The synthesis of symplectic geometry, the calculus of variations and control theory offered in this book provides a cruc

*388*
*26*
*4MB*

*English*
*Pages 436*
*Year 2016*

*Table of contents : Contents......Page 5Acknowledgments......Page 10Introduction......Page 11Chapter 1 The Orbit Theorem and Lie determined systems......Page 221.1 Vector fields and differential forms......Page 231.2 Flows and diffeomorphisms......Page 271.3 Families of vector fields: the Orbit theorem......Page 301.4 Distributions and Lie determined systems......Page 352.1 Control systems and families of vector fields......Page 402.2 The Lie saturate......Page 46Chapter 3 Lie groups and homogeneous spaces......Page 503.1 The Lie algebra and the exponential map......Page 523.2 Lie subgroups......Page 553.3 Families of left-invariant vector fields and accessibility......Page 603.4 Homogeneous spaces......Page 624.1 Symplectic vector spaces......Page 654.2 The cotangent bundle of a vector space......Page 684.3 Symplectic manifolds......Page 705.1 Poisson manifolds and Poisson vector fields......Page 765.2 The cotangent bundle of a Lie group: coadjoint orbits......Page 78Chapter 6 Hamiltonians and optimality: the Maximum Principle......Page 856.1 Extremal trajectories......Page 866.2 Optimal control and the calculus of variations......Page 896.3 The Maximum Principle......Page 926.4 The Maximum Principle in the presence of symmetries......Page 1026.5 Abnormal extremals......Page 1066.6 The Maximum Principle and Weierstrass’ excess function......Page 1127.1 Hyperbolic geometry......Page 1177.2 Elliptic geometry......Page 1227.3 Sub-Riemannian view......Page 1277.4 Elastic curves......Page 1347.5 Complex overview and integration......Page 1368.1 Lie groups with an involutive automorphism......Page 1398.2 Symmetric Riemannian pairs......Page 1418.3 The sub-Riemannian problem......Page 1528.4 Sub-Riemannian and Riemannian geodesics......Page 1568.5 Jacobi curves and the curvature......Page 1598.6 Spaces of constant curvature......Page 163Chapter 9 Affine-quadratic problem......Page 1689.1 Affine-quadratic Hamiltonians......Page 1739.2 Isospectral representations......Page 1769.3 Integrability......Page 180Chapter 10 Cotangent bundles of homogeneous spaces as coadjoint orbits......Page 18810.1 Spheres, hyperboloids, Stiefel and Grassmannian manifolds......Page 18910.2 Canonical affine Hamiltonians on rank one orbits: Kepler and Newmann......Page 19810.3 Degenerate case and Kepler’s problem......Page 20010.4 Mechanical problem of C. Newmann......Page 20810.5 The group of upper triangular matrices and Toda lattices......Page 213Chapter 11 Elliptic geodesic problem on the sphere......Page 21911.1 Elliptic Hamiltonian on semi-direct rank one orbits......Page 22011.2 The Maximum Principle in ambient coordinates......Page 22411.3 Elliptic problem on the sphere and Jacobi’s problem on the ellipsoid......Page 23211.4 Elliptic coordinates on the sphere......Page 234Chapter 12 Rigid body and its generalizations......Page 24112.1 The Euler top and geodesic problems on SOn(R)......Page 24212.2 Tops in the presence of Newtonian potentials......Page 252Chapter 13 Isometry groups of space forms and affine systems: Kirchhoff’s elastic problem......Page 25913.1 Elastic curves and the pendulum......Page 26213.2 Parallel and Serret–Frenet frames and elastic curves......Page 26913.3 Serret–Frenet frames and the elastic problem......Page 27213.4 Kichhoff’s elastic problem......Page 277Chapter 14 Kowalewski–Lyapunov criteria......Page 28314.1 Complex quaternions and SO4(C)......Page 28514.2 Complex Poisson structure and left-invariant Hamiltonians......Page 29114.3 Affine Hamiltonians on SO4(C) and meromorphic solutions......Page 29414.4 Kirchhoff–Lagrange equation and its solution......Page 308Chapter 15 Kirchhoff–Kowalewski equation......Page 31715.1 Eulers’ solutions and addition formulas of A. Weil......Page 32515.2 The hyperelliptic curve......Page 33115.3 Kowalewski gyrostat in two constant fields......Page 33416.1 The curvature problem......Page 34716.2 Elastic problem revisited – Dubins–Delauney on space forms......Page 35216.3 Curvature problem on symmetric spaces......Page 37116.4 Elastic curves and the rolling sphere problem......Page 379Chapter 17 The non-linear Schroedinger’s equation and Heisenberg’s magnetic equation–solitons......Page 38917.1 Horizontal Darboux curves......Page 39017.2 Darboux curves and symplectic Fr ´ echet manifolds......Page 39717.3 Geometric invariants of curves and their Hamiltonian vector fields......Page 40717.4 Affine Hamiltonians and solitons......Page 420Concluding remarks......Page 425References......Page 427Index......Page 434*

C A M B R I D G E S T U D I E S I N A DVA N C E D M AT H E M AT I C S 154 Editorial Board ´ S , W. F U LTO N , A . K ATO K , F. K I RWA N , B. BOLLOBA P. S A R NA K , B . S I M O N , B . TOTA RO

Optimal Control and Geometry: Integrable Systems The synthesis of symplectic geometry, the calculus of variations, and control theory offered in this book provides a crucial foundation for the understanding of many problems in applied mathematics. Focusing on the theory of integrable systems, this book introduces a class of optimal control problems on Lie groups whose Hamiltonians, obtained through the Maximum Principle of optimality, shed new light on the theory of integrable systems. These Hamiltonians provide an original and unified account of the existing theory of integrable systems. The book particularly explains much of the mystery surrounding the Kepler problem, the Jacobi problem, and the Kowalewski Top. It also reveals the ubiquitous presence of elastic curves in integrable systems up to the soliton solutions of the non-linear Schroedinger’s equation. Containing a useful blend of theory and applications, this is an indispensable guide for graduates and researchers, in many fields from mathematical physics to space control.

p r o f e s s o r j u r d j e v i c is one of the founders of geometric control theory. His pioneering work with H. J. Sussmann was deemed to be among the most influential papers of the century and his book, Geometric Control Theory, revealed the geometric origins of the subject and uncovered important connections to physics and geometry. It remains a major reference on non-linear control. Professor Jurdjevic’s expertise also extends to differential geometry, mechanics and integrable systems. His publications cover a wide range of topics including stability theory, Hamiltonian systems on Lie groups, and integrable systems. He has spent most of his professional career at the University of Toronto.

CAMBRIDGE TRACTS IN MATHEMATICS GENERAL EDITORS ´ B. BOLLOB AS, W. FULTON, A. KATOK, F. KIRWAN, P. SARNAK, B. SIMON, B. TOTARO All the titles listed below can be obtained from good booksellers or from Cambridge University Press. For a complete series listing visit: http://www.cambridge.org/mathematics 113. A. Kirillov, Jr An introduction to Lie groups and Lie algebras 114. F. Gesztesy et al. Soliton equations and their algebro-geometric solutions, II 115. E. de Faria & W. de Melo Mathematical tools for one-dimensional dynamics 116. D. Applebaum L´evy processes and stochastic calculus (2nd Edition) 117. T. Szamuely Galois groups and fundamental groups 118. G. W. Anderson, A. Guionnet & O. Zeitouni An introduction to random matrices 119. C. Perez-Garcia & W. H. Schikhof Locally convex spaces over non-Archimedean valued fields 120. P. K. Friz & N. B. Victoir Multidimensional stochastic processes as rough paths 121. T. Ceccherini-Silberstein, F. Scarabotti & F. Tolli Representation theory of the symmetric groups 122. S. Kalikow & R. McCutcheon An outline of ergodic theory 123. G. F. Lawler & V. Limic Random walk: A modern introduction 124. K. Lux & H. Pahlings Representations of groups 125. K. S. Kedlaya p-adic differential equations 126. R. Beals & R. Wong Special functions 127. E. de Faria & W. de Melo Mathematical aspects of quantum field theory 128. A. Terras Zeta functions of graphs 129. D. Goldfeld & J. Hundley Automorphic representations and L-functions for the general linear group, I 130. D. Goldfeld & J. Hundley Automorphic representations and L-functions for the general linear group, II 131. D. A. Craven The theory of fusion systems ¨ anen ¨ 132. J. V a¨ an Models and games 133. G. Malle & D. Testerman Linear algebraic groups and finite groups of Lie type 134. P. Li Geometric analysis 135. F. Maggi Sets of finite perimeter and geometric variational problems 136. M. Brodmann & R. Y. Sharp Local cohomology (2nd Edition) 137. C. Muscalu & W. Schlag Classical and multilinear harmonic analysis, I 138. C. Muscalu & W. Schlag Classical and multilinear harmonic analysis, II 139. B. Helffer Spectral theory and its applications 140. R. Pemantle & M. C. Wilson Analytic combinatorics in several variables 141. B. Branner & N. Fagella Quasiconformal surgery in holomorphic dynamics 142. R. M. Dudley Uniform central limit theorems (2nd Edition) 143. T. Leinster Basic category theory 144. I. Arzhantsev, U. Derenthal, J. Hausen & A. Laface Cox rings 145. M. Viana Lectures on Lyapunov exponents ˝ 146. J.-H. Evertse & K. Gy ory Unit equations in Diophantine number theory 147. A. Prasad Representation theory 148. S. R. Garcia, J. Mashreghi & W. T. Ross Introduction to model spaces and their operators 149. C. Godsil & K. Meagher Erd˝os-Ko-Rado theorems: Algebraic approaches 150. P. Mattila Fourier analysis and Hausdorff dimension 151. M. Viana & K. Oliveira Foundations of ergodic theory 152. V. I. Paulsen & M. Raghupathi An introduction to the theory of reproducing kernel Hilbert spaces 153. R. Beals & R. Wong Special functions and orthogonal polynomials 154. V. Jurdjevic Optimal control and geometry: integrable systems

Optimal Control and Geometry: Integrable Systems VELIMIR JURDJEVIC University of Toronto

University Printing House, Cambridge CB2 8BS, United Kingdom Cambridge University Press is part of the University of Cambridge. It furthers the University’s mission by disseminating knowledge in the pursuit of education, learning and research at the highest international levels of excellence. www.cambridge.org Information on this title: www.cambridge.org/9781107113886 © Cambridge University Press 2016 This publication is in copyright. Subject to statutory exception and to the provisions of relevant collective licensing agreements, no reproduction of any part may take place without the written permission of Cambridge University Press. First published 2016 Printed in the United Kingdom by Clays, St Ives plc A catalogue record for this publication is available from the British Library Library of Congress Cataloguing in Publication data Names: Jurdjevic, Velimir. Title: Optimal control and geometry : integrable systems / Velimir Jurdjevic, University of Toronto. Description: Cambridge : Cambridge University Press, 2016. | Series: Cambridge studies in advanced mathematics ; 154 | Includes bibliographical references and index. Identifiers: LCCN 2015046457 | ISBN 9781107113886 (Hardback : alk. paper) Subjects: LCSH: Control theory. | Geometry, Differential. | Hamiltonian systems. | Lie groups. | Manifolds (Mathematics) Classification: LCC QA402.3 .J88 2016 | DDC 515/.642–dc23 LC record available at http://lccn.loc.gov/2015046457 ISBN 978-1-107-11388-6 Hardback Cambridge University Press has no responsibility for the persistence or accuracy of URLs for external or third-party internet websites referred to in this publication, and does not guarantee that any content on such websites is, or will remain, accurate or appropriate.

Contents

Acknowledgments Introduction

page ix xi

Chapter 1 The Orbit Theorem and Lie determined systems 1.1 Vector fields and differential forms 1.2 Flows and diffeomorphisms 1.3 Families of vector fields: the Orbit theorem 1.4 Distributions and Lie determined systems

1 2 6 9 14

Chapter 2 Control systems: accessibility and controllability 2.1 Control systems and families of vector fields 2.2 The Lie saturate

19 19 25

Chapter 3 Lie groups and homogeneous spaces 3.1 The Lie algebra and the exponential map 3.2 Lie subgroups 3.3 Families of left-invariant vector fields and accessibility 3.4 Homogeneous spaces

29 31 34 39 41

Chapter 4 Symplectic manifolds: Hamiltonian vector fields 4.1 Symplectic vector spaces 4.2 The cotangent bundle of a vector space 4.3 Symplectic manifolds

44 44 47 49

Chapter 5 Poisson manifolds, Lie algebras, and coadjoint orbits 5.1 Poisson manifolds and Poisson vector fields 5.2 The cotangent bundle of a Lie group: coadjoint orbits

55 55 57

Chapter 6 Hamiltonians and optimality: the Maximum Principle 6.1 Extremal trajectories 6.2 Optimal control and the calculus of variations

64 65 68

v

vi

6.3 6.4 6.5 6.6

Contents

The Maximum Principle The Maximum Principle in the presence of symmetries Abnormal extremals The Maximum Principle and Weierstrass’ excess function

71 81 85 91

Chapter 7 Hamiltonian view of classic geometry 7.1 Hyperbolic geometry 7.2 Elliptic geometry 7.3 Sub-Riemannian view 7.4 Elastic curves 7.5 Complex overview and integration

96 96 101 106 113 115

Chapter 8 Symmetric spaces and sub-Riemannian problems 8.1 Lie groups with an involutive automorphism 8.2 Symmetric Riemannian pairs 8.3 The sub-Riemannian problem 8.4 Sub-Riemannian and Riemannian geodesics 8.5 Jacobi curves and the curvature 8.6 Spaces of constant curvature

118 118 120 131 135 138 142

Chapter 9 Affine-quadratic problem 9.1 Affine-quadratic Hamiltonians 9.2 Isospectral representations 9.3 Integrability

147 152 155 159

Chapter 10 10.1 10.2 10.3 10.4 10.5

Cotangent bundles of homogeneous spaces as coadjoint orbits Spheres, hyperboloids, Stiefel and Grassmannian manifolds Canonical affine Hamiltonians on rank one orbits: Kepler and Newmann Degenerate case and Kepler’s problem Mechanical problem of C. Newmann The group of upper triangular matrices and Toda lattices

Chapter 11 Elliptic geodesic problem on the sphere 11.1 Elliptic Hamiltonian on semi-direct rank one orbits 11.2 The Maximum Principle in ambient coordinates 11.3 Elliptic problem on the sphere and Jacobi’s problem on the ellipsoid 11.4 Elliptic coordinates on the sphere

167 168 177 179 187 192 198 199 203 211 213

Contents

Chapter 12 Rigid body and its generalizations 12.1 The Euler top and geodesic problems on SOn (R) 12.2 Tops in the presence of Newtonian potentials

vii

220 221 231

Chapter 13

Isometry groups of space forms and affine systems: Kirchhoff’s elastic problem Elastic curves and the pendulum Parallel and Serret–Frenet frames and elastic curves Serret–Frenet frames and the elastic problem Kichhoff’s elastic problem

238 241 248 251 256

Chapter 14 Kowalewski–Lyapunov criteria 14.1 Complex quaternions and SO4 (C) 14.2 Complex Poisson structure and left-invariant Hamiltonians 14.3 Affine Hamiltonians on SO4 (C) and meromorphic solutions 14.4 Kirchhoff–Lagrange equation and its solution

262 264 270 273 287

Chapter 15 Kirchhoff–Kowalewski equation 15.1 Eulers’ solutions and addition formulas of A. Weil 15.2 The hyperelliptic curve 15.3 Kowalewski gyrostat in two constant fields

296 304 310 313

13.1 13.2 13.3 13.4

Chapter 16 16.1 16.2 16.3 16.4

Elastic problems on symmetric spaces: the Delauney–Dubins problem The curvature problem Elastic problem revisited – Dubins–Delauney on space forms Curvature problem on symmetric spaces Elastic curves and the rolling sphere problem

326 326 331 350 358

Chapter 17 17.1 17.2 17.3 17.4

The non-linear Schroedinger’s equation and Heisenberg’s magnetic equation–solitons Horizontal Darboux curves Darboux curves and symplectic Fr´echet manifolds Geometric invariants of curves and their Hamiltonian vector fields Affine Hamiltonians and solitons Concluding remarks

386 399 404

References Index

406 413

368 369 376

Acknowledgments

This book grew out of the lecture notes written during the graduate courses that I gave at the University of Toronto in the mid 2000s. These courses made me aware of the need to bridge the gap between mainstream mathematics, differential geometry, and integrable systems, and control theory, and this realization motivated the initial conception of the book. I am grateful to the University of Toronto for imposing the mandatory retirement that freed my time to carry out the necessary research required for the completion of this project. There are several conferences and workshops which have had an impact on this work. In particular I would like to single out the Conference of Geometry, Dynamics and Integrable systems, first held in 2010 in Belgrade, Serbia and second in Sintra – Lisbon, Portugal in 2011, the INDAM meeting on Geometric Control and sub-Riemannian Geometry, held in Cortona, Italy in 2012, the IV Ibaronamerican Meeting on Geometric Mechanics and Control held in Rio de Janeiro in 2014, and the IMECC/Unicamp Fourth School and Workshop on Lie theory held in Campinas, Brazil in 2015. I thank the organizers for giving me a chance to present some of the material in the book and benefit from the interaction with the scientific community. I also would like to thank the editorial staff of Cambridge University Press for providing an invaluable assistance in transforming the original manuscript to its present form. In particular, my thanks go to David Hemsley for his painstaking work in unraveling the inconsistencies and ambiguities in my original submission. I feel especially indebted to my wife Ann Bia for taking care of the life around us during my preoccupation with this project. Thank you Ann.

ix

Introduction

Upon the completion of my book on geometric control theory, I realized that this subject matter, which was traditionally regarded as a domain of applied mathematics connected with the problems of engineering, made important contributions to mathematics beyond the boundaries of its original intent. The fundamental questions of space control, starting with the possibility of navigating a dynamical system from an initial state to a given final state, all the way to finding the best path of transfer, inspired an original theory of differential systems based on Lie theoretic methods, and the quest for the best path led to the Maximum Principle of optimality. This theory, apart from its relevance for the subject within which it was conceived, infuses the calculus of variations with new and fresh insights: controllability theory provides information about the existence of optimal solutions and the Maximum Principle leads to the solutions via the appropriate Hamiltonians. The new subject, a synthesis of the calculus of variations, modern symplectic geometry and control theory, provides a rich foundation indispensable for problems of applied mathematics. This recognition forms the philosophical underpinning for the book. The bias towards control theoretic interpretations of variational problems provides a direct path to Hamiltonian systems and reorients our understanding of Hamiltonian systems inherited from the classical calculus of variations in which the Euler–Lagrange equation was the focal point of the subject. This bias also reveals a much wider relevance of Hamiltonian systems for problems of geometry and applied mathematics than previously understood, and, at the same time, it offers a distinctive look at the theory of integrable Hamiltonian systems. This book is inspired by several mathematical discoveries in the theory of integrable systems. The starting point was the discovery that the mathematical formalism initiated by G. Kirchhoff to model the equilibrium configurations of a thin elastic bar subjected to twisting and bending torques at its ends can be xi

xii

Introduction

reformulated as an optimal control problem on the orthonormal frame bundle of R3 , with obvious generalizations to any Riemannian manifold. On threedimensional spaces of constant curvature, where the orthonormal frame bundle coincides with the isometry group, this generalization of Kirchhoff’s elastic model led to a left-invariant Hamiltonian H on a six-dimensional Lie group whose Hamiltonian equations on the Lie algebra showed remarkable similarity with the equations of motion for the heavy top (a rigid body fixed at a point and free to move around this point under the gravitational force). Further study revealed an even more astonishing fact, that the associated control Hamiltonian system is integrable precisely in three cases under the same conditions as the the heavy top [JA; Jm]. This discovery showed that the equations of the heavy top form an invariant subsystem of the above Hamiltonian system and that the solvability of the equations for the top is subordinate to the integrability of this Hamiltonian system (and not the other way around as suggested by Kirchhoff and his “kinetic analogue” metaphor [Lv]. More importantly, this discovery suggested that integrability of mechanical tops is better understood throgh certain left-invariant Hamiltonians on Lie groups, rather than through conventional methods within the confines of Newtonian physics. The above discovery drew attention to a larger class of optimal control problems on Lie groups G whose Lie algebra g admits a Cartan decomposition g = p ⊕ k subject to [p, p] ⊆ k, [p, k] = p, [k, k] ⊆ k.

(I.1)

These optimal problems, defined by an element A ∈ p and a positive definite quadratic form , on k, consist of finding the solutions of the affine control system dg = g(A + U(t)), U(t) ∈ k, dt

(I.2)

that conform to the given boundary conditions g(0) = g0 and g(T) = g1 , for 1 T which the integral 2 0 U(t), U(t) dt is minimal. This class of optimal control problems is called affine-quadratic. We show that any affine-quadratic problem is well defined for any regular element A in p in the sense that for any any pair of points g0 and g1 in G, there exists a time T > 0, and a control U(t) on [0, T] that generates a solution T g(t) in (2) with g(0) = g0 and g(T) = g1 , and attains the minimum of 12 0 U(t), U(t) dt. Remarkably, the Hamiltonians associated with these optimal problems reveal profound connections with integrable systems. Not only do they link mechanical tops with geodesic and elastic problems, but also reveal the hidden

Introduction

xiii

symmetries, even for the most enigmatic systems such as Jacobi’s geodesic problem on the ellipsoid, and the top of Kowalewski. This book lays out the mathematical foundation from which these phenomena can be seen in a unified manner. As L. C. Young notes in his classical book on the calculus of variations and optimal control [Yg], problems of optimality are not the problems to tackle with bare hands, but only when one is properly equipped. In the process of preparing ourselves for the tasks ahead it became necessary to amalgamate symplectic and Poisson geometry with control theory. This synthesis forms the theoretic background for problems of optimality. Along the way, however, we discovered that this theoretic foundation also applies to the classic theory of Lie groups and symmetric spaces as well. As a result, the book turned out to be as much about Lie groups and homogeneous spaces, as is about the problems of the calculus of variations and optimal control. The subject matter is introduced through the basic notions of differential geometry, manifolds, vector fields, differential forms and Lie brackets. The first two chapters deal with the accessibility theory based on Lie theoretic methods, an abridged version of the material presented earlier in [Jc]. The orbit theorem of this chapter makes a natural segue to the chapters on Lie groups and Poisson manifolds, where it is used to prove that a closed subgroup of a Lie group is a Lie group and that a Poisson manifold is foliated by symplectic manifolds. The latter result is then used to show that the dual of a Lie algebra is a Poisson manifold, with its Poisson structure inherited from the symplectic structure of the cotangent bundle, in which the symplectic leaves are the coadjoint orbits. This material ends with a discussion of left-invariant Hamiltonians, a prelude to the Maximum Principle and differential systems with symmetries. The chapter on the Maximum Principle explains the role of optimal control for problems of the calculus of variations and provides a natural transition to the second part of the book on integrable systems. The Maximum Principle is presented through its natural topological property as a necessary condition for a trajectory to be on the boundary of the reachable set. The topological view of this principle allows for its strong formulation over an enlarged system, called the Lie saturate, that includes all the symmetries of the system. This version of the Maximum Principle is called the Saturated Maximum Principle. It is then shown that Noether’s theorem and the related Moment map associated with the symmetries are natural consequences of the Saturated Maximum Principle. This material forms the theoretic background for the second part of the book, which, for the most part, deals with specific problems. This material begins with a presentation of the non-Euclidean geometry from the Hamiltonian point

xiv

Introduction

of view. This choice of presentation illustrates the relevance of the above formalism for the problems of geometry and also serves as the natural segue to the chapter on Lie groups G with an involutive automorphism σ and to the geometric problems on G induced by the associated Cartan decomposition g = p⊕k of the Lie algebra g of G. In these situations the Cartan decomposition then yields a splitting Fp ⊕ Fk of the tangent bundle TG, with Fp and Fk the families of left-invariant vector fields on G that take values in p, respectively k, at the group identity e. The distributions defined by these families of vector fields, called vertical and horizontal form a basis for the class of variational problems on G described below. Vertical distribution Fk is involutive and its orbit through the group identity e is a connected Lie subgroup K of G whose Lie algebra is k. This subgroup is contained in the set of fixed points of σ and is the smallest Lie subgroup of G with Lie algebra equal to k, and can be regarded as the structure group for the homogeneous space M = G/K. Horizontal family Fp is in general not involutive. We then use the Orbit theorem to show that on semi-simple Lie algebras the orbit of Fp through the group identity is equal to G if and only if [Fp , Fp ] = Fk , or, equivalently, if and only if [p, p] = k. This controllability condition, translated to the language of the principal bundles, says that [p, p] = k is a necessary and sufficient condition that any two points in G can be connected by a horizontal curve in G, where a horizontal curve is a curve that is tangent to Fp . The aforementioned class of problems on G is divided into two classes each treated somewhat separately. The first class of problems, inspired by the Riemannian problem on M = G/K defined by a positive-definite, AdK invariant quadratic form , on p is treated in Chapter 8. In contrast to the existing literature on symmetric spaces, which introduces this subject matter through the geodesic symmetries of the underlying symmetric space [Eb; Hl], the present exposition is based on the pioneering work of R. W. Brockett [Br1; Br2] and begins with the sub-Riemannian problem of finding a horizontal T dg −1 g−1 (t) dg curve g(t) in G of minimal length 0 dt , g (t) dt dt that connects given points g0 and g1 under the assumption that [p, p] = k. We demonstrate that this intrinsic sub-Riemannian problem is fundamental for the geometry of the underlying Riemannian symmetric space G/K, in the sense that all of its geometric properties can be extracted from g, without ever descending onto the quotient space G/K. We show that the associated Hamiltonian system is completely integrable and that its solutions can be written in closed form as g(t) = g0 exp t(A + B) exp (−tB), A ∈ p, B ∈ h.

(I.3)

Introduction

xv

The projection of these curves on the underlying manifold G/K coincides with the curves of constant geodesic curvature, with B = 0 resulting in the geodesics. We then extract the Riemannian curvature tensor κ(A, B) = [[A, B], A], B, A ∈ p, B ∈ p.

(I.4)

from the associated Jacobi equation. The chapter ends with a detailed analysis of the Lie algebras associated with symmetric spaces of constant curvature, the setting frequently used in the rest of the text. The second aforementioned class of problems, called affine-quadratic, presented in Chapter 9, in a sense is complementary to the sub-Riemannian case mentioned above, and is most naturally introduced in the language of control theory as an optimal control problem over an affine distribution D(g) = {g(A+U) : U ∈ k defined by an element A in p and a positive-definite quadratic form Q(u, v) defined on k. The first part deals with controllability, as a first step to the well-posedness of the problem. We first note a remarkable fact that any semi-simple Lie algebra g, as a vector space, carries two Lie bracket structures: the semi-simple Lie algebra and the semi-direct product Lie algebra induced by the adjoint action of K on p. This means that the affine-quadratic problem on a semi-simple Lie group G then admits analogous formulation on the semidirect product Gs = p K. Hence, the semi-direct affine-quadratic problem is always present behind every semi-simple affine problem. We refer to this semi-direct affine problem as the shadow problem. We then show that every affine system is controllable whenever A is a regular element in p. This fact implies that the corresponding affine-quadratic problem is well posed for any positive-definite quadratic form on k. On semi-simple Lie groups G with K compact and with a finite center, the Killing form is negative-definite on k and can be used to define an AdK invariant, positive-definite bilinear form , on k. The corresponding optimal control system is AdK -invariant and hence can be regarded as the canonical affine-quadratic problem on G. It is then natural to consider the departures from the canonical case defined by a quadratic form Q(u), v for some linear transformation Q on k which is positive-definite relative to , . Any such affine-quadratic problem induces a left-invariant affine Hamiltonian H=

1 −1 Q (Lk ), Lk + A, Lp 2

(I.5)

on the Lie algebra g = p ⊕ k obtained by the Maximum Principle, where Lk and Lp denote the projections of an element L ∈ g on the factors k and p. The Hamiltonians which admit a spectral representation of the form

xvi

Introduction dLλ = [Mλ , Lλ ] with dt Mλ = Q−1 (Lk ) − λA, and Lλ = −Lp + λLh + (λ2 − s)B

(I.6)

for some matrix B, are called isospectral. In this notation s is a parameter, equal to zero in the semi-simple case and equal to one in the semi-direct case. The spectral invariants of Lλ = Lp −λLk +(λ2 −s)B are constants of motion and are in involution with each other relative to the Poisson structure induced by either the semi-simple Lie algebra g or by the semi-direct product gs = pk (see [Rm], also [Bv; RT]). We show that an affine Hamiltonian H is isospectral if and only [Q−1 (Lk ), A] = [Lk , B]

(I.7)

for some matrix B ∈ p that commutes with A. In the isospectral case every solution of the homogeneous part dLk = [Q−1 (Lk ), Lk ] dt

(I.8)

is the projection of a solution Lp = sB of the affine Hamiltonian system (I.6) and hence admits a spectral representation dLk = [Q−1 (Lk ) − λA, Lk − λB]. dt

(I.9)

The above shows that the fundamental results of A. T. Fomenko and V. V. Trofimov [Fa] based on Manakov’s seminal work on the n-dimensional Euler’s top are subordinate to the isospectral properties of affine Hamitonian systems on g, in the sense that the spectral invariants of Lk − λB are always in involution with a larger family of functions generated by the spectral invariants of Lλ = −Lp + λLh + (λ2 − s)B on g. The spectral invariants of Lλ belong to a larger family of functions on the dual of the Lie algebra whose members are in involution with each other, and are sufficiently numerous to guarantee integrability in the sense of Liouville on each coadjoint orbit in g∗ [Bv]. We then show that the cotangent bundles of space forms, as well as the cotangent bundles of oriented Stiefel and oriented Grassmannian manifolds can be realized as the coadjoint orbits in the space of matrices having zero trace, in which case the restriction of isospectral Hamiltonians to these orbits results in integrable Hamiltonians on the underlying manifolds. In particular, we show that the restriction of the canonical affine Hamiltonian to the

Introduction

xvii

cotangent bundles of non-Euclidean space forms (spheres and hyperboloids) is given by H=

1 1 ||x||2 ||y||2 − (Ax, x) , = ±1. 2 2

(I.10)

This Hamiltonian governs the motion of a particle on the space form under a quadratic potential V = 12 (Ax, x) . We then show that all of these mechanical systems are completely integrable by computing the integrals of motion generated by the spectral invariants of the matrix Lλ . These integrals of motion coincide with the ones presented by J. Moser in [Ms2] in the case of C. Newmann’s system on the sphere. Remarkably, the degenerate case A = 0 provides a natural explanation for the enigmatic discovery of V.A. Fock that the solutions of Kepler’s problem move along the geodesics of the space forms [Fk; Ms1; O1; O2]. We show that the stereographic projections from the sphere, respectively the hyperboloid, can be extended to the entire coadjoint orbit in such a way that the extended map is a symplectomorphism from the coadjoint orbit onto the cotangent bundle of Rn /{0} such that H = 12 (x, x) (y, y) is mapped onto 1 and the energy level H = 2h 2 is mapped onto the energy E = 12 ||p||2 − ||q||

level E = − 12 h2 . Therefore E < 0 in the spherical case and E > 0 in the hyperbolic case. The Euclidean case E = 0 is obtained by a limiting argument when is regarded as a continuous parameter which tends to zero. This correspondence also identifies the angular momentum and the Runge– Lenz vector associated with the problem of Kepler with the moment map associated with the Hamiltonian H. The chapter on the matrices in sln+1 (R) also includes a discussion of a leftinvariant geodesic problem on the group of upper triangular matrices that is relevant for the solutions of a Toda lattice system. Our exposition then turns to Jacobi’s geodesic problem on the ellipsoid and the origins of its integrals of motion. We show that there is a surprising and beautiful connection between this classical problem and isospectral affine Hamiltonians on sln (R) that sheds much light on the symmetries that account for the integrals of motion. The path is somewhat indirect: rather than starting with Jacobi’s problem on the ellipsoid x · D−1 x = 1, we begin instead with a geodesic T √ problem on the sphere in which the length is given by the elliptic metric 0 (Dx(t) · x(t) dt. It turns out that the Hamiltonian system corresponding to the elliptic problem on the sphere is symplectomorphic to the Hamiltonian system associated with the geodesic problem of Jacobi on the ellipsoid, but in contrast to Jacobi’s problem, the Hamiltonian system on the sphere can be represented as a coadjoint orbit. It turns out that the Hamiltonian system associated with the elliptic problem on

xviii

Introduction

the sphere is equal to the restriction of an isospectral affine Hamiltonian system H on sln+1 (R), and hence inherits the integrals of motion from the spectral matrix Lλ . In fact, the Hamiltonian is given by H=

1 −1 D Lk D−1 , Lk + D−1 , Lp , 2

(I.11)

and its spectral matrix by Lλ = Lp − λLk + (λ2 − s)D. This observation reveals that the mechanical problem of Newmann and the elliptic problem on the sphere share the same integrals of motion. This discovery implies not only that all three problems – the mechanical problem of Newmann, Jacobi’s problem on the ellipsoid and the elliptic problem on the sphere – are integrable, but it also identifies the symmetries that account for their integrals of motion. These findings validate Moser’s speculation that the symmetries that account for these integrals of motion are hidden in the Lie algebra sln+1 (R) [Ms3]. The material then shifts to the rigid body and the seminal work of S. V. Manakov mentioned earlier. We interpret Manakov’s integrability results in the realm of isospectral affine Hamiltonians, and provide natural explanations for the integrability of the equations of motion for a rigid body in the presence of a quadratic Newtonian field (originally discovered by O. Bogoyavlensky in 1984 [Bg1]. We then consider the Hamiltonians associated with the affine-quadratic problems on the isometry groups SE3 (R), SO4 (R) and SO(1, 3). These Hamiltonians contain six parameters: three induced by the left-invariant metric and another three corresponding to the coordinates of the drift vector. The drift vector reflects how the tangent of the curve is related to the orthonormal frame along the curve. To make parallels with a heavy top, we associate the metric parameters with the principal moments of inertia and the coordinates of the drift vector with the coordinates of the center of gravity. Then we show that these Hamiltonians are integrable precisely under the same conditions as the heavy tops, with exactly three integrable cases analogous to the top of Euler, top of Lagrange and the top of Kowalewski. The fact that the Lie algebras so4 (R) and so(1, 3) are real forms for the complex Lie algebra so4 (C) suggests that the Hamiltonian equations associated with Kirchhoff’s problem should be complexified and studied on so4 (C) rather than on the real Lie algebras. This observation seems particularly relevant for the Kowalewski case. We show that the Hamiltonian system that corresponds to her case admits four holomorphic integrals of motion, one of which is of the form 1 2 ¯ 1 2 λ 2 λ ¯2 z − bw1 + s b z − bw2 + s b , I4 = 2λ 1 2 2λ 2 2

Introduction

xix

where s = 0 corresponds to the semi-direct case and s = 1 to the semi-simple case. For s = 0 and λ = 1 this integral of motion coincides with the one obtained by S. Kowalewski in her famous paper of 1889 [Kw]. The passage to complex Lie algebras validates Kowalewska’s mysterious use of complex variables and also improves the integration procedure reported in [JA]. Our treatment of the above Hamiltonians reveals ubiquitous presence of elastic curves in these Hamitonians. Elastic curves are the projections of 1 T 2 extremal curves associated with the functional 2 0 κ (s) ds. In Chapter 16 we consider this problem in its own right as the curvature problem. Parallel to the curvature problem we also consider the problem of finding a curve of shortest length among the curves that satisfy fixed tangential directions at their ends and whose curvature is bounded by a given constant c. This problem is referred to as the Dubins–Delauney problem. Our interest in Delauney–Dubins problem is inspired by a remarkable paper of L. Dubins of 1957 [Db] in which he showed that optimal solutions exist in the class of continuosly differentiable curves having Lebesgue integrable second derivatives, and characterized optimal solutions in the plane as the concatenations of arcs of circles and straight line segments with the number of switchings from one arc to another equal to at most two. We will show that the solutions of n-dimensional Dubins’ problem on space forms are essentially three dimensional and are characterized by two integrals of motion I1 and I2 . Dubins’ planar solutions persist on the level I2 = 0, while on I2 = 0 the solutions are given by elliptic functions obtained exactly as in the paper of J. von Schwarz of 1934 in her treatment of the problem of Delaunay [VS]. Our solutions also clarify Caratheodory’s fundamental formula for the problem of Delauney at the end of his book on the calculus of variations. [Cr, p. 378]. This chapter also includes a derivation of the Hamiltonian equation associated with the curvature problem on a general symmetric space G/K corresponding to the Riemannian symmetric pair (G, K). The corresponding formulas show clear dependence of this problem on the Riemannian curvature of the underlying space. We then recover the known integrability results on the space forms explained earlier in the book, and show the connections with rolling sphere problems discovered in [JZ]. The book ends with with a brief treatment of infinite-dimensional Hamiltonian systems and their relevance for the solutions of the non-linear Schroedinger equation, the Korteveg–de Vries equation and Heisenberg’s magnetic equation. This material is largely inspired by another spectacular property of the elastic curves – they appear as the soliton solutions in the non-linear Schroedinger equation. We will be able to demonstrate this fact by

xx

Introduction

introducing a symplectic structure on an infinite-dimensional Fr´echet manifold of framed curves of fixed length over a three-dimensional space form. We will then use the symplectic form to identify some partial differential equations of mathematical physics with the Hamiltonian flows generated by the functionals defined by the geometric invariants of the underlying curves, such as the curvature and the torsion functionals. Keeping in mind the reader who may not be familiar with all aspects of this theory we have made every effort to keep the exposition self-contained and integrated in a way that minimizes the gap between different fields. Unavoidably, some aspects of the theory have to be taken for granted such as the basic knowledge of manifolds and differential equations,

1 The Orbit Theorem and Lie determined systems

Let us begin with the basic concepts and notations required to set the text in motion starting from differentiable manifolds as the point of departure. Throughout the text, manifolds will be generally designated by the capital letters M, N, O, . . . and their points by the lower case letters x, y, z, . . . . Unless otherwise stated, all manifolds will be finite dimensional, smooth, and second countable, that is, can be covered by countably many coordinate charts. Local charts on M will be denoted by (U, φ) with U a coordinate neighborhood and φ a coordinate map on U. For each point x ∈ U the coordinates φ(x) in Rn will be denoted by (x1 , x2 , . . . , xn ). For notational simplicity, we will often write x = (x1 , . . . , xn ), meaning that (x1 , . . . , xn ) is the coordinate representation of a point x in some coordinate chart (φ, U). The set of smooth functions f defined on open subsets of M will be denoted by C∞ (M). It is a ring with respect to pointwise addition ( f + g)(x) = f (x) + g(x) and pointwise multiplication ( fg)(x) = f (x)g(x), both defined on the intersection of their domains. On smooth manifolds tangent vectors v can be regarded both as the equivalence classes of curves and as the derivations. The first case corresponds to the notion of an “arrow”: a tangent vector v at x is defined as the equivalence class of parametrized curves σ (t) defined in some open interval I containing 0 such that σ (0) = x with σ1 ∼ σ2 if and only if in each coordinate chart (U, φ) d d dt φ ◦ σ1 |t=0 = dt φ ◦ σ2 |t=0 . We shall follow the usual custom and write dxn d dx1 ,..., , v = φ ◦ x(t)|t=0 = dt dt dt where x(t) is any representative in the equivalence class of curves that defines v. In the second case, tangent vector are defined by their action on functions resulting in directional derivatives. As such, tangent vectors v at a point x are linear mappings from C∞ (M) into R that satisfy the Leibnitz formula, 1

2

1 The Orbit Theorem and Lie determined systems

v( fg) = f (x)v(g) + g(x)v( f ), for any functions f and g. In this context, tangent vectors in local coordinates will be written as v = ni=1 vi ∂x∂ i , where ∂ ∂ ∂ ∂x1 , . . . ∂xn denotes the usual basis of tangent vectors defined by ∂xi f = f ∈ C∞ (M). These two notions of tangent vectors are reconciled through the pairing

∂f ∂xi ,

∂f d vi . v( f ) = f ◦ x(t)|t=0 dt ∂xi n

i=1

We will use Tx M to denote the tangent space at x, and use TM to denote the tangent bundle of M; TM is a smooth manifold whose dimension is equal to 2dim(M). Each coordinate chart (U, φ) in M induces a coordinate chart ˜ in TM with φ(v) ˜ (TU, φ) = (x1 , . . . , xn , v1 , . . . , vn ) for every v ∈ Tx (U) and x ∈ U. Recall that cotangent vectors at x are the equivalence classes of functions in ∞ C (M) that vanish at x, with f and g in the same equivalence class if and only if f and g coincide in some open neighborhood of x and dtd f ◦ σ (t)|t=0 = dtd g ◦ σ (t)|t=0 for every curve σ on M such that σ (0) = x. The pairing [f ], [σ ] = d dt f ◦ σ (t)|t=0 reflects the duality between tangent and cotangent vectors and identifies cotangent vectors as linear functions on Tx (M). In local coordinates, cotangent vectors will be written as the sums n ∂f i i=1 ∂xi dx , where dx, , . . . , dxn denotes the dual basis relative to the basis ∂ ∂ ∗ ∂x1 , . . . , ∂xn . The cotangent space at x will be denoted by Tx M and the ∗ ∗ cotangent bundle ∪{T (M) : x ∈ M} will be denoted by T M. The cotangent bundle is also a smooth manifold whose dimension is twice that of the underlying manifold M.

1.1 Vector fields and differential forms Since these objects are fundamental for this study, it is essential to be precise about their meanings. Recall that a vector field X on M is a smooth mapping from M into TM such that π ◦ X = I, where π denotes the natural projection from TM onto M. Thus X(x) belongs to Tx (M) for each point x in M. In the language of vector bundles, vector fields are sections of the tangent bundle. The space of smooth vector fields will be denoted by V ∞ (M). In local coordinates vector fields X will be represented either by the arrow vector (X1 (x1 , . . . , xn ), . . . , Xn (x1 , . . . , xn )), or by the expression X(x1 , . . . , xn ) =

n i=1

Xi (x1 , . . . , xn )

∂ , ∂xi

1.1 Vector fields and differential forms

3

∂f depending on the context. The function ni=1 Xi (x1 , . . . , xn ) ∂x (x1 , . . . , xn ) will i be denoted by Xf . This action of vector fields on functions identifies vector fields with derivations on M, that is, it identifies V ∞ (M) with the linear mappings D on C∞ (M) that satisfy D( fg)(x) = f (x)(D(g)(x)) + g(x)(D( f )(x))

(1.1)

for all functions f and g in C∞ (M). The space V ∞ (M) has a rich mathematical structure. To begin with, it is a module over the ring of smooth functions under the operations: (i) (( f X)g)(x) = f (x)(Xg)(x) for any function g and all x in M. (ii) (X + Y)f = Xf + Yf for all functions f . Secondly, V ∞ (M) is a Lie algebra under the addition defined by (ii) above and the Lie bracket [X, Y] = Y ◦ X − X ◦ Y, where [X, Y] means that [X, Y]f = Y(Xf ) − X(Yf ) for every f ∈ C∞ (M). The reader can easily show that in local coordinates [X, Y] is given by Z = ni=1 Zi ∂x∂ i with Zi =

n ∂Xi j=1

∂xj

Yj −

∂Yi Xj . ∂xj

(1.2)

There seems to be no established convention about the sign of the Lie bracket. In some books the Lie bracket is taken as the negative of the one defined above (for instance, [AM] or [Hl]). Differential forms are geometric objects dual to vector fields. They are defined analogously, as the smooth mappings ω from M into T ∗ M such that π ◦ ω = I, where now π is the natural projection from T ∗ M onto M. In local coordinates, ω will be written as ω(x1 , . . . , xn ) = ni=1 ωi (x1 , . . . , xn )dxi for some some smooth functions ω1 , . . . , ωn . Differential forms act on vector fields to produce functions ω(X) given by ω(X) = ni=1 ωi Xi in each chart (U, φ). Differential forms are contained in the complex of exterior differential forms in which functions in C∞ (M) are considered as the forms of degree 0, and the differential forms defined above as the forms of degree 1. Differential forms of degree k can be defined in several ways [BT; Ar]. For our purposes, it will be convenient to define them through the action on vector fields. A differential form ω of degree k is any mapping ω : V ∞ (M) × · · · × V ∞ (M) → C∞ (M)

k

that satisifies ωx (X1 , . . . Xi−1 , f Xi + gWi , Xi+1 , . . . , Xk ) = f ω(X1 , . . . , Xn ) + g(ω(X1 , . . . , Wi , . . . Xk ),

4

1 The Orbit Theorem and Lie determined systems

for each i ∈ {1, . . . , k} and each function f and g, and ω(X1 , . . . Xi , . . . , Xj , . . . , Xk )) = −ω((X1 , . . . Xj , . . . , Xi , . . . , Xk )), for each index i and j. Forms of degree k will be denoted by k (M). It follows that forms of degree k are k-multilinear and skew-symmetric mappings over V ∞ (M). The skewsymmetry property implies that k (M) = 0, for k > dim(M). Alternatively, differential forms can be defined through the wedge products. The wedge product ω1 ∧ ω2 of 1-forms ω1 and ω2 is a 2-form defined by (ω1 ∧ ω2 )(X, Y) = ω1 (X)ω2 (Y) − ω1 (Y)ω2 (X). Any 2-form ω van be expressed as a wedge product of 1-forms. To demon strate, let X = ni=1 X i ∂x∂ i and Y = ni=1 Y i ∂x∂ i . It follows that n n ∂ ∂ ∂ ∂ i j i j j i = . XYω , (X Y − X Y )ω , ω(X, Y) = ∂xi ∂xj ∂xi ∂xj i, j

But then (dxi ∧ dxj )(X, Y) = ω

ω=

∂ ∂ ∂xi , ∂xj

n

i>j

(X i Y j − X i Y j ). Hence,

ωij (dxi ∧ dxj ),

i, j

where ωij are the functions ω ∂x∂ i , ∂x∂ j . We now come to another indispensable theoretic ingredient, the exterior derivative. Definition 1.1 The exterior derivative d is a mapping from k (M) into

k+1 (M) defined by df (X) = X( f ), when k = 0, dω(X1 , . . . , Xk+1 ) =

k+1 (−1)i+1 Xi ω(X1 , . . . , Xˆ i , . . . , Xk+1 ) i=1

−

(−1)i+j ω([Xi , Xj ], . . . , Xˆ i , . . . , Xk+1 ) i 0, where the hat above an entry indicates the absence of that entry from the expression. For instance, ω(X1 , Xˆ 2 , X3 ) = ω(X1 , X3 ), ω(X1 , X2 , Xˆ 3 ) = ω(X1 , X2 ). In particular, the exterior derivative of a 1-form ω is given by dω(X1 , X2 ) = X1 ω(X2 ) − X2 ω(X1 ) + ω([X1 , X2 ]).

(1.3)

1.1 Vector fields and differential forms

5

To show the exterior derivative in more familiar terms [BT], let X1 = ∂ ∂ to a system of coordinates ∂x1 , . . . , Xn = ∂xn denote the standard basis relative n functions x1 , . . . , xn . If f is a function, then df = i=1 ωi dxi for some n ∂f ∂f , and df = ω1 , . . . , ωn . It follows that ωi = df (Xi ) = Xi ( f ) = ∂x i=1 ∂xi dxi . i Therefore, the exterior derivative of f coincides with the directional derivative. n Consider now the exterior derivative of a 1-form ω = i=1 ωi (x)dxi . It follows that dω(Xi , Xj ) = Xi ω(Xj ) − Xj ω(Xi ) =

∂ωj ∂ωi − , ∂xj ∂xi

since [Xi , Xj ] = 0. Hence, n ∂ωj ∂ωi dxi ∧ dxj . dω = − ∂xj ∂xi i=1, j=1

The exterior derivative of a 2-form ω = ω1 dx2 ∧dx3 +ω2 dx3 ∧dx1 +ω3 dx1 ∧ dx2 in R3 is given by ∂ω1 ∂ω2 ∂ω3 dω(X, Y, Z) = (X · (Y ∧ Z), + + ∂x1 ∂x2 ∂x3 where X · (Y ∧ Z) denotes the signed volume defined by the vectors X = (X1 , X2 , X3 ), Y = (Y1 , Y2 , Y3 ) and Z = (Z1 , Z2 , Z3 ). Differential forms in R3 are intercheangeably identified with vector fields via the following identification: (w ∈ 1 (R3 ) ⇐⇒ W ∈ V ∞ (R3 )) ⇐⇒ w(X) = (W · X), X ∈ V ∞ (R3 ). In particular, if w = df then the corresponding vector field is called the gradient of f and is usually denoted by grad(f ). ∂ω3 ∂ω3 ∂ω3 ∂ω2 ∂ω2 ∂ω1 ∂ω2 1 ∂ω1 The expressions ∂x3 − ∂x2 , ∂x1 − ∂ω ∂x3 , ∂x2 − ∂x1 and ∂x1 + ∂x2 + ∂x3 are known as the curl and the divergence of a vector field ω1 ∂x∂ 1 + ω2 ∂x∂ 2 + ω3 ∂x∂3 . . In this terminology, d2 = 0 coincides with the well-known formulas of vector calculus curl(grad) = 0

and

div(curl) = 0.

Differential forms ω for which the exterior derivative is equal to 0 are called closed. The forms ω for which ω = dγ for some form γ are called exact. It can be shown that d2 = 0, therefore exact forms are automatically closed. The quotient of k-closed forms over the exact k-forms is called the kth de Rham cohomology of M.

6

1 The Orbit Theorem and Lie determined systems

1.2 Flows and diffeomorphisms Let us now consider differential equations dσ (t) = X(σ (t)), (1.4) dt defined by a vector field X in a manifold M. Solution curves σ (t) are called integral curves of X. In local coordinates, integral curves are the solutions of a system of ordinary differential equations dσi (t) = Xi (σ1 (t), . . . , σn (t)), i = 1, . . . , n. dt It then follows from the basic theory of differential equations that for each initial point x there exists integral curves σ (t) of X defined on an open interval I = (−, ) such that σ (0) = x. Any such curve can be extended to a maximal open interval Ix = (e− (x), e+ (x)) whose end points are called the negative and the positive escape time. All integral curves of X that pass through a common point x have the same negative and positive escape time. The solution curve σ (t), t ∈ (e− (x), e+ (x)) is called the integral curve of X through x. Definition 1.2 Let X be a vector field and let = {(x, t) : x ∈ M, t ∈ (e− (x), e+ (x))}. The mapping φ : → M defined by φ(x, t) = σ (t) will be called the flow, or a dynamical system induced by X. The theory of ordinary differential equations concerning the existence and uniqueness of solutions and their smooth dependence on the initial conditions can be summarized by the following essential properties: 1. φ(x, 0) = x for each x ∈ M. 2. φ(x, s + t) = φ(φ(x, s), t) = φ(φ(x, t), s) for all (x, s), (x, t) and (x, s + t) in . 3. φ is smooth. ∂ φ(x, t) = X ◦ φ(x, t). 4. ∂t Conversely, any smooth mapping φ : → M with an open subset of M × R, and a neighborhood of M × {0} that satisfies properties (1), (2), (3), necessarily satisfies property (4) with X = ∂φ ∂t (x, t)|t=0 . Vector field X is called the infinitesimal generator of the flow. The set {φ(x, t) : t ∈ R} is called the trajectory through x, or the motion through x. Definition 1.3 A mapping F from a manifold M onto a manifold N is called a diffeomorphism if F is invertible with both F and its inverse smooth. Manifolds M and N are said to be diffeomorphic if there is a diffeomorphism between them.

1.2 Flows and diffeomorphisms

7

Flows of vector fields induce diffeomorphisms in the following sense. Let U be an open set in M whose closure is compact. Then there is an open interval I = (−a, a) such that U × I is contained in . The mapping t (x) = φ(x, t) is a diffeomorphism from U onto t (U). Indeed, −1 t = −t . A vector field X is said to be complete if = M × R, i.e., if each integral curve of X is defined for all t ∈ R. Complete vector fields induce global flows φ : M × R → M. If X is a complete vector field then the corresponding family of diffeomorphisms { t : t ∈ R} is called the one-parameter group of diffeomorphisms induced by X. Indeed, { t : t ∈ R} is a group under = −t . If a vector field is the composition with t ◦ s = t+s and −1 t not complete then its flow defines a local group of diffeomorphisms in some neighborhood of each point in M. It is known that all vector fields on compact manifolds are complete. The shift in perspective from dynamical systems to groups of diffeomorphisms suggests another name for the trajectories. The trajectory through x becomes the orbit through x under the one-parameter group of diffeomorphisms. Each name evokes its own orientation, hence both will be used depending on the context.

1.2.1 Duality between points and linear functionals The two ways of seeing vector fields, as arrows or as derivations, calls for further notational distinctions that elucidate the calculations with their flows. If X is a vector field then it is natural to write X(q) for the induced tangent vector at q, seen as the arrow with its base at q and the direction X(q), We will use a different notation for the same tangent vector seen as a derivation; qˆ ◦X will denote the same tangent vector defined by (ˆq ◦ X)( f ) = X( f )(q) for any function f , where now qˆ : C∞ (M) → R denotes the evaluation of f at q, i.e., qˆ ( f ) = f (q). It is easy to verify that for each point q ∈ M, qˆ is a homomorphism from C∞ (M) into R, the latter viewed as the ring under multiplication and addition, that is, qˆ is a mapping from C∞ (M) into R that satisfies: 1. qˆ (αf + βg) = α qˆ ( f ) + β qˆ (g) for all real numbers α and β and all functions f and g, and 2. qˆ ( fg) = qˆ ( f )ˆq(g) for all functions f and g. Conversely, for any non-trivial homomorphism φ from C∞ (M) into R there exists a unique point q in M such that qˆ = φ ([AS]). Therefore, the correspondence q → qˆ identifies points in M with linear functionals on C∞ (M) that satisfy (1) and (2) above.

8

1 The Orbit Theorem and Lie determined systems

The dualism between points and linear functionals carries over to diffeomorphisms. If F is any diffeomorphism on M then Fˆ will denote the pull back ˆ f ) = f ◦ F. It follows that Fˆ is a ring on functions in C∞ (M) defined by F( ∞ automorphism on C (M). Conversely, any ring automorphism on C∞ (M) is of the form Fˆ for some diffeomorphism F, as can be easily shown by the same proof as above. We will now extend this notation to the flows t induced by vector fields ˆ t for the flow { t : t ∈ R} induced by X. It on M, and let exp tX denote follows that exp(t + s)X = exp tX ◦ exp sX = exp sX ◦ exp tX, s, t ∈ R.

(1.5)

The above implies that exp tX|t=0 = I and exp −tX = (exp tX)−1 . Moreover, d exp tX = X ◦ exp tX = exp tX ◦ X. (1.6) dt Let us now note an important fact that will be useful in the calculations below. Suppose that { t : t ∈ R} is the flow induced by a vector field X. Then, {F ◦ t ◦ F −1 : t ∈ R} is a one-parameter group of diffeomorphisms on M for any diffeomorphism F, and hence is generated by some vector field Y. It follows that Y = F∗ X ◦ F −1 , where F denotes the tangent map induced by F. d Recall that F∗ (v) = w, where v = dσ dt |t=0 , σ (0) = q, and w = dt (F(σ (t))|t=0 . −1 −1 Since F ◦ t ◦ F acts on points, Y = F∗ X ◦ F is the arrow representation of the infinitesimal generator of {F ◦ t ◦ F −1 : t ∈ R}. As a derivation, Y = Fˆ −1 ◦ exp tX ◦ Fˆ by the following calculation: d ˆ f )) ◦ F −1 = (Fˆ −1 ◦ X ◦ F)( ˆ f ), f ◦ (F ◦ t ◦ F −1 ) = X(F( dt and therefore Yf =

ˆ exp tY = Fˆ −1 ◦ exp tX ◦ F.

(1.7)

Equation (1.6) yields the following asymptotic formula: exp tX ≈ I + tX +

t2 2 tn X + ··· + + ··· . 2 n!

Then, exp tY ◦ exp tX ◦ exp −tY ◦ exp −tX t2 t2 = I + tY + Y 2 + · · · ◦ I + tX + X 2 + · · · 2 2 2 t2 2 t 2 ◦ I − Yt + Y + · · · ◦ I − tX + X + · · · 2 2

(1.8)

1.3 Families of vector fields: the Orbit theorem

9

t2 = I + t(X + Y) + (X 2 + 2X ◦ Y + Y 2 ) + · · · 2 t2 2 2 ◦ (I − t(X + Y) + (X + 2X ◦ Y + Y ) + · · · 2 = I + t2 (Y ◦ X − X ◦ Y) + · · · , which in turn yields an important formula √ √ √ √ d exp tY ◦ exp tX ◦ exp − tY ◦ exp − tX|t=0 = [X, Y]. dt

(1.9)

Similar calculations show that the Lie bracket [X, Y] can alternatively be defined by the formula ∂2 exp −tX ◦ exp sY ◦ exp tX|t=s=0 = [X, Y]. ∂t∂s

(1.10)

The preceeding formula can be seen in slightly more general terms according to the following definitions. Definition 1.4 If X is a vector field then adX : V ∞ (M) → V ∞ (M) denotes the mapping adX(Y) = [X, Y], Y ∈ V ∞ (M). If F is a diffeomorphism on M, then AdF : V ∞ (M) → V ∞ (M) is defined as AdF (X) = Fˆ −1 ◦ X ◦ Fˆ for all X in V ∞ (M). It then follows from (1.10) that d Adexp tX = Adexp tX ◦ adX = adX ◦ Adexp tX . dt

(1.11)

1.3 Orbits of families of vector fields: the Orbit theorem It is well known that each orbit of a one-parameter group of diffeomorphisms { t } generated by a vector field X is a submanifold of the ambient manifold M; the orbit through a critical point of X is zero dimensional, otherwise an orbit is one dimensional. These orbits are often referred to as the leafs of X, in which case M is said to be foliated by the leaves of X. There are two pertinent observations about these orbits that are relevant for the text below: 1. The orbits are not of the same dimension whenever X has critical points. 2. It may happen that an orbit is an immersed rather than an embedded submanifold of M. Recall that a submanifold is called embedded if its

10

1 The Orbit Theorem and Lie determined systems

topology coincides with the relative topology induced by the topology of the ambient manifold. For immersed submanifolds, all relatively open sets are open in the submanifold topology, but there may be other open sets which are not in this class. For instance, each orbit of the flow t (z, w) = {z exp tθ , w exp tφ, z ∈ C, w ∈ C, |z|2 = 1, |w|2 = 1} is dense on the torus T 2 = {z ∈ C : |z|2 = 1} × {w ∈ C : |w|2 = 1} whenever the ratio φθ is irrational. The sets { t (z, w) : t ∈ (t0 , t1 )} are open in the orbit topology, but are not equal to the intersections of open sets in T 2 with the orbit. Remarkably, the manifold structure of orbits generated by one vector field extends to arbitrary families of vector fields, and that is the content of the Orbit theorem. To be more precise, let F be an arbitrary family of vector fields (finite or infinite) which, for simplicity of exposition only, will be assumed to consist of complete vector fields. For each X ∈ F, Xt will denote the one-parameter group of diffeomorphisms on M generated by X, and G(F) will denote the group of diffeomorphisms generated by ∪{ Xt : X ∈ F, t ∈ R}. A typical element in G(F) is of the form X

X

p−1 ◦ · · · Xt11 g = tpp ◦ tp−1

(1.12)

for a subset {X1 , . . . , Xp } of F and some numbers t1 , t2 , . . . , tp , or gˆ = exp t1 X1 ◦ exp t2 X2 ◦ exp tp−1 Xp−1 ◦ exp tp Xp .

(1.13)

Definition 1.5 The set {g(x) : g ∈ G(F)} will be called the orbit of F through x and will be denoted by OF (x). The orbit of F through a point x can be defined analogously on the space of functions as the set of automorphisms φ on C∞ (M) of the form φ = xˆ ◦ exp t1 X1 ◦ exp t2 X2 ◦ exp tp−1 Xp−1 ◦ exp tp Xp . with the understanding that the automorphism xˆ ◦ F is identified with the point F(x). Proposition 1.6 The Orbit theorem (possibly immersed) submanifold of M.

Each orbit OF (x) is a connected

This theorem, well known in the control community, has not yet found its proper place in the literature on geometry and therefore may not be so familiar to the general reader. Partly for that reason, but mostly because of the importance for the subsequent applications, we will outline the most important

1.3 Families of vector fields: the Orbit theorem

11

features of its proof (for more detailed proofs the reader can either consult the original sources [Sf; Ss] or the books [AS; Jc]. The prerequisite for the proof is a manifold version of the implicit function theorem, known as the constant rank theorem. The constant rank theorem Let N and M be manifolds and let F : N → M be a smooth mapping whose rank of the tangent map F∗ (x) is constant as x varies over the points of N. Let k denote this rank and let n denote the dimension of M. Then, (a) F −1 (y) is an (n − k)-dimensional embedded submanifold of N for each y in the range of F. (b) Each point x in N has a neighborhood U such that F(U) is an embedded submanifold of M of dimension k. Sketch of the proof of the Orbit theorem Let N denote the orbit of F through a point x in M. We then have the following: (i) The orbit topology The topology on N as the strongest topology under which all mappings X

X

p−1 ◦ · · · Xt11 (y) {t1 , t2 , . . . , tp } → tpp ◦ tp−1

(1.14)

are continuous as the mappings from Rp into N, where y is an arbitrary point of N, and {X1 , . . . , Xp } an arbitrary finite subset of F. Both the choice of the vector fields in F and their number is arbitrary. Since all such mappings are continuous, the topology of N if finer than the relative topology induced by the topology of the ambient manifold. In particular, N is Hausdorff because M is Hausdorff. (ii) Local charts To define local charts at a point z in N, let F denote a mapping of the form (1.14) that satisfies: 1. F(ˆs) = z for some point sˆ in Rp , and 2. The rank of the tangent map F∗ (ˆs) is maximal among all the mappings given by (1.14). Let Rk(z) denote the rank of F∗ (ˆs) defined by (1) above. It turns out that Rk(z) is constant as z varies over the points of N. Let k denote the common value of Rk(z). For each zˆ in N let F denote any mapping of the form (1.14) that satisfies conditions (1) and (2) above. Then the rank of F is constant in some neighborhood of sˆ in Rp . By the constant rank theorem there exists a neighborhood U of tˆ in Rp such that F(U) is an embedded k-dimensional submanifold of M. Any such set F(U) will be

12

1 The Orbit Theorem and Lie determined systems

referred to as a local integral manifold of F at zˆ. Then it can be shown that each local integral manifold is open in the orbit topology of N. (iii) N is second countable We will supply a complete proof of this assertion since it was omitted in the proof in [Jc]. The fact that each point of N is contained in a coordinate neighborhood that is an embedded submanifold of M implies that N is second countable by the following topological arguments based on the notion of paracompactness. Recall that a topological space is paracompact if every open cover of the space has a locally finite refinement. A collection of subsets of a topological space is said to be locally finite if every point of the space has a neighborhood that meets only finitely many sets of the collection and an open covering V of a topological space is a refinement of a covering U if every V ∈ V is a subset of an element of U . The following fact is crucial for our proof: a locally compact topological space is paracompact if and only if it is second countable. The proof of this assertion can be found in Royden [Ro]. So it suffices to show that each orbit is paracompact. To show that N is paracompact let U be an open cover of N and let z be a point of N. Designate by U an open neighborhood of z that is an embedded submanifold of N. Let {Uα } denote the open sets in U that intersect U. Since U is embedded manifold, each Uα ∩ U is relatively open, and therefore Uα ∩ U = Oα ∩ U for some open set Oα . Since ∪Oα is an open submanifold of M, it is second countable and hence paracompact. Let {Vα } denote a locally finite refinement of {Oα }, and let V be an open neighborhood of z that meets only finitely many Vα . Then V ∩ U is a neighborhood of z that meets only finitely many Vα ∩ Uα , and {Vα ∩ Uα } is an open refinement of {Uα }. Therefore, N is paracompact. (iv) Tangency properties of the orbits A vector field X on M is said to be tangent to a submanifold N if X(x) belongs to Tx (N) for each point x in N. This notion extends to families of vector fields. We will say that a family F is tangent to N if every vector field X in F is tangent to N. An integral curve σ (t) of a vector field X that is tangent to N remains in N for t in some open interval (−, ) whenever σ (0) ∈ N (this fact is an immediate consequence of the existence of solutions to the Cauchy problem in differential equations). Conversely, if each integral curve σ (t) of X that initiates on N remains in N for t in an open neighborhood of 0 then X is tangent to N. Evidently F is tangent to each of its orbits. Moreover, for each y in an orbit N of F the curve σt (s) = yˆ ◦ exp −tX ◦ exp sY ◦ exp tX is contained in N and

1.3 Families of vector fields: the Orbit theorem

13

satisfies σt (0) = y. Therefore, dσ ds |s=0 = yˆ ◦ exp −tX ◦ Y ◦ exp tX belongs to Ty N for all t. Since each tangent space Ty (N) is closed, dtd (ˆy ◦ exp −tX ◦ Y ◦ exp tX)|t=0 = [X, Y](y) belongs to Ty (N). Therefore, each Lie bracket [X, Y] is tangent to the orbits of F. The same applies to the brackets of higher orders. To take full advantage of these observations we need a few definitions. If F is a family of vector fields then and x ∈ M, then Fx will denote the set of tangent vectors {X(x) : X ∈ F}. The set Fx will be called the evaluation of F at x. Definition 1.7 If F is any family of vector fields, then the Lie algebra generated by F is the smallest Lie algebra, in the sense of set inclusion, that contains F. It will be denoted by Lie(F). Its evaluation at any point x will be denoted by Liex (F). Typical elements in Lie(F) are linear combinations of the Lie brackets of the form [Xm , [Xm−1 , . . . , [X2 , X1 ]] . . . ], with {X1 , X2 , . . . , Xm } ⊆ F. An easy induction on the number of iterated Lie brackets shows that Lie(F) is tangent to the orbits of F. Hence Liey (F) ⊆ Ty (OF (x)) for all y in an orbit OF (x). In general, however, the tangent space of an orbit of F at a point y is generated by the tangent maps associated with to the linear span of tangent vectors of the form yˆ ◦ Adexp tp Xp ◦ Adexp tp−1 Xp−1 ◦ · · · ◦ Adexp t1 X1 (X)

(1.15)

associated with the mappings yˆ ◦ exp tp Xp ◦ exp tp−1 Xp−1 ◦ exp t1 X1 ◦ exp tX. In fact, vectors in (1.15) are tangent to the curve σ (t) = qˆ ◦ exp tp Xp ◦ · · · ◦ exp t1 X1 exp tX at t = 0. The example below shows that yˆ ◦ Adexp tY (X) need not be in Liey (F) and that the dimension of N exceeds the dimension of Liey (F). Example 1.8 Let M = R2 and let F = {X, Y} with X(x, y) =

∂ ∂ ∂ , and Y(x, y) = + φ(x) , ∂x ∂x ∂y

where φ is a smooth function that satisfies φ(x) = 0 for x ≤ 0 and φ(x) > 0 for ∂ ∂ and [Y, adX(Y)] = (φ (1) )2 − φφ (2) ) ∂y . Thus x > 0. Then adk X(Y) = φ (k) ∂y Lie(F) is an infinite-dimensional Lie algebra whenever the ring of functions generated by φ (k) , k = 0, 1 . . . is infinite dimensional.

14

1 The Orbit Theorem and Lie determined systems

It follows that Lieq (F) is one-dimensional subspace of the tangent space at q for points q = (x, y) with x ≤ 0. For all other points Lieq (F) = R2 . The reader can easily show that there is only one orbit of F equal to the entire plane R2 .

1.4 Distributions and Lie determined systems The Orbit theorem can be recast as a theorem in the theory of distributions [Sh; St]. Loosely speaking, a k-dimensional distribution is a collection of k-dimensional subspaces D(x) of each tangent space Tx M which varies smoothly with the base point x. One-dimensional distributions reduce to line bundles in TM. Distributions can be seen as generalized differential equations in which the ordinary differential equation dσ dt = X(σ (t)) is replaced by a more general dσ equation dt ∈ D(σ (t)). A curve σ (t) that is a solution of the above inclusion equation is called an integral curve of D. The problem is to find conditions on D that guarantee the existence of integral curves through each initial point x. A submanifold N that contains the point x and also contains every integral curve of D that passes through x is called an integral manifold of D through x. Distribution D is called integrable if integral manifolds exist through each point x ∈ M and are of the same dimension as D. Below we shall consider slight generalizations of the distributions that appear in the literature on differential geometry. For our purposes: Definition 1.9 A distribution D is a subset of the tangent bundle TM consisting of linear subspaces D(q) of Tq (M) for q ∈ M. A distribution is said to be smooth if for each q there exist smooth vector fields X1 , . . . , Xk in a neighborhood U of q such that: (a) X1 (q), . . . , Xk (q) is a basis for D(q), and (b) {X1 (x), X2 (x), . . . , Xk (x)} ⊆ D(x) for all points x in U. In what follows all distributions are assumed to be smooth. A vector field X is said to be tangent to a distribution D if X(x) ∈ D(x) for all x ∈ M. Alternatively, smooth distributions could have been defined in terms of vector fields that are tangent to the distribution for the following reasons. If X1 , . . . , Xk and U are as in Definition 1.9, then let U0 be an open neighborhood of q such that its closure is a compact subset of U. Then there exists a smooth function α on M such that α = 1 on U0 , and α = 0 outside U [Hl]. This implies that the modified vector fields X˜ 1 = αX1 , . . . , X˜ k = αXk are tangent to D and are a basis for D(q).

1.4 Distributions and Lie determined systems

15

The reader should note that the above definition allows for the possibility that the dimension of the distribution may vary with the base point q. We will also make another small departure from the existing literature on this subject, and extend admissible solutions to absolutely continuous curves, in which case we will say that an absolutely continuous curve x(t) is an integral curve of a distribution D if dx dt ∈ D(σ (t)) for almost all points t in the domain of x. Recall that a parametrized curve x(t), t ∈ [t0 , t1 ] in Rn is said to be absolutely continuous if dx dt exists almost t dx everywhere in [t0 , t1 ] and is integrable, and, dτ holds for all s, t in [t0 , t1 ]. Then a curve moreover, x(t) − x(s) = s dτ x(t) on a manifold M is said to be absolutely continuous if it is absolutely continuous in every system of coordinates. It follows that an absolutely continuous curve x(t), t ∈ [t0 , t1 ] is an integral curve of a distribution D whenever there exist vector fields X1 , . . . , Xm tangent to D(x(t)) and measurable and bounded functions u1 (t), . . . , um (t) such that dx = ui (t)Xi (x(t)) dt m

(1.16)

i=1

for each point t at which x is differentiable. The existence theory of differential equations then guarantees that for each choice of vector fields X1 , . . . , Xm that are tangent to a distribution D, and each choice of bounded and measurable functions u1 (t), . . . , um (t) on an interval [0, T] there exist solution curves x(t) of (1.16) defined on some interval [0, s), s > 0. If the interval [0, s) is maximal, then x(t) is a unique integral curve that satisfies the above differential equation and passes through a fixed point at t = 0. A distribution is said to be involutive if the Lie bracket of vector fields tangent to the distribution is also tangent to the distribution. A distribution D is said to be integrable if for each x in M there exists a submanifold Nx that contains x and in addition satisfies Ty (Nx ) = D(y) for all points y ∈ Nx . Such a manifold is called an integral manifold of D through x. An integral manifold that is connected and is not a subset of any other connected integral manifold is said to be maximal. The example below shows that the distributions need not be of constant rank. Example 1.10 Let Mn (R) denote the space of all n×n matrices and let D(x) = {Ax : AT = −A, A ∈ Mn (R)}, for each x ∈ Rn . Then D(0) = {0} and D(x) = {y : x · y = 0} for each x = 0, because any vector y which is orthogonal to x can be written as y = ni=1 yxij (ei ∧ ej )x (here xj denotes the coordinate of x that is not equal to zero). Evidently, D is smooth, since it is defined by linear vector fields X(x) = Ax.

16

1 The Orbit Theorem and Lie determined systems

It is easy to see that the distribution D in Example 1.8 is integrable. The maximal integral manifold through each point x0 = 0 is the sphere ||x|| = ||x0 ||. that are tangent to D. Each distribution D defines a family of vector fields D The following proposition relates integrable manifolds of distributions to the orbits of D. Proposition 1.11 If a distribution D is integrable then it is necessarily invo coincide with the maximal integral manifolds of D. lutive and the orbits of D Proposition 1.12 The Frobenius theorem Suppose that D is an involutive distribution such that the rank of D(x) is constant for all x ∈ M. Then D is inte coincide with the maximal integral manifolds of D. grable and the orbits of D Let us first prove the Frobenius theorem. through a point q. We will show that Proof Let N denote an orbit of D dim(N) = dimD(q). Suppose that F = {X1 , . . . , Xk } is an involutive family such that X1 (q), . . . , Xk (q) are linearly independent on an of vector fields in D open set U in M. It suffices to show that each tangent vector y˙ ◦ Adexp tXj (Xi ) belongs to the linear span of vectors in Fy for each y ∈ U, because then the tangent space Ty N is the linear span of y˙ ◦ Adexp tXj (Xi ) (expression (1.15)). Since F is involutive there exist smooth functions αij(m) on U such that [Xi , Xj ](y) =

k

(m)

αij (y)Xm (y)

m=1

for all y in U. If t denote the flow of Xi , then let σj (t) = Ad t (Xj (y)). Then, k k d (m) (m) σj (t) = t∗ [Xi , Xj ] −t (y) = αij ( −t y) t∗ Xm −t (y) = αij ( −t y)σm . dt m=1

m=1

This system is a linear system of equations k dσj = Ajm (t)σm (t), dt

(1.17)

m=1

(m)

where A(t) is the matrix with entries Ajm (t) = αij ( −t y). Let now φj (t) = k m=1 xjm (t)Xm (y), where xij (t) satisfy the following system of linear differential equations: dxjm = Aji (t)xim (t), xjm (0) = δjm , j = 1, . . . , k. dt k

i=1

1.4 Distributions and Lie determined systems

17

These solutions exist and are unique. Then k k k k dφj dxjm = Xm (y) = Aji (t)xim (t)Xm (y) = Aji (t)φi (t). dt dt m=1

m=1 i=1

i=1

So both φ1 (t), . . . , φk (t) and σ1 (t), . . . , σk (t) are the solutions of the same differential equation with the same initial conditions at t = 0, hence, must be equal to each other. This shows that each curve σj (t) belongs to the linear span of F(y). Let us now turn to the proof of Proposition 1.11. N denote Proof Suppose that N is a connected integral manifold of D. Let D to N. The orbits of D N partition N. Since Ty (N) = D(y) = the restriction of D N is open in N. Since D(y) for all points y in N it follows that each orbit of D N is connected there is only one orbit and therefore, N is equal to the orbit of N . This argument shows that N ⊆ O (x). The same argument shows that D D D N is tangent to its orbit, is involutive because the Lie algebra generated by D hence is tangent to N. through x is of the same dimension as It remains to show that the orbit of D N. Let q be any point of N. Since N is an integral manifold of D there exists and a neighborhood U of q such that a basis of vector fields X1 , . . . , Xk in D X1 (y) . . . , Xk (y) are linearly independent at each point y ∈ U. Let F(y) be the linear span of (X1 (y), . . . , Xk (y) at each point y ∈ U. Since F is involutive, the theorem of Frobenius (Proposition 1.11) applies. Therefore, N and the orbit through any point of N are of the same dimension and consequently the orbit through any point of N is the maximal integral manifold of D. Definition 1.13 A family of vector fields F is said to be Lie determined if the distribution defined by Lie(F) is integrable. Corollary 1.14 Any family F of vector fields, such that Liex F is of constant rank at all points of an orbit is Lie determined. In view of the preceding theorem, F is Lie determined if Liex (F) = Tx (Ox (F)) for all x in M, that is, if the tangent spaces of the orbits of F coincide with the evaluation of Lie(F) at the points of each orbit. The following cases of Lie determined systems are well known. Example 1.15 The distribution D = Lie(F) defined by the family of vector fields in Example 1.8 is involutive, but not of constant rank. It does not have integral manifolds anywhere along the y axis, and hence, is not integrable.

18

1 The Orbit Theorem and Lie determined systems

On the other hand, for analytic systems the structure of orbits is simple, thanks to this theorem. Proposition 1.16 The Hermann–Nagano theorem Suppose that M is an analytic manifold. Then any analytic family of vector fields F is Lie determined. Proof

When X and Y are analytic then ∞ k t x ◦ adk X(Y) x ◦ Adexp tX (Y) = k! k=1

for small t, and therefore, Adexp tX (Y) belongs to Lie(F). Consequently, the tangent spaces of an orbit of F coincide with the evaluation of Lie(F) at the points of the orbit. Example 1.17 Let F = {X, Y} be the family in R3 defined by two linear vector fields X(q) = Aq and Y(q) = Bq with A and B the following matrices: ⎞ ⎛ ⎞ ⎛ 0 −1 0 1 0 0 A = ⎝ 1 0 0 ⎠ , B = ⎝ 0 −1 0 ⎠ . 0 0 1 0 0 0 The Lie bracket of linear fields X(q) = Aq and Y(q) = Bq is linear and is given by the matrix C = AB − BA. It follows that Lie(F) is equal to the linear span of the matrices: A, B, C, D, where A and B are as defined above, and ⎞ ⎛ ⎞ ⎛ 0 1 0 1 0 0 C = ⎝ 1 0 0 ⎠ , D = ⎝ 0 −1 0 ⎠ . 0 0 0 0 0 0 An easy calculation shows that ⎧ 3 R for x2 + y2 = 0 and z = 0, ⎪ ⎪ ⎨ 2 for x2 + y2 = 0 and z = 0, R Lieq (F) = ⎪ R for x2 + y2 = 0 and z = 0, ⎪ ⎩ 0 for x2 + y2 + z2 = 0, ⎛ ⎞ x at a point q = ⎝ y ⎠ . Since the vector fields are analytic, F is Lie determined z by the Hermann–Nagano theorem. Indeed, the orbits of F are zero dimensional through the origin, one dimensional through non-zero points along the z-axis, two dimensional in the punctured plane z = 0 and x2 + y2 = 0, and threedimensional at all points x2 + y2 = 0 and z = 0.

2 Control systems: accessibility and controllability

2.1 Control systems and families of vector fields A control system on a manifold M is any differential system on M of the form dx = F(x(t), u(t)), dt

(2.1)

where u(t) = (u1 (t), . . . , um (t)). Functions u1 (t), . . . , um (t) are called controls. They are usually assumed to be bounded and measurable on compact intervals [t0 , t1 , T] and take values in a prescribed subset U of Rm . A trajectory is any absolutely continuous curve x(t) defined on some interval I = [t0 , t1 ] which satisfies (2.1) almost everywhere in [t0 , t1 ] for some control function u(t). We will assume that F is sufficiently regular that the Cauchy problem x(τ ) = x0 associated with the time-varying vector field Xu (t) = F(x, u(t)) admits a unique solution x(t) on some interval [t0 , t1 ] that varies smoothly relative to the initial point x0 . It is well known in the theory of differential equations [CL] that these properties will be fulfilled under the following conditions: 1. The vector field Xu (x) = F(x, u) is smooth for each u in U. 2. The mapping (x, u) → F(x, u) from M × U¯ into TM is continuous. 3. The mapping (x, u) ∈ M × U¯ → ∂F ∂x (x, u) is continuous in any choice of local coordinates. Under these conditions, each initial point x0 and each control function u(t) give rise to a unique trajectory x(t) that emanates from x0 at the initial time t = 0. Control theory is fundamentally concerned with the following questions: 1. Controllability: Given two states x0 and x1 in M is there a control u(t) that steers x0 to x1 in some (or a priori fixed) positive time T? 19

20

2 Control systems: accessibility and controllability

2. Motion planning: If the answer to the first question is affirmative, then what is the trajectory that provides the given transfer? 3. Optimality: What is an optimal way of steering x0 to x1 ? Alternatively, these questions could be reinterpreted as questions about the nature of points reachable by the trajectories of the system, and this shift in emphasis identifies the reachable sets as the basic objects of study associated with any control system. This chapter is devoted to the qualitative properties of the reachable sets. There are essentially three kinds of reachable sets from a given initial point x0 : points in M reachable in exactly T units of time, points reachable in at most T units of time, and points reachable in any positive time. These sets will be denoted respectively by A(x0 , T), A(x0 , ≤ T), and A(x0 ). Evidently, A(x0 , T) ⊆ A(x0 , ≤ T) ⊆ A(x0 ). The following notion is basic. Definition 2.1 System (2.1) is said to be controllable if A(x0 ) = M for each initial point x0 . It is said to be strongly controllable if A(x0 , T) = M for each x0 in M and all T > 0. This subject matter originated with linear systems in Rn , dx = Ax + ui (t)bi , dt m

(2.2)

i=1

where A is an n × n matrix and b1 , . . . , bm are vectors in Rn , and the following theorem: Proposition 2.2 If U = Rm , then (2.2) is strongly controllable if and only if the linear span of ∪k≥0 {Ak bi , i = 1, . . . , m} is equal to Rn , that is, if and only if the rank of the matrix C = (B AB · · · An−1 B),

(2.3)

where B denotes the matrix with columns b1 , . . . , bm , is equal to n. This proposition is remarkable in the sense that it characterizes the nature of the reachable sets in terms of an algebraic condition that completely bypasses the need to solve the differential equation. It turns out that this controllability criterion lends itself to Lie algebraic interpretations applicable to non-linear situations as well.

2.1 Control systems and families of vector fields

21

Before elaborating on this point further, let us first note that linear systems belong to a particular class of systems dx = X0 (x) + ui (t)Xi (x) dt m

(2.4)

i=1

for some choice of vector fields X0 , . . . , Xm on M. Such systems are known as control affine systems, where X0 is called the drift. The remaining vector fields X1 , . . . , Xm are called controlled vector fields. Apart from their interest for control theory, affine systems figure prominently in mechanics and geometry. For instance, variational problems involving geometric invariants of curves, such as problems involving curvature and torsion, are naturally formulated on the orthonormal frame bundle via the Serret–Frenet differential systems. Then the Serret–Frenet system can be regarded as an affine control system on the group of motions of Rn with the curvatures κ1 , κ2 , . . . , κn−1 playing the role of controls. We shall see later that affine systems also play an important part in the theory of mechanical tops. Any attempt to bridge control theory with mechanics and geometry, however, requires some preliminary remarks justifying the passage to absolutely continuous trajectories, a generalization that might seem totally alien to both a geometer and a physicist. For a control practitioner, the need for measurable controls is dictated by the solutions of optimal problems involving inequality constraints. Even the simplest optimal problems with bounds on controls result in chattering controls that take the solutions outside the realm of the usual Euler–Lagrange equation. For such a person measurable controls are indispensable. This bias towards measurable controls, however, resulted in an important side effect: it drew attention to piecewise constant controls and led to a new paradigm in which a control system was replaced by a family of vector fields. This paradigm shift identified a control system with a polysystem, a generalization of a dynamical system consisting of a single vector field, and uncovered the Lie bracket as a basic tool for studying its geometric properties. With these remarks in mind let us return to control system (2.1) and its family of vector fields F = {X : X(x) = F(x, u), u ∈ U}

(2.5)

generated by the constant controls. For each X in F, the semi-orbit {x : x =

Xt (x0 ), t ≥ 0} is contained in the reachable set A(x0 ). The set of points reachable by the piecewise constant controls from x0 at some positive time

22

2 Control systems: accessibility and controllability

T consist of points x which can be written as X

X

p−1 · · · ◦ Xt11 (x0 ), , t1 ≥ 0, . . . , tp ≥ 0 x = tpp ◦ tp−1

(2.6)

for some vector fields X1 , . . . , XP in F, or dually as functions xˆ = xˆ0 ◦ exp t1 X1 ◦ exp t2 X2 ◦ · ◦ exp tp Xp , t1 ≥ 0, t2 ≥ 0, . . . , , tp ≥ 0. (2.7) The concatenation of flows by the elements of F admits a geometric interpretation in the group of diffeomorphisms of M through the following objects. Definition 2.3 If F is a family of vector fields then G(F) denotes the group of diffeomorphisms generated by {exp tX : t ∈ R}, X ∈ F, and S(F) will denote the semigroup generated by {exp tX : t ≥ 0}, X ∈ F. Any diffeomorphism in G(F) is of the form

= exp t1 X1 ◦ exp t2 X2 ◦ · · · ◦ exp tp Xp for some vector fields X1 , . . . , Xk in F and numbers t1 , . . . , tk , while a diffeomorphism in S(F) is of the same form except that the numbers t1 , . . . , tp are all non-negative. The reachable sets A(x0 , T), AF (x0 , ≤ T) and A(x0 ) of F are defined completely analogously to the reachable sets of a control system, with obvious extensions to arbitrary families of vector fields and not just the family induced by (2.1). Of course, the reachable sets defined by F correspond to the points reachable by piecewise controls, only when F is induced by a system (2.1). The fact that the piecewise constant controls are dense in the class of bounded and measurable controls implies that the topological closure of the reachable sets of F and the control system F are the same. It follows that the reachable sets of F can be interpreted either as the orbits of the semi-group S(F), or as the semi-orbits of the group G(F) through the point x0 . The theory of accessiblity emanates from the following: Proposition 2.4 The Accessibility theorem Let F be any Lie determined family of vector fields on a manifold M, and let N denote an orbit of F through a point x0 . Then the interior of AF (x0 , ≤ T) relative to the orbit topology of N is dense in AF (x0 , ≤ T) for any T > 0. For a proof see [Jc]. Definition 2.5 A family of vector fields Fon M is said to have the accessibility property at x if AF (x, ≤ T) has a non-empty interior in M.

2.1 Control systems and families of vector fields

23

Definition 2.6 A family of vector fields on M is controllable if AF (x) = M for each x ∈ M. It is strongly controllable if AF (x, ≤ T) = M for all T > 0 and all x ∈ M. Corollary 2.7 A Lie determined family of vector fields F on a manifold M has the accessibility property at x if and only if Liex F = Tx M. Moreover, the accessibility property at a single point implies the accessibility property at all points of M whenever M is connected. Proof If F has the accessibility property at x then the orbit of F through x is of the same dimension as the ambient manifold M. Since F is Lie determined, Liex F = Tx M. Then F has the accessibility property at each point of the orbit of F by the Accessibility theorem (Proposition 2.4). The orbit through x is both open and closed, hence it is equal to M, whenever M is connected. Proposition 2.8 Let F be a family of vector fields on M such that Liex (F) = Tx M for all x ∈ M. If AF (x) is dense in M for some x, then AF (x) = M. Proof The fact that Liex (F) = Tx M for each x implies that each orbit of F is open in M. Since AF (x) is connected and dense implies that there is only one orbit of F. Let −F = {−X : X ∈ F}. The orbits of F and −F are the same. Then A−F (y, ≤ T) contains an open set in M for each y ∈ M and each T > 0 by the Accessibility theorem. To say that z ∈ A−F (y, ≤ T) is the same as saying that y ∈ AF (z, ≤ T). To show that each y ∈ M is reachable from x by F, let O be an open set contained in A−F (y, ≤ T) and let z ∈ O ∩ AF (x). Then y ∈ AF (z, ≤ T) and AF (z, ≤ T) ⊆ AF (x) imply that y ∈ AF (x). Example 2.9 Positive semi-orbits of the flow (eitθ z, eitφ w) are dense on the torus T 2 = {(z, w) : |z| = |w| = 1} whenever the ratio φθ is irrational. So it may happen that the reachable set is dense without being equal to the entire space. Let us now return to the linear systems and give a Lie theoretic proof for Proposition 2.2. Proof We want to show that (2.2) is strongly controllable if and only if the linear span of {bi , Abi , . . . , An−1 bi , i = 1, . . . , m} is equal to Rn [Kl]. This control system induces an affine distribution m ui Xi , u = (u1 , . . . , um ) ∈ Rm F = X0 + i=1

24

2 Control systems: accessibility and controllability

in Rn with X0 a linear field X0 (x) = Ax and each controlled vector field Xi constant and equal to bi . Then, adk X0 (Xi )(x) = Ak bi and [Xi , Xj ] = 0 for any i, j ≥ 1. It follows that Liex (F) is equal to the linear span of Ax and vectors Ak bi , i = 1, . . . , m, k = 0, 1, . . . . In particular, the evaluation of Lie(F) at the origin is equal to the linear span of Ak bi , i = 1, . . . , m, k = 0, 1, . . . . Since Ak b for k ≥ n is linearly dependent on {b, Ab, A2 b, . . . , An1 b}, it suffices to consider powers Ak bi , i = 1, . . . , m, k = 0, 1, . . . , n − 1. Therefore, Lie0 (F) is equal to the range space of the controllability matrix C. If (2.2) is strongly controllable, then the orbit of F through the origin is equal to Rn , and by the Hermann–Nagano theorem, the dimension of each orbit is the same as the rank of the Lie algebra. Hence, the rank of the controllability matrix must be n. Conversely, if the dimension of Lie0 (F) is equal to n then the reachable sets AF (0, ≤ T) and A−F (0, ≤ T) contain open sets in Rn for each T > 0 by Proposition 2.4. But each of these reachable sets are also linear subspaces of Rn , hence each of them is equal to Rn . So any point x0 can be steered to the origin in an arbitrarily short amount of time by a trajectory of (2.2) and also the origin can be steered to any point x in arbitrarily short amount of time by a trajectory of (2.2). Hence (2.2) is strongly controllable. Let us now consider the affine distributions m ui Xi , u = (u1 , . . . , um ) ∈ Rm F = X0 +

(2.8)

i=1

associated with a control affine system (2.3) consisting of analytic vector fields X0 , . . . , Xm . All analytic systems are Lie determined as a consequence of the Hermann–Nagano theorem. When X0 = 0 then F reduces to the distribution defined in Chapter 1. In that case, m = = Proposition 2.10 Control system dx i−1 ui Xi (x) with u dt (u1 , . . . , um ) ∈ Rm is strongly controllable if and only if Liex F = Tx M, for all x ∈ M. Proof

This proposition is a paraphrase of the Orbit theorem (Proposition 1.6).

For affine systems, with non-zero drift, there are no natural conditions that guarantee controllability, even when Lie(F is of full rank at all points of M. For instance, consider:

2.2 The Lie saturate

25

Example 2.11 dx = Ax + uBx dt

(2.9)

in M = R2 /(0) with A and B matrices and u(t) a scalar control. Let us first take 0 1 1 0 A= ,B= . 1 0 0 −1 It is easy to verify that Lie(F) consist of linear fields Cx with C an arbitrary 2 × 2 matrix with zero trace. Hence, Liex (F) = R2 for each x = 0. However, (2.9) is not controllable because d ((x1 (t)x2 (t)) = x12 (t) + x22 (t), dt and x1 (t)x2 (t) ≥ 0 whenever x1 (0)x2 (0) ≥ 0. On the other hand, (2.9) becomes 0 1 controllable on R2 /(0) when A is replaced by A = . For then, −1 0 the flow of X0 = Ax is periodic. Therefore, {±X0 + uX1 } and {X0 + uX1 } have the same reachable sets, and hence, the reachable sets coincide with the orbits. Since Liex F = R2 at all points x = 0 there is only one orbit of F, and therefore, (2.9) is controllable.

2.2 The Lie saturate Even though there is no general theory that guarantees controllability, there are some controllability criteria that can be applied in certain situations to obtain positive results. One such criterion is based on the notion of the Lie saturate, It is defined as follows. Definition 2.12 The Lie saturate of a family of vector fields F is the largest set Fˆ in Lie(F), in the sense of set inclusion, such that A¯ F (x) = A¯ Fˆ (x) for each x ∈ M. The Lie saturate will be denoted by LS(F). Its significance is described by the following: Proposition 2.13 The controllability criterion A Lie determined family F is controllable on M if and only if LS(F) = Lie(F).

26

2 Control systems: accessibility and controllability

Proof If LS(F) = Lie(F) then LS(F) is controllable. Since ALS (F ) (x) is contained in the closure of AF (x), AF (x) is dense in M. But then F is controllable by Proposition 3. In general, there is no constructive procedure to go from F to its Lie saturate. Nevertheless, there are some constructive steps that can be taken to enlarge a given family of vector fields without altering the closure of its reachable sets. The first step involves a topological closure of the family, which, in turn, requires a topology on the space of vector fields. The most convenient topology V ∞ (M) is that defined by the smooth uniform convergence on compact subsets of M. This means that a sequence of vector fields X (m) converges to a vector field X if for any chart (U, φ) and any compact set C in M (m)

∂ k Xi

∂x1i1 · · · ∂xnin

→

∂ k Xi ∂x1i1 · · · ∂xnin

uniformly on φ(C) in Rn for each k and each multi-index i1 + · · · + ik = n. (m) (m) Here, (X1 , . . . , Xn ) and (X1 , . . . , Xn ) denote the coordinate vectors of X m and X. The following theorem is a paraphrase of well-known facts from the theory of differential equations (see also [Jc]) Proposition 2.14 Suppose that a sequence of complete vector fields X m converges to a vector field X and suppose that x(t) is an integral curve of X defined on an interval [0, T]. Let yk be any sequence of points in M that converges to x(0) and let xk(m) (t) denote the integral curves of X m with xkm (0) = yk . Then there exists an integer k0 such that for k > k0 each curve xkm (t) is defined on [0, T] and converges uniformly to x(t) on [0, T]. This proposition easily implies Corollary 2.15

Let F¯ denote the topological closure of F. Then, A¯ F¯ (x) = A¯ F (x)

for each x ∈ M, where A¯ denotes the topological closure of the set A. Corollary 2.16

The Lie saturate is a closed family of vector fields.

The next step makes use of weak limits in the space of controls according to the following: Proposition 2.17 Let u(k) be a sequence in L∞ ([0, T], Rm ) that converges (k) weakly to u∞ in L∞ ([0, T], Rm ), and let Zk (t) = X0 + m i=1 ui (t)Xi and m ∞ Z(t) = X0 + i=1 ui (t)Xi . Suppose that x(t) is an integral curve of Z(t)

2.2 The Lie saturate

27

defined on the interval [0, T]. Then for all sufficiently large k, the integral curves x(k) (t) of Zk (t) with x(k) (0) = x(0) are defined on [0, T], and converge uniformly to x(t) on [0, T]. For a proof see [Jc, p. 118]. We just remind the reader that a sequence of functions u(k) in a Banach space B converges weakly to a point u in B if L(u(k) ) converges to L(u) for every linear function L in the dual of B. L∞ ([t0 , t1 ], Rm ) is its In the case that B is equal to L1 ([t0 , t1 ], Rm ), then m t1 (k) dual, and weak convergence means that i=1 t0 ui (t)fi (t) dt converges m t1 to i=1 t0 ui (t)fi (t) dt for every choice of functions f = (f1 , . . . , fm ) in L∞ ([t0 , t1 ], Rm ). Proposition 2.18 Let F = {X, Y} where X and Y are smooth vector fields. Let x(t) denote an integral curve of the convex combination λX + (1 − λ)Y. If x(t) is defined on an interval [0, T], then x(T) ∈ cl(AF (x(0), T)). Consequently, xˆ ◦ exp (λX + (1 − λ)Y)(T) ∈ cl(A(x, T)) for each x ∈ M for which xˆ ◦ exp (λX + (1 − λ)Y)(T) is defined. Proof Let 0 = t0 < t! < · · · < t2n = t be an equidistant partition of the interval [0, T]. If Ik denotes the interval (tk−1 , tk ], then let uk denote the characteristic function of the set I2 ∪ I4 ∪ · · · ∪ I2n and let vk denote the characteristic function of the set I1 ∪ I3 ∪ · · · ∪ I2n−1 . Then T T 1 T 1 T lim uk (τ )f (τ ) dτ = f (τ ) dτ and lim vk (τ )f (τ ) dτ = f (τ ) dτ k→∞ 0 k→∞ 0 2 0 2 0 for every function f in L∞ ([0, t], R). Therefore, both sequences of functions converge weakly to the constant function u = v = 12 . Let zn denotes the integral curve of Zn (t) = 2λun (t)X + 2(1 − λ)(1 − un (t))Y that satisfies zn (0) = x(0). It follows that zn (T) ∈ cl(AF (x(0), T) for each n, and that Zn converges weakly to Z = λX + (1 − λ)Y. But then zn converges uniformly to x(t) on the interval [0, T]. Hence, x(T) ∈ cl(AF (x(0), T)). Corollary 2.19 cl(AF (x, T)) = cl(Ach(F ) (x, T)), where ch((F) denotes the convex hull of F. Corollary 2.20 cl(AF (x) = cl(Aco(F ) (x), where co(F) denotes the the positive convex cone generated by F. Proof The reachable sets AF (x) are invariant under positive reparametrizations of vector fields in F. In addition, the reachable sets are unaltered by symmetries. The following notion is basic.

28

2 Control systems: accessibility and controllability

Definition 2.21 A diffeomorphism is a normalizer for F if

−1 (AF ( (x)) ⊆ A¯ F (x) and ( −1 ∗ F )(x) ⊆ Liex (F) for all x ∈ M. Definition 2.22 For each diffeomorphism and each vector field X, (X) will denote the vector field ∗ (X) ◦ −1 , and (F) = { (X) : X ∈ F}. Then: Proposition 2.23 family F. Then

Suppose that N (F) denotes the set of normalizers for a { (X) : X ∈ F, ∈ N (F)} ⊆ LS(F).

The proof is obvious. It will be convenient for future reference to assemble all these facts into the following: Proposition 2.24 The Lie saturate LS(F) is a closed set of vector fields, invariant under its normalizer, and also invariant under the following enlargements: 1. If Y1 , Y2 , . . . , Yp are any set of vector fields in LS(F) then the positive affine hull {α1 Y1 + · · · + αp Yp : αi ≥ 0, i = 1, . . . , p} is also contained in LS(F). 2. If V is a vector space of vector fields in LS(F), then Lie(V) is in LS(F). 3. If ±Y is in LS(F), then ( Yλ )∗ X Y−λ is in LS(F) for any λ ∈ R, and any X ∈ LS(F). Here, Yλ ∗ denotes its tangent map of the flow Yλ : λ ∈ R induced by Y. Let us now return to the affine control systems (2.4) with these symmetry tools at our disposal. It follows that α(X0 + ui Xi ) ∈ LS(F) for each α > 0 by 1 in Proposition 2.24. Then, λXi = limα→0 (α(X0 + λ αu Xi ) is in LS(F) for each number λ. But then the vector space generated by the controlled vector fields X1 , . . . , Xm is in the Lie saturate of F by 2 in Proposition 2.24. Hence, {Adexp λi Xi (X0 ) is in the Lie saturate of F for each λi . We will be able to show that in some situations −X0 is in the positive convex cone spanned {Adexp λi Xi (X0 ), (λ1 , . . . , λm ) ∈ Rm , i = 1, . . . , m}, which then implies controllability whenever Lie(F) is of full rank on M. We will come back to this theory when discussing the Maximum Principle. In the meantime, we will first integrate this material with other geometric structures, with Lie groups at the core.

3 Lie groups and homogeneous spaces

The mathematical formalism developed so far has natural applications to the theory of Lie groups. To highlight these contributions, and also to better orient the subject for further use, it seems best to start at the beginning and develop the relevant concepts in a self contained manner. Definition 3.1 A group G is called a real Lie group if G is a real analytic manifold and the group operations (x, y) → xy and x → x−1 are real analytic, the first as a mapping from G × G into G and the second as a mapping from G into G. A group G is called a complex Lie group if G is a complex manifold and the group operations are holomorphic. We will assume that all Lie groups are real unless explicitly stated otherwise. Prototypical Example: the general linear group The group GL(E) of all linear automorphisms of a finite-dimensional real vector space E is a Lie group. Proof Each basis e1 , . . . , en in E identifies automorphisms T in GL(E) with n×n matrices X(T) = (xij (T)) defined by Tei = nj=1 xji ej . Then the entries of 2

the matrix X(T) provide a correspondence between points of Rn and elements of GL(E). If φ denotes this correspondence, then GL(E) is topologized by the finest topology such that φ is a homeomorphism. Under this topology the 2 coordinate neighborhoods U are the inverse images of open sets in Rn defined by Det(X(T) = 0. Matrices that correspond to different bases are conjugate to each other, hence are defined by analytic (rational) functions.

29

30

3 Lie groups and homogeneous spaces

The composition of elements in GL(E) corresponds to the products of matrices. Since the entries of the product depend polynomially on the entries of the matrices, the group multiplication is analytic. Similarly the entries of the inverse of a matrix are rational functions of the entries of the matrix, hence analytic. It is a common practice to denote GL(Rn ) by GLn (R), a convention that will be adopted in this text. Any Lie group G has two distinguished groups of diffeomorphisms, the group of left translations {Lg : g ∈ G} and the group of right translations {Rg : g ∈ G}. The left translations are defined by Lg (x) = gx and the right translations by Rg (x) = xg. We shall follow the convention established earlier and use Fˆ to denote the automorphism on C∞ (M) induced by a diffeomorphism F. Definition 3.2 Vector field X on a Lie group G is called left-invariant if Lˆ g−1 ◦ X ◦ Lˆ g = X for any g ∈ G. Right-invariant vector fields are defined by Rˆ −1 g ◦ ˆ X ◦ Rg = X. It follows that the left (right)-invariant vector fields satisfy (Lg )∗ X(x) = X(Lg (x)) = X(gx) ((Rg )∗ X(x) = X(Rg (x)) = X(xg) ), for all g ∈ G. Since X(g) = (Lg )∗ X(e), both left- and right-invariant vector fields are determined by their values at the group identity e. Moreover, each tangent vector A in Te (G) determines a unique left (right)-invariant vector field XA defined by XA (g) = (Lg )∗ A (respectively, XA (g) = (Rg )∗ A). Proposition 3.3 The Lie bracket of left (right) invariant vector fields is left (right) invariant vector field. Proof ˆ Fˆ −1 ◦ Y ◦ F] ˆ = Fˆ −1 ◦ [X, Y] ◦ F. ˆ [Fˆ −1 ◦ X ◦ F,

We will use Fl and Fr to designate the Lie algebras of left (right) invariant vector fields on G. Proposition 3.4

Let F : G → G denote the diffeomorphism F(x) = x−1 Then, Xl → Fˆ −1 ◦ Xr ◦ Fˆ

is an isomorphism from Fr onto Fl . Moreover, if A = Xr (e) then (F ◦ Xr ◦ F −1 )(e) = −A.

3.1 The Lie algebra and the exponential map

Proof

31

ˆ Then, Let Y = Fˆ −1 ◦ Xr ◦ F.

Lˆ g−1 ◦ Y ◦ Lˆ g = Lˆ g−1 Fˆ −1 ◦ Xr ◦ Fˆ Lˆ g = Fˆ −1 Rˆ g ◦ Xr ◦ Rˆ g Fˆ = Fˆ −1 ◦ Xr ◦ Fˆ = Y because Lˆ g−1 Fˆ −1 = Fˆ −1 Rˆ g . Therefore Y is left-invariant. To prove the second statement let σ () denote a curve in G such d −1 ()|=0 = − A, and therefore that σ (0) = e and dσ d (0) = A. Then d σ F∗ (A) = −A. It follows from the preceeding proposition that both of these algebras are finite dimensional and that their common dimension is equal to the dimension of G.

3.1 The Lie algebra and the exponential map Let g denote the tangent space of G at the group identity e. Invariant vector fields on G induce a Lie bracket on g by the following rule. Definition 3.5 If A and B are elements of the Lie algebra g then their Lie bracket [A, B] is defined by [A, B] = [XA , XB ](e) where XA and XB denote the left-invariant vector fields such that XA (e) = A and XB (e) = B. Remark 3.6 The Lie bracket could be defined also by right-invariant vector fields in which case the Lie bracket [A, B] would be the negative of the one defined by the left-invariant vector fields. The Jacobi identity on g follows from the Jacobi identity in the Lie algebra of vector fields. In fact, [XA , [XB , XC ]](e) + [XC , [XA , XB ]](e) + [XB , [XC , XA ]](e) = 0 reduces to [A, [B, C]] + [C, [A, B]] + [B, [C, A]] = 0. It follows that g is a Lie algebra under the vector addition and the Lie bracket defined above. It is called the Lie algebra of G and will be denoted by g. Proposition 3.7 The Lie algebra gln (R) of GLn (R) is equal to Mn (R), the vector space of all n × n matrices with real entries, with the Lie bracket given by [A, B] = BA − AB.

32

3 Lie groups and homogeneous spaces

Proof Since GLn (R) ⊂ Mn (R), gln (R) ⊆ Mn (R). Any matrix A in Mn (R) defines a one-parameter group of left translations t (g) = getA . Its infinitesimal generator is the left-invariant vector field X(g) = gA. This shows that gln (R) = Mn (R). We showed in (1.10) in Chapter 1 that the Lie bracket conforms to the following formula: ∂2 exp −tX ◦ exp sY ◦ exp tX|t=s=0 = [X, Y]. ∂t∂s So if X(g) = gA and Y(g) = gB are any left-invariant fields then according to the above formula ∂ 2 −tA sB tA e e e |t=s=0 = [X, Y](e) = [A, B] ∂t∂s The above yields [A, B] = (BA − AB). Remark 3.8 The commutator [A, B] of matrices A and B is usually denoted by AB − BA rather than its negative. Our choice is dictated by the choice of the sign in the definition of the Lie bracket. Each choice is natural in some situations and less natural in others. For instance, in the case of linear fields X(x) = Ax and Y(x) = Bx in Rn the choice ]X, Y] = Y ◦ X − X ◦ Y implies that [X, Y](x) = (AB − BA)x, which seems more natural than its negative. It is imperative to maintain the conventions initially chosen, otherwise confusion is bound to follow. Proposition 3.9 Each left (right)-invariant vector field X is analytic and complete. The flow : G × R → G of a left-invariant vector field X(g) is given by

(g, t) = ge(t), where e(t) is a one-parameter abelian subgroup of G. Proof The fact that invariant vector fields are analytic follows directly from the analyticity of the group multiplication. To show completeness, let X be a left-invariant vector field and let ξ(t) be the left-translate Lg (σ (t)) of an integral curve σ (t) of X. Then, dσ d ξ(t) = (Lg )∗ = (Lg )∗ X ◦ σ (t) = X ◦ Lg (σ (t)) = X ◦ ξ(t). dt dt Therefore ξ(t) is also an integral curve of X. The fact that the integral curves of X are invariant under the left translations has several direct consequences. Firstly, it implies that the integral curve through any point g is equal to the left-multiple ge(t) of the integral curve e(t) through the group identity e. Secondly, it implies that the integral curves are defined for all t for the following reasons:

3.1 The Lie algebra and the exponential map

33

Let I = [−a, a] denote an interval such that e(t) is defined for all t ∈ I. Such an interval exists as a consequence of the local existence of solutions of differential equations. Let e+ denote the positive escape time for e(t). Suppose that e+ < ∞. Let T be any number such that 0 < e+ − T < a2 . Then, e(t) , t ≤ T σ (t) = e(T)e(t) , t ∈ I is an integral curve of X that is defined on the interval [0, e+ + a2 ]. By the uniqueness of solutions of differential equations, e(t) = σ (t), contradicting the finiteness of e+ . The last part follows from the general property t+s (e) =

t ◦ s (e) = s ◦ t (e). If X is a left-invariant vector field X(g) = (Lg )∗ A then its integral curve through the group identity will be denoted by eAt . Since {eAt : t ∈ R} is a one-parameter abelian subgroup of G, eAt agrees with the exponential of a matrix when G is a linear group of matrices, that is, t2 2 tn A + · · · + An + · · · . (3.1) 2 n! For a general Lie group G the above formula is to be interpreted functionally in terms of the convergent series eAt = I + At +

eˆ ◦ exp tX = e + e ◦ tX + . . . e ◦

tn n X + ··· . n!

Since e ◦ X n f = e ◦ An ( f ) for all n we may write t2 2 tn A + · · · + An + · · · , 2 n! with the understanding that both sides of the formula act on functions. e ◦ exp tX = etA = I + At +

Definition 3.10 If h is a Lie subalgebra of g then exp h = {eA : A ∈ h} is called the exponential of h. Proposition 3.11 (a) If g = k ⊕ p then (exp k)(exp p) contains a neighborhood U of the identity in G. (b) There is an open neighborhood U of e such that U ⊆ exp g. (c) If G is compact and connected then exp g = G. Proof Let A1 , . . . , Ak denote a basis in k, and let B1 , . . . , Bp denote a basis in p. The mapping F(s1 , . . . , sp , t1 , . . . , tk ) = eA(t1 ,...,tk ) eB(s1 ,...,sp )

34

3 Lie groups and homogeneous spaces

p where A(t1 , . . . , tk ) = ki=1 ti Ai and B(s1 . . . , sp ) = i=1 si Bi satisfies ∂F ∂F s=t=0 = Ai , i = 1, . . . , k and s=t=0 = Bi , i = 1, . . . , p ∂ti ∂si It follows from the inverse function theorem that the range of F covers a neighborhood U of F(0, 0) = e. Therefore, parts (a) and (b) follow. The proof of Part (c) will be deferred to other sections of the book. −1 a Example 3.12 Let B = where a = 0. Then there is no matrix A 0 −1 1

1

such that eA = B for the following reason: e 2 A e 2 A = eA , hence eA is a square of a matrix. It is easy to checkthat B is not a square of any matrix. However, if 0 π a = 0, then A = is such that eA = B. −π 0

3.2 Lie subgroups Definition 3.13 A subgroup H of a Lie group G is a Lie subgroup of G if H is a submanifold of G and the inclusion map i : H → G is a group homomorphism. Proposition 3.14 Let F denote any family of left-invariant vector fields on a Lie group G. Then the orbit of F through the group identity is a Lie subgroup H of G. The orbit of F through any other point g is the left coset gH. The Lie algebra h of H is equal to the evaluation Liee (F). Proof This proposition is a paraphrase of the Orbit theorem for families of left-invariant vector fields. When F is a family of left-invariant vector fields, then Lie(F) is a finite-dimensional algebra of left-invariant vector fields in G. Since the left-invariant vector fields are analytic, the tangent space at e of the orbit OF (e) is equal to h = Liee (F). If H = OF (e), then the points of H are of the form g = eAm tm eAm−1 tm−1 · · · eA1 t1 where A1 , . . . , Am are in h and where t1 , . . . , tm are arbitrary real numbers. Evidently, H is a Lie subgroup of G and OF (g) = gOF (e). Corollary 3.15 Let G be a Lie group with its Lie algebra g. Each subalgebra h of g is the Lie algebra of a connected Lie subgroup H of G. Proof Let F denote the family of left-invariant vector fields with values h at e. Then H is equal to the orbit of F through e.

3.2 Lie subgroups

35

Proposition 3.16 Any closed subgroup H of a Lie group G is a Lie subgroup of G with its topology equal to the relative topology inherited from G. Proof Let h = {A : exp tA ∈ H}. For each element A ∈ h, let XA denote the associated left-invariant vector field and let F denote the family vector fields XA with A ∈ h. Then the orbit OF (e) of F is a Lie subgroup in G by Proposition 3.14. Let K denote the orbit OF (e). Evidently, K ⊆ H. Let k denote the Lie algebra of K. For each element A ∈ k, exp tA belongs to K hence, belongs to H for all t. Since each A in h belongs to k, Lie(h) also belongs to k. Therefore, Lie(h) = h = k. It remains to show that K is open in H when H is topologized by the relative topology inherited from G. Let H0 denote the connected component of H through the group identity. Since K is connected, K ⊆ H0 . We will now show that K is open in H0 from which it would follow that K = H0 . Recall that the orbit topology is finer than the relative topology inherited from G. It suffices to show that for any neighborhood O of the identity, open in its orbit topology of K, there exists an open neighborhood U of the identity in G such that U ∩ H is contained in O. For then, every neighborhood of e that is open in the orbit topology would be also open in the relative topology, and the same would be true for any other open set in K since it is a group translate of an open neighborhood of the group identity. Let p denote any subspace of g that is transversal to h, that is g = h ⊕ p. If V in h and W in p are any neighborhoods of the origin then, according to Proposition 3.11, (exp V)(exp W) contains an open neighborhood of the identity in G. Moreover, V can be chosen so that exp V ⊆ O. Suppose that exp V does not contain any sets of the form H ∩ U with U a neighborhood of e in G. Let Vn and Wn denote sequences of open sets in V and W that shrink to 0. Then there exist sequences of points hn in Vn and pn in Wn such that pn = 0 and exp hn exp pn belongs to H for all n. Since exp hn belongs to H, exp pn belongs to H as well. To get the contradiction we will follow the argument used in [Ad] and assume that p is equipped with a norm || , || (any vector space can be equipped with a Euclidean norm). The sequence ||ppnn || contains a convergent subsequence. There is no loss in generality if we assume that the sequence pn ||pn || itself converges to a point p with ||p|| = 1. The fact that {pn } converges to 0 implies that for any t = 0 there exist integers mn such that 1 ||pn || 1 < . < mn t mn + 1

36

3 Lie groups and homogeneous spaces

Evidently mn → ∞ and hence, limn→∞ ||pn ||mn = t. But then (exp pn )mn = exp mn pn = exp mn ||pn ||

pn → exp tp. ||pn ||

This means that exp tp belongs to H, since (exp pn )mn is in H, and H is closed. This contradicts the fact that p is in p. Therefore, exp V contains a relatively open set H ∩ U. Since exp V ⊆ O, H ∩ U ⊆ O, and hence, the orbit topology on K coincides with the relative topology inherited from G. This implies that K is both open and closed in H0 , and consequently it must be equal to H0 . But then the connected component through any other point g ∈ H is equal to gK, and our proof is finished. As a corollary of the preceeding proposition the following linear groups are Lie subgroups of the group GLn (R): 1. The special linear group SLn (R). Let SL(Rn ) denote the group of all volume preserving linear automorphism of the Euclidean space Rn and let Sln (R) be the subgroup of Gln (R) of non-singular matrices T such that Det(T) = 1. Any oriented basis e1 , . . . , en sets up an isomorphism between SL(Rn ) and SLn (R). Since Det(eAt ) = eTr(A)t , it follows that the Lie algebra sln (R) of Sln (R) is equal to the vector space of all n × n matrices of zero trace. 2. The orthogonal group On (R). Assume that E is a Euclidean vector space with x, y denoting its Euclidean scalar product. The orthogonal group denoted by O(E) is a subgroup of all linear automorphisms T of E that satisfy Tx, Ty = x, y for all x and y in E. Any orthonormal basis in E sets up an isomorphism between O(E) and the orthogonal group On (R) consisting of n×n matrices T with real entries that satisfy T −1 = T ∗ , where T ∗ denotes the matrix transpose of T. Each matrix T ∈ On (R) satisfies 1 = Det(TT −1 ) = Det(TT ∗ ) = Det(T)Det(T ∗ ) = Det2 (T) The subgroup of matrices with determinant equal to 1 is called the special group of rotations and is denoted by SOn (R). It is equal to the connected component of On (R) through the identity matrix I. The Lie algebra of On (R) is denoted by son (R) and consists of the n × n skew-symmetric matrices. Therefore, the dimension of O(E) is equal to n(n−1) 2 . 3. The Lorentzian group SO(p, q). The Lorentzian group, denoted by SO(p, q) is the subgroup of GLn (R) that leaves the Lorentzian quadratic p form x, y = i xi pi − ni=p+1 xi pi invariant. If Jp,q is the matrix

3.2 Lie subgroups

Ip 0

0 Iq

37

, q = n − p,

where Ip and Iq denote p- and q-dimensional identity matrices, then T belongs to SO(p, q) if and only if T ∗ JT = J. The Lie algebra of SO(p, q) is denoted by so(p, q). Matrices A belong to so(p, q) if and only if A∗ J + JA = 0. An easy calculation shows that so(p, q) consist of the matrices a b A= bT c with a and c skew-symmetric p × p and q × q matrices and b an arbitrary p × q matrix. Here, bT denotes the matrix transpose of b. 4. The symplectic group Spn . The quadratic form [ , ] in R2n defined by [(x, p), (y, q)] = ni=1 qi xi − pi yi is called symplectic. The symplectic form can be related to the Euclidean quadratic form ( , ) via the formula 0 I [(x, p), (y, q)] = ((x, p), J(y, q)), J = . −I 0 The subgroup of Gl2n (R) that leaves the symplectic form invariant is called symplectic and is denoted by Spn . It follows that a matrix M is in Spn if and only if M T JM = J. A matrix A belongs to the Lie algebra spn (R) if and only if eAt ∈ Spn , or equivalently, if and only if AT J + JA = 0. Therefore, matrices A in the Lie algebra spn have the following block form: a b , A= c −aT with a an arbitrary n × n matrix and b and c symmetric n × n matrices. Hence the dimension of Spn is equal to n2 + n(n + 1) = 2n2 + n. 5. The unitary group Un . Let E denote a dimensional complex vector space with a Hermitian inner product z, w. The unitary group U(E) is the subgroup of complex linear automorphisms T that satisfy Tz, Tw = z, w for all z, w in E. Every orthonormal basis sets up an isomorphism between U(E) and the space of n × n matrices T with complex entries that satisfy T −1 = T ∗ , where T ∗ denotes the Hermitian transpose of T. That is, the entries of T −1 are given by T¯ ji , where Tij denote the entries of T and where a¯ stands for the complex conjugate of a. Then it follows that

38

3 Lie groups and homogeneous spaces 1 = Det(TT −1 ) = Det(TT ∗ ) = Det(T)Det(T ∗ ) ¯ = |Det(T)|2 . = Det(T)Det(T) Therefore, the determinant of each member of the group lies on the unit circle, which accounts for its name. The group of complex n × n matrices whose inverses are equal to their Hermitian transposes is denoted by Un . The Lie algebra un consists of skew-Hermitian matrices A, that is, matrices A such that A∗ = −A with A∗ denoting the Hermitian transpose of A. Any such matrix can be written as A = B+iC with B equal to the real part of A and C equal to the imaginary part of A. It follows that B is skew-symmetric and C is symmetric. Note that un is a real Lie algebra and not a complex Lie algebra (iB is Hermitian, for any skew-Hermitian matrix B, hence does not belong to un ). The dimension + n(n+1) = n2 . of un is equal to n(n−1) 2 2 Elements of Un can be also represented by 2n × 2n matrices with real entries. This representation, denoted by U2n (R) is obtained by treating Cn as R2n . In fact, e1 , . . . , en , ie1 , . . . , ien is a basis for the real vector space obtained from E by restricting the scalars to real numbers. Then each point v of E is represented by 2n coordinates (x1 , . . . , xn , p1 . . . , pn ) written more compactly as (x, p). Then the Hermitian product v, w of v = (x, p) and w = (y, q) is equal to v, w = (u, v) + i[v, w], with ( , ) the Euclidean quadratic form and [ , ] the symplectic form. Each complex linear transformation T is real linear, and if T preserves the Hermitian product on E then its matrix Tˆ relative to the basis e1 , . . . , en , ie1 , . . . , ien preserves both of the above quadratic forms. It then follows that U2n (R) = O2n (R) ∩ Sp2n .

The subgroup of Un of elements with determinant equal to 1 is denoted by SUn and is called the special unitary group. Its Lie algebra is denoted by sun . The case n = 2 is particularly special since then SU2 can be 3 identified with the sphere S . The identification is simple: each matrix z w in SU2 is identified with a point (z, w) on S3 via the −w¯ z¯ determinant |z|2 + |w|2 = 1. 6. Semi-direct products: the Euclidean group of motions SEn (R). Any Lie group K that acts linearly on a finite-dimensional topological vector space

3.3 Families of left-invariant vector fields and accessibility

39

V defines the semi-direct product G = V K consisting of points in V × K and the group operation (v, S)(w, T) = (v + Sw, ST) for all (v, S) and (w, T) in V × K. It follows that G is a group with the group identity e equal to (0, I), where I is the identity in K, and the group inverse (v, S)−1 = (−S−1 v, S−1 ). Topologized by the product topology of V × K, G becomes a Lie group whose dimension is equal to dim(V) + dim(K). Each semi direct product G = V K acts on V by (v, S)(x) = v + Sx. The semi-direct product of a Euclidean space En with the group of rotations O(En ) is called the group of motions of En because its action preserves the n each Euclidean distance. An orthonormal basis e1 , . . . , en in ⎞ ⎛ E identifies x1 ⎟ ⎜ point x of En with a column vector of coordinates ⎝ ... ⎠ in Rn , and xn # $ with the matrix Rij in On (R) via the formula

identifies each R ∈ n Rij ej for i = 1, . . . , n. Rei = O(En )

j=1

The correspondence between elements (x, R) of En O(En ) and the column vectors in Rn and matrices in On (R) defines an isomorphism between En O(En ) and the semi-direct product Rn On (R). This n O(En ) into GL isomorphism defines an embedding of E n+1 (R) by 1 0 identifying (x, R)) with a matrix g = in GLn+1 (R). The x R connected component of En O(En ) that contains the group identity is equal to En SO(En ) and is denoted by SEn (R).

3.3 Families of left-invariant vector fields and accessibility Suppose now that F is a family of either right- or left-invariant vector fields in a Lie group G. The evaluation of F at the identity defines a set = {X(e) : X ∈ F} in the Lie algebra g, which is called the trace of F. We will use Lie() to denote the Lie algebra generated by . It follows that Liee (F) = Lie(). Then the orbit of F through the group identity e consists of elements g that can be written as g = ett Ap · · · et2 A2 et1 A1 , for some elements A1 , A2 , . . . , Ap in and real numbers t1 , t2 , . . . , tp .

(3.2)

40

3 Lie groups and homogeneous spaces

This orbit is a connected Lie subgroup K whose Lie algebra is Lie(). The orbit of F through any other point g0 is either the right-translate Kg0 or the left-translate g0 K depending whether F is right- or left-invariant. It follows that there is one orbit of F whenever Lie() = g, and G is connected. For the purposes of controllabiity, however, it is the action of the semigroup S(F) generated by {exp tX : t ≥ 0, X ∈ F} that is relevant. The reachable set AF (e) is the semigroup S() consisting of points g ∈ G that can be written in the form (3.2) but with t1 ≥ 0, t2 ≥ 0, . . . , tp ≥ 0 and the reachable set AF (g0 ) is either the right or the left translate of g0 by S() (depending whether F is right- or left-invariant). It follows that F is controllable if and only if S() = G. This means that Lie() = g is a necessary condition of controllability. This question of sufficiency can be paraphrased in slightly more geometric terms by passing to the Lie saturate of F, which means that could be replaced by LS() = LSe (F). On the basis of Proposition 2.23 in Chapter 2 we could assume that LS() enjoys the following properties: 1. + ⊆ LS() and R+ ⊆ LS(), and 2. If V is any vector subspace of LS() then ead(V) () ⊆ LS(). This property implies that the largest vector subspace in LS() is a Lie subalgebra of g. Subsets of Lie algebras with these properties are called wedges in the literature on Lie semi-groups [Hg]. The largest vector subpace in a wedge is called an edge. The reachable sets e are the semigroups in G generated by {eX : X ∈ }. These semigroups will be denoted by S(). It follows that controllability questions of families of invariant vector fields reduce to finding conditions on such that S() = G. The simplest cases are given by our next theorem. Proposition 3.17 If G is either a compact and connected Lie group, or if G is a semi-direct product of a vector space V and a compact and connected Lie group K that admits no fixed non-zero points in V, then S() = G for any subset such that Lie() = g. We will only prove the compact case. For the semi-direct products the reader is referred to [BJ]. Proof Assume first that G is compact and connected. The positive limit set

+ (A) of an element A ∈ g is defined to be the closure of the set of all points X in G such that etn A → X for some sequence {tn } such that tn → ∞. The positive limit set is non-empty when G is compact.

3.4 Homogeneous spaces

41

Let h = limtn →∞ etn A . There is no loss in generality if {tn } is chosen so that tn+1 − tn → ∞. Then, lim e(tn+1 −tn ))A = lim etn+1 A lim etn A = h−1 h = e.

n→∞

n→∞

n→∞

Hence, e ∈ + (A). It follows that e−tA = e−tA e = e−tA limtn →∞ etn A = limn→∞ e(tn −t) A. Therefore, the closure of exp contains exp ±, and the latter is equal to G whenever Lie() = g. Hence, S() is dense in G and therefore equal to G by Proposition 2.8 in Chapter 2.

3.4 Homogeneous spaces A Lie group G is said to act on a manifold M if there exists a smooth mapping φ : G × M → M such that: 1. φ(e, x) = x for all x ∈ M, where e denotes the group identity in G, and 2. φ(g2 g1 , x) = φ(g2 , φ(g1 , x) for all g1 , g2 in G and all x ∈ M. Alternatively, group actions may be considered as the groups of transformations { g , g ∈ G} where g (x) = φ(g, x), for x ∈ M, subject to ( g )−1 =

g− 1 and g · h = gh for all g and h in G. The set { g (x) : g ∈ G} is called the orbit through x. The group G is said to act transitively on M if there is only one orbit, i.e., if for any pair of points x and y in M there is an element g in G such that g (x) = y. A manifold M together with a transitive action by a Lie group G is called homogeneous. Homogeneous spaces can be identified with the orbit through an arbitrary point x in which case M can be considered as the quotient spaces G/K where K is the isotropy group of the base point x, i.e., K = {g ∈ G : g (x) = x}. The base space M is termed homogeneous because the isotropy groups corresponding to different base points are conjugate. A Lie group G acts on any subgroup K by either the right action (g, h) → hg−1 or by the left action (g, h) → gh. So homogenous spaces can be identified with the quotients G/K with K a closed subgroup of G.

3.4.1 Examples Example 3.18 Positive-definite matrices The left action of G = SLn (R) on K = SOn(R) identifies the space of left cosets gK with the quotient SLn (R)/SOn (R). The space of left cosets can be identified with positive matrices Pn by the following argument. Every matrix S ∈ G can be written

42

3 Lie groups and homogeneous spaces

in polar form√as S = PR, where P is a positive matrix and R is a rotation. In fact, P = SS∗ and R = P−1 S. Suppose now that S1 K = S2 K. Then −1 ∗ ∗ S1 S2−1 is a rotation. Hence, (S1 S2−1 )−1 = (S1 S√ 2 ) which implies that S1 S1 = ∗ ∗ S2 S2 . Therefore the correspondence SK → SS is one to one. Then Pn is topologized so that this correspondenceis a homeomorphism. a b The case n = 2 is special. Elements ∈ SL2 (R) act on the points c d in the upper-half plane by the Moebius tranformations az + b a b . (z) = c d cz + d It follows that SO2 (R) is the isotropy & subgroup associated with z = i. % √ y √xy Since x + iy = S(i) with S = , the action is transitive. The √1 0 y above realizes the upper-half plane {z ∈ C : Im(z) > 0} as the quotient SL2 (R)/SO2 (R). Example 3.19 The generalized upper-half plane The generalized upper-half plane Hn+ consists of n × n complex matrices Z of the form Z = X + iY with X and Y n × n matrices with real entries , X symmetric, and Y positive. It can be shown that the symplectic group Spn acts transitively on H+n by the fractional transformations A B (Z) = (AZ + B)(CZ + D)−1 C D and that the isotropy group of Z0 = iI is equal to SUn . In this notation, g ∈ Spn A B is written in the block form g . Therefore, C D Hn+ = Spn /SUn . For n = 2, Sp2 = SL2 (R) and SU2 = SO2 (R). Hence, this formula agrees with the one obtained above. Example 3.20 The spheres and the hyperboloids Any subgroup G of GLn+1 (R) acts on column vectors in Rn+1 by the matrix multiplications on the left. In the case that G = SOn+1 (R) then this action preserves the Euclidean norm and therefore, the orbit through any non-zero vector x0 is the sphere which case the isotropy ||x|| = ||x0 ||. Most commonly, x0 is taken to be e1 , in 1 0 , where 0 denotes the group K consists of matrices of the form 0T Q

3.4 Homogeneous spaces

43

zero row vector in Rn , 0T the column vector Q an arbitrary matrix in SOn (R). Evidently, the isotropy group is isomorphic to SOn (R), and thus Sn = SOn+1 (R)/SOn (R). If the Euclidean inner product is replaced by the Lorentzian inner product 2 n x, y = x1 y1 − n+1 i=2 xi yi then ||x|| = x, x = 1 is the hyperboloid H = {x : n+1 x12 − i=2 = 1}. The group SO(1, n) acts transitively on Hn and the isotropy subgroup of x0 = e1 is isomorphic to SOn (R), hence Hn = SO(1, n)(R)/SOn (R). Example 3.21 The Grassmannians The set of k-dimensional subspaces in an n-dimensional Euclidean space is denoted by G(n, k). Each S in G(n, k) can be identified with the orthogonal reflection PS defined by PS (x) = x, , x ∈ S, PS (x) = −x, x ∈ S⊥ . Then G(n, k) is topologized by the finest topology in which the correspondence S → PS is a homeomorphism. The group of rotations On (R) acts on the reflections PS by the conjugation, that is, (T, PS ) → TPS T ∗ . The action is transitive and the isotropy {T ∈ On (R) : TPS T ∗ = PS } consists of the rotations that satisfy T(S) = S and T(S⊥ ) = S⊥ . This subgroup is isomorphic with On−k × Ok (R), hence G(n, k) = On (R)/On−k × Ok (R). Example 3.22 The Steifel manifolds Str The Stiefel manifold Str consist of ordered orthonormal vectors x1 , . . . , xr in a Euclidean vector space En . The points of Str can be represented by r × n matrices X with columns the coordinates of x1 , . . . , xr relative to a fixed orthonormal basis in Rn . Such matrices form an nr − 12 r(r + 1)-dimensional manifold considered as points in Rnr subject to r2 symmetric conditions xi xj = δij . If R is an element of SOn (R) and if X is an n × r matrix of orthonormal vectors, then RX is also a matrix of orthonormal vectors. This action is transitive, since any two n × r matrices X and Y can be completed to elements X¯ and Y¯ in SOn (R) in which ¯ Y) ¯ −1 takes X onto Y. Therefore, Str can be identified with the orbit case R = X( of SOn (R) through the point X0 = (e1 , . . . , er ) modulo the isotropy group of X0 . Evidently, this isotropy group is isomorphic to SOn−r (R), hence Str = SOn (R)/SOn−r (R).

4 Symplectic manifolds: Hamiltonian vector fields

It is a common practice in applied mathematics to mix Hamiltonian vector fields with the underlying Riemannian structure and express Hamiltonian vector fields as the “skew-gradients” of functions. This tendency to identify tangent bundles with the cotangent bundles via the metric obscures the role of symplectic structure in Hamiltonian systems and also obscures the nature of symmetries present in Hamiltonian systems. For instance, on a given manifold there are many Riemannian structures, but there is only one canonical symplectic form on its cotangent bundle. So the passage from functions to Hamiltonian fields via the symplectic form is intrinsic, while the passage from functions to skew-gradients is not. For that reason we will be somewhat more formal in our introduction of Hamiltonian systems and will limit our discussion at first to vector spaces before going on to general manifolds.

4.1 Symplectic vector spaces A finite-dimensional vector space V together with a bilinear form ω : V × V → R is called symplectic if: (i) ω is skew-symmetric, i.e., ω(v, w) = ω(w, v) for all v and w in V. (ii) ω is non-degenerate, i.e., ω(v, w) = 0 for all w ∈ V can hold only for v = 0. n n If a1 , a2 , . . . , an is a basis in V then ω(v, w) = i=1 j=1 wi vj ω(ai , aj ) n n where w = i=1 wi ai and v = i=1 vi ai . The matrix A with entries ω(ai , aj ) is skew-symmetric and hence its eigenvalues are imaginary. It follows that any symplectic space V must be even dimensional since A must be non-singular because of the non-degeneracy of ω. 44

4.1 Symplectic vector spaces

45

The symplectic form provides means of identifying V with its dual. The correspondence v ∈ V → ω(v, ·) ∈ V ∗ is a linear isomorphism between V and V ∗ . A linear subspace S of a symplectic vector space V is called isotropic if ω vanishes on S, i.e., if ω(v, w) = 0 for all v and w in S. Any one-dimensional subspace of V is isotropic. A linear subspace L of V is called Lagrangian if L is isotropic and not a proper linear subspace of any isotropic subspace of V. A basis a1 , . . . , an , b1 , . . . , bn is called symplectic if ω(ai , aj ) = ω(bi , bj ) = 0, and ω(ai , bj ) = δij for all i and j. We will use a1 , . . . , ak to designate the linear span of any vectors a1 , . . . , ak . It then follows that each of the spaces L1 = a1 , . . . , an and L2 = b1 , . . . , bn is Lagrangian corresponding to any symplectic basis a1 , . . . , an , b1 , . . . , bn . A linear subspace S is said to be a symplectic subspace of V if the restriction of the symplectic form ω to S is non-degenerate. If S is any subset of V then S⊥ = {v : ω(v, w) = 0, for all w ∈ S}. Proposition 4.1 Let S be any linear subspace of V. Then dim(S⊥ )+ dim(S) = dim(V), and (S⊥ )⊥ = S. Proof The mapping L : v ∈ V → ω(v, ·)|S , ∈ S∗ is surjective, and its kernel is S⊥ . Since dim(ker(L)) + dim(Range(L)) = dim(V), dim(S⊥ ) + dim(S∗ ) = dim(S⊥ ) + dim(S) = dim(V) = dim((S⊥ )⊥ ) + dim(S⊥ ), it follows that dim(S) = dim(S⊥ )⊥ . Since S ⊆ (S⊥ )⊥ , S = (S⊥ )⊥ . Proposition 4.2 Suppose that a1 , . . . , ak , b1 , . . . , bk are any vectors in a symplectic space V that satisfy ω(ai , aj ) = ω(bi , bj ) = 0, ω(ai , bj ) = δij for all i ≤ k and j ≤ k. Let W denote the linear span of a1 , . . . , ak , b1 , . . . , bk . Then each of W and W ⊥ are symplectic subspaces of V and V = W ⊕ W ⊥. Proof

If w =

k

i=1 αi ai

+ βi bi belongs to W ∩ W ⊥ then

αi = ω(w, bi ) = 0 and βi = ω(ai , w) = 0

46

4 Symplectic manifolds: Hamiltonian vector fields

for all i. Hence, W ∩ W ⊥ = {0}. Then it follows from Proposition 4.1 that V = W ⊕ W ⊥ . Evidently, the symplectic form must be non-degenerate on each factor and hence each of W and W ⊥ are symplectic. Proposition 4.3 Suppose that S is any isotropic subspace of V. Let a1 , a2 , . . . , ak be any basis in S. Then there are vectors b1 , b2 , . . . , bk in V such that ω(ai , aj ) = ω(bi , bj ) = 0, and ω(ai , bj ) = δij for all i and j. Proof Let S1 denote the linear span of a2 , . . . , ak . Then a1 cannot belong to (S1⊥ )⊥ by Proposition 4.1. Therefore there exists b ∈ S1⊥ such that ω(a1 , b) = 0. Let b1 = ω(a11 ,b) b. By Proposition 4.2, V = W1 ⊕ W1⊥ where W1 = a1 , b1 . Since S1 is contained in W1⊥ and W1⊥ is symplectic, the above procedure can be repeated with S2 = a3 , . . . , ak to obtain b2 . The procedure stops at the k stage. Corollary 4.4 Each Lagrangian subspace of a 2n-dimensional symplectic space V is n dimensional. Corollary 4.5 Every basis a1 , . . . , an of a Lagrangian space L corresponds to a symplectic basis a1 , . . . , an , b1 , . . . , bn . Every symplectic basis a1 , . . . , an , b1 , . . . , bn gives rise to symplectic coor dinates (x1 , . . . , xn , p1 . . . , pn ) defined by v = ni=1 xi ai + pi bi for any vector v ∈ V. If (y1 , . . . , yn , q1 . . . , qn ) denote the symplectic coordinates of a vector w in V then ω(v, w) =

n

xi qi − yi pi .

(4.1)

i=1

We will use (x, p) and (y, q) to denote the symplectic coordinates (x1 , . . . , xn , p1 , . . . , pn ) and (y1 , . . . , yn , q1 , . . . , qn ) of any two points in V. If ((x, p), (y, q)) designates the standard Euclidean product in R2n , then (4.1) can be written as ω(v, w) = ((x, p), J(y, q)) where 0 I J= , −I 0 with I equal to the n × n identity matrix. The matrix J corresponds to the linear transformation defined by J(ai ) = bi and J(bi ) = −ai for each i. Then J 2 = −I and hence J is the complex structure on V. Thus V together with J becomes an n-dimensional complex space with

4.2 The cotangent bundle of a vector space

47

(α +iβ)v = αv+βJ(v) for any complex number z = α +iβ and any vector v ∈ V. Then a1 , . . . , an is a basis for the complexification of V and x1 +ip1 , . . . , xn + ipn are the complex coordinates of v relative to this basis. The symplectic form then can be identified with the imaginary part of the Hermitian inner product on V given by x + ip, y + iq =

n

(x j y j + pi q j ) + i(q j x j − p j y j ).

(4.2)

j=1

Symplectic vector spaces arise naturally in the following setting. Let E∗ designate the dual of a real n-dimensional vector space E. Then V = E × E∗ has a natural symplectic structure given by ω((x, f ), (y, g)) = g(x) − f (y)

(4.3)

for all (x, f ) and (y, g) in V. Then both E and E∗ can be embedded in V as E × {0} and {0} × E∗ , in which case they become Lagrangian subspaces, called horizontal and vertical Lagrangians. Any basis a1 , . . . , an in E gives rise to the dual basis a∗1 , . . . , a∗n in E∗ , and together they constitute a symplectic basis in V.

4.2 The cotangent bundle of a vector space Any n-dimensional real vector space E is an n-dimensional manifold with (x1 , . . . , xn ) the global coordinates of a point x induced by a basis a1 , . . . , an . n Then it is natural to identify the tangent vectors at x = i=1 xi ai with n dxi v = i=1 dt ai , for if x() is a curve such that x(0) = x0 then the induced i tangent vector v at x0 is identified with v = ni=1 dx d (0)ai . In this context, the tangent space Tx (E) is identified with the affine space {(x, v) : v ∈ E}. In this identification Tx (E) = (x, 0) + T0 (E), with α(x, v) + β(x, w) = (x, αv + βw). Hence, the tangent space at the origin plays a special role: tangent vectors at x need to be first translated to the origin before they can be scaled or added. Since T0 (E) is equal to (0) × E it is then natural to identify the cotangent space T0∗ (E) with (0)×E∗ and the cotangent space at any point x with the affine space {(x, p) : p ∈ E∗ }. If a∗1 , . . . , a∗n denotes the dual basis corresponding to

48

4 Symplectic manifolds: Hamiltonian vector fields

the basis a, . . . , an then every cotangent vector (0, p) can be written as p = n ∗ i=1 pi ai , in which case the natural pairing between vectors and covectors is given by (x, p), (x, v) =

n

pi vi .

i−1

The above implies that the tangent bundle T(E) and the cotangent bundle T ∗ (E) are identified with E×E and E×E∗ . With this formalism at our disposal, then the tangent bundle T(T ∗ (E)) is identified with (E × E∗ ) × (E × E∗ ), with the understanding that the first two factors designate the base point (x, p) ˙ p˙ ) at this in T ∗ (E) and the second factors designate the tangent vectors (x, point. There is a natural pairing between covectors p and tangent vectors x˙ as well as between tangent vectors p˙ and x. ˙ This pairing gives rise to the following differential forms on T ∗ E. Definition 4.6 Let (x˙1 , p˙ 1 ) and (x˙2 , p˙ 2 ) designate arbitrary tangent vectors at ˙ p˙ ) = p(x) ˙ is a point (x, p) in T ∗ (E). Differential form θ defined by θ(x,p) (x, called the Liouville form. If dθ denotes its exterior derivative, then ω = −dθ is called the symplectic form on T ∗ E. Coordinates x1 , . . . , xn , p1 . . . , pn relative to the bases a1 , . . . , an , a∗1 , . . . , a∗n are called symplectic. This choice of coordinates induces coordinates on n n ˙ = ˙ i a∗i . These tangent vectors x˙ and p˙ with x˙ = i=1 x˙i ai and p i=1 p n ∂ relations could be expressed also by dual notation as x˙ = i=1 x˙i ∂xi and n p˙ = i=1 p˙ i ∂p∂ i . It follows that in these coordinates θ=

n i=1

pi dxi and ω =

n

dxi ∧ dpi ,

(4.4)

i=1

which further implies that ˙ − p˙ (y). ˙ ω(x,p) (v, w) = q˙ (x)

(4.5)

for any tangent vectors v = (x, ˙ p˙ ) and w = (y, ˙ q˙ ) at a point (x, p) in E × E∗ . Expression (4.5) shows that ω is non-degenerate over each point (x, p) ∈ E × E∗ . Hence, ω defines a symplectic structure in the vector space of tangent vectors E × E∗ over each base point (x, p). It is clear that the linear span of (a1 , 0), . . . , (an , 0) is a Lagrangian subspace of

4.3 Symplectic manifolds

49

E × E∗ and so is the linear span of (0, a∗1 ), . . . , (0, a∗n ). Combined vectors (a1 , 0), . . . , (an , 0), (0, a∗1 ), . . . , (0, a∗n ) form a symplectic basis for E × E∗ . Definition 4.7 Vector field H is called Hamiltonian vector fields if there is a v) for all tangent vectors v. function H such that dH(v) = ω(H, Every function H induces a Hamiltonian vector field H given by H =

n ∂H ∂ ∂H ∂ − ∂pi ∂pi ∂xi ∂xi i=1

in each choice of symplectic coordinates. The integral curves (x(t), p(t)) of H are conveniently written as ∂H dp ∂H dx (t) = (x(t), p(t)), (t) = − (x(t), p(t)) dt ∂p dt ∂x

(4.6)

with the understanding that it is a shorthand notation for the differential system dxi ∂H (x1 (t), . . . , xn (t), p1 (t), . . . , pn (t)) and (t) = dt ∂pi ∂H dpi (t) = − (x1 (t), . . . , xn (t), p1 (t), . . . , pn (t)) dt ∂xi

(4.7)

for all i = 1, . . . , n.

4.3 Symplectic manifolds We now extend the previous formalism to arbitrary cotangent bundles. The fundamental setting is the same as it was for cotangent bundles of vector spaces; the symplectic form is the exterior derivative of the differential form of Liouville. To explain in more detail, let π denote the natural projection from T ∗ M onto M carrying each covector ξ at x to its base point x. Then for each curve (t) in T ∗ M that originates at ξ at t = 0, the projected curve σ (t) = π ◦ (t) originates at x = π ◦ ξ . The tangent map π∗ is a linear mapping from Tξ (TM) onto Tx M that carries d dσ ∗ dt (0) onto dt (0). The Liouville form θ is equal to the dual mapping π : ∗ ∗ T M → T(T M), defined by d dσ d (0) = ξ ◦ π∗ (0) = ξ ◦ (0). π ∗ (ξ )) dt dt dt Thus θ maps a covector ξ at Tx∗ M onto a covector θξ at Tξ∗ (T ∗ M) such that d dσ (0) = ξ (0) θξ dt dt

50

4 Symplectic manifolds: Hamiltonian vector fields

∗ for each tangent vector d dt (0) in Tξ (T M). This somewhat abstract definition of θ takes on the familiar form when expressed in local coordinates. Then, θξ = ni=1 vi pi , where v1 , . . . , vn denote the coordinates of a tangent vector v relative to the basis ∂x∂ 1 , . . . , ∂x∂ n and p1 , . . . , pn denote the coordinates of a covector at x relative to the dual basis dx1 , . . . , dxn . Therefore, θ = ni=1 pi dxi as in the previous section.

Definition 4.8 The symplectic form ω on T ∗ M is equal to the negative of the exterior derivative dθ . It then follows from (1.3) in Chapter 1 that ωξ (V(ξ ), W(ξ )) = W(θ (V) − V(θ (W) − θ ([V, W])

(4.8)

for any vector fields V and W on T ∗ M. In symplectic coordinates, vector fields ∂x∂ 1 , . . . , ∂x∂ n , ∂p∂ 1 , . . . , ∂p∂ n form a basis for tangent vectors on an open set on T ∗ M. Hence, it is sufficient to n ∂ ˙ i ∂p∂ i consider pairs of vector fields (V, W) of the form V = i=1 x˙i ∂xi + p n ∂ ∂ and W = ˙ i ∂pi for some points (x1 , . . . , xn , p1 , . . . , pn ) and i=1 y˙i ∂xi + q (y1 , . . . , yn , q1 , . . . , qn ) in R2n . Then [V, W] = 0, and moreover, θ (V) = n n n ˙ i x˙i . Similarly, V(θ (W)) = ˙ i y˙i . i=1 pi x˙i and W(θ (V)) = i=1 q i=1 p Hence, ω in (4.8) is given by ω=

n

q˙ i x˙i − p˙ i y˙i ,

i=1

which agrees with expression (4.5) of the previous section. defined by Definition 4.9 If h is any function on M, then vector field h, ), X(ξ )) dhξ (X(ξ ) = ωξ (h(ξ for all tangent vectors X(ξ ) ∈ Tξ M, is the Hamiltonian vector field generated by h. The above can be stated more succinctly in terms of the interior product ih as dh = ih . The interior product denoted by iX is a contraction of a differential form ω with a vector field X. It maps k-exterior forms k (M) into (k−1)-forms in k−1 (M) by the formula (iX ω)(X1 , . . . , Xk−1 ) = ω(X, X1 , . . . , Xk−1 ), with the understanding that the contraction of a zero form is zero.

(4.9)

4.3 Symplectic manifolds

51

It follows that Hamiltonian vector fields h are given by h =

n ∂h ∂ ∂h ∂ − i i i i ∂p ∂x ∂x ∂p

(4.10)

i=1

in any choice of local coordinates, from which it follows that the integral curve of h are the solutions of ∂h dpi ∂h dxi = i, = − i , i = 1, . . . , n. dt ∂p dt ∂x

(4.11)

Cotangent bundles are a particular case of a more general class of manifolds called symplectic manifolds [Ar]. Definition 4.10 A manifold M together with a smooth, non-degenerate, closed 2-form ω on M is called symplectic. Any such 2-form is called symplectic. As in the case of the cotangent bundle, a vector field h on M is called Hamiltonian if there exists a function h on M such that dh = ih . The set of Hamiltonian vector fields on M will be denoted by Ham(M). Every symplectic manifold is even dimensional as a consequence of the nondegeneracy of the symplectic form. Proposition 4.11 Darboux’s theorem Let (M, ω) denote a symplectic 2n-dimensional manifold. At each point of M there exists a coordinate neighborhood U with coordinates (x1 , . . . , xn , p1 , . . . , pn ) such that vectors ∂ ∂ ∂ ∂ ,..., , ,..., ∂x1 ∂xn ∂p1 ∂pn

(4.12)

form a symplectic basis on U. In this basis, the Hamiltonian vector fields are given by f =

n ∂f ∂ ∂f ∂ − . ∂pi ∂xi ∂xi ∂pi i=1

Proof We need to show that at any point q of M there exist a neighborhood U of q and a basis of vector fields E1 , . . . , En , En+1 , . . . , E2n on U such that ω(Ei , Ej ) = ω(En+i , En+j ) = 0, 1 ≤ i, j ≤ n, ω(Ei , En+j ) = δij . Then, x1 , . . . , xn , p1 , . . . , pn are the coordinates defined by the contractions iEi ω = dxi , iEn+i = dpi , i = 1, . . . , n, relative to which the vector fields are as in (4.12), and ω is given by ω = dx1 ∧ dp1 + · · · + dxn ∧ dpn .

52

4 Symplectic manifolds: Hamiltonian vector fields

The proof is by induction on the dimension of M. Assume that the proposition is true for dim(M) ≤ 2(n − 1) and let M be a 2n-dimensional manifold with a symplectic form ω. The proof is the same as in Proposition 4.3 in Section 4.1. Let E1 be any vector field such that E1 (q) = 0. Then there exists a vector field V such that ω(Ei , V) = 0, by the non-degeneracy condition. Let En+1 = ω(E11 ,V) V. If n = 1, the proof is finished. Otherwise, pass to the skew-orthogonal complement S⊥ = {V : ω(E1 , V) = ω(V, En+1 ) = 0}. Since ω is non-degenerate on S⊥ , the induction hypothesis then implies that S⊥ admits a basis with the above properties. It then follows that the Hamiltonian vector fields that correspond to the coordinate functions fi = pi , fn+i = −xi , i = 1, . . . , n, span the tangent space at point of U. Therefore, each orbit of Ham(M) is open and coincides with M whenever the latter is connected. Our next propositions make use of the following generalization of the Lie derivative. Definition 4.12 Let X denote a vector field on M and ω a k form in k (M), with k > 0. Designate by the one-parameter group of diffeomorphisms induced by X. Then the Lie derivative LX ω of ω along X is defined by the following formula: d ω (ξ ) ( ) ∗ v1 , . . . , ∗ vk )|=0 . d Evidently, the Lie derivative coincides with the directional derivative of a function along X. For general forms, the following remarkable formula holds. (LX ω)ξ (v1 , . . . , vk ) =

Proposition 4.13 Cartan’s formula

The Lie derivative LX is given by

LX = d ◦ iX + iX ◦ d, where d denotes the exterior derivative and iX is the contraction along X. The proof of this proposition can be found in [AS]. Definition 4.14 If F is a diffeomorphism M and if ω is any form in k (M), ˆ is the form defined by then Fω ˆ ξ (v1 , . . . , vk ) = ωF(ξ ) (F∗ v1 , . . . , F∗ vk ). (Fω) ˆ = ω. A form ω is said to be invariant under a diffeomorphism F if Fω Definition 4.15 Let (M, ω) denote a symplectic manifold. A diffeomorphism that leaves the symplectic form ω invariant is called a symplectomorphism. The set of symplectomorphisms on M will be denoted by Symp(M).

4.3 Symplectic manifolds

53

Evidently, Symp(M) is a subgroup of the group of diffeomorphisms Diff (M). Proposition 4.16

Let (M, ω) be a symplectic manifold. Then,

(a) exp th is a symplectomorphism for each h in C∞ (M) and for each t in R. for each F ∈ Symp(M) and each function g. (b) F ◦ g ◦ F −1 = Fg g(exp belongs to Ham(M) for each t ∈ R and all functions f (c) (exp th) −th) and g. Proof Part (a) follows from Cartan’s formula Lh ω = d ◦ ih ω + iX ◦ dω = d ◦ dh = 0. Part (b) follows from the calculation below: ˆ ξ (X, F gF −1 ) = Fω −1 ) ωξ (X, F gF = ωF(ξ ) (F −1 ◦ X ◦ F, g) = F ◦ (F −1 ◦ X ◦ F)g = X(Fg) = ωξ (X, (Fg)) belongs to Part (c) is evident since (exp th) Therefore, F ◦ g ◦ F −1 = Fg. Symp(M). Definition 4.17 The Poisson bracket { f , g} of any two functions f and g in C∞ (M) is defined by d (exp tg)f (ξ )|t=0 = (gf )(ξ ) = ωξ (f(ξ ), g(ξ )), ξ ∈ M. dt Evidently, the Poisson bracket is skew-symmetric. Furthermore, { f , g}(ξ ) =

f1 f2 )(ξ ) = f1 (ξ ){ f2 , g}(ξ ) + f2 (ξ ){ f1 , g}(ξ ). { f1 f2 , g}(ξ ) = g( Therefore, the Poisson bracket acts on functions as a derivation. Proposition 4.18 If F ∈ Symp(M) then F◦{ f , g} = {Ff , Fg} for any functions f and g. Proof ˆ F ◦ { f , g} = F ◦ ω(f, g) = Fω(F ◦ f ◦ F −1 , F ◦ g ◦ F −1 ) = ω((F ◦ f ◦ F −1 , F ◦ g ◦ F −1 ) , Fg) = {Ff , Fg} = ω(Ff f , g}|t=0 = Then dtd (exp th){ identity:

d dt {(exp th)f , (exp th)g}|t=0

reduces to the Jacobi

{ f , {g, h}} + {h, { f , g}} + {g, {h, f }} = 0.

54

4 Symplectic manifolds: Hamiltonian vector fields

Corollary 4.19 C∞ (M) is a Lie algebra under the Poisson bracket, that is, the Poisson bracket is bilinear, skew-symmetric, and satisfies the Jacobi’s identity. Proposition 4.20 Ham(M) is a Lie subalgebra of Vec(M), and the correspondence f → f is a Lie algebra homomorphism from C∞ (M) onto Ham(M). The kernel of this homomorphism consists of constant functions, whenever M is connected. Proof

Let f and g be arbitrary functions on M. Then,

[f, g]h = g ◦ f(h) − f ◦ g(h) = {{h, f }, g} − {{h, g}, f } = {{ f , g}, h} = { f, g}h. Therefore, [f, g] = { f, g}. This shows that Ham(M) is a Lie algebra, and the correspondence f → f a Lie algebra homomorphism. In any system of symplectic coordinates, f is given by f =

n ∂f ∂ ∂f ∂ − . ∂pi ∂xi ∂xi ∂pi i=1

If f = 0, then = 0 and = 0 and therefore f is constant in each such coordinate neighborhood. But then f is constant when M is connected. ∂f ∂pi

∂f ∂xi

5 Poisson manifolds, Lie algebras, and coadjoint orbits

The correspondence between functions and vector fields is articulated more naturally by the Poisson bracket rather than the symplectic form. This subtle distinction leads to the notion of a Poisson manifold [Wn]. The shift from symplectic to Poisson structure leads to remarkable discoveries fundamental for the theory of integrable systems and the geometry of Lie groups. Let us begin with the basic concepts.

5.1 Poisson manifolds and Poisson vector fields Definition 5.1 A Poisson bracket on a manifold M is a mapping { , } : C∞ (M) × C∞(M) → C∞ (M) that is bilinear and skew-symmetric, and satisfies { fg, h} = f {g, h} + g{ f , h}, { f , {g, h}} + {{h, { f , g}} + {g, {h, f }} = 0, for all functions f , g, h in C∞ (M). A manifold M together with a Poisson bracket { , } is called a Poisson manifold. Definition 5.2 If (M, {, }) is a Poisson manifold then vector field f defined by f(g) = {g, f } will be called a Poisson vector field. The set of all Poisson vector fields will be denoted by Poiss(M). It is an easy consequence of the Jacobi identity that [ f, g] = { f, g} for all functions f and g, which implies that the Poisson vector fields form a Lie subalgebra of Vec(M). Moreover, the mapping f → f 55

56

5 Poisson manifolds, Lie algebras, and coadjoint orbits

is a homomorphism from the Poisson algebra C∞ onto the Lie subalgebra Poiss(M) in V ∞ (M). Proposition 5.3 Let (M, {, }) be a symplectic manifold. and let h be a Poisson vector field on M. Then, f , (exp th)g}, exp th { f , g} = {(exp th) for all functions f and g. Proof

, (exp th)g}. Let φ(t) = exp −th({(exp th)f Then,

dφ dt

is equal to

+ (exp th)g}) h{(exp + (exp th)g} + d {(exp th)f exp −th(− th)f dt h{(exp + (exp th)g} + h{(exp + (exp th)g}) = exp −th(− th)f th)f =0 because of the Jacobi identity. Therefore, φ(t) = { f , g} for all t. Proposition 5.4 Let f and g be any functions on a Poisson manifold M. Then (exp tf)g(exp −tf) is a Poisson vector field that corresponds to (exp tf)g. Proof

Let h denote a function on M. Then, (exp tf)g(exp −tf)h = exp tf{(exp −tf)h, g} = {h, (exp tf)g}.

Proposition 5.5 Let (M, {, }) be a Poisson manifold and let F denote the family of all Poisson vector fields on M. Then each orbit of F is a symplectic submanifold of M. Proof Let N denote an orbit of F. It follows from (1.15) in Chapter 1 that the tangent space at each point z of N is a linear combination of vectors exp tm hm ◦ exp tm−1 hm−1 ◦ · · · exp t1 h1 ) zˆ ◦ (exp −t1 h1 ◦ · · · exp −tm hm )g(◦ for some functions g, h1 , . . . , hm on M. According to Proposition 5.4, (exp −t1 h1 ◦ · · · exp −tm hm )g(◦ exp tm hm ◦ exp tm−1 hm−1 ◦ · · · exp t1 h1 ) = h, where h = (exp −t1 h1 ◦ · · · exp −tm hm )g. Hence each tangent space Tz N of N is equal to the evaluation Fz of F at z. To show that N is symplectic, let {, }N denote the restriction of the Poisson bracket to N. Recall that all smooth functions on N are the restrictions of smooth function on M. It follows (N, {, }N ) is a Poisson submanifold of (M, {, }). Define ωξ ( f(ξ ), g(ξ )) = { f , g}N (ξ ) (5.1) for all points ξ in N.

5.2 The cotangent bundle of a Lie group: coadjoint orbits

57

To verify that ω is non-degenerate, suppose that ωξ ( f(ξ ), g(ξ )) = 0 for )) = { f , g}N (ξ ) = some function g, and all functions f on N. Since ωξ ( f(ξ ), g(ξ g( f )(ξ ) it follows that g f = 0 for any function on N. But this means that g = 0 and consequently, ω is non-degenerate. Since the Jacobi identity holds on N, ω is closed. The above proposition may be restated in the language of distributions as follows. Proposition 5.6 The distribution D defined by D(ξ ) = { f(ξ ) : f ∈ C∞ (M)} is integrable and each maximal integral manifold of D is symplectic. We will refer to the maximal integral manifolds of D as the symplectic leaves of M. Corollary 5.7 A connected Poisson manifold M is symplectic if and only if the family of all Poisson vector fields has only one orbit in M.

5.2 The cotangent bundle of a Lie group: coadjoint orbits Let us now incorporate certain group symmetries into the symplectic structure of the cotangent bundle of a Lie group G. Each Lie group G admits a global frame of either left- or right-invariant vector fields, which implies that both the tangent bundle TG and the cotangent bundle T ∗ G can be written as the products G × g and G × g∗ , where g∗ denotes the dual of the Lie algebra g of G. In fact, every basis A1 , . . . , An in the Lie algebra g of a Lie group G defines a global frame of either left- or right-invariant vector fields X1 , . . . , Xn such that Xi (e) = Ai for each i = 1, . . . , n. We will work with left-invariant vector fields and identify the tangent bundle TG with G × g. This identification is done as follows: points (g, A) ∈ G × g are identified with vectors V(g) ∈ Tg (G) via the formula V(g) = (Lg )∗ (A). Similarly, points (g, ) of G × g∗ are identified with points ξ ∈ Tg∗ (G) via the formula Lg∗ ξ = , that is, ξ and are related by ξ(V(g)) = (Lg −1 ∗ V(g)) for every tangent vector V(g) at g. In particular, ξ(V(g)) = (A) for left-invariant vector fields V(g) = (Lg )∗ (A), and ξ(V(g)) = ((Lg −1 ∗ Rg ∗ (A)) for any rightinvariant vector field V(g) = Rg ∗ (A). Each product G × g and G × g∗ is a product of a Lie group with a vector space, and hence is a Lie group in its own right. On G × g∗ the group operation is given by

58

5 Poisson manifolds, Lie algebras, and coadjoint orbits (g1 , 1 )((g2 , 2 ) = (g1 g2 , 1 + 2 ).

Having identified T ∗ G with the product G × g∗ , the tangent bundle of T ∗ G will be identified with the product (G × g∗ ) × (g × g∗ ), with the understanding that the second product denotes the tangent vectors at the base points described by the first product. Relative to this decomposition, vector fields on T ∗ G will be written as pairs (X(g, ), Y(g, )) with X(g, ) ∈ g, Y(g, ) ∈ g∗ , and (g, ) the base point in G × g∗ . In particular, the left-invariant vector fields V on G × g∗ are then represented by the pairs (A, ) = V(e, 0). It follows that

t (g0 , 0 ) = {(g0 etA , 0 + t) : (g0 , 0 ) ∈ G × g∗ } is the one-parameter group generated by a left-invariant vector field V on G×g∗ whose value at the identity is equal to (A, ). Let us now evaluate the canonical symplectic form on T ∗ G in the representation G × g∗ . In this representation the Liouville’s differential form θ is given by θ(g,) ((X(g, ), Y(g, )) = (X(g, )). When V is left-invariant, then θ(g,) (V) = (A), where A denotes the projection of V(e, 0) on g. If V1 and V2 are any left-invariant fields such that V1 (e, 0) = (A1 , 1 ) and V2 (e, 0) = (A2 , 2 ) then exp Vi t(g0 , 0 ) = (g0 exp tAi , 0 + ti ), i = 1, 2. Therefore, V2 (θ (V1 ))(g, ) =

d d (θexp tV2 (g,) (V1 )|t=0 = ( + t2 )(A1 )|t=0 = 2 (A1 ), dt dt

and similarly, V1 θ (V2 )(g, ) = 1 (A2 ). Moreover, θ(g,) ([V1 , V2 ]) = ([A1 , A2 ]). Let us now return to the symplectic form ω = −dθ . We have ω(g,) (V1 , V2 ) = −dθ(g,) (V1 , V2 ) = V2 (θ (V1 ))(g, ) − V1 (θ (V2 ))(g, ) − θ(g,) ([V1 , V2 ]) = 2 (A1 ) − 1 (A2 ) − ([A1 , A2 ]) We will refer to ω(g,) ((A1 , l1 ), (A2 , l2 )) = l2 (A1 ) − l1 (A2 ) − ([A1 , A2 ])

(5.2)

as the left-invariant representation of ω. The left-invariant representation is different from the one obtained through ∂H d ∂H the canonical coordinates, hence the usual formulas dg dt = ∂ , dt = − ∂g are no longer valid. The correct expressions are obtained by identifying a function

5.2 The cotangent bundle of a Lie group: coadjoint orbits

59

H on G × g∗ with its Hamiltonian vector field H((g, )) = (A(g, ), a(g, )) through the formula ω(g,) (A(g, ), a(g, )) , (B, b)) = dH(g,) (B, b).

(5.3)

Then, dHg, (B, 0) =

d H(g exp (Bt), )|t=0 = ∂Hg ◦ Lg ∗ (B) dt

and d H(g, + tb))|t=0 = ∂Hl (b). dt Since ∂Hl (g, ) is a linear function on g∗ , it is naturally identified with an element of g. It follows that dHg, (0, b) =

A(g, ) = ∂Hl (g, ) and l(g, ) = −∂Hg (g, ) ◦ Lg ∗ − ad∗ ∂Hl (g, )(), and therefore the integral curves (g(t), (t)) of H are the solution curves of dg (t) = g(t)∂Hl (g(t), (t)) , dt d (t) = −∂Hg (g(t), (t)) ◦ Lg ∗ − ad∗ ∂Hl (g(t), (t))((t)) (5.4) dt where ad∗ A : g∗ → g∗ is defined by (ad∗ A())(B) = [A, B] for all B ∈ g. Definition 5.8 Functions F on G × g∗ are said to be left-invariant if F(Lg (h), ) = F(h, ) for all g and h in G and all in g∗ . It follows that the left-invariant functions on G × g∗ are in exact correspondence with the functions on g∗ . The integral curves of the Hamiltonian vector fields generated by the left invariant functions are the solutions of the following differential equations d dg (t) = g(t)dH(t)) , (t) = −ad∗ dH((t) ((t)). (5.5) dt dt Proposition 5.9 [Ki] The dual g∗ of a Lie algebra g is a Poisson manifold with the Poisson bracket { f , h}() = −([df , dh]) for any functions f and h on g∗ . Proof Functions on g∗ coincide with the left-invariant functions on G × g∗ . Hence, = ω(g,) ((df , 0), (dh, 0)) = −ad∗ ([df , dg]) = −([df , dh]). ω(g,) ( f, h)

60

5 Poisson manifolds, Lie algebras, and coadjoint orbits

It follows that { , } is the restriction of the canonical Poisson bracket on G × g∗ to the left-invariant functions. Hence, it automatically satisfies the properties of a Poisson manifold. Remark 5.10 In the literature on integrable systems, the Poisson bracket { f , h}() = ([df , dh]) is often referred as the Lie–Poisson bracket [Pr]. We have taken its negative so that the Poisson vector fields agree with the projections to g∗ of the Hamiltonian vector fields generated by the leftinvariant functions. This choice of sign is also dictated by the choice of the sign for our Lie bracket [Jc]. It follows that each function H on g∗ defines a Poisson vector field H through f ), and it also defines a Hamiltonian vector field on the formula { f , H} = H( ∗ G × g . Each integral curve (g(t), (t)) of the Hamiltonian field associated with H is a solution of (5.5) and the projection (t) is an integral curve of and conversely, each solution of H is the projection of a solution of (5.5). H, Therefore, integral curves of H are the solutions of d (t) = −ad∗ dH((t) ((t)). dt

(5.6)

Solutions of equation (5.6) are intimately linked with the coadjoint orbits of G. To explain in more detail, let us introduce the relevant concepts first. −1 satisfies (e) = e, and The diffeomorphism g (x) = Lg ◦ R−1 g (x) = gxg

gh = g ◦ h for all g and h. The tangent map of g at the group identity is denoted by Adg . It follows that Adg is a linear automorphism of g. Moreover, Adgh = Adg ◦Adh , and therefore g → Adg is a representation of G into Gl(g). It is called the adjoint representation of G. The dual representation Ad∗ , called the coadjoint representation, is a representation of G into Gl(g∗ ). It is defined by Adg∗ () = ◦ Adg−1 , ∈ g∗ .

(5.7)

Definition 5.11 The set {Adg∗ () : g ∈ G} is called the coadjoint orbit of G through . The following proposition is of central importance in the theory of integrable systems. Proposition 5.12 Let F denote the family of Poisson vector fields on g∗ and let M be any coadjoint orbit of G. Then the orbit of F through each point in M is an open submanifold of M.

5.2 The cotangent bundle of a Lie group: coadjoint orbits

61

Proof Let h be a function on g∗ . The integral curves (t) of the associated Poisson field h are the solutions of equation (5.6) and (t) is the projection of a solution (g(t), (t)) of (5.5), that is, g(t) is a solution of dg dt = Lg (t)∗ (dh((t)). Let Adg(t) denote the curve in Gl(g) defined by the solution g(t) of dg dt = Lg(t) ∗ (dh((t)) with g(0) = e. Then according to (1.11) of Chapter 1, d (0 ◦ Adg(t) (X) = −(0 ◦ Adg(t) [dh, X] dt for any X ∈ g and any 0 ∈ g∗ . This means that Adg∗−1 (t) (0 ) satisfies equation (5.6). Hence, (t) = Adg∗−1 (t) (0 ), which implies that (t) evolves on the coadjoint orbit through 0 . It follows that the orbit of F through a point in M is contained in M. It remains to show that the orbit of F through contains an open neghborhood of in M. Let F0 denote the subfamily of F consisting of linear functions hA () = (A) defined by an element A in g. An integral curve (t) of hA is of the form (t) = Ad∗(eAt ) (0 ) for some 0 . If h1 , . . . , hm is any subset of linear functions defined by A1 , . . . , Am in g, then ∗ ◦ exp tm hm ◦ · · · exp t1 h1 = Ade∗Am tm ◦ · · · ◦ Ade∗A1 t1 () = Ad(e Am tm ···eA1 t1 ) ().

Our proof is now finished since the transformation (t1 , . . . , tm ) → eAm tm · · · eA1 t1 covers an open neighborhood of the identity for any basis A1 , . . . , Am of g. Corollary 5.13 Each coadjoint orbit of G is a symplectic submanifold of g∗ . The tangent space at a point on a coadjoint orbit M consists of vectors ◦ adX, X ∈ g and ω ( ◦ adX, ◦ adY) = [X, Y]

(5.8)

is the symplectic form on M. This corollary is a direct consequence of Proposition 5.5. Corollary 5.14 Proof

Each coadjoint orbit of G is even dimensional.

Symplectic manifolds are even-dimensional.

In contrast to the coadjoint orbits, adjoint orbits can be odd dimensional, as the following example shows. a b Example 5.15 Let G be the group of the upper-diagonal matrices 0 c with ac = 0. The Lie algebra g of G consists of 2 × 2 matrices X = x1 x2 . Let A∗1 , A∗2 , A∗3 be the dual basis in g∗ relative to the basis 0 x3

62

5 Poisson manifolds, Lie algebras, and coadjoint orbits A1 =

1 0

0 0

, A2 =

0 0

1 0

, A3 =

0 0 0 1

.

Then, 2 ' b b b x1 + x2 + x3 A2 + x3 A3 : ac = 0 AdG (X) = x1 A1 − ac a a and ∗ (l1 A∗1 AdG

+ l2 A∗2

+ l3 A∗3 )

' b2 ∗ ∗ ∗ = l1 − l2 A1 + cl2 A2 + l3 A3 : ac = 0 . ac

Evidently, AdG (X) is one dimensional for all matrices X with x2 x3 = 0, but AdG (A∗ ) is either zero dimensional when l2 = 0, or two dimensional in other cases. Definition 5.16 A non-degenerate quadratic form , on g is said to be invariant if [A, B], C = A, [B, C] for all elements A, B, C in g. Proposition 5.17 Adjoint orbits are symplectic on any Lie algebra g that admits an invariant form , . The symplectic form ω is given by ωL (adA(L), adB(L)) = L, [A, B]) for all tangent vectors adA(L) and adB(L) at L. Proof Since , is non-degenerate it can be used to identify g∗ with g via the formula ∈ g∗ ⇔ L ∈ g whenever (X) = L, X for all X ∈ g. ∗ () is identified with Ad (L) and tangent vectors ad ∗ A at Ad () are Then AdG G g identified with tangent vectors ad(A) at Adg (L). It follows that the symplectic form ω (ad∗ A(), ad∗ B()) = ([A, B]) is taken to ωL (ad(A)(L), ad(B)(L)) = L, [A, B]. Definition 5.18 Let g be any Lie algebra. The bilinear form Kl(A, B) = Tr(ad(A) ◦ ad(B)), where Tr(X) denotes the trace of a matrix X, is called the Killing form [Hl]. The Killing form is a symmetric quadratic form on g that enjoys several noteworthy properties. To begin, it is invariant relative to any automorphism

5.2 The cotangent bundle of a Lie group: coadjoint orbits

63

φ on g, that is Kl(A, B) = Kl(φ(A), φB)) for any A, B in g. The argument is simple: Kl(φ(A), φ(B)) = Tr(ad(φ(A)) ◦ ad(φ(B))) = Tr(φ ◦ ad(A) ◦ φ −1 ◦ φ ◦ ad(B) ◦ φ −1 ) = Kl(A, B). In particular, Kl is AdG invariant. Then, d Kl(Adexp tA (B), Adexp tA (C))|t=0 = Kl([A, B], C + Kl(B, [A, C]), dt and therefore, Kl is invariant in the sense of Definition 5.5. 0=

Definition 5.19 A Lie algebra is called semi-simple if its Killing form is nondegenerate. It follows that in semi-simple Lie groups every adjoint orbit is symplectic. Most studies of coadjoint orbits have been carried out in a semi-simple setting (apart from three-dimensional Lie groups, where the coadjoint orbits can be computed explicitly). One of the more outstanding results in this theory is that every coadjoint (adjoint) orbit is K¨ahler on simple compact Lie groups [Bo]. It was the work of A. Kirillov [Ki] that initiated the interest in coadjoint orbits. This interest was further amplified with the discovery of solitons in the Korteveg de Vries equation and their relation to Toda lattices [Ks; Sy]. Remarkably, many coadjoint orbits are symplectomorphic [Al] to the cotangent bundles of other manifolds, and that observation sheds new light on the theory of integrable systems. This observation with its implications deserves special attention and will be deferred to Chapter 10. In the meantime, we will first make another detour to control theory and its Maximum Principle as a bridge between Hamiltonian systems and optimal control.

6 Hamiltonians and optimality: the Maximum Principle

Let us now return to control systems and their reachable sets with a more detailed interest in their boundary properties. Recall the basic setting initiated in Chapter 2: the control system dx = F(x(t), u(t), u(t) ∈ U (6.1) dt on a manifold M and the associated family of vector fields F defined by the constant controls. We will assume that the control system (6.1) conforms to the same assumptions as in Chapter 2, so that the Cauchy problem x(τ ) = x0 associated with the time-varying vector field Xu (t) = F(x, u(t)) admits a unique solution x(t) that satisfies dx dt (t) = Xu (t)(x(t)) for almost all t for each admissible control u(t), and varies smoothly relative to the initial point x0 . We will use U to denote the space of bounded and measurable control functions on compact intervals [t0 , t1 ] that take values in the control set U in Rm . Such controls will be referred to as admissible controls. As before, A(x, T), A(x, ≤ T) and A(x) will denote the reachable sets (at exactly T units of time, up to T units of time, and reachable at any time) by the trajectories generated by the control functions in U . Analogously, AF (x, T), AF (x, ≤ T), and AF (x) will denote the reachable sets by the elements in F, i.e., points reachable by piecewise constant controls in U . Since each control function in U is the pointwise limit of a sequence of piecewise constant controls, the reachable sets defined by the trajectories generated by the elements of U have the same closure as the reachable sets generated by the family F. In addition, we will use A¯ to denote the closure of set A, ∂A to denote the boundary of A, and A0 or int(A) to denote the interior of A.

64

6.1 Extremal trajectories

65

6.1 Extremal trajectories In this chapter we will be interested in the boundary points of the reachable sets. To avoid the situation in which every point is a boundary point, we will assume that the interiors of the reachable sets are not empty. Of course, if the interior of A(x0 , T) is not empty, then the interior of A(x0 ) is not empty, but the converse may not be true, as the following simple example shows. Example 6.1 The reachable set at exactly T units of time of the control system dx 2 dt = u(t)e1 + (1 − u(t))e2 in R is equal to the line x2 (T) − x2 (0) + x1 (T) − x1 (0) = T, while A(x0 ) is the half space x2 (T) − x2 (0) + x1 (T) − x1 (0) ≥ 0. We have already seen that the interior of A(x0 ) is not empty whenever Lie(F) is of full rank at x0 , in which case the set of interior points of AF (x0 ) is dense in its closure. In this regard, an even stronger statement is given by the following: Proposition 6.2

If dim(Liex (F)) = dim(M) then int(AF (x)) = int(A¯ F (x)).

For the proof see [Jc, p. 68]. To state the conditions under which the interior of A(x, T) is not empty, we need to introduce another geometric object, called the zero-time ideal I(F). Definition 6.3 The vector space spanned by the iterated Lie brackets of F, called the derived algebra of F, will be denoted by D(F). The vector space spanned by D(F) and all differences X − Y of elements in F is called the zero-time ideal of F and is denoted by I(F). It is easy to see that Lie(F) = RX + I(F), where X is any fixed vector field in F and RX is the line through X, because any linear combination k α1 X1 + · · · + αk Xk of elements in F can be written as i=1 αi X + α1 (X1 − X) + · · · + αk (Xk − X), and Lie(F) = F + D(F) where F denotes the linear span of F. The above implies that either dim(Liex (F)) − 1 = dim(Ix (F)), or dim (Liex (F) = dim(Ix (F). The second case occurs whenever X(x) belongs to Ix (F) for some X ∈ F. Definition 6.4 The zero-time orbit of F through x0 will be denoted by N0 (x0 ). It consists of points x which can be written as xˆ = xˆ0 ◦ exp tk Xk ◦ · · · exp t2 X2 ◦ exp t1 X1 ,

66

6 Hamiltonians and optimality: the Maximum Principle

for some elements X1 , . . . , Xk in F and (t1 , . . . , tk ) ∈ Rk that satisfy t1 + t2 + · · · + tk = 0. Zero-time orbits are associated with the group of diffeomorhisms of the form exp tk Xk ◦ · · · exp t2 X2 ◦ exp t1 X1 for some elements X1 , . . . , Xk in F and (t1 , . . . , tk ) ∈ Rk that satisfy t1 + t2 + · · · + tk = 0. This group constitutes a normal subgroup of the group G(F) generated by {exp tX : X ∈ F, t ∈ R}. Suppose now that X is a complete vector field in a family F and suppose that N0 (x) is the zero-time orbit of F through a point x. Then, each time translate Nt = exp tX(N0 (x)) is the zero-time orbit through yˆ = xˆ ◦ exp tX(x). The relevant properties of the zero-time orbits are assembled in the proposition below. Proposition 6.5 Suppose that F is a family of vector fields such that the zerotime ideal I(F) is Lie determined, and suppose that F contains at least one complete vector field X. Then, (a) the orbit of I(F) through a point x coincides with the zero-time orbit through x. (b) If N is an orbit of F then the dimension of Ix (F) is constant over N. This dimension is either equal to the dimension of Liex (F), in which case the zero-time orbit through any point in N coincides with N, or it is equal to dim(Liex (F)) − 1, in which case, each zero-time manifold is s submaniold of N of codimension 1 and N is foliated by the time translates of N0 (x). (c) Each reachable set A(x, T) is contained in NT and has a non-empty interior in the topology of NT . Moreover, the set of interior points in A(x, T) is dense in the closure of A(x, T). We refer the reader to [Jc] for the proof of this theorem. Corollary 6.6 Tx M.

A(x, T) has a non-empty interior in M if and only if Ix (F) =

¯ T), it is not true in Even though the interior of A(x, T) is dense in A(x, ¯ T) is equal to the interior of A(x, T) as the general that the interior of A(x, following example shows. Example 6.7 Consider v(t) v(t) dy dx =( =( x − u(t)y, y + u(t)x, 2 2 dt dt x +y x2 + y2 1 in the punctured plane M = R2 − {0}, where |v(t)| ≤ 2π and |u(t)| < 1. In ( y −1 2 2 polar coordinates, r = x + y , φ = tan x , the above system takes on a

6.1 Extremal trajectories

67

paticularly simple form: dφ dr = v(t), = u(t). dt dt The reachable set from r = 1, φ = 0 at time T = π is given by {(r, φ) : 12 ≤ r ≤ 32 , −π < φ < π }. The interior of the closure of this set includes the open segment − 32 < x < − 12 which is not reachable at time T = π . With these preliminary considerations behind us, let us now come to the main issues of this chapter, the extremal trajectories, the Maximum Principle, and the relations to optimal control and the calculus of variations. To begin with: Definition 6.8 A trajectory x(t) that emanates from x0 at t = 0 and belongs to the boundary of A(x0 ) at some time T > 0 is called extremal. The corresponding control is also called extremal. Let us now introduce typical optimal control problems and then relate them to the extremal properties of the trajectories. Optimal control problems are defined by the control system (6.1) and a cost functional f : M × U → R. There are two optimal problems defined by this data. The first problem is defined over the trajectories of indefinite time duration; it consists of finding a control u(t) ¯ ∈ U that generates a trajectory x(t) ¯ on some interval [0, T] ¯ = x1 and attains the minimum value for the that satisfies ¯ = x0 x(T) S x(0) integral 0 f (x(t), u(t)) dt relative to any other trajectory (x(t), u(t)) on [0, S] that satisfies the same boundary conditions x(0) = x0 and x(S) = x1 . That is, T S f (x(t), ¯ u(t)) ¯ dt ≤ f (x(t), u(t)) dt. 0

0

The prototype of these problems are time optimal problems in which f = 1 and T is the least positive time that x1 can be reached from x0 by a trajectory of the system. The second optimal problem is defined over the trajectories of fixed time duration. Its formulation is similar to the one above, except that all the competing trajectories are defined over the same time interval [0, T]. Conceptually, these two problems are the same, since the optimal problem with fixed time duration can be turned into optimal problems over indefinite time duration by dx adding an extra equation dτ dt = 1 to the dynamics dt = F(x(t), u(t)). In the extended space M × R, the point (x1 , T) can be reached from (x0 , 0) only at time T. A control ¯ that generates a trajectory x(t) ¯ that attains the minimum S u(t) value for 0 f (x(t), u(t)) dt is called optimal and so is the corresponding

68

6 Hamiltonians and optimality: the Maximum Principle

trajectory x(t). ¯ Sometimes it is convenient to refer to (x(t), ¯ u(t)) ¯ as an optimal pair. Thus an optimal control problem is a two point boundary value problem T of finding the minimum of the cost functional 0 f (u(t), x(t)) dt over the trajectories x(t) that satisfy the given boundary conditions. Suppose now that the dynamics F, the control set U and the cost functional f are given. The pair ˜ = R × M with the trajectories ( f , F) defines another control system F˜ on M x˜ (t) = (y(t), x(t)) the solutions of dx dy (t) = f (x(t), u(t)), (t) = F(x(t), u(t)), u(t) ∈ U. (6.2) dt dt System (6.2) will be referred to as the cost-extended control system. The passage to the cost-extended system turns the optimal problem into an extremal problem, in the sense that every optimal pair (x(t), ¯ u(t)) ¯ generates a costextended trajectory (¯y(t), x(t)) ¯ which at the terminal time T is on the boundary of the cost-extended reachable set, because the line segment {(α, x(T)) ¯ :0< ˜ α < y(T)} does not intersect the cost-extended reachable set A(˜x(0)). The Maximum Principle then provides a necessary condition that a trajectory be extremal. This necessary condition states that an extremal trajectory is the projection of an integral curve of the associated Hamiltonian lift in the cotangent bundle T ∗ M. We will come back to this statement with complete details, but in the meantime let us digress briefly into the classical theory of the calculus of variations and make some parallels between its basic concerns and the problems of optimal control.

6.2 Optimal control and the calculus of variations Let us go back to the central question of the calculus of variations of finding the conditions underwhich a given curve x(t), ¯ t ∈ [t0 , t1 ] in Rn yields a minimum t1 dx ¯ subject to for the integral t0 f (t, x(t), dt ) dt in some neighborhood of x(t), fixed boundary conditions [Cr] and [GF]. In this context, f (t, x, u) is a given real-valued function on R × Rn × Rn . Let us first consider the most common case in which the minimum is taken in the weak sense, that is, over continuously differentiable curves x(t) on the ¯ 0 ) = x(t0 ) and x(t ¯ 1) = interval [t0 , t1 ] that satisfy the boundary conditions x(t x(t1 ) and conform to ) ) ' ) dx dx¯ ) ) ) sup ||x(t) − x(t)|| ¯ + ) (t) − (t)) : t ∈ [t0 , t1 ] < dt dt for some > 0. Then the most common necessary condition of optimality, based on the premise that a local minimum of a given functional occurs at the points where the derivative is zero, leads to the Euler–Lagrange equation

6.2 Optimal control and the calculus of variations

69

∂f dx d ∂f (t, x(t), u(t)) + (t, x(t), u(t)) = 0, u(t) = (t). (6.3) dt ∂u ∂x dt n ∂f 2 In the situations where the strong Legendre condition (t, x, u) ij ∂ui ∂uj dx¯ ui uj > 0 holds in some neighborhood of the solution curve x(t), ¯ dt , the equation ∂f (t, x, u) ∂u can be solved to yield u = φ(t, x, p). In that case, (x(t), ¯ p¯ (t)) is a solution of the Hamiltonian equation p=

∂H dp ∂H dx = , =− , (6.4) dt ∂p dt ∂x associated with H = −f (t, x, φ(t, x, p)) + nj=1 pj φj (t, x, p), where p¯ (t) = ∂f dx¯ ¯ ∂u t, x(t), dt . Equations (6.3) and (6.4) are very different from each other even though at a first glance they may seem completely equivalent. On manifolds, the Euler– Lagrange equation is an equation on the tangent bundle, and as a second-order differential equation it requires some notion of a connection or a covariant derivative for its formulation. In contrast, the Hamiltonian equations seem more natural, in the sense that, once the Hamiltonian is obtained, its equations are written intrinsically on the cotangent bundle thanks to the canonical symplectic form on the cotangent bundle. On the other hand, the Euler–Lagrange equation has a conceptual advantage, in the sense that it corresponds to the critical point of the functional. In the case of the Hamiltonian equations, the connection to the original variational problem is less clear. This conceptual discrepancy left its imprint on the existing literature. For the physicist, the Legendre transform was good enough since it led to the Hamiltonian that represents the total energy of the system. For that reason, the Hamiltonian formulation was always preferable in the physical world. That is not so much the case for a geometer, who prefered to see his geodesics through the Euler–Lagrange equation [DC; Gr]. Leaving such philosophical matters aside, there is another issue that sharply divides the calculus of variations – the contributions of C. Weierstrass and his theory of strong extrema [Cr; Yg]. Let us now go briefly into this theory and to the “excess function” of Weierstrass, which, according to L. C. Young, “revolutionized the calculus of variations” [Yg]. In the theory of strong extrema, parametrized curves defined on an interval I = [t0 .t1 ] are called addmisible if they are continuous and differentiable almost everywhere on I with the derivative belonging to L∞ ([t0 , t1 ]). It then

70

6 Hamiltonians and optimality: the Maximum Principle

follows that admissible curves satisfy the Lipschitz condition ||x(t) − x(s)|| ≤ K|t − s| for all tsdxand t in I, and therefore are absolutely continuous, that is, x(t) − x(s) = s dτ (τ ) dτ [Ro]. On the other hand, every absolutely continuous curve x(t) on I is differentiable almost everywhere on I and is admissible when || dx dt || is bounded on I. Hence, admissible curves coincide with absolutely continuous curves with bounded derivatives, i.e., they coincide with the Lipschitzian curves. A strong neighborhood of an admissible curve x(t), t ∈ [t0 , t1 ], consists of admissible curves y(t) all defined on the interval [t0 , t1 ] such that ni=1 |xi (t) − yi (t) < for some > 0 and all t ∈ [t0 , t1 ]. Every strong neighborhood of n a curve x(t) defines an open neighborhood O (x) of R defined by O (x) = n n y ∈ R : supt∈[0,T] i=1 |yi − xi (t)| < . t Any smooth function L on R × Rn × Rn defines a functional t01 L(t, x(t), over the admissible curves x(t) in Rn . The basic problem of the calculus of variations (in the strong sense) is to find conditions under which a t curve x(t) ¯ on an interval I = [t0 , t1 ] attains a minimum of t01 L(t, x(t), dx dt (t)) dt over all admissible curves x(t) that satisfy the same boundary conditions as x(t) ¯ in some strong neighborhood of x(t). ¯ This problem admits a natural reformulation in control theoretic terms: addmissible curves in O (x) are the solutions in O (x) of the “control” system dx dt (t)) dt

dx = ui (t)ei , dt n

(6.5)

i=1

where e1 , . . . , en denote the standard basis in Rn , and where the controls functions u1 (t), . . . , un (t) are measurable curves on the interval [t0 , t1 ]. Every ¯ is a trajectory of (6.5) generated by the control admissible curve in O (x) dx u(t) = dt (t), but the converse may not be true, since a solution of (6.5) that ¯ at t0 may not remain in O (x) ¯ for all t ∈ [t0 , t1 ]. originates in O (x) The problem of finding conditions under which x(t) ¯ yields a minimum for t1 dx t0 L t, x(t), dt dt over all admissible curves in some strong neighborhood O (x) ¯ of x(t) ¯ subject to the boundary conditions x(t0 ) = x(t ¯ 0 ) and x(t1 ) = ¯ = ddtx¯ x(t ¯ 1 ), is the same as the problem of finding conditions under which u(t) is an optimal control for the control system (6.5) on O (x) ¯ relative to the ) = x(t ¯ ), x(t ) = x(t ¯ ) and the cost functional boundary conditions x(t 0 0 1 1 t1 L(t, x(t), u(t))) dt. t0 Weierstrass’ revolutionary discovery was that along an optimal trajectory x(t) ¯ the following inequality must hold: dx¯ ∂L dx¯ dx¯ − L(t, x(t), ¯ u(t)) − − u(t) · t, x(t), ¯ ≥ 0, (6.6) L t, x(t), ¯ dt dt ∂u dt

6.3 The Maximum Principle

71

for any curve u(t) ∈ Rn . The function E(t, x, y, u) = L(t, x, y) − L(t, x, u) − (y − u)

∂L (t, x, u)|u=y ∂u

became known as the excess function of Weierstrass [Cr]. So, in contrast to the Euler–Lagrange equation, which appears as a necessary condition for optimality in the weak sense, the inequality of the excess function of Weierstrass is a necessary condition of optimality in the strong sense. As far as I can tell, the two conditions have never been properly integrated in the classical literature on this subject. [Cr; Yg]. We will presently show that the inequality of Weierstrass may be considered as a harbinger of the Maximum Principle. In regard to weak versus strong optimality, it is clear that an optimal curve in the strong sense that is continuously differentiable, is also optimal in the weak sense, but the converse may not be true. For instance, x(t)= 0 is weakly dx 2 2 optimal for the functional defined by L = ||x(t)|| 1 − || dt || , because L ≥ 0 in a sufficiently small weak neighborhood of x(t) = 0. That is no longer the case in the strong sense, no matter how small the neighborhhood.

6.3 The Maximum Principle The Maximum Principle states that each extremal trajectory of a control system is the projection of an integral curve of a certain Hamiltonian system on the cotangent bundle of M. In order to be more explicit, we will restore the notations and the basic material introduced in the previous chapter concerning the symplectic structure of the cotangent bundle. Then θ will denote the Liouville form, and ω = −dθ will denote the symplectic form on T ∗ M. Points of T ∗ M will be generally denoted by ξ and their projections on M by x. The starting point is the observation that each control system can be lifted to a Hamiltonian system on the cotangent bundle. This lifting is based on the following: Definition 6.9 The Hamiltonian lift of a time-varying vector field Xt on a manifold M is the Hamiltonian vector field ht on T ∗ (M) that corresponds to the function ht (ξ ) = θξ (Xt (x)) = ξ ◦ Xt (x) for ξ ∈ Tx∗ M. Remark 6.10 Time-varying vector fields make a smoother transition to control systems.

72

6 Hamiltonians and optimality: the Maximum Principle

Let us first examine this definition in terms of the canonical coordinates. If Xt = ni=1 X i (t, x1 , . . . , xn ) ∂x∂ i then ht (ξ ) = ξ ◦ Xt (x) is given by h(t, x1 , . . . , xn , p1 , . . . , pn ) =

n

pi X i (t, x1 , . . . , xn ).

i=1

It follows that the integral curves of ht are the solutions of the following system of equations: dpi dxi (t) = X i (t, x1 (t) . . . , xn (t)), (t) dt dt n ∂ pj (t) X j (t, x1 (t) . . . , xn (t)), i = 1, . . . , n =− ∂xi

(6.7)

j=1

This system is the adjoint companion of the variational equation dxi dvi ∂ (t) = Xti , (x1 (t) . . . , xn (t)), (t) = vj (t) X i (t, x1 (t) . . . , xn (t)). dt dt ∂xj n

j=1

(6.8) We now highlight some facts which will be relevant for the general situation. We first note that the projection of the variational equation on the second factor is a linear time-varying system dv dt = A(t)v(t), with A(t) the matrix with ∂ i entries Aij (t) = ∂xj X (t, x(t) . It then follows from the theory of differential equations that v(t) is defined over the same time interval as the solution curve x(t) [CL]. Moreover, the companion curve p(t) in (6.7), being the solution of ∗ the adjoint system dp dt (t) = −A (t)p(t), is also defined over the same interval as the underlying curve x(t). Hence the integral curves of ht exist over the same intervals as the integral curves of Xt . Secondly, the solutions v(t) of the variational equation and the solutions p(t) of the Hamiltonian equation are related through a simple relation n

pi (t)vi (t) = constant.

(6.9)

i=1

Condition (6.9) gives rise to the “forward-backward” principle: The solutions of the variational equations propagate forward in time, while the solutions of the adjoint equation propagate backward in time, and their scalar product remains constant for all times. Let us now elevate these observations to the general setting independent of the particular choice of coordinates. Let t,τ denote the time-varying flow of Xt defined by x(t) = t,τ (x0 ), where x(t) denotes the solution of dx dt = Xt (x(t))

6.3 The Maximum Principle

73

with x(τ ) = x0 . The uniqueness theorem of solutions with given initial data implies that the flow satisfies the following composition rule:

t,τ = t,s ◦ s,τ ,

(6.10)

whenever this composition is well defined. Conversely, every time-varying flow that satisfies (6.10) is generated by a time-varying vector field Xt . The tangent map ( τ ,t )∗ maps Tx(τ ) M onto Tx(t) M, and the dual mapping ∗ M onto T ∗ M. Let ∗ ( τ ,t )∗ maps Tx(t) τ ,t = t,τ . Since the order of the x(τ ) ∗ ∗ variables is reversed, τ ,t maps Tx(τ ) M onto Tx(t) M. Moreover, τ ,s ◦ s,t (ξ ) = τ ,t (ξ ◦ ( t,s )∗ = ξ ◦ ( t,s )∗ ( s,τ )∗ = ξ ◦ ( t,τ )∗ = τ ,t (ξ ). It follows that τ ,t is a time-varying flow on T ∗ M, and hence is generated by a time-varying vector field on T ∗ M. Proposition 6.11

τ ,t is the flow of the Hamiltonian lift ht of Xt .

Proof Let Vt denote the infinitesimal generator of the flow τ ,t , i.e., Vt = ∂ ∗ ∂t τ ,t |τ =t . Then π(τ ,t (ξ )) = x(t) for every ξ ∈ Tx(τ ) , where π denotes the ∗ natural projection from T M onto M. Therefore, π∗ (Vt (ξ )) = Xt (π(ξ )), ξ ∈ Tx∗ M. ∗ (M) then (w(s)) belongs If w(s) is any curve in T ∗ M such that w(s) ∈ Ty(s) τ ,t to the cotangent space of M at τ ,t (y(s)). Hence the projection π(τ ,t w(s)) is equal to τ ,t (y(s)). It follows that

π∗ ((τ ,t )∗ W) = ( τ ,t )∗ v, where W = dw ds (0) and τ ,t (w(0)). Then,

dy ds (0)

= v. Let now ξ(t) denote the curve t →

θξ(t) ((τ ,t )∗ W) = ξ(t) ◦ π∗ ((τ ,t ))∗ W) = w(0) ◦ t,τ ∗ ◦ { τ ,t )∗ v = w(0)(v) and, hence, dtd (θξ(t) ((τ , )∗ W)|t=τ = 0. Since dtd (θξ(t) ((τ , )∗ W)|t=τ is the same as the the Lie derivative LVt (θ )(W), we can use Cartan’s formula to write the preceding equality as 0 = LVt (θ ) = d ◦ iVt θ + iVt ◦ dθ . But iVt θξ = ξ(Xt (x) for all ξ ∈ Tx∗ (M) because π∗ (Vt ) = Xt . Hence dht = iVt ◦ ω.

74

6 Hamiltonians and optimality: the Maximum Principle

Let us now consider the Hamiltonian lift of the control system (6.11). Each control functions u(t) defines a time-varying Hamiltonian hu(t) , with hu(t) (ξ ) = ξ ◦ F(x, u(t)), ξ in Tx∗ M, which, in turn, induces the Hamiltonian vector field = {hu(t) : u ∈ U } hu(t) . The family of time-varying Hamilonian vector fields H conforms is the Hamiltonian lift of the control system (6.1). Every flow in H to the same regularity conditions as the underlying system (6.1), and hence, dξ (t) = hu(t) (ξ(t)) dt admits a unique solutions ξ(t) through each initial point ξ0 , with the understanding that the differential equation is satisfied only at the points where ξ(t) is differentiable. Theorem 6.12 Maximum Principle–version 1 Suppose that x(t) ¯ is a trajectory generated by a control u(t) ¯ such that x(T) ¯ belongs to ∂A(x0 ). Then x(t) ¯ is the projection of a curve ξ¯ (t) in T ∗ M defined on [0, T] such that: 1. 2. 3. 4.

dξ¯ dt (t)

= hu(t) ¯ (ξ¯ (t)) for almost all t ∈ [0, T]. ¯ξ (t) = 0 for any t ∈ [0, T]. hu(t) ¯ (ξ¯ (t)) = 0 for almost all t ∈ [0, T]. hu (ξ¯ (t)) ≤ 0 for any u in U for almost all t ∈ [0, T].

Curves ξ(t) in T ∗ M that satisfy conditions (1), (2), (3), and (4) of the Maximum Principle are called extremals. The Maximum Principle then says that each exremal trajectory x(t) in the base space M is the projection of an extremal curve in ξ(t) in T ∗ M. Proof The proof of the Maximum Principle can be divided into three essentially independent parts. The first part deals with the infinitesimal variations along an extremal curve and the cone of attainability at the terminal point x(T). ¯ The second part consists in showing that the cone of attainability cannot ¯ belongs to the boundary of be the entire tangent space Tx(T) ¯ M when x(T) the attainable set. The third, and final part, consists of choosing an integral that satisfies the Maximum Principle. We break up the proof curve ξ¯ of hu(t) ¯ accordingly, into three separate parts. Throughout the proof we will assume that x(t) ¯ is a trajectory generated by u(t) ¯ that belongs to the boundary of A(x0 ) at some time T > 0. Then ¯ = x(t) ¯ will denote the local flow of the time-varying vector field t,s (x(s)) = F( x(t), ¯ u(t)). ¯ We will make use of the following auxiliary facts: Xu(t) ¯ (a) There exists a tubular neighborhood ∪{Os : 0 ≤ s ≤ T} with each Os an open neighborhood of x(s) ¯ such that x(t) = t,s y is defined for all t ∈ [0, T], for each y ∈ Os (this fact follows from a theorem on the

6.3 The Maximum Principle

75

continuity of solutions of differential equations with respect to the initial data [CL]). (b) The next statement concerns the nature of regular points along the curve x(t). ¯ A point τ is a regular point of x(t) ¯ if x(t) ¯ is differentiable at t = τ and dλ d (x(τ ¯ + λ(s))|s=0 = (0)Xu(τ ¯ )) ¯ ) (x(τ ds ds for any differentiable curve λ(s) ∈ R defined in an open interval around zero such that λ(0) = τ . Every Lebesgue point of u(t) ¯ is a regular point of x(t). ¯ Since Lebesgue points of any integrable function have full measure on any compact interval, it follows that almost every point t ∈ [0, T] is a regular point of x(t). ¯ In particular, regular points are dense in [0, T]. Recall that a point τ is called a Lebesgue point of an integrable curve τ 1 u(t) in Rm if limt→τ |t−τ | t |ui (s) − ui (τ )| ds = 0, i = 1, . . . , m. Then, τ d dτ a u(s) ds = u(τ ) at every Lebesgue point of u(t). But then it can be shown that 1 τ 1 τ lim Xu(s) ¯ ds = lim F(x(s), ¯ u(s)) ¯ ds = Xu(τ ¯ ). ¯ (x(s)) ¯ ) (x(τ →0 τ − →0 τ − (The basic theory of Lebesgue points can be found in Theory of Functions of a Real Variable by I. P. Natanson [Na]. The relevant facts are also discussed in the original publication on the Maximum Principle [Pt]. Part 1: The cone of attainability We shall use the family F = {Xu : Xu (x) = F(x, u), x ∈ M, u ∈ U} to generate the perturbations of of the flow along x(t). ¯ Vector fields Xu in F u are time invariant, hence, their flows { s : s ∈ R} are defined by a single parameter s, in contrast to the flow defined by a time-varying control u(t). Each Xu in F defines a perturbed trajectory t → σ (t, τ , u, λ1 , λ, s) with ¯ − sλ1 ), σ (t, τ , u, λ1 , λ, s) = t−sλ,τ ◦ uλ1 s ◦ x(τ

(6.11)

for s sufficiently small and positive, and arbitrary numbers λ and λ1 with λ1 positive. The perturbed trajectory follows x(t) ¯ up to τ − sλ units of time, then switches to Xu , which it follows for sλ1 units of time, and then it switches back to Xu(t) ¯ , which it follows for the rest of the time. At each time t > τ , this perturbed trajectory is in the reachable set from x0 and, moreover, σ (t, τ , u, λ, s) → x(t) ¯ uniformly as s tends to zero. At regular points τ and t, the curve s → σ (t, τ , u, λ, s) is differentiable at s = 0 and defines a tangent vector at Tx(t) ¯ M given by ¯ + λ1 t,τ ∗ (Xu − Xu(τ ¯ )). v(t, τ , u) = λXu(t) ¯ (x(t)) ¯ ) )(x(τ

(6.12)

76

6 Hamiltonians and optimality: the Maximum Principle

This vector is called an elementary perturbation vector of x(t). ¯ These vectors are a subset of a more general class of tangent vectors obtained by differentiating more complex perturbations. These perturbed trajectories are of the form xs (τ ) = τ −sλ,τm ◦ (s), where (s) is a local diffeomorphism generated by the composition of elementary perturbations

(s) = usλmm ◦ τm −sλm ,τm−1 ◦ · · · ◦ τ3 −sλ3 ,τ2 ◦ usλ2 2 ◦ τ2 −sλ2 ,τ1 ◦ usλ1 1 ◦ x(τ ¯ 1 − sλ1 ),

(6.13)

at regular points 0 < τ1 ≤ τ2 · · · ≤ τm , τm ≤ τ . The perturbed trajectory xs (t) follows x(t) ¯ for τ1 −sλ1 units of time, then switches to Xu1 for sλ1 units of time. The terminal point of this piece of trajectory is then moved along Xu(t) ¯ until the time t = τ2 − sλ2 , after which the switch is made to Xu2 for sλ2 units of time. Then the flow reverts back to Xu(t) ¯ for t = τ2 − sλ3 units of time, and then the trajectory switches to Xu3 for sλ3 units of time. This process is repeated in time succession until the last switch to Xum is completed, upon which the trajectory follows Xu(t) ¯ until the terminal time τ − sλ. The terminal point xs (τ ) is in A(x0 ) for any positive numbers λ1 , . . . , , λm and any number λ, provided that s is a sufficiently small positive number. The tangent vector v at x(τ ¯ ), obtained by differentiating xs (τ ) with respect to s at s = 0, is of the form ¯ )) + v = λXu(τ ¯ ) (x(τ

m

λi τ ,τi ∗ (Xui − Xu(τ ¯ i )). ¯ i ) )(x(τ

(6.14)

i=1

as can be easily verified by induction on the number of regular points. The set of vectors v given by (6.14) may be regarded as the positive convex cone generated by the elementary perturbation vectors (for the same terminal point τ ). We shall use C(τ ) to denote its topological closure. Since ¯ 2 )) = τ2 ,τ1 ∗ Xu(τ ¯ 1 )), Xu(τ ¯ 2 ) (x(τ ¯ 1 ) (x(τ τ2 ,τ1 ∗ C(τ1 ) ⊆ C(τ2 ),

(6.15)

for any regular points τ1 and τ2 with τ1 < τ2 . * Definition 6.13 The cone K(T) = {C(t), t ≤ T, t regular} is called the cone of attainability at x(T). ¯ Lemma 6.14

K(T) is a closed convex cone.

Proof If v ∈ K(T), then v ∈ C(t) for some regular time t. Since C(t) is a cone, αv ∈ C(t) for all α ≥ 0. Hence, αv ∈ K(T). If v1 = T,τ1 ∗ c1 for some

6.3 The Maximum Principle

77

c1 ∈ C(t1 ) and if v2 = T,τ2 ∗ c2 for some c2 ∈ C(t2 ) with τ1 < τ2 , then v1 = T,τ2 ∗ τ2 ,τ1 ∗ c1 . Now both vectors c2 and τ2 ,τ1 ∗ c1 belong to C(τ2 ) hence, their sum belongs to C(τ2 ). This implies that v1 + v2 belongs to K(T). Part 2 Lemma 6.15 The cone of attainability K(T) is not equal to Tx(T) ¯ M whenever x(T) ¯ ∈ ∂A(x0 ). Proof Suppose that K(T) = Tx(T) ¯ M. Let w1 , . . . , vn be any points in K(T) such that the convex cone generated by these points is Tx(T) ¯ M. Then there exist regular points τ1 , . . . , τn and vectors c1 , . . . , cn , such that each ci ∈ C(τi ) and wi = (T,τi )∗ ci = wi . There is no loss in generality in assuming that τ1 ≤ τ2 ≤ · · · ≤ τn . Then vectors vi = (τn ,τi )∗ ci , i = 1, . . . , n all belong to C(τn ) and the positive convex cone generated by these vectors is equal to Tx(τ ¯ n ) M. The preceding argument shows that K(T) = Tx(T) ¯ M implies that C(τ ) = Tx(τ ¯ ) M for some regular point τ ≤ T. Since C(τ ) is the closure of the convex cone spanned by the elementary perturbation vectors, there exist elementary vectors v1 , . . . , vm in C(τ ) such that the convex cone generated by v1 , . . . , vm is equal to Tx(τ ¯ ) M. Let τ1 , . . . , τm denote regular points in the interval [0, τ ] such that vi = λXu(τ ¯ )) + λi (τ ,τi )∗ (Xui − Xu(τ ¯ i )), i = 1, . . . , m. ¯ ) (x(τ ¯ i ) )(x(τ There is no loss in generality in assuming that all these regular points are distinct, and renumbered, so that 0 = τ0 < τ1 , τ2 < · · · < τm . Let Fi (si ) =

usii τi −si ,τi−1 , i = 1, . . . , m, and then let F(s1 , . . . , sm ) = τ +s1 +···+sm ,τm ◦ Fm (sm ) ◦ Fm−1 (sm−1 ) ◦ · · · ◦ F1 (s1 )(x0 ). (6.16) It follows that F(0) = x(T) ¯ and that there is neighborhood O of 0 in Rm such m m that F(O ∩ R+ ) ⊆ A(x0 ), where Rm + denotes the orthant {s ∈ R : si ≥ 0, i = 1, . . . , m}. ∂F (0) = vi , i = 1, . . . , m, and since the convex cone spanned by Since ∂s i m v1 , . . . , vm is equal to Tx(τ ¯ ) M, dF0 (R+ ) = Tx(τ ¯ ) M, where dF0 denotes the tangent map of F at s = 0. But then, according to the Generalized Implicit Function theorem [[AS], Lemma 12.4, p. 172], x(T) ¯ = F(0) ∈ intF(O ∩ Rm + ), which contradicts our assumption that x(T) ¯ is on the boundary of cA(x0 ).

78

6 Hamiltonians and optimality: the Maximum Principle

Part 3: The extremal curve We will now revert to the notations established earlier, with t,τ the flow ¯ defined by (t,τ )∗ (v) = induced by Xu(t) ¯ , (t,τ )∗ its tangent flow along by x(t), d ( )(σ (s)| , where σ (s) is a curve that satisfies s(0) = x(τ ¯ ) and dσ t,τ s=0 ds ds (0) = ∗ v, and t,τ the dual of t,τ . Since K(T) is a proper closed convex cone there exists a covector ξ1 ∈ ∗ M such that ξ (v) ≤ 0 for each v ∈ K(T). Moreover, ξ (X ¯ = 0, Tx(T) 1 1 u(T) ¯ (x(T)) ¯ ¯ is in K(T) for any real number α. since αXu(T) ¯ (x(T)) Let ξ¯ (t) = ∗T,t (ξ1 ). It then follows from Proposition 6.11 that ξ¯ (t) is an ¯ It remains to show integral curve of hu(t) ¯ whose projection on M is equal to x(t). that ξ¯ (t) satisfies the conditions of the Maximum Principle. If t is a regular point and if v ∈ C(t), then vT = T,t ∗ (v) belongs to K(T) and hence, ξ¯ (t)(v) = ξ¯ (T)(vT ) ≤ 0 by the “forward-backward principle.” This shows that ξ¯ (t)(Xu − ¯ ≤ 0 and that ξ¯ (t)Xu(t) ¯ = 0 for any regular point t. Since almost Xu(t) ¯ (x(t)) ¯ (x(t)) all points are regular, 0 = hu(t) ¯ (ξ¯ (t)) ≥ hu (t)(ξ¯ )(t), u ∈ U almost everywhere in [0, T]. There is another version of the Maximum Principle pertaining to the trajectories x(t) whose terminal point x(T) belongs to the boundary of A(x0 , T). This version is given by our next theorem. Theorem 6.16 The Maximum Principle–version 2 Suppose that x(t) is a trajectory generated by a control u(t) ¯ such that the terminal point x(T) belongs to ∂A(x0 , T). Then x(t) is the projection of a curve ξ(t) in T ∗ M defined on [0, T] such that: ¯ (ξ(t)) for almost all t ∈ [0, T]. 1. dξ dt (t) = hu(t) 2. ξ(t) = o for any t ∈ [0, T]. 3. hu(t) ¯ (ξ(t)) is constant on [0, T]. 4. hu(t) ¯ (ξ(t)) ≥ hu (ξ(t)) for any u in U for almost all t ∈ [0, T]. ˜ = M × R, and then consider an Proof Let (x, y) denote the points of M ˜ given by extended control system on M dy dx = F(x(t), u(t)), = 1, (6.17) dt dt over the bounded and measurable control functions that take values in a given ˜ x0 ) denote the set of points in M ˜ that are reachable from set U in Rm . Let A(˜ x˜ 0 = (x0 , 0) by the trajectories of (6.17). A point x˜ 1 = (x1 , T) is reachable by a trajectory x˜ (t) = (x(t), y(t)) from x˜ 0 if and only if x1 = x(T). Morever, ˜ x0 ) whenever x(T) ∈ ∂A(x0 , T). x˜ (T) ∈ ∂ A(˜

6.3 The Maximum Principle

79

Therefore, Theorem 6.12 is applicable to the trajectory x˜ (t) = (x(t), y(t)) ˜ Since T ∗ M ˜ = generated by the control u(t). ¯ Let ξ˜ denote the points of T ∗ M. ∗ ∗ ∗ ∗ ∗ ˜ T M × T R, ξ = (ξ , λ) with ξ ∈ T M and λ ∈ T R. If T R is represented by R∗ × R, then each point λ ∈ Ty∗ R can be written as λ = (p, y). Then ˜ u(t) h˜ u(t) ¯ (ξ˜ ) = hu(t) ¯ (ξ ) + p. Let ξ˜ (t) = (ξ(t), λ(t)) denote an integral curve of h ¯ that satisfies the Maximum Principle in Theorem 6.12. Then, dξ ∂ h˜ u(t) dp ¯ = hu(t) =− = 0. ¯ (ξ(t)), dt dt ∂y ˜ u(t) Thus, p is constant. But then hu(t) ¯ (ξ(t)) is constant, since h ¯ (ξ˜ (t)) = 0. ˜ ˜ ˜ ˜ Finally, hu(t) ¯ (ξ (t)) ≥ hu (ξ (t)) implies that hu(t) ¯ (ξ(t)) ≥ hu (ξ(t)) almost everywhere. Alternatively, one could have proved this version of the Maximum Principle by modifying the perturbations in the proof of Theorem 6.12 so that the endpoint of the perturbed trajectory traces a curve in A(x0 , T), as was done in [AS]. Simply restrict to λ = 0 in (6.11), and (6.13). Then, the proof in Theorem 6.12 leads to the inequality hu(t) ¯ (ξ(t)) ≥ hu (ξ(t)) for any u in U for almost all t ∈ [0, T]. The remaining part of the proof follows from the following proposition. Proposition 6.17 Let ξ(t) be an integral curve of the Hamiltonian hu(t) defined in the interval [0, T] by a bounded and measurable control u(t). If ξ(t) satisfies hu(t) (ξ(t)) ≥ hv (ξ(t)), v ∈ U, at all points t of [0, T] where hu(t)) (ξ(t)) is differentiable, then hu(t) (ξ(t)) is constant in [0, T]. Proof Let K equal to the closure of {u(t) : t ∈ [0, T]}. Since u(t) is bounded, K is compact. By the usual estimates on vector fields over compact domains one can show that function φ(ξ ) = Maxu∈K {hu (ξ ) : u ∈ K} is Lipschitzian as a function on T ∗ M. Therefore, hu(t) (ξ(t)) is a Lipschitzian curve and hence, absolutely continuous. We will next show that dtd φ(ξ(t)) = 0 a.e. in [0, T]. Since φ(ξ(t)) = hu(t) (ξ(t)) almost everywhere, this would imply that dtd hu(t) (ξ(t)) is equal to 0 almost everywhere, which in turn would imply that hu(t) (ξ(t)) is constant. Let τ denote a Lebesgue point for u(t). Then φ(ξ(t)) is differentiable at t = τ and φ(ξ(τ )) = hu(τ ) (ξ(τ )). For points t > τ , φ(ξ(t)) − φ(ξ(τ )) hu(τ ) (ξ(t)) − hu(τ ) (ξ(t)) ≥ . t−τ t−τ

80

6 Hamiltonians and optimality: the Maximum Principle

After taking the limit as t → τ , the above yields dtd φ(ξ(t))|t=τ ≥ dtd φu(τ ) (ξ(t))|t=τ = {hu(τ ) , hu(τ ) }(ξ(t)) = 0. The same argument with t < τ shows that dtd φ(ξ(t))|t=τ ≤ 0. Therefore, dtd φ(ξ(t))|t=τ = 0. We will now adapt the Maximum Principle to problems of optimal control. The Hamiltonian lift of the cost-extended system dx dy (t) = f (x(t), u(t)), (t) = F(x(t), u(t)) dt dt ˜ where M ˜ = R × M. Then T ∗ M ˜ = T ∗ R × T ∗ M. Points ξ˜ takes place in T ∗ M, ∗ ˜ will be written as ξ˜ = (ξ0 , ξ ). In these notations, the Hamiltonian lift in T M is given by ∗ ˜ ˜ hu(t) ¯ (ξ ) = ξ0 fu(t) (x) + ξ Fu(t) (x), ξ˜ ∈ T(y,x) M.

Since this Hamiltonian does not depend explicitly on the variable y, ξ0 (t) is constant along each integral curve (ξ0 (t), ξ(t)) of hu(t) . According to the Maximum Principle, x˜ (t) is the projection of an extremal curve ξ˜ (t) defined by the Hamiltonian lift ˜ hu(t) ¯ (ξ ) = ξ0 fu(t) + ξ Fu(t) ∗ M. ˜ For optimal control problems in which the cost functional for ξ˜ ∈ T(y,x) is minimized ξ0 ≤ 0. Since the Hamiltonian is a homogeneous function, this constant can be reduced to −1 whenever it is not equal to zero. The Hamiltonian hu(t)(ξ˜ ) = ξ0 fu(t) + ξ Fu(t) is usually identified with its projection on T ∗ M in which ξ0 is regarded as a parameter. The values of ξ0 give rise to two kinds of extremal curves – normal and abnormal. Extremal curves which are generated by the Hamiltonian vector field that correspond to hu with ξ0 = −1 are called normal and the extremals generated by hu (ξ ) = ξ ◦ Fu are called abnormal. In this context the Maximum Principle lends itself to the following formulation.

Proposition 6.18 The Maximum Principle of optimality The optimal trajectories (x(t), u(t)) are either the projections of normal extremal curves ξ(t) generated by the Hamiltonian hu (ξ ) = −f (x, u) + ξ ◦ Fu (x), or they are the projections of abnormal extremal curves generated by the Hamiltonian hu (ξ ) = ξ ◦Fu (x). The abnormal extremal curves cannot be equal to zero at any time t ∈ [0, T]. In either case, the extremal curves conform to the maximality condition hu(t) (ξ(t)) = ξ0 f (x(t), u(t)) + ξ(t) ◦ Fu (t)(x(t)) ≥ ξ0 f (x(t), v) + ξ(t) ◦ Fv (x(t)) (6.18)

6.4 The Maximum Principle in the presence of symmetries

81

for any v ∈ U and almost all t ∈ [0, T]. For optimal problems with a variable time interval, hu(t) (ξ(t)) = 0, a.e. on [0, T]. The Maximum Principle also covers the case where a two-point boundary value optimal problem is replaced by an optimal control problem of finding a trajectory (x(t), u(t)) that originates on a submanifold S0 of M at t = 0 and terminates on a submanifold S1 of M at t = T and minimizes the T integral 0 f (x(t), u(t)) dt among all such trajectories. For such problems, the Maximum Principle imposes additional conditions known as the transversality conditions: If (x(t), u(t)) is the projection of an extremal curve ξ(t) then ξ(0)(v) = 0 for all tangent vectors v in the tangent space of S0 at x(0) and ξ(0)(v) = 0 for all tangent vectors v in the tangent space of S1 at x(T).

6.4 The Maximum Principle in the presence of symmetries For Lie determined systems, the perturbations could be extended to vector fields in the Lie saturate to obtain the following version of the Maximum Principle. Theorem 6.19 The saturated Maximum Principle Suppose that F consisting of vector fields Xu (x) = F(x, u), u ∈ U is Lie determined, and suppose that ¯ u(t)) ¯ such that x(T) ∈ ∂A(x0 ). Then x(t) ¯ is x(t) ¯ is a solution of ddtx¯ = F(x(t), ∗ ¯ the projection of a curve ξ (t) in T M defined on [0, T] such that: ¯ 1. ddtξ (t) = hu(t) ¯ (ξ¯ (t)) for almost all t ∈ [0, T]. 2. ξ¯ (t) = o for any t ∈ [0, T]. 3. hu(t) ¯ (ξ¯ (t)) = 0 for all t ∈ [0, T]. 4. hX (ξ¯ (t)) ≤ 0 for any X in LS(F) for almost all t ∈ [0, T] where hX denotes the Hamiltonian lift of X.

Proof If the cone of attainability K(T) corresponding to the perturbations in the Lie saturate were equal to Tx(t) ¯ M, then, as in the proof of the Maximum Principle, that would imply that the terminal point x(T) ¯ is in the interior of the ¯ is in the interior of A(x0 ), closure of A(x0 ), which then would imply that x(T) as a consequence of Proposition 6.2. This implication violates the original assumption that x(T) ¯ belongs to the boundary of A(x0 ). In general, it is not true that cl(int(A(x, T))) = int(A(x, T)), and therefore, cl(A(x, T)) is not a useful object to generate an analogous extension of the Maximum Principle–version 2. However, in the presence of symmetries,

82

6 Hamiltonians and optimality: the Maximum Principle

condition (4) of the Maximum Principle–version 2 could be extended to a larger family to give additional information about the extremal trajectory. To elaborate, let us first define the basic terms. Definition 6.20 A diffeomorphism is a symmetry for control system (6.1) if −1 (A( (x), T)) = A(x, T), x ∈ M. If is a symmetry, then −1 A(x, T) = A( −1 x, T), which in turn implies that −1 is a symmetry as well. We will use Sym(F) to denote the group of symmetries associated with control system (6.1). Evidently, any diffeomorphism which is a symmetry also satisfies a weaker condition

−1 (A( (x))) = A(x), x ∈ M; hence, Sym(F) is contained in the group of normalizers for F. Therefore, { (F); ∈ Sym(F), X ∈ F} ⊆ LS(F). However, a normalizer need not be a symmetry, as the following example shows. Example 6.21 Let M be the groupof motions of the plane. Then each matrix 1 0 g in M can be represented as g = , x ∈ R2 , R ∈ SO2 (R). Consider x R the Serret–Frenet system ⎛ ⎞ 0 0 0 dg = g(t) ⎝ 1 0 −u(t) ⎠ , dt 0 u(t) 0

1 0 with u(t) playing the role of control. For any solution g(t) = , x(t) R(t) T is the length of x(t) in the interval [0, T] and u(t) is the signed curvature of x(t). It follows that A(g0 , T) is not equal to M for any T > 0. Since A(g0 ) = M for any g0 ∈ M, any diffeomorphism is a weak symmetry for the system, but only the left-translations are the symmetries of the system. Proposition 6.22 Suppose that the control system admits a group of symmetries Sym(F). Then each extremal curve ξ(t) in Theorem 6.16 satisfies the strengthened condition; hu(t) ¯ (ξ(t)) ≥ hX (ξ(t)),

(6.19)

for any X ∈ { (F); ∈ Sym(F), X ∈ F}, and for almost all t ∈ [0, T]. The proof is analogous to the proof of the saturated Maximum Principle and will be omitted.

6.4 The Maximum Principle in the presence of symmetries

83

Definition 6.23 A vector field X is said to be a symmetry for system (6.1) if its one-parameter group of diffeomorphisms { t : t ∈ R} is a subgroup of Sym(F)). Proposition 6.24 Noether’s theorem Suppose that a vector field X is a symmetry for system (6.1). Then the Hamiltonian hX associated with the Hamiltonian lift of X is constant along each extremal curve ξ(t) defined by Theorem 6.16. Proof Let ξ(t) denote an extremal curve generated by a control function u(t) ¯ that projects onto the trajectory x(t) and let { s : s ∈ R} denote the group of diffeomorphisms generated by X. As before, ( s ) Y = ( s )∗ Y −s . The strengthened maximality condition implies that hu(t) ¯ (ξ(t) ≥ h( s ) Y (ξ(t)), a.e. ∈ [0, T],

(6.20)

for any Y ∈ F and any s. This inequality is preserved if Y is replaced by any time-varying vector field Yu(t) (x) = F(x, u(t)), and in particular it is preserved ¯ when u(t) = u(t). ¯ Let Xt (x) = F(x, u(t)). The resulting inequality shows that h s Xt (ξ(t)) attains its maximum for d h s Xt (ξ(t))|s=0 = 0. This means that s = 0 and therefore, ds 0= Since

d h X (ξ(t))|s=0 = ξ(t)([X, Xt ](x(t) = {hX , hu(t) ¯ }(ξ(t). ds s t

d dt hX (ξ(t))

= {hX , hu(t) ¯ }(ξ(t)) = 0, hX (ξ(t)) is constant.

To recognize Noether’s theorem in its more familiar form we need to return to optimal control problem and the cost-extended system (6.2) dx dy (t) = f (x(t), u(t)), (t) = F(x(t), u(t)) dt dt

T associated with the optimal control problem of minimizing 0 f (x(t), u(t)) dt over the trajectories of dx dt = F(x(t), u(t)), u(t) ∈ U that satisfy the given boundary conditions on a fixed time interval [0, T]. The following proposition is a corollary of the preceding proposition. Proposition 6.25 Suppose that a vector field X on M with its flow { s } satisfies ( s )∗ Fu ◦ −s = Fus for some us ∈ U and all s ∈ R, and f ( s (x), us ) = f (x, u) for all s ∈ R and all u ∈ U. Then the vector field X˜ = (0, X) is a symmetry for the extended system (6.2). Consequently, the Hamiltonian hX (ξ0 , ξ ) = ξ ◦ X(x) is constant along each extremal curve (ξ0 , ξ(t)). ˜ s } denote the flow for the extended vector field X. ˜ If z(t) = Proof Let { ˜ (y(t), x(t)) is any trajectory of (6.2), let w(t) ˜ = (w0 (t), w(t)) = s (z(t). Then

84

6 Hamiltonians and optimality: the Maximum Principle

dw0 dt

= f (x(t), u(t)) = f ( s (x(t), us (t))) and dw dt = s∗ F( −s (x(t), u(t))) = F(w(t), us (t)). ˜ s (z(0)). It follows that Therefore, w(t) ˜ is a solution of (6.2) with w(0) ˜ = ˜ ˜ ˜

s A(˜x0 , T) ⊆ A( s x˜ 0 , T). Hence, X is a symmetry for (6.2). Since hX˜ (ξ0 , ξ ) = ξ(X(x) = hX (ξ ), hX (ξ(t)) is constant along any extremal curve ξ˜ (t) = (ξ0 , ξ(t)) by the previous proposition. Suppose now that G is a Lie group which acts on a manifold M. If φ : G × M → M denotes this action, let g denote the induced diffeomorphism

g (x) = φ(g, x). The mapping g → g is a homomorphism from G into the group of diffeomorphisms on M, that is, h ◦ g = hg for any h, g in G. If g denotes the Lie algebra of G, the above implies that { etA : t ∈ R} is a one-parameter group of diffeomorphisms on M for each exponental etA in G defined by A ∈ g. Let XA to denote its infinitesimal generator. It follows that XA is a complete vector field on M. In the terminology of [Jc], XA is a vector field subordinated to G. It is easy to show that the correspondence A → XA satisfies XαA+βB = αXA + βXB and X[A,B] = [XA , XB ]. Let hA denote the Hamiltonian lift of XA , that is, hA (ξ ) = ξ(XA (π(ξ ))), ξ ∈ For a fixed ξ , the mapping A → hA (ξ ) is linear, hence can be identified with an element (ξ ) in the dual g∗ of g.

T ∗ M.

Definition 6.26 The mapping : T ∗ M → g∗ is called the moment map [AM]. Proposition 6.27 Suppose that G is a symmetry group for a control system dx dt = F(x(t), u(t))), u(t) ∈ U. Then the moment map is constant along each extremal curve ξ(t) defined by Theorem 6.16. Proof A paraphrase of the proof in Proposition 6.24 shows that hA (ξ(t)) is constant along each extremal curve ξ(t) for any A ∈ g. This means that (ξ(t)) is constant. The action of G on M can be extended to an action on T ∗ M by the formula = ξ ◦ g−1 ∗ for each ξ ∈ Tx∗ M, where g ∗ denotes the tangent map of

g . It follows that ∗h ◦ ∗g = ∗hg , and hence the mapping (g, ξ ) ∈ G×Tx M → ( ∗g (ξ ) ∈ Tx∗ M is an action under G. The moment map carries the following equivariance property:

∗g (ξ )

( ∗g (ξ )) = Adg∗ ()(ξ ), where Adg∗ denotes the coadjoint action of G on g∗ . The above implies that G is a symmetry group under the right action for any left-invariant control system F. Then G itself is a symmetry group for F under

6.5 Abnormal extremals

85

the right action. The fact that F is left-invariant means that F(x, u) = Lx∗ F(e, u) for each x ∈ G and each u ∈ U. In this notation, Lx denotes the left translation by x, Lx∗ is its tangent map, and e is the group identity in G. This implies that A(x, T) = Lx A(e, T). Hence, the group of left translations is a symmetry group for any leftinvariant control system. If g = Lg then etA x = xetA . Hence, XA is the right-invariant vector field with XA (e) = A. To preserve the left-invariant symmetries, T ∗ G will be realized as G × g∗ as explained in the previous chapter. In this representation, points ξ in T ∗ G are represented by pairs (g, ); the Hamiltonians of left-invariant vector fields XA = Lg ∗ (A) are given by hA = (A) and hA (g, ) = (Lg−1 ∗ ◦ Rg ∗ (A)) are the Hamiltonians of right-invariant fields XA (g) = Rg ∗ (A). It then follows from Proposition 6.24 that (t)(Lg−1 (t) ∗ ◦ Rg (t)∗ (A)) is constant along each extremal curve ξ(t) = (g(t), )t)). But this means that ∗ ((t)) = Ad ∗ ((0)), where Ad ∗ denotes the coadjoint action of G, or, Adg(t) g(0) equivalently, (t) = Adg∗−1 (t)g(0) ((0). The above can be assembled into the following proposition. Proposition 6.28 The projection (t) of each extremal curve ξ(t) = (g(t), (t)) generated by a left-invariant control system evolves on the coadjoint orbit of G through (0).

6.5 Abnormal extremals For optimal control problems in which the dimension of the control space is less than the dimension of the state space it may happen that an optimal trajectory is not a projection of a normal extremal curve. The presence of “abnormal curves” for the problems of the calculus of variations was noticed by A. G. Bliss in the early 1920s [Bl]. Carath´eodory rediscovered these curves in the late 1930s in his treatment of Zermelo’s navigational problem, which he called anomalous [Cr]. Much later, L. C. Young drew attention to the “rigid curves,” curves that do not admit variations and hence can not be detectable by the classical methods of the calculus of variations [Yg]. In some instances, particularly in differential geometry literature, abnormal extremals are ignored which resulted in a number of false claims [Rn; Sz]. It was not until the late 1990s that it became clear that there was nothing “abnormal” about abnormal extremals and that they may occur in many

86

6 Hamiltonians and optimality: the Maximum Principle

instances for the following reasons. Each optimal trajectory x(t) that minimizes T the cost 0 f (x(t), u(t)) dt defines an extended trajectory x˜ = (y(t), x(t)) ˜ x0 , T) at which terminates on the boundary of extended reachable set A(˜ T time T because y(T) = 0 f (x(t), u(t)) dt is minimal. Therefore, the cone of attainability along an optimal trajectory cannot be the entire tangent space. In the simplest of all situations, the cone of attainability has a non-empty interior in the extended tangent space and its projection on Tx M is equal to Tx M. Then there are no abnormal extremals. However, it may happen that the projection of the cone of attainability on Tx M is a proper closed convex cone, possibly of lower dimension, even though the reachable set A(x0 , T) is open for each x0 ∈ M and any T > 0. Since the projection of the cone of attainability on T∗ M corresponds to the zero multiplier in front of the cost, the projection is independent of the cost and depends only on the control system. This is the situation that results in abnormal extremals. In some cases the projection of these extremals is optimal for a given cost functional, and not optimal for some other cost functionals. Determining the optimality status of the projected trajectory is in general a difficult problem [Ss1]. Below we will illustrate the general situation with some problems of sub-Riemannian geometry initiated by a famous paper of R. Montgomery [AB; BC; Mt]. Let D denote the kernel of = dz − A(y)dx, where x, y, z denote the coordinates of a point q ∈ R3 and where A is a function with a single nondegenerate critical point at y = 0, that is, A is a function with dA dy (0) = 0 and d2 A (0) = 0. Then D is a two-dimensional distribution spanned by vector fields dy2 ∂ ∂ ∂ X1 = ∂x + A(y) ∂z and X2 = ∂y The trajectories of D are the solutions of

dq = u(t)X1 (q) + v(t)X2 (q), (6.21) dt where u(t) and v(t) are bounded and measurable control functions. In coordinates, dx dy dz = u(t), = v(t), = A(y)u(t). dt dt dt

(6.22) 2

The paper of Agrachev et al. deals with the Martinet form = dz − y2 dx [AB]. In the paper of Montgomery [Mt], θ is given by = dz−A(r)dθ ( , where r, θ , z are the cylindrical coordinates x = r cos θ , y = r sin θ , r = x2 + y2 and A is a function of r that has a single non-degenerate critical point at r = 1. If we introduce new coordinates x = θ , y = r + 1, and redefine A(y) as A(y + 1), then the Martinet distribution becomes a special case of the Montgomery distribution.

6.5 Abnormal extremals

87

∂ ∂ It is easy to verify that [X1 , X2 ] = A (y)) ∂z and that [X2 , [X1 , X2 ]] = −A ∂z . 3 Hence, Lieq {X1 , X2 } = Tq R at all points q, and therefore, any pair of points in R3 can be connected by a trajectory of (6.21). A two-dimensional distribution in a three-dimensional ambient space is called contact if for any point q in the ambient space M and any choice of vector fields X1 and X2 that span D in a neighborhood of a point q, X1 (q), X2 (q), [X1 , X2 ](q) are linearly independent. The distribution defined by is not contact since [X1 , X2 ](q) = 0 on the plane y = 0. This plane is called the Martinet plane. Since dz = A(y)dx on D, any metric on D can be written in the form a(q)dx2 + 2b(q)dxdy + c(q)dy2 for some functions a, b, c. Over the solutions dq dq q(t) of (6.21) such a metric can be written as dt , dt = a(q(t))u2 (t) +

terms of this notation, the length of q(t) 2b(q(t))u(t)v(t) + c(q(t))v2 (t)). In T dq dq on an interval [0, T] is given by 0 dt , dt dt, while its energy is given by 1 T dq dq 2 0 dt , dt dt. Let us now consider the extremals associated with the sub-Riemannian problem of finding a trajectory of shortest length that connects two given points in R3 . This problem can be phrased either as a time optimal problem over the dq solutions of (6.21) subject to the constraint dq dt , dt ≤ 1, or as the problem of minimizing the energy over a fixed time interval [0, T]. The latter results in T the optimal control problem of minimizing 0 (a(q(t)u2 (t) + 2b(q(t)u(t)v(t) + c(q(t)v2 (t)) dt over the trajectories of (6.21) that satisfy q(0) = q0 and q(T) = q1 . For simplicity of exposition we will assume that b = 0, and we will also assume that a and c depend only on x and y, and are of the form a(x, y) = 1 + yf (x, y) and c(x, y) = 1 + g(x, y) for some smooth functions f and g. The case where a, b, c do not depend on z is called isoperimetric in 2 [AB]. The same paper shows that in the case A(y) = y2 any sub-Riemannian isoperimetric problem can be reduced to the above normal form. The metric in this normal form is said to be flat if a = c = 1. In the paper of Montgomery [Mt], the metric dr2 + r2 dθ 2 corresponds to a = 1 + y(2 + y) and c = 1. Let us now consider the abnormal extremals associated with (6.21). As we already remarked, those extremals depend on the distribution only, and not on the metric. Let px , py , pz denote the dual variables corresponding to the coordinates x, y, z. Then the Hamiltonian lift h of (6.22) is given by h = u(t)h1 + v(t)h2 ,

88

6 Hamiltonians and optimality: the Maximum Principle

where h1 = px + A(y)pz and h2 = py . An abnormal extremal is a solution of the following constrained Hamiltonian system: dy dz dx = u(t), = v(t), = A(y)v(t), dt dt dt dpy dpx dA dpz = 0, = − pz , = 0, py = 0, px + A(y)pz = 0. dt dt dy dt

(6.23)

It is easy to see that x(t) = x0 + t, y(t) = 0, z(t) = z0 + A(0)t, px = −A(0)pz , py = 0, pz = c, with c an arbitrary constant, are the only solutions of the above system. The projections of these abnormal curves are straight lines x(t) = x0 + t, z(t) = z0 + A(0)t in the Martinet plane y = 0; they are generated by u(t) = 1 and v(t) = 0 (in the cylindrical coordinates of Montgomery, these lines are helices r(t) = 1, θ (t) = θ0 = t, z(t) = z0 + A(1)t). The normal extremals are generated by the Hamiltonan 1 1 1 2 2 H= (px + A(y)pz ) + py , 2 a c defined by the extremal controls u = 1a (px + A(y)pz ) and v = 1c py . It follows that they are the solutions of py dz 1 1 dy dx = (px + A(y)pz ), = , = A(y) (px + A(y)pz ), dt a dt c dt a dpx ax cx 2 2 = 2 (px + A(y)pz ) + 2 py , (6.24) dt a c ay dpy cy A (y)pz dpz = 2 (px + A(y)pz )2 − (px + A(y)pz ) + 2 p2y , = 0. dt a dt a c Definition 6.29 The projections of the extremals on R3 are called geodesics. A geodesic is called abnormal (respectively normal) if it is a projection of an abnomal (respectively normal) extremal curve. It is called strictly abnormal if it is a projection of an abnormal extremal curve but not the projection of a normal one. Proposition 6.30 Lines x(t) = x0 + t, z(t) = z0 + A(0)t in the Martinet plane are strictly abnormal if and only if f (x, 0) = 0. Proof The above lines in the Martinet plane are the projections of normal extremal curves if and only if px +A(0)pz = 1 and ay (x, 0) = 0. Since a(x, y) = 1 + yf (x, y), a(x, 0) = 1 and ay (x, 0) = f (x, 0)). It follows that ay (x, 0) = 0 if and only if f (x, 0) = 0. It follows that the abnormal geodesics are not strictly abnormal in the flat case, but that they are strictly abnormal in Montgomery’s example. The abnormal geodesics are optimal in the flat case, and remarkably, they are

6.5 Abnormal extremals

89

also optimal (at least for small time intervals) in Montgomery’s example. In both cases, point q1 = (x0 + c, 0, A(0)c) with c > 0 can be reached from q0 = (x0 , 0, z0 ) in time t = c with controls that satisfy u2 + v2 ≤ 1 only with u(t) = 1 and v(t) = 0. Hence, this trajectory is optimal for any optimal problem and not just for the sub-Riemannian one. This observation is evident in the flat case, but is not so obvious in the example of Montgomery. The proposition below provides the necessary arguments. Proposition 6.31 There exists an interval [0, T] with T < 1 such that the helix q(t) = (1, t, A(1)t) is optimal on this interval. Proof It will be convenient to work with the cylindrical coordinates. Then 1 ∂ ∂ ∂ are othonormal vector fields for the X1 (r, θ ) = r ∂θ + A(r) ∂z and X2 = ∂r metric dr2 + r2 dθ 2 . Consider now the reachable set A(q0 , T) associated with dr dθ 1 dz 1 = v(t), = u(t), = A(r)u(t), u2 (t) + v2 (t) ≤ 1 (6.25) dt dt r dt r with q0 = (1, 0, 0). Since the control set U is compact and convex, each velocity set {uX1 (x) + bX2 (x), u2 + v2 ≤ 1} is compact and convex. Moreover, each trajectory that originates at q0 at t = 0 exists for all t in [0, T] with T < 1 (for T = 1 the trajectory with u = −1 escapes M at time 1 since r(1) = 0). Therefore, A(x0 , T) is closed, as a consequence of Fillipov’s theorem [Jc, p. 120; AS, p. 141]. It is also compact, but that is not relevant for the computation below. We will now consider an auxiliary problem of finding a trajectory q(t) that originates at q0 , terminates at r(T) = 1, θ (T) = T and minimizes z(T). Such a trajectory exists since the reachable set is closed. Moreover, the terminal point q(T) is on the boundary of A(q0 , T). Our claim is that the helix is the only solution for this auxiliary problem. The following observation is crucial: if q(t) = (r(t), θ (t), z(t)) is any trajectory that solves the auxiliary problem such that r(t) is not equal to 1 for all t, then dθ dt must be negative somewhere in the interval [0, T]. The argument is simple: since A(1) is a global mimimum for A(r), A(r) = A(1)+B(r) for some positive function B(r). Then T T dθ dθ A(r(t)) dt = A(1)T + B(r(t)) dt. z(T) = dt dt 0 0 Hence, z(T) > A(1)T unless dθ dt is negative on a set of positive measure. According to the Maximum Principle, x(t) is the projection of an extremal curve ξ(t) = (r(t), θ (t), pr (t), pθ (t)) associated with the Hamiltonian λ 1 h(u(t),v(t)) = − A(r)u(t) + pr v(t) + pθ u(t), λ = 0, 1. r r

90

6 Hamiltonians and optimality: the Maximum Principle

In the abnormal case λ = 0, p2r + p2θ > 0 and hence the extremals are + 1 generated by the Hamiltonian h = p2r + r12 p2θ . But then dθ dt = r2 h pθ and pθ is constant. The projections of these curves cannot meet the boundary conditions in view of the above remark, and hence are ruled out. We now pass to the normal case λ = 1. There are two possibilities: either p2r (t) + (pθ − A(r))2 > 0,

(6.26)

and + the extremal curves are generated by the Hamiltonian h = p2r + r12 (pθ − A(r))2 corresponding to the extremal controls u(t) = 1 hr (pθ

− A(r)) and v(t) =

pr h,

or

pr (t) = 0 and pθ − A(r) = 0, a.e. on [0, T].

(6.27)

In the latter case the extremal solutions are given by dr 1 1 dpθ dθ 1 dpr 1 = v(t), = u(t), = − 2 A(r) + A (r) + 2 pθ )u(t), = 0. dt dt r dt r dt r r (6.28) Since pθ is constant, A(r(t)) = A(1) and pθ = A(1). It follows that r(t) = 1, and hence, v(t) = 0. But then u(t) = 1, and the resulting solution is the helix. We will show that the helix is the + only solution by showing that the projections of the integral curves of h =

p2r +

pr dθ pθ − A(r) dpr pθ − A dr = , = = , 2 dt h dt dt r h hr2

1 (p r2 θ

− A(r))2 given by

pθ − A dpθ + A (r) , =0 r dt (6.29)

cannot meet the given boundary conditions. System (6.29) is integrable. In fact, pθ is constant and , dr (pθ − A(r))2 = ± h2 − . dt r2

(6.30)

For dθ dt to change sign, pθ − A(1) must be positive and there must be an interval [t0 , t1 ] such that pθ − A(r(t)) < 0 on (t0 , t1 ) with pθ − A(r(t0 )) = pθ −A(r) h , or pθ − A(r(t1 )) = 0. On this interval, dθ dr = ± hr2 2 h2 −

θ =

(p − A(r)) dr . + 2 r2 h2 − (pθ −A(r)) 2 r

(pθ −A(r)) r2

6.6 The Maximum Principle and Weierstrass’ excess function

91

φ Let pθ −A(r) = −h cos φ. Then, dr = h cossin r φ−A (r) dφ, and the above integral becomes cos φ dφ . (6.31) θ = − r(h cos φ − A (r)) π On the interval [t0 , t1 ], − −π 2 ≤ φ ≤ 2 and θ is given by π 2 cos φ dφ − . π r(h cos φ − A (r)) −2

For small T, r is close to 1 and A (r) is close to 0, hence the above integral is close to π . Hence θ (T) cannot be equal to T. As we have already remarked, the above distributions are not contact, because [X1 , X2 ] is linearly dependent on X1 and X2 on the Martinet plane. The fact that abnormal geodesics are confined to the Martinet plane is a particular case of a more general situation described by the following proposition. Proposition 6.32 Suppose that a distribution D is spanned by two vector fields X1 , X2 in an open set M. Suppose further that [X1 , X2 ](x) is linearly independent from X1 (x) and X2 (x) at each x ∈ M. Then no integral curve x(t) of D in M is the projection of an abnormal extremal curve. Proof If x(t) is an integral curve of D then dx dt = u(t)X1 (x(t)) + v(t)X2 (x(t)) for some functions u(t) and v(t). Every abnormal extremal curve ξ(t) in T ∗ M is an integral curve of the Hamiltonian vector field h associated with h(ξ ) = u(t)h1 (ξ ) + v(t)h2 (ξ ), such that ξ(t) = 0 and is subject to the constraints h1 (ξ(t)) = h2 (ξ(t)) = 0. In this notation h1 and h2 denote the Hamiltonian lifts of X1 and X2 , that is, h1 (ξ ) = ξ(X1 (x)), h2 (ξ ) = ξ(X2 (x)) for all ξ ∈ Tx∗ M. But then 0 = dtd h1 (ξ(t)) = {h1 , h}(ξ(t) = h[X1 ,u(t)X1 +v(t)X2 ] (ξ(t) = v(t) h[X1 ,X2 ] (ξ(t)). Similarly, 0 = dtd h2 (ξ(t)) = u(t)h[X2 ,X1 ] (ξ(t)). Since x(t) is not a stationary curve, either u(t) = 0 or v(t) = 0. Therefore, h[X1 ,X2 ] (ξ(t)) = 0. But that contradicts ξ(t) = 0 since X1 (x), X2 (x), [X1 , X2 ](x) span Tx M.

6.6 The Maximum Principle and Weierstrass’ excess function Having the Maximum Principle at our disposal, let us now return to the necessary conditions for the strong extrema in the problem of the calculus of

92

6 Hamiltonians and optimality: the Maximum Principle

variations. To make an easy transition to the control problem, let M denote a ˜ = R×M, strong neighborhood of x(t). ¯ Secondly, pass to the extended space M by treating time as another spacial variable x0 , in order to make L dependent on the space variables only. Additionally, use x˜ (t) = (t + t0 , x(t + t0 )) to shift ¯ 0 )) and x˜ (t1 − t0 ) = (t1 , x(t ¯ 1 )). the boundary conditions to x˜ (0) = (t0 , x(t To convert to optimal control problem, let X1 , . . . , Xn be any linearly independent set of vector fields that span Tx M at each point x ∈ M. Then n every admissible curve x(t) is a solution of dx i=1 ui (t)X(x(t)), and every dt = extended curve x˜ (t) = (x0 (t), x(t)) is a solution of the “affine control system” dx0 dx ui (t)Xi (x(t), = 1, = dt dt n

(6.32)

i=1

with controls u(t) taking values in U = Rn . If a trajectory x˜ (t) satisfies ¯ 0 )) and x˜ (s) = (t1 , x(t ¯ 1 )) for some the boundary conditions x˜ (0) = (t0 , x(t − t . This implies that if x(t) ¯ is to attain the minimum s > 0 then s = t 1 0 t dt subject to the above boundary conditions, then the of t01 L t, x(t), dx dt extended trajectory x˜ (t) = (t+t0 , x(t+t ¯ 0 )) is a solution for the optimal problem T defined by 0 L(˜x(t), u(t)) dt with variable time interval [0, T]. Let u(t) ¯ denote the control that generates x(t). ¯ Then the optimal trajectory (˜x(t), u(t)) ¯ is the projection of an extremal curve (λ, (ξ0 (t), ξ(t)), where (ξ0 , ξ ) ˜ realized as the product T ∗ R × T ∗ M, and λ = 0, −1, is a point in T ∗ M, depending on whether ξ(t) is normal or abnormal. This extremal curve is an integral curve of hu(t) ¯ associated with hu(t) x, u(t)) ¯ + ξ0 + ¯ (ξ0 (t), ξ(t)) = λL(˜

n

u¯i (t)Xi (x), (ξ0 , ξ ) ∈ Tx∗0 R × Tx∗ M.

i=1

(6.33) Moreover, hu(t) ¯ (ξ0 (t), ξ(t)) = 0, for almost all t along the extremal curve. There are two immediate consequences of the Maximum Principle. First, there are no abnormal extremals. For, if λ = 0, then hu (ξ0 , ξ(t)) ≤ 0 for any u ∈ Rn would imply that ξ(t) = 0. But then ξ0 = 0 since hu(t) ¯ ((ξ0 , ξ(t)) = ξ0 . However, (ξ0 , ξ(t)) = 0 violates the non-degeneracy condition of the Maximum Principle. Second, ξ0 (t) is constant, because x0 (t) is a cyclic coordinate, that is, the vector fields in (6.32) do not depend explicitly on x0 . Consequently, we may pass to the reduced Hamiltonian Hu(t) ¯ + ¯ (ξ ) = −L(t, x(t), u(t))

n i=1

u¯i (t)hi (ξ ),

(6.34)

6.6 The Maximum Principle and Weierstrass’ excess function

93

where hi denotes the Hamiltonian lift of Xi for i = 1, . . . , n. This Hamiltonian is the projection of the original Hamiltonian (6.33). As such, it is constant almost everywhere along each extremal curve ξ(t), and is maximal, in the sense that Hu(t) ¯ (ξ(t)) = Maxu∈Rn Hu (ξ(t)) a.e. in [t0 , t1 ]. It is important to note that the Hamiltonian in (6.34) is intrinsic, that is, does not depend on the choice of the frame X1 , . . . , Xn . For if Y1 , . . . , Yn is another frame on M, then every point y ∈ Tx M is represented by the sum ni=1 vi Yi (x) n for some numbers v1 , . . . , vn . It then follows that Yi (x) = j=1 Uij (x)Xj (x) for ¯ some transformation U(x) ∈ GLn (R) and that ui = nj=1 Uji vj . Along x(t), n n dx¯ ¯ vj (t) where dt = i=1 v¯ i (t)Yi (x(t)). ¯ Furthermore, u¯i (t) = j=1 Uji (x(t))¯ % Hu(t) ¯ (ξ ) = −L t, x, = −L ⎝t, x, = L t, x,

& u¯i (t)Xi (x) +

i=1

⎛

%

n

i=1

n i=1

n

n

u¯i (t)ξ(Xi (x)) ⎞

Uji (x)¯vj (t)Xi (x)⎠ +

ij

Uji v¯ j (t)ξ(Xi (x))

ij

&

v¯ i (t)Yi (x) +

i=1

n

n

v¯ i (t)ξ(Yi (x)) = Hv¯ (t) (ξ ).

i=1

v¯ (t) have the same integral curves. Hence H u(t) ¯ and H Since Hu(t) ¯ is maximal along an extremal curve ξ(t) 0=−

∂L (t, x(t), u) + hi (ξ(t)), i = 1, . . . , n, ∂ui

(6.35)

2

L and ∂u∂i ∂u (t, x(t), ¯ u(t)) ¯ ≥ 0. This condition is known as the Legendre conj dition in the literature on the calculus of variations. As we have already 2L remarked earlier, when ∂u∂i ∂u (t, x(t), ¯ u(t))0 ¯ is strictly positive-definite in some j neighborhood of x(t), ¯ t ∈ [t0 , t1 ], then equation (6.35) is solvable, in the sense that there is a function φ defined in some neighborhood of the graph {(t, ξ(t)) : t ∈ [t0 , t1 ]} such that

u(t) ¯ = φ(t, ξ(t)).

(6.36)

Under these conditions, each extremal curve ξ(t) is an integral curve of a single Hamiltonian vector field defined by H(t, ξ ) = −L(t, x, φ(t, ξ )) +

n i=1

φi (t, ξ )

∂L (t, x, φ(t, ξ )). ∂ui

(6.37)

94

6 Hamiltonians and optimality: the Maximum Principle

Suppose now that the Hamiltonian H given by (6.37) is defined on the entire cotangent bundle T ∗ M. That is, suppose that the equation 0=−

∂L (t, x, u) + hi (ξ ), i = 1, . . . , n ∂ui

admits a global smooth solution u = φ(t, ξ ), ξ ∈ Tx∗ M. This condition is equivalent to saying that L induces a transformation : R × T ∗ M → R × TM such that (t, ξ ) = ni=1 φi (t, ξ )Xi (x) for each (t, ξ ) ∈ R × Tx∗ M. It follows that H(t, ξ ) = Maxu∈Rn {Hu (t, ξ ), ξ ∈ T ∗ M}.

(6.38)

The famous condition of Weierstrass, E(t, x(t)), ¯ u(t), ¯ u) = L(t, x(t), ¯ u(t)) ¯ − L(t, x(t), ¯ u) +

n i=1

(u¯i (t) − ui )

∂L (t, x(t), ¯ u(t)) ¯ ≥ 0, ∂u

is nothing more than a paraphrase of the maximality condition Hu(t) ¯ (ξ(t)) ≥ n Hu (ξ(t)), u ∈ R . In the literature on the calculus of variations, the excess function of Weierstrass is usually associated with the sufficient conditions of optimality under some additional assumptions [Cr; Yg]. The first assumption is that the Hilbert–Cartan form −Hdt + , with the Liouville form ξ ◦ π∗ , is exact in some neighborhood of an extremal curve ξ¯ (t), t ∈ [t0 , t1 ]. The Hilbert–Cartan form is usually written in the canonical coordinates as −H(t, x, p)dt+pdx [C2]. To say that this form is exact, means that there exists a function S(t, x) such that ∂S = pi and ∂S the differential dS is equal to −Hdt + pdx. But then, ∂x ∂t = −H. i That is, S(t, x) is a solution of the Hamilton–Jacobi equation ∂S ∂S + H t, x, = 0. (6.39) ∂t ∂x The second assumption is that some neighborhood of ξ¯ (t), t ∈ [t0 , t1 ] projects diffeomorphically onto a neighborhood of the projected curve x(t) ¯ = π(ξ¯ (t)). This assumption implies that every curve x(t) in this neighborhood is the projection of of a unique curve ξ(t) in T ∗ M. Moreover, ξ(t) and ξ¯ (t) have the same initial and terminal points whenever x(t) and x(t) ¯ have the same terminal points. Under these assumptions, x(t) ¯ is optimal relative to the curves x(t) in this neighborhood that satisfy the same boundary conditions x(t0 ) = ¯ 1 ). x(t ¯ o ), x(t1 ) = x(t

6.6 The Maximum Principle and Weierstrass’ excess function The proof is simple and goes as follows. Let u(t) = t1 (L(t, x(t), u(t)) − L(t, x(t), ¯ u(t))) ¯ dt t0

t1

= ≥

t0 t1 t0 t1

dx dt

and u(t) ¯ =

dx¯ dt .

γ¯

Then,

(Hu(t) ¯ p¯ (t)) − p¯ (t)u(t) ¯ − Hu(t) (t, x(t), p(t)) + p(t)u(t)) dt ¯ (t, x(t), (Hu(t) ¯ p¯ (t)) − p¯ (t)u(t) ¯ − Hu(t) ¯ (t, x(t), ¯ (t, x(t), p(t)) + p(t)u(t)) dt

dx¯ dx H(t, x(t), ¯ p¯ (t)) − p¯ (t) − H(t, x(t), p(t)) + p(t) dt dt dt t0 = (Hdt − pdx) − (Hdt − pdx) = dS − dS = 0, =

95

γ

γ¯

γ

where γ¯ (t) = (t, (x(t), ¯ p¯ (t))) and γ (t) = (t, (x(t), p(t))).

7 Hamiltonian view of classic geometry

We have now arrived at a transition point at which the preceding theory begins to be directed to problems of geometry and applied mathematics. We will begin with the applications to the classic geometry, as a segue to further investigations into the geometry of symmetric spaces and the problems of mechanics. We will proceed in the spirit of Felix Klein’s Erlangen Program of 1782, and start with a Lie group G, which acts on a manifold M with the aim of arriving at the natural geometry on M induced by the structure of G.

7.1 Hyperbolic geometry a b Consider first the group G = SU(1, 1) = : a, b in C, |a|2 − b¯ a¯ ' 2 |b| = 1 , and its action on the unit disk D = {z ∈ C : |z| = 1} by the Moebius transformations gz = 4|dz|2

az+b ¯ a. bz+¯

Our aim is to show that the classic hyperbolic

metric 1−|z|2 on D is an immediate consequence of the canonical subRiemannian problem on G. Since the above action is transitive, D can be regarded as the homogeneous space G/K where K is the isotropy group of afixed point ' z0 . It is convenient a 0 to take z0 = 0, in which case, K = ,|a| = 1 ; the isotropy group 0 a¯ through any other point z0 is the conjugate group gKg−1 , where g ∈ G such that g(0) = z0 .

96

7.1 Hyperbolic geometry

97

iα β The Lie algebra g of G consists of the matrices ,α ∈ R β¯ −iα ' and β ∈ C . Then g can be regarded as the sum g = p ⊕ k, where k = ' ' i 0 0 β α : α ∈ R and p = : β ∈ C . It is easy to check 0 −i β¯ 0 that [A1 , A2 ] ∈ k for any matrices A1 and A2 in p. Consider now the bilinear form A, B = 2Tr(AB), where Tr denotes the trace ¯ ¯ of a matrix. It followsthat A1 , A2 = − α1 α2 + β1 β 2 + β1 β2 for any matrices β1 β2 iα1 iα1 and A2 = . Hence , is positiveA1 = β¯1 −iα1 β¯2 −iα2 definite on p and can be used to define a left-invariant metric on the space of curves g(t) in G that satisfy g−1 (t) dg dt (t) = A(t), with A(t) ∈ p. We can put this observation in the sub-Riemannian context by introducing the matrices 1 A1 = 2

0 −i

i 0

1 , A2 = 2

0 1

1 0

1 , A3 = 2

−i 0

0 i

.

(7.1)

Matrices A1 and A2 form an orthonormal basis for p, hence+the length of any T curve g(t) in G that satisfies g−1 (t) dg u21 (t) + u22 (t) dt, dt (t) ∈ p is given by 0 where u1 (t) and u2 (t) are defined by g−1 (t) dg dt = u1 (t)A1 + u2 (t)A2 . Since [A1 , A2 ] = A3 any two points g0 and g1 in G can be connected by a horizontal curve g(t), that is, a curve that satisfies g−1 (t) dg dt ∈ p. So there is a natural time optimal control problem on G that consists of finding the trajectories g(t) of dg = g(t)(u1 (t)A1 + u2 (t)A2 ), u21 (t) + u22 (t) ≤ 1 dt that connect g0 to g1 in the least possible time T. To see the connection with the hyperbolic metric on M, consider the lifting of curves in D to the in G. Any curve z(t) = x(t) + iy(t) lifts horizontal curves a(t) b(t) to a curve g(t) = in G via the formula ¯ b(t) a¯ (t) z(t) =

b(t) a(t)(0) + b(t) . = ¯ a¯ (t) b(t)(0) + a¯ (t)

An easy calculation shows that g(t) = where θ (t) is an arbitrary curve.

√ 1 1−|z(t)|2

1 z(t) e−iθ(t) 0 z¯(t) 1

, eiθ(t) 0

98

7 Hamiltonian view of classic geometry

Proposition 7.1

a. g−1 (t) dg dt (t) ∈ p for all t if and only if dx dθ 1 dy = y(t) − x(t) . dt dt 1 − |z(t)|2 dt

If θ (t) is any curve that satisfies this condition (they differ by an initial condition) then, dx 2 dy dg −1 = cos 2θ − sin 2θ A1 g (t) dt dt dt 1 − |z(t)2 dy dx (7.2) sin 2θ + cos 2θ A2 . + dt dt The energy functional associated with g(t) is given by 2 . dz 1 −1 dg −1 dg 2 g (t) , g (t) =# (7.3) $2 . 2 dt dt dt 1 − |z(t)|2 −iθ ( 0 1 z e , so that Proof Let α = 1 − |z|2 , Z = , and = 0 eiθ z¯ 1 1 g(t) = α(t) Z(t)(t). A short calculation shows that α˙ 1 dg −zz˙¯ z˙e2iθ −i 0 ˙ =− I+ 2 + θ g−1 (t) . z˙¯e−2iθ −¯zz˙ 0 i dt α α 0 z˙e2iθ dg 1 −1 whenever θ˙ = α12 (xy ˙ − xy). ˙ Upon Hence, g dt = α 2 z¯˙e−2iθ separating the real and the imaginary parts of z˙e2iθ we get the formula above. The above shows that the sub-Riemannian metric on G induces the hyper|dz| bolic metric ds = 2 1−|z| 2 on the unit disk. The lifting to horizontal curves in G 1 uncovers the differential form dθ = 1−|z| 2 (ydx − xdy) relevant for evaluating the hyperbolic area enclosed by closed curves in M. There are two other models in the classical hyperbolic geometry: the upper half plane P = {z ∈ C :∈ z ≥ 0} and the hyperboloid H3 = {x ∈ R3 : −x02 + x12 + x22 = 1, x0 > 0}. All of these models can be seen in a unified way through the actions of Lie groups. In the first case, P can be realized as the quotient SL2 (R)/SO2 (R), while in the second case H3 can be realized as the quotient SO(1, 2)/SO2 (R). In the first case, SL2 (R) acts on P via the Moebius transformations. If g(z) = az+b cz+d then g(i) = i if and only if g ∈ SO2 (R). Hence, points z ∈ P can be identified with matrices g in SL2 (R) via the formula g(i) = z up to the matrices in SO2 (R). Since any matrix g in SL2 (R) can be written as the product

7.1 Hyperbolic geometry

99

a b α β with α 2 + β 2 = 1, the above identification can be 0 c −β α reduced to the upper triangular matrices in SL2 (R). It then follows that any curve z(t) in P is identified with & % 1 √ √ 1 x(t) y(t) cos θ (t) sin θ (t) y(t) y(t) , g(t) = √1 − sin θ (t) cos θ (t) 0 y(t)

where θ (t) is an arbitrary curve. In this case, 1 1 0 1 0 1 1 0 A1 = , A2 = , A3 = 2 0 −1 2 1 0 2 1

−1 0

,

(7.4)

is the basis that is isomorphic to the matrices in (7.1). For then, A1 , A2 , A3 conform to the same Lie bracket table as in the case of su(1, 1), and the trace form A, B = 2Tr(AB) is positive-definite on the linear span of A1 and A2 , with A1 and A2 orthonormal relative to , . As in the first case, the sub-Riemannian metric on the two-dimensional leftinvariant distribution spanned by matrices A1 and A2 induces the hyperbolic metric on P. It is easy to check that z(t) lifts to a horizontal curve g(t) if and 1 dy only if dθ dt (t) = y(t) dt . The sub-Riemannian metric on the space of horizontal curves coincides with the hyperbolic metric on P in the sense that ,. ( 2 x˙ + y˙2 dg −1 dg −1 g (t) , g (t) = . (7.5) dt dt y These two hyperbolic models are isometric as can be easily demonstrated i 1 through the Cayley transform C = √1 . The Moebius transfor2 −1 −i −1 mation defined by C takes D onto P. Hence, the unitdisk D for C gC acts on a b p q any g ∈ SL2 (R). In fact if g = then C−1 gC = , where c d q¯ p¯ 1 1 (a + d + i(b − c)), q = (−b − c + i(d − a)). 2 2 p q Since C ∈ SL2 (C), |p|2 − |q|2 = 1 and hence belongs to SU(1, 1). q¯ p¯ This conjugation provides the desired isomorphism between SU(1, 1) and SL2 (R) and the corresponding Lie algebras sl2 (R) and su(1, 1). One can easily check that if A˜ i = CAi C−1 , where A1 , A2 , A3 are the matrices in (7.4), then A˜ 1 = −A1 , A˜ 2 = −A2 , A˜ 3 = A3 , where now A1 , A2 , A3 are the matrices in (7.1). Since A˜ 1 and A˜ 2 remain orthonormal relative to the trace p=

100

7 Hamiltonian view of classic geometry

metric, the above shows that the sub-Riemannian problems on SL2 (R) and SU(1, 1) are isometric. To complete this hyperbolic landscape, let us shift to the hyperboloid H2 = {x ∈ R3 : x12 − x22 − x32 = 1, x1 > 0}. Since G = SO(1, 2) is the isometry group for the Lorentzian metric (x, y) = x1 y1 − x2 y2 − x3 y3 , G acts on H2 by the matrix multiplication (g, x) → gx. Then H2 can be represented as the quotient G/K where K is the isotropy group Ke1 = e1 . The matrices in K are ⎛ ⎞ 1 0 0 of the form ⎝ 0 a b ⎠ with a2 + b2 = 1. Hence, K is isomorphic with 0 −b a SO2 (R). ⎛ ⎞ 0 α β The Lie algebra g of G consist of the matrices ⎝ α 0 −γ ⎠. Hence, β γ 0 ⎛ ⎞ ⎛ ⎞ ⎛ ⎞ 0 1 0 0 0 1 0 0 0 A1 = ⎝ 1 0 0 ⎠ , A2 = ⎝ 0 0 0 ⎠ , A3 = ⎝ 0 0 −1 ⎠ (7.6) 0 0 0 1 0 0 0 1 0 form a basis for g. The reader can readily check that the trace form A, B = 1 2 Tr(AB) is positive-definite on the linear span of A1 and A2 . Analogous to the previous cases, A1 and A2 define a two-dimensional, leftinvariant distribution D whose integral curves g(t), again called horizontal, are 2 the solutions of dg dt = g(t)(u1 (t)A1 + u2 (t)A2 ). Any curve x(t) in H can be lifted to a horizontal curve g(t) via the formula g(t)e1 = x(t). Then, dx dt = dg dt e1 = g(t)(u1 A1 + u2 A2 )e1 and consequently, −x˙21 + x˙22 + x˙23 = −(g(t)(u1 A1 + u2 A2 )e1 , g(t)(u1 A1 + u2 A2 )e1 ) = u21 + u22 . So the hyperbolic metric coincides with the sub-Riemannian metric defined by the trace form. All three of these Lie algebras are isomorphic and conform to the following Lie bracket table: [A1 , A2 ] = A3 , [A1 , A3 ] = A2 , [A2 , A3 ] = −A1 , but the corresponding groups are not all isomorphic: SL2 (R) is a double cover of SO(1, 2). In fact, any matrix any matrix X ∈ sl2 (R) can be written asthe sum of a sym 0 −x1 −x3 x2 and a skew-symmetric matrix , in metric matrix x2 x3 x1 0 which case, 1 X, X = − Tr(X 2 ) = x12 − x22 − x32 . 2

7.2 Elliptic geometry

101

Hence, sl2 (R), as a vector space, can be regarded as a three-dimensional Lorentzian space. It follows that Adg (X) = gXg−1 belongs to SO(1, 2) for any g ∈ SL2 (R). The correspondence g ⇒ Adg , called the adjoint representation of SL2 (R), shows that SL2 (R) is a double cover of SO(1, 2) since Ad±I = I. It also shows that sl2 (R) and so(1, 2) are isomorphic, with −ad(a1 A1 + a2 A2 + a3 A3 ) ⎛ ⎞ 0 a1 a2 corresponding to the matrix ⎝ a1 0 −a3 ⎠. a2 a3 0

7.2 Elliptic geometry There are two ways to proceed. One can either begin with the action of SO3 (R) on the sphere S2 , and then identify the sphere with the quotient SO3 (R)/SO2 (R), or alternatively, one may start with the action of SU2 on the projective complex line CP1 . These methods, similar in principle, but different in appearance, are in fact isometric, as can be seen through the stereographic projection of the sphere S2 onto the extended complex plane CP1 . In the first case, SO3 (R) acts on the sphere S2 by the matrix multiplication (g, x) → gx. Then SO2 (R) is the isotropy group of the point x0 = e1 . As in the hyperbolic situations, x(t) on the sphere can be lifted to horizontal curves g(t) in SO3 (R) via the formula g(t)e1 = x(t). The horizontal curves are defined relative to the left-invariant distribution spanned by the ⎛ ⎞ ⎛ ⎞ 0 0 1 0 −1 0 matrices A1 = ⎝ 0 0 0 ⎠ and A2 = ⎝ 1 0 0 ⎠. It follows that −1 0 0 0 0 0 ⎛ ⎞ 0 0 0 [A1 , A2 ] = − A3 , where A3 = ⎝ 0 0 −1 ⎠. 0 1 0 Matrices A1 , A2 , A3 are orthonormal relative to the bilinear form A, B = − 12 Tr(AB). If g(t) is a solution of dg dt = g(t)(u1 (t)A1 + u2 (t)A2 ), then x˙21 + x˙22 + x˙23 = u22 (t) + u22 (t). The reader may verify this formula by a calculation similar to the one used on the hyperboloid. Hence, the sub-Riemannian metric on the horizontal curves coincides with the spherical metric on the sphere inherited from the Euclidean metric in the ambient space R3 . a b Let us now turn to SU2 and its action , z → az+b ¯ a on bz+¯ −b¯ a¯ CP 1 . It is easy to verify that this action is transitive and that K =

102

7 Hamiltonian view of classic geometry

' a 0 : |a| = 1 is the isotropy group of z = 0 (K is also the isotropy 0 a¯ group of z = ∞). Evidently, K is isomorphic to S1and SO2 (R). iα β The Lie algebra g of SU2 consists of matrices with α ∈ R −β¯ −iα and β ∈ C. Let now A, B = −2Tr(AB). Relative to this form, the Pauli matrices 1 1 0 i 1 −i 0 0 1 , A2 = , A3 = (7.7) A1 = i 0 0 i 2 −1 0 2 2 form an orthonormal basis in su2 . Since SU2 is a double cover of SO3 (R), su2 and so3 (R) are isomorphic. In fact, matrices A1 , A2 , A3 on so3 (R) are in exact correspondence with the Pauli matrices on su2 . 1 z 1 √ Analogous to the situation on the disk, g(t) = 1+|z(t)|2 z¯ 1 iθ(t) 0 e is a lift of a curve z(t) in CP 1 . The proposition below 0 e−iθ(t) describes the horizontal lifts. Proposition 7.2 Let p denotes the real linear span of A1 and A2 . Then g−1 (t) dg dt (t) ∈ p for all t if and only if dx 1 dy dθ = y(t) − x(t) . dt dt 1 + |z(t)|2 dt If θ (t) is any curve that satisfies this condition (they differ by an initial condition), then dx 2 dy dg −1 g (t) = cos 2θ − sin 2θ A1 dt dt dt 1 + z(t)2 dy dx (7.8) sin 2θ + cos 2θ A2 . + dt dt The energy functional associated with g(t) is given by . 1 −1 dg −1 dg 2 g (t) , g (t) = 2 dt dt 1 + |z(t)|2

2 dz . dt

(7.9)

Traditionally, these two realizations of the elliptic metric are reconciled through the stereographic projection of S2 onto CP 1 , in which points x = (x1 , x2 , x3 ) on S2 are projected onto points z in CP 1 via the formula x1 =

2 2 |z|2 − 1 x, y = y, x = . 3 1 + |z|2 1 + |z|2 1 + |z|2

7.2 Elliptic geometry

103

Then an easy calculation shows that x˙21 + x˙22 + x˙23 =

4|dz|2 . 1 + |z|2

Let us now extract the essential properties of the sub-Riemannian problems uncovered by these geometric models. We have shown that each non-Euclidean space M is a homogenous space G/K. In the hyperbolic case, G is either SU(1, 1), SL2 (R), or SO0 (1, 2), and in the elliptic case G is either SO3 (R) or SU2 . In the first case the groups are non-compact, while in the second case they are compact. In each case, a scalar multiple of the trace form Tr(AB) is positive-definite on the two-dimensional vector space p in the Lie algebra g of G and defines a metric , on the corresponding left-invariant distribution D. Then matrices A1 , A2 , A3 denote a basis in g such that A1 , A2 form an orthonormal basis in p and [A1 , A2 ] = −A3 with = 1 in the elliptic case and = −1 in the hyperbolic case. Curves g(t) in G are called horizontal if g(t)−1 dg dt ∈ p for all t. Horizontal curves are the solutions of a “control system” dg = g(t)(u1 A1 + u2 (t)A2 ) dt

(7.10)

where u1 (t) and u2 (t) are arbitrary bounded and measurable functions playing the role of controls. Vertical curves are the curves that satisfy g−1 dg dt = u(t)A3 for some function u(t). Vertical curve that satisfy g(0) = I define a group K, which i is isomorphic to SO2 (R) in all cases. We have shown that each curve in M = G/K is a projection of a horizontal curve and that any two horizontal curves g1 and g2 that project onto the same curve x(t) in M satisfy g2 (t) = g1 (t)h for some h ∈ K. Any two points of G can be connected by a horizontal curve, since T + 2 [A1 , A2 ] = ±A3 , and any horizontal curve has length 0 u1 (t) + u22 (t) dt T and energy 12 0 (u21 (t) + u22 (t)) dt. The fundamental sub-Riemannian problem consists of finding a horizontal curve of shortest length that connects two given points in G. We have shown that the projection of the sub-Riemannian metric in G coincides with the Riemannian metric in M. We have also shown that both the hyperbolic and the elliptic models are isometric, so it suffices to pick a representative in each group. We will pick the simply connected models, G = SU(1, 1) in the hyperbolic case, and G = SU2 in the elliptic case. These groups are double covers of SO0 (1, 2) and SO3 (R).

104

7 Hamiltonian view of classic geometry

We will use G to denote these two groups with G = SU(1, 1) for = −1 and G = SU2 for = 1. On occasions we will refer to these groups as the non-compact case ( = −1) and compact case ( = 1). Then g will denote the Lie algebra of G . If A1 , A2 , A3 denote the bases in g given by (7.1) and (7.7) then these matrices conform to the Lie bracket Table 7.1. Table 7.1 [,] A1 A2 A3

A1 0 A3 −A2

A2 −A3 0 A1

A3 A2 −A1 0

Since Tr(AB) = Tr(BA), and since the sub-Riemannian metric is a scalar multiple of the trace form, A, [B, C] = [A, B], C , A, B, C in g . The fact that [p , k ] = p implies that gAg−1 ∈ p for A ∈ p and any g ∈ K. This implies that the sub-Riemannian metric is invariant under K, that is, gAg−1 , gBg−1 = A, B , g ∈ K. The precise relation between the sub-Riemannian problem on G and the induced Riemannian problem on M is described by the following proposition. Proposition 7.3 Suppose that g(t) is the shortest horizontal curve that connects a coset g0 K to the coset g1 K. Then the projected curve x(t) in M is the shortest curve that connects x0 = g0 K to x1 = g1 K. Conversely, if x(t) be the shortest curve in M that connects x0 and x1 , then x(t) is the projection of a horizontal curve g(t) of minimal sub-Riemannian length that connects the coset x0 = g0 K to the coset x1 = g1 K. Proof

The proof is obvious.

It follows that the geodesic problem on M is equivalent to finding a horizontal curve of minimal length that connects the initial manifold S0 = g0 K to the terminal manifold S1 = g1 K. Next to the geodesic problem we will also consider another Tnatural geometric problem on M , the problem of minimizing the integral 12 0 κ 2 (t) dt, where κ(t) denotes the geodesic curvature of a curve x(t) in M that satisfies the given tangential conditions at t = 0 and t = T. We will refer to this problem as the elastic problem for the following historical reasons. This problem originated with a study of Daniel Bernoulli, who in 1742 suggested to L. Euler that the differential equation describing the equilibrium

7.2 Elliptic geometry

105

shape of a thin inextensible beam subject to bending torques at its ends can be found by making the integral of the square of the curvature along the beam a minimum. Euler, acting on this suggestion, obtained the differential equation for this problem in 1744 and was able to describe its solutions, known since then as the elasticae [E; Lv]. Its modern treatment was initiated by P. Griffiths, who obtained the Euler–Lagrange equation on simply connected spaces of constant curvature [Gr]. This study inspired several other papers [BG; LS; Je], each of which tackled the problem through its own, but somewhat disparate methods. The elastic problem can be naturally formulated as a two point boundary value problem in the isometry group G as an optimal control problem over the Serret–Frenet system dg = g(t)(A1 + u(t)A3 ) (7.11) dt T in G that seeks the minimum of the integral 12 0 u2 (t) dt among the solutions that satisfy the boundary conditions g(0) = g0 and g(T) = g1 . This formulation draws attention to a more general class of variational problems, called affine-quadratic, that plays an important role in the theory of integrable systems and also lends itself to quick and elegant solutions. The focus on Lie groups, rather than on the spaces on which the groups act, reveals a curious twist in our understanding of the classical geometry. Our literature on geometry, dominated by its historical origins, tends to favor the Euclidean geometry over its non-Euclidean analogues. In this mindset, the leap to non-Euclidean geometries is drastic, and requires a great deal of intellectual adjustment. However, had the history started with non-Euclidean geometry first, the passage to Euclidean geometry would require only a minor algebraic shift. To be more explicit, note that K acts linearly and irreducibly on p by conjugation. Let G0 denote the semi-direct product p K. Evidently, G0 is isomorphic to SE2 (R), the group of motions of the plane. The restriction of , to p is K-invariant and hence defines a Euclidean metric on p . Thus p together with , is a two-dimensional Euclidean space E2 . Then p can be realized as the quotient space G0 /K through the action (x, R)(y) = x+Ry of G0 on p : p is the orbit of G0 through the origin in p , and K is the isotropy group. If g0 denotes the Lie algebra of G0 , then its points are pairs (A, B), A ∈ p and B ∈ k and [(A1 B1 ), (A2 , B2 )] = (adB1 (A2 ) − adB2 (A1 ), [B1 , B2 ]) is the Lie bracket in g0 . Then (A, B) in g0 can be i identified with A + B in g , in which case the semi-direct Lie bracket is given by (adB1 (A2 ) − adB2 (A1 ), [B1 , B2 ]) = [B1 , A2 ] − [B2 , A1 ] + [B1 , B2 ].

(7.12)

106

7 Hamiltonian view of classic geometry

Of course, in this case k is a one-dimensional Lie algebra and [B1 , B2 ] = 0. With this identification g , as a vector space, carries a double Lie algebra g and g0 . However, relative to the semi-direct Lie bracket, any two elements in p commute. The reader may easily check that A1 , A2 , A3 in the semi-direct case conform to the Lie bracket table (Table 7.1) with = 0. As in the non-Euclidean setting, horizontal curves in G0 are the curves that dg satisfy g(t)−1 dg dt ∈ p , or dt = g(t)(u1 (t)A1 + u2 (t)A2 ) for some control functions u1 (t) and u2 (t). If we denote g(t) by the pair (x(t), R(t)) ∈ p K, then dR dx = u1 (t)AdR (A1 ) + u2 (t)AdR (A2 ), = 0, dt dt 2 2 2 and || dx dt || = u1 + u2 . An analogous argument applies to the Euclidean elastic problem, and therefore we may conclude that for any left-invariant optimal problem on G , there is a corresponding Euclidean problem on G0 = p K.

7.3 Sub-Riemannian view We will now show that the above classical problems are easily solvable by the theoretic ingredients provided by the previous chapters. Let h1 , h2 , h3 denote the Hamiltonian lifts of the left-invariant vector fields Xi (g) = gAi , i = 1, 2, 3. In the left-invariant realization of T ∗ G as the product G × g∗ , these functions are given by hi () = (Ai ), i = 1, 2, 3, where ∈ g∗ . Functions h1 , h2 , h3 may be regarded as the coordinates of a point ∈ g∗ relative to the dual basis A∗1 , A∗2 , A∗3 . The sub-Riemannian problem of finding the horizontal curves of minimal length leads to the energy Hamiltonian H=

1 2 (h + h22 ). 2 1

(7.13)

Recall now that any function H on g∗ may be considered as a left-invariant Hamiltonian on G in which case its Hamiltonian equations can be written d dg = g(t)dH, = −ad∗ dH((t))((t) dt dt with (g, ) ∈ G × g , where dH = proposition is basic.

∂H ∂h1 A1

+

∂H ∂h2 A2

+

∂H ∂h3 A3 .

(7.14) The following

Proposition 7.4 C = h21 + h22 + h23 is an integral of motion for any leftinvariant Hamiltonian H.

7.3 Sub-Riemannian view

Proof

107

It suffices to prove that the Poisson bracket {I, H} = 0:

{C, H}() = ([dI, dH]) 0 / ∂H ∂H ∂H = h1 A1 + h2 A2 + h3 A3 , A1 + A2 + A3 ∂h1 ∂h2 ∂h3 ∂H ∂H ∂H ∂H = −h1 A3 + h1 A2 + h2 A3 − h2 A1 ∂h2 ∂h3 ∂h1 ∂h3 ∂H ∂H +h3 − A2 + ∂h1 ∂h2 ∂H ∂H ∂H ∂H + h1 h2 + h2 h3 − h1 h2 ∂h2 ∂h3 ∂h1 ∂h3 ∂H ∂H = 0. + h3 −h2 + h1 ∂h1 ∂h2

= −h1 h3

Function C is called a Casimir or an invariant function. It follows that the Hamiltonian H = 12 (h21 + h22 ) is a scalar multiple of the Casimir function in the Euclidean case, and hence, Poisson commutes with any function on g∗0 . But this means that each of h1 , h2 , h3 are constant along the flow of H. Since the extremal controls satisfy u1 = h1 and u2 = h2 , the solutions g(t) = (x(t), R(t)) ∈ p K are the solutions of dx dR = AdR(t) (h1 A1 + h2 A2 ), = 0. dt dt Hence, R(t) is constant and x(t) = x0 + th1 AdR (A1 ) + th2 AdR (A1 ) are straight lines in p . In the non-Euclidean cases the solutions reside on the intersection of the energy cylinder H = 12 (h21 + h22 ) and the hyperboloid C = h21 + h22 − h23 , in the hyperbolic case, and the sphere C = h21 + h22 + h23 , in the elliptic case. This implies that h3 is constant along the flow of H (which we already knew from the symmetry considerations, since the metric is invariant under K). The extremals which project onto the geodesics in M satisfy the transversality conditions (0)(k ) = 0 and (T)(k ) = 0. This means that h3 = 0. Equations (7.14) imply that dh1 dh2 = h3 h2 , = −h3 h1 . dt dt If we now write h = h1 + ih2 then dh = −ih3 h. dt

(7.15)

(7.16)

108

7 Hamiltonian view of classic geometry

It follows that h(t) = ae−h3 t , with a = h(0). Then, dg 1 0 h(t) = g(t)(h1 (t)A1 + h2 (t)A2 ) = g(t) ¯ 0 dt 2 − h(t) & & % i % i 1 0 0 0 a e 2 h3 e− 2 h3 (7.17) = g(t) i i 2 − a¯ 0 0 e− 2 h3 0 e 2 h3 The preceding equality implies that g(t) = g0 e(P +Q)t e−Qt ,

(7.18)

0 a and Q = h3 A3 , are the solutions to the sub− a¯ 0 Riemannian problem. The sub-Riemannian geodesics satisfy an additional condition H = 12 , which implies that |a|2 = a21 + a22 = 1 The above formula can be simplified by calculating the exponentials of a −ia3 , where h3 is replaced by a3 for the matrices: let L = 12 − a¯ −ia3 with P =

1 2

uniformity of notation. Then, L2 = − a23 = C. Therefore,

|a|2 +a23 I= 4

2

− α4 I, where α 2 = |a|2 +

t2 t3 eLt = I + Lt + L2 + L3 + · · · 2! 3! & % 2 4 6 t t 1 t 2 1 3 1 = I 1 + (−) α + (− ) α + (−) α ··· 2! 2 4! 2 6! 2 % & 3 5 t 1 t t 2 1 α − α + (−)2 α + ··· . + L α 2 3! 2 5! 2 In the elliptic case, # $ sin 2t α t α I+ 2L, e = cos 2 α

Lt

(7.19)

while, in the hyperbolic case, there are three distinct possibilities depending on the sign of C. If C > 0 then α > 0 and # $ sinh 2t α t Lt α I+ 2L. e = cosh 2 α

7.3 Sub-Riemannian view

109

√ √ In the case that C < 0 then α = i −C and cosh (tα) = cos (t −C). The remaining case, C = 0, yields eLt = I + Lt. When |a|2 = 1, then + + t 1 t 1 − a23 I + + sinh 1 − a23 L, 1 − a23 > 0, eLt = cosh 2 2 2 1 − a3 eLt = I + Lt, 1 − a23 = 0, + + t 1 t eLt = cos a23 − 1I + + a23 − 1L, 1 − a23 < 0. sin 2 2 2 a −1 3

If we write eLt =

u(t) v(t)

v(t) u(t)

for some complex numbers u(t), v(t), then,

it t ia3 t a it t α− sinh α , v(t) = e 2 a3 sinh α, α 2 = 1 − a23 , u(t) = e 2 a3 cosh 2 α 2 α 2 it it at it (7.20) u(t) = e 2 1 − a3 , v(t) = e 2 , a23 = 1, 2 2 it t ia3 t a it t α− sin α , v(t) = e 2 a3 sin α, α 2 = a23 − 1. u(t) = e− 2 a3 cos 2 α 2 α 2 The projected curves z(t) on the unit disk D are given by ⎧ + ⎪ a sinh 2t 1−a23 ⎪ ⎪ + + + , 1 − a23 > 0, ⎪ ⎪ t ⎪ ⎨ 1−a23 cosh 2 1−a23 +ia3 sinh t 1−a23 at , a23 = 1, z(t) = 1+it + ⎪ ⎪ ⎪ a sin 2t a23 −1 ⎪ 2 ⎪ ⎪ ⎩ +a2 −1 cos t +a2 −1+ia sin t+a2 −1 , 1 − a3 < 0. 3

2

3

3

(7.21)

3

This family of curves in the unit disk move along the generalized circles (circles and lines) through the origin with centers at c = −ia 2a3 . For each curve z(t) in the above family, a3 is the geodesic curvature of z(t). Curves with a23 < 1 satisfy + sinh2 2t 1 − a23 . |z(t)|2 = + sinh2 2t 1 − a23 + 1 − a23 Since the center c of these circles is greater than 12 , z(t) moves along the a −a to z(∞) = α+ia , with α 2 = 1 − a23 . circular segment from z(∞) = α−ia 3 3 The limiting case a3 = 0 results in the line t (7.22) z(t) = a tanh , 2 which is the hyperbolic geodesic through the origin.

110

7 Hamiltonian view of classic geometry

The nilpotent case, a23 = 1 corresponds to the circle with center at ia2 . This special class of circles, which are tangential to the boundary |z| = 1 are called the class of horocycles. The point on the boundary on such a circle is identified with z(±∞). The remaining circles, a23 > 1 are contained entirely in D. Along + . such circles, z(t) undergoes circular periodic motion with period T = (2n+1)π a23 −1 u(t) v(t) In the elliptic case, eLt = and the solutions are given by −¯v(t) u(t) ¯ it t ia3 t a it t a3 2 cos α− sin α , v(t) = e 2 a3 sin α, α 2 = 1 + a23 . u(t) = e 2 α 2 α 2 (7.23) The projected curves z(t) =

a sin 2t α B = α cos 2t α + ia3 sin 2t α A¯

are circles through the origin, having centers at geodesic, which is given by the line t z(t) = a tan , 2

ia 2a3 ,

with the exception of the

(7.24)

when a3 = 0. Let us now comment on some phenomena in sub-Riemannian geometry that are not present in Riemannian geometry. First, some definitions which are common to both of these geometries. The set of all points g ∈ G whose distance from the origin is T is called the sub-Riemannian sphere of radius T centered at I and will be denoted by ST (I). The sub-Riemannian ball of radius T, denoted by BT (I), is the set of all points whose distance from I is less or equal than T. Similar to these notions is the notion of the exponential map, or the wave front, which assigns to each sub-Riemannian geodesic its endpoint. In our situation, the exponential map at T is the mapping from the cylinder |a|2 = 1, a3 ∈ R onto the point g = g0 e(P +Q)t e−Qt where P and Q are the matrices defined by (7.18). It will be denoted by WI (T). In the problems of Riemannian geometry, the exponential map coincides with the sphere of radius T for sufficiently small T. In the problems above, the Riemannian geodesics are given by (7.22) in the hyperbolic case and by (7.23) in the elliptic case. In the hyperbolic case, T is the shortest distance from I and z(T) = a tanh 2t ; hence the wave front coincides with the Riemannian sphere for any T > 0. In the elliptic case, the Riemannian sphere of radius T coincides with the wave front up to T = π , beyond which the geodesic stops being optimal. This is evident on the Riemann sphere: the geodesic in (7.24) traces

7.3 Sub-Riemannian view

111

the great circle through the south pole of the Riemann sphere and reaches the north pole at T = π . Every point beyond the north pole is reachable by the geodesic traced in the opposite direction in time less than π . In problems of sub-Riemannian geometry the wave front WI (T) is never equal to SI (T), no matter how small T. The argument is essentially