematical Physics. vol. and B. Quantum states and measures on the spectral presheaf. Deep Beauty (H. Classical and quantum probabilities as truth values. Doring ¨ and C. Intuitionistic quantum logic of an n-level system. [16] . 37 (1936). [5] M. [24] . Proceedings of the 6th International Workshop on Quantum Physics and Logic (QPL 2009). vol. Journal of Mathematical Physics. vol. vol. Sci. 719–752. 2. pp. 053515. Bohriﬁcation of von Neumann algebras and quantum logic. Cosmology and Black Holes”. A topos perspective on the Kochen-Specker theorem: II. Deep Beauty (H. 207–219. editor). Topos quantum logic and mixed states. [12] A. eprint available at arXiv:1111. and E. pp.. editor). Selinger. Heunen. J. L. Isham and J. [14] . 46 (2005). pp. Basel. A topos for algebraic quantum theory. Isham. 270. [7] A. Springer. vol. 207–238. 49 (2008). pp. 95. 49 (2008). The physical interpretation of daseinisation. 2011. [18] C. Caspers. 2009. P. Journal of Mathematical Physics. Journal of Mathematical Physics. pp. Lecture Notes in Physics. vol. p. Formal languages for physics. Synthese. Heidelberg. Spitters. [21] C. The logic of quantum mechanics. P. [15] . Daseinisation and the liberation of quantum theory. Is it true. London. vol. [19] . 053516. 731–759. Electronic Notes in Theoretical Computer Science. 2669–2733. pp. Isham. [10] . Contemporary Physics. Spitters. Birkhoﬀ and J.

E-mail: andreas. Isham. 25–101. Specker. [29] S. [36] C. pp. toposes and modalities. vol. A topos perspective on the Kochen-Specker theorem: III. pp. Rauszer. 72 (1995). vol. Majid. International Journal of Theoretical Physics. vol. 56–140. Berlin. [34] . Lawvere. Como 1990 (A. The problem of hidden variables in quantum mechanics. TOPOS-BASED LOGIC FOR QUANTUM SYSTEMS AND BI-HEYTING ALGEBRAS 173 [25] . Makkai and G. editor). pp. W. Heidelberg. [30] F. 1992. Von Neumann algebras as the base category. Rosolini. pp. vol. Sheaves in Geometry and Logic: A First Introduction to Topos Theory. pp. Lecture Notes in Mathematics. [33] S. Springer. I. 1983. Reyes and H. 1995. J. Ringrose. 83 (1973). 2008. Intrinsic co-heyting boundaries and the leibniz rule in certain toposes. Introduction. Some possible roles for topos theory in quantum theory and quantum gravity. Moerdijk. [27] C. Foundations of Physics. vol. 219–249. II. Springer. R. 36 (1977). International Journal of Theoretical Physics. New York. E. editors). Mac Lane and I. J. pp. New York. vol. C. 39 (2000). 1174. Proceedings. P. pp. Academic Press. vol. and J. 1413–1436. Interval valuations. Butterﬁeld. Majid. Category Theory. [32] S. New York. Kochen and E. Journal of Mathematics and Mechanics. Carboni. pp. 59–87. Semi-boolean algebras and their applications to intuitionistic logic with dual operations. On Space and Time (S. Completeness results for intuitionistic and modal logic in a categorical setting. 1707–1735. Heidelberg. Springer. vol. 1488. vol. Berlin. vol. [31] . [26] . 41 (2002). Heidelberg. 1986. 1991. Pedicchio. Zolfaghari. [28] R. Quantum spacetime and physical reality. 73–87. 25–43. V. 17 (1967). M. 613–639. A topos perspective on the Kochen-Specker theorem: IV. Bi-heyting algebras. Tokyo. Lecture Notes in Mathematics. Cambridge University Press. Hamilton. [35] M. E. and G. Journal of Philosophical Logic.doering@posteo. Fundamentals of the Theory of Operator Algebras. pp. Berlin. [38] G. 279–281. [37] . Model theory for an extension of intuitionistic logic. 30 (2000). Cambridge University Press. Fundamenta Mathematicae. Kadison and J. 25 (1996). Annals of Pure and Applied Logic. pp. Categories in Continuum Physics (Buﬀalo 1982). vol. Studia Logica. 1–16.de . Foundations of Quantum Group Theory. pp. Reyes. New York.

. Their ‘quantum logic’ was cast in order-theoretic terms. which arguably makes it the most technologically successful theory of physics ever. some 75 years later. However. and that it supports probabilistic inference. Association for Symbolic Logic 174 . . Quantum theory underpins so many things in our daily lives including chemical industry. following Schro. The physics and the logic of quantum-ish logic. In 1932 John von Neu- mann formalized Quantum Mechanics in his book “Mathematische Grund- lagen der Quantenmechanik”. then matter . 45 c 2016. the most prominent practical use of logic. A. §1. Logic and Algebraic Structures in Quantum Computing Edited by J. This gives rise to an intrinsically quantitative kind of logic. The work presented here is supported by the British Engineering and Physical Research Council (EPSRC). “Selling categories to the masses”. which truly deserves the name ‘logic’ in that it also models meaning in natural language. THE LOGIC OF QUANTUM MECHANICS – TAKE II BOB COECKE Abstract. von Neumann wrote in a letter to American mathematician Garrett Birkhoﬀ: “I would like to make a confession which may seem immoral: I do not believe absolutely in Hilbert space no more. . in 1935. the US Oﬃce of Naval Research (ONR) and the Foundational Questions Institute (FQXi). as proposed by Birkhoﬀ and von Neumann. Eskandarian and V. Soon thereafter they published a paper entitled “The Logic of Quantum Mechanics” [13]. rather than its particular propositional structure due to the existence of superpositions. energy production and information technology. “In the beginning God created tensor . Chubb.” (sic)—for more details see [73]. that it supports automation. “How computer science helps bringing quantum mechanics to the masses”. has re- mained the same. very much in the spirit of the then reigning algebraic view of logic. This was eﬀectively the oﬃcial birth of the quantum mechanical formalism which until now. with the distributive law being replaced with a weaker (ortho)modular law. The content of this paper reﬂects a series of seminars in 2010–2012 with as titles: “Monoidal categories as an axiomatic foundation”. Harizanov Lecture Notes in Logic. or the actual title of this paper itself. We put forward a new take on the logic of quantum mechanics. We thank David Corﬁeld and Pascal Vaudrevange for feedback on a previous version. then speech”. ¨ dinger’s point of view that it is composition which makes quantum theory what it is. merely three years after the birth of his brainchild. . the latter being the origin of logic.

von Neumann turned his attention to the actual physical concepts behind quantum theory. Similarly. arbitrary suprema and inﬁma exist. one does not ﬁnd a trace of this activity neither in the mainstream physics. His quantum logic gave the resulting ‘algebra of physical properties’ a privileged role. Hence. 2 Distributivity means that for any elements a.1. etc. . In order-theoretic terms this means. From this it easily follows that the distributive law2 breaks down: given atom3 r = p. and the large body of research that has been produced in the area. 3 An atomis an element p = 0 which is such that whenever a < p then a = 0. The use of linear algebra and complex numbers in so many diﬀerent scientiﬁc areas. a partially ordered set with a minimal element 0 and maximal element 1. mathematics. and in which each pair of elements has a supremum and an inﬁmum. we can also model any movie by means of the data stream that runs through your cables when watching it. . What went wrong? 1. i. The mathematics of it. a vector space over it. Let us consider the raison d’ˆetre for the Hilbert space formalism. 1 I. the continuum structure. clearly show that quite a bit of modeling can be done using Hilbert spaces. 75 years later quantum logic did not break through. All of this leads us to . . THE LOGIC OF QUANTUM MECHANICS – TAKE II 175 This resulted in a research community of quantum logicians [68. he simply used it because it happened to be ‘available’. inner-product structure. b. 47. 75 years later one may want to conclude that this activity was a failure. the notion of a physical property and the structure imposed on these by the peculiar nature of quantum observation. So why would one need all this ‘Hilbert space stuﬀ’. i. the physics of it. Birkhoﬀ and von Neumann as well as many others believed that understanding the deep structure of superposition is the key to obtaining a better understanding of quantum theory as a whole. On the other hand. there are physical resons for assuming that this lattice is complete [71.e.e. superposition means that the strongest property which is true for two distinct states is also true for states other than the two given ones.2. nor logic literature. Why? According to von Neumann. Birkhoﬀ and von Neumann crafted quantum logic in order to emphasize the notion of quantum superposition In terms of states of a physical system and properties of that system. 1. But as already mentioned. .e. we should rather turn our attention to the stuﬀ that is being taught at drama schools and directing schools. However. q with r < p ∨ q we have r ∧ (p ∨ q) = r while (r ∧ p) ∨ (r ∧ q) = 0 ∨ 0 = 0. . 71. as well as results in model theory. 30]. the ﬁeld structure of complex numbers. more speciﬁcally. . In fact. 69]. But does this mean that these data streams make up the stuﬀ that makes a movie? Clearly not. that the join p ∨ q of two atoms p and q is also above other atoms. c of the lattice we have that a ∧ (b ∨ c) = (a ∧ b) ∨ (a ∧ c) and that a ∨ (b ∧ c) = (a ∨ b) ∧ (a ∨ c). despite von Neumann’s reputation. representing states by the atoms of a lattice1 of properties [69].

It is called symmetric monoidal category[11]. how do we describe multiple quantum systems given that we know how to describe the individual quantum ¨ systems. which we denote by ⊗. Schrodinger pushed forward the idea that the stuﬀ which truly characterizes quantum behavior is precisely the manner in which quantum systems compose [74]. In other words. . most notably exponential quantum computational speed-up which led to the quantum computing paradigm [79]. • Task 1. given that the structure deduced in Task 0 applies to a wide range of theories (as we shall see below in Section 2) what extra structure do we need to add such that the resulting framework allows us to derive typical quantum behaviors. On the other hand. Over the past 30 years or so we have seen ample evidence for this claim. . we take on the challenge to ﬁnd this same structure elsewhere in what we usually conceive as ‘our classical reality’. The game plan. we want to axiomatize composition of systems. And. In fact. 32]. that is. The solution to this has been around for quite a while. Now reversing the roles. . Next we investigate which additional assumptions on ⊗ are needed in order to deduce experimentally observed phenomena? That is. But one can show that an arbitrary symmetric monoidal category is always ‘categorically . also in 1935. without any reference to underlying spaces. we want to know what remains of the Hilbert space formalism if we ‘remove all of its structure except for the manner in which systems compose’. while set theory based models such as the Hilbert space model are typically non-strict. which invokes so-called ‘coherence conditions’ [67] between ‘natural transformations’ [45]. First we want to solve: tensor product structure = ??? the other Hilbert space stuﬀ that is. including superposition? 1. here are the resulting outcomes: • Outcome 0: That was an easy one. • Task 2. This may involve looking at this classical reality through a ‘novel pair of glasses’. physical processes themselves form a strict symmetric monoidal category. and the focus on quantum information processing has revealed a wide range of quantum phenomena which all crucially depend on the manner in which quantum systems compose. rather than explaining all of quantum theory in terms of superposition. can we maybe explain all of quantum theory in term of the manner in which quantum systems compose.3. Here is the list of tasks we’ve set ourselves: • Task 0.176 BOB COECKE The Achilles’ heel of quantum logic is the fact that it fails to elegantly capture ‘composition of quantum systems’. as discused in [23. Once this ‘typically quantum’ structure has been identiﬁed. So-called ‘quantum non- locality’ was experimentally conﬁrmed.

• Outcome 1c: The diagrammatic framework underpinning strict symmet- ric monoidal categories has meanwhile been adopted by several leading researchers in quantum foundations e. in terms of their graphical language [70. that is. [49. and in the case that one could derive something physically relevant one had to work really hard. and provided novel insights in the nature of quantum non-locality [27.g. and the required computations are utterly trivial.” As we shall see below. exposing this structure has already helped to solve standing open problems in quantum information. here we will only spell out strict symmetric monoidal categories. up to isomorphisms. which means that.g. This is in sharp contrast with Birkhoﬀ-von Neumann quantum logic where one couldn’t derive much. with very little additional structure one can already derive a wide range of quantum phenomena. [44. 26]. 57. THE LOGIC OF QUANTUM MECHANICS – TAKE II 177 equivalent’ to a strict symmetric monoidal category. . a language which is such that an equational statement holds in it if and only if it follows from the axioms of a strict symmetric monoidal category. whatever one can do with a non-strict one. • Outcome 1b: Moreover. . 54]. Hence. 14].] we join the quantum picturalism revolution [23]”. one can do with a strict one too. 17. e. • Outcome 2a: Observe the following similar looking pictures: . quoting Lucien Hardy in [54]: “[. • Outcome 1a: Quoting Princeton philosopher Hans Halvorson in his editorial to the volume Deep Beauty: Understanding the Quantum World through Mathematical Innovation which marked 75 years since the pub- lications of von Neumann’s quantum formalism [52]: “What is perhaps most striking about Coecke’s approach is the sheer ratio of results to assumptions. 59].

and in doing so produced important new applications in each of these areas: . statements like this are still tightly related to a truth-concept. f) ∨ Eat(b. P vs. then got sick. . how do meanings of words interact to form meanings of sentences? • Use: automated reasoning. The graphical calculi are in each case very similar. Could this be pointing at the existence of some sort of quantitative logic. quantiﬁers. Note in particular that in each case the data of interest is of a fundamentally quantitative nature. but also generate them. Clearly there is a lot more to the meaning of a sentence than it either being true or false. f)) ⇒ Sick(a). and predicates Sick(person) and Eat(person. . 23]. Sick(b) However.4. which are key to modern methods for verifying the correctness of new software and hardware. variable f referring to food. more speciﬁcally. is that it is about ‘arguments in natural language’. . This leads us to the following questions: What do we mean by meaning? What is the logic governing meaning. 20. 1. which is not typical to these areas but of a more universal nature? So let us now consider . Consider for example the sentence:“Alice and Bob either ate everything or nothing. tracing back to Aristotle. some kind of food) we can formalize this as follows: (∀f : Eat(a. which points at a common reasoning system in each of these very distinct areas.” By using connectives. the logic of it. 37]. a linguistics paper on how to compute the meaning of a given the meaning of its words [19. Rather than making a case for one or another logical paradigm we will take a pragmatic stance and conceive logic in terms of its origin and its most prominent practical use: • Origin: structure in natural language. constants a(lice) and b(ob). where one does not only try to automatically prove theorems. Even more adventurous is automated theory exploration.178 BOB COECKE These are respectively taken from a physics paper on the ﬂow of infor- mation in quantum protocols [1. Logic also controls robot behaviors in artiﬁcial intelligence. NP)—see also Figure 1. we classify statements in terms of these either being true or not. . What is logic? The previous century has known a huge proliferation of logics of various kinds. that is. The origin of logic. Our diagrammatic framework appeals to both of these senses of logic. . and there probably are as many opinions of what logic actually is. f) ∧ Eat(b. f)) ∨ ¬(∃f : Eat(a. which is a much harder task (cf. and a probability theory paper that axiomatizes Bayesian inference [38]. Logic now forms the foundation for ﬁelds like automated proof checking and automated theorem proving in computer science.

• The diagrammatic formalism underpins the automated reasoning software quantomatic developed at Oxford and Google by Dixon. represented by boxes. The data of a minimal process logic consists of processes. Sadrzadeh and the author in [19. We explain this framework in Section 4. 42. 2. We refer the reader to [23. details are in [40. It was a cover heading feature in New Scientist [5] and meanwhile greatly improved performance of several natural language processing tasks [51]. Kissinger. 41] and on the quantomatic website. We explain this structure in terms of its graphical language. 37]. building further on the work done at Edinburgh [58]. as well as its structural relationship the graphical quantum formalism. We could as well have given a symbolic presentation. THE LOGIC OF QUANTUM MECHANICS – TAKE II 179 Figure 1. introduced by Clark. More recently. Soloviev and Frot—see also Figure 2.1. • The above depicted framework for modeling how meanings of words interact to form meanings of sentences. Graphical language. §2. The theory[mine] website which allows one to buy an automatically generated theorem and name it after someone. We won’t discuss this here. Merry. By a process logic we simply mean any strict symmetric monoidal category. each of which takes some type of systems as . Duncan. Minimal process logic. 32] for such a symbolic presentation. and by minimal we mean that at this stage we consider no structure (yet) other than the strict symmetric monoidal structure. and how they compose to make up recipes—[23] also discusses how a process logic explains why tigers have stripes while lions don’t. is the ﬁrst to do so based on a clear conceptual underpinning. It is a novelty gift spin-oﬀ from the automated theory exploration expertise at Edinburgh University—see [15] for the science. exempliﬁed for the speciﬁc case of cooking processes. work on automated theory exploration of graphical theories also started at Oxford [61].

grammatical types. Screenshot of the quantomatic software de- veloped in a collaboration between Oxford and Google. or a sentence.google. Those with neither an input type nor an output type are called values. A process without an output type is called a valuation. verb.e. represented by (an) input wire(s). classical data of a certain size.180 BOB COECKE Figure 2. representing ‘no system’: one system n sub-systems no system -. its input. and some type of system as its output: These types may be compound./0 1 n 0 Examples of types could be a particular quantum system./0 .com/site/ quantomatic/. e. etc.g./ 0 -. or trivial. A process with no input wire is called a state—one can think of these as ‘preparation processes’. There are two modes: sequential or causal or connected composition. the type of a noun. The connectives of a minimal process logic constitute composition of processes. . which can be downloaded from http://sites. . i.

Note that sequential composition requires the output type of f to be equal to the input type of g while no such restriction exists for parallel composition . respectively depicted as: So by post-composing a state with a valuation one obtains a value. Hence computation proceeds by topological deformations: . 28]. The formal paradigm underpinning minimal process logic is a topological one: The topology captures ‘what interacts with what’. THE LOGIC OF QUANTUM MECHANICS – TAKE II 181 Figure 3. 34.g. with the help of some new graphical elements. 36. and. a wire standing for interaction while no wire stands for no interaction. Examples of quantum mechanical concepts that can be expressed in purely topological terms. parallel or acausal or disconnected composition . They are taken from [22. see Figure 3 for some topologically characterized quantum mechanical concepts. It is surprising how many concepts can be expressed purely in these topological terms—e. 25. The computational content of minimal process logic boils down to the simple intuitive rule that topologically equivalent diagrams are equal.

3. k after h’ is the same as ‘g while k. the ‘bifunctoriality equation’ (g ◦ f) ⊗ (k ◦ h) = (g ⊗ k) ◦ (f ⊗ h) of monoidal categories becomes: In terms of processes this means that ‘g after f. f while h’. Note that this is exactly how the highly successful Dirac notation [39] works: a ket | can be turned into a bra |.182 BOB COECKE There is no additional equational content to a minimal process logic. The ﬁrst bit of extra structure will induce some kind of metric on the states. Quantum process logic . since a strict symmetric monoidal category is subject to a number of axioms. This may sound surprising.1. identiﬁed by Abramsky and the author in [1. an inner-product.e. applying this valuation to any other state will yield a value. Our next goal is to derive some non- trivial quantum phenomena by endowing a minimal process logic with a tiny bit of extra structure. Dagger compact structure. we will ask that each state can be turned into a valuation. after. §3. For example. and when composing | with another ket |φ we obtain a bra-ket |φ i. namely. denoting sequential composition by ◦ and parallel composition by ⊗. The explanation is that in the graphical language all these equations become tautologies. while. we will allow for the inputs and the outputs of any process to be ‘ﬂipped’: ∀ ∃! .Take IIa. 2]. Since states may themselve arise by composing processes other than states.

for which there exists a measurement that yields a particular outcome with certainty. Note in particularly that for non-Boolean lattices an orthocomplementation is a structure. diagrammatically: quantum = classical That is. an order-reversing involution. the bad way to ﬁx an old fashion fuse by means of a copper wire. 5 Note that a state of two systems doesn’t have inputs. so ﬂipping is involutive. However. while it reverses sequential composition . THE LOGIC OF QUANTUM MECHANICS – TAKE II 183 Note that ﬂipping twice yields the original box. as there can exist many diﬀerent ones on the same lattice. . in quantum theory this already occurs for states on which there is no uncertainty. and it is also clear that it preserves parallel composition . we will assert that pure quantum states admit entanglement. a quantum state of two systems can in general not be described by describing the state of its parts. That is. We refer to ﬂipping as the adjoint or dagger4 So far we haven’t said anything speciﬁc about the parallel composition . Now ¨ we will truly follow Schrodinger’s path and specify in which manner quantum systems compose diﬀerently than classical systems. that is. such a cup-shaped state for example obeys: (1) The equivalent symbolic expression for this equation would be: ( ⊗ 1) ◦ (1 ⊗ )=1 4 From the perspective of Birkhoﬀ-von Neumann quantum logic. not a property.g. can also not be described by independently describing the state of each system. namely f † (a) = (f ∗ (a )) [46. So how do we provide a constructive witness for the fact that the state of two systems does not ‘disconnect’ in two separate one-system states? Simply by explicitly introducing a special two system state which is obtained by (internally) connecting its two outputs with a cup-shaped wire:5 Sticking to our topological paradigm. Note that this is also not the case for probabilistic classical data: a situation of two systems which comes with the promise that the states of the system are the same but unknown. In other words. In lattice theoretic terms the linear algebraic adjoint indeed arises as an expression involving Galois adjoints (−)∗ and orthocomplementation (−) . one could conceive this as the analog to an orthocomplementation on the lattice structure. 29]. so this is more like ”internal wiring” of its outputs e.

and cap-shaped wires also enable us to ‘deﬁne’ the transpose which we depict by rotating a box 180o : It then immediately follows that we have: that is. Cup. something to which we refer as unitarity Hence: . 3.e. that is. Now for some physics. 2]. We have: and we choose f such that its composite with its adjoint yields the identity.and cap-shaped wires.184 BOB COECKE where 1 stands for a single straight wire and is obtained simply by ﬂipping i. Deriving physical phenomena. We assumed the existence of an adjoint for any box and represent it via ﬂipping. we can slide boxes across cup. We obtain a strict dagger-compact category [1. its adjoint.2. Going berserk. we can treat the entire graphical calculus for dagger compact categories in terms of beads which slide on wires.

Logical completeness wrt Hilbert spaces. THE LOGIC OF QUANTUM MECHANICS – TAKE II 185 Introducing agents Alice and Bob yields quantum teleportation: Note that. the entanglement swapping protocol [81]: So how much quantum mechanics can we derive in this calculus? 3. so one had to rely on sheer luck to discover it. Selinger [60. Evidently there are many dagger compact categories.3. . 77]). A more detailed discussion of this graphical derivation and its physical interpretation is in [20. The diagrammatic language presented above is directly related to the symbolic notion of a dagger compact category as follows: Theorem 1 (Kelly-Laplaza. to mention two: • Wires represent ﬁnite dimensional Hilbert spaces. given that the quantum mechanical formalism was born in 1932. Similarly we derive another quantum mechanical feature. that this phenomenon took 60 years to be discovered [12]. and the parallel composition of wires is the tensor product while parallel composition of boxes is the Kronecker product. 23]. An equational statement be- tween expressions in the dagger compact categorical language holds if and only if it is derivable in the above described graphical calculus. the dagger is the linear algebraic adjoint. The standard quan- tum mechanical formalism provides no indication whatsoever that something like this would be possible. sequential composition is ordinary function composition . boxes linear maps.

g. • Bell-state. An equational statement between expressions in dagger compact categorical language holds if and only if it is derivable in the dagger compact category of ﬁnite dimensional Hilbert spaces. . holds in quantum theory if and only if it can be derived in the graphical language. . • adjoints (e. On the other hand. boxes relations. . linear maps. • complete positivity (cf. Natural language process logic. positivity.e. Before continuing with the further de- velopment of quantum process logic. there doesn’t have to be an exact match. the dagger is the relational con- verse. . conjugation. Search engines such as google and other natural language processing tools also have an understanding of meanings of words which they use to provide us with the most relevant outputs for our queries. the from- word-meaning-to-sentence-meaning process. • inner-product. in particular. Hence one could wonder how much one can actually derive in (the graphical calculus for) dagger compact categories. for example. . and parallel composition is the cartesian product. §4.1. Hilbert-Schmidt norm. Theorem 2 (Hasegawa-Hofmann-Plotkin. Evidently these two examples have very diﬀerent spaces and one would evidently not associate sets and relations with quantum processes.186 BOB COECKE • Wires represent sets. The description of the compact structure for each of these as well as some more examples can be found in [32]. From word meaning to sentence meaning. sequential composition is composition of relations. . projections. provided we understand all of its words. [77]). There is a technological side to this. since surely. 78]). The model of word meaning which these engines employ enables them to produce outputs that include words that are closely related to the words in our query. To put this in more quantum physics related terms. eﬀects. So how do we know what a sentence means? There must be some kind of mechanism. Meaning here manifestly goes beyond simply assigning truth values to sentences. any equation involving: • states. . linear-algebraic trace. This can mean many things.Selinger [55. operations. one has a dictionary available. . for transforming the meaning of words into the meaning of a sentence. . . . used by all of us. transposition. we turn our attention on something completely diﬀerent: meaning in natural language. . . Bell-eﬀect. we all understand sentences that we may have never heard before in our lives. Consider as given the mean- ings of words. tensor product and linear algebraic adjoints. i. self-adjointness and unitarity). . The answer is surprising. . . there are no dictionaries for entire sentences. 4.

Neither of these point me in the direction of appropriate vehicles that would serve my purpose. “How to Sprint Faster: 6 steps . . we would like to have a ﬁxed type for the meaning of all sentences.]”. Hence there is some process. which transforms the meanings of the string of words in the meaning of the sentence made up from these: What drives this process? That is. given a string of words.Retired Police Oﬃcer & [. the overall wire structure. . . what mediates their interaction? The answer is obvious: since grammatically incorrect sentences have no clear meaning anyway. depends on the grammatical structure of the sentence. Now. searching on Google for “I want something that allows me to go faster than when I only use my legs” returns among its top hits: “Dif- ference Between Oxycontin and Oxycodone”. THE LOGIC OF QUANTUM MECHANICS – TAKE II 187 However. representing grammatical types of words by wires and their meanings by state-boxes we can depict a string of words as: But the overall type. The reason is the lack of a theory that produces the meaning of a sentence from the meanings of its words.wikiHow’.com . . “What are good ways for a girl to [XXX]”. We can now describe the problem for from-word-meaning-to-sentence- meaning processes in more precise terms: • Given a theory of word meaning. the from-word-meaning-to-sentence-meaning process.Onelegtim. However. and given a theory of grammar. “My Story .e. sentences with diﬀerent grammatical structure may have the same meaning. so clearly there is no understanding of the meaning of my query.]” and “Golf Swing Power: What Your Legs Should Be Doing [. i. and more generally. whatever the manner is in which we describe the meaning of words. how can we combine these into an algorithm which produces the meaning of sentences from the meanings of its words? .

all novels. 4. and works as follows. It takes inspiration from Wittgensteins’ philosophy of ‘meaning is use’ [80]. etc. the internet. the corpus. Let’s stay at an abstract level a bit longer.natcorp. a small 6 This can be accesses at http://www.188 BOB COECKE As already mentioned. it would be suﬃcient for one person to hate another person in order for everyone to hate everyone. whereby meanings of words can be determined from their context. all editions of a certain newspaper.ox. and considers an n dimensional vector space with chosen basis where each basis vector represents one of the context words. The current dominant theory of word meaning for natural language processing tasks is the so-called distributional or vector space model of meaning [75]. Then one selects a huge body of written text. One ﬁxes a collection of n words. but here we already anticipate the description of grammatical structure that we discuss below. before we will describe concrete theories of word meaning and grammar. which could have dramatic consequences. to turn inputs into outputs and represent the verb as: You may ask where these cups suddenly come from. Note in particular also that for these kinds of word-states we again have: since otherwise. . this problem was addressed by Clark. as deﬁned above. For example.g. for the case of a transitive verb.1. What is a verb? A transitive verb is something that requires an object and a subject in order to yield a grammatically correct sentence. the meaning of the sentence would not depend on the meanings of the nouns. Sadrzadeh and the author in [19. and one producing the sentence: Since we rather represent a verb as a state we can use transposition. E. considering the verb ‘hate’. two respectively requiring an object and a subject.uk/. the British National Corpus6 which is a 100 million word collection of samples of written and spoken language from a wide range of sources. 37].1. So we can think of a transitive verb as a process with three wires. the context words. A theory for word meaning.ac. that is. Next one decides on a scope.

Algebraic gadgets that govern grammatical types have been around for quite a bit longer [4. and these would allow one to derive correct statements such as n · tv · n ≤ s. and. Now we have to assign grammatical types to the elements of a pregroup. while others like transitive verbs will be assigned compound types. so the string of types ‘noun transitive verb noun’ indeed makes up a grammati- cally correct sentence. The caps here represent the equations x · −1 (x) ≤ 1 and (x)−1 · x ≤ 1. . Nn (a)). There are several variants available. THE LOGIC OF QUANTUM MECHANICS – TAKE II 189 integer k. there are additional operations subject to certain laws which make up the actual structure of the algebra. 18. hence ! ! ! n · tv · n = n · −1 (n) · s · (n)−1 · n = n · −1 (n) · s · (n)−1 · n ≤ 1 · s · 1 = s. these additional operations are a left inverse −1 (−) and a right inverse (−)−1 . z has as its overall type t. . A theory of grammar. These algebras also have a relation ≤ where a · · · · · z ≤ t means that the string of types a . and for each context word x one counts how many times Nx (a) a word a to which one wants assign a meaning occurs at a distance of at most k words from x. as well as to 1 ≤ −1 (x) · x and 1 ≤ x · (x)−1 . . the meaning vector of a. . . (n)−1 and n. Now. . Philosophically. here we will focus on Lambek’s pregroups [65]. a transitive verb and a noun make up a sentence s.2. For example. (n)−1 and n: Then. . in particular. one can simply compute the inner-product of their meaning vectors. indecomposable. in order to compare meanings of words. n · tv · n ≤ s expresses the fact that a noun. One obtains a vector (N1 (a). Concretely.1. . . each with their pro’s and con’s. how closely their meanings are related. φn (a)). . . this is not just an analogy with the graphical language of compact categories. this is only manifest in that these algebras all have a composition operation that allows to build larger strings of words from smaller strings of words. cancel out: so that at the end the only remaining system is the sentence type. However. we use caps to indicate that n and −1 (n). For the speciﬁc case of pregroups. Finally. −1 (n). s. i. which one normalizes in order to obtain (φ1 (a). In fact. 64]. Some will be atomic. these algebraic gadgets trace back to Freges’ principle that the meaning of a sentence is a function of the meaning of its parts [48].e. any partial order is a category. 4. subject to x · −1 (x) ≤ 1 and (x)−1 · x ≤ 1 where 1 is the unit for the composition operation. tv = −1 (n) · s · (n)−1 . Pregroups are in fact compact categories! To see this. We start with ﬁve systems of respective types n. 10. We can depict this computation graphically as follows.

here we are dealing with a non-symmetric tensor 4. which have a clear ‘logical’ meaning. the verb now acts as a fuction on the object and the subject: The meanings of all sentences live in the same vector space so we can again simply use the inner-product to measure their similarity. Using cups we can model the meaning of . and while equations x · −1 (x) ≤ 1 and (x)−1 · x ≤ 1 provide caps. substitute the sentence types by meaning vectors of the particular words we are interested in. as in Section 4. By rewriting this using transposition. and the vector spaces for word meaning when organized as a dagger compact category. and substitute the caps by the vector space caps. More details on this are in [37]. The reason that there are two kinds of caps and cups is the fact that we are not allowed to change the order of words in a sentence while two physical systems do not come with some ordering.2. Here is an example of this: As above. Grefenstette and Sadrzadeh have recently exploited this theory for standard natural processing tasks and their method outperforms all existing ones [51]. The structural similarity between the pregroup theory of grammar. In category-theoretic terms. Combining theories.190 BOB COECKE the composition provides the tensor. is exactly what we will exploit to explicitly construct the from-word-meaning-to-sentence-meaning process. What about the cups? They can be used to model ‘special words’ like “does” and “not”.1. the wire structure here is obtained from the types of these words according to the pregroup grammar. equations 1 ≤ −1 (x) · x and 1 ≤ x · (x)−1 provide cups. We consider the graphical representation of the proof of grammatical correctness of a sentence. so we obtain: where the dotted line indicates the linear map that when applied to the vector −−−→ −−−→ −−→ Alice ⊗ hates ⊗ Bob produces the vector that we take to be the meaning of a sentence.

7]. mainly within topology. THE LOGIC OF QUANTUM MECHANICS – TAKE II 191 ‘does’. . and obtain a grammatical quantum ﬁeld 7 If the ﬁeld has a non-trivial involution then this category has a dagger too ( = transposition). to the compact category of vector spaces over some ﬁeld K.3. An interesting analogy arises. one can now ask the question: why are there algebraic gadgets that describe grammatical correctness. rather than taking a category of topological structures as domain. that is. to interact. This example also shows how the wires are mediating the ‘ﬂow’ of word meaning in sentences. While it takes its inspiration from quantum ﬁeld theory. which was ﬁrst observed by Louis Crane. it has become an area of research in its own right. i. They allow for the words Alice and like. i. An aside: quantizing grammar. and which is discussed in detail in [72].e.e. ‘does nothing really’. Turning things upside-down. we can take a pregroup as domain. for which we use an input-output not-box that does just that: and then we can simply use homotopy to compute: which is exactly what we would expect the meaning of ‘Alice does not like Bob’ to be: the negation of Alice liking Bob. that is. while far apart in the sentence. The object of study is a monoidal functor: F : nCob → FVectK :: → V from the compact category of closed (n − 1)-dimensional manifolds with diﬀeomorphism classes of n-dimensional manifolds connecting the closed (n − 1)-dimensional manifolds as morphisms. 4. ‘negates meaning’. 8. An important area of contemporary mathematics is the study of Topological Quantum Field Theory (TQFT) [6. why do these even exist. and ‘not’.7 Now. Our theory of word meaning explains this: they witness the manner of how word meanings interact to form the meaning of a sentence. a category of grammatical structures.

and in particular. and re- lationships between these e. This was established in a series of papers by Pavlovic. Dagger compact categories cap- ture a substantial number of quantum mechanical concepts. 27. . 26].(1)? Similarly to ‘however one bends a wire. it still remains just a wire that acts as an identity’. observables themselves. But they are by no means universal with respect to quantum theory. independent of how the web is build up. . Examples of FdHilb-processes not expressible in dagger compact language are basic quantum computa- tional gates such as the CNOT-gate. complementarity. The calculus was also rich enough to address a number of concrete quantum computational and quantum foundational problems e. Quantum process logic . 25].3. 14. 57. and the dagger compact category FdHilb related to the von Neumann model described in Section 3. and the author in [34. see [44. and.g. eq.m So what is the analogue of the topological calculus with cups and caps. 33. We refer to these dots as . . is again the .Take IIb.192 BOB COECKE theory: F : Pregroup → FVectR+ :: → V §5. Duncan. phase-gates etc. • that the language is not rich enough to describe all processes in FdHilb. ⎧ m ⎫ ⎪ ⎪/ 0. 24. We will now present an extended graphical language which does capture all of these. any web of spiders with the same overall number of inputs and outputs. Examples of concepts that are not captured by dagger compact language are the classical data obtained in measurements.g. by which we can mean two diﬀerent things: • that they do not capture all quantum mechanical concepts. . is complete with respect to them. Paquette. Rather than only allowing for wires we allow for ‘dots’ at which wires branch into multiple wires./ 0⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ n ⎪ ⎪ ⎪ ⎩ ⎪ ⎭ n.⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎨ ⎪ ⎬ ‘spiders’ = ⎪ ⎪ ⎪ ⎪ ⎪ ⎪. or none.

Since bases allow to represent observables and classical data. except for the fact that quantum theory only becomes interesting if we consider several ‘incompatible’ bases. . 63. they fuse together. This in particular implies that for the speciﬁc spiders: 2 0 / 0. . 31]. THE LOGIC OF QUANTUM MECHANICS – TAKE II 193 same. . Again in other words. . . . we almost reached our goal. . . the rule governing spider calculus is that if two spiders ‘shake legs’. which is a non-trival result./ 0 2+0−1 so reasoning with spiders strictly generalizes reasoning with wires In FdHilb a family of spiders of the above kind on-the-nose captures an orthonormal basis. . it only matters what is connected to what. . Next one shows that these dagger special commutative Frobenius algebras in FdHilb are the same thing as orthonormal bases [35]. one can show that reasoning with those spiders is equivalent to working with a so-called dagger special commutative Frobenius algebra [62. So for any k > 0: m+m −k / 0./ 0 n+n −k Hence./ 0 ./ 0 0 2 we obtain eq. but not the manner in which this connection is realized. and . . Firstly. / 0.(1): 0+2−1 / 0.

8 We do not subscribe (anymore) to conceiving the diagrammatic logic as a ‘degenerate hyper-deductive variant’ of standard logic in categorical form since this does not recognize the quantitative nor the process content. What happens if a dark gray and a light gray spider which represent complementary observables ‘shake legs’? Well. 76]) which is degenerate in the sense that conjunction and disjunction coincide [43. 25. Rather.194 BOB COECKE So now we consider two diﬀerent families of spiders. 3. The remaining challenge. Such a pair of diﬀerently colored spider families that interact in this manner forms the basis of a rich calculus with many more extra features than the ones described here. 56]. . §6. But since this still belongs to the world of speculation. In this paper we pushed forward the idea that the diagrammatic languages describing quantum phenomena as well as meaning-related linguistic phenomena may constitute some new kind of quantitative logic. Bayesian inversion boiling down to nothing but transposition for appropriately chosen cups and caps: This was established by Spekkens and the author in [38]. 8 There is also ongoing work on relating traditional quantum logic with dagger compact categories or related structures at a purely structural level e. we leave this to future writings. to which we refer for details. The compact structure can then be seen as a resource sensitive variant (as in Linear Logic [50. 26] for more details and concrete applications. We refer the interested reader to [23. their ‘legs fall oﬀ’: This was shown by Duncan and the author in [25]. 21].g. the generalization in [33] of Carboni and Walters’ axiomatization of the category of relations [16]. for example. The same logic also governs Bayesian inference. So where does traditional logic ﬁt into this picture? One perspective is to start with standard categorical logic [66. we would like to conceive the quantitative diagrammatic logic as ‘the default thing’ from which traditional qualitative logic arises via some kind of structural collapse. represented by a diﬀerent gray scale. [53. 9]. There are several results that could be taken as a starting point in this direction.

French. I. pp. 84 (2011). pp. pp. AAAI. [23] . pp. Rickles. Carboni and R. topology. vol. vol. M. vol. A compositional distributional model of meaning. Dixon. Johansson. Topological quantum ﬁeld theories. [21] .E. AIP Press. Baez. Comptes Rendus des S´eances de l’Acad´emie des Sciences. pp. Axiomatic description of mixed states from Selinger’s CPM-construction. pp. vol. 133–140. Aron. Crepeau. 3–94. 14(6) (2005). pp. Informational derivation of quantum theory. New Structures for Physics (B. vol. D’Ariano. 120402. Ajdukiewicz. 47–58. vol. 1–27. vol. vol. [3] S. F. 012311. Clark. Atiyah. and P. [8] J. [2] . 31–38. 415– 425. pp. [14] S. vol. Springer. 2008. and W. editors). Physics. Physical Review Letters. Journal of Mathematical Physics. Paris. Abramsky and B. B. 1 (1937). 1. Automated quantum reasoning: Non logic – semi-logic – hyper-logic. Khrennikov.R. arXiv:q-alg/9503002. Die syntaktische Konnexit¨at. 37 (1936). M. Brassard. Abramsky and N. p. vol. Heunen. R. arXiv:quant-ph/0404040. and M. vol. pp. vol. IEEE Computer Society. Cartesian bicategories I. Saatsi. Bennett. New Structures for Physics (B. Coecke. pp. [20] B. C. F. pp. Physical Review Letters. Studia Philosophica. Publications Math´ematiques de l’IHES. logic and computation: a Rosetta stone. pp. Annals of Mathematics. Coecke. G. McCasland. Bar-Hillel. Stay. Coecke. editor). C. [6] M. 11–32. [11] J. Quantum links let computers read. 68(1) (1988). Oxford University Press. 70(13) (1993). Proceedings of the 19th Annual IEEE Symposium on Logic in Computer Science (LICS). . pp. 59–83. Quantum Theory: Reconsid- erations of the Foundations III (A. [13] G. THE LOGIC OF QUANTUM MECHANICS – TAKE II 195 REFERENCES [1] S. Tree models for the description of language. Benabou. H. [22] . IT-2 (1956). [12] C. L. 208(2790) (2010). [16] A. Peres. Higher-dimensional algebra and topological quantum ﬁeld theory. Physical Review A. pp. AAAI Spring Symposium: Quantum Interaction. Theory and Applications of Categories. A categorical semantics of quantum protocols. 2005. G. p. Electronic Notes in Theoretical Computer Science. Jozsa. 1887–1890. Entangled and sequential quantum protocols with dephasing. arXiv:quant-ph/0510032. Abstract physical traces. 175–186. Lecture Notes in Physics. Baez and J. p. 49 (1987). Lecture Notes in Physics. 823–843. 210 (2008). 240–266. and R. Birkhoﬀ and J. Walters. 113–124. Tzevelekos. The Structural Foundations of Quantum Gravity (D. pp. Chomsky. [9] J. 51 (2009). [18] N. Sadrzadeh. 36 (1995). Cavallo. Categories avec multiplication. [17] G. 2011. arXiv:0908. [19] S. Introduction to categories and categorical logic. 10–11. Contemporary Physics. 256 (1963). Extended version: arXiv:quant-ph/0402130. Proceedings of the Second Quantum Interaction Symposium (QI-2008). 2006. New Scientist. K. [7] J. [4] K. editor). [10] Y. Quantum picturalism. The logic of quantum mechanics. 2004. C. 108 (2012). Journal of Pure and Applied Algebra. 6073. no. Dolan. The Theory behind TheoryMine. pp. Chiribella. Transactions on Information Theory. pp. editor). S. Springer. [5] J. pp. 2011. Coecke. Kindergarten quantum mechanics — lecture notes. Perinotti. 111–124. C. vol. 81–98. A quasiarithmetical notation for syntactic description. pp. 29 (1953). Bundy. [15] A. ´ vol. Wootters. Quantum quandaries: a category-theoretic perspective. and J.1787. Teleporting an unknown quantum state via dual classical and einstein-podolsky-rosen channels. vol. 1895–1899. von Neumann. 2007. 3–13. Coecke. T. A. Boixo and C. Baez and M. C. 95–172. Language.

Oxford University Press. arXiv:1003.1037. and M. Languages and Programming (ICALP). Semantic Techniques in Quantum Computation (S. [42] L. [43] R. Operational Galois adjunctions.uk/projects/quantomatic/. O. editors).inf. J. (2011). [34] B.1598. pp. 2011. Coecke. Phase groups and the origin of non-locality for qubits. M. Wilce. Wilce. pp. 567–604. B. vol. Coecke. Fundamental Theories of Physics. http://dream. [29] B. Extended version: arXiv:1002.2368. 2000. Synthese. [41] L. Operational quantum logic: An overview. Coecke and E. Coecke and D. Clark. Kauﬀman G. D. Kissinger. Cambridge University Press. arXiv:0905. Spekkens. 297–308. [39] P. 2006. vol. 23–42. 2011. Chen and S. Coecke. A. A. 230– 244.5005. [38] B. Coecke and A. Proceedings of the 37th International Colloquium on Automata. Moore. Ph. Moore. vol. 2000. M. arXiv:1102. Categories and Languages (B. Proceedings of the 19th EACSL Annual Conference on Computer Science Logic (CSL). [25] . 2011. 345–384. Duncan. [27] B. 270(2) (2011). Lecture Notes in Physics. Dixon and A. 2010. R. Springer. [33] B. A. The principles of quantum mechanics (third edition). 1–46. O. J. arXiv:1203. 2011. Moore. Current Research in Operational Quantum Logic: Algebras. Duncan. Dirac.2540. 2010. [26] B. A new description of orthogonal bases. p. Lamonaco. [30] B. 15–31. Kissinger. Mathematical Structures in Computer Science. quantomatic. Interacting quantum observables: categorical algebra and diagrammatics. Dixon and R. Coecke. pp. Duncan. Open graphs and monoidal theories. Sadrzadeh. Springer.4114. 2008. Paquette. pp.196 BOB COECKE [24] B. 29–69. pp. Taylor and Francis. W. Oxford University. Perdrix. 043016. Duncan. Interacting quantum observables. editors). Spekkens. editor). Coecke and A. 111. and R. Soloviev. J. Wilce. Picturing classical and quantum Bayesian inference. Pavlovic. E. New Structures for Physics (B. [40] L.D. [31] B.4988. Strong complementarity and non-locality in categorical quantum mechanics. pp. Gay and I. vol. 167–271. Springer-Verlag. R. Wang. Graphical reasoning in compact closed categories for quantum computation. 6247.1997. Coecke. [36] B. and S. pp.3010. and A. 2012. vol. Coecke and R. Linguistic Analysis. D. Environment and classical channels in categorical quantum mechanics. Categories for the practicing physicist. Electronic Notes in Theoretical Computer Science. B. editors). 36 (2010). 2007. Duncan. arXiv:0904. Mathematical foundations for a compositional distributional model of meaning.ed. pp. Mackie. Lecture Notes in Computer Science. J. The compositional structure of multipartite quantum entan- glement. and Q. Pavlovic. Coecke and D. vol. W. 13 (2011). Fundamental Theories of Physics. . editors). 1–36. Languages and Programming. Coecke and S. Types for Quantum Computation. IEEE Computer Society.ac. and J. 2010. Lecture Notes in Computer Science. Pavlovic. 111. Quantum measurements without sums. Mathematics of Quantum Computing and Technology (L. Lecture Notes in Computer Science. Dixon. Edwards. Current Research in Opera- tional Quantum Logic: Algebras. Moore B. 195–218. arXiv:quant-ph/0608035. pp. Kissinger. [37] B. vol. 56(1) (2009). arXiv:quant-ph/0008019. Proceedings of the 27th Annual IEEE Symposium on Logic in Computer Science (LICS). Paquette. 1947. [28] B. New Journal of Physics. A. Coecke. Classical and quantum structuralism. and D. D. arXiv:1011. pp. Electronic Notes in Theoretical Computer Science. arXiv:quant-ph/0810. Merry. Coecke and R. Frot. Coecke. Kissinger. thesis. Coecke. 210 (2008). Automata. Annals of Mathematics and Artiﬁcial Intelligence. Springer-Verlag. vol. Categories and Languages (D. POVMs and Naimark’s theorem without sums. [32] . and A. pp. to appear. arXiv:quant-ph/09064725. Vicary. Mathemat- ical Structures in Computer Science. arXiv:quant-ph/0608072. [35] B. Extended version: arXiv:1004. to appear.

4800. pp. [50] J. pp. Quantum logic in dagger kernel categories. W.6079. p. Cam- bridge University Press. [57] C. The Mathematical Foundations of Quantum Mechanics. 1–101. vol. arXiv:1005. Lambek and P. J. J. Bundy. Mac Lane.4722. vol. [56] C. vol. ¨ [48] G. 769–802. 095011. [60] G. Natural associativity and commutativity. [45] S. Conjecture synthesis for inductive theories. 285–296. 13 (2004). vol. Probabilistic theories with puriﬁcation. 81 (2010). [58] M. [51] E. p. 6. 1667–1675. Kelly and M. vol. 1988. 59.-Y. A formalism-local framework for general probabilistic theories including quantum theory. Plotkin. Experimental support for a categorical compositional distributional model of meaning. and A. (2010). The geometry of tensor calculus I. N. Grefenstette and M. and C. Randall. editors). EMNLP. 062348. Scott. Journal of Mathematical Physics. Coherence for compact closed categories. 1–27. M. The mathematics of sentence structure. [52] H. no. Horsman. 88 (1991).5164. Hofmann. L. Perinotti. Springer. Perdrix. [65] . pp. pp. American Mathematics Monthly. Kissinger. Theoretical Computer Science. 55–112. [46] Cl-. . pp. vol. M. Springer. Basic concepts. Theory and Applications of Categories. Foulis and C. [64] J. New York. Halvorson.A. pp. 48(3) (2009). 25–50. New Journal of Physics. Joyal and R. Transactions of the American Mathematical Society. [62] J. Faure. Journal of Pure and Applied Algebra. Deterministic evolutions and Schr¨odinger ﬂows. Kock. Harding. 13 (2011). Lack. Type grammar revisited. M. Pillars of Computer Science (A. pp. Synthesising graphical theories. 13(11) (1972). Benjamin. Order. vol. Avron. Duncan and S. arXiv:1202. Lecture Notes in Computer Science. 367–385. [63] S. pp. D’Ariano G. 65 (1958). Deep Beauty: Understanding the Quantum World Through Mathematical Innovation. Jacobs. p. 28–46. L. vol. 2010. Lecture Notes in Computer Science. Physical Review A. [49] G. 154–170. Piron. [55] M. [54] L. [66] J. Linear logic. 1007 (1892). Eilenberg and S. Lecture Notes in Computer Science. Quantum picturalism for topological cluster-state computing. [59] A. D. Hasegawa. Rabinovich. Frege. 68(2) (1995). International Journal of Theoretical Physics. vol. vol. Finite dimensional vector spaces are complete for traced symmetric monoidal categories. 2008. (2012). 50(1) (1987). pp. [67] S. Logical Aspects of Computational Linguistics. pp. Sadrzadeh. Advances in Mathematics. THE LOGIC OF QUANTUM MECHANICS – TAKE II 197 [44] R. Composing PROPs. 251–289. Hardy. Operational statistics. Cambridge University Press. Heunen and B. 2004. 1963. [53] J. Dixon. Mackey. 2011. Dershowitz. Girard. Proceedings of ICALP. J. pp. vol. 177–212. Laplaza. 27(2) (2010). H. ACL. 49(4) (1963). [61] A. Lambek. Introduction to Higher Order Categorical Logic. 58(2) (1945). arXiv:1101. and G. 2011. pp. vol. vol. Journal of Automated Reasoning. pp. Frobenius Algebras and 2D Topological Quantum Field Theories. 1582. vol. 1394–1404. pp. vol. 19 (1980). General theory of natural equivalences. 47(3) (2011). vol. pp. [68] G. Cambridge University Press. Uber Sinn und Bedeutung. Mac Lane. and A. M. A link between quantum logic and categorical quantum mechanics. The Rice University Studies. Rewriting measurement-based quantum computations with generalised ﬂow. Chiribella and P. vol. Moore. Street. Helvetica Physica Acta. vol. Zeitschrift f¨ur Philosophie und Philosophische Kritik. I. 147–163. pp. [47] D. 150–157. 231. D. A. Johansson. 193–213. 1999.

J. Piron. Electronic Notes in Theoretical Computer Science. pp. Foundations of Quantum Physics. [81] M. Moore. 221–244. Zukowski. Seely. Electronic Notes in Theoretical Computer Science. On state spaces and property lattices. DEPARTMENT OF COMPUTER SCIENCE. Sadrzadeh. K. Combinatorial Mathematics and its Applications. Basil & Blackwell. pp. Why John von Neumann did not like the Hilbert space formalism of quantum mechanics (and what he liked instead ). Penrose. 97–123. [74] E. pp. QUANTUM GROUP E-mail: coecke@cs. A.uk . [71] C. Applications of negative dimensional tensors. Zeilinger. [78] . 141–153. Wittgenstein. 27(4) (1996). Computational Linguistics. [80] L. 113– 119. Schrodinger. 92 (1989). pp. Bell states and negative sentences in the distributed model of meaning. Dagger compact closed categories and completely positive maps. Electronic Notes in Theoretical Computer Science. ¨ Discussion of probability relations between separated systems. ∗-autonomous categories and cofree algebras. 61–83. Contemporary Mathematics. 371–382. pp. vol.ox. [79] P. pp. Preller and M. 493–510. [73] M. Studies in History and Philosophy of Modern Physics.ac. Philosophical Investigations. Horne. A. 270(1) (2011). Studies in History and Philosophy of Modern Physics. 24(1) (1998). ¨ Automatic word sense discrimination. A. A. 1972. vol. Event-ready-detectors Bell experiment via entanglement swapping. vol. [77] P. 1484–1509. SIAM Journal on Computing. G. Shor. UNIVERSITY OF OXFORD. Polynomial-time algorithms for prime factorization and discrete logarithms on a quantum computer. Schutze. [72] A. [76] R. 26(5) (1997). and A. vol. Finite dimensional Hilbert spaces are complete for dagger compact closed categories (extended abstract). W. vol. 1971. Ekert.198 BOB COECKE [69] D. [70] R. Academic Press. 139–163. vol. [75] H. vol. Selinger. pp. 170 (2007). 4287–4290. W. Linear logic. pp. 31 (1935). Cambridge Philosophical Society. vol. vol. Benjamin. Redei. pp. 1976. pp. pp. 555–563. 30(1) (March 1999). 71 (1993). 270(2) (2011). Physical Review Letters. vol. M.

this calculus is used to depict and reason about information ﬂows in entangled quantum states modeled in tensor spaces. Hence we can compare meanings of diﬀerent language constructs and enhance the applicability of the theory. MEHRNOOSH SADRZADEH. whereby the vector space foundations of quantum mechanics were recasted in a higher order language and quantum protocols such as teleportation found succinct conceptual proofs. 35]. by hiding the underlying additive vector space structure. and the latter are used to model classical data in quantum protocols by Coecke-Pavlovic-Vicary. REASONING ABOUT MEANING IN NATURAL LANGUAGE WITH COMPACT CLOSED CATEGORIES AND FROBENIUS ALGEBRAS DIMITRI KARTSAKLIS. Eskandarian and V. A. applied to formalizing the grammatical structure of natural language. However. as well as quantum. Speciﬁcally. AND BOB COECKE Abstract. in previous work Coecke-Clark- Sadrzadeh used the product category of pregroups with vector spaces and provided a distributional model of meaning for sentences. We report on experimental results on a number of language tasks and verify the theoretical predictions. Most quantum protocols rely on classical. We recast this theory in terms of strongly monoidal functors and advance it via Frobenius algebras over vector spaces. data ﬂow. Compact closed categories have found applications in modeling quantum informa- tion protocols by Abramsky-Coecke. phrases. The pictorial calculus revealed the multi-linear algebraic level needed for proving quantum information protocols and simpliﬁed the reasoning thereof to a great extent. Chubb. the phenomena that were considered to be mysteries of quantum mechanics and the Achilles heel of quantum logic [4]. Harizanov Lecture Notes in Logic. Compact closed categories were ﬁrst introduced by Kelly [19] in early 1970’s. Logic and Algebraic Structures in Quantum Computing Edited by J. The Frobenius algebras enable us to work in a single space in which meanings of words. 45 c 2016. In the work of [1]. Compact closed categories are complete with regard to a pictorial calculus [19. Introduction. STEPHEN PULMAN. Some thirty years later they found applications in quantum mechanics [1]. this classical data ﬂow was modeled using bi-products deﬁned over a compact closed category. and sentences of any structure live. the pictorial calculus could not Support by EPSRC grant EP/F042728/1 is acknowledged by the ﬁrst two authors. They also provide semantics for Lambek’s pregroup algebras. Association for Symbolic Logic 199 . and are implicit in a distribu- tional model of word meaning based on vector spaces. §1. The former are used to formalize topological quantum ﬁeld theories by Atiyah and Baez-Dolan.

The other application domain. PULMAN. comparing meanings of transitive and intransitive sentences. The theoretical predictions of the model were made concrete in [17]. §2. Compact closed categories have also found applications in two completely orthogonal areas of computational linguistics: formalizing grammar and reasoning about lexical meanings of words. In this article. and their categorical axiomatization was not as clear as the built-in monoidal tensor of the category. for instance see [24]. 26] and in particular to automatic word-synonymy detection [10]. The solution was based on a cartesian product of the pregroup category and the category of ﬁnite dimensional vector spaces. Distributional models have been widely studied and successfully applied to a variety of language tasks [34. we ﬁrst recast the theoretical setting of [9] using a succinct functorial passage from a free pregroup of basic types to the category of ﬁnite dimensional vector spaces. and a new term/deﬁnition classiﬁcation task. As a result. 25. a property that only holds for. These models consist of vector spaces whose basis are sets of context words and whose vectors represent meanings of target words. We start by recalling some def- initions. the distributional models do not scale to meanings of phrases and sentences. which is associative. Later. Then. A monoidal category [19] is a category C with a monoidal tensor ⊗.200 D. Recalling some categorical deﬁnitions. originally used in group theory [14] and later widely applied to other ﬁelds of mathematics and physics such as topological quantum ﬁeld theory (TQFT) [2. hence can be used to axiomatize. The former application is through Lambek’s pregroup grammars [23]. can be empirically veriﬁed by performing three experiments: the disambiguation task of [15]. Frobenius algebras. AND B. B. proved useful. Whereas the type-logical approaches to language do not provide a convincing model of word meaning. SADRZADEH. we are able to compare meanings of phrases and sentences with diﬀerent structures. The long standing challenge of combining these two models was addressed in previous work [6. and in particular the Frobenius algebraic constructions. S. referred to as distributional models of meaning. classical states [8]. KARTSAKLIS. COECKE extend well to bi-products. which are compact closed categories [31] and have been applied to formalizing grammars of a wide range of natural languages. 9. and moreover compare these with lexical vectors of words. 32]. we further advance the theory and show how Frobenius algebras over vector spaces provide solutions for the problem of the concrete construction of linear maps for predicative words with complex types. That is. we have that . 3]. C ∈ C. then implemented and veriﬁed in [15]. M. for all objects A. 21. formalizes meanings of words regardless of their grammatical roles and via the context of their occurrence [13]. This enhances the domain of application of our model: we show how the theoretical predictions of the model. It turned out that the operations of such algebras on vector spaces with orthonormal basis correspond to a uniform copying and deleting of the basis.

Furthermore. that is we have A∗ := Al = Ar and the above four equalities collapse to the following two: (A ⊗ 1A ) ◦ (1A ⊗ A ) = 1A (1A∗ ⊗ A ) ◦ ( A ⊗ 1A∗ ) = 1A∗ A functor F from a monoidal category C to a monoidal category D is a monoidal functor [20]. ·. REASONING ABOUT MEANING IN NATURAL LANGUAGE 201 A ⊗ (B ⊗ C ) ∼ = (A ⊗ B) ⊗ C . To see this. which serves as the unit of the tensor. (−)l . 1. These isomorphisms need to satisfy the usual coherence conditions. the following morphisms exist: Ar : A ⊗ Ar → I Ar : I → Ar ⊗ A Al : Al ⊗ A → I Al : I → A ⊗ Al and they satisfy the following yanking conditions: (1A ⊗ Al ) ◦ ( Al ⊗ 1A ) = 1A (Ar ⊗ 1A ) ◦ (1A ⊗ Ar ) = 1A (Al ⊗ 1Al ) ◦ (1Al ⊗ Al ) = 1Al (1Ar ⊗ Ar ) ◦ ( Ar ⊗ 1Ar ) = 1Ar In a symmetric compact closed category. a monoidal category is compact closed whenever any object A ∈ C has a left Al and a right adjoint Ar . This is a partially ordered monoid where each element of the partial order has a left and a right adjoint. The case for the right adjoint is similar. A⊗I ∼ =A∼ = I ⊗A. we refer to this category by Preg. that is. A monoidal functor is strongly monoidal [20]. and since adjoints are unique. it follows that F (Al ) must be left adjoint to F(A). A strongly monoidal functor on two compact closed categories C and D preserves the compact structure. A monoidal category is called symmetric whenever we have A ⊗ B ∼ = B ⊗ A. An example of a compact closed category is a Lambek pregroup [23]. that is. whenever F is a functor and moreover there exists a morphism I → F(I ) and the following is a natural transformation: F(A) ⊗ F(B) → F(A ⊗ B) satisfying the corresponding coherence conditions. (−)r ). that is F(Al ) = F(A)l and F(Ar ) = F (A)r . for which we have the following two compositions of morphisms: F(Al ) ⊗ F(A) → F(Al ⊗ A) → F(I ) → I I → F(I ) → F(A ⊗ Al ) → F(A) ⊗ F(Al ) From these. whenever the above morphism and natural transforma- tion are invertible. denoted by (P. the left and right adjoints collapse into one. which are the partial order versions of the yanking conditions of a compact closed category: p · p r ≤ 1 ≤ pr · p pl · p ≤ 1 ≤ p · pl . consider the case of the left adjoint. again satisfying the standard conditions. that is we have the following inequalities. Moreover there exists an object I ∈ C. ≤.

This is a representation of category of manifolds and cobordisms Cob (representing. The and maps. The left and right adjoints are deﬁned using the standard deﬁnition of adjoints and in terms of the min and max operations on the integers as follows. given by the inner product and maximally entangled states or Bell pairs. According to these rules. 3. We brieﬂy review two orthogonal models of meaning in computational linguistics: pregroup grammars and distributional models of meaning. COECKE An example of a pregroup is the set of all unbounded monotone functions on integers. SADRZADEH. and we show how one can interpret the former in the latter using a strongly monoidal functor. M.1. The rules on the left describe the formation of a grammatical sentence S in terms of other non-terminals. we refer to this category by FVect.202 D. by ﬁxing a basis {ri }i . house The above rules are referred to as generative rules. Category theory in linguistics. AND B. for f ∈ ZZ and m. The adjoint of each vector space is its dual. and assigns a vector space of states to each manifold and a linear operation to each cobordism. This representation is formalized using a strongly monoidal functor from Cob to FVect by Baez and Dylon [3]. S. KARTSAKLIS. are deﬁned as follows: A : A ⊗ A → R given by cij ri ⊗ rj → cij ri | rj ij ij A : R → A ⊗ A given by 1 → ri ⊗ ri i An example of a monoidal functor is Atiyah’s deﬁnition of a topological quantum ﬁeld theory (TQFT). n ∈ Z: f r (n) = {m ∈ Z | f(m) ≤ n} f l (n) = {m ∈ Z | n ≤ f(m)} An example of a symmetric compact closed category is the category of ﬁnite dimensional vector spaces and linear maps over a ﬁeld (which for our purposes we take to be the set of real numbers R). with function composition as the monoidal tensor and the identity function as its unit. PULMAN. which. becomes isomorphic to the vector space itself. Consider the simple grammar gener- ated by the following set of rules: S → Np Vp itV → smile Vp → tV Np | N tV → build Np → Adj Np | N Adj → strong N → man. The monoidal tensor is the tensor product of vector spaces whose unit is the ﬁeld. that is we have A∗ ∼ = A (note that this isomorphism is not natural). woman. Type-logical pregroup grammars. a sentence is a noun phrase Np followed by a verb . §3. respectively possible choices of space and spacetime) over the category of ﬁnite dimensional vector spaces FVect.

or a pregroup grammar for short. This type reduction represents the grammatical structure of the sequence. For . we take to be {n. the non-terminals of the above grammar (except for S) are interpreted as unary or binary predicates to produce meaning for phrases and sentences. For instance. n is the type representing a noun phrase and s is the type representing a sentence. There are various options when interpreting these non-terminals: for instance. and ‘house’ are nouns. We adhere to the more popular (among computational linguistics) verb-centric view and follow the former option. and outputs a sentence. ‘smile’ is an intransitive verb. A pregroup type-logical grammar. form an algebra of types. Finally. These types are then assigned to the vocabulary of a language. The types of the resulting predicates. for the purpose of this paper. Explicit in this type is also the fact that the intransitive verb has to be on the right hand side of the noun phrase. ‘build’ is a transitive verb. ‘strong’ is an adjective. s}. it has to be to the left of its input noun phrase. furthermore. According to these. REASONING ABOUT MEANING IN NATURAL LANGUAGE 203 phrase Vp. n r · s · n l is the type of a transitive verb. a unary predicate that inputs a noun phrase and outputs another noun phrase. or we can interpret a noun phrase as a binary predicate that inputs a verb phrase and outputs a sentence. In a predicative approach. The complex types of this pregroup represent the predicates. and a noun phrase is an adjective Adj followed either by a Np or a noun N . via a relation referred to as a type dictionary. We refer to this free pregroup grammar by PregF . obtained by recursively unfolding the rules. n r · s is the type of an intransitive verb. where a verb phrase itself is a transitive verb tV followed either by a Np or a noun N . monoid. and adjunction axioms to the multiplication of the types of the words of the sequence. The rules on the right instantiate all but one (S) of the non-terminals to terminals. to which we refer to by αw1 w1 ···wn . and ‘man’. This fact is succinctly expressed by the adjoint r of the type n. which is a binary predicate that inputs two noun phrases. we can either interpret a verb phrase as a binary predicate that inputs a noun phrase and outputs a sentence. ‘woman’. We treat these words as lemmas and take freedom in conjugating them in our example sentences. Here. Similarly. that is to the non-terminals of the generative rules. according to the ﬁrst rule. interpreted as a unary predicate that inputs a noun phrase and outputs a sentence. referred to as a type-logical grammar. Our example type dictionary is as follows: man woman houses strong smiled built n n n n·n l n ·s r n · s · nl r Every sequence of words w1 w2 · · · wn from the vocabulary has an associated type reduction. a type reduction is the result of applying the partial order. In a pregroup grammar. is the pregroup freely generated over a set of basic types which. has to be to the right of one and to the left of the other. n · n l is the type of an adjective in attributive position.

Similarly.2. It is not so clear what is the denotation of the adjective ‘strong’ or the verb ‘build’. then applying to it the adjunction and monoid axioms. which in principle can be the set of all lemmatized words of a corpus of documents or a dictionary. they can be topics obtained from a dimensionality reduction algorithm such as single value decomposition (SVD). Distributional models of word meaning. 3. SADRZADEH. The problem is resolved by adhering to a meaning-as-use model of meaning. Alternatively. We refer to such a vector space with an orthonormal . KARTSAKLIS. AND B. we will see how distributional models of meaning address this problem. For instance. COECKE example. For instance. and even worse. denoted by tensors and compositions of the and identity maps. words with basic types are only interpreted as atomic symbols. the basis vectors are often restricted to the few thousands most occurring words of the corpus. the type reduction αstrong house associated to the sequence ‘strong house’ is computed by multiplying the types of ‘strong’ and ‘house’. In practice. Meanings of some words can be determined by their denotations. the morphisms corresponding to the above adjective-noun phrase and sentence are as follows: strong man strong man built houses 1n ⊗ nl (nr ⊗ 1s ) ◦ (1n ⊗ nl ⊗ 1n r ·s ⊗ nl ) The generative rules formalize the grammar of a natural language and their consequent type-logical grammars provide a predicative interpretation for the words with complex types. all the words with the same type have the same interpretation. according to the context in which they often appear. hence obtaining n · n l · n ≤ n.g. A basis for this vector space is a set of target words. In the next section. e. However. the type reduction is a morphism of the category PregF . Matters get complicated when it comes to words with complex types such as adjectives and verbs. PULMAN. meaning of the word ‘house’ can be the set of all houses or their images. First formalized by Firth in 1957 [13] and about half a century later implemented and applied to word sense disambiguation by Schutze ¨ [34]. that is n ·n l ·n. a music dictionary. whereby one can assign meaning to all words. or a set of specialized words depending on the application domain.204 D. This context-based approach to meaning is the main idea behind the distributional models of meaning. distri- butional models of meaning interpret words as vectors in a highly dimensional (but ﬁnite) vector space with a ﬁxed orthonormal basis over real numbers. M. S. regardless of their grammatical type. the type reduction of the sentence ‘strong man built houses’ is as follows: αstrong man built houses : n · nl · n · nr · s · nl · n ≤ n · nr · s ≤ s In categorical terms. and the answer to the question ‘what is a house?’ can be provided by pointing to a house.

In this model. no matter how it is built. We provide a mapping of the free pregroup grammar PregF to FVectW via a strongly monoidal functor F . This functor . often in the form of Tf-Idf values which show how important is a speciﬁc basis word by taking into account not only the number of times it has occurred to the document. As an example. which serves as the meaning of the word. and to FVect restricted to tensor powers of W as FVectW . ‘house’ and ‘build’ often appear close to ‘brick’. human 6 man woman . This number is usually normalized. we will see how a combination of type-logical and distributional models overcome both of their corresponding shortcomings. but also the number of documents in which it appears in the corpus. The weights of this vector are obtained by counting how many times the word has appeared ‘close’ to a basis word. where ‘close’ is a window of n (usually equal to 5) words. A toy distributional model of meaning. The set {human.3. ‘woman’. The cosine of the angle between the word vectors has proved to be a good measure in predicting synonymy of words [10]. 3.mortal 9 build house brick Figure 1. to each word is associated a vector. whereas ‘man’ and ‘woman’ often appear close to ‘mortal’ and ‘human’. consider the toy vector space of Figure 1. mortal. as our basic distributional vector space W . Quantizing the grammar. For instance. brick} is the basis of this vector space and the words ‘man’. The words that have often appeared in the same context have a smaller angle between their vectors. ‘house’ and ‘build’ each have a vector assigned to them. as there is more to a language than the contexts of its words and these models on their own do not scale up to the interpretations of phrases and sentences. the distributional models of meaning cannot serve as the deﬁnite models of natural language. In the next section. REASONING ABOUT MEANING IN NATURAL LANGUAGE 205 basis {wi }i . Despite these good predictions.

that is: F(n · n l ) = F(n r · s) = W ⊗ W F(n r · s · n l ) = W ⊗ W ⊗ W Similarly. we have: F(n) = F(s) = W By functoriality. The meaning of a sequence of words w1 w2 · · · wn with type reduction αw1 w2 ···wn is: Deﬁnition(*) F(α ) (− →⊗− w → ⊗ ··· ⊗ − w w1 w2 ···wn 1 w→) 2 n As an example. COECKE assigns the basic vector space W to both of the basic types. for instance intransitive or transitive. since for x ∈ {n. we need to run some experiments. motivated by the above mentioned fact that W ∗ ∼ = W: F(x l ) = F(x r ) = F(x) Since W ∗∗ ∼ = W∗ ∼ = W . S. the partial orders between the basic types (for example those presented in [23]) are mapped to linear maps from W to W . will be a vector in W . for instance the type reduction of a transitive sentence is mapped as follows: F(αsbj verb obj ) = F(nr ⊗ 1s ⊗ nl ) = W ⊗ 1W ⊗ W : W ⊗ (W ⊗ W ⊗ W ) ⊗ W −→ W Now we can use the deﬁnition of [9] to provide a meaning for phrases and sentences of our grammar.206 D. whereas we know very well how to build vectors in W for words with basic types such . the iterated adjoint types are also mapped to W : F(x ll ) = F(x rr ) = F(x) The complex types are mapped to tensor products of vector spaces. However. M. In order to determine that this measure of synonymy provides good predictions. AND B. PULMAN. SADRZADEH. take: −−→ houses − −−→ = men → − cimen − w houses = ck → built = w built − cijk →⊗− (w →⊗− w →) w i k i j k i k ijk Substituting these in Deﬁnition(*). The adjoints of basic types are also mapped to W . we obtain the following for the meaning of the sentence ‘men built houses’: ' (' −−−→( F nr ⊗ 1s ⊗ nl − −→ ⊗ built ⊗ − men houses ' −−−→( = (W ⊗ 1W ⊗ W ) − −→ ⊗ built ⊗ − men houses built −−→ − men| →− → −−−−→ − → = cijk w i wk |houseswj ijk This deﬁnition ensures that the interpretations of noun phrases and sentences of any grammatical structure. that is. the type reductions are mapped to the compositions of tensor products of identity and maps of FVectW . s} we have the following. hence we can measure the cosine of the angle between them and compute their synonymy. KARTSAKLIS.

they have found applications in other ﬁelds of mathematics and physics. . Frobenius algebras were originally introduced in 1903 by F. Frobenius algebras. A Frobenius algebra over a symmetric monoidal category (C. Since then. ⊗.g. I ) is a tuple (F. in topological quantum ﬁeld theories [21] and in categorical quantum mechanics [8]. The general categorical deﬁnitions recalled below are due to Carboni and Walters [5]. In the next section we show how the notion of a Frobenius algebra over a vector space can be of use in addressing this matter. Frobenius in the context of proving representation theorems for group theory [14]. e. Their concrete instantiations to algebras over vector spaces were developed in [8]. and there is no known standard procedure to construct these. REASONING ABOUT MEANING IN NATURAL LANGUAGE 207 as ‘man’ and ‘house’. our method further requires interpretations of words with complex types to be in tensor spaces. §4. G.

. where for an F object of C the triple (F.. ). . .

satisfying the coalgebraic associativity and unit conditions: .. the following are morphisms of C. ) is an associative coalgebra. that is.

) is an associative algebra. the above . . the following are morphisms of C. satisfying the algebraic associativity and unit conditions: : F ⊗ F → F : I → F Moreover. that is.: F → F ⊗F : F → I The triple (F.

and morphisms satisfy the following Frobenius condition: ( ⊗ 1F ) ◦ (1F ⊗ .

) = .

◦ = (1F ⊗ ) ◦ (.

⊗ 1F ) A Frobenius Algebra is commutative if it satisﬁes the following two conditions for : X ⊗ Y → Y ⊗ X . ⊗. the symmetry morphism of (C. I ): ◦.

=.

a Frobenius Algebra is isometric or special if it satisﬁes the following condition: ◦ . ◦ = Finally.

explicitly given as follows: v → − . = Id In the category FVect. any vector space V with a ﬁxed basis {− → vi }i has a commutative special Frobenius algebra over it.

the coalgebra and algebra operations relate to each other via the equation . :: − → i →v ⊗−→ v i i :: − → v → 1 i :: − → vi ⊗ − → vi → − → vi :: 1 → − → vi In a Frobenius algebra over an orthonormal vector space.

equal to the transpose for vector spaces over reals. where † is the adjoint. the operation . † = . In such Frobenius algebras.

corresponds to copying and its unit corresponds to deleting of the vectors. They enable one to faithfully .

In linear algebraic terms. SADRZADEH. for v ∈ W . PULMAN. S. M. such as W ⊗ W. we have that .208 D. W ⊗ W ⊗ W. COECKE encode vectors of W into spaces with higher tensor ranks. AND B. · · · . KARTSAKLIS.

that is. Each tensor is represented by a triangle. The objects of this fragment are the tensors of multi-linear algebra. Speciﬁcally. We brieﬂy introduce the fragment of this calculus that we are going to use in this paper. for z ∈ W ⊗ W . The tensor products of vectors are represented by juxtaposing their corresponding triangles. adjectives and intransitive verbs are rank-2 tensors. is depicted as follows: Men built houses W WWW W Computations with Frobenius algebras can also be represented within the more general diagrammatic calculus of symmetric monoidal categories. hence losing the information encoded in the non-diagonal part. and a 3d-array is a rank-3 tensor. we have that (z) is a vector consisting only of the diagonal elements of z. Words are represented by tensors that correspond to their meaning: subjects and objects are rank-1 tensors (vectors). a matrix is a rank-2 tensor. The framework of compact closed categories comes with a complete diagrammatic calculus that allows convenient graphical representations of the derivations. ﬁrst formalized in [18]. referred to as string diagrams. Pictorial calculi. For example. whereas the identity morphism is a vertical straight line. §5. In linear algebraic terms. whose rank can be determined by its wires. a vector is a rank-1 tensor. the linear maps of the coalgebra and algebra are depicted by: (. it loses some information when encoding a higher rank tensor into a lower rank space. and transitive verbs are rank-3 tensors.(v) is a diagonal matrix whose diagonal elements are weights of v. the meaning of a transitive sentence. The operation is referred to as uncopying. The maps are depicted as cups. following Deﬁnition(*).

) = ( .. ) = The Frobenius condition is depicted by: = = The commutativity conditions are shown as: .

one can extend these 0/1 entries to real numbers and model words with complex types as weighted predicates. tells us which noun phrases are related to other noun phrases. . ··· For an informal introduction to compact closed categories. §6. see [7]. and their diagrammatic calculi. For instance. predicates can be modeled as matrices (or equivalently. Building such linear maps from a corpus turns out to be a non-trivial task. linear maps ). provided that the diagram of computation is connected. where a j is the vector/linear map associated to the argument aj and the index i counts the number of times each word aj has appeared in the corpus as the argument of p. the verb becomes a linear map that moreover tells us to what extent these noun phrases are related to each other. For instance. in the type-logical model.. 15] we argue that such a linear map can be constructed by taking the sum of the tensor products of the vectors/linear maps of its arguments. a transitive verb is a binary predicate that. the linear map representing an n-ary predicate p with arguments a1 to an is i a 1 ⊗ · · · ⊗ a n . over the semiring of booleans. the Frobenius conditions guarantee that any diagram depicting a Frobenius algebraic computation can be reduced to a normal form that only depends on the number of input and output wires of the nodes. . In a vector space model. Frobenius algebras. = . The type-logical models of meaning treat words with complex types as predicates. REASONING ABOUT MEANING IN NATURAL LANGUAGE 209 = = The isometry condition is depicted by: = Finally.. In vector spaces over reals. In previous work [17. Following . These are predicates that not only tell us which instantiations of their arguments are related to each other. In a matrix calculus. but also that to what extent these are related to each other. This justiﬁes depicting computations with Frobenius algebras as spiders. Building tensors for words with complex types. referring to the right hand side diagram below: ··· .

KARTSAKLIS. In order to encode them in W ⊗ W . SADRZADEH.210 D. M. we use the Frobenius . Adjectives and intransitive verbs. but we need a linear map in W ⊗ W ⊗ W . AND B. S. We use the pictorial calculi of the compact closed categories and Frobenius algebras to depict the resulting linear maps and sentence vectors. This problem is overcome by using the operations of a Frobenius algebra over vector spaces. COECKE this method. the linear maps corresponding to the predicates of our simple grammar are as follows: intransitive verb transitive verb adjective − → − → −→ −−→ i sbji (sbji ⊗ obji ) i nouni i There is a problem: this method provides us with a linear map in a space whose tensor rank is one less than the rank of the space needed by Deﬁnition(*). PULMAN. For instance.1. 6. The linear maps of adjectives and intransitive verbs are elements of W . the linear map of the transitive verb ends up being in W ⊗ W .

The ﬁrst option is to copy the “row” dimension of the linear map corresponding to the verb. that is. when substituted in Deﬁnition(*). Transitive verbs. this dimension encodes the information of the subjects of the verb from the corpus. which geometrically speaking provide us with diﬀerent ways of “diagonally” placing a plane into a cube.2. The linear map of a transitive verb is an element of W ⊗ W . the above will result in the left hand side vector below. 6. We face a few options here. operation and obtain the following linear map: For the intransitive verb. this has to be encoded in W ⊗ W ⊗ W . CpSbj. which is then normalized to the right hand side vector. when applied to its subject. In the left hand side diagram below we see how . the order of the above application is swapped: the triangle of the adjective will change place with the triangle of the subject of the intransitive verb. = When an adjective is applied to a noun.

we obtain the diagram in the right hand side: . Once substituted in Deﬁnition(*). transforms the verb in this way.

REASONING ABOUT MEANING IN NATURAL LANGUAGE 211 Verb: Sentence: In this case. the .

map transforms the matrix of the verb as follows: .

which encodes the information about the objects of the verb from the corpus: Verb: Sentence: Now the . Our other option is to copy the “column” dimension of the matrix. :: cij (− → ni ⊗ − n→ j) → cii (− → ni ⊗ − → ni ⊗ −n→ j) ij iij CpObj.

-map does the following transformation: .

since they suggest a closed form formula for each case. :: c (−→ n ⊗− ij n→) → i j c (−→ n ⊗− n→ ⊗ − jj n→) i j j ij ijj The diagrams above simplify the calculations involved. (b) the result of this interaction serves as input for the . Taking as an example the diagram of the copy-subject method. we see that: (a) the object interacts with the verb.

(c) one wire of the output of . map.

− → − → Linear algebraically. this corresponds to the computation . interacts with the subject. while the other branch delivers the result.

which expresses the fact that the meaning of a sentence is obtained by ﬁrst applying the meaning of the verb to the meaning of the object. then applying the (.(verb × obj)T × sbj.

**version of the) result to the meaning of the subject. This computation
**

results in the Equation 1 below:

−−−−−−−→ − → −→

sbj verb obj = sbj ) (verb × obj) (1)

This order of application is the exact same way formalized in the generative

rules of the language. On the contrary, the meaning of a transitive sentence for

the copy-object results is given by Equation 2 below, which expresses the fact

that the meaning of a sentence is obtained by ﬁrst applying the (transposed)

meaning of the verb to the meaning of the subject and then applying the result

to the meaning of the object:

−−−−−−−→ − → T −

→

sbj verb obj = obj ) (verb × sbj) (2)

Note that equipped with the above closed forms we do not need to create or

manipulate rank-3 tensors at any point of the computation, something that

would cause high computational overhead.

Purely syntactically speaking, in a pregroup grammar the order of application

of a transitive verb does not matter: it is applied to its subject and object in

212 D. KARTSAKLIS, M. SADRZADEH, S. PULMAN, AND B. COECKE

**parallel. Semantically, as originated in the work of Montague [28], a transitive
**

verb is ﬁrst applied to its object and then to its subject. In the more modern

approaches to semantics via logical grammars, this order is some times based

on the choice of the speciﬁc verb [12]. Our work in this paper is more inline

with the latter approach, where for the speciﬁc task of disambiguating the verbs

of our dataset, ﬁrst applying the verb to the subject then to the object seems

to provide better experimental results. According to our general theoretical

setting, the linear map corresponding to the transitive verb should be a rank-3

tensor, but at the moment, apart from work in progress which tries to conjoin

eﬀorts with Machine Learning to directly build these as rank-3 tensors, we do

not have the technology to do other than described in this paper. However, in

the ideal case that the linear maps of words are already in the spaces allocated

to them by the theory, these choices will not arise, as the compact nature of

the matrix calculus implies that the application can be done in parallel in

all the cases that parallel applications are prescribed by the syntax. From a

linear-algebra perspective, fully populated rank-3 tensors for verbs satisfy the

following equality:

−→ −→ T −→ −→

subj verb obj = (verb × obj)T × subj = (verb × subj) × obj

**which shows that the order of application does not actually play a role.
**

MixCpDl. We can also use a mixture of

and

maps. There are three

reasonable options here, all of which start by applying two

**’s to the two wires
**

of the linear map of the verb (that is, one for each of the dimensions). Then

one can either apply a to one of the copies of the ﬁrst wire, or a to one of

the copies of the second wire. These two options are depicted as follows:

**The ﬁrst diagram has the same normal form as the copy-subject option, and
**

the second one has the same normal form as the copy-object option.

Finally, one can apply a

to one wire from each of the copied wires of the

verb, the result of which is depicted in the following left hand side diagram.

When substituted in Deﬁnition(*), we obtain the following right hand side

diagram for the meaning of the transitive sentence:

Verb: Sentence:

REASONING ABOUT MEANING IN NATURAL LANGUAGE 213

**The normal form of the diagram of the sentence is obtained by collapsing the
**

three dots and yanking the corresponding wires, resulting in the following

diagram:

−−→ −−→

Linear algebraically, the spider form of the verb is equivalent to i (sbji ) obji ).

A verb obtained in this way will only relate the properties of its subjects and

objects on identical bases and there will be no interaction of properties across

bases. For instance, for a certain verb v, this construction will result in a

vector that only encodes to what extent v has related subjects and objects

with property −→, and has no information about to what extent v has related

w 1

subjects with property −w→ to objects with property −

1

→. The closed form of the

w 2

above diagram is: −→ −→ ! −

−

→ →

sbj ) (sbji ) obji ) ) obj

i

**6.3. Encoding the existing non-predicative models. Apart from the predica-
**

tive way of encoding meanings of words with complex types, there exists two

other approaches in the literature, who simply work with the context vectors

of such words [27, 16]. These two approaches are representable in our setting

using the Frobenius operations.

Multp. To represent the model of [27] in our setting, in which the meaning

of a sentence is simply the point-wise multiplication of the context vectors of

−−→

the words, we start from the context vector of the verb, denoted by verb, and

apply three

**’s and then one
**

to it. The result is depicted in the left hand side

diagram below; once this verb is substituted in Deﬁnition(*), we obtain the

right hand side diagram below as the meaning of a transitive sentence:

Verb: Sentence:

**The normal form of the diagram of the sentence and its closed linear algebraic
**

form are as follows:

−

→ −−→ − →

= sbj ) verb ) obj

**Kron. In the model of [16], the tensor of a transitive sentence is calculated
**

as the Kronecker product of the context vector of the verb with itself, so we

214 D. KARTSAKLIS, M. SADRZADEH, S. PULMAN, AND B. COECKE

−−→ −−→

have verb = verb ⊗ verb. To encode this, we start from the Kronecker product

of the context vector of the verb with itself, apply one

an object. A dataset for this task was originally developed in [27] for intransitive sentences. Disambiguation. Experiments. a subject. and later extended to transitive sentences in [15]. classiﬁed into a hundred million diﬀerent lexical tokens. given the context (subject and object) of that verb. and a landmark verb used for the comparison. The models. Table 1. The weights of each vector are set to the ratio of the probability of the context word given the target word to the probability of the context word overall. ¨ which is an extension of Schutze’s original disambiguation task from words to sentences. The diﬀerent options presented in Section 6 and summa- rized in Table 1 provide us a number of models for testing our setting.1. which when substituted in the equation of Deﬁnition(*) results in a normal form (depicted in the right hand side below) very similar to the normal form of the Multp model: Verb: Sentence: Linear algebraically. Model Description CpSbj Copy subject on relational matrices CpObj Copy object on relational matrices MixCpDl Diagonalize on relational matrices Kron Diagonalize on direct matrices Multp Multiplicative model 7. we use the latter. We use the set of its 2000 most frequent lemmas as a basis of our basic distributional vector space W . to each one of the vectors and then a to both of them jointly. The result is the following left hand side verb. The goal is to assess how well a model can discriminate between the diﬀerent senses of an ambiguous verb. As our similarity measure we use the cosine distance between the vectors. We train our vectors on the British National Corpus (BNC). which has about six million sentences and one million words. The entries of this dataset consist of a target verb. One such entry . the above normal form is equivalent to: − → −−→ −−→ − → sbj ) verb ) verb ) obj (3) §7. We ﬁrst test our models on a disambiguation task.

but a letter or a shopping list can only be written. High and Low refer to average scores for high and low landmarks. Similarly.000 Kron 0.95 0. where the meaning of a transitive sentence is computed as the addition of the relevant context vectors. for example. Table 2. Up- perBound refers to agreement between annotators.67 0. On the other hand. Results. we can imagine that the crucial factor to disambiguate between the verbs “write”. book” these results should be reversed.168 CpSbj 0. The subject in all these cases is not so important.143 CpObj 0. REASONING ABOUT MEANING IN NATURAL LANGUAGE 215 for example is.89 0.90 0. Model High Low Addtv 0.80 2. which measures the degree of correlation of the cosine distance with the judgements of 25 human evaluators.90 0.050 Multp 0. pupil. spell”. takes place along the dimension of subjects (rows). The results of this experiment are shown in Table 2. indicating that the most successful model for this task is the copy-object model. The evaluation of this experiment is performed by calculating Spearman’s . as this was calculated in [15]. name. The better performance of this model against the copy-subject approach provides us some insights about the role of subjects and objects in disambiguating our verbs. Hence.77 0. meaning that the resulting vector will bring much more information from the objects than the subjects (this is also suggested by Equation 2).90 0. Disambiguation results. to “pupil publish name”.75 0.620 . but a book can only be written. “write. “publish” and “spell” is more the object than the subject: a book or a paper can be both published and written. the fact that this vector performs better than the one of the copy-subject method provides an indication that the object of some ambiguous verbs (which turns out to be the case for our dataset) can be more important for disambiguating that verb than the subject. Intuitively.172 UpperBound 4. A good model should be able to understand that the sentence “pupil write name” is closer to the sentence “pupil spell name” than.49 0.60 0. who has been asked to assess the similarity of each pair of sentences using a scale from 1 to 7. given the context “writer. respectively. a word can be spelled and written. By copying the dimension associated with the object.21 0.31 0. As our baseline we use a simple additive model (Addtv). the compression of the original sentence matrix.163 MixCpDl 0.95 0.

We also create intransitive versions using the high and the low landmarks. we compare each transitive sentence coming from the target verbs with all the other intransitive sentences. Results of the comparison between transitive and intransitive sentences. M. KARTSAKLIS. S. su ) > sim(str . Table 3. Our transitive sentence here is str = “pupil write name”. To present a concrete example. The similar performance of these two is not a surprise. Since the meaning of a target verb and the high landmark verb were speciﬁcally . expecting that the highest similarity would come from its own intransitive version. If the similarity between two sentences s1 and s2 is given by sim(s1 . for example. for 100 target verbs. and so on. the bad performance of the model (MixCpDl) that is obtained by the application of the uncopying map conforms to the predictions of the theory. sit ) > sim(str . Comparing transitive and intransitive sentences. the intransitive version of this is sin = “pupil write”. PULMAN. pupil. su ) where su represents an unrelated intransitive version coming from a target verb diﬀerent than the one of str . The results of this experiment are shown in Table 3 below. in the context of an experiment aiming to compare transitive and intransitive sentences. then. COECKE The copy-object model is followed closely by the (Kron) and the Multp models. In order to do that.5 sim(str . sit ) 7 of 93 7. We see. sit ) 36 of 9900 0. as these were expressed in Section 6. In this section we will examine the potential of the above approach in practice. sit ) 6 of 93 5. getting shi = “pupil spell” and slo = “pupil publish”.216 D. given their almost identical nature.1) to conduct the following simple experiment: We create intransitive versions of all the transitive sentences from target verbs and their high and low landmarks by dropping the object. SADRZADEH. slo ) > sim(str . name. Case Errors % sim(str . s2 ).4 The outcome follows indeed our expectations for this task. shi ) > sim(str . we would expect that: sim(str . AND B. shi ) > sim(str . spell. publish”. 7. the next higher similarity would come from the intransitive version that uses the corresponding high landmark verb. consider the entry “write. slo ) > sim(str .2.6 sim(str . Finally. that the highest error rate comes from cases where the intransitive sentence of the high landmark verb is closer to a transitive sentence than the intransitive version coming from the sentence itself (ﬁrst row of the table). we use the dataset of the previous verb disambiguation task (see detailed description in Section 7.

are not so diﬀerent semantically.g. the low-landmark intransitive version “tribunal test” has almost identical meaning with the normal intransitive version “tribunal try”.4% error rate. since in many cases dropping the object leads to semantically identical expressions. Deﬁnition classiﬁcation. Errors in the ﬁrst category of comparisons. this is naturally the most error-prone category. For each term we added two more deﬁnitions. 7. The ability of reliably comparing the meaning of single words with larger textual fragments. either by using entries of WordNet for the term or by simple . phrases or even sentences. with only a 0. such as deﬁnition classiﬁcation. so it is easier to be “mistakenly” selected by the model as the one closest to the original transitive sentence. can be an invaluable tool for many challenging NLP tasks. REASONING ABOUT MEANING IN NATURAL LANGUAGE 217 selected to be very similar given the context (subject and object). the model performs really well for cases when an unrelated intran- sitive sentence is compared with a transitive one.3. For the transitive sentence “tribunal try crime”. Our dataset consists of 112 terms (72 nouns and 40 verbs) extracted from a junior dictionary together with their main deﬁnition. In this task we examine the extend to which our models can correctly match a number of terms (single words) with a number of deﬁnitions. Table 4. Here many of the misclassiﬁcations can also be attributed to the increased ambiguity of the involved verbs when the object is absent. or even the simple everyday search on the internet. sentiment analysis. the similarity between “man draw sword” and “man draw” is considered smaller than the similarity of the ﬁrst sentence with “man write”. Finally. where the similarity of the involved intransitive versions is apparent. for example. Although this is an obvious error. str sin shi people run round people run people move boy meet girl boy meet boy visit management accept responsibility management accept management bear patient accept treatment patient accept patient bear table draw eye table draw table attract claimant draw beneﬁt claimant draw claimant attract tribunal try crime tribunal try tribunal judge The six cases of the second category (where an intransitive sentence from a low-landmark gives higher similarity than the normal intransitive version) are quite similar. “man draw” and “man write”. The seven misclassiﬁed cases are presented in Table 4. For example. so the error was not completely unjustiﬁed. e. we should acknowledge that the two intransitive sentences.

First. AND B.30 0. we use verb = i obji . this task did not require human annotation and we evaluate the results by calculating separate F1-scores for each term.17 0. In all cases a deﬁnition for a noun-term is a noun phrase.25 0.31 0. verb-terms in the bottom part).28 0. we calculate the distance between each deﬁnition and every term in the dataset. where terms have the role of classes. M. A sample of the dataset entries can be found in Table 5. we construct our verb vectors by summing over all context vectors of −−→ objects with which the verb appears in the corpus.21 0.27 Results. Table 6.22 0.26 Reltn 0. The results are presented in Table 6. Results of the term/deﬁnition comparison task. Table 5. COECKE paraphrase of the main deﬁnition using a thesaurus.21 0. Due to the nature the dataset. . S. and getting their average as an overall score for the whole model. Nouns Verbs Model P R F1 P R F1 Addtv 0.28 0.32 0. whereas the deﬁnitions for the verb-terms consist of verb phrases. where the subject is missing.16 0. PULMAN. Sample of the dataset for the term/deﬁnition com- parison task (noun-terms in the top part. getting a total of three deﬁnitions per term. KARTSAKLIS.21 0.19 0. 1 Alternative def. Since this experiment includes verb phrases. SADRZADEH. that is.24 0. The deﬁnition is “classiﬁed” to the term that gives the higher similarity. 2 blaze large strong ﬁre huge potent ﬂame substantial heat husband married man partner of a woman male spouse foal young horse adolescent stallion juvenile mare horror great fear intense fright disturbing feeling apologise say sorry express regret or acknowledge sadness shortcoming or failing embark get on a ship enter boat or vessel commence trip vandalize break things cause damage produce destruction regret be sad or sorry feel remorse express dissatisfaction We approach this evaluation problem as a classiﬁcation task.22 0. Term Main deﬁnition Alternative def.23 Multp 0.218 D.

which was based on the product of the category of a pregroup type-logic with the category of ﬁnite dimensional vector spaces.0 2-5 10 25. Some interesting other cases are presented in Table 8. The multiplicative model performs similarly. which is also correct. Table 7. which again were very close for the two models. An error analysis shows that for the noun-term set the relational model returns the correct main deﬁnition in 25 of the 72 cases.1 25 34.3 5 6. the deﬁnition we originally assigned to the term ‘jacket’ was ‘short coat’. since a mixing of the two sets would be inconsistent. Results of the term/deﬁnition comparison task based on the rank of the main deﬁnition. For example.0 4 10. we ramiﬁed their applications to natural language syntax and semantics.5 §8. In summary. Furthermore.0 11-40 9 22. since the challenge that this task poses to a machine is great. . The relational model delivers again the best performance.8 20 27. whereas in 47 cases (65%) the correct deﬁnition is in the top-ﬁve list for that term (Table 7). All models perform better on the verb terms than the noun part of the dataset.8 1 15 37.7 2-5 20 27.0 13 32. yet in general F-scores tend to be low. some of the “misclassiﬁed” cases can also be considered as somehow “correct”. after a brief review of the deﬁnitions of compact closed categories and monoidal functors. and F-score considers anything but the perfect result (every deﬁnition assigned to the correct term) as unacceptable.9 11-72 15 20. For this experiment we also calculated Mean Reciprocal Ranks values.5 8 20.5 15 37. Multp Reltn Rank Count % Count % 1 26 36. We evaluate separately the performance on the noun terms and the performance on the verb terms. and better for the verb-term set. however. This is natural.5 Verbs 6-10 6 15. Additive model serves again as our baseline. and is compared with the multiplicative model.6 Nouns 6-10 11 15. We recasted the categorical setting of [9]. REASONING ABOUT MEANING IN NATURAL LANGUAGE 219 This is referred to as the relational model (Reltn). the system “preferred” the deﬁnition ‘waterproof cover’. Conclusion and future work.8 22 30. although the diﬀerence from the multiplicative model is small.

220 D. Term Original deﬁnition Assigned deﬁnition rod ﬁshing stick round handle jacket short coat waterproof cover mud wet soil wet ground labyrinth complicated maze burial chamber in terms of a monoidal functor from the former to the latter. In recent work [33] we have shown how the functorial passage to FVect can be extended from a pregroup algebra to the Lambek Calculus [22]. This passage is similar to the vector space representation of category of manifolds in a topological quantum ﬁeld theory. SADRZADEH. This instantiation resulted in meanings of all sentences living in a basic vector space W . sentences with nested structures such as ‘Mary saw John reading a book’ could not be assigned a meaning. A sample of ambiguous cases where the model assigned a diﬀerent deﬁnition than the original. As a consequence. M. AND B. hence the logical and the concrete types did not match. hence we became able to compare their meanings with one another and also with meanings of single words and phrases. and will help us separate the two distinct tasks of composition and disambiguation that currently are interwoven in a single step. one direction is to start from type-logical grammars that are more expressive than pregroups. which has a monoidal (rather than compact) structure. where each word is associated with one or more sense vectors. S. such as Lambek-Grishin algebras . meanings of phrases and sentences with diﬀerent grammatical structure lived in diﬀerent vector spaces. It overcomes the shortcomings of our ﬁrst implementation [15]. so a direct comparison of them was not possible. We developed experiments based on this instantiation and evaluated the predictions of our model in a number of sentence and phrase synonymy tasks. A model like this will avoid encoding diﬀerent meanings of words in one vector. From the theoretical perspective. An experimental future direction is a higher order evaluation of the deﬁnition classiﬁcation task using an unambiguous vector space along the lines of [34]. KARTSAKLIS. furthermore. We showed how the operations of Frobenius algebras over vector spaces provide a concrete instantiation of this setting and used their pictorial calculus to simplify the multi-linear algebraic computations. We conclude that the concrete setting of this paper provides a robust and scalable base for an implementation of the framework of [9]. COECKE Table 8. PULMAN. whose main problem was that the vector space representation of the atomic type s was taken to be the tensor space W ⊗ W (for a transitive sentence). It remains to show how this passage can be extended to more expressive versions of Lambek calculi. ready for further experimentation and applications.

Annals of Mathematics. de Groote and F.D. Coecke. 2007. Sitzungsberichte der Preußischen Akademie der Wissenschaften zu Berlin. 415–425. M. 167–271. vol. [15] E. Clark. [16] . Carboni and R. 2010. 2011. 6073D0–6105. Abramsky and B. More speciﬁcally. Topological quantum ﬁeld theories. Proceedings of Conference on Empirical Methods in Natural Language Processing (EMNLP). 175–186. B. 2011. G. 1 (2008). de Groote. Sadrzadeh. where any context- dependant method fails to succeed. 2004. pp. calculi of discontinuity [30]. 2011. 2001. The logic of quantum mechanics. 2011. vol. Journal of Pure and Applied Algebra. and W. [12] P. 35 (1936). Grefenstette and M. Sadrzadeh. Proceedings of Workshop on Geometrical Models of Natural Language Semantics (GEMS). Pulman. Mathematical Structures in Computer Science. Mathematical foundations for distributed compo- sitional model of meaning. pp. pp. Experimenting with Transitive Verbs in a DisCoCat. 345–384. hence these might turn out to be promising in this aspect too. Categories for the practicing physicist. Walters. Publications Math´ematique de l’Institut des ´ Hautes Etudes Scientiﬁques. Firth. . 36. 355–388. [4] G. 813. Coecke. and J. [2] M. Coecke. Lamarche. vol. Studia Logica. pp. Concrete compositional sentence spaces for a compositional distributional model of meaning. Dolan. New Structures for Physics (B. Atiyah. or abstract categorial grammars [11]. [8] B. Linguistic Analysis. Frobenius. [3] J. Buszkowski. Pulman. Pavlovic. Sadrzadeh. pp. pp. International Conference on Computational Semantics (IWCS’11). editor). Vicary. [7] B. Studies in Linguistic Analysis. [11] P. editors). M. M. Proceedings of AAAI Spring Symposium on Quantum Interaction. From Distributional to Semantic Similarity. [5] A. A synopsis of linguistic theory 1930-1955. Towards abstract categorial grammars. REASONING ABOUT MEANING IN NATURAL LANGUAGE 221 [29]. S. REFERENCES [1] S. (1957). thesis. Combining symbolic and distributional models of meaning. D. Springer. Curran. Clark and S. Moortgat. (1903). vol. arXiv:quant-ph/0402130. A categorical semantics of quantum protocols. R. AAAI Press. and S. [10] J. Birkhoﬀ and J. Lambek Festschrift (J. The abstract operations of Frobenius algebras have been used to model a limited class of logical operations on vector spaces [8]. von Neumann. Lecture Notes in Physics. Higher-dimensional algebra and topological quantum ﬁeld theory. 2004. Theorie der hyperkomplexen Gr¨oßen. 49 (1987). 71 (2002). Experimental support for a categorical compositional distributional model of meaning. Coecke and E. Proceedings of Association for Computational Linguistic. van Benthem. C. Another venue to explore is meanings of logical words. A new description of orthogonal bases. Proceedings of the 19th Annual IEEE Symposium on Logic in Computer Science. Baez and J. [17] E. Paquette. University of Edin- burgh. vol. Cartesian bicategories I. Coecke. F. [13] J. Ph. vol. 68 (1989). C. [14] F. vol. Classical non-associative Lambek calculus. Journal of Mathematical Physics. pp. [6] S. one wants to investigate if the new operations and axioms of these extensions are preserved by the vector space semantic functor. vol. 36 (1995). Coecke. 148–155. Clark. [9] B. and S. 823–843. Grefenstette. pp.

uk E-mail: coecke@cs. Psychological Review. Selinger. Springer-Verlag. Annals of Pure and Applied Logic.ac. 17 (2007). Raghavan. Moortgat. Sadrzadeh. 275–337. 270. 2008. vol. The geometry of tensor calculus I. pp. pp. OX1 3QD E-mail: stephen.pulman@cs. Cambridge University Press. [30] G. ¨ Automatic word sense discrimination.kartsaklis@qmul. Joyal and R. KARTSAKLIS. pp. 2011. vol. and B. (2009). ¨ Introduction to Information Retrieval.uk E-mail: mehrnoosh. M. Lambek. and representation of knowledge.uk . [24] J. Mathematical Structures in Com- puter Science. [33] M. 113–120. vol. SADRZADEH. Bell states and negative sentences in the distributed model of meaning. [19] G.uk DEPARTMENT OF COMPUTER SCIENCE UNIVERSITY OF OXFORD WOLFSON BUILDING. 2008. pp.ox. 1079–1100. 18 (1995). pp. 236–244. G. Formal Philosophy. COECKE [18] A. Proceedings of the 6th QPL Workshop on Quantum Physics and Logic. 141–153. Coecke. Electronic Notes in Theoretical Computer Science. P. Linguistics and Philosophy. Preller and M. 66–105. [21] J. 681–710. Polimetrica. 175–219. 189–223. 2008. Lecture Notes in Mathematics. vol. E. SCHOOL OF ELECTRONIC ENGINEERING AND COMPUTER SCIENCE QUEEN MARY UNIVERSITY OF LONDON E-mail: d.222 D. 65 (1958). M. 154–170. Coecke. Advances in Mathematics. Cambridge University Press. PARKS ROAD OXFORD. Lambek. Kelly. pp. pp. vol. [28] R. A survey of graphical languages for monoidal categories. Mitchell and M. Computational Linguistics. PULMAN. vol. Lambek vs. pp. pp. The mathematics of sentence structure. [25] T. Casadio. Grefenstette. Schutze. vol. Morrill. New Haven. Kelly. Lewis. MacLane. 164 (2013). Polimetrica. (1997). AND B. [35] P. [23] . Yale University Press. Proceedings of the 46th Annual Meeting of the Association for Computational Linguistics. English as a formal language. induction. Preller and J. Vector-based models of semantic composition. 97–123. [26] C. Dumais. [20] A.sadrzadeh@qmul. 1972. A solution to Plato’s problem: The latent semantic analysis theory of acquisition. [34] H. Springer. Manning. Schutze. [32] A. 88 (1991). pp. Lapata. Milan. vol. editors). [27] J. Symmetric categorial grammar. Laplaza. Archiv der Mathematik. pp. D. Lambek: Vector Space Semantics and String Diagrams for Lambek Calculus.ac. and S. 2003. pp. M. London Math- ematical Society Student Texts. 309–340. vol. 1974. Many-variable functorial calculus (I ). Kock. Coherence in Categories (G. M. editor). Computational Algebraic Approaches to Natural Language. [31] A.ac. 23 (1972). Strong functors and monoidal monads. and H. S. Milan. From Word to Sentence. Frobenius Algebras and 2D Topological Quantum Field Theories. Kock. Lambek and C. 2011. New Structures for Physics (B. Sadrzadeh. Landauer and S. American Mathematics Monthly.ox. 2006. [22] J.ac. Journal of Philosophical Logic. 24 (1998). Montague. [29] M. Discontinuity in categorial grammar. 281. Street. Free compact 2-categories.

one based in Temperley-Lieb recoupling theory. algebras of fermions. and we show how the operation of nega- tion in logic. This paper is an introduction to relationships between quantum topology and quantum computing. can generate the fusion algebra for a Majorana fermion. These methods are rooted in the bracket state sum model for the Jones polynomial. the other quite elementary and also based on the Temperley-Lieb algebra. Introduction. showing how knots are related not just to braiding and quantum operators. can generate the fusion algebra for a Majorana fermion. quantum entanglement and topologi- cal entanglement. the quaternions and the braid group representations related to Majorana fermions. This paper is an introduction to relationships between quantum topology and quan- tum computing. We take a foundational approach. but to quantum set theoretical foundations and algebras of fermions. the quaternions and the braid group representations related to Ma- jorana fermions. seen as both a value and an operator. It then discusses unitary solutions to the Yang-Baxter equation that are universal quantum gates. We call negation in this mode the mark. Chubb. The paper begins with these fundamentals. Harizanov Lecture Notes in Logic. its relationship with topological quantum ﬁeld theory and applies these methods to produce unitary representations of the braid groups that are dense in the unitary groups. as it operates on itself to change from marked to unmarked states. a particle that is its own anti-particle and interacts with itself either to annihilate itself or to produce itself. We show how knots are related not just to braiding and quantum operators. perception and measurement. Results are applied to give quantum algorithms for the computation of the col- ored Jones polynomials for knots and links. seen as both a value and an operator. but to quantum set theoretical foundations. A. Two constructions are given for the Fibonacci model. This paper begins an exploration of quantum epistemology in relation to the structure of discrimination as the underpinning of basic logic. They provide a conceptual key to Logic and Algebraic Structures in Quantum Computing Edited by J. and the Witten-Reshetikhin-Turaev invariant of three manifolds. Association for Symbolic Logic 223 . The paper begins with these fundmentals. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING WITH MAJORANA FERMIONS LOUIS H. We show how the operation of negation in logic. A self-contained study of the quantum universal Fibonacci model is given. KAUFFMAN Abstract. as it operates on itself to change from marked to unmarked states. We call negation in this mode the mark. 45 c 2016. Eskandarian and V. and gives an exposition of knot-theoretic recoupling theory. The mark viewed recursively as a simplest discrete dynamical system naturally generates the fermion algebra. The mark viewed recursively as a simplest discrete dynamical system naturally generates the fermion algebra. §1.

The Sections 3. the mark . the Fibonacci model for topological quantum computing is seen to be based on the fusion rules for a Majorana fermion and these in turn are the interaction rules for the mark seen as a logical particle. Temperley-Lieb algebra and topological quantum ﬁeld theory §13. Spin networks and Temperley-Lieb recoupling theory §15. This is explained in Sections 3. In particular. Laws of Form §6. cobordism categories. Knot logic §4. §1. 4 and 5. Introduction §2. underlying Boolean logic. An outline of the parts of this paper is given below. The oirginal formal structure of the mark gives the fusion algebra for the Majorana fermion. SU (2) representations of the Artin braid group §11. is composed of a “logical particle”. Braiding and topological quantum ﬁeld theory §14. Majorana fermions and algebraic knot sets §5. Fermions. A direct construction of the Fibonacci model Much of what is new in this paper proceeds from thinking about knots and sets and distinctions. Quantum mechanics and quantum computation §7. that interacts with itself to either produce itself or to cancel itself. A remark about EPR entanglement and Bell’s inequality §9. Braiding operators and universal quantum gates §8. The quaternions emerge naturally from the reentering mark. Fibonacci particles §16. KAUFFMAN many of the models that appear later in the paper. 4 and 5 are self-contained and self-explanatory. The Aravind hypothesis §10. In Section 5 we show that this iconic representation of the particle is directly . Knots and braids §3. All these models have their roots in unitary representations of the Artin braid group to the quaternions. These sections show how a formal system discovered by Spencer-Brown [92]. The bracket polynomial and the Jones polynomial §12. −→ −→ In this sense the mark is a formal model of a Majorana fermion. It requires a shift in viewpoint to see that the operator of negation can also be seen as a logical value. Quantum computation of colored Jones polynomials and the Witten-Reshetikhin-Turaev invariant §18. Quantum topology.224 LOUIS H. The Fibonacci recoupling model §17.

The patterns of interaction and braiding of such a particle P give rise to the Fibonacci model. One has fermion annihilation operators and their conjugate creation operators † . In particular we weave the topology with the algebra of fermions and in order to clarify this development. Majoranas are related to standard fermions as follows: The algebra for Majoranas is c = c † and cc = −c c if c and c are distinct Majorana fermions with c 2 = 1 and c = 1. 4. Fermion algebra. . then they anti-commute: φ = −φ. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 225 related to modeling with surface cobordisms and this theme occurs throughout the paper. The Majorana fermions c satisfy c † = c so that they are their own anti-particles. then P can interact with itself to either produce itself or to annihilate itself. There is a fundamental commutation relation † + † = 1. Recall fermion algebra. In Section 5 we also show that the mark. One can make a standard fermion from two 2 Majoranas via √ = (c + ic )/ 2. A group of researchers [78] claims. at this writing. possibly related to collective electronic excitations. generates the Cliﬀord algebra associated with a Majorana fermion and we end this section by showing how this iterant viewpoint leads naturally to the Dirac equation using the approach of [86]. (A line of fermions could have a Majorana fermion happen non-locally from one end of the line to the other. 5. This is part of the contents of the Sections 3. we give a quick summary of that algebra and a quick summary of topological quantum computing in the rest of this introduction. This is the simple “fusion algebra” for this particle. We weave this material with the emergence of unitary braid group represen- tations that are signiﬁcant for quantum information theory. The algebra of fermions is directly relevant to this knot set theory and can be formulated in terms of the Cliﬀord algebra of Majorana fermions. One can write P 2 = P + 1 to denote the two possible self-interactions the particle P. If P is a Majorana fermion particle.) The Fibonacci model that we discuss is also based on Majorana particles. to have found quasiparticle Majorana fermions in edge eﬀects in nano-wires. If you have more than one of them say and φ. √ † = (c − ic )/ 2. In these sections we examine relationships with knots as models of non-standard set theory. There is a lot of interest in these as quasi-particles and they are related to braiding and to topological quantum computing. One has 2 = 0 = ( † )2 . viewed as a generator of a discrete dynamical system.

c gives rise to a representation of the quaternion group. . c2 . cn } −→ Span{c1 . cn } via Tk (x) = k xk−1 . Let I = ba. . This is a special case of the braid group representation described above for an arbitrary list of Majorana fermions. and Tk is the identity otherwise. √ k−1 = (1 − ck+1 ck )/ 2. giving the quaternions. This gives a very nice unitary representaton of the Artin braid group and it deserves better understanding. . c3 . . This is a generalization of the well-known association of Pauli matrices and quaternions. In Section 5 we show how the dynamics of the reentering mark leads to two (algebraic) Majorana fermions e and that correspond to the spatial . . . KAUFFMAN Similarly one can mathematically make two Majoranas from any single fermion. ACA = CAC. . J = cb. It is worth noting that a triple of Majorana fermions say a. Then I 2 = J 2 = K 2 = IJK = −1. . cn } then there are natural braiding operators that act on the vector space with these ck as the basis. The operators √ A = (1/ 2)(1 + I ) √ B = (1/ 2)(1 + J ) √ C = (1/ 2)(1 + K) braid one another: ABA = BAB. Then the braiding operators are Tk : Span{c1 . . Tk (ck+1 ) = −ck . b. We have a 2 = b 2 = c 2 = 1 and they anticommute. . . Now if you take a set of Majoranas {c1 . . c2 . BCB = CBC. c2 .226 LOUIS H. These braiding operators are entangling and so can be used for universal quantum computation. K = ac. The braiding is simply: Tk (ck ) = ck+1 . . . but they give only partial topological quantum computation due to the interaction with single qubit operators not generated by them. The operators are mediated by algebra elements √ k = (1 + ck+1 ck )/ 2.

will be explored in subsequent publications. In quantum computing. the simplest non-trivial example of the Temperley- Lieb recoupling Theory gives the so-called Fibonacci model. Nevertheless. and can be used to model quantum computation universally in terms of representations of the braid group. it is fruitful to move back and forth between quantum topological methods and the techniques of quantum information theory. discuss analogies (such as topological entanglement and quantum entanglement). generalizes the Penrose theory of spin networks and is inherently topological. and the form of the invariants is closely related to the form of the computation of amplitudes in quantum mechanics. links and three dimensional manifolds have been born of this interaction. show direct correspondences between certain topological operators (solutions to the Yang-Baxter equation) and universal quantum gates. we outline the basics of the Temperley-Lieb Recoupling Theory. roughly speaking. √ √ = (e + i )/ 2 and † = (e − i )/ 2. We sketch the background topology. based on knot logic and Laws of Form. The recoupling theory yields representations of the Artin braid group into unitary groups U (n) where n is a Fibonacci number. This suggestive point of view. the approach given here is elementary. Consequently. 39] for the Jones polynomial. This paper describes relationships between quantum topology and quantum computing as a modiﬁed version of Chapter 14 of the book [12] and an expanded version of [60] and an expanded version of a chapter in [62]. The same structure can be explained in terms of the SU (2)q quantum group. These representations are dense in the unitary group. topological quantum ﬁeld theory and quantum computing. This gives a model of a fermion creation operator as a point in a non- commutative spacetime. It is built in terms of diagrammatic combinatorial topology. Temperley-Lieb recoupling Theory is based on the bracket polynomial model [35. Witten-Reshetikhin-Turaev invariants of three manifolds. We describe the background for topological quantum computing in terms of Temperley-Lieb (we will sometimes abbreviate this to TL) recoupling theory. The corresponding standard fermion operators are then given by the formulas below. In this paper. The structure is built from simple beginnings and this structure and its recoupling language can be applied to many things including colored Jones polynomials. and show explicitly how the Fibonacci model arises from . and has relationships with functional integration and Witten’s approach to topological quantum ﬁeld theory. Hence the term: topological quantum computation. Quantum topology is. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 227 and temporal aspects of this recursive process. Topological quantum computing. Many invariants of knots. that part of low-dimensional topology that interacts with statistical and quantum physics. This is a recoupling theory that generalizes standard angular momentum recoupling theory.

228 LOUIS H. the ﬁrst two P’s interact to produce a P. and the P interacts with P to produce P. than the representations we have just given. The relationship of the work here with the mathematics of Chern-Simons theory and conformal ﬁeld theory occurs through the work of Witten. leading to the Fibonacci model and its generalizations. b. we have that PP may be equal to P or to ∗ in a given situation. In another possibility. This part of the story will await a sequel to the present exposition. This is a completely elementary construction independent of the recoupling theory of the previous sections. Thus we begin with (PP)P. because they encompass a dense collection of all unitary transformations. One obtains a representation of the three strand Artin braid group on V [(ab)c] by assigning appropriate . The particles are placed in the positions a. One considers process spaces where a row of particles labeled P can successively interact. c. and the ∗ interacts with P to produce P. and wants to see how these dense unitary braid group repre- sentations arise from it. Moore and Seiberg and Moore and Read [76]. subject to the restriction that the end result is P. Here is a very condensed presentation of how unitary representations of the braid group are constructed via topological quantum ﬁeld theoretic methods. including single qubit transformations needed for universal quantum computing. The diagrammatic computations in the sections 12 to 18 are completely self-contained and can be used by a reader who has just learned the bracket polynomial. One has a mathematical particle with label P that can interact with itself to produce either itself labeled P or itself with the null label ∗. When ∗ interacts with P the result is always P. It follows from this analysis that the space of linear combinations of processes V [(ab)c] is two dimensional. Thus for our particle P. and this may partake of a number of possibilities. One can compare the mathematical techniques of the present paper with the physics of the quantum Hall eﬀect and its possibilities for topological quantum computing. KAUFFMAN it. but it is understood that the “value” of PQ is the result of the interaction. We studied this construction in [61] and a version of it has been used in [89]. We shall denote the interaction of two particles P and Q by the expression PQ. (PP)P −→ (∗)P −→ P. In the ﬁnal section of the paper we give a separate construction for the Fibnacci model that is based on 2 × 2 complex matrix representations of the Temperley-Lieb algebra. the ﬁrst two P’s interact to produce a ∗. In a typical sequence of interactions. When ∗ interacts with ∗ the result is always ∗. The two processes we have just described can be taken to be the qubit basis for this space. in principle. These representations are more powerful. (PP)P −→ (P)P −→ P. For example the space V [(ab)c] denotes the space of interactions of three particles labeled P.

In this scheme. Figure 1. These issues are illustrated in Figure 1. where the parenthesization of the particles is indicated by circles and by also by trees. The trees can be taken to indicate patterns of particle interaction. The other operator for this representation corresponds to the interchange of b and c. This interchange is accomplished by a unitary change of basis mapping F : V [(ab)c] −→ V [a(bc)]. A full representation of the Artin braid group on each space is deﬁned in terms of the local interchange phase gates and the . Braiding anyons. If A : V [(ab)c] −→ V [(ba)c] is the ﬁrst braiding operator (corresponding to an interchange of the ﬁrst two particles in the association) then the second operator B : V [(ab)c] −→ V [(ac)b] is accomplished via the formula B = F −1 RF where the R in this formula acts in the second vector space V [a(bc)] to apply the phases for the interchange of b and c. vector spaces corresponding to associated strings of particle interactions are interrelated by recoupling transformations that generalize the mapping F indicated above. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 229 phase changes to each of the generating processes. See also Figure 50 for an illustration of the braiding B = F −1 RF . where two particles interact at the branch of a binary tree to produce the particle product at the root. One can think of these phases as corresponding to the interchange of the particles labeled a and b in the association (ab)c.

A knot is an embedding of a circle in three-dimensional space. 31. Knots and braids. In our approach the structure of phase gates and recoupling transformations arise naturally from the structure of the bracket model for the Jones polynomial. I have woven this work into the present paper in a form that is coupled with recent and previous work on relations with logic and with Majorana fermions. Two recent books contain material relevant to the context of this paper. Then again long molecules such as rubber molecules and DNA molecules can be knotted and linked. There are many applications of the theory of knots. a problem of studying the topological forms that can be made by placing one space inside another. 47. 90]. the braiding of quasi- particles (collective excitations) leads to non-trival representations of the Artin braid group. 19. 61. taken up to ambient isotopy. These gates and transformations have to satisfy a number of identities in order to produce a well-deﬁned representation of the braid group. These previous papers are an exploration of the foundations of knot theory in relation to Laws of Form. KAUFFMAN recoupling transformations. 34. Such particles are called Anyons. More work needs to be done in all these domains. 73. These identities were discovered originally in relation to topological quantum ﬁeld theory. Topology is a background for the physical structure of real knots made from rope of cable. The purpose of this section is to give a quick intro- duction to the diagrammatic theory of knots. 46. 10. The interested reader should examine these approaches to fundamental physics. 45. At the level of discrete dynamical systems the papers are related to foundations of physics. recursion and discrete dynamical systems. Lomonaco in the papers [56. 63. As a result. the ﬁeld of practical knot tying is a ﬁeld of applied topology that existed well before the mathematical discipline of topology arose. In the case of knot theory we consider the placements of a circle inside three dimensional space. The braiding in these models is related to topological quantum ﬁeld theory. non-standard set theory. 62]. 65. 51. In modeling the quantum Hall eﬀect [94. .230 LOUIS H. 33. 52]. 56. links and braids. 58. 57. It is planned to use this paper and other joint work as a springboard for a book [55] on topological quantum information theory and for a book that expands on the foundational issues raised in this paper and the previous papers of the author. They are [87] and [86]. 71. 50. It is hoped that the mathematics we explain here will form a bridge between theoretical models of anyons and their applications to quantum computing. 40. 59. The problem of deciding whether two knots are isotopic is an example of a placement problem. §2. Acknowledgements. Thus we obtain a knot-theoretic basis for topological quantum computing. 32. Much of this paper is based upon joint work with Samuel J. The relations with logic stem from the following previous papers of the author [67. 54. 64. 72.

Figure 3. 85. In this paper we will study how knot theory can be used to produce unitary representations of the braid group. That is. Figure 2 illustrates a diagram for a knot. A link is an embedding of a disjoint collection of circles. Many invariants such as the Jones polynomial are constructed via partition functions and generalized quantum amplitudes. Knot theory is closely related to theoretical physics as well with applications in quantum gravity [91. As a result. The Reidemeister moves. Quantum topology is the study and invention of topological invariants via the use of analogies and techniques from mathematical physics. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 231 There have been a number of intense applications of knot theory to the study of DNA [17] and to polymer physics [42]. Such representations can play a fundamental role in quantum computing. Figure 2. The diagram is regarded both as a schematic picture of the knot. These moves are illustrated . and as a plane graph with extra structure at the nodes (indicating how the curve of the knot passes over or under itself by standard pictorial conventions). A knot diagram. Two knots are regarded as equivalent if one embedding can be obtained from the other through a continuous family of embeddings of circles in three- space. Ambient isotopy is mathematically the same as the equivalence relation generated on diagrams by the Reidemeister moves. taken up to ambient isotopy. 53] and many applications of ideas in physics to the topological structure of knots themselves [39]. one expects to see relationships between knot theory and physics.

The Reidemeister moves are of great use for analyzing the structure of knot invariants and they are closely related to the Artin braid group. Figure 5. in Figure 3. The Reidemeister moves are useful in doing combinatorial topology with knots and links. notably in working out the behaviour of knot invariants. A braid is an embedding of a collection of strands that have their ends in two rows of points that are set one above the other with respect to a choice of vertical. which we discuss below. The strands are not individually knotted and they are disjoint from one another. KAUFFMAN Figure 4. Braid generators.232 LOUIS H. identical numbers). Closing braids to form knots and links. Each move is performed on a local part of the diagram that is topologically identical to the part of the diagram illustrated in this ﬁgure (these ﬁgures are representative examples of the types of Reidemeister moves) without changing the rest of the diagram. Figure 5 and Figure 6 for illustrations of braids and . A knot invariant is a function deﬁned from knots and links to some other mathematical object (such as groups or polynomials or numbers) such that equivalent diagrams are mapped to equivalent objects (isomorphic groups. See Figure 4. identical polynomials.

. 2. Thus the theory of braids is critical to the theory of knots and links. A key theorem of Alexander states that every knot or link can be represented as a closed braid. n − 2. See Figure 4 for an illustration of the elementary braids and their relations. Braids are diagrammed vertically as in Figure 4. Let Bn denote the Artin braid group on n strands. ﬁxing the endpoints. one sees the identities s1 s1−1 = 1 (where the identity element in B4 consists in four vertical strands). where a braid is an intertwining of strands that lead from one set of n points to another set of n points. In Figure 4 we illustrate the form of the basic generators of the braid group. and ﬁnally s1 s3 = s3 s1 . . The product of two braid diagrams is accomplished by adjoining the top strands of one braid to the bottom strands of the other braid. s1 s2 s1 = s2 s1 s2 . Figure 5 illustrates how to close a braid by attaching the top strands to the bottom strands by a collection of parallel arcs. Braids are a key structure in mathematics. and then the inverse of the ﬁrst generator is drawn. Braids can be multiplied by attaching the bottom row of one braid to the top row of the other braid. We recall here that Bn is generated by elementary braids {s1 . Following this. . The braid generators si are represented by diagrams where the i-th and (i + 1)-th strands wind around one another by a single half-twist (the sense of this turn is shown in Figure 4) and all other strands drop straight to the bottom. sn−1 } with relations 1. Borromean rings as a braid closure. . . si sj = sj si for |i − j| > 1. Figure 6 illustrates the famous Borromean Rings (a link of three unknotted loops such that any two of the loops are unlinked) as the closure of a braid. si si+1 si = si+1 si si+1 for i = 1. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 233 Figure 6. and the products are taken in order from top to bottom. the braids form a group under this notion of multiplication. . From the algebraic point of view the braid groups Bn are important extensions of the symmetric . and the form of the relations among these generators. moves on braids. In Figure 4 we have restricted the illustration to the four-stranded braid group B4 . . It is not just that they are a collection of groups with a vivid topological interpretation. Note that the braid group has a diagrammatic topological interpretation. Taken up to ambient isotopy. In that ﬁgure the three braid generators of B4 are shown.

A curve or segment with no curves passing underneath it is the empty set. n − 1. The entities a and b that are in the relation a ∈ b are diagrammed as segments of lines or curves. . 3. . This mutuality is diagrammed as topological linking. and and sets can be members of each other as in Figure 11. . with the a-curve passing underneath the b-curve. we indicate two sets. 2. In the next sections we shall show how representations of the Artin braid group are rich enough to provide a dense set of transformations in the unitary groups. consisting of a mark crossing over another mark. The ﬁrst (looking like a right-angle bracket that we refer to as the mark) is the empty set. We have an exact sequence of groups 1 −→ Pn −→ Bn −→ Sn −→ 1 exhibiting the Artin braid group as an extension of the pure braids Pn (inducing the identity permutation). Figure 7. si2 = 1 for i = 1. Recall that the symmetric group Sn of all permutations of n distinct objects has presentation as shown below. In the Figure 8. si sj = sj si for |i − j| > 1. si si+1 si = si+1 si si+1 for i = 1.234 LOUIS H. . building the von Neumann construction of the natural numbers in this notation as in Figure 9. More about this point of view can be found in the author’s paper “Knot Logic” [41]. We can continue this construction. This notation allows us to also have sets that are members of themselves as in Figure 10. This leads to the question . §3. Membership. Membership is represented by under-passage of curve segments. . KAUFFMAN groups Sn . We shall use knot and link diagrams to represent sets. n − 2. We write a ∈ S to say that a is a member of the set S. is the set whose only member is the empty set. Set theory is about an asymmetric relation called membership. Thus the braid groups are in principle fundamental to quantum computation and quantum information theory. by the symmetric group. Thus Sn is obtained from Bn by setting the square of each braiding generator equal to one. Knot logic. The second. In this section we shall diagram the membership relation as in Figure 7. 1. .

KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 235

Figure 8. Von Neumann 1.

Figure 9. Von Neumann 2.

**beyond ﬂatland: Is there a topological interpretation for this way of looking at
**

set-membership?

Figure 10. Omega is a member of Omega.

Figure 11. Mutual membership.

**Consider the example in Figure 12, modiﬁed from the previous one. The
**

link consisting of a and b in this example is not topologically linked. The two

236 LOUIS H. KAUFFMAN

Figure 12. Cancellation.

**components slide over one another and come apart. The set a remains empty,
**

but the set b changes from b = {a, a} to empty. This example suggests the

following interpretation.

Regard each diagram as specifying a multi-set (where more than one instance

of an element can occur), and the rule for reducing to a set with one representative

for each element is: Elements of knot sets cancel in pairs. Two knot sets are said

to be equivalent if one can be obtained from the other by a ﬁnite sequence of pair

cancellations.

This equivalence relation on knot sets is in exact accord with the ﬁrst

Reidemeister move as shown in Figure 13.

Figure 13. Reidemeister 2.

**There are other topological moves, and we must examine them as well. In
**

fact, it is well-known that topological equivalence of knots (single circle embed-

dings), links (mutltiple circle embeddings) and tangles (arbitrary diagrammatic

embeddings with end points ﬁxed and the rule that you are not allowed to move

strings over endpoints) is generated by three basic moves (the Reidemeister

moves) as shown in Figure 14. See [39].

It is apparent that move III does not change any of the relationships in the

knot multi-sets. The line that moves just shifts and remains underneath the

other two lines. On the other hand move number one can change the self-

referential nature of the corresponding knot-set. One goes, in the ﬁrst move,

between a set that indicates self-membership to a set that does not indicate self-

membership (at the site in question). See Figure 15 This means that in knot-set

theory every set has representatives (the diagrams are the representatives of

KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 237

Figure 14. Reidemeister moves.

**the sets) that are members of themselves, and it has representatives that are
**

not members of themselves. In this domain, self-membership does not mean

inﬁnite descent. We do not insist that

a = {a}

implies that

a = {{{{. . . }}}}.

Rather, a = {a} just means that a has a little curl in its diagram. The Russell

set of all sets that are not members of themselves is meaningless in this domain.

**Figure 15. Reidemeister I: Replacing self-membership with
**

no self-membership.

Figure 16. Trefoil is an empty knotset.

238 LOUIS H. KAUFFMAN

Figure 17. Chain.

Figure 18. Borromean rings.

**We can summarize this ﬁrst level of knot-set theory in the following two
**

equivalences:

1. Self-Reference: a = {b, c, . . . } ⇐⇒ a = {a, b, c, . . . }

2. Pair Cancellation: S = {a, a, b, c, . . . } ⇐⇒ S = {b, c, . . . }

With this mode of dealing with self-reference and multiplicity, knot-set theory

has the interpretation in terms of topological classes of diagrams. We could

imagine that the ﬂatlanders felt the need to invent three dimensional space and

topology, just so their set theory would have such an elegant interpretation.

But how elegant is this interpretation, from the point of view of topology?

Are we happy that knots are equivalent to the empty knot-set as shown in

Figure 16? For this, an extension of the theory is clearly in the waiting. We

are happy that many topologically non-trivial links correspond to non-trivial

knot-sets. In the Figure 17 , a chain link becomes a linked chain of knot-sets.

But consider the link shown in Figure 18. These rings are commonly called the

Borromean Rings. The Rings have the property that if you remove any one of

them, then the other two are topologically unlinked. They form a topological

tripartite relation. Their knot-set is described by the three equations

a = {b, b}

b = {c, c}

c = {a, a}.

**Thus we see that this representative knot-set is a “scissors-paper-stone”
**

pattern. Each component of the Rings lies over one other component, in a

cyclic pattern. But in terms of the equivalence relation on knot sets that we

have used, the knot set for the Rings is empty (by pair cancellation).

KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 239

**In order to go further in the direction of topological invariants for knots and
**

links it is necessary to use more structure than the simple membership relation

that motivates the knots-sets. Viewed from the point of view of the diagrams

for knots and links there are a number of possible directions. For example, one

can label all the arcs of the diagram and introduce algebraic relations at each

crossing. This leads to the fundamental group and the quandle [39]. One can

also label all the arcs of the diagram from an index set and view this labeling

as a state in analogous to a state of a physical system in statistical mechanics.

Then evaluations of these states and summations of the evaluations over all

the states give the class of knot invariants called quantum invariants for knots

and links [39]. These include the Jones polynomial and its generalizations.

In this paper we will explain and use the Jones polynomial and the so-called

colored Jones polynomials. See Section 17 for this development. The purpose

of this section has been to introduce the subject of knot and link diagrams

in the context of thinking about foundations of mathematics. However, it

is worthwhile adding structure to the knot set theory so that it can at least

see the higher order linking of the Borommean rings. We do this in the next

subsection by keeping track of the order in which sets are encountered along

the arc of a given component, and by keeping track of both membership and

co-membership where we shall say that A is co-member of B if B is a member

of A. As one moves along an arc one sequentially encounters members and

co-members.

3.1. Ordered knot sets. Take a walk along a given component. Write down

the sequence of memberships and belongings that you encounter on the walk

as shown in Figure 19.

Figure 19. An ordered knot set.

**In this notation, we record the order in which memberships and “co-member-
**

ships” (a is a co-member of b if and only if b is a member of a) occur along

the strand of a given component of the knot-set. We do not choose a direction

of traverse, so it is ok to reverse the total order of the contents of a given

component and to take this order up to cyclic permutation. Thus we now have

the representation of the Borromean Rings as shown in Figure 20.

With this extra information in front of us, it is clear that we should not allow

the pair cancellations unless they occur in direct order, with no intervening

co-memberships. Lets look at the revised Reidemeister moves as in Figure 21.

240 LOUIS H. KAUFFMAN

Figure 20. Borromean rings as ordered knot set.

Figure 21. Reidemeister moves for ordered knot sets.

**As is clear from the above diagrams, the Reidemeister moves tell us that we
**

should impose some speciﬁc equivalences on these ordered knot sets:

1. We can erase any appearance of a[a] or of [a]a inside the set for a.

2. If bb occurs in a and [a][a] occurs in b, then they can both be erased.

3. If bc is in a, ac is in b and a[b] is in c, then we can reverse the order of

each of these two element strings.

We take these three rules (and a couple of variants suggested by the diagrams)

as the notion of equivalence of ordered knot-sets. One can see that the ordered

knot-set for the Borromean rings is non-trivial in this equivalence relation.

In this sense we have a a proof that the Borromean rings are linked, based

on their scissors, paper, stone structure. The only proof that I know for the

KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 241

**non-triviality of the Borommean ordered knot set uses the concept of coloring
**

discussed in the next subsection.

Knots and links are represented by the diagrams themselves, taken up the

equivalence relation generated by the Reidemeister moves. This calculus of

diagrams is quite complex and it is remarkable, the number and depth of

diﬀerent mathematical approaches that are used to study this calculus and its

properties. Studying knots and links is rather like studying number theory. The

objects of study themselves can be constructed directly, and form a countable

set. The problems that seem to emanate naturally from these objects are

challenging and fascinating. For more about knot-sets, see [40]

3.2. Quandles and colorings of knot diagrams. There is an approach to

studying knots and links that is very close to our ordered knot sets, but starts

from a rather diﬀerent premise. In this approach each arc of the diagram

receives a label or “color”. An arc of the diagram is a continuous curve in the

diagram that starts at one under crossing and ends at another under crossing.

For example, the trefoil diagram is related to this algebra as shown in Figure 22.

Figure 22. The quandle for the trefoil knot.

**Each arc corresponds to an element of a “color algebra” IQ(T ) where T
**

denotes the trefoil knot. We have thatIQ(T ) is generated by colors a,b and

c with the relations c ∗ b = a, a ∗ c = b, b ∗ a = c, a ∗ a = a. Each of these

relations is a description of one of the crossings in T . These relations are

speciﬁc to the trefoil knot. If we take on an algebra of this sort, we want its

coloring structure to be invariant under the Reidemeister moves. This implies

the following global relations:

x∗x =x

(x ∗ y) ∗ y = x

(x ∗ y) ∗ z = (x ∗ z) ∗ (y ∗ z)

for any x, y and z in the algebra (set of colors) IQ(T ). An algebra that satisﬁes

these rules is called an Involutory Quandle (See [39]), hence the initials IQ.

These global relations are really expressions of the concept of self-crossing and

iterated crossing in the multiplicity of crossings that are available in a calculus

of boundaries where the notation indicates the choice of interpretation, where

Knot theory can be seen as a natural articulation not of three dimensional space (a perfectly good interpretation) but of the properties of interactions of boundaries. d }. By the same token. the coloring proofs can be transferred to ordered knot sets in the case of links. c = {bd }. Majorana fermions and algebraic knot sets.242 LOUIS H. c}. The ordered knot set corresponding to a link can be colored or not colored in the same manner as a link diagram. The spaces between the letters in the ordered code of the knot set can be assigned colors in the same way as the arcs of a link diagram. Fermions. b = {a. We . Each boundary can be regarded as that boundary transgressed by another boundary. where we remove the parentheses and write the contents of each set as a algebraic product. then two diagrams that are related by the Reidemeister moves will have isomorphic algebras. For example. In this way of thinking. We leave the details of this analysis of link sets to another paper. §4. KAUFFMAN one boundary is seen to cross (over) the other boundary. c = {b. b ∗ b = b. c = {c}. Three-coloring turns out to be quite useful for many knots and links. Thus we have seen that the trefoil knot is knotted due to its having a non-trivial three-coloring. but go back to the ordinary knot sets that just catalog memberships. a ∗ a = a. Lets take this view. the Borommean rings must be linked. c = {c}. Since there is no such coloring. one can see that the Borommean rings are linked by checking that they do not have a non-trivial three-coloring! This fact is easy to check by directly trying to color the rings. c ∗ c = c. This coloring of the unlinked rings would then induce a coloring of the Borommean rings. Then the knot set is a ordered list of the memberships that are encountered along the weave. b ∗ a = c. They will also inherit colorings of their arcs from one another. Thus the calculation of the algebra IQ(K ) for a knot or link K has the potentiality for bringing forth deep topological structure from the diagram. The choice of who is the transgressed and who transgresses is the choice of a crossing. In the last part of our discussion of knot sets we added order and co-membership to the structure. If we adopt these global relations for the algebra IQ(K ) for any knot or link diagram K . In the case of the trefoil. c. In this way. the knot set is an ordered sequence of memberships and co-memberships that are encountered as one moves along the strand of that part of the weave. a ∗ c = b. That uncolorability implies that the rings are linked follows from the fact that there is a non-trivial coloring of three unlinked rings (color each ring by a separate color). a ∗ a = a. and this would become the algebraic statements a = {b}. b = {ac}. forming the three-color quandle. the choice of membership in the context of knot-set theory. We have the complete set of relations c ∗ b = a. in Figure 17 we have a = {b}. b. one can see that the algebra actually closes at the set of elements a.

. CA = B. . and we see that the rule x 2 = 1 should be obeyed by this algebra of products of set members. 8. By introducing the Cliﬀord algebra with x 2 = 1 and xy = −yx for generators. Then we can deﬁne X ΔY for sets X = {α} and Y = {} by the equation X ΔY = {α} where α represents the product of the members of X and Y taken together. Then we have A2 = B 2 = C 2 = {1} where it is understood that {1} = {} represents the empty set. suppose A = {yx}. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 243 retain the brackets in order to continue to diﬀerentiate the set from its contents.) Furthermore we have AB = C. This commutative law would disregard the ordering. adding a notion of sign to the algebraic representation of the knot sets. What shall we do about a = {bcdc}? We could decide that xy = yx for all x and y in a given knot set. This example suggests that we could change the algebraic structure so that members satisfy xy = −yx. . In order to explain this association. 70]. We then get the pattern of the quaternion group: A2 = B 2 = C 2 = ABC = −1 where −1 denotes the “negative” empty set. mi mj + mj mi = 0 when i = j. in the algebra 1 represents the empty word. For example. There is a natural association of fermion algebra to knot sets. B = {zy}. In a standard collection of fermion operators m1 . and we would have {bcdc} = {bccd } = {bd }. Then we would have that {bccd } = {bd } since repetitions are eliminated. The relations in this example are very close to the quaternions. (That is. 26. 28. The generators of this Cliﬀord algebra represent fermions that are their own anti-particles. . we ﬁrst give a short exposition of the algebra of fermion operators. mk one has that each mi is a linear operator on a Hilbert space with an adjoint operator mi† (corresponding to the anti-particle for the particle created by mi ) and relations mi2 = 0. The simplest algebraic version of the knot sets is to have a commutative algebra with x 2 = 1 for all members. BC = A. The operation X ΔY represents the union of knot sets and corresponds to exclusive or in standard set theory. mi mi† + mi† mi = 1. More recently. it has been suggested that Majorana fermions may occur in collective electronic phenomena [74. we bring the knot sets into direct correspondence with an algebra of Majorana fermions. For a long time it has been conjectured that neutrinos may be Majorana fermions. . C = {xz}.

. y. B = zy. √ s2 s1 s2 = (1/2 2)(1 + B)(1 + A)(1 + B) √ = (1/2 2)(1 + A + B + BA)(1 + B) √ = (1/2 2)(1 + A + B + BA + B + AB + B 2 + BAB) √ = (1/2 2)(1 + A + B + BA + B − BA − 1 + A) √ = (1/ 2)(A + B). . Let A = yx. C represent the quaternions. then formal real and imaginary parts of m yield a mathematical pair of Majorana fermions. z be three Majorana fermions. It is then easy to see that a 2 = b 2 = 1 and ab = −ba imply that m and m † form a fermion in the sense that m 2 = (m † )2 = 0 and mm † + m † m = 1. The non-local nature of this pair is promising for creating topologically protected qubits. Let x. Thus pairs of Majorana fermions can be construed as ordinary fermions. B. if m is an ordinary fermion.244 LOUIS H. s2 = (1 + B)/ 2. Given two Majorana fermions a and b with a 2 = b 2 = 1 and ab = −ba. ck and ci2 = 1 while ci cj = −cj ci for all i = j. . For example. There is a algebraic translation between the fermion algebra and Majorana fermion algebra. conceived in this way can give rise to a chain of Majorana fermions with a non-localized pair corresponding to the distant ends of the chain. deﬁne √ m = (a + ib)/ 2 and √ m † = (a − ib)/ 2. Now deﬁne √ √ √ s1 = (1 + A)/ 2. KAUFFMAN There is another brand of Fermion algebra where we have generators c1 . Here is an example that shows how the topology comes in. √ s1 s2 s1 = (1/2 2)(1 + A)(1 + B)(1 + A) √ = (1/2 2)(1 + A + B + AB)(1 + A) √ = (1/2 2)(1 + A + B + AB + A + A2 + BA + ABA) √ = (1/2 2)(1 + A + B + AB + A − 1 − AB + B) √ = (1/ 2)(A + B). s3 = (1 + C )/ 2. Similarly. . It is easy to see that si and sj satisfy the braiding relation for any i = j. A chain of electrons in a nano-wire. These are the Majorana fermions. and there is at this writing an experimental search for evidence for the existence of such end-eﬀect Majorana fermions. C = xz. Conversely. j = 2. We now see that it is exactly the Majorana fermion algebra that matches the properties of the knot sets. here is the veriﬁcation for i = 1. We have already seen that A.

The Cliﬀord algebra for Majorana fermions makes a speciﬁc choice in the matter and in this way ﬁxes the representation of the braiding. The topology alone tells us only the relative change of phase between the two particles. These braiding operators can be seen to act on the vector space over the complex numbers that is spanned by the fermions x. y} and{x. z}. we remark that linear combinations of products in the Cliﬀord algebra can be regarded as superpositions of the knot sets. consider 1 + yx s= √ . see [39]. Superposition of sets suggests that we are creating a species of quantum set theory and indeed Cliﬀord algebra based quantum set theories have been suggested (see [18]) by David Finkelstein and others. But braiding invariance of certain linear combinations of sets is a relationship with knotting at a second level. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 245 Thus s1 s2 s1 = s2 s1 s2 . On interchange. Finally. This braid group representation is signiﬁcant for quantum computing as we shall see in Section 7. This requires more investigation. Now view Figure 23 where we have illustrated a topological interpretation for the braiding of two fermions. It is also clear that this Cliﬀord algebraic quantum set theory should be related to our previous constructions for quantum knots [58. the braid group representation shows that the Cliﬀord algebraic representation for knot sets is related to topology at more than one level. 2 1 + yx 1 − yx T (p) = sps −1 = ( √ )p( √ ). 73. and so a natural braid group representation arises from the Majorana fermions. z. (For more information on this topological interpretation of 2 rotation for fermions. 2 2 and verify that T (x) = y and T (y) = −x.) Without a further choice it is not evident which particle of the pair should receive the phase change. . It may come as a surprise to a quantum set theorist to ﬁnd that knot theoretic topology is directly related to this subject. y. For the purpose of this discussion. 72. 63]. the belt becomes twisted by 2 . Thus xy + xz is a superposition of the sets with members {x. To see how this works. 71. invariant under the Reidemeister moves (up to a global sign). and it suggests that knot theory and the theory of braids occupy a fundamental place in the foundations of quantum mechanics. taken as products of generators. We will make one more remark here. In the topological interpretation a twist of 2 corresponds to a phase change of −1. In the topological interpretation the two fermions are connected by a ﬂexible belt. and reserve further analysis for a subsequent paper. This multiple relationship certainly deserves more thought. The relation x 2 = 1 for generators makes the individual sets.

Inside and outside. The reason we introduce this notation is that in the calculus of indications the mark can interact with itself in two possible ways. the mark: . The patterns of this mark and its self-interaction match those of a Majorana fermion as discussed in the previous section. As is evident from Fgure 24. Laws of Form. Braiding action on a pair of fermions. but fundamentally simpler than the usual Boolean arithmetic of 0 and 1 with its two binary operations and one unary operation (negation). This calculus is a study of mathematical foundations with a topological notation based on one symbol. In the calculus of indications one takes a step in the direction of simplicity. the mark is regarded as a shorthand for a rectangle drawn in the plane and dividing the plane into the regions inside and outside the rectangle. §5. Figure 24. This single symbol represents a distinction between its own inside and outside. The resulting formalism becomes a version of Boolean arithmetic. A Majorana fermion is a particle . In this section we discuss a formalism due the G.246 LOUIS H. Spencer- Brown [92] that is often called the “calculus of indications”. and also a step in the direction of physics. KAUFFMAN Figure 23.

KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 247 that is its own anti-particle. in this paper. non- intersecting rectangles) are called expressions. we have illustrated both the rectangle and the marked version of the expression. In an expression you can say deﬁnitively of any two marks whether one is or is not inside the other. that by adding braiding to the calculus of indications we arrive at the Fibonacci model. But in the normal use of negation there is no way that the negation sign combines with itself to produce itself.e. A fusion rule represents all of the diﬀerent particle interactions in the form of a set of equations. The relationship between two marks is . The fusion rule is P 2 = 1 + P. In Figure 25. This is the bare minimum that we shall need. This represents the fact that P can interact with itself to produce the neutral particle (represented as 1 in the fusion rule) or itself (represented by P in the fusion rule). [74]. Thus the possible interactions are PP −→ ∗ and PP −→ P. Here we describe the particle directly in terms of its interactions. This appears to ruin the analogy between negation and the Majorana fermion. In the calculus of indications patterns of non-intersecting marks (i. In the previous section we described Majorana fermions in terms of their algebra of creation and annihilation operators. that can in principle support quantum computing. the mark. And so we might write ∼∼ −→ ∗ where ∗ is a neutral linguistic particle. Is there a linguistic particle that is its own anti-particle? Certainly we have ∼∼Q = Q for any proposition Q (in Boolean logic). This is part of a general scheme called “fusion rules” [76] that can be applied to discrete particle interacations. The bare bones of the Majorana fermion consist in a particle P such that P can interact with itself to produce a neutral particle ∗ or produce itself P. Remarkably. the calculus of indications provides a context in which we can say exactly that a certain logical particle. We will later see. We shall come back to the combinatorics related to this fusion equation. an identity operator so that ∗Q = Q for any proposition Q. can act as negation and can interact with itself to produce itself. For example in Figure 25 we see how patterns of boxes correspond to patterns of marks.

248 LOUIS H. KAUFFMAN

Figure 25. Boxes and marks.

**either that one is inside the other, or that neither is inside the other. These two
**

conditions correspond to the two elementary expressions shown in Figure 26.

**Figure 26. Translation between boxes and marks.
**

The mathematics in Laws of Form begins with two laws of transformation

about these two basic expressions. Symbolically, these laws are:

1. Calling: = ,

2. Crossing: = .

The equals sign denotes a replacement step that can be performed on instances

of these patterns (two empty marks that are adjacent or one mark surrounding

an empty mark). In the ﬁrst of these equations two adjacent marks condense

to a single mark, or a single mark expands to form two adjacent marks. In

the second equation two marks, one inside the other, disappear to form the

unmarked state indicated by nothing at all. That is, two nested marks can be

replaced by an empty word in this formal system. Alternatively, the unmarked

state can be replaced by two nested marks. These equations give rise to a natural

calculus, and the mathematics can begin. For example, any expression can be

reduced uniquely to either the marked or the unmarked state. The following

example illustrates the method:

= =

= = .

The general method for reduction is to locate marks that are at the deepest

places in the expression (depth is deﬁned by counting the number of inward

crossings of boundaries needed to reach the given mark). Such a deepest mark

must be empty and it is either surrounded by another mark, or it is adjacent to

an empty mark. In either case a reduction can be performed by either calling

or crossing.

KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 249

**Laws of Form begins with the following statement. “We take as given the
**

idea of a distinction and the idea of an indication, and that it is not possible

to make an indication without drawing a distinction. We take therefore the

form of distinction for the form.” Then the author makes the following two

statements (laws):

1. The value of a call made again is the value of the call.

2. The value of a crossing made again is not the value of the crossing.

The two symbolic equations above correspond to these statements. First

examine the law of calling. It says that the value of a repeated name is the

value of the name. In the equation

=

one can view either mark as the name of the state indicated by the outside of

the other mark. In the other equation

= .

**the state indicated by the outside of a mark is the state obtained by crossing
**

from the state indicated on the inside of the mark. Since the marked state is

indicated on the inside, the outside must indicate the unmarked state. The Law

of Crossing indicates how opposite forms can ﬁt into one another and vanish

into nothing, or how nothing can produce opposite and distinct forms that ﬁt

one another, hand in glove. The same interpretation yields the equation

=

where the left-hand side is seen as an instruction to cross from the unmarked

state, and the right hand side is seen as an indicator of the marked state.

The mark has a double carry of meaning. It can be seen as an operator,

transforming the state on its inside to a diﬀerent state on its outside, and it

can be seen as the name of the marked state. That combination of meanings is

compatible in this interpretation.

From the calculus of indications, one moves to algebra. Thus

A

**stands for the two possibilities
**

= ←→ A =

= ←→ A =

**In all cases we have
**

A = A.

250 LOUIS H. KAUFFMAN

**By the time we articulate the algebra, the mark can take the role of a unary
**

operator

A −→ A .

But it retains its role as an element in the algebra. Thus begins algebra with

respect to this non-numerical arithmetic of forms. The primary algebra that

emerges is a subtle precursor to Boolean algebra. One can translate back and

forth between elementary logic and primary algebra:

1. ←→ T

2. ←→ F

3. A ←→ ∼A

4. AB ←→ A ∨ B

5. A B ←→ A ∧ B

6. A B ←→ A ⇒ B

The calculus of indications and the primary algebra form an eﬃcient system

for working with basic symbolic logic.

By reformulating basic symbolic logic in terms of the calculus of indications,

we have a ground in which negation is represented by the mark and the mark is

also interpreted as a value (a truth value for logic) and these two intepretations

are compatible with one another in the formalism. The key to this compatibility

is the choice to represent the value “false” by a literally unmarked state in the

notational plane. With this the empty mark (a mark with nothing on its inside)

can be interpreted as the negation of “false” and hence represents “true”. The

mark interacts with itself to produce itself (calling) and the mark interacts

with itself to produce nothing (crossing). We have expanded the conceptual

domain of negation so that it satisﬁes the mathematical pattern of an abstract

Majorana fermion.

Another way to indicate these two interactions symbolically is to use a

box,for the marked state and a blank space for the unmarked state. Then one

has two modes of interaction of a box with itself:

1. Adjacency:

and

2. Nesting: .

With this convention we take the adjacency interaction to yield a single box,

and the nesting interaction to produce nothing:

=

=

KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 251

**We take the notational opportunity to denote nothing by an asterisk (*). The
**

syntatical rules for operating the asterisk are Thus the asterisk is a stand-in for

no mark at all and it can be erased or placed wherever it is convenient to do so.

Thus

= ∗.

At this point the reader can appreciate what has been done if he returns to

the usual form of symbolic logic. In that form we that

∼∼X = X

for all logical objects (propositions or elements of the logical algebra) X . We

can summarize this by writing

∼∼ =

as a symbolic statement that is outside the logical formalism. Furthermore,

one is committed to the interpretation of negation as an operator and not as

an operand. The calculus of indications provides a formalism where the mark

(the analog of negation in that domain) is both a value and an object, and so

can act on itself in more than one way.

The Majorana particle is its own anti-particle. It is exactly at this point

that physics meets logical epistemology. Negation as logical entity is its

own anti-particle. Wittgenstein says (Tractatus [96] 4.0621) “. . . the sign ‘∼’

corresponds to nothing in reality.” And he goes on to say (Tractatus 5.511)

“How can all-embracing logic which mirrors the world use such special catches

and manipulations? Only because all these are connected into an inﬁnitely ﬁne

network, the great mirror.” For Wittgenstein in the Tractatus, the negation sign

is part of the mirror making it possible for thought to reﬂect reality through

combinations of signs. These remarks of Wittgenstein are part of his early

picture theory of the relationship of formalism and the world. In our view,

the world and the formalism we use to represent the world are not separate.

The observer and the mark are (formally) identical. A path is opened between

logic and physics.

The visual iconics that create via the boxes or half-boxes of the calculus of

indications a model for a logical Majorana fermion can also be seen in terms

of cobordisms of surfaces. View Figure 27. There the boxes have become

circles and the interactions of the circles have been displayed as evolutions in an

extra dimension, tracing out surfaces in three dimensions. The condensation

of two circles to one is a simple cobordism betweem two circles and a single

circle. The cancellation of two circles that are concentric can be seen as the

right-hand lower cobordism in this ﬁgure with a level having a continuum of

critical points where the two circles cancel. A simpler cobordism is illustrated

above on the right where the two circles are not concentric, but nevertheless

are cobordant to the empty circle. Another way of putting this is that two

topological closed strings can interact by cobordism to produce a single string

252 LOUIS H. KAUFFMAN

**or to cancel one another. Thus a simple circle can be a topological model for a
**

Majorana fermion.

Figure 27. Calling, crossing and cobordism.

**In Sections 15 and 16 we detail how the Fibonacci model for anyonic
**

quantum computing [68, 81] can be constructed by using a version of the two-

stranded bracket polynomial and a generalization of Penrose spin networks.

This is a fragment of the Temperley-Lieb recoupling theory [41].

5.1. The square root of minus one is an eigenform and a clock. So far we

have seen that the mark can represent the fusion rules for a Majorana fermion

since it can interact with itself to produce either itself or nothing. But we have

not yet seen the anti-commuting fermion algebra emerge from this context of

making a distinction. Remarkably, this algebra does emerge when one looks at

the mark recursively.

Consider the transformation

F (X ) = X .

If we iterate it and take the limit we ﬁnd

G = F (F (F (F (. . . )))) = ...

**an inﬁnite nest of marks satisfying the equation
**

G= G .

With G = F (G), I say that G is an eigenform for the transformation F . See

[47] for more about this point of view. See Figure 28 for an illustration of

this nesting with boxes and an arrow that points inside the reentering mark

to indicate its appearance inside itself. If one thinks of the mark itself as a

Boolean logical value, then extending the language to include the reentering

mark G goes beyond the boolean. We will not detail here how this extension

can be related to non-standard logics, but refer the reader to [40]. Taken at face

KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 253

**value the reentering mark cannot be just marked or just unmarked, for by its
**

very deﬁnition, if it is marked then it is unmarked and if it is unmarked then it is

marked. In this sense the reentering mark has the form of a self-contradicting

paradox. There is no paradox since we do not have to permanently assign

it to either value. The simplest interpretation of the reentering mark is that

it is temporal and that it represents an oscillation between markedness and

unmarkedness. In numerical terms it is a discrete dynamical system oscillating

between +1 (marked) and −1 (not marked).

... =

Figure 28.

With the reentering mark in mind consider now the transformation on real

numbers given by

T (x) = −1/x.

This has the ﬁxed points i and −i, the complex numbers whose squares are

negative unity. But lets take a point of view more directly associated with the

analogy of the recursive mark. Begin by starting with a simple periodic process

that is associated directly with the classical attempt to solve for i as a solution

to a quadratic equation. We take the point of view that solving x 2 = ax + b is

the same (when x = 0) as solving

x = a + b/x,

and hence is a matter of ﬁnding a ﬁxed point. In the case of i we have

x 2 = −1

and so desire a ﬁxed point

x = −1/x.

There are no real numbers that are ﬁxed points for this operator and so we

consider the oscillatory process generated by

T (x) = −1/x.

The ﬁxed point would satisfy

i = −1/i

and multiplying, we get that

ii = −1.

254 LOUIS H. KAUFFMAN

**On the other hand the iteration of T yields
**

1, T (1) = −1, T (T (1)) = +1, T (T (T (1))) = −1, +1, −1, +1, −1, . . . .

The square root of minus one is a perfect example of an eigenform that occurs

in a new and wider domain than the original context in which its recursive

process arose. The process has no ﬁxed point in the original domain.

Looking at the oscillation between +1 and −1, we see that there are naturally

two phase-shifted viewpoints. We denote these two views of the oscillation by

[+1, −1] and[−1, +1]. These viewpoints correspond to whether one regards

the oscillation at time zero as starting with +1 or with −1. See Figure 29. We

shall let the word iterant stand for an undisclosed alternation or ambiguity

between +1 and −1. There are two iterant views: [+1, −1] and [−1, +1] for

the basic process we are examining. Given an iterant [a, b], we can think of

[b, a] as the same process with a shift of one time step. The two iterant views,

[+1, −1] and [−1, +1], will become the square roots of negative unity, i and

−i.

... +1, -1, +1, -1, +1, -1, +1, -1, ...

[-1,+1] [+1,-1]

Figure 29.

**We introduce a temporal shift operator
such that
**

[a, b]
=
[b, a]

and

= 1

for any iterant [a, b], so that concatenated observations can include a time step

of one-half period of the process

. . . abababab . . . .

We combine iterant views term-by-term as in

[a, b][c, d ] = [ac, bd ].

We now deﬁne i by the equation

i = [1, −1]
.

This makes i both a value and an operator that takes into account a step in

time.

KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 255 We calculate ii = [1. c = ie and we can construct the quaternions I = ba = e. −1][−1. participating in the algebraic structure of the complex numbers. 1] = 1 and e = [1. 2 2 2 and can continue as we did in Section 4. C = √ (1 + K ). −1] = [1. B = √ (1 + J ). −1] = −1. In fact the corresponding algebra structure of linear combinations [a. Now we can make contact with the algebra of the Majorana fermions. but it is striking how this Marjorana fermion algebra emerges from an analysis of the recursive nature of the reentering mark. b]+[c. Recall from Section 4 that a pair of Majorana fermions can be assembled to form a single standard fermion. 1] = − e. In this view i represents a discrete oscillating temporal process and it is an eigenform for T (x) = −1/x. We treat this generalization elsewhere [46. We have that (e )2 = −1 and so regard this as a natural construction of the square root of minus one in terms of the phase synchronization of the clock that is the iteration of the reentering mark. K = ac = i . Let e = [1. J = cb = ie. d ] is isomorphic with 2×2 matrix algebra and iterants can be used to construct n × n matrix algebra. b = . 50]. letting it commute with the other operators. −1]. while the fusion algebra for the Majorana fermion emerges from the distinctive properties of the mark itself. With the quaternions in place. We see how the seeds of the fermion algebra live in this extended logical context. Thus we have e 2 = 1. Thus we have constructed a square root of minus one by using an iterant viewpoint. We can regard e and as a fundamental pair of Majorana fermions. This is a formal correspondence. −1] = [−1. Then we have the (ie )2 = +1 and so we have a triple of Majorana fermions: a = e. Note how the development of the algebra works at this point. In our case we have the two Marjorana fermions e and . and e = − e. There is one more comment that is appropriate for this section. 2 = 1. Once we have the square root of minus one it is natural to introduce another one and call this one i. 1] = [−1. −1] [1. Then we have e 2 = [1. we have the braiding operators 1 1 1 A = √ (1 + I ).

but in fact the algebra underlying this equation has the same properties as the creation and annihilation algebra for fermions. Let O = i∂/∂t + iα∂/∂x − m so that the Dirac equation takes the form O(x. We will get to this operator by ﬁrst taking the case where p is a scalar (we use one dimension of space and one dimension of time. This is our familiar Cliﬀord algebra pattern and we can use the iterant algebra generated by e and if we wish. Dirac constructed his equation by looking for an algebraic square root of p 2 + m 2 so that he could have a linear operator for E that would take the same role as the Hamiltonian in the Schrodinger equation. we have the Dirac equation i∂/∂t = −iα∂/∂x + m. in that the fermions arise from solving the Dirac equation. we have constructed fermion algebra and quaternion algebra. √ √ = (e + i )/ 2 and † = (e − i )/ 2. Now note that Oe i(px−Et) = (E − αp + m)e i(px−Et) . Relativity and the Dirac equation. associative algebra.256 LOUIS H. This may sound circular. then energy E. momentum p and mass m are related by the (Einstein) equation E 2 = p2 + m2 . We can now go further and construct the Dirac equation. because the quantum operator for momentum is −i∂/∂x and the operator for energy is i∂/∂t.2. Then. so it is by way of this algebra that we will come to the Dirac equation. If the speed of light is equal to 1 (by convention). Hence we will satisﬁy E 2 = p 2 + m 2 if α 2 = 2 = 1 and α + α = 0. there is further analysis needed of the relation of the physics and the logic. i. This will be taken up in a separate paper. 5. retrieved from the way that we decomposed the process into space and time. Let E = αp + m where α and are elements of a a possibly non-commutative. t) = 0. Then E 2 = α 2 p 2 + 2 m 2 + pm(α + α). Since all this is initially built in relation to extending the Boolean logic of the mark to a non-boolean recursive context. Starting with the algebra structure of e and and adding a commuting square root of −1. Since e represents a spatial view of the basic discrete oscillation and is the time-shift operator for this oscillation it is of interest to note that the standard fermion built by these two can be regarded as a quantum of spacetime. KAUFFMAN and the corresponding standard fermion annihilation and creation operators are then given by the formulas below.

Since the equation can have real solutions. this is the Majorana representation for the Dirac equation (compare [65]). In fact for the speciﬁc above we will now have D(Ue i(px−Et) ) = U 2 e i(px−Et) = 0. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 257 and that if U = (E − αp + m)α = αE + p + αm. We can rewrite this as ∂/∂t = α∂/∂x + im. then U 2 = −E 2 + p2 + m 2 = 0. We see that if i is real. where we represent −1 0 e= 0 1 and 0 1 = 1 0 as matrix versions of the iterants associated with the reentering mark. we can take the equation ∂/∂t = e∂/∂x + e m. obtaining the operator D = Oα = iα∂/∂t + i∂/∂x + αm. As the reader can check. then we can write a fully real version of the Dirac equation. and the equivalent Dirac equation D = 0. For example. i∂/∂t = −iα∂/∂x + m. We will explore this relationship with the Rowlands formulation in a separate paper. . the corresponding Rowland nilpotent U is given by the formula U = −i E + ie p + em. For the case of one dimension of space and one dimension of time. this calculation suggests that we should multiply the operator O by α on the right. In fact. This way of reconﬁguring the Dirac equation in relation to nilpotent algebra elements U is due to Peter Rowlands [86]. Return now to the original version of the Dirac equation. from which it follows that = Ue i(px−Et) is a (plane wave) solution to the Dirac equation. these are their own complex conjugates and correspond to particles that are their own anti-particles.

One of the details required for any speciﬁc quantum problem is the nature of the unitary evolution. the superposition of a state with itself is again itself. B ∈ H . States are “really” in the projective space associated with H . Note that since Uv |Uw = v |U † U |w = v |w = when U is unitary. starting with the mark as a logical and recursive particle. Once again. The dual element to A corresponds to the conjugate transpose A† of the vector A.258 LOUIS H. it is most convenient to regard the states |v and |w as vectors in a vector space. This information is used to choose an appropriate Hamiltonian through which the unitary operator is constructed via a correspondence principle that replaces classical variables with appropriate quantum operators.) One needs to know certain aspects of classical physics to solve any speciﬁc quantum problem. A key concept in the quantum information viewpoint is the notion of the superposition of states. the ket |B is identiﬁed with the vector B ∈ H . The quantum information context encapsulates a concise model of quantum theory: The initial state of a quantum process is a vector |v in a complex vector space H . In the present paper we have given a picture of how. (In the path integral approach one needs a Langrangian to construct the action on which the path integral is based. Quantum mechanics and quantum computation. it follows that probability is preserved in the course of a quantum process. one can tell a story that reaches the Dirac equation and its algebra. KAUFFMAN For eﬀective application to the topics in this paper. Measurement returns basis elements of H with probability | |v|2 /v |v where v |w = v † w with v † the conjugate transpose of v. He also separated the parts of the bracket into the bra A | and the ket |B. If a quantum system has two distinct states |v and |w. This is speciﬁed by knowing appropriate information about the classical physics that supports the phenomena. There is only one superposition of a single state |v with itself. one needs to use two dimensions of space and one dimension of time. On the other hand. This will be explored in another paper. then it has inﬁnitely many states of the form a|v + b|w where a and b are complex numbers taken up to a common multiple. §6. Thus A |B = A | |B In this interpretation. We than take it as part of the procedure of dealing with states to normalize them to unit length. A physical process occurs in steps |v −→ U |v = |Uv where U is a unitary linear transformation. Dirac [15] introduced the “bra-(c)-ket” notation A |B = A† B for the inner product of complex vectors A. and . We shall quickly indi- cate the basic principles of quantum mechanics. while the bra < A | is regarded as the element dual to A in the dual space H ∗ .

One runs the computer by repeatedly initializing it. |Cn } is an orthonormal basis for H . If the intermediate states |Ci are a complete set of orthonormal alternatives then we can assume that Ci |Ci = 1 for each i and that Σi |Ci Ci | = 1. . In conventional notation. If {|C1 . . not a scalar. and then measuring the result of applying the unitary transformation U to the initial state. This becomes the path integral expression for the amplitude B|A. and Pi = |Ci Ci |. together with an initial state and a choice of measurement basis. Let {W0 . a composition U of unitary transformations. This identity now corresponds to the fact that 1 is the sum of the probabilities of an arbitrary state being projected into one of these intermediate states. . Wn } . projecting to the subspace of H that is spanned by the vector |A. and we have the following formula for the square of P = |AB | : P 2 = |AB ||AB | = A(B † A)B † = (B † A)AB † = B |AP. What is a quantum computer? A quantum computer is. The probability for this event is equal to |B |A|2 .1. |C2 . . The reader should consult [79] for more speciﬁc examples of quantum algorithms. abstractly. . Dirac can write the “ket-bra” |AB | = AB † . . The standard example is a ket-bra P = |A A| where A |A = 1 so that P 2 = P. . 6. Then P is a projection matrix. The results of these measurements are then analyzed for the desired information that the computer was set to determine. The key to using the computer is the design of the initial state and the design of the composition of unitary transformations. This can be reﬁned if we have more knowledge. Let H be a given ﬁnite dimensional vector space over the complex numbers C . W1 . for any vector |B we have P|B = |AA | |B = |AA |B = A |B|A. then for any vector |A we have |A = C1 |A|C1 + · · · + Cn |A|Cn . Hence B |A = B |C1 C1 |A + · · · + B |Cn Cn |A One wants the probability of starting in state |A and ending in state |B. Having separated the bra and the ket. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 259 the inner product is expressed in conventional language by the matrix product A† B (which is a scalar since B is a column vector). In fact. the ket-bra is a matrix. If there are intermediate states between the intermediate states this formu- lation can be continued until one is summing over all possible paths from A to B.

and more speciﬁcally by using solutions to the Yang-Baxter equation [7]. and later work out factorizations into qubit transformations. and the probabilistic nature of the result that characterizes quantum computation. a vector in the space is called a qubit. A quantum computation consists in the application of a unitary transfor- mation U to an initial qunit = a0 |0 + · · · + an |n with ||2 = 1. The result of observation is to put the system into one of the basis states. ﬁrst discovered in relation to 1 + 1 dimensional quantum ﬁeld theory. A measurement of U returns the ket |i with probability |i|U|2 . One can start with a given space. A class of invariants of knots and links called quantum invariants can be constructed by using representations of the Artin braid group. When the dimension of the space H is two (n = 1). one obtains either the ket |0 or the ket |1. In speaking of an idealized quantum computer. we do not specify the nature of measurement process beyond these probability postulates. A qubit represents one quantum of binary information. Given a vector v in H let |v|2 := v|v. then the probability that it will return the state |j is |j|U |i|2 . Such computation could be carried out by an idealized quantum mechanical system. then the ket |0 is observed with probability |α|2 . In particular. we will call the vectors in H qunits. In the case of general dimension n of the space H . An measurement of v returns one of the coordinates |i of v with probability |i|v|2 . It is hoped that such systems can be physically realized. On measurement. §7. and 2 dimensional statistical mechanics.260 LOUIS H. Note that i|v is the i-th coordinate of v. we have i|j = ij where ij denotes the Kronecker delta (equal to one when its indices are equal to one another. . and the ket |1 is observed with probability ||2 . KAUFFMAN be an orthonormal basis for H so that with |i := |Wi denoting Wi and i| denoting the conjugate transpose of |i. This constitutes the binary distinction that is inherent in a qubit. It is the necessity for writing a given computation in terms of unitary transformations. plus an measurement of U. It is quite common to use spaces H that are tensor products of two-dimensional spaces (so that all computations are expressed in terms of qubits) but this is not necessary in principle. if we start the computer in the state |i. and equal to zero otherwise). Braiding operators and universal quantum gates. This model of measurement is a simple instance of the situation with a quantum mechanical system that is in a mixed state until it is observed. If the qubit is | = α|0 + |1. Note however that the information obtained is probabilistic.

KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 261

**Braiding operators feature in constructing representations of the Artin braid
**

group, and in the construction of invariants of knots and links.

A key concept in the construction of quantum link invariants is the asso-

ciation of a Yang-Baxter operator R to each elementary crossing in a link

diagram. The operator R is a linear mapping

R : V ⊗ V −→ V ⊗ V

deﬁned on the 2-fold tensor product of a vector space V, generalizing the

permutation of the factors (i.e., generalizing a swap gate when V represents

one qubit). Such transformations are not necessarily unitary in topological

applications. It is useful to understand when they can be replaced by unitary

transformations for the purpose of quantum computing. Such unitary R-

matrices can be used to make unitary representations of the Artin braid

group.

A solution to the Yang-Baxter equation, as described in the last paragraph

is a matrix R, regarded as a mapping of a two-fold tensor product of a vector

space V ⊗ V to itself that satisﬁes the equation

(R ⊗ I )(I ⊗ R)(R ⊗ I ) = (I ⊗ R)(R ⊗ I )(I ⊗ R).

From the point of view of topology, the matrix R is regarded as representing

an elementary bit of braiding represented by one string crossing over another.

In Figure 30 we have illustrated the braiding identity that corresponds to the

Yang-Baxter equation. Each braiding picture with its three input lines (below)

and output lines (above) corresponds to a mapping of the three fold tensor

product of the vector space V to itself, as required by the algebraic equation

quoted above. The pattern of placement of the crossings in the diagram

corresponds to the factors R ⊗ I and I ⊗ R. This crucial topological move

has an algebraic expression in terms of such a matrix R. Our approach in this

section to relate topology, quantum computing, and quantum entanglement is

through the use of the Yang-Baxter equation. In order to accomplish this aim,

we need to study solutions of the Yang-Baxter equation that are unitary. Then

the R matrix can be seen either as a braiding matrix or as a quantum gate in a

quantum computer.

R I I R

R I I R

I R = R I

R I I R

Figure 30. The Yang-Baxter equation.

**The problem of ﬁnding solutions to the Yang-Baxter equation that are
**

unitary turns out to be surprisingly diﬃcult. Dye [16] has classiﬁed all such

262 LOUIS H. KAUFFMAN

**matrices of size 4 × 4. A rough summary of her classiﬁcation is that all
**

4 × 4 unitary solutions to the Yang-Baxter equation are similar to one of the

following types of matrix:

⎛ √ √ ⎞

1/ 2 0√ 0√ 1/ 2

⎜ 0 1/√2 −1/√ 2 0 ⎟

R=⎜ ⎝

⎟

0√ 1/ 2 1/ 2 0√ ⎠

−1/ 2 0 0 1/ 2

⎛ ⎞

a 0 0 0

⎜ 0 0 b 0 ⎟

R =⎜

⎝ 0 c 0 0 ⎠

⎟

0 0 0 d

⎛ ⎞

0 0 0 a

⎜ 0 b 0 0 ⎟

R = ⎜

⎝ 0 0 c 0 ⎠

⎟

d 0 0 0

**where a, b, c, d are unit complex numbers.
**

For the purpose of quantum computing, one should regard each matrix as

acting on the stamdard basis {|00, |01, |10, |11} of H = V ⊗ V, where V is

a two-dimensional complex vector space. Then, for example we have

√ √

R|00 = (1/ 2)|00 − (1/ 2)|11,

√ √

R|01 = (1/ 2)|01 + (1/ 2)|10,

√ √

R|10 = −(1/ 2)|01 + (1/ 2)|10,

√ √

R|11 = (1/ 2)|00 + (1/ 2)|11.

**The reader should note that R is the familiar change-of-basis matrix from the
**

standard basis to the Bell basis of entangled states.

In the case of R , we have

R |00 = a|00, R |01 = c|10,

R |10 = b|01, R |11 = d |11.

**Note that R can be regarded as a diagonal phase gate P, composed with a
**

swap gate S.

⎛ ⎞

a 0 0 0

⎜ 0 b 0 0 ⎟

P=⎜ ⎝ 0 0 c 0 ⎠

⎟

0 0 0 d

**KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 263
**

⎛ ⎞

1 0 0 0

⎜ 0 0 1 0 ⎟

S=⎜

⎝ 0

⎟

1 0 0 ⎠

0 0 0 1

Compositions of solutions of the (Braiding) Yang-Baxter equation with the

swap gate S are called solutions to the algebraic Yang-Baxter equation. Thus

the diagonal matrix P is a solution to the algebraic Yang-Baxter equation.

Remark 1. Another avenue related to unitary solutions to the Yang-Baxter

equation as quantum gates comes from using extra physical parameters in this

equation (the rapidity parameter) that are related to statistical physics. In [99]

we discovered that solutions to the Yang-Baxter equation with the rapidity

parameter allow many new unitary solutions. The signiﬁcance of these gates

for quatnum computing is still under investigation.

7.1. Universal gates. A two-qubit gate G is a unitary linear mapping G :

V ⊗ V −→ V where V is a two complex dimensional vector space. We say

that the gate G is universal for quantum computation (or just universal) if G

together with local unitary transformations (unitary transformations from V

to V ) generates all unitary transformations of the complex vector space of

dimension 2n to itself. It is well-known [79] that CNOT is a universal gate.

(On the standard basis, CNOT is the identity when the ﬁrst qubit is |0, and it

ﬂips the second qbit, leaving the ﬁrst alone, when the ﬁrst qubit is |1.)

A gate G, as above, is said to be entangling if there is a vector

|α = |α ⊗ | ∈ V ⊗ V

such that G|α is not decomposable as a tensor product of two qubits. Under

these circumstances, one says that G|α is entangled.

In [11], the Brylinskis give a general criterion of G to be universal. They prove

that a two-qubit gate G is universal if and only if it is entangling.

Remark 2. A two-qubit pure state

|φ = a|00 + b|01 + c|10 + d |11

is entangled exactly when (ad − bc) = 0. It is easy to use this fact to check

when a speciﬁc matrix is, or is not, entangling.

Remark 3. There are many gates other than CNOT that can be used as

universal gates in the presence of local unitary transformations. Some of these

are themselves topological (unitary solutions to the Yang-Baxter equation,

see [57]) and themselves generate representations of the Artin braid group.

Replacing CNOT by a solution to the Yang-Baxter equation does not place

the local unitary transformations as part of the corresponding representation

of the braid group. Thus such substitutions give only a partial solution to

creating topological quantum computation. In this paper we are concerned

264 LOUIS H. KAUFFMAN

**with braid group representations that include all aspects of the unitary group.
**

Accordingly, in the next section we shall ﬁrst examine how the braid group on

three strands can be represented as local unitary transformations.

Theorem 1. Let D denote the phase gate shown below. D is a solution to the

algebraic Yang-Baxter equation (see the earlier discussion in this section). Then

D is a universal gate.

⎛ ⎞

1 0 0 0

⎜ 0 1 0 0 ⎟

D=⎜ ⎝ 0 0 1 0 ⎠

⎟

0 0 0 −1

Proof. It follows at once from the Brylinski Theorem that D is universal.

For a more speciﬁc proof, note that CNOT = QDQ −1 , where Q = H ⊗ I , H

is the 2 × 2 Hadamard matrix. The conclusion then follows at once from this

identity and the discussion above. We illustrate the matrices involved in this

proof below:

√ 1 1

H = (1/ 2)

1 −1

⎛ ⎞

1 1 0 0

√ ⎜ 1 −1 0 0 ⎟

Q = (1/ 2) ⎜⎝ 0 0 1 1 ⎠

⎟

0 0 1 −1

⎛ ⎞

1 0 0 0

⎜ 0 1 0 0 ⎟

D=⎜ ⎝ 0 0 1 0 ⎠

⎟

0 0 0 −1

⎛ ⎞

1 0 0 0

⎜ 0 1 0 0 ⎟

QDQ −1 = QDQ = ⎜ ⎟

⎝ 0 0 0 1 ⎠ = CNOT

0 0 1 0

**Remark 4. We thank Martin Roetteles [84] for pointing out the speciﬁc
**

factorization of CNOT used in this proof.

Theorem 2. The matrix solutions R and R to the Yang-Baxter equation,

described above, are universal gates exactly when ad − bc = 0 for their internal

parameters a, b, c, d . In particular, let R0 denote the solution R (above) to the

Yang-Baxter equation with a = b = c = 1, d = −1.

⎛ ⎞

a 0 0 0

⎜ 0 0 b 0 ⎟

R = ⎜ ⎝ 0 c 0 0 ⎠

⎟

0 0 0 d

**KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 265
**

⎛ ⎞

1 0 0 0

⎜ 0 0 1 0 ⎟

R0 = ⎜

⎝ 0

⎟

1 0 0 ⎠

0 0 0 −1

Then R0 is a universal gate.

Proof. The ﬁrst part follows at once from the Brylinski Theorem. In fact,

letting H be the Hadamard matrix as before, and

√ √ √ √

1/√ 2 i/ √2 1/√ 2 1/ √2

= , =

i/ 2 1/ 2 i/ 2 −i/ 2

(1 − i)/2 (1 + i)/2

= .

(1 − i)/2 (−1 − i)/2

Then

CNOT = ( ⊗

)(R0 (I ⊗

)R0 )(H ⊗ H ).

This gives an explicit expression for CNOT in terms of R0 and local unitary

transformations (for which we thank Ben Reichardt).

Remark 5. Let SWAP denote the Yang-Baxter Solution R with a = b =

c = d = 1.

⎛ ⎞

1 0 0 0

⎜ 0 0 1 0 ⎟

SWAP = ⎜ ⎝ 0 1 0 0 ⎠

⎟

0 0 0 1

SWAP is the standard swap gate. Note that SWAP is not a universal gate. This

also follows from the Brylinski Theorem, since SWAP is not entangling. Note

also that R0 is the composition of the phase gate D with this swap gate.

Theorem 3. Let

⎛√ √ ⎞

1/ 2 0√ 0√ 1/ 2

⎜ 0 1/√2 −1/√ 2 0 ⎟

R=⎜

⎝

⎟

⎠

0√ 1/ 2 1/ 2 0√

−1/ 2 0 0 1/ 2

be the unitary solution to the Yang-Baxter equation discussed above. Then R is a

universal gate. The proof below gives a speciﬁc expression for CNOT in terms

of R.

Proof. This result follows at once from the Brylinksi Theorem, since R is

highly entangling. For a direct computational proof, it suﬃces to show that

CNOT can be generated from R and local unitary transformations. Let

√ √

1/√2 1/ √2

α=

1/ 2 −1/ 2

. .2. Thus we assume that ck2 = 1 and ci cj = −cj ci for each k = 1 . Here we generalize this construction and show how the Marjorana fermions give rise to univer- sal topological gates. cn } via the linear transformations Tk : Vn −→ Vn deﬁned by Tk (v) = sk vsk−1 . . See [57] for more information about these calculations. Then the braiding operators Tk each satisfy the Yang-Baxter equation and so we have universal gates (in the presence of single . Furthermore. Then by the same algebra as we explored in Section 4 it is easy to verify that sk+1 sk sk+1 = sk sk+1 sk and that si sj = sj si whenever |i − j| > 1. Tk (ck+1 ) = −ck and that Tk is the identity otherwise. Majorana fermions generate universal braiding gates. cn denote n Majorana fermion creation operators. . . Note that sk−1 = √1 (1 2 − ck+1 ck ). n − 1. Let c1 . Remark 6. 7. . . . KAUFFMAN √ √ −1/√ 2 1/√ 2 = i/ 2 i/ 2 √ √ 1/√2 i/ √2 = 1/ 2 −i/ 2 −1 0 = 0 −i Let M = α ⊗ and N = ⊗ . For universality. It is then easy to verify that Tk (ck ) = ck+1 . n and whenever i = j. Then deﬁne operators 1 sk = √ (1 + ck+1 ck ) 2 for k = 1 . . Thus the si give a representation of the n-strand braid group Bn . it is easy to see that a speciﬁc representation is given on the complex vector space Vn with basis {c1 . . c2 .266 LOUIS H. take n = 4 and regard each Tk as operating on V ⊗ V where V is a single qubit space. This completes the proof. c2 . Then it is straightforward to verify that CNOT = MRN. Recall that in Sec- tion 4 we showed how to construct braid group representations by using Majorana fermions in the special case of three particles.

Such states turn out to be related to subtle nonlocality in quantum physics. if Alice has L and Bob has R and Alice performs a local unitary transformation on “her” tensor factor. One can transport L and R individually and we shall write S =L∗R to denote that they are the “parts” (but not tensor factors) of S. These “states” L and R together comprise the EPR state S. Thought experiments of the sort we are about to describe were ﬁrst devised by Einstein. A remark about EPR. √ R = (|0{|1} + |1{|0})/ 2. For example. is said to be entangled if it cannot be written as a tensor product of vectors from non-trivial factors of H ⊗n . In the right state R. transporting L from one place to another. we think of two “parts” of this state that are separated in space. which also uses Majorana fermions. On the other hand. then it is possible that quantum computers based on these topological unitary representations will be constructed. entanglement and Bell’s inequality. but they are accessible individually just as are the two photons in the usual thought experiement. an observer can only observe the right hand factor. referred henceforth as EPR. an observer can only observe the left hand factor. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 267 qubit unitary operators) from Majorana fermions. §8. A state | ∈ H ⊗n . and so it makes it a bit more apparent what EPR were concerned about. and a diﬀerent subtler representation of the braid groups that is also promising for topological quantum computing. where H is the qubit space. . This is also a “spooky action at a distance” whose consequence does not appear until a measurement is made. as in the original experiment where the photons separate. After all. lots of things that we can do to L or R do not aﬀect S. In the later sections of this paper we will describe the Fibonacci model. In the left state L. Consider the entangled state √ S = (|0|1 + |1|0)/ 2. The curious thing about this formalism is that it includes a little bit of macroscopic physics implicitly. Podolosky and Rosen. If experimental work shows that Majorana fermions can be detected and controlled. We want a notation for these parts and suggest the following: √ L = ({|0}|1 + {|1}|0)/ 2. In an EPR thought experiment. It helps to place this algebraic structure in the context of a gedanken experiment to see where the physics comes in. this applies to both L and R since the transformation is actually being applied to the state S.

we look at the structure of the Bell inequalities using the Clauser. To this end. 1 0 1 −1 −1 √ S= / 2. b. S and T equal to ±1). Classical probability calculation with random variables of value ±1 gives the value of QS + RS + RT − QT = ±2 (with each of Q. d are real numbers. To simplify the results of this calculation we shall here assume that the coeﬃcients a. 1 0 Q= . The classic case is that of the Bell state √ φ = (|01 − |10)/ 2. Shimony. . repsectively. Holt formalism (CHSH) as explained in the book by Nielsen and Chuang [79]. −1 1 2 1 −1 √ T = / 2. That quantum expectation is not classical is embodied in the fact that Δ can be greater than 2. in the sense of tensor indecomposability. has to do with the structure of the EPR thought experiment. c. For this we use the following observables with eigenvalues ±1. R. 0 −1 1 0 1 R= . KAUFFMAN To go a bit deeper it is worthwhile seeing what entanglement. of a quantum state of the form φ = a|00 + b|01 + c|10 + d |11. Hence the classical expectation satisﬁes the Bell inequality E(QS) + E(RS) + E(RT ) − E(QT ) ≤ 2. −1 −1 2 The subscripts 1 and 2 on these matrices indicate that they are to operate on the ﬁrst and second tensor factors. ﬁnding that √ Δ = (2 − 4(a + d )2 + 4(ad − bc))/ 2. In general we see that the following inequality is needed in order to violate the Bell inequality √ (2 − 4(a + d )2 + 4(ad − bc))/ 2 > 2. This is equivalent to √ ( 2 − 1)/2 < (ad − bc) − (a + d )2 . We calculate the quantity Δ = φ|QS|φ + φ|RS|φ + φ|RT |φ − φ|QT |φ. Horne.268 LOUIS H. Here √ Δ = 6/ 2 > 2.

We take the stance that topological properties of systems are properties that remain invariant under . while projecting to |0 yields an entangled state. On the other hand. One’s intuition suggests that it is this sort of entanglement that should have a topological context. §9. Consider for example the state | = |001 + |010 + |100. and entanglement in the sense of Bell inequality violation for a given choice of Bell operators are not equivalent concepts. Measurement in any factor of the GHZ yields an unentangled state. so that projecting to |1 in the ﬁrst coordinate yields an unentangled state. We point out that there are states whose entanglement after an measurement is a matter of probability (via quantum amplitudes). this shows that an unentangled state cannot violate the Bell inequality. Measurement in any coordinate yields probabilistically an entangled or an unentangled state. √ sense the Borromean rings are analogous to the GHZ state |GHZ = In this (1/ 2)(|000 + |111). This formula also shows that it is possible for a state to be entangled and yet not violate the Bell inequality. Link diagrams can be used as graphical devices and holders of information. The Aravind hypothesis. A key example is the Borromean rings. The relationship of topology and physics needs to be examined carefully. but any two of them are unentangled. We see from this calculation that entanglement in the sense of tensor indecomposability. For example. The Borromean rings are entangled. But one needs to go deeper in this consideration. Deleting any component of the Boromean rings yields a remaining pair of unlinked rings. For example | = |0(|01 + |10) + |1|00. Benjamin Schumacher has pointed out [88] that any entangled two-qubit state will violate Bell inequalities for an appropriate choice of operators. but φ is an entangled state. if φ = (|00 − |01 + |10 + |11)/2. New ways to use link diagrams must be invented to map the properties of such states. One direction is to consider appropriate notions of quantum knots so that one can formlate superpositions of topological types as in [58]. Aravind points out that this property is basis dependent. The Bell inequality violation is an indication of quantum mechanical entanglement. then Δ(φ) satisﬁes Bell’s inequality. See Figure 18. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 269 Since we know that φ is entangled exactly when ad − bc is non-zero. This deepens the context for our question of the relationship between topological entanglement and quantum entanglement. Measurement of a link would be modeled by deleting one component of the link. In this vein Aravind [5] proposed that the entanglement of a link should correspond to the entanglement of a state.

. −w¯ z¯ where z and w are complex numbers. We begin with the structure of SU (2). Thus if z = a + bi and w = c + di where a. which represents all possible ways of moving the rope around (without cutting the rope. and z¯ denotes the complex conjugate of z. and so U (2) is usually regarded as the group of local unitary transformations in a quantum information setting. 0 1 0 −i −1 0 i 0 . 72. an element |K lying in an appropriate Hilbert space Hn . and M † is the conjugate transpose of M . The knot type of a quantum knot |K is simply the orbit of the quantum knot under the action of the ambient group An . Accordingly. In making quantum physical models. KAUFFMAN certain transformations that are identiﬁed as “topological equivalences”. such as quantum superposition and quantum entanglement. Associated with a quantum knot system is a group of unitary transformations An . Thus it is a fact that local unitary transformations can be “generated by braids” in many ways. and without letting the rope pass through itself. To be in SU (2) it is required that Det(M ) = 1 and that M † = M −1 where Det denotes determinant. and i 2 = −1. we have formulated a model for quantum knots [71. d are real numbers. unlike a classical closed piece of rope. a quantum knot can exhibit non-classical behavior. A quantum knot system represents the “quantum embodiment” of a closed knotted physical piece of rope.e. One regards the groups SU (2) and U (2) as acting on a single qubit. If one is looking for a coherent way to represent all unitary transformations by way of braids. b.270 LOUIS H. i. 63] that meets these requirements. these equivalences should correspond to unitary transformations of an appropriate Hilbert space. SU (2) representations of the Artin braid group.e. the particular spatial conﬁguration of the knot tied in the rope. This leads to new questions connecting quantum computing and knot theory. then U (2) is the place to start. Here we will show that there are many representations of the three-strand braid group that generate a dense subset of U (2). represents the state of such a knotted closed piece of rope. c. A matrix in SU (2) has the form z w M = . The purpose of this section is to determine all the representations of the three strand Artin braid group B3 to the special unitary group SU (2) and concomitantly to the unitary group U (2). A quantum knot (i. It is convenient to write 1 0 i 0 0 1 0 i M =a +b +c +d .. 73. called the ambient group.) Of course. then a + bi c + di M = −c + di a − bi with a 2 + b 2 + c 2 + d 2 = 1. as a state of this system. §10.

and u × v is the vector cross product of u and v. s. v are pure quaternions. 4. 3. 6. A quaternion of the form rI + sJ + tK for real numbers r. Note that if q = a + bI + cJ + dK (as above). We shall use this identiﬁcation. Thus a general quaternion has the form q = a + bu where u is a pure quaternion of unit length and a and b are arbitrary real numbers. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 271 and to abbreviate this decomposition as M = a + bI + cJ + dK where 1 0 i 0 0 1 0 i 1≡ . and all the above properties are consequences of this deﬁnition. A general quaternion has the form q = a + bI + cJ + dK where the value of qq † =a 2 + b 2 + c 2 + d 2 . Note that quaternion multiplication is associative. J. s. Thus the unit quaternions are identiﬁed with SU (2) in this way. J ≡ . A unit quaternion (element of SU (2)) has the addition property that a 2 +b 2 = 1. In fact. 2. K is called the quaternions after William Rowan Hamilton who discovered this algebra prior to the discovery of matrix algebra. is not ﬁxed to unity. I. t) of real numbers R3 . one can take the deﬁnition of quaternion multiplication as (a + bu)(c + dv) = ac + bc(u) + ad (v) + bd (−u · v + u × v). First we recall some facts about the quaternions. and some facts about the quaternions to ﬁnd the SU (2) representations of braiding. The algebra of 1. then uv = −u · v + u × v whre u · v is the dot product of the vectors u and v. then u 2 = −1. I ≡ . The length of q is by deﬁnition qq † . Note that the set of pure unit quaternions forms the two-dimensional sphere S 2 = {(r. . 1. If u. 5. s. t is said to be a pure quaternion. then q † = a −bI −cJ −dK so that qq † = a 2 + b 2 + c 2 + d 2 = 1. If u is a pure unit length quaternion. t)|r 2 + s 2 + t 2 = 1} in R3 . K≡ 0 1 0 −i −1 0 i 0 so that I 2 = J 2 = K 2 = IJK = −1 and IJ = K JK = I KI = J JI = −K KJ = −I IK = −J. We identify the set of pure quaternions with the vector space of triples (r.

then ghg −1 = c + dφg (v) and h −1 gh = a + bφh −1 (u). for P any point in R3 . and φg (v) = φh −1 (u). φg is a rotation about the axis u by the angle . Theorem 4. in the case where there is a minus sign we have g = a + bu and h = a − bv = a + b(−v). Since s1 s2 s1 = s2 s1 s2 is the generating relation for B3 . and φg (v) = φh −1 (u). By the same token. Thus it follows from the braiding relation that a = c. If u = v then g = h and the braid relation is trivially satisﬁed. KAUFFMAN 7. Then (without loss of generality). The mapping φ : SU (2) −→ SO(3) is a two-to-one surjective map from the special unitary group to the rotation group. regarded as a pure quaternion. Suppose that g = a + bu and h = c + dv where u and v are unit pure quaternions so that a 2 + b 2 = 1 and c 2 + d 2 = 1. However. Then φg is an orientation preserving rotation of R3 (hence an element of the rotation group SO(3)). We have proved the ﬁrst sentence of the Theorem in the discussion prior to its statement. . This means that we want a homomorphism : B3 −→ SU (2). Thus we can now prove the following Theorem. Speciﬁcally. b = ±d. b = sin(/2) for a chosen angle . the braid relation ghg = hgh is true if and only if h = a + bv. Proof. and hence we want elements g = (s1 ) and h = (s2 ) in SU (2) representing the braid group generators s1 and s2 . and that φg (v) = ±φh −1 (u). we have φh −1 (u) = h −1 uh = (a 2 − b 2 )u + 2ab(u × −v) + 2(u · (−v))b 2 (−v) = (a 2 − b 2 )u + 2ab(v × u) + 2(v · u)b 2 (v). h = a + bv. In quaternionic form. Let u and v be pure unit quaternions and g = a + bu and h = c + dv have unit length. The speciﬁc formula for φg (P) as shown below: φg (P) = gPg −1 = (a 2 − b 2 )P + 2ab(P × u) + 2(P · u)b 2 u. Let g = a + bu be a unit length quaternion so that u 2 = −1 and a = cos(/2). and analyze its meaning in the unit quaternions. given that g = a + bu and h = a + bv. this result was proved by Hamilton and by Rodrigues in the middle of the nineteeth century. the only requirement on g and h is that ghg = hgh. We want a representation of the three-strand braid group in SU (2). Furthermore. Deﬁne φg : R3 −→ R3 by the equation φg (P) = gPg † . We rewrite this relation as h −1 gh = ghg −1 . Therefore assume that g = a + bu. We have already stated the formula for φg (v) in the discussion about quaternions: φg (v) = gvg −1 = (a 2 − b 2 )v + 2ab(v × u) + 2(v · u)b 2 u. the condition φg (v) = φh −1 (u) is satisﬁed if and only if u · v = a 2b−b2 when u = 2 2 v.272 LOUIS H.

J and K are mutually orthogonal) the three operators 1 A = √ (1 + I ). 2 1 B = √ (1 + J ). Then we can rewrite g and h in matrix 2 2 form as the matrices G and H . Let h = a + b[(c 2 − s 2 )I + 2csK ] where c 2 + s 2 = 1 and c 2 − s 2 = a 2b−b2 . For deﬁniteness. The Fibonacci example. 2b 2 The Majorana fermion example. we write H = FGF † where F is an element of SU (2) as shown below. then this implies that a2 − b2 u·v = . This equation is equivalent to 2(u · v)b 2 (u − v) = (a 2 − b 2 )(u − v). 2 Each pair satisﬁes the braiding relation so that ABA = BAB. Then the theorem tells us that √ we need a − b = 0 2 2 2 2 and since a + b = 1. Note the case of the theorem where g = a + bu. then we have for the braiding generators (since I . Let g = e I = a + bI where a = cos() and b = sin(). 2 1 C = √ (1 + K). ACA = CAC . we conclude that a = 1/ 2 and b likewise. i e 0 G= 0 e −i ic is F = is −ic . h = a + bv. Suppose that u · v = 0. If u = v. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 273 Hence we require that (a 2 − b 2 )v + 2(v · u)b 2 u = (a 2 − b 2 )u + 2(v · u)b 2 (v). We have already met this braiding triplet in our discussion of the construction of braiding operators from Majorana fermions in Section 4. This shows (again) how close Hamilton’s quaternions are to topology and how braiding is fundamental to the structure of fermionic physics. Instead of writing the explicit form of H. BCB = CBC .

h} such that ghg = hgh. Consider representations of B3 into SU (2) produced by the method of this section. e aI e bJ e cI = cos(b)e I (a+c) + sin(b)e I (a−c) J . The simplest example is given by g = e 7 I/10 √ f = I + K h = fgf −1 where 2 + = 1. Proof. c. and such that the axes of A and B are linearly independent in R3 . Lemma 1. that the Fibonacci representation described above is dense in SU (2). invariant under Reidmeister moves II and III. and the powers of B are dense in the rotations about its axis. can be normalized to give an invariant of all three . It follows for example. if u and v are linearly independent unit vectors in R3 . In fact we shall see just such representations constructed later in this paper. c. Hence any element of SU (2) can be written in the form e aI e bJ e cI for appropriate choices of angles a. §11. We now discuss the Jones polynomial. has the possibility for generalization to representations of braid groups (on greater than three strands) to SU (n) or U (n) for n greater than 2. H ] of SU (2) generated by a pair of elements {g. b. then any element of SU (2) can be written in the form e au e bv e cu for appropriate choices of the real numbers a. That is consider the subgroup SU [G. In fact. We shall call this the Fibonacci representation of B3 to SU (2). The bracket polynomial and the Jones polynomial. We wish to understand when such a representation will be dense in SU (2). and indeed the generic representation of B3 into SU (2) will be dense in SU (2). while the other generator H = FGF † is derived from G by conjugation by a unitary matrix. Then by the Lemma the set of elements Aa+c B b Aa−c are dense in SU (2). We shall construct the Jones polynomial by using the bracket state summation model [35]. For this we need more topology. Density. We need the following lemma. KAUFFMAN This representation of braiding where one generator G is a simple matrix of phases. This Lemma can be used to verify the density of a representation. See [59] for the details of this proof. Then g and h satisfy ghg = hgh and generate a representation of the three-strand braid group that is dense in SU (2). b. by ﬁnding two elements A and B in the representation such that the powers of A are dense in the rotations about its axis. by using a version of topological quantum ﬁeld theory.274 LOUIS H. Our next task is to describe representations of the higher braid groups that will extend some of these unitary representations of the three-strand braid group. The bracket polynomial.

We take the convention that the letter chi. < K > satisﬁes the following formulas < > = A < + > +A−1 <)(> < > = A−1 < + > +A <)(>. denotes a crossing where the curved line is crossing over the straight segment. The barred letter denotes the switch of this crossing. A-1 A A A-1 -1 A A < > =A < > + A-1 < > < > = A-1 < > +A < > Figure 31. It is easy to see that Properties 2 and 3 deﬁne the calculation of the bracket on arbitrary link diagrams. assigns to each unoriented link diagram K a Laurent polynomial in the variable A. is the Jones polynomial [29. then < K > = < K >. If K * O denotes the disjoint union of K with an extra unknotted and unlinked component O (also called ‘loop’ or ‘simple closed curve’ or ‘Jordan curve’). with a change of variable. See Figure 31 for a graphic illustration of this relation. then < K * O > = < K >. This normalized invariant. where the small diagrams represent parts of larger diagrams that are identical except at the site indicated in the bracket. where the curved line is undercrossing the straight segment. 3. If K and K are regularly isotopic diagrams. The bracket polynomial . The Jones polynomial was originally discovered by a diﬀerent method than the one given here. . The choices of coeﬃcients (A and A−1 ) and the value . such that 1. 30]. Bracket smoothings. < K > = < K > (A). where = −A2 − A−2 . KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 275 Reidemeister moves. 2. and an indication of the convention for choosing the labels A and A−1 at a given crossing.

Remember that it is necessary to keep track of the diagrams up to regular isotopy (the equivalence relation generated by the second and third Reidemeister moves). as also seen in this ﬁgure. + - + + or + . Note that in these conventions the A-smoothing of is +. In computing the bracket. one ﬁnds the following behaviour under Reide- meister move I: < >= −A3 <> and < >= −A−3 <> where denotes a curl of positive type as indicated in Figure 32. The small arcs on the right hand side of these formulas indicate the removal of the curl from the corresponding diagram. . Our convention of signs is also given in Figure 32. The convention for crossing signs is shown in Figure 32. One useful consequence of these formulas is the following switching formula A < > −A−1 < >= (A2 − A−2 ) < + > . where we chose an orientation for K . View Figure 33. Here is an example. . Note that the type of a curl does not depend on the orientation we choose. or - Figure 32. Properly interpreted. w(K ) is called the writhe of K . while the A- smoothing of is )(. the switching formula above says that you can switch a crossing and smooth it either way and obtain a three diagram relation. The type of a curl is the sign of the crossing when we orient it locally. KAUFFMAN of make the bracket invariant under the Reidemeister moves II and III. The bracket is invariant under regular isotopy and can be normalized to an invariant of ambient isotopy by the deﬁnition fK (A) = (−A3 )−w(K) < K > (A). Crossing signs and curls. and indicates a curl of negative type.276 LOUIS H. This is useful since some computations will simplify quite quickly with the proper choices of switching and smoothing. and where w(K) is the sum of the crossing signs of the oriented link K . Thus Property 1 is a consequence of the other two properties.

and thus there are 2N states of a diagram with N crossings. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 277 K U U' Figure 33. we now describe it as a state summation. denoted < K|S > . Since the trefoil diagram K has writhe w(K ) = 3. Figure 33 shows a trefoil diagram K . an unknot diagram U and another unknot diagram U . One of the neatest applications is to simply compute. a fact that is much harder to prove by classical methods. There are two evaluations related to a state. as we have done. Thus < K >= −A5 − A−3 + A−7 . In order to obtain a closed formula for the bracket. we have A−1 < K > −A < U >= (A−2 − A2 ) < U > and < U >= −A3 and < U >= (−A−3 )2 = A−6 . Deﬁne a state. The label is called a vertex weight of the state. Applying the switching formula. The ﬁrst one is the product of the vertex weights. In a state we label each smoothing with A or A−1 according to the left-right convention discussed in Property 3 (see Figure 31). denoted ||S||. fK (A) for the trefoil knot K and determine that fK (A) is not equal to fK (A−1 ) = f−K (A). There are two choices for smoothing a given crossing. The second evaluation is the number of loops in the state S. The bracket model for the Jones polynomial is quite useful both theoretically and in terms of practical computations. Hence A−1 < K >= −A4 + A−8 − A−4 . Trefoil and two relatives. of K to be a choice of smoothing for each crossing of K . . The state summation. Let K be any unoriented link diagram. Thus A−1 < K > −A(−A3 ) = (A−2 − A2 )A−6 . This shows that the trefoil is not ambient isotopic to its mirror image. S. we have the normalized polynomial fK (A) = (−A3 )−3 < K >= −A−9 (−A5 − A−3 + A−7 ) = A−4 + A−12 − A−16 . This is the bracket polynomial of the trefoil diagram K .

where they express the summation over all states of the physical system of probability weighting functions for the individual states. of those states with an A-type smoothing and those with an A−1 -type smoothing at that crossing. Hence this state summation produces the bracket polynomial as we have described it at the beginning of the section. in this survey paper. < O > = 1. or performing other mathematical operations on it. Such physical partition functions contain large amounts of information about the corresponding physical system. Remark 7. 1 Remark 8. The second and the third equation are clear from the formula deﬁning the state summation. As a result. < K * O > = < K >. The reader is referred . Some of this information is directly present in the properties of the function. Partition functions are ubiquitous in statistical mechanics. useful for producing systematic solutions to the Yang-Baxter equation. The subject is intertwined with the algebraic structure of Hopf algebras and quantum groups. We have chosen. In fact Hopf algebras are deeply connected with the problem of constructing invariants of three-dimensional manifolds in relation to invariants of knots. by the formula < K >= < K|S > ||S||−1 . The ﬁrst equation expresses the fact that the entire set of states of a given diagram is the union. By a change of variables one obtains the original Jones polyno- mial. VK (t). but rather to proceed to Vassiliev invariants and the relationships with Witten’s functional integral. S It follows from this deﬁnition that < K > satisﬁes the equations < > = A < + > +A−1 <)(>. with respect to a given crossing. KAUFFMAN Deﬁne the state summation. The bracket polynomial provides a connection between knot theory and physics. such as the location of critical points and phase transition. there are many ways to deﬁne partition functions of knot diagrams that give rise to invariants of knots and links. < K >. to not discuss the details of these approaches. There is much more in this connection with statistical mechanics in that the local weights in a partition function are often expressed in terms of solutions to a matrix equation called the Yang-Baxter equation. in that the state summation expression for it exhibits it as a generalized partition function deﬁned on the knot diagram.278 LOUIS H. that turns out to ﬁt perfectly invariance under the third Reidemeister move. for oriented knots and links from the normalized bracket: VK (t) = fK (t − 4 ). Some of the information can be obtained by diﬀerentiating the partition function.

29. 4. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 279 to [35. 39. the key point is that Lie algebras can be used to construct invariants of knots and links. 44. This quantum computer will probabilistically (via quantum amplitudes) compute the values of the states in the state sum for ZK . and |CAP is the composition of caps. In order to view ZK as a quantum computation. of particles arising from the vacuum. 83. and the vacuum- vacuum diagram for the knot is interpreted as a quantum computer. the presently successful quantum algorithms for computing knot invariants indeed use such representations of the braid group. 41. In fact. Such models can be formulated in terms of the Yang-Baxter equation [35. 66. with time as the vertical dimension. 56]. . 37. and so corresponding quantum algorithms may shed light on the relationship of this level of computational complexity with quantum computing (See [23]). we indicate how topological braiding plus maxima (caps) and minima (cups) can be used to conﬁgure the diagram of a knot or link. In this way of looking at things. Can the invariants of knots and links such as the Jones polynomial be conﬁgured as quantum computers? This is an important question because the algorithms to compute the Jones polynomial are known to be NP-hard. 39. 30. 93] for more information about relationships of knot theory with statistical mechanics. In Figure 34. 36. This also can be translated into algebra by the association of a Yang-Baxter matrix R (not necessarily the R of the previous sections) to each crossing and other matrices to the maxima and minima. 43. This is the case when the R-matrices (the solutions to the Yang- Baxter equation used in the model) are unitary. One can use unitary representations of the braid group that are constructed in other ways. M is the composition of elementary braiding matrices. however. and we shall see this below. and |CAP as the measurement of this state. M must be a unitary operator. For topology. The next paragraph explains how this comes about. interacting (in a two-dimensional space) and ﬁnally annihilating one another. Quantum computation of the Jones polynomial. 82.1. 27. We regard CUP| as the preparation of this state. the knot diagram can be viewed as a picture. Thus we can write the amplitude in the form ZK = CUP|M |CAP where CUP| denotes the composition of cups. Each R-matrix is viewed as a a quantum gate (or possibly a composition of quantum gates). 11. Hopf algebras and quantum groups. that it is not necessary that the invariant be modeled via solutions to the Yang-Baxter equation. 37. The invariant takes the form of an amplitude for this process that is computed through the association of the Yang-Baxter solution R as the scattering matrix at the crossings and the minima and maxima as creation and annihilation operators. There are models of very eﬀective invariants of knots and links such as the Jones polynomial that can be put into this form [44]. We should remark.

A knot quantum computer. s2 −→ tQ + uI. and showing how the quantum computation devolves to ﬁnding the trace of a unitary transformation. 3. See [44. Let |v and |w be two qubits in V. Q 2 = |w|2 Q. The idea behind the construction of this representation depends upon the algebra generated by two single qubit density matrices (ket-bras). Quantum algorithms for computing the Jones polynomial have been dis- cussed elsewhere. P and Q generate a representation of the Temperley-Lieb algebra (See Section 12 of the present paper).280 LOUIS H. it is useful to point out this analogy between the structure of the knot invariants and quantum computation. KAUFFMAN <CAP| (measurement) M (unitary braiding) |CUP> (preparation) Z K = <CAP| M |CUP> Figure 34. as an example. s. . 56. 97]. a complex vector space of dimension two over the complex numbers. we give a local unitary representation that can be used to compute the Jones polynomial for closures of 3-braids. 2. 57. QPQ = |v|w|2 Q. Let P = |vv| and Q = |ww| be the corresponding ket-bras. We analyze this representation by making explicit how the bracket polynomial is computed from it. Here. Note that P 2 = |v|2 P. t. where I is the identity mapping on V and r. PQP = |v|w|2 P. u are suitably chosen scalars. Note also that this is a local unitary representation of B3 to U (2). One can adjust parameters to make a representation of the three-strand braid group in the form s1 −→ rP + sI. Nevertheless. In the following we use this method to adjust such a representation so that it is unitary.

The sum of products over the generators U1 and U2 of the Temperley-Lieb algebra comes from expanding this expression. let I (b) denote the sum of the exponents in the braid word that expresses b. 2 /3] * [5 /6. we have d = −2 cos(2). Note that the traces of these matrices are given by the formulas tr(U1 ) = tr(U2 ) = d while tr(U1 U2 ) = tr(U2 U1 ) = 1. With A = e i . To see it. Lemma 2. 0). /6] * [ /3. For b a three-strand braid. 7 /6] * [4 /3. Φ(s2 ) = AI + A−1 U2 . Here is a speciﬁc representation depending on two symmetric matrices U1 and U2 with d 0 U1 = = d |ww| 0 0 and −1 √ d 1 − d −2 U2 = √ = d |vv| 1 − d −2 d − d −1 √ where w = (1. If b is any braid. it follows that Φ(b) = AI (b) I + Π(b) where I is the 2 × 2 identity matrix and Π(b) is a sum of products in the Temperley-Lieb algebra involving U1 and U2 . Suppose that b = s1 s2−1 s1 . KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 281 We leave it as an exersise for the reader to verify that it ﬁts into our general classiﬁcation of such representations as given in section 10 of the present paper. Thus a specialization of a more general represention of the braid group gives rise to a continuous family of unitary representations of the braid group. For any A with d = −A2 − A−2 these formulas deﬁne a representation of the braid group. consider an example. Note that U12 = dU1 and U22 = dU1 . The desired representation of the Artin braid group is given on the two braid generators for the three strand braid group by the equations: Φ(s1 ) = AI + A−1 U1 . It is a calculation. . and v = (d −1 . U1 U2 U1 = U1 and U2 U1 U2 = U1 . Then Φ(b) = Φ(s1 s2−1 s1 ) = Φ(s1 )Φ(s2−1 )Φ(s1 ) = (AI + A−1 U1 )(A−1 I + AU2 )(AI + A−1 U1 ). 1 − d −2 ). Here I denotes the 2 × 2 identity matrix. We ﬁnd a speciﬁc range of angles in the following disjoint union of angular intervals ∈ [0. 44]. Moreover. 2 ] that give unitary representations of the three-strand braid group. assuming the entries of v are real. We omit the proof of this Lemma. 5 /3] * [11 /6. This is an example of a speciﬁc representation of the Temperley- Lieb algebra [35.

U2 . 2 Here CU denotes controlled U. We measure the expectation for the ﬁrst qubit |0 of the resulting state 1 (H |0 ⊗ | + H |1 ⊗ U |) 2 1 = ((|0 + |1) ⊗ | + (|0 − |1) ⊗ U |) 2 1 = (|0 ⊗ (| + U |) + |1 ⊗ (| − U |)). 2 2 2 . U1 U2 and U2 U1 . one can use the Hadamard test to obtain the diagonal matrix elements |U | of U . It follows from this calculation that the question of computing the bracket polynomial for the closure of the three-strand braid b is mathematically equivalent to the problem of computing the trace of the unitary matrix Φ(b). The result is the equation < b >= AI (b) d 2 + tr(Π(b)) where b denotes the standard braid closure of b. acting as U when the control bit is |1 and the identity mapping when the control bit is |0. can be calculated directly from the trace of this representation. denoted < b >. The trace is then the sum of these matrix elements as | runs over an orthonormal basis for the vector space. it follows that the value of the bracket polynomial of the closure of the braid b. 2 This expectation is 1 1 1 (| + |U † )(| + U |) = + Re|U |.282 LOUIS H. except for the part involving the identity matrix. In order to (quantum) compute the trace of a unitary matrix U . We ﬁrst obtain 1 1 + Re|U | 2 2 as an expectation by applying the Hadamard gate H 1 H |0 = √ (|0 + |1) 2 1 H |1 = √ (|0 − |1) 2 to the ﬁrst qubit of 1 CU ◦ (H ⊗ 1)|0| = √ (|0 ⊗ | + |1 ⊗ U |. KAUFFMAN Since the Temperley-Lieb algebra in this dimension is generated by I . From this we see at once that < b >= tr(Φ(b)) + AI (b) (d 2 − 2). The Hadamard test.U1 . and the sharp brackets denote the bracket polynomial.

That is (writing just F for FO and FM ). then composition is associative. Deﬁnition 1. F (1A ) = 1F (A) . Letting A denote the domain of f and B denote the codomain of f. then a functor F : Cat1 −→ Cat2 consists in functions FO : Obj(Cat1 ) −→ Obj(Cat2 ) and FM : Morph(Cat1 ) −→ Morph(Cat2 ) such that identity morphisms and composition of morphisms are preserved under these mappings. Obj(Cat). F (g ◦ f) = F (g) ◦ F (f). and g ◦ 1A = g for any morphism g with domain A. Morph(Cat). the objects of Cat. Note that the Hadamard test enables this quantum computation to estimate the trace of any unitary matrix U by repeated trials that estimate individual matrix entries |U |. If Cat1 and Cat2 are two categories. the domain of f and the codomain of f. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 283 The imaginary part is obtained by applying the same procedure to 1 √ (|0 ⊗ | − i|1 ⊗ U | 2 This is the method used in [3]. Quantum topology. 3. 2. Each morphism f is associated to two objects of Cat. It is useful in this regard to have available the concept of category. B and C are objects of Cat. the morphisms of Cat. The purpose of this section is to discuss the general idea behind topological quantum ﬁeld theory. 2. and we shall begin the section by discussing this far-reaching mathematical concept. and 2. That is (h ◦ g) ◦ f = h ◦ (g ◦ f). To each object A of Cat there is a unique identity morphism 1A : A −→ A such that 1A ◦ f = f for any morphism f with codomain A. Given f : A −→ B and g : B −→ C where A. 3. satisfying the following axioms: 1. it is customary to denote the morphism f by the arrow notation f : A −→ B. 4. . Given three morphisms f : A −→ B. then there exists an associated morphism g ◦ f : A −→ C called the composition of f and g. cobordism categories. A category Cat consists in two related collections: 1. 1. F (f : A −→ B) = F (f) : F (A) −→ F (B). §12. Temperley-Lieb algebra and topological quantum ﬁeld theory. and the reader may wish to contemplate its eﬃciency in the context of this simple model. and to illustrate its application to basic quantum mechanics and quantum mechanical formalism. We shall return to quantum algorithms for the Jones polynomial and other knot polynomials in a subsequent paper. g : B −→ C and h : C −→ D.

In all the examples mentioned so far. . The Artin braid group Bn can be regarded as a category whose single object is an ordered row of points [n] = {1.284 LOUIS H. and as its morphisms. . the morphisms in the category are restrictions of mappings in the category of sets. A given ordered row of points is interpreted as the starting or ending row of points at the bottom or the top of the braid. n}. For example. The morphisms from ∗ to itself are the elements of the group and composition is group multiplication. 2. KAUFFMAN A functor F : Cat1 −→ Cat2 is a structure preserving mapping from one category to another. The reader may wish to have morphisms between objects with diﬀerent n. Each morphism produces a permutation of the ordered row of points (corresponding to the begiinning and ending points of the individual braid strands). the morphisms have both external and internal structure. . has as its objects smooth manifolds of dimension n. 3. The notion of category is a broad mathematical concept. and weaving of the braid is extra structure beyond the object that is its domain and codomain. Cat(G). It is often convenient to think of the image of the functor F as an interpretation of the ﬁrst category in terms of the second. the fundamental group is a functor from the category of topological spaces with base point. ∂M n+1 . but this is not necessarily the case. For example. We regard M n+1 as a morphism from L(M n+1 ) to R(M n+1 ) M n+1 : L(M n+1 ) −→ R(M n+1 ). Thus one has the category of sets where the objects are sets (collections) and the morphisms are mappings between sets. The morphisms are the braids themselves and composition is the multiplication of the braids. We shall use this terminology below and sometimes refer to an interpretation without specifying all the details of the functor that describes it. Cob[n]. In this example. One has the category of groups where the objects are groups and the morphisms are homomorphisms of groups. In the case of the braid category. smooth manifolds M n+1 of dimension n +1 with a partition of the boundary. . Cat(B). The n-Cobordism Category. into two collections of n-manifolds that we denote by L(M n+1 ) and R(M n+1 ). for this example. with one object ∗. We will have this shortly in the Temperley-Lieb category and in the category of tangles. the object has no internal structure and all the complexity of the category is in the morphisms. One has the category of topological spaces where the objects are spaces and the morphisms are continuous mappings of topological spaces. whose objects are all ordered rows of points [n]. encompassing many ﬁelds of mathematics. Finally. any group G can be regarded as a category. and whose morphisms are of the form b : [n] −→ [n] where b is a braid in Bn . to the category of groups. we can take all the braid groups Bn (n a positive integer) under the wing of a single category. Functors are structure preserving mappings from one category to another. .

Lets begin with Cob[0]. Cob[0] is directly related to the Dirac notation of bras and kets and to the Temperley-Lieb algebara. and within unoriented manifolds there are those that are orientable and those that are not orientable. We shall concentrate in this section on these cobordism categories. S : p −→ q See Figure 35. the empty set in the category of manifolds). and their relationships with quantum mechanics. We take p to be an object of this category and also ∗. This is a place where notation and mathematical structure share common elements. where ∗ denotes the empty manifold (i. then the only morphism is the identity morphism. symbolized by p ⊗ p ⊗ · · · ⊗ p ⊗ p.e. these cobordism categories are highly signiﬁcant for quantum mechanics. The object ∗ occurs in Cob[n] for every n. p p f: p p Identity p p * p p * Figure 35. In general the objects of Cob[0] consist in the empty object ∗ and non-empty rows of points. One can choose to consider either oriented or non-oriented manifolds. with the standard deﬁnition of topological quantum ﬁeld theory. A line segment S with boundary points p and q is a morphism from p to q. Zero dimensional manifolds are just collections of points. the manifolds will be oriented. Elementary cobordisms. and the simplest one. Thus if we look at the subcategory of Cob[0] whose only object is p. but we shall not specify an orientation. The reader will note that Cob[0] and the usual arrow notation for morphisms are very closely related. In the next section. since it is possible that either the left set or the right set of a morphism is empty. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 285 As we shall see. In this ﬁgure we have illustrated the morphism from p to p. The deﬁnitions of the cobordism categories for oriented manifolds go over mutatis mutandis. The simplest convention for this category is to take this morphism to be the identity. Two points occur as the boundary of an interval. The simplest zero dimensional manifold is a single point p. In this section we will implicitly discuss only orientable manifolds. Figure 35 also contains a morphism p ⊗ p −→ ∗ .

The second represents a cobordism from the empty set to two points. we have indicated more morphisms in Cob[0]. and if a manifold M n+1 can be written as a union of two submanifolds Ln+1 and Rn+1 so that that an n-manifold W n is their common boundary: M n+1 = Ln+1 ∪ Rn+1 with Ln+1 ∩ Rn+1 = W n then. and we have named the morphisms just discussed as |Ω : p ⊗ p −→ ∗. We shall discuss this after describing the similarities with quantum mechanical formalism. This cobordism is topologically a circle and. Θ| : ∗ −→ p ⊗ p. In general. then a linear mapping from C to C is determined by the image of 1. In order to interpret the notion of scalar we would have to map the cobordism category to the category of vector spaces and linear mappings. and hence is characterized by the scalar that is the image of 1. . in the Dirac formalism is interpreted as a scalar. In Figure 36. Thus in Figure 36 we have that Θ| ◦ |Ω = Θ|Ω : ∗ −→ ∗ represents a cobordism from the empty manifold to itself. Nevertheless.286 LOUIS H. and M n+1 will be a scalar (morphism that commutes with all other mor- phisms) in the category Cob[n]. It is therefore assumed that in Cob[0] the composition with the morphism Θ|Ω commutes with any other morphism. The ﬁrst represents a cobordism of two points to the empty set (via the bounding curved interval). KAUFFMAN and the morphism ∗ −→ p ⊗ p. an n + 1 manifold without boundary behaves as a scalar in Cob[n]. In that way Θ|Ω behaves like a scalar in the cobordism category. The point to notice is that the usual conventions for handling Dirac bra-kets are essentially the same as the compostion rules in this topological category. the reader should note that if V is a vector space over the complex numbers C. In this sense a mapping C −→ C can be regarded as a possible image in vector spaces of the abstract structure Θ|Ω : ∗ −→ ∗. we can write M n+1 = Ln+1 ∪ Rn+1 = Ln+1 |Rn+1 .

Now view Figure 37. the bra–ket and ket–bra formalism is seen as patterns of connection of the one-manifolds that realize the cobordisms. Bras. This Figure illustrates a morphism S in Cob[0] that requires two crossed line segments for its planar representation. note how the zero dimensional cobordism category has structural parallels to the Dirac ket–bra formalism U = |ΩΘ| UU = |ΩΘ|ΩΘ| = Θ|Ω|ΩΘ| = Θ|ΩU. it is clear that . In the cobordism category. I S 2= I S SU = US = U Figure 37. kets and projectors. From this example. and S 2 = I where I denotes the identity morphisms for a two-point row. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 287 Identity |Ω > = <Θ| = < Θ |Ω > |Ω > < Θ | U U = |Ω > < Θ |Ω > < Θ | =U = < Θ | Ω > |Ω > < Θ | = < Θ | Ω> U Figure 36. { |Ω > < Θ | } 1 =P 1 { |Ω > <Θ | } = Q {1 <Θ|} {|Ω> 1} Figure 38. Permutations. Projectors in tensor lines and elementary topology. Thus S can be regarded as a non-trivial permutation. Getting back to the contents of Figure 36.

For algebraic purposes the loop ∗ −→ ∗ is taken to be a scalar algebraic variable that commutes with all elements in the algebra. Thus the equation UU = Θ|ΩU. AlgTL[n] is generated by the morphisms in TL[n] that go from [n] to itself. Note that with the empty object ∗. KAUFFMAN Cob[0] contains the structure of all the syymmetric groups and more. . In fact. Then the algebra AlgTL[n] is generated by {In . U1 . For example. but a diﬀerent commutative ring of coeﬃcients can be chosen and the value of the loop may be taken in this ring. Up to multiplication by the loop. and connects i to i +1 in each row. In this sense. Let Ui denote the morphism from [n] to [n] that connects k with k for k < i and k > i +1 from one row to the other. see [48]. The multiplicative structure of AlgTL[n] can be described by generators and relations as follows: Let In denote the identity morphism from [n] to [n].288 LOUIS H. We shall call this crossingless subcategory of Cob[0] the Temperley-Lieb Category and denote it by CatTL. This part can be characterized by those morphisms that can be represented by planar diagrams without crossings between any of the line segments (the one-manifolds). one could call Cob[0] the Brauer category. In CatTL we have the subcategory TL[n] whose only objects are the row of n points and the empty object ∗. we shall be concentrating on the part of Cob[0] that does not involve permutations. if we take the subcateogry of Cob[0] consisting of all morphisms from [n] to [n] for a ﬁxed positive integer n. U2 . The Temperley-Lieb Algebra. the morphism whose diagram is a single loop appears in TL[n] and is taken to commute with all other morphisms. . In the algebra we are allowed to add morphisms formally and this addition is taken to be commutative. For a proof that these are indeed all the relations. We shall return to this point of view later. In this section. Initially the algebra is taken with coeﬃcients in the integers. becomes UU = U in the algebra. We leave the commuting relation for the reader to draw in the case where n is four or greater. These relations are illustrated for three strands in Figure 38. then this gives the well-known Brauer algebra (see [9]) extending the symmetric group by allowing any connections among the points in the two rows. Un−1 } with relations Ui2 = Ui Ui Ui+1 Ui = Ui Ui Uj = Uj Ui : |i − j| > 1. for quantum mechanical applications it is natural to work over the complex numbers. . the product (composition) of two such morphisms is another ﬂat morphism from [n] to itself. . . and whose morphisms can all be represented by conﬁgurations that embed in the plane as in the morphisms P and Q in Figure 38.

Speciﬁcally. . and the relation PQP = P which is the same as U1 U2 U1 = U1 Note the composition at the bottom of the Figure 39. . This is exactly what we expect in interpretations. Thus Figure 39 illustrates the morphisms P and Q obtained by such tensoring. In this way there is a correspondence between morphisms p ⊗ p −→ ∗ and morphims p −→ p. we associate Θ : V −→ V in this way. Comparing with the diagrammatic for the category Cob[0]. as mappings from the generating object to itself. and a multiplicity of morphisms can correspond to a single line segment. one can envisage a more complex interpretation of the morphisms where each one-manifold (line segment) has a label. |m} be a basis for V . we can interpret the line segment [1] −→ [1] as a mapping from a vector space V to itself. . let {|0. . by straightening. For example. Here we see a composition of the identity tensored with a ket. Then [1] −→ [1] is the diagrammatic abstraction for V −→ V. we obtain the beginnings of the Temperley-Lieb algebra and the Temperley-Lieb category. In Figure 40 we have illustrated the generalization of the straightening proce- dure of Figure 39. Then Θ : V −→ V is determined by Θ|i = Θij |j (where we have used the Einstein summation convention on the repeated index j) corresponds to the bra Θ| : V ⊗ V −→ C deﬁned by Θ|ij = Θij . and there are many instances of linear mappings from V to V . By tensoring the ket–bra on one side or another by identity morphisms. At the vector space level there is a duality between mappings V ⊗ V −→ C and linear maps V −→ V . Nevertheless. The diagrammatic for this association involves “straightening” the curved structure of the morphism to a straight line. we say that Θ : V −→ V is obtained by straightening the mapping Θ| : V ⊗ V −→ C. Given Θ| : V ⊗ V −→ C. We have denoted these corresponding morphisms by Θ and Ω respectively. followed by a bra tensored with the identity. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 289 Figure 38 and Figure 39 indicate how the zero dimensional cobordism category contains structure that goes well beyond the usual Dirac formalism. . pointing out that in this category each of the morphisms Θ| and |Ω can be seen. In Figure 39 the straightening occurs because the connection structure in the morphism of Cob[0] does not depend on the wandering of curves in diagrams for the morphisms in that category. In Figure 40 we have elaborated this situation even further.

resulting in |φ. Figure 40 illustrates the staightening of |Θ and Ω|. The key to teleportation. If we interpret [2] as a single vector space W. KAUFFMAN Note that in this interpretation. |Θ > Θ <Ω| Ω |φ > |Θ > |ψ > <Ω | |φ > |ψ > Θ Ω |φ > = Θ Ω |ψ > Figure 40. .290 LOUIS H. the bras and kets are deﬁned relative to the tensor product of V with itself and [2] is interpreted as V ⊗ V . The resulting single qubit state. { |Ω > < Θ | } 1 =P 1 { |Ω > <Θ | } = Q {1 <Θ|} {|Ω> 1} =R = R = 1 = PQP = P Figure 39. then the usual formalisms of bras and kets still pass over from the cobordism category. and the straightening of a composition of these applied to |. In the left-hand part of the bottom of Figure 40 we illustrate the preparation of the tensor product |Θ ⊗ | followed by a successful measurement by Ω| in the second two tensor factors. as seen by straightening. is |φ = Θ ◦ Ω|. The basic Temperley-Lieb relation.

transform a state | via a combination of preparation and measurement just so long as the straightenings of the preparation and measurement (Θ and Ω) are each invertible (unitary). This is the key to teleportation [49. In general we will have |φ = Ω|. In Figure 41 we show typical cobordisms (morphisms) in Cob[1] from two circles to one circle and from one circle to two circles. The purpose of this section is to discuss in a very general way how braiding is related to topological . 13. For now. There is much more to say about the category Cob[0] and its relationship with quantum mechanics. then the transmitted state |φ will be equal to |. Later in this paper. As we go to higher dimensions the structure of cobordisms becomes more interesting and more complicated. Their composition is a surface of genus one seen as a morphism from two circles to two circles. 1]. Braiding and topological quantum ﬁeld theory. we shall use these ideas in formulating our representations of the braid group. and invite the reader to explore further. indeed unitarily. One can then choose a basis of measurements |Ω. In the standard teleportation procedure one chooses the preparation Θ to be (up to normalization) the 2 dimensional identity matrix so that | = |00 + |11. we point out how things look as we move upward to Cob[n] for n > 0. each corresponding to a unitary transformation Ω so that the recipient of the transmission can rotate the result by the inverse of Ω to reconsitute | if he is given the requisite information. If the successful measurement Ω is also the identity. §13. The bottom of the ﬁgure indicates a ket-bra in this dimension in the form of a mapping from one circle to one circle as a composition of a cobordism of a circle to the empty set and a cobordism from the empty set to a circle (circles bounding disks). KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 291 From this. Cobordisms of 1-manifolds are surfaces. It is remarkable that there is so much structure in the lowest dimensions of these categories. This is the basic design of the teleportation procedure. We will stop here. Figure 41. we see that it is possible to reversibly. These are often called “pairs of pants”.

a vector Z(Y ) ∈ Z(Σ) for each compact. The ideas in the subject of topological quantum ﬁeld theory (TQFT) are well expressed in the book [6] by Michael Atiyah and the paper [95] by Edward Witten. oriented (d + 1)-dimensional manifold Y with boundary Σ. The remarkable fact is that the case of three-dimensions is also related to quantum theory. and to the lower-dimensional versions of the TQFT. KAUFFMAN quantum ﬁeld theory. in this view a TQFT is basically a functor from the cobordism categories deﬁned in the last section to Vector Spaces over the complex numbers. Z(Σ1 ∪ Σ2 ) = Z(Σ1 ) ⊗ Z(Σ2 ) where ∪ denotes disjoint union. Y2 is a cobordism from Σ2 to Σ3 and Y is the composite cobordism Y = Y1 ∪Σ2 Y2 . Z(φ) = C (C denotes the complex numbers) for the empty manifold φ. This gives a signiﬁcant way to think about three-manifold invariants in terms of lower dimensional patterns of interaction. oriented d - dimensional manifold Σ. We have already seen that in the lowest dimensional case of cobordisms of zero-dimensional manifolds. then Z(Y ) = Z(Y2 ) ◦ Z(Y1 ) : Z(Σ1 ) −→ Z(Σ2 ) is the composite of the corresponding linear mappings. Note that. 4. The functor satisﬁes the following axioms. in itself a . 3. The surface is divided up into trinions as illustrated in Figure 42. If Y1 is a cobordism from Σ1 to Σ2 . Here is Atiyah’s deﬁnition: Deﬁnition 2. A trinion is a surface with boundary that is topologically equivalent to a sphere with three punctures. a ﬁnite dimensional vector space Z(Σ) to each compact. 2. Regard the three-manifold as a union of two handlebodies with boundary an orientable surface Sg of genus g. In the section to follow. this gives rise to a rich structure related to quatum mechanics and quantum information theory. A TQFT in dimension d is a functor Z(Σ) from the cobordism category Cob[d ] to the category Vect of vector spaces and linear mappings which assigns 1. a linear mapping Z(Y ) : Z(Σ1 ) −→ Z(Σ2 ) when Y is a (d + 1)-manifold that is a cobordism between Σ1 and Σ2 (whence the boundary of Y is the union of Σ1 and −Σ2 . we will use the Temperley-Lieb recoupling theory to produce specﬁc unitary representations of the Artin braid group. Z(Σ × I ) is the identity mapping on Z(Σ). 2. The trinion constitutes. Here follows a brief description.292 LOUIS H. With Σ × I (where I denotes the unit interval) denoting the identity cobordism from Σ to Σ. 1. 5. Z(Σ† ) = Z(Σ)† where Σ† denotes the manifold Σ with the opposite orientation and Z(Σ)† is the dual vector space. 3.

distinct graphs can correspond to topologically identical cobordisms of circles. Below this isotopy we indicate the corresponding graphs. The pattern of a trinion is a trivalent graphical vertex. Given a surface S (possibly with boundary) and a decomposition of that surface into triions. as illustrated in Figure 44. t) where t denotes the particular trinion decomposition. so that the category of trivalent graphs (as morphisms from ordered sets of points to ordered sets of points) has an image in the category of cobordisms of compact one-dimensional manifolds. In that ﬁgure we show the trivalent vertex graphical pattern drawn on the surface of the trinion. we associate to it a trivalent graph G(S. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 293 cobordism in Cob[1] from two circles to a single circle. . This can be accomplished as an ambient isotopy of the embeddings in three dimensional space that are indicated by this ﬁgure. or from a single circle to two circles. In Figure 45 we illustrate another feature of the relationship betweem surfaces and graphs. In the graph category there will have to be a transformation between a braided and an unbraided trivalent vertex that corresponds to this homeomorphism. as illustrated in Figure 42. and that it is extraordinarily useful to articulate transformations between the graphs that correspond to the homeomorphisms of the corresponding surfaces.B and C ) ﬁxed while undoing the internal twist. At the top of the ﬁgure we indicate a homeomorphism between a twisted trinion and a standard trinion. or from three circles to the empty set. forming a graphical pattern for this combordism. It turns out that the graphical structure is important. Decomposition of a surface into trinions. The homeomorphism leaves the ends of the trinion (denoted A. The beginning of this structure is indicated in the bottom part of Figure 44. Trinion Figure 42. In this correspondence. It should be clear from this ﬁgure that any cobordism in Cob[1] can be diagrammed by a trivalent graph.

Trinion associativity. and the diagrams of braiding and . A B A B = C C Figure 45. From the point of view that we shall take in this paper. the key to the mathematical structure of three-dimensional TQFT lies in the trivalent graphs. Trivalent vectors. We can think of these braided graphs as representing idealized Feynman diagrams. KAUFFMAN a b ε V( ) c d a b e f ε V( ) c Figure 43. = Figure 44. Tube twist. with the trivalent vertex as the basic particle interaction vertex. In this view one thinks of the particles as moving in a two-dimensional medium. including the braiding of grapical arcs. and the braiding of lines representing an interaction resulting from an exchange of particles.294 LOUIS H.

This vector space is the space of processes associated with the graph G. These are both included in the category of braided trivalent graphs. then the set of vector spaces and linear mappings associated to the surfaces can consitute a functor from the category of cobordisms of one-manifolds to vector spaces. any trivalent graph will have a set of admissible labelings. Similarly. Given a surface S and a decomposition t of the surface into trinions. We take the set of admissible labelings of a given graph G as a basis for a vector space V (G) over the complex numbers. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 295 trivalent vertex interactions as indications of the temporal events in the system. We shall detail the requirements below. Thinking of the basic trivalent vertex as the form of a particle interaction there will be a set of particle states that can label each arc incident to the vertex. and the particle with spin c is to the left of the particle with spin d . t)). Here it suﬃces to note that there will be some restrictions on these labels. one takes the convention that the particle with spin a is to the left of the particle with spin b. by deﬁnition a lableled trivalent graph in a category of graphs that satisfy the properties outlined in the previous paragraph. N. and hence gives rise to a one-dimensional topological quantum ﬁeld theory. The simplest case of this idea is C. then there is an underlying permutation that is obtained by following strands from the bottom to the top of the diagram (thinking of time as moving up the page). Yang’s original interpretation of the Yang-Baxter equation [98]. so that a trivalent vertex has a set of possible labelings. with time indicated in the direction of the morphisms in the category. A spin network is. In the next two sections we will see speciﬁc rules for labeling such states. Yang articulated a quantum ﬁeld theory in one dimension of space and one dimension of time in which the R-matrix giving the scattering ampitudes for an interaction of two particles whose (let us say) spins cd corresponded to the matrix indices so that Rab is the amplitude for particles of spin a and spin b to interact and produce particles of spin c and d . we have the associated graph G(S. . In Figure 43 we illustrate the labeling of the trivalent graphs by such particle states. If this can be accomplished. Adding such graphs to the category of knots and links is an extension of the tangle category where one has already extended braids to allow any embedding of strands and circles that start in n ordered points and end in m ordered points. If one follows the concatenation of such interactions. The tangle category includes the braid category and the Temperley-Lieb category. It is desirable to have this vector space independent of the particular decomposition into trinions. Yang designed the Yang-Baxter equation for R so that the amplitudes for a composite process depend only on the underlying permutation corresponding to the process and not on the individual sequences of interactions. Since these interactions are between particles in a line. To this end we need some properties of the particle interactions that will be described below. These are the possible particle processes that this graph can support. t) and hence a vector space of processes V (G(S.

Figure 47 illustrates the Yang-Baxter equation. The extra spatial dimension is taken in displacing the woven strands perpendicular to the page. These are the trivalent vertices discussed above. Taking this picture to heart.296 LOUIS H. Matrix elements corresponding to trivalent vertices can represent these interactions. Creation and fusion. It is natural to assume that braiding intertwines with creation as shown in Figure 49 (similarly with fusion).and over-crossings as modeling events in a spacetime with two dimensions of space and one dimension of time. we can use the same interpretation. but think of the diagrams with their under. This corresponds to the associativity at the level of trinion combinations shown in Figure 44. There will be a matrix expression for the compositions of braiding and fusion or creation as indicated in Figure 25. and allows us to use braiding operators R and R−1 as scattering matrices. The actual formalism of such an operator will parallel the mathematics of . since it indicates that the diagrams can be interpreted as embeddings of graphs in three-dimensional space. Once one introduces trivalent vertices for fusion and creation. Figure 46. The intertwining identity is an assumption like the Yang-Baxter equation itself. there is the question how these interactions will behave in respect to the braiding operators. R I I R R I I R I R = R I R I I R Figure 47. KAUFFMAN In taking over the Yang-Baxter equation for topological purposes. Here we will restrict ourselves to showing the diagrammatics with the intent of giving the reader a ﬂavor of these structures. one can add other particle properties to the idealized theory. and it ﬁts with our interpretation of the vertices in terms of trinions. See Figure 46. that simpliﬁes the mathematical structure of the model. This intertwining identity is clearly the sort of thing that a topologist will love. In particular one can add fusion and creation vertices where in fusion two particles interact to become a single particle and in creation one particle changes (decays) into two particles. Yang-Baxter equation. It is to be expected that there will be an operator that expresses the recoupling of vertex interactions as shown in Figure 50 and labeled by Q.

One says that the interactions satisfy the hexagon identity if this composition is the identity. Finally there is a hexagonal cycle of interactions between braiding. . If one just considers the abstract structure of recoupling then one sees that for trees with four branches (each with a single root) there is a cycle of length ﬁve as shown in Figure 51. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 297 =R Figure 48. Recoupling. F F F F F Figure 51. recoupling for angular momentum. F Figure 50. Braiding. One can start with any pattern of three vertex interactions and go through a sequence of ﬁve recouplings that bring one back to the same tree from which one started. Intertwining. See for example [41]. Pentagon identity. It is a natural simplifying axiom to assume that this composition is the identity mapping. = Figure 49. This axiom is called the pentagon identity. recoupling and the intertwining identity as shown in Figure 52.

For the case . but a sketch is in order. Thus ABA = R(R−1 F −1 R)R = F −1 R2 = (F −1 RF )F −1 R = (R−1 F −1 R)F −1 R = (R−1 F −1 R)R(R−1 F −1 R) = BAB. the pentagon identity and the hexagon identity. The hexagon identity tells us that R−1 FRF −1 RF = I where I is the identity mapping on the process space for trees with three branches. Letting A=R and B = F −1 RF. we see that the hexagon identity is equivalent to the statement B = R−1 F −1 R. the intertwining identity.298 LOUIS H. implies that A and B satisfy the braiding relation. A graphical three-dimensional topological quantum ﬁeld theory is an algebra of interactions that satisﬁes the Yang-Baxter equation. It is worth pointing out how these identities are related to the braiding. Thus the hexagon relation in this context. The combination of the hexagon and pentagon relations ensures that the braid group representations that are generated are well-deﬁned and ﬁt together as we include smaller numbers of strands in larger numbers of strands. We omit the further details of the veriﬁcation of this statement. Remark 9. There is not room in this summary to detail the way that these properties ﬁt into the topology of knots and three-dimensional manifolds. KAUFFMAN F R F = R R F Figure 52. Hexagon identity.

an invariant of the three-manifold M3 . K ) can be formally compared with the Witten [95] integral Z(M . 41] for more information on this approach. Such TQFT models would describe applicable physics. up to normalization. 91]. then the inner product I (M ) = M− |M+ is. at base. and a surface of genus three that is decomposed into four trinions. See [39. K ) of knots and links in orientable three-manifolds. The invariant I (M 3 . 69.). By these graphical reformulations. Such particles are called Anyons. Aspects of the quantum Hall eﬀect may be related to topological quantum ﬁeld theory [94]. hexagon and pentagon identities. Z(M. One can study a physics in two dimensional space where the braiding of particles or collective excitations leads to non-trival representations of the Artin braid group. Furthermore. so that one obtains a source of invariants I (M 3 . It turns out that the vector space V (Sg ) = V (G(Sg . 77]. Here we see the uses of the relationships that occur in the higher dimensional cobordism categories. With the deﬁnition of graphical topological quantum ﬁeld theory given above. and deﬁned in terms of the graphical topological quantum ﬁeld theory. if a closed three-manifold M 3 is decomposed along a surface Sg into the union of M− and M+ where these parts are otherwise disjoint three-manifolds with boundary Sg . Now return to Figure 42 where we illustrate trinions. One expects that physical situations involving 2 + 1 spacetime will be approximated by such an idealized theory. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 299 of topological quantum ﬁeld theory related to the group SU (2) there is a construction based entirely on the combinatorial topology of the bracket polynomial (See Sections 11 to 18 of this paper. K ) and I (M 3 . K) = DAe (ik/4 )S(M. shown in relation to a trivalent vertex. as descirbed in the previous section. does not depend upon the choice of trinion decomposition. . t)) to a surface with a trinion decomposition as t described above. One can then associate a well-deﬁned vector |M in V (Sg ) whenenver M is a three manifold whose boundary is Sg . knots and links can be incorporated as well. K ) are essentially equivalent for appropriate choice of gauge group and corresponding spin networks. 39. The reader interested in the SU (2) case of this structure and its implications for invariants of knots and three manifolds can consult [41. There are also applications to 3 + 1 quantum gravity[53.A) WK (A). This independence is guaranteed by the braiding. a highly simpliﬁed theory of point particle interactions in 2 + 1 dimensional spacetime. 14. It can be used to articulate invariants of knots and links and invariants of three manifolds. 3 It can be shown that up to limits of the heuristics. One can think about applications of anyons to quantum computing along the lines of the topoological models described here. 85. a three-dimensional TQFT is.

based on the composition of recoupling with the elementary braiding at a vertex. 23. In the next section we show how certain natural deformations [41] of Penrose spin networks [80] can be used to produce these unitary representations of the Artin braid group and the corresponding models for anyonic topological quantum computation.300 LOUIS H. There we show a more complex braiding operator. In this technology a symmetrizer is a sum of tangles on n strands (for a chosen integer n). Jones- Wenzl projector) is constructed on the basis of the bracket polynomial expan- sion. The result of this generalization is a structure that satisﬁes all the properties of a graphical TQFT as described in the previous section. 25]. In this section we discuss a combinatorial construction for spin networks that generalizes the original construction of Roger Penrose. A more complex braiding operator. (This structure is implicit in the Hexagon identity of Figure 52. and specializes to classical angular momentum recoupling the- ory in the limit of its basic variable. §14.) The new braiding operator is a source of unitary representations of braid group in situations (which exist mathematically) where the recoupling transformations are themselves unitary. View Figure 54 and Figure 55. In Figure 54 we indicate how the basic projector (symmetrizer. Spin networks and Temperley-Lieb recoupling theory. A complete de- scription of this theory can be found in the book “Temperley-Lieb Recoupling Theory and Invariants of Three-Manifolds” by Kauﬀman and Lins [41]. KAUFFMAN F R -1 F -1 B = F RF Figure 53. The “q-deformed” spin networks that we construct here are based on the bracket polynomial relation. This kind of pattern is utilized in the work of Freedman and collaborators [22. The construction is based on the properties of the bracket polynomial (as already described in Section 11). 21] and in the case of classical angular momentum formalism has been dubbed a “spin-network quantum simlator” by Rasetti and collaborators [75. 24. 20. The tangles are made by summing over braid lifts of . A key point in the application of TQFT to quantum information theory is contained in the structure illustrated in Figure 53.

. 2 = = −1/δ n 11 n 1 1 = −Δ n /Δ n+1 n 1 Δ -1 = 0 Δ0 = 1 Δ n+1 = δ Δ n ..Δ n-1 Figure 55. The projectors have the property that the concatenation of a projector with itself is just that projector. and if you tie two lines on the top or the bottom .A = d -1 = A + A . Two strand projector. Each elementary braid is then expanded by the bracket polynomial relation as indicated in Figure 54 so that the resulting sum consists of ﬂat tangles without any crossings (these can be viewed as elements in the Temperley-Lieb algebra). n n n = = . permutations in the symmetric group on n letters. n strands t(σ) {n}! = Σ(A -4 ) σ ε Sn = 0 n t(σ) = (1/{n}!) Σ σε (A -3 ) ~ σ Sn Figure 54. Basic projectors. as indicated in Figure 54... KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 301 ~ 2 -2 = = -A .

In order to accomplish this interaction. There is a basic orthogonality of three vertices as shown in Figure 57. the left hand side of the equation becomes a Theta graph and the right hand side becomes a multiple of a “delta” where Δa denotes the bracket polynomial evaluation of the a-strand loop with a projector on it. with d = −A2 − A−2 is the loop value for the bracket polynomial. b. Here if we tie two three-vertices together so that they form a “bubble” in the middle. Vertex. k so that a = i + j. See Figure 56. It is the sum of two parallel arcs and two turn-around arcs (with coeﬃcient −1/d. b.302 LOUIS H. b = j + k. The interactions of these particles are governed by how they can be tied together into three-vertices. For example a + b ≥ c. together to form a three-vertex. c is greater than or equal to the third. is usually known as the Jones-Wenzl projector. One can think of the vertex as a possible particle interaction where [a] and [b] interact to produce [c]. any two of the legs of the vertex can be regarded as interacting to produce the third leg. j. then the resulting network with labels a and b on its free ends is a multiple of an a-line (meaning a line with an a-projector on it) or zero (if a is not equal to b). we must share lines between them as shown in that ﬁgure so that there are non-negative integers i. Here the formula for that projector is particularly simple. . as shown in the ﬁgure. This recursion formula is due to Jones and Wenzl and the projector in this form. That is. of a. This general deﬁnition of projectors is very useful for this theory. The multiple is compatible with the results of closing the diagram in the equation of Figure 57 so the two free ends are identiﬁed with one another. This is equivalent to the condition that a +b+c is even and that the sum of any two of a. b. KAUFFMAN a b a b j i k i+j=a c j+k=b c i+k=c Figure 56. b. The two-strand projector is shown in Figure 55. As such the reader can think of them as “particles”. The Θ(a. The projectors are combinatorial analogs of irreducible representations of a group (the original spin nets were based on SU (2) and these deformed nets are based on the corresponding quantum group to SU(2)). c strands respectively. developed as a sum in the Temperley-Lieb algebra (see Section 12 of this paper). Figure 55 also shows the recursion formula for the general projector. then the evaluation is zero. c = i + k. In Figure 56 we show how to tie three projectors. c) denotes the bracket evaluation of a theta graph made from three trivalent vertices and labeled with a. c on its edges. On closure. of a projector together.

Finally. Recoupling formula. . a b c i d k = Tet [ a b i c d k ] Figure 59. The tetrahedral graph is shown in Figure 59. recoupling coeﬃcients that can be expressed. there is the braiding relation. Here we are indicating only the relationships and external logic of these objects. a b a b i =Σ{ a b i c d j } j c d j c d Figure 58. Orthogonality of trivalent vertices. c . These can be found in [41]. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 303 There is a recoupling formula in this theory in the form shown in Figure 58. Tetrahedron network. a a = a = a = Δa c d a = Θ( a . c . as illustrated in Figure 36. One derives the formulas for these coeﬃcients directly from the orthogonality relations for the trivalent vertices by closing the left hand side of the recoupling formula and using orthogonality to evaluate the right hand side. Here there are “6-j symbols”. The reader should be advised that there are speciﬁc calculational formulas for the theta and tetrahedral nets. d ) c d = Δa a δb b Figure 57. in terms of tetrahedral graph evaluations and theta graph evaluations. as shown in Figure 58. This is illustrated in Figure 60. d ) a a Θ( a .

1. In this section we discuss the structure of the evaluations for Δn and the theta and tetrahedral networks. this q-deformed spin network theory satisﬁes the pentagon. d . b. as illustrated in Figure 57. 14. k) Figure 60. Tetrahedron formula for recoupling coeﬃcients. k ) Δk Tet [ a b i c d k ] Δk { a b i c d k } = Θ( a . KAUFFMAN a b =Σ{ a b c i d k j a b i c d j } c j d k = Σ{ a b i c d j } Θ( a . Evaluations. d . k ) Θ( c . We refer to [41] for the details behind these formulas. One can apply the theory to many diﬀerent situations. A2 − A−2 . b. All these identities follow naturally from the basic underlying topological construction of the bracket polynomial. one ﬁnds that A2n+2 − A−2n−2 Δn = (−1)n . a b a b ab = λc c c ab (a+b-c)/2 (a'+b'-c')/2 λc = (-1) A x' = x(x+2) Figure 61. For the bracket variable A. d . b . With the braiding relation in place. hexagon and braiding naturality identities needed for a topological quantum ﬁeld theory. j ) Δ δ k Δ j Δ j j j j ={ a b i c d k } Θ( a . Recall that Δn is the bracket evaluation of the closure of the n-strand projector.304 LOUIS H. Local braiding formula. k ) Θ( c . j ) Θ( c .

Note that (a + b + c)/2 = m + n + p. We shall see particular applications of this viewpoint later in the paper. Then we have the formula √ Δa Δb Δc ModVert[a. b. sin( /r) Here the corresponding quantum integer is sin(n /r) [n] = . . b. c = n + p. the recoupling theory becomes ﬁnite with the restriction that only three-vertices (labeled with a. b. c] denote the modiﬁed vertex. b. b. 14. . Let ModVert[a. . . sin( /r) Note that [n + 1] is a positive real number for n = 0. 2. Θ(a. By multiplying all the vertices by an appropriate factor.2. c) are admissible when a + b + c ≤ 2r − 4. then the recoupling transformations are unitary. 1. c]. c) . This is a very useful fact. b. c] denote the original 3-vertex of the Temperley-Lieb recoupling theory. in the sense that its transpose is equal to its inverse. Let Vert[a. c] = Vert[a. r − 2 and that [r − 1] = 0. It means that when the resulting matrices are real. The evaluation of the theta net is expressed in terms of quantum integers by the formula [m + n + p + 1]![n]![m]![p]! Θ(a. When A = e i /2r . All the summations in the formulas for recoupling are restricted to admissible triples of this form. then sin((n + 1) /r) Δn = (−1)n . b. The formula for the recoupling coeﬃcients given in Figure 60 has less symmetry than is actually inherent in the structure of the situation. we can reconﬁgure the formulas in this theory so that the revised recoupling transformation is orthogonal. Figure 62 illustrates this modiﬁcation of the three-vertex. A2 − A−2 If A = e i /2r where r is a positive integer. Symmetry and unitarity. b = m + n. c) = (−1)m+n+p [m + n]![n + p]![p + m]! where a = m + p. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 305 One sometimes writes the quantum integer A2n − A−2n [n] = (−1)n−1 Δn−1 = .

Proof. b. a b i M [a. c) is positive real. c) = (−1)(a+b+c)/2 Θ(a. b. We rewrite the recoupling formula in this new basis and emphasize that the recoupling coeﬃcients can be seen (for ﬁxed external labels a. c) is real. d ]ij = ModTet / Δa Δb Δc Δd c d j . and can be taken to be a positive real number for (a. c) ˆ where Θ(a. and Δa Δb Δc = (−1)(a+b+c) [a + 1][b + 1][c + 1] where the quantum integers in this formula can be taken to be positive real. b. c) = . In Figure 63 we show how this modiﬁcation of the vertex aﬀects the non-zero term of the orthogonality of trivalent vertices (compare with Figure 57).e. b. c) showing that this factor can be taken to be positive real. KAUFFMAN Lemma 3. b. c.” The coeﬃcient in the modiﬁed bubble identity is : : Δb Δc (b+c−a)/2 [b + 1][c + 1] = (−1) Δa [a + 1] where (a. c. It follows from this that : [a + 1][b + 1][c + 1] f(a. In Figure 64. b. b.using the matrix notation M [a. d ]ij for the modiﬁed recoupling coeﬃcients. c) admissible (i. By the results from the previous subsection. ˆ Θ(a. In particular b + c − a is even and hence this factor can be taken to be real. a + b + c ≤ 2r − 4).306 LOUIS H. Figure 65 and Figure 66 we have shown the form of this transformation. ˆ Θ(a. The result shown in Figure 64 and Figure 65 is the following formula for the recoupling matrix elements. The proof of this formula follows directly from trivalent–vertex orthogonality (See Figure 57 and Figure 60. b. b. In Figure 64 we derive an explicit formula for these matrix elements.). For the bracket evaluation at the root of unity A = e i /2r the factor √ Δa Δb Δc f(a. b. c) form an admissible triple. b. c) = Θ(a. c. b. d ) as a matrix transforming the horizontal “double-Y ” basis to a vertically disposed double-Y basis. and is given in Figure 64. We refer to this as the “modiﬁed bubble identity.

b. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 307 √ where Δa Δb Δc Δd is short-hand for the product : : Δa Δb Δc Δd Δj Δj Δj : : [a + 1][b + 1] [c + 1][d + 1] = (−1)(a+b−j)/2 (−1)(c+d −j)/2 (−1)j [j + 1] [j + 1] [j + 1] = (−1)(a+b+c+d )/2 [a + 1][b + 1][c + 1][d + 1] In this form. we see that this coeﬃent can be taken to be real. Fibonacci particles. d. a. Thus we obtain families of unitary representations of the Artin braid group from the recoupling theory at these roots of unity. b. d ]T = M [b. j) are admissible triples. c]. d. c. It follows from this formula that M [a. The proof is given the discussion above. In Figure 67 we illustrate the formula M [a. d. §15. The matrix M [a. b. c ) c Figure 62. d ]T = M [a. b. b. a b a b Δa Δ b Δ c = c Θ( a . In Section 16 we shall show explictly how these methods work in the case of the Fibonacci model where A = e 3i /5 . c. b . 81] can be constructed by using a version of the two-stranded bracket polynomial and a generalization of Penrose spin networks. c. real-valued matrix. a. d ]−1 . d ]−1 = M [b. Theorem 5. d ] is real-valued. c]. Modiﬁed three vertex. In this section and the next we detail how the Fibonacci model for anyonic quantum computing [68. and its value is independent of the choice of i and j. b. c. In the Temperley-Lieb theory we obtain unitary (in fact real orthogonal ) recoupling transformations when the bracket variable A has the form A = e i /2r for r a positive integer. Proof. c. Hence M [a. c. This is a fragment of the Temperley-Lieb recoupling . since (a. It follows from Figure 58 (turn the diagrams by ninety degrees) that M [a. b. d ] is an orthogonal. j) and (c.

The Fibonacci model is a TQFT that is based on a single “particle” with two states that we shall call the marked state and the unmarked state. c ) a a a a b c Δb Δc = Δa Figure 63. We already gave in the preceding sections a general discussion of the theory of spin networks and their relationship with quantum computing. Derivation of modiﬁed recoupling coeﬃcients. b . Modiﬁed bubble identity. theory [41]. The particle in the marked state can interact with itself either to produce a single particle in . b . KAUFFMAN a a Θ( a . c ) b c = Δa a a a Δa Δ b Δ c b c = b c Θ( a .308 LOUIS H. a b =Σ a b i a b j k j c d i k c d k c d =Σ a b c d ik Δ a Δb Δj Δ c Δd Δj Δ δk j j k Δ a Δb = a b c d ij Δj Δ c Δd Δj Δ j a b c d ij = a i b j = ModTet [ a b i c d j ] c d ΔaΔbΔc Δ d Δ a Δb Δ c Δd Δ Δj Δj j Figure 64.

a b b d i j i j c d = a c ΔaΔbΔc Δ d ΔaΔbΔc Δ d T -1 = a b c d = a b c d Figure 67. Modiﬁed recoupling formula. Modiﬁed matrix transpose. One way to indicate these two interactions symbolically is to use a box.b. Nesting: .c. .d] = ij c d ij Figure 66. or to produce a single particle in the unmarked state. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 309 a b a b i =Σ a b c d ij j c d j c d Figure 65. the marked state. a b a b c d ij = c i d j ΔaΔbΔc Δ d a b M[a. Modiﬁed recoupling matrix. Then one has two modes of interaction of a box with itself: 1.for the marked state and a blank space for the unmarked state. The particle in the unmarked state has no inﬂuence in interactions (an unmarked state interacting with any state S yields that state S). Adjacency: and 2.

. Fibonacci trees. From this point of view.310 LOUIS H. The arithmetic of combining boxes (standing for acts of distinction) according to these rules has been studied and formalized in [92] and correlated with Boolean algebra and classical logic. but it is worth noting some of its purely combinatorial properties ﬁrst. P P P P * particle interaction. KAUFFMAN With this convention we take the adjacency interaction to yield a single box. and the nesting interaction to produce nothing: = = We take the notational opportunity to denote nothing by an asterisk (*). Thus = ∗. Fibonacci P We shall make a recoupling theory based on this particle. Figure 68. Here within and next to are ways to refer to the two sides delineated by the given distinction. there are two modes of relationship (adjacency and nesting) that arise at once in the presence of a distinction. The syntatical rules for operating the asterisk are Thus the asterisk is a stand-in for no mark at all and it can be erased or placed wherever it is convenient to do so. P P P P dim(V111) = 1 0 * P P P P P P P P * P dim(V 1111 ) = 2 P 0 |0> * |1> * Figure 69.

. . we shall denote this tree by V0111. See Figure 69 for an illustration of the simplest cases. . . In the case where all the labels are marked at the top and the bottom label is unmarked. More generally one can consider a left-associated tree with n upward branches and one root. . P −→ P In Figure 69 we indicate in small tree diagrams the two possible interactions of the particle P with itself. We see from Figure 69 that the dimension of V0(3) is 1.an associated with the tree. an unmarked state at the bottom). 1. In Figure 69. . and they constitute our formalism for this very elementary particle. . 1. a2 .a2 . In the ﬁrst interaction the particle vanishes. . Given the tree T (1. It follows from this that dim(V0(n) ) = fn−2 where fk denotes the k-th Fibonacci number: f0 = 1. In each case we demand that the particles interact successively to produce an unmarked particle in the end. . Thus the two possible interactions of P with itself are as follows. f3 = 3. Let T (a1 . all marked and then four branches all marked.. Each such sequence is regarded as a basis vector in a complex vector space Vba1 . . In the second interaction the particle a single copy of P is produced.. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 311 From here on we shall denote the Fibonacii particle by the letter P. P −→ ∗ 2. P. . an on the top and root label b at the bottom of the tree. f2 = 2..11 = V0(n) where n denotes the number of upward branches in the tree. 1 : 0) (n marked states at the top. We consider all possible processes (sequences of particle interactions) that start with the labels at the top of the tree. at the root of the tree. P. f4 = 5. we have indicated the diﬀerent results of particle processes where we begin with a left-associated tree structure with three branches. . This means that V0(4) is a natural candidate in this context for the two-qubit space.. a process basis vector in V0(n) is in direct correspondence with a string of boxes and asterisks (1’s and 0’s) of length n − 2 with no repeated asterisks and ending in a marked state. f1 = 1. 1. .. . . an : b) denote such a tree with particle labels a1 . producing the asterix. . These are the two basic actions of a single distinction relative to itself.. and that dim(V0(4) ) = 2. f5 = 8. and end with the labels at the bottom of the tree.

the present section is self-contained. . list its two main properties (the operator is idempotent and a self-attached strand yields a zero evaluation) and give diagrammatic proofs of these properties. While we have outlined the general recoupling theory based on the bracket polynomial in earlier sections of this paper. We now show how to make a model for recoupling the Fibonacci particle by using the Temperley Lieb recoupling theory and the bracket polynomial. There is a third possibility. KAUFFMAN where fn+2 = fn+1 + fn . The Fibonacci recoupling model. using only basic information about the bracket polyonmial. P * PP P * *P PPP PP P * * P * PP * P* Tree of squences with no occurence of Figure 70. Everything we do in this section will be based on the 2-projector. In this ﬁgure we state the deﬁnition of the 2-projector. We could remark at the outset. The Fibonaccie particle is. that the 4-projector will be zero if we choose the bracket polynomial variable A = e 3 /5 . where two 2-projectors interact to produce a 4-projector. there are two basic interactions of the 2-projector with itself. one giving a 2-projector. in this mathematical model. and the essential properties of the 2-projector as shown in Figure 71. This is the pattern of self-iteraction of the Fibonacci particle. depicted in Figure 72. As the reader can see from Figure 72. we show the essence of the Temperley-Lieb recoupling model for the Fibonacci particle. the other giving nothing. Fibonacci sequence. Rather than start there.312 LOUIS H. identiﬁed with the 2-projector itself. This fact is illustrated in Figure 70. we will assume that the 4-projector is forbidden and deduce (below) that the theory has to be at this root of unity. its properties and evaluations based on the bracket polynomial model for the Jones polynomial. The dimension formula for these spaces follows from the fact that there are fn sequences of length n − 1 of marked and unmarked states with no repetition of an unmarked state. In Figure 72. ** §16.

We will discuss the resulting algebra below. . In Figures 73–78 we have provided complete diagrammatic calculations of all of the relevant small nets and evaluations that are useful in the two-strand theory that is being used here. with a solid strand corresponding to the marked particle. A dark vertex indicates either an interaction point. Forbidden Process = Figure 72. or it may be used to indicate the single strand is shorthand for two ordinary strands. Note that in Figure 72 we have adopted a single strand notation for the particle interactions. here is a guided tour. Remember that these are all shorthand expressions for underlying bracket polynomial calculations. The 2-projector. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 313 = − 1/δ = 0 = = − 1/δ = −(1/δ)δ = 0 = − 1/δ = Figure 71. Figure 73 illustrates three basic nets in case of two strands. a dotted strand (or nothing) corresponding to the unmarked particle. Fibonacci particle as 2-projector. The reader may wish to skip directly to Figure 79 where we determine the form of the recoupling coeﬃcients for this theory. For the reader who does not want to skip the next collection of ﬁgures.

delta and tetrahedron nets. the delta by Δ. = = = −1/δ = 0 Figure 74. The theta net is denoted Θ. KAUFFMAN These are the theta. delta and tetrahedron. The Tetrahedron net will be similarly decomposed in Figure 77 and Figure 78. . In this ﬁgure we have shown the decomposition on the theta and delta nets in terms of 2-projectors. In Figure 74 we illustrate how a pedant loop has a zero evaluation. Note how.314 LOUIS H. LoopEvaluation – 1. and A is the bracket polynomial parameter. Θ = = = Δ = = Τ = Figure 73. The ﬁgure begins with a calculation of the result of closing a single strand of the 2-projector. Figure 76 illustrates the explicit calculation of the delta and theta nets. In Figure 75 we use the identity in Figure 74 to show how an interior loop (formed by two trivalent vertices) can be removed and replaced by a factor of Θ/Δ. We then ﬁnd that Δ = 2 − 1 and Θ = ( − 1/)2 − Δ/ = ( − 1/)( 2 − 2). and the tetrahedron by T . Theta. The result is a single stand multiplied by ( − 1/) where = −A2 − A−2 . line two proves that one network is a multiple of the other. in this ﬁgure. while line three determines the value of the multiple by closing both nets.

KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 315 = = = Θ/Δ = x +y = x = x Θ =xΔ = x x = Θ/Δ Figure 75. At the end of Figure 78 we obtain the formula for the tetrahedron T = ( − 1/)2 ( 2 − 2) − 2Θ/. and simpliﬁed. LoopEvaluation – 2. Calculate Theta. . The reader should note the ﬁrst line of Figure 77 where the tetradedral net is translated into a pattern of 2-projectors. Figure 77 and Figure 78 illustrate the calculation of the value of the tetrahedral network T . = − 1/δ = (δ − 1/δ) Δ = = (δ − 1/δ) = (δ − 1/δ) δ Δ = δ2 − 1 Θ = = − 1/δ Θ = (δ − 1/δ) 2 δ − Δ/δ Figure 76. Delta. The rest of these two ﬁgures are a diagrammatic calculation. using the expansion formula for the 2-projector.

316 LOUIS H. Figure 79 is the key calculation for this model. In this ﬁgure we assume that the recoupling formulas involve only 0 and 2 strands. Τ = − (1/δ) (δ − 1/δ) 2 δ − Θ/δ = − 1/δ 2 − (δ − 1/δ) − Θ/δ 3 2 = (δ − 1/δ) δ − (1/δ)Θ − (δ − 1/δ) − Θ/δ 2 = (δ − 1/δ) (δ 2 − 2) − 2Θ/δ Figure 78. with 0 corresponding to the null particle and 2 corresponding to the 2-projector. KAUFFMAN Τ = = = = − 1/δ = − Θ/δ = − 1/δ − Θ/δ = − (1/δ) (δ − 1/δ) 2 δ − Θ/δ Figure 77. (2 + 2 = 4 is forbidden as in Figure 72. Calculate tetrahedron – 1.) From this assumption we calculate that the recoupling matrix is given by a b 1/Δ Δ/Θ F = = c d Θ/Δ2 T Δ/Θ2 Figure 80 and Figure 81 work out the exact formulas for the braiding at a three-vertex in this theory. When the 3-vertex has three marked lines. Calculate Tetrahedron – 2. then the .

then the braiding operator is multiplication by A8 . braiding operator is multiplication by −A4 . Recoupling for 2-projectors. + + = -1 = A = -- = −1/δ + + +(2/δ 2 ) 4 A -1 = -A 3 = -A Figure 80. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 317 = a + b = c + d = a a = 1/Δ Θ = b Θ 2 /Δ = b b = Δ/Θ = c c = Θ/Δ 2 = d d = Τ Δ/Θ 2 Figure 79. Notice that it follows from the symmetry of the diagrammatic recoupling formulas of Figure 79 that the square of the recoupling matrix F is equal to the . Braiding at the three-vertex. as in Figure 58. When the 3-vertex has two marked lines. as shown in Figure 81.

That is. a quadratic equation whose solutions are √ Δ = (1 ± 5)/2. We then take A = e 3 i/5 . We shall now specialize to the case where √ Δ = = (1 + 5)/2. Hence Δ2 = Δ + 1 = 2 .318 LOUIS H. identity. Furthermore. This is equivalent to saying that Δ2 = 1 + Δ. leaving the other cases for the exploration of the reader. KAUFFMAN = = = − 1/δ 3 − 1/δ = -A 3 − 1/δ = -A = 6 = A − 1/δ = -A3 8 2 = A ( − 1/δ ) = A -4 + (1 . Braiding at the null-three-vertex. Θ/Δ3 + T/(ΔΘ) 1/Δ + Δ2 T 2 /Θ4 Thus we need the relation 1/Δ + 1/Δ2 = 1.A ) 8 = A Figure 81. we know that Δ = 2 − 1 from Figure 76. 1 0 2 1/Δ Δ/Θ 1/Δ Δ/Θ =F = 0 1 Θ/Δ2 T Δ/Θ2 Θ/Δ2 T Δ/Θ2 1/Δ2 + 1/Δ 1/Θ + T Δ2 /Θ3 = .

The ﬁnal adjustment. becomes 1/Δ Δ/α 2 Θ α 2 Θ/Δ2 −1/Δ For symmetry we require Δ/(α 2 Θ) = α 2 Θ/Δ2 . We take √ α2 = Δ3 /Θ. With this choice of α we have √ √ Δ/(α 2 Θ) = ΔΘ/(Θ Δ3 ) = 1/ Δ. Since the Θ has two vertices. so long as Δ2 = Δ + 1. Other properties of the model will remain unchanged. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 319 so that √ = −A2 − A−2 = −2 cos(6 /5) = (1 + 5)/2. Consider the result of replacing each trivalent vertex (with three 2-projector strands) by a multiple by a given quantity α. Note that T = −Θ2 /Δ2 . √ This proves that we can satisfy this model when Δ = = (1 + 5)/2. A ﬁnal adjustment of the model gives this desired symmetry. the tetradhedron T will be multiplied by α 4 . The new recoupling matrix. Hence the new symmetric F is given by the equation √ √ F = 1/Δ √ 1/ Δ = √ 1/ Δ −1/Δ − . Our last version of F suﬀers from a lack of symmetry. and hence not unitary. Similarly. Thus Θ = ( − 1/)2 − Δ/ = − 1. from which it follows immediately that F 2 = I. after such an adjustment is made. it will be multiplied by α 2 . For this specialization we see that the matrix F becomes 1/Δ Δ/Θ 1/Δ Δ/Θ 1/Δ Δ/Θ F = = = Θ/Δ2 T Δ/Θ2 Θ/Δ2 (−Θ2 /Δ2 )Δ/Θ2 Θ/Δ2 −1/Δ This version of F has square equal to the identity independent of the value of Θ. The Δ and the will be unchanged. It is not a symmetric matrix. and T = ( − 1/)2 ( 2 − 2) − 2Θ/ = ( 2 − 2) − 2( − 1)/ = ( − 1)( − 2)/ = 3 − 5. Note that − 1/ = 1.

0) ( Δ a) 2 a = 0 if b = 0 b Figure 82. we have that the local braiding matrix for the model is given by the formula below with A = e 3 i/5 . This gives the Fibonacci model. The matrices S1 and S2 are both unitary. Evaluation of the plat closure of a braid.y) x y x . In that ﬁgure we illustrate the calculation of the evalutation of the (a)-colored bracket polynomial for the plat closure P(B) of a braid B. In this section we make some brief comments on the quantum computation of colored Jones polynomials.0) 0 0 0 0 = B(0.).y 0 a a a a 0 0 0 a a = = Σ x . 0 A8 0 −e 2 i/5 The simplest example of a braid group representation arising from this theory is the representation of the three strand braid group generated by S1 = R and S2 = FRF (Remember that F = F T = F −1 . and they generate a dense subset of the unitary group U (2). supplying the ﬁrst part of the transformations needed for quantum computing. Quantum computation of colored Jones polynomials and the Witten- Reshetikhin-Turaev invariant. First. consider Figure 82. This material will be expanded in a subsequent publication. Using Figure 80 and Figure 81. §17. The reader can infer the deﬁnition of the plat closure from Figure 82. KAUFFMAN where Δ is the golden ratio and = 1/Δ. . 4 i/5 −A4 0 e 0 R= = .y B(x.y) x y 0 a a a a 0 0 = B(0. B P(B) a a a Σ a = = B(x.320 LOUIS H.

0) = Σx1 . We then see that we can use our discussion of the Temperley-Lieb recoupling theory as in sections 14. .. .. we therefore can consider quantum computation of the matrix entries B(0. . . We can obtain these matrix entries by using the Hadamard test as described in section 11. In the general case we will get n/2 < PB >a = B(0..a and take the case of the action on the vector v whose process space coordinates are all zero. In order to consider quantumn computation of the colored bracket or colored Jones polynomials. 0). . xn )v(x1 . KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 321 One takes a braid on an even number of strands and closes the top strands with each other in a row of maxima. Here n is even and equal to the number of braid strands.a . xn ) runs over a basis for the space V0a.. ..xn B(x1 . These matrix entries in the case of the roots of unity A = e i /2r and for the a = 2 Fibonacci model with A = e 3i /5 are parts of the diagonal entries of the unitary transformation that represents the braid group on the process space V0a. . . .a. . Then... . . . We reserve discussion of these issues to a subsequent publication. and it is proved in the case a = 2 in Figure 74. .. . when we close the top of the braid action to form PB. the bottom strands are closed with a row of minima.a . . as the ﬁgure shows.15 and 16 to compute the value of the colored bracket polynomial for the plat closure PB. Similarly.. . . diﬀering just by a normalization factor. .. The calculation simpliﬁes to this degree because of the vanishing of loops in the recoupling graphs. . xn ) where B(x1 . It is not hard to see that any knot or link can be represented as the plat closure of some braid. . The colored Jones polynomials are normalized versions of the colored bracket polymomials. is the evaluation of that link where each single strand has been replaced by a parallel strands and the insertion of Jones-Wenzl projector (as discussed in Section 14). Note that in this ﬁgure we indicate the action of the braid group on the process spaces corresponding to the small trees attached below the braids.. but for braids of arbitrary size..a. The vanishing result is stated in Figure 82. . 0)Δa . Then the action of the braid takes the form Bv(0... . . in essentially the same framework as we described in section 11. . we cut the sum down to the evaluation of just one term.a. . denoted < L >a . The computational complexity of these models is essentially the same as the models for the Jones polynomial discussed in [3]. The (a)-colored bracket polynonmial of a link L. xn ) denotes the matrix entries for this recoupling transforma- tion and v(x1 .. .. . As shown in Figure 82. . we regard the braid as acting on a process space V0a. As a result we get relatively eﬃcient quantum algorithms for the colored Jones polynonmials at these roots of unity. In the ﬁgure we illustrate with n = 4.

) 8 = A Figure 83.322 LOUIS H. Above the box. a quantum algorithm for the computation of the Witten functional integral [95] via this knot-theoretic combinatorial topology. and the Fibonacci model can be used to perform quantum computation for the values of this invariant. See Figure 83 for diagammatics that resolve this fact. It would be very interesting to understand a more direct approach to such a computation via quantum ﬁeld theory and functional integration. Finally. A -4 )( . but also for computing the Witten-Reshetikhin-Turaev (WRT) invariants at the above roots of unity. This means that we have. A direct construction of the Fibonacci model. in principle. Dubrovnik polynomial specialization at two strands. §18. we give elementary constructions for unitary representations of the three strand . It is worth remarking here that these algorithms give not only quantum algorithms for computing the colored bracket and Jones polynomials. the (2)-colored bracket polynomial is a special case of the Dubrovnik version of the Kauﬀman polynomial [38]. in unnormalized form is given as a ﬁnite sum of colored bracket polynomials: WRT(L) = Σr−2 a=0 Δa < L >a . we show how the double strands with projectors reproduce this relation. ) . the natural underlying knot polynomial is a special evaluation of the Dubrovnik polynomial. KAUFFMAN 4 -4 = A +A +δ -4 4 =A + A +δ - 4 = (A . A -4 )( . The reason for this is that the WRT invariant. = (A 4 . we note that in the case of the Fibonacci model. and so the same computation as shown in Figure 82 applies to the WRT. This observation means that in the Fibonacci model. In section 10 of this paper. The skein relation for the Dubrovnik polynomial is boxed in this ﬁgure.

In section 11 we show how to use unitary representations of the three strand brand group to devise a quantum computation for the Jones polynomial. without using the recoupling theory that we have explained in the previous sections of the paper. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 323 braid group in U (2). These rules of interaction are illustrated in Figure 68. Local braiding. and show how to construct the Fibonacci model by elementary means. In this section we return to these considerations. This ﬁnal approach is signiﬁcant in that it shows an even closer relationship of the Fibonacci model with the Temperley Lieb algebra representation associated with the Jones polynomial. Recall that in the Fibonacci model we have a (mathematical) particle P that interacts with itself either to produce P or to produce a neutral particle ∗. Figure 70 and Figure 84. While we do not assume the recoupling theory of the previous sections. The Fibonacci particle P. Thus ∗ acts as an identity trasformation. we essentially reconstruct its patterns for the particular purposes of the Fibonacci model. If X is any particle then ∗ interacts with X to produce X . . Figure 69. P P P P P * P * P * P * * * * Figure 84. The constructions in this section are based on the combinatorics of the Fibonacci model. P P P P = μ * * P P P P = λ P P Figure 85.

The change of basis is shown in Figure 88 and has matrix F as shown below. |P} for this space of particle interactions. In particular. then the braiding gives a phase factor of . This process space has dimension two and can support a second braiding generator for the second two strands on the top of the tree. The possible pathways are illustrated in Figure 86. If P interacts to produce ∗. We want to make this braiding matrix part of a larger representation of the braid group. One should visualize these particles as moving in a plane and the diagrams of interaction are either creations of two particles from one particle. In Figure 85 we illustrate braiding of P with itself in relation to the two possible interactions of P with itself. In order to articulate the second braiding we change basis to the process space corresponding to P(PP) as shown in Figure 87 and Figure 88. If P interacts to produce P. This space starts with three P particles and considers processes associated in the patttern (PP)P with the stipulation that the end product is P. We assume at the outset that and are unit complex numbers. Thus we have a braiding matrix for these “local” particle interactions: 0 R= 0 written with respect to the basis {|∗.324 LOUIS H. we want a representation of the three-strand braid group on the process space V3 illustrated in Figure 6. They correspond to (PP)P −→ (∗)P −→ P and (PP)P −→ (P)P −→ P. We want a unitary representation of three-strand braids so that (. or fusions of two particles to a single particle (depending on the choice of temporal direction). then the braiding gives a phase factor of . KAUFFMAN The braiding of two particles is measured in relation to their interaction.

1 ) = R and (.

The symmetry of the change of basis formula essentially demands that F 2 = I . a b F = b −a where a 2 + b 2 = 1 with a and b real. If F is real. We begin by noting that 0 0 − 0 0 −1 0 R= = + = + 0 0 0 0 0 0 0 0 where = ( − ). Thus. symmetric and F 2 = I . This form of the matrix for the basis change is determined by the requirement that F is symmetric with F 2 = I . Now we try to simultaneously construct an F and construct a representation of the Temperley-Lieb algebra. See Figure 88. . Since R is unitary we see that S = FRF is also unitary. For the Temperley-Lieb representation. Thus R = I + −1 U where U = so that 0 0 U 2 = U . then F is unitary.2 ) = S = F −1 RF . we want = −2 − −2 . if F is constructed in this way then we obtain a unitary representation of B3 . We take the form of the matrix F as follows.

b −a 0 0 b −a ab b2 Thus V 2 = V and since V = |vv| and U = |ww| with w = (1. we have the Temperley-Lieb representation and the corresponding unitary braid group representation for 2-strand braids and the 2-strand Temperley-Lieb algebra. P P P P P P P P P F P a * + b * P P P P P P P P P P P P F P b + -a P * P P P Figure 87. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 325 Hence we need −2 − −2 = ( − ). Recoupling formula. Now we can go on to B3 and TL3 via S = FRF = I + −1 V with V = FUF . Similarly UVU = 2 a 2 U . We must examine V 2 . We ﬁnd that V 2 = FUFFUF = FU 2 F = FUF = V. UVU and VUV . With this restriction on . b)T (T denotes transpose). we need 2 a 2 = 1 and so we shall take a = −1 . Thus. 0)T and v = Fw = (a. we see that VUV = 3 |vv|ww|vv| = 3 a 2 |vv| = 2 a 2 V. we have a representation of the Temperley-Lieb algebra TL3 so that . P P P |x> : |*> or |P> x P Figure 86. With this choice. which implies that = −−3 . as desired and a b 0 a b a2 ab V = FUF = = . Three strands at dimension two.

1 = AI + A−1 U and .

2 = AI + A−1 V√gives a unitary representation of the braid group when A = = e i and b = 1 − −2 is real. 4 which is satisﬁed for inﬁnitely many values of in the ranges [0. 2 /3] ∪ [5 /6. This last reality condition is equivalent to the inequality 1 cos 2 (2) ≥ . . 7 /6] ∪ [4 /3. /6] ∪ [ /3. 5 /3].

V = = . With these choices we have √ a b √ 1/ 1 − −2 F = = b −a 1 − −2 −1/ real and unitary. 2 0 a ab a b U = . Here we have .326 LOUIS H. 0 0 ab b 2 b b 2 Now examine Figure 89. |P}. Change of basis. KAUFFMAN P P P P P P R x λ(x) P x P P P P P P P F a4 P -1 S = F RF R P P P P P P -1 F P P Figure 88. Here we illustrate the action of the braiding and the Temperley-Lieb Algebra on the ﬁrst Fibonacci process space with basis {|∗. and for the Temperley-Lieb algebra.

.1 = R.

2 = FRF and U1 = U. we have arrived at exactly the 3-strand braid representations that we used in our papers [44. In this paper we are working in the context of the Fibonacci process spaces and so we wish to see how to make a representation of the Temperley-Lieb algebra to this model as a whole. not restricting ourselves to only three strands. So far. The generic case to consider is the . 56] giving a quantum algorithm for the Jones polynomial for three-strand braids. U2 = V as described above. Thus we have a representation of the braid group on three strands and a representation of the Temperley-Lieb algebra on three strands with no further restrictions on .

In Figure 91 we have illustrated the triplets from the previous ﬁgure as part of a possibly larger tree and have drawn the strings horizontally rather than diagonally. Here we have denoted this action as U3 because it connotes the action on the third and fourth vertical strands in the sequences shown in Figure 91. Note that in a larger sequence we can recognize Uj by examining the triplet surrounding the j − 1-th element in the sequence. Multiply by δ. Use V. * Use λ. P Use F. U3 | ∗ PP = 0. U3 |PPP = b|P ∗ P + b 2 |PPP. In this ﬁgure we have listed the eﬀects of braiding the vertical strands 3 and 4. U3 | ∗ P∗ = | ∗ P∗. Use V. P P P P P |xyz>: |PPP> |P *P > x |* P* > y |*PP > z |PP* > P Figure 90. action of the Temperley-Lieb algebra on process spaces of higher dimension as shown in Figure 90 and Figure 91. Multiply by 0. * Use F. A ﬁve dimensional process space. We see from this ﬁgure that the action of the Temperley-Lieb algebra must be as follows: U3 |P ∗ P = a|P ∗ P + b|PPP. P Figure 89. Algebra for a two dimensional process space. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 327 Two Dimensional Process Space P P P |x> P x Braiding Temperley-Lieb Use μ. U3 |PP∗ = 0. just as the pattern above is .

Note that in a sequence for the Fibonacci process there are never two consecutive appearances of the neutral element ∗. we have only indicated three elements in the sequences above. * P * |* P P> Use λ. Algebra for a ﬁve dimensional process space. Thus we see that in order for U4 U3 U4 = U4 . It is convenient to leave out the ﬂanking P’s when notating the sequence. * P P |P P *> Use λ. P P * Figure 91. governed by the elements surrounding the second element in the sequence. it is convenient to assume that the ends of the sequence are ﬂanked by P as in Figure 90 and Figure 91 for sequences of length 3. In working with this representation of the braid group and Temperley-Lieb algebra. Multiply by 0. . Using these formulas we can determine conditions on such that this is a representation of the Temperley-Lieb algebra for all Fibonacci sequences. P P P |* P *> Use μ. we need that 2 b 4 = 1.328 LOUIS H. Consider the following calculation: U4 U3 U4 |PPPP = U3 U2 (b|PP ∗ P + b 2 |PPPP) = U4 (bU3 |PP ∗ P + b 2 U3 |PPPP) = U4 (0 + b 2 (b|P ∗ PP + b 2 |PPPP) = b 2 (bU4 |P ∗ PP + b 2 U4 |PPPP) = 2 b 4 U4 |PPPP. KAUFFMAN Five Dimensional Process Space |xyz> x y z Braiding Temperley-Lieb |P * P> Use F. Multiply by 0. P * P |PPP> Use F. Thus |PP ∗ P ∗ P ∗ P is a Fibonacci sequence. Use V. Multiply by δ. For simplicity. We shall refer to a sequence of ∗ and P as a Fibonacci sequence if it contains no consecutive appearances of ∗. Use V.

U1 |Px2 x3 . . . xn . Then the Temperley-Lieb algebra on n + 2 strands with loop value acts on Vn via the formulas given below. xi−3 PPPxi+1 . Ui |x1 . . . xn } where each xi equals either P or ∗ and there do not occur two consecutive appearances of ∗ in the sequence {x1 . √ When − 1/ = 1. . . Theorem 6 (Fibonacci Theorem). xn + b|PPx3 . . Note that 2 b 4 = 2 (1 − −2 )2 = ( − 1/)2 . U2 | ∗ Px3 . . . Ui |x1 . Ui |x1 . xn . . xn = b| ∗ Px3 . . xn = 0. . we have the solutions = 1± 5 . We refer to this basis for Vn as the set of Fibonacci sequences of length n. . . When √ − 1/ = −1. Then the dimension of Vn is equal to fn+1 where fn is the n-th Fibonacci number: √ f0 = f1 = 1 and fn+1 = fn + fn−1 . Let Vn+2 be the complex vector space with basis {|x1 x2 . However. Let a = 1/ and 1+ 5 √ b = 1 − a 2 . xi−3 ∗ P ∗ xi+1 . . . . . xn = 0. U2 |P ∗ x3 . . . xn = a| ∗ Px3 . . . . First we give the left-end actions. U2 |PPx3 . . xi−3 P ∗ Pxi+1 . This leaves only = ±φ √ where φ = 1+2 5 (the Golden Ratio) as possible values for that satisfy the reality condition for F . xi−3 PPPxi+1 . . . . . . . U1 | ∗ x2 x3 . xn . . xn = 0. Thus. Ui |x1 . . . . . xn . xn = b|x1 . xn + b 2 |x1 . ruling out the choice = 1−2 5 . . . . xi−3 P ∗ Pxi+1 . xn = 0. xn . . up to a sign we have arrived at the well-known value of = φ (the Fibonacci model) as essentially the only way to have an extension of this form of the representation of the Temperley-Lieb algebra for n strands. . xn . xn = | ∗ x2 x3 . . . . . . . xi−3 P ∗ Pxi+1 . Let’s state this positively as a Theorem. . we have the solutions = −1±2 5 . Let = ±φ where φ = 2 . . . . xi−3 PPPxi+1 . . . . . xn = |x1 . for the reality 2 √ −2 of F we require that 1 − ≥ 0. xn + b|x1 . . xn = a|x1 . xi−3 ∗ PPxi+1 . . . xn + b 2 |PPx3 . . Thus we require that − 1/ = ±1. xn }. xi−3 ∗ P ∗ xi+1 . . KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 329 It is easy to see that 2 b 4 = 1 is the only remaining condition needed to make sure that the action of the Temperley-Lieb algebra extends to all Fibonacci Model sequences. . . Then we give the general action for the middle strands. . xi−3 PP ∗ xi+1 . . . . Ui |x1 . .

Note that the left and right end Temperley-Lieb actions depend on the same basic pattern as the middle action. . we give the right-end action. . The Fibonacci sequences |x1 x2 . . xn−2 P∗ + b 2 |x1 . . we have a unitary represen- tation of the Artin Braid group Bn+2 to TLn+2 . . Un+1 |x1 . xn should be regarded as ﬂanked left and right by P’s just as in the special cases discussed prior to the proof of the Fibonacci Theorem. .330 LOUIS H. KAUFFMAN Finally. Corollary 1. xn−2 ∗ P = 0. Un+1 |x1 . . . . xn−2 P∗ = 0. xn−2 PP = b|x1 . With the hypotheses of Theorem 2. Remark 10. . . : Bn+2 −→ TLn+2 given by the formulas (. Un+1 |x1 . xn−2 PP. .

i ) = AI + A−1 Ui . (.

The local braiding matrix is given by the formula below with A = e 3 i/5 . In the original Fibonacci model [60]. 8 4 i/5 A 0 e 0 R= = . The Theorem and Corollary give the original parameters of the Fibonacci model and shows that this model admits a unitary representation of the braid group via a Jones representation of the Temperley-Lieb algebra. 0 −A4 0 −e 2 i/5 √ This is exactly what we get from our method by using = 1+2 5 and A = e 3 i/5 . Remark 11. √ √ 1/ √ 1/ √ F = = 1/ −1/ − √ where = 1+2 5 is the golden ratio and = 1/. where A = e 3 i/5 where the Ui connote the representation of the Temperley-Lieb algebra on the space Vn+2 of Fibonacci sequences as described in the Theorem above. Just as we have explained earlier in this paper. there is a basic non-trivial recoupling matrix F .i−1 ) = A−1 I + AUi . the simplest example of a braid group representation arising from this theory is the representation of the three strand braid group generated by .

1 = R and .

2 = FRF (Remember that F = F T = F −1 . The matrices .).

1 and .

we gave . In our earlier paper [60] and in the previous sections of the present work. and they generate a dense subset of U (2). supplying the local unitary transformations needed for quantum computing. The full braid group representation on the Fibonacci sequences is computationally universal for quantum computation.2 are both unitary.

Akutsu and M. 615–640. 2002. pp. Aharonov and I. 183–204.. 2 (2002/3). Soc. Freedman. Chen. Principles of Quantum Mechanics. Cambridge University Press. arXiv: 1112. Mathematical Proceedings of the Cambridge Philosophical Society. vol. Search for majorana fermions in superconductors. Quantum computation and the localization of modular functors. Jones. Amsterdam. Atiyah. 1998. Commuting actions – a tale of two groups. [13] B. Springer-Verlag. 1997. Topological views on computational complexity. 489–515. F. Unitary solutions to the yang-baxter equation in dimension four. Kauﬀman. Acad.1950. Categorical quantum mechanics. Universal quantum gates. Mathematics of Quantum Computation (R. Foundations of Computational Mathematics. 95 (2005). Quantum Relativity: A Synthesis of the Ideas of Einstein and Heisenberg. [3] D. Abramsky and B. Ernst and D.. vol. W. Chapman & Hall/CRC Press. Coecke. Handbook of Quantum Logic and Quantum Structures: Quantum Logic. 2009. J. 117–151. quant-ph/0605181. 53–59. Baxter. 1996. Press. quant-ph/0511096. UIUC. Chen. pp. [20] M. 4 pp. L. American Mathematical Society. S. [16] H. In this section. Lie algebras and their representations (Seoul 1995). vol. 108 (1990). Lomonaco (editors). Florida. Aharonov. 4/25/2005. 135 (1991). [11] J. Japan. Braid topologies for quantum computation. pp. [9] G. 453–464. pp. Brylinski and G. Elsevier/North-Holland. [7] R. 2006. Brylinski. Math. Oxford University Press. and S. Borromean entanglement of the GHZ state. quant-phy/0402014. 427–436. Realizing non-abelian statistics in time-reversal invariant systems. [5] P. Sumners. Contemp. pp. 1982. Benkart. REFERENCES [1] S. 1990. Arad. A. [2] D. Brylinski and R. J. Wadati. 2007. [8] C. 1–46. Bonesteel. Fradkin and P. The Geometry and Physics of Knots. H. R. 1958. A. pp. The logic of entanglement. [15] P. M. Beenakker. K. ACM. Hormozi. editors). vol. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 331 a construction for the Fibonacci model based on Temperely-Lieb recoupling theory. pp. editors). and S. A calculus for rational tangles: Applications to DNA recombination. Fendley. [18] D. Physics Department. quant-ph/0505665. 56 (1987). Chapman & Hall/CRC. Dirac. Landau. Physical Review Letters. quant-ph/0003128. [19] E. Exactly Solved Models in Statistical Mechanics. 2-d physics and 3-d topology. 1996. Cohen et al. The BQP-hardness of approximating the Jones polynomial. pp. pp. Extra Volume. [4] Y. Process. L. vol. Knot invariants and critical statistical systems. Phys. Theory Seminar. J. vol. [21] . 1 (2001). Zikos. [17] C. Series. Entanglement and Passion-at-a-Distance (R. Quantum Inf. Potentiality. Finkelstein. Documenta Mathematica. vol. W. Mathematics in Quantum Computation and Quantum Technology. Communications in Mathematical Physics. pp. Aravind. 194. Boca Raton. 839–842. Dye. STOC’06: Proceedings of the 38th Annual ACM Symposium on Theory of Computing (New York). [14] L. . 261–323. Simon. 140503. G. we have reconstructed the Fibonacci model on the more elementary grounds of the representation of the Temperley-Lieb algebra summarized in the statement of the Fibonacci Theorem and its Corollary. ICM. V. [12] G. Crane. E. arXiv:quant-ph/0211050. pp. F. Kluwer. Coecke. [6] M. and Z. A polynomial quantum algorithm for approximating the Jones polynomial. L. [10] N.

[27] V. [35] . Topological quantum computation with the universal R matrix for Ising anyons. Dallas.. p. With an appendix by F. 126 (1987). Amer. Amaroso. Reﬂexivity and foundations of physics. Garnerone. Fasc. [30] . braid group and link polynomials. vol. Kauﬀman (editor). Doebner and V. 118–164. Paciﬁc J. [36] . H. Math. 3. Ivanov. 1994. 114. Self-reference and recursive forms. 318 (1990). pp. World Scientiﬁc Pub. Providence. Dobrev.. Co. A. Kauﬀman.Turaev. 234 (2003). On knot invariants related to some statistical mechanics models.332 LOUIS H. Freedman. Communications in Mathematical Physics. Non-abelian statistics of half-quantum vortices in p-wave superconductors. 256–265. Math. Lie Theory and Its Applications in Physics VII (Soﬁa) (H. 311–334. 92. Goodman and H. Imperial College. [42] L. Space and time in computation. 1–110. pp. A. Knot logic. R. World Scientiﬁc Publishers. New invariants in the theory of knots. 335–338. Fourth Edition (2012). and Z. Topology.) (R. vol. editors). Series.. KAUFFMAN [22] . Hecke algebra representations of braid groups and link polynomials. vol. A magnetic model with a possible chern-simons phase. 12-14 July 2010 (Melville. Texas. RI. vol. H. 1985. and M. 587–603. [33] . [41] . 10 (1987). 1996. editor). P. pp. and S. 1 Feb 2000. [37] . AMS Contemp. [40] . pp. K. 26 (1987). London. 195–242. 395–407. Second Edition (1993). Math. A modular functor which is universal for quantum computation. Amer. American Institute of Physics Pub. State models and the Jones polynomial. [43] L. [38] . pp. N. The Interface of Knots and Physics.. 44–53. pp. Wang. 86 (2001). pp. vol. 268. 417– 471.Y. vol. Third Edition (2002). vol. Annals Studies. Heron Press. Temperley-Lieb Recoupling Theory and Invariants of Three-Manifolds. vol. Nov. [23] M. vol. editors). Monthly. Kauﬀman. World Scientiﬁc Pub. Search for Fundamental Theory. Religious Experience and Scientiﬁc Paradigms. quant-ph/0110060v1 9 Oct 2001. 527–553. AIP. Princeton University Press. AMS PSAPM. H. Jeﬀers. 51. Kauﬀman (editor). Quantum automata. vol. The VIIth Intenational Symposium Honoring French Mathematical Physicist Jean-Pierre Vigier. pp. An invariant of regular isotopy. H. G. vol. Knots and Physics. Knots and Applications. 605–622. vol. quant-ph/0601169. 227 (2002). Rowlands. Journal of Social and Biological Structures. 53–72. [28] D. 78 (1989). Communications in Mathe- matical Physics. D. Wenzl.. Sign and space. pp. [32] . [34] . [39] . Trans. pp. Georgiev. 2008. Marzuoli. [24] M. and Z. 1996. Stony Brook (New York). Math. 227 (2002). PhysComp ’94. Communications in Mathematical Physics. 1994. Inventiones mathematicae. [26] L. 137 (1989). 48–89. USSR. Freedman. [25] S. pp. vol. LOMI preprint E-3-87. [31] L. Proceedings of the Workshop on Physics and Computation. Soc. Rasetti. vol. Wang. Physical Review Letters. pp. The Yang-Baxter equations and invariants of links. pp. topology and discrete physics. Proceedings of the 1982 IASWR Conference. IEEE Computer Society Press. Annals of Mathematics. pp. Simulation of topological ﬁeld theories by quantum computers. A. Leningrad. 263–297. Kitaev. pp. Institute of Advanced Study of World Religions. American Mathematical Society. [29] V. 1991. 95 (1988). UK. M. Jones. 1994. quant-ph/0001071. F. 129–183. Larsen. pp. Statistical mechanics and the Jones polynomial. Steklov Institute. quant-ph/0001108v2. S. Knots and Applications (L. .

vol. 361. 68. 012034.1–173. vol. Pirich. Eigenforms. 2009 (S. 4 (2002). 452 (1996). 8057. 2004. [58] . vol. CASYS. Topological Quantum Information Theory. J. 80570T–1–80570T–15. 267–332. Donkor. hep-th/0505069. 12 -14 April 2004 (E. editor). H. Knot theory and a physical state of quantum gravity. 2002. R. Grossing. Liege. pp. Glafka-2004: non-commutative worlds. 81–95. R. Proceedings of the Royal Society of London A. New Journal of Physics. [45] . 2010. K. pp. vol. quant- ph/0407224. Quantum knots. J.QA/0105255. 6 (2004). pp. D. E. Braiding operators are universal quantum gates. May 2004). 16 (2007). [48] . 4(3) (2009). Kybernetes. and A. vol. April 2011 (H. knot polynomials and anyonic topological quantum computation. [55] . J. pp.C.1–73. Non-commutative worlds. [46] .. [62] . Pirich. editor). Opt. [65] . A. Proceedings of Symposia in Applied Mathematics. p. editor). pp. vol. Kauﬀman and T. [49] . V. Brandt. 5065–5080. January 3-4. [52] .GN/0410329. 2008. 2011. American Mathematical Society. editor). Handbook of Knot Theory (Menasco and Thistlethwaite. [60] . editors). pp. Bowden (ed. 5105. (in the Proceedings of the 2004 Byelorus Conference on Quantum Optics). Liko. Lie Theory and Its Applications in Physics VII. [53] L. [47] .). pp. 1–39. Fifth International Conference. The Fibonacci Model and the Temperley-Lieb Algebra. . Proceedings of Spie.47. Proceedings of Spie. 45 (2006). 139–146. R. The International Journal of Systems and Cybernetics. Journal of Physics: Con- ference Series. [57] . 129–150. [61] . vol. pp. H. Reﬂexivity and eigenform: The shape of process. pp. Entanglement criteria: Quantum and topolog- ical. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 333 [44] . [50] . A. vol. Quantum Computation and Infor- mation (S. 173. pp. vol. Quantizing knots groups and graphs.. B. 2005. Eigenform. Discrete physics and the dirac equation. Belgium (2001) (D. Soﬁa.18. Proceedings of ANPA 25”. q-deformed spin networks. 9 (2005). [54] L. Pub. p. Dubois. 233–318. Donkor. Kauﬀman and P. 101–137. pp. math. FL (E. Noyes. math. E. Topological quantum information theory. Teleportation topology. vol. and H. vol. 227–232. 218 (1996). vol. editors). E. Quantum computing and the Jones polynomial. Vienna Austria 11-13 November 2011 (G. Quantum entanglement and topological entanglement. Spie Proceedings. Time imaginary value. pp. vol.. Quantum Information and Computation. 23 (2006). 268–284. 627. vol. Lomonaco Jr. Physics Letters A. Brandt. [51] . Spin networks and quantum computation. 21-22 April. Amsterdam. [56] . pp.. 51–58. 2012. Proceedings of the AMS Short Course in Quantum Computation and Quantum Information held in Washington. 2003. editors). Knot Theory Ramiﬁcations. Heron Press. R63. [59] . Elsevier B. pp. and H. AMS CONM/305. pp. pp. vol. New Journal of Physics. American Mathematical Society. Kauﬀman and S. vol.. New Journal of Physics. Modern Phys. IOP Publishing. Spie Proceedings. Quantum Information and Computation IX. paradox sign and space. pp. (Short version in “Spin. 103–176. 73. editors). International Journal of Theoretical Physics. Kybernetes. [64] L. International J. Lomonaco Jr. EmerQuM11: Emergent Quantum Mechanics 2011 (Heinz von Foerster Congress). Discrete physics and the derivation of electromagnetism from the formalism of quantum mechanics. SPIE. 225–239. Lomonaco Jr. 22 (2008). Brandt. Quantum Information and Computation II. 1443–1470. Donkor. E. AIP Conference Proceedings. vol. 6 (2004). Pirich. Orlando. Computing Anticipatory Sys- tems. Knot diagrammatics. (Book in preparation). 34 (2005). Spectrosc. discrete processes and quantum processes. [63] . H. 2002. Classical and Quantum Gravity.

Kauﬀman.D. . Ph. pp. pp. vol. vol. Majorana. Proc. 805702–1– 805702–14. vol. P. 171–184. RI). http://arxiv. Schmeikal. arXiv. 1969. editors). H. and L. Nova Publishers. Introduction to topological superconductivity and Majorana ferminons. Math. Noabelions in the fractional quantum Hall eﬀect. Sympos. Donkor. Smolin. vol. H. Classical and quantum conformal ﬁeld theory. Marzuoli and M. Leijnse and K. Proceedings of Spie. Chuang. vol. [80] R. Topological computing for beginners. Nielsen and I. E. Zuo. Invariants of three manifolds via link polynomials and quantum groups. E. A. [84] M. L. Schumacher. Physical Review D. Bastin. arXiv:1206. 171–206. [82] N. 2007. http://www. R. Quantum knots and lattices. 85–115. 127 (1990). 2010. Cambridge University Press. Nuclear Physics B. Spin network quantum simulator. World Scientiﬁc Publishing Company. Rowlands. 8057. Flensberg. Physics Letters A. Spin networks and quantum gravity. Bakkers. [75] A. Kouwenhuven. [71] S. Kohno. pp. Kitaev. American Mathematical Society. Varela. vol. [87] B. and A. Brandt. [83] .. vol. pp. E. 68. vol. vol. arXiv: 1204. pp. 7 (2008).edu/∼preskill/ph219. Form dynamics.caltech. [76] G. vol. Math. Spie Proceedings. Amer. Zero to Inﬁnity: The Foundations of Physics. [88] B. 1998. 123 (1989). [86] P. 52 (1995). Turaev. Kauﬀman and D. vol. Rasetti. 210. vol. Lomonaco and L. [67] L. A symmetric theory of electrons and positrons. Quantizing braids and other mathematical structures: the general quantization procedure. (slide presentation). S. 5743–5759. J.334 LOUIS H. Pirich. M. thesis. AMS Translations of Mathematical Monographs. Correlation. [81] J. 14 (1937). [70] M. pp. 2011. 1–26. pp. 209–276. pp. 4 (1995). Journal of Quantum Information Processing. Anyons in an exactly solved model and beyond. Quantum Computation and Quantum Information. [79] M. 360 (1991). Communication. S. Quantum Theory and Beyond (T. Plissard. G. Primordial Space: Point-free Space and Logic Case. 131– 162. pp. 547–597. Soc. Moore and N. Mourik. vol. 177–254. Annals of Physics. 2012. [68] A. [69] T. Invariants of 3-manifolds derived from ﬁnite dimensional Hopf algebras. Appl. pp. Quantum Information and Computation IX. [77] G.iqi. 306 (2002). Preskill. E. Moore and N. Conformal Field Theory and Topology. Inventiones mathematicae. Uni- versity of Texas at Austin. Quantum Information Science and its Contributions to Mathematics (Providence. 124003. fall 2003). (Private conversation. pp. [74] E. Series on Knots and Every- thing.2792. Frolov. 79–87. Roetteles. April 2011 (H. Communications in Mathematical Physics. 2–111. Y. p. 2000. [73] . SPIE. 41. 321 (2006). [85] C. A. editor).org/abs/0805.0339. Technical report. 103 (1991). Caltech. Cambridge University Press. vol..cond-mat/0506438 v1 17 June 2005. Read. Semiconductor Science and Technology. [78] V. Journal of Knot Theory and its Ramiﬁcations. KAUFFMAN [66] L. Ribbon graphs and their invariants derived from quantum groups. Reshetikhin and V. M. Signatures of Majorana fermions in hybred superconductor-semiconductor devices. Penrose. [72] . pp. Journal of Social and Biological Structures. Kauﬀman and F. H. or a blueprint for quantum systems that do rope tricks. Seiberg. Angular momentum: An approach to combinatorial spacetime. 27 (2012).1736. Radford. Rovelli and L. pp. (1980). vol. and Complementarity. K. 1990. P. Quantum knots and mosaics. I Nuovo Cimento. R. 362–396. Communications in Mathematical Physics.

Freedman. P. E. Witten. pp. [92] G. Inc. George Allen and Unwin Ltd. [97] P. L. H. Topological quantum computing with only one mobile quasiparticle. London: Kegan Paul. KNOT LOGIC AND TOPOLOGICAL QUANTUM COMPUTING 335 [89] P. Spencer-Brown. World Scientiﬁc Publishing Company. Trench. Fractional Statistics and Anyon Superconductivity. Wocjan and J. and L. Communications in Mathe- matical Physics. and M. 1990. 351–399. 1091–1112. 1312. STATISTICS AND COMPUTER SCIENCE (M/C 249) UNIVERSITY OF ILLINOIS AT CHICAGO 851 SOUTH MORGAN STREET CHICAGO. [91] L. 121 (1989). 1969. pp. G. 19 (1967). Yard. 31 (1992). 96 (2006). Wilczek. State sum invariants of 3-manifolds and quantum 6j symbols. Modern Physics Letters A. pp. Shor and S. Yang. [99] Y. W. vol. [90] S. 4 (1989). 8 (2008). H. vol. DEPARTMENT OF MATHEMATICS. 070503. Turaev and O. Wittgenstein. p. Simon. Some exact results for the many-body problem in one dimension with repulsive delta-function interaction. Laws of Form.. [96] L. ILLINOIS 60607-7045 E-mail: kauﬀman@uic. Quantum Information Processing. Estimating Jones polynomials is a complete problem for one clean qubit. Physical Review Letters. Tractatus Logico-Philosophicus. pp. [93] V. Brace and Company. Ltd.. Kauﬀman. 681–714. N. 4 (2005). vol. Link polynomials and critical points of the chern-simons path integrals. Jordan. Topology. M.edu . vol. [94] F. N. [95] E. Petrovic. vol.. H. 4 pp. pp. 865–902. Hormozi. Trubner and Co. vol. Yang-Baxterizations. 159–197. The Jones polynomial: quantum algorithms and applications in quantum complexity theory. Bonesteel. quant-ph/0509175. L. universal quantum gates and Hamiltonians. vol. Ge. Quantum Information & Computation. [98] C. Viro. Zhang. Physical Review Letters. N. quant-ph/0603069. Smolin. Quantum ﬁeld Theory and the Jones Polynomial. 1922. London. pp. New York: Harcourt..

.

13 box. 11 generators. 37–39 aﬃne map. 9 triangle inequalities. 259. 39 annihilation operator. 155. 310 biextensional collapse. 123. 230. 33. 13. 140. 161. 230. 140. 124. 320. 258. 232–234. 142. 164 biproduct. 273. 194 group. 323–326. 163. 140. 228. 252. 125–128. 320. 330 337 . 270. Bell measurement. 162 arrow. 72. sub. 247. 134. 266. 79 Bloch sphere. 20 bivalence. 328. 230. 256. 225–228. 201–203. 226. boson. 10. 293–295. 300. 74. 245. 284. 233. 125. 282. 269 Artin group. 307. 309. 267. 277. 139. 183–187 associativity. 21. 247. see also particle. 154. self-adjoint bimorphism. 229. 178. basic distributional vector space. 31. 25 braid. 267–269 Abelian group. 136 206. 279. 321. 324–326 basic state. 299–302. 178. 159. 219 Bell. see particle. 299. 307. 256. 8. 310 atomic Borel set. 144. 281. 129–131. 69 Alice & Bob. 261. 14 263. 153 Aspect. 272. 154. 158. 153. 163. 300. 330 bounded operator. Bell pairs. 108 167. subatomic 263. 34. 320–322 barrier penetration. 136–138. 33–35. 279 Boolean algebra. 165. 10. 8. 202 291. 179 312–314 colored. 13. 155. 144. 191 atoms (of a lattice). 134. 238–240. 252. 274. 178 275. 234. see also bilinear map. 130. 300. John Stewart. 135 Bell’s Theorem. 138. 225. 91 adjoint. Bell state. 138. 245. bipartite system. 299. 280. 270 block universe. 127. 226–230. 304. 321. 284. 224. 73 adjacency. 110 143. 281. 179. 143. 179 bracket polynomial. 9. 272. 278. 152. 232–234. 157–160. 72. 178 285–291 reasoning. 146–148. 12–14. biﬁbration. 231. 152. 183–186. 71. 280. 280–282. 143. 233. 250. 111. 112. 232–234. 108 ambient group. 9. 289 automated bra-ket notation. 191. 170 Bell’s inequality. 136 operator. 162 antimatter. 175 bra. 227. theorem proving. 298. 126. 165 anyon. proof checking. 214 284. 260. 270. 141 input-output. 224. 156. 13. theory exploration. 230. 215. 205. 178. 54 complete Boolean algebra. Alain. 143 adjunction. 281. 147 bit. 132. 258. INDEX Abelian algebra. 274. 272. 140 dual. 68. Boolean logic. 73 additive model. see also morphism Borromean rings. 207 bifunctor. 324 Bayesian inference. 292. 223. 17 self-adjoint. 259. 147 black hole. 264.

142. 36 planar. 136 clopen subobject.338 Index three-strand group. 96. 288 choice function. 47. 283. 114 coarse-graining. 225. 89. 176. 123. 93–98. 11. 115 clopen sets. 22 complement. 260. 125. 134. 208 dagger. 153. 185. 151. 18 complementary observables. 165 cocomplete. 112. 262. 111. 12 Carroll. 219 co-measurable physical quantities. 207. 330 tangle. 98–102. 291–293. 154. 255. 135 190 operational. 15 326 strict. 169. 323 tensor. 13. 94. 107. see also Hilbert space. 156 Chu space. 100. 285–287. 300. 22 CNOT gate. 31 classical computer. 21 complementarity. 88. 16. 153 braided. of modules. 266. 136. 201. 152 of convex sets. 304. 95–101. 279. theorem. 20 commutativity. 15 clock. 114 braiding operator. 15 293 homomorphisms of. 125. 176. 186 . 291. 105. 225. 186 classical perspective. 179. 158 dagger symmetric. 192–194 170 distributional. 135. 19 symmetric. 118 Brauer algebra. 160 Brouwer algebra. 147 coalgebra. 141 dagger. 226. 176. 21 category-theoretic. 15. 139 complex number. 50 Cartesian. 113. 25. 18. 296. 113. 20. 288. complex vector. 284–292. 112. 66. 209. 112. 105. 21. 208 complex matrix. 189. 62 dagger braided. 180. 261. 152. 159. 184–186. 270. ﬁnite-dimensional 329 of Hilbert spaces. 185. 15 universal gate. 104–111 of systems. 324. 159. 295 local. 117. 105–108. 45. 15 coherence. 123–125. 20. 153 complete. 200. 177. 189. 22 compact space. 21 commutant. 263–266 compact closed. 115. 162 autonomous. 167–170 dagger. 180–183. of vector spaces. 91. 135. 201 colimit. 316–318. 252. 311. traced. 154 category. 245. 210. conditions. 169. 52. 230. 106 coequaliser of algebras. 299 Kleisli. 258. 266. 21 composition. 139. 274. 272. 19. 263. 58. 139. 101. 280–282. 226. 20 completeness theorem. 21. 94 classical computation. 243. 136. 186. 190. 112. symmetric. 251. Lewis. 271–274. 114–116 261. 199–201. 18. 135. 266 traceable. 100. 14–16. 165. Cliﬀord algebraic representation. 284. 153–155. 63 cartesian product. 160. 303. 191 165. 15 parallel. 244–247. 116. 245. 170 left autonomous. 129. 15. 182. 136. 182. 223. 20. 16. 184 braiding. 108. 176 110–117. 229. 229. 326. 136 locally small. 102. 291– dual. 294. 179. 283–286. 296–300. dagger-compact. 22 commutator. 115. 280. 135. 155. 12 right autonomous. 152. 263. 181–183. 143 174. 177. 225. 135. 327. 223. 92. 16. 16. 135. 139. 126. 289. 157. 317 Cauchy-Schwarz inequality. 185 of sets. 320. 157. 175 of ﬁnite-dimensional Hilbert spaces. 142. 108 cobordism. 177. 192. 207. 182 304. 124. 176. 176 opposite. 159– compact. 16 category. 295 256 axioms for. 91–101. 22 179. 295. 162. small. 134. 19. 17–20. 89. 273. 68 monoidal. 157. 95.

214–216. 217. 138. 140. 132. Podolosky. 100. 167. 133. 141–143 convex set. 255 gauge. 200. 220 EPR. 204. 126 Dirac. 194 conditional probability. see also adjunction. 97. 127. 200. 147. 124. 83 Equivalence Principle. see context-based approach. 185 outer. 67