Classical and Quantum Nonlinear Integrable Systems

Theory and Applications

Edited by

A Kundu

Saha Institute of Nuclear Physics

Calcutta, India

Institute of Physics Publishing

Bristol and Philadelphia

Copyright © 2003 IOP Publishing Ltd.



c IOP Publishing Ltd 2003

All rights reserved. No part of this publication may be reproduced, stored in

a retrieval system or transmitted in any form or by any means, electronic,
mechanical, photocopying, recording or otherwise, without the prior permission
of the publisher. Multiple copying is permitted in accordance with the terms
of licences issued by the Copyright Licensing Agency under the terms of its
agreement with Universities UK (UUK).

British Library Cataloguing-in-Publication Data

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

ISBN 0 7503 0959 8

Library of Congress Cataloging-in-Publication Data are available

Commissioning Editor: Tom Spicer

Production Editor: Simon Laurenson
Production Control: Sarah Plenty
Cover Design: Victoria Le Billon
Marketing: Nicola Newey and Verity Cooke

Published by Institute of Physics Publishing, wholly owned by The Institute of

Physics, London
Institute of Physics Publishing, Dirac House, Temple Back, Bristol BS1 6BE, UK
US Office: Institute of Physics Publishing, The Public Ledger Building, Suite 929,
150 South Independence Mall West, Philadelphia, PA 19106, USA

Typeset by Sunrise Setting Ltd, Torquay, Devon, UK

Printed in the UK by MPG Books Ltd, Bodmin, Cornwall

Copyright © 2003 IOP Publishing Ltd.

 Contents

Preface

PART I
Classical Systems
1 A journey through the Korteweg–de Vries equation
M Lakshmanan
1.1 Introduction
1.2 Nonlinear dispersive waves: Scott Russell phenomenon and
solitary waves
1.2.1 KdV equation and cnoidal waves and the solitary waves
1.3 The Fermi–Pasta–Ulam (FPU) numerical experiments on
anharmonic lattices
1.3.1 The FPU lattice and recurrence phenomenon
1.4 The KdV equation again!
1.4.1 Asymptotic analysis and the KdV equation
1.5 Numerical experiments of Zabusky and Kruskal: the birth of
solitons
1.5.1 Periodic boundary conditions
1.5.2 Initial condition with just two solitary waves
1.6 Hirota’s bilinearization method: explicit soliton solutions
1.6.1 One-soliton solution
1.6.2 Two-soliton solution
1.6.3 N-soliton solutions
1.6.4 Asymptotic analysis
1.7 The Miura transformation and linearization of KdV: the Lax pair
1.7.1 The Miura transformation
1.7.2 Galilean invariance and the Schrödinger eigenvalue
problem
1.7.3 Linearization of the KdV equation
1.7.4 Lax pair
1.8 Lax pair and the method of inverse scattering
1.8.1 The IST method for the KdV equation
1.9 Explicit soliton solutions
1.9.1 One-soliton solution (N = 1)

Copyright © 2003 IOP Publishing Ltd.

 1.9.2 Two-soliton solution
1.9.3 N-soliton solution
1.9.4 Soliton interaction
1.9.5 Non-reflectionless potentials
1.10 Hamiltonian structure of KdV equation: complete integrability
1.10.1 KdV as a Hamiltonian dynamical system
1.10.2 Complete integrability of the KdV equation
1.11 Infinite number of conserved densities
1.12 Bäcklund transformations
1.13 The Painlevé property for the KdV equations
1.14 Lie and Lie–Bäcklund symmetries
1.15 Conclusion
2 The Painlevé methods
R Conte and M Musette
2.1 The classical programme of the Painlevé school and its
achievements
2.2 Integrability and Painlevé property for partial differential equations
2.3 The Painlevé test for ODEs and PDEs
2.3.1 The Fuchsian perturbative method
2.3.2 The non-Fuchsian perturbative method
2.4 Singularity-based methods towards integrability
2.4.1 Linearizable equations
2.4.2 Auto-Bäcklund transformation of a PDE: the singular
manifold method
2.4.3 Single-valued solutions of the Bianchi IX cosmological
model
2.4.4 Polynomial first integrals of a dynamical system
2.4.5 Solitary waves from truncations
2.4.6 First-degree birational transformations of Painlevé
equations
2.5 Liouville integrability and Painlevé integrability
2.6 Discretization and discrete Painlevé equations
2.7 Conclusion
3 Discrete integrability
K M Tamizhmani, A Ramani, B Grammaticos and T Tamizhmani
3.1 Introduction: who is afraid of discrete systems?
3.2 The detector gallery
3.2.1 Singularity confinement
3.2.2 The perturbative Painlevé approach to discrete integrability
3.2.3 Algebraic entropy
3.2.4 The Nevanlinna theory approach
3.3 The showcase
3.3.1 The discrete KdV and its de-autonomization

Copyright © 2003 IOP Publishing Ltd.

 3.3.2 The discrete Painlevé equations
3.3.3 Linearizable systems
3.4 Beyond the discrete horizon
3.4.1 Differential-difference systems
3.4.2 Ultra-discrete systems
3.5 Parting words
4 The dbar method: a tool for solving two-dimensional integrable
evolution PDEs
A S Fokas
4.1 Introduction
4.1.1 The dbar method
4.1.2 Coherent structures
4.1.3 Organization of this chapter
4.2 The KPI equation
4.3 The DSII equation
4.3.1 The defocusing DS equation
4.4 Summary
5 Introduction to solvable lattice models in statistical and mathematical
physics
Tetsuo Deguchi
5.1 Introduction
5.2 Solvable vertex models
5.2.1 The six-vertex model
5.2.2 The partition function and the transfer matrix
5.2.3 Diagonalization of the transfer matrix
5.2.4 The free energy of the six-vertex model
5.2.5 Critical singularity in the antiferroelectric regime near
the phase boundary
5.2.6 XXZ spin chain and the transfer matrix
5.2.7 Low-lying excited spectrum of the transfer matrix and
conformal field theory
5.3 Various integrable models on two-dimensional lattices
5.3.1 Ising model and Potts model
5.3.2 Chiral Potts model
5.3.3 The eight-vertex model
5.3.4 IRF models
5.4 Yang–Baxter equation and the algebraic Bethe ansatz
5.4.1 Solutions to the Yang–Baxter equation
5.4.2 Algebraic Bethe ansatz
5.5 Mathematical structures of integrable lattice models
5.5.1 Braid group
5.5.2 Quantum groups (Hopf algebras)
Appendix. Commuting transfer matrices and the Yang–Baxter equations

Copyright © 2003 IOP Publishing Ltd.

 PART II
Quantum Systems
6 Unifying approaches in integrable systems: quantum and statistical,
ultralocal and non-ultralocal
Anjan Kundu
6.1 Introduction
6.2 Integrable structures in ultralocal models
6.2.1 List of well-known ultralocal models
6.3 Unifying algebraic approach in ultralocal models
6.3.1 Generation of models
6.3.2 Fundamental and regular models
6.3.3 Fusion method
6.3.4 Construction of classical models
6.4 Integrable statistical systems: vertex models
6.5 Directions for constructing new classes of ultralocal
models
6.5.1 Inhomogeneous models
6.5.2 Hybrid models
6.5.3 Non-fundamental statistical models
6.6 Unified Bethe ansatz solution
6.7 Quantum integrable non-ultralocal models
6.7.1 Braided extensions of QYBE
6.7.2 List of quantum integrable non-ultralocal models
6.7.3 Algebraic Bethe ansatz
6.7.4 Open directions in non-ultralocal models
6.8 Concluding remarks
7 The physical basis of integrable spin models
Indrani Bose
7.1 Introduction
7.2 Spin models in one dimension
7.3 Ladder models
7.4 Concluding remarks
8 Exact solvability in contemporary physics
Angela Foerster, Jon Links and Huan-Qiang Zhou
8.1 Introduction
8.2 Quantum inverse scattering method
8.2.1 Realizations of the Yang–Baxter algebra
8.3 Algebraic Bethe ansatz method of solution
8.3.1 Scalar products of states
8.4 A model for two coupled Bose–Einstein condensates
8.4.1 Asymptotic analysis of the solution
8.5 A model for atomic–molecular Bose–Einstein condensation

Copyright © 2003 IOP Publishing Ltd.

 8.5.1 Asymptotic analysis of the solution
8.5.2 Computing the energy spectrum
8.6 The BCS Hamiltonian
8.6.1 A universally integrable system
8.6.2 Asymptotic analysis of the solution
9 The thermodynamics of the spin- 12 XXX chain: free energy and low-
temperature singularities of correlation lengths
Andreas Klümper and Christian Scheeren
9.1 Introduction
9.2 Lattice path integral and quantum transfer matrix
9.2.1 Mapping to a classical model
9.2.2 Bethe ansatz equations
9.3 Manipulation of the Bethe ansatz equations
9.3.1 Derivation of nonlinear integral equations
9.3.2 Integral expressions for the eigenvalue
9.4 Numerical results
9.5 Low-temperature asymptotics
9.5.1 Calculation to order O(β) and O(1)
9.5.2 O(1/β) corrections
9.5.3 O(1/β) corrections to the nonlinear integral equations
9.5.4 O(1/β) corrections to the eigenvalue
9.6 Summary and discussion
Appendix
10 Reaction–diffusion processes and their connection with integrable
quantum spin chains
Malte Henkel
10.1 Reaction–diffusion processes
10.2 Quantum Hamiltonian formulation
10.3 Hecke algebra and integrability
10.4 Single-species models
10.5 The seven-vertex model
10.6 Further applications
10.6.1 Spectral integrability
10.6.2 Similarity transformations
10.6.3 Free fermions
10.6.4 Partial integrability
10.6.5 Multi-species models
10.6.6 Diffusion algebras
10.7 Outlook: local scale-invariance

Copyright © 2003 IOP Publishing Ltd.

 Preface

Though the historic observation of the Great Wave of Translation by the

British naval engineer J Scott Russell in a canal of Edinburgh in 1834 may
be considered as the first recorded evidence in the investigation of nonlinear
dynamics governed by evolution equations like the KdV equation, the theory of
nonlinear integrable systems took considerably longer to reach its present stage
of accomplishment (see M Lakshmanan’s review in chapter 1). The modern age
of soliton physics perhaps starts from numerical experiments on such systems
as well as the formulation of analytical methods such as the inverse scattering
method (ISM) for exact solutions of them, based mainly on the works of
Zabusky–Kruskal, Gardner–Green–Kruskal–Miura (GGKM), Ablowitz–Kaup–
Newell–Segur (AKNS), Zakharov, Novikov, Manakov and others. However, the
foundation of this theory has already been laid by stalwarts like Liouville,
Poincaré, Painlevé and Kovalewskaya. The applications of this theory to diverse
fields, e.g. fluid dynamics, nonlinear optics, plasma physics, electrical network
and even biological systems, has attracted enormous and immediate interest in
this subject.
Another breakthrough came when integrable systems were raised to the
quantum level. Although the celebrated Bethe ansatz for exactly solving the
eigenvalue problem of the Hamiltonian for the isotropic Heisenberg spin chain
was introduced way back in 1931, its extension, application to other physical
models and true recognition took quite some time and were achieved gradually
through the pioneering works of Yang, Lieb, Wu, Mattis, Sutherland, Baxter
and others. However, a more general and powerful algebraic formulation of the
Bethe ansatz, which may also be considered the quantum ISM, was developed
mostly by Faddeev’s group in Leningrad exploiting the Yang–Baxter equation.
In later years, the deep connection of a host of other subjects, e.g. statistical
models, conformal field theory, quantum group, knot theory etc, with the theory
of quantum integrable systems and its exciting application to many problems in
condensed matter physics and other fields were revealed. The upsurge in work
in low-dimensional physics, stimulated by its connection with string theory and
the possible link with high-Tc superconductivity as well as the realization of
its exact solutions in such applicable fields as the reaction–diffusion equation,
cellular automata has triggered renewed interest in recent years to the theory

Copyright © 2003 IOP Publishing Ltd.

 and applications of integrable systems in a much wider and diverse sense. This,
in turn, has aroused the interest of chemists, biologists and other professionals
apart from physicists and mathematicians in this subject raising it almost to the
level of a scientific culture, the basic notions of which every scientist must know.
Therefore, there is a well-felt need for a collection of review articles covering the
basic and contemporary areas of this subject with updated information written
by specialists, active in their respective fields, but aimed primarily at scientists in
general.
The present collection is a modest attempt towards this goal. It aims to report
on the recent advances in the theory and applications of nonlinear integrable
systems in both classical and quantum domains. It tries to provide lucid exposition
of the basic theories, placing more emphasis on the underlying ideas rather than
the technicalities, as far as is possible in this sophisticated subject of mathematical
physics, and to indicate the challenging unsolved problems in this fascinating
field. This book on integrable systems consists of ten chapters, broadly divided
into two major parts—classical and quantum—each containing five reviews
devoted to these two broad divisions which cover various important aspects of
the subject.
The classical part opens with a pedagogical review ‘A journey through the
KdV equation’ by M Lakshmanan (Trichy, India). After a brief description of the
historical background and logical development of soliton physics, Lakshmanan
introduces the essential and powerful methods in the theory of classical integrable
systems illustrating them lucidly through the single example of the KdV equation.
Thus, we come to know about the first documentary evidence of an encounter with
the solitary wave and the historical Fermi–Pasta–Ulam computer experiment—
a cornerstone in the development of nonlinear theory—about the birth of the
soliton and its exact solution through Hirota’s bilinearization as well as by the
ISM. Continuing our journey, we learn more about other necessary aspects of
integrable systems like Lax pair formalism, Hamiltonian method, Lie and Lie–
Bäcklund symmetries of the nonlinear equation, Bäcklund transformation etc,
again from the simple example of the celebrated KdV equation. Identifying the
symmetries with independent conserved quantities, one gets to the origin of
the infinite number of involutive integrals of motion, the central notion of an
integrable system.
A deep analytic method, which is considered to be the most important
criterion for detecting the integrability of a nonlinear system, is known as the
Painlevé method. This method and the related properties are analysed in depth
in the second chapter ‘The Painlevé method’ by R Conte (Saclay, France) and
M Musette (Brussel, Belgium). The basic idea centres around the absence of
moving critical singularities in the general solution, which are related to singular
points for an ordinary differential equation (ODE) but to singular manifolds for
partial differential equations (PDE). The authors give the historical background by
taking us through the evolution of this idea and revealing glimpses of the conflicts
and collaborations of great minds in the creation of this beautiful method. In this

Copyright © 2003 IOP Publishing Ltd.

 systematic exposure, it is shown through numerous examples how the Painlevé
criterion can be applied through an algorithmic and step-by-step approach to
ODEs as well as PDEs, some of which are also of significant physical interest.
A contemporary direction in the development of classical integrable systems,
namely the integrability of discrete systems, is studied in detail in the next chapter
‘Discrete integrability’ by a team of authors (K M Tamizhmani (Pondicherry,
India), A Ramani (Palaiseau, France), B Grammaticos (Paris, France) and
T Tamizhmani (Karaikal, India)). The integrability criteria formulated for
differential equations (see chapter 2) must be generalized when dealing with
discrete systems. Four different such criteria, namely singularity confinement,
the perturbative Painlevé approach, algebraic entropy and Nevanlinna, are
introduced carefully, bringing out the basic ideas behind each of them together
with details of their application with simple examples. Important classes of
discrete equations including the discrete KdV and Painlevé equations along with
linearizable mapping, the differential difference equation and cellular automaton
are introduced and analysed.
A promising extension of the ISM for solving integrable evolution equations
in two special dimensions is known as the dbar method. The basic theory and
applications of this important method are presented in chapter 4 ‘The dbar
method: a tool for solving two-dimensional integrable evolution PDEs’ by A S
Fokas (Cambridge, UK), one of the pioneers in its development. Kadomtsev–
Petviashvili and Davey–Stewartson equations, the most celebrated integrable
equations defined on the plain, are analysed in detail to demonstrate how the
dbar method can be used systematically to extract the analogues of solitons in
two dimensions, namely the lump, the dromion and the line-soliton solutions
for these systems. The dbar method is also initiated on a simple example
of linearized Davey–Stewartson II equation, which though apparently a trivial
problem, provides an excellent pedagogical introduction for this involved method.
Progress in integrable classical statistical systems defined on a 2D lattice
is reviewed in chapter 5 ‘Introduction to solvable lattice models statistical and
mathematical physics’ by T Deguchi (Tokyo, Japan). Such important models as
the six- and eight-vertex models, the Ising, Potts and chiral Potts models and
the IRF and RSOS models etc are presented, focusing in particular on the six-
vertex model due to its prototypical nature—and still being the simplest one.
Introducing the coordinate Bethe ansatz for exact diagonalization of the transfer
matrix, its free energy is calculated which also helps to detect analytically the
critical singularity at the phase transition. Deguchi shows how the finite-size
analysis for this model gives information about the related conformal field theory,
how the Yang–Baxter equation is solved for the model through the algebraic Bethe
ansatz and its connection with rich mathematical structures such as quantum and
braid groups along with some new symmetries. The importance of the graphical
approach especially for integrable statistical models is emphasized.
Since 2D classical statistical models have a deep connection with 1D
quantum systems, this brings us logically to quantum integrable systems.

Copyright © 2003 IOP Publishing Ltd.

 The quantum part opens with the chapter ‘Unifying approaches in integrable
systems: quantum and statistical, ultralocal and non-ultralocal’ by A Kundu
(Calcutta, India). The aim of this review is to present the list of the, by now,
significant collection of quantum integrable models, both ultralocal and non-
ultralocal, in a systematic way, stressing their underlying algebraic structures. A
unifying scheme based on an ancestor model is presented for generating ultralocal
models along with their related statistical vertex models restricted to a 2 × 2 Lax
operator with a trigonometric and rational quantum R-matrix. The algebraic Bethe
ansatz formulation for the exact solution of the models has also been shown to
follow this unifying trend. Along with the known integrable models, possible
directions for investigations in this field and the generation of new models are
suggested. The ultralocal models are classified through their associated quantum
algebra and governed by the Yang–Baxter equation (YBE), while non-ultralocal
models, the theory of which is still in the stage of development, allow their
systematization through the braided extension of the YBE. It needs mentioning
that, unfortunately in most standard reviews, the class of non-ultralocal models,
which includes important models such as quantum KdV, the non-abelian Toda
chain, WZWN model etc, is generally ignored.
The subsequent chapters aim to focus mainly on important and contemporary
developments in the application of quantum integrable theory and to make contact
with the experiments, which are now becoming possible especially in condensed
matter physics due to technological advances.
Chapter 7 ‘The physical basis of integrable spin models’ by I Bose (Calcutta,
India), brings out this applicable aspect of integrable systems to physical models
and reviews systematically various exact results with experimental importance.
The main emphasis here is to reveal the basic concepts of the theory as well
as the applications to real systems without indulging much in the technical
details. After becoming acquainted with the idea of the coordinate Bethe ansatz
(BA) for solving the integrable spin chain, we get glimpses of the achievements
and central results of this spin model for both ferro- and antiferro-magnetic
interactions, linking them with real magnetic materials. Exact ground states,
low-lying excitations and dynamical correlation functions for the spin- 12 chain
are carefully introduced here, indicating their importance in neutron scattering
experiments. The significance of resonant valence bond states with spin–charge
separation in their excitations for high-Tc superconductors and their similarity
with the BA result for Luttinger liquid-like models are highlighted in an exciting
manner. Many other integrable models such as the Haldane–Shastry long-range
spin model, higher-spin quantum chains, Dzyaloshinskii–Morriya model together
with a few others like the AKLT model with partial exact results are presented,
together with their physical relevance. Finally, the spin and t–J ladder models,
both the integrable and partially solvable, are introduced and analysed. It is a
pleasure to mention that all models described here come with useful practical and
theoretical information and with an extensive reference list.

Copyright © 2003 IOP Publishing Ltd.

 The next chapter ‘Exact solvability in contemporary physics’ by A Foerster
(Porto Alegre, Brazil), J Links and H Q Zhou (Queensland, Australia), is
devoted to recent achievements in applying the algebraic Bethe ansatz for finding
exact results in some exciting fields of contemporary physics including those
in the nanoscale domain. Apparently at the nano-level, due to large quantum
fluctuations, the mean field approximation fails, enhancing the importance
of the exact treatment provided by integrable systems. Recounting the basic
theory of quantum integrable systems and algebraic Bethe ansatz, the review
shows how this procedure can be applied for the analysis of three important
models: Bose–Einstein (BE) condensates coupled via Josephson tunnelling, an
atomic–molecular BE condensation and the BCS pairing model, relevant for
superconducting metallic nanograins.
Chapter 9 ‘The thermodynamics of the spin- 12 XXX chain: free energy
and low-temperature singularities of correlation lengths’ by A Klümper and
C Scheeren (Dortmund, Germany) introduces in detail the theory and application
of the quantum transfer matrix (QTM) method for the simplest example of a
spin- 12 Heisenberg chain. In this review using the lattice path integral formulation
and the QTM method the low-temperature asymptotics of the free energy and
correlation lengths are studied. The need to combine numerical computation with
exact results is demonstrated nicely in calculating such physical quantities as
the specific heat and magnetic susceptibility at finite temperature for the spin
system. This method, though also based on exact Bethe ansatz (BA) results
for integrable systems, unlike the thermodynamic BA does not depend on the
string hypothesis. Its central point is to derive a set of coupled nonlinear integral
equations from the BA equations and apply them for an efficient analysis of the
thermodynamics of the model. The authors carefully and systematically show
how to determine various excitations of the spin model and how to find low-
temperature corrections to the eigenvalues and correlation lengths, which are
compatible with the results of finite-size analysis and the predictions of conformal
field theory. It may be mentioned that Andreas Klümper is one of the pioneers in
inventing and developing this field of research.
Finally chapter 10 ‘Reaction–diffusion processes and their connection
with integrable quantum spin chains’ by M Henkel (Nancy, France) gives a
pedagogical account of this physically important subject using many diagrams
and graphical representations and exposing its relation with integrable systems.
The long-time behaviour of such processes is strongly influenced by fluctuations
in low dimensions, which makes the usual mean field approximation inapplicable
and requires a truly microscopic approach to their description. Therefore, the
mapping of the reaction–diffusion (RD) system to integrable magnetic chains,
rather unexpected for such non-equilibrium stochastic processes, not only
demonstrates the diverse applicability of the integrable system but also shows
how its powerful machinery such as the Bethe ansatz can be used for exact
and microscopic analysis of the RD processes. Useful methods like spectral
and partial integrability, free fermions, similarity transformations and diffusion

Copyright © 2003 IOP Publishing Ltd.

 algebras are also reviewed here with several concrete examples. It has been shown
how the recent concept of local scale-invariance could be used in describing non-
equilibrium aging phenomena giving particular emphasis on the kinetic Ising
model with Glauber dynamics.
We sincerely hope that this collection of reviews will be successful in serving
its purpose, proving accessible to the widest range of readership and stimulating
interest in this fascinating subject in all its readers.

Anjan Kundu
April 2003

Copyright © 2003 IOP Publishing Ltd.

 PART I

CLASSICAL SYSTEMS

Copyright © 2003 IOP Publishing Ltd.

 Chapter 1

A journey through the Korteweg–de Vries

equation
M Lakshmanan
Centre for Nonlinear Dynamics, Department of Physics,
Bharathidasan University, Tiruchrapalli, India

1.1 Introduction
All around us, Nature is abundant with innumerable phenomena which can be
described by dispersive wave propagation: hydrodynamic waves, acoustic waves,
electromagnetic waves including optical waves, plasma waves, waves on strings
and rods, etc are just some examples. Such dispersive waves are all characterized
by the appropriate dispersion relations. If the dispersion relation is independent
of the amplitude of the underlying waves so that the frequency is a function of
wavenumber alone, ω = ω(k), we have linear dispersive systems described by
linear partial differential equations. In this case, the system can admit wavepackets
or wavegroup solutions which are linear superpositions of a large number of
elementary waves. Since the group velocity vg , in general, differs from the wave
velocity vp (except for the dispersionless case ω(k) = ck, c = constant), the waves
of the group disperse and die down over distance.
However, in nature, not all waves are so gentle that they disperse and
diminish over distance—there can be permanent waves where the dispersion
relation is amplitude-dependent, ω = ω(k, A), where A is the amplitude of the
wave, corresponding to nonlinear dispersive wave propagation. In this case,
for example, there may be a possibility for solitary waves which can travel
without change of speed and shape over long distances to form. In fact, such a
phenomenon was actually observed as far back as in 1834 by the British naval
architect John Scott Russell in the Union Canal connecting the Scottish cities of
Edinburgh and Glasgow and reported in the scientific literature in 1844 [1].
The Korteweg–de Vries (KdV) equation appeared as a sound basic
formulation of the Scott Russell phenomenon in the year 1895 when the Dutch

Copyright © 2003 IOP Publishing Ltd.

 physicists Korteweg and de Vries [2] derived it from first principles and showed
explicitly that it admits amplitude-dependent cnoidal nonlinear dispersive wave
solutions, a limiting form of which is the solitary wave. The modern theory of
solitons and of completely integrable infinite-dimensional nonlinear dynamical
systems has its origin when the same KdV equation re-occurred in the asymptotic
analysis of the dynamics of the now famous Fermi–Pasta–Ulam (FPU) nonlinear
lattice in the investigations of Kruskal and Zabusky [3]. In fact, it was the
numerical analysis [4] by Zabusky and Kruskal, which showed the remarkable
elastic collision property of solitary waves, that leads ultimately to the notion
of solitons. Kruskal and his coworkers [5] went on to formulate a sound new
technique to solve the Cauchy initial value problem of the KdV equation, which
is now called the inverse scattering transform (IST) method. This pedagogical
review essentially aims to give a brief and elementary account of these historical
developments and of the further understanding of the various solitonic and
complete integrability properties of the KdV equation, which then serves as a
prototypical example for other integrable soliton systems discussed in this special
issue.
The organization of the review is as follows. In section 1.2, we briefly
introduce Scott Russell’s observation and point out how the KdV solitary wave
represents the Scott Russell phenomenon. The unexpected results of the FPU
numerical experiments and the deduction of KdV equation as an asymptotic limit
by Kruskal and Zabusky are discussed in sections 1.3 and 1.4, respectively. This
is followed in section 1.5 by a description of the now well-known numerical
analysis of the KdV equation by Zabusky and Kruskal, signalling the birth of
the concept of a soliton, whose explicit construction is demonstrated by the
Hirota bilinearization method in section 1.6. The Miura transformation and the
identification of Lax pair for the KdV is demonstrated in section 1.7, leading to
the development of IST analysis (section 1.8), which provides explicit soliton
solutions (section 1.9). The Hamiltonian structure and complete integrability
aspects of the KdV equation are discussed in section 1.10 and the existence of
an infinite number of conservation laws and constants of motion is pointed out
in section 1.11. Then the existence of the Bäcklund transformation (section 1.12)
and the Painlevé property (section 1.13) are pointed out. Finally in section 1.14, a
brief account of the connection between the existence of symmetries, invariance
and integrability of the KdV equation is given. Section 1.15 is then devoted to
conclusions.

1.2 Nonlinear dispersive waves: Scott Russell phenomenon

and solitary waves

As noted in the introduction, in nature there can be waves of permanence, which

arise purely due to nonlinear effects. In the 1830s the Scottish naval architect,
John Scott Russell, was carrying out investigations on the shapes of the hulls

Copyright © 2003 IOP Publishing Ltd.

 of ships and the speed and forces needed to propel them for the Union Canal
Company. In August 1834, riding on horseback, Scott Russell observed the
‘Great Wave of Translation’ in the Union canal connecting the Scottish cities of
Edinburgh and Glasgow, where he was carrying out his experiments. He reported
his observations to the British Association in his 1844 ‘Report on Waves’ in the
following delightful description [1].

I believe I shall best introduce the phenomenon by describing the

circumstances of my own first acquaintance with it. I was observing the
motion of a boat which was rapidly drawn along the narrow canal by a
pair of horses, when the boat suddenly stopped not so the mass of water
in the channel which it had put in motion; it accumulated round the
prow of the vessel in a state of violent agitation, then suddenly leaving
it behind, rolled forward with great velocity, assuming the form of a
large solitary elevation, a rounded, smooth and well-defined heap of
water, which continued its course along the canal apparently without
change of form or diminution of speed. I followed it on horse back, and
overtook it still rolling on at a rate of some eight or nine miles an hour,
preserving its original figure some thirty feet long and a foot to a foot
and a half in height. Its height gradually diminished and after a chase of
one or two miles I lost it in the windings of the canal. . .

Like a tsunami wave, this rolling pile of water, a solitary wave, also somehow
maintained its shape and speed which was much larger than conventional linear
dispersive waves. Scott Russell immediately realized that their distinct feature is
their longevity and that they have so much staying power that he could use them to
pump water uphill, which is not possible ordinarily. Scott Russell also performed
some laboratory experiments generating solitary waves by dropping a weight at
one end of a water channel. He was able to deduce empirically that the volume of
water in the wave is equal to the volume of water displaced and further that the
speed, c, of the solitary wave is obtained from the relation

c2 = g(h + a) (1.1)

where a is the amplitude of the wave, h is the undisturbed depth of water and g is
the acceleration due to gravity. A consequence of (1.1) is that taller waves travel
faster!
To put Russell’s formula on a firmer footing, both Boussinesq and Lord
Rayleigh [6] assumed that a solitary wave has a length scale much greater than
the depth of the water. They deduced from the equations of motion for an inviscid
and incompressible fluid, Russell’s formula, (1.1), for c. In fact, they also showed
that the wave profile is given by

u(x, t) = a sech2 [β(x − ct)] (1.2a)

Copyright © 2003 IOP Publishing Ltd.

 where
4h3 (h + a)g
β −2 = (1.2b)
3a
for any a > 0, although the sech2 profile is strictly correct only if a/ h  1.

1.2.1 KdV equation and cnoidal waves and the solitary waves
The ultimate explanation of the Scott Russell phenomenon was provided by two
Dutch physicists Korteweg and de Vries in 1895 [2]. Starting from the basic
principles of hydrodynamics and considering unidirectional wave propagation in a
long but shallow channel, they deduced the celebrated wave equation responsible
for the phenomenon, which now goes by their names. The KdV equation is a
simple nonlinear dispersive wave equation (for details of the actual derivation
see, for example, [6, 12]). In its modern version, it reads as

ut + 6uux + uxxx = 0. (1.3)

Let us look for elementary wave solutions of (1.3) in the form

u = 2f (x − ct) = 2f (ξ ) ξ = x − ct. (1.4)

Then equation (1.3) can be reduced to an ordinary differential equation (ODE)

−cfξ + 12ffξ + fξ ξ ξ = 0. (1.5)

The solution of (1.5) can be expressed in terms of a Jacobian elliptic function

as
√
f (ξ ) = f (x − ct) = α3 − (α3 − α2 )sn2 [ α3 − α1 (x − ct), m] (1.6a)

where the arbitrary parameters α1 , α2 , α3 and c are related to the three integration
constants of equation (1.5) and are also interrelated as
c α3 − α2
(α1 + α2 + α3 ) = m2 = . (1.6b)
4 α3 − α1
Equation (1.6) represents, in fact, the so-called cnoidal wave for obvious reasons.

Special cases
(i) m ≈ 0: harmonic wave. When m ≈ 0, (1.6) leads to elementary progressing
harmonic wave solutions. This can be verified by taking the limit m → 0
(corresponding to the linearized version of (1.3)) in (1.6).
(ii) m = 1: solitary wave. When m = 1, we can write
√
f = α2 + (α3 − α2 ) sech2 [ α3 − α1 (x − ct)]. (1.7)

Copyright © 2003 IOP Publishing Ltd.

 2

´  ¼µ
1

0
-4 -2 0 2 4

Figure 1.1. Solitary wave solution (1.9) of the KdV equation (1.3). Here c = 4.

Choosing now α2 = 0, α1 = 0, α3 = c, we have

√ 
c c
f = sech 2
(x − ct) . (1.8)
4 2
Substituting (1.8) into (1.4), the solution can be written as

√ 
c 2 c (x − ct)
u(x, t) = 2f = sech . (1.9)
2 2
This is, of course, the Scott Russell solitary wave, as can be seen after
suitable rescaling.
The characteristic feature of this solitary wave is that the velocity of the
wave (v = c) is directly proportional to the amplitude (a = c/2): the larger
the wave is, the faster it moves. Unlike the progressing wave, it is fully
localized, decaying exponentially fast as x → ±∞ (see figure 1.1). We will
find in the following sections that this solitary wave is a remarkably stable
entity so we are able to ascribe a particle property to it. It is a purely nonlinear
effect.
(iii) 0 < m < 1: cnoidal waves. When 0 < m < 1, we have the amplitude-
dependent elliptic function solution (1.6). Rewriting it in the form of f =
f (ωt − kx), we can easily check from (1.6) that the dispersion relation
is now amplitude-dependent: ω = ω(k, a). For example, from the solution
(1.6), we can identify f = f (ωt − kx), with
√ √
k = α3 − α1 ω = α3 − α1 c = ck c = 4(α1 + α2 + α3 ).
Using the relations in (1.6), one can establish the amplitude-dependent dispersion
relation mentioned earlier.

1.3 The Fermi–Pasta–Ulam (FPU) numerical experiments on

anharmonic lattices
Not until more than half a century later, the KdV equation and its solitary
wave received their rightful recognition by physicists and mathematicians. The

Copyright © 2003 IOP Publishing Ltd.

 Figure 1.2. The FPU nonlinear lattice.

breakthrough came in an entirely different context—this time in the study of wave

propagation in nonlinear lattices. For more details, see, for example, J Ford [7].

1.3.1 The FPU lattice and recurrence phenomenon

In the early 1950s, E Fermi, J Pasta, and S Ulam were set to make use of
the MANIAC-I analogue computer at Los Alamos Laboratory, USA in solving
important problems in physics. In particular, they were interested in checking
the widely held concepts of ergodicity and equipartition of energy in irreversible
statistical mechanics. They considered for this purpose the dynamics of a chain of
weakly coupled nonlinear oscillators. The chain contains 32 (or 64) mass points,
which interacted through nonlinear forces (see figure 1.2).
Then the equation of motion of the lattice for the displacements yi , i =
0, 1, 2, . . . , N, can be written as

d2 yi
m = f (yi+1 − yi ) − f (yi − yi−1 ) i = 1, 2, . . . , N − 1 (1.10)
dt 2
with y0 = 0 and yN = 0, where FPU assumed the following specific forms for
f (y):

(a) quadratic nonlinearity: f (y) = y + αy 2

(b) cubic nonlinearity: f (y) = y + βy 3

γ1 y |y| < d
(c) broken (piecewise) linearity: f (y) =
γ2 y + δ |y| > d
(α, β, δ, d, γ1 , γ2 are constants).

When there is no nonlinearity (for example, α = 0 in case (a) or β = 0 in

case (b)), it is easy to check that the equation of motion (1.10) is separable into
linear normal modes and that there will be no energy sharing among them.
However, when one of the weakly nonlinear interactions is switched on, the
modes become coupled and one would expect the energy to flow back and forth in
the original normal modes and eventually that equipartition of energy would occur
among the modes. Numerical analysis should confirm this expectation. FPU’s
results are contained in Los Alamos Report Number 1940 of the year 1955. To

Copyright © 2003 IOP Publishing Ltd.

 Figure 1.3. A plot of the normal mode energies Ek = 12 (ȧk2 + 2k ak2 ) for N = 32 and
α = 0.25 in (1.10) [7]. The numbers on the curves represent the modes.

their great surprise, FPU found that no equipartition of energy occurred. When
the energy was assigned to the lowest mode, as time went on only the first few
modes were excited and even this energy returned to the lowest mode after a
characteristic time called the recurrence time.
Figure 1.3 contains the of FPU’s results for the case N = 32 with α = 0.25
in case (a). Starting with an initial shape at t = 0 in the form of a half of a sine
wave given by yj = sin(j π/32), so that only the fundamental harmonic mode was
excited with an initial amplitude a1 = 4 and energy E1 = 0.077 . . . , the figure
depicts the evolution of the first four normal mode energies, Ek , k = 1, 2, 3, 4.
During the time interval 0 ≤ t ≤ 160 in figure 1.3, where t is measured in periods
of the fundamental mode, modes 2, 3 and 4 etc, sequentially begin to absorb
energy from the initially dominant first mode, as one would expect from a standard
analysis. After this, the pattern of energy sharing undergoes a dramatic change.
Energy is now exchanged primarily only among modes 1 through 6 with all the
higher modes getting very little energy. In fact, the motion is almost periodic, with
a recurrence period (the so-called FPU recurrence) at about t = 157 fundamental
periods. The energy in the fundamental mode returns to within 3% of its value at
t = 0.
The unexpected recurrence phenomenon in the FPU experiments stimulated
a great variety of research into the following domains:

(1) the statistical behaviour of nonlinear oscillators, mixing, ergodicity, etc,

(2) the theory of normal mode coupling,
(3) the investigation of nonlinear normal modes, and integrable systems

Copyright © 2003 IOP Publishing Ltd.

 and so on. In fact, the FPU experiment is considered to be the trendsetter of the
modern era of nonlinear dynamics.

1.4 The KdV equation again!

The entirely unexpected results of the FPU experiments motivated many scientists
to try to understand nonlinear phenomena more deeply. Martin Kruskal and
Norman Zabusky from Princeton Plasma Physics Laboratory set out to understand
the FPU recurrence phenomenon through a combination of analytical and
numerical investigations. Their approach, which is based on an asymptotic
analysis, is as follows.

1.4.1 Asymptotic analysis and the KdV equation

Consider the equation of motion (1.10) of the nonlinear lattice with the combined
nonlinear force
f = y + αy 2 + βy 3 . (1.11)
Then, consider the continuous limit
a→0
yn (t) −→ y(na, t) = y(x, t) (1.12)

where a is the lattice parameter. We may also write

a→0
yn±1 (t) −→ y((n ± 1)a, t) = y(x ± a, t)
∂y a2 ∂ 2y a 3 ∂ 3y a4 ∂ 4y
=y±a + ± + +··· . (1.13)
∂x 2! ∂x 2 3! ∂x 3 4! ∂x 4
Substituting (1.13) into (1.10) with f in the form of (1.11), one can obtain the
equation of motion for the continuous case (by retaining terms up to order a 4 ) as
  2 
1 ∂ 2y ∂ 2y ∂y 2 ∂y
= 2 1 + 2αa + 3βa + higher order terms (1.14)
c2 ∂t 2 ∂x ∂x ∂x

where c2 = a 2 /m. Then we can consider the following four cases.

(i) Linear case: α = 0, β = 0. Equation (1.14) is nothing but the linear
dispersionless wave equation.
(ii) Nonlinear case: α = 0, β = 0. This is a hyperbolic equation (when higher-
order terms are omitted). By using the method of characteristics, one can
show that the solution develops multi-valuedness or shocks!
(iii) Addition of a fourth derivative term: Physically one does not expect shocks
to occur in a nonlinear lattice (certainly the experiments did not reveal any!).
So Zabusky and Kruskal added a fourth derivative term (1/12)∂ 4y/∂x 4 to
the right-hand side of (1.14) so as to obtain the final form of equation of

Copyright © 2003 IOP Publishing Ltd.

 motion as
  2 
1 ∂ 2y ∂ 2y ∂y ∂y a 2 ∂ 4y
= 1 + 2αa + 3βa 2 + . (1.15)
c2 ∂t 2 ∂x 2 ∂x ∂x 12 ∂x 4

Now considering unidirectional waves (moving to the right), one can make
a change of variables,

ξ = x − ct τ = a 2 ct y = v/a. (1.16)

Equation (1.15) can then be rewritten as

 2
∂ 2v ∂v ∂ 2 v 3 ∂ 2v ∂v 1 ∂ 4v a 2 ∂ 2v
+α + β + = (1.17)
∂ξ ∂τ ∂ξ ∂ξ 2 2 ∂ξ 2 ∂ξ 24 ∂ξ 4 2 ∂τ 2

after redefining (α/a 2 ) as α and (β/a 2 ) as β. Then in the continuous limit,

a → 0, and with the redefinition
∂v
u= (1.18)
∂ξ
one finally obtains
 
∂u ∂u 3 2 ∂u 1 ∂ 3u
+ αu + βu + = 0. (1.19)
∂τ ∂ξ 2 ∂ξ 24 ∂ξ 3

√ β = 0, α = 0√and if τ and ξ are replaced by the standard notation

When
t = 24 τ and x = 24 ξ , we have

ut + αuux + uxxx = 0 (1.20)

which is nothing but the KdV equation (1.3) with (α = 6), for a suitable
choice of α but which has now occurred in an entirely new context!
Similarly for α = 0, β = 0, we have
3
ut + βu2 ux + uxxx = 0 (1.21)
2
which we may call as the modified KdV equation or, briefly, the MKdV
equation.
(iv) Scale change: Note that the KdV equation can always be written in the form

ut + puux + quxxx = 0 (1.22)

under a suitable scale change. Or, in other words, making a change of scales
of the variables t, x and u and redefining the variables, (1.22) can always be
written in the form (1.20). We will use this freedom to choose the coefficients

Copyright © 2003 IOP Publishing Ltd.

 p and q as per convenience in our further analysis. This is also true for the
MKdV equation.
Thus, one may conclude that wave propagation in the FPU lattice with a
quadratic nonlinear force may be described in a non-trivial way by the KdV
equation and the lattice with a cubic nonlinear force may be described by
the MKdV equation. Then what do these equations have to do with the FPU
recurrence phenomenon reported by Fermi, Pasta and Ulam in their well-
known experiments?

1.5 Numerical experiments of Zabusky and Kruskal: the birth

of solitons
Zabusky and Kruskal recalled that the KdV equation admits a solitary wave
solution, which has a distinct nonlinear character. Further, if nonlinear normal
modes exist leading to recurrence and non-energy-sharing phenomena, then the
solitary wave should play an important role. So they initiated a many-faceted and
deep numerical study of the KdV equation, the results of which were reported in
the year 1965 [4]. The KdV equation which Zabusky and Kruskal considered in
their numerical analysis had the form

ut + uux + δ 2 uxxx = 0. (1.23)

Their study mainly focused on the following two aspects:

(1) What will be the type of solution the system admits for a chosen initial
condition, particularly for a spatially periodic initial condition, u(x, 0) =
cos πx, 0 ≤ x ≤ 2, so that u, ux , uxx are periodic on [0, 2] with u(x, t) =
u(x + 2, t), etc?
(2) How do solitary waves of the KdV equation interact mutually, particularly
when two such waves of differing amplitudes (and so of different velocities)
interact?
For their numerical analysis, Zabusky and Kruskal converted the KdV equation
(1.3) into a difference equation on a rectangular mesh with periodic boundary
conditions [4]. The outcome of these numerical experiments may be summarized
as follows.

1.5.1 Periodic boundary conditions

(1) As δ 2 is small, the nonlinearity dominates over the third derivative term. As
a consequence the wave steepens in regions where it has a negative slope.
(2) As the wave steepens, the δ 2 uxxx term becomes important and balances the
nonlinear term uux .
(3) At a later time the solution develops a train of eight well-defined (solitary)
waves with different amplitudes each like sech2 functions, with the

Copyright © 2003 IOP Publishing Ltd.

 Figure 1.4. Zabusky–Kruskal’s numerical experimental results [4]: solution of the KdV
equation (1.3) with δ = 0.022 and u(x, 0) = cos πx for 0 ≤ x ≤ 2. The dotted curve
represents u at t = 0. The dashes are the solution at t = 1/π. The continuous curve gives
u at t = 3.6/π.

faster (taller) waves catching up and overtaking the slower (short) waves
(figure 1.4). These nonlinear waves interact strongly and then continue
thereafter almost as if there had been no interaction at all.
(4) Each of the solitary wave pulses moves uniformly at a rate which is linearly
proportional to its amplitude. Thus, the solitons spread apart. Because of
the periodic boundary condition, two or more solitons eventually overlap
spatially and interact nonlinearly (figure 1.4). Shortly after the interaction,
they reappear virtually unaffected in size and shape.
(5) There exists a period TR , the so-called recurrence time at which all the
solitons arrive almost in the same phase and almost reconstruct the initial
state through nonlinear interactions, thereby explaining the FPU recurrence
phenomenon qualitatively.
(6) The persistence of the solitary waves led Zabusky and Kruskal to coin the
name soliton (after names such as the photon, phonon, etc) to emphasize the
particle-like character of these waves which seem to retain their identities in
a collision.

1.5.2 Initial condition with just two solitary waves

These observations can be better understood by considering an initial condition
consisting of just two solitary waves of differing amplitudes as shown in
figure 1.5. Suppose that at time t → −∞ (say t = −800 units), two such waves
are given, which are well separated and with the bigger one to the right as in
figure 1.5. Then as the system evolves as per the KdV equation, after a sufficient
time the waves overlap and interact (the bigger one catches up with the smaller
one). Following the process still longer, one finds that the bigger one separates
from the smaller one, after overtaking it, and asymptotically (as t → ∞) the wave

Copyright © 2003 IOP Publishing Ltd.

 
 
0.06
0.03 800
0
400
-100 0
-50
-400

0

50
100 -800

Figure 1.5. Two-soliton

√ interaction of the KdV equation. The parameters in (1.37) are fixed
(0) (0)
as k1 = 0.2, k2 = 3 k1 , ω1 = k13 , ω2 = k23 and η1 = 0, η2 = 0.

t 2 ∆ / k1

2 ∆ / k2

u2

u1

Figure 1.6. Phase shifts of two interacting solitons.

solution regains its initial shape and, hence, the two solitary waves also regain
their velocities. The only effect of the interaction is a phase shift, that is the centre
of each wave is at a different position than where it would have been if each one
of them were travelling alone (figure 1.6). Again, because of the analogy with
the elastic collision property of particles, Zabusky and Kruskal referred to these
solitary waves as solitons.
Thus, the major new concepts that emerge from the Zabusky–Kruskal
experiments are:
(1) when the nonlinearity suitably balances the linear dispersion as in the KdV
equation, solitary waves can arise;
(2) these solitary waves in appropriate nonlinear systems can interact elastically
like particles without changing their shapes or velocities;
(3) the solitons can constitute the general solution of the initial value problem of
a class of nonlinear dispersive wave equations like the KdV equation.
Naturally, the next obvious question to arise is whether exact analytical forms
of the soliton solutions of the KdV equation, beyond the solitary wave solution,
can be obtained which can correspond to all these numerical results. In fact,

Copyright © 2003 IOP Publishing Ltd.

 Martin Kruskal and his coworkers went on further to completely integrate the
initial value problem (IVP) of the KdV equation and, in this process, also invented
a new method to solve the IVP of a class of nonlinear evolution equations. The
method is now called the inverse scattering transform (IST) method, which may
be considered as a natural generalization of the Fourier transform method that is
applicable to linear dispersive systems.

1.6 Hirota’s bilinearization method: explicit soliton solutions

Before introducing the IST method (see section 8) for the KdV equation, let us
consider the so-called direct or bilinearization method, which was introduced by
R Hirota in 1971 [8]. By using this method, we will obtain explicitly the two-
soliton solution of the KdV equation which, in fact, corresponds to Zabusky–
Kruskal’s numerical experimental result on two-solitary-wave scattering as
depicted in figure 1.5. We will also indicate the method to obtain more general
soliton solutions.
Let us consider the KdV equation (1.3). If it is the aim to obtain soliton
solutions alone and not the much more general task of solving the IVP, then we
can use the algorithmic bilinearization method of Hirota mentioned earlier. The
main ingredient in this method is to introduce a bilinearizing transformation so
that the given evolution equation can be written in the so-called bilinear form:
each term in the transformed equation has a total degree two. Thus, with the
transformation
∂2
u = 2 2 log F (1.24)
∂x
the KdV equation (1.3) takes the form
Fxt F − Fx Ft + Fxxxx F − 4Fxxx Fx + 3Fxx
2
= 0. (1.25)
Now expanding F in a formal power series in terms of a small parameter  (which
one can always introduce into equation (1.25)) as
F = 1 + f (1) +  2 f (2) + · · · . (1.26)
Equating each power of  separately to zero, we get a system of linear partial
differential equations (PDEs). Up to O( 3 ), we can write them as
O( 0 ) : 0=0 (1.27a)
O() : fxt(1) + fxxxx
(1)
=0 (1.27b)
O( ) :
2
fxt(2) + fxxxx
(2)
= fx(1) ft(1) + 4fxxx fx − 3(fxx
(1) (1) (1) 2
) (1.27c)
(3) (1) (2) (2) (1) (1) (2) (2) (1)
O( 3 ) : fxt + fxxxx
(3)
= fx ft + fx ft − fxt f − fxt f
− fxxxx
(1)
f (2) − fxxxx
(2)
f (1) + 4fxxx
(1) (2)
fx
+ 4fxxx fx − 6fxx fxx .
(2) (1) (1) (2)
(1.27d)

Copyright © 2003 IOP Publishing Ltd.

 We can then successively solve this set of linear PDEs. To start with, we can easily
write the solution of (1.27b) as

N
(0) (0)
f (1) = eη i ηi = ki x − ωi t + ηi ωi = ki3 ηi = constant.
i=1
(1.28)
Substituting this into the right-hand side of (1.27c), we can solve for f (2) . This
procedure can be repeated further to find f (3) , f (4) , . . . successively. In practice,
one finds the solution for N = 1, 2, 3 and then hypothesizes it for arbitrary N
which should then be proved by induction. It turns out that for every given value
of N in the summation in (1.28) we get a soliton of order N as discussed later.
The task of obtaining the forms of the right-hand sides of (1.27) can be
simplified enormously by introducing the so-called Hirota’s bilinear D-operator.
The associated algebra has been developed by Hirota. However, we will not
introduce it here and the interested reader may refer, for example, to [9].

1.6.1 One-soliton solution

For example, for N = 1,

f (1) = eη1 η1 = k1 x − ω1 t + η1(0) (1.29)

with ω1 = k13 and

fxt(2) + fxxxx
(2)
= 0. (1.30)
So we can choose = 0. Then one can easily prove that all
f (2) f (i) = 0, i ≥ 3.
Thus, the solution to (1.26) becomes

F = 1 + eη 1 η1 = k1 x − k13 t + η1(0) . (1.31)

Substituting this into the transformation (1.24), we finally obtain the one-soliton
solution
k2 1 (0)
u(x, t) = 1 sech2 (k1 x − k13 t + η1 ) (1.32)
2 2
which√is the same as the solitary wave solution (1.12), with the identification
k1 = c .

1.6.2 Two-soliton solution

Proceeding in a similar way for N = 2, one has
(0) (0)
f (1) = eη1 + eη2 η1 = k1 x − ω1 t + η1
η2 = k2 x − ω2 t + η2 .
(1.33)
Substituting (1.33) into the right-hand side of (1.27c), and solving, we obtain

f (2) = eη1 +η2 +A12 eA12 = [(k1 − k2 )/(k1 + k2 )]2 . (1.34)

Copyright © 2003 IOP Publishing Ltd.

 Using this in (1.26), we then obtain

F = 1 + eη1 + eη2 + eη1 +η2 +A12 . (1.35)

Substituting this in (1.24), we ultimately obtain the two-soliton solution

 
1 2 k22 cosech2 (η2 /2) + k12 sech2 (η1 /2)
u = (k2 − k1 ) 2
. (1.36)
2 (k2 coth(η2 /2) − k1 tanh(η1 /2))2

The solution (1.36) when plotted has exactly the same form as in figure 1.5,
thereby showing the soliton nature of the solitary wave as discussed in the
previous section.

1.6.3 N -soliton solutions

One can proceed as before for the general case, with the choice

f (1) = eη1 + eη2 + · · · + eηN (1.37)

and then solve successively for f (2) , . . . , f (N) and finally obtain F and u.
Explicit expressions can be written down with some effort, which we desist from
doing so here due to its somewhat complicated nature. For more details, see, for
example, [9].

1.6.4 Asymptotic analysis

Let us consider the two-soliton solution (1.36) and analyse the limits t → −∞
and t → +∞ separately, so as to understand the interaction of two one-solitons
centred around η1 ≈ 0 or η2 ≈ 0. Without loss of generality, let us assume that
(0)
k2 > k1 . Then we can see that in the limits t → ±∞, η1 = k1 x − k13 t + η1 and
(0)
η2 = k2 x − k23 t + η2 take the following limiting values:
(i) t → −∞
η1 ≈ 0, η2 → ∞ η2 ≈ 0, η1 → −∞.
(ii) t → +∞
η1 ≈ 0, η2 → −∞ η2 ≈ 0, η1 → ∞.
Substituting these limiting values into the two-soliton expression (1.36) and
with simple algebra, we can easily show that we have the following solutions for
t → +∞ and t → −∞:
(i) t → −∞
Soliton 1 (η1 ≈ 0)
   
1 η1 − k2 + k1
u(x, t) = k12 sech2 = log . (1.38)
2 2 k2 − k1

Copyright © 2003 IOP Publishing Ltd.

 Soliton 2 (η2 ≈ 0)
1
u(x, t) = k22 sech2 [(η2 + )/2]. (1.39)
2
(ii) t → ∞
Soliton 1 (η2 ≈ 0)
1
u(x, t) = k12 sech2 [(η1 + )/2]. (1.40)
2
Soliton 2 (η1 ≈ 0)
1
u(x, t) = k22 sech2 [(η2 − )/2]. (1.41)
2
Using this analysis, we can readily interpret the two-soliton solution of the
KdV equation given by (1.36) in the following way. Two individual solitary waves
(one-solitons) of differing amplitudes, k12 /2 and k22 /2 (k2 > k1 ), with the smaller
one positioned to the right of the larger one, travel to the right with speeds k1 and
k2 respectively. The larger one soon catches up with the smaller one, undergoes
a nonlinear interaction while overtaking it and, ultimately, soliton 1 and soliton
2 are interchanged. The net effect is merely a total phase shift 2 suffered by
the solitons without any change in shape, amplitude or speed. Or, in other words,
the solitary waves of the KdV equation undergo elastic collisions, reminiscent of
particle collisions, as demonstrated by the numerical experiments of Zabusky and
Kruskal. So they are, indeed, solitons of the KdV equation.
This analysis can be extended to N-soliton solutions also but we will refrain
from doing so here. However, in the next section, we will discuss the more general
method of solving the Cauchy initial value problem of the KdV equation, namely
the IST method, which can also lead to explicit N-soliton solutions.

1.7 The Miura transformation and linearization of KdV:

the Lax pair
1.7.1 The Miura transformation
It is known that the nonlinear Burgers equation, which is a nonlinear heat
equation,
ut + uux = νuxx (1.42)
where ν is a constant parameter, can be transformed into the standard linear heat
equation
vt = νvxx (1.43)
under the so-called Cole–Hopf transformation
vx
u = −2ν . (1.44)
v

Copyright © 2003 IOP Publishing Ltd.

 In the year 1968, R M Miura [10], who was working in Martin Kruskal’s
group at Princeton, noted that the KdV equation
ut − 6uux + uxxx = 0 (1.45)
and the modified KdV equation
vt − 6v 2 vx + vxxx = 0 (1.46)
are related to each other through the transformation
u = v 2 + vx . (1.47)
Note the change in sign in the second term of the KdV equation (1.45) which
can be obtained by a scale change from (1.3) and is chosen for convenience. The
transformation (1.47) is now called the Miura transformation in the literature,
which essentially relates two nonlinear equations to each other.
Now in analogy with the Cole–Hopf transformation (1.44) for the Burgers
equation, one can think of a transformation
ψx
v= (1.48)
ψ
for the MKdV equation (1.46). Then in view of the Miura transformation (1.47),
we have the following transformation for the KdV equation:
ψxx
u= . (1.49)
ψ

1.7.2 Galilean invariance and the Schrödinger eigenvalue problem

At this point, we note that the KdV equation (1.45) is form invariant under the
Galilean transformation
x
= x − λt t
= t u
= u + λ. (1.50)
Consequently, in the new frame of reference, we have
ψx
x

u
= + λ. (1.51)
ψ
Equation (1.51) can be re-expressed as
ψx
x
+ (λ − u
)ψ = 0. (1.52)
Omitting the primes for convenience hereafter, we finally obtain the time-
independent Schrödinger-type linear eigenvalue problem
ψxx + (λ − u)ψ = 0 (1.53)
in which the unknown function u(x, t) of the KdV equation appears as a
‘potential’, while (1.53) defining the transformation function ψ(x, t) itself is
linear.

Copyright © 2003 IOP Publishing Ltd.

 1.7.3 Linearization of the KdV equation
Treating (1.53) as a linearizing transformation for the KdV equation (1.45), we
can straightforwardly obtain an evolution equation for the function ψ(x, t), which
reads
ψt = −4ψxxx + 6uψx + 3ux ψ. (1.54)
Thus, one can conclude that the nonlinear KdV equation (1.45) is equivalent
to two linear differential equations, namely the Schrödinger-type eigenvalue
equation (1.53) and the associated linear time evolution equation (1.54)
for the eigenfunction ψ(x, t). Note that in both these linear equations the
unknown function u(x, t) (and its derivatives) of the KdV equation occurs as a
coefficient. One can also easily check the converse, namely given the two linear
systems (1.53) and (1.54), they are equivalent to the nonlinear KdV equation.

1.7.4 Lax pair

The possibility of the linearization of the nonlinear KdV equation in terms of the
linear systems (1.53) and (1.54) can be rephrased in the following elegant way, as
formulated by P D Lax [11]. Consider the linear eigenvalue problem

Lψ = λψ (1.55a)

where, in the present problem, the linear differential operator is

∂2
L=− + u(x, t) (1.55b)
∂x 2
and λ = λ(t) is the eigenvalue at time t. Let the eigenfunction ψ(x, t) evolve as

ψt = Bψ (1.56a)

where, in the case of the KdV equation, the second linear differential operator is
 
∂3 ∂ ∂
B = −4 + 3 u + u . (1.56b)
∂x 3 ∂x ∂x

Then with the requirement that the eigenvalue λ does not change with time, that
is
λ(t) = λ(0) = constant (1.57)
the compatibility of (1.55b) and (1.56b) leads to the Lax equation or Lax
condition:
Lt = [B, L] = (BL − LB). (1.58)
For the specific forms of L and B given by (1.55b) and (1.56b), the Lax equation
is, indeed, equivalent to the KdV equation (1.45), provided λ is unchanged in

Copyright © 2003 IOP Publishing Ltd.

 time. For any other suitable choice of L and B, a different nonlinear evolution
equation will be obtained.
One may say that the eigenvalue problem is isospectral. The Lax
condition (1.58) is the isospectral condition for the Lax pair L and B. The Lax
condition has, indeed, played a very important role in soliton theory, not only for
the KdV equation but also for the other soliton systems as well. In fact, it is now
an accepted fact that the existence of a Lax pair is, indeed, a decisive hallmark of
integrable systems [9, 12].

1.8 Lax pair and the method of inverse scattering: a new

method to solving the initial value problem
We are interested in solving the initial value problem (IVP) of the KdV equation,
that is: ‘Given the initial value u(x, 0) at t = 0, how does the solution of the
x→±∞
KdV equation (1.45) evolve for the given boundary conditions, say u(x, t) −→
0?’. Also how does the linearization property discussed in the previous section
help in this regard? In the following we briefly point out that indeed the
linearization property leads to a new method of integrating the nonlinear evolution
equation (1.45) through a three-step process. The procedure was originally
developed by Gardner, Greene, Kruskal and Miura. This method, now called the
inverse scattering transform (IST) method, may be considered as a nonlinear
Fourier transform method. It will be now described as applicable to the KdV
equation.

1.8.1 The IST method for the KdV equation

The analysis proceeds in three steps similar to the case of the Fourier transform
method applicable for linear dispersive systems:
(i) direct scattering transform analysis,
(ii) analysis of time evolution of scattering data and
(iii) IST analysis.
The method is schematically shown in figure 1.7. The details are as follows.

1.8.1.1 Direct scattering analysis and scattering data at t = 0

The given information is the initial data u(x, 0), which has the property that it
vanishes sufficiently fast as x → ±∞. Now considering the Schrödinger spectral
problem (1.55) at t = 0,
|x|→∞
ψxx + [λ − u(x, 0)]ψ = 0 u −→ 0 (1.59)

it is well known from linear spectral theory (and one-dimensional quantum

mechanics) that the system (1.59) admits (for details see, for example, [9, 12])

Copyright © 2003 IOP Publishing Ltd.

 Figure 1.7. Schematic diagram of the inverse scattering transform method.

(1) a finite number of bound states with eigenvalues

λ = −κn2 n = 1, 2, . . . , N (1.60)

and normalization constants Cn (0) of the associated eigenstates and

(2) a continuum or scattering states with the continuous eigenvalues,

λ = k2 −∞ < k < ∞. (1.61a)

They are further characterized by the reflection coefficient R(k, 0) and
transmission coefficient T (k, 0) such that

|R(k, 0)|2 + |T (k, 0)|2 = 1. (1.61b)

Thus, from the given potential (initial data) u(x, 0) with the boundary
conditions u → 0 as x → ±∞, one can carry out a direct scattering analysis
of (1.59) to obtain the scattering data at t = 0:

S(0) = {κn , Cn (0), R(k, 0), T (k, 0), n = 1, 2, . . . , N, −∞ < k < ∞}.
(1.62)
A typical example to illustrate this is the model potential u(x) =
−A sech2 αx [12].

1.8.1.2 Time evolution of scattering data

Now as the potential u(x, t) evolves from its initial value u(x, 0) so that it satisfies
the KdV equation, how does the corresponding scattering data given by (1.62)
evolve from S(0) to S(t)? In order to understand this, one can use the time

Copyright © 2003 IOP Publishing Ltd.

 evolution equation of the eigenfunction (1.54),
 
∂ 3ψ ∂ ∂
ψt = −4 + 3 u + u ψ. (1.63)
∂x 3 ∂x ∂x

Since the scattering data is intimately associated with the asymptotic (x → ±∞)
behaviour of the eigenfunction, where the potential u(x, t) → 0, it is enough if
we confine the analysis to this region. Thus as x → ±∞, (1.63) can be written as

∂ 3ψ
ψt = −4 x → ±∞. (1.64)
∂x 3

(a) Scattering states. Without loss of generality, we can write the asymptotic
form of the general solution to (1.64) as
x→−∞
ψ(x, t) −→ a+ (k, t) eikx + a− (k, t) e−ikx (1.65a)
x→+∞ −ikx
−→ b+ (k, t) e ikx
+ b− (k, t) e . (1.65b)

Substituting these asymptotic forms into (1.64), we obtain

da± db±
= ±4ik 3 a± = ±4ik 3 b± . (1.66)
dt dt
On integration, we get

a± (k, t) = a± (k, 0) e±4ik

3t
(1.67a)
±4ik 3 t
b± (k, t) = b± (k, 0) e . (1.67b)

Consider now the standard type of scattering solutions (involving incident,

reflected and transmitted waves)
1 x→−∞
ψ −→ eikx + R(k, t)e−ikx (1.68a)
a+ (k, t)
x→+∞
−→ T (k, t) eikx . (1.68b)

Comparing (1.66) and (1.68) and making use of (1.67), it is straightforward to see
that
a− (k, t)
= R(k, 0) e−8ik t
3
R(k, t) = (1.69a)
a+ (k, t)
b+ (k, t)
T (k, t) = = T (k, 0) (1.69b)
a+ (k, t)

with b− (k, t) taken as zero.

Copyright © 2003 IOP Publishing Ltd.

 (b) Bound states. By construction, the eigenvalues λ = −κn2 , n =
1, 2, . . . , N, do not change with time:

λ(t) = λ(0) =⇒ κn (t) = κn (0). (1.70)

Correspondingly, the eigenfunctions of these discrete states satisfy the time
evolution equation
ψn,t = −4ψn,xxx x → ±∞. (1.71)
Then we have
ψn (x, t) −→ eκn x x → −∞ (1.72a)
−κn x
−→ Cn (t) e x → +∞ (1.72b)

with the normalization condition

∞
|ψn (x, t)|2 dx = 1. (1.72c)
−∞

Substituting (1.72) into (1.71), we see that the normalization constants Cn (t)
evolve as
dCn
= 4κn3Cn (1.73)
dt
so that
3
Cn (t) = Cn (0) e4κn t . (1.74)
Thus, at an arbitrary future instant of time ‘t’, the scattering data S(t)
corresponding to the potential u(x, t) evolves from Sn (0) of the initial data
u(x, 0):
3
S(t) = {κn (t) = κn (0), Cn (t) = Cn (0) e4κn t , n = 0, 1, 2, . . . , N,
R(k, t) = R(k, 0) e−8ik t , −∞ < k < ∞}.
3
(1.75)

1.8.1.3 Inverse scattering analysis

Now given the scattering data S(t) as in (1.75) at time ‘t’, can one invert the data
and obtain uniquely the potential u(x, t) of the Schrödinger spectral problem
(1.55), in which the time variable ‘t’ enters only as a parameter? The answer
is yes, and it can be done by solving a linear Volterra-type singular, integral
equation called the Gelfand–Levitan–Marchenko integral equation [9, 12]. The
scattering data S(t), given by (1.75), is given as input into this integral equation
which, when solved, gives the solution u(x, t) of the KdV equation. The linear
integral equation reads as follows:

K(x, y, t) + F (x + y, t)
∞
+ F (y + z, t)K(x, z, t) dz = 0 y>x (1.76a)
x

Copyright © 2003 IOP Publishing Ltd.

 where


N ∞
1
F (x + y, t) = Cn2 (t) e−κn (x+y) + R(k, t) eik(x+y) dk. (1.76b)
n=1
2π −∞

Note that in this equation the time variable ‘t’ enters only as a parameter and
that all the information about the scattering data are contained in the function
F (x + y, t). Solving (1.76), we finally obtain the potential

d
u(x, t) = −2 K(x, x + 0, t). (1.77)
dx
For details, see, for example, [9, 12]. Thus, the initial value problem of the KdV
equation stands solved.

1.9 Explicit soliton solutions

Now we are in a position to obtain all the soliton solutions and the properties
associated with them as discussed in the previous section. As seen earlier, for
solving the general initial value problem, one has to solve the Gelfand–Levitan–
Marchenko integral equation (1.76), with the full set of scattering data S(t).
Although this is possible in principle, in practice this may not be completely
feasible analytically. However, for the special, but important, class of the so-called
reflectionless potentials, characterized by the condition

R(k, t) = 0 (1.78)

it is possible to solve fully the Gelfand–Levitan–Marchenko integral equation.

An example of the reflectionless case is, again, the potential u(x) =
−A sech2 αx, A > 0. Then if there are N bound states, the corresponding
solution, u(x, t), turns out to be the N-soliton solution. First, let us obtain the
one- and two-soliton solutions explicitly and then generalize the results to the
N-soliton solution.

1.9.1 One-soliton solution (N = 1)

Consider the special case of reflectionless potential (R(k, t) = 0) with only one
bound state, N = 1, specified by (example: u(x) = −2 sech2 x has got one bound
state, λ = −1)
C1 (t) = C(t) = C(0) e+4κ
3t
κ1 = κ (1.79)
so that in (1.76b)
3 t −κ(x+y) 3 t −κ(x+y)
F (x + y, t) = C 2 (0) e8κ = C02 e8κ (1.80)

Copyright © 2003 IOP Publishing Ltd.

 and the Gelfand–Levitan–Marchenko integral equation becomes
∞
2 8κ 3 t −κ(x+y)
K(x, y, t) + C0 e 2 8κ 3 t
+ C0 e e−κ(y+z)K(x, z, t) dz = 0.
x
(1.81)
Then it is straightforward to check that

∂K
= −κK. (1.82)
∂y

On solving, we have
K(x, y, t) = e−κy h(x, t) (1.83)
where the function h(x, t) is to be determined. Substituting (1.83) back into (1.81)
and simplifying, we can find that
3 t −κx
−C02 e8κ
h(x, t) = 3 t −2κx
. (1.84)
[1 + (C02 /2κ) e8κ ]

Then from (1.83), we have

3 t −κ(x+y)
−C02 e8κ
K(x, y, t) = 3 t −2κx
. (1.85)
[1 + (C02 /2κ) e8κ ]

So the corresponding solution to the KdV equation can be obtained from (1.77)
as
d
u(x, t) = −2 K(x, x + 0, t)
dx
= −2∂K(x, y, t)/∂x|y=x − 2∂K(x, y, t)/∂y|y=x
e−2κ(x−4κ
2 t )−2δ
1
= −2κ 2 δ= log(2κ/C02 )
[1 + e−2κ(x−4κ t )−2δ ]2
2
2
= −2κ sech2 [κ(x − 4κ 2t) + δ].
2
(1.86)

Expression (1.86) is, indeed, the one-soliton solution (1.32) obtained by the Hirota
method. (Note the scale change and negative sign in (1.86), due to the difference
in the coefficients in front of the nonlinear term in (1.45) and (1.3), and also a
redefinition of the parameters.)

1.9.2 Two-soliton solution

Again, let us consider a reflectionless potential such that R(k, t) = 0 but now with
two bound states (example: u(x) = −6 sech2 x has two bound states with λ1 =
−4, λ2 = −1), specified by the discrete values κ1 and κ2 and the corresponding

Copyright © 2003 IOP Publishing Ltd.

 normalization constants C1 (t) and C2 (t). Then we have the Gelfand–Levitan–
Marchenko integral equation in the form
2 8κ1 t −κ1 (x+y)
3 2 8κ2 t −κ2 (x+y)3
K(x, y, t) + C10 e e + C20 e e
∞
e−κ1 (y+z) K(x, z, t) dz
3
+ C10
2 8κ1 t
e
x
∞
2 8κ23 t
+ C20 e e−κ2 (y+z) K(x, z, t) dz = 0 (1.87)
x

where C10 and C20 are the normalization constants corresponding to the two
bound states. Let

K(x, y, t) = e−κ1 y h1 (x, t) + e−κ2 y h2 (x, t). (1.88)

Using (1.88) in (1.87) and equating the coefficients of e−κ1 y and e−κ2 y to zero
separately, we can obtain two algebraic equations for h1 and h2 . Solving them,
one obtains h1 (x, t) and h2 (x, t). Using them in (1.88), we obtain from (1.77)
that
d −κ1 y
u(x, t) = −2 [e h1 (x, t) + e−κ2 y h2 (x, t)]
dx
κ 2 cosech2 γ2 + κ12 sech2 γ1
= −2(κ22 − κ12 ) 2 (1.89a)
(κ2 coth γ2 − κ1 tanh γ1 )2
where
 2 
1 Ci0 (κ2 − κ1 )
γ i = κi x − 4κi3t − δi δi = log i = 1, 2. (1.89b)
2 2κi (κ2 + κ1 )
One can easily check that the form (1.89) is, indeed, the two-soliton solution of
the KdV equation discussed in section 1.6, with appropriate scale change and
redefinition of parameters.

1.9.3 N -soliton solution

Considering now reflectionless potentials with N-bound states, we can write the
expression (1.76b) as


N 
N
F (x + y, t) = Cn2 (t) e−κn (x+y) = Cn e−κn x Cn e−κn y
n=1 n=1

= gn (x, t)gn (y, t) gn (x, t) = Cn (t) e−κn x . (1.90)

Then defining

N
K(x, y) = ωn (x)gn (y) (1.91)
n=1

Copyright © 2003 IOP Publishing Ltd.

 (here the t dependence is suppressed for convenience) and substituting into the
Gelfand–Levitan–Marchenko integral equation (1.76), we obtain


N ∞
ωm (x) + gm (x) + ωn (x) gm (z)gn (z) dz = 0. (1.92)
n=1 x

Defining now the matrices

∞
Pmn (x) = δmn + gm (z)gn (z) dz (1.93a)
x
ω(x) = (ω1 (x), ω2 (x), . . . , ωN (x))T g(x) = (g1 (x), g2 (x), . . . , gN (x))T
(1.93b)

(1.93) can be rewritten as the matrix equation

P(x)ω(x) = −g(x). (1.94)

Using (1.91), we have

K(x, x) = gT (x)ω(x) = −gT (x)P−1 (x)g(x).

Also from (1.93), we have

dPmn
= −gm (x)gn (x).
dx
Then
−1
K(x, x) = − tr(gm Pmn gn )
 
dP
= tr P−1
dx
  Pml dPlm
=
l m |P| dx
1 d|P| d
= = log |P| (1.95)
|P| dx dx
where Pml is the cofactor matrix and |P| is the determinant of P. In (1.95),
standard properties of matrices and determinants have been used.
Finally, we can write

d d2
u(x, t) = −2 K(x, x + 0, t) = −2 2 log |P| (1.96)
dx dx
as the required N-soliton solution of the KdV equation. It can also be obtained by
the Hirota method as described in section 1.6. One may note the similarity in the
form of (1.96) and the bilinearizing transformation (1.24).

Copyright © 2003 IOP Publishing Ltd.

 1.9.4 Soliton interaction
As described in the introduction, the solitons of the KdV equation undergo only
elastic collisions without any change in shape or speed, except for the phase shifts.
We have seen in section 1.6 from the two-soliton solution expression (1.36) that
the larger and smaller solitons undergo phase shifts + and − respectively given
by  
+ − κ1 − κ2
 = − = log < 0. (1.97)
κ1 + κ2
Extending this analysis to the N-soliton case, (1.96), assuming that κ1 > κ2 >
· · · > κN > 0, then for fixed γn , for t → ±∞
u(x, t) ∼ −2κn2 sech2 (γn + ±
n) γn = κn (x − 4κn2 t + δn ) (1.98)
so that the nth soliton undergoes a phase shift given by
n = + −
n − n

N   
n−1  
κn − κm κm − κn
= log − log . (1.99)
m=n+1
κn + κm m=1
κm + κn

1.9.5 Non-reflectionless potentials

Let us now consider the case in which the initial state of the KdV equation
is such that the potential is non-reflectionless, that is R(k, 0) = 0. Then as
discussed in section 1.8, the reflection coefficient evolves according to (1.69).
Correspondingly, the contribution to F (x + y) in (1.76b) comes from both the
bound states and continuum states. Solving the Gelfand–Levitan–Marchenko
integral equation exactly in this case becomes impossible. However, it is possible
to carry out a perturbative analysis for sufficiently large t and to show that
asymptotically the solution of the KdV equation consists of N individual solitons
in the background of small amplitude dispersive propagating waves which, in due
course, disperse and die down (see, for example, [9]).
To see the nature of the dispersive waves, let us consider the case in which
there is no bound state at all so that (1.76b) becomes
∞
1
R(k, 0) e−8ik t eikx dk.
3
F (x, t) = (1.100)
2π −∞
Substituting this into the Gelfand–Levitan–Marchenko equation, one can solve it
and show that the solution of the KdV equation for sufficiently large t is
∞
1
4ikR(k, 0) e−8ik t e−2ikx dk
3
u(x, t) ≈ (1.101a)
2π −∞
so that
∞
1
u(x, t) = F (k) e−i(ωt −k̂x) dk ω = −k̂ 3 , k̂ = −2k. (1.101b)
2π −∞

Copyright © 2003 IOP Publishing Ltd.

 Equation (1.101) is nothing but the wave packet solution of the linearized KdV
equation, which have been discussed in the earlier sections.

1.10 Hamiltonian structure of KdV equation: complete

integrability
Having solved the initial value problem of the KdV equation, we now turn our
attention to establish the integrability aspects of it. As a prelude, in this section
we wish to bring out the Hamiltonian nature of the KdV equation and obtain
appropriate Hamiltonian form for it. We will also point out the fact that the
KdV equation may be considered as an infinite-dimensional completely integrable
dynamical system in the Liouville sense. In fact, all soliton possessing systems
belong to the class of such completely integrable dynamical systems.

1.10.1 KdV as a Hamiltonian dynamical system

Let us consider the Lagrangian density

L = [ 12 ψx ψt − ψx3 − 12 ψxx
2
]. (1.102)

Then, the Euler–Lagrange equation of motion for the ψ field becomes

ψxt − 6ψx ψxx + ψxxxx = 0. (1.103)

Defining
u = ψx (1.104)
(1.103) is seen to reduce to the KdV equation (1.45). Thus (1.102) may be
considered as the Lagrangian of the KdV equation through the potential field
function ψ(x, t).
Defining now the canonically conjugate momentum

∂L ψx
π= = (1.105)
∂ψt 2

the Hamiltonian density becomes


1 2 1 2
H = ψxx + ψx3 = πx2 + ψxx + 2π 2 ψx + πψx2 . (1.106)
2 4

Then the Hamiltonian of the KdV equation is


1 2
H= πx + ψxx + 2π ψx + πψx dx.
2 2 2
(1.107)
4

Copyright © 2003 IOP Publishing Ltd.

 Now using the expression (1.106) for the Hamiltonian density into the Hamilton’s
equation of motion, we can readily derive

ψt = ψx2 + 4ψx π − 2πxx (1.108a)

πt = 4ππx + 2πx ψx + 2πψxx − 1
2 ψxxxx . (1.108b)

Then with the substitution π = ψx /2, one can easily check that the evolution
equation for ψx or π is identical from both (1.108a) and (1.108b). It also coincides
with the KdV ψ field equation (1.103) as it should be. One can, thus, give both a
Lagrangian and Hamiltonian description for the KdV equation and conclude that
it is a Hamiltonian continuous system in the dynamical sense.
One can also give an alternative Hamiltonian description, by writing (1.106)
into terms of the KdV field function u = ψx :
∞  
1 2 3
1 2 3
H = ux + u H= ux + u dx. (1.109)
2 −∞ 2

Then writing the Hamiltonian equation of motion for a single field in the form
(for further details see, for example, [9])

∂ δH
ut = (1.110)
∂x δu
we obtain the KdV equation, after using the definition for the functional
derivative.

1.10.2 Complete integrability of the KdV equation

In order to understand the complete integrability property of the KdV equation, it
is more convenient to use the Hamiltonian equation (1.110) than the standard
form. Correspondingly the definition of the Poisson bracket between two
functionals U and V (for the KdV equation) can be defined [9]
∞
δU ∂ δV
{U, V } = dx . (1.111)
−∞ δu(x) ∂x δu(x)

We know that any transformation from one set of canonical variables (p, q) to a
new set (P , Q) is canonical provided the Poisson brackets of the new set satisfy
the relations

{P , P } = 0 {Q, Q} = 0 {P , Q} = δ(x − x
). (1.112)

Or in other words, if such a transformation exists, then the relation (1.112) ensures
that P and Q are, indeed, canonical variables.
It so happens that for the KdV equation one can find a suitable canonical
transformation from the continuous field variable u(x) to a new set of canonical

Copyright © 2003 IOP Publishing Ltd.

 variables (Pi , Qi ) and (P (k), Q(k)), i = 1, 2, . . . , N and −∞ < k < ∞, so that
the latter are infinite in number. More interestingly, one can prove that the P s
and Qs are not only canonical variables but also the action and angle variables,
respectively, of the KdV system. Consequently, the Hamiltonian (1.109) can be
written purely as a function of the action variables, Pi s and P (k)s, alone. The
resulting equation of motion can be obviously integrated trivially and, in this
Liouville sense, the KdV system becomes a completely integrable but, infinite-
dimensional (or degrees of freedom), nonlinear dynamical system.
Now, how can one find such a canonical transformation (CT)? Indeed one
finds that the direct scattering transform, which we discussed in section 1.8, does
correspond to such a CT and that the scattering data S(t) provides the necessary
coordinates to construct the required action and angle variables. Although the
analysis is somewhat involved, but direct, it has been successfully performed by
Zakharov and Faddeev in 1981 [13], see also Ablowitz and Clarkson [9] for fuller
details.
Further, one can also show that in terms of these new canonical variables the
Hamiltonian H
given by (1.109) becomes


N ∞

= − 32
H
5/2
Pj +8 k 3 P (k) dk. (1.113)
5 j =1 −∞

Thus, the Hamiltonian is a function of the action variables (momenta) only. So

the resultant equations of motion can be trivially integrated and solved for the
variables Qi (t), Pi (t), Q(k, t) and P (k, t), i = 1, 2, . . . , N, in terms of their
initial values. As a result, the KdV equation can be considered as a completely
integrable infinite-dimensional dynamical system. Finally it has been realized in
recent times that there is need for some modifications to the Poisson bracket
structure given by (1.111) due to certain technical difficulties in connection
with the satisfaction of Jacobi identity for certain class of functionals, see, for
example, [14].

1.11 Infinite number of conserved densities

Using the KdV equation (1.45), one can write the following conservation laws
easily:

ut + (−3u2 + uxx )x = 0 (1.114a)

(u )t + (−4u + 2uuxx − u2x )x = 0
2 3
(1.114b)
( 12 u2x + u3 )t + (3u2 uxx + ux uxxx − 92 u4 − 6uu2x − 12 u2xx )x = 0. (1.114c)

Note that (1.114b) is obtained by multiplying with u throughout the KdV

equation. Similarly (1.114c) can be obtained. These equations are in the so-called
conservative form and they correspond to conservation laws, because they can be

Copyright © 2003 IOP Publishing Ltd.

 written as
∂P ∂Q
+ =0 (1.115)
∂t ∂x
where P and Q are functions of u, ux , . . . , such that they vanish at x →
|x|→∞
±∞, since u −→ 0 sufficiently fast. If P and Q are connected by a gradient
relationship, that is P = Fx , then (1.103) gives Q = −Ft . Integrating (1.115), we
have ∞
∂
P dx = 0. (1.116)
∂t −∞
In other words, each of ∞
I= P dx (1.117)
−∞
constitutes a conserved quantity of the KdV equation. In particular from (1.117)
we see that
∞ ∞ ∞  
1 2
I1 = u dx I2 = u2 dx I3 = ux + u3 dx (1.118)
−∞ −∞ −∞ 2

are specific integrals of motion. Note that I3 = H

is the Hamiltonian of the
system, see (1.109).
Interestingly, the KdV equation possesses many more conservation laws and
constants of motion: in fact they are infinitely many in number. In order to realize
them, we can proceed as follows.
Introducing the so-called Gardner transformation [15]

u = ω + ωx +  2 ω2 (1.119)

where  is a small parameter and substituting it into the KdV equation, one can
show that u is a solution of the KdV equation provided

ωt − 6(ω +  2 ω2 )ωx + ωxxx = 0. (1.120)

Expressing ω now formally as a power series in ,


∞
ω(x, t; ) =  n ωn (x, t) (1.121)
n=0

and substituting it into (1.120), one can equate each power of  separately to zero.
Then one obtains the following conservation laws:

O( 0 ) : (ω0 )t = (3ω02 − ω0xx )x (1.122a)

O( ) : (ω1 )t = (6ω0 ω1 − ω1xx )x
1
(1.122b)
O( ) :
2
(ω2 )t = (3ω12 + 6ω0 ω 2
+ 2ω02 − ω2xx )x (1.122c)

Copyright © 2003 IOP Publishing Ltd.

 and so on. However, from (1.119), again comparing the coefficients of powers of
, one finds

ω0 = u ω1 = ux ω2 = uxx + u2 etc. (1.123)

Substituting (1.123) into the conservation laws (1.122), one can obtain the
previous conservation laws (1.114) again, as well as further conservation laws
which are infinite in number.
Finally, one can also check that the infinite number of integrals of motion
arising from this are functionally independent and involutive as the Poisson
brackets among them vanish:
∞
δIn ∂ δIm
{In , Im } = dx = 0. (1.124)
−∞ δu(x) ∂x δu(x)
This is yet another property indicative of the complete integrability of the KdV
equation. One can also proceed further and show that the existence of these infinite
number of conserved quantities is intimately related to the existence of an infinite
number of generalized symmetries, the so-called Lie–Bäcklund symmetries. For
more details, see, for example, [16].

1.12 Bäcklund transformations

Next, we point out another more important property of the KdV equation, namely
that it admits the so called (auto) Bäcklund transformation (BT). A BT is a
transformation which connects the solutions of two differential equations. If the
transformation connects two distinct solutions of the same equation, then it is
called an auto-Bäcklund transformation. The existence of such an auto-Bäcklund
transformation is indicative of the existence of soliton solutions and, in some
sense, integrability of the system as well. There are several ways of obtaining
such Bäcklund transformations but we will consider the BT for KdV equation
without its actual derivation. For some details, see, for example, [17, 18].
Now introducing the transformation u = ψx , the KdV equation be-
comes (1.103). Integrating it with respect to x and taking the integration ‘constant’
to be zero without loss of generality, we obtain

ψt − 3ψx2 + ψxxx = 0. (1.125)

Equation (1.125) is often called the potential KdV equation. If ω and ω are any
two solutions of the potential KdV equation (1.125), then the auto-Bäcklund
transformation of it is

ωx + ωx + 2κ 2 + 12 (ω − ω)2 = 0 (1.126a)
ωt + ωt − 3(ωx − ω x )(ωx + ωx ) + ωxxx − ωxxx = 0 (1.126b)

where κ is a real parameter. Then the two equations are compatible with (1.125).

Copyright © 2003 IOP Publishing Ltd.

 The use of such a Bäcklund transformation is immediately obvious. If ω = 0,
the trivial solution to (1.125), then solving (1.126), we obtain the one-soliton
solution
ω(x, t) = −2κ tanh{κ(x − 4κ 2 t) + δ} (1.127a)
so that

u(x, t) = ωx = −2κ 2 sech2 {κ(x − 4κ 2t) + δ} (1.127b)

which is nothing but the one-soliton solution of the KdV equation. One can then
use the one-soliton solution in (1.126) as the new ‘seed’ solution and obtain two-
soliton solution and the process can be continued to obtain higher-order solitons.
For more details, we refer to [9, 17, 18].

1.13 The Painlevé property for the KdV equations

It is now well recognized that a systematic approach to determine whether a
nonlinear PDE is integrable or not is to investigate the singularity structure of
the solutions, namely the Painlevé property. This approach, which was originally
suggested by Weiss, Tabor and Carnevale [19] (WTC), aims to determine the
presence or absence of movable non-characteristic manifolds (of branching type,
both algebraic and logarithmic, and essential singular). When the system is free
from movable critical singular manifolds so that the solution is single-valued,
the Painlevé property holds, suggesting its integrability. Otherwise, the system is
non-integrable.
It has been shown by WTC [19] that the KdV equation is, indeed, free from
movable critical singular manifolds by expanding it locally as a Laurent series
and analysing its structure. From such an analysis one can also establish the other
integrability properties such as the Lax pair, Hirota bilinearization, Bäcklund
transformation, etc. (For details see chapters 2 and 3 on P-analysis in this book
and also [9, 12].)

1.14 Lie and Lie–Bäcklund symmetries

The various integrability properties of the KdV equation discussed earlier can also
be traced to the existence of various symmetry transformations under which the
KdV equation remains form invariant. The simplest among them is the set of one-
parameter continuous group of Lie point symmetries [16,21], whose infinitesimal
generators can be expressed in the form
∂ ∂ ∂
X = ξ(t, x, u) + τ (t, x, u) + η(t, x, u) (1.128)
∂x ∂t ∂u
where ξ, τ and η are the infinitesimals associated with the variables x, t and u,
respectively, of equation (1.3). One can easily show that by solving the underlying

Copyright © 2003 IOP Publishing Ltd.

 system of linear partial differential equations for ξ , τ and η, arising from the
form invariance of (1.3), the KdV equation admits a set of four one-parameter Lie
symmetries corresponding to time and space translations, scaling and Galilean
invariance. Consequently, one can reduce the KdV equation into ODEs in terms of
appropriate similarity variables by solving the associated characteristic equations.
In this way one obtains the travelling wave form (1.5) as well as reduction of
the KdV equation to first or second Painlevé transcendental equations (for more
details, see for example, [16, 21]).
One can go beyond the Lie point symmetries and show that the KdV equation
admits generalized Lie–Bäcklund symmetries involving not only the variables t, x
and u but also the derivatives ux , uxx , uxxx , etc. In fact, one can show that an
infinite number of Lie–Bäcklund symmetries, whose infinitesimal generators are
of the form
∂ ∂
X = ξ(t, x, u, ux , uxx , . . .) + τ (t, x, u, ux , uxx , . . .)
∂x ∂t
∂
+ η(t, x, u, ux , uxx , . . .) (1.129)
∂u
also exist. The first few of them are, for example, given by
∂
X1 = ux (1.130)
∂u
∂
X2 = −(6uux + uxxx ) (1.131)
∂u
∂
X3 = (uxxxxx + 10uuxxx + 10ux uxx + 30u2 ux ) (1.132)
∂u
and further symmetries can be identified through a non-local recursion
operator [16]
x
R = −Dx2 − 4u − 2ux Dx−1 Dx−1 = dx (1.133)
−∞

such that RXi = Xi+1 , which is a new generator of symmetries. Associated with
each of these symmetries, one can identify an independent conserved quantity
and thereby clarify the origin of the existence of the infinite number of involutive
integrals of motion. The existence of Lax pair and Bäcklund transformations
can also be related to the existence of recursion operator and Lie–Bäcklund
symmetries, thereby giving a group theoretical interpretation of the complete
integrability of the KdV equation.

1.15 Conclusion
We have come a long way from Scott Russell’s observation of solitary waves
in 1834 to the modern-day methods of identification of completely integrable

Copyright © 2003 IOP Publishing Ltd.

 infinite-dimensional nonlinear dynamical systems. In particular, the KdV
equation is a prototypical example of a soliton possessing a completely integrable
system and, from a physical point of view, it is ubiquitous. It possesses many
remarkable features: Lax pair, N-soliton solutions, IST solvability, Hamiltonian
structure, infinite number of conservation laws, symmetries and constants of
motion, Bäcklund transformation, Hirota bilinearization and Painlevé property
to name but a few. Many other geometrical and group theoretical properties
can also be ascribed to it. For example, it possesses a four-parameter group of
Lie symmetries, which one can use to identify interesting similarity variables.
In turn, the KdV equation can be reduced to ODEs for wave solutions and
Painlevé transcendental equations for similarity solutions [21]. Moreover, the
KdV equation possesses an infinite number of Lie–Bäcklund symmetries and
making use of them the infinite number of constants of motion can also be found
in a systematic way.
It would have been entirely fortuitous if KdV equation were to be an isolated
case possessing all these remarkable properties. Fortunately, now we know that a
very large class of nonlinear evolution equations including sine-Gordon, nonlinear
Schrödinger, modified KdV, Heisenberg spin, Toda lattice, etc equations in
(1 + 1) dimensions also possess the same general features as the KdV equation [9,
11, 15]. Even (2 + 1)-dimensional systems such as the Kadomtsev–Petviashvili,
Davey–Stewartson, Ishimori, Novikov–Nizhnik–Veselov, etc equations [20] have
been found to be integrable. The list is getting extended in different directions
in (1 + 1)- and (2 + 1)-dimensional nonlinear partial differential equations,
differential-difference equations, difference equations and their corresponding
quantum versions, some of which are discussed in this volume. All these
developments make the field of integrable systems highly exciting and vibrant.
One can confidently state that all these developments became possible due to our
understanding of the KdV equation, which continues to play a pivotal role in
nonlinear dynamics.

Acknowledgments
I wish to thank Mr T Kanna for help in preparing this article and the Department
of Science and Technology, Government of India for support.

References
[1] Scott Russell J 1844 Report on Waves British Association Reports
[2] Korteweg D J and de Vries G 1895 Phil. Mag. 39 422
[3] Kruskal M D and Zabusky N J Progress on the Fermi–Pasta–Ulam nonlinear
string problem Princeton Plasma Physics Laboratory Annual Report MATT-Q-
21, Princeton, pp 301–8
[4] Zabusky N J and Kruskal M D 1965 Phys. Rev. Lett. 15 240

Copyright © 2003 IOP Publishing Ltd.

 [5] Gardner C S, Greene J M, Kruskal M D and Miura R M 1967 Phys. Rev. Lett. 19
1095
[6] Bullough R K 1988 ‘The wave par excellence’, The solitary progressive great wave of
equilibrium of the fluid: an early history of the solitary wave Solitons: Introduction
and Applications ed M Lakshmanan (New York: Springer)
Bullough R K and Caudrey P J 1995 Acta Appl. Math. 39 193
[7] Ford J 1992 Phys. Reports 213 271
[8] See, for example, Hirota R 1976 Bäcklund Transformations ed R M Miura (New
York: Springer)
[9] Ablowitz M J and Clarkson P A 1992 Solitons, Nonlinear Evolution Equations and
Inverse Scattering (Cambridge: Cambridge University Press)
[10] Miura R M 1968 J. Math. Phys. 9 1202
[11] Lax P D 1968 Commun. Pure Appl. Math. 21 467
[12] Lakshmanan M and Rajasekar S 2003 Nonlinear Dynamics: Integrability, Chaos and
Patterns (New York: Springer)
[13] Zakhrov V E and Faddeev L D 1971 Funct. Anal. Appl. 5 280
[14] Kundu A and Basu-Mallick B 1990 J. Phys. Soc. Japan 59 1560
Kundu A and Basu-Mallick B 1990 J. Phys. A 23 L709
[15] Gardner C S 1971 J. Math. Phys. 12 1548
[16] Bluman G W and Kumei S 1989 Symmetries and Differential Equations (New York:
Springer)
[17] Scott A C 1999 Nonlinear Science: Emergence and Dynamics of Coherent Structures
(Oxford: Oxford University Press)
[18] Rogers C and Shadwick W F 1982 Bäcklund Transformations and Applications (New
York: Academic Press)
[19] Weiss J, Tabor M and Carnevale G 1983 J. Math. Phys. 24 522
[20] Konopelchenko B G 1993 Solitons in Multidimensions (Singapore: World Scientific)
[21] Lakshmanan M and Kaliappan P 1983 J. Math. Phys. 24 795

Copyright © 2003 IOP Publishing Ltd.

 Chapter 2

The Painlevé methods

R Conte† and M Musette‡
† Service de physique de l’état condensé (URA 2464),

CEA–Saclay, Gif-sur-Yvette, France

‡ Dienst Theoretische Natuurkunde, Vrije Universiteit Brussel,

Belgium

2.1 The classical programme of the Painlevé school and its

achievements
It is impossible to understand anything to the Painlevé property without keeping
in mind the original problem as stated by L Fuchs, Poincaré and Painlevé: to
define new functions from ordinary differential equations (ODEs). This simply
formulated problem implies selecting those ODEs whose general solution can be
made single-valued by some uniformization procedure (cuts, Riemann surface),
so as to fit the definition of a function. This property (the possibility to uniformize
the general solution of an ODE), nowadays called the Painlevé property (PP), is
equivalent to the more practical definition.

Definition 2.1 The Painlevé property of an ODE is the absence of movable

critical singularities in its general solution.

Let us recall that a singularity is said to be movable (as opposed to fixed) if its
location depends on the initial conditions, and critical if multi-valuedness takes
place around it. Other definitions of the PP excluding, for instance, the essential
singularities or replacing ‘movable critical singularities’ by ‘movable singularities
other than poles’ or ‘its general solution’ by ‘all its solutions’ are incorrect. Two
examples taken from Chazy [1] explain why this is so. The first example is the
celebrated Chazy class III equation

− 2uu

+ 3u
2 = 0 (2.1)

Copyright © 2003 IOP Publishing Ltd.

 whose general solution is only defined inside or outside a circle characterized by
the three initial conditions (two for the centre, one for the radius); this solution
is holomorphic in its domain of definition and cannot be analytically continued
beyond it. This equation, therefore, has the PP and the only singularity is a
movable analytic essential singular line which is a natural boundary.
The second example [1, p 360] is the third-order second-degree ODE

(u

− 2u
u

)2 + 4u

2 (u

− u
2 − 1) = 0 (2.2)

whose general solution is single-valued,

ec1 x+c2 c2 − 4
u= + 1 x + c3 (2.3)
c1 4c1

but which also admits a singular solution (envelope solution) with a movable
critical singularity,
u = C2 − log cos(x − C1 ). (2.4)

For more details, see the arguments of Painlevé [2, section 2.6] and Chazy [2,
section 5.1].
The PP is invariant under an arbitrary homography on the dependent variable
and an arbitrary change of the independent variable (homographic group)

α(x)U (X) + β(x)

(u, x) → (U, X) u(x) = X = ξ(x)
γ (x)U (X) + δ(x) (2.5)
(α, β, γ , δ, ξ ) functions αδ − βγ = 0.

Every linear ODE possesses the PP since its general solution depends
linearly on the movable constants so, in order to define new functions, one must
turn to nonlinear ODEs in a systematic way: first-order algebraic equations, then
second-order, etc. The current achievements are the following.
First-order algebraic ODEs (polynomial in u, u
, analytic in x) define only
one function, the Weierstrass elliptic function ℘, new in the sense that its ODE

u
2 − 4u3 + g2 u + g3 = 0 (g2 , g3 ) arbitrary complex constants (2.6)

is not reducible to a linear ODE. Its only singularities are movable double poles.
Second-order algebraic ODEs (polynomial in u, u
, u

, analytic in x) define
six functions, the Painlevé functions Pn, n = 1, . . . , 6, new because they are
not reducible to either a linear ODE or a first-order ODE. This question of
irreducibility, the subject of a long dispute between Painlevé and Joseph Liouville,
has been rigorously settled only recently [3]. The canonical representatives of

Copyright © 2003 IOP Publishing Ltd.

 P1–P6 in their equivalence class under the group (2.5) are:

P1 : u

= 6u2 + x
P2 : u

= 2u3 + xu + α
u
2 u
αu2 + γ u3 β δ
P3 : u

= − + 2
+ +
u x 4x 4x 4u
u
2 3 β
P4 : u

= + u3 + 4xu2 + 2x 2u − 2αu +
2u 2 u


2


1 1
2 u (u − 1) β u u(u + 1)
P5 : u = + u − + αu + +γ +δ
2u u − 1 x x 2 u x u−1



1 1 1 1 1 1 1
P6 : u

= + + u
2 − + + u

2 u u−1 u−x x x−1 u−x


u(u − 1)(u − x) x x−1 x(x − 1)
+ α+β 2 +γ +δ
x 2 (x − 1)2 u (u − 1)2 (u − x)2
in which α, β, γ , δ are arbitrary complex parameters. Their only singularities
are movable poles (in the ex complex plane for P3 and P5, in the x-plane for
the others), with, in addition, three fixed critical singularities for P6, located at
x = ∞, 0, 1.
Third- and higher-order ODEs [1, 4–6] have not yet defined new functions.
Although there are some good candidates (the Garnier system [7], several fourth-
order ODEs [5,8], which all have a transcendental dependence on the constants of
integration), the question of their irreducibility (to a linear, Weierstrass or Painlevé
equation) is very difficult and still open. To understand the difficulty, it is sufficient
to consider the fourth-order ODE for u(x) defined by

u = u1 + u2 u

1 = 6u21 + x u

2 = 6u22 + x. (2.7)

This ODE (easy to write by the elimination of u1 , u2 ) has a general solution which
depends transcendentally on the four constants of integration and it is reducible.
The master equation P6 was first written by Picard in 1889 in a particular
case, in a very elegant way. Let ϕ be the elliptic function defined by
ϕ
dz
ϕ : y → ϕ(y, x) y= √ (2.8)
∞ z(z − 1)(z − x)
and let ω1 (x), ω2 (x) be its two half-periods. Then the function

u : x → u(x) = ϕ(2c1 ω1 (x) + 2c2 ω2 (x), x) (2.9)

with (c1 , c2 ) arbitrary constants, has no movable critical singularities and it

satisfies a second-order ODE which is P6 in the particular case α = β = γ =
δ − 1/2 = 0. The generic P6 was found simultaneously from two different
approaches: the nonlinear one of the Painlevé school as stated earlier [9]; and

Copyright © 2003 IOP Publishing Ltd.

 the linear one of R Fuchs [10] as an isomonodromy condition. In the latter, one
considers a second-order linear ODE for ψ(t) with four Fuchsian singularities
of cross-ratio x (located, for instance, at t = ∞, 0, 1, x) with, in addition, as
prescribed by Poincaré for the isomonodromy problem, one apparent singularity
t = u,

2 d2 ψ A B C D 3
− = 2 + + + +
ψ dt 2 t (t − 1) 2 (t − x) 2 t (t − 1) 4(t − u)2
a b
+ + (2.10)
t (t − 1)(t − x) t (t − 1)(t − u)
(A, B, C, D denote constants and a, b parameters). The requirement that the
monodromy matrix (which transforms two independent solutions ψ1 , ψ2 when
t goes around a singularity) be independent of the non-apparent singularity x
results in the condition that u, as a function of x, satisfies P6.
A useful by-product of this search for new functions is the construction
of several exhaustive lists (classifications) of second [11–15], third [1, 4, 6],
fourth [4,5] or higher-order [16] ODEs, whose general solution is explicitly given
because they have the PP. Accordingly, if one has an ODE in such an already
well studied class (e.g. second-order second-degree binomial-type ODEs [14]
u

2 = F (u
, u, x) with F rational in u
and u, analytic in x), and which is
suspected to have the PP (for instance, because one has been unable to detect any
movable critical singularity, see section 2.3), then two cases are possible: either
there exists a transformation (2.5) mapping it to a listed equation, in which case
the ODE has the PP and is explicitly integrated; or such a transformation does not
exist and the ODE does not have the PP.

2.2 Integrability and Painlevé property for partial differential

equations
Defining the PP for PDEs is not easy but this must be done for future use in
sections 2.3 (the Painlevé test) and 2.4 (proving the PP). Such a definition must
involve a global, constructive property, which excludes the concept of a general
solution. Indeed, it is only in non-generic cases like the Liouville equation that
the general solution of a PDE can be built explicitly. This is where the Bäcklund
transformation comes in. Let us first recall the definition of this powerful tool
(for simplicity, but this is not a restriction, we give the basic definitions for a
PDE defined as a single scalar equation for one dependent variable u and two
independent variables (x, t)).

Definition 2.2 [17, vol III ch XII, 18]. A Bäcklund transformation (BT) between
two given PDEs,

E1 (u, x, t) = 0 E2 (U, X, T ) = 0 (2.11)

Copyright © 2003 IOP Publishing Ltd.

 is a pair of relations

Fj (u, x, t, U, X, T ) = 0 j = 1, 2 (2.12)

with some transformation between (x, t) and (X, T ), in which Fj , depends on

the derivatives of u(x, t) and U (X, T ), such that the elimination of u (resp. U )
between (F1 , F2 ) implies E2 (U, X, T ) = 0 (resp. E1 (u, x, t) = 0). When the two
PDEs are the same, the BT is also called the auto-BT.
Under a reduction PDE → ODE, the BT reduces to a birational
transformation (also with the initials BT!), which is not involved in the definition
of the PP for ODEs. Therefore, one needs an intermediate (and quite important)
definition before defining the PP.

Definition 2.3 A PDE in N independent variables is integrable if at least one of

the following properties holds.
(1) Its general solution can be obtained, and it is an explicit closed form
expression, possibly presenting movable critical singularities.
(2) It is linearizable.
(3) For N > 1, it possesses an auto-BT which, if N = 2, depends on an arbitrary
complex constant, the Bäcklund parameter.
(4) It possesses a BT to another integrable PDE.
Examples of these various situations√ are, respectively: the PDE ux ut +
uuxt = 0 with the general solution u = f (x) + g(t), which presents movable
critical singularities and can be transformed into the d’Alembert equation; the
Burgers PDE ut + uxx + 2uux = 0, linearizable into the heat equation ψt +
ψxx = 0; the KdV PDE ut + uxxx − 6uux = 0, which is integrable by the
inverse spectral transform (IST) [19]; and the Liouville PDE uxt + eu = 0, which
possesses a BT to the d’Alembert equation ψxt = 0.
We now have enough elements to give a definition of the PP for PDEs which
is indeed an extrapolation of the one for ODEs.

Definition 2.4 The Painlevé property (PP) of a PDE is its integrability

(definition 2.3) and the absence of movable critical singularities near any non-
characteristic manifold.
From this, one can see that the PP is a more demanding property than mere
integrability.
The PP for PDEs is invariant under the natural extension of the homographic
group (2.5) and classifications similar to those of ODEs have also been performed
for PDEs, in particular second-order first-degree PDEs [20,21]; however, only the
already known PDEs (Burgers, Liouville, sine-Gordon, Tzitzéica, etc) have been
isolated. Classifications based on other criteria, such as the existence of an infinite
number of conservation laws [22], isolate more PDEs, which are all probably

Copyright © 2003 IOP Publishing Ltd.

 integrable in the sense of definition 2.3; it would be interesting to check that,
under the group of transformations generated by Bäcklund transformations and
hodograph transformations, each of them is equivalent to a PDE with the PP.
If one performs a hodograph transformation (typically an exchange of the
dependent and independent variables u and x) on a PDE with the PP, the
transformed PDE possesses a weaker form of the PP in which, for instance, all
leading powers and Fuchs indices become rational numbers instead of integers.
Details can be found, for example, in [23]. For instance, the Harry-Dym,
Camassa–Holm [24] and DHH equations [25] can all be mapped to a PDE with
the PP by some hodograph transformation.
If definition 2.4 of the PP for PDEs is really an extrapolation of the one for
ODEs, then, given a PDE with the PP, every reduction to an ODE which preserves
the differential order (i.e. a non-characteristic reduction) yields an ODE which
necessarily has the PP. This proves the conjecture by Ablowitz et al [26], provided
definition 2.4 really extrapolates that for ODEs and this is the difficult part of this
question.
There are plenty of examples of such reductions, for instance the self-dual
Yang–Mills equations admit reductions to all six Pn equations [27].
Given a PDE, the process to prove its integrability is twofold. One must first
check whether it may be integrable, for instance by applying the Painlevé test, as
will now be explained in section 2.3. Then, in the case of a non-negative answer,
one must build explicitly the elements which are required to establish integrability,
for instance by using methods described in section 2.4.
Although partially integrable and non-integrable equations, i.e. the majority
of physical equations, admit no BT, they retain part of the properties of (fully)
integrable PDEs and this is why the methods presented here apply to both cases
as well. One such example is included for information, see section 2.4.5.

2.3 The Painlevé test for ODEs and PDEs

The test is an algorithm providing a set of necessary conditions for the equation
to possess the PP. A full, detailed version can be found in [2, section 6]; here
we only give a short presentation of its main subset, known as the method of
pole-like expansions, due to Kowalevski (1889) and Gambier [11]. The generated
necessary conditions are a priori not sufficient: this was proven by Picard (1893),
who exhibited the example of the ODE with the general solution u = ℘ (λ log(x −
c1 ) + c2 , g2 , g3 ), namely

u
2  g2  u
2
u

− 6u 2
− − =0 (2.13)
4u − g2 u − g3
3 2 λ 4u3 − g2 u − g3

which has the PP iff 2πiλ is a period of the elliptic function ℘, a transcendental
condition on (λ, g2 , g3 ) impossible to obtain in a finite number of algebraic steps

Copyright © 2003 IOP Publishing Ltd.

 such as the Painlevé test. Therefore, it is wrong to issue statements like ‘The
equation passes the test, therefore it has the PP’. The only way to prove the PP is:
• for an ODE, either to explicitly integrate with the known functions (solutions
of linear, elliptic, hyperelliptic (a generalization of elliptic) or Painlevé
equations) or, if one believes a new function has been found, to prove both
the absence of movable critical singularities and irreducibility [3] to known
functions; and
• for a PDE, to build the integrability elements of definition 2.4 explicitly.
Let us return to the test itself. It is sufficient to present the method of pole-
like expansions for one Nth-order equation
E(x, t, u) = 0 (2.14)
in one dependent variable u and two independent variables x, t. Movable
singularities lie on a codimension-one manifold
ϕ(x, t) − ϕ0 = 0 (2.15)
in which the singular manifold variable ϕ is an arbitrary function of the
independent variables and ϕ0 an arbitrary movable constant. Basically [28], the
WTC test consists in checking the existence of all possible local representations,
near ϕ(x, t) − ϕ0 = 0, of the general solution (whatever its definition, a difficult
problem for PDEs) as a locally single-valued expression, for example the Laurent
series
+∞
u= uj χ j +p −p∈N
j =0
(2.16)

+∞
j +q
E= Ej χ −q ∈N
j =0
with coefficients uj , Ej independent of the expansion variable χ. The natural
choice χ = ϕ − ϕ0 [28] generates lengthy expressions uj , Ej . Fortunately, there
is much freedom in choosing χ, the only requirement being that it vanishes as
ϕ − ϕ0 and one should have a homographic dependence on ϕ − ϕ0 so as not to
alter the structure of the movable singularities; hence, the result of the test. The
unique choice which minimizes the expressions and puts no constraint on ϕ is [29]


ϕ − ϕ0 ϕx ϕxx −1
χ= = − ϕx = 0 (2.17)
ϕx − (ϕxx /2ϕx )(ϕ − ϕ0 ) ϕ − ϕ0 2ϕx
in which x denotes an independent variable whose component of grad ϕ does
not vanish. The expansion coefficients uj , Ej are then invariant under the six-
parameter group of homographic transformations
a
ϕ + b

ϕ → a
d
− b
c
= 0 (2.18)
c
ϕ + d

Copyright © 2003 IOP Publishing Ltd.

 in which a
, b
, c
, d
are arbitrary complex constants and these coefficients only
depend on the following elementary differential invariants and their derivatives:
the Schwarzian  
ϕxxx 3 ϕxx 2
S = {ϕ; x} = − (2.19)
ϕx 2 ϕx
and one other invariant per independent variable t, y, . . .

C = −ϕt /ϕx K = −ϕy /ϕx , . . . . (2.20)

The two invariants S, C are linked by the cross-derivative condition

X ≡ ((ϕxxx )t − (ϕt )xxx )/ϕx = St + Cxxx + 2Cx S + CSx = 0 (2.21)

identically satisfied in terms of ϕ.

For the practical computation of (uj , Ej ) as functions of (S, C) only, i.e.
what is called invariant Painlevé analysis, the variable ϕ disappears and the only
information required is the gradient of the expansion variable χ,
S 2 1
χx = 1 + χ χt = −C + Cx χ − (CS + Cxx )χ 2 . (2.22)
2 2
with the constraint (2.21) between S and C.
Consider, for instance, the Kolmogorov–Petrovskii–Piskunov (KPP)
equation [30, 31]

E(u) ≡ but − uxx + γ uux + 2d −2 (u − e1 )(u − e2 )(u − e3 ) = 0 (2.23)

with (b, γ , d 2 ) real and ej real and distinct, encountered in reaction–diffusion

systems (the convection term uux [32] is quite important in physical applications
to prey–predator models).
The first step, to search for the families of movable singularities u ∼
u0 χ p , E ∼ E0 χ q , u0 = 0, results in the selection of the dominant terms Ê(u):

Ê(u) ≡ −uxx + γ uux + 2d −2 u3 (2.24)

which provide two solutions (p, u0 ):

p = −1 q = −3 − 2 − γ u0 + 2d −2 u20 = 0. (2.25)

The necessary condition that all values of p be integer is satisfied.

The second step is, for every selected family, to compute the linearized
equation,

Ê(u + εw) − Ê(u)

(Ê
(u))w ≡ lim
ε→0 ε
= (−∂x + γ u∂x + γ ux + 6d −2 u2 )w = 0
2
(2.26)

Copyright © 2003 IOP Publishing Ltd.

 then its Fuchs indices i near χ = 0 as the roots of the indicial equation

P (i) = lim χ −i−q (−∂x2 + γ u0 χ p ∂x + γpu0 χ p−1 + 6d −2 u20 χ 2p )χ i+p (2.27)

χ→0

= −(i − 1)(i − 2) + γ u0 (i − 2) + 6d −2 u20 (2.28)

= −(i + 1)(i − 4 − γ u0 ) = 0 (2.29)

and finally to enforce the necessary condition that, for each family, these two
indices be distinct integers [33, 34]. Considering each family separately would
produce a countable number of solutions, which is incorrect. Considering the two
families simultaneously, the diophantine condition that the two values i1 , i2 of
the Fuchs index 4 + γ u0 be integer has a finite number of solutions, namely [35,
appendix I]

γ 2d 2 = 0 (i1 , i2 ) = (4, 4) u0 = (−d, d) (2.30)

γ d =2
2 2
(i1 , i2 ) = (3, 6) γ u0 = (−1, 2) (2.31)
2 2
γ d = −18 (i1 , i2 ) = (−2, 1) γ u0 = (−6, −3). (2.32)

It would be wrong at this stage to discard negative integer indices. Indeed, in

linear ODEs such as (2.26), the single-valuedness required by the Painlevé test
restricts the Fuchs indices to integers, whatever their sign. Let us proceed with
the first case only, γ = 0 (the usual KPP equation).
The recurrence relation for the next coefficients uj ,

∀j ≥ 1 : Ej ≡ P (j )uj + Qj ({ul | l < j }) = 0 (2.33)

depends linearly on uj and nonlinearly on the previous coefficients ul .

The third and last step is then to require, for any admissible family and any
Fuchs index i, that the no-logarithm condition

∀i ∈ Z P (i) = 0 : Qi = 0 (2.34)

holds true. At index i = 4, the two conditions, one for each sign of d [36],

Q4 ≡ C[(bdC + s1 − 3e1 )(bdC + s1 − 3e2 )(bdC + s1 − 3e3 )

− 3b2 d 3 (Ct + CCx )] = 0 (2.35)
s1 = e1 + e2 + e3

are not identically satisfied, so the PDE fails the test. This ends the test.
If, instead of the PDE (2.23), one considers its reduction u(x, t) = U (ξ ), ξ =
x − ct to an ODE, then C = constant = c, and the two conditions Q4 = 0 select
the seven values c = 0 and c2 = (s1 − 3ek )2 (bd)−2 , k = 1, 2, 3. For all these
values, the necessary conditions are then sufficient since the general solution U (ξ )
is single-valued (equation number 8 in Gambier list [11] reproduced in [37]).

Copyright © 2003 IOP Publishing Ltd.

 It frequently happens that the Laurent series (2.16) only represents a
particular solution, for instance because some Fuchs indices are negative integers,
for example the fourth-order ODE [4, p 79]
u

+ 3uu

− 4u
2 = 0 (2.36)
which admits the family
p = −2, u0 = −60 Fuchs indices (−3, −2, −1, 20). (2.37)
The series (2.16) depends on two, not four, arbitrary constants, so two are missing
and may contain multi-valuedness. In such cases, one must perform a perturbation
in order to represent the general solution and to test the missing part of the solution
for multi-valuedness.
This perturbation [34] is close to the identity (for brevity, we skip the t
variable)

+∞ 
+∞
x unchanged u= εn u(n) : E = εn E (n) = 0 (2.38)
n=0 n=0

where, as in the Painlevé α-method, the small parameter ε is not in the original
equation.
Then, the single equation (2.14) is equivalent to the infinite sequence
n=0 E (0) ≡ E(x, u(0) ) = 0 (2.39)

∀n ≥ 1 E (n) ≡ E
(x, u(0) )u(n) + R (n) (x, u(0) , . . . , u(n−1) ) = 0 (2.40)

with R (1) identically zero. From a basic theorem of Poincaré [2, theorem II,
section 5.3], necessary conditions for the PP are:
• the general solution u(0) of (2.39) has no movable critical points;
• the general solution u(1) of (2.40) has no movable critical points; and
• for every n ≥ 2, there exists a particular solution of (2.40) without movable
critical points.
Order zero is just the original equation (2.14) for the unknown u(0) , so one
takes for u(0) the already computed (particular) Laurent series (2.16).
Order n = 1 is identical to the linearized equation
E (1) ≡ E
(x, u(0) )u(1) = 0 (2.41)
and one must check the existence of N independent solutions u(1) locally single-
valued near χ = 0, where N is the order of (2.14).
The two main implementations of this perturbation are the Fuchsian
perturbative method [34] and the non-Fuchsian perturbative method [38]. In this
example (2.36), both methods indeed detect multi-valuedness, at perturbation
order n = 7 for the first one and n = 1 for the second one (details later).

Copyright © 2003 IOP Publishing Ltd.

 2.3.1 The Fuchsian perturbative method
Adapted to the presence of negative integer indices in addition to the ever present
value −1, this method [33, 34] generates additional no-log conditions (2.34).
Denoting by ρ the lowest integer Fuchs index, ρ ≤ −1, the Laurent series for


+∞
uj χ j +p
(1)
u(1) = (2.42)
j =ρ

represents a particular solution containing a number of arbitrary coefficients equal

to the number of Fuchs indices, counting their multiplicity. If this number equals
N, it represents the general solution of (2.41). Two examples will illustrate the
method [2, section 5.7.3].
The equation
u

+ 4uu
+ 2u3 = 0 (2.43)
possesses the single family
(0) (0) (0)
p = −1 E0 = u0 (u0 − 1)2 = 0 indices (−1, 0) (2.44)
(0)
with the puzzling fact that u0 should be both equal to 1, according to the equation
(0)
E0 = 0, and arbitrary, according to the index 0. The necessity of performing a
perturbation arises from the multiple root of the equation for u(0)
0 , responsible for
the insufficient number of arbitrary parameters in the zeroth-order series u(0) . The
application of the method provides

u(0) = χ −1 (the series terminates) (2.45)

E (x, u ) = ∂x2 + 4χ −1 ∂x + 2χ −2
(0)
(2.46)
−1
u(1) = u(1)
0 χ u(1)
0 arbitrary (2.47)

(0) (1)2 (1) (1)

(2)
E = E (x, u
(0) (2)
)u + 6u u + 4u u
(1)2
= χ −2 (χ 2 u(2))

+ 2u0 χ −3 = 0 (2.48)
(1)2
u(2) = −2u0 χ −1 (log χ − 1). (2.49)

The movable logarithmic branch point is, therefore, detected in a systematic way
at order n = 2 and index i = 0. This result was, of course, found long ago by the
α-method [39, section 13, p 221].
Equation (2.36) possesses the two families:

u0 = −60, indices (−3, −2, −1, 20), Ê = u

+ 3uu

− 4u
2
(0)
p = −2
(2.50)

2
p = −3 u(0)
0 arbitrary, indices (−1, 0), Ê = 3uu − 4u . (2.51)

Copyright © 2003 IOP Publishing Ltd.

 The second family has a Laurent series (p : +∞) which happens to terminate
[34]:
u(0) = c(x − x0 )−3 − 60(x − x0 )−2 (c, x0 ) arbitrary. (2.52)
For this family, the Fuchsian perturbative method is then useless, because the two
arbitrary coefficients corresponding to the two Fuchs indices are already present
at zeroth order.
The first family provides, at zeroth order, only a two-parameter expansion
and, when one checks the existence of the perturbed solution
 

+∞ +∞
(n) j −2−3n
u= ε n
uj χ (2.53)
n=0 j =0
(0) (1) (1) (1)
one finds that coefficients u20 , u−3 , u−2 , u−1 can be chosen arbitrarily and, at
order n = 7, one finds two violations [34]:
7 2 6
Q(7) (0) (1)
−1 ≡ u20 u−3 = 0 Q(7) (0) (1) (1)
20 ≡ u20 u−3 u−2 = 0 (2.54)
implying the existence of a movable logarithmic branch point.

2.3.2 The non-Fuchsian perturbative method

Whenever the number of indices is less than the differential order of the equation,
the Fuchsian perturbative method fails to build a representation of the general
solution, thus possibly missing some no-log conditions. The missing solutions of
the linearized equation (2.41) are then solutions of the non-Fuchsian type near
χ = 0.
In section 2.3.1, the fourth-order equation (2.36) has been shown to fail the
test after a computation practically untractable without a computer. Let us prove
the same result without computation at all [38]. The linearized equation
E (1) = E
(x, u(0) )u(1) ≡ [∂x4 + 3u(0)∂x2 − 8u(0)
x ∂x + 3uxx ]u
(0) (1)
=0 (2.55)
is known globally for the second family because the two-parameter solution (2.52)
is closed form; therefore, one can test all the singular points χ of (2.55). These
are χ = 0 (non-Fuchsian) and χ = ∞ (Fuchsian) and the key to the method is
the information obtainable from χ = ∞. Let us first lower by two units the order
of the linearized equation (2.55), by ‘subtracting’ the two global single-valued
solutions u(1) = ∂x0 u(0) and ∂c u(0), i.e. u(1) = χ −4 , χ −3 ,
u(1) = χ −4 v : [∂x2 − 16χ −1 ∂x + 3cχ −3 − 60χ −2 ]v

= 0. (2.56)
Then the local study of χ = ∞ is unnecessary, since one recognizes the Bessel
equation. The two other solutions in global form are:

c = 0 : v1

= χ −3 0 F1 (24; −3c/χ) = χ 17/2 J23 ( 12c/χ) (2.57)

v2 = χ 17/2
N23 ( 12c/χ) (2.58)

Copyright © 2003 IOP Publishing Ltd.

 where the hypergeometric function 0 F1 (24; −3c/χ) is single-valued and
possesses an isolated essential singularity at χ = 0, while the Neumann function
N23 is multi-valued because of a log χ term.

2.4 Singularity-based methods towards integrability

In this section, we review a variety of singularity-based methods able to provide
some global elements of integrability. The singular manifold method of Weiss
et al [28] is the most important of them but it is not the only one.
A prerequisite notion is the singular part operator D,
log ϕ → D log ϕ = uT (0) − uT (∞) (2.59)
in which the notation uT (ϕ0 ), which emphasizes the dependence on ϕ0 , stands for
the principal part (T -truncated) of the Laurent series (2.16),
−p

uT (ϕ0 ) = uj χ j +p . (2.60)
j =0

In our KPP example (2.25) with γ = 0, this operator is D = d∂x .

2.4.1 Linearizable equations

When a nonlinear equation can be linearized, the singular part operator defined in
(2.59) directly defines the linearizing transformation.
For instance, the Kundu–Eckhaus PDE for the complex field U (x, t) [40,41]
 
β2
iUt + αUxx + |U |4 + 2b eiγ (|U |2 )x U = 0, (α, β, b, γ ) ∈ R
α
(2.61)
with αβb cos γ = 0, passes the test iff [42, 43] b2 = β 2 . Under the parametric
representation
√
U = ux eiθ (2.62)
the equivalent fourth-order PDE for u [43]
α β 2 − (b sin γ )2 4
(uxxxx u2x + u3xx − 2ux uxx uxxx ) + 2 ux uxx
2 α
1
+ 2(b cos γ )u3x uxxx + (ut t u2x + uxx u2t − 2ut ux uxt ) = 0 (2.63)
2α
admits two families, namely in the case b2 = β 2 ,
1
u log ψ indices (−1, 0, 1, 2) (2.64)
2β cos γ
3
u log ψ indices (−3, −1, 0, 2) (2.65)
2β cos γ

Copyright © 2003 IOP Publishing Ltd.

 in which (log ψ)x is the χ of the invariant Painlevé analysis. When the test is
satisfied (b2 = β 2 ), the linearizing transformation [40] is provided by [43] the
singular part operator of the first family, which maps the nonlinear PDE (2.61) to
the linear Schrödinger equation for V obtained by setting b = β = 0 in (2.61),
Kundu–Eckhaus (U ) ⇐⇒ Schrödinger (V ) iVt + αVxx = 0 (2.66)
√ √ log ϕ
U = ux eiθ V = ϕx eiθ u= . (2.67)
2β cos γ

2.4.2 Auto-Bäcklund transformation of a PDE: the singular manifold

method
Widely known as the singular manifold method or truncation method because it
selects the beginning of a Laurent series and discards (‘truncates’) the remaining
infinite part, this method was introduced by Weiss et al [28] and later improved
in many directions [44–49,53]. Its most recent version can be found in the lecture
notes of a CIME school [50, 51], to which we refer for further details.
The goal is to find the BT or, if a BT does not exist, to generate some exact
solutions. Since the BT is itself the result of an elimination [52] between the
Lax pair and the Darboux involution, the task splits into the two simpler tasks of
deriving these two elements. Let us take one example.
The modified Korteweg–de Vries equation (mKdV)
mKdV(w) ≡ bwt + (wxx − 2w3 /α 2 )x = 0 (2.68)
is equivalently written in its potential form
p-mKdV(r) ≡ brt + rxxx − 2rx3 /α 2 + F (t) = 0 w = rx (2.69)
a feature which will shorten the expressions used later. This last PDE admits two
opposite families (α is any square root of α 2 ):
p = 0− q = −3 r  α log ψ indices (−1, 0, 4) D = α (2.70)
and the results to be found are:
• the Darboux involution
r = D log Y + R (2.71)
a relation expressing the difference of the two solutions r and R of p-mKdV
as the logarithmic derivative D log Y , in which D = α is the singular part
operator of either family and Y is a Riccati pseudopotential equivalent to the
Lax pair (see next item),
• the Lax pair, written here in its equivalent Riccati representation,
 
yx 1 W
=λ −y −2 (2.72)
y y α
   
yt 1 W W2 Wx
b = −4λ + 2 2 + 2 − 4λ2 y (2.73)
y y α α α
x

Copyright © 2003 IOP Publishing Ltd.

 in which W satisfies the mKdV equation (2.68) and λ is the spectral
parameter.
• the BT, by some elimination between the previous two items.
This programme is achieved by defining the truncation [49],

r = α log Y + R (2.74)

in which r satisfies the p-mKdV, R is an as yet unconstrained field, Y is the most

general homographic transform of χ which vanishes as χ vanishes,

Y −1 = B(χ −1 + A) (2.75)

A and B are two adjustable fields and the gradient of χ is (2.22). The left-hand
side (lhs) of the PDE is then


6
p-mKdV(r) ≡ Ej (S, C, A, B, R)Y j −3 (2.76)
j =0

and the system of determining equations to be solved is

∀j Ej (S, C, A, B, R) = 0. (2.77)

This choice of Y (2.75) is necessary to implement the two opposite families

feature of mKdV. The general solution of the determining equations introduces an
arbitrary complex constant λ and a new field W [49]

W = (R − α log B)x A = W/α

bC = 2Wx /α − 2W /α + 4λ2
2 2

S = 2Wx /α − 2W 2 /α 2 − 2λ2 (2.78)

and the equivalence of the cross-derivative condition (Yx )t = (Yt )x to the mKdV
equation (2.68) for W proves that one has obtained a Darboux involution and a
Lax pair, with the correspondence y = BY.
The auto-BT of mKdV is obtained by the elimination of Y , i.e. by the
substitution
log BY = α −1 (w − W ) dx (2.79)

in the two equations (2.72) and (2.73) for the gradient of y = BY .

The singular manifold equation, defined [28] as the constraint put on ϕ for
the truncation to exist, is obtained by the elimination of W between S and C,

bC − S − 6λ2 = 0 (2.80)

and it is identical to that of the KdV equation.

Copyright © 2003 IOP Publishing Ltd.

 Remark. The fact that, in the Laurent series (2.60), uT (0) (the ‘lhs’) and uT (∞)
(the ‘constant level coefficient’) are both solutions of the same PDE is not
sufficient to define a BT, since any non-integrable PDE also enjoys this feature. It
is necessary to exhibit both the Darboux involution and a good Lax pair.
Most (1 + 1)-dimensional PDEs with the PP have been successfully
processed by the singular manifold method, including the not so easy Kaup–
Kupershmidt [53] and Tzitzéica [54] equations.
The extension to (2 + 1)-dimensional PDEs with the PP has also been
investigated ([55] and references therein).

2.4.3 Single-valued solutions of the Bianchi IX cosmological model

Sometimes, the no-log conditions generated by the test provide some global
information, which can then be used to integrate.
The Bianchi IX cosmological model is a six-dimensional system of three
second-order ODEs:

(log A)

= A2 − (B − C)2 and cyclically = d/dτ (2.81)

or, equivalently,

(log ω1 )

= ω22 + ω32 − ω22 ω32 /ω12

(2.82)
A = ω2 ω3 /ω1 ω12 = BC and cyclically.

One of the families [56, 57]

A = χ −1 + a2 χ + O(χ 3 ) χ = τ − τ2
B = χ −1 + b2 χ + O(χ 3 ) (2.83)
−1
C =χ + c2 χ + O(χ ) 3

has the Fuchs indices −1, −1, −1, 2, 2, 2 and the Gambier test detects no
logarithms at the triple index 2. The Fuchsian perturbative method


N 
2+N−n
A = χ −1 εn aj(n) χ j χ = τ − τ2 and cyclically (2.84)
n=0 j =−n

then detects movable logarithms at (n, j ) = (3, −1) and (5, −1) [57] and the
enforcement of these no-log conditions generates the three solutions:

(b2(0) = c2(0) and b−1

(1) (1)
= c−1 ) or cyclically (2.85)
a2(0) = b2(0) = c2(0) = 0 (2.86)
(1) (1) (1)
a−1 = b−1 = c−1 . (2.87)

Copyright © 2003 IOP Publishing Ltd.

 These are constraints which reduce the number of arbitrary coefficients to,
respectively, four, three and four, thus defining particular solutions which may
have no movable critical points.
The first constraint (2.85) implies the equality of two of the components
(A, B, C) and, thus, defines the four-dimensional subsystem B = C [58], whose
general solution is single-valued,

k1 k22 sinh k1 (τ − τ1 )
A= B =C = . (2.88)
sinh k1 (τ − τ1 ) k1 sinh2 k2 (τ − τ2 )
The second constraint (2.86) amounts to suppressing the triple Fuchs index 2,
thus defining a three-dimensional subsystem with a triple Fuchs index −1. One
can, indeed, check that the perturbed Laurent series (2.84) is identical to that of
the Darboux–Halphen system [59, 60]

ω1
= ω2 ω3 − ω1 ω2 − ω1 ω3 and cyclically (2.89)

whose general solution is single-valued.

The third and last constraint (2.87) amounts to suppressing two of the three
Fuchs indices −1, thus defining a four-dimensional subsystem whose explicit
writing is yet unknown. With the additional constraint
(0) (0) (0)
a 2 + b 2 + c2 = 0 (2.90)

the Laurent series (2.83) is identical to that of the three-dimensional Euler system
(1750) [61], describing the motion of a rigid body around its centre of mass

ω1
= ω2 ω3 and cyclically (2.91)

whose general solution is elliptic [61]:

ωj = (log(℘ (τ − τ0 , g2 , g3 ) − ej ))
j = 1, 2, 3, (τ0 , g2 , g3 ) arbitrary
(2.92)
℘
2 = 4(℘ − e1 )(℘ − e2 )(℘ − e3 ) = 4℘ 3 − g2 ℘ − g3 . (2.93)

The four-dimensional subsystem (the one without (2.90)) defines an extrapolation

to four parameters of this elliptic solution, quite probably single-valued, whose
closed form is still unknown.
One thus retrieves by the analysis all the results of the geometric assumption
of self-duality [62], even slightly more.

2.4.4 Polynomial first integrals of a dynamical system

A first integral of an ODE is, by definition, a function of x, u(x), u
(x), . . . which
takes a constant value at any x, including the movable singularities of u. Consider,

Copyright © 2003 IOP Publishing Ltd.

 for instance, the Lorenz model
dx dy dz
= σ (y − x) = rx − y − xz = xy − bz(x − y). (2.94)
dt dt dt

First integrals in the class P (x, y, z) eλt , with P polynomial and λ constant,
should not be searched for with the assumption P the most general polynomial
in three variables. Indeed, P must have no movable singularities. The movable
singularities of (x, y, z) are

x ∼ 2iχ −1 y ∼ −2iσ −1 χ −2 z ∼ −2σ χ −2indices (−1, 2, 4)

(2.95)
therefore, the generating function of admissible polynomials P is built from the
singularity degrees of (x, y, z) [63]:

1
= 1 + αx + α 2 (x 2 + y + z) + α 3 (x 3 + xy + xz)
(1 − αx)(1 − α 2 y)(1 − α 2 z)
+ α 4 (x 4 + x 2 y + x 2 z + yz + z2 + y 2 ) + · · ·
(2.96)

defining the basis, ordered by singularity degrees,

(1) (x) (x 2 , y, z) (x 3 , xy, xz)

(x 4 , x 2 y, x 2 z, yz, z2 , y 2 ) . . . .
(2.97)
The candidate of lowest degree is a linear combination of (x 2 , y, z), which indeed
provides a first integral [64]

K1 = (x 2 − 2σ z) e2σ t b = 2σ. (2.98)

Six polynomial first integrals are known [65] with a singularity degree at most
equal to four and these are the only ones in the polynomial class [66].

2.4.5 Solitary waves from truncations

If the PDE is non-integrable or if one only wants to find particular solutions,
the singular manifold method of section 2.4.2 still applies, it simply produces
less results. For autonomous partially integrable PDEs, the typical output is a set
of constant values for the unknowns S, C, A, B, R in the determining equations
(2.77). In such a case, quite generic for non-integrable equations, the integration
of the Riccati system (2.22) yields the value

k k
χ −1 = tanh (ξ − ξ0 ) ξ = x − ct k 2 = −2S c=C (2.99)
2 2
the singular part operator D has constant coefficients; therefore, the solutions r
in (2.71) are solitary waves r = f (ξ ), in which f is a polynomial in sech kξ and

Copyright © 2003 IOP Publishing Ltd.

 tanh kξ . This follows immediately from the two elementary identities [67]

1  π 1  π
tanh z − = −2i sech 2z + i tanh z + = 2 tanh 2z + i .
tanh z 2 tanh z 2
(2.100)
In the simpler case of a one-family PDE, the (degenerate) Darboux
involution is
u = D log ψ + U ∂x log ψ = χ −1 (2.101)
and this class of solitary waves r = f (ξ ) reduces to the class of polynomials in
tanh(k/2)ξ . In the example of the chaotic Kuramoto–Sivashinsky (KS) equation

ut + uux + µuxx + buxxx + νuxxxx = 0 ν = 0 (2.102)

one finds [68]

15(16µν − b 2 )
D = 60ν∂x3 + 15b∂x2 + ∂x (2.103)
76ν
k
u = D log cosh (ξ − ξ0 ) + c (c, ξ0 ) arbitrary (2.104)
2
in which b2 /(µν) only takes the values 0, 144/47, 256/73, 16, and k is not
arbitrary. In the quite simple form (2.104), much more elegant than a third-degree
polynomial in tanh, the only nonlinear item is the logarithm, D being linear and
cosh solution of a linear system. This displays the enormous advantage of taking
into account the singularity structure when searching for such solitary waves.
The correct method to obtain all the trigonometric solitary waves of
autonomous PDEs and their elliptic generalization has been recently built [69].

2.4.6 First-degree birational transformations of Painlevé equations

At first glance, it seems that the truncation procedure described in section 2.4.2
should be even easier when the PDE reduces to an ODE. This is not the case
because, in addition to the Riccati variable χ or Y of the truncation, there
exists a second natural Riccati variable and, therefore, a homographic dependence
between the two Riccati variables, which must be taken into account under penalty
of failure of the truncation.
Indeed, any Nth-order first-degree ODE with the Painlevé property is
necessarily [70, pp 396–409] a Riccati equation for U (N−1) , with coefficients
depending on x and the lower derivatives of U , for example in the case of P6,

= A2 (U, x)U
2 + A1 (U, x)U
+ A0 (U, x, A, B, , ). (2.105)

Then the Riccati variable of the truncation (denote it Z) is linked to U

by some
homography,
(U
+ g2 )(Z −1 − g1 ) − g0 = 0 g0 = 0 (2.106)

Copyright © 2003 IOP Publishing Ltd.

 in which g0 , g1 , g2 are functions of (U, x) to be found. Implementing this
dependence in the truncation [71] provides a unique solution for P6, which is
the unique first-degree birational transformation, first found by Okamoto [72],
N x(x − 1)U
0 1 x − 1
= + + + (2.107)
u−U U (U − 1)(U − x) U U −1 U −x
x(x − 1)u
θ0 θ1 θx − 1
= + + + (2.108)
u(u − 1)(u − x) u u−1 u−x
∀j = ∞, 0, 1, x : (θj2 + 2j − (N/2)2 )2 − (2θj j )2 = 0 (2.109)

N= (θk2 − 2k ) (2.110)

with the classical definition for the monodromy exponents,

2
θ∞ = 2α θ02 = −2β θ12 = 2γ θx2 = 1 − 2δ (2.111)
2∞ = 2A 20 = −2B 21 = 2 2x = 1 − 2 . (2.112)

The equivalent affine representation of (2.109)–(2.110) is

   
1  1 1  1
θj = j − k + j = θj − θk + (2.113)
2 2 2 2
 
N =1− k = −1 + θk = 2(θj − j ) j = ∞, 0, 1, x (2.114)

in which j, k = ∞, 0, 1, x.
The well-known confluence from P6 down to P2 then allows us to recover
[73] all the first-degree birational transformations of the five Painlevé equations
(P1 admits no such transformation because it does not depend on any parameter),
thus providing a unified picture of these transformations.

2.5 Liouville integrability and Painlevé integrability

A Hamiltonian system with N degrees of freedom is said to be Liouville-
integrable if it possesses N functionally independent invariants in involution. In
general, there is no correlation between Liouville-integrability and the Painlevé
property, as seen in the two examples with N = 1,

p2
H (q, p, t) = − 2q 3 − tq (2.115)
2
p2
H (q, p, t) = + q5 (2.116)
2
in which the first system is Painlevé-integrable and not Liouville-integrable
and vice versa for the second system. However, given a Liouville-integrable
Hamiltonian system which, in addition, passes the Painlevé test, one must try
to prove its Painlevé integrability by explicitly integrating.

Copyright © 2003 IOP Publishing Ltd.

 Such an example is the cubic Hénon–Heiles system

H ≡ 12 (p12 + p22 + c1 q12 + c2 q22 ) + αq1 q22 − 13 βq13 + 12 c3 q2−2 α = 0

(2.117)
q1

+ c1 q1 − βq12 + αq22 = 0 (2.118)

q2

+ c2 q2 + 2αq1 q2 − c3 q2−3 =0 (2.119)

which passes the Painlevé test in three cases only,

(SK): β/α = −1 c1 = c2 (2.120)

(K5): β/α = −6 (2.121)
(KK): β/α = −16 c1 = 16c2 . (2.122)

In these three cases, the general solution q1 (hence q22 ) is, indeed, single-valued
and expressed with genus two hyperelliptic functions. This was proven by Drach
in 1919 for the second case, associated to KdV5 and, only recently [74], in the
two other cases. This proof completes the result of [75], who found the separating
variables (a global object) by just considering the Laurent series (a local object),
following a powerful method due to van Moerbeke and Vanhaecke [76].

2.6 Discretization and discrete Painlevé equations

This quite important subject (the integrability of difference equations) is reviewed
elsewhere in this volume [77], so we will just write a few lines about it, for
completeness.
Let us consider the difference equations or q-difference equations (we skip,
for brevity, the elliptic stepsize [78]),

∀x ∀h: E(x, h, {u(x + kh), k − k0 = 0, . . . , N}) = 0 (2.123)

∀x ∀q: E(x, q, {u(xq ), k − k0 = 0, . . . , N}) = 0
k
(2.124)

algebraic in the values of the field variable, with coefficients analytic in x and the
stepsize h or q. As compared to the continuous case, the main missing item is an
undisputed definition for the discrete Painlevé property. The currently proposed
definitions are:

(1) there exists a neighbourhood of h = 0 (resp. q = 1) at every point of which

the general solution x → u(x, h) (resp. x → u(x, q)) has no movable critical
singularities [79]; and
(2) the Nevanlinna order of growth of the solutions at infinity is finite [80];

but none is satisfactory. Indeed, the first one says nothing about discrete equations
without continuum limit, and the second one excludes the continuous P6 equation.

Copyright © 2003 IOP Publishing Ltd.

 Despite the lack of consensus on this definition, a discrete Painlevé test
has been developed to generate necessary conditions for these properties. Of
exceptional importance at this point is the singularity confinement method [81],
which tests with great efficiency a property not yet rigorously defined but
which for sure will be an important part of the good definition of the discrete
Painlevé property. The approach developed by Ruijsenaars [82] for linear discrete
equations, namely to require as much analyticity as possible, should be interesting
to transpose to nonlinear discrete equations.
Just for consistency, an interesting development would be to display a
discrete version of (2.13) escaping all the methods of the discrete test.
Let us say a word about the discrete analogue of the Painlevé and Gambier
classifications. These second-order first-degree continuous equations all have a
precise form (u

is a second-degree polynomial in u
, the coefficient of u
2 is the
sum of, at most, four simple poles in u, etc), directly inherited from the property
of the elliptic equations isolated by Briot and Bouquet. In the discrete counterpart,
the main feature is the existence of an addition formula for the elliptic function ℘
of Weierstrass. As remarked earlier by Baxter and Potts (see references in [83]),
this formula defines an exact discretization of (2.6). Then, all the autonomous
discrete second-order first-degree equations with the (undefined!) discrete PP
have a precise form resulting from the most general discrete differentiation
of the addition formula and the non-autonomous ones simply inherit variable
coefficients as in the continuous case. Of course, the second-order higher-degree
(mostly multi-component) equations are much richer, see details in the review
[84].
Another open question concerns the continuum limit of the contiguity
relation of the ODEs which admit such a relation. The contiguity relation of
the (linear) hypergeometric equation has a continuum limit which is not the
hypergeometric equation but a confluent one. In contrast, the contiguity relation
of the (linearizable) Ermakov equation has a continuum limit which is again an
Ermakov equation [85]. One could argue that the latter depends on a function and
the former only on a finite number of constants. Nevertheless, this could leave
the hope to upgrade from P5 to P6 the highest continuum limit for the contiguity
relation of P6 [73, 86, 87].

2.7 Conclusion

The allowed space forced us to skip quite interesting developments, such as

the relation with differential geometry [88] or the way to obtain the nonlinear
superposition formula from singularities [89], or the weak Painlevé property [70,
Leçons 5–10, 13, 19; 90, 91].
For applications to non-integrable equations, not covered in this volume, the
reader can refer to tutorial presentations such as those in [50, 92].

Copyright © 2003 IOP Publishing Ltd.

 Acknowledgments
The financial support of Tournesol grant T99/040, IUAP Contract No P4/08
funded by the Belgian government and the CEA is gratefully acknowledged.

References
[1] Chazy J 1911 Acta Math. 34 317–85
[2] Conte R 1999 The Painlevé Property, One Century Later (CRM Series in
Mathematical Physics) ed R Conte (New York: Springer) pp 77–180 solv-
int/9710020
[3] Umemura H 1990 Nagoya Math. J. 119 1–80
[4] Bureau F J 1964 Ann. Mat. Pura Appl. LXVI 1–116
[5] Cosgrove C M 2000 Stud. Appl. Math. 104 1–65
(http://www.maths.usyd.edu.au:8000/res/Nonlinear/Cos/1998-22.html)
[6] Cosgrove C M 2000 Stud. Appl. Math. 104 171–228
(http://www.maths.usyd.edu.au:8000/res/Nonlinear/Cos/1998-23.html)
[7] Garnier R 1912 Ann. Éc. Norm. 29 1–126
[8] Kudryashov N A and Soukharev M B 1998 Phys. Lett. A 237 206–16
[9] Painlevé P 1906 C. R. Acad. Sci. Paris 143 1111–17
[10] Fuchs R 1905 C. R. Acad. Sci. Paris 141 555–8
[11] Gambier B 1910 Acta. Math. 33 1–55
[12] Bureau F J 1972 Ann. Mat. Pura Appl. XCI 163–281
[13] Cosgrove C M 1993 Stud. Appl. Math. 90 119–87
[14] Cosgrove C M and Scoufis G 1993 Stud. Appl. Math. 88 25–87
[15] Cosgrove C M 1997 Stud. Appl. Math. 98 355–433
[16] Exton H 1971 Rend. Mat. 4 385–448
[17] Darboux G 1894 Leçons sur la théorie générale des surfaces et les applications
géométriques du calcul infinitésimal (4 vol) (Paris: Gauthier-Villars) Reprinted
1972 Théorie générale des surfaces (New York: Chelsea) Reprinted 1993 (Paris:
Gabay)
[18] Matveev V B and Salle M A 1991 Darboux Transformations and Solitons (Springer
Series in Nonlinear Dynamics) (Berlin: Springer)
[19] Ablowitz M J, Kaup D J, Newell A C and Segur H 1974 Stud. Appl. Math. 53 249–315
[20] Cosgrove C M 1993 Stud. Appl. Math. 89 1–61
[21] Cosgrove C M 1993 Stud. Appl. Math. 89 95–151
[22] Mikhailov A V, Shabat A B and Sokolov V V 1991 What is Integrability?
ed V E Zakharov (Berlin: Springer) pp 115–84
[23] Clarkson P A, Fokas A S and Ablowitz M J 1989 SIAM J. Appl. Math. 49 1188–209
[24] Camassa R and Holm D D 1993 Phys. Rev. Lett. 71 1661–4
[25] Degasperis A, Holm D D and Hone A N W 2002 Theor. Math. Phys. 133 1461–72,
nlin.SI/0205023
[26] Ablowitz M J, Ramani A and Segur H 1980 J. Math. Phys. 21 715–21, 1006–15
[27] Mason L J and Woodhouse N M J 1993 Nonlinearity 6 569–81
Kowalevski S 1889 Acta. Math. 12 177–232
Picard E 1893 Acta. Math. 17 297–300
[28] Weiss J, Tabor M and Carnevale G 1983 J. Math. Phys. 24 522–6

Copyright © 2003 IOP Publishing Ltd.

 [29] Conte R 1989 Phys. Lett. A 140 383–90
[30] Kolmogorov A N, Petrovskii I G and Piskunov N S 1937 Bull. Univ. État Moscou,
série internationale A 1 1–26
[31] Newell A C and Whitehead J A 1969 J. Fluid Mech. 38 279–303
[32] Satsuma J 1987 Topics in Soliton Theory and Exact Solvable Nonlinear Equations
ed M J Ablowitz, B Fuchssteiner and M D Kruskal (Singapore: World Scientific)
pp 255–62
[33] Fordy A P and Pickering A 1991 Phys. Lett. A 160 347–54
[34] Conte R, Fordy A P and Pickering A 1993 Physica D 69 33–58
[35] Bureau F J 1964 Ann. Mat. Pura Appl. LXIV 229–364
[36] Conte R 1988 Phys. Lett. A 134 100–4
[37] Ince E L 1926 Ordinary Differential Equations (London: Longmans Green).
Reprinted 1956 (New York: Dover). Russian translation (GTIU, Khar’kov, 1939)
[38] Musette M and Conte R 1995 Phys. Lett. A 206 340–6
[39] Painlevé P 1900 Bull. Soc. Math. France 28 201–61
[40] Kundu A 1984 J. Math. Phys. 25 3433–8
[41] Calogero F and Eckhaus W 1987 Inverse Problems 3 229–62
[42] Clarkson P A and Cosgrove C M 1987 J. Phys. A: Math. Gen. 20 2003–24
[43] Conte R and Musette M 1994 Theor. Math. Phys. 99 543–8
[44] Musette M and Conte R 1991 J. Math. Phys. 32 1450–7
[45] Estévez P G, Gordoa P R, Martı́nez Alonso L and Medina Reus E 1993 J. Phys. A:
Math. Gen. 26 1915–25
[46] Garagash T I 1993 Nonlinear Evolution Equations and Dynamical Systems ed V G
Makhankov, I V Puzynin and O K Pashaev (Singapore: World Scientific) pp 130–3
[47] Musette M and Conte R 1994 J. Phys. A: Math. Gen. 27 3895–913
[48] Conte R and Musette M 1996 Nonlinear Physics: Theory and Experiment
ed E Alfinito, M Boiti, L Martina and F Pempinelli (Singapore: World Scientific)
pp 67–74
[49] Pickering A 1996 J. Math. Phys. 37 1894–927
[50] Conte R 2003 Direct and Inverse Methods in Solving Nonlinear Evolution Equations
ed A Greco (Berlin: Springer) nlin.SI/0009024
[51] Musette M 2003 Direct and Inverse Methods in Solving Nonlinear Evolution
Equations ed A Greco (Berlin: Springer)
[52] Chen H H 1974 Phys. Rev. Lett. 33 925–8
[53] Musette M and Conte R 1998 J. Math. Phys. 39 5617–30
[54] Conte R, Musette M and Grundland A M 1999 J. Math. Phys. 40 2092–106
[55] Estévez P G 2001 Inverse Problems 17 1043–52
[56] Contopoulos G, Grammaticos B and Ramani A 1993 J. Phys. A: Math. Gen. 25
5795–9
[57] Latifi A, Musette M and Conte R 1994 Phys Lett. A 194 83–92
Latifi A, Musette M and Conte R 1995 Phys. Lett. A 197 459–60
[58] Taub A H 1951 Ann. Math. 53 472–90
[59] Darboux G 1878 Ann. Éc. Norm. 7 101–50
[60] Halphen G-H 1881 C. R. Acad. Sci. Paris 92 1101–3. Reprinted 1918 Œuvres vol 2
(Paris: Gauthier-Villars) pp 475–7
[61] Belinskii V A, Gibbons G W, Page D N and Pope C N 1978 Phys. Lett. A 76 433–5
[62] Gibbons G W and Pope C N 1979 Commun. Math. Phys. 66 267–90
[63] Levine G and Tabor M 1988 Physica D 33 189–210

Copyright © 2003 IOP Publishing Ltd.

 [64] Segur H 1982 Topics in Ocean Physics ed A R Osborne and P Malanotte Rizzoli
(Amsterdam: North-Holland) pp 235–77
[65] Kuś M 1983 J. Phys. A: Math. Gen. 16 L689–91
[66] Llibre J and Zhang Xiang 2002 Invariant algebraic surfaces of the Lorenz system
J. Math. Phys. 43 1622–45
[67] Conte R and Musette M 1993 Physica D 69 1–17
[68] Kudryashov N A 1988 Prikl. Mat. Mekh. 52 465–70 (Engl. Transl. 1988 J. Appl.
Math. Mech. 52 361–5)
[69] Musette M and Conte R 2003 Physica D 181 70–9
[70] Painlevé P 1897 Leçons sur la théorie analytique des équations différentielles
(Leçons de Stockholm, 1895) (Paris: Hermann). Reprinted 1973 Œuvres de Paul
Painlevé vol I (Paris: Éditions du CNRS)
[71] Conte R and Musette M 2002 Physica D 161 129–41
[72] Okamoto K 1986 J. Fac. Sci. Univ. Tokyo, Sect. IA 33 575–618
[73] Conte R and Musette M 2001 J. Phys. A: Math. Gen. 34 10 507–22
[74] Verhoeven C, Musette M and Conte R 2002 J. Math. Phys. 43 1906–15
[75] Ravoson V, Gavrilov L and Caboz R 1993 J. Math. Phys. 34 2385–93
[76] Vanhaecke P 1996 Integrable systems in the Realm of Algebraic Geometry (Lecture
Notes in Mathematics vol 1638) (Berlin: Springer)
[77] Tamizhmani K M, Ramani A, Grammaticos B and Tamizhmani T, this volume
[78] Sakai H 2001 Comm. Math. Phys. 220 165–229
[79] Conte R and Musette M 1996 Phys. Lett. A 223 439–48
[80] Ablowitz M J, Halburd R and Herbst B M 2000 Nonlinearity 13 889–905.
[81] Grammaticos B, Ramani A and Papageorgiou V 1991 Phys. Rev. Lett. 67 1825–8
[82] Ruijsenaars S N M 2001 J. Nonlinear Math. Phys. 8 256–87
[83] Conte R and Musette M Theory of Nonlinear Special Functions: The Painlevé
Transcendents ed L Vinet and P Winternitz (New York: Springer) to appear
solv-int/9803014.
[84] Grammaticos B, Nijhoff F W and Ramani A 1999 The Painlevé Property, One
Century Later (CRM Series in Mathematical Physics) ed R Conte (New York:
Springer) pp 413–516
[85] Hone A N W 1999 Phys. Lett. A 263 347–54
[86] Okamoto K 1987 Ann. Mat. Pura Appl. 146 337–81
[87] Nijhoff F W, Ramani A, Grammaticos B and Ohta Y 2001 Stud. Appl. Math. 106
261–314 solv-int/9812011
[88] Bobenko A I and Eitner U 2000 Lecture Notes in Mathematics vol 1753 (Berlin:
Springer) http://www-sfb288.math.tu-berlin.de
[89] Musette M and Verhoeven C 2000 Physica D 144 211–20
[90] Ramani A, Dorizzi B and Grammaticos B 1982 Phys. Rev. Lett. 49 1539–41
[91] Grammaticos B, Dorizzi B and Ramani A 1984 J. Math. Phys. 25 3470–3
[92] Musette M 1999 The Painlevé Property, One Century Later (CRM Series in
Mathematical Physics) ed R Conte (New York: Springer) pp 517–72

Copyright © 2003 IOP Publishing Ltd.

 Chapter 3

Discrete integrability
K M Tamizhmani† and A Ramani† , B Grammaticos‡ and
T Tamizhmani‡
† CPT, Ecole Polytechnique, CNRS, Palaiseau, France
‡ GMPIB, Université Paris VII, France

3.1 Introduction: who is afraid of discrete systems?

For the past few centuries, physicists and mathematicians have been familiar
with differential systems. The fundamental equations used in the modelling of
natural phenomena are cast in differential form based on the underlying (often
tacit) assumption that spacetime is continuous. The power of this differential
description lies in the richness of existing tools (although this is an a posteriori
statement) and the success of this approach is undeniable [1]. Integrable systems
hold a privileged position among the differential family. They are, according
to Calogero [2], both ‘universal’ and ‘widely applicable’. Given this situation
why should a physicist be interested in discrete difference systems? Well, for
one, difference systems are ubiquitous in physical modelling. As soon as one
attempts a numerical simulation of a differential system, one has to transcribe it
into an algorithm which is invariably cast in discrete form. It is not exaggerated to
state that our knowledge of the physical universe based on numerical simulations
relies on discrete equations. Still, while numerical algorithms are often considered
as inaccurate approximations of a continuous (differential) ‘reality’, there exist
domains where discrete systems arise naturally. For instance, when some physical
quantity depends on a particular parameter, there exist cases where one can
establish recursion relations thus reducing the computation of this quantity
to that of some basic ‘seed’ one and the solution of the recursion. Solvable
recursion relations are one example of integrable discrete systems. In recent
years, the domain of discrete integrability has undergone a real revolution. As
a consequence discrete analogues have been proposed for most well-known

Copyright © 2003 IOP Publishing Ltd.

 integrable differential systems. Moreover, specific tools for the investigation of
discrete integrable systems have been developed.
In this short review we shall present a selection of results on integrable
discrete systems. We shall start with a presentation of the various approaches
proposed for the detection of integrability, and then present results which will
illustrate the parallel existing between integrable discrete and continuous systems.
Finally, we shall present results on systems which either lie between the discrete
and continuous ones or go beyond the discrete systems. We shall not attempt
any rigorous definition of the notion of discrete integrability. Our experience on
this point is that there exist as many brands of integrability as of integrability
specialists. We shall rather present a plethora of examples which will (hopefully)
serve as a guide for the reader to develop his/her own understanding of the subject.

3.2 The detector gallery

While, as we shall argue in what follows, discrete systems are fundamental

entities, for historical and practical reasons everybody is more familiar with
continuous systems. In the domain of integrability, the use of complex analysis
has made possible the development of specific and efficient tools for the prediction
and actual integration of systems expressed as (ordinary or partial) differential
equations. According to Poincaré, to integrate a differential equation is to find,
for the general solution, a finite expression, possibly multi-valued, in terms of a
finite number of functions. The word ‘finite’ indicates that integrability is related
to global rather than local knowledge of the solution. However, this definition is
not very useful unless one defines more precisely what is meant by ‘function’.
By extending the solution of a given ordinary differential equation (ODE) in the
complex domain, one has the possibility, instead of asking for a global solution
for an ODE, to look for solutions locally and obtain a more global result by
analytic continuation. If we wish to define a function, we must find a way to
treat branch points, i.e. points around which two (at least) determinations are
exchanged. This can be done through various uniformization procedures provided
the branch points are fixed. Linear ODEs are such that all the singularities of their
solutions are fixed and are, thus, considered integrable. In the case of nonlinear
ODEs, the situation is not so simple due to the fact that the singular points in
this case may depend on the initial conditions: they are movable. The approach
of Painlevé [3] and his school, which, to be fair, was based on ideas from Fuchs
and Kovalevskaya, was simple: they decided to look for those nonlinear ODEs
the solutions of which were free from movable branch points. Painlevé managed
to take up Picard’s challenge and determine the functions defined by the solutions
of second-order nonlinear equations. The success of this approach is well known:
the Painlevé transcendents were discovered in that way and their importance in
mathematical physics is ever growing. The Painlevé property, i.e. the absence of

Copyright © 2003 IOP Publishing Ltd.

 movable branch points, has since been used with great success in the detection of
integrability [4].
We must stress one important point here. The Painlevé property as introduced
by Painlevé is not just a predictor of integrability but practically a definition of
integrability. As such, it becomes a tautology rather than a criterion. It is thus
crucial to make the distinction between the Painlevé property and the algorithm
for its investigation. The latter can only search for movable branch points within
certain assumptions [5]. The search can thus lead to a conclusion of which the
validity is questionable: if we find that the system passes what is usually referred
to as the Painlevé test (in one of its several variants), this does not necessarily
mean that the system possesses the Painlevé property. Thus, at least as far as its
usual practical application is concerned, the Painlevé test may not be sufficient for
integrability. The situation becomes further complicated if we consider systems
that are integrable through quadratures and/or cascade linearization. If we extend
the notion of integrability in order to include such systems, it turns out that
the Painlevé property is no longer related to it. Thus, the criterion based on the
singularity structure is not a necessary one in this case.
Despite these considerations, the Painlevé test has been of great heuristic
value for the study of the integrability of continuous systems, leading to the
discovery of a host of new integrable systems. The question thus naturally arose
as to whether these techniques could be transposed mutatis mutandis to the study
of discrete systems. The discrete systems we are referring to here (and which play
an important role in physical applications) are systems that are cast into a rational
form, perhaps after some transformation of the dependent variable. Since these
systems have singularities, it is natural to assume that singularities would play
an important role in connection with integrability. While this is quite plausible,
an approach based on singularities would be unable to deal with polynomial
mappings which do not possess any. Still, one would not expect all polynomial
mappings to be integrable, in particular in the view of the fact that many of them
exhibit chaotic behaviour. Moreover, any argument based on singularities in the
discrete domain can only bear a superficial resemblance to the situation in the
continuous case. One cannot hope to relate the singularities of mappings directly
to those of ODEs for the simple reason that there exist discrete systems which do
not have any non-trivial continuous limit.
Having set the framework we can now present a review of the various
discrete integrability detectors.

3.2.1 Singularity confinement

As we explained earlier, it is far from clear how one can relate the singularities
of discrete systems to those of continuous ones. Still, the notions of singularity
and single-valuedness can be transposed from the continuum to the discrete
setting. In the continuous case, a singularity that introduces multi-valuedness
is considered incompatible with integrability. Analogous ideas were introduced

Copyright © 2003 IOP Publishing Ltd.

 in the discrete case and one expects singularities to play an important role in
the study of the integrability of discrete systems. In this spirit, we introduced
the notion of singularity confinement [6]. Let us illustrate the main idea by an
example. Consider the mapping
a 1
xn+1 + xn−1 = + 2. (3.1)
xn xn
Obviously, a singularity appears whenever the value of xn becomes 0. Iterating
this value, one obtains the sequence {0, ∞, 0} and then the indeterminate form
∞ − ∞. As Kruskal points out, the real problem lies in the latter, while the
occurrence of a simple infinity is something that can easily be dealt with by
going to projective space. The way to treat this difficulty is to use an argument of
continuity with respect to the initial conditions and introduce a small parameter
. In this case, if we assume that xn =  we obtain for the first values of x:
xn+1 ≈ 1/ 2 , xn+2 ≈ −. When we carefully compute the next one we find that
not only is it finite but it also contains the memory of the initial condition xn−1 .
The singularity has disappeared.
This is the property that we have dubbed singularity confinement and after
having analysed a host of discrete systems we concluded that it was characteristic
of a system that was integrable through spectral methods. Through a bold move,
singularity confinement has been elevated to the rank of an integrability criterion.
In what follows, we shall comment on its necessary and sufficient character.
Several questions had to be answered for singularity confinement to be really
operative. The first, the one we have just encountered, was the one related to the
fact that the iteration of a mapping may not be defined uniquely in both directions.
Thus, we proposed the criterion of preimage non-proliferation [7], which had
the advantage of eliminating en masse all polynomial nonlinear mappings. One
remark at this point related to the existence of integrable mappings involving two
variables is unavoidable. A typical example is what we call asymmetric discrete
Painlevé equations. It can be argued that, in a such a mapping, one of the variables
can always be eliminated leading to a single mapping for the other one. However,
for a generic second-order system, the resulting mapping will be one where the
variables xn+1 and xn−1 appear at powers higher than unity. Its evolution leads, in
general, to an exponential number of images and preimages, of the initial point.
This non-single-valued system cannot be integrable. This is not in contradiction to
the fact that we can obtain one solution of the mapping, namely the one furnished
by the evolution of the two-variable system. This is the only solution that we know
how to describe, while the full system, with exponentially increasing number of
branches, eludes a full description.
The second point is that the notion of ‘singularity’ had to be refined. Clearly,
the simple appearance of an infinity in the iteration of a mapping is not really a
problem. What is crucial is that a mapping may, at some point, ‘lose a degree of
freedom’. In a mapping of the form xn+1 = f (xn , xn−1 ), this means simply that
∂xn+1 /∂xn−1 = 0 and the memory of the initial condition xn−1 disappears from

Copyright © 2003 IOP Publishing Ltd.

 the iteration. What does ‘confinement’ mean in this case? Clearly, the mapping
must recover the lost degree of freedom and the only way to do this is through
the appearance of an indeterminate form 0/0, ∞ − ∞, etc, in the subsequent
iterations.
An interesting application of the singularity confinement method has been
the derivation of non-autonomous integrable mappings [8]. Let us illustrate our
approach here. Our starting point is an autonomous integrable mapping. The
way to apply singularity confinement for de-autonomization is to start from the
(confined) singularity pattern of an autonomous (integrable) mapping and ask
for the non-autonomous extension with exactly the same singularity pattern. The
previous example (3.1) will help us make things clearer. As we have seen, the
singularity pattern is {0, ∞, 0}. Now we assume that a is no longer a constant but
may depend on n. The singularity analysis can be performed in a straightforward
way. Assuming that xn = , we obtain xn+1 ≈ 1/ 2 , xn+2 ≈ − and requiring
xn+3 to be finite, we obtain the constraint an+2 − 2an+1 + an = 0, i.e. an is of the
form an = αn + β. Thus, the non-autonomous form of (3.1), compatible with the
confinement property, is
αn + β 1
xn+1 + xn−1 = + 2 (3.2)
xn xn

Mapping (3.2) is presumably integrable and it turns out that indeed it is. As we
have shown in [9], it does possess a Lax pair. Moreover, it is the contiguity relation
of the solutions of the one-parameter PIII equation [10]. Its continuous limit is PI ,
so (3.2) can be considered as its discrete analogue.

3.2.2 The perturbative Painlevé approach to discrete integrability

The main idea of the perturbative Painlevé approach is as follows. Suppose that
we start from a discrete system which contains a small parameter. Typically,
one considers the lattice spacing δ as small, whenever one is interested in the
continuous limit of the mapping. Next, we expand in a power series in this small
parameter. If the initial mapping is integrable, then the equations obtained at each
order of the series must equally be integrable. (This is something that we learned
from J Satsuma [11] but no practical application of this idea was produced at the
time.)
The ones to rediscover this idea (which, in fact, goes back to Poincaré) and
use it in practice were Conte and Musette [12]. The way to do this was to work
with expansions in the lattice spacing, obtain a sequence of coupled differential
equations and investigate the integrability of the latter using the Painlevé
algorithm. The advantage of this approach is that one can treat polynomial
mappings on the same footing as rational ones. Let us illustrate this approach
through an example. We choose the well-known logistic map

xn+1 = λxn (1 − xn ). (3.3)

Copyright © 2003 IOP Publishing Ltd.

 We introduce the lattice parameter δ and expand everything in a power
series in it. We have λ = λ0 + δλ1 + δ 2 λ2 + · · · and xn = w0 + δw1 + δ 2 w2 +
· · · . Similarly we have xn+1 = (w0 + δw0
+ δ 2 w0

/2 + · · · ) + δ(w1 + δw1
+
δ 2 w1

/2 + · · · ) + δ 2 (w2 + · · · ) + · · · where the continuous variable is t = nδ.

In this particular case, we take w0 ≡ 0, λ0 = 1 and the first equation (at order δ 2 )
is nonlinear in terms of the quantity w1 :

w1
= −w12 + λ1 w1 (3.4)

which is a Riccati equation and has movable poles as its only singularities. Then,
at order δ 3 , we have an equation for w2 :
 
λ2
λ
w2
= −w2 (λ1 − 2w1 ) − w13 + w12 + λ2 − 1 w1
1
(3.5)
2 2

and similarly at higher orders. Notice that (3.5) is linear for w2 . The same applies
to all subsequent equations. Indeed, at order δ n+1 , we find a differential equation
for the new quantity wn in terms of the ws that have been obtained before. Since
this equation is linear, it cannot have movable singularities when considered as
an equation for wn , everything else being supposed known. However, when we
consider the whole cascade of equations, the subsequent objects will, in general,
have singularities whenever the earlier ones are singular and these singularities
are movable in terms of the whole cascade. Moreover, they are not poles. Already
equation (3.5) shows that, in the neighbourhood of a pole of w1 , where w1 ≈ 1/s
with s = t − t0 (t0 being the location of the movable singularity of w1 ), w2
has logarithmic singularities w2 ≈ −log(s)/s 2 . This singularity is a critical one
which must be considered as movable in terms of the cascade and, therefore, the
perturbative Painlevé property is not satisfied. This is consistent with the fact that
the logistic map is known to be non-integrable.
Although the perturbative Painlevé approach is powerful enough, it is not
without drawbacks. The main critique is based on the fact that not all discrete
systems possess non-trivial continuous limits. In this case, if one does not
have a valid starting point, the whole approach collapses. Moreover, the direct
discretization of Conte and Musette consists in discretizing a given continuous
equation by introducing some freedom and using the perturbative Painlevé
approach in order to pinpoint the integrable subcases. However, this method is
only as good as one’s imagination and if the proposed discretization is not rich
enough, one may miss very interesting cases. In particular, the cases where,
starting from a single equation, the singularity confinement leads to terms of the
form (−1)n , suggesting that the natural form of the mapping is that of a system,
are almost impossible to guess in the direct discretization approach. In [13], we
have presented another example of a mapping:

xn+1 − xn + xn+1
2
+ xn2 + axn+1 xn = 0 (3.6)

Copyright © 2003 IOP Publishing Ltd.

 which does pass the perturbative Painlevé approach but which cannot be
integrable (unless a = −2), according to us, since it violates the preimage
proliferation condition. To put it in a nutshell, the perturbative Painlevé approach
to the integrability of discrete systems must be used as every other integrability
detector: with caution and discernment.

3.2.3 Algebraic entropy

While examining a family of mappings which did satisfy the singularity
confinement criterion, Hietarinta and Viallet [14] found that the mapping

1
xn+1 + xn−1 = xn + (3.7)
xn2

which has a confined singularity pattern {0, ∞, ∞, 0}, behaves chaotically.

Moreover, they pointed out that one can construct whole families of non-
integrable mappings which satisfy the confinement criterion. This has led them
to propose a new discrete integrability criterion. The approach is based on the
relation of discrete integrability and the complexity of the evolution introduced
by Arnold and Veselov. According to Arnold [15], the complexity (in the case
of mappings of the plane) is the number of intersection points of a fixed curve
with the image of a second curve obtained under the mapping at hand. While the
complexity grows exponentially with the iteration for generic mappings, it can be
shown [16] to grow only polynomially for a large class of integrable mappings.
As Veselov points out, ‘integrability has an essential correlation with the weak
growth of certain characteristics’.
The notion of complexity was further extended in the works of Viallet and
collaborators who focused on rational mappings [17]. They introduced what
they called algebraic entropy, which is a global index of the complexity of
the mapping. The main idea is that there exists a link between the dynamical
complexity of a mapping and the degree of its iterates. If we consider a mapping
of degree d, then the nth iterate will have a degree d n , unless common factors
lead to simplifications. It turns out that when the mapping is integrable such
simplifications do occur in a massive way leading to a degree growth which is
polynomial in n, instead of exponential. Thus, while the generic non-integrable
mapping has exponential degree growth, a polynomial growth is an indication of
integrability.
Let us illustrate this approach by a practical application on a mapping that
we have already encountered:

a 1
xn+1 + xn−1 = + 2. (3.8)
xn xn

In order to compute the degree of the iterates, we introduce the homogeneous

coordinates by taking x0 = p, x1 = q/r, assigning to p the degree zero and

Copyright © 2003 IOP Publishing Ltd.

 computing the degree of homogeneity in q and r at every iteration. We could,
of course, have introduced a different choice for x0 but it turns out that the
choice of a zero-degree x0 considerably simplifies the calculations. We thus obtain
the degrees: 0, 1, 2, 5, 8, 13, 18, 25, 32, 41, . . . . Clearly the degree growth is
polynomial. We have d2m = 2m2 and d2m+1 = 2m2 + 2m + 1. This is in perfect
agreement with the fact that the mapping (3.8) is integrable (in terms of elliptic
functions), being a member of the QRT [18] family of integrable mappings. (A
remark is necessary at this point. In order to obtain a closed-form expression for
the degrees of the iterates, we start by computing a sufficient number of them.
Once the expression of the degree has been heuristically established we compute
several more and check that they agree with the analytical expression predicted.)
As a matter of fact, the precise values of the degrees are not important: they are
not invariant under coordinate changes. However, the type of growth is invariant
and can be used as an indication of whether the mapping is integrable or not.
Let us show what happens in the case of a non-integrable mapping. We
choose one among those examined in some previous work of ours [19]:
xn−1
xn+1 = a + . (3.9)
xn
Again we take x0 = p, x1 = q/r, and compute the degree of homogeneity in q and
r. We find the sequence of degrees dn : 0, 1, 1, 2, 3, 5, 8, 13, 21, . . . . This is clearly
a Fibonacci sequence obeying the recursion dn+1 = d√ n + dn−1 and thus leading
to an exponential growth with asymptotic ratio (1 + 5)/2. As a consequence
the mapping (3.9) is not expected to be integrable, which is in agreement with the
findings of [19].
Let us examine again the problem of de-autonomization of mapping (3.1) in
the light of an algebraic entropy approach. We start with the mapping

a 1
xn+1 + xn−1 = + 2 (3.10)
xn xn

where a is an a priori arbitrary function of n. We compute the iterates of (3.10)

and we obtain the sequence

r 2 + a1 qr − pq 2 qQ4
x2 = x3 =
q2 r(r 2 + a1 qr − pq 2 )2
(r 2 + a1 qr − pq 2 )Q7 qQ4 Q12
x4 = x5 =
qQ24 r(r + a1 qr − pq 2 )Q27
2

where the Qk s are homogeneous polynomials in q, r of degree k. The computation

of the degrees of xn leads to 0, 1, 2, 5, 9, 17, 30, 54, 95,. . . . The growth is
exponential with ratio of the order of 1.76, a clear indication that the mapping is
not integrable in general. The simplifications that do occur are insufficient to curb
the exponential growth. As a matter of fact, if we follow a particular factor we can

Copyright © 2003 IOP Publishing Ltd.

 check that it keeps appearing either in the numerator or the denominator (where
its degree is alternatively 1 and 2). This corresponds to the unconfined singularity
pattern {0, ∞2 , 0, ∞, 0, ∞2 , 0, ∞, . . . }. Already at the fourth iteration, the
degrees differ in the autonomous and non-autonomous cases. Our approach
consists in requiring that the degree in the non-autonomous case be identical to
the one obtained in the autonomous one. If we implement the requirement that
d4 be 8 instead of 9, we find the condition an+1 − 2an + an−1 = 0, i.e. precisely
the one obtained through singularity confinement. Here this condition means that
q divides Q7 exactly. Moreover, once this condition is satisfied, the subsequent
degrees of the non-autonomous case coincide with that of the autonomous one.
For example, both q and r 2 + a1 qr − pq 2 divide Q12 exactly, leading to d5 = 13
instead of 17 etc. Thus, the mapping leads to polynomial growth in agreement
with its integrable character.

3.2.4 The Nevanlinna theory approach

As we explained in the previous subsection, we expect the integrability of a
mapping to be conditioned by the behaviour of its solutions when the independent
variable goes to infinity. The tools for the study of the growth of a given function
are furnished by the theory of meromorphic functions [20]. The reason why
such an approach would apply to discrete systems is due to the formal identity
which exists between discrete systems and delay equations [21]. Thus, one starts
from a difference equation and considers it as a delay equation in the complex
plane of the independent variable [22]. The natural framework for the study of
the behaviour near infinity of the solutions of a given mapping is Nevanlinna
theory [23]. This theory provides tools for the study of the value distribution
of meromorphic functions. In particular, it introduces the notion of order. The
latter is infinite for very-fast-growing functions, while a finite order indicates
a moderate growth. It would be reasonable to surmise that an infinite order is
an indication of non-integrability for discrete systems. (We do not make any
statement here concerning continuous systems and, in fact, it is well known
that integrable differential equations may have solutions of infinite order.) The
Nevanlinna theory provides an estimation of the growth of the solutions of a
given discrete system. However, since the order may depend on the precise
coefficients of the equation and their dependence on the independent variable,
the starting point of our application of the Nevanlinna theory is that the mappings
are autonomous.
The main tool for the study of the value distribution of entire and
meromorphic functions is the Nevanlinna characteristic (and various quantities
related to the latter). The Nevanlinna characteristic of a function f , denoted by
T (r; f ), measures the ‘affinity’ of f for the value ∞. It is usually represented
as the sum of two terms: the frequency of poles and the contribution from the
arcs |z| = r where |f (z)| is large [23]. From the characteristic, one can define the
order of a meromorphic function: σ = lim supr→∞ log T (r; f )/ log r. When f

Copyright © 2003 IOP Publishing Ltd.

 z
is rational, T (r; f ) ∝ log r and σ = 0. A fast growing function like ee leads to
T ∝ er and thus σ = ∞.
In what follows, we shall introduce the symbols ,  and ≺ which denote
equality, inequality and strict inequality respectively up to a function of r which
remains bounded when r → ∞. The two basic relations which reproduce the
statement on the affinity of f for ∞, 0 or a are

T (r; 1/f )  T (r; f ) (3.11)

T (r; f − a)  T (r; f ). (3.12)

Using these two identities, we can easily prove that the characteristic function of a
homographic transformation of f (with constant coefficients) is equal to T (r; f )
up to a bounded quantity. From a theorem due to Valiron [24], we have
 
P (f )
T r;  sup(p, q)T (r; f ) (3.13)
Q(f )
where P and Q are polynomials in f with constant coefficients, of degrees p and
q respectively, provided the rational expression P /Q is irreducible.
Let us also give some useful classical inequalities:

T (r; fg)  T (r; f ) + T (r; g) (3.14)

T (r; f + g)  T (r; f ) + T (r; g). (3.15)

Another inequality, which was proven in [25], is

T (r; fg + gh + hf )  T (r; f ) + T (r; g) + T (r; h). (3.16)

One last property of the Nevanlinna characteristic was obtained by Ablowitz

et al [22]. In our notation, it reads

T (r; f (z ± 1))  (1 + )T (r + 1; f (z)). (3.17)

This relation (which is valid for r large enough for any given ) makes it possible
to have access to the characteristic, and thus the order, of the solution of some
difference equations.
As an application of the Nevanlinna approach, we shall examine a mapping
which is related to the q-PIII family:
P (xn )
xn+1 xn−1 = . (3.18)
Q(xn )
As we have shown in [25], all the solutions of (3.18) are of infinite order (except
a finite number of constant solutions) if the maximum of the degrees of P , Q
exceeds 2. The main ingredient in the proof of this result is the inequality
T (r; xn+1 xn−1 )  T (r; xn+1 ) + T (r; xn−1 ), leading to

2(1 + )T (r + 1; x)  wT (r; x) (3.19)

Copyright © 2003 IOP Publishing Ltd.

 where w is the sup of the degrees of P , Q. From (3.19), we have

w
T (r + 1; x)  T (r; x). (3.20)
2(1 + )

Now if w > 2, for r large enough one can always choose  small enough so
that λ ≡ w/2(1 + ) becomes strictly greater than unity. The precise meaning of
(3.20) is that, for r large enough, we have

T (r + 1; x) ≥ λT (r; x) − C (3.21)

for some C independent of r. The case C negative is trivial: T (r + k; x) ≥

λk T (r; x). For positive C, we have
 
C C
T (r + 1; x) − ≥ λ T (r; x) − . (3.22)
λ−1 λ−1

Thus, whenever T (r; x) is an unbounded growing function of r (i.e. T  0), then

for some r large enough the right-hand side of this inequality becomes strictly
positive. This will be the case unless x is a constant solution of the mapping,
which is not true for the generic solution. Iterating (3.22), we see that T (r + k; x)
diverges at least as fast as λk , thus log T (r; x) > r log λ and the order σ of x
is infinite. According to our hypothesis, the mapping cannot be integrable. The
general form of (3.18) with quadratic P , Q is

ηxn2 + ζ xn + µ
xn+1 xn−1 = . (3.23)
αxn2 + βxn + γ

Once this degree constraint is obtained, we proceed to the second step by

implementing the singularity confinement criterion. Our assumption is that
once the mappings with very-fast-growing solutions are eliminated, singularity
confinement is sufficient for integrability. At this second step, we are again
dealing with autonomous systems. The third and final step consists in the de-
autonomization, again using the singularity confinement criterion. This three-
tiered approach is easy to implement and has a practically wide range of
applicability. The application of singularity confinement to (3.23) results in
the QRT constraint η = γ or η = α = 0. The de-autonomization of this form,
presented in [26], leads to the q-PIII equation as well as mappings which are
q-discrete forms of PII and PI .
The various discrete integrability detectors we have presented here do not
exhaust the topic: this is a domain under active investigation and more results
are being regularly obtained. We will just mention here the approach recently
proposed by Costin and Kruskal [27] as well a whole line of research based on
algebraic geometry approaches [28].

Copyright © 2003 IOP Publishing Ltd.

 3.3 The showcase
In this section we will review a (short) selection of results obtained in the domain
of discrete integrable systems. In one decade the bulk of accumulated results is
such that several conferences are devoted to the topic (and just as an insider’s
advertisement let us mention here the CIMPA school to be held in 2003 in
Pondicherry organized by some of the coauthors of this article).

3.3.1 The discrete KdV and its de-autonomization

Let us start with the examination of the equation that serves as a paradigm in all
integrability studies, namely KdV, the discrete form of which is [29]

m+1 1 1
Xn+1 = Xnm + m − m+1 . (3.24)
Xn+1 Xn

(Incidentally, this is precisely the equation we have studied in [6], while

investigating the singularity confinement property.) The study of the degree
growth of the iterates in the case of a two-dimensional lattice is substantially
more difficult than that of the one-dimensional case. It is, thus, very important to
make the right choice from the outset. Here are the initial conditions we choose:
on the line m = 0 we take Xn0 of the form Xn0 = pn /q, while on the line n = 0 we
choose X0m = rm /q (with r0 = p0 ). We assign to q and the ps, rs the same degree
of homogeneity. Then we compute the iterates of X using (3.24) and calculate the
degree of homogeneity in p, q, r at the various points of the lattice. Here is what
we find:
.. .. .. .. .. .. .
. . . . . . ..

1 7 19 31 41 51 · · ·

1 5 13 19 25 31 · · ·

1 3 5 7 9 11 · · ·


m 1 1 1 1 1 1 ···
−−→
n
At this point, we must indicate how the analytical expression for the degree can
be obtained. First we compute several points on the lattice which allow us to
have a good guess at how the degree behaves. In the particular case of a two-
dimensional discrete equation relating four points on an elementary square like
(3.24) and with the present choice of initial conditions (and given our experience
on one-dimensional mappings), we can reasonably surmise that the dominant
behaviour of the degree will be of the form dnm ∝ mn. Moreover, the sub-dominant

Copyright © 2003 IOP Publishing Ltd.

 terms must be symmetric in m, n and, at most, linear. With those indications, it
is possible to ‘guess’ the expression dnm = 4mn − 2 max(m, n) + 1 (for mn = 0)
and subsequently calculate some more points in order to check its validity. This
procedure will be used throughout this chapter.
So the lattice KdV equation leads, quite expectedly, to a polynomial growth
in the degrees of the iterates. Let us now turn to the more interesting question of
de-autonomization. The form (3.24) of KdV is not very convenient and thus we
shall study its potential form [30]:

m+1 znm
xn+1 = xnm + . (3.25)
xnm+1 − xn+1
m

(The name ‘potential’ is given here in analogy to the continuous case: the
dependent variable x of equation (3.25) is related to the dependent variable X
of equation (3.24) through xnm+1 − xn+1 m = X m and (3.24) is recovered exactly
n
if zn = 1.) The de-autonomization we are referring to consists in finding an
m

explicit m, n dependence of znm which is compatible with integrability. Let us

first compute the degrees of the iterates for constant z:

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

1 4 7 10 13 16 · · ·

1 3 5 7 9 11 · · ·

1 2 3 4 5 6 ···


m 1 1 1 1 1 1 ···
−−→
n

The degree dnmis given simply by dnm = mn + 1. Assuming a generic (m, n)

dependence for z, we obtain the following successive degrees:

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

1 4 10 20 35 56 · · ·

1 3 6 10 15 21 · · ·

1 2 3 4 5 6 ···


m 1 1 1 1 1 1 ···
−−→
n

Copyright © 2003 IOP Publishing Ltd.

 We note readily that the degrees form a Pascal triangle, i.e. they are identical to the
binomial coefficients, leading to an exponential growth at least on a strip along the
diagonal. The way to obtain an integrable de-autonomization is to require that the
degrees obtained in the autonomous and non-autonomous cases be identical. The
first constraint can be obtained by reducing the degree of x22 from six to five. As
a matter of fact, starting from the initial conditions xn0 = pn /q, x0m = rm /q (with
r0 = p0 ), we obtain x11 = (p1 p0 − p0 r1 − z00 q 2 )/(q(p1 − r1 )), x12 = Q3 /(qQ2 )
where Qk is a polynomial of degree k, and a similar expression for x21 . Computing
x22 , we find x22 = Q6 /(q(p1 − r1 )Q4 ). It is impossible for q to divide Q6 for
generic initial conditions. However, requiring (p1 − r1 ) to be a factor of Q6 , we
find the constraint z11 − z01 − z10 + z00 = 0. The relation of this result to singularity
confinement is quite easy to perceive. The singularity corresponding to q = 0 is,
indeed, a fixed singularity: it exists for all (n, m)s where either n or m are equal
to zero. However, the singularity related to p1 − r1 = 0 appears only at a certain
iteration and is, thus, movable. The fact that with the proper choice of znm the
denominator factors out is precisely what one expects for the singularity to be
confined.
Requiring that z satisfy
m+1
zn+1 − znm+1 − zn+1
m
+ znm = 0 (3.26)

suffices to reduce the degrees of all higher xs to those of the autonomous case. The
solution of (3.26) is znm = f (n) + g(m) where f , g are two arbitrary functions.
This form of znm is precisely the one obtained in the analysis of convergence
acceleration algorithms [31] using singularity confinement. The integrability of
the non-autonomous form of (3.25) (and its relation to cylindrical KdV) has been
discussed by Nagai and Satsuma [32] in the framework of the bilinear formalism.

3.3.2 The discrete Painlevé equations

In order to fix the ideas, let us state at the outset what we mean by the term
discrete Painlevé equation (d-P) [33]. A discrete Painlevé equation is an integrable
(second-order, non-autonomous) mapping which, at the continuous limit, goes
over to one of the continuous Painlevé equations, the latter not necessarily in
canonical form. Thus, the d-Ps constitute the discretizations of the continuous
Painlevé equations. Numerous methods for the derivation of discrete Painlevé
equations do exist. These approaches can be cast roughly into four major classes:
(i) The ones related to some inverse problem. The discrete AKNS method, the
methods of orthogonal polynomials, of discrete dressing, of non-isospectral
deformations, etc belong to this class.
(ii) The methods based on some reduction. Similarity reduction of integrable
lattices is the foremost among them but this class contains the methods
based on limits, coalescences and degeneracies of d-P’s as well as stationary
reductions of non-autonomous differential-difference equations.

Copyright © 2003 IOP Publishing Ltd.

 (iii) The contiguity relations approach. Discrete Ps can be obtained from the auto-
Bäcklund, Miura and Schlesinger transformations of both continuous and
discrete Painlevé equations.
(iv) The direct constructive approach. Two methods fall under this heading. One
is the construction of discrete Painlevé equations from the geometry of some
affine Weyl group. The other is the method of de-autonomization using the
singularity confinement approach.
Our standard approach for the construction of discrete Painlevé equations is
to start from the QRT mapping [18]. The latter (in its symmetric form) is an
autonomous mapping of the form
xn+1 xn−1 f3 (xn ) − (xn+1 + xn−1 )f2 (xn ) + f1 (xn ) = 0 (3.27)
where the fi are specific quartic polynomials involving five parameters. The
solutions of (3.27) can be expressed in terms of elliptic functions. We then
allow the parameters to depend on the independent variable n and single out the
integrable cases. The rationale behind this choice is that, since the continuous
Painlevé equations are the non-autonomous extensions of the elliptic functions,
the discrete Ps should follow the same pattern in the discrete domain. Let us show
how this works in a specific example. We start with the QRT mapping
g(xn − a)(xn − b)
xn+1 xn−1 = (3.28)
(xn − c)(xn − d)
which we have already encountered in the previous section. A careful application
of the singularity confinement criterion leads to the following results: c, d are
parity-dependent constants while a, b depend exponentially on the independent
variable with a even–odd dependence. This thus leads to a natural rewriting of
the equation by separation of even and odd x calling, for instance, the odd ones
y with the redefinition x2n → xn , x2n+1 → yn , and similarly for the (a, b, c, d)
which at odd n’s will now e called (p, r, s, t). Because of the redefinition of n,
what was previously called xn+2 is now just xn+1 . We find two coupled equations
for x and y:
st (xn − an )(xn − bn )
yn yn−1 = (3.29a)
(xn − c)(xn − d)
cd(yn − pn )(yn − rn )
xn+1 xn = (3.29b)
(yn − s)(yn − t)
where c, d, s, t are constants and a, b, p, r are proportional to λn . One constraint
among the coefficients does exist:
pn rn cd = qan bn st. (3.30)
The system (3.29) defines the discrete PVI equation [34]. The interesting result is
the fact that the dependence on the independent variable is exponential. Thus, the
mapping is not a difference equation but rather a q-difference one.

Copyright © 2003 IOP Publishing Ltd.

 We could have derived equation (3.29) directly if we had started from
the asymmetric QRT mapping which is, in fact, a second-order system of two
equations:

xn+1 xn f3 (yn ) − (xn+1 + xn )f2 (yn ) + f1 (yn ) = 0 (3.31a)

yn yn−1 g3 (xn ) − (yn + yn−1 )g2 (xn ) + g1 (xn ) = 0. (3.31b)

The direct de-autonomization of a form such as (3.31) can lead to systems

which are strongly asymmetric, i.e. systems where the two equations do not have
the same overall functional form. Examples of such d-Ps do exist, of course.
Moreover, systems of two first-order mappings do not exhaust all the possible
second-order d-Ps. On several instances, it turns out that the discrete Painlevé
equation can be written as a system involving several variables where all but two
of the equations are local rational relations of the dependent variables.
We cannot hope to give the complete list of discrete Painlevé equations
since, in principle, there is an infinite number of them. Still the basis for
their classification does exist. When discrete Painlevé equations were first
systematically derived, we established what we called the ‘standard’ d-Ps
which fall into a degeneration cascade, i.e. an equation with a given number
of parameters can be obtained from one with more parameters through the
appropriate coalescence procedure. (The word ‘standard’ is to be understood
here as a terminology of the authors, introduced in order to distinguish the
first list of discrete Painlevé equations obtained in the paper [33] from the
remaining equations obtained in subsequent works.) This list comprised three-
point mappings for one dependent variable and was initially incomplete since the
discrete ‘symmetric’ (in the QRT terminology) form of PVI was missing. This gap
has been recently filled in [35] and we can now give the full list of standard d-Ps
(with the necessary caveat as to the meaning of the word ‘standard’ as explained
earlier):

zn
δ-PI xn+1 + xn−1 = −xn + +1
xn
zn xn + a
δ-PII xn+1 + xn−1 =
1 − xn2
(xn − aqn )(xn − bqn )
q-PIII xn+1 xn−1 =
(1 − cxn )(1 − dxn )
(xn2 − a 2 )(xn2 − b 2 )
δ-PIV (xn+1 + xn )(xn + xn−1 ) =
(xn − zn )2 − c2
(xn − a)(xn − 1/a)(xn − b)(xn − 1/b)
q-PV (xn+1 xn − 1)(xn xn−1 − 1) =
(1 − xn qn /c)(1 − xn qn /d)

Copyright © 2003 IOP Publishing Ltd.

 (xn + xn+1 − zn − zn+1 )(xn + xn−1 − zn − zn−1 )
δ-PV
(xn + xn+1 )(xn + xn−1 )
((xn − zn )2 − a 2 )((xn − zn )2 − b 2 )
=
(xn2 − c2 )(xn2 − d 2 )
(xn xn+1 − qn qn+1 )(xn xn−1 − qn qn−1 )
q-PVI
(xn xn+1 − 1)(xn xn−1 − 1)
(xn − aqn )(xn − qn /a)(xn − bqn )(xn − qn /b)
=
(xn − c)(xn − 1/c)(xn − d)(xn − 1/d)

where zn = αn + β, qn = q0 λn and a, b, c, d are constants. We note that both δ

(difference) and q equations figure in this list. The degeneration pattern for these
equations is:

q-PVI / δ-PV / q-PIII

 
δ-PV / δ-PIV / δ-PII / δ-PI .

Although this list is appealingly simple, it is somewhat misleading since it

deals with equations the degrees of freedom of which have been artificially
amputated. In fact, all these equations have richer ‘asymmetric’ forms. However,
one cannot just straightforwardly transpose this degeneration pattern to the case
of asymmetric d-Ps. Thus, we resort to the only sure guide, namely the affine
Weyl groups and the equations related to them [36]:
q
Ee8 A
u: 1
uuu
uu
uu
 uu
E8
q / Eq / Eq / Dq / Aq / (A2 + A1 )q / (A1 + A1 )q / Aq .
7 6 5 4 1

      
Eδ8 / Eδ / Eδ / Dc / Ac / (2A1 )c / Ac
7 6 4 N3N NN 1
NN NN
NN NN
NN NN
NN NN
NN  NNN
N& c & 
A
2
/ Ac
1

In this diagram, we assign to a Weyl group an upper index e if it supports a discrete

equation involving elliptic functions, an upper index q if the equation is of q-type,
an upper index δ if it is a difference equation not explicitly related to a continuous
equation and an upper index c if it is a difference equation which is explicitly the
contiguity relation of one of the (continuous) transcendental Painlevé equations,
namely PVI for D4 , PV for A3 , PIV for A2 , (full) PIII for 2A1 (which means the

Copyright © 2003 IOP Publishing Ltd.

 direct product of twice A1 in a self-dual way), PII for the A1 on the last line
and finally the one-parameter PIII for the A1 on the line above last. Neither PI
nor the zero-parameter PIII appear here, since having no parameter they have no
contiguity relations, hence no discrete difference equation related to them.
It is beyond the scope of this chapter to present examples of discrete Painlevé
equations associated to each of the affine Weyl groups in the degeneration
diagram. These results can be found in [36]. We will limit ourselves to just two
q
examples related to the groups D5 and Dc4 . The notation we will use is qn = q0 λn ,
√
ρn = qn / λ, zn = z0 + nδ and ζn = zn − δ/2. In the first group, we have the
equations which we have encountered already:

(yn − aqn )(yn − qn /a)

xn+1 xn =
(yn − c)(yn − 1/c)
(xn − bρn )(xn − ρn /b)
yn yn−1 =
(xn − d)(xn − 1/d)

where a, b, c, d are four constants. This is a discrete form of PVI derived first by
Jimbo and Sakai [34]. In the second group, Dc4 , we have [37]

ζn+1 zn byn + (1 − yn2 )(zn /2 + c)

+ = zn + a +
1 − xn+1 yn 1 − yn xn (1 + dyn )(1 + yn /d)
zn ζn bxn + (1 − xn2 )(zn /2 − c)
+ = ζn + a +
1 − yn xn 1 − xn yn−1 (1 + dxn )(1 + xn /d)

where a, b, c and d are four constants. In the same space, one also has, in a
different direction [38],

(ym − zm )2 − a
xm+1 xm = 2 −b
ym
ζm − c ζm + c
ym + ym−1 = +
1 + dxm 1 + xm /d

where a, b, c and d are four constants.

The discrete Painlevé equations have several properties which are perfectly
parallel to those of their continuous analogues. Prominent among these is the
existence of special solutions. Let us sketch here the method for their derivation.
We start from an equation that is given in the QRT [18] asymmetric form:

xn+1 xn f3 (yn ) − xn+1 f4 (yn ) − xn f2 (yn ) + f1 (yn ) = 0 (3.32a)

yn yn−1 g3 (xn ) − yn−1 g4 (xn ) − yn g2 (xn ) + g1 (xn ) = 0 (3.32b)

where fi , gi are specific quartic polynomials of y and x, respectively, with

coefficients which may depend on the independent variable n. As we have
explained in [39], the special-function-type solutions of the Painlevé equations

Copyright © 2003 IOP Publishing Ltd.

 are obtained whenever the latter can be solved through a Riccati equation.
Transposing this to the discrete case, we seek a solution of system (3.32) in the
form [40]
αyn + βγ
xn+1 = (3.33a)
γ yn + δ
xn + ζ
yn = (3.33b)
ηxn + θ

i.e. in the form of a homographic mapping, since the latter is the discrete analogue
of the Riccati equation. The coefficients α, β, . . . , θ appearing in (3.33) depend,
in general, on the dependent variable n. The existence of a solution in the form of
(3.33) is possible only when some special relation exists between the parameters
of the discrete Painlevé equation. We shall refer to this relation as ‘linearizability
constraint’. Eliminating yn between (3.33a) and (3.33b), one can obtain a Riccati
relation between xn+1 and xn of the form

an xn + bn
xn+1 = . (3.34)
cn x n + d n
The linearization we referred to earlier is one obtained through a Cole–Hopf
transformation x = u/v. Substituting into equation (3.34), we obtain a linear
equation for v:

cn cn+2 vn+2 − (dn+1 cn + an cn+1 )vn+1 + (an dn − bn cn )vn = 0. (3.35)

We give here the linearization of the two examples we have already presented. For
q
the D5 , asymmetric q-PIII , the linearizability condition has already been obtained
by Jimbo and Sakai [34]. The constraint reads

cdµ = ab (3.36)
√
where µ = λ. The homographic system
yn − aqn
xn+1 = (3.37a)
d(yn − c)
xn − bρn
yn = (3.37b)
c(xn − d)
leads, with x = u/v, to the linear equation

vn+2 + (qn (ac + bdµ) − c2 d 2 − 1)vn+1 + cd(aqn − c)(bρn − d)vn = 0.

(3.38)
The latter was identified by Jimbo and Sakai as the equation for the
q-hypergeometric 2 φ1 function. Through the use of an appropriate gauge in v,

Copyright © 2003 IOP Publishing Ltd.

 this equation can be arranged so as to have linear dependence in q for all three
coefficients.
For the Dc4 , discrete PV we consider the second of the two equations given
earlier. This equation can be obtained as two different coalescence limits, from the
asymmetric δ-PIV and q-PIII [38] and also as a contiguity relation of the solutions
of the continuous Painlevé VI [37]:

(yn − zn )2 − p2
xn+1 xn = (3.39a)
yn2 − a 2
ζn − r ζn + r
yn + yn−1 = + (3.39b)
1 − bxn 1 − xn /b

where a, b, p and r are four constants. The linearizability constraint can be

obtained from either of the limits (and also from the relation to PVI ):

a + p + r + δ/2 = 0. (3.40)

The discrete Riccati

yn − zn + p
xn+1 = b (3.41a)
yn − a
zn + p + abxn
yn = (3.41b)
1 − bxn
leads to the linear equation

vn+2 + (zn (b2 + 1) + (b 2 − 1)(a − p) + δ)vn+1 + b 2 (zn − a − p)vn = 0.

(3.42)
Equation (3.42) can be further transformed through the gauge transformation
vn = φn wn , where φn+1 = −(zn + a + p)φn :

(zn+1 + a + p)wn+2 − (zn (b2 + 1) + (b 2 − 1)(a − p) + δ)wn+1

+ b 2 (zn − a − p)wn = 0. (3.43)

It can be easily shown that (3.43) is satisfied by the Gauss hypergeometric

function wn = F ((zn + p + a)/δ, 2p/δ; 1 + 2(a + p)/δ; 1 − 1/b 2). Equation
(3.43) is just a contiguity relation [41] of the latter.

3.3.3 Linearizable systems

The third case study we are going to present here is that of linearizable systems.
While soliton equations, like KdV and discrete Painlevé equations, can be
integrated through spectral methods, there exists a whole family of systems the
integration of which is considerably simpler: they can be reduced to a linear
mapping through a local transformation.

Copyright © 2003 IOP Publishing Ltd.

 While studying the growth properties of linearizable mappings, we
discovered a most interesting property: when a second-order mapping is
linearizable, its degree growth is linear at maximum. Let us illustrate this through
an example:
xn (xn − yn − a)
xn+1 = (3.44a)
xn2 − yn
(xn − yn )(xn − yn − a)
yn+1 = (3.44b)
xn2 − yn
where a was taken constant. We start by assuming that a is an arbitrary function
of n and compute the growth of the degree. We find dxn = 0, 1, 2, 3, 4, 5,
6, 7, 8, . . . and dyn = 1, 2, 3, 4, 5, 6, 7, 8, 9, . . . , i.e. again a linear growth.
This is an indication that (3.44) is integrable for arbitrary an and indeed it
is. Dividing the two equations, we obtain yn+1 /xn+1 = 1 − yn /xn , i.e. yn /xn =
1/2 + k(−1)n , whereupon (3.44) is reduced to a homographic mapping for x.
We turn now to the three-point mapping we have studied in [42, 43] from
the point of view of integrability, in general, and linearizability, in particular. The
generic mapping studied in [43] was one trilinear in xn , xn+1 , xn−1 . Several cases
can be considered. Our starting point is the mapping
xn+1 xn xn−1 + βxn xn+1 + ζ ηxn+1 xn−1 + γ xn xn−1
+ βγ xn + ηxn−1 + ζ xn+1 + 1 = 0. (3.45)
We start with the initial conditions x0 = r, x1 = p/q and compute the
homogeneous degree in p, q at every n. We find dn = 0, 1, 1, 2, 3, 5, 8, 13, . . . ,
√ = dn + dn−1 leading to exponential growth of dn
i.e. a Fibonacci sequence dn+1
with asymptotic ratio (1 + 5)/2. Thus mapping (3.45) is not expected to be
integrable in general. However, as shown in [43], integrable sub-cases do exist.
We start by requiring that the degree growth be less rapid and as a drastic decrease
in the degree, we demand that d3 = 1 instead of 2. We find that this is possible
when either β = ζ = 0, in which case the mapping reduces to
η 1
xn+1 = −γ − − (3.46)
xn xn xn−1
or γ = η = 0, giving a mapping identical to (3.46) after x → 1/x. In this case the
degree is dn = 1 for n > 0. The linearization of (3.46) can be obtained in terms
of a projective system [43], i.e. a system of three linear equations, a fact which
explains the constancy of the degree.
The trilinear three-point mapping possesses also many non-generic sub-
cases, some of which are integrable. The first non-generic case can be written
as
xn (γ xn−1 + ) + (xn+1 + 1)(ηxn−1 + 1) = 0. (3.47)
The degrees of the iterates of mapping (3.47) again form a Fibonacci sequence
even in the case  = 0 or η = 0. The only case that presents slightly different

Copyright © 2003 IOP Publishing Ltd.

 behaviour is the case γ = 0:

(xn+1 + 1)(ηxn−1 + 1) + xn = 0. (3.48)

In the generic case, the degree of the iterate behaves like dn = 0, 1, 1, 1, 2, 2,

3, 4, 5, 7, 9, 12, 16, 21, 28, 37, 49, . . . , satisfying the recursion relation dn+1 =
dn−1 + dn−2 leading to an exponential growth with asymptotic ratio
  1/3   1/3
1 23 1 23
+ + − .
2 108 2 108

Although the mapping is generically non-integrable, it does possess integrable

sub-cases. Requiring, for example, that d4 = 1, we obtain the constraint
 = η = 1 and the mapping becomes periodic with period 5. If we require
d5 = 1, we obtain n = −ηn+1 (ηn − 1) and ηn+1 ηn ηn−1 − ηn+1 ηn + ηn+1 −
1 = 0, leading again to a periodic mapping with period 8. In these cases,
the degree of the iterates exhibits, of course a periodic behaviour. A
more interesting result is obtained if we require d9 < 7. We find that the
condition η = 1 and  an arbitrary constant leads to a non-exponential
degree growth dn = 0, 1, 1, 1, 2, 2, 3, 4, 5, 6, 7, 9, 10, 12, 14, 15, 18, 20, 22,
25, 27, 30, 33, 36, 39, 42, 46, 49, . . . . Although the detailed behaviour of dn is
pretty complicated, one can see that the growth is quadratic: we have, for example,
d4m+1 = m(m + 1) for m > 0. Thus, this mapping is expected to be integrable and
it is indeed a member of the QRT family. Its constant of motion is given by
   
yn+1 yn 1 1 2
K = yn+1 + yn −  + + ( + 1) + −
yn yn+1 yn yn+1 yn yn+1

where yk = xk + 1. The second non-generic case is

γ xn xn−1 + δxn+1 xn−1 + xn + ζ xn+1 = 0. (3.49)

A study of the degree √ growth always leads to exponential growth with

asymptotic ratio (1 + 5)/2, except when γ = 0, in which case the degrees
obey the recurrence dn+1 = dn−1 + dn−2 . No integrable sub-cases are expected
for mapping (3.49). The last non-generic case we shall examine is

γ xn xn−1 + xn+1 xn−1 + xn + ηxn−1 = 0. (3.50)

Again the degree sequence is a Fibonacci one except when γ = 0 or η = 0, in

which case we have the recursion dn+1 = dn−1 + dn−2 , or when n = γn ηn−2 . In
the latter case the degree growth follows the pattern dn = 0, 1, 1, 2, 2, 3, 3, . . . ,
i.e. a linear growth. Thus, we expect this case to be integrable. This is
precisely what we found in [43]. Assuming η = 0, we can scale it to η =
1 and, thus,  = γ . The mapping can then be integrated to the homography

Copyright © 2003 IOP Publishing Ltd.

 (xn−1 + 1)(xn + 1) = kaxn−1 , where k is an integration constant and a is related
to γ through γn = −an+1 /an . Thus, in this case mapping (3.50) is a discrete
derivative of a homographic mapping.
At this point, a most interesting question can be formulated: what is the
influence of singularity confinement on linearizability. As we have shown in [44],
a mapping which is linearizable does not necessarily possess the singularity
confinement property. Several mappings derived in [45] as special limits of
discrete Painlevé equations can be linearized in this way. For instance, the
nonlinear equation
  
xn+1 + xn − a xn xn−1 + xn − a xn xn2
− − − =M (3.51)
zn+1 ζn zn ζn ζn2

with a a constant, where zn and ζn are defined from a single arbitrary function g
of n through zn = gn+1 + gn−1 , ζn = gn+1 + gn , can be solved through the linear
equation
An xn+1 + Bn (xn − a) + An+1 xn−1
=K (3.52)
zn xn+1 + (zn+1 + zn )(xn − a) + zn+1 xn−1
where An = gn2 (gn+1 + gn−1 ) and Bn = −(gn+1 + gn )gn+2 gn−1 − (gn+2 +
gn−1 )gn+1 gn . This mapping, while linearizable, is generically non-confining
unless g is a constant.
However, singularity confinement still plays an important role. As a matter
of fact, while a generic linearizable mapping has linear growth a confining
linearizable mapping has zero degree growth. The simplest example of this is
projective mapping but more complicated examples do exist.

3.4 Beyond the discrete horizon

Difference equations, be they ordinary or partial, do not exhaust, and by far, the
domain of applicability of discrete systems. In what follows we shall present two
classes of systems which extend the discrete ones (albeit in opposite directions).

3.4.1 Differential-difference systems

In this section we shall focus on equations of the form un+1 = F (un−1 un ,
u
n , n, t) with F homographic in un−1 , rational in un , u
n and analytic in n, t and
where the prime denotes the derivative with respect to t. Given an equation of
this form, we iterate an initial conditions in homogeneous coordinates u0 = p,
u1 = q/r, where p, q, r are functions of t. We assign to p (and t) the degree
zero, the degree 1 to q and r and their derivatives and compute the degree dn
of homogeneity of the numerator and denominator of un at every iteration. A
different choice of u0 could have been possible but it turns out that the present
choice of zero-degree u0 considerably simplifies the calculations.

Copyright © 2003 IOP Publishing Ltd.

 We shall start with two well-known integrable systems. The first is the Kac–
Moerbeke equation [46], also known as the Lotka–Volterra or semi-discrete KdV
equation:
u

un+1 = un−1 + n . (3.53)

un
Let us explain how the degree growth is computed. We start from u0 = p,
u1 = q/r and compute the first few iterates of (3.53). We thus obtain

pqr − q
r − qr

u2 =
qr
q 2 (pqr − q
r − qr
+ r
2 − rr

) + r 2 (p
q 2 + qq

− q
2 )
u3 =
(pqr − q
r − qr
)qr
and so on. Since p and t are of degree 0 and q and r of degree 1, we find that
the homogeneity degrees of the numerator and denominator of u2 and u3 are,
respectively, d2 = 2 and d3 = 4. Computing the degree of the successive iterates,
we find dn = 0, 1, 2, 4, 7, 11, 16, 22, . . . , i.e. given by dn = (n2 − n + 2)/2 for
n > 0. The fact that the degree growth is polynomial is not astonishing given that
the Kac–Moerbeke system is integrable. The second system we shall examine is
the semi-discrete mKdV equation [47]:

u
n
un+1 = un−1 + . (3.54)
u2n − 1

Again we find a polynomial growth dn = 0, 1, 2, 5, 8, 13, 18, 25, . . . . We have,

indeed, d2m = 2m2 and d2m+1 = 2m2 + 2m + 1. Again non-exponential growth
is expected since the integrability of (3.54) is well established.
Once our approach has passed this basic test, it is natural to ask how we
can generalize equations (3.53) and (3.54). In order to keep this search for
generalizations manageable, we shall limit ourselves to equations of the form

αu
n + βu2n + γ un + δ
un+1 = un−1 + (3.55)
κu
n + ζ u2n + ηun + θ

where α, . . . , θ, κ are functions of n and t. Our approach will be based on a

dual singularity confinement/low-growth requirement strategy. We shall start by
reducing the possible integrable forms of (3.55) using the necessary criterion of
singularity confinement and then analyse the reduced form through the study of
the degree growth. We start by supposing that κ = 0 (we take κ = 1), in which
case by translation we can put α = 0. Let us assume that un is such that u
n +
ζ u2n + ηun + θ has a simple zero for some t = t0 . In this case, un+1 will have a
simple pole, un+1 ∝ 1/(t − t0 ). This singularity will propagate indefinitely, i.e.
un+3 , un+5 etc will also have poles unless the following conditions are fulfilled:
β = γ = ζ = 0; and ηn+1 = ηn−1 , θn+1 = θn−1 , δn+1 − 2δn + δn−1 = 0. Thus, δ

Copyright © 2003 IOP Publishing Ltd.

 is linear in n while η and θ are n-independent with even/odd dependence, which
means that we have ηe (t), θe (t) for even n and  different ηo (t), θo (t) for odd
n. Introducing u = ξe,o v, where ξe,o = exp(− ηe,o dt), we can transform the
equation to
δn
ξe ξo (vn+1 − vn−1 ) =
. (3.56)
vn + θe,o /ξe,o

The factor ξe ξo can be absorbed in δ and a translation of v, by θe,o /ξe,o dt,
allows us to put θ in the denominator of (3.56) to zero. Thus, we arrive finally at
the equation
λ(t)n + µ(t)
vn+1 − vn−1 = (3.57)
vn

where, moreover, it is possible, through a suitable redefinition of time, to take

λ = 1. This equation, as we explained earlier, is a candidate for integrability.
Once this reduced form is obtained through singularity confinement, we can
apply the non-exponential growth criterion. We start by considering the equation
vn+1 − vn−1 = a(n, t)/vn
where a is a priori an arbitrary function of n and t.
We compute, as in the case of systems (3.53) and (3.55), the degree growth
starting from v0 = p and v1 = q/r and obtain the exponentially growing sequence
dn = 0, 1, 2, 4, 8, 16, . . . , i.e. dn = 2n−1 for n > 0. Next we ask how is it
possible to curb this growth and it turns out that we can, for n = 4, obtain a
condition for the degree to be six rather than eight. This condition is an+1 −
2an + an−1 = 0, i.e. a must be a linear function of n, in perfect agreement with
the singularity confinement criterion. Implementing this constraint, we can now
compute the degree growth for equation (3.57). We now obtain the sequence
dn = 0, 1, 2, 4, 6, 9, 12, 16, . . . , i.e. d2m−1 = m2 and d2m = m(m + 1), which
are precisely the same values as the ones obtained when a is a constant (in
both n and t). Thus, the low-growth requirement criterion confirms the possibly
integrable character of (3.57). Although this is not a proof of its integrability,
the fact that this new criterion is satisfied strengthens the argument in favour of
integrability. We shall come back again to this equation and show that it can be
transformed into a known integrable system. For the time being, we compute its
continuous limit. Equation (3.57) is another differential-difference form of the
potential KdV equation. Introducing the continuous variables x = (n + t), s =
 3 t and taking v(n, t) = n − t + w(x, s), a(n, t) = 2(−1 +  4 (b
(s)x + c(s))),
we find, at the limit  → 0, the equation
ws + wx2 − 16 wxxx = b
(s)x + c(s). (3.58)
This is indeed a potential form of KdV. Differentiating once with respect to x, we
obtain for the quantity W = wx − b(s) the equation
Ws + 2W Wx − 16 Wxxx = −2b(s)Wx . (3.59)

Introducing the new variables T = s and X = x − 2 b(s) ds, we find finally
WT + 2W WX − 16 WXXX = 0 (3.60)

Copyright © 2003 IOP Publishing Ltd.

 i.e. the KdV equation. We can now show how equation (3.57) can be integrated
[48]. Starting from (3.57), we introduce wn = vn+1 − vn−1 . We then have for w
the equation
wn
an+1 an−1
= − . (3.61)
wn wn wn+1 wn wn−1
Next we introduce the variable un = −1/wn wn−1 and using (3.61) we recover the
non-autonomous extension to the Kac–Moerbeke equation:

vn

an+1 vn+1 − an−2 vn−1 = + (an−1 − an )vn

vn
which was introduced in [49] by Cherdantsev and Yamilov.

3.4.2 Ultra-discrete systems

In this section, we shall present another extension of discrete systems: ultra-
discrete ones. This name is used to designate systems where the dependent
variables, as well as the independent ones, take only discrete values. In this respect
ultra-discrete systems are generalized cellular automata. The name of ultra-
discrete is reserved for systems obtained from discrete ones through a specific
limiting procedure introduced in [50], by the Tokyo–Kyoto group.
Before introducing the ultra-discrete limit, let us first consider the question
of nonlinearity. How simple can a nonlinear system be and still be genuinely
nonlinear. The nonlinearities we are accustomed to, i.e. ones involving powers,
are not necessarily the simplest. It turns out (admittedly with hindsight) that
the simplest nonlinear function of x one can think of is |x|. It is indeed linear
for both x > 0 and x < 0 and the nonlinearity comes only from the different
determinations. Thus, one would expect the equations involving nonlinearities
only in terms of absolute values to be the simplest. The ultra-discrete limit does
just that, i.e. it converts a given (discrete) nonlinear equation to one where only
absolute-value nonlinearities appear. The key relation is the following limit:

lim  log(1 + ex/ ) = max(0, x) = (x + |x|)/2. (3.62)

→0+

Other equivalent expressions exist for this limit and the notation that is often
used is the truncated power function (x)+ ≡ max(0, x). It is easy to show that
lim→0+  log(ex/ + ey/ ) = max(x, y) and the extension to n terms in the
argument of the logarithm is straightforward.
Two remarks are in order at this point. First, since the function (x)+ takes
only integer values when the argument is integer, the ultra-discrete equations
can describe generalized cellular automata, provided one restricts the initial
conditions to integer values. This approach has already been used in order to
introduce cellular automata (and generalized cellular automata) related to many
interesting evolution equations [50]. Second, the necessary condition for the

Copyright © 2003 IOP Publishing Ltd.

 procedure to be applicable is that the dependent variables be positive, since we
are taking a logarithm and we require that the result take values in Z. This
means that only some solutions of the discrete equations will survive in the ultra-
discretization.
As an illustration of the method and a natural introduction to ultra-discrete
Painlevé equations, let us consider the following discrete Toda system [51]:

utn+1 − 2utn + utn−1 = log(1 + δ 2 (eun+1 − 1)) − 2 log(1 + δ 2 (eun − 1))

t t

t
+ log(1 + δ 2 (eun−1 − 1)) (3.63)

which is the integrable discretization of the continuous Toda system:

d2 rn
= ern+1 − 2 ern + ern−1 . (3.64)
dt 2
For the ultra-discrete limit, one introduces w through δ = e−L/2 , wnt = utn − L
and takes the limit  → 0. Thus, the ultra-discrete limit of (3.63) simply becomes

wnt +1 − 2wnt + wnt −1 = (wn+1

t
)+ − 2(wnt )+ + (wn−1
t
)+ . (3.65)

Equation (3.65) is the cellular automaton analogue of the Toda system (3.64).
Let us now restrict ourselves to a simple periodic case with period two, i.e.
rn+2 = rn and similarly wn+2 = wn . Calling r0 = x and r1 = y, we have from
(3.64) the equation ẍ = 2 ey − 2 ex and ÿ = 2 ex − 2 ey , resulting in ẍ + ÿ = 0.
Thus, x + y = µt + ν and we obtain, after some elementary manipulations,

ẍ = a eµt e−x − 2 ex . (3.66)

Equation (3.66) is a special form of the Painlevé PIII equation. Indeed, putting
v = ex−µt /2, we find that
v̇ 2
v̈ = + eµt /2 (a − 2v 2 ). (3.67)
v
The same periodic reduction can be performed on the ultra-discrete Toda equation
(3.65). We introduce w0t = Xt , w1t = Y t and have, in perfect analogy to the
continuous case, Xt +1 − 2Xt + Xt −1 = 2(Y t )+ − 2(Xt )+ and Y t +1 − 2Y t +
Y t −1 = 2(Xt )+ − 2(Y t )+ . Again, 2t (Xt + Y t ) = 0 and we can take Xt + Y t =
mt + p (where m, t, p take integer values). We thus find that X obeys the ultra-
discrete equation:

Xt +1 − 2Xt + Xt −1 = 2(mt + p − Xt )+ − 2(Xt )+ . (3.68)

This is the ultra-discrete analogue of the special form (3.67) of the Painlevé PIII
equation.
In order to construct the ultra-discrete analogues of the Painlevé equations,
we must start with the discrete form that allows the ultra-discrete limit to be taken.

Copyright © 2003 IOP Publishing Ltd.

 The general procedure is to start with an equation for x, introduce X through
x = eX/ and then take appropriately the limit  → 0. Clearly the substitution
x = eX/ requires x to be positive. This is a stringent requirement that limits
the exploitable form of the d-Ps to multiplicative ones. Fortunately, many such
forms are known for the discrete Painlevé transcendents. We have, for instance,
for d-PI−1 the multiplicative q forms:
λn 1
d-PI−1 : xn+1 xn−1 = + 2 (3.69)
xn xn
1
d-PI−2 : xn+1 xn−1 = λn + (3.70)
xn
d-PI−3 : xn+1 xn−1 = λn xn + 1. (3.71)

From them it is straightforward to obtain the canonical forms of the ultra-

discrete PI :

u-PI−1 : Xn+1 + Xn−1 + 2Xn = (Xn + n)+ (3.72)

u-PI−2 : Xn+1 + Xn−1 + Xn = (Xn + n)+ (3.73)
u-PI−3 : Xn+1 + Xn−1 = (Xn + n)+ . (3.74)

Ultra-discrete forms have been derived for all Painlevé equations [52]. Moreover,
we have shown that their properties are perfectly parallel with those of their
discrete and continuous analogues (degeneration through coalescence, existence
of special solutions, auto-Bäcklund and Schlesinger transformations).

3.5 Parting words

In this short review, we have tried to present a selection of results on discrete
integrable systems. This review is far from being exhaustive: the domain has
simply mushroomed over the past decade and any attempt at exhaustiveness is
bound to fail. Thus, we have preferred to focus on two topics which have been
among the main themes of our work: integrability detectors and discrete Painlevé
equations. It would have been fair, to the more mathematically oriented at least,
to present a clear definition of what is meant by discrete integrability. Certainly
by now the reader must have deduced that the term is a blanket one covering
both integrability through spectral methods (and its sub-case, where constants of
motion do exist making the reduction of the system possible) and linearizability.
As we saw, the two types of discrete integrability are not associated with the same
behaviour with respect to integrability detectors.
Before closing this review, we wish to stress one important point. Discrete
systems are fundamental entities, more fundamental than continuous ones. As a
matter of fact, continuous systems can be obtained as limits of discrete ones. What
is even more important for integrability practitioners is that the continuous limit
generically preserves the invariances and symmetries of the discrete system. Thus,

Copyright © 2003 IOP Publishing Ltd.

 one expects to obtain integrable continuous systems when taking the appropriate
limits of integrable discrete ones. In contrast, the discretization of an integrable
continuous system leads, in general, to a non-integrable one. (This explains
why the study of discrete integrability is a highly refined art.) In the opposite
direction, ultra-discrete cellular-automaton-like systems can also be obtained
starting from discrete ones. Again, the ultra-discretization preserves the invariants
of the discrete system.
Finally, despite the arguments presented in both the introduction and the
conclusion, one can still wonder about the physical relevance of discrete systems.
Thus, one can ask the ultimate question concerning the discrete nature of the
physical world. But how can we refer to the world we know as ‘discrete’ (be
it only for modelling purposes) when our experience (and several centuries
of physical theories) have accustomed us to thinking in terms of continuous
spacetime. But what is the evidence that the world is, indeed, continuous? What
our senses, and the extended senses that constitute the physical instruments, tell
us is that the world looks continuous all the way down to the measurement limits
(and there is no sign that this continuity may disappear with further increase in
the precision of the measurements). Still, that is all that the experiments can give:
upper limits to the lattice length of a supposedly discrete spacetime. There exists,
indeed, serious speculation that spacetime may be discrete at lengths way beyond
our experimental possibilities, typically the Planck length (10−32 cm) [53]. If this
were true, this would mean that the true equations of motion of the world would
be discrete equations. The continuous ones, with which we are familiar, would
thus appear as limiting cases of the more fundamental, discrete ones and the
invariances (that play such a major role in physics) only approximate properties
that do not survive discretization, unless the dynamical (discrete) equations also
have invariances and symmetries themselves. This is par excellence the domain of
integrable discrete systems. Thus, studies of discrete integrability, apart from the
purely mathematical interest they present, may play an important role in forging
the appropriate tools for the investigation of the physical world.

Acknowledgments
This review was made possible thanks to invitations from Paris VII University for
T Tamizhmani and from Ecole Polytechnique for KM Tamizhmani.

References
[1] Grammaticos B and Ramani A 1997 Integrability of Nonlinear Systems (Lect.
Notes Phys. vol 495) ed Y Kosmann-Schwarzbach, B Grammaticos and K M
Tamizhmani (Berlin: Springer) p 30
[2] Calogero F 1990 What is Integrability? ed V Zakharov (Berlin: Springer) p 1
[3] Painlevé P 1902 Acta Math. 25 1
[4] Ramani A, Grammaticos B and Bountis T 1989 Phys. Rep. 180 159

Copyright © 2003 IOP Publishing Ltd.

 [5] Ablowitz M J, Ramani A and Segur H 1978 Lett. Nuovo Cimento 23 333
[6] Grammaticos B, Ramani A and Papageorgiou V 1991 Phys. Rev. Lett. 67 1825
[7] Grammaticos B, Ramani A and Tamizhmani K M 1994 J. Phys. A: Math. Gen. 27
559
[8] Grammaticos B, Nijhoff F and Ramani A 1999 Discrete Painlevé equations The
Painlevé Property: One Century Later (CRM Series in Mathematical Physics)
ed R Conte (Berlin: Springer) p 413
[9] Papageorgiou V G, Nijhoff F W, Grammaticos B and Ramani A 1992 Phys. Lett. A
164 57
[10] Ramani A, Grammaticos B, Tamizhmani T and Tamizhmani K M 1999 J. Phys. A:
Math. Gen. 32 1
[11] Satsuma J 1992 Private communication
[12] Conte R and Musette M 1996 Phys. Lett. A 223 439
[13] Grammaticos B and Ramani A 2000 Chaos Solitons Fractals 11 7
[14] Hietarinta J and Viallet C 1998 Phys. Rev. Lett. 81, 325
[15] Arnold V I 1990 Bol. Soc. Bras. Mat. 21 1
[16] Veselov A P 1992 Commun. Math. Phys. 145 181
[17] Bellon M P and Viallet C-M 1999 Commun. Math. Phys. 204 425
Bellon M P, Maillard J-M and Viallet C-M 1991 Phys. Rev. Lett. 67 1373
[18] Quispel G R W, Roberts J A G and Thompson C J 1989 Physica D 34 183
[19] Grammaticos B, Ramani A and Lafortune S 1998 Physica A 253 260
[20] Yanagihara N 1985 Arch. Ration. Mech. Anal. 91 169
[21] Grammaticos B, Ramani A and Moreira I 1993 Physica A 196 574
[22] Ablowitz M J, Halburd R and Herbst B 2000 Nonlinearity 13 889
[23] Hille E 1976 Ordinary Differential Equations in the Complex Domain (New York:
Wiley)
[24] Valiron G 1931 Bull. Soc. Math. France 59 17
[25] Grammaticos B, Tamizhmani T, Ramani A and Tamizhmani K M 2001 J. Phys. A:
Math. Gen. 34 3811
[26] Ramani A and Grammaticos B 1996 Physica A 228 160
[27] Costin O and Kruskal M D 2002 Equivalent of the Painlevé property for certain
classes of difference equations and study of their solvability Preprint
[28] Takenawa T 2001 J. Phys. A: Math. Gen. 34 L95
[29] Hirota R 1977 J. Phys. Soc. Japan 43 1424
Papageorgiou V, Nijhoff F W and Capel H 1990 Phys. Lett. A 147 106
[30] Capel H, Nijhoff F W and Papageorgiou V 1991 Phys. Lett. A 155 337
[31] Papageorgiou V, Grammaticos B and Ramani A 1993 Phys. Lett. A 179 111
[32] Nagai A and Satsuma J 1995 Phys. Lett. A 209 305
[33] Ramani A, Grammaticos B and Hietarinta J 1991 Phys. Rev. Lett. 67 1829
[34] Jimbo M and Sakai H 1996 Lett. Math. Phys. 38 145
[35] Grammaticos B and Ramani A 1999 Phys. Lett. A 257 288
[36] Grammaticos B and Ramani A 2000 Reg. Chaotic Dyn. 5 53
[37] Nijhoff F W, Ramani A, Grammaticos B and Ohta Y 2000 Stud. Appl. Math. 106 261
See also Okamoto K 1987 Ann. Mat. Pura Applicata 146 337, where the birational
tranformations of PVI are given. Using these Schlesinger transformations one can
derive the discrete equations related to Dc4
[38] Grammaticos B, Ohta Y, Ramani A and Sakai H 1998 J. Phys. A: Math. Gen. 31 3545

Copyright © 2003 IOP Publishing Ltd.

 [39] Tamizhmani K M, Ramani A, Grammaticos B and Kajiwara K 1998 J. Phys. A: Math.
Gen. 31 5799
[40] Tamizhmani T, Tamizhmani K M, Grammaticos B and Ramani A 1999 J. Phys. A:
Math. Gen. 32 4553
[41] Abramowitz A and Stegun I 1965 Handbook of Mathematical Functions (New York:
Dover)
[42] Ramani A, Grammaticos B and Karra G 1992 Physica A 180 115
[43] Ramani A, Grammaticos B, Tamizhmani K M and Lafortune S 1998 Physica A 252
138
[44] Ramani A, Grammaticos B and Tremblay S 2000 J. Phys. A: Math. Gen. 33 3045
[45] Ramani A, Grammaticos B, Ohta Y and Grammaticos B 2000 Nonlinearity 13 1073
[46] Kac M and van Moerbeke P 1975 Adv. Math. 16 160
[47] Ablowitz M J and Ladik J F 1977 Stud. Appl. Math. 57 1
[48] Tamizhmani K M, Ramani A, Grammaticos B and Ohta Y 1999 J. Phys. A: Math.
Gen. 32 6679
[49] Cherdantsev I and Yamilov R 1995 Physica D 87 140
[50] Tokihiro T, Takahashi D, Matsukidaira J and Satsuma J 1996 Phys. Rev. Lett. 76 3247
[51] Grammaticos B, Ohta Y, Ramani A, Takahashi D and Tamizhmani K M 1997 Phys.
Lett. A 226 53
[52] Grammaticos B, Ohta Y, Ramani A and Takahashi D 1998 Physica D 114 185
[53] Einstein A 1950 Physics and Reality (Essays in Physics) (New York: Philosophical
Library)
Feynman R P 1982 Int. J. Theor. Phys. 21 467
Witten E 1996 Reflections on the fate of spacetime Physics Today April 24

Copyright © 2003 IOP Publishing Ltd.

 Chapter 4

The dbar method: a tool for solving

two-dimensional integrable evolution PDEs
A S Fokas
Department of Applied Mathematics and Theoretical Physics,
University of Cambridge, UK

4.1 Introduction
There exists a large class of nonlinear evolution PDEs in one space variable
which can be treated analytically. Such equations are called integrable and
the method for solving the initial value problem on the infinite line for such
equations is called the inverse scattering method or inverse spectral method.
The most well-known integrable equations are the nonlinear Schrödinger, the
Korteweg–deVries and the sine-Gordon equations. The inverse scattering method
is based on a certain mathematical problem in the theory of functions of
one complex variable called the Riemann–Hilbert (RH) problem (see [1] for
an introduction). Some integrable evolution equations in one space dimension
possess particular solutions, which are localized in space and which retain their
shape upon interaction with any other localized disturbance. Such solutions are
called solitons; they are important not because they are exact solutions but because
they characterize the long-time behaviour of the solution. Indeed, it can be shown
that the large-time asymptotics of the solution of integrable evolution equations
in one space dimension is dominated by solitons [2]. Solitons appear in a large
number of physical circumstances, including fluid mechanics, nonlinear optics,
plasma physics, quantum field theory, relativity, elasticity, biological models,
nonlinear networks, etc [3]. This is a consequence of the fact that a soliton is
the realization of a certain physical coherence which is natural to a variety of
nonlinear phenomena.

4.1.1 The dbar method

Every integrable nonlinear evolution equation in one spatial dimension has several
integrable versions in two spatial dimensions. Two such integrable physical

Copyright © 2003 IOP Publishing Ltd.

 generalizations of the Korteweg–deVries equation are the so-called Kadomtsev–
Petviashvili I (KPI) and II (KPII) equations. In the context of water waves, they
arise in the weakly nonlinear, weakly dispersive, weakly two-dimensional limit
and in the case of KPI when the surface tension is dominant. The nonlinear
Schrödinger equation also has two physical integrable versions known as the
Davey–Stewartson I (DSI) and the DSII equations. They can be derived from
the classical water wave problem in the shallow water limit and govern the time
evolution of the free surface envelope in the weakly nonlinear, weakly two-
dimensional, nearly monochromatic limit. The KP and DS equations have several
other physical applications.
The fact that integrable nonlinear equations both in one and two space
dimensions appear in a wide range of physical applications is not an accident but a
consequence of the fact that these equations express a certain physical coherence
which is natural, at least asymptotically, to a variety of nonlinear phenomena.
Indeed Calogero and Eckhaus have shown that large classes of nonlinear evolution
PDEs, characterized by a dispersive linear part and a largely arbitrary nonlinear
part, after rescaling yield asymptotically equations (for the amplitude modulation)
having a universal character [4]. These ‘universal’ equations are, therefore, likely
to appear in many physical applications. Many integrable equations are precisely
these ‘universal’ models.
A method for solving the Cauchy problem for decaying initial data for
integrable evolution equations in two spatial dimensions emerged in the early
1980s. This method is sometimes referred to as the ∂¯ (dbar) method. Recall that
the inverse spectral method for solving nonlinear evolution equations on the line
is based on a matrix RH problem. This problem expresses the fact that there exist
solutions of the associated x-part of the Lax pair which are sectionally analytic.
Analyticity survives in some multi-dimensional problems: it was shown formally
by Manakov and by Fokas and Ablowitz [5] that KPI gives rise to a non-local
RH problem. However, for other multi-dimensional problems, such as the KPII,
the underlying eigenfunctions are nowhere analytic and the RH problem must be
replaced by the ∂¯ (dbar) problem. Actually, a ∂¯ problem had already appeared
in the work of Beals and Coifman [6] where the RH problem appearing in the
analysis of one-dimensional systems was considered as a special case of a ∂¯
problem. Soon thereafter, it was shown in [7] that KPII required the essential
use of the ∂¯ problem. The situation for the DS equations is analogous to that of
the KP equations.

4.1.2 Coherent structures

There exist two types of localized coherent structures associated with integrable
evolution equations in two spatial variables: the lumps and the dromions.
These solutions play a role similar to the role of solitons, namely they also
characterize the long time behavior of integrable evolution equations in two space
dimensions [5, 8].

Copyright © 2003 IOP Publishing Ltd.

 We now give some examples of lumps and dromions.

4.1.2.1 Lumps
The KPI equation is
∂x [qt + 6qqx + qxxx ] = 3qyy . (4.1)
The 1-lump solution of this equation is given by
 
1
q(x, y, t) = 2∂x ln |L(x, y, t)| + 2
2 2
L = x − 2λy + 12λ2 t + a
4λI
(4.2)
λ = λR + iλI λI > 0
where λ and a are complex constants. Several types of multi-lump solutions are
given in [9].
The focusing DSII equation is
iqt + qzz + qz̄z̄ − 2q(∂z̄−1|q|2z + ∂z−1 |q|2z̄ ) = 0 (4.3)

where z = x + iy, and the operator ∂z̄−1 is defined by

1 f (ζ, ζ̄ )
(∂z̄−1 f )(z, z̄) = dζ ∧ dζ̄ . (4.4)
2iπ R2 ζ − z
The 1-lump solution of this equation is given by

β ei(p +p̄ )t +pz−p̄z̄

2 2

q(z, z̄, t) = (4.5)

|z + α + 2ipt|2 + |β|2
where α, β, p are complex constants. A typical 1-lump solution is depicted in
figure 4.1.

4.1.2.2 Dromions
The DSI equation is
iqt + (∂x2 + ∂y2 )q + qu = 0
(4.6)
uxy = 2(∂x2 + ∂y2 )|q|2.
The 1-dromion solution of this equation is given by

ρ eX−Ȳ
q(x, y, t) =
α eX+X̄ + β e−Y −Ȳ + γ eX+X̄−Y −Ȳ + δ (4.7)
X = px + ip t 2
Y = qy + iq t 2
|ρ| = 4(Re p)(Re q)(αβ − γ δ)
2

where p, q are complex constants and α, β, γ , δ are positive constants. Several

types of multi-dromion solutions are given in [10].

Copyright © 2003 IOP Publishing Ltd.

 1

0.75
10

0.5

5
0.25

0
-10
0

-5

0 -5

-10
10

Figure 4.1. A typical 1-lump solution.

4.1.2.3 Fusion of a lump and a line-soliton

We conclude this section by noting that there exists another type of
generalized solitons in two dimensions, namely the so-called line-solitons. These
solutions can be constructed from the usual solitons by adding an appropriate
y-dependence. However, there exist certain lines in the x–y plane where these
solutions do not decay. A solution describing the fusion of 1-lump and 1-line-
soliton for the KPI equation is given by [11]
 
1 b θ(x,y,t ) c −θ(x,y,t )
q(x, y, t) = 2∂x ln |L(x, y, t)| + 2 + bc +
2 2
e + e
4λI 2λI 2λI
(4.8)
where L, λ are defined in (4.2), b, c are non-negative real constants and

θ (x, y, t) = 2λI [x − 2λR y − 4(λ2I − 3λ2R )t].

The fusion of a 1-lump and a 1-line soliton is depicted in figure 4.2.

4.1.3 Organization of this chapter

In section 4.2, we discuss the KPI equation. In section 4.3, we discuss both the
focusing and defocusing DSII equations.

Copyright © 2003 IOP Publishing Ltd.

 10

8
|U|2

0
10

5 10
5
0
0
Ŧ5
Ŧ5
Ŧ10 Ŧ10
Y
X

Figure 4.2. The interaction of a 1-lump and a 1-line soliton for KPI.

4.2 The KPI equation

The KPI equation (4.1) is the compatibility condition of the following Lax pair:

iψy + ψxx + q(x, y, t)ψ = 0 (4.9)

ψt + 4ψxxx + 6qψx + 3i(∂x−1 q)y ψ + 3qx ψ = 0. (4.10)

We first freeze t and consider equation (4.9).

Let
µ(x, y, k) = e−ikx+ik y ψ(x, y, k).
2
(4.11)
Then µ satisfies
iµy + µxx + 2ikµx + qµ = 0. (4.12)
Let µ+
and µ−
denote the particular solutions of equation (4.12) which are
defined as follows:
 ∞ ∞ ∞ y ∞ 0 
i
µ+ (x, y, k) = 1 + − dη dξ dm + dη dξ dm
2π y −∞ 0 −∞ −∞ −∞
× eim(x−ξ )−im(m+2k)(y−η)(qµ+ )(ξ, η, k). (4.13)+

Copyright © 2003 IOP Publishing Ltd.

 µ− is defined by a similar equation where the integrals with respect to dm are
interchanged. The kernel of equation (4.13)+ is analytic for Im k > 0. It is shown
in [12] (see also [13]) that if the L1 norm of the Fourier transform of q exists,
then equation (4.13)+ is a Fredholm integral equation, thus if µ+ exists, then µ+
is analytic in k for Im k > 0. This is indeed the case if the previous L1 norm is
sufficiently small. Similar considerations are valid for µ− for Im k < 0.
Equations (4.12) can be derived by noting that
∞ ∞
µ= dx
dy
G(x − x
, y − y
, k)(qµ)(x
, y
, k)
−∞ −∞

where G(x, y, k) satisfies

∞ ∞
1
iGy + Gxx + 2ikGx = δ(x)δ(y) = dp1 dp2 e−ip1 x−ip2 y .
(2π)2 −∞ −∞

Thus,
∞ ∞
1 e−ip1 x−ip2 y
G(x, y, k) = dp1 dp2 .
(2π)2 −∞ −∞ p2 − p1 (p1 − 2(kR + ikI ))
Equations (4.13) follow from the identity
∞
dp2 e−ip2 y
= −2πi e−i(a+ib)y (H (y) − H (b)) a, b ∈ R
−∞ p2 − (a + ib)

where H denotes the usual Heaviside function.

Defining ψ ± in terms of µ± by equations (4.11), it follows that ψ ± satisfy
 ∞ ∞ ∞
i
ψ + = eikx−ik y +
2
− dη dξ dm
2π y −∞ k
y ∞ k 
dm eim(x−ξ )−im (y−η) qψ +
2
+ dη dξ (4.14)+
−∞ −∞ −∞

and similarly for ψ − .

Let ψ L be defined by
 y ∞ ∞ 
2 i 2
ψ L = eikx−ik y + dη dξ dm eim(x−ξ )−im (y−η)qψ L .
2π −∞ −∞ −∞
(4.15)
We denote equations (4.14)± and (4.15) by

ψ ± = eikx−ik + G± ±
2y
k qψ
2y
(4.16)
ψ L = eikx−ik + GL qψ L .

We emphasize that GL is independent of k.

Copyright © 2003 IOP Publishing Ltd.

 Equations (4.16) imply that ψ ± are simply related to ψ L :
∞
ψ+ − ψL = dm T + (k, m)ψ L (x, y, m) (4.17a)
k
k
−
ψ −ψ = L
dm T − (k, m)ψ L (x, y, m) (4.17b)
−∞

where
∞ ∞
i
T ± (k, m) = − dη e−imξ +im η q(ξ, η)ψ ± (ξ, η, k).
2
dξ (4.18)±
2π −∞ −∞

Indeed,

ψ + − ψ L = G+ + + + +
k ψ − G ψ = (Gk − G )ψ + G (ψ − ψ )
L L L L L

or ∞
(ψ + − ψ L ) = dm eimx−im y T + (k, m) + GL (ψ + − ψ L ).
2

This equation and the definition of ψ L immediately imply (4.17a); similarly for
equation (4.17b).
Equations (4.17) together with the analyticity properties of ψ ± yield the
following linear integral equation for µL :
k
− −ikx+ik 2 y
dm eimx−im y T − (k, m)µL (x, y, m)
2
µ (x, y, k) = 1 + P e
L
−∞
∞
+ −ikx+ik 2 y
dm eimx−im y T + (k, m)µL (x, y, m)
2
−P e
k
(4.19)

where ∞
1 f (l) dl
P ± f (k) = k ∈ R. (4.20)
2iπ −∞ l − (k ± i0)
Indeed, multiplying equations (4.17) by e = exp(−ikx + ik 2 y) and using the fact
that P ∓ (µ± − 1) = 0, we find
∞
− −
P (eψ − 1) + P e
L
dm = 0
k
k
P + (eψ L −) + P + e dm = 0
−∞
∞ k
where k dm and −∞ dm denote the right-hand sides of equations (4.17a) and
(4.17b) respectively. Subtracting these equations and using

(P + − P − )(eψ L − 1) = eψ L − 1

Copyright © 2003 IOP Publishing Ltd.

 we find an equation for ψ L . Rewriting this equation for µL , which is defined in
terms of ψ L by equation (4.11), we find equation (4.19).
Equation (4.19) expresses µL in terms of T ± (k, m). Note that this equation
implies that
 
µL1 (x, y) 1
µ =1+
L
+O .
k k2
Substituting this expression into equation (4.12) we find

q = −2i∂x (µL
1 (x, y)). (4.21)

Since q depends on t, T ± also depend on t. It turns out that if q evolves

according to a KPI equation, then the time evolution of T ± is simple. This implies
the following scheme for integrating KPI.

Theorem 4.1 [13]. Let q0 (x, y) ∈ S(R2 ) satisfy

∞
dx q0 (x, y) = 0 (4.22)
−∞
∞
dy dξ (1 + ξ 2 )q̂0 (ξ, y)  1 (4.23)
−∞

where S denotes the space of Schwartz functions and q̂0 (ξ, y) denotes the Fourier
transform of q0 (x, y) in the x variable.
Given q0 (x, y), define ψ ± (x, y, k) by equations (4.14)± where q is replaced
by q0 . Given ψ ± (x, y, k), define T ± (k, m) by equation (4.18)± where q is
replaced by q0 . Given T ± , define µL (x, y, t, k) by equation (4.19) where T ±
are replaced by T ± (k, m) exp[i(k 3 + m3 )t]. Given µL (x, y, t, k) define q by

q(x, y, t) = −2i∂x lim k(µL (x, y, t, k) − 1).

k→∞

Then q(x, y, t) satisfies the KPI equation (4.1) with q(x, y, 0) = q0 (x, y).

Remark 4.1 (1) The KP equation without the zero mass assumption (4.22) is
studied in [14, 15].
(2) If the small norm assumption (4.23) is violated then equations (4.13)±
can have homogeneous solutions. These homogeneous solutions give rise to
lumps: lumps were formally incorporated into the inverse scattering scheme in [5].

4.3 The DSII equation

Before solving DSII (4.3), we introduce the main ideas of the ∂¯ method by solving
the linearized version of this equation. It is, of course, elementary to solve this
equation using a Fourier transform in x and y. Thus, the reason for solving this

Copyright © 2003 IOP Publishing Ltd.

 equation by the ∂¯ method is a pedagogical one. Indeed, the steps used for the
solution of the DSII equation are similar to the ones used here.
In order to solve this equation using a Lax pair formulation, we note it is
the compatibility condition of the following pair of linear equations for the scalar
function µ(z, z̄, t, kR , kI ) [16],

µz̄ − kµ = q (4.24)

µt = i(µzz + k 2 µ + kq + qz̄ ) k ∈ C. (4.25)

Indeed, these equations imply

iqt + qzz + qz̄z = i(µz̄t − µt z̄ ) = 0. (4.26)

We define a solution of (4.24) bounded for all k ∈ C. The Green function of

the operator ∂/∂ z̄ is 1/πz. Therefore, the Green function of the left-hand side of
equation (4.24) is (c/πz) ek z̄ where c is independent of z̄. In order for this Green
function to be bounded for all k ∈ C, we take c = e−k̄z . Hence, we define µ as the
following solution of equation (4.24),

1 ek(z̄−ζ̄ )−k̄(z−ζ )
µ(x, y, t, k) = dξ dη q(ξ, η, t) ζ
ξ + iη. (4.27)
π R2 z−ζ
The large z behaviour of µ involves α(kR , kI ) where

1
α(kR , kI )
dx dy e2i(kR y−kI x) q(x, y) k = kR + ikI . (4.28)
π R 2

Indeed,
lim (z ek̄z−k z̄ µ(x, y, k)) = α(kR , kI ). (4.29)
z→∞
Equation (4.27) implies
∂µ
= ek z̄−k̄z α (4.30a)
∂ k̄
as well as  
1
µ=O as k → ∞. (4.30b)
k
These equations define a ∂¯ problem for the function µ. The unique solution of
equations (4.30) is given by

1 e2i(lIx−lR y) α(lR , lI )
µ(x, y, k) = dlR dlI l = lR + ilI . (4.31)
π R2 k−l
Given α, equation (4.31) yields µ, which then implies q through equation (4.24):

1
q(x, y, t) = − dkR dkI e2i(kI x−kR y) α(kR , kI ). (4.32)
π R2

Copyright © 2003 IOP Publishing Ltd.

 Equations (4.28) and (4.32), are the usual formulae for the two-dimensional direct
and inverse Fourier transforms.
If q depends on t, α also depends on t. Equations (4.25) and (4.29) imply
αt = i(k̄ 2 + k 2 )α. (4.33)
In this way, one recovers the usual scheme for solving equation (4.26) through the
two-dimensional Fourier transform.
If q0 (z, z̄) ∈ L1 ∩ L∞ then equation (4.27) is well defined. If in addition
∂q0 /∂z, ∂q0 /∂ z̄ ∈ L1 ∩ L∞ then µ = O(1/k). Since q0 ∈ L1 , α(kR , kI , 0) ∈
L∞ . Also if the first three derivatives of q0 ∈ L1 , then α(kR kI , 0) ∈ L1 and the
∂¯ problem (4.30) is uniquely solvable.
Theorem [17]. Let q0 (z, z̄) ∈ S(R2 ). Assume that the L1 and L∞ norms of
q0 (z, z̄) and of its Fourier transform q̂0 (k, k̄) satisfy
π q̂0 ∞ q̂0 1 π
q0 ∞ q0 1 < < (4.34)
2 1 − τ2 2
where
1
τ=√ q̂0 1 q0 1 .
2π 3
Given q0 , define ν1 (z, z̄, k, k̄) by

1 dζ ∧ dζ̄
ν1 (z, z̄, k, k̄) = 1 − q0 (ζ, ζ̄ )ν2 (ζ, ζ̄ , k, k̄)
2iπ R2 ζ − z

1 dζ ∧ dζ̄
ν2 (z, z̄, k, k̄) = q̄0 (ζ, ζ̄ )ν1 (ζ, ζ̄ , k, k̄) e−ik(z−ζ )−ik̄(z̄−ζ̄ ) .
2iπ R 2 ζ̄ − z̄
(4.35)
Given ν1 , define T (k, k̄) by

1
T (k, k̄) = dz ∧ dz̄ q̄0 (z, z̄)ν1 (z, z̄, k, k̄) ei(kz+k̄z̄) . (4.36)
2π R2

Given T , define µ1 (z, z̄, t, k, k̄) by

∂µ1 ∂µ2
e = e−i(kz+k̄z̄)−i(k +k̄ )t .
2 2
= −T eµ̄2 = T eµ̄1 (4.37)
∂ k̄ ∂ k̄
Given µ1 and T , define q by

1
dk ∧ dk̄ T̄ (k, k̄)µ1 (z, z̄, t, k, k̄) ei(kz+k̄z̄)+i(k +k̄ )t .
2 2
q(z, z̄, t) =
2π R 2
(4.38)
Then q solves the defocusing DSII equation (4.3) with q(z, z̄, 0) = q0 (z, z̄).

Remark 4.2 If the assumption (4.34) is violated, then equations (4.35) can
have homogeneous solutions. These solutions were formally incorporated in the
inverse scattering scheme in [18] and [19].

Copyright © 2003 IOP Publishing Ltd.

 4.3.1 The defocusing DS equation
This equation is similar to equation (4.3) but there is a plus sign in front of 2q.
In this case the analogues of equations (4.35) and (4.37) can be solved without
a small norm assumption [20]. For example, the analogue of equations (4.37) is
now given by

∂µ1 ∂µ2
e = e−i(kz+k̄z̄)−i(k
2 +k̄ 2 )t
= T eµ̄2 = T eµ̄1 .
∂ k̄ ∂ k̄
These equations simply

∂(µ1 ± µ2 )
= T e(µ1 ± µ2 ). (4.39)
∂ k̄
Thus, the functions µ1 ± µ2 are generalized analytic functions [21]; therefore, the
solution of equations (4.39) exists without the need for small norm assumption
on T .

4.4 Summary
The dbar method is an effective tool for solving the Cauchy problem on the
plane for two-dimensional integrable nonlinear PDEs. However, up to now it has
not been possible to extend this method to evolution PDEs in higher than two
dimensions. In fact, even the question of the existence of integrable evolution
equations in three or higher dimensions remains open.

Acknowledgment
This work was partially supported by the EPSRC.

References
[1] Deift P 1999 Orthogonal Polynomials and Random Matrices. A Riemann–Hilbert
Approach (Courant Institute Lecture Notes) (New York: New York University)
[2] Deift P and Zhou X 1993 Ann. Math. 137 245
[3] Crighton D 1995 Applications of KdV in KdV’95 ed M Hazewinkel, H Capel and
E de Jager (Amsterdam: Kluwer) pp 2977–84
[4] Calogero F 1993 Important Developments in Soliton Theory ed A S Fokas and
V Zakharov (Berlin: Springer)
[5] Fokas A S and Ablowitz M J 1983 Stud. Appl. Math. 69 211–28
[6] Beals R and Coifman R 1985 Proc. Symp. Pure Math. 43 45
[7] Ablowitz M J, BarYaakov D and Fokas A S 1983 Stud. Appl. Math. 69 135–42
[8] Fokas A S and Santini P M 1990 Physica D 44 99–130
[9] Ablowitz M J and Villaroel J 1997 Phys. Rev. Lett. 78 570–3

Copyright © 2003 IOP Publishing Ltd.

 [10] J Hietarinta 2001 Scattering of solitons and dromions Scattering ed P Sabatier and
E Pike (New York: Academic Press)
[11] Fokas A S and Pogrebkov A K 2003 Nonlinearity 16 771–83
[12] Segur H 1982 AIP Conf. Proc. 88 211
[13] Zhou X 1990 Commun. Math. Phys. 128 551
[14] Sung L Y and Fokas A S 1999 Proc. Cambr. Phil. Soc. 125 113
[15] Boiti M, Pempinelli F and Pogrebkov A 1994 Inverse Problems 10 505
Boiti M, Pempinelli F and Pogrebkov A 1994 J. Math. Phys. 35 4683
[16] Fokas A S and Gel’fand I M 1994 Lett. Math. Phys. 32 189
[17] Fokas A S and Sung L Y 1992 Inverse Problems 673
[18] Fokas A S and Ablowitz M J 1984 J. Math. Phys. 25 2494
[19] Arkadiev A, Pogrebkov A K and Polivanov M C 1989 Physica D 36 1896
[20] Beals R and Coifman R 1989 Inverse Problems 5 87
[21] Vekua I N 1962 Generalized Analytic Functions (New York: Pergamon)

Copyright © 2003 IOP Publishing Ltd.

 Chapter 5

Introduction to solvable lattice models in

statistical and mathematical physics
Tetsuo Deguchi
Department of Physics, Ochanomizu University, Tokyo, Japan

5.1 Introduction
We introduce the six-vertex model defined on a two-dimensional square lattice.
We describe the model in detail, since it gives an important prototype of many
solvable lattice models defined on two-dimensional lattices [1]. The transfer
matrix of the six-vertex model generalizes the XXZ quantum spin chain which
plays a central role among integrable quantum spin chains [2,3]. The eight-vertex
model, which generalizes the six-vertex model directly, may be considered as
the most important exactly solvable model in statistical mechanics [4]. Moreover,
many mathematical theories such as the algebraic Bethe ansatz [7] and quantum
groups [5, 6] are closely related to the six-vertex model. Starting from the
six-vertex model, one may have a wide viewpoint on various physical and
mathematical topics related to solvable models. There are quite a large number
of topics related to exactly solvable models in physics and mathematics [1, 8–29].
We explain in section 5.2 some features of the six-vertex model defined on
a square lattice. We introduce the Boltzmann weights and the transfer matrix
for the six-vertex model. We review a method for diagonalizing the transfer
matrix, which is called the coordinate Bethe ansatz, and we give the expressions
of the free energy per site in the ferroelectric, the antiferroelectric and the
disordered phases, respectively [30, 31]. The disordered phase is gapless, while
the ferroelectric and the antiferroelectric phases have gaps [32, 33]. We derive a
critical singularity appearing at the phase transition from the antiferroelectric to
the disordered phase. We review the calculation of the singular part of the free
energy through analytic continuation, as shown in [1]. The critical singularity
is very weak and has an essential singularity similar to the Kosterlitz–Thouless
(KT) transition. We have, thus, derived the KT-like singularity through exact

Copyright © 2003 IOP Publishing Ltd.

 calculation. After reviewing the finite-size analysis of conformal invariance
[34–37], we discuss how the massless phase of the six-vertex model is related
to conformal field theory (CFT) with c = 1 which has U (1) symmetry. The
c = 1 CFT has a critical line where critical exponents change continuously
with respect to some parameter of the model [38–40]. There are quite a few
papers on the finite-size corrections of integrable models [41–45]. (For a review,
see [16,46,47].) The critical line is also characteristic of the Tomonaga–Luttinger
liquid [48]. The existence of a critical line was first discovered by RJ Baxter
through the exact solution of the eight-vertex model [49].
In section 5.3, we review various integrable models in statistical
mechanics [1]. We briefly introduce the Ising model [8, 50–52], the Potts
models [53–56] and the chiral Potts model [57–68] and then the eight-vertex
model [4, 49, 69, 70] and the IRF models [71–78]. In section 5.4, we solve
explicitly the Yang–Baxter equations for the six-vertex model. We introduce
the algebraic Bethe ansatz [7, 16, 79–81]. Here we show that the Yang–Baxter
equation of the algebraic Bethe ansatz can be expressed by graphs. In section 5.5,
we discuss some mathematical theories associated with integrable models such as
the braid group [82, 83] and the quantum groups [84].
There have been novel developments in the mathematical physics associated
with the six-vertex model [20, 84]. The integrable vertex models associated
with various Lie algebras which generalize the six-vertex model have been
obtained [85, 86]. The crystal basis of the quantum groups is derived from a
mathematical analysis of the corner transfer matrix which is fundamental for
calculating the one-point functions of the vertex and IRF models [87,88]. Through
the q-vertex operators, the correlation functions of the XXZ spin chain or the
six-vertex model are obtained [20]. The dynamical Yang–Baxter equation [89]
and the elliptic quantum groups [90–92] have also been extensively discussed. In
fact, we can derive the R-matrix of the eight-vertex model systematically from
the elliptic quantum group through the twists [92]. Furthermore, the correlation
functions of the XXZ model calculated with the q-vertex operators have been re-
derived for large but finite chains through the algebraic Bethe ansatz with Drinfeld
twists [93, 94]. Here we note that the q-vertex operator can be defined only on
the infinite chain, while the algebraic Bethe ansatz with the Drinfeld twists can
be applied to any finite chain. By taking the thermodynamic limit, it has been
shown that the two approaches indeed give the same results. These papers indeed
illustrate non-trivial physical applications of the Drinfeld twists. It has recently
been found that the symmetry of the six-vertex model is enhanced at some
particular coupling constants: the transfer matrix commutes with the generators
of the sl2 loop algebra for the six-vertex model at the roots of unity [95].
Let us discuss some physical motivations for the six-vertex model. The
exact solution of the six-vertex model was originally introduced for studying
the statistical mechanics of ferroelectrics such as the residual entropy and
the ferroelectric transitions [30, 31]. However, it seems that the physical
motivation of the six-vertex model for ferroelectricity has decreased. There are,

Copyright © 2003 IOP Publishing Ltd.

 however, many different physical applications of the six-vertex model. Here we
consider a few examples: domain wall theory [97], crystal growth [98–103]
and the thermodynamics of the XXZ spin chain through the quantum transfer
matrix [104–109]. The crystal growth on surfaces has been discussed by applying
exact solutions of the six-vertex model [98–102] and some extensions [103]. The
free energy of the six-vertex model gives the equilibrium crystal shapes [98, 99].
The finite-temperature thermodynamics of the XXZ spin chain has been studied
extensively through the quantum transfer matrix, which is a version of the
inhomogeneous six-vertex transfer matrix [105–109]. The quantum transfer
matrix is obtained by regarding one direction of the square lattice as the
imaginary time or inverse temperature [104]. There have been considerable efforts
to evaluate thermal quantities analytically or numerically. Several functional
equations on the eigenvalues of the transfer matrix have been devised [108, 109].
Finally, we remark that a universal relation between the dispersion curve and the
ground-state correlation length in quantum spin chains is discussed by using the
exact solutions of the vertex models [110].
In section 5.2, we employ mainly the notation of Baxter’s textbook [1] except
for the transfer matrix. In section 5.4, however, we briefly show how the notation
in statistical mechanics is related to the notation of the algebraic Bethe ansatz
or the quantum inverse scattering method. The graphical illustration should be
useful. Finally, in section 5.5, we discuss many connections of the six-vertex
model to several mathematical developments such as the quantum groups.

5.2 Solvable vertex models

5.2.1 The six-vertex model

5.2.1.1 Ice rule

Let us consider a square lattice as a model of a two-dimensional ferroelectric

crystal. Molecules are placed on the vertices of the lattice. Arrows are placed
on the edges of the lattice and these correspond to directions of dipole moments
of hydrogen bonds. As a crystal with hydrogen bonding, we may consider ice,
i.e. the crystal of water molecules. In this review, however, we simplify the
molecular background of the model (for instance, see [9]). We assume that the
dipole moments defined on edges take only two values: ±1.
At a vertex in the lattice, there are four edges. There are 16 possible
configurations of the four edges around the vertex since each of the edges takes
two values: ±1.
Let symbols α, β, γ , δ denote the values of the dipole moments around the
vertex. Due to charge neutrality, they should satisfy the following condition:

α + β = γ + δ. (5.1)

Copyright © 2003 IOP Publishing Ltd.

 6
γ

α δ
-

Figure 5.1. Configuration of the polarizations around a vertex: α, β, γ , δ. The Boltzmann

weight is expressed by w(α, β|γ , δ). The positive directions of dipole moments are given
by the upward or rightward arrows.

(1) (3) (5)

6 ? 6
- - - - - 

6 ? ?

(2) (4) (6)

? 6 ?
     -
? 6 6

Figure 5.2. Vertex configurations satisfying the ice rule. They have the Boltzmann
weights w(α, β|γ , δ) as follows: (1) w(1, 1|1, 1); (2) w(2, 2|2, 2); (3) w(1, 2|2, 1); (4)
w(2, 1|1, 2); (5) w(1, 2|1, 2); (6) w(2, 1|2, 1). Configurations (1) and (2) are for the
weight a, (3) and (4) for b and (5) and (6) for c.

For an illustration, let us consider the case when α, β, γ and δ are given by +1.
Then, α and β give +2 to the vertex, while γ and δ remove +2 from it, so that the
net charge around the vertex is kept neutral: α + β − γ − δ = 0.
There are only six configurations satisfying the condition. The other
configurations that do not satisfy the condition are not allowed in thermal
equilibrium. We call this the six-vertex model. Condition (5.1) is sometimes
called the ice rule, since ice as a crystal consists of water molecules connected
by hydrogen bonding.
We denote by 1 and 2 the values of polarization 1 and −1, respectively. The
symbols 1 and 2 are useful for matrix notation. Let p denote the notation of ±1
and k 1 and 2. Then, they are related by the relation: k = 1 + (1 − p)/2.

Copyright © 2003 IOP Publishing Ltd.

 5.2.1.2 Boltzmann weights
It is a key idea in exactly solvable models that we define the model by the
Boltzmann weights not by the energies of configurations. Let us introduce the
Boltzmann weights for configurations around a vertex. For a vertex configuration
α, β, γ , δ, we denote the energy at the vertex by (α, β|γ , δ). Then, the
Boltzmann weight for a temperature T is given by

w(α, β|γ , δ) = exp(−(α, β|γ , δ)/kB T ). (5.2)

Under the ice rule, there are only six configurations allowed round a vertex. Here,
it is assumed that the energy of a configuration violating the ice rule should be
infinite. We denote by j the energy of the j th vertex configuration shown in
figure 5.2.
Under no external field, the Boltzmann weights must be invariant when
reversing all the polarizations simultaneously. Thus, we have 1 = 2 , 3 = 4
and 5 = 6 , when there is no external field. We denote the Boltzmann weights
as follows:
w(1, 1|1, 1) = w(2, 2|2, 2) = w1 = a
w(1, 2|2, 1) = w(2, 1|1, 2) = w2 = b (5.3)
w(1, 2|1, 2) = w(2, 1|2, 1) = w3 = c.
The Boltzmann weights of the zero-field six-vertex model have essentially
only two parameters. For instance, we may choose a/c and b/c. Note that the
probability for the vertex configuration of a is given by a/(a + b + c), which
does not change by replacing a, b and c with ρa, ρb and ρc.
Let us consider π/2 rotation of the square lattice. If we rotate vertex
configuration (1) of figure 5.2 by the angle π/2 in the counterclockwise direction,
then it becomes vertex configuration (4). Under the π/2 rotation, the weight a is
exchanged with the weight b, while the weight c does not change.

5.2.2 The partition function and the transfer matrix

Let us discuss the partition function of the system. We now set the boundary
conditions. Here, we consider the periodic boundary conditions for the two-
dimensional lattice. We take a product of the Boltzmann weights over all the
vertices of the lattice and sum the product over all allowed configurations of the
arrows on the lattice:
 
Z= w(aj , bj |cj , dj ). (5.4)
config j: vertices

The partition function of the square lattice can be formulated as the trace of
the products of the transfer matrices. Let us define the transfer matrix τ of the
six-vertex model. The matrix elements of the transfer matrix τ acting on N lattice

Copyright © 2003 IOP Publishing Ltd.

 a1 a2 aN
···
c1 c2 cN c1
···

b1 b2 bN

a ,...,a
Figure 5.3. Matrix elements of the transfer matrix τb 1,...,b N .
1 N

sites are given by

a ,...,a

τb11,...,bNN = w(c1 , b1 |a1 , c2 )w(c2 , b2 |a2 , c3 ) · · · w(cN , bN |aN , c1 ).
c1 ,...,cN
(5.5)
Under periodic boundary conditions, the partition function ZNN
of N × N

lattice is given by the trace of the N

th power of the transfer matrices:


N
N
N

ZNN
= Tr(τ N ) = ,...,aN = 1 + 2 + · · · + 2N .
(τ N )aa11 ,...,a N (5.6)
a1 ,...,aN

Here j denotes the eigenvalue of the transfer matrix τ .

The free energy per site f is given by

f = −kB T log ZNN

/(NN
). (5.7)

In the thermodynamic limit N, N

→ ∞, the free energy per site is given by the
largest eigenvalue max of the transfer matrix τ .

5.2.3 Diagonalization of the transfer matrix

5.2.3.1 The Yang–Baxter relations for six-vertex model
Let us consider three sets of Boltzmann weights: (w1 , w2 , w3 ) = (a, b, c),
(a
, b
, c
) and (a

, b

, c

). We denote by τ
and τ

the transfer matrices

constructed from the sets of Boltzmann weights (a
, b
, c
) and (a

, b

, c

),
respectively. If the three sets of Boltzmann weights satisfy the Yang–Baxter
equation

w(α, γ |a1 , a2 )w
(β, b3 |γ , a3 )w

(b1 , b2 |α, β)
α,β,γ

= w

(β, α|a2 , a3 )w
(b1 , γ |a1 , β)w(b2 , b3 |γ , α) (5.8)
α,β,γ

then the transfer matrices τ

and τ

commute. The derivation of the commutation

relation is given in the appendix. We note that a graphical presentation of the
Yang–Baxter equation (5.8) will be shown in figure 5.6.

Copyright © 2003 IOP Publishing Ltd.

 Let us define the parameter as follows:

a 2 + b2 − c2
= . (5.9)
2ab
For the zero-field six-vertex model, we can show that if the two sets of Boltzmann
weights have the same value for the parameter , then their transfer matrices
commute. We shall explicitly discuss in section 5.4 that it is indeed derived from
the Yang–Baxter equations (5.8).

5.2.3.2 The coordinate Bethe ansatz

a ,...,a
Let us consider the matrix element τb11,...,bNN of the transfer matrix τ . Due to the
ice rule, we may express the suffix a1 , . . . , aN by the positions of the value
2, as follows. Suppose that there are n suffices given by the value 2 among
the N suffices a1 , . . . , aN . The n suffices are expressed as ax1 , ax2 , . . . , axn
where the xj s are in increasing order: x1 < x2 < · · · < xn . Then, the entry
a1 , . . . , aN is equivalent to the set of xj s: x1 , . . . , xn . For an illustration, let
us consider the case N = 5 and n = 3. Then, (x1 , x2 , x3 ) = (1, 3, 4) corresponds
to (a1 , a2 , a3 , a4 , a5 ) = (2, 1, 2, 2, 1). Thus, the matrix element τba11,...,b ,...,aN
N
can be
denoted briefly by τyx11,...,y
,...,xn
n .
Let us now discuss how to solve the secular equation τg = g. Here, the
transfer matrix τ is a 2N × 2N matrix, g is a 2N -dimensional eigenvector with
eigenvalue . In terms of matrix elements, the secular equation can be written as

τyx11,...,y
,...,xn
n
g(y1 , . . . , yn ) = g(x1 , . . . , xn ). (5.10)
y1 ,...,yn

Here, g(x1 , . . . , xn ) denotes the matrix element of vector g for the entry
(x1 , . . . , xn ). In the coordinate Bethe ansatz, we assume the following form for
the matrix element of the possible eigenvector g:

g(x1 , . . . , xn ) = AP exp(kP 1 x1 + · · · + kP n xn ). (5.11)
P ∈Sn

Here, Sn denotes the symmetric group of order n and P is a permutation of n

letters, 1, 2, . . . , n, where P maps j into Pj . The expression (5.11) is called the
Bethe ansatz wavefunction. If the vector g whose elements are of the form (5.11)
is an eigenvector of the transfer matrix, then we call it a Bethe ansatz eigenvector.
For general n, the vector (5.11) becomes an eigenvector of the transfer
matrix, if the wavenumbers kj satisfy the Bethe ansatz equations. They are given
by the following expression:


n
exp(iNkj ) = (−1)n−1 exp(−i(kj , k )) for j = 1, . . . , n (5.12)
=1

Copyright © 2003 IOP Publishing Ltd.

 where (p, q) is defined by

1 − 2 eip + ei(p+q)
exp(−i(p, q)) = . (5.13)
1 − 2 eiq + ei(p+q)
For the solutions kj to the Bethe ansatz equations, the eigenvalue  of the transfer
matrix is given by

(k1 , . . . , kn ) = a N L(z1 ) · · · L(zn ) + bN M(z1 ) · · · M(zn ) (5.14)

where zj = exp(ikj ) for j = 1, . . . , n and the functions L(z) and M(z) are
defined by

ab + (c2 − b 2 )z a 2 − c2 − abz
L(z) = M(z) = . (5.15)
a(a − bz) b(a − bz)
When we discuss the spectrum of an integrable model through the coordinate
Bethe ansatz, we often assume that all the eigenvectors of the transfer matrix
are characterized by the Bethe ansatz wavefunction (5.11). However, it is not
certain whether the assumption is valid or not. Thus, we have to check it by
other methods. In fact, there are several numerical studies on the validity of the
completeness of the Bethe ansatz for some integrable models.
However, there is no doubt about the mathematical structure of the Bethe
ansatz wavefunction. We can derive the expression (5.11) by the algebraic Bethe
ansatz through the ‘two-site’ model [16,79]. (There is an instructive note in [111].)
It was shown that the matrix elements of the product of the B operators acting on
the vacuum are given by the Bethe ansatz wavefunction (5.11) with the kj being
generic.
The Yang–Baxter relation leads not only to the integrability of the six-vertex
model but also to the systematic construction of the eigenvectors. In fact, we shall
see in section 5.4 that the algebraic Bethe ansatz is solely based on the Yang–
Baxter equation.

5.2.3.3 An example of the eigenvector

For an illustration, let us consider the eigenvector g for the case of n = 1:
g(x) = A exp(ikx). Through a direct calculation, we have


N 
x−1 
N
τyx g(y) = a x−y−1bN−x+y−1 c2 g(y) + a N+x−y−1 by−x−1c2 g(y)
y=1 y=1 y=x+1

a N−1 bN−x c2 z
= (a N L(z) + bN M(z))g(x) + (1 − zN ). (5.16)
a − bz

Therefore, g(x) becomes an eigenvector if the Bethe ansatz equation zN = 1 is

satisfied.

Copyright © 2003 IOP Publishing Ltd.

 b/c

II

III
1
@
@
@
IV @
I
@
@ a/c
0 1

Figure 5.4. Phase diagram of the six-vertex model: regimes I and II are ferroelectric,
regime III is disordered and regime IV is antiferroelectric.

5.2.4 The free energy of the six-vertex model

5.2.4.1 Three phases of the six-vertex model
There are three phases for the zero-field six-vertex model. They are given by the
regions of the parameter : a ferroelectric phase when > 1; an antiferroelectric
phase when < −1; and a disordered phase when −1 < < 1. It is found that
the disordered phase (−1 < < 1) is gapless (massless), while the ferroelectric
( > 1) and antiferroelectric phases ( < −1) are gapful (massive).
Let us consider the phase diagram of the six-vertex model shown in
figure 5.4. We recall that the ratios a/c and b/c determine the model. Here we
note that the number of independent parameters is given by two, since the overall
normalization factor is arbitrary. Regimes I and II give the ferroelectric phase.
Regimes III and IV are antiferroelectric and disordered, respectively. In terms of
the Boltzmann weights, regime I is given by a > b + c, regime II by b > a + c
and regime IV by a + b < c. Then, regime III is given by a < b + c, b < c + a
and c < a + b.
Let us derive the three phases through an intuitive argument. For the
ferroelectric regime of > 1, we have a 2 + b 2 − c2 > 2ab, which leads to the
inequality |a − b| > c. When a > b, we have a > b + c. Thus, the configuration
where all the Boltzmann weights are given by the weight a should be the largest
contribution to the partition function Z. In fact, when n = 0 in equation (5.14),
we have  = a N + b N . When a > b, the free energy per site, f , is given by 1 :
f = −kB T log a. When b > a, we have f = 3 , similarly.
For the antiferroelectric regime of < −1, we have the inequality a + b <
c. Thus, the vertex configurations for c should be more favourable than those of
a or b. In fact, it is shown that the transfer matrix has the largest eigenvalue when
n = N/2. Furthermore, if we send | | to infinity, all the vertex configurations
should be given by those of c. The phase is, thus, called antiferroelectric.

Copyright © 2003 IOP Publishing Ltd.

 We may explain the reason why it is called antiferroelectric. Let us consider
configuration (5) in figure 5.2. We see that the arrow coming from the left goes
upward, while the arrow coming from the right goes downward. If all the vertex
configurations on the square lattice are given by (5), then the lines coming from
South West to North East and the lines coming from North East to South West
occupy the lattice alternatively. This gives the antiferroelectric order.

5.2.4.2 Parametrization of the Boltzmann weights

It is not trivial to parametrize the Boltzmann weights. Recall that there are two
independent parameters for the six-vertex model. Thus, if we consider as a
parameter, there is only one other one. As we shall see later, it is related to the
spectral parameter.
We recall that the phases of the zero-field six-vertex model are classified as
follows: < −1, −1 < < 1 and 1 < .
(1) Antiferroelectric phase. For < −1, we define a real parameter λ by

= − cosh λ (0 < λ) (5.17)

and we parametrize the Boltzmann weights as

   
λ−v λ+v
a = ρ sinh b = ρ sinh
2 2 (5.18)
c = ρ sinh λ (−λ < v < λ)

where ρ is the normalization factor. Let us define the rapidity α for the
wavenumber k as
eλ − e−iα sin 12 (α − iλ)
exp(ik) = = − . (5.19)
eλ−iα − 1 sin 12 (α + iλ)

By replacing the wavenumbers p and q with the rapidities α and β in

equation (5.13), the phase factor (p, q) can be written as

e2λ − e−i(α−β) sin 12 ((α − β) − 2iλ)

exp(−i(p, q)) = = − . (5.20)
e2λ−i(α−β) − 1 sin 12 ((α − β) + 2iλ)

(2) Disordered phase. For −1 < < 1, we define a positive real parameter µ
by
= − cos µ (0 < µ < π) (5.21)
and we parameterize the Boltzmann weights as follows:
   
µ−w µ+w
a = ρ sin b = ρ sin c = ρ sin µ (−µ < w < µ).
2 2
(5.22)

Copyright © 2003 IOP Publishing Ltd.

 Here ρ is the normalization factor. We define the rapidity α for the
wavenumber k by

eiµ − eα sinh 12 (α − iµ)

exp(ik) = = − . (5.23)
eiµ+α − 1 sinh 12 (α + iµ)

In terms of rapidities α and β, the phase factor (p, q) is expressed as

e2iµ − eα−β sinh 12 ((α − β) − 2iµ)

exp(−i(p, q)) = = − . (5.24)
e2iµ+α−β − 1 sinh 12 ((α − β) + 2iµ)

(3) Ferroelectric phase. For > 1, we define the real parameter λ by

= cosh λ (0 < λ) (5.25)

and we may parameterize the Boltzmann weights as follows. When a > b

(regime I),
   
λ−v λ+v
a = ρ sinh b = −ρ sinh c = ρ sinh(λ) (v < −λ).
2 2
(5.26)
When a < b (regime II),
   
λ−v λ+v
a = −ρ sinh b = ρ sinh c = ρ sinh(λ) (v > λ).
2 2
(5.27)
Here we recall that ρ is the normalization factor. The wavenumber k is
related to the rapidity α by

eλ − e−iα sin 12 (α − iλ)

exp(ik) = − = . (5.28)
eλ−iα − 1 sin 12 (α + iλ)

The phase factor (p, q) can be written as

e2λ − e−i(α−β) sin 12 ((α − β) − 2iλ)

exp(−i(p, q)) = = − . (5.29)
e2λ−i(α−β) − 1 sin 12 ((α − β) + 2iλ)

Let us consider curved lines given by changing the spectral parameter w

or v continuously. All the curves pass through the points (a/c, b/c) =
(1, 0) and (0, 1). Except for the two points, however, any points on the
horizontal axis b/c = 0 or the vertical axis a/c = 0 are never reached by
these parametrizations with finite values.

Copyright © 2003 IOP Publishing Ltd.

 5.2.4.3 Expressions for the free energy
(1) Antiferroelectric phase. When < −1, the system is in the antiferroelectric
phase. The free energy per site f is given by
 
λ+v  ∞
e−mλ sinh m(λ + v)
f = −kB T log a − kB T + (5.30)
2 m=1
m cosh mλ
 
λ−v  ∞
e−mλ sinh m(λ − v)
= −kB T log b − kB T + (5.31)
2 m=1
m cosh mλ

for −λ < v < λ.

(2) Disordered phase. When −1 < < 1, we have
∞
sinh(µ + w)x sinh(π − µ)x
f = −kB T log a − kB T dx (5.32)
−∞ 2x sinh πx cosh µx
∞
sinh(µ − w)x sinh(π − µ)x
= −kB T log b − kB T dx (5.33)
−∞ 2x sinh πx cosh µx

for −µ < w < µ.

(3) Ferroelectric phase. When a > b (regime I), we have f = −kB T log a,
while when b > a (regime II), we have f = −kB T log b.

5.2.4.4 Correlation length

The correlation length has been calculated for the six-vertex model [1,32]. In fact,
it is calculated for the eight-vertex model. In the ferroelectric and antiferroelectric
regimes of the six-vertex model, the correlation length is finite. It becomes very
large near their phase boundaries with the disordered phase. In the disordered
phase of the six-vertex model, the correlation length diverges throughout the
regime. Thus, the disordered phase is critical.

5.2.5 Critical singularity in the antiferroelectric regime near the phase

boundary
We now discuss the singular behaviour of the free energy at the phase transition
from the antiferroelectric to the disordered phase [1]. Here we note that the
former one has a gap, while the latter is gapless. Recall that the phase boundary
between regimes III and IV is given by a + b = c. When T > Tc , the system
should be in the disordered regime (−1 < < 1), while when T < Tc it is in the
antiferroelectric regime ( < −1). At the lower temperature, the system should
be ordered.
Let us calculate the analytic continuation of the high-temperature free energy
(5.32) into the low-temperature phase, so that we can single out the singularity
of the free energy near Tc and for T < Tc . First, we reformulate the integral of

Copyright © 2003 IOP Publishing Ltd.

 equation (5.32) as follows:
∞ ∞
sinh(µ + w)x sinh(π − µ)x sinh(µ + w)x exp(π − µ)x
dx = P dx.
−∞ 2x sinh πx cosh µx −∞ 2x sinh πx cosh µx
(5.34)
Here P denotes the principal value integral. Second, we set λ to be a very small
positive real number, and take a value v satisfying −λ < v < λ. We consider the
path in the complex µ-plane:
µ = λ exp(−iθ ) for 0 ≤ θ ≤ π/2. (5.35)
Along the path, we calculate the analytic continuation of the high-temperature
free energy (5.32). Here, we also consider the path of w: w = v exp(−iθ ) for
0 ≤ θ ≤ π/2. Then, we have
∞
−i sin(λ + v)x exp(π + iλ)x
P dx
−∞ 2x sinh πx cos λx
λ+v  ∞
e−mλ sinh m(λ + v)
= +
2 m=1
m cosh mλ
∞
(−1)m cos((m − 1/2)πv/λ) e−(m−1/2)π /λ
2

−i . (5.36)
m=1
(m − 1/2) sinh((m − 1/2)π 2 /λ)
The real part of the analytic continuation corresponds to the expression for the
antiferroelectric free energy. Therefore, we obtain the singular part of the free
energy
∞
(−1)m cos((m − 1/2)πv/λ) exp(−(m − 1/2)π 2 /λ)
fsing = ikB T . (5.37)
m=1
(m − 1/2) sinh((m − 1/2)π 2 /λ)
We define the reduced temperature t by
t = (a + b − c)/c. (5.38)
In the low-temperature phase (T < Tc ) and near Tc , t is given by
t ≈ − 18 (λ2 − v 2 ). (5.39)
Approximately, t is given by t ≈ −λ2 /8. Near Tc , we have
 πv 
fsing ≈ −4ikB T e−π /λ cos
2
. (5.40)
2λ
Thus, we have  
constant
fsing ∝ exp − √ . (5.41)
−t
Near Tc , the free energy has an essential singularity.
The singularity of the free energy is very close to that of the Kosterlitz–
Thouless transition. In fact, calculating
 exactly,
√ we  can show that the correlation
length ξ diverges at Tc as ξ ∝ exp constant/ −t , when T approaches Tc in the
antiferroelectric phase.

Copyright © 2003 IOP Publishing Ltd.

 5.2.6 XXZ spin chain and the transfer matrix
The logarithmic derivative of the transfer matrix of the six-vertex model gives the
Hamiltonian of the XXZ spin chain
d d
log τ |v=−λ = τ −1 τ ∝ HXXZ + constant (5.42)
dv dv
where HXXZ is given by

L
HXXZ = J (σjX σjX+1 + σjY σjY+1 + σjZ σjZ+1 ). (5.43)
j =1

Intuitively, we may express it by τ6V (v) ≈ exp(−vHXXZ ).

The XXZ spin chain and the six-vertex transfer matrix have the same
eigenvectors in common thanks to equation (5.42). Taking the logarithm of the
Bethe ansatz equations, we have

M
Nkj = 2πIj − (kj , k ) for j = 1, . . . , M (5.44)
=1

where M is the number of down spins. (We assume 2M ≤ N.) Here Ij is an

integer if M is odd and half an integer if M is even.
The ground state of the XXZ spin chain for < 1 was obtained by Yang and
Yang [3]. The ground state is specified by the integers Ij = j − (M + 1)/2 for
j = 1, . . . , M. When N is very large, the distribution of kj s becomes continuous.
The number of kj s between k and k + dk can be approximated by Nρ(k) dk.
Thus, we have the integral equation of ρ(k):
Q
∂(k, k
)
2πρ(k) = 1 + ρ(k
) dk
(5.45)
−Q ∂k
where Q is determined by the normalization condition
Q
ρ(k) dk = M/N. (5.46)
−Q

The integral equation (5.45) can be solved by changing the variable k to rapidity
α and then by taking the Fourier transform for the half-filling case M/N = 1/2.
When M/N is close to 1/2, the integral equation can be solved by the Wiener–
Hopf method [3].

5.2.7 Low-lying excited spectrum of the transfer matrix and conformal field
theory
In this section, we assume that the low-lying excited spectra of the transfer matrix
of gapless models should be characterized by conformal invariance, if the system

Copyright © 2003 IOP Publishing Ltd.

 size is large enough. The assumption is not rigorous: however, there are many
studies which confirm it numerically for integrable models. We review the finite-
size corrections for the XXZ spin chain and the six-vertex model.

5.2.7.1 Finite-size corrections

Let us consider a conformally invariant field theory defined in the two-
dimensional Euclidean space with coordinates r1 and r2 . The energy–momentum
tensor Tµν for µ, ν = 1, 2 should be symmetric and traceless due to the conformal
symmetry. Introducing the complex coordinates, z = r1 + ir2 , z̄ = r1 − ir2 , we
define the chiral operator T = (T11 − T22 − 2iT12 )/4, and the antichiral operator
T̄ = (T11 − T22 + 2iT12 )/4. The operator T (or T̄ ) depends only on the variable
z (or z̄).
The energy–momentum tensor has the operator product expansion

c/2 2T (z2 ) ∂T (z2 )

T (z1 )T (z2 ) = + + +··· . (5.47)
(z1 − z2 )4 (z1 − z2 )2 z1 − z2

Here c iscalled the central charge. We define the operators Ln by the expansion
T (z) = ∞ n=−∞ Ln z
−n−2 . The operator product expansion (5.47) corresponds to

the Virasoro algebra

c
[Ln , Lm ] = (m − n)Lm+n + (m3 − m)δm+n,0 . (5.48)
12
Under a conformal transformation z → w, the energy–momentum tensor is
transformed as
 
dw 2 c
T (z) = T̃ (w) + {w, z} (5.49)
dz 12
where the symbol {w, z} denotes the Schwarzian derivative: (d3 w/dz3 )/(dw/dz)
−(3/2)(d2w/dz2 )2 /(dw/dz)2 .
Let us consider the conformal mapping from the z-plane to a cylinder of
circumference L:
L
z→w= log z. (5.50)
2π
Here w = τ − ix with imaginary time τ = it. The Hamiltonian Ĥ on the cylinder
is given by the space integral of the (1,1) component of the energy–momentum
tensor (Tcyl )µν
L
1
Ĥ = dx (Tcyl (w) + T̄cyl (w))
2π 0
2π πc
= (L0 + L̄0 ) − . (5.51)
L 6L

Copyright © 2003 IOP Publishing Ltd.

 Here we have used (5.49). For the momentum operator on the cylinder, we have
L
1
P̂ = dx (Tcyl (w) − T̄cyl (w))
2π 0
2π
= (L0 − L̄0 ). (5.52)
L
Let us now discuss the application of the formulas (5.51) and (5.52) to the
quantum spin chains. We assume that the low-lying excited energies should be
gapless and conformally invariant. In other words, we assume that the excitations
near the ground state have a linear dispersion relation. Let v denote the velocity
of the linear dispersion. Then, for the ground-state energy E0 , we have
πvc
E0 = Le∞ − (5.53)
6L
and for the excited energy Eex and the momentum Pex , we have

2πv
Eex − E0 = (h + h̄ + N + N̄)
L (5.54)
2π
Pex − P0 = (h − h̄ + N − N̄ )
L
where h and h̄ are the conformal weights related to the zero modes of the field
and the eigenvalues of N and N̄ are given by non-negative integers.
There is another viewpoint on finite-size scaling. Let us consider the t-axis
as the space axis for an infinitely long quantum spin chain, and the x-axis as the
imaginary time axis. Here we assume that L = vβ = v/T . Thus, our system now
becomes the quantum spin chain in the finite temperature T . Replacing E0 with
vβf , where f denotes the finite-temperature free energy of the spin chain, we
have f = −πcT 2 /6v. Thus, we may calculate the specific heat C by the formula
C = −T ∂ 2 f/∂T 2 and we have
πc
C= T. (5.55)
3v

5.2.7.2 The free boson: CFT with c = 1

Let us consider a free Bose field ϕ(x, t) defined on a cylinder of circumference
L. The Lagrangian is given by

1
L = g dx {(∂t ϕ)2 − (∂x ϕ)2 } (5.56)
2

We define the mode ϕn by the Fourier expansion ϕ(x, t) = ϕn (t)×
exp(−2πinx/L). From the canonical quantization, we have the conjugate
momentum πn = gLϕ̇−n and the commutation relation [ϕn , πm ] = iδnm . With the

Copyright © 2003 IOP Publishing Ltd.

 operators an and ān for n = 0 satisfying

[an , am ] = nδn+m [an , ām ] = 0 [ān , ām ] = nδn+m (5.57)

√
the Fourier mode is expressed as ϕn = i(an − ā−n )/(n 4πg), for n = 0. The
Hamiltonian is given by

1 2 2π 
H= π + (a−n an + ā−n ān ). (5.58)
2gL 0 2L n=0

Hereafter, we assume g = 1/4π. The convention is consistent with the

conformally invariant partition functions.
We now discuss the compactification of the boson with radius R. Suppose
that the field operator ϕ takes its value only on the circle of radius R. In
other words, we may identify ϕ with ϕ + 2πR. Then, the eigenvalue of the
momentum π0 conjugate to ϕ0 is given by n/R for an integer n. Here we recall
that the wavenumber of a one-dimensional system of size L is given by 2πn/L
(n ∈ Z), and also that the range of ϕ0 is given by 2πR, which corresponds to
L. Furthermore, we may assign on the operator ϕ the boundary condition for an
integer m:
ϕ(x + L, t) = ϕ(x, t) + 2πmR. (5.59)

Here, the integer m is called the winding number. The mode expansion of ϕ is
given by

4π 2πRm 1
ϕ(x, t) = ϕ0 + π0 t + x+i (an e2πin(x−t )/L − ā−n e2πin(x+t )/L).
L L n=0
n
(5.60)
In terms of the coordinates z = exp(2π(τ − ix)/L) and z̄ = exp(2π(τ + ix)/L),
ϕ(x, t) is given by the sum of the holomorphic and antiholomorphic parts:
ϕ(z, z̄) = φ(z) + φ̄(z̄) . Here, they are given by

ϕ0 1
φ(z) = − ia0 log(z) + i ak z−k
2 k=0
k
1 (5.61)
ϕ0
φ̄(z̄) = − iā0 log(z̄) + i āk z̄−k
2 k=0
k

with a0 = n/R + mR/2 and ā0 = n/R − mR/2. Here we can show that the
operator J (z) = i∂φ(z)/∂z is the U (1) current operator.
Making use of Noether’s theorem, we have
 2
1 ∂φ(z) 1
T (z) = − : : T̄ (z̄) = − : (∂¯ φ̄(z̄))2 : . (5.62)
2 ∂z 2

Copyright © 2003 IOP Publishing Ltd.

 Here : : denotes a proper normal ordering. Then, through the Laurent expansion
of powers of z, we have

1 ∞
L0 = a02 + a−n an . (5.63)
2 n=1

Thus, the conformal weights hnm and h̄nm are given by

 2  2
1 n 1 1 n 1
hn,m = + mR h̄n,m = − mR . (5.64)
2 R 2 2 R 2

5.2.7.3 The XXZ spin chain and CFT with c = 1

We discuss the finite-size corrections to the XXZ spin chain. The finite-size
corrections to the ground-state energy are calculated in [42–45] using the Euler–
MacLaurin formula [41]. (For a review, see [16, 46, 47].) The result is
 
πv 2πv 2 2 ( M)2
Eex = Le∞ − + ( D) ξ + + N + N̄ (5.65)
6L L 4ξ 2
2π
Pex − P0 = 2kF D + ( D M + N − N̄ ). (5.66)
L

Here the term e∞ denotes the ground-state energy per site. The Fermi velocity is
obtained by the derivative of the dressed energy [16] with respect to the rapidity
at the Fermi level.
The central charge c is given by 1. D and M are integers. M denotes
the change in the number of down spins and D the number of particles jumping
over the Fermi sea through the backscattering. We note that the difference of
the conformal weights (5.64) is given by hnm − h̄nm = nm. Thus, M and D
correspond to the n and m of the c = 1 CFT, respectively. N and N̄ are derived
from particle–hole excitations near the Fermi surface. The Fermi wavenumber kF
is given by kF = πM/L where M is the number of down spins. If the dispersion
is linear, kF is consistent with the number of particles M.
The parameter ξ is given by the dressed charge, which is defined by an
integral equation. We note that the sum of the conformal weights (5.64) is given
by hn,m + h̄n,m = n2 /R 2 + m2 R 2 /4. Thus, the dressed charge ξ corresponds to
the radius R of the c = 1 CFT, ξ = R/2. Under zero magnetic field, the dressed
charge ξ or the radius R is given by [33]
 −1/2
1 1
R= − cos−1 ( ) . (5.67)
2 2π

Copyright © 2003 IOP Publishing Ltd.

 5.3 Various integrable models on two-dimensional lattices
5.3.1 Ising model and Potts model
5.3.1.1 Ising model
Let us consider the Ising model defined on a square lattice. Each lattice site has a
spin variable which takes the two values ±1. We denote by σj the spin variable
of lattice site j . The Hamiltonian of the Ising model is given by

H = −J σi σj (5.68)
"i,j #

where the symbol "i, j # denotes that sites i and j are nearest neighbours and we
take the sum over all the pairs of adjacent sites on the lattice.
There have been many papers written on the two-dimensional Ising model
[8]. However, it should be noted that correlation functions are calculated exactly
for the two-dimensional Ising model in the scaling limit [50, 51].
We note that exact solutions are discussed for the Ising model defined on
various two-dimensional lattices such as the Kagome lattice (for a review, see
[52]).

5.3.1.2 Self-dual Potts model

The Potts model generalizes the Ising model into a p-state model with p > 2
[1, 53, 54]. Let us consider the three-state Potts model defined on a square lattice
[53]. Each lattice site has a spin variable which takes three values 1, 2, 3. We
denote by σj the spin variable of lattice site j . The Hamiltonian of the Potts
model is given by 
H = −J δ(σi , σj ). (5.69)
"i,j #

Here the symbol δ(a, b) denotes the Kronecker delta.

In general, the Potts model is not solvable. However, at its criticality, it is
equivalent to some variant of the six-vertex model and is solvable. Thus, the self-
dual Potts model is integrable [55, 56].

5.3.1.3 Ashkin–Teller model

With each site i we associate two spins: si and ti . They take values ±1. The
Hamiltonian of the model is given by

H=− [K2 (si sj + ti tj ) + K4 si sj ti tj ]. (5.70)
"i,j #

It is known that the Ashkin–Teller model and the eight-vertex model are in the
same universality class [38]. The universality class is described by c = 1 CFT
with the twisted boson [40].

Copyright © 2003 IOP Publishing Ltd.

 5.3.2 Chiral Potts model
5.3.2.1 General case
There is another version of the Potts model [1, 53]. We may assign the chiral
symmetry, or Z/pZ symmetry on the Potts model, which we shall call the chiral
Potts model.
We now explain the most general chiral Potts model defined on a square
lattice [64]. Let a and b denote the spin variables defined on two nearest-
neighbouring sites. The interaction energy between the spins depends on the
difference n = a − b (mod N) as

N−1
E(n) = Ej ω j n ω = e2π/N . (5.71)
j =1

We note that E(N + n) = E(n). The parameter Ej can be written as


Ej N
= −Kj ω j
for j = 1, . . . , (5.72)
kB T 2
where Kj and j constitute N − 1 independent variables, and the symbol [[·]]
denotes the Gaussian symbol. For a real number x, [[x]] denotes the biggest
integer not larger than x. We also assume that EN−j is complex conjugate to
Ej : EN−j = Ej∗ . When N is odd, we have
[[(N−1)/2]]
  
E(n) 2π ((−1)N + 1)
− = 2Kj cos (j n + j ) + KN/2 (−1)n .
kB T j =1
N 2
(5.73)

5.3.2.2 Integrable chiral Potts model

Let us discuss the integrable restriction of the most general chiral Potts model
defined on the square lattice. It has horizontal and vertical couplings. Suppose
that spin variables a and b are located on two neighbouring sites connected
by a horizontal line. When the line goes rightward from a to b, the horizontal
coupling has energy Epq (a − b) and Boltzmann weight Wpq (n). Here p and q
are ‘rapidity’ parameters. For spin variables c and d located on two neighbouring
sites connected by a vertical line, the vertical coupling has energy Ēpq (c − d) and
Boltzmann weight W̄pq (c − d) when the vertical line goes upward from c to d.
The model is called solvable if the Boltzmann weights W and W̄ satisfy the
star–triangle equation

N
W̄qr (b − d)Wpr (a − d)W̄pq (d − c)
d=1
= Rpqr Wpq (a − b)W̄pr (b − c)Wqr (a − c). (5.74)

Copyright © 2003 IOP Publishing Ltd.

 The solution to this equation is given by
 

n
µp y q − x p ω j
Wpq (n) = Wpq (0)
j =1
µq y p − x q ω j
  (5.75)

n
ωxp − xq ωj
W̄pq (n) = W̄pq (0) µp µq
j =1
yq − yp ω j

with a constant Rpqr depending on the three rapidity variables. The constraint
that the Boltzmann weight should have periodicity modulo N: Wpq (n + N) =
Wpq (n) gives, for all rapidity pairs p and q,

 N
µp ypN − xqN yqN − ypN
= (µp µq )N = . (5.76)
µq yqN − xpN xpN − xqN

We can define k and k

such that

k
1 − kypN
µp = = (5.77)
1 − kxpN k

xpN + ypN = k(1 + xpN ypN ) (5.78)

where k 2 + (k
)2 = 1. Thus, the rapidities are placed on a curve of genus g > 1 of
Fermat type.
We make some comments. Multiplying the two equations of (5.76) and
noting that p and q are independent, we can show equation (5.77) and then
equation (5.78). The star–triangle equation is proven illustratively in the appendix
of [60]. We note that the vertex-type formulation of the chiral Potts model
is related to the tetrahedron equation [63]. The integrable chiral Potts model
generalizes the self-dual ZN model given by Fateev and Zamolodchikov [65]. An
elliptic extension of the self-dual ZN model is introduced in [66], and the Yang–
Baxter equation is proven for the model in [67]. Some non-trivial connections
among higher-rank chiral Potts models, elliptic IRF models and Belavin’s ZN
symmetric model are explicitly discussed in [68].

5.3.3 The eight-vertex model

Let us explain the eight-vertex model which was solved by Baxter in 1972 [4].
The Boltzmann weights w(α, β|γ , δ; u) are non-zero if the charge is conserved
modulo 2: α + β = γ + δ (mod) 2. We have the six configurations around a vertex
shown in figure 5.2 and the two in figure 5.5. We assume that there is no external
field. The weights, therefore, become symmetric: 1 = 2 , 3 = 4 , 5 = 6 and

Copyright © 2003 IOP Publishing Ltd.

 (7) (8)
6 ?
 - - 
? 6

Figure 5.5. Vertex configurations for the eight-vertex model. (7) w(2, 2|1, 1);
(8) w(1, 1|2, 2).

7 = 8 .
w(1, 1|1, 1) = w(2, 2|2, 2) = w1 = a8V
w(1, 2|2, 1) = w(2, 1|1, 2) = w2 = b8V
(5.79)
w(1, 2|1, 2) = w(2, 1|2, 1) = w3 = c8V
w(1, 1|2, 2) = w(2, 2|1, 1) = w4 = d8V.
We give a parametrization of the Boltzmann weights. We define the theta function

∞
θ (z; τ ) = 2p1/4 sin πz (1 − p2n )(1 − p2n exp(2πiz))(1 − p2n exp(−2πiz))
n=1
(5.80)
where the nome p is related to the parameter τ by p = exp(πiτ ) with Im τ >
0. We also define the theta functions θ0 (z) and θ1 (z) satisfying θα (z + 1) =
(−1)α θα (z) and θα (z + τ ) = i e−πi(z+τ/2) θ1−α (z) for α = 0, 1, where we define
θ1 (z) by θ1 (z; τ ) = θ (z; 2τ ). The Boltzmann weights a8V (z), b8V (z), c8V (z) and
d8V (z) are expressed as
θ0 (z)θ0(2η) θ1 (z)θ0 (2η)
a8V(z) = b8V (z) =
θ0 (z − 2η)θ0 (0) θ1 (z − 2η)θ0 (0)
(5.81)
θ0 (z)θ1 (2η) θ1 (z)θ1 (2η)
c8V (z) = − d8V(z) = − .
θ1 (z − 2η)θ0(0) θ0 (z − 2η)θ0 (0)

5.3.4 IRF models

5.3.4.1 Unrestricted 8V SOS model
We introduce unrestricted solid-on-solid (SOS) models. They are also called
interaction-round-a-face (IRF) models [1].
To each site i of a two-dimensional square lattice, a spin ai is associated.
Let i, j, k, and  be the lattice sites surrounding a face (or a square), where
i, j, k,  are placed counterclockwise from the southwest corner. We assume that
an elementary configuration is given by that of the four spin variables around the
face, and the probability of having ai , aj , ak , a is denoted by the Boltzmann
weight w(ai , aj , ak , a ; z). Here, the variable z is called the spectral parameter.
For unrestricted 8V SOS model, ai can take any integer. When sites i and j
are nearest neighbouring, then the states ai and aj are said to be admissible if and

Copyright © 2003 IOP Publishing Ltd.

 only if |ai − aj | = 1 [1, 69]. The Yang–Baxter equations are given by

w(a, b, g, f ; z − w)w(f, g, d, e; z)w(g, b, c, d; w)
g

= w(f, a, g, e; w)w(a, b, c, g; z)w(g, c, d, e; z − w) (5.82)
g

where the summation of the variable g is taken over all the admissible states.
The Boltzmann weights are given by
θ (2η − z)
w(d + 1, d + 2, d + 1, d; z, w0 ) = w(d, d − 1, d − 2, d − 1; z) =
θ (2η)
w(d − 1, d, d + 1, d; z, w0 ) = w(d + 1, d, d − 1, d; z, w0 )
θ (z)
=
θ (2η)
√
θ (2η(d + 1) + w0 )θ (2η(d − 1) + w0 )
×
θ (2ηd + w0 )
θ (z + 2ηd + w0 )
w(d + 1, d, d + 1, d; z, w0 ) =
θ (2ηd + w0 )
θ (z − 2ηd − w0 )
w(d − 1, d, d − 1, d; z, w0 ) = . (5.83)
θ (2ηd + w0 )

5.3.4.2 RSOS models

Let us explain restricted solid-on-solid models (RSOS models) [71]. Let s denote
the number of elements in S. Consider an s × s matrix C satisfying the following
conditions [76, 77]:
(i) Cab = Cba = 0 or 1,
(ii) Caa = 0 and
(iii) for each a ∈ S, there should exist b ∈ S such that Cab = 1.
For such a choice of C, we impose a restriction that two states a and b can occupy
the neighbouring lattice sites if and only if Cab = 1. We call such a pair of the
states (a, b) admissible. For the case of unrestricted models, the infinite matrix C
satisfies the conditions (i)–(iii) with an infinite set S.
For an illustration, let us consider the restricted eight-vertex solid-on-solid
model (the restricted 8V SOS model), which we also call the ABF model [71].
For the N-state case, we have S = {1, 2, . . . , N}. The non-zero matrix elements
of C are given by Cj,j +1 = Cj +1,j = 1 for j = 1, 2, . . . , N − 1; other matrix
elements such as C1,N and CN,1 are given by zero. Setting w0 = 0, then we have
the Boltzmann weights of the ABF model. The Boltzmann weights satisfy the
Yang–Baxter relations (5.82) with the finite set S = {1, 2, . . . , N}.
Let us explain CSOS models, another type of RSOS model [75–77]. Here
we assume that 2Nη = m1 , where integer m1 has no common divisor with N.

Copyright © 2003 IOP Publishing Ltd.

 If we set w0 = 0, then we have the Boltzmann weights of the cyclic SOS model
(CSOS model). We can show that the Boltzmann weights satisfy the Yang–
Baxter relations with the finite set S = {1, 2, . . . , N} and the cyclic admissible
conditions C1,N = CN,1 = 1.
We note that the connection between the 6j symbols and the Boltzmann
weights of the IRF models was first discussed in [78].

5.3.4.3 Fusion IRF models and ABCD IRF models

The 8V SOS model was generalized into the fusion IRF models [72]. IRF models
associated with A(1)
n Lie algebra have also been constructed [73]. The IRF models
associated with the B (1) C (1) D (1) type Lie algebra have also been obtained [74].
In the IRF models, the one-point function, which is the magnetization per
site, can be calculated by the corner transfer matrix method invented by Baxter
[1].

5.3.4.4 Gauge transformations

It is sometimes convenient to employ a gauge transformation
gc
w(a, b, c, d; z) → w(a, b, c, d; z) . (5.84)
ga

The transformed Boltzmann weights also satisfy

√ the Yang–Baxter relations (5.82).
For instance, we may set ga = exp(πia/2) θ (2ηa + w0 ) (a ∈ Z).

5.4 Yang–Baxter equation and the algebraic Bethe ansatz

5.4.1 Solutions to the Yang–Baxter equation
5.4.1.1 Derivation of a solution for the six-vertex model
Let us solve the Yang–Baxter equation for the six-vertex model. We recall that it
is given by

w(α, γ |a1 , a2 )w
(β, b3 |γ , a3 )w

(b1 , b2 |α, β)
α,β,γ

= w

(β, α|a2 , a3 )w
(b1 , γ |a1 , β)w(b2 , b3 |γ , α). (5.85)
α,β,γ

The Yang–Baxter equation is illustrated in figure 5.6.

There are 23 × 23 = 64 cases for the entries (a1 , a2 , a3 ) and (b1 , b2 , b3 ).
Due to the ice rule, however, the Yang–Baxter equation is trivial unless a1 + a2 +
a3 = b1 + b2 + b3 . We have thus only 20 entries: (3 C0 )2 + (3 C1 )2 + (3 C2 )2 +
(3 C3 )2 = 1 + 9 + 9 + 1 = 20.

Copyright © 2003 IOP Publishing Ltd.

 a1 a2 a3 a2
YH
H


*
HH 6 6

HH γ a3 a1 β 
u H

v
H *
 YH
H
HH
 H

α H
 = H
H α
H  H
 u + vHH  u + v HH


 β b3 b1 γ H
 HH
v HH
u H

b1 b2 b2 b3

Figure 5.6. The Yang–Baxter equations for vertex models. The spectral parameters are
shown by the angles between pairs of straight lines.

Let us consider symmetries of the Boltzmann weights given in (5.3). If we

exchange 1 and 2, the Boltzmann weights do not change. Thus, we have reduced
20 cases into 10 cases:
(a1 , a2 , a3 ; b1 , b2 , b3 ) = (1, 1, 1; 1, 1, 1),
(1, 1, 2; 1, 1, 2), (1, 1, 2; 1, 2, 1), (1, 1, 2; 2, 1, 1),
(1, 2, 1; 1, 1, 2), (1, 2, 1; 1, 2, 1), (1, 2, 1; 2, 1, 1),
(2, 1, 1; 1, 1, 2), (2, 1, 1; 1, 2, 1), (2, 1, 1; 2, 1, 1). (5.86)
We now recall that the Boltzmann weights (5.3) have the symmetry
w(α, β|γ , δ) = w(γ , δ|α, β) = w(β, α|δ, γ ) (5.87)
Combining these symmetries we can show that the Yang–Baxter equation for
the two entries (1) and (2) are equivalent: (1) (a1 , a2 , a3 ; b1 , b2 , b3 ); and (2)
(b3 , b2 , b1 ; a3 , a2 , a1 ). Precisely, the lhs (or rhs) of the Yang–Baxter equation
of case (1) corresponds to the rhs (or lhs) of equation (5.85). Thus, we have the
following three cases:
(1, 1, 2; 1, 1, 2) (1, 1, 2; 1, 2, 1) (1, 2, 1; 1, 1, 2). (5.88)
For the three cases, the Yang–Baxter equations are given by
ac
a

= bc
b

+ ca
c

ab
c

= ba
c

+ cc
b

(5.89)
cb
a

= ca
b

+ bc
c

.
A non-trivial solution (a

, b

, c

) exists only if the determinant vanishes:


  
ac −bc
−ca


2
2



(a ) + (b ) − (c )

2 a 2 + b2 − c2
0




cc
ba − ab  = abca b c a
b

− .
cb −ca −bc
 ab
(5.90)

Copyright © 2003 IOP Publishing Ltd.

 We define the parameter by

a 2 + b2 − c2
= . (5.91)
2ab
The condition that the determinant vanishes is given by

=
. (5.92)

Thus, the transfer matrices τ and τ

commute, if the two sets of weights (a, b, c)
and (a
, b
, c
) have the same .
In terms of the spectral parameter u, we can parametrize the three
Boltzmann weights a, b and c. We express the weight as w(α, β|γ , δ; u). Let
u and v be arbitrary. We denote w(α, β|γ , δ), w
(α, β|γ , δ) and w

(α, β|γ , δ)
as w(α, β|γ , δ; u), w(α, β|γ , δ; u + v) and w(α, β|γ , δ; v), respectively. The
Yang–Baxter equations are depicted in figure 5.7. As a solution, we may have
(a, b, c) = (ρ sinh(u + 2η), ρ sinh u, ρ sinh 2η). Here we set = cosh(2η).
The transfer matrices τ (u) and τ (v) commute: τ (u)τ (v) = τ (v)τ (u).

5.4.1.2 Gauge transformations for vertex models

Let us suppose that the ws satisfy the Yang–Baxter equation. Then we can show
that transformed weights w̃s defined by

w̃(α, β|γ , δ; u) = ()α+γ exp(κ(α + γ − β − δ)u)w(α, β|γ , δ; u) (5.93)

also satisfy the Yang–Baxter equations [82, 83]. Here  = ±1, and the number κ
is arbitrary.
The gauge transformation is important in the derivation of the Jones
polynomial from the symmetric Boltzmann weights of the six-vertex model under
zero field [82]. It is also quite useful when we discuss the relation of the six-vertex
model to the quantum group, as we shall see in section 5.5 (see also the appendix
in [95]).

5.4.2 Algebraic Bethe ansatz

5.4.2.1 R-matrix and the L-operator
Let us diagonalize the transfer matrix using the algebraic Bethe ansatz [16,80,81].
Let us introduce the notation of the matrix tensor product. We define the
j
direct product A ⊗ B of matrices A and B. Let Ak denote the matrix element for
the entry of column j and row k of the matrix A. Then, the matrix element of
column (j1 , j2 ) and (k1 , k2 ) is defined by
j ,j j j
(A ⊗ B)k11 ,k22 = Ak11 Bk22 . (5.94)

Copyright © 2003 IOP Publishing Ltd.

 ab
We now define the R-matrix of the XXZ spin chain. The element Rcd
corresponds to the entry of column (a, b) and row (c, d):
 
R(z)11
11 R(z)11
12 R(z)11
21 R(z)11
22  
  a(z) 0 0 0
R(z)12 R(z)12 R(z)12 22 
R(z)12  0 c(z) b(z) 0 
R(z) =  = .
11 12 21
   0 b(z) c(z) 0 
R(z)21
11 R(z)21
12 R(z)21
21
21
R(z)22 
0 0 0 a(z)
R(z)22
11 R(z)22
12 R(z)22
21 R(z)22
22
(5.95)
Here a(z), b(z) and c(z) are given by

a(z) = sinh(z + 2η) b(z) = sinh z c(z) = sinh 2η. (5.96)

Here, the functions a(z), b(z) and c(z) are equivalent to the Boltzmann weights
a, b and c in section 5.2.
We now introduce the L-operators. We write the matrix element for the L-
j
operator with entry (j, k) as (Ln (z))j k or Ln (z)k . The L-operator for the XXZ
spin chain is given by
   
Ln (z)11 Ln (z)12 sinh(zIn + ησnz ) sinh 2ησn−
Ln (z) = = . (5.97)
Ln (z)2 Ln (z)2 sinh 2ησn+ sinh(zIn − ησnz )
1 2

Here In and σna (n = 1, . . . , L) are acting on the nth vector space Vn . The
L-operator is an operator-valued matrix which acts on the auxiliary vector
space V0 .
The symbols σ ± denote σ + = E12 and σ − = E21 , and σ x , σ y , σ z are the
Pauli matrices.
In terms of the R-matrix and L-operators, the Yang–Baxter equation can be
expressed as

R(z − t)(Ln (z) ⊗ Ln (t)) = (Ln (t) ⊗ Ln (z))R(z − t). (5.98)

Here the tensor symbol in Ln (z) ⊗ Ln (t) denotes the tensor product of the
auxiliary spaces.
The Yang–Baxter equation (5.98) gives the relation between the two products
of 4 × 4 matrices. For an illustration, we consider the lhs of (5.98):
a ,a
 c ,c
[R(z − t) · Ln (z) ⊗ Ln (t)]b11 ,b22 = R(z − t)ac11,c
,a2
2
(Ln (z) ⊗ Ln (t))b11 ,b22
c1 ,c2

= R(z − t)ac11,c
,a2
L (z)cb11 × Ln (t)cb22 . (5.99)
2 n
c1 ,c2

Here the symbol × denotes the product of matrices acting on the nth space Vn .
Expressing the operator products in the nth space Vn , the lhs of (5.98) can be

Copyright © 2003 IOP Publishing Ltd.

 6
6αn
(1) (2) a1 @
I  a2
@
@
z−η @ z
b - a @
@
@
@
b1 @ b2
βn

a ,a
Figure 5.7. (1) Ln (z)ab |αn ,βn ; (2) R(z)b1 ,b2 .
1 2

αn t −η αn
b1 HH 6
6 6 
6 * a1
HH z − η 
HH 
H c1 a1 b1 c1 
HH HH 
H *
 
γn H  H  γn
H
 z−t = z−t H
H
 HH  H
t − η  j
H  HH
c2 a2 b2 c2 HH
  HH z − η
  HH
b2  β β j a2
H
n n

Figure 5.8. The Yang–Baxter equation: R(z − t)(L(z) ⊗ L(t)) = (L(t) ⊗ L(z))R(z − t),
equation (5.98).

written as
a ,a 
[R(z − t) · Ln (z) ⊗ Ln (t)]b11 ,b22 αn ,βn
 c1 ,c2 
= R(z − t)ac11,c
,a2
(L n (z) ⊗ L n (t)) b1 ,b2 α
2 n ,βn
c1 ,c2
 c  c 
= R(z − t)ac11,c
,a2
L (z)b11 αn ,γn Ln (t)b22 γn ,βn .
2 n
(5.100)
c1 ,c2 γn

5.4.2.2 Monodromy matrix and the construction of the eigenvector

We define the monodromy matrix by the product of the L-operators (see also
figure 5A.1):
T (z) = LN (z) · · · L2 (z)L1 (z). (5.101)

The transfer matrix τ6V (z) of the six-vertex model is given by the trace of T (z):
 
A(z) B(z)
τ6V (z) = tr T (z) = A(z) + D(z) where T (z) = . (5.102)
C(z) D(z)

Copyright © 2003 IOP Publishing Ltd.

 The Yang–Baxter equation (5.98) leads to the commutation relation: R(z − t)×
(T (z) ⊗ T (t)) = (T (t) ⊗ T (z))R(z − t), from which we have many relations
among the operators A, B, C and D. For instance, we have B(z)B(t) = B(t)B(z).
Furthermore, we have

a(t − z) c(t − z)
A(z)B(t) = B(t)A(z) − B(z)A(t) (5.103)
b(t − z) b(t − z)
a(z − t) c(z − t)
D(z)B(t) = B(t)D(z) − B(z)D(t). (5.104)
b(z − t) b(z − t)

We define the ‘vacuum’ by

N
$ %& '
|0# = |↑#1 |↑#2 · · · |↑#N . (5.105)

Multiplying A(z) and D(z) on the vacuum, we have

A(z)|0# = a(z − η)N |0# D(z)|0# = b(z − η)N |0#. (5.106)

Let us consider the vector generated by the product of B operators:

|M# = B(t1 ) · · · B(tM )|0#. (5.107)

Then, through the commutation relations such as (5.103) and (5.104) we can
show [7,16] that the vector |M# gives an eigenvector of the transfer matrix τ6V (z)
if rapidities t1 , t2 , . . . , tM satisfy the set of equations
 N 
M  
a(tj − η) c(tk − tj ) b(tj − tk ) a(tj − tk ) b(tk − tj )
=−
b(tj − η) b(tk − tj ) c(tj − tk ) k=1;k=j
b(tj − tk ) a(tk − tj )
for j = 1, . . . , M. (5.108)

These are the Bethe ansatz equations (5.12) with a different parametrization. For a
set of solutions, t1 , t2 , . . . , tM to equations (5.108), the eigenvalue of the transfer
matrix τ6V (z) is given by

M
a(tj − z) M
a(z − tj )
(z; t1 , t2 , . . . , tM ) = a(z − η)N + b(z − η)N
j =1
b(tj − z) j =1
b(z − tj )
M
sinh(tj − z + 2η)
= sinhN (z + η)
j =1
sinh(tj − z)
M
sinh(tj − z − 2η)
+ sinhN (z − η) . (5.109)
j =1
sinh(tj − z)

Copyright © 2003 IOP Publishing Ltd.

 5.4.2.3 Connection to the coordinate Bethe ansatz result
Let us compare the result of the coordinate Bethe ansatz in section 5.2. We
consider the disordered phase: −1 < < 1. First, we change the variables w
and α defined in section 5.2 into u and ζ by w = 2u − µ and α = 2ζ − iµ,
respectively. Thus, from the expressions (5.22) we have (a, b, c) = (sin(µ − u),
sin u, sin µ) = i(sinh(−iµ + iu), sinh(−iu), sinh(−iµ)) and
ab + (c2 − b2 )zj sinh((αj − iw − 2iµ)/2)
LBaxter (zj ) = =−
a(a − bzj ) sinh((αj − iw)/2)
sinh(−(ζj − iu) + iµ)
= . (5.110)
sinh(ζj − iu)
Here we recall that the symbol LBaxter (z) has been defined in equation (5.15) in
order to denote the eigenvalue of the transfer matrix of the six-vertex model. The
expression (5.110) can be derived from the formula (5.109) of the algebraic Bethe
ansatz as follows. First, we take the gauge transformation b(z) → −b(z), which
corresponds to the case  = −1 and κ = 0 in equation (5.93). Then, we replace
the variables z, tj and 2η in (5.109) by z − η → −iu, tj − η → −ζj and 2η → iµ,
respectively. We have the following:
a(tj − z) a(tj − z) sinh(tj − z + 2η) sinh(−(ζj − iu) + iµ)
→− =− →
b(tj − z) b(tj − z) sinh(tj − z) sinh(ζj − iu)
(5.111)
and sinhN (z + η) → sinhN (−iu + iµ) . Thus, the formula (5.109) reproduces
the expression (5.14) for the eigenvalues of the transfer matrix except for the
normalization factor iN .

5.5 Mathematical structures of integrable lattice models

5.5.1 Braid group
5.5.1.1 The Yang–Baxter equation in operator formalism
The Yang–Baxter equation in section 5.2 gives a sufficient condition for the
existence of commuting transfer matrices. However, there are other viewpoints
on the Yang–Baxter equation.
j
Let Ek denote the matrix given by
j
(Ek )ab = δa,j δb,k a, b = 1, 2. (5.112)
We define operators Xj (u) by
⊗(j −1) ⊗(N−j −1)
 $ %& ' $ %& '
Xj (z) = w(a, b|c, d; z) I ⊗ · · · ⊗ I ⊗Eac ⊗ Ebd ⊗ I ⊗ · · · ⊗ I
a,b,c,d
for j = 1, . . . , N − 1. (5.113)

Copyright © 2003 IOP Publishing Ltd.

 c,d
Here, the symbol w(a, b|c, d; z) corresponds to Ra,b (z), and the essential part of
Xj (z) is given by

X(z) = w(a, b|c, d; z)Eac ⊗ Ebd (5.114)
a,b,c,d

which is equivalent to R(z). In terms of the Xj (z), the Yang–Baxter equation can
be expressed as

Xj (z)Xj +1 (z + t)Xj (t) = Xj +1 (t)Xj (z + t)Xj +1 (z) for j = 1, . . . , N − 1.

(5.115)

5.5.1.2 The braid group

The braid group BN for N strings is an infinite group which is generated by the
generators b1 , . . . , bN−1 satisfying the defining relations

bj bj +1 bj = bj +1 bj bj +1
bi bj = bj bi for |i − j | > 1. (5.116)

Let us assume that the limit limz→∞ Xj (z) exists. Then, equation (5.115)
becomes

Xj (∞)Xj +1 (∞)Xj (∞) = Xj +1 (∞)Xj (∞)Xj +1 (∞)

for j = 1, . . . , N − 1. (5.117)

This is nothing but the defining relations of the braid group. Thus, the Boltzmann
weights of solvable models expressed in terms of the spectral parameter lead to
representations of the braid group.
We now show that from a given exactly solvable model, one can derive two
different representations of the braid group [82]. Here, the gauge transformation
(5.93) plays a central role. This technical point is quite fundamental when we
make connections between exactly solvable models and quantum groups and the
Temperley–Lieb algebra (for instance, see the appendix in [95]).
We first consider the Boltzmann weights of the zero-field six-vertex model
given by equations (5.96). Taking the infinite limit to them, we have
 
1 0 0 0
0 0 exp(−2η) 0
lim X(z)/ sinh(z + 2η) =  
. (5.118)
z→∞ 0 exp(−2η) 0 0
0 0 0 1

The representation of the braid group (5.118) leads to a link polynomial equivalent
to the linking number.
Let us now apply the gauge transformation (5.93) with  = 1 and κ = 1/2
to the weights given by equations (5.96) [82]. Then, from the transformed

Copyright © 2003 IOP Publishing Ltd.

 Boltzmann weights we have
 
1 0 0 0
0 0 exp(−2η) 0
lim X̃(z)/sinh(z + 2η) = 
0 exp(−2η) 1 − exp(−4η)
. (5.119)
z→∞ 0
0 0 0 1

This matrix representation of the braid group leads to the Jones polynomial with
q = exp(2η) [82].
We note that based on the representations of the braid group which are
derived from the Boltzmann weights of exactly solvable models, we can construct
various invariants of knots and links (see, for reviews, [83, 112, 113]).

5.5.2 Quantum groups (Hopf algebras)

From the quantum groups (the Hopf algebras), we can systematically construct
representations of the braid group such as that derived from the six-vertex model.
Almost all the solutions of the Yang–Baxter equations can be constructed in some
framework of quantum groups. Furthermore, the connection between solvable
models and quantum groups is useful for investigating the non-trivial properties
of integrable models.
Let us introduce the quantum group Uq (sl2 ), which is a q-analogue of the
universal enveloping algebra of sl2 . Generators X± , H satisfy

K − K −1
KX± K −1 = q ±2 X± [X+ , X− ] = . (5.120)
q − q −1

We may express K as K = q H . Taking the limit of q to unity, the relations are

reduced into the commutation relations of sl2 .
The tensor product is defined by the following:

(K) = K ⊗ K
(X+ ) = X+ ⊗ I + K ⊗ X+ (5.121)
(X− ) = X− ⊗ K −1 + I ⊗ X− .

The operation (·) is called the comultiplication. In the quantum group, the
comultiplication does not commute with the exchange operator σ defined by
σ a ⊗ b = b ⊗ a. However, there is an operator R which satisfies

R (x) = σ (x)R for x ∈ Uq (sl2 ). (5.122)

Thus, the tensor product V1 ⊗ V2 can be related to V2 ⊗ V1 through the R-matrix

of the quantum group.
In the Uq (sl2 ), the R-matrix can be constructed in the operator formalism. If
operators X± and H satisfy the defining relations of Uq (sl2 ), then the operator R

Copyright © 2003 IOP Publishing Ltd.

 defined by the following satisfies the intertwining relation (5.122)

R = q −H ⊗H/2 expq (−(q − q −1 )K −1 X+ ⊗ X− K) (5.123)

where expq x denotes the infinite series


∞
xn
expq x = q −n(n−1)/2 . (5.124)
n=0
[n]!

The operator R is called the universal R-matrix.

The representation (5.119) of the braid group corresponds to the
representation of the universal R-matrix on the tensor product of two fundamental
representations.
We note that the universal R-matrix can be constructed canonically through
Drinfeld’s quantum double construction (for instance, see [24]). This is similar to
the Sugawara construction which derives the energy–momentum tensor from the
current operator.

Appendix. Commuting transfer matrices and the Yang–Baxter

equations
We show that if two given sets of Boltzmann weights of the six-vertex model
satisfy the Yang–Baxter relation, then their transfer matrices commute. We
consider three sets of Boltzmann weights: (w1 , w2 , w3 ) = (a, b, c), (a
, b
, c
)
and (a

, b

, c

). Let us denote by τ
and τ

the transfer matrices constructed from

the sets of Boltzmann weights (a
, b
, c
) and (a

, b

, c

), respectively. Then, we
can show that if the three sets of Boltzmann weights satisfy the Yang–Baxter
equations given in section 5.2.3, then the transfer matrices τ
and τ

commute.
Let us now explicitly discuss the commutation relation. We first introduce the
monodromy matrix. It is an N ranked tensor, whose (α, β) elements are defined
as follows:

(Tα,β )ab11 ,...,a
,...,bN =
N
w(β, b1 |a1 , c2 )w(c2 , b2 |a2 , c3 ) · · · w(cN , bN |aN , α).
c2 ,...,cN
(5A.1)
The transfer matrix is given by the trace of the monodromy matrix

τ = tr(T ) = Tα,α . (5A.2)
α=1,2

In terms of the matrix elements, we have


(τ )ab11 ,...,a
,...,bN =
N
(Tα,α )ab11 ,...,a
,...,bN
N
(5A.3)
α=1,2

Copyright © 2003 IOP Publishing Ltd.

 a1 a2 aN
···
β c2 cN α
···

b1 b2 bN

a ,...,a
Figure 5A.1. The matrix element (α, β) of the monodromy matrix (Tα,β )b1 ,...,bN .
1 N

Let us denote by T
and T

the monodromy matrices for the sets of

Boltzmann weights (a
, b
, c
) and (a

, b

, c

), respectively. We consider the

product of the matrix elements of the two monodromy matrices: Tα,β
T

. The
γ ,δ

entry of (a1 , . . . , aN ) and (b1 , . . . , bN ) of the product Tα,β Tγ ,δ is given by

(Tγ
1 ,γN+1 Tδ

1 ,δN+1 )b11 ,...,bNN

a ,...,a

(Tγ
1 ,γN+1 )ae11,...,e (Tδ

1 ,δN+1 )b11 ,...,bNN

e ,...,e
= ,...,aN
N
e1 ,...,eN
 
= w
(γ1 , e1 |a1 , c2 )w
(c2 , e2 |a2 , c3 ) · · · w
(cN , eN |aN , γN+1 )
e1 ,...,eN c2 ,...,cN

× w

(δ1 , b1 |e1 , d2 )w

(d2 , b2 |e2 , d3 ) · · · w

(dN , bN |eN , δN+1 )

d2 ,...,dN
  
= (w
(γ1 , e1 |a1 , c2 )w

(δ1 , b1 |e1 , d2 ))
c2 ,...,cN d2 ,...,dN e1 ,...,eN

· (w
(c2 , e2 |a2 , c3 )w

(d2 , b2 |e2 , d3 )) · · ·
· (w
(cN , eN |aN , γN+1 )w

(dN , bN |eN , δN+1 ))

  γ ,c c ,c c ,γN+1
= S(a1 , b1 )δ11,d22 · S(a2 , b2 )d22 ,d33 · · · S(aN , bN )dNN ,δN+1 . (5A.4)
c2 ,...,cN d2 ,...,dN

cj ,c
Here S(aj , bj )dj ,dj+1
j+1
has been defined by
cj ,c 
S(aj , bj )dj ,dj+1
j+1
= w
(cj , ej |aj , cj +1 )w

(dj , bj |ej , dj +1 ). (5A.5)

ej

c ,c
We define the matrix element Md00,d11 as follows:
c ,c
Md00,d11 = w(d0 , d1 |c0 , c1 ). (5A.6)
c ,c
Here we assume that Md00,d11 denotes the matrix element for column (c0 , d0 ) and
row (c1 , d1 ) of the matrix M. Multiplying the matrix M to the product Tα,β
T

γ ,δ
and applying the Yang–Baxter relation N times, we can derive the following:
 γ ,c 
Mδ00,d11 Tc
1 ,γN+1 Td

1 ,δN+1 = Tc

1 ,cN+1 Td
1 ,dN+1 MdNN ,δN+1
c ,γN+1
. (5A.7)
c1 ,d1 cN ,dN

Copyright © 2003 IOP Publishing Ltd.

 a1 a2 ··· aN a1 a2 ··· aN
γ0 c1 c c2 c c γN+1 γ0 s c1 c2 c cγN+1
@ @
@ @
@ ··· = @ ···
@ s s s c @ s s
δ0 d1 d2 δN+1 δ0 d1 d2 δN+1
b1 b2 ··· bN b1 b2 ··· bN
·
·
·
a1 a2 ··· aN
γ0 s c1 s s cN γN+1
@
= @
··· @
c c c @
δ0 d1 dN δN+1
b1 b2 ··· bN

Figure 5A.2. Pictorial proof of the commutation relation MT

T

= T

T
M. Open and
closed circles denote the Boltzmann weights w
and w

, respectively. The summation over

variables c1 , . . . , cN and d1 , . . . , dN is assumed.

Let us briefly discuss the derivation of relation (5A.7). It is depicted in

figure 5A.2. In the first equality of figure 5A.2, we have applied the Yang–Baxter
relation formulated as follows:
 γ ,c 
S
(a1 , b1 )δ00,d11 Mdc11,d
γ ,c
Mδ00,d11 S(a1 , b1 )cd11,c
,d2 =
2 ,c2
2
. (5A.8)
c1 ,d1 c1 ,d1
cj ,c
Here the symbol S
(aj , bj )dj ,dj+1
j+1
has been defined by
c ,c 
S
(aj , bj )djj ,dj+1
j+1
= w

(cj , ej |aj , cj +1 )w
(dj , bj |ej , dj +1 ). (5A.9)
ej

We also note that the lhs of (5A.7) corresponds to the sum

  γ ,c c2 ,c3 cN ,γN+1
Mδ00,d11 S(a1 , b1 )cd11,c
,d2 · S(a2 , b2 )d2 ,d3 · · · S(aN , bN )dN ,δN+1 .
2

c1 ,c2 ,...,cN d1 ,d2 ,...,dN

(5A.10)
Let us consider the inverse of the matrix M:
(M −1 M)cd11,c −1 c1 ,c2
,d2 = (MM )d1 ,d2 = δc1 ,c2 δd1 ,d2 .
2
(5A.11)

Multiplying the inverse M −1 to both sides of (5A.7), we have

MT
T

M −1 = T

T
. (5A.12)
Noting tr(MT
T

M −1 ) = tr(T
T

), we obtain the commutation relation of the

transfer matrices
τ
τ

= τ

τ
. (5A.13)

Copyright © 2003 IOP Publishing Ltd.

 References
[1] Baxter R J 1982 Exactly Solved Models in Statistical Mechanics (London:
Academic Press)
[2] Bethe H A 1931 Z. Phys. 71 205
[3] Yang C N and Yang C P 1966 Phys. Rev. 150 321
Yang C N and Yang C P 1966 Phys. Rev. 150 327
Yang C N and Yang C P 1966 Phys. Rev. 151 258
[4] Baxter R 1972 Ann. Phys. 70 193.
[5] Drinfeld V G 1985 Sov. Math. Dokl. 32 254
Drinfeld V G 1987 Proc. Int. Congress of Mathematics (Berkeley, 1986) ed A M
Gleason (Providence, RI: American Mathematical Society) pp 798–820
[6] Jimbo M 1985 Lett. Math. Phys. 10 63
[7] Takhtajan L and Faddeev L 1979 Russ. Math. Survey 34 11
[8] McCoy B M and Wu T T 1973 The Two-Dimensional Ising Model (Cambridge, MA:
Harvard University Press)
[9] Ziman J M 1979 Models of Disorder (Cambridge: Cambridge University Press)
[10] Gaudin M 1983 La fonction de l’onde de Bethe pour les modèles exacts de la
méchanique statistique (Paris: Masson)
[11] Itzykson C, Saleur H and Zuber J-B (ed) 1988 Conformal Invariance and
Applications to Statistical Mechanics (Singapore: World Scientific)
[12] Itzykson C and Drouffe J-M 1989 Statistical Field Theory vol I and II (Cambridge:
Cambridge University Press)
[13] Yang C N and Ge M-L 1989 Braid Groups, Knot Theory and Statistical Mechanics
(Singapore: World Scientific)
[14] Jimbo M (ed) 1990 Yang–Baxter Equation in Integrable Systems (Singapore: World
Scientific)
[15] Martin P 1991 Potts Models and Related Problems in Statistical Mechanics
(Singapore: World Scientific)
[16] Korepin V E, Bogoliubov N M and Izergin A G 1993 Quantum Inverse Scattering
Method and Correlation Functions (Cambridge: Cambridge University Press)
[17] Lusztig G 1993 Introduction to Quantum Groups (Boston: Birkhäuser)
[18] Chari V and Pressley A 1994 A guide to Quantum Groups (Cambridge: Cambridge
University Press)
[19] Korepin V E and Essler F H (ed) 1994 Exactly Solvable Models of Strongly
Correlated Electrons (Singapore: World Scientific)
[20] Jimbo M and Miwa T 1995 Algebraic Analysis of Solvable Lattice Models (Rhode
Island: American Mathematical Society)
[21] Tsvelik A M 1995 Quantum Field Theory in Condensed Matter Physics
(Cambridge: Cambridge University Press)
[22] Majid S 1995 Foundations of Quantum Group Theory (Cambridge: Cambridge
University Press)
[23] Di Francesco P, Mathieu P and Sénéchal D 1997 Conformal Field Theory (New
York: Springer)
[24] Etingof P and Schiffmann O 1998 Lectures on Quantum Groups (Cambridge, MA:
International Press)
[25] Gogolin A O, Nersesyan A A and Tsvelik A M 1998 Bosonization and Strongly
Correlated Systems (Cambridge: Cambridge University Press)

Copyright © 2003 IOP Publishing Ltd.

 [26] Takahashi M 1999 Thermodynamics of One-Dimensional Solvable Models
(Cambridge: Cambridge University Press)
[27] Sachdev S 1999 Quantum Phase Transitions (Cambridge: Cambridge University
Press).
[28] Kashiwara M and Miwa T (ed) 2000 Physical Combinatorics (Boston: Birkhäuser)
[29] Kashiwara M and Miwa T (ed) 2002 MathPhys Odyssey 2001 (Boston: Birkhäuser)
[30] Lieb E H 1967 Phys. Rev. 162 162
Lieb E H 1967 Phys. Rev. Lett. 18 1046
Lieb E H 1967 Phys. Rev. Lett. 19 108
[31] Lieb E H and Wu F Y 1972 Phase Transitions and Critical Phenomena vol 1, ed
C Domb and M S Green (London: Academic Press) p 331
[32] Johnson J D, Krinsky S and McCoy B M 1973 Phys. Rev. A 8 2526.
[33] Luther A and Peschel I 1975 Phys. Rev. B 12 3908
[34] Cardy J L 1984 J. Phys. A: Math. Gen. 17 L385
Cardy J L 1984 J. Phys. A: Math. Gen. 17 L961
Cardy J L 1986 Nucl. Phys. B 270 186
[35] Blöte H W J, Cardy J L and Nightingale M P 1986 Phys. Rev. Lett. 56 742
[36] Affleck I 1986 Phys. Rev. Lett. 56 746
[37] Affleck I 1990 Fields, Strings and Critical Phenomena ed E Brezin and J Zinn-Justin
(Amsterdam: North-Holland) p 563
[38] Kadanoff L P and Brown A C 1979 Ann. Phys. 121 318
[39] Cardy J L 1987 J. Phys. A: Math. Gen. 20 L891
[40] Yang S-K 1988 Nucl. Phys. B 285 153
[41] de Vega H J and Woynarovich F 1985 Nucl. Phys. B 251 439
[42] de Vega H J and Karowski M 1987 Nucl. Phys. B 285 [FS 19] 619
[43] Pokrovskii S V and Tsvelik A M 1987 Sov. Phys.–JETP 66 1275
Pokrovskii S V and Tsvelik A M 1987 Zh. Eksp. Teor. Fiz. 93 2232
[44] Bogoliubov N M, Izergin A G and Yu Reshetikhin N 1987 J. Phys. A: Math. Gen.
20 6023
[45] de Vega H J 1987 J. Phys. A: Math. Gen. 20 6023
[46] de Vega H J 1989 Int. J. Mod. Phys. 4 2371
[47] Suzuki J, Nagao T and Wadati M 1992 Int. J. Mod. Phys. B 6 119
[48] Haldane F D M 1981 Phys. Rev. Lett. 47 1840
Haldane F D M 1981 J. Phys. C: Solid State Phys. 14 2585
[49] Baxter R 1971 Phys. Rev. Lett. 26 852
[50] Wu T T, McCoy B M, Tracy C A and Barouch E 1976 Phys. Rev. B 13 316
[51] Sato M, Miwa T and Jimbo M 1977 Proc. Jap. Acad. 53A 147, 153, 183
Sato M, Miwa T and Jimbo M 1978 Publ. RIMS, Kyoto University 14 223
Sato M, Miwa T and Jimbo M 1979 Publ. RIMS, Kyoto University 15 201, 577, 871.
[52] Syôji I 1972 Phase Transitions and Critical Phenomena vol 1, ed C Domb and M S
Green (London: Academic Press) p 269
[53] Potts R B 1952 Proc. Cambr. Phil. Soc. 48 106
[54] Kihara T, Midzuno Y and Shizume T 1954 J. Phys. Soc. Japan 9 681
[55] Temperly H N V and Lieb E H 1971 Proc. R. Soc. London A 322 251
[56] Baxter R J 1982 J. Stat. Phys. 28 1
[57] Au-Yang H, McCoy B M, Perk J H H, Tang S and Yan M-L 1987 Phys. Lett. A 123
219
[58] Baxter R J, Perk J H H and Au-Yang H 1988 Phys. Lett. A 128 138

Copyright © 2003 IOP Publishing Ltd.

 [59] Albertini G, McCoy B M and Perk J H H 1989 Adv. Stud. Pure Math. 19 1
[60] Au-Yang H and Perk J H H 1989 Adv. Stud. Pure Math. 19 57
[61] Bazhanov V V and Stroganov Yu G 1990 J. Stat. Phys. 59 799
[62] von Gehlen G and Rittenberg V 1985 Nucl. Phys. B 257 351
[63] Sergeev S M, Mangazeev V V and Stroganov Yu G 1996 J. Stat. Phys. 82 31
[64] Au-Yang H and Perk J H H 1997 Int. J. Mod. Phys. B 11 11
[65] Fateev V A and Zamolodchikov A B 1982 Phys. Lett. A 92 37
[66] Kashiwara M and Miwa T 1986 Nucl. Phys. B 275 121
[67] Hasegawa K and Yamada Y 1990 Phys. Lett. A 146 387
[68] Yamada Y September 2002 Proc. Sixth Int. Workshop CFT and IM, Russia, Int. J.
Mod. Phys. A to appear
See also Yamada Y 1999 Commentarii Mathematici, Universitatis Sancti Pauli 48
49–76
[69] Baxter R 1973 Ann. Phys. 76 1
Baxter R 1973 Ann. Phys. 76 25
Baxter R 1973 Ann. Phys. 76 48
[70] Belavin A A 1981 Nucl. Phys. B 180 189
[71] Andrews G E, Baxter R J and Forrester P J 1984 J. Stat. Phys. 35 193
[72] Date E, Jimbo M, Kuniba A, Miwa T and Okado M 1988 Adv. Stud. Pure Math. 16
17
[73] Jimbo M, Miwa T and Okado M 1988 Nucl. Phys. B 300 74
[74] Jimbo M, Miwa T and Okado M 1988 Commun. Math. Phys. 116 353
[75] Pearce P A and Seaton K A 1988 Phys. Rev. Lett. 60 1347
[76] Kuniba A and Yajima T 1987 J. Stat. Phys. 52 829
[77] Akutsu Y, Deguchi T and Wadati M 1988 J. Phys. Soc. Japan 57 1173
[78] Pasquier V 1988 Commun. Math. Phys. 118 507
[79] Korepin V E 1982 Commun. Math. Phys. 86 381
[80] Faddeev L D and Takhtajan L 1984 J. Sov. Math. 24 241
[81] Takhtajan L A 1985 Introduction to Algebraic Bethe Ansatz (Lecture Notes in
Physics vol 242) (Berlin: Springer) pp 175–219
[82] Akutsu Y and Wadati M 1987 J. Phys. Soc. Japan 56 3039
[83] Wadati M, Deguchi T and Akutsu Y 1989 Phys. Rep. 180 247
[84] Jimbo M 1992 Nankai Lectures on Mathematical Physics (Singapore: World
Scientific) pp 1–61
[85] Jimbo M 1986 Commun. Math. Phys. 102 537
[86] Bazhanov V V 1987 Commun. Math. Phys. 113 471
[87] Kashiwra M 1990 Commun. Math. Phys. 133 249
[88] Lusztig G 1990 Proc. Amer. Math. Soc. 3 447
Lusztig G 1990 Prog. Theor. Phys. 102 (Supplement) 175
[89] Babelon O, Bernard D and Billey E 1996 Phys. Lett. B 375 89
[90] Felder G 1994 Proc. Int. Congr. Mathematicians (Zürich) (Basel: Birkhäuser)
p 1247
[91] Felder G and Varchenko A 1996 Commun. Math. Phys. 181 741
Felder G and Varchenko A 1996 Nucl. Phys. B 480 485
[92] Jimbo M, Konno H, Odake S and Shiraishi J 1999 Transformation Groups 4 303
[93] Maillet J M and Sanchez de Santos J 2000 Amer. Math. Soc. Transl. (2) 201 137–78
[94] Kitanine N, Maillet J M and Terras V 2000 Nucl. Phys. B 567 [FS] 554
[95] Deguchi T, Fabricius K and McCoy B M 2001 J. Stat. Phys. 102 701

Copyright © 2003 IOP Publishing Ltd.

 [96] Pokrovskii V L and Talapov A L 1980 Sov. Phys. JETP 51 134.
[97] den Nijs M 1988 Phase Transitions and Critical Phenomena, vol 12, ed C Domb
and J L Lebowitz (London: Academic Press) p 219
[98] van Beijeren H 1977 Phys. Rev. Lett. 38 993
[99] Jayaprakash C, Saam W F and Teitel S 1983 Phys. Rev. Lett. 50 2017
Jayaprakash C and Saam W F 1984 Phys. Rev. B 30 3916
[100] Akutsu Y, Akutsu N and Yamamoto T 1988 Phys. Rev. Lett. 61 424
Saam W F 1989 Phys. Rev. Lett. 62 2636
Akutsu Y, Akutsu N and Yamamoto T 1989 Phys. Rev. Lett. 62 2637
[101] Noh J D and Kim D 1996 Phys. Rev. E 53 3225
[102] Abraham D B, Essler F H L and Latrémolière F T 1999 Nucl. Phys. B 556 411
[103] Akutsu Y, Akutsu N and Yamamoto T 2001 Phys. Rev. B 64 085415
[104] Suzuki M 1985 Phys. Rev. B 31 2957
Suzuki M and Betsuyaku H 1986 Phys. Rev. B 34 1829
[105] Suzuki M and Inoue M 1987 Prog. Theor. Phys. 78 787
[106] Koma T 1987 Prog. Theor. Phys. 78 1213
[107] Suzuki J, Akutsu Y and Wadati M 1990 J. Phys. Soc. Japan 59 1357
[108] Klümper A 1993 Z. Phys. B 91 507
[109] Shiroishi M and Takahashi M 2002 Phys. Rev. Lett. 89 117201
[110] Okunishi K, Akutsu Y, Akutsu N and Yamamoto T 2001 Phys. Rev. B 64 104432
[111] Deguchi T 2001 J. Phys. A: Math. Gen. 34 9755
[112] Kauffman L H 1991 Knots and Physics (Singapore: World Scientific)
[113] Wu F Y 1992 Rev. Mod. Phys. 64 1099

Copyright © 2003 IOP Publishing Ltd.

 PART II

QUANTUM SYSTEMS

Copyright © 2003 IOP Publishing Ltd.

 Chapter 6

Unifying approaches in integrable systems:

quantum and statistical, ultralocal and
non-ultralocal
Anjan Kundu
Saha Institute of Nuclear Physics, Calcutta, India

6.1 Introduction
By quantum integrable systems, we will mean systems with a sufficient number of
higher conserved quantities including the Hamiltonian of the model. Such a notion
of integrability in the Liouville sense allows the system to be described through
action-angle variables with the conserved quantities, which are now operators,
playing the role of action variables. For integrable systems, the conserved
quantities, being functionally independent, should form a commuting set of
operators [cn , cm ] = 0, n, m = 1, 2, . . . , N, such that their total number matches
with the degrees of freedom of the system. For example, a one-dimensional lattice
model of l-sites describing d-mode pseudoparticles should have the number of
conserved quantities N = dl. Note that, for spin- 12 chains, we have d = 1, while
spin-1 and electron models account for d = 2. In this review, we will stick to
single mode d = 1 systems for simplicity and consider mainly periodic lattice
models with N < ∞, where the algebraic structures can be seen in their exact
form. At the lattice constant → 0, the field models will be generated from their
exact lattice versions, whenever possible. Integrable field models with N → ∞
consequently need to have an infinite number of conservation laws.
Integrable systems, therefore, are restrictive systems with a very rich
symmetry. The beauty of such models is that they allow exact solutions for
the eigenvalue problem simultaneously for all conserved operators including
the Hamiltonian. Moreover, such one-dimensional quantum systems are also
related to the corresponding two-dimensional classical statistical models with a
fluctuating variable. Therefore, parallel to a quantum mechanical model, one can,

Copyright © 2003 IOP Publishing Ltd.

 in principle, also exactly solve a related vertex-type model on a two-dimensional
lattice using almost the same techniques and similar results [1]. Celebrated
examples of such interrelated integrable quantum and statistical systems are the
XYZ quantum spin- 12 chain and the eight-vertex statistical model, the XXZ spin
chain and the six-vertex model, spin-1 chain and the 19-vertex model etc.
To describe an integrable system with such an involved structure, naturally
one can no longer start from the Hamiltonian of the model as is customary
in physics, since now the Hamiltonian is merely one among many commuting
conserved charges. Certain abstractions which are formalized by the quantum
inverse scattering method and the algebraic Bethe ansatz, therefore, need to
be adopted (see [2, 3]). Though we could use the same language, we take
here a slightly different view point since we intend to describe integrable
systems belonging to both ultralocal and non-ultralocal classes. For an effective
description of integrable systems, it is convenient to define a generating function
called the transfer matrix τ (λ), which depends on some extra parameter λ known
as the spectral parameter, such that one can recover the infinite number of
conserved quantities
 as the expansion coefficients of τ (λ) or any function of it
like ln τ (λ) = j cj λj . The crucial integrability condition may then be defined
in a compact form as
[τ (λ), τ (µ)] = 0 (6.1)
from which the commutativity of the cj s follows immediately by comparing the
coefficients of the different powers of λ, µ.
However, for solving the eigenvalue problem as well as for identifying the
structure of the model, we require a more general matrix formulation, from where
the integrability condition may be derived. At the same time, we need to transit
from a global to a local description defined at each lattice point, where some
individual properties of a model are well expressed. At this local level, as we
see now, the differences between the ultralocal and the non-ultralocal models
become prominent. An integrable system allowing the necessary abstraction may
be represented by an unusual type of matrix called the Lax operator Laj (λ)
defined at each site j in a one-dimensional discretized lattice. The index a defines
the matrix or the auxiliary space, while j designates the quantum space. The
matrix elements of the Lax operator, unlike in usual matrices, are operators acting
on some Hilbert space. The models with Lax operators commuting at different
lattice sites:

Laj (λ)Lbk (µ) = Lbk (µ)Laj (λ) a = b, j = k (6.2)

are known as ultralocal models, while the integrable models for which
this ultralocality condition does not hold are classified as non-ultralocal
models. Note that in expressions like (6.2), different auxiliary spaces
mean different tensor products like L1j (λ) = Lj (λ) ⊗ I and L2j (µ) = I ⊗
Lj (µ). The ultralocal property (6.2) generally reflects the involvement of
canonical operators with commutation relations like [u(x), p(y)] = iδ(x − y) or

Copyright © 2003 IOP Publishing Ltd.

 [ψ(x), ψ † (y)] = δ(x − y) in the Lax operator giving a trivial commutator at
points x = y. In non-ultralocal models, however, the basic fields may be of non-
canonical type, e.g. [j1 (x), j1 (y)] = δx
(x − y), or derivatives of the canonical
fields may appear in their Lax operators violating the ultralocal condition and
bringing additional complexities, which might not always be resolved. For this
reason, the theory and application of non-ultralocal models are still in the process
of development and are far from completion. In spite of many important models
belonging to this class, it is rather disappointing to note that this category of
models has not received the required attention in the literature.

6.2 Integrable structures in ultralocal models

We focus first on the ultralocal systems due to their relative simplicity and
formulate a unifying scheme for generating such quantum and statistical
integrable models. For ensuring the integrability of an ultralocal model, it is
sufficient to impose a certain matrix commutation relation known as the quantum
Yang–Baxter equation (QYBE) on its representative Lax operator in the form

Rab (λ − µ)Laj (λ)Lbj (µ) = Lbj (µ)Laj (λ)Rab (λ − µ) (6.3)

defined at each lattice site j = 1, 2, . . . , N. This QYBE actually expresses the

commutation relations among different matrix elements of the L-operator, given
in a compact matrix form, where the structure constants are determined by the
spectral-parameter-dependent c-number elements of the R(λ − µ)-matrix. The
R-matrix, in turn, should satisfy a similar but simpler YBE:

Rab (λ − µ)Rac (λ − γ )Rbc (µ − γ ) = Rbc (µ − γ )Rac (λ − γ )Rab (λ − µ).

(6.4)
Since our intention is to establish the integrability which is a global property, we
have to switch from this local picture at each site j to a global one by defining a
matrix, known as the monodromy matrix:


N  
A(λ), B(λ)
Ta (λ) = Laj (λ) T (λ) ≡ . (6.5)
C(λ), D(λ)
j =1

Multiplying, therefore, the QYBE (6.3) for j = 1, 2, . . . , N and using the

ultralocality condition (6.2), thanks to which one can treat the objects at different
lattice points as commuting objects as in the classical case and drag Laj (λ)
through all Lbk (µ)s for k = j, b = a to arrive at the global QYBE

Rab (λ − µ)Ta (λ) Tb (µ) = Tb (µ)Ta (λ)Rab (λ − µ). (6.6)

Note that the local and global QYBEs have exactly the same structural form.
Invariance of the algebraic form for the tensor product of the algebras, as revealed
here, indicates the occurrence of the coproduct related to a deep Hopf algebra

Copyright © 2003 IOP Publishing Ltd.

 structure underlying all integrable systems [4]. We will see later that, for non-
ultralocal models, such a structure is slightly modified to include additional
braiding relations. For the periodic ultralocal models, further defining the transfer
matrix as τ (λ) = tra Ta (λ), taking the trace from both sides of the global YBE
(6.6) and cancelling the R-matrices due to the cyclic rotation of matrices under
the trace we reach finally for τ (λ) the trace identity (6.1) defining the quantum
integrability of the system. Therefore, we may conclude that the local QYBE
(6.1) in association with the ultralocality condition (6.2) is the sufficient condition
for quantum integrability of an ultralocal system. Consequently, we may define
such an integrable system by its representative Lax operator together with the
associated R-matrix satisfying these criteria. Note that we are concerned here
only with systems with periodic boundary conditions. For models with open
boundaries, the QYBE should, however, be modified with the inclusion of a
reflection matrix, which was introduced in detail in [5].

6.2.1 List of well-known ultralocal models

To have a concrete picture before us, we furnish a list of well-known ultralocal
models together with their L-operators and R-matrices. We will, however, for
simplicity restrict ourselves here only to the quantum models with 2 × 2 matrix
Lax operators associated with 4 × 4 R-matrices. We show in the next section
how these Lax operators can be generated in a systematic way confirming their
αβ
integrability. The Rγ δ -matrix that satisfies the YBE relation (6.2), with the indices
taking only the values 1, 2, can be given in a simple form by defining its non-
trivial elements [6]:
11 22 12 21 12 21
R11 = R22 = a(λ) R12 = R21 = b(λ) R21 = R12 = c. (6.7)

These elements may be expressed explicitly through trigonometric functions in

spectral parameters:

a(λ) = sin(λ + α) b(λ) = sin λ c = sin α (6.8)

or as its α → 0, λ → 0 limit, through rational functions:

a(λ) = λ + α b(λ) = λ c = α. (6.9)

Moreover, under a twisting transformation,

Fab (θ ) = eiθ(σa −σb )

3 3
R(λ) → R̃(λ, θ ) = F (θ )R(λ)F (θ ) with (6.10)

one gets the twisted trigonometric and rational R-matrix solution of (6.2), which
12 = b(λ) eiθ , R 21 = b(λ) e−iθ . Apart
may be given by (6.7) with the difference R12 21
from these R-matrices, there can be an elliptic R-matrix solution, for example
that related to the XYZ spin chain and the eight-vertex model [1]. All the
models we consider here, however, are associated with trigonometric or rational

Copyright © 2003 IOP Publishing Ltd.

 R-matrices and in the list presented here we group them accordingly, denoting
the Hamiltonian by H and the Lax operator related to the field (lattice) models by
L (Ln ).

6.2.1.1 Models associated with the trigonometric R-matrix (q = eiα , ξ = eiλ )

(i) Field models.

(1) Sine-Gordon model [7]:

m2
ut t − uxx =
sin(αu)
 α  (6.11)
ip m sin(λ − αu)
L(λ) = p = u̇.
m sin(λ + αu) −ip

(2) Liouville model [8]:

ut t − uxx = eiαu
 
p ξ eiαu
L(ξ ) = i  1 iαu  (6.12)
e −p
ξ
[u(x), p(y)] = iδ(x − y).

(3) A derivative nonlinear Schrödinger (DNLS) model [9]:

iψt − ψxx + 4iψ † ψψx = 0

 1 2 
− 4 ξ + k− N ξ ψ†
L(ξ ) = i (6.13)
4 ξ − k+ N
1 2
ξψ
N = ψ†ψ [ψ(x), ψ † (y)] = δ(x − y).

(4) Massive Thirring (bosonic) model (MTM) [6]:

H = dx [−iψ̂ † (σ 3 ∂x + σ 2 )ψ̂ + 2ψ (1)† ψ (2)† ψ (2) ψ (1) ]

ψ̂ † = (ψ (1)† , ψ (2)† ) [ψ (a) (x), ψ †(b) (y)] = δab δ(x − y)

 1 
f + (ξ, N (a) ) ξ ψ (1)† + ψ (2)†
 ξ  (6.14)
L(ξ ) = i  1 
ξ ψ (1) + ψ (2) f − (ξ, N (a) )
ξ
   
± 1 1
f (ξ, N ) = ±
(a)
− ξ + k∓ N − k± N
2 (1) (2)
.
4 ξ2

Copyright © 2003 IOP Publishing Ltd.

 (ii) Lattice models.
(1) Anisotropic XXZ spin chain [10]:

N
H= σn1 σn+1
1
+ σn2 σn+1
2
+ cos ασn3 σn+1
3
n (6.15)
Ln (ξ ) = sin(λ + ασ 3 σn3 ) + sin α(σ + σn− + σ − σn+ ).
(2) Lattice sine-Gordon model [11]:
 
g(un ) eipn m sin(λ − αun )
Ln (λ) =
m sin(λ + αun ) e−ipn g(un ) (6.16)
g (un ) = 1 + m cos α(2un + 1).
2 2 2

(3) Lattice Liouville model [8]:

 
eipn f (un ) ξ eiαun
Ln (ξ ) =  iαun 
e f (un ) e−ipn (6.17)
ξ
f 2 (un ) = 1 + 2 eiα(2un+1) .
(4) Lattice DNLS model [12]:
1 
q −Nn − iξ q Nn +1 κA†n
ξ 
Ln (ξ ) =  1 Nn 
−(Nn +1)
κAn q + iξ q (6.18)
ξ
cos α(2Nn + 1)
[An , A†m ] = δnm , An -q boson.
cos α
(5) Lattice MTM [13]:
Exact lattice version of MTM (6.14).
Lax operator: Ln = L(1) (2)
n L̃n (each factor is a realization of (6.18) for a
bosonic mode).
(6) Discrete-time or relativistic quantum Toda chain [14]:

H= (cosh 2αpi + α 2 cosh α(pi + pi+1 ) e(ui −ui+1 ) )
i
 
1 αpn (6.19)
e − ξ e−αpn α eun 
Ln (ξ ) =  ξ .
−α e−un 0

6.2.1.1a Models associated with the twisted trigonometric R-matrix

(6a) Quantum Suris discrete-time Toda chain [14, 15]:
 
1 2αpk
e − ξ α e uk
Lk (ξ ) =  ξ . (6.20)
−α e 2αp k −u k 0

Copyright © 2003 IOP Publishing Ltd.

 (7) Ablowitz–Ladik model [6, 16]:

ibj,t + (1 + αbj† bj )(bj +1 + bj −1 ) = 0

 
1 † (6.21)
bk 
Lk (ξ ) =  ξ [bk , bl†] = δkl (1 − bk† bk ).
bk ξ

6.2.1.2 Models associated with the rational R-matrix

(i) Field models.
(1) Nonlinear Schrödinger equation (NLS):
 
λ ψ
iψt + ψxx + (ψ † ψ)ψ = 0 L(λ) = . (6.22)
ψ† −λ
(ii) Lattice models.
(1) Isotropic XXX spin chain [10]:

N
H= σ'n · σ'n+1 Lan (λ) = λI + αPan
n (6.23)
1
Pan = (I + σ'a · σ'n ).
2
(2) Gaudin model [18]: In the simplest case the Hamiltonians

N
1
Hk = σk · σ'l ),
(' k = 1, 2, . . . , N
l=k
k − l (6.24)
Lak (λ) = (λ − k )I + αPak .
(3) Lattice NLS model [11]:
 1 1

λ + s − Nn 2 (2s − Nn ) 2 ψn†
Ln (λ) = 1 1
ψ(2s − Nn ) 2
2 λ − s + δNn (6.25)
Nn = ψn† ψn [ψk , ψl† ] = δkl .
(4) Simple lattice NLS [19]:
 
λ + s − Nn ψn†
Ln (λ) = . (6.26)
ψn −1
(5) Discrete self-trapping dimer model [20]:

 
1 2
H =− (sa − N ) + (ψ ψ + ψ ψ )
(a) 2 †(1) (2) †(2) (1)
2 a (6.27)
[ψ (a)
,ψ †(b)
] = δab a, b = 1, 2.

Copyright © 2003 IOP Publishing Ltd.

 Lax operator L(λ) = L(1) (λ)L(2) (λ) (each factor as (6.26) for each of
two bosonic modes).
(6) Toda chain (non-relativistic) [6]:
 1  
pn − λ eun

H= pi2 + e(ui −ui+1 ) Ln (λ) = . (6.28)
2 −e−un 0
i

6.3 Unifying algebraic approach in ultralocal models

Though the QYBE itself represents a unifying approach for all ultralocal models,
we intend to specify here a common algebraic structure independent of the
spectral parameter that will not only systematize the models including those
previously listed but also identify their common integrable origin, establishing
naturally the quantum integrability of all of them, simultaneously. From the
previous list of models, one may observe that different integrable models have
their representative Lax operators in diverse forms with a varied dependence on
the spectral parameter as well as on the basic operators like the spin, bosonic
or the canonical operators. However, the R-matrices associated with all of them
are given by the same form (6.7) with known trigonometric (6.8) or its limiting
rational (6.9) solutions. To explain this intriguing observation, we may look for
a common origin for the Lax operators linked with a general underlying algebra
free from spectral parameters, though derivable from the QYBE. We propose to
take the Lax operator of such an ancestor model in the form [17]
 
ξ c1+ eiαS + ξ −1 c1− e−iαS + S −
3 3
(anc)
Ltrig (ξ ) =
− S + ξ c2+ e−iαS + ξ −1 c2− eiαS
3 3
(6.29)
ξ = eiαλ ± = 2 sin αξ ±1

where S' and ca± , a = 1, 2 are some operators, the algebraic properties of
which are specified later. The structure of (6.29) becomes clearer if we notice
the decomposition Ltrig (ξ ) = ξ L+ + ξ −1 L− , where L± are spectral-parameter
(anc)

ξ -independent upper and lower triangular matrices similar to the construction

in [4]. Inserting (6.29) in the QYBE together with its associated R-matrix (6.7)
with trigonometric solution (6.8) and matching the different powers of ξ , we
obtain the underlying general algebra as

[S 3 , S ± ] = ±S ±
1
[S + , S − ] = (M + sin(2αS 3 ) + M − cos(2αS 3 )) (6.30)
sin α
[M ± , ·] = 0
√
with M ± = ± 12 ±1(c1+ c2− ± c1− c2+ ) behaving as central elements with arbitrary
values of cs. As we have previously mentioned, the integrable systems are

Copyright © 2003 IOP Publishing Ltd.

 associated with an important Hopf algebra A, exhibiting the properties like
(1) coproduct (x) : A → A ⊗ A; (2) antipode or ‘inverse’ S : A → A; (3) counit
 : A → k; (4) multiplication M : A ⊗ A → A; and (5) unit α : k → A. It can be
shown that all these properties hold also for (6.30) defining it as a Hopf algebra.
Referring the interested readers to the original works [21] for a more mathematical
treatment of the non-cocommutative Hopf algebra, we give here only some simple
and intuitive arguments in its constructions. For example, the coproduct (x),
the most important of these characteristics, can be derived for algebra (6.30) by
exploiting a QYBE property that the product of two Lax operators Laj Laj +1 is
again a solution of the QYBE and may be given in explicit form as

(S + ) = c1+ eiαS ⊗ S + + S + ⊗ c2+ e−iαS

3 3

(S − ) = c2− eiαS ⊗ S − + S − ⊗ c1− e−iαS

3 3
(6.31)
(S 3 ) = I ⊗ S 3 + S 3 ⊗ I (ci± ) = ci± ⊗ ci± .
The multiplication property mentioned earlier is also in agreement with the
ultralocality condition, which is used for the transition from local to global QYBE
following a multiplication such as
(A ⊗ B)(C ⊗ D) = (AC ⊗ BD) (6.32)
with A = Li (λ), B = Li (µ), C = Li+1 (λ), D = Li+1 (µ). Note that (6.30) is a
q-deformed algebra and a generalization of the well-known quantum algebra [21]
Uq (su(2)).
In fact, different choices of the central elements ca± reduce this algebra to
the q-spin, q-boson as well as various other q-deformed algebras along with
their undeformed limits. Therefore, we can easily obtain the coproduct for these
algebras, whenever admissible, from their general form (6.31) in a systematic way
by taking the corresponding values of cs.

6.3.1 Generation of models

We know that the well-known integrable models listed earlier were discovered at
different points of time, mostly in an isolated way and generally by quantization
of the existing classical models. However, as we will see, they can actually
be generated in a systematic way through various realizations of the same Lax
operator (6.29) giving a unifying picture of integrable ultralocal models. For this
we first find a representation of (6.30) such as
S3 = u S + = e−ip g(u) S − = g(u) eip (6.33)
in physical variables with [u, p] = i, where the operator function
g(u) = (κ + sin α(s − u)(M + sin α(u + s + 1)
1
+ M − cos α(u + s + 1))) 2
1
(6.34)
sin α

Copyright © 2003 IOP Publishing Ltd.

 containing free parameters κ and s. We demonstrate now that the Lax operator
(6.29), which represents a generalized lattice SG-like model for (6.33) may serve
as an ancestor model (with possible realizations in other physical variables such as
the bosonic ψ, ψ † or spin s ± , s 3 operators) for generating all integrable ultralocal
quantum as well as statistical systems. As an added advantage, the Lax operators
of these models are derived automatically from (6.29), while the R-matrix is
simply inherited. The underlying algebras of the models are also given by the
corresponding representations of the ancestor algebra (6.30), which, being a direct
consequence of the QYBE, ensures the quantum integrability of all its descendent
models that we construct here. It should be stressed that due to the symmetry of
solution (6.7)—[R(λ − µ), σ a ⊗ σ a ] = 0, a = 1, 2, 3—the Lax operator (6.29)
as a solution of QYBE may be right or left multiplied by any σ a . We shall use
this freedom in our following constructions, whenever needed.
Note that we may also generate the quantum field models by taking the
continuum limit of their lattice variants with the lattice spacing → 0 properly.
Though, in general, such transitions to the field limit might be tricky and
problematic, we suppose their validity by assuming the lattice operators go
smoothly to the field operators pj → p(x), ψj → ψ(x), with the corresponding
commutators: [ψj , ψk† ] = (1/ )δj k → [ψ(x), ψ(y)] = δ(x − y) etc. The lattice
Lax operator, therefore, should reduce to its field counterpart L(x, λ) as Lj (λ) →
I + i L(x, λ) + O( 2 ). The associated R-matrix, however, remains the same,
since it does not contain the lattice constant . Thus integrable field models like
the sine-Gordon, Liouville, NLS or the derivative NLS models can be recovered
from their exact lattice versions and with the same quantum R-matrix, though all
discrete models may not always have such a direct field limit.

6.3.1.1 Models belonging to the trigonometric class

(1) Choosing all central elements simply as ca± = 1, a = 1, 2, which gives
M− = 0, M + = 1, (6.30) reduces clearly to the well-known quantum algebra
Uq (su(2)) [21] given by

[S 3 , S ± ] = ±S ± [S + , S − ] = [2S 3 ]q (6.35)

with the known form of its coproduct recovered easily from (6.31). The simplest
representation S' = 12 σ' for this case derives from (6.29) the integrable XXZ spin
chain (6.15). On the other hand, representation (6.33) with the corresponding
reduction of (6.34) as
1 1
g(u) = [1 + cos α(2u + 1)] 2
2 sin α
with a suitable choice of parameters s, κ recovers the Lax operator of the lattice
sine-Gordon model (6.16) directly from (6.29) and at its field limit the field Lax
operator (6.11). Note that the spectral dependence in ± appearing in (6.29) can

Copyright © 2003 IOP Publishing Ltd.

 be easily removed through a simple gauge transformation [4] and, therefore, we
ignore them in our construction and use the freedom of translational symmetry of
the spectral parameters λ → λ + constant, whenever needed.
(2) An unusual exponentially deformed algebra can be generated from
(6.30) by fixing the elements as c1+ = c2− = 1, c1− = c2+ = 0, which gives M ± =
√
± 12 ±1 and
3
± ± + − e2iαS
[S , S ] = ±S
3
[S , S ] = (6.36)
2i sin α
and reduces (6.34) to
1
(1 + eiα(2u+1)) 2
g(u) = √ .
2 sin α
This algebra and its corresponding realization yields clearly from (6.29) the Lax
operator of the lattice Liouville model (6.17) and at its field limit that of the
Liouville field model (6.12).
It is interesting to observe here that although the underlying algebraic
structure and, hence, its realization are fixed by the choice of M ± , the Lax
operator (6.29) depends explicitly on the set of cs and, therefore, may take
different forms for the same model. For example, in the present case c1− = 0 would
again give the same value for M ± but a different Liouville Lax operator [22]
which is more convenient for the Bethe ansatz solution.
This, therefore, opens up, interesting possibilities for obtaining systemati-
cally different useful Lax operators for the same integrable model, as well as for
constructing new non-ultralocal models [23].
(3) Recall that the well-known q-bosonic algebra may be given by [24]
[A, N] = A, [A† , N] = −A† , AA† − q −2 A† A = q 2N or in its conjugate form
with q → q −1 . Combining these two forms, we can easily write the commutator
of such q-bosons:
cos(α(2N + 1))
[A, N] = A [A† , N] = −A† [A, A† ] = . (6.37)
cos α
It is interesting to find that for the choice of the central elements c1+ = c2+ = 1,
c1− = −iq, c2− = i/q compatible with M + = 2 sin α, M − = 2i cos α, we may get
a realization

S + = −κA S − = κA†
1
S 3 = −N κ = −i(cot α) 2 (6.38)

with (6.31) reducing directly to the relation (6.37), which thus gives a new
integrable q-boson model. It is important to note now that either using (6.33)
which simplifies (6.34 ) to g 2 (u) = [−2u]q or directly taking the mapping of the
q-bosons to standard bosons:
 1
[2N]q 2
A=ψ N = ψ † ψ,
2N cos α

Copyright © 2003 IOP Publishing Ltd.

 we may convert (6.29) with (6.38) to an exact lattice version of the quantum
derivative nonlinear Schrödinger (QDNLS) equation (6.18) and, consequently, to
the QDNLS field model (6.13). The QDNLS is related also to the interacting Bose
gas with a derivative δ-function potential [25].
(4) Since the matrix product of Lax operators with each factor representing
different Lax operator realization for the same model should again give a QYBE
solution, we can construct multi-mode integrable extensions by taking the product
of single-mode Lax operators. Using this trick, i.e. by combining the two
QDNLS models constructed earlier as L(c1± , c2± , ψ (1) )L(c2∓ , c1∓ , ψ (2) ) = L(λ),
we can create an integrable exact lattice version of the massive Thirring model
[13]. At the continuum limit, it goes to the bosonic massive Thirring model
introduced in [6], the field Lax operator (6.14) of which can be given simply
by the superposition L = L(1) (ξ, k± , ψ (1) ) + σ 3 L(2) (1/ξ, k∓ , ψ (2) )σ 3 , where
1 ± ik± sin α = e±iα/2 and the constituing operators L(a) is given clearly by the
DNLS Lax operator (6.13) for each of its two bosonic modes.
(5) Since the general algebra permits trivial eigenvalues for central elements,
one may choose both M ± = 0, which might correspond to different sets of
choices, e.g. (i) ca+ = 1, a = 1, 2; (ii) ca− = 1, a = 1, 2; (iii) c1∓ = ±1; or (iv) c1+ =
1, with the rest of the cs being zeros. It is easy to see that all of these sets lead to
the same underlying algebra:

[S + , S − ] = 0 [S 3 , S ± ] = ±S ± (6.39)

though generating different Lax operators from (6.29).

As here, (6.34) gives simply g(u) = constant, canonically interchanging
u → −ip, p → −iu, from (6.33), one gets

S 3 = −ip S ± = α e∓u (6.40)

which evidently generates from the same general Lax operator (6.29) the discrete-
time or relativistic quantum Toda chain (6.19). Note that, (iii) and (iv) give the
two different Lax operators found in [14] and [26] for the relativistic Toda chain.
Cases (i) and (ii), however, could be used for constructing non-ultralocal quantum
models, namely the light-cone SG and the mKdV models [23].

6.3.1.1a Models in the twisted trigonometric class

Under twisting, when the R-matrix changes as in (6.10) the associated Lax
operator is similarly transformed: Ln (λ) → L̃n (λ, θ ) = Fn (θ )Ln (λ)Fn (θ ), with
Fan (θ ) = eiθ(σa −Sn ) . As a result, the ancestor model (6.29) associated with the
3 3

trigonometric twisted R-matrix (6.10) is deformed with a change in its operator

elements:

ca± → ca± e−iθSk Sk± → S̃k± = e−i 2 θSk Sk± e−i 2 θSk
3 1 3 1 3
(6.41)

Copyright © 2003 IOP Publishing Ltd.

 and, as a consequence, the diagonal elements of the twisted Lax operator take the
3
form ei(θ±α)Sk , with obvious preference for the choice θ = ±α.
(6) We may generate the quantum analogue of the Suris discrete-time Toda
chain belonging to the twisted class by starting from the ancestor model with
the change (6.41) but by fixing the parameter θ = −α (an equivalent model is
obtained by the choice θ = α). Using the same realization (6.40) we arrive now at
the explicit form (6.20).
(7) However if we start from the same twisted ancestor model with the same
value θ = −α for the twisting parameter but take the central elements as c1+ =
√
c2− = 0 with c1− = c2+ = 1 giving M ± = 12 ±1 (compare this with the Liouville
case!), all non-commuting operators clearly vanish from the diagonal elements
of the resulting Lax operator. Moreover, renaming the deformed operators as
bk = 2 sin α S̃k+ , we get their modified algebra as a type of q-boson—[bk , bl† ] =
δkl (1 − bk† bk )—and, thus, finally generate the exact form of the Ablowitz–Ladik
model (6.21).
The domain of the models considered can, therefore, be considerably
extended if we use twisting or some other allowed transformations [27] that
preserve integrability.

6.3.1.2 Models belonging to the rational class

One of the crucial parameters built into both the R-matrix and the previous Lax
operators is the deformation parameter q = eiα , the physical meaning of which
is as anisotropic or relativistic parameter. We now consider the undeformed limit
q → 1 or α → 0 related to isotropic or non-relativistic models belonging to the
rational class, which reduces various α-dependent objects as S ± → is ± , {ca± } →
{cai }, M + → −m+ , M − → −αm− , ξ → 1 + iαλ. This transforms (6.30) to a
q-independent algebra with

[s + , s − ] = 2m+ s 3 + m− [s 3 , s ± ] = ±s ± (6.42)

where m+ = c10 c20 , m− = c11 c20 + c10 c21 and cai , i = 0, 1 are central to (6.42). Note
that (6.42) is a generalization of spin as well as the bosonic algebra and its
coproduct can be obtained as a limit of (6.31). Consequently, the general Lax
operator (6.29) is converted into
 0 
(anc) c1 (λ + s 3 ) + c11 s−
Lrat (λ) = (6.43)
s+ c20 (λ − s 3 ) − c21

and the quantum R-matrix (6.7) is reduced to its rational form (6.9).
We would see that the ultralocal integrable systems belonging to the rational
class can be generated in a similar way now from the Lax operator (6.43) with
algebra (6.42), all sharing the same rational R-matrix (6.9). It is not difficult
to check by a variable change (u, p) → (ψ, ψ † ) that at the limit α → 0 (6.33)

Copyright © 2003 IOP Publishing Ltd.

 reduces to a generalized Holstein–Primakov transformation (HPT):
s3 = s − N s + = g0 (N)ψ s − = ψ † g0 (N)
(6.44)
g02 (N) = m− + m+ (2s − N) N = ψ †ψ
which is also an exact realization of (6.42). Therefore, Lax operator (6.43) with
such a realization may be considered to be a generalized lattice NLS, which serves
as a generating model for all quantum integrable models belonging to the rational
class.
(1) The choice m+ = 1, m− = 0 clearly reduces (6.42) to the su(2) algebra
[s + , s − ] = 2s 3 , [s 3 , s ± ] = ±s ± . A compatible choice ca0 = 1, ca1 = 0 yields
from (6.43) for the spin- 12 representation the Lax operator of the XXX spin
chain (6.23).
Taking spin- 12 and spin-1 realizations alternatively along the lattice we can
now construct the integrable alternate spin model discovered in [29].
Note that a slightly different choice c10 = −c20 = 1, ca1 = 0, giving m+ =
−1, m− = 0, generates the corresponding model with the su(1, 1) algebra.
The bosonic realization (6.44) in the present cases with m+ = ±1, m− =
0 is simplified to the standard HPT with g02 (N) = ±(2s − ψ † ψ), which
reproduces from (6.43) the exact lattice NLS model (6.25) and at the
continuum limit the more familiar NLS field model (6.22), with the +(−)
sign in the HPT corresponding to the attractive (repulsive) interaction.
(2) A complementary choice m+ = 0, m− = 1, however, converts (6.44) to
s + = ψ, s − = ψ † , s 3 = s − N due to g0 (N) = 1 and reduces (6.42) directly
to the standard bosonic relations [ψ, N] = ψ, [ψ † , N] = −ψ † , [ψ, ψ † ] = 1.
Remarkably, (6.43) with this realization generates yet another simple lattice
NLS model with Lax operator (6.26).
(3) Combining two such bosonic Lax operators (6.26), constructed earlier—
L(1) (λ)L(2) (λ) = L(λ)—and considering them to be inserted at a single
site, we can construct the Lax operator of an integrable model involving
two-bosonic modes, which yields the quantum discrete self-trapping model
(6.27).
(4) Note that the trivial choice m± = 0 again gives algebra (6.39) and, hence, the
realization (6.40). This, however, yields from (6.43) the Lax operator of the
non-relativistic Toda chain (6.28) associated with the rational R-matrix. It is
interesting to note that in [30] Lax operators like (6.29) and (6.43) appeared
in their bosonic realization and were shown to be the most general possible
form within their respective class.
Therefore, these quantum Lax operators are at the core of ultralocal
integrable models, both discrete and continuum ones, which can be
constructed from them in a unified way. Models belonging to the
trigonometric and rational classes are generated from (6.29) and its limiting
form (6.43) respectively, and, therefore, inherit the same corresponding
R-matrices (6.8) and (6.9).

Copyright © 2003 IOP Publishing Ltd.

 6.3.2 Fundamental and regular models
The Lax operator Laj (λ), in general, acts on the product space Va ⊗ hj ,
of the common auxiliary space Va and the quantum space hj at site j .
The models with Va isomorphic to all hj , j = 1, . . . , N, and given by the
fundamental representation are called fundamental models. For such models,
the finite-dimensional matrix representations of auxiliary and quantum spaces
become equivalent and may lead to Lal (λ) ≡ Ral (λ). However, for clarifying
a misconception prevalent in the literature, we should stress that a model is
represented by its Lax operator only, the associated R-matrix accounts for
the commutation property of the elements of this Lax operator through the
QYBE. Therefore, even for a fundamental model, the Lax operator may differ
from its R-matrix. We may, however, demand for the fundamental models a
useful additional property: Lal (0) = Pal , known as the regularity condition given
through the permutation operator, which may be expressed in the general case as

an n2 × n2 matrix Pal(n) = nβα=1 Eαβ a El , E
βα αβ being a matrix with its (α, β)
element as 1 and the rest 0.
Recall that the global operator τ (λ) is constructed from the local Lax
operators Laj (λ), j = 1, . . . , N, as τ (λ) = tra (La1 (λ) . . . LaN (λ)), where the
transfer matrix τ (λ) acts on the total quantum space H = ⊗N j =1 hj . Therefore,
from a knowledge of Lax operators, it is possible to derive all conserved
quantities including the Hamiltonian, which, in general, would be non-local
objects. The regularity condition on Lax operators, however, allows this difficulty
to be overcome and Hamiltonians with nearest-neighbour (NN) interactions to be
obtained. Let us pay some special attention to this specific group of models since,
as we will see later, the most important integrable models applicable to the physics
of condensed matter problems are given by the regular models with n = 2, 3, 4
etc. For this reason, though our main concern in this chapter is the 2 × 2 auxiliary
matrix space, we describe here the more general n cases and demonstrate that the
Hamiltonians of such different physical models interestingly have similar forms
when expressed through the permutation operator Pjj +1 . Such a permutation
operator exhibits a space-interchanging property Paj Lak = Lj k Paj , along with
P 2 = 1 and tra (Paj ) = 1. For all regular and periodic models, using the freedom
of cyclic rotation of matrices under the trace, we can express the transfer matrix as
τ (0) = tra (Paj Paj +1 · · · PaN Pa1 · · · Paj −1 )
= (Pjj +1 · · · Pj N Pj 1 · · · Pjj −1 ) tra (Paj ) (6.45)
for any j and, as its derivative with respect to λ, we similarly get

N
τ
(0) = tra (Paj L
aj +1 (0) · · · PaN Pa1 · · · Paj −1 )
j =1

N
= (L
jj +1 (0) · · · Pj N Pj 1 · · · Pjj −1 ) tra (Paj ) (6.46)
j =1

Copyright © 2003 IOP Publishing Ltd.

 where the periodic boundary condition LaN+j = Laj is assumed. Defining
now H = c1 = dλ d
ln τ (λ)|λ=0 = τ
(0)τ −1 (0) and using (6.45), (6.46), we may
construct the related Hamiltonian:

N
H= L
jj +1 (0)Pjj +1 (6.47)
j =1

with only NN interactions, where all non-local factors are cancelled out due to
the relevant properties of the permutation operator. Similarly taking the higher
derivatives of ln τ (λ), the higher conserved quantities cj , j = 2, 3, . . . , N, can
be constructed for these regular models. Note that the conserved operator cj
involves interactions with j + 1 neighbours.
For the simplest case of n = 2, we may take the Lax operator as the R-matrix
given by (6.7), which satisfies clearly the regularity condition L(0) ≡ R(0) = P
for both trigonometric and rational cases. Moreover, for (6.8), the part L
jj +1 (0)
in (6.47) introduces anisotropy reproducing the Hamiltonian of the XXZ spin
chain (6.15). However, since for the rational case (6.9) L
jj +1 (0) = 1, using the
(2)  j j +1
expression Pjj +1 = 2βα=1 Eαβ Eβα ≡ 12 (Hjj σ
+1 + 1), where Hjj +1 = σ
σ 'j σ'j +1 ,

(6.47) is clearly reduced to the isotropic spin- 2 Hamiltonian H = j Hjj
1 σ σ
+1
(6.23).
It is intriguing to note that, for the rational class, the same form of the
 (n)
Hamiltonian H = N j =1 Pjj +1 with several higher values of n describes most of
the important integrable models, though their physical forms are given mainly
through various representations of the permutation operator. For example, for
n = 3 corresponding to the SU (3) group, we can express the permutation operator
3 j +1
(3) j '
Pjj +1 = βα=1 Eαβ Eβα through spin-1 operators S giving a variant of the
integrable spin-1 model:

H= S'j S'j +1 + (S'j S'j +1 )2 with  = +1. (6.48)
j

Considering a supersymmetric invariant gl(1, 2) case, i.e. realizing the corre-

(1,2) †
sponding graded permutation operator Pjj +1 using fermionic (caj , caj ), a =↑, ↓
'
and spin S operators, we may again construct, from the Hamiltonian density
(1,2)
+1 = (2Pjj +1 − 1), the well-known integrable t–J model [31]:
tJ
Hjj

    
†
H tJ
= tJ
Hjj +1 = −tP cσj cσj +1 + H.C. P
j j σ =↑↓
 
1
+ J Sj Sj +1 − nj nj +1 + nj + nj +1 (6.49)
4

with J = 2t = 2, where P projects out the double occupancy states.

Copyright © 2003 IOP Publishing Ltd.

 A different four-dimensional realization of the fermion operators, in contrast,
converts the same Hamiltonian to an integrable correlated electron model
proposed in [32].
(4) (2) (2)
Similarly for n = 4, i.e. for SU (4), realizing Pjj +1 = Pjj +1 ⊗ Pjj +1 in the
factorized form, we can get
 (σ ) (τ )
H= Pjj +1 ⊗ Pjj +1
j
1 σ
= (Hjj +1 + 1)(Hjj
τ
+1 + 1)
4 j
1 
= [H σ + H τ + σ
(Hjj +1 Hjj +1 + 1)]
τ
(6.50)
4 j

with H σ,τ representing isotropic spin- 12

Hamiltonians (6.23). If we now add the
interaction along the rung—Hrung = J j σ'j τ'j , with [H, Hrung] = 0 to (6.50),
where σ, τ represent the spins along two legs of the ladder—we may construct a
model which is nothing but the integrable spin- 12 ladder discovered recently [33].
However, from the same form of Hamiltonian but by considering a
(2,2)
supersymmetric extension SU (2, 2), we may realize Pjj +1 again through fermion
†
operators (caj , caj ), a =↑, ↓, to construct an integrable extension of the Hubbard
model proposed in [34].
One can repeat this construction of the spin–ladder model to generate
the integrable t–J ladder model, introduced in [35], which would, therefore,
correspond to a similar construction in n = 6 with Hamiltonian
 (2,4)
 (1,2) (1,2) 1  (1)t J (2)t J
H= Pjj +1 = Pjj +1 ⊗ Pjj +1 = (Hjj +1 + 1)(Hjj +1 + 1)
j j
4 j
1  (1)t J (2)t J
= [H (1)t J + H (2)t J + (Hjj +1 Hjj +1 + 1)] (6.51)
4 j

where H (a)t J , a = 1, 2 are t–J Hamiltonians (6.49) along two legs. Adding a
suitable Hrung to (6.51) with [H, Hrung] = 0, defining the interaction along the
rung, we finally obtain the integrable t–J ladder model.
Apart from these applications of integrable systems with a similar structure
from the algebraic point of view, we should mention some other important models
like the Hubbard model and the Kondo problem, which also fall into the class of
exactly solvable problems in one dimension [36]. Employing further twisting and
gauge transformations on multi-fermion or multi-spin integrable models, one can
generate another type of integrable model of current interest [37]. The importance
of solvable models in physical systems, their relevance to experiments and related
issues are discussed in [38]. For detailed and involved applications of the Bethe
ansatz technique including that for the theory of correlation functions to various

Copyright © 2003 IOP Publishing Ltd.

 integrable systems like the δ-bose gas, NLS, sine-Gordon etc, readers are referred
to [3].

6.3.3 Fusion method

We have constructed spin and boson (including q-spin and q-boson) models
through the realization of particular Lax operators with inequivalent auxiliary
and quantum spaces. However, in the case of finite-dimensional higher-rank spin
representations, there exists an intriguing method, known as the fusion method,
for obtaining higher-spin models by fusing elementary R-matrices such as (6.7).
Thus, by fusion of only the quantum spaces, one can construct spin-s Lax
operators with a spin- 12 auxiliary space, which is also directly obtainable from
(6.29) as a particular realization. Fusing the auxiliary spaces further, the higher-
spin Lax operator with a spin-s auxiliary space may be constructed as


s 
s
Lab = (Pa+ ⊗ Pb+ ) Raj bk (λ + iα(2s − k − j ))(Pa+ ⊗ Pb+ )g2s (λ)
j =1 k=1
(6.52)
+
with Pa(b) as the symmetrizer in the fused spin-s space a(b) and g2s some
normalizing factor [39]. For the rational R-matrix corresponding to (6.9), one
obtains, from (6.52), the integrable spin-s Babujian–Takhtajan model [39] which,
for s = 1, may be given in the same form as Hamiltonian (6.48) but with  = −1.
Similarly, for the trigonometric case (6.8), the fused model would correspond to
an integrable anisotropic higher-spin chain.
It should, however, be stressed that such a fusion technique, as far as we
know, has not yet been formulated for bosonic and q-bosonic models. Such an
extension, at least for the restricted values of q, therefore, needs more attention.

6.3.4 Construction of classical models

The systematic procedure for constructing quantum integrable models as various

reductions of the same ancestor model, as described here, is also, applicable
naturally to the corresponding classical models by taking the classical limit

→ 0. At this limit, all field operators would be transformed to ordinary functions
with their commutators reducing to the Poisson brackets. Note also that the
parameter α appearing in the R-matrix is actually scaled as
α, which yields
the classical r-matrix R(λ) = I +
r(λ) + O(
2 ) and reduces QYBE (6.1) to its
classical limit {Lai (λ), Lbj (µ)} = δij [rab (λ − µ), Lai (λ)Lbj (µ)]. The classical
Lax operator reduced from (6.29) would, however, remain almost in the same
form, though the corresponding quantum algebras would change into their
corresponding Poisson algebras. This aspect of classical integrable systems is
given in great detail in the excellent monograph [40]. Using these classical
analogues of quantum systems, one can, therefore, also apply the algebraic

Copyright © 2003 IOP Publishing Ltd.

 scheme formulated here for generating quantum models in the classical context
and systematically construct the corresponding classical integrable models [41].

6.4 Integrable statistical systems: vertex models

D-dimensional quantum systems are known to be related to (1 + D)-dimensional
classical statistical models, which is naturally also true for D = 1, where the
integrability of models might be manifested. Interestingly, the integrable quantum
spin chain and the corresponding vertex model share the same quantum R-matrix
and have the same representation for the transfer matrix, the commutativity of
which ([τ (λ), τ (µ)] = 0) guarantees their integrability. However, while the spin
chain Hamiltonian Hs is expressed through the transfer matrix as ln τ (λ) =
I + λHs + O(λ2 ), the partition function Z of the vertex model is constructed
from τ (λ) as Z = tr(τ (λ)M ). The known integrable vertex models are usually
related to the quantum fundamental models described earlier.
In conventional vertex models, each bond connecting N × M arrays in a
two-dimensional lattice can take n different possible random values with certain
probabilities, which for a configuration i, j ; k, l of bonds meeting at each vertex
point is given by the Boltzmann weights wij,kl . These Boltzmann weights may
ij
be assigned as matrix elements wij,kl (λ) = Rkl (λ) of an R-matrix (though it
might be of a more general L-operator, as we will see later), which for integrable
models must satisfy the Yang–Baxter equation (6.1) and correspond to a quantum
integrable
 model.(The partition function of these vertex models may be expressed
as Z = config a,b,j,k ωa,j ;b,k (λ).
The simplest among the vertex models for n = 2 is the six-vertex model [1],
which corresponds to the XXZ spin chain and may be defined on a square lattice
with a random direction on each bond (left or right on the horizontal, up or down
on the vertical), constrained by the ice rule, by which the number of incoming
and outgoing arrows at each vertex have to be the same. This leaves only six
possible configurations and the corresponding Boltzmann weights may be given
by six non-trivial matrix elements of the R-matrix (6.7) with (6.8). It is fascinating
that this model may describe the possible configurations of hydrogen (H) ions
around oxygen (O) atoms in an ice crystal having two different (close–removed)
positions of the H-ions relative to the O-atom in the H-bonding, while the ice rule
corresponds to the charge neutrality of the water molecule.
A more general six-vertex model may be obtained if instead we assign its
Boltzmann weights directly to the spin- 12 matrix representation of the general
ancestor Lax operator (6.29). The parameters c1+ = −c1− = ρ+ , c2+ = −c2− = ρ−
present in the Lax operator may be combined to serve as the horizontal h ∼
ln ρ+ ρ− and vertical v ∼ ln ρ+ /ρ− fields acting on the model, which amazingly
recovers the most general six-vertex model proposed many years ago [42] through
a different construction. This also confirms the fact that the Lax operator (6.29) is,
indeed, in the core of integrable quantum as well as statistical models. Using the

Copyright © 2003 IOP Publishing Ltd.

 twisting transformation, one can also recover the 6V(1) vertex model introduced
in [27].
We may consider higher vertex models with n > 2, which may be obtained
from the R-matrix (or the Lax operators) of the corresponding quantum
integrable fundamental models with higher-dimensional auxiliary spaces. The
well-known examples are the 19-vertex model [43] related to the Babujian–
Takhtajan integrable spin-1 model [39], the Boltzmann weights of which may
be given by the matrix elements of Lax operator (6.52) with s = 1. Similarly,
one may construct the vertex models which are equivalent to the Hubbard model,
supersymmetric t–J model, Bariev chain etc [44].
In a following section, a new type of vertex model will be constructed from
our ancestor Lax operator using non-fundamental representations.

6.5 Directions for constructing new classes of ultralocal

models

The same unified scheme as that described in section 6.3 for constructing
integrable models may also be used to indicate various directions for generating
new integrable classes of quantum and statistical models.

6.5.1 Inhomogeneous models

In all previous constructions, the central elements in the ancestor models (6.29)
or (6.43) are chosen as constant parameters. However if they are chosen to be
site-dependent (or even time-dependent) functions, we can get an inhomogeneous
class of models. In these cases, the cs would be attached with site indices cj in the
Lax operators and, similarly, in (6.34) the Mj± would appear as functions, leading
to the corresponding inhomogeneous extensions of known integrable models, e.g.
the inhomogeneous lattice sine-Gordon, Liouville, Toda chain or the NLS model.
However, since the local algebra remains the same as in the original model,
they have the same quantum R-matrices. Although similar, inhomogeneous Toda
chain, NLS models etc were originally proposed as classical systems, they seem
to be new and have not been studied so far as quantum models. Recall that the
impurity models proposed earlier [45] fall into this class and are obtained by a
particular choice of inhomogeneous cj s which amounts to shifting the spectral
parameter. Implementing the same idea on the XXX spin chain, we notice that,
if in its construction along with ca0 = 1 we also choose c21 = −c11 = j resulting
again in m+ = 1, m− = 0, we get the same form of Lax operator but with a
shift Lj (λ − j ), resulting in the Gaudin model (6.24). Similarly higher-spin
representations as well as the su(1, 1) variant would yield other generalizations
of the same model. The commuting set of Hamiltonians for the Gaudin model

Copyright © 2003 IOP Publishing Ltd.

 may be generated from its transfer matrix at the limit α → 0 [18] as
1
Hj = (N τ (λ → j ) j = 1, 2, . . . , N.
α2 ( k (λ − k ))

Remarkably, the Gaudin model may be mapped into the integrable BCS model,
which is of immense contemporary interest [28].
Physically, such inhomogeneities may be interpreted as impurities, varying
external fields, incommensuration etc.

6.5.2 Hybrid models

Another way of constructing new models is to use different realizations of the
algebras (6.30) or (6.42) at different lattice sites, depending on the type of
R-matrix. For example, one may consider spin- 12 and spin-1 representations of
su(2) at alternate lattice sites, which was actually realized in [29]. However,
we can build more general inhomogeneous integrable models by considering
different underlying algebras and different Lax operators at differing sites.
The basic idea is that the Lax operators representing different models that
are descended from the same ancestor model and share the same R-matrix
can be combined together to build various hybrid models preserving quantum
integrability. For example, we may consider fermion–boson or spin–boson
interacting models by alternatively inserting spin- 12 and bosonic (or q-bosonic)
Lax operators at alternate sites. One such physical construction would be the
celebrated Jaynes–Cummings model. It is also possible to construct some exotic
hybrid integrable models, an example of which could be a hybrid sine-Gordon–
Liouville model where, for x ≥ 0, it would follow sine-Gordon dynamics while,
for x < 0, Liouville dynamics!

6.5.3 Non-fundamental statistical models

Vertex models, as already mentioned, are described generally by the R-matrix
of a regular quantum integrable model. However, one can construct a new class
of integrable vertex models by exploiting a richer variety of non-fundamental
systems, where we define the Boltzmann weights as matrix elements of the
j,k
generalized Lax operator (6.29)—Lab (u) = ωa,j ;b,k (u)—with the use of the
explicit matrix representation for the basic operators S ± , S 3 :

"s, m̄|S 3 |m, s# = mδm,m̄ "s, m̄|S ± |m, s# = fs± (m)δm±1,m̄ . (6.53)

Here fs+ (m) = fs− (m + 1) ≡ g(m) is defined as in (6.34). Such general

Boltzmann weights would now represent an ancestor vertex model analogous to
the quantum case and would generate, through various reductions, a new series
of vertex models, linked to q-spin and q-boson models with generic q, q roots of
unity and q → 1 [46]. In all these models, by generalizing the usual approach, the

Copyright © 2003 IOP Publishing Ltd.

 horizontal (h) and vertical (v) links may become inequivalent and independent at
every vertex point. The h-links, which are related to the auxiliary space admit two
values, while the v-links, which correspond to the quantum space, may have richer
possibilities with j, k ∈ [1, D], D being the dimension of the non-fundamental
matrix representation of the q-algebras. The familiar ice rule is generalized here
as the ‘colour’ conservation a + j = b + k for determining non-zero Boltzmann
weights. Note that, alternatively, finite-dimensional higher-spin and q-spin vertex
models can also be constructed using the fusion technique [39].
An interesting possibility for regulating the dimension of the matrix
representation opens up at q p = ±1, when a variety of new q-spin and q-boson
vertex models with finite-dimensional representation can be generated [46].
As in quantum models, we can also construct here a rich collection of hybrid
models by combining different vertex models of the same class and inserting their
defining Boltzmann weights along the vertex points l = 1, 2, . . . , N in each row,
in any but in the same manner. Due to the association with the same R-matrix, the
integrability of such statistical models is naturally preserved.

6.6 Unified Bethe ansatz solution

In physical models, our aim usually is to solve the eigenvalue problem for the
Hamiltonian only. Solvable models allow such exact solutions H |m# = Em |m#
through coordinate formulation of the Bethe ansatz (CBA) [47], which was used
successfully in many condensed matter physics problems like the spin chain, the
attractive and repulsive δ-Bose gas, the Hubbard model etc [48]. Nevertheless,
the CBA depends heavily on the structure of the Hamiltonian of individual
models and consequently lacks the unified approach of its algebraic formulation.
We would focus here briefly only on the algebraic Bethe ansatz (ABA) [2, 3],
which under certain conditions can solve the eigenvalue problem for the spectral-
parameter-dependent transfer matrix τ (λ)|m# = m (λ)|m# and, hence, through
its expansion the eigenvalue problem for the whole set of conserved operators,
simultaneously. Moreover, the ABA, due to its predominantly model-independent
features, which we will demonstrate later, appears to be a fairly universal method.
Since the eigenvectors are common for all commuting conserved operators,
by expanding ln m (λ) simply as

c1 |m# = 
m (0)−1
m (0)|m# c2 |m# = (
m (0)−1

m (0)) |m# (6.54)

etc, we obtain their respective values, where one may take H = c1 or any other
combination of cs as the Hamiltonian, depending on the concrete model. This
powerful method which is applicable to both integrable quantum and statistical
systems, requires, however, explicit knowledge of the associated Lax operator
and the R-matrix.
It may be noted that the off-diagonal element B(λ) (C(λ)) of the
monodromy matrix (6.5) acts generally as a creation (annihilation) operator

Copyright © 2003 IOP Publishing Ltd.

 for the pseudoparticles, induced by the local creation (annihilation) operator
as the matrix elements in Lj (λ) acting on the quantum space at j . Therefore,
the m-particle state |m#
( may be created by acting m times with B(λa ) on the
pseudovacuum |0# = N j |0#j , giving |m# = B(λ1 )B(λ2 ) . . . B(λm )|0#, where
we suppose the crucial annihilation condition C(λa )|0# = 0.
Now, for solving the eigenvalue problem of τ (λ) = A(λ) + D(λ) exactly,
we have to drag this operator through the string of B(λa )s without spoiling
their structure and finally hit the pseudovacuum giving A(λ)|0# = α(λ)|0#
and D(λ)|0# = β(λ)|0#. For this purpose, therefore, one requires commutation
relations between the elements of (6.5), which for ultralocal models may be
derived from the QYBE (6.6). This, apart from ensuring the integrability of the
system, is another important role played by (6.6), yielding the relations

A(λ)B(λa ) = f (λa − λ)B(λa )A(λ) − f1 (λa − λ)B(λ)A(λa )

(6.55)
D(λ)B(λa ) = f (λ − λa )B(λa )D(λ) − f1 (λ − λa )B(λ)D(λa )

together with the trivial commutators

[A(λ), A(µ)] = [B(λ), B(µ)] = [D(λ), D(µ)] = [A(λ), D(µ)] = 0

etc, where f (λ) = a(λ)/b(λ), f1 (λ) = c(λ)/b(λ) are combinations of the

elements from the R(λ)-matrix (6.7). We note that (6.55) are almost the right kind
of relations but for the second terms on both the rhss, where the argument of B
has changed spoiling the structure of the eigenvector. However, if we put the sum
of all such unwanted terms = 0, we should be able to achieve our goal. In field
models, such unwanted terms vanish automatically, while in lattice models their
removal amounts to the Bethe equations, which may be induced independently by
the periodic boundary condition, giving
 N  f (λa − λb )
α(λa )
= a = 1, 2, . . . , m. (6.56)
β(λa ) b=a
f (λb − λa )

Therefore, the ABA finally solves the eigenvalue problem for τ (λ), yielding
m   m 
m (λ) = f (λa − λ) α(λ) + f (λ − λa ) β(λ) (6.57)
a=1 a=1

where the Bethe equation (6.56), which is also equivalent to the singularity-free
condition of the eigenvalue (6.57) serves, in turn, as the set of equations for
determining the parameters λa .
Note that in both these equations, α(λ) = ("0|L̂11j (λ)|0#)
N and β(λ) =
22 N
("0|L̂j (λ)|0#) are the only model-dependent parts given by the action of
the upper and lower diagonal operator elements L̂ii j (λ), i = 1, 2, of the Lax
operator of the model on the pseudovacuum. For vertex models, for which the
ABA formulation runs in parallel, the Lax operator elements in these equations

Copyright © 2003 IOP Publishing Ltd.

 should be replaced by their matrix representations expressed through the
Boltzmann weights as "0|L̂11 j (λ)|0# = ω+,1;+,1 (λ), "0|L̂j (λ)|0# = ω−,1;−,1 (λ).
22

It is remarkable that the rest of the terms in (6.57) and (6.56) are given solely
through the R-matrix elements f (λ) and, therefore, depend only on the related
class (6.8) or (6.9). Recall that in integrable models, as described in section 6.3,
the R-matrix remains the same for all models belonging to a particular class, while
the L-operators differ and may be obtained through various reductions from the
same ancestor Lax operator.
Therefore, taking the Lax operator elements in (6.57) and (6.56) as
those from the general Lax operator (6.29), one may consider the previous
eigenvalue and the Bethe equation to be the unifying equations for exact solution
of all integrable ultralocal quantum and statistical models constructed here.
Consequently, models like the DNLS, SG, Liouville and the XXZ chain together
with the six-vertex model, belonging to the trigonometric class (6.8) should share
similar eigenvalue relations with individual differences appearing only in the form
of the α(λ) and β(λ) coefficients. Thus, this deep-rooted universality in integrable
systems helps us to solve the eigenvalue problem for the whole class of models
and for the full hierarchy of their conserved currents in a systematic way. Let us
present the explicit example of the XXZ chain with Lax operator (6.15), defining
|0# as all spin-up state which gives α(λ) = sinN (λ + α), β(λ) = sinN λ in Bethe
equation (6.56) (with a shift λ → λ + α/2) resulting in
 
sin(λa + α/2) N  m
sin(λa − λb + α)
= (6.58)
sin(λa − α/2) b=a
sin(λa − λb − α)

for a = 1, 2, . . . , m. Similarly, (6.57) gives the eigenvalue

m
sin(λa − λ + α/2)
m (λ) = sin (λ + α)
XXZ N

a=1
sin(λa − λ − α/2)
m
sin(λ − λa + 3(α/2))
+ sinN λ (6.59)
a=1
sin(λ − λa + α/2)

yielding for Hxxz = c1 , the energy spectrum


m
1
(m)
Exxz = m (λ)
−1
m (λ)|λ=0 = sin α + N cot α.
a=1
sin(λa − α/2) sin(λa + α/2)
(6.60)
At the limit α → 0, sin λ → λ, when the R-matrix along with its associated
models reduce to the rational class, one can derive the corresponding Bethe ansatz
results by taking the rational limit of these equations. For example, the relevant
equations for the isotropic XXX chain can be obtained directly from those for
the XXZ chain presented here. Intriguingly, the corresponding result for the NLS
lattice model, which belongs to the same rational class, should also show a close
similarity to that of the XXX chain.

Copyright © 2003 IOP Publishing Ltd.

 6.7 Quantum integrable non-ultralocal models
Though many celebrated classical integrable models, e.g. the KdV, mKdV,
nonlinear σ -model, derivative NLS, belong to the class of non-ultralocal models,
successful quantum generalization could be made only for a handful of them.
The reason, as mentioned already, is the violation of the ultralocality condition.
Recall that this condition helps the transition from the local QYBE to its
global form and, consequently, establishes the integrability of ultralocal systems.
Therefore, the key equations and the related formulation for the integrability
theory of the non-ultralocal models must be suitably modified.

6.7.1 Braided extensions of QYBE

For understanding the algebraic structures underlying the non-ultralocal systems,
first we have to note that the trivial multiplication property (6.32) valid
for ultralocal models needs to be generalized here as (A ⊗ B)(C ⊗ D) =
ψBC (A(C ⊗ B)D) where the braiding ψBC takes into account the non-
commutativity of B2 , C1 . In spite of such braided extension of the multiplication
rule, the associated coproduct structure of the underlying Hopf algebra, crucial
for transition to the global QYBE, must be preserved. Such a braided extension
of the Hopf algebra [49, 50] was implemented in formulating the integrability
theory of non-ultralocal models through an unified approach [51]. The basic idea
is to complement the commutation rule for the Lax operators at the same site with
their braiding property at different lattice sites. Note, however, that, in general, the
braiding may differ widely and with arbitrarily varying ranges, the picture might
become too complicated for an explicit description. Therefore, let us first limit
ourselves to the nearest-neighbour (NN)-type braiding
−1
L2j +1 (µ)Z21 L1j (λ) = L1j (λ)L2j +1 (µ) (6.61)
assuming that ultralocality holds starting from the next neighbours. A pictorial
description of this condition is given in figure 6.1(a).
The local QYBE at the same time must also be generalized to incorporate
the braiding relations, such that the transition to its global form becomes possible
again. Such braided extension of the QYBE (BQYBE) compatible with (6.61)
takes the form (see figure 6.1(b))
−1 −1
R12 (λ − µ)Z21 L1j (λ)L2j (µ) = Z12 L2j (µ)L1j (λ)R12 (λ − µ). (6.62)
We list here the known non-ultralocal integrable models that can be described
by these braided equations. Note that the quantum R-matrix appearing here is
the same (6.7) as for the ultralocal systems. However, the additional braiding
matrix Z, unlike the R-matrix, seems to be model-dependent and generally
independent of the spectral parameter; though similar to the R-matrix, it satisfies
the YBE-like equations and might also become spectral parameter dependent for
specific models [51].

Copyright © 2003 IOP Publishing Ltd.

 L1j
L2j 1 L1j

2 1
1 L2j 1
Z 21
1

2
a)

L 1j L 2j
1
1
1 1

2 1 2 1
R12 Z 21 Z 12 R12
2
2
L 2j b) L 1j
[1,k]
[1,k]
T1 T2
1
1
1 1
1 2 k 1 2 k
1
1
2 Z 12
2
R12 Z 21 R12 2
2
1 2 k
1 2 k
[1,k] C) [1,k]
T2 T1

Figure 6.1. Pictorial description of (a) braiding relation (6.61), (b) local braided
QYBE (6.62) for the Lax operators Laj (λa ) and (c) global braided QYBE for
(
Ta[1,k] (λa ) = kj =1 Laj (λa ), k < N. Note that putting Z = 1, i.e. removing the braiding
by undoing the crossing of the broken lines, one can recover the corresponding pictures
for the ultralocal models [1], namely ultralocality condition (6.2) and the local (6.3) and
global QYBE (6.6), respectively.

The next step is the global extension of the BQYBE for the monodromy
matrix (6.5) and it is not difficult to check that due to the braiding relation
[k,j ]
(6.61),
(k the form of the BQYBE is preserved for global matrices like Ta (λ) =
j =1 Laj (λ) (see figure 6.1(c)). However, since for the periodic boundary
condition, one imposes LaN+1 (λ) = La1 (λ), the Lax operators Laj (λ) for j = 1
and j = N again become NN entries and, hence, modify the equation due to the
appearance of an extra Z-matrix from the braiding relation (6.61), leading finally
to the global BQYBE:
−1 −1 −1 −1
R12 (λ − µ)Z21 T1 (λ)Z12 T2 (µ) = Z12 T2 (µ)Z21 T1 (λ)R12 (λ − µ). (6.63)

Though this equation is similar to (6.6), the commutation of the transfer

matrices ensuring the integrability of the systems through factorization of the
trace identity becomes problematic due to the presence of the Z-matrix. A
detailed discussion of this problem and classification of the Z-matrices allowing

Copyright © 2003 IOP Publishing Ltd.

 factorization is given in [51]. Some non-ultralocal systems were investigated from
a different angle in [52]. It is easy to see that from the corresponding equations
for the non-ultralocal models presented here, one can recover the known relations
for the ultralocal models by supposing the braiding matrix Z = 1 (see also the
caption in figure 6.1).

6.7.2 List of quantum integrable non-ultralocal models

Non-ultralocal models are mostly non-fundamental systems with infinite

dimensional representations defined in some Hilbert space. They may correspond
to integrable models with spectral-parameter dependent Lax operator and R(λ)-
matrix or may describe only non-ultralocal algebras with a spectral parameter-less
L-operator and an R(λ)λ→+∞ → Rq+ -matrix. Nevertheless, the non-ultralocal
quantum models listed here should be described through the same braided
relations (6.61)–(6.63) or their corresponding spectral-parameter-less form in a
systematic way. Therefore, we present only the explicit form of their braiding
matrix Z and the L-operator, indicating the class of R-matrix to which they
belong. These inputs should be enough to obtain all individual equations and
derive the related results.

6.7.2.1 Systems with spectral-parameter-less R-matrix

(1) Current algebra in the WZWN model [53]. The model involves the non-
ultralocal current algebra
γ
{L1 (x), L2 (y)} = [C, L1 (x) − L2 (y)]δ(x − y) + γ Cδ
(x − y) (6.64)
2

with C12 = 2P12 − 1, where P12 is the permutation operator, L = 12 (J0 + J1 )

with Jµ = ∂µ gg −1 is the current and g ∈ SU (N) the chiral field. The discretized
and quantum versions of this algebra may be cast as the spectral-parameter-free
limit of the previous braided YBE relations with Rq+ as the R-matrix, current L
+
as the Lax operator and Z12 = Rq21 as the braiding matrix, which takes the form

+ + + −1
Rq21 L1j L2j = L2j L1j Rq12 L1j L2j +1 = L2j +1 (Rq12 ) L1j . (6.65)

For the details and an interesting quantum group relation of this model, the readers
are referred to the original works [53].
(2) Coulomb gas picture of conformal field theory [54]. The Drinfeld–
Sokolov linear problem—Qx = L(x)Q—describing this system may be given
in the simplest case by the linear operator L(x) = v(x)σ 3 − σ + with a non-
ultralocal property due to the current-like relation {v(x), v(y)} = δ
(x − y).
Discretized and quantized forms of the current-like operator defined through the

Copyright © 2003 IOP Publishing Ltd.

 commutation relations
α
= ±i (δk,l+1 − δk+1,l )
2 (6.66)
+ − α
[vk , vl ] = i (δk+1,l − 2δk,l + δk,l+1 )
2
− 3
construct the corresponding discretized linear operator as Lk = e−ivk σ +
ivk+ +
e σ , which, similar to the previous case, satisfies the spectral-parameter-
free braided YBE and other relations with Rq+ as R and Z = q −σ ⊗σ as the
3 3

braiding matrix. Generalization of this model for SU (N) has also been similarly
constructed in [54].

6.7.2.2 Models with rational R(λ)-matrix

(3) Non-Abelian Toda chain [55]. The Lax operator of the model given by
 
λ − Ak −Bk−1
Lk (λ) =
I 0 (6.67)
−1 −1
Ak = ġk gk Bk = gk+1 gk gk ∈ SU (N)
represents a non-ultralocal integrable model and solves all braided relations
including the BQYBE with spectral-parameter-dependent rational R(λ) = P −
ihλI and the braiding matrix Z12 = 1 + ih(e22 ⊗ e12 )π, where P and π are
permutation operators. For further details of this model including its gauge
relation with an ultralocal model, we refer to the original work [55].
(4) Non-ultralocal quantum mapping [56]. The system is described by the
Lax operator Ln = V2n V2n−1 , with Vn = λn σ − + σ + + 12 vn (1 + σ 3 ), where the
discretized operator vk ≡ vk− involves non-ultralocal algebra (6.66) and, at the
continuum limit, → 0 yields the current-like field vk → i v(x). This non-
ultralocal quantum integrable model satisfies again integrable braided relations
with the spectral-parameter-dependent rational R(λ1 − λ2 )-matrix similar to the
previous case but now with a spectral-parameter-dependent braiding matrix
Z12 (λ2 ) = I − (h/λ2 )σ − ⊗ σ + and Z21 (λ1 ). For generalization of this model to
higher rank groups and other details, we refer again to the original work [56].

6.7.2.3 Models with trigonometric R(λ)-matrix

(5) Quantum mKdV model [57]. This well-known non-ultralocal model may
be raised to the quantum level with discrete Lax operator
 
(Wk− )−1 i ξ Wk+
Lk (ξ ) = (6.68)
−i ξ(Wk+ )−1 Wk−
±
where Wj± = eivj with vk± obeying non-ultralocal relations (6.66). R-matrix (6.8)
1 3
and the braiding matrix Z12 = Z21 = q − 2 σ ⊗σ are associated with this non-
3

ultralocal integrable system [57]. The Bethe ansatz solution of the quantum

Copyright © 2003 IOP Publishing Ltd.

 mKdV and its generalizations can be found in detail in [58]. It is seen easily
that one can recover the well-known Lax operator of the mKdV field model—
U (x, ξ ) = (i/2)(iv(x)σ 3 + ξ σ 2 ) from (6.68) at the field limit when vk∓ →
√
∓ v(x), as Lk = I + U (x, ξ ) + O( 2 ).
(6) Quantum light-cone sine-Gordon model. It is known that this well-
known equation, ∂+− 2 u = 2 sin 2u, may be represented by the zero curvature

condition ∂− U+ − ∂+ U− + [U+ , U− ] = 0 of the Lax pair U± with U− (x) =

(i/2)∂− u(x)σ 3 + ξ(e−iu(x)σ + + eiu(x)σ − ) and, similarly, for U+ (x). Recently,
quantum as well as exact lattice versions of the non-ultralocal Lax operator have
been constructed [23] which, in particular for U− (x), may be given in the form

(λ) = ei(pj −α∇uj )σ + ξ(e−i(pj +αuj+1 ) σ + + ei(pj +αuj+1 ) σ − )

(−)lcsg 3
Lj
(6.69)
∇uj ≡ uj +1 − uj .

It may be shown also that (6.69) obeys exactly the previous BQYBE and the
braiding relation with the trigonometric R-matrix (6.8) and the braiding matrix
= eiασ ⊗σ and, consequently, represents a genuine quantum integrable non-
(−) 3 3
Z12
ultralocal model.
Some other non-ultralocal models known in the literature need an
introduction to braiding beyond NN, the basic formulation of which can be found
in [50, 51]. Examples of such models with the same braiding between any two
different sites are: (i) integrable model on moduli space [59], (ii) supersymmetric
models [51, 60], (iii) braided algebra [49], (iv) non-ultralocal extension of the
YBE [61] etc. A unified description of them can be found in [51, 62].

6.7.3 Algebraic Bethe ansatz

The solution of the eigenvalue problem for integrable non-ultralocal models by
diagonalizing the transfer matrix may be formulated through the algebraic Bethe
ansatz in exact analogy with the ultralocal models, whenever the factorization of
the trace problem, as previously mentioned, could be resolved. The key equation
that is to be used for non-ultralocal models for finding the commutation relations
analogous to (6.55) in the ABA scheme should naturally be given by BQYBE
(6.63). However, we skip all details of this ABA formulation for non-ultralocal
models, which can be found in explicit form in the example of the non-ultralocal
quantum mKdV model in [57, 58].

6.7.4 Open directions in non-ultralocal models

Since some of the previously described non-ultralocal models, e.g. the non-
Abelian Toda chain, the WZWN current algebra, the mKdV etc, can be connected
to ultralocal models through operator-dependent local gauge transformations, it

Copyright © 2003 IOP Publishing Ltd.

 would be challenging to discover similar relations, if any, for the rest of the
quantum integrable non-ultralocal models [23].
Other challenging problems undoubtedly are the possible quantum integrable
formulation of the well-known non-ultralocal models, e.g. the nonlinear σ -model,
the complex sine-Gordon model, the derivative NLS equation etc through the
braided YBE.
As we know, there is a remarkable interconnection between the integrable
quantum and statistical models. However, this connection has been discovered
as yet only for ultralocal models as we have also seen here. Therefore, a new
direction of study would be to investigate whether there could be any meaningful
statistical model corresponding to the integrable non-ultralocal models described
here.
Another problem worth looking into would be to formulate fundamental non-
ultralocal models, if they exist, which could then be possibly used to generalize
spin and electron models with non-ultralocality.
Anyway since this vast branch of integrable systems has received
significantly insufficient attention, we may hope to find much hidden excitement
in it.

6.8 Concluding remarks

Quantum integrable systems can be divided into two broad classes: ultralocal
and non-ultralocal. We have presented here a brief description of such models
with references for further details and demonstrated that the models belonging
to both these classes can be described systematically through a set of algebraic
relations signifying the integrability of these systems. For ultralocal models, these
relations are the ultralocality condition and the QYBE involving Lax operator
L and the R-matrix, while for non-ultralocal models they are extended to the
braiding relation and the braided QYBE with the additional entry of the braiding
matrix Z. The L-operator representing an individual model is naturally model-
dependent and the same also seems to be true for the Z-matrix. The R-matrix, on
the other hand is mainly of two types (the elliptic case has not been considered
here)—trigonometric and rational—depending on the class of models that are
associated with the q-deformed and undeformed algebras, respectively. This also
induces a significantly model-independent approach in the ABA method for
solving the eigenvalue problem. For ultralocal systems, the theory of which is
more developed, one can go further and prescribe a unifying algebraic scheme
for generating individual Lax operators realized from a single ancestor model in a
systematic way. It would be a challenge to extend the formulation of this scheme
to non-ultralocal models. The integrable statistical vertex models can be related to
the corresponding quantum models which, as a rule, belong to ultralocal systems.
Systematic extension of such relations to non-ultralocal systems would be another
challenging problem.

Copyright © 2003 IOP Publishing Ltd.

 References

[1] Baxter R 1981 Exactly Solved Models in Statistical Mechanics (New York: Academic
Press)
Deguchi T 2003 Chapter 5 in this book
[2] Faddeev L D 1980 Sov. Sci. Rev. C 1 107
de Vega H J 1989 Int. J. Mod. Phys. A 4 2371
Jimbo M (ed) 1989 Yang–Baxter Equation in Integrable System (Singapore: World
Scientific)
Foerster A, Links J and Zhou H Q 2003 Chapter 8 in this book
[3] Korepin V E, Bogoliubov N M and Izergin A G 1993 QISM and Correlation
Functions (Cambridge: Cambridge University Press)
[4] Reshetikhin N Yu, Takhtajan L A and Faddeev L D 1989 Algebra Analysis 1 178
Faddeev L D 1995 Int J. Mod. Phys. A 10 1845
[5] Sklyanin E 1988 J. Phys. A: Math. Gen. 21 2375
[6] Kulish P and Sklyanin E 1982 Lecture Notes in Physics vol 151 (Berlin: Springer)
[7] Sklyanin E, Takhtajan L and Faddeev L 1979 Teor. Mat. Fiz. 40 194
[8] Faddeev L D and Takhtajan L A, 1986 Lecture Notes Physics vol 246, ed H de Vega
et al (Berlin: Springer)
[9] Kundu A and Basu-Mallick B 1993 J. Math. Phys. 34 1252
[10] Takhtajan L A and Faddeev L D 1979 Russ. Math. Surveys 34 11
[11] Izergin A G and Korepin V E 1982 Nucl. Phys. B 205 [FS5] 401
[12] Kundu A and Basu-Mallick B 1992 Mod. Phys. Lett. A 7 61
[13] Basu-Mallick B and Kundu A 1992 Phys. Lett. B 287 149
[14] Kundu A 1994 Phys. Lett. A 190 73
[15] Suris Yu B 1990 Phys. Lett. A 145 113
[16] Ablowitz M J and Ladik J F 1976 J. Math. Phys. 17 1011
[17] Kundu A 1999 Phys. Rev. Lett. 82 3936
[18] Sklyanin E K 1989 J. Sov. Math. 47 2473
[19] Kundu A and Ragnisco O 1994 J. Phys. A: Math. Gen. 27 6335
[20] Enol’skii V, Salerno M, Kostov N and Scott A 1991 Phys. Scr. 43 229
Enol’skii V, Kuznetsov V and Salerno M 1993 Physica D 68 138
[21] Drinfeld V G 1986 Proc. Int. Cong. Mathematicians (Berkeley) vol 1 p 798
Chari V and Pressley A 1994 A Guide to Quantum Groups (Cambridge: Cambridge
University Press)
[22] Faddeev L D and Tirkkonen O 1995 Nucl. Phys. B 453 647
[23] Kundu A 2002 Phys. Lett. B 550 128
[24] Macfarlane A J 1989 J. Phys. A: Math. Gen. 22 4581
Biederharn L C 1989 J. Phys. A: Math. Gen. 22 L873
[25] Shnirman A G, Malomed B A and Ben-Jacob E 1994 Phys. Rev. A 50 3453
[26] Inoue R and Hikami K 1998 J. Phys. Soc. Japan 67 87
[27] Sogo K, Uchinomi M, Akutsu Y and Wadati M 1982 Prog. Theor. Phys. 68 508
Wadati M, Deguchi T and Akutsu Y 1989 Phys. Rep. 180 247
[28] Dukelsky J and Schuck P 2001 Phys. Rev. Lett. 86 4207
Delft J V and Poghossian R 2002 Phys. Rev. B 66 134502
Foerster A, Links J and Zhou H Q 2003 Chapter 8 in this book
[29] de Vega H and Woynarovich F 1992 J. Phys. A: Math. Gen. 25 4499

Copyright © 2003 IOP Publishing Ltd.

 [30] Tarasov V O 1985 Teor. Mat. Fiz. 63 175
Izergin A G and Korepin V E 1984 Lett. Math. Phys. 8 259
[31] Essler F and Korepin V E 1992 Phys. Rev. B 46 9147
[32] Bracken A J, Gould M D, Links J R and Zhang Y Z 1995 Phys. Rev. Lett. 74 2768
[33] Wang Y 1999 Phys. Rev. B 60 9236
Bose I 2003 Chapter 7 in this book
[34] Essler F, Korepin V E and Schoutens K 1992 Phys. Rev. Lett. 68 2960
[35] Frahm H and Kundu A 1999 J. Phys.: Condens. Matter 27 L557
[36] Izumov Yu A and Skryabin Yu N 1988 Statistical Mechanics of Magnetically Ordered
Systems (New York: Consultants Bureau)
[37] Kundu A. 2001 Nucl. Phys. B 618 500 and references therein
[38] Gogolin A O, Narsesyan A A and Tsvelik A M 1999 Bosonization and Strongly
Correlated Systems (Cambridge: Cambridge University Press)
Sachdev S 2000 Quantum Phase Transition (Cambridge: Cambridge University
Press)
See also Reviews by Bose I 2003 Chapter 7 in this book and Foerster A, Links J and
Zhou H Q 2003 Chapter 8 in this book
[39] Babujian H M 1983 Nucl. Phys. B [FS7] 317
Babujian H M 1982 Phys. Lett. 90 A 479
Takhtajan L A 1982 Phys. Lett. 87 A 479
[40] Faddeev L D and Takhtajan L A 1987 Hamiltonian Methods in the Theory of Solitons
(Berlin: Springer)
[41] Kundu A 1994 Theor. Math. Phys. 99 428
Kundu A 2002 Unifying scheme for generating discrete integrable systems including
inhomogeneous and hybrid models Preprint nlin.SI/0212004 (to appear in J. Math.
Phys.)
[42] Sutherland B, Yang C P and Yang C N 1967 Phys. Rev. Lett. 19 588
[43] Zamolodchikov A B and Fateev V A 1980 Sov. J. Nucl. Phys. 32 293
[44] Olmedilla E, Wadati M and Akutsu Y 1987 J. Phys. Soc. Japan 87 2298
Foerster A and Karowski M 1992 Phys. Rev. B 46 9234
Zhou H Q 1996 J. Phys. A: Math. Gen. 29 5504
[45] Eckle H, Punnoose A and Römer R 1997 Europhys. Lett. 39 293
[46] Kundu A 2002 J. Phys. A: Math. Gen. 35 L1
Kundu A 2003 Physica A 318 144
[47] Bethe H 1931 Z. Phys. 71 205
[48] Yang C N and Yang C P 1966 Phys. Rev. 150 321, 327
Mattis D C (ed) 1993 Encyclopedia of Exactly Solved Models in One Dimension
(Singapore: World Scientific)
Lieb E and Liniger W 1963 Phys. Rev. 130 1605
McGuire J B 1964 J. Math. Phys. 5 622
Yang C N 1967 Phys. Rev. Lett. 19 1312
Lieb E H and Wu F Y 1968 Phys. Rev. Lett. 20 1445
[49] Majid S 1991 J. Math. Phys. 32 3246
[50] Hlavaty L 1994 J. Math. Phys. 35 2560
Freidel L and Maillet J M 1991 Phys. Lett. B 262 278
Freidel L and Maillet J M 1991 Phys. Lett. B 263 403
[51] Hlavaty L and Kundu A 1996 Int. J. Mod. Phys. 11 2143

Copyright © 2003 IOP Publishing Ltd.

 [52] Reshetikhin N Yu and Semenov-Tian-Shansky M 1990 Lett. Math. Phys. 19 133
Chu M, Goddard P, Halliday I, Olive D and Schwimmer A 1991 Phys. Lett. B 266 71
Volkov A Yu 1996 Commun. Math. Phys. 177 381
[53] Faddeev L D 1990 Commun. Math. Phys. 132 131
Alekseev A, Faddeev L D, Semenov-Tian-Shansky M and Volkov A 1991 The
unraveling of the quantum group structure in the WZWN theory Preprint CERN-
TH-5981/91
[54] Babelon O and Bonora L 1991 Phys. Lett. B 253 365
Babelon O 1991 Commun. Math. Phys. 139 619
Bonora L and Bonservizi V 1993 Nucl. Phys. B 390 205
[55] Korepin V E 1983 J. Sov. Math. 23 2429
[56] Nijhoff F W, Capel H W and Papageorgiou V G 1992 Phys. Rev. A 46 2155
[57] Kundu A 1995 Mod. Phys. Lett. A 10 2955
[58] Fioravanti D and Rossi M 2002 J. Phys. A: Math. Gen. 35 3647
[59] Alexeev A Yu 1993 Integrability in the Hamiltonian Chern–Simons theory Preprint
hep-th/9311074
[60] Chaichian M and Kulish P 1990 Phys. Lett. B 234 72
Liao L and Song X 1991 Mod. Phys. Lett. 6 959
[61] Schwiebert C 1994 Generalized quantum inverse scattering method Preprint hep-
th/9412237
[62] Kundu A 1996 Quantum integrable system: construction solution algebraic aspect
Preprint hep-th/9612046 (unpublished)

Copyright © 2003 IOP Publishing Ltd.

 Chapter 7

The physical basis of integrable spin models

Indrani Bose
Department of Physics, Bose Institute, Calcutta, India

7.1 Introduction
The study of integrable models constitutes an important area of theoretical
physics. Integrable models in condensed matter physics describe interacting
many-particle systems. The most prominent examples are interacting spin and
electron systems which include several real materials of interest. Integrable
models, because of their exact solvability, provide a complete and unambiguous
understanding of the variety of phenomena exhibited by real systems. Integrability
in the quantum case implies the existence of N conserved quantities where N is
the number of degrees of freedom of the system. The corresponding operators
including the Hamiltonian commute with each other. More specifically, integrable
models are also described as exactly solvable since the ground-state energy and
the excitation spectrum of the models can be determined exactly. Historically,
the first example of the exact solvability of a many-body problem was that of
a spin- 12 quantum spin chain [1]. The technique used to solve the eigenvalue
problem is now known as the Bethe ansatz (BA) named after Hans Bethe who
formulated it. The demonstration of integrability, namely the existence of N
commuting operators can be made in the more general mathematical framework
of the quantum inverse scattering method (QISM) [2]. The BA has been used
extensively to obtain exact results for several quantum models in one dimension.
Examples include the Fermi and Bose gas models in which particles on a
line interact through delta-function potentials [3], the Hubbard model [4], one-
dimensional (1D) plasma which crystallizes as a Wigner solid [5], the Lai–
Sutherland model which includes the Hubbard model and a dilute magnetic model
as special cases [6], the Kondo model [7], the single impurity Anderson model [8],
the supersymmetric t–J model (J = 2t) etc [9]. The BA method has further
been applied to derive exact results for classical lattice statistical models in two
dimensions.

Copyright © 2003 IOP Publishing Ltd.

 The BA denotes a particular form for the many-particle wavefunction. In
a 1D system with pairwise interactions, a two-particle scattering conserves the
momenta individually due to the energy and momentum conservation constraints
peculiar to one dimension. Hence, the scattering particles can either retain their
original momenta or exchange them. In the case of two particles (N = 2), the
wavefunction has the form

ψ(x1 , x2 ) = A12 ei(k1 x1 +k2 x2 ) + A21 ei(k2 x1 +k1 x2 ) (7.1)

where x1 , x2 denote the locations of the two particles and k1 , k2 are the momentum
variables. The wavefunction can alternatively be written as

ψ(x1 , x2 ) = ei(k1 x1 +k2 x2 ) + eiθ12 ei(k2 x1 +k1 x2 ) (7.2)

where θ12 is the scattering phase shift. The BA generalizes the wavefunction
(equation (7.1)) to the case of N particles and is given by

  N 
ψ= A(P ) exp i kPj xj x1 < x2 < · · · < xN . (7.3)
P j =1

The sum over P is a sum over all permutations of 1, . . . , N. The amplitude

A(P ) is factorizable. Each A(P ) is a product of factors eθij corresponding to
each exchange of ki s required to go from the ordering 1, . . . , N to the ordering
P . An overall sign factor may arise depending on the parity of the permutation.
The unknown variables, θij and ki , are obtained as solutions of coupled nonlinear
equations. The factorizability condition is at the heart of the exact solvability
of the eigenvalue problem. In the more general QISM approach, the so-called
Yang–Baxter equation provides the condition for factorization of a multi-particle
scattering matrix in terms of two-particle scattering matrices.
The traditional BA (equation (7.3)) is known as the coordinate Bethe ansatz
(CBA). Over the years, the BA method has been generalized in different ways.
The nested BA technique [3, 10] has been applied to study a system of particles
with internal degrees of freedom. The state of a system of electrons is specified in
terms of both the spatial positions as well as the spin indices of the electrons. The
asymptotic Bethe ansatz [11] deals with a class of models in which the interaction
between a pair of particles falls off as the inverse square of the distance between
the particles. The thermodynamic Bethe ansatz method [12] is used to calculate
thermodynamic quantities and is a finite-temperature extension of the BA method.
The algebraic Bethe ansatz (ABA) [13] has been developed in the powerful
mathematical framework of the QISM. The ABA and CBA are equivalent in the
sense that both lead to the same results for the energy eigenvalues. The CBA,
however, does not provide knowledge of the correlation functions as the structure
of the wavefunction is not sufficiently explicitly known. The QISM allows the
calculation of the correlation functions in some cases [14]. The mathematical
formalism is also much more systematic and general. One can further establish

Copyright © 2003 IOP Publishing Ltd.

 the existence of an infinite number (N → ∞) of mutually commuting operators.
The QISM, moreover, provides a prescription for the construction of integrable
models. In this review, we will not discuss the mathematical aspects of integrable
models for which a good number of reviews already exist [2,15–17]. We focus on
the physical basis of some integrable spin models in condensed matter physics and
the useful physical insights derived from the solution of these models. The review
is not meant to be exhaustive and should be supplemented by the references
quoted at the end.

7.2 Spin models in one dimension

The interest in 1D spin models arises from the fact that there are several real
magnetic materials which can be described by such models. The spins interact
via the Heisenberg exchange interaction and in many compounds the exchange
interaction within a chain of spins is much stronger than that between chains.
Thus, the compounds effectively behave as linear chain systems. The most general
exchange interaction Hamiltonian describing a chain of spins in which only
nearest-neighbour (nn) spins interact is given by


N
y y
HXY Z = [Jx Six Si+1
x
+ Jy Si Si+1 + Jz Siz Si+1
z
] (7.4)
i=1

where Siα (α = x, y, z) is the spin operator at the lattice site i, N is the total
number of sites and Jα denotes the strength of the exchange interaction. Consider
the spins to be of magnitude 12 . The eigenvalue problems corresponding to
the isotropic chain (Jx = Jy = Jz = J ) and the longitudinally anisotropic chain
(Jx = Jy = Jz ) were originally solved using the CBA. Later, the same solutions
were obtained using the formalism of QISM [13, 15]. Baxter [18] calculated the
ground-state energy of the fully anisotropic model (equation (7.4)) and Johnson
et al [19] found the excitation spectrum. The results were derived on the basis of a
special relationship between the transfer matrix of the exactly-solved 2D classical
lattice statistical eight-vertex model and the fully anisotropic quantum spin
Hamiltonian HXY Z . Later, the same results were obtained by the ABA approach
of the QISM. The Ising (Jx = Jy = 0) and the XY (Jz = 0) Hamiltonians are
special cases of HXY Z .
Consider the isotropic Heisenberg exchange interaction Hamiltonian in one
dimension:
N
H =J S'i · S'i+1 (7.5)
i=1
with periodic boundary conditions. The sign of the exchange interaction
determines the favourable alignment of the nn spins. J > 0 corresponds to the
antiferromagnetic (AFM) exchange interaction due to which nn spins tend to
be antiparallel. If J < 0 (equivalently, replace J by −J in equation (7.5) with

Copyright © 2003 IOP Publishing Ltd.

 J > 0), the exchange interaction is ferromagnetic (FM) favouringa parallel
alignment of nn spins. One can include a magnetic field term −h N z
i=1 Si in
the Hamiltonian (equation (7.5)), where h is the strength of the field. Given
a Hamiltonian, the quantities of interest are the ground-state energy and the
low-lying excitation spectrum. Knowledge of the latter enables one to calculate
thermodynamic quantities like the magnetization, specific heat and susceptibility
at low temperatures. In the case of the FM Heisenberg Hamiltonian, the exact
ground state has a simple structure. All the spins are parallel, i.e. they align in the
same direction. The lowest excitation is a spin wave or magnon. The excitation
is created by deviating a spin from its ground-state arrangement and letting it
propagate. For more than one spin deviation, one has continua of scattering states
as well as bound complexes of magnons. In a bound complex, the spin deviations
preferentially occupy nn lattice positions. The r-magnon bound-state energy can
be calculated using the BA [1] and the energy (in units of J) measured with respect
to the ground-state energy is
1
 = (1 − cos K) (7.6)
r
where K is the centre-of-mass momentum of the r-magnons. The spin wave
excitation energy is obtained for r = 1. The results can be generalized to the
longitudinally anisotropic XXZ Hamiltonian. The multi-magnon bound states
were first detected in the quasi-1D magnetic system CoCl2 .2H2 O [20]. Later
improvements made it possible to observe even 14-magnon bound states [21].
In the case of the AFM isotropic Heisenberg Hamiltonian, the ground state
is a singlet and the ground-state wavefunction is a linear combination of all
possible states in which half the spins are up and the other half down. The AFM
ground state can be obtained from the FM ground state by creating r = N/2
magnons with momenta ki and negative energies −J (1 − cos ki ). Remember
that the sign of the exchange integral changes in going from ferromagnetism to
antiferromagnetism. The highest energy state in the FM case (r = N/2) becomes
the ground state in the AFM case. The BA equations can be recast in terms of the
variables zi ≡ cot(ki /2) [22]:
 z z 
i− j
N arctan zi = πIi + arctan i = 1, 2, . . . , r. (7.7)
j =i
2

The Bethe quantum numbers Ii s are integers (half-integers) for odd (even) r. For
a state specified by {I1 , . . . , Ir }, the solution (z1 , . . . , zr ) can be obtained from
equation (7.7). The energy and momentum wavenumber of the state are given by
E − EF r
2
=− (7.8)
i=1 1 + zi
J 2

2π 
r
k = πr − Ii (7.9)
N i=1

Copyright © 2003 IOP Publishing Ltd.

 Figure 7.1. A two-spinon configuration in an AFM chain.

with EF = JN/4. For the AFM ground state, the Bethe quantum numbers are given
by  
N 1 N 3 N 1
{Ii } = − + , − + , . . . , − . (7.10)
4 2 4 2 4 2
In the thermodynamic limit N → ∞, the exact ground-state energy has been
computed as
Eg = NJ (− ln 2 + 14 ). (7.11)
The AFM ground state serves as the physical vacuum for the creation of
elementary excitations. These excitations are not the spin-1 magnons but
spin- 12 spinons [23]. The spinons can be generated systematically by suitable
modifications of the vacuum array of the BA quantum numbers (equation (7.10))
(for details see [22, 23]). For even N, spinons are always created in pairs, each
such pair originating from the removal of one magnon from the ground state.
Since the spinons are spin- 12 objects, the lowest excitations consisting of a pair
of spinons are four-fold degenerate, three triplet (S = 1) and one singlet (S = 0)
excitations. The energy can be written as E(k1 , k2 ) = (k1 ) + (k2 ) where the
spinon spectrum (ki ) = (π/2) sin ki and the total momentum k = k1 + k2 . At a
fixed total momentum k, one gets a continuum of scattering states. The lower
boundary of the continuum is given by (π/2)| sin k| with one of the ki
s equal
to zero. The upper boundary is obtained for k1 = k2 = k/2 and is given by
π| sin k/2|. Figure 7.1 gives an example of a two-spinon configuration.
The BA results are obtained in the thermodynamic limit. In this limit, the
energies and the momenta of the spinons just add up, showing that they do not
interact. Since the spinons are excited in pairs, the total spin of the excited state is
an integer. Inelastic neutron scattering study of the linear chain S = 12 Heisenberg
AFM (HAFM) compound KCuF3 has confirmed the existence of unbound spinon
pair excitations [24]. It is to be noted that in the case of a ferromagnet, the
low-lying excitation spectrum consists of a single magnon branch whereas the
AFM spectrum is a two-spinon continuum with well-defined lower and upper
boundaries.
The dynamical properties of a magnetic system are governed by the time-
dependent pair correlation functions or their spacetime double Fourier transforms
known as the dynamical correlation functions. An important time-dependent
correlation function is
G(R, t) = "S'R (t) · S'0 (0)#. (7.12)

Copyright © 2003 IOP Publishing Ltd.

 The corresponding dynamical correlation function is the quantity measured in
inelastic neutron scattering experiments. The differential scattering cross-section
in such an experiment is given by

d2 σ 1  iq'·R' +∞ µ µ
∝ S µµ ('
q , ω) = e dt eiωt "SR (t)S0 (0)# (7.13)
d dω N R −∞

where q' and ω are the momentum wavevector and energy of the spin excitation
and µ = x, y, z. For a particular q' , the peak in S µµ ('
q , ω) occurs at a value of ω
which gives the excitation energy. At T = 0,
 µ
q , ω) =
S µµ (' Mλ δ(ω + Eg − Eλ ) (7.14)
λ

Eg (Eλ ) is the energy of the ground (excited) state and

µ
Mλ = 2π|"G|S µ ('
q )|λ#|2 (7.15)

is the transition rate between the singlet (Stot = 0) ground state |G# and the triplet
(Stot = 1) states |λ# [25]. Exact calculation of the dynamical correlation functions
in the BA formalism is not possible. Bougourzi et al [26] have used an alternative
approach, based on the algebraic analysis of the completely integrable spin
chain and have calculated the exact two-spinon part of the dynamical correlation
function S xx (q, ω) for the 1D S = 12 AFM XXZ model. In this model, the
Ising part of the XXZ Hamiltonian provides the dominant interaction. Karbach
et al [27] have calculated the exact two-spinon part of S zz (q, ω) for the isotropic
Heisenberg Hamiltonian. In both cases, the size of the chain is infinite. The
exact form of the two-spinon contribution to the dynamical correlation function
S xx (q, ω) of the S = 12 XXZ HAFM chain is complicated and is given by
 ) 
ω ω 2 − χ 2 ω2  ϑ 2 (β c ) |tan(q/2)|−c
0 0 A −
xx
S(2) (q, ω) = 1+ (7.16)
8I ω ω − ω0
2 2 2 c
c=± ϑd (β− ) Wc

where
JK πK
1 − k

I= sinh χ≡ and k, k
≡ 1 − k2
π K 1 + k

are the moduli of the elliptic integrals K ≡ K(k), K

≡ K(k
). The anisotropy
parameter q = − exp(−πK
/K) with −Jz /Jx = = (q + q −1 )/2. Also,
)
 2
ω04 2 T
W± = χ − ± cos q (7.17)
ω4 ω2
* *
T = ω2 − χ 2 ω02 ω2 − ω02 (7.18)

Copyright © 2003 IOP Publishing Ltd.

 2I sin(q)
ω0 = (7.19)
1+χ
   
1+χ 2I ωWc
c
β− (q, ω) = F arcsin ,χ (7.20)
2 χ(1 + χ)ω02
(F is the incomplete elliptic integral)
  ∞ γl 
e cosh(2γ l) cos(tγ l) − 1
ϑA (β) = exp −
2
(7.21)
l=1
l sinh(2γ l) cosh(γ l)

γ = πK
/K, t ≡ 2β/K
and ϑd (x) is a Neville theta-function. The derivation of
xx (q, ω) involves generating the two-spinon states from the spinon vacuum,
S(2)
namely the AFM ground state, with the help of spinon creation operators
and expressing the spin fluctuation operator S µ (q) in terms of the spinon
creation operators. The two-spinon part is expected to provide the dominant
contribution to the dynamical correlation function (7.14). For example, in the
case of the isotropic Heisenberg Hamiltonian, the two-spinon excitations account
for approximately 73% of the total intensity in S zz (q, ω). The two-spinon
triplet excitations play a significant role in the low-temperature spin dynamics
of quasi-1D AFM compounds like KCuF3 ,Cu(C6 D5 COO)2 .3D2 O, Cs2 CuCl4
and Cu(C4 H4 N2 (NO3 )2 ) [24, 28]. These excitations can be probed via inelastic
neutron scattering and, hence, a knowledge of the exact dynamical correlation
function is useful. The two-spinon singlet excitations cannot be excited in neutron
scattering because of selection rules (the spinon vacuum |G# is a singlet and
the excited state |λ# in equation (7.15) is a triplet). Linear chain compounds
like CuGeO3 exhibit the spin-Peierls transition [29]. The transition gives rise to
lattice distortion and consequently to a dimerization of the exchange interaction.
Exchange interactions between successive pairs of spins alternate in strength.
There is a tendency for the formation of dimers (singlets) across the strong bonds.
One can construct an appropriate dynamical correlation function in which the
dimer fluctuation operator (DFO) replaces the spin fluctuation operator S µ (q).
The DFO connects the AFM ground state to the two-spinon singlet and not to the
two-spinon triplet.
Two well-known physical realizations of the 1D S = 12 Ising–Heisenberg
compounds are CsCoCl3 and CsCoBr3 . Several inelastic neutron scattering
measurements have been carried out on these compounds to probe the low-
temperature spin dynamics [30]. In these compounds, the Ising part of the
XXZ Hamiltonian is significantly dominant so that perturbation calculations
around the Ising limit are feasible. Near the Ising limit, the exact two-spinon
dynamical correlation function S xx (q, ω) is identical in the lowest order to the
first-order perturbation result of Ishimura and Shiba (IS) [31]. The IS calculation
provides physical insight on the nature of spinons. The Ising part of the XXZ
Hamiltonian is the unperturbed Hamiltonian and the XY part constitutes the
perturbation. The two-fold degenerate Néel states are the ground states of the

Copyright © 2003 IOP Publishing Ltd.

 Ising Hamiltonian. These two states serve as the ‘spinon vacuua’. An excitation
is created by flipping a block of adjacent spins from the spin arrangement in the
Néel state. For example, in figure 7.1, a block of seven spins is flipped in the
Néel state. The block of overturned spins gives rise to two parallel spin pairs at
its boundary with the unperturbed Néel configuration. It is these domain walls
or kink solitons which are the equivalents of spinons. A two-spinon excited
z
state (Stot = 1) is obtained as a linear superposition of states in which an odd
number ν (ν = 1, 3, 5, . . . ) of spins is overturned in the Néel configuration.
In each such state, both the domain walls have equal spin orientations with the
spins pointing up. The excitation continuum of two spinons is obtained in first-
order perturbation theory. The lineshapes of S xx (q, ω) observed in experiments
are highly asymmetric with a greater concentration of intensity near the spectral
threshold and a tail extending to the upper boundary of the continuum. The exact
two-spinon part of S xx (q, ω) has also an asymmetric shape in agreement with
experimental data. The first-order perturbation-theoretic result of IS for S xx (q, ω)
fails to reproduce the asymmetry. A second-order perturbation calculation leads
to greater asymmetry in the lineshapes [32]. Furthermore, in the framework of a
first-order perturbation theory, the effects of full anisotropy (Jx = Jy = Jz ), next-
nearest-neighbour coupling, interchain coupling and exchange mixing have been
shown to give rise to asymmetry in lineshapes [33].
Recently, a large number of studies have been carried out on a class of
models in which the interaction between spins falls off as the inverse square of
the distance between them. A lattice model which belongs to this class is known
as the Haldane–Shastry model [34], the Hamiltonian of which is given by
 Pij
H =J (7.22)
i<j
d(i − j )2

where d(l) = (N/π)| sin(πl/N)| is the chord distance between the pair of spins
separated by l sites on a ring with N equally spaced spins. Pij is the spin
exchange operator, Pij = (2S'i · S'j + 12 ). The model is exactly solvable and the
key results are: the ground state has a form similar to the fractional quantum
Hall ground state, the ground state is a quantum spin liquid (QSL) and the
elementary excitations are the spin- 12 spinons obeying fractional statistics, the
thermodynamics as well as the various dynamical correlation functions can be
calculated exactly. The latter calculations are possible because of the simple
structure of the eigenspectrum.
A correct analysis of the BA equations for the S = 12 HAFM in one
dimension gave rise to the concept of spinons which has subsequently been
verified in experiments. Approximate methods like spin wave theory fail to predict
the spinon continuum, thus pointing to the importance of integrable models in
providing the correct physical picture. The existence of spinons in dimensions
greater than one is a highly debatable issue. No precise statement can be made due
to the lack of exact results in d > 1. The issue is of considerable significance in

Copyright © 2003 IOP Publishing Ltd.

 connection with the resonating-valence-bond (RVB) theory of high-temperature
superconductivity. In a valence bond (VB) state, pairs of spins are in singlet spin
configurations (a singlet is often termed a VB). The RVB state is a coherent linear
superposition of VB states. In 1973, Anderson [35] in a classic paper suggested
that the ground state of the S = 12 HAFM on the frustrated triangular lattice is an
RVB state. The RVB state is a singlet (total spin is zero) and is often described
as a QSL since translational as well as rotational symmetries are preserved in the
state. The RVB state is spin disordered and the two-spin correlation function has
an exponential decay as a function of the distance between the spins. Interest in
the RVB state revived after the discovery of high-temperature superconductivity
in 1986 [36]. The common structural ingredient of the high-TC cuprate systems is
the copper-oxide (CuO2 ) plane which ideally behaves as a S = 12 HAFM defined
on a square lattice. It is largely agreed that the ground state (T = 0) has AFM
long range order (LRO). The low-lying excitations are the conventional S = 1
magnons. In the spinon picture, a magnon is a pair of confined spinons. The
spinons cannot move apart from each other unlike in one dimension. The cuprates
exhibit a rich phase diagram as a function of the dopant concentration. On doping,
positively charged holes are introduced in the CuO2 plane. The holes are mobile
in a background of antiferromagnetically interacting spins. The motion of holes
acts against antiferromagnetism and the AFM LRO is rapidly destroyed as the
concentration of holes increases. The resulting spin disordered state has been
speculated to be an RVB state. In close analogy with the S = 12 HAFM chain,
the low-lying spin excitations in the RVB state are pairs of spinons. The spinons
are created by breaking a VB. The spinons are not confined as in the case of an
ordered ground state but separate via a rearrangement of the VBs. The spinons
have spin- 12 and charge 0. The charge excitations in an RVB state are known
as holons with charge +e and spin 0. Holons are created on doping the RVB
state, i.e. replacing electrons by holes. Spinons and holons are best described as
topological excitations in a QSL. The key feature of the doped RVB state is that of
spin–charge separation, i.e. the spin and charge excitations are decoupled entities.
Spin–charge separation can be rigorously demonstrated in the case of interacting
electron systems in one dimension known by the general name of Luttinger liquids
(LLs). The Hubbard model in one dimension is the most well-known example of
an LL. The model is integrable and the BA results for the excitation spectrum
confirm that spinons and holons are the elementary excitations [36, 37].
Coming back to the RVB state, there has been an intensive search for
spin models in two dimensions with RVB states as exact ground states. Recent
calculations show that there is AFM LRO in the ground state of the S = 12
HAFM on the triangular lattice, contrary to Anderson’s original conjecture [38].
Frustrated spin models with nn as well as non-nn exchange interactions have
been constructed for which the RVB states are the exact ground states in certain
parameter regimes [39]. These are short-ranged RVB states with the VBs forming
between nn spin pairs. The spinon excitation spectrum in this case is gapped.
A model which captures the low-energy dynamics in the RVB scenario is the

Copyright © 2003 IOP Publishing Ltd.

 quantum dimer model (QDM) [40]. The Hamiltonian of the model defined on a
square lattice is given by
{
HQDM = {− t ( ) + v(
+ H.C. + ) (7.23)

where the full lines represent dimers (VBs) and the sum runs over all the
plaquettes of the lattice. The first term of the Hamiltonian is the kinetic part
representing the flipping of a pair of parallel dimers on the two bonds of a
plaquette to the other possible orientation, i.e. from horizontal to vertical and
vice versa. The second term counts the number of flippable pairs of dimers in any
dimer configuration and is analogous to the potential term of the Hamiltonian.
The ground state of the QDM on the square lattice is not, however, a QSL except
at the special point t = V . Moessner and Sondhi [41] have studied the QDM on
the triangular lattice and shown that, in contrast to the square lattice case, the
ground state is an RVB state with deconfined, gapped spinons in a finite range
of parameters. Recently, some microscopic models of 2D magnets have been
proposed [42], the low-lying excitations of which are of three types: spinons,
holons and ‘vortex-like’ excitations with no spin and charge, dubbed as visons.
Some of these models are related to the QDM. Two integrable models [42, 43]
which share common topological features with the microscopic models in two
dimensions have been constructed and have applications in fault-tolerant quantum
computation. The models, however, cannot resolve the issue of spinons in two
dimensions as quantum numbers like the total S z are not conserved in these
models. The search for microscopic models in two dimensions, with spinons as
the elementary excitations, acquires particular significance in the light of recent
experimental evidence of the spinon continuum in the 2D frustrated quantum
antiferromagnet Cs2 CuCl4 [44]. The ground state of this compound is expected
to be a QSL with spinons and not magnons as the elementary excitations. Exactly
solvable models in two dimensions are needed for a clear understanding of the
origin of the experimentally observed spinon continuum.
Real materials are often anisotropic in character. The anisotropy may be
present in the exchange interaction Hamiltonian itself or there may be additional
terms in the Hamiltonian corresponding to different types of anisotropy. A
well-known anisotropic interaction, present in many AFM materials, is the
Dzyaloshinskii–Moriya (DM) interaction with the general form

' · (S'i × S'j ).

HDM = D (7.24)

Moriya [45] provided the microscopic basis of the DM interaction by extending

Anderson’s superexchange theory to include the spin–orbit interaction. The DM
coupling acts to cant the spins because the coupling energy is minimized when the
two spins are perpendicular to each other. Some examples of materials with DM
interaction include the quasi-2D compound Cs2 CuCl4 [44], the CuO2 planes of
the undoped cuprate system La2 CuO4 [46], the quasi-1D compound Cu-benzoate

Copyright © 2003 IOP Publishing Ltd.

 [47] etc. The DM canting of spins is responsible for the small ferromagnetic
moment of the CuO2 planes even though the dominant in-plane exchange
interaction is AFM in nature. Alcaraz and Wreszinski [48] have shown that
the XXZ quantum Heisenberg chain (both FM and AFM) with DM interaction
is equivalent to the XXZ Hamiltonian with modified boundary conditions and
anisotropy parameter Jz /Jx . The DM interaction is assumed to be of the form

 N
y y x
HDM ( ) = − (σ x σ − σi σi+1 ) (7.25)
2 i=1 i i+1

where the vector√D ' in equation (7.24) is in the z-direction. The new anisotropy
parameter is δ/ 1 + 2 where δ is the anisotropy parameter of the original
XXZ Hamiltonian. With changed boundary conditions, the model is still BA
solvable. In fact, in the thermodynamic limit (N → ∞), the boundary conditions
do not affect the critical behaviour. Thus, the Hamiltonian, which includes both
the XXZ Hamiltonian and the DM interaction, has the same critical properties
and
√the phase diagram as the XXZ Hamiltonian with the anisotropy parameter
δ/ 1 + 2 .
We next turn our attention to spin-S (S > 12 ) quantum spin chains. The spin-S
Heisenberg exchange interaction Hamiltonian in one dimension is not integrable.
A family of Heisenberg-like models has been constructed for S = 1, 32 , 2, 52 , . . .
etc for which the spin-S quantum Hamiltonian is given by

Hs = Q(S'i · S'i+1 ) (7.26)
i

where Q(x) is a polynomial of degree 2S [49]. With this generalization, the

spin-S quantum spin chains are integrable. The integrable models, however, do
not distinguish between half-odd integer and integer spins. In both cases, the
integrable models have gapless excitation spectrum. For half-odd integer AFM
Heisenberg spin chains (with only the bilinear exchange interaction term), the
Lieb–Schultz–Mattis (LSM) theorem [50] states that the excitation spectrum is
gapless. The theorem cannot be proved for AFM integer spin chains. Haldane in
1983 pointed out the difference between the half-odd integer and integer AFM
Heisenberg spin chains and made the conjecture that integer spin chains have a
gap in the excitation spectrum [51]. Integer spin quantum antiferromagnets in one
dimension have been widely studied analytically, numerically and experimentally
and Haldane’s conjecture has turned out to be true. There are several examples of
quasi-1D S = 1 AFM materials which exhibit the Haldane gap (HG). Some of the
most widely studied materials are CsNiCl3 , Ni(C2 H8 N2 )2 NO2 (ClO4 ) (NENP),
Y2 BaNiO5 etc. Recently, experimental evidence of an S = 2 antiferromagnet
which exhibits the Haldane gap has been obtained. In this compound the
manganese ions form effective S = 2 spins and are coupled in a quasi-1D
chain [52]. Integrable models of integer spin chains do not reproduce the HG
but are of considerable interest since they provide exact information about the

Copyright © 2003 IOP Publishing Ltd.

 phase diagram of generalized integer spin models. Consider the generalized
Hamiltonian for an AFM S = 1 chain:

H= [cos θ (S'i · S'i+1 ) + sin θ (S'i · S'i+1 )2 ] (7.27)
i

with θ varying between 0 and 2π. The biquadratic term has been found to be
relevant in some real integer-spin materials. There are two gapped phases: the
Haldane phase for − 14 π < θ < 14 π and a dimerized phase for − 34 π < θ < − 14 π
[53]. At θ = − 14 π, the model is integrable and the gap vanishes to zero. This point
separates the two gapped phases, Haldane and dimerized, which have different
symmetry properties. Thus, a quantum phase transition occurs at ϑ = − 14 π from
the Haldane to the dimerized phase. The integrable model provides the exact
location of the transition point. The point θ = 14 π corresponds to the Hamiltonian
which is a sum over the permutation operators and is again exactly solvable.
The Haldane phase includes the isotropic Heisenberg chain (θ = 0) and the
Affleck–Kennedy–Lieb–Tasaki (AKLT) Hamiltonian (tan θVBS = 13 ) [54]. The
latter model is not integrable but the ground state is known exactly. The ground
state is described as a valence bond solid (VBS) state in which a VB (singlet)
covers every link of the chain. Since the gap does not become zero for 0 ≤ θ ≤
θVBS , there is no phase transition in going from one limiting Hamiltonian to the
other. Thus, the isotropic Heisenberg and AKLT chains are in the same phase.
The doped cuprate systems exhibit a variety of novel phenomena in their
insulating, metallic and superconducting phases. A full understanding of these
phenomena is, as yet, lacking. There is currently a strong research interest in
doped spin systems. The idea is to look for simpler spin systems in which the
consequences of doping can be studied in a less ambiguous manner. The spin-1
HG nickelate compound Y2 BaNiO5 can be doped with holes on replacing the off-
chain Y3+ ions by Ca2+ ions. Inelastic neutron scattering (INS) measurements on
the doped compound provide evidence for the appearance of new states in the HG
[55]. The structure factor S(q), obtained by integrating the dynamical correlation
function S(q, ω) over ω, acquires an incommensurate, double-peaked form in
the doped state [56]. Frahm et al [57] have constructed an integrable model
describing a doped spin-1 chain. In the undoped limit, the spectrum is gapless and
so the HG of the integer spin system is not reproduced. It is, however, possible to
reintroduce a gap in the continuum limit where a field-theoretical description of
the model is possible. The model has limited relevance in explaining the physical
features of the doped nickelate compound. Another interesting study relates to the
appearance of magnetization plateaus in the doped S = 1 integrable model [58].
The location of the plateaus depends on the concentration of holes. Experimental
evidence of this novel phenomenon has not been obtained so far.
An electron in a solid, localized around an atomic site, has three degrees
of freedom: charge, spin and orbital. The orbital degree of freedom is relevant
to several transition metal oxides which include the cuprate and manganite
systems. The latter compounds on doping exhibit the phenomenon of colossal

Copyright © 2003 IOP Publishing Ltd.

 magnetoresistance in which there is a huge change in electrical resistivity on
the application of a magnetic field. The manganites like the cuprates have a rich
phase diagram as a function of the dopant concentration [59]. We now give a
specific example of the orbital degree of freedom. The Mn3+ ion in the manganite
compound LaMnO3 has four electrons in the outermost 3d energy level. The
electrostatic field of the neighbouring oxygen ions splits the 3d energy level into
two sublevels, t2g and eg . Three of the four electrons occupy the three t2g orbitals
dxy , dyz, dzx and the fourth electron goes to the eg -sublevel containing the two
orbitals dx 2 −y 2 and d3z2 −r 2 . The fourth electron, thus, has an orbital degree of
freedom as it has two possible choices for occupying an orbital. The four electrons
have the same spin orientation to minimize the electrostatic repulsion energy
according to Hund’s rule. The total spin is, thus, S = 2. The orbital degree of
freedom is described by the pseudospin T' such that Tz = 12 (− 12 ) when the dx 2 −y 2
(d3z2 −r 2 ) orbital is occupied. The three components of the pseudospin satisfy
commutation relations similar to those of the spin components. The eg doublet is
further split into two hyperfine energy levels due to the well-known Jahn–Teller
(JT) effect. In concentrated systems, the JT effect can lead to orbital ordering
below an ordering temperature. In the antiferromagnetically ordered Néel state,
the spins are, alternately, up and down. Similarly, in the case of antiferro-orbital
ordering, the occupied orbitals alternate in type at successive sites of the lattice.
The orbital degree of freedom is frozen as a result. Apart from the JT mechanism
of orbital ordering, there is an exchange mechanism which may lead to orbital
order. The exchange mechanism is a generalization of the usual superexchange to
the case of orbital degeneracy. Starting from the degenerate Hubbard model, in
which there are two degenerate orbitals at each site, one can derive the following
generalized exchange Hamiltonian [60]:

H= {J1 S'i · S'j + J2 T'i · T'j + J3 (S'i · S'j )(T'i · T'j )}. (7.28)
ij

Consider the case J1 = J2 = J . For J3 = 0, two independent Heisenberg-like

Hamiltonians are obtained which are BA solvable. At the Kolezhuk–Mikeska
point, J3 /J = 43 , the ground state is exactly known [61]. The point J3 /J = 4
is integrable and there are three gapless excitation modes. The compounds
Na2 Ti2 Sb2 O and NaV2 O5 are examples of materials in one dimension with
coupled spin and orbital degrees of freedom [62]. These systems have been
described by anisotropic versions of the Hamiltonian in equation (7.28) but
without adequate agreement with experiments. The elementary excitations in the
orbital sector are the orbital waves or ‘orbitons’. An excitation of this type is
created in the orbitally ordered state by changing the occupied orbital at a site
and letting the defect propagate in the solid. The excitations are analogous to the
spin waves or magnons in a magnetically ordered solid. Experimental evidence
of orbital waves has recently been obtained in the manganite compound LaMnO3
through Raman scattering measurements [63,64]. As discussed before, integrable
spin models provide important links between theory and experiments. A similar

Copyright © 2003 IOP Publishing Ltd.

 scenario in the case of systems with coupled spin and orbital degrees of freedom
is yet to develop.

7.3 Ladder models

The simplest ladder model consists of two chains coupled by rungs (figure 7.2).
In general, the ladder may consist of n chains coupled by rungs. In the spin ladder
model, each site of the ladder is occupied by a spin (in general of magnitude 12 )
and the spins interact via the Heisenberg AFM exchange interaction. In the doped
spin ladder model, some of the sites are empty, i.e. occupied by holes. The holes
can move in the background of interacting spins. There are two major reasons
for the considerable research interest in ladders. Powerful techniques like the BA
and bosonization are available for the study of 1D many-body systems whereas
practically very few rigorous results are known for 2D systems. Ladders provide
a bridge between 1D and 2D physics and are ideally suited to study how the
electronic and magnetic properties change as one goes from a single chain to the
square lattice. The unconventional properties of the CuO2 planes of the cuprate
systems are the main reason for the significant interest in 2D many-body systems.
Many of these properties are ascribed to strong correlation effects. Ladders are
simpler systems in which some of the issues associated with strong correlation
can be addressed in a more rigorous manner. The second motivation for the study
of ladder systems is that several such systems have been discovered in the recent
past. In the following, we describe in brief some of the major physical properties
of ladders. There are two exhaustive reviews on ladders which provide more
detailed information [65, 66].
Consider a two-chain spin ladder described by the AFM Heisenberg
exchange interaction Hamiltonian

H= Jij S'i · S'j . (7.29)
"ij #

The nn intra-chain and the rung exchange interactions are of strength J and
JR respectively. When JR = 0, one obtains two decoupled AFM spin chains for
which the excitation spectrum is known to be gapless. For all JR /J > 0, a gap

JR JR

Figure 7.2. A two-chain ladder. The rung and intra-chain nn exchange interactions are of
strength JR and J respectively.

Copyright © 2003 IOP Publishing Ltd.

 (the so-called spin gap (SG)) opens up in the spin excitation spectrum. The result
is easy to understand in the simple limit in which the exchange coupling JR along
the rungs is much stronger than the coupling J along the chains. The intra-chain
coupling may, thus, be treated as a perturbation. When J = 0, the exact ground
state consists of singlets along the rungs. The ground-state energy is −3JR N /4,
where N is the number of rungs in the ladder. The ground state has total spin
S = 0. In first-order perturbation theory, the correction to the ground-state energy
is zero. An S = 1 excitation may be created by promoting one of the rung singlets
to an S = 1 triplet. The weak coupling along the chains gives rise to a propagating
S = 1 magnon. In first-order perturbation theory, the dispersion relation is

ω(k) = JR + J cos k (7.30)

where k is the momentum wavevector. The SG, defined as the minimum excitation
energy, is given by
SG = ω(π)  (JR − J ). (7.31)
The two-spin correlations decay exponentially along the chains showing that the
ground state is a QSL. The magnons can further form bound states. Experimental
evidence of two-magnon bound states has been obtained in the S = 12 two-
chain ladder compound Ca14−x Lax Cu24 O41 (x = 5 and 4) [67]. The family of
compounds Srn−1 Cun+1 O2n consists of planes of weakly-coupled ladders of
(n + 1)/2 chains [68]. For n = 3 and 5, respectively, one gets the two-chain
and three-chain ladder compounds SrCu2 O3 and Sr2 Cu3 O5 respectively. For
the first compound, experimental evidence of the SG has been obtained. The
latter compound has properties similar to those of the 1D Heisenberg AFM
chain [69]. A recent example of a spin ladder belonging to the organic family
of materials is the compound (C5 H12 N)2 CuBr4 , a ladder system with strong
rung coupling (JR /J  3.5) [70]. The phase diagram of the AFM spin ladder
in the presence of an external magnetic field is particularly interesting. In the
absence of the magnetic field and at T = 0, the ground state is a QSL with a
gap in the excitation spectrum. At a field Hc1 , there is a transition to a gapless
LL phase (gµB Hc1 = SG , the spin gap, µB is the Bohr magneton and g the
Landé splitting factor). There is another transition at an upper critical field Hc2
to a fully polarized FM state. Both Hc1 and Hc2 are quantum critical points.
The quantum phase transition from one ground state to another is brought
about by changing the magnetic field. At small temperatures, the behaviour
of the system is determined by the crossover between two types of critical
behaviour: quantum critical behaviour at T = 0 and classical critical behaviour
at T = 0. Quantum effects are persistent in the crossover region at small finite
temperatures and such effects can be probed experimentally. In the case of the
ladder system (C5 H12 N)2 CuBr4 , the magnetization data obtained experimentally
exhibit universal scaling behaviour in the vicinity of the critical fields Hc1 and
Hc2 . In the gapless regime Hc1 < H < Hc2 , the ladder model can be mapped onto
an XXZ chain, the thermodynamic properties of which can be calculated exactly

Copyright © 2003 IOP Publishing Ltd.

 by the BA. The theoretically computed magnetization M versus magnetic field h
curve is in excellent agreement with the experimental data. Organic spin ladders
provide ideal testing grounds for the theories of quantum phase transitions. For
inorganic spin ladder systems, the value of Hc1 is too high to be experimentally
accessible.

Bose and Gayen [71] have studied a frustrated two-chain spin model with
diagonal couplings. The intra-chain and diagonal spin–spin interactions are of
equal strength J . It is easy to show that for JR ≥ 2J the exact ground state
consists of singlets (dimers) along the rungs with the energy Eg = −3JR N /4
where N is the number of rungs. Xian [72] later pointed out that as long as
JR /J > (JR /J )c  1.401, the rung dimer state is the exact ground state. At
JR /J = (JR /J )c , there is a first-order transition from the rung dimer state to the
Haldane phase of the S = 1 chain. Kolezhuk and Mikeska [73] have constructed
a class of generalized S = 12 two-chain ladder models for which the ground
state can be determined exactly. The Hamiltonian H is a sum over plaquette
Hamiltonians and each such Hamiltonian contains various two-spin as well as
four-spin interaction terms. They have further introduced a toy model which has
a rich phase diagram in which the phase boundaries can be determined exactly.

The standard spin ladder models with bilinear exchange are not integrable.
For integrability, multispin interaction terms have to be included in the
Hamiltonian. Some integrable ladder models have already been constructed [74].
We discuss one particular model proposed by Wang [75]. The Hamiltonian is
given by

J1 
N
J2 N
H= ['
σj · σ'j +1 + τ'j · τ'j +1 ] + σ'j · τ'j
4 i=1 2 j =1

U1 N
+ σj · σ'j +1 )('
(' τj · τ'j +1 )
4 j =1

U2 N
+ σj · τ'j )('
(' σj +1 · τ'j +1 ) (7.32)
4 j =1

where σ'j and τ'j are the Pauli matrices associated with the site j of the upper
and lower chains respectively. N is the total number of rungs in the system. The
ordinary spin ladder Hamiltonian is obtained from equation (7.32) when the four
spin terms are absent, i.e. U1 = U2 = 0. For general parameters J1 , J2 , U1 and U2 ,
the model is non-integrable. The integrable cases correspond to U1 = J1 , U2 = 0
or U1 = J1 , U2 = −J1 /2. Without loss of generality, one can put J1 = U1 = 1,

Copyright © 2003 IOP Publishing Ltd.

 J2 = J and U2 = U . For U = 0, the Hamiltonian (7.32) reduces to

1 N
H= (1 + σ'j · σ'j +1 )(1 + τ'j · τ'j +1 )
4 j =1
 
J  N
1 1
+ σj · τ'j − 1) +
(' J− N. (7.33)
2 j =1 2 2

Three quantum phases are possible. For J > J+c = 2, the system exists in the
rung dimerized phase. The ground state is a product of singlet rungs. The SG
is given by SG = 2(J − 2). For J+c > J > J−c , a gapless phase is obtained with
three branches of gapless excitations. J+c is the quantum critical point at which
a QPT from the dimerized phase to the gapless phase occurs. In the vicinity
of the quantum critical point, the susceptibility and the specific heat can be
calculated using the thermodynamic BA. From the low-temperature expansion
of the thermodynamic BA equation, one obtains
1 1
C ∼T 2 χ ∼ T −2 (7.34)

which are typical of quantum critical behaviour. In the presence of an external

magnetic field h, the magnetic field can be tuned to drive a QPT at the quantum
critical point hc = 2(J − 2) from the gapless phase to a gapped phase. The third
quantum phase (h = 0) is obtained for
π ln 3
J < J−c = − √ + .
4 3 4

This is a gapless phase with two branches of gapless excitations. For U = −12 , a
similar phase diagram is obtained. Note that the ladder model may equivalently
be considered as a spin-orbital model with σ' and τ' representing the spin and the
pseudospin.
Doped ladder models are toy models of strongly correlated systems [65].
In these systems, the double occupancy of a site by two electrons, one with
spin up and the other with spin down, is prohibited due to strong Coulomb
correlations. In a doped spin system, there is a competition between two
processes: hole delocalization and exchange energy minimization. A hole moving
in an antiferromagnetically ordered spin background, say the Néel state, gives
rise to parallel spin pairs which raise the exchange interaction energy of the
system. The questions of interest are: whether a coherent motion of the holes
is possible; whether two holes can form a bound state; the development of
superconducting (SC) correlations; the possibility of phase separation of holes etc.
Some of these issues are of significant relevance in the context of doped cuprate
systems in which charge transport occurs through the motion of holes [76]. In the
SC phase, the holes form bound pairs with possibly d-wave symmetry. Several
proposals have been made so far on the origin of hole binding but there is, as

Copyright © 2003 IOP Publishing Ltd.

 yet, no general consensus on the actual binding mechanism. The doped cuprate
systems exist in a ‘pseudogap’ phase before the SC phase is entered. In fact,
some cuprate systems also exhibit SG. As already mentioned, the doped two-
chain ladder systems are characterized by a SG. The issue of how the gap evolves
on doping is of significant interest. The possibility of binding hole pairs in a two-
chain ladder system was first pointed out by Dagotto et al [77]. In this case,
the binding mechanism is not controversial and can be understood in a simple
physical picture. Again, consider the case JR * J , i.e. a ladder with dominant
exchange interactions along the rungs. In the ground state, the rungs are mostly
in singlet spin configurations. On the introduction of a single hole, a singlet spin
pair is broken and the corresponding exchange interaction energy is lost. When
two holes are present, they prefer to be on the same rung to minimize the loss in
the exchange interaction energy. The holes thus form a bound pair. In the more
general case, detailed energy considerations show that the two holes tend to be
close to each other and effectively form a bound pair. For more than two holes,
several calculations suggest that considerable SC pairing correlations develop
in the system on doping. True superconductivity can be obtained only in the
bulk limit. Theoretical predictions motivated the search for ladder compounds
which can be doped with holes. Much excitement was created in 1996 when the
ladder compound Sr14−x Cax Cu24 O41 was found to become SC under pressure at
x = 13.6 [78]. The transition temperature Tc is ∼12 K at a pressure of 3 GPa. As
in the case of cuprate systems, bound pairs of holes are responsible for charge
transport in the SC phase. Experimental results on doped ladder compounds point
out strong analogies between the doped ladder and cuprate systems [65].
The strongly correlated doped ladder system is described by the t–J
Hamiltonian
 
Ht –J = − ++ C
+ Jij (S'i · S'j − 14 ni nj ).
tij (Ciσ j σ + H.C.) + (7.35)
"ij #,σ "ij #

The C++ and C+iσ are the electron creation and annihilation operators which act in
iσ
the reduced Hilbert space (no double occupancy of sites),
++ = C + (1 − ni−σ )
Ciσ iσ
(7.36)
+iσ = Ciσ (1 − ni−σ )
C
where σ is the spin index and ni , nj are the occupation numbers of the ith and j th
sites respectively. The first term in equation (7.35) describes the motion of holes
with hopping integrals tR and t for motion along the rung and chain respectively.
In the standard t–J ladder model, i and j are nn sites. The second term contains
the usual AFM Heisenberg exchange interaction Hamiltonian. The t–J model,
thus, describes the motion of holes in a background of antiferromagnetically
interacting spins. A large number of studies have been carried out on t–J
ladder models. These are reviewed in [65, 66]. We describe briefly some of the
major results. The SG of the undoped ladder changes discontinuously on doping.

Copyright © 2003 IOP Publishing Ltd.

 Remember that the SG is the difference in energies of the lowest triplet excitation
and the ground state. In the doped state, there are two distinct triplet excitations.
One triplet excitation is that of the undoped ladder obtained by exciting a rung
singlet to a rung triplet. A new type of triplet excitation is possible when at least
two holes are present. On the introduction of two holes in two rung singlets, a pair
of free spin- 12 s is obtained which combines to give rise to a singlet (S = 0) or a
triplet (S = 1) state. The triplet configuration of the two free spins corresponds to
the second type of triplet excitation. The SG of this new excitation is unrelated
to the SG of the magnon excitation. The true SG is the one which has the lowest
value in a particular parameter regime.
The low-energy modes of a ladder system are characterized by their spin.
Singlet and triplet excitations correspond to charge and spin modes respectively.
In each sector, the hole may further be in a bonding or antibonding state with
opposite parities. We consider only the even-parity sector to which the lowest-
energy excitations belong. In both the S = 0 and S = 1 sectors, an excitation
continuum with well-defined boundaries is present. The S = 0 and S = 1 continua
are degenerate in energy. A bound-state branch with S = 0 splits off below
the continuum, the lowest energy of which corresponds to the centre-of-mass
momentum wavevector K = 0 [79, 80]. Thus the two-hole ground state is in
the singlet sector and corresponds to a bound state of two holes with K = 0.
The bound state has d-wave type symmetry. Within the bound-state branch,
excitations with energy infinitesimally close to the ground state are possible.
These excitations are the charge excitations since the total spin is still zero and
the charge excitation spectrum is gapless. The lowest spin excitations in a wide
parameter regime are between the S = 0 ground state and the lowest energy state
in the S = 1 continuum [81]. The continuum does not exist in the undoped ladder
and so the SG evolves discontinuously on doping in this parameter regime. A
suggestion has, however, been made that the lowest triplet excitation is a bound
state of a magnon with a pair of holes [82]. In summary, the two-chain ladder
model has the feature that the charge excitation is gapless but the spin excitation
has a gap. This is the Luther–Emery phase and is different from the LL phase in
which both the spin and charge excitations are gapless.
Bose and Gayen [83] have derived several exact, analytical results for the
ground-state energy and the low-lying excitation spectrum of the frustrated t–J
ladder doped with one and two holes. The undoped ladder model has already
been described. In the doped case, the hopping integral has the value tR for
hole motion along the rungs and the intra-chain and diagonal hopping integrals
are of equal strength t. The latter assumption is crucial for the exact solvability
of the eigenvalue problem in the one- and two-hole sectors. Though the model
differs from the standard t–J ladder model (the diagonal couplings are missing
in the latter), the spin and charge excitation spectra exhibit similar features. In
particular, the dispersion relation of the two-hole bound-state branch is obtained
exactly and the exact ground state is shown to be a bound state of two holes
with K = 0 and d-wave type symmetry. The ladder exists in the Luther–Emery

Copyright © 2003 IOP Publishing Ltd.

 phase. There is no spin charge separation, as in the case of a LL. In the exact
hole eigenstates, the hole is always accompanied by a free spin- 12 . The hole–hole
correlation function can also be calculated exactly. When JR * J , the holes of a
bound pair are predominantly on the same rung. For lower values of JR , the holes
prefer to be on nn rungs so that energy gain through the delocalization of a hole
along the rung is possible.
The t–J ladder model constructed by Bose and Gayen is not integrable.
Frahm and Kundu [84] have constructed a t–J ladder model which is integrable.
The Hamiltonian is given by
 (a)
H= Ht –J + Hint + Hrung − µ

n. (7.37)
a

The two chains of the ladder are labelled by a = 1, 2 and µ is the chemical
(a)
potential coupling to the number of electrons in the system. Ht −J is the t–J
(a) (a) (a)
Hamiltonian (7.35) for a chain plus the terms nj + nj +1 where nj is the total
number of electrons on site j :
 (1) (2)
Hint = − [Ht –J ]jj +1 [Ht –J ]jj +1 . (7.38)
j

Hrung includes the t–J Hamiltonian (7.35) corresponding to a rung and a Coulomb

interaction term V j n(1) (2)
j nj . The possible basis states of a rung are the
following. When no hole is present, a rung can be in a singlet or a triplet spin
configuration. When a single hole is present, the rung is in a bonding (|σ+ #)
or antibonding (|σ− #) state with |σ± # ≡ √1 (|σ 0# ± |0σ #) and σ =↑ or ↓. The
2
rung can further be occupied by two holes. Frahm and Kundu have studied the
phase diagram of the ladder model at low temperatures and in the strong coupling
regime JR * 1, V * µ + |tR | near half-filling. In this regime, the triplet states
are unfavourable. By excluding the triplet states and choosing J = 2t = 2, the
Hamiltonian H (7.37) can be rewritten as
 
5
H =− jj +1 − Al Nl + constant (7.39)
j l=1

where Nl , l = 1, 2 (3, 4) is the number of bonding (antibonding) single-hole rung

states with spin ↑, ↓ and N5 is the number of empty
rungs. If L is the total number
of rungs in the ladder, the remaining N0 = L − l Nl rungs are in singlet spin
configurations. The permutation operator j k interchanges the states on rungs j
and k. If both the rungs are singly occupied by a hole, an additional minus sign is
obtained on interchanging the rung states. The potentials Al s are
A1 = A2 ≡ µ+ = tR − µ + V (7.40)
A3 = A4 ≡ µ− = −tR − µ + V (7.41)
A5 ≡ V+ = −2µ + V . (7.42)

Copyright © 2003 IOP Publishing Ltd.

 The nature of the ground state and the low-lying excitation spectrum depends on
the relative strengths of the potentials Al s. The Hamiltonian (7.39) is BA solvable.
The phase diagram V versus the hole concentration nh has been computed for
µ+ = µ− , i.e. tR = 0. For large repulsive V , the ground state can be described
as a Fermi sea of single-hole states |σ± # propagating in a background of rung
dimer states |s#. The double-hole rung states |d# are energetically favourable for
sufficiently strong attractive rung interactions. In the intermediate region, both
types of hole rung states are present. In the frustrated t–J ladder model studied
by Bose and Gayen [83], the exact two-hole ground state is a linear combination
of single-hole and double-hole rung states propagating in a background of rung
dimer states. The single-hole rung states are the bonding states.
In a remarkable paper, Lin et al [85] have considered the problem of
electrons hopping on a two-chain ladder. The interaction between the electrons is
sufficiently weak and finite-ranged. At half filling, a perturbative renormalization
group (RG) calculation shows that the model scales onto the Gross–Neveu (GN)
model which is integrable and has SO(8) symmetry. At half filling, the two-
chain ladder is in the Mott insulating phase with d-wave pairing correlations.
The insulating phase is, furthermore, a QSL. The integrability has been utilized
to determine the exact energies and quantum numbers of all the low-energy
excitations which constitute the degenerate SO(8) multiplets. The lowest-lying
excitations can be divided into three octets all with a non-zero gap (mass
gap) m. Each excitation has a dispersion 1 (q) = m2 + q 2 where q is the
momentum variable measured with respect to the minimum energy value. One
octet consists of two-particle excitations: two charge ±2e Cooper pairs around
zero momentum, a triplet of S = 1 magnons around momentum (π, π) and three
neutral S = 0 particle–hole pair excitations. SO(8) transformations rotate the
components of the vector multiplet into one another unifying the excitations in
the process. The SO(5) subgroup which rotates only the first five components of
the vector is the symmetry proposed by Zhang [86] to unify antiferromagnetism
and superconductivity in the cuprates. The vector octet is related by a triality
symmetry to two other octets with mass gap m. The 16 particles of these two
octets have the features of quasi-electrons and quasi-holes. Above the 24 √ states
with mass gap m, there are other higher-lying ‘bound’ states with mass gap 3m.
Finally, the continuum of scattering states occurs above the energy 2m. Lin et al
have also studied the effects of doping a small concentration of holes into the
Mott insulating phase. In this limit, the effect of doping can be incorporated in
the GN model by adding a term −µQ to the Hamiltonian, µ being the chemical
potential and Q the total charge. Integrability of the GN model is not lost as Q is
a global SO(8) generator. Doping is possible only for 2µ > m when Cooper pairs
enter the system. The doped ladder exists in the Luther–Emery phase, whereas
in the half-filled insulating limit both the spin and charge excitations are gapped.
In the doped phase, the Cooper pairs can transport charge and quasi-long-range
d-wave SC pairing correlations develop in the system. The other features of the
standard t–J ladder model, e.g. the discontinuous evolution of the SG on doping,

Copyright © 2003 IOP Publishing Ltd.

 are reproduced. The lowest triplet excitation is a bound state of an S = 1 magnon
with a Cooper pair. As mentioned before, a similar result has been obtained
numerically in the case of the standard t–J ladder [82].√The triplet excitation
belongs to the family of 28 excitations with mass gap 3m. If x denotes the
dopant
√ concentration, then the SG jumps from S (x = 0) = m to S (x = 0+ ) =
( 3 − 1)m upon doping. The integrability of the weakly-interacting two-chain
ladder model has yielded a plethora of exact results which illustrate the rich
physics associated with undoped and doped ladders.

7.4 Concluding remarks

Integrable models have a dual utility. They serve as testing grounds for
approximate methods and techniques. Also, they are often models of real systems
and provide rigorous information about the physical properties of such systems.
Integrable models are sometimes more general than what is necessary to describe
real systems. In such cases, an integrable model corresponds to an exactly solvable
point in the general phase diagram. The point may be a quantum critical point
at which transition from one quantum phase to another occurs or the integrable
model may be in the same phase as a more realistic model. In the latter case,
the physical properties of the two models are similar. In this review, we have
discussed the physical basis of some integrable spin models with special focus on
the relevance of the models to real systems. The Heisenberg spin chain is probably
the best example of the essential role played by exact solvability in correctly
interpreting the experimental data. The concept of spinons owes its origin to
the exact analysis of the BA equations. The theoretical prediction motivated
the search for real spin systems in which experimental confirmation could be
made. In this review, examples are also given of systems for which the links
between integrable models and experimental results are not well established. A
major portion of the review is devoted to physical systems which exhibit rich
phenomena, like the systems with both spin and orbital degrees of freedom and
undoped and doped spin ladder systems, where the need for integrable systems is
particularly strong. These systems exhibit a variety of novel phenomena, a proper
understanding of which should be based on rigorous theory. Two-dimensional
spin systems with QSL ground states have been specially mentioned to explain the
recent interest in constructing integrable models of such systems. The review is
meant to be an elementary introduction to the genesis and usefulness of integrable
models vis-à-vis physical spin systems. Future challenges are also highlighted to
motivate further research on integrable models.
There are some AFM spin models which are not integrable but for which the
ground states and, in some cases, the low-lying excited states are known exactly.
The most prominent amongst these are the Majumdar–Ghosh (MG) chain [87]
and the AKLT [54] model respectively. The MG Hamiltonian is defined in one
dimension for spins of magnitude 12 . The Hamiltonian includes both nn as well

Copyright © 2003 IOP Publishing Ltd.

 as nnn interactions. The strength of the latter is half that of the former. The exact
ground state is doubly degenerate and the states consist of singlets along alternate
links of the lattice. The excitation spectrum is not exactly known and has been
calculated on the basis of a variational wavefunction [88]. Generalizations of
the MG model to two dimensions with exactly-known ground states are possible
[39,89–91]. The Shastry–Sutherland model [89] is of much current interest due to
the recent discovery of the compound SrCu2 (BO3 )2 which is well described by the
model [92]. Some of these models including the AKLT model have been reviewed
in [93–96] from which more information about the models can be obtained.
These models incorporate physical features of real systems and provide valuable
insight into the magnetic properties of low-dimensional quantum spin systems.
The models supplement integrable models in obtaining exact information and
provide motivation for the construction of integrable generalizations.

References
[1] Bethe H 1931 Z. Phys. 71 205
See also Mattis D C (ed) 1993 The Many Body Problem: An Encyclopedia of Exactly
Solved Models in One Dimension (Singapore: World Scientific) for an English
translation of Bethe’s paper
[2] Takhtajan L A and Faddeev L D 1979 Russian Math. Surveys 34 11
Faddeev L D 1980 Sov. Sci. Rev. C 1 107
[3] Lieb E and Liniger W 1963 Phys. Rev. 130 1605
Lieb E 1963 Phys. Rev. 130 1616
Gaudin M 1967 Phys. Lett. A 24 55
Yang C N and Yang C P 1969 J. Math. Phys. 10 1115
[4] Lieb E and Wu F Y 1968 Phys. Rev. Lett. 20 1445
[5] Sutherland B 1975 Phys. Rev. Lett. 34 1083
Sutherland B 1975 Phys. Rev. Lett. 35 185
[6] Sutherland B 1975 Phys. Rev. B 12 3795
[7] Andrei N 1980 Phys. Rev. Lett. 45 379
[8] Wiegmann P B 1981 J. Phys. C: Solid State Phys. 14 1463
Filyov V M, Tsvelick A M and Wiegmann P G 1981 Phys. Lett. A 81 175
[9] Bares P A and Blatter G 1990 Phys. Rev. Lett. 64 2567
[10] González T, Martin-Delgado M A, Sierra G and Vozmediano A H 1995 Quantum
Electron Liquids and High-Tc Superconductivity (Berlin: Springer) ch 10
[11] Sutherland B 1985 Exactly Solvable Problems in Condensed Matter and Relativistic
Field Theory ed B S Shastry, S S Jha and V Singh (Berlin: Springer) p 1
[12] Gaudin M 1971 Phys. Rev. Lett. 26 1301
[13] Takhtajan L A 1985 Exactly Solvable Problems in Condensed Matter and Relativistic
Field Theory ed B S Shastry, S S Jha and V Singh (Berlin: Springer) p 175
[14] Bogoliubov N M, Izergin A G and Korepin V E 1985 Exactly Solvable Problems in
Condensed Matter and Relativistic Field Theory ed B S Shastry, S S Jha and V
Singh (Berlin: Springer) p 220
See also Korepin V E, Bogoliubov N M and Izergin A G 1993 QISM and Correlation
Functions (Cambridge: Cambridge University Press)

Copyright © 2003 IOP Publishing Ltd.

 [15] Izyumov Yu A and Skryabin Yu N 1988 Statistical Mechanics of Magnetically
Ordered Systems (New York: Consultants Bureau) ch 5 and references therein
[16] Thacker H B 1981 Rev. Mod. Phys. 53 253
[17] Kundu A 1998 Indian J. Phys. B 72 283
[18] Baxter R J 1972 Ann. Phys. 70 323
[19] Johnson J D, Krinsky S and McCoy B M 1973 Phys. Rev. A 8 2526
[20] Torrance J B and Tinkham M 1969 Phys. Rev. 187 587
Torrance J B and Tinkham M 1969 Phys. Rev. 187 59
[21] Nicoli D F and Tinkham M 1974 Phys. Rev. B 9 3126
[22] Karbach M, Hu K and Müller G 1998 Comput. Phys. 12 565 Preprint cond-
mat/9809163
[23] Faddeev L D and Takhtajan L A 1981 Phys. Lett. A 85 375
See also Majumdar C K 1985 Exactly Solvable Problems in Condensed Matter and
Relativistic Field Theory ed B S Shastry, S S Jha and V Singh (Berlin: Springer)
p 142
[24] Tennant D A, Perring T G, Cowley R A and Nagler S E 1993 Phys. Rev. Lett. 70 4003
[25] Müller G 1982 Phys. Rev. B 26 1311
Mohan M and Müller G 1983 Phys. Rev. B 27 1776
[26] Bougourzi A H, Karbach M and Müller G 1998 Phys. Rev. B 57 11429
[27] Karbach M et al 1997 Phys. Rev. B 55 12510
[28] Tennant D A, Cowley R A, Nagler S E and Tsvelik A M 1995 Phys. Rev. B 52 13368
Dender D C et al 1996 Phys. Rev. B 53 2583
Coldea R et al 1997 Phys. Rev. Lett. 79 151
Hammar P R et al 1999 Phys. Rev. B 59 1008
[29] Arai M et al 1996 Phys. Rev. Lett. 77 3649
Fabricius K et al 1998 Phys. Rev. B 57 1102
[30] Nagler S E, Buyers W J L, Armstrong R L and Briat B 1983 Phys. Rev. B 27 1784
Nagler S E, Buyers W J L, Armstrong R L and Briat B 1983 Phys. Rev. 28 3873
Buyers W J, Hogan M J, Armstrong R L and Briat B 1986 Phys. Rev. B 33 1727
[31] Ishimura N and Shiba H 1980 Prog. Theor. Phys. 63 743
[32] Bose I and Chatterjee S 1983 J. Phys. C: Solid State Phys. 16 947
[33] Bose I and Ghosh A 1996 J. Phys.: Condens. Matter 8 351
Matsubara F and Inawashiro S 1991 Phys. Rev. B 43 796
Goff J P, Tennant D A and Nagler S E 1995 Phys. Rev. B 52 15992
[34] Haldane F D M 1988 Phys. Rev. Lett. 60 635
Shastry B S 1988 Phys. Rev. Lett. 60 639
See also Ha Z N C 1996 Quantum Many-Body Systems in One Dimension (Singapore:
World Scientific)
[35] Anderson P W 1973 Mater. Res. Bull. 8 153
See also Fazekas P and Anderson P W 1974 Phil. Mag. 30 432
[36] Anderson P W 1997 The Theory of Superconductivity in the High-Tc Cuprates
(Princeton: Princeton University Press)
[37] Korepin V E and Essler F H L (ed) 1994 Exactly Solvable Models of Strongly
Correlated Electrons (Singapore: World Scientific)
[38] Huse D A and Elser V 1988 Phys. Rev. Lett. 60 2531
Bernu B, Lhuillier C and Pierre L 1992 Phys. Rev. Lett. 69 2590
[39] Bose I 1992 Phys. Rev. B 45 13072
Bose I and Ghosh A 1997 Phys. Rev. B 56 3149

Copyright © 2003 IOP Publishing Ltd.

 [40] Rokhsar D S and Kivelson S 1988 Phys. Rev. Lett. 61 2376
[41] Moessner R and Sondhi S 2001 Phys. Rev. Lett. 86 1881
[42] Nayak C and Shtengel K 2001 Phys. Rev. B 64 064422
Balents L, Fisher M P A and Girvin S M 2002 Phys. Rev. B 65 224412
[43] Kitaev Yu A 2003 Annals Phys. 303 2
[44] Coldea R et al. 2001 Phys. Rev. Lett. 86 1335
Coldea R et al 2002 Phys. Rev. Lett. 88 137203
[45] Moriya T 1963 Magnetism ed G T Rado and H Suhl (New York: Academic)
[46] Cheong S W, Thompson J D and Fisk Z 1989 Phys. Rev. B 39 4395
[47] Oshikawa M and Affleck I 1997 Phys. Rev. Lett. 79 2883
[48] Alcaraz F C and Wreszinski W F 1990 J. Stat. Phys. 58 45
[49] Takhtajan L A 1982 Phys. Lett. A 87 479
Babujian H M 1982 Phys. Lett. A 90 479
[50] Lieb E, Schultz T D and Mattis D C 1961 Ann. Phys. 16 407
[51] Haldane F D M 1983 Phys. Rev. Lett. 50 1153
Haldane F D M 1983 Phys. Lett. A 93 464
[52] Granroth G E et al 1996 Phys. Rev. Lett. 77 1616
[53] Mila F and Zhang F C 2000 Preprint cond-mat/0006068
[54] Affleck I, Kennedy T, Lieb E H and Tasaki H 1987 Phys. Rev. Lett. 59 799
[55] Di Tusa J F et al 1994 Phys. Rev. Lett. 73 1857
[56] Xu G et al 2000 Science 289 419
See also Bose I and Chattopadhyay E 2001 Int. J. Mod. Phys. B 15 2535
[57] Frahm H, Pfannmüller M P and Tsvelik A M 1998 Phys. Rev. Lett. 81 2116
[58] Frahm H and Sobiella C 1999 Phys. Rev. Lett. 83 5579
[59] Tokura Y and Nagaosa N 2000 Science 288 462
Khomskii D I and Sawatzky G A 1997 Solid State Commun. 102 87
[60] Khomskii D I 2001 Int. J. Mod. Phys. B 15 2665
[61] Kolezhuk A K and Mikeska H J 1998 Phys. Rev. Lett. 80 2709
[62] Axtell E, Ozawa T, Kauzlarich S and Singh R R P 1997 J. Solid State Chem. 134 423
Isobe M and Ueda Y 1996 J. Phys. Soc. Japan 65 1178
Fujii Y et al 1997 J. Phys. Soc. Japan 66 326
[63] Allen P B and Perebeinos V 2001 Nature 410 155
[64] Saitoh E et al 2001 Nature 410 180
[65] Dagotto E 1999 Rep. Prog. Phys. 62 1525
[66] Dagotto E and Rice T M 1996 Science 271 618
[67] Windt M et al 2001 Phys. Rev. Lett. 87 127002
[68] Rice T M, Gopalan S and Sigrist M 1993 Europhys. Lett. 23 445
[69] Azuma M et al 1994 Phys. Rev. Lett. 73 3463
[70] Watson B C et al 2001 Phys. Rev. Lett. 86 5168
[71] Bose I and Gayen S 1993 Phys. Rev. B 48 10653
[72] Xian Y 1995 Phys. Rev. B 52 12485
[73] Kolezhuk A K and Mikeska H J 1998 Int. J. Mod. Phys. B 12 2325
[74] Frahm H and Rödenbeck C 1996 Europhys. Lett. 33 47
Frahm H and Rödenbeck C 1997 J. Phys. A: Math. Gen. 30 4467
Albeverio S, Fei S M and Wang Y 1999 Europhys. Lett. 47 364
Batchelor M T and Maslen M 1999 J. Phys. A: Math. Gen. 32 L377
Tonel A P et al 2001 Preprint cond-mat/0105302 and references therein
[75] Wang Y 1999 Phys. Rev. B 60 9236

Copyright © 2003 IOP Publishing Ltd.

 [76] Ornstein J and Millis A J 2000 Science 288 468
[77] Dagotto E, Riera J and Scalapino D 1992 Phys. Rev. B 45 5744
[78] Uehara M et al 1996 J. Phys. Soc. Japan 65 2764
[79] Tsunetsugu H, Troyer M and Rice T M 1994 Phys. Rev. B 49 16078
[80] Troyer M, Tsunetsugu H and Rice T M 1996 Phys. Rev. B 53 251
[81] Jurecka C and Brenig W 2001 Preprint cond-mat/0107365
[82] Poilblanc D et al 2000 Phys. Rev. B 62 R14633
[83] Bose I and Gayen S 1994 J. Phys.: Condens. Matter 6 L405
Bose I and Gayen S 1999 J. Phys.: Condens. Matter 11 6427
[84] Frahm H and Kundu A 1999 J. Phys.: Condens. Matter 11 L557
[85] Lin H, Balents L and Fisher M P A 1998 Phys. Rev. B 58 1794
[86] Zhang S C 1997 Science 275 1089
[87] Majumdar C K and Ghosh D K 1969 J. Math. Phys. 10 1388, 1399
Majumdar C K 1970 J. Phys. C: Solid State Phys. 3 911
[88] Shastry B S and Sutherland B 1981 Phys. Rev. Lett. 47 964
[89] Shastry B S and Sutherland B 1981 Physica B 108 1069
[90] Bose I and Mitra P 1991 Phys. Rev. B 44 443
See also Bhaumik U and Bose I 1995 Phys. Rev. B 52 12489
Ghosh A and Bose I 1997 Phys. Rev. B 55 3613
[91] Siddharthan R 1999 Phys. Rev. B 60 R9904
Kumar B 2002 Phys. Rev. B 66 024406
[92] Miyahara S and Ueda K 1999 Phys. Rev. Lett. 82 3701
[93] Bose I 2001 Field Theories in Condensed Matter Physics ed S Rao (India: Hindustan
Book Agency) p 359
[94] Auerbach A 1994 Interacting Electrons and Quantum Magnetism (New York:
Springer)
[95] Affleck I 1989 J. Phys.: Condens. Matter 1 3047
[96] Bose I 2001 Quantum magnets: a brief overview Preprint cond-mat/0107399

Copyright © 2003 IOP Publishing Ltd.

 Chapter 8

Exact solvability in contemporary physics

Angela Foerster† , Jon Links‡ and Huan-Qiang Zhou‡
† Instituto
de Fı́sica da UFRGS, Porto Alegre, Brasil
‡ Department of Mathematics, The University of Queensland,

Australia

8.1 Introduction
The current realization of nanotechnology as a viable industry is presenting a
wealth of challenging problems in theoretical physics. Phenomena such as Bose–
Einstein condensation, entanglement and decoherence in the context of quantum
information, superconducting correlations in metallic nanograins, soft condensed
matter, the quantum Hall effect, nano-optics, the Kondo effect and Josephson
tunnelling phenomenon are all emerging to paint a vast canvas of interwoven
physical theories which provide hope and expectation that the emergence of new
nanotechnologies will be rapid in the short-term future. A significant tool in the
evolution of the theoretical aspects of these studies has been the development
and application of potent mathematical techniques, which are becoming ever
increasingly important as our understanding of the complexities of these physical
systems matures.
One approach that has recently been raised to prominence in this regard is
that of the exact solution of a physical model. The necessity of studying the exact
solution has been demonstrated through the experimental research on aluminium
grains with dimensions at the nanoscale level. The work of Ralph, Black and
Tinkham (RBT) [1] in 1996 detected the presence of superconducting pairing
correlations in metallic nanograins which manifest as a parity effect in the energy
spectrum dependent on whether the number of valence electrons on each grain
is even or odd. A naı̈ve approach to describe these systems theoretically is to
apply the theory of superconductivity due to Bardeen, Cooper and Schrieffer
(BCS) [2]. Indeed, the BCS model is the appropriate model for these systems
but the associated mean field treatment fails. This is because a mean field

Copyright © 2003 IOP Publishing Ltd.

 theory approximates certain operators in the model by an average value. At the
nanoscale level, the quantum fluctuations are sufficiently large enough that this
approximation is invalid. In fact, there had been a long harboured notion that
superconductivity would break down for systems where the mean single particle
energy level spacing, which is inversely proportional to the volume, is comparable
to the superconducting gap, as in the case of metallic nanograins. This was
conjectured by Anderson [3] in 1959 on the basis of the BCS theory but the
experiments by RBT show this to not be the case. Consequently, an exact solution
is highly desired, a view that has been promoted in [4].
The study of exact solutions of quantum mechanical models has its origins
in the work of Bethe in 1931 on the Heisenberg model [5]. The field received a
tremendous impetus in the 1960s with the work of McGuire [6], Yang [7], Baxter
[8] and Lieb and Wu [9] and it has prospered ever since. The work of RBT cited
earlier has brought the discipline to a new audience, when it was realized that the
exact solution of the BCS model had been obtained, although largely ignored, by
Richardson in 1963 [10]. The reason that Richardson’s work was overlooked for
so long is because the theory that had been proposed by BCS was so spectacularly
successful that there had never been a need to use an alternative approach. Once
the results of RBT were communicated however, it was clear that a new viewpoint
was needed. When the condensed matter physics community became aware of
Richardson’s work, his results were promptly adopted and it was shown that the
analysis of the exact solution gave agreement with the experiments [11]. A concise
yet informative account of the developments is given in [12].
In this review, we will recount the quantum inverse scattering method and
the associated algebraic Bethe ansatz method for the exact solution of integrable
quantum Hamiltonians. We then show how this procedure can be applied for
the analysis of three models which are the focus of many current theoretical
studies: a model for two Bose–Einstein condensates coupled via Josephson
tunnelling, a model for atomic–molecular Bose–Einstein condensation and the
BCS model. In each case, we undertake an asymptotic analysis of the solution
and demonstrate how this can be applied to extract the asymptotic behaviour of
certain correlation functions at zero temperature through use of the Hellmann–
Feynman theorem [13].

8.2 Quantum inverse scattering method

First we will review the basic features of the quantum inverse scattering method
[14, 15]. The theory of exactly solvable quantum systems in this setting relies on
the existence of a solution R(u) ∈ End(V ⊗ V ), where V denotes a vector space,
which satisfies the Yang–Baxter equation acting on the three-fold tensor product
space V ⊗ V ⊗ V :

R12 (u − v)R13 (u)R23 (v) = R23 (v)R13 (u)R12 (u − v). (8.1)

Copyright © 2003 IOP Publishing Ltd.

 Here Rj k (u) denotes the matrix in End(V ⊗ V ⊗ V ) acting non-trivially on the
j th and kth spaces and as the identity on the remaining space. The R-matrix
solution may be viewed as the structural constants for the Yang–Baxter algebra
which is generated by the monodromy matrix T (u) whose entries generate the
algebra
R12 (u − v)T1 (u)T2 (v) = T2 (v)T1 (u)R12 (u − v). (8.2)
We note that as a result of (8.1) the Yang–Baxter algebra is necessarily associative.
In component form, we may write
 pq j
 p q jl
Rik (u − v)Tp (u)Tql (v) = Tk (v)Ti (u)Rqp (u − v)
p,q p,q

so the Rijkl (u) give the structure constants of the algebra.

Here, we will only concern ourselves with the su(2) invariant R-matrix
which has the form
1
R(u) = (u · I ⊗ I + ηP )
u+η
 
1 0 0 0
0 b(u) c(u) 0
=
0 c(u) b(u) 0
 (8.3)
0 0 0 1

with b(u) = u/(u + η) and c(u) = η/(u + η). Here, P is the permutation
operator which satisfies

P (x ⊗ y) = y ⊗ x ∀ x, y ∈ V .

In this case, the Yang–Baxter algebra has four elements which we express as
 
A(u) B(u)
T (u) = . (8.4)
C(u) D(u)

Next, suppose that we have a representation, which we denote π, of the

Yang–Baxter algebra. For later convenience, we set

L(u) = π(T (u))

which we refer to as an L-operator. Defining the transfer matrix through

t (u) = π tr((T (u))) = π(A(u) + D(u)) (8.5)

it follows from (8.1) that the transfer matrices commute for different values of the
spectral parameters; viz.

[t (u), t (v)] = 0 ∀u, v. (8.6)

Copyright © 2003 IOP Publishing Ltd.

 There are two significant consequences of (8.6). The first is that t (u) may be
diagonalized independently of u, that is the eigenvectors of t (u) do not depend on
u. Secondly, taking a series expansion

t (u) = c k uk
k
it follows that
[ck , cj ] = 0 ∀ k, j.
Thus, for any Hamiltonian which is expressible as a function of the operators ck
only, then each ck corresponds to an operator representing a constant of the motion
since it will commute with the Hamiltonian. When the number of conserved
quantities is equal to the number of degrees of freedom of the system, the model
is said to be integrable.
An important property of the Yang–Baxter algebra is that it has a co-
multiplication structure which allows us to build tensor product representations.
In particular, given two L-operators LU , LW acting on V ⊗ U and V ⊗ W
respectively, then L = LU LW is also an L-operator as can be see from
R12 (u − v)L1 (u)L2 (v) = R12 (u − v)LU W U W
1 (u)L1 L2 (v)L2 (v)
= R12 (u − v)LU U W W
1 (u)L2 (v)L1 (u)L2 (v)
= LU
2 (v)L1 (u)R12 (u − v)L1 (u)L2 (v)
U W W

2 (v)L1 (u)L2 (v)L1 (u)R12 (u − v)

= LU U W W

= LU
2 (v)L2 (v)L1 (u)L1 (u)R12 (u − v)
W U W

= L2 (v)L1 (u)R12 (u − v).

Furthermore, if L(u) is an L-operator, then so is L(u + α) for any α since the
R-matrix depends only on the difference of the spectral parameters.

8.2.1 Realizations of the Yang–Baxter algebra

In order to construct a specific model, we must address the question of
determining a realization of the Yang–Baxter algebra. Here we will present
several examples which will all be utilized later. The first realization comes
from the R-matrix itself, since it is apparent from (8.1) that we can make the
identification L(u) = R(u) such that a representation of (8.2) is obtained. This
is the realization used in the construction of the Heisenberg model [14, 15]. A
second realization is given by L(u) = G (c-number realization), where G is an
arbitrary 2 × 2 matrix whose entries do not depend on u. This follows from the
fact that [R(u), G ⊗ G] = 0.
There is a realization in terms of canonical boson operators b, b† with the
relations [b, b† ] = 1 which reads [16, 17, 31]:
 
u + ηN̂ b
Lb (u) = (8.7)
b† η−1

Copyright © 2003 IOP Publishing Ltd.

 where N̂ = b † b. There also exists a realization in terms of the su(2) Lie algebra
with generators S z and S ± [14, 15]:
 
1 u − ηS z −ηS +
L (u) =
S
(8.8)
u −ηS − u + ηS z
with the commutation relations [S z , S ± ] = ±S ± , [S + , S − ] = 2S z . It is worth
noting that in the case when the su(2) algebra takes the spin- 12 representation,
the resulting L-operator is equivalent to that given by the R-matrix. Another is
realized in terms of the su(1, 1) generators K z and K ± [18, 19]:
 
u + ηK z ηK −
LK (u) = (8.9)
−ηK + u − ηK z
with the commutation relations [K z , K ± ] = ±K ± , [K + , K − ] = −2K z .
Here we will use these realizations to construct a variety of exactly solvable
models. First, however, we will introduce the algebraic Bethe ansatz which
provides the exact solution.

8.3 Algebraic Bethe ansatz method of solution

For a given realization of the Yang–Baxter algebra, the solution to the problem of
finding the eigenvalues of the transfer matrix (8.5) via the algebraic Bethe ansatz
is obtained by utilizing the commutation relations of the Yang–Baxter algebra.
We have from the defining relations (8.2) that (among other relations)
[A(u), A(v)] = [D(u), D(v)] = 0
[B(u), B(v)] = [C(u), C(v)] = 0
u−v+η η (8.10)
A(u)C(v) = C(v)A(u) − C(u)A(v)
u−v u−v
u−v−η η
D(u)C(v) = C(v)D(u) + C(u)D(v).
u−v u−v
A key step in successfully applying the algebraic Bethe ansatz approach is
finding a suitable pseudovacuum state, |0#, which has the properties
A(u)|0# = a(u)|0#
B(u)|0# = 0
C(u)|0# = 0
D(u)|0# = d(u)|0#
where a(u) and d(u) are scalar functions.
Assuming the existence of such a pseudovacuum state, choose the Bethe
state

M
|v# ≡ |v1 , . . . , vM # = C(vi )|0#. (8.11)
i=1

Copyright © 2003 IOP Publishing Ltd.

 Note that because [C(u), C(v)] = 0, the ordering is not important in (8.11). The
approach of the algebraic Bethe ansatz is to use the relations (8.10) to determine
the action of t (u) on |v#. The result is
t (u)|v# = (u, v)|v#
 
N
ηa(vi ) 
M
vi − vj + η
− |v1 , . . . , vi−1 , u, vi+1 , . . . , vM #
i
u − vi j =i vi − vj
 
M
ηd(vi ) 
M
vi − vj − η
+ |v1 , . . . , vi−1 , u, vi+1 , . . . , vM #
α u − vi j =i vi − vj
(8.12)
where
M
u − vi + η M
u − vi − η
(u, v) = a(u) + d(u) . (8.13)
i=1
u − vi i=1
u − vi
This shows that |v# becomes an eigenstate of the transfer matrix with eigenvalue
(8.13) whenever the Bethe ansatz equations

a(vi ) M
vi − vj − η
= i = 1, . . . , M (8.14)
d(vi ) j =i vi − vj + η

are satisfied. Note that in the derivation of the Bethe ansatz equations, it is required
that vi = vj , ∀ i, j. This is a result of the Pauli principle for Bethe ansatz solvable
models as developed in [20] for the Bose gas. We will not reproduce the proofs
for the present cases, as they follow essentially the same argument as that in [20].

8.3.1 Scalar products of states

One of the important applications of the previous discussion is that there exists a
formula due to Slavnov [14, 21, 22] for the scalar product of states obtained via
the algebraic Bethe ansatz for the R-matrix (8.3). The formula reads
S(v : u) = "0|B(v1 ) . . . B(vM )C(u1 ) . . . C(uM )|0#
det F (u : v)
=
det V (u : v)
where
∂ 1
Fij = (uj , v) Vij =
∂vi uj − vi
the parameters {vi } satisfy the Bethe ansatz equations (8.14) and {uj } are
arbitrary. The significance of this result is that it opens up the possibility of
determining form factors and correlation functions for any model which can be
derived in this manner. Although we will not go into any details here, we wish
to point out that explicit results for two of the models which we will discuss
subsequently can be found in [23, 24].

Copyright © 2003 IOP Publishing Ltd.

 8.4 A model for two coupled Bose–Einstein condensates
Experimental realization of Bose–Einstein condensates in dilute atomic alkali
gases has stimulated a diverse range of theoretical and experimental research
activity [25–29]. A particularly exciting possibility is that a pair of Bose–Einstein
condensates (such as a Bose–Einstein condensate trapped in a double-well
potential) may provide a model tunable system in which to observe macroscopic
quantum tunnelling. Here we will show that a model Hamiltonian for a pair
of coupled Bose–Einstein condensates admits an exact solution. The model is
also realizable in Josephson coupled superconducting metallic nanoparticles [30],
which has applications in the implementation of solid-state quantum computers.
The canonical Hamiltonian which describes tunnelling between two Bose–
Einstein condensates takes the form [27]
K µ EJ †
H= (N1 − N2 )2 − (N1 − N2 ) − (b b2 + b2† b1 ). (8.15)
8 2 2 1
where b1† , b2† denote the single-particle creation operators in the two wells and
N1 = b1† b1 , N2 = b2† b2 are the corresponding boson number operators. The total
boson number N1 + N2 is conserved and set to the fixed value of N. The physical
meaning of the coupling parameters for different realizable systems may be
found in [27]. It is useful to divide the parameter space into three regimes: Rabi
(K/EJ  N −1 ), Josephson (N −1  K/EJ  N) and Fock (N  K/EJ ). There
is a correspondence between (8.15) and the motion of a pendulum [27]. In the
Rabi and Josephson regimes, this motion is semiclassical, unlike the case of the
Fock regime. For both the Fock and Josephson regimes, the analogy corresponds
to a pendulum with fixed length, while in the Rabi regime the length varies. An
important problem is to study the behaviour in the crossover regimes, which is
accessible through the exact solution. The exact solvability of (8.15) which we
discuss here follows from the fact that it is mathematically equivalent to the
discrete self-trapping dimer model studied by Enol’skii et al [31], who solved
the model through the algebraic Bethe ansatz. We will describe this construction
below.
The co-multiplication behind the Yang–Baxter algebra allows us to choose
the following representation of the monodromy matrix:
L(u) = Lb1 (u + ω)Lb2 (u − ω)
 
(u + ω + ηN1 )(u − ω + ηN2 ) + b2†b1 (u + ω + ηN1 )b2 + η−1 b1
= .
(u − ω + ηN2 )b1† + η−1 b2† b1† b2 + η−2
(8.16)
Defining the transfer matrix as before through t (u) = tr(L(u)), we have explicitly
in the present case

t (u) = u2 + uηN̂ + η2 N1 N2 + ηω(N2 − N1 ) + b2†b1 + b1† b2 + η−2 − ω2 .

Copyright © 2003 IOP Publishing Ltd.

 Then 

dt 
t (0) = = ηN̂
du u=0
and it is easy to verify that the Hamiltonian is related to the transfer matrix t (u)
by
H = −κ(t (u) − 14 (t
(0))2 − ut
(0) − η−2 + ω2 − u2 )
where the following identification has been made for the coupling constants:

K κη2 µ EJ
= = −κηω = κ.
4 2 2 2
An explicit representation of (8.4) is obtained from (8.16) with the
identification

A(u) = (u + ω + ηN1 )(u − ω + ηN2 ) + b2† b1

B(u) = (u + ω + ηN1 )b2 + η−1 b1
C(u) = (u − ω + ηN2 )b1† + η−1 b2†
D(u) = b1† b2 + η−2 .

Choosing the Fock vacuum as the pseudovacuum, which satisfies B(u)|0# = 0 as

required by the Bethe ansatz procedure, the eigenvalues a(u) and d(u) of A(u)
and D(u) on |0# are

a(u) = (u + ω)(u − ω)
d(u) = η−2 .

The Bethe ansatz equations are then explicitly

N
vi − vj − η
η2 (vi2 − ω2 ) = (8.17)
v − vj + η
j =i i

with the eigenstates of the form (8.11) with C(u) given as before. From the Bethe
ansatz equations, we may derive the useful identity

m m N
vi − vj − η
η2 (vi2 − ω2 ) = (8.18)
i=1
v − vj + η
i=1 j =m+1 i

which will be used later.

It is clear that the Bethe states are eigenstates of N̂ with eigenvalue N. As
N is the total number of bosons, we expect N + 1 solutions of the Bethe ansatz
equations. As mentioned earlier, we must exclude any solution in which the roots
of the Bethe ansatz equations are not distinct. For example, the solution
*
vj = ± ω2 − (−1)N η−2 ∀j (8.19)

Copyright © 2003 IOP Publishing Ltd.

 of (8.17) is invalid, except when N = 1. (Note the error in [24].) For a given valid
solution of the Bethe ansatz equations, the energy of the Hamiltonian is obtained
from the transfer matrix eigenvalues (8.13) and reads as
 N 
 
−2 η2 N 2
η
E = −κ η 1+ − uηN − u2
−
i=1
vi − u
4
N  
η
− η−2 + ω2 + (u2 − ω2 ) 1− . (8.20)
i=1
vi − u

Note that this expression is independent of the spectral parameter u which can be
chosen arbitrarily. The formula simplifies considerably with the choice u = ω, by
employing (8.18), which yields a polynomial form:
 N 
η2 N 2
E = −κ η−2 η2 (vi − ω + η)(vi + ω) − − ηωN − η−2 .
i=1
4

However, for the purpose of an asymptotic analysis in the Rabi regime, it is more
convenient to choose u = 0, while for the Fock regime we use u = η2 .

8.4.1 Asymptotic analysis of the solution

Here we will recall the asymptotic analysis of the exact solution that was
conducted in [32]. We start the analysis with the Rabi regime where η2 N  1.
From the Bethe ansatz equations, it is clear that η2 vi2 → 1 as η → 0, so that
vi ≈ ±η−1 . However, when η = 0 we know that the Hamiltonian is diagonalizable
by using the Bogoliubov transformation, from which we can deduce that the
solution of the Bethe ansatz equations corresponding to the ground state must
have vi ≈ η−1 . Therefore, it is reasonable to consider the asymptotic expansion

vi ≈ η−1 + i + ηδi . (8.21)

Excitations correspond to changing the signs of the leading terms in the Bethe
ansatz roots. To study the asymptotic behaviour for the mth excited state, we set

vi ≈ −η−1 + i + ηδi i = 1, . . . , m
−1
(8.22)
vi ≈ η + i + ηδi i = m + 1, . . . , N

with the convention that the ground state corresponds to m = 0.

From the leading terms of the Bethe ansatz equations for vi , i ≤ m, we find


m
1
i = (8.23)
j =i
i − j

Copyright © 2003 IOP Publishing Ltd.

 which implies

m 
m
m(m − 1)
i = 0 i2 = .
i=1 i=1
2
In a similar fashion, we have, for m < i ≤ N,


N
1
i = − (8.24)
j =m+1
i − j
j =i

which implies


N 
N
(N − m)(N − m − 1)
i = 0 i2 = − .
i=m+1 i=m+1
2

It is clear from (8.23) and (8.24) why the Pauli exclusion principle applies in
the present case. In the asymptotic expansion for vi , i is assumed finite. However,
if vi = vj for some i, j , then i = j and (8.23) and (8.24) imply that i , j are
infinite which is a contradiction. Hence, vi must be distinct for different i. Note
also that for this approximation to be valid, we require η−1 * i . However, we
see that |i | is of the order of N 1/2 . Thus, our approximation will be valid for
ηN 1/2  1, which is precisely the criterion for the Rabi region and, consequently,
N cannot be arbitrarily large for fixed η or vice versa.
Now we go to the next order. From (8.18), we find

m
m(m − 1) m(m − N) mω2
δi = − + −
i=1
4 2 2

N
(N − m)(N − m − 1) m(m − N) (N − m)ω2
δi = − + +
i=m+1
4 2 2

which using (8.20) leads us to the result

Em η2 ω2 (N − 2m) η2 N η2
≈ −N + 2m − + + m(N − m).
κ 2 4 2
The energy level spacings m = Em − Em−1 are, thus,
 
η2
m ≈ κ 2 + η ω + (N − 2m + 1) .
2 2
2

One may check that m /N is of the order of N −1 . This indicates that the Rabi
regime is semiclassical [27]. This value for the gap between the ground and first
excited state agrees, to leading order in η2 N, with the Gross–Pitaevskii mean-field
theory [33] giving a Josephson plasma frequency of ωJ = 2κ(1 + η2 N/2)1/2 .

Copyright © 2003 IOP Publishing Ltd.

 Now we look at the asymptotic behaviour of the Bethe ansatz equations in the
Fock regime η2 * N. It is necessary to distinguish the following cases: (i) ω = 0
and (ii) ω = 0.
(i) ω = 0. In this case, it is appropriate to consider the permutation operator
P which interchanges the labels 1 and 2 in (8.15). For ω = 0, P commutes with
the Hamiltonian and any eigenvector of the Hamiltonian is also an eigenvector of
P with eigenvalue ±1. Therefore, the Hilbert space splits into the direct sum of
two subspaces corresponding to the symmetric and antisymmetric wavefunctions.
From now on, we restrict ourselves to the case when N is even, i.e. N = 2M,
although a similar calculation is also applicable to the case when N is odd. A
careful analysis leads us to conclude that the ground state lies in the symmetric
subspace. The asymptotic form of the roots of the Bethe ansatz equations for the
ground state takes the ‘string’-like structure
j
CM
vj ± ≈ −(M − j )η ± i η−(2j −1) + M(M + 1)η−3 δj 1 j = 1, . . . , M
(j − 1)!
j
where CM is a binomial coefficient. For this asymptotic ansatz to be valid,
we require that any term in the asymptotic expansion should be much smaller
than those preceding it. This yields η2 * N which coincides with the defining
condition for the Fock region. Throughout, the Pauli exclusion principle has been
taken into account to exclude any possible spurious solutions of the Bethe ansatz
equations.
This structure clearly indicates that in the ground state the N bosons fuse into
M ‘bound’ states and excitations correspond to a breakdown of these bound states.
Specifically, the first and second excited states correspond to the breakdown of
the bound state at −(M − 1)η, with the first excited state in the antisymmetric
subspace and the second excited state in the symmetric subspace. Explicitly, we
can write down the spectral parameter configurations for the first two excited
states:

v1+ ≈ −Mη + a1+ η−3 v1− ≈ −(M − 1)η + a1− η−3

vj ± ≈ −(M − j )η + aj ± η−(2j −1) j = 2, . . . , M

with
M +1 M(M + 1)
a1+ = − a1− =
2 √2
−(M − 1) ± (M − 1) 13M 2 + 10M + 1
2
a2± =
√12
(M − 1)(M − 2) 2M(M + 1)
a3± = ±
24
M −j +1
aj ± = √ aj −1,± j = 3, . . . , M
(j + 1)j (j − 1)(j − 2)

Copyright © 2003 IOP Publishing Ltd.

 for the (antisymmetric) first excited state and

(M + 1)(2M + 1) M(M + 1)
a1+ = − a1− = −
2 √ 2
−(M − 1)2 ± i(M − 1) 11M 2 + 14M − 1
a2± =
√12
(M − 1)(M − 2) 2M(M + 1)
a3± = ±i
24
M −j +1
aj ± = √ aj −1,± j = 3, . . . , M
(j + 1)j (j − 1)(j − 2)

for the (symmetric) second excited state. The breakdown of the bound state at
−(M − j )η, j = 2, . . . , M results in the higher excited states.
Substituting these results into (8.20) leads us to the asymptotic ground-state
energy
E0 ≈ −2κη−2M(M + 1)
while for the first and second excited states, we have

M2 + M − 2
E1 ≈ κη2 − κη−2
3
5M 2 + 5M + 2
E2 ≈ κη2 + κη−2 .
3
In contrast to the Rabi regime, the Fock regime is not semiclassical, as the ratio
of the gap and N is of finite order when N is large.
We can perform a similar analysis for odd N. In this case, the gap between
the ground and the first excited states is proportional to κη−2 instead of κη2 .
Furthermore, the ground-state root structure is different in the odd case since not
all the bosons can be bound in pairs. This indicates there is a strong parity effect
in the Fock regime, in contrast to the Rabi regime.
(ii) ω = 0. In this case the root structure is somewhat more complicated
than for ω = 0, so we will not present the details. We remark, however, that our
calculations show that up to order η−2 the ground-state energy eigenvalue takes
the same form as in the case ω = 0. Actually, the leading contribution arising from
the ω term appears only as ω2 η−4 . This means that the results presented here are
applicable for all values of ω (or, equivalently, µ).
Although it is difficult to define rigorously [27, 34], the relative phase
between Bose–Einstein condensates is useful in understanding interference
experiments [28, 29, 35]. Recall that in Josephson’s original proposal [36]
for Cooper pair tunnelling through an insulating barrier between macroscopic
superconductors, the current is a manifestation of the relative phase between
the wavefunctions of the superconductors. By definition, the relative phase ! is
conjugate to the relative number of atoms in the two condensates n ≡ N1 − N2 .

Copyright © 2003 IOP Publishing Ltd.

 Using the Hellmann–Feynman theorem, we find that
 
∂E0 ∂E0 2
" n # = 8
2
−4 .
∂K ∂ µ
For the ground state in the limit of strong tunnelling (i.e. the Rabi regime),
" n2 # ≈ N − ( µN/EJ )2 . In the case of weak tunnelling (i.e. the Fock regime),
" n2 # ≈ 2N(N + 2)(EJ /K)2 . The degree of coherence between the two Bose–
Einstein condensates can be discussed in terms of [27]
1 † 1 ∂E0
α≡ "a1 a2 + a2† a1 # = − .
2N N ∂EJ
In the strong coupling limit, α ≈ 1 − N −1 ( µ)2 /(8EJ )2 , indicating very close
to full coherence in the ground state. In the opposite limit, we have α ≈ 2(N +
2)EJ /K  1, indicating the absence of coherence. These results give the first-
order corrections to the results presented in [28, 37] for the number fluctuations
and the coherence factor at zero temperature.

8.5 A model for atomic–molecular Bose–Einstein

condensation
After the experimental realization of Bose–Einstein condensation in dilute
alkali gases, many physicists started to consider the possibility of producing a
molecular Bose–Einstein condensate from photoassociation and/or the Feshbach
resonance of an atomic Bose–Einstein condensate of a weakly interacting dilute
alkali gas [38, 39]. This novel area has attracted considerable attention from
both experimental and theoretical physicists and, in particular, it has recently
been reported that a Bose–Einstein condensate of rubidium has been achieved
comprised of a coherent superposition of atomic and molecular states [40, 41].
As stressed in [42], even in the ideal two-mode limit, mean field theory fails
to provide long-term predictions due to strong interparticle entanglement near
the dynamically unstable molecular mode. The numerical results have shown
that the large-amplitude atom–molecular coherent oscillations are damped by
the rapid growth of fluctuations near the unstable point, which contradicts the
mean field theory predictions. In order to clarify the controversies raised by these
investigations, one can appeal to the exact solution of the two-mode model, the
derivation of which we will now present.
The two-mode Hamiltonian takes the form
ω † 
H= a a + (a † a † b + b† aa) (8.25)
2 2
where a † and b† denote the creation operators for atomic and molecular modes
respectively. Note that the total atom number operator N̂ = Na + 2Nb where
Na = a † a, Nb = b † b provides a good quantum number since [H, N̂ ] = 0.

Copyright © 2003 IOP Publishing Ltd.

 In order to derive this Hamiltonian through the quantum inverse scattering
method, we take the following L-operator

L(u) = GLb (u − δ − η−1 )LK (u)

with the matrix G given by

 −1 
−η 0
G= .
0 η−1

This gives us the explicit realization of the Yang–Baxter algebra:

A(u) = −η−1 (u + ηK z )(u − δ − η−1 + ηNb ) + bK +

B(u) = −K − (u − δ − η−1 + Nb ) − η−1 b(u − ηK z )
C(u) = η−1 b† (u + ηK z ) − η−1 K +
D(u) = b † K − + η−2 (u − ηK z )

and
t (0) = δK z + b † K − + bK + − ηK z Nb . (8.26)
Let |0# denote the Fock vacuum state and let |k# denote a lowest weight
state of the su(1, 1) algebra with weight k, i.e. K z |k# = k|k#. On the product state
|"# = |0#|k#, it is clear that B(u)|"# = 0 and

a(u) = −η−1 (u + ηk)(u − δ − η−1 )

d(u) = η−2 (u − ηk).

We can immediately conclude that the eigenvalues of (8.26) are given by

M
vi − η M
vi + η
(0) = k(δ + η−1 ) − kη−1 (8.27)
i=1
vi i=1
vi

subject to the Bethe ansatz equations

(vi + ηk)(1 − ηvi + ηδ)  M

vi − vj − η
= . (8.28)
(vi − ηk) v − vj + η
j =i i

Realizing the su(1, 1) algebra in terms of canonical boson operators through

(a †)2 a2 2Na + 1
K+ = K− = Kz =
2 2 4
we then find that the Hamiltonian (8.25) is related to (8.26) through

H = lim (t (0) − δ/4)

η→0

Copyright © 2003 IOP Publishing Ltd.

 with ω = δ. Note that, in this case, the possible lowest weight states for the
su(1, 1) algebra are

|k = 1/4# ≡ |0# |k = 3/4# ≡ a † |0#.

Moreover, we have N = 2M + 2k − 1/2.

It is worth mentioning at this point that another realization of the su(1, 1)
algebra is given in terms of two sets of boson operators by
Na + Nc + 1
K + = a † c† K − = ac Kz =
2
with J = Na − Nc a central element commuting with the su(1, 1) algebra in this
representation. Due to the symmetry a † ↔ c† we may assume J ≥ 0. For this case
we define the Hamiltonian

H = lim (t (0) − δ/2) + βJ

η→0

= αNa + γ Nc + (a † c† b + b †ac) (8.29)

with α = δ/2 + β and γ = δ/2 − β. This model has a natural interpretation

for atomic–molecular Bose–Einstein condensation for two distinct atomic species
which can bond to form a di-atomic molecule. In this case, the possible lowest
weight states for the su(1, 1) algebra are

|k = (m + 1)/2# ≡ (a † )m |0#

and J = 2k − 1. A detailed analysis of this model through the exact solution will
be given at a later date.
For the exact solution of the Hamiltonian (8.25) it is necessary to take the
quasi-classical limit η → 0 in the Bethe ansatz equations (8.28). The resulting
Bethe ansatz equations take the form

2k M
1
δ − vi + =2 . (8.30)
vi v − vi
j =i j

Also, in this limit the corresponding energy eigenvalue is


M
E = ω(M + k − 1/4) −  vi
i=1
M
1
= ω(k − 1/4) − 2k . (8.31)
v
i=1 i

The equivalence of the two energy expressions can be deduced from (8.30). The
eigenstates too are obtained by this procedure. Consider the following class of

Copyright © 2003 IOP Publishing Ltd.

 states:

M
|v1 , . . . , vM # = c(vi )|"# (8.32)
i=1

where c(v) = (vb † − a † a † /2), |"# = |0# for k = 1/4 and |"# = a †|0# for k =
3/4. In the case when the set of parameters {vi } satisfy the Bethe ansatz equations
(8.30), then (8.32) are precisely the eigenstates of the Hamiltonian.

8.5.1 Asymptotic analysis of the solution

In the limit of large |δ|, we can perform an asymptotic analysis of the Bethe ansatz
equations to determine the asymptotic form of the energy spectrum. We choose
the following ansatz for the Bethe roots:

δ −1 µi i≤m
vi ≈ −1
δ + i + δ µi i > m.

For i > m, we obtain, from the zero-order terms in the Bethe ansatz equations,


M
1
i = 2
j =m+1
i − j
j =i

which implies

M
i = 0.
i=m+1

From the terms in δ −1 , we find

M
µj − µi
µi = 2(k + m) + 2
j=m+1
(j − i )2
j =i

and, thus,

M
µi = 2(k + m)(M − m).
i=m+1
Next we look at the Bethe ansatz equations for i ≤ m. The terms in δ give

2k m
1
1+ =2
µi µ
j =i j
− µi

which implies

m
µi = −2km − m(m − 1).
i=1

Copyright © 2003 IOP Publishing Ltd.

 This gives the energy levels


M
2 
M
Em ≈ ω(M + (k − 1/4)) − ω(M − m) −  i − µi
i=m+1
ω i=1
2
= ω(m + k − 1/4) + (3m2 − m + 4km − 2kM − 2mM).
ω
The level spacings are

m = Em − Em−1
22
≈ω− (M + 2 − 3m − 2k)
ω
from which we conclude that, in this limit, the model is semi-classical.
Let E denote the ground-state energy (E = E0 for δ * 0, E = EM for
δ  0) and the gap to the first excited state. Employing the Hellmann–
Feynman theorem, we can determine the asymptotic form of the following zero-
temperature correlations
∂E ∂E
"Na # = 2 θ = −2
∂ω ∂
where θ = −"a † a † b + b†aa# is the coherence correlator. For large N, we
introduce the rescaled variables
δ "Na # θ
δ∗ = ∗ = "Na #∗ = θ∗ = . (8.33)
N 1/2 N 1/2 N N 3/2
We then have, for δ ∗ * 0,
1
∗ ≈ δ ∗ − "Na #∗ ≈ 0 θ∗ ≈ 0
δ∗
while, for δ ∗  0,
2 1 1
∗ ≈ −δ ∗ − "Na #∗ ≈ 1 − θ∗ ≈ − .
δ∗ 2(δ ∗ )2 δ∗
This shows that the model has scale invariance in the asymptotic limit. The
scaling properties actually hold for a wide range of values of the scaled detuning
parameter δ ∗ , which is established through numerical analysis [43].

8.5.2 Computing the energy spectrum

For this model, there is a convenient method to determine the energy spectrum
without solving the Bethe ansatz equations (cf [19]). This is achieved by
introducing the polynomial function whose zeros are the roots of the Bethe ansatz

Copyright © 2003 IOP Publishing Ltd.

 equations, i.e.

M
G(u) = (1 − u/vi ).
i=1
It can be shown from the Bethe ansatz equations that G satisfies the differential
equation

uG

− (u2 − δu − 2k)G
+ (Mu − E/ + δ(k − 1/4))G = 0 (8.34)

subject to the initial conditions

E − ω(k − 1/4)
G(0) = 1 G
(0) = .
2k
In order to show this, we set

F (u) = uG

− (u2 − δu − 2k)G
.

As a result of the Bethe ansatz equations (8.30), it is deduced that F (vi ) = 0.

Given that F (u) is a polynomial of degree (M + 1), we then conclude that
F (u) = (αu + β)G(u) for some constants α, β, which are determined by the
asymptotic limits u → 0and u → ∞. Equation (8.34) then follows.
By setting G(u) = n gn un , the recurrence relation
E − ω(n + k − 1/4) n−M −1
gn+1 = gn + gn−1 (8.35)
(n + 1)(n + 2k) (n + 1)(n + 2k)
is readily obtained. It is clear from this relation that gn is a polynomial in E of
degree n. We also know that G is a polynomial function of degree M and so
we must have gM+1 = 0. The (M + 1) roots of gM+1 are precisely the energy
levels Em . Moreover, the eigenstates (8.32) are expressible as (up to overall
normalization)
M  † † n
† (M−n) a a
|v1 , . . . , vM # = gn (b ) |"#.
n=1
2

The recurrence relation (8.35) can be solved as follows (cf [19]). Setting

n
gn+1 = g0 xj yj
j =0

with
E − ω(j + k − 1/4)
xj =
(j + 1)(j + 2k)
and substituting into the recurrence relation (8.35), we have
j −M −1
xj xj −1 yj −1 (yj − 1) = .
(j + 1)(j + 2k)

Copyright © 2003 IOP Publishing Ltd.

 This yields yj = 1 + cj −1 /yj −1 with

2 (j + 1)(j + 2k)(j − M)
cj =
(E − ω(j + k + 3/4))(E − ω(j + k − 1/4))
which means yj can be expressed as a continued fraction. The requirement that
G is a polynomial function of order M decrees yM = 0, in turn implying

2 M(M + 2k − 1)
yM−1 =
(E − ω(M + k − 1/4))(E − ω(n + k − 5/4))
which is an algebraic equation that determines the allowed energy levels Em . This
procedure can easily be employed to determine the energy spectrum numerically,
without resorting to solving the Bethe ansatz equations. Explicit results can be
found in [43].

8.6 The BCS Hamiltonian

The experimental work of Ralph, Black and Tinkham [1] on the discrete energy
spectrum in small metallic aluminium grains generated interest in understanding
the nature of superconducting correlations at the nanoscale level. Their results
indicate significant parity effects due to the number of electrons in the system. For
grains with an odd number of electrons, the gap in the energy spectrum reduces
with the size of the system, in contrast to the case of a grain with an even number
of electrons, where a gap larger than the single-electron energy levels persists.
In the latter case, the gap can be closed by a strong applied magnetic field. The
conclusion drawn from these results is that pairing interactions are prominent
in these nanoscale systems. For a grain with an odd number of electrons, there
will always be at least one unpaired electron, so it is not necessary to break a
Cooper pair in order to create an excited state. For a grain with an even number
of electrons, all excited states have at least one broken Cooper pair, resulting in
a gap in the spectrum. In the presence of a strongly applied magnetic field, it is
energetically more favourable for a grain with an even number of electrons to
have broken pairs and, hence, in this case there are excitations which show no gap
in the spectrum.
The physical properties of a small metallic grain are described by the reduced
BCS Hamiltonian [11]
L
 L
 † †
H= j nj − g ck+ ck− cj − cj + . (8.36)
j =1 j,k

Here, j = 1, . . . , L labels a shell of doubly degenerate single-particle energy

levels with energies j and nj is the fermion number operator for level j . The
operators cj ± , cj†± are the annihilation and creation operators for the fermions at
level j . The labels ± refer to time-reversed states.

Copyright © 2003 IOP Publishing Ltd.

 One of the features of the Hamiltonian (8.36) is the blocking effect. For any
unpaired electron at level j , the action of the pairing interaction is zero since only
paired electrons are scattered. This means that the Hilbert space can be decoupled
into a product of paired and unpaired electron states in which the action of the
Hamiltonian on the subspace for the unpaired electrons is automatically diagonal
in the natural basis. In view of the blocking effect, it is convenient to introduce
hard-core boson operators bj = cj − cj + , bj† = cj†+ cj†− which satisfy the relations

(bj† )2 = 0 [bj , bk† ] = δj k (1 − 2bj† bj ) [bj , bk ] = [bj† , bk† ] = 0

on the subspace excluding single-particle states. In this setting, the hard-core
boson operators realize the su(2) algebra in the pseudo-spin representation, which
will be utilized later.
The original approach of BCS [2] to describe the phenomenon of
superconductivity was to employ a mean field theory using a variational
wavefunction for the ground state which has an undetermined number of
electrons. The expectation value for the number operator is then fixed by means
of a chemical potential term µ. One of the predictions of the BCS theory is
that the number of Cooper pairs in the ground state of the system is given by
the ratio /d where is the BCS ‘bulk gap’ and d is the mean level spacing
for the single-electron eigenstates. For nanoscale systems, this ratio is of the
order of unity, in seeming contradiction with the experimental results discussed
earlier. The explanation for this is that the mean field approach is inappropriate
for nanoscale systems due to large superconducting fluctuations.
As an alternative to the BCS mean field approach, one can appeal to the exact
solution of the Hamiltonian (8.36) derived by Richardson [10] and developed by
Richardson and Sherman [44]. It has also been shown by Cambiaggio et al [45]
that (8.36) is integrable in the sense that there exists a set of mutually commutative
operators which commute with the Hamiltonian. These features have recently
been shown to be a consequence of the fact that the model can be derived in
the context of the quantum inverse scattering method using the L-operator (8.8)
with a c-number L-operator [23, 46], which we will now explicate.

8.6.1 A universally integrable system

In this case, we use a c-number realization G of the L-operator as well as (8.8) to
construct the transfer matrix
t (u) = tr0 (G0 L0L (u − L ) · · · L01 (u − 1 )) (8.37)
which is an element of the L-fold tensor algebra of su(2). Here tr0 denotes
the trace taken over the auxiliary space labelled 0 and G = exp(−αησ ) with
σ = diag(1, −1). Defining
u − j
Tj = lim t (u)
u→j η2

Copyright © 2003 IOP Publishing Ltd.

 for j = 1, 2, . . . , L, we may write, in the quasi-classical limit, Tj = τj + o(η)
and it follows from the commutivity of the transfer matrices that [τj , τk ] =
0, ∀ j, k. Explicitly, these operators read as

L
 θj k
τj = 2αSjz + (8.38)
k=j
j − k

with θ = S + ⊗ S − + S − ⊗ S + + 2S z ⊗ S z .
We define a Hamiltonian through

L L L L
1 1  1  1 
H =− j τj + 3 τj τk + 2 τj − Cj (8.39)
α j =1 4α j,k=1 2α j =1 2α j =1
L
 L
1 
=− 2j Sjz − S−S+ (8.40)
j =1
α j,k=1 j k

where
C = S + S − + S − S + + 2(S z )2

is the Casimir invariant for the su(2) algebra. The Hamiltonian is universally
integrable since it is clear that [H, τj ] = 0, ∀j irrespective of the realizations of
the su(2) algebra in the tensor algebra.
In order to reproduce the Hamiltonian (8.36), we realize the su(2) generators
through the hard-core boson (spin- 12 ) representation, i.e.

Sj+ = bj Sj− = bj† Sjz = 12 (I − nj ). (8.41)


In this instance, one obtains (8.36) (with the constant term − L j j ) where
g = 1/α as shown by Zhou et al [23] and von Delft and Poghossian [46].
For each index k in the tensor algebra in which the transfer matrix acts, and
accordingly in (8.40), suppose that we represent the su(2) algebra through the
irreducible representation with spin sk . Thus {Sk+ , Sk− , Skz } act on a (2sk + 1)-
dimensional space. In employing the method of the algebraic Bethe ansatz
discussed earlier, we find that

L
u − k − ηsk
a(u) = exp(−αη)
k=1
u − k
L
u − k + ηsk
d(u) = exp(αη)
k=1
u − k

Copyright © 2003 IOP Publishing Ltd.

 which gives the eigenvalues of the transfer matrix (8.37) as
L
u − k + ηsk M
u − vj − η
(u) = exp(αη)
k=1
u − k j =1
u − vj
L
u − k − ηsk M
u − vj + η
+ exp(−αη) .
k=1
u − k j =1
u − vj

The corresponding Bethe ansatz equations read as

L M
vl − k + ηsk vl − vj + η
exp(2αη) =− .
v
k=1 l
− k − ηsk v − vj − η
j =1 l

The eigenvalues of the conserved operators (8.38) are obtained through

the appropriate terms in the expansion of the transfer matrix eigenvalues in the
parameter η. This yields the following result for the eigenvalues λj of τj :
 
 L  M
2sk 2
λj = 2α + − sj (8.42)
 − k
k=j j
 − vi
i=1 j

such that the parameters vj now satisfy the Bethe ansatz equations
L
 
M
2sk 2
2α + = . (8.43)
k=1
vj − k i=j vj − vi

Through (8.42) we can now determine the energy eigenvalues of (8.40). It is useful
to note the following identities:

M M  L
vj sk
2α vj + 2 = M(M − 1)
j =1
v
j =1 k=1 j
− k
 L
M 
sk
αM + =0
j =1 k=1
vj − k
M  L 
M  L L
vj sk sk k
− =M sk .
v − k j =1 k=1 vj − k
j =1 k=1 j k=1

Employing these, it is deduced that

L
 L

λj = 2α sj − 2αM
j =1 j =1
L
 L
 L 
 L L
 
M
j λj = 2α j sj + sj sk − 2M sk − 2α vj + M(M − 1)
j =1 j =1 j =1 k=j k=1 j =1

Copyright © 2003 IOP Publishing Ltd.

 which, combined with the eigenvalues 2sj (sj + 1) for the Casimir invariants Cj ,
yields the energy eigenvalues


M
E=2 vj . (8.44)
j =1

From this expression, we see that the quasi-particle excitation energies are given
by twice the Bethe ansatz roots {vj } of (8.43). In order to specialize this result to
the case of the BCS Hamiltonian (8.36), it is a matter of setting sk = 1/2, ∀ k.
Finally, let us remark that in the quasi-classical limit the eigenstates assume the
form
M  L bj†
|"# = |0#.
v − j
i=1 j =1 i

The construction given here can also be applied on a more general level.
Taking higher spin representations of the su(2) algebra produces models of BCS
systems which are coupled by Josephson tunnelling, as described in [47, 48].
One can also employ higher-rank Lie algebras, such as so(5) [49] and su(4)
[50,51] which produce coupled BCS systems which model pairing interactions in
nuclear systems. For the general case of an arbitrary Lie algebra, we refer to [52].
Finally, let us mention that if one reproduces this construction with the su(1, 1)
L-operator (8.9) in place of the su(2) L-operator (8.8), the pairing model for
bosonic systems introduced by Dukelsky and Schuck [53] is obtained.

8.6.2 Asymptotic analysis of the solution

In the limit g → 0, we can easily determine the ground-state energy of (8.36): it is
given by filling the Fermi sea. Here, we will assume that the number of fermions
is even. Thus, for small g > 0, it is appropriate to consider the asymptotic solution

vi ≈ i + gδi + g 2 µi i = 1, . . . , M.

Substituting this into (8.43) and equating the different orders in g yields
 
L M
g g2 1 1
vi ≈ i − + −
2 4 k=m+1 j − k i=j j − i

which immediately gives us the asymptotic ground-state energy

 L
M
g2 
M  1
E0 ≈ 2 j − gM + .
j =1
2 j =1 k=M+1 j − k

Next we look at the first excited state. In the g = 0 case, this corresponds to
breaking the Cooper pair at level M and putting single unpaired electrons in the

Copyright © 2003 IOP Publishing Ltd.

 levels M and M+1 . Now these two levels become blocked. Solving equations
(8.43) for this excited state is the same as for the ground state except that there are
now (M − 1) Cooper pairs and we have to exclude the blocked levels. We can,
therefore, write down the energy


M−1   L
g 2 M−1 1
E1 ≈ M + M+1 + 2 j − g(M − 1) + .
j =1
2 j =1 k=M+2 j − k

The gap is found to be

 

M−1 L

g2 1 1
≈ M+1 − M +g+ + .
2 j =1
M+1 − j k=M+1
k − M

As in previous examples, we can calculate some asymptotic correlation

functions for zero temperature by using the Hellmann–Feynman theorem. In
particular,
∂E0
"ni # =
∂i
which, for i ≤ M, gives
L

g2 1
"ni # ≈ 2 −
2 k=M+1
(i − k )2

while, for i > M, we get

g2 
M
1
"ni # ≈ .
2 j =1 (j − i )2

We can also determine the asymptotic form of the Penrose–Onsager–Yang off-

diagonal long-range order parameter [54, 55] to be
L
1  1 ∂E0
"b† bj # = −
L i,j =1 i L ∂g
L
M g M  1
≈ − .
L L j =1 k=M+1 j − k

Acknowledgments
We are deeply indebted to Ross McKenzie, Mark Gould, Xi-Wen Guan and
Katrina Hibberd for their collaborations on these topics. Financial support
from the Conselho Nacional de Desenvolvimento Cientı́fico e Tecnológico and
Australian Research Council is gratefully accepted.

Copyright © 2003 IOP Publishing Ltd.

 References
[1] Black C T, Ralph D C and Tinkham M 1996 Phys. Rev. Lett. 76 688
Ralph D C, Black C T and Tinkham M 1997 Phys. Rev. Lett. 78 4087
[2] Cooper L N, Bardeen J and Schrieffer J R 1957 Phys. Rev. 108 1175
[3] Anderson P W 1959 J. Chem. Solids 11 28
[4] Héritier M 2001 Nature 414 31
[5] Bethe H 1931 Z. Phys. 71 205
[6] McGuire J B 1964 J. Math. Phys. 5 622
[7] Yang C N 1967 Phys. Rev. Lett. 19 1312
[8] Baxter R J 1982 Exactly Solved Models in Statistical Mechanics (New York:
Academic Press)
[9] Lieb E H and Wu F Y 1968 Phys. Rev. Lett. 20 1445
[10] Richardson R W 1963 Phys. Lett. 3 277
Richardson R W 1963 Phys. Lett. 5 82
[11] von Delft J and Ralph D C 2001 Phys. Rep. 345 61
[12] Sierra G 2002 Statistical Field Theories, NATO Sc. Ser. II: Math. Phys. Chem. vol. 73
ed A Cappelli and G Mussardo (Dordrecht: Kluwer Academic Publishers)
[13] Hellmann H 1937 Einfuhrung in Die Quantumchemie (Leipzig: Franz Deutsche)
Feynman R P 1939 Phys. Rev. 56 340
[14] Korepin V E, Bogoliubov N M and Izergin G 1993 Quantum Inverse Scattering
Method and Correlation Functions (Cambridge: Cambridge University Press)
[15] Faddeev L D 1995 Int. J. Mod. Phys. A 10 1845
[16] Enol’skii V Z, Kuznetsov V and Salerno M 1993 Physica D 68 138
[17] Kundu A and Ragnisco O 1994 J. Phys. A: Math. Gen. 27 6335
[18] Jurco B 1989 J. Math. Phys. 30 1739
[19] Rybin A, Kastelewicz G, Timonen J and Bogoliubov N 1998 J. Phys. A: Math. Gen.
31 4705
[20] Izergin A G and Korepin V E 1982 Lett. Math. Phys. 6 283
[21] Slavnov N A 1989 Theor. Math. Phys. 79 502
[22] Kitanine N, Maillet J M and Terras V 1999 Nucl. Phys. B 554 647
[23] Zhou H-Q, Links J, McKenzie R H and Gould M D 2002 Phys. Rev. B 65 060502(R)
[24] Links J and Zhou H-Q 2002 Lett. Math. Phys. 60 275
[25] Parkin A S and Walls D F 1998 Phys. Rep. 303 1
[26] Dalfovo F, Giorgini S, Pitaevskii L P and Stringari S 1999 Rev. Mod. Phys. 71 463
[27] Leggett A J 2001 Rev. Mod. Phys. 73 307
[28] Orzel C, Tuchman A K, Fenselau M L, Yasuda M and Kasevich M A 2001 Science
291 2386
[29] Cataliotti F S, Burger S, Fort C, Maddaloni P, Minardi F, Trombettoni A, Smerzi A
and Inguscio M 2001 Science 293 843
[30] Makhlin Y, Schön G and Shnirman A 2001 Rev. Mod. Phys. 73 357
[31] Enol’skii V Z, Salerno M, Kostov N A and Scott A C 1991 Phys. Scr. 43 229
[32] Zhou H-Q, Links J, McKenzie R H and Guan X-W 2003 J. Phys. A: Math. Gen. 36
L113
[33] Smerzi A, Fantoni S, Giovanazzi S and Shenoy S R 1997 Phys. Rev. Lett. 79 4950
Paraoanu G-S, Kohler S, Sols F and Leggett A J 2001 J. Phys. B: At. Mol. Opt. 34
4689
[34] Yu S-X 1997 Phys. Rev. Lett. 79 780

Copyright © 2003 IOP Publishing Ltd.

 [35] Hall D S, Mathews M R, Wieman C E and Cornell E A 1998 Phys. Rev. Lett. 81 1539
[36] Josephson B D 1962 Phys. Lett. 1 251
Josephson B D 1974 Rev. Mod. Phys. 46 251
[37] Pitaevskii L and Stringari S 2001 Phys. Rev. Lett. 87 180402
[38] Wynar R, Freeland R S, Han D J, Ryu C and Heinzen D J 2000 Science 287 1016
[39] Inouye S, Andrews M R, Stenger J, Miesner H J, Stamper-Kurn D M and Ketterle W
1998 Nature 392 151
[40] Zoller P 2002 Nature 417 493
[41] Donley E A, Clausen N R, Thompson S T and Wieman C E 2002 Nature 417 529
[42] Vardi A, Yurovsky V A and Anglin J R 2001 Phys. Rev. A 64 063611
[43] Zhou H-Q, Links J and McKenzie R H 2002 Preprint cond-mat/0207540
[44] Richardson R W and Sherman N 1964 Nucl. Phys. 52 221
Richardson R W and Sherman N 1964 Nucl. Phys. 52 253
[45] Cambiaggio M C, Rivas A M F and Saraceno M 1997 Nucl. Phys. A 624 157
[46] von Delft J and Poghossian R 2002 Phys. Rev. B 66 134502
[47] Links J, Zhou H-Q, McKenzie R H and Gould M D 2002 Int. J. Mod. Phys. B 16
3429
[48] Links J and Hibberd K E 2002 Int. J. Mod. Phys. B 16 2009
[49] Links J, Zhou H-Q, Gould M D and McKenzie R H 2002 J. Phys. A: Math. Gen. 35
6459
[50] Guan X-W, Foerster A, Links J and Zhou H-Q 2002 Nucl. Phys. B 642 501
[51] Guan X-W, Foerster A, Links J and Zhou H-Q 2002 Exact results for BCS systems
Proc. Workshop on Integrable Theories, Solitons and Duality ed L Ferreira,
J Gomes and A Zimerman PRHEP-unesp 2002/016
[52] Asorey M, Falceto F and Sierra G 2002 Nucl. Phys. B 622 593
[53] Dukelsky J and Schuck P 2001 Phys. Rev. Lett. 86 4207
[54] Penrose O and Onsager L 1956 Phys. Rev. 104 576
[55] Yang C N 1962 Rev. Mod. Phys. 34 694

Copyright © 2003 IOP Publishing Ltd.

 Chapter 9

The thermodynamics of the spin- 12 XXX

chain: free energy and low-temperature
singularities of correlation lengths
Andreas Klümper and Christian Scheeren
Theoretische Physik I, Universität Dortmund, Germany

9.1 Introduction
Much effort has been devoted to the study of integrable quantum spin chains
such as the Heisenberg model [1, 2], t–J [3–5] and Hubbard models [6, 7]
and many more. The appealing feature of these systems is the availability of
exact data for the spectrum and other physical properties despite the truly
interacting nature of the spins (resp. particles). The computational basis for the
work on integrable quantum chains is the Bethe ansatz yielding a set of coupled
nonlinear equations for one-particle wavenumbers (Bethe ansatz roots). Many
studies of the Bethe ansatz equations have been directed at the ground state of
the considered system and have revealed interesting non-Fermi liquid properties
such as algebraically decaying correlation functions with non-integer exponents
and low-lying excitations of different types, i.e. spin and charge with different
velocities constituting so-called spin and charge separation, see [8–10].
A very curious situation arises in the context of the calculation of the
partition function from the spectrum of the Hamiltonian. Despite the validity of
the Bethe ansatz equations for all energy eigenvalues of the previously mentioned
models, the direct evaluation of the partition function is rather difficult. In contrast
to ideal quantum gases, the eigenstates are not explicitly known, the Bethe ansatz
equations just provide implicit descriptions that pose problems of their own kind.
Yet, knowing the behaviour of quantum chains at finite temperature is important
for many reasons. As a matter of fact, the strict ground state is inaccessible
due to the very fundamentals of thermodynamics. Therefore, the study of finite

Copyright © 2003 IOP Publishing Ltd.

 temperatures is relevant for theoretical as well as experimental reasons, see also
section 9.4. At high temperature, quantum systems show universal but trivial
properties without correlations. Lowering the temperature, the systems enter a
large regime with non-universal correlations and finally approach the quantum
critical point at exactly zero temperature showing again universal yet now non-
trivial properties with divergent correlation lengths governed by conformal field
theory [11].
In this chapter we want to review the various techniques developed for the
study of the thermodynamics of integrable systems at the example of the simplest
one, namely the spin- 12 Heisenberg chain. In fact, most of the techniques were
first developed for this model and only later were they generalized to other
systems. Very early Bethe [12] constructed the eigenstates for the isotropic spin- 12
Heisenberg chain. Only much later was the fully anisotropic spin- 12 XY Z chain
discovered to be integrable [13, 14], see also [15] and references therein for the
partially isotropic XXZ chain.
The thermodynamics of the Heisenberg chain was studied in [1, 2, 16] by
an elaborate version of the method used in [17]. Here, the partition function
was evaluated in the thermodynamic limit in which many of the states do
not contribute. The macro-state for a given temperature T is described by
a set of density functions formulated for the Bethe ansatz roots satisfying
integral equations obtained from the Bethe ansatz equations. In terms of the
density functions, expressions for the energy and the entropy were derived. The
minimization of the free energy functional yields what is nowadays known as the
thermodynamical Bethe ansatz (TBA).
There are two ‘loose ends’ in the briefly sketched procedure. First, the
description of the spectrum of the Heisenberg model was built on the so-called
‘string hypothesis’ according to which admissible Bethe ansatz patterns of roots
are built from regular building blocks. This hypothesis was criticized a number
of times and led to activities providing alternative access to the finite-temperature
properties [18–26]. The central idea of these works was a lattice path-integral
formulation of the partition function of the Hamiltonian and the definition of a
suitable ‘quantum transfer matrix’ (QTM), see also sections 9.2, 9.3 and 9.4.
The two apparently different approaches, the combinatorial TBA and the
operator-based QTM, are not at all independent! In the latter approach, there
are several quite different ways of analysing the eigenvalues of the QTM. In the
standard (and most economical) way, see later, a set of just two coupled nonlinear
integral equations (NLIE) is derived [25, 26]. Alternatively, an approach based on
the ‘fusion hierarchy’ leads to a set of (generically) infinitely many NLIEs [25,27]
that are identical to the TBA equations though completely different reasoning has
been applied!
Very recently [28], yet another formulation of the thermodynamics of the
Heisenberg chain has been developed. At the heart of this formulation is just
one NLIE with a structure very different from that of the two sets of NLIEs
discussed so far. Nevertheless, this new equation has been derived from the ‘old’

Copyright © 2003 IOP Publishing Ltd.

 NLIEs [28, 29] and is certainly an equivalent formulation. In the first applications
of the new NLIE, numerical calculations of the free energy have been performed
with excellent agreement with the older TBA and QTM results. Also, analytical
high-temperature expansions up to order 100 have been carried out on this basis.
The second ‘loose end’ within TBA concerns the definition of the entropy
functional. In [1, 2, 16, 17], the entropy is obtained within a combinatorial
evaluation of the number of micro-states compatible with a given set of density
functions of roots. As such, it is a lower bound to the total number of micro-
states falling into a certain energy interval. However, this procedure may be
viewed as a kind of saddle-point evaluation in the highly dimensional subspace
of all configurations falling into the given energy interval. Hence, the result is
correct in the thermodynamic limit and the ‘second loose end’ can actually be
tied up. Interestingly, the ‘second loose end’ of the TBA approach was motivation
for a ‘direct’ evaluation [30] of the partition function of integrable quantum
chains. This combinatorial approach is based on the string hypothesis but avoids
the definition of an entropy expression. A straightforward (though involved)
calculation leads to the single NLIE of [28].
The purpose of this chapter is to review the approach to thermodynamics of
integrable quantum chains that we believe is the most efficient one, namely the
QTM approach described in sections 9.2, 9.3 and 9.4. The NLIE of the isotropic
Heisenberg chain may be obtained as a special case from the thermodynamics of
the XY Z chain published in [25, 26]. However, the derivation we are going to
present is new and much more elegant and elementary than that of [25, 26]. The
main strength of the QTM-based NLIEs is the usefulness in the entire temperature
range from high to extremely low temperatures. As a demonstration of this, we
present in section 9.5 an analysis of the correlation lengths of the Heisenberg
chain at low temperatures (and arbitrary fields). This analysis leads to very explicit
results for the correlation lengths diverging like 1/T at low T . Mathematically,
we find the structure of the dressed energy and dressed charge formalism known
from finite-size analysis of the Hamiltonian at exactly T = 0.

9.2 Lattice path integral and quantum transfer matrix

In the following we consider the isotropic Heisenberg chain with Hamiltonian HL


L
HL = S'j S'j +1 (9.1)
j =1

with periodic boundary conditions on a chain of length L. The local interaction

of the spin- 12 objects is of exchange type and has antiferromagnetic character.

Copyright © 2003 IOP Publishing Ltd.

 9.2.1 Mapping to a classical model
In order to deal with the thermodynamics in the canonical ensemble, we have to
construct exponentials of the Hamiltonian. These operators are obtained from the
row-to-row transfer matrix T (λ) of the six-vertex model in the Hamiltonian limit
(small spectral parameter λ) [15]

T (λ) = eiP −λHL +O(λ

2)
(9.2)

with P denoting the momentum operator. We further introduce an auxiliary

transfer matrix T (λ) [7, 25] adjoint to T (λ) with Hamiltonian limit

T (λ) = e−iP −λHL +O(λ ) .

2
(9.3)

With these settings the partition function ZL of the quantum chain at finite
temperature T reads as

ZL = Tr e−β HL = lim ZL,N (9.4)

N→∞

where β = 1/T and ZL,N is defined by

ZL,N := Tr[T (τ )T (τ )]N/2 . (9.5)

The rhs of this equation can be interpreted as the partition function of a staggered
six-vertex model with alternating rows corresponding to the transfer matrices
T (τ ) and T (τ ), see figure 9.1. We are free to evaluate the partition function of this
classical model by adopting a different choice of transfer direction. A particularly
useful choice is based on the transfer direction along the chain and corresponding
transfer matrix T QTM defined for the columns of the lattice yielding

ZL,N = Tr(T QTM )L . (9.6)

In the remainder of this chapter, we will refer to T QTM as the ‘quantum transfer
matrix’ of the quantum spin chain because T QTM is the closest analogue to the
transfer matrix of a classical spin chain. Due to this analogy, the free energy f
per lattice site is given just by the largest eigenvalue max of the QTM

f = −kB T lim log max . (9.7)

N→∞

Note that the eigenvalue depends on the argument τ = β/N which vanishes in the
limit N → ∞ requiring a sophisticated treatment.
The main difference to classical spin chains is the infinite dimensionality of
the space in which T QTM is living (for N → ∞). In formulating (9.7) we have
implicitly employed the interchangeability of the two limits (L, N → ∞) and the
existence of a gap between the largest and the next-largest eigenvalues of T QTM
for finite temperature [31, 32].

Copyright © 2003 IOP Publishing Ltd.

 imaginary time direction β = N . τ
−τ

N
−τ

chain length L

Figure 9.1. Illustration of the two-dimensional classical model onto which the quantum
chain at finite temperature is mapped. The square lattice has width L identical to the chain
length and height identical to the Trotter number N. The alternating rows of the lattice
correspond to the transfer matrices T (τ ) and T (τ ), τ = β/N. The column-to-column
transfer matrix T QTM (quantum transfer matrix) is of particular importance to the
treatment of the thermodynamic limit. The arrows placed on the bonds indicate the type of
local Boltzmann weights, i.e. R- and R-matrices alternating from row to row. (The arrows
do not denote local dynamical degrees of freedom.)

The next-leading eigenvalues give the exponential correlation lengths ξ of

the equal time correlators at finite temperature:
 
1  max 
= lim ln  . (9.8)
ξ N→∞  

Finally, we want to comment on the study of thermodynamics of the quantum

L in the presence of an external magnetic field h coupling to the spin S =
chain
z
j =1 Sj , where Sj denotes a certain component of the j th spin, for instance Sj .
Of course, this changes (9.5) only trivially:

ZL,N := Tr{[T (τ )T (τ )]N/2 · eβhS }. (9.9)

On the lattice, the equivalent two-dimensional model is modified in a simple

way by a horizontal seam. Each vertical bond of this seam carries an individual
Boltzmann weight e±βh/2 if Sj = ±1/2 which indeed describes the action of the
operator

L
eβhS = eβhSj . (9.10)
j =1

Consequently, the QTM is modified by an h dependent boundary condition. It

is essential that these modifications can still be treated exactly as the additional
operators acting on the bonds belong to the group symmetries of the model.

Copyright © 2003 IOP Publishing Ltd.

 Figure 9.2. Sketch of the distribution of Bethe ansatz roots vj for finite N. Note that the
distribution remains discrete in the limit of N → ∞ for which the origin turns into an
accumulation point.

9.2.2 Bethe ansatz equations

The eigenvalues of T QTM are obtained by application of a Bethe ansatz. As the
QTM of the Heisenberg chain is just a staggered row-to-row transfer matrix, the
result can be obtained from [15] yielding
λ1 (v) + λ2 (v)
(v) = (9.11)
[(v − i(2 − τ ))(v + i(2 − τ ))]N/2
as a function of the so-called spectral parameter v. The terms λ1,2 (v) are defined
by
q(v + 2i)
λ1 (v) := e+βh/2 φ(v − i)
q(v)
(9.12)
−βh/2 q(v − 2i)
λ2 (v) := e φ(v + i)
q(v)
where φ(v) is simply

φ(v) := [(v − i(1 − τ ))(v + i(1 − τ ))]N/2 τ := β/N (9.13)

and q(v) is defined in terms of yet to be determined Bethe ansatz roots vj :


q(v) := (v − vj ). (9.14)
j

Note that we are mostly interested in  which is obtained from (v) simply by
setting v = 0. Nevertheless, we are led to the study of the full v-dependence since
the condition fixing the values of vj is the analyticity of λ1 (v) + λ2 (v) in the
complex plane. This yields
a(vj ) = −1 (9.15)
where the function a(v) is defined by
λ1 (v) φ(v − i)q(v + 2i)
a(v) = = eβh . (9.16)
λ2 (v) φ(v + i)q(v − 2i)

Copyright © 2003 IOP Publishing Ltd.

 Algebraically, we are dealing with a set of coupled nonlinear equations similar
to those occurring in the study of the eigenvalues of the Hamiltonian [15].
Analytically, there is a profound difference as here (9.16) the ratio of φ-functions
possesses zeros and poles converging to the real axis in the limit N → ∞. As
a consequence, the distribution of Bethe ansatz roots is discrete and shows an
accumulation point at the origin, see figure 9.2. Hence, the treatment of the
problem by means of linear integral equations for continuous density functions
[33] is not possible in contrast to the Hamiltonian case.

9 .3 Manipulation of the Bethe ansatz equations

The eigenvalue expression (9.11) under the subsidiary condition (9.15) has to
be evaluated in the limit N → ∞. This limit is difficult to take as an increasing
number N/2 of Bethe ansatz roots vj has to be determined. As the distribution of
these parameters is discrete, the standard approach based on continuous density
functions is not possible.

9.3.1 Derivation of nonlinear integral equations

The main idea of our treatment is the derivation of a set of integral equations for
the function a(v). This function possesses zeros and poles related to the Bethe
ansatz roots vj , see figure 9.3. Next we define the associated auxiliary function
A(v) by
A(v) = 1 + a(v). (9.17)
The poles of A(v) are identical to those of a(v). However, the set of zeros is
different. From (9.15) we find that the Bethe ansatz roots are zeros of A(v)
(depicted by open circles in figure 9.4). There are additional zeros farther away
from the real axis with imaginary parts close to ±2. For the sake of completeness,

Figure 9.3. Distribution of zeros (◦) and poles (×) of the auxiliary function a(v). All zeros
and poles vj ∓ 2i are of first order, the zeros and poles at ±(2i − iτ ), ±iτ are of order N/2.

Copyright © 2003 IOP Publishing Ltd.

 v j +2i
L
vj
Ŧi W
Ŧ2i+i W

Figure 9.4. Distribution of zeros and poles of the auxiliary function A(v) = 1 + a(v). Note
that the positions of zeros (◦) and poles (×) are directly related to those occurring in the
function a(v). There are additional zeros (
) above and below the real axis. The closed
contour L, by definition, surrounds the real axis and the zeros (◦) as well as the pole at
−iτ .

these zeros are depicted in figure 9.4 (open squares) but for a while they are not of
prime interest to our reasoning. Next we are going to formulate a linear integral
expression for the function log a(v) in terms of log A(v). To this end, we consider
the function
1 1
f (v) := log A(w) dw (9.18)
2πi L v − w
defined by an integral with closed contour L surrounding the real axis, the
parameters vj and the point −iτ in anticlockwise manner, see figure 9.4. Note
that the number of zeros of A(v) surrounded by this contour is N/2 and, hence,
identical to the order of the pole at −iτ . Therefore, the integrand log A(w) does
not show any non-zero winding number on the contour and consequentially the
integral is well defined. By use of standard theorems, we see that the function
f (v) is analytic in the complex plane away from the real axis with asymptotic
behaviour equal to zero. Due to Cauchy’s theorem, there is a jump discontinuity at
the real axis identical to the discontinuity of log A(v) itself. This discontinuity, in
turn, is identical to that of the function log[q(v)/(v + iτ )N/2 ] whose asymptotic
behaviour is equal to 0. We, therefore, have the identity
q(v)
f (v) = log (9.19)
(v + iτ )N/2
which is proved by noting the three properties of the difference function of the
left- and right-hand sides: (i) analyticity on the complex plane with a possible
exception at the real axis, (ii) continuity at the real axis (hence we have analyticity
everywhere) and (iii) zero asymptotics (from this follows boundedness—due to
Liouville’s theorem this bounded function is constant and, of course, equal to
zero).
Thanks to (9.18) and (9.19), we have a linear integral representation of
log q(v) in terms of log A(v). Because of (9.16) the function log a(v) is a linear

Copyright © 2003 IOP Publishing Ltd.

 combination of log q and explicitly known functions leading to
 
(v − iτ )(v + 2i + iτ ) N/2
log a(v) = βh + log
(v + iτ )(v + 2i − iτ )


1 1 1
+ − log A(w) dw (9.20)
2πi L v − w + 2i v − w − 2i
& '$ %
2 1
− .
π L (v − w)2 + 4
This expression for a(v) is remarkable as it is a NLIE of convolution type. It is
valid for any value of the Trotter number N which only enters in the driving (first)
term on the rhs of (9.20). This term shows a well-defined limiting behaviour for
N → ∞:
 
N (v − iτ )(v + 2i + iτ ) iβ iβ 2β
log →− + = (9.21)
2 (v + iτ )(v + 2i − iτ ) v v + 2i v(v + 2i)
leading to a well-defined NLIE for a(v) in the limit N → ∞:

2β 2 1
log a(v) = βh + − log A(w) dw. (9.22)
v(v + 2i) π L (v − w)2 + 4
This NLIE allows for a numerical (and, in some limiting cases, analytical)
calculation of the function a(v) on the axes Im(v) = ±1. About the historical
development, we would like to note that NLIEs very similar to (9.22) were
derived for the row-to-row transfer matrix in [34, 35]. These equations were then
generalized to the related cases of staggered transfer matrices (QTMs) of the
Heisenberg and RSOS chains [25, 26] and the sine-Gordon model [36].

9.3.2 Integral expressions for the eigenvalue

In (9.20), and (9.22) we have found integral equations determining the function a
for finite and infinite Trotter number N, respectively. The remaining problem is
the derivation of an expression for the eigenvalue  in terms of a or A.
From (9.11), we see that λ1 (v) + λ2 (v) is a rational function, and due to
the BA equations without poles. Hence, λ1 (v) + λ2 (v) is a polynomial and the
degree of this polynomial is N. Any polynomial is determined by its zeros and
the asymptotic behaviour. The zeros of λ1 (v) + λ2 (v) are solutions to a(v) =
λ1 (v)/λ2 (v) = −1, i.e. solutions to the BA equations or zeros of A(v) = 1 + a(v)
that do not coincide with BA roots! These zeros are so-called hole-type solutions
to the BA equations which we label by wl , l = 1, . . . , N. The holes are located
in the complex plane close to the axes with imaginary parts ±2, see zeros in
figure 9.4 depicted by . In terms of wl the function λ1 (v) + λ2 (v) reads as

λ1 (v) + λ2 (v) = (e+βh/2 + e−βh/2 ) (v − wl ). (9.23)
l

Copyright © 2003 IOP Publishing Ltd.

 Thanks to Cauchy’s theorem, we find for v not too far from the real axis that

1 1 1 N/2
[log A(w)]
dw = − (9.24)
2πi L v − w − 2i j
v − v j − 2i v + iτ − 2i

as the only singularities of the integrand surrounded by the contour L are the
simple zeros vj and the pole −iτ of order N/2 of the function A. Also, we obtain
  1
1 1 1 N/2
[log A(w)]
dw = − +
2πi L v − w j
v − vj − 2i l
v − wl v + 2i − iτ
(9.25)
where the evaluation of the integral is done by use of the singularities outside
of the contour L. We deform the contour such that the upper (lower) part of L
is closed into the upper (lower) half-plane where the relevant singularities are
the simple poles vj + 2i, the zeros wl and the pole iτ − 2i of order N/2 of the
function A.
Next, we take the difference of (9.24) and (9.25), perform an integration by
parts with respect to w and finally integrate with respect to v:


1 1 1
− log A(w) dw
2πi L v − w v − w − 2i
[(v − i(2 − τ ))(v + i(2 − τ ))]N/2
= log ( + constant. (9.26)
l (v − wl )

The constant is determined from the asymptotic behaviour of the lhs for v →
∞ with the result constant = − log A(∞) = − log(1 + exp(βh)). Combining
(9.23), (9.26) and (9.11), we find that

1 log A(w)
log (v) = −βh/2 + dw. (9.27)
π L (v − w)(v − w − 2i)
These formulas, (9.27) and (9.22), are the basis of an efficient analytical and
numerical treatment of the thermodynamics of the Heisenberg chain. There are,
however, variants of these integral equations that are somewhat more convenient
for the analysis, especially for magnetic fields close to zero [37].

9.4 Numerical results

By numerical integration and iteration, the integral equation (9.22) can be
solved on the axes Im(v) = ±1 defining functions a ± (x) := a(x ± i). Choosing
appropriate initial functions, the series ak± with k = 0, 1, 2, . . . converges
rapidly. In practice, only a few steps are necessary to reach a high-precision
result. Moreover, using the well-known fast Fourier transform algorithm, we can
compute the convolutions very efficiently.

Copyright © 2003 IOP Publishing Ltd.

 1.0
0.3
0.8
0.2

C
0.6 0.1

0.0
0.0 0.5 1.0 1.5 2.0
C/T

0.4 T

0.2

0.0
0.0 0.5 1.0 1.5 2.0
T

Figure 9.5. Specific heat coefficient c(T )/T versus temperature T for the spin- 12 XXX
chain. In the inset the specific heat c(T ) versus T is shown.

In order to calculate derivatives of the thermodynamical potential with

respect to temperature T and magnetic field h, one can avoid numerical
differentiations by utilizing similar integral equations guaranteeing the same
numerical accuracy as for the free energy. The idea is as follows. Consider the
function
∂ ∂ 1 ∂a a
laβ := log a with log(1 + a) = = laβ
∂β ∂β 1 + a ∂β 1+a
we directly obtain from (9.22) a linear integral equation for laβ if we regard
the function a as given. Once the integral equation (9.22) is solved for a, the
integral equation for laβ associated with (9.22) can be solved. In this manner, we
may continue to any order of derivatives with respect to T (and h). However, in
practice only the first and second orders matter. Here we restrict our treatment
to the derivation of the specific heat c(T ) and the magnetic susceptibility χ(T )
(derivatives of second order with respect to T and h), see figures 9.5 and 9.6. Note
the characteristic behaviour of c(T ) and χ(T ) at low temperatures. The linear
behaviour of c(T ) and the finite ground-state limit of χ(T ) are manifestations of
the linear energy–momentum dispersion of the low-lying excitations (spinons)
of the isotropic antiferromagnetic Heisenberg chain. In the high-temperature
limit, the asymptotics of c(T ) and χ(T ) are 1/T 2 and 1/T . This and the finite-
temperature maxima are a consequence of the finite-dimensional local degree
of freedom, i.e. the spin per lattice site. In figure 9.6 the numerical data down
to extremely low temperatures are shown providing evidence of logarithmic
correction terms, see [38] and later lattice studies [37, 39] confirming the field
theoretical treatment by [40]. These terms are responsible for the infinite slope

Copyright © 2003 IOP Publishing Ltd.

 0.108

0.106 0.14

0.12
0.104
χ

0.10
0.102
0.0 0.5 1.0 1.5 2.0
0 0.001 0.002 0.003 0.004 0.005
T

Figure 9.6. Magnetic susceptibility χ at low temperature T for the spin- 12 XXX chain. In
the inset χ(T ) is shown on a larger temperature scale.

of χ(T ) at T = 0 despite the finite ground-state value χ(0) = 1/π 2 . Precursors

of such strong slopes have been seen in experiments down to relatively low
temperatures, see, e.g., [41]. Unfortunately, most quasi one-dimensional quantum
spin systems undergo a phase transition at sufficiently low temperatures driven
by residual higher dimensional interactions. Hence, the onset of quantum critical
phenomena of the Heisenberg chain at T = 0 becomes visible but cannot be
identified beyond all doubts.
Another application of thermodynamical data of quantum spin chains is the
microscopic modelling of magnetic systems such as spin-Peierls compounds,
ladder systems etc. Quite generally, the microscopic interactions of a given
substance are determined by a comparative analysis of the experimental results
for the susceptibility and the theoretical data obtained for a model. In such a way,
the spin-Peierls compound CuGeO3 was found to be poorly described by a strictly
nearest-neighbour Heisenberg chain [42–44] (unlike the system mentioned earlier
[41]) but extremely well by a chain with nearest (J1 ) and next-nearest neighbour
interactions (J2 ). For CuGeO3 , the best agreement of experimental and theoretical
data was found for the ratio of exchange constants J2 /J1 = 0.35 and J1 = 79 K.

9.5 Low-temperature asymptotics

Next we would like to study the leading low-temperature behaviour (the
previously mentioned logarithmic terms are next-leading corrections). We will
closely follow [45]. The goal of this calculation is a comparison with the
conformal field theory results for the free energy as well as the correlation lengths.
As the system exhibits a spin-reversal symmetry, we may choose either sign of
the magnetic field h without changing the physical properties. (This is obvious

Copyright © 2003 IOP Publishing Ltd.

 for the Hamiltonian and also for the defining relation of the partition function. It
is probably not so obvious for the NLIE (9.22). Nevertheless, it can be shown that
+h ↔ −h is equivalent to a ↔ (1)/a.)
For our purpose it is the negative sign of h that simplifies the analysis of
(9.22) as a(v) is exponentially small on the axis Im(v) = +1, i.e. it is smaller
than exp(−Cβ) with some positive constant C. Therefore, ln A = ln(1 + a) on
the contour Im(v) = +1 is also exponentially small. Keeping just the contribution
due to the lower contour Im(v) = −1 and abusing the notation by writing a(x)
for a(v) with v = x − i we find that
∞
log a(x) = −β (x) +
0
K(x − y) log A(y) dy (9.28)
−∞

where we have employed the shorthand notation

2 1 2
K(x) := −  0 (x) := −h − . (9.29)
π x2 + 4 x2 +1
In a similar manner, we obtain for the eigenvalue
∞
1
log  = −βh/2 + ρ0 (x) log A(x) dx ρ0 (x) = . (9.30)
−∞ π(x 2 + 1)

These integral equations in the low-temperature limit allow us to make contact

with the dressed charge formulation. The key to this is the linearization of the
NLIE up to terms of order O(1/β). The necessary calculations are carried out
order by order (O(β), O(1), O(1/β)).
We want to do these calculations at the same time for the leading eigenvalue
max as well as for the next-leading eigenvalues  of the QTM. The NLIE
(9.22) (and also (9.28)) as written are valid just for the largest eigenvalue
max . The next-leading eigenvalues can be shown to satisfy very similar integral
equations with identical algebraic structure, but deformed integration contours.
At intermediate temperatures, the eigenvalues show involved crossover scenarios
involving complicated contours [46].
In the low-temperature limit, the distribution of roots simplifies [46]. In the
lower part of the complex plane, the roots (◦ in figure 9.4) and holes ( in lower
part of figure 9.4) bend towards each other. The extremal points touch each other
on the axis Im(v) = −1 with real parts that are denoted by ±x0 , see figure 9.7.
In order to compare the distribution of roots and holes at arbitrary temperature
T (and zero h) in figure 9.4 and the situation at low temperature (and h < 0)
in figure 9.7, the (◦, ) symbols have to be replaced by (•, ◦). Note that, in
figure 9.7, a reflection at the axis Im(v) = −1 has been performed in order to
have the sea of roots (•) below and the holes (◦) above.
A result of [46] is a change in the nature of the QTM excitations from
high to low temperature. In the intermediate temperature regime, we have the

Copyright © 2003 IOP Publishing Ltd.

 Im(v) roots
holes

x0 x0
Re(v)

Figure 9.7. Depiction of the root (•) and hole (◦) distribution at low temperature (and
h < 0). Note that both distributions touch the axis Im(v) = −1 at the points with real parts
±x0 . For purposes of illustration, the picture has been turned upside down (imaginary axis
is directed downwards) in order to have the holes above the roots.

so-called 1-strings, 2-strings etc but at low temperature we simply have particle–
hole-type excitations in very much the same way as in (free) Fermi systems.
Hence, excitations are constructed by changing roots into holes and vice versa.
The points ±x0 play the role of Fermi points and the number of roots is related
to the spin of the eigenstate of the QTM. The integral equations derived earlier
for the largest eigenvalue are still valid in the general case only if the integration
contour is designed such that it separates the roots from the holes.

9.5.1 Calculation to order O(β) and O(1)

We proceed stepwise by first allowing patterns physically describing excitations
of d many roots from one Fermi point to the other without changing the total spin
(s = 0). In the second step, we allow for a change in the total spin characterized
by the number s but no excitations from one Fermi point to the other (d = 0).
Finally, we admit any combination of d and s.

9.5.1.1 Particle–hole excitations from left to right Fermi point d = 0 (s = 0)

At the left (right) Fermi point a number d of holes/roots is circumvented in
a clockwise (counterclockwise) manner. At low T , the integral equation is
linearized by replacing log A by 0 for |x| ≥ x0 , and by log a + 2πid for |x| ≤ x0 :
log a = −β 0 + K ∗ (log a + 2πid) + O(1/β) (9.31)

where the symbol ∗ denotes convolution of two functions a ∗ b(x) = I a(x −
y)b(y) dy with integration interval I = [−x0, x0 ].

Copyright © 2003 IOP Publishing Ltd.

 root
hole

Figure 9.8. Spinless excitation.

We define the dressed energy  and dressed charge ξ by solutions to

 :=  0 + K ∗  ξ := 1 + K ∗ ξ. (9.32)

In terms of  and ξ (9.31) has the explicit solution

log a = −β + 2πid(ξ − 1) + O(1/β). (9.33)

From the last equation and the definition of the Fermi points ±x0 , i.e. (x0 ) = 0,
we have log a(x0 ) = [ξ(x0 ) − 1]2πid or

log a(x0 ) = 2πid(Z − 1) with Z := ξ(x0 ). (9.34)

9.5.1.2 Excitations with spin s = 0 (d = 0)

In contrast to the previous section where the roots and holes at different Fermi
points were surrounded in an opposite manner, in the present case the two Fermi
points are surrounded with an identical orientation. At low T , the integral equation
is linearized by replacing log A by −2πis for x < −x0, by log a for |x| < x0 and
by +2πis for x > x0 . Hence, we find

log a = −β 0 + 2πisK ◦ σ + K ∗ log +O(1/β) (9.35)

with an odd step function σ taking values −1, 0, +1 for x < −x0 , |x| < x0 ,
x > x0 , where ◦ denotes the convolution of two functions a ◦ b(x) = I a(x −
y)b(y) dy with integration interval I = ]−∞, −x0 ] ∪ [x0 , ∞[. Defining the

root
hole

Figure 9.9. Excitation with spin.

Copyright © 2003 IOP Publishing Ltd.

 function η as the solution to the integral equation

η=K ◦σ +K ∗η (9.36)

log a can be written as

log a = −β + 2πisη (9.37)

where  is the dressed energy. The function η remains to be determined. To this

end, we take the derivative of (9.36) yielding

η
(x) = K(x − x0 )(1 − η(x0 )) + K(x + x0 )(1 + η(−x0 )) + K ∗ η
(x)
= f0
+ K ∗ η
(x) (9.38)

with
f0
= [K(x − x0 ) + K(x + x0 )][1 − η(x0 )] (9.39)

where we have used η(−x) = −η(x), i.e. η is odd. Using this property once again,
we find

2η(x0 ) = η
(x) dx = 1 · η
= ξ · f0
(9.40)
|x|<x0 |x|<x0 |x|<x0
 
where we have employed the ‘dressed function trick’, i.e. ab0 = a0 b.
Substituting (9.39) into the last term of (9.40) and observing the definition
of ξ in (9.32), we obtain

2η(x0) = 2[1 − η(x0 )][ξ(x0 ) − 1] (9.41)

and, hence,
η(x0 ) = [1 − ξ −1 (x0 )]. (9.42)

By equation (9.37), therefore,

log a(x0) = 2πis(1 − Z −1 ). (9.43)

9.5.1.3 General case: arbitrary d, s

Due to linearity, we find

log a = −β + 2πid(ξ − 1) + 2πisη + O(1/β)

(9.44)
log a(±x0 ) = 2πid(Z − 1) ± 2πis(1 − Z −1 ).

Copyright © 2003 IOP Publishing Ltd.

 9.5.2 O(1/β) corrections
In order to determine the O(1/β) corrections in equation (9.44), we study the
explicit form of the nonlinear integral equations. As an example we take a term
like
∞
ψ(x − y) log A(y) dy = ψ(x − y) log(1 + a(y)) dy
−∞ |y|>x0
 
1
+ ψ(x − y) log a(y) dy + ψ(x − y) log 1 + dy
|y|<x0 |y|<x0 a(y)
(9.45)
where ψ is some function, for instance K or ρ0 . The linear term has already been
considered so we concentrate on the next two. We divide the integration over the
real line into two intervals [−∞, 0] ∪ [0, ∞] and first consider the negative real
axis:
−x0 0  
1
ψ(x − y) log(1 + a(y)) dy + ψ(x − y) log 1 + dy.
−∞ −x0 a(y)
For large β, the derivative of log a(x) is dominated by −β(x), where  takes
negative (positive) values inside (outside) the interval [−x0 , +x0]. Changing the
variable of integration to z = − log a(y), noting that y contributes only in a
vicinity of −x0 with width 1/β, and dz/dy  −(log a)
(−x0 ) = −β
(x0 ) leads
to
 ∞
1
−→
ψ(x + x0 ) log(1 + e−z ) dz
β (x0 ) − log a(−x0 )
− log a(−x0 ) 
+ log(1 + ez ) dz . (9.46)
−∞
The integration contour for z winds a certain yet-to-be-determined number n
times around the singularity of the log-function in the integrand. Its explicit
integration (indicated in the appendix) gives
 
1 π2 1
ψ(x + x0 ) + [log a(−x0) + 2πin] .
2
(9.47)
β
(x0 ) 6 2
Collecting the terms on the negative and positive parts of the real axis, we arrive
at
1

[+ ψ(x − x0 ) + − ψ(x + x0 )] (9.48)

β (x0 )
with
π2 1 π2 1
± = + [log a(±x0) + 2πin± ]2 = + [2πi(Z · d ∓ Z −1 · s)]2
6 2 6 2
(9.49)
where the number n± is identical to d ∓ s and we have used (9.44).

Copyright © 2003 IOP Publishing Ltd.

 9.5.3 O(1/β) corrections to the nonlinear integral equations
We linearize (9.28) in the general case of arbitrary values of d and s and account
for all O(β k ) terms with k = 1, 0, −1:
1
log a = −β 0 + 2πisK ◦ σ + [+ K(x − x0 ) + − K(x + x0 )]
β
(x 0)
+ K ∗ (log a + 2πid) (9.50)

where the O(1/β) term follows from (9.48) with setting ψ = K.

9.5.4 O(1/β) corrections to the eigenvalue

Next, we linearize (9.30) and keep terms of order O(β k ) with k = 1, 0, −1:
[+ ρ0 (x0 ) + − ρ0 (−x0 )]
log  = −βh/2 +
β
(x0 )
x0
+ ρ0 (x)[log a(x) + 2πid] dx (9.51)
−x0

where we have again ignored the contribution of the σ term in the integral as it is
an odd function. With the definition of the dressed density ρ satisfying

ρ = ρ0 + K ∗ ρ (9.52)
 
and using the ‘dressed function trick’, ab0 = a0 b, we obtain, with (9.50),
x0
ρ0 (x)[log a(x) + 2πid] dx
−x0
x0 x0
= −β ρ(x)0 (x) dx + 2πid ρ(x) · 1 dx
−x0 −x0

x0 
1
+
ρ(x)[+ K(x − x0 ) + − K(x + x0 )] (9.53)
β (x0 ) −x0
& '$ %
= + [ρ(x0 ) − ρ0 (x0 )] + − [ρ(−x0 ) − ρ0 (−x0 )]
where we have again ignored the contribution of the σ term in (9.50) as it is an
odd function. Combining the last equation with (9.51), we obtain
x0
log  = −βh/2 − β ρ(x)0 (x) dx
−x
& 0 '$ %
=: E
x0
ρ(x0 )
+ 2πid ρ(x) dx +(+ + − )
. (9.54)
−x0 β (x0 )
& '$ %
=: ρ

Copyright © 2003 IOP Publishing Ltd.

 The (sound) velocity of the elementary excitations is given by

v = 
/2πρ|x0 . (9.55)

For the largest eigenvalue (d = s = 0), we find

πT
−βf = log max = −β(h/2 + E) + (9.56)
6 v
and, for arbitrary integers d and s,
max 2πT
log = −2πiρd + [(Zd)2 + (Z −1 s)2 ]. (9.57)
 v
Finally, we would like to note that the density ρ is related to the Fermi momentum
kF = πρ. We find that log(max /), in general, is complex. With view to
(9.8), we see the real part defines the (reciprocal) correlation length 1/ξ and
the imaginary part is responsible for oscillations with wavevector 2dkF . The
characteristic divergence of all ξ s at low temperature is in quantitative accordance
with conformal field theory predictions, see [47, 48] and references therein.

9.6 Summary and discussion

We have reviewed a treatment of the thermodynamical properties of the isotropic
Heisenberg chain based on a lattice path-integral formulation and the definition
of the so-called quantum transfer matrix (QTM). A transparent analysis of its
eigenvalues has been given, resulting in a set of nonlinear integral equations
(NLIE) from which the free energy and the correlation lengths of the system
were derived. From a numerical solution of these NLIEs at arbitrary temperature,
the specific heat and magnetic susceptibility data were obtained. At very low
temperature, a linearization of the NLIE was performed yielding systematically
the leading terms of the low-temperature expansion. In particular, the O(1/β)
terms coincide with conformal-field-theory predictions. Also the mathematical
structure of dressed energy and dressed charge functions known from finite-size
studies of quantum chains at zero temperature was found. This may not look
surprising, but we would like to stress that the analysis of NLIEs for a lattice
system for finite T and infinite L is quite different from the analysis of T = 0
and finite L. Finally, we note that the derivation of the low-T results has been
presented for the special case of the isotropic Heisenberg chain. However, the
entire calculation can be directly carried over to the general XXZ case and, with
modifications, to multi-component systems.

Acknowledgments
The authors acknowledge financial support by the Deutsche Forschungsgemein-
schaft under grants Kl 645/3-3, 645/4-1 and the Schwerpunktprogramm SP1073.

Copyright © 2003 IOP Publishing Ltd.

 Im a

Re a
1 x

Figure 9A.1. Integration contour of variable a.

Appendix
Here we want to sketch the calculation of integrals of type (9.46)

x
log

log(1 + ez ) dz (9A.1)
−∞
-
with the contour along the real axis and at the end encircling a certain number n
of odd multiples of πi. We substitute z = log a and obtain the integral
x
log(1 + a)
In (x) := da (9A.2)
a
0

now with a contour starting at the origin and (A) surrounding the origin n times
in narrow loops in a clockwise manner, (B) followed by n larger loops in a
counterclockwise manner around the origin as well as −1 finally ending at x.
Obviously the first part (A) of the contour can be dropped. The simplest case with
n = 1 is illustrated in figure 9A.1.
We want to express In in terms of I0 . The explicit relation we are going to
prove is
In (x) = I0 (x) − 2π 2 n2 + 2πin log x. (9A.3)
The essential ingredient of our computation is the analytic continuation of I0 (x)
in x surrounding 0 and −1 exactly one time in a counterclockwise manner, see
figure 9A.1, giving I1 (x):
x −1 x
log(1 + a) log(1 + a) log(1 + a) + 2πi
I1 (x) = da = da + da
a 0 a −1 a
0 & '$ %
& '$ % ‘straight’
one loop
x
2πi
= I0 (x) + da = I0 (x) − 2π 2 + 2πi log x. (9A.4)
−1 a

Copyright © 2003 IOP Publishing Ltd.

 Now we can prove (9A.3) by induction. Obviously, it is correct for n = 0. Finally,
we consider (9A.3) valid as written and perform the analytic continuation of In (x)
in x (by surrounding 0 and −1 in a counterclockwise loop) leading to In+1 (x):

In+1 (x) = continuation of {In (x)}

= continuation of {I0 (x) − 2π 2 n2 + 2πin log x}
= I0 (x) − 2π 2 + 2πi log x − 2π 2 n2 + 2πin(log x + 2πi)
= I0 (x) − 2π 2 (n + 1)2 + 2πi(n + 1) log x (9A.5)

which is (9A.3) with n replaced by n + 1.

Finally we note that

0 
∞ zn 
∞ 0
ne ezn
log(1 + e ) dz =
z
(−1) dz = (−1) n
dz
n=1
n n=1
n
−∞ −∞

∞
1 π2
= (−1)n = (9A.6)
n=1
n2 12

from which we obtain

π2 1
I0 (x) + I0 (1/x) = + (log x)2 (9A.7)
6 2
which is verified by (i) inserting the special value x = 1 and (ii) taking derivatives
of both sides with respect to x.

References
[1] Takahashi M 1971 Phys. Lett. A 36 325–6
[2] Gaudin M 1971 Phys. Rev. Lett. 26 1301–4
[3] Schlottmann P 1992 J. Phys. Cond. Mat. 4 7565–78
[4] Jüttner G, Klümper A and Suzuki J 1997 Nucl. Phys. B 486 650
[5] Suga S and Okiji A 1997 Physica B 237 81–3
[6] Kawakami N, Usuki T and Okiji A 1989 Phys. Lett. A 137 287–90
[7] Jüttner G, Klümper A and Suzuki J 1998 Nucl. Phys. B 522 471
[8] Schulz H J 1993 Correlated Electron Systems vol 9 (Singapore: World Scientific)
p 199
[9] Eßler F H L and Korepin V E 1994 Exactly Solvable Models of Strongly Correlated
Electrons (Singapore: World Scientific)
[10] Lieb E H 1995 Advances in Dynamical Systems and Quantum Physics (Singapore:
World Scientific) p 173
[11] Belavin A A, Polyakov A M and Zamolodchikov A B 1984 Nucl. Phys. B 241 333
[12] Bethe H 1931 Z. Phys. 71 205–26
[13] Sutherland B 1970 J. Math. Phys. 11 3183–6
[14] Baxter R J 1971 Phys. Rev. Lett. 26 834

Copyright © 2003 IOP Publishing Ltd.

 [15] Baxter R J 1982 Exactly Solved Models in Statistical Mechanics (London: Academic
Press)
[16] Takahashi M 1971 Prog. Theor. Phys. 46 401–15
[17] Yang C N and Yang C P 1969 J. Math. Phys. 10 1115–22
[18] Koma T 1987 Prog. Theor. Phys. 78 1213–18
[19] Suzuki M and Inoue M 1987 Prog. Theor. Phys. 78 787–99
[20] Bariev R Z 1982 Theor. Math. Phys. 49 1021
[21] Truong T T and Schotte K D 1983 Nucl. Phys. B 220[FS8] 77–101
[22] Suzuki J, Akutsu Y and Wadati M 1990 J. Phys. Soc. Japan 59 2667–80
[23] Suzuki J, Nagao T and Wadati M 1992 Int. J. Mod. Phys. B 6 1119–80
[24] Takahashi M 1991 Phys. Rev. B 43 5788
Takahashi M 1992 Phys. Rev. B 44 12382
[25] Klümper A 1992 Ann. Phys., Lpz 1 540–53
[26] Klümper A 1993 Z. Phys. B 91 507
[27] Kuniba A, Sakai K and Suzuki J 1998 Nucl. Phys. B 525 597
[28] Takahashi M 2000 Preprint cond-mat/0010486, ISSP Technical Report A3579
[29] Takahashi M, Shiroishi M and Klümper A 2001 J. Phys. A: Math. Gen. 34 L187–94
[30] Kato G and Wadati M 2002 J. Math. Phys 43 5060
[31] Suzuki M and Inoue M 1987 Prog. Theor. Phys. 78 787
[32] Suzuki J, Akutsu Y and Wadati M 1990 J. Phys. Soc. Japan 59 2667
[33] Hulthén L 1938 Ark. Mat. Astron. Fys. A 26 1–105
[34] Klümper A and Batchelor M T 1990 J. Phys. A: Math. Gen. 23 L189–95
[35] Klümper A, Batchelor M T and Pearce P A 1991 J. Phys. A: Math. Gen. 24 3111–33
[36] Destri C and de Vega H J 1992 Phys. Rev. Lett. 69 2313–17
[37] Klümper A 1998 Eur. Phys. J. B 5 677
[38] Eggert S, Affleck I and Takahashi M 1994 Phys. Rev. Lett. 73 332
[39] Klümper A and Johnston D C 2000 Phys. Rev. Lett. 84 4701
[40] Lukyanov S 1998 Nucl. Phys. B 522 533
[41] Takagi S, Deguchi H, Takeda K, Mito M and Takahashi M 1996 J. Phys. Soc. Japan
65 1934–7
[42] Riera J and Dobry A 1995 Phys. Rev. B 51 16098
[43] Castilla G, Chakravarty S and Emery V J 1995 Phys. Rev. Lett. 75 1823
[44] Fabricius K, Klümper A, Löw U, Büchner B and Lorenz T 1998 Phys. Rev. B 57
1102
[45] Scheeren C 1998 Diploma Thesis Cologne University
[46] Klümper A, Martı́nez J R, Scheeren C and Shiroishi M 2001 J. Stat. Phys. 102 937–
51
[47] Bogoliubov N M and Korepin V E 1989 Int. J. Mod. Phys. B 3 427–39
[48] Korepin V E, Bogoliubov N and Izergin A 1993 Quantum Inverse Scattering Method
and Correlation Functions (Cambridge: Cambridge University Press)

Copyright © 2003 IOP Publishing Ltd.

 Chapter 10

Reaction–diffusion processes and their

connection with integrable quantum spin
chains
Malte Henkel
Laboratoire de Physique des Matériaux (CNRS UMR 7556)
et Laboratoire Européen de Recherche Universitaire
Sarre-Lorraine, Université Henri Poincaré Nancy I, France

10.1 Reaction–diffusion processes

The understanding of non-equilibrium statistical physics is still much more
incomplete than that of equilibrium theory, due to the absence of an analogue
of the Boltzmann–Gibbs approach and in spite of considerable recent progress
[1]. Therefore, non-equilibrium systems have to be specified by some defining
dynamical rules which are then analysed. The topic has received a lot of attention
and many reviews exist, e.g. [2–9]. Exactly solvable systems far from equilibrium
have been recently reviewed in a nice way [10]. Here a pedagogically-minded
introduction to the application of a few standard tools from one-dimensional
integrable quantum systems to non-equilibrium statistical mechanics is presented.
After recalling why standard descriptions such as kinetic or reaction–diffusion
differential equations are, in general, insufficient in one dimension, we remind
the reader in section 10.2 on the quantum Hamiltonian formulation of non-
equilibrium processes which, in turn, is based on the master equation. Section 10.3
recalls a few basic facts about Hecke algebras. These building blocks are used
in sections 10.4 and 10.5 to show explicitly the integrability of certain single-
species reaction–diffusion processes, through their relation to integrable vertex
models. Section 10.6 reviews some further methods such as spectral and partial
integrability, the free-fermion technique, similarity transformations or diffusion
algebras. We close in section 10.7 with an outlook on how the recently introduced

Copyright © 2003 IOP Publishing Ltd.

 Figure 10.1. Microscopic reactions in the diffusive pair-annihilation process.

concept of local scale-invariance might become useful in the description of non-

equilibrium ageing phenomena.
No effort has been made to provide a complete bibliography. This may be
found in the excellent reviews quoted earlier.
A class of non-equilibrium models which are particularly simple to formulate
are the so-called reaction–diffusion processes. Consider the following example:
particles of a single kind (species) move on a lattice (figure 10.1). Each site of the
lattice may be either empty (denoted by ◦) or else be occupied by a single particle
(denoted by •). The particles are allowed to undergo the following movements,
see figure 10.1, which involve the states of two nearest-neighbour sites:

◦+•↔•+◦ diffusion, with rate D

(10.1)
•+•→◦+◦ pair-annihilation, with rate 2α.

The first of the allowed movements in (10.1) is reversible while the second is
not. A typical question is then for the long-time behaviour of such quantities
as the mean particle density n(t). Trivially n(t) decreases with increasing time
t but different long-time asymptotic behaviours such as n(t) ∼ t −y or n(t) ∼
e−t /τ are conceivable. The oldest approach to this problem was introduced by
Smoluchowski [11] and consists of writing down kinetic equations, e.g. for the
spatially averaged density n(t) and one obtains ∂t n(t) = −λn(t)2 for the problem
at hand, where λ = 4α. With the initial condition n(0) = n0 , the solution n(t) =
n0 (1 + n0 λt)−1  (λt)−1 is easily found and apparently answers the physical
question. A slightly more involved version of this argument does allow for spatial
variation of the density n = n(r, t) and considers a reaction–diffusion equation

∂t n(r, t) = D∇ 2 n(r, t) − λn(r, t)2 . (10.2)

While the analysis of such nonlinear partial differential equations is a formidable

problem in its own right, these equations do not yet capture the essential physics
in low-spatial dimensions, as we now show. Rather, they must be considered as
an approximation of mean-field type.
In order to understand the approximative nature of equations such as (10.2),
and following [12], consider the mean particle density in a large volume V :

1
n̄(t) = dr n(r, t). (10.3)
V V

Copyright © 2003 IOP Publishing Ltd.

 It then follows that

1
∂t n̄(t) = dr [D∇ 2 n(r, t) − λn(r, t)2 ]
V V

D λ
= dσ · ∇n(r, t) − dr n(r, t)2
V ∂V V V
 2
1
≤ −λ dr n(r, t) . (10.4)
V V
In the second line, Gauss’ theorem was used where dσ is normal to the surface
∂V (the flow through ∂V vanishes for large volumes V → ∞). The last line
follows from the Cauchy–Schwarz inequality. Together with the initial condition
n̄(0) = n0 , the inequality ∂t n̄(t) ≤ −λn̄(t)2 yields the bound
n0 1
n̄(t) ≤ ≤ (10.5)
1 + n0 λt λt
for all times t ≥ 0. However, the model just defined can be solved exactly in
one spatial dimension (in a setting defined precisely in section 10.3), provided
D = 2α. The exact mean particle density is given by [13]


∞ 
n̄(t) = n0 e−4Dt I0 (4Dt) + 2(1 − n0 ) (1 − 2n0 )k−1 Ik (4Dt)
k=1
1
√ t −1/2 t →∞ (10.6)
8πD
where Ik is the modified Bessel function of order k. Clearly, for large times, the
exact mean density n̄(t) decreases considerably slower than even the upper bound
(10.5) derived from the reaction–diffusion equation (10.2).
The failure of equation (10.2) to describe correctly the long-time behaviour
can be understood from the following heuristic argument [14]. In the long-time
limit, the particle density should already be very low and it is conceivable that,
at most, one annihilation reaction takes place at a given time. Let L = L(t) be
the typical distance between two particles. Then, the time needed by diffusive
motion to overcome this distance is of the order t ∼ L(t)2 . However, the mean
particle density is n̄(t) ∼ L(t)−d in d spatial dimensions and this argument would
give n̄(t) ∼ t −d/2 . Therefore, the assumption implicit in equations such as (10.2)
that diffusive motion can render the system sufficiently homogeneous fails in low
dimensions (in our model for d < 2) and one has instead the long-time behaviour
[14] 
t −d/2 if d < 2
n̄(t) ∼ −1 (10.7)
t if d > 2.
Therefore, d ∗ = 2 is the upper critical dimension of the diffusive pair-annihilation
process. For dimensions d > d ∗ , reaction–diffusion equations should be expected

Copyright © 2003 IOP Publishing Ltd.

 Table 10.1. Measured decay exponent y of the mean exciton density n̄(t) ∼ t −y on
polymer chains. The error bar for TMMC comes from averaging over the results of [17]
for different initial particle densities.

Substance y Reaction(s) Reference


◦◦
C10 H8 0.52–0.59 •• → [15]
•◦

◦◦
P1VN/PMMA film 0.47(3) •• → [16]
•◦
TMMC 0.48(4) •• → •◦ [17]

to give qualitatively correct results and entire branches of physical chemistry

are built on this. However, for low-dimensional structures with d < d ∗ , as might
occur, for example, in nanodevices, fluctuation effects become dominant.
The importance of fluctuations in low-dimensional reaction–diffusion
processes has also been confirmed experimentally. An effectively one-
dimensional setting can be achieved by studying the kinetics of excitons (localized
electronic excitations) along polymer chains (other examples are reviewed in
[6, 10]). For details, consult the reviews by Kroon and Sprik and Kopelman and
Lin in [4]. The only purpose of the polymer chains is to provide a carrier for the
excitons. Schematically, single excitons may hop from one monomer to the next
(thus modelling a diffusive motion) while a reaction occurs if two excitons meet.
One may have one or both of the reactions •• → ◦◦ or •• → •◦, see table 10.1.
We shall show in section 10.6 that for any branching ratio

B = (•• → ◦◦)/ (•• → •◦)

the long-time behaviour is still described by the model (10.1), with a renormalized
rate. For late times, one expects the mean exciton density to fall off as a power
law n̄(t) ∼ t −y . The excitons are unstable, with lifetimes of the order 10−3 s.
Their decay produces light of a characteristic frequency whose intensity can be
used to measure n̄(t) while light with a different frequency is emitted if excitons
decay through a pair reaction. This allows one to measure ∂t n̄(t) as well, through
time-resolved experiments down to the picosecond scale. Table 10.1 gives some
results for the exponent y in different materials (the branching ratio B  10% in
the first two lines of table 10.1). Clearly y  1/2 as expected from (10.6) and far
from unity. This is strong evidence in favour of strong fluctuation effects in these
systems and against their description through a reaction–diffusion equation (10.2).
Another aspect becomes apparent if we now briefly consider the triple
annihilation process • • • → ◦ ◦ ◦ combined with single-particle diffusion. The

Copyright © 2003 IOP Publishing Ltd.

 reaction–diffusion equation reads as
∂t n(r, t) = D∇ 2 n(r, t) − λn(r, t)3 . (10.8)
Following the same lines as before but using now the Hölder inequality, the
differential inequality ∂t n̄(t) ≤ −λn̄(t)3 leads to the bound
n0 1
n̄(t) ≤ * ≤ √ t −1/2 . (10.9)
1 + 2n20 λt 2λ

This is of the same order of magnitude as the long-time behaviour expected

from diffusive motion in one spatial dimension. Therefore, and in agreement with
scaling arguments showing that d ∗ = 1 [18], already for the triple annihilation
process, fluctuation effects should not play a major role in any physically
realizable dimension (d > 1)1.
In conclusion, the description of reaction–diffusion processes with pair
reactions in low dimensions requires a truly microscopic approach beyond kinetic
reaction–diffusion equations while these equations may well turn out to be
adequate for multi-particle reactions. For that reason, we shall, in the following,
consider the master equation formulation of reaction–diffusion processes with
pair-reaction terms only.

10.2 Quantum Hamiltonian formulation

We now review the reformulation of a non-equilibrium stochastic system defined
by some master equation in terms of the spectral properties of an associated
quantum Hamiltonian H and which goes back at least to the classic paper by
Glauber [20]. To be specific, we consider in this section only systems defined on
a chain with L sites and two allowed states per site. We represent the states of
the system in terms of spin configurations σ = {σ1 , σ2 , . . . , σL } where σ = +1
corresponds to an empty site and σ = −1 corresponds to a site occupied by a
single particle. We are interested in the probability distribution function P (σ ; t)
of the configurations σ . Our starting point is the master equation for the P (σ ; t):

∂t P (σ ; t) = [w(τ → σ )P (τ ; t) − w(σ → τ )P (σ ; t)] (10.10)
τ =σ

where w(τ → σ ) are the transition rates between the configurations τ and σ and
are assumed to be given from the phenomenology. In order to rewrite this as a
matrix problem, one introduces a state vector

|P # = P (σ ; t)|σ # (10.11)
σ
1 This does not imply, however, that models containing both binary and multi-site reaction terms
could not have a non-trivial behaviour. For example, the phase structure of the pair contact process
(•• → ◦◦, • • ◦ → • • •) with single-particle diffusion (•◦ ↔ ◦•) is presently controversial and under
very active study [19].

Copyright © 2003 IOP Publishing Ltd.

 and equation (10.10) becomes

∂t |P # = −H |P # (10.12)

where the matrix elements of H are given by


"σ |H |τ # = −w(τ → σ ) + δσ,τ w(τ → υ). (10.13)
υ

The operator H describes a stochastic process since allthe elements of

the columns add up to zero. Conservation of probability σ P (σ ; t) = 1 is
equivalent to the relation
"s|H = 0 (10.14)

where "s| = σ "σ | is a left eigenvector of Hwith eigenvalue 0. Correspondingly,
H has at least one right eigenvector |s# = σ Ps (σ )|σ # with eigenvalue 0, that
is H |s# = 0. Such a vector does not evolve in time and, therefore, corresponds to
a steady-state distribution of the system. Since, in general, H is not symmetric,
this steady-state vector may be highly non-trivial. Note that all this is completely
general and applies to any stochastic process defined by a master equation. With a
view to the processes to be studied later one calls H a quantum Hamiltonian and
this formulation of the master equation the quantum Hamiltonian formalism (see
[6, 10] for recent reviews). The reason for this choice of language is the fact that
for the processes studied later (and, in fact, many other processes as well) H is the
Hamiltonian of some quantum system such as the Heisenberg XXZ Hamiltonian.
The steady state |s# of a stochastic system corresponds in this mapping to the
ground state of the quantum system. The probabilistic interpretation is guaranteed
by the following classical result.

Theorem 10.1 (Hyver–Keizer–Schnackenberg [21]) For a quantum

 Hamiltonian
H which satisfies the master equation (10.12) and has "s| = σ "σ | as a left
eigenstate such that "s|H = 0, the following holds.
(i) There is a stationary state

|s# = Ps (σ )|σ # (10.15)
σ

such that H |s# = 0.

(ii) Consider the eigenvalue problem H |n# = En |n#, with n = 0, 1, 2 . . . . Then

Re En ≥ E0 = 0. (10.16)

(iii) Let |P0 # = |P (0)# be the initial state such that the weights of the individual
configurations satisfy 0 ≤ P (σ ; 0) ≤ 1 and "s|P0 # = 1. Then, for all times
t ≥ 0, one has

0 ≤ P (σ ; t) ≤ 1 and "s|P # = 1. (10.17)

Copyright © 2003 IOP Publishing Ltd.

 (iv) Let H : Rn → Rn be a linear map such that for the elements Hij of H holds

n
Hij ≤ 0 Hij = 0 ∀j ∈ {1, . . . , n}. (10.18)
i=1
Then H is a ‘quantum Hamiltonian’ of a Markov process described by the
master equation (10.12).
Time-dependent averages of an observable F are given by the matrix element

"F #(t) = F (σ )P (σ ; t) = "s|F |P # = "s|F exp(−H t)|P0 # (10.19)
σ
and we see that equation (10.16) means that the system, indeed, evolves towards
the steady-state |s#, thus Ps (σ ) = P (σ ; ∞).
In what follows, we shall be mainly interested in averages of particle
numbers nj at site j and their correlators. These can be expressed in the quantum
spin formulation in terms of the projector
 
1 z 0 0
ñj = (1 − σj ) = (10.20)
2 0 1 j

and one-point and two-point functions of the nj are then expressed as2
C1 (j ; t) = "ñj #(t) = "s|ñj |P # C2 (j, ; t) = "ñj ñ #(t) = "s|ñj ñ |P #.
(10.21)
Two basic situations are readily distinguished from the spectrum of H . If in
the limit of infinite lattice size L → ∞, the lowest excited states have a finite gap
 to the ground-state energy E0 = 0, then the averages (10.21) will approach their
steady-state values exponentially fast on the time-scale τ = 1/ . However, if
there is, in the L → ∞ limit, a continuous spectrum down to E0 = 0, one expects
an algebraic decay of the correlators as the system approaches the steady state.

10.3 Hecke algebra and integrability

Before we shall write down quantum Hamiltonians for certain reaction-diffusion
systems explicitly, we need some background information on Hecke algebras in
order to make contact with integrability. The Hecke algebra Hn (q) is spanned by
n generators ei :
Hn (q) = {e1 , e2 , . . . , en } (10.22)
which satisfy the following relations
ei ei±1 ei − ei = ei±1 ei ei±1 − ei±1
ei ej = ej ei if |i − j | ≥ 2 (10.23)
−1
ei2 = (q + q )ei
2 We stress that the structure of these matrix elements is quite distinct from expectation values "0|F |0#
in ordinary quantum mechanics.

Copyright © 2003 IOP Publishing Ltd.

 where q ∈ C is a parameter. The representations of Hn (q) and the relationship to
equilibrium statistical mechanics are discussed in great detail in [22]. We are not
only interested in Hn (q) but also in some quotients, denoted by (P , M)Hn (q)
[23]. Two specific examples will be of interest to us. The first such quotient
is the celebrated Temperley–Lieb algebra (2, 0)Hn (q), where, in addition, to
equation (10.23) the additional relations
ei ei±1 ei − ei = 0 ei±1 ei ei±1 − ei±1 = 0 (10.24)

hold. However, the quotient (1, 1)Hn (q) is defined through the condition [23]
(ei ei+2 )ei+1 (q + q −1 − ei )(q + q −1 − ei+2 ) = 0 i = 1, 2, . . . . (10.25)

For the definition of more general quotients, we refer to [23].

Consider the N × N matrices E ab where a, b = 0, 1, . . . , N − 1. The only
non-vanishing element of E ab is the one on the ath line and the bth column and
this element is equal to unity. Define further
Eiab = 1 ⊗ · · · ⊗ 1 ⊗ E ab ⊗ 1 ⊗ · · · ⊗ 1 (10.26)
where E ab occurs on position i and i = 1, . . . , L runs over the sites of a chain.
Explicit realizations of the Hecke algebra or one of its quotients may be found in
the Perk–Schultz models [24], whose Hamiltonian is of the form

L−1
H (P ,M) = ei(P ,M) . (10.27)
i=1

The importance of this observation becomes clear from the following.

L−1
Theorem 10.2 (Jones [25]) If H = i=1 ei , where the ei are the generators of
the Hecke algebra HL−1 (q), H is integrable through the Baxterization procedure.
The Baxterization procedure allows to define, starting from H , in a
systematic way Boltzmann weights which satisfy the Yang–Baxter equation. We
shall illustrate this in the example of the seven-vertex model in section 10.5.
In many practical applications, the following result is useful.
L−1
Theorem 10.3 (Martin and Rittenberg [23]) If H = i=1 ei , where the ei are
the generators
L−1 of the Hecke algebra quotient (P , M)H L−1 (q) and, furthermore,
H
= i=1 fi , where the fi are different generators of the same quotient
(P , M)HL−1 (q), then H and H
have the same eigenvalues, up to degeneracies.
We finish this section by writing explicit examples for the quotients (2, 0)
and (1, 1) in the case N = 2 [26]. Then the matrices E ab can be expressed through
Pauli matrices
     
0 1 0 −i 1 0
σ =
x
σ =
y
σ =
z
(10.28)
1 0 i 0 0 −1

Copyright © 2003 IOP Publishing Ltd.

 x,y,z
and we define σi by analogy with (10.26). Set


(2,0) 1 y y q + q −1 z z
ei = − σix σi+1
x
+ σi σi+1 + (σi σi+1 − 1)
2 2

q − q −1 z z
− (σi − σi+1 ) . (10.29)
2
Therefore, the Hamiltonian [27, 28]

L−1
(2,0)
H (2,0) = ei (10.30)
i=1

is integrable. In addition, it satisfies a quantum group invariance, since

[H (2,0), S z ] = [H (2,0), S ± ] = 0 (10.31)

where, recalling also that σ ± = 12 (σ x ± iσ y ),

1 L L
Sz = σiz S± = Si±
2 i=1 i=1
    (10.32)
ln q 
i−1
ln q L
Si± = exp σz σi± exp − σz
2 =0 2 =i+1

which, in turn, obey the commutation relations of Uq (su(2)), namely

q 2S − q −2S
z z

[S z , S ± ] = ±S ± [S + , S − ] = . (10.33)
q − q −1
However, set

+ σi σi+1 + qσiz + q −1 σi+1 − q − q −1 ].

(1,1) y y z
ei = − 12 [σix σi+1
x
(10.34)
L−1 (1,1)
Here, the associated integrable Hamiltonian [29] H (1,1) = i=1 ei is also
invariant under the supersymmetric quantum group Uq (su(1|1)), since

[H (1,1), S z ] = [H (1,1), T ± ] = 0 (10.35)

where
 

L
iπ 
i−1
T ±
=q (1−L)/2
q i−1
exp (σ + 1) σi±
z
(10.36)
i=1
2 =1 

and
q L − q −L
[S z , T ± ] = ±T ± {T + , T − } = . (10.37)
q − q −1

Copyright © 2003 IOP Publishing Ltd.

 10.4 Single-species models
We are now ready to study explicit examples of stochastic quantum Hamiltonians.
The classical example merely considers particles of a single species (•) which
may hop randomly onto an empty nearest-neighbour site (◦), thereby modelling
the reversible reaction •◦ ↔ ◦• with rate D. This process is often called the
symmetric exclusion process. The quantum Hamiltonian reads as


D L−1 y y
H =− [σ x σ x + σj σj +1 + (σjz σjz+1 − 1)] (10.38)
2 j =1 j j +1

and coincides with the (ferromagnetic) XXX Heisenberg quantum chain [30].
Certainly, one may now use the Bethe ansatz solution of HXXX to rederive
known results on simple diffusion. The recent interest in this set-up comes from
the insight that the integrability of the associated quantum chains allows us to
make contact with the pre-established algebraic techniques for the treatment of
these [26, 31]. Independently, integrability was also observed to occur in the
transfer matrices for discrete-time dynamics [32, 33]. The enormous possibilities
for non-trivial applications then triggered an ongoing wave of activity, see, e.g.,
[2, 4, 10, 26, 34–39] and references therein.
Following [26, 31], we now give more examples of integrable quantum
Hamiltonians of stochastic systems, restricting ourselves for simplicity to a single
species of particles and to binary reactions only (see section 10.1). The reaction
rates are defined in table 10.2, using the convention of various authors but
unfortunately there is no standard notation. While we prefer a light notation
(slightly modified from [40]) and shall use it here3 , other authors often opt for
a systematic though heavier notation with several indices.
For the time being and for purposes of illustration, let us consider besides
diffusion only those reactions which irreversibly reduce the number of particles
(that is βL,R = νL,R = σ = 0). Define
√ √ )  )
γL γR δL δR DL γL δL
D = DL DR , γ = , δ= , q= = =
D D DR γR δR
(10.39)
= 12 (q + q −1 )(1 + δ − γ ) − α/D, h = 12 (2α/D + γ (q + q −1 )). (10.40)

Note that the ratio of the left and right rates is taken to be the same for diffusion,
coagulation and death processes. We first consider an open chain with L sites.
Then the quantum Hamiltonian becomes

H = D(HXXZ (h, , δ) + Hα + Hγ + Hδ ) (10.41)

3 The letter ν is inspired by naissance (French for birth) and σ comes from Schöpfung (German for
creation). The letter β might have come from branching.

Copyright © 2003 IOP Publishing Ltd.

 Table 10.2. Two-sites reaction–diffusion processes of a single species and their rates as
denoted by various authors.

diffusion to the left ◦• → •◦ DL a32 w32 w1,1 (1, 0) 01

10
diffusion to the right •◦ → ◦• DR a23 w23 w1,1 (0, 1) 10
01
pair annihilation •• → ◦◦ 2α a14 w14 w1,1 (0, 0) 11
00
coagulation to the right •• → ◦• γR a24 w24 w1,0 (0, 1) 11
01
coagulation to the left •• → •◦ γL a34 w34 w0,1 (1, 0) 11
10
death at the left •◦ → ◦◦ δL a13 w13 w1,0 (0, 0) 10
00
death at the right ◦• → ◦◦ δR a12 w12 w0,1 (0, 0) 01
00
decoagulation to the left ◦• → •• βL a42 w42 w1,0 (1, 1) 01
11
decoagulation to the right •◦ → •• βR a43 w43 w0,1 (1, 1) 10
11
birth at the right ◦◦ → ◦• νR a21 w21 w0,1 (0, 1) 00
01
birth at the left ◦◦ → •◦ νL a31 w31 w1,0 (1, 0) 00
10
pair creation ◦◦ → •• 2σ a41 w41 w1,1 (1, 1) 00
11

Rates defined after reference [40] [35] [10] [31] [41]

where HXXZ (h, , δ) is the standard XXZ quantum chain, including bulk and
boundary magnetic fields


1 L−1 y y
HXXZ (h, , δ) = − σ x σ x + σj σj +1 + (σjz σjz+1 − 1)
2 j =1 j j +1

1
+ h(σjz + σjz+1 − 2) − (1 − δ)(q − q −1 )(σjz − σjz+1 )
2
(10.42)
which contains the diagonal and diffusive matrix elements while the particle
annihilation terms are contained in

L−1
Hα = −2α q −2j −1 σj+ σj++1
j =1

L−1
Hγ = −γ q −j (ñj σj++1 + q −1 σj+ ñj +1 ) (10.43)
j =1

L−1
Hδ = −δ q −j (q −2 (1 − ñj )σj++1 + qσj+ (1 − ñj +1 ))
j =1

Copyright © 2003 IOP Publishing Ltd.

 and σ ± = 12 (σ x ± iσ y ) are the one-particle annihilation/creation operators.
For a physical understanding of this we consider two special cases.

(1) Consider pure asymmetric diffusion, that is α = γ = δ = 0, also referred

to as the asymmetric exclusion process. Then H = DH (2,0) as given
by equations (10.29) and (10.30). We thus have√a very clear physical
interpretation of the quantum-group parameter q = DR /DL [31]. Besides
simple biased diffusion, this model is related, e.g. to the 1D Kardar–Parisi–
Zhang equation or to the noisy Burgers equation [42]. The quantum group
may be used for the calculation of correlation functions [43].
(2) In addition to diffusion, add annihilation such that 2α = DL + DR , that is
= 0 and keep γ = δ = 0. Then H = D(H0 + H1 ), where the Hermitian
part H0 = H (1,1) is given by equation
(10.34) and this part alone is, therefore,
L−1
supersymmetric. However, H = D i=1 fi , where fi ∈ (1, 1)HL−1(q), but
the fi are no longer symmetric. This was the first example of a non-
symmetric realization of a Hecke algebra [31]. We note that besides the
already established integrability, this system is also soluble through free-
fermion techniques.

Proposition 10.1 [31] The spectrum of H in equation (10.41) is independent of

the particle-reaction terms contained in equation (10.43), that is

spec(H ) = spec(DHXXZ (h, , δ)). (10.44)

To see this, recall that the XXZ Hamiltonian conserves the number of
particles while the reaction terms irreversibly decrease the total particle number.
Thus, H can be written in a block diagonal form
 
N0 Xδ Xα
 N1 Xγ ,δ Xα 
 
H =  . . 
 (10.45)
 N2 Xγ ,δ .

.. ..
. .

where Nn refers to the n-particle states and X are the reaction matrix elements.
Because of the identity
 
A X
det = det A det B (10.46)
0 B

it follows that the elements of (10.43) do not enter into the characteristic
polynomial of H . 
Therefore, the phase diagram for the full Hamiltonian H can be read off
from the well-known spectrum of HXXZ (h, , δ) [44]. For our purposes, we need
the following [31]. From (10.40), only the portion of the phase diagram where

Copyright © 2003 IOP Publishing Ltd.

 h + ≥ 1 is important for us. First, the spectrum always has a finite gap when
h + > 1, which is realized whenever δ = 0 or q = 1. Then the ground state
of HXXZ is a trivial ferromagnetic frozen state which corresponds to the empty
state ◦ ◦ · · · ◦ ◦. The energy gap  = E1 − E0 is finite. Second, the spectrum
is gapless for + h = 1, where the system undergoes a Pokrovsky–Talapov
transition. This situation occurs for δ = 0 and q = 1. We have thus identified the
cases where the model approaches the steady state exponentially (non-vanishing
gap) or algebraically (gapless).
At this point, it is of interest to discuss the role of the boundary conditions
and we consider now a periodic chain, for simplicity just for the asymmetric
exclusion process (that is α = γ = δ = 0). We stress that if q = 1, Hper cannot
be read from (10.41) by simply taking periodic boundary conditions. Rather, we
have
 
1 L
q + q −1 z z
+ − −1 − +
Hper = − qσi σi+1 + q σi σi+1 + (σi σi+1 − 1)
q + q −1 i=1 4
(10.47)
±
together with the periodic boundary conditions σL+1 = σ1± and σL+1
z
= σ1z . The
hopping terms can be brought back to the familiar XXZ form of equation (10.41)
through a similarity transformation Hper
= UH U −1 with the matrix U =
per
L z
exp(πg =1 σ ) with q = e 2πg such that [45]
 
1  L
q + q −1

y y
Hper =− σ x σ x + σi σi+1 + (σiz σi+1
z
− 1)
2(q + q −1 ) i=1 i i+1 2
(10.48)
which looks the same as DH (2,0) but we now have the non-periodic boundary
±
conditions σL+1 = q ∓L σ1± , σL+1
z
= σ1z . As a consequence, spec(Hper ) has no gap
even for q = 1. While for a finite number r of particles and long chains L → ∞,
it is easy to see that Re Eper ∼ L−2 [45], for finite densities n = r/L, an elaborate
Bethe ansatz calculation shows that Re Eper ∼ L−3/2 [42]. Therefore, and quite
in distinction from equilibrium systems, a change in the boundary conditions may
well induce a phase transition in the long-time behaviour (observed first in driven
diffusive systems [46]). What happens is easily understood in this particular
example. For an open chain, the particles get stuck at one end of the chain and
a non-trivial position-dependent steady-state density profile ρs (i) builds up. The
time needed for this should be of the order of the time the particles need to move
from one end to the other which is finite for q = 1. However, for a periodic chain,
the particles keep chasing each other forever and a steady-state particle current
will be observed. By going to a reference frame co-moving with the mean velocity
of that current, one is back to the case q = 1 of unbiased diffusion.
The energy gaps  can now be found by concentrating on the spectrum-
generating part HXXZ . For h + > 1, the gaps are finite and are easily found. We
concentrate here on the case of unbiased diffusion, when h + = 1 and we are

Copyright © 2003 IOP Publishing Ltd.

 at the Pokrovsky–Talapov transition. The low-lying energy gaps are given by the
following proposition.

Proposition 10.2 On the Pokrovsky–Talapov line h + = 1, δ = 0, and for

L large, the low-lying eigenvalues of H (equation (10.41)) are for periodic
boundary conditions
 2
2π 8π 3 D  r
(Ij − I )2 + O(L−4 )
(per)
Er =D (I12 + · · · + Ir2 ) −
L L3 1 − j,=1
(10.49)
where the Ij are pairwise distinct integers (half-integers) when r is odd (even).
For an open chain
 π 2  
2(r − 1)
Er(free) = D (I12 + · · · + Ir2 ) 1 − + O(L−4 ) (10.50)
L L 1−
where the Ij are pairwise distinct non-negative integers. The integer r =
0, 1, 2, . . . gives the number of particles in the sectors of HXXZ .
The finite-size amplitudes limL→∞ L2 Er are independent of , for either
periodic or open chains. This proves an old conjecture [31] based on numerical
calculations. For = 0, the well-known free-fermion solution is reproduced. To
leading order in L−1 , eigenvalues with the same I12 + · · · + Ir2 are degenerate. We
observe that for periodic boundary conditions, this degeneracy is already broken
by the first correction in 1/L, while for free boundary conditions, the leading
correction keeps that symmetry.
Equations (10.49) and (10.50) are easily found from the Bethe ansatz. We
first consider periodic boundary conditions. The XXZ chain may be broken into
sectors containing only the states with r particles. Performing the Bethe ansatz as
usual [47], one has for the energies
Er = 2D(r − cos k1 − · · · − cos kr ) (10.51)
where the quasi-momenta k1 , . . . , kr are solutions of the Bethe ansatz equations

r
Lkj = 2πIj − (kj , k ) j = 1, . . . , r (10.52)
=1

where the Ij are pairwise distinct integers (half-integers) when r is odd (even)
and
sin((k − k
)/2)
(k, k
) = 2 arctan . (10.53)
cos((k + k
)/2) − cos((k − k
)/2)
We are interested in the leading finite-size corrections when L → ∞ with r fixed.
The ansatz
2π aj
kj = Ij + 2 + · · · (10.54)
L L

Copyright © 2003 IOP Publishing Ltd.

 gives
2π
(k1 , k2 )  (I1 − I2 ) + O(L−2 )
1− L

and
2π  r
aj = (I − Ij ).
1 − =1

Then (10.49) follows. Second, for free boundary conditions, the Bethe ansatz [48]
reproduces equation (10.51) for the energies, while the Bethe ansatz equations for
the quasi-momenta now take the form
1
Lkj = πIj − [(kj , k ) − (−kj , k )] (10.55)
2 =j

for j = 1, . . . , r and where the Ij are pairwise distinct non-negative integers.

Equation (10.54) leads to

aj = − (r − 1)πIj2
1−
and we arrive at (10.50). 

10.5 The seven-vertex model

Having seen for some single-species reaction–diffusion processes how the
relationship with the Bethe ansatz solution XXZ chain could be used to infer
certain physical properties, we present in this section an example of the
Baxterization procedure [25]. That procedure permits one to associate to a
stochastic quantum Hamiltonian H related to a Hecke algebra the Boltzmann
weights of a corresponding two-dimensional vertex model and, thus, prove its
integrability. In principle, the model is then solved through the Bethe ansatz.
For the six- and eight-vertex models, the completeness of the Bethe ansatz has
recently been proven [49].
Following [31], we consider the pair-annihilation model equation (10.41)
already defined in section 10.4 with γ = δ = 0 and, furthermore, we take
2α = DR + DL , thus = 0. We call  = 2α/D. After having performed the
canonical transformation Eiab → (−1)a−b Eiab , only at even sites i, the quantum
Hamiltonian takes, in the basis given by equation (10.28), the simple form


L−1
H = −D ei (10.56)
i=1

Copyright © 2003 IOP Publishing Ltd.

 where
 
0 0 0 
0 q −1 q 0 
ei = 11 ⊗ · · · ⊗ 1i−1 ⊗ 
0 q −1
 ⊗ 1i+2 ⊗ · · · (10.57)
q 0 
0 0 0 q + q −1

and 1i are 2 × 2 unit matrices attached to the site i.

While we could certainly solve this particular model through a Jordan–
Wigner transformation followed by a canonical transformation, see, e.g., [40],
we are interested in generic approaches of a more general validity than in those
cases reducible to a free-fermion description.
We have already seen in section 10.4 that the ei satisfy the Hecke algebra
(10.23). We now construct a two-dimensional vertex model corresponding to H
having a row-to-row transfer matrix T (θ ) depending on the spectral parameter
θ . This transfer matrices will satisfy the Yang–Baxter equations [50] which
imply the commutation relations [T (θ ), T (θ
)] = 0 if θ = θ
. The construction
is based on the matrix Ři (θ ), i = 1, 2, . . . , L − 1, which depends on the spectral
parameter θ . The Baxterization procedure for Hecke algebras [25] gives

sinh θ sinh(η − θ )
Ři (θ ) = ei + q = eη (10.58)
sinh η sinh η

which, for our model (10.56), (10.57), leads to

Ři (θ ) = 11 ⊗ · · · ⊗ 1i−1 ⊗ Ri,i+1 ⊗ 1i+2 ⊗ · · · (10.59)

with the non-vanishing elements of the matrix Ri,i+1

 
sinh(η − θ )  sinh θ
1   e−θ sinh η eη sinh θ 
.
Ri,i+1 :=
sinh η  e−η sinh θ eθ sinh η 
sinh(η + θ )
(10.60)
The relations (10.23) imply that these matrices satisfy the spectral parameter-
dependent braid-group relations

Ři (θ )Ři±1 (θ + θ
)Ři (θ
) = Ři±1 (θ
)Ři (θ + θ
)Ři±1 (θ )
[Ři (θ ), Řj (θ
)] = 0 |i − j | ≥ 2 (10.61)

which are equivalent to the Yang–Baxter equations.

In a 2D vertex model with vertex configurations labelled by (k, , m, n), the
kn are obtained from
Boltzmann weights S,m

k,n
Ři (θ ) = S,m 11 ⊗ · · · ⊗ 1i−1 ⊗ E mk ⊗ E n ⊗ 1i+2 ⊗ · · · . (10.62)

Copyright © 2003 IOP Publishing Ltd.

 ◦ •
@
5. @ -? - e−η sinh θ
@@
• ◦ ?

• ◦ 6
6. @@   eη sinh θ
@ 6
@
◦ •

◦ ◦
@ -?

7. @  sinh θ
@@ 6
• •

Figure 10.2. Diffusion and pair-annihilation of particles in the seven-vertex model and
their Boltzmann weights, for the vertices 5 to 7.

This implies that the vertex model associated with equation (10.56) is a seven-
vertex model. In a vertex model, arrows are attached to the bonds of a square
lattice [50]. In the stochastic model, we associate a particle (•) with an arrow
pointing up/right and no particle (◦) with an arrow pointing down/left. In
figure 10.2 we list together the chemical reactions, the vertex configurations and
their Boltzmann weights. The vertices usually labelled 1 to 4 correspond to no
reaction and are not shown (see [31]). Vertices 5 and 6 correspond to diffusion to
the right and to the left and vertex 7 to pair annihilation. In the leftmost column
of figure 10.2, the state of the particles before the reaction is given as the lower
pair of symbols while the state after the reaction is given by the upper pair of
symbols. The middle column gives the corresponding vertex configuration and
the right column the Boltzmann weight. The Hamiltonian equation (10.56) may
be recovered from H = − dθ d
ln T (θ )|θ=0 .

10.6 Further applications

We have studied in some detail the pair-annihilation process and its integrability.
Still, extracting explicitly information about the long-time behaviour (or the

Copyright © 2003 IOP Publishing Ltd.

 steady-state in more complicated models) is not yet trivial. In this section, we
briefly review some approaches which may be useful.

10.6.1 Spectral integrability

We have already seen that in certain cases, the quantum Hamiltonian H = H0 +
H1 such that spec(H ) = spec(H0 ), independently of the precise form of H1 . It
may happen that although H0 is integrable, H is not. Such a model is said to be
spectrally integrable. If only binary interactions are present, it is convenient to
express H in terms of a two-site matrix Hi,i+1 acting on the sites i and i + 1


L
H= 11 ⊗ · · · ⊗ 1j −1 ⊗ Hj,j +1 ⊗ 1j +2 · · · ⊗ 1L . (10.63)
j =1

For the model (10.41) with left–right symmetry, that is q = 1, we have

 
0 −δ −δ −2α
0 1 + δ −1 −γ 
Hj,j +1 = D 0 −1 1 + δ
. (10.64)
−γ 
0 0 0 2(α + γ )

One can always rescale time such that D = 1. Then the parameters of the XXZ
chain H0 = HXXZ become = 1 + δ − γ − α and h = α + γ .
In equations (10.63) and (10.64), H is only integrable for δ = 0.
However, the special case α = 0, γ = δ simply corresponds to the radioactive
decay of diffusively moving particles. While for δ = 0, one is back to the
critical Pokrovsky–Talapov line h + = 1, the associated quantum spin chain
 ground state for δ = 0. 
has a frozen The one- and two-particle correlators
C1 (t) = L j =1 C1 (j ; t) and C2;n (t) =
L
j =1 C(j, j + n; t) with n fixed, see
equation (10.21), only imply the sectors with r = 1 and r = 2 particles of HXXZ,
respectively. Their long-time behaviour is easily worked out from the results of
section 10.4 and is collected in table 10.3 [40]. Of course, we implicitly assume
that the corresponding amplitudes do not accidentally vanish. At first sight, one
might have expected a simple exponential factor e−2kδt for the k-point correlator
Ck and we already observe that eventual algebraic prefactors are not readily
predicted from the spectrum of H alone. The more complicated form of the
relaxation time for δ > α + γ comes from a bound state in the two-particle sector
of HXXZ, with energy 4δ + 4 − 2 − 2/ , see [50, 51]. More general initial
conditions are discussed in [40].

10.6.2 Similarity transformations

In trying to extract explicit information on certain reaction–diffusion systems,
the integrability of the quantum Hamiltonian H plays a central role. Since it is
difficult to realize the constraints (10.18) of stochasticity and integrability at the

Copyright © 2003 IOP Publishing Ltd.

 Table 10.3. Generic long-time behaviour of the one- and two-point correlators C1 (t) and
C2;n (t) for n finite, in the system equation (10.64) and with a translation-invariant initial
state and a finite initial particle density.

δ C1 (t) C2;n (t)

0 t −1/2 t −3/2
<α+γ exp(−2δt) t −1/2 exp(−4δt)
>α+γ exp(−2δt) exp [−4δt + 2( + 1/ − 2)t]

same time, it is of interest to see whether there exist systematic transformations

of an integrable quantum Hamiltonian H towards a new stochastic Hamiltonian
+.
H
Specifically, we shall consider the transformation
,
L
+ = BH B −1
H B= Bj (10.65)
j =1

where Bj = 1 ⊗ · · · ⊗ 1 ⊗ B ⊗ 1 ⊗ · · · ⊗ 1 is the transformation matrix B

+ are said to be similar
acting on the site j . Then the systems described by H and H
to each other. An interesting alternative, which has not yet been systematically
studied, is to consider an enantiodromy transformation [10]
+ = BH TB −1
H (10.66)

where H T is the transpose of H .

From now on, we shall focus on translationally invariant systems and
consider periodic boundary conditions. The effect of the transformation B on H is
completely given by its effect on the two-particle Hamiltonian Hj,j +1 in (10.64).
A stochastic similarity transformation arises if both H and H + represent stochastic
systems. For a simple example, consider the symmetric annihilation-coagulation
process (10.63), (10.64) with δ = 0. If C(t|α, γ ) = C1 (t) is the spatially averaged
particle density with the rates α and γ , respectively, a stochastic similarity
transformation shows that [52–54]
α+γ
C1 (t|α, γ ) = C1 (t|0, α + γ ). (10.67)
2α + γ
Similar results hold for any k-point correlator Ck (t). So far, explicit methods
to find the time-dependent correlators are only available for either the pure
coagulation model α = 0 through empty-interval methods (see later) or the pure
annihilation model γ = 0 through free-fermion techniques, see [10, 40, 55] and
references therein. Equation (10.67) allows to reduce any symmetric annihilation-
coagulation process to pure coagulation, for any initial density C1 (0|0, γ ).

Copyright © 2003 IOP Publishing Ltd.

 Table 10.4. Single-species processes with space-independent reaction rates and which are
similar via (10.65) to a free-fermion model. The reaction rates are defined in table 10.2.

Model Reactions Conditions

A •• ↔ ◦◦ •◦ ↔ ◦• 2(α + σ ) = DL + DR
B •• → ◦◦ •• → ◦•, •◦ •◦ ↔ ◦• 2α + γL + γR = DL + DR
C ◦◦ → •• ◦◦ → ◦•, •◦ •◦ ↔ ◦• 2σ + νL + νR = DL + DR
D •• → •◦, ◦• •◦, ◦• → •• •◦ ↔ ◦• γ L = DL γ R = DR
E ◦◦ → •◦, ◦• •◦, ◦• → ◦◦ •◦ ↔ ◦• νL = DL νR = DR
F •◦, ◦• → ◦◦ •◦, ◦• → •• δ
 R L/δ = βR /βL
δR = βR γR = νL
G •◦, ◦• ↔ ••, ◦◦
δL = βL γL = νR

δL = δR = γL = γR
H •→◦ ◦→•
βL = βR = νL = νR

This also explains the experimental results in table 10.1. The known stochastic
similarity transformations of the form (10.65) leave the parameters and h of the
XXZ chain invariant, but the results of proposition 10.2 suggest that a stochastic
similarity transformation between systems with different values of might exist.
See [53] for the extensions to δ > 0 and q = 1.
Equation (10.67) allows long-time behaviour C(t|α, γ ) ∼ t −1/2 to be
recovered from a simple heuristic argument. For pure coagulation, one particle
always remains, thus C(∞|0, γ ) = 1/L in the steady state. Therefore, the steady-
state density C(∞|α, γ ) ∼ L−1 . However, from the spectrum of H , the leading
relaxation time τ =  −1 ∼ L2 ∼ ξ 2 , where ξ is identified as the characteristic
spatial length scale. Therefore, C(∞|α, γ ) ∼ ξ −1 ∼ τ −1/2 . The asserted time-
dependent behaviour, therefore, might have been anticipated on dimensional
grounds.

10.6.3 Free fermions

For the pure annihilation model with 2α = DR + DL , one has = 0. In this case,
H may be diagonalized through a Jordan–Wigner transformation followed by a
canonical transformation [55]. In order for this to work, H may only contain pairs
of particle creation and annihilation operators. For space-independent reaction
rates, the complete list of reaction–diffusion process whose quantum Hamiltonian
H+ is similar through (10.65) to a free-fermion Hamiltonian H is as follows
[10, 40, 53] and shown in table 10.4. Since the transformation (10.65) is spatially
local, these correspondences actually hold in any dimension but free-fermion
methods are only available in one dimension.

Copyright © 2003 IOP Publishing Ltd.

 Of these models, only models A (diffusive pair-annihilation and creation,
solved exactly in one dimension in [55]), G (kinetic Ising model with Glauber
dynamics) and H (free decay and creation of particles) are reversible and have
an equilibrium steady state. Their quantum Hamiltonian is, therefore, similar
to a symmetric matrix. The similarity of the kinetic Ising model with Glauber
dynamics (model G) to a free-fermion model was obtained long ago through a
duality transformation [56] and, more recently, as a true similarity transformation
[53, 57]. This suggests studying a more general type of relationship, based on
domain-wall dualities, see [10, 57] for details. Models C and E are obtained by
a particle–hole permutation • ↔ ◦ from models B and D, respectively. Model
B is the biased annihilation-coagulation process, while model D is the diffusive
coagulation-process with arbitrary decoagulation. Finally, model F is the doubly
biased voter model (in space and in the preference between • and ◦) and in [58]
some correlators are found from the free-fermion form. The physical behaviour
of all these models can be treated in a single calculation. For example, the mean
particle density depends on a single parameter h such that [40]

t −1/2 if h = 1
C1 (t) − C1 (∞) ∼ −3/2 (10.68)
t exp(−t/τ ) if 0 < h < 1

where C1 (∞) is the steady-state density and τ = 1/(4 − 4h) is the relaxation time
(see [10] and references therein for more information on solved free-fermion
models).
This kind of analysis was generalized to find those reaction–diffusion
systems which are similar, via a transformation of the type (10.65), to the XXZ
chain [40]. While the full result is too complex to be re-stated here, an interesting
special case is given by the conditions

γR + βL + 2α + DL = νL + δL + 2σ + DR
(10.69)
γL + βR + 2α + DR = νR + δR + 2σ + DL .
In this case, the (usually infinite) hierarchy of equations of motion for the k-point
particle-density correlators Ck (t) = "n1 (t) . . . nk (t)# closes naturally, such that
C˙k (t) only depends on the C (t) with  ≤ k. In principle, the equations of motions
for the Ck can then be solved iteratively [35].

10.6.4 Partial integrability

The previous sections have shown that constructing integrable stochastic systems
which go beyond mere free diffusion is a non-trivial exercise. One might wonder
whether the condition of full integrability is not too strong. After all, from a
practical point of view it would be enough to identify a set {Q1 , . . . , QM }
of observables such that these satisfy a closed set of equations, say Q̇i =
fi (Q1 , . . . , QM ), with i = 1, . . . , M. Such a partial integrability may be
enough for many practical needs. Indeed, such an approach is available through

Copyright © 2003 IOP Publishing Ltd.

 1.00

O=1

C1(t) O=10

O=0

0.10

0.01
1 10 100 1000
t

Figure 10.3. Evolution of the mean particle density C1 (t) in the symmetric coagulation
model with the production reaction • ◦ • → • • • for several production rates λ. For long
√
times, the asymptotic behaviour is C1 (t) ∼ 1/ t for all values of λ (after [59]).

the empty-interval method [7, 12]. Consider a periodic chain with L sites and
lattice spacing a. Let In (t) be the probability that at time t, n consecutive sites are
empty. Then the mean particle density is [12, 52]
C1 (t) = (1 − I1 (t))/a. (10.70)
In order to illustrate the method, we consider the left–right symmetric pure
coagulation model and also take the free-fermion condition γ = D of model D
in table 10.4, but we now add a three-site production reaction • ◦ • → • • • with
rate 2Dλ [59]. The equations of motion for the In (t) read as
I˙1 (t) = 2D(I0 (t) − 2I1 (t) + I2 (t)) − 2Dλ(I1 (t) − 2I2 (t) + I3 (t))
(10.71)
I˙n (t) = 2D(In−1 (t) − 2In (t) + In+1 (t)) 2≤n≤L−1
together with the boundary conditions I0 (t) = 1 and IL (t) = 0 (assuming that
there is at least one particle in the system). The solution of these equations
is straightforward. For example, one obtains for the leading relaxation time
τ −1 =  = 2Dπ 2 L−2 + O(L−4 ), in agreement with the results of section 10.4.
The effect of the production term is only transient,√ as illustrated in figure 10.3
for the mean density C1 (t). For λ = 0, C1 (t) ∼ 1/ t is of course expected from
equations (10.6) and (10.67). While the free-fermion condition γ = D is essential
for the method to work, we also see that the presence of the production term poses
no problem at all for the closure of the equations of motion (10.71)4.
4 This term cannot, e.g. by a similarity transformation, be turned into a term treatable by either free-
fermion or full integrability methods.

Copyright © 2003 IOP Publishing Ltd.

 Accepting the free-fermion condition γL,R = DL,R , one can extend the
treatment to the more general model D of table 10.4 and may even extend this
further to include the processes ◦◦ → •• and ◦◦ → ◦•, •◦ with rates 2σ, νR , νL ,
respectively [7, 12]. Let us call this system model D
which depends on the
seven parameters DL,R , βL,R , νL,R , σ . In a remarkable paper [41], the idea of
the empty-interval method was translated into the Hamiltonian formalism and
several new sets of observables were defined which generalize the variables In (t)
and lead again to closed equations of motion. It turns out that the spectrum of
relaxation times of model D
is given by the Hamiltonian of the Wannier–Stark
ladder [41]
L
+ σn σn+1 + (h + h
n)σnz ]
y y
H =− [σnx σn+1
x
(10.72)
n=−L

where h and h
are constants. In this case, the couplings in H are space-
dependent. The extension of the similarity/enantiodromy approach to this more
general setting remains to be done. Extensions of the empty-interval method to
interactions on more than two sites are studied in [60].
While the empty-interval method, as such, does not work for the pair-
annihilation process, the method has been generalized recently [61]. We briefly
explain the idea using the left–right symmetric pair-annihilation process with the
free-fermion condition α = D (model A or B) as example. Let Gn (t) be the
probability that, at time t, one has on n consecutive sites an even number of
particles. The mean particle density is C1 (t) = (1 − G1 (t))/a. Furthermore, let
Fn (t) (Hn (t)) be the probability that a segment of n consecutive sites with an
even (odd) number of particles is followed by the presence of a particle at the
(n + 1)th site. From the relations

2Fn (t) = (1 − G1 ) + (Gn − Gn+1 ) 2Hn (t) = (1 − G1 ) − (Gn − Gn+1 )

(10.73)
and the boundary condition G0 (t) = 1, the equations of motion

Ġn (t) = 2D(Fn−1 − Hn−1 + Hn − Fn ) = 2D(Gn−1 (t) − 2Gn (t) + Gn+1 (t))
(10.74)
follow. They can be solved by standard methods. Reaction terms parametrized
by σ, νL,R , βL,R (see table 10.2) and even the reaction ◦ • ◦ → • • • can be
added [61]. Correlators are studied in [62].
In view of the practical success of these techniques it is perhaps not
completely futile to ask whether there might a be a systematic way to identify
these ‘empty-interval’ or related variables?

10.6.5 Multi-species models

We consider chains with N states per site. One of them is taken to be the empty
site (◦) and the other states are labelled An , n = 1, . . . , N − 1. Finding integrable
stochastic systems becomes more difficult when N increases. Several examples

Copyright © 2003 IOP Publishing Ltd.

 Table 10.5. Some integrable reaction–diffusion processes of N − 1 species and their Hecke
algebra quotient [26], see text for the definition of the rates.

Model Reactions Quotient

A An ◦ ↔ ◦An (2, 0)
B An ◦ ↔ ◦An An Am ↔ Am An (N, 0)
C An ◦ ↔ ◦An An Am ↔ Am An A1 A1 → ◦A1 (N − 1, 1)
D An ◦ ↔ ◦An An Am ↔ Am An Ar Ar → Ar±1 Ar (N − 2, 2)
E An ◦ ↔ ◦An Ar As → ◦Ar+s (1, 1)
F An ◦ ↔ ◦An Ar As → ◦Ar+s An An → An±1 An±1 (2, 1)

were found [26] through the quotients (P , M)HL−1 (q) as realized through the
Perk–Schultz model. They are collected in table 10.5. The following conventions
apply.

(1) For the first reaction in all models and the second reaction in models B, C, D
(with n < m understood), the reaction to the right (left) occurs with rate
R (L ).
(2) For models E, F the sum r + s has to be taken modN. If in this case
r + s = 0 mod N, the rate is L + R . If r + s = 0 mod N as well as for
the third reaction in models C, D the rate is R . In model D, it is also
assumed that, in the third reaction, pairs (r, s = r ± 1) never have an element
in common. If the products on the right are interchanged (e.g. A1 A1 → A1 ◦
in model C), the rate is L .
(3) In the third reaction in model F, the rates are ± respectively such that
+ + − = L + R .
√ √
One defines D = L R = 1 and q = R / L . The Hecke algebra
quotient (P , M)HL−1 (q) according to the realization as a Perk–Schultz quantum
chain equation (10.27) [24, 26] is also indicated.
From table 10.5 and theorem 10.3, we see that the simple diffusion model A
has, up to degeneracies, the same spectrum as the XXZ chain used in section 10.4
to describe biased diffusion of a single species of particles •. In the same way, the
spectrum of model E is, up to degeneracies, the same as the one found for pair
annihilation in section 10.4, with 2α = DL + DR .
For illustration, we briefly consider model E for N = 3. Each site may
contain either a particle of type A or B or be empty (◦). Single particles may
diffuse to the right A◦ → ◦A, B◦ → ◦B with a rate R or similarly to the
left with rate L . On encounter, between like particles the reactions AA → ◦B
and AA → B◦ occur with rates R and L , respectively, and similarly BB →
◦A, A◦. Two unlike particles react AB → ◦◦ with rate L + R . In the left–right
symmetric case, the identity of the spectra of H(E) and the one of pair annihilation,

Copyright © 2003 IOP Publishing Ltd.

 up to degeneracies, was checked directly [31]. Furthermore, in the spirit of the
empty-interval method, a closed system of equations of motion was found, whose
solution leads to the mean particle densities n̄A (t) ∼ n̄B (t) ∼ t −1/2 [63].
In [64], Bethe ansatz solutions of the master equation for N-species models
with particle-numbers conservation are studied. In particular, model B with
N = 3, 5 was rediscovered. The models in [64] are found from solutions of
quantum Yang–Baxter equations. Further study might reveal a relationship to
diffusion algebras [39], see later.
For periodic boundary conditions and N > 2, the diffusion bias leads after a
similarity transformation to a generalized Dzialoshinsky–Moriya interaction [65].
A sufficient criterion for integrability was derived in an attempt to look more
systematically for integrable many-species models [66].
Finally, a different generalization from section 10.4 is to consider integrable
stochastic models on ladders [71], rather than chains.

10.6.6 Diffusion algebras

For certain integrable systems, there exist algebraic methods which allow one to
find the steady state |s# such as the celebrated matrix product states [2, 67]. Time-
dependent problems are treated in [68].
Behind this seemingly technical and ad hoc method, there is a new
and general mathematical structure. We shall explain here the main idea
using reaction–diffusion systems with N states per site labelled by An , n =
0, 1, . . . , N − 1 (where A0 = ◦) moving on a periodic chain with L sites
but generalizations to different boundary conditions are possible. The allowed
reactions are An Am → Am An with rate gnm (in particular, model B from
table 10.5 is a special case of this). The un-normalized steady-state probability
distribution is [2]

Ps (σ ) = P (σ1 , . . . , σL ) = Tr(Dσ1 Dσ2 · · · DσL ) (10.75)

where the matrices Dσ satisfy the quadratic relations

gσρ Dσ Dρ − gρσ Dρ Dσ = xρ Dσ − xσ Dρ (10.76)

where σ < ρ and σ, ρ ∈ {1, . . . , N} and gσρ ∈ R\{0}, gρσ ∈ R and the xσ are
complex parameters. If, in addition, the set A of generators Dσ admits a linear
Poincaré–Birkhoff–Witt (PBW) basis of ordered monomials Dσk11 Dσk22 · · · Dσknn
with σ1 > σ2 > · · · > σn and kn ∈ N, A is called a diffusion algebra [39].
These conditions imply certain constraints on the gσρ and the xσ , quite
analogously to the Jacobi identities of a Lie algebra. These constraints can be
fully solved and a classification of all diffusion algebras for N species is obtained
[39, 69]. The representation theory of N-species diffusion algebras is just getting
started, see [70].

Copyright © 2003 IOP Publishing Ltd.

 10.7 Outlook: local scale-invariance
We finish with a discussion on how the scale invariance of many reaction–
diffusion systems might be turned into a dynamical symmetry. For example, the
symmetric pair-annihilation process is on the Pokrovsky–Talapov critical line.
One has the covariance

"n(r1 , t1 ) · · · n(rp , tp )# = b−(x1 +···+xp ) "n(r1

, t1
) · · · n(rp
, tp
)# (10.77)

of the p-point correlators under the dilatation r → r

= br, t → t
= bz t of the
space and time coordinates r, t respectively, where z is the dynamical exponent
and x1 , . . . , xp are scaling dimensions. In the cases at hand, z = 2.
This is reminiscent of the situation at equilibrium critical points. In those
systems, it is known that under fairly general conditions, the covariance of the p-
point correlators under global-scale transformations r → br can be extended to
conformal transformations. In addition, in two dimensions the energy–momentum
tensor of local conformally invariant field theories becomes an analytic function
T = T (z) of the complex coordinate z such that not only T (z) itself but also
all powers T n (z), n = 1, 2, 3 . . . , are conserved [72, 73]. This signals the
integrability of 2D conformally invariant field theories. Is it possible to generalize
the spacetime dilatations encountered for critical reaction–diffusion systems in a
similar way?
This question has been recently addressed in the context of kinetic spin-
systems [74, 75]. We have already seen above that the kinetic Ising model with
Glauber dynamics [20] may be obtained through a similarity transformation of
the quantum Hamiltonian from a certain single-species reaction–diffusion system,
see model G from table 10.4. We now concentrate on this system. In the Glauber–
Ising model, the transition rates in the master equation are chosen such that
the steady state |s# is given by the equilibrium probability distribution
 Ps (σ ) ∼
e−H[σ ]/T with the classical Ising model Hamiltonian H = − (i,j ) σi σj where
T is the temperature. Glauber dynamics may be realized through the discrete-time
heat-bath rule σi (t) → σi (t + 1) such that

σi (t + 1) = ±1 with probability 12 [1 ± tanh(hi (t)/T )] (10.78)


with the local field hi (t) = h + j (i) σj (t). With the choice (10.78), the master
equation can be solved exactly in one dimension [20]. The time-dependent spin–
spin correlators and their approach towards equilibrium are, thus, determined.
In contrast to equilibrium statistical mechanics, where fine-tuning the model
parameters is needed to reach a critical point, dynamical scaling is often found
to occur in large regions of the model’s parameter space. For example, prepare
the system initially in a disordered state and then quench the temperature to a
final temperature T < Tc below the critical temperature Tc > 0.5 Although the
5 In the 1D Glauber–Ising model, T = 0 leads to certain modifications of the ageing as described
c
from the point of view of local scale-invariance [76].

Copyright © 2003 IOP Publishing Ltd.

 Figure 10.4. Snapshot of the coarsening of ordered domains in the 2D Glauber–Ising
model, after a quench to T = 1.5 < Tc from a totally disordered state and at times t = 25
(left) and t = 275 (right) after the quench.

steady state of the model is not critical, the relaxation towards it occurs through
domain coarsening and is very slow, the typical length scale varying with time as
L(t) ∼ t 1/z , see figure 10.4. Typically, the observables depend algebraically on
the time t passed since the quench, see [8, 9] for (a collection of) recent reviews.
Here we concentrate on the two-time spatio-temporal response function R(t, s; r)
of the time-dependent spin σr (t) at site r with respect to an external magnetic
field h0 (s) applied at the origin 0 at an earlier time s < t. Generically, two-time
quantities such as R(t, s; r) depend on both times t and s and not merely on the
difference τ = t − s. This breaking of time-translation invariance is called ageing.
For ageing systems, an extension of dynamical scaling is possible and allows
one to fix the form of the two-time response function. Specifically, it can be shown
that for a dynamical exponent z = 2 [74, 75, 77]
  1+a−λR /2  
δ"σr (t)#  t −1−a M r2
R(t, s; r) = = r0 (t − s) exp − .
δh0 (s) h=0 s 2 t −s
(10.79)
Here a and λR are non-equilibrium exponents to be determined which
characterize the ageing universality class [78, 83]. Finally, r0 and M are non-
universal constants. We first present evidence that the response function of the
Glauber–Ising model in 2D and 3D is, indeed, given by (10.79). Then we discuss
where this presumably exact result comes from.
We first consider the autoresponse R(t, s) = R(t, s; 0). While R itself is
too noisy to be measured directly, integrated response functions are accessible
through simulations, see [79] and references therein for the details which we skip
over here. For the example, the integrated autoresponse
s
ρ(t, s) = du R(t, u) ∼ s −a fM (t/s) (10.80)
0

Copyright © 2003 IOP Publishing Ltd.

 1 1.5 0.30
s=81
s=100 s=25 s=100
s=200 s=36 s=121
s=400 s=49 s=144
0.5
s=800 s=64 s=169
0 s=1600 s=81 0.26
s=196
s=100 s=225

ρ (x,µ)
ln(fM(x))

−0.5

(2)
−1 0.22
−1.5

(a) (b) (c)

−2 −2.5 0.18
0 1 2 0 1 2 0 4 8 12
ln(x) ln(x) x

Figure 10.5. Scaling form of the integrated magnetic response in the Glauber–Ising model
as a function of x = t/s below criticality. The symbols correspond to different waiting
times s. The integrated autoresponse is shown in (a) two dimensions at T = 1.5 and in
(b) three dimensions at T = 3. An example of the integrated spatio-temporal response in
two dimensions at T = 1.5 and with µ = 2 is shown in (c). The full curves are obtained
from (10.79). After [79].

is relatively easy to measure, whereas the scaling function fM (x) can be

calculated explicitly from (10.79). In the Glauber–Ising model, the exponent
a = 1/z = 1/2 (see [80] for a detailed discussion) and λR  1.26 and 1.6 in two
and three dimensions, respectively. In figure 10.5(a) and (b), the scaling function
fM (x), as obtained from large-scale simulations, is shown for several values of the
waiting time s. In both two and three dimensions, a nice scaling behaviour is found
and the form of the scaling function agrees very well with the prediction from
equation (10.79). Next, the r-dependence of R(t, s; r) is tested by measuring the
spatio-temporally integrated response
s √
µs
du dr r d−1 R(t, u; r) ∼ s d/2−a ρ (2) (t/s, µ)
0 0

where µ is a control parameter. We stress that the scaling function ρ (2) no longer
contains the free non-universal parameter [79]. As an example, we compare in
figure 10.5(c) data from two dimensions taken with µ = 2 with equation (10.79).
Besides the expected scaling, the functional form of the scaling function neatly
follows the prediction. We stress that the position, the height and the width of
the maximum of ρ (2) in figure 10.5(c) are completely fixed. Similar results have
been obtained for other values of µ and in three dimensions as well. This provides
strong evidence that equation (10.79) is exact, at least in this model [79]. Tests of
(10.79) in different universality classes are described in [75].

Copyright © 2003 IOP Publishing Ltd.

 e2 e2 e2

e1 e1 e1

(a) (b) (c)

Figure 10.6. Root space of the complexified conformal Lie algebra conf3 , indicated by the
full and the open points. The double circle in the centre denotes the Cartan subalgebra. The
generators belonging to the three non-isomorphic parabolic subalgebras [82] are indicated
- 1 , (b) a.
by the full points, namely (a) sch -1 .
ge1 and (c) alt

In order to derive (10.79), consider the diffusion equation

(2M∂t − ∂r · ∂r )φ(t, r) = 0. (10.81)

For fixed M, the Schrödinger group is the maximal invariance group on the space
of solutions of equation (10.81). It is defined by the spacetime transformations
(R is a rotation matrix)
αt + β Rr + vt + a
t −→ t
= r −→ r
= αδ − βγ = 1 (10.82)
γt + δ γt + δ

and acts projectively on the solutions φ(t, r) [81]. Let schd be the Lie algebra of
(10.82). Time translations occur in schd and are parametrized by β. If we treat
the ‘mass’ M not as a constant but as another variable, the embedding schd ⊂
confd+2 for the complexified Lie algebras follows [82], where confd+2 is the Lie
algebra of the conformal group in d + 2 dimensions. From the classification of the
parabolic subalgebras of confd+2 we obtain several new subalgebras, called age -
- [82]. For the 1D case, we illustrate in figure 10.6 their definition through
or alt
the root space of conf3 ∼= B2 . These subalgebras still contain the generator for the
dilatations t → b2 t, r → br (which is in the Cartan subalgebra of conf3 ) but no
longer contain time translations (which is in the lower left corner of figure 10.6).
They are candidates for a dynamic symmetry algebra of ageing systems. If we
assume that the two-time response function transforms covariantly under the
- a set of linear differential equations for R(t, s; r) is
- or alt,
action of either age
obtained. Matching their solution with the expected [78] scaling behaviour of R,
we recover equation (10.79) in the special case z = 2.
The functional form of R depends on the fact that the Galilei transformation
of (10.82) is identical to the well-known one of a free particle. It is not trivial at all
that the response function of an interacting field theory such as the Glauber–Ising

Copyright © 2003 IOP Publishing Ltd.

 model in d > 1 dimensions should be recovered from a dynamical symmetry of
the equation of motion of a free-field theory.
There exist infinite-dimensional Lie algebras which contain schd as
subalgebras. For example, the Schrödinger group (10.82) is a subgroup of the
group defined by the transformations t → t
and r → r
where
*
t
= β(t), r
= r β̇(t) or else t
= t, r
= r − α(t) (10.83)

and β and α are arbitrary functions. Whether this has a bearing on the
ageing behaviour of non-equilibrium spin systems is still open. Local scale-
transformations generalizing the Schrödinger group (10.82) to general values
of the dynamical exponent z = 2 exist [75]. It can be shown that R(t, s; r) =
R(t, s; 0)!(r(t − s)−1/z ), such that equation (10.79) holds for the autoresponse
R(t, s; 0) if λR /2 is replaced by λR /z and !(v) is given as the solution of a linear
differential equation of fractional order [75].

Acknowledgments
I thank M Pleimling and J Unterberger for their pleasant collaboration on local
scale-invariance and the members of the Groups de Travail Mathématiques/
Physique for keeping my interest in integrable systems alive. This work
was supported by CINES Montpellier (projet pmn2095) and the Bayerisch-
Französisches Hochschulzentrum (BFHZ).

References
[1] Derrida B, Lebowitz J L and Speer E R 2002 Phys. Rev. Lett. 89 030601
[2] Derrida B, Janowsky S A, Lebowitz J L and Speer E R 1993 J. Stat. Phys. 73 813.
Derrida B 1998 Phys. Rep. 301 65
[3] Schmittmann B and Zia R K P 1995 Phase Transitions and Critical Phenomena vol
17, ed C Domb and J Lebowitz (London: Academic Press)
[4] Privman V (ed) 1996 Nonequilibrium Statistical Mechanics in One Dimension
(Cambridge: Cambridge University Press)
[5] Marro J and Dickman R 1999 Nonequilibrium Phase Transitions in Lattice Models
(Cambridge: Cambridge University Press)
[6] Hinrichsen H 2000 Adv. Phys. 49 815
[7] ben-Avraham D and Havlin S 2000 Diffusion and Reactions in Fractals and
Disordered Systems (Cambridge: Cambridge University Press)
[8] Cates M E and Evans M R (ed) 2000 Soft and Fragile Matter (Bristol: IOP Press)
[9] Cugliandolo L F 2002 Preprint cond-mat/0210312
[10] Schütz G M 2000 Phase Transitions and Critical Phenomena vol 19, ed C Domb
and J Lebowitz (London: Academic Press)
[11] Smoluchowski M v 1917 Z. Phys. Chem. 92 129
[12] ben-Avraham D, Burschka M A and Doering C R 1990 J. Stat. Phys. 60 695
[13] Spouge J L 1988 Phys. Rev. Lett. 60 871
Spouge J L 1988 Phys. Rev. Lett. 60 1885 (erratum)

Copyright © 2003 IOP Publishing Ltd.

 [14] Toussaint D and Wilczek F 1983 J. Chem. Phys. 78 2642
[15] Prasad J and Kopelman R 1989 Chem. Phys. Lett. 157 535
[16] Kopelman R, Li C S and Shi Z -Y 1990 J. Luminescence 45 40
[17] Kroon R, Fleurent H and Sprik R 1993 Phys. Rev. E 47 2462
[18] Cornell S and Droz M 1993 Phys. Rev. Lett. 70 3824
[19] See Henkel M and Hinrichsen H 2003 for a forthcoming review
[20] Glauber R J 1963 J. Math. Phys. 4 294
[21] Hyver C 1972 J. Theor. Biol. 36 133
Keizer J 1972 J. Stat. Phys. 6 67
Schnakenberg J 1976 Rev. Mod. Phys. 48 571
[22] Martin P P 1991 Potts Model and Related Problems in Statistical Mechanics
(Singapore: World Scientific)
[23] Martin P P and Rittenberg V 1992 Int. J. Mod. Phys. A 7 (Suppl. 1B) 707
Martin P P and Rittenberg V 1992 Int. J. Mod. Phys. B 4 792
[24] Schultz C L 1981 Phys. Rev. Lett. 46 629
Schultz C L 1983 Physica 122 71
Perk J H H and Schultz C L 1981 Phys. Lett. A 84 407
Perk J H H and Schultz C L 1981 Non-Linear Integrable Systems, Classical Theory
and Quantum Theory ed M Jimbo and T Miwa (Singapore: World Scientific)
[25] Jones V R 1990 Int. J. Mod. Phys. B 4 701
[26] Alcaraz F C and Rittenberg V 1993 Phys. Lett. B 314 377
[27] Kirrilov A N and Reshetikhin N Yu 1988 LOMI Preprint
[28] Pasquier V and Saleur H 1990 Nucl. Phys. B 330 523
[29] Saleur H 1989 Proc. Trieste Conf. on Recent Developments in Conformal Field
Theories
[30] Alexander S and Holstein T 1978 Phys. Rev. B 18 301
[31] Alcaraz F C, Droz M, Henkel M and Rittenberg V 1994 Ann. Phys. 230 250
[32] Kandel D, Domany E and Nienhuis B 1990 J. Phys. A: Math. Gen. A 23 L557
[33] Schütz G 1993 J. Stat. Phys. 71 471
[34] Alcaraz F C, Arnaudon D, Rittenberg V and Scheunert M 1994 Int. J. Mod. Phys.
A 9 3473
[35] Schütz G M 1995 J. Stat. Phys. 79 243
Aghamohammadi A and Khorrami M 2001 J. Phys. A: Math. Gen. 34 7431
[36] Albeverio S and Fei S-M 1998 Rev. Math. Phys. 10 723
[37] Alcaraz F C, Dasmahapatra S and Rittenberg V 1998 J. Phys. A: Math. Gen. 31 845
[38] Grosskinsky S, Schütz G M and Spohn H 2003 Preprint cond-mat/0302079
[39] Isaev A P, Pyatov P N and Rittenberg V 2001 J. Phys. A: Math. Gen. 34 5815
[40] Henkel M, Orlandini E and Santos J 1997 Ann. Phys. 259 163
[41] Peschel I, Rittenberg V and Schultze U 1994 Nucl. Phys. B 430 633
[42] Gwa L-H and Spohn H 1992 Phys. Rev. A 46 844
Kim D 1995 Phys. Rev. E 52 3512
[43] Sandow S and Schütz G 1994 Europhys. Lett. 26 7
[44] Johnson J D and McCoy B M 1972 Phys. Rev. A 6 1613
Takanishi M 1973 Prog. Theor. Phys. 50 1519
[45] Henkel M and Schütz G M 1994 Physica A 206 187
[46] Krug J 1991 Phys. Rev. Lett. 67 1882
[47] Alcaraz F C, Barber M N and Batchelor M T 1988 Ann. Phys. 182 280

Copyright © 2003 IOP Publishing Ltd.

 [48] Alcaraz F C, Barber M N, Batchelor M T, Baxter R J and Quispel G R W 1987
J. Phys. A: Math. Gen. 20 6397
[49] Baxter R J 2002 J. Stat. Phys. 108 1
[50] Baxter R J 1982 Exactly Solved Models in Statistical Mechanics (London:
Academic Press)
[51] Gaudin M 1983 La fonction d’onde de Bethe (Paris: Masson)
[52] Krebs K, Pfannmüller M P, Wehefritz B and Hinrichsen H 1995 J. Stat. Phys. 78
1429
[53] Henkel M, Orlandini E and Schütz G M 1995 J. Phys. A: Math. Gen. 28 6335
[54] Simon H 1995 J. Phys. A: Math. Gen. 28 6585
[55] Lushnikov A A 1986 Sov. Phys.–JETP 64 811
Lushnikov A A 1987 Phys. Lett. A 120 135
[56] Siggia E 1977 Phys. Rev. B 16 2319
[57] Santos J E 1997 J. Phys. A: Math. Gen. 30 3249
[58] Aghamohammadi A and Khorrami M 2000 J. Phys. A: Math. Gen. 33 7843
[59] Henkel M and Hinrichsen H 2001 J. Phys. A: Math. Gen. 34 1561
[60] Khorrami M, Aghamohammadi A and Alimohammadi M 2003 J. Phys. A: Math.
Gen. 36 345
[61] Masser T and ben-Avraham D 2001 Phys. Rev. E 63 066108
[62] Masser T and ben-Avraham D 2001 Phys. Rev. E 64 062101
[63] Privman V 1992 Phys. Rev. A 46 R6140
[64] Alimohammadi M and Ahmadi N 2002 J. Phys. A: Math. Gen. 35 1325
[65] Dahmen S R 1995 J. Phys. A: Math. Gen. 28 905
[66] Popkov V, Fouladvand M E and Schütz G M 2002 J. Phys. A: Math. Gen. 35 7187
[67] Derrida B, Evans M R, Hakim V and Pasquier V 1993 J. Phys. A: Math. Gen. 26
1493
[68] Popkov V and Schütz G M 2002 Mat. Fis. Anal. Geom. 9 401
Stinchcombe R B and Schütz G M 1995 Phys. Rev. Lett. 75 140
[69] Pyatov P N and Twarock R 2002 J. Math. Phys. 43 3268
[70] Twarock R 2002 Proc. Quantum Theory and Symmetries ed E Kapuscik and
A Horzela (Singapore: World Scientific) p 615
[71] Albeverio S and Fei S-M 2001 J. Phys. A: Math. Gen. 34 6545
[72] Belavin A A, Polyakov A M and Zamolodchikov A B 1984 Nucl. Phys. B 241 333
[73] Zamolodchikov A B 1989 Adv. Stud. Pure Math. 19 641
[74] Henkel M, Pleimling M, Godrèche C and Luck J-M 2001 Phys. Rev. Lett. 87 265701
[75] Henkel M 2002 Nucl. Phys. B 641 405
[76] Picone A and Henkel M in preparation
[77] Henkel M 1994 J. Stat. Phys. 75 1023
[78] Godrèche C and Luck J-M 2002 J. Phys: Condens. Matter 14 1589
[79] Henkel M and Pleimling M 2003 Preprint cond-mat/0302482
[80] Henkel M, Paessens M and Pleimling M 2003 Europhys. Lett. 62 664
[81] Niederer U 1972 Helv. Phys. Acta 45 802
[82] Henkel M and Unterberger J 2003 Nucl. Phys. B 660 407
[83] Picone A and Henkel M 2002 J. Phys. A: Math. Gen. 35 5575

Copyright © 2003 IOP Publishing Ltd.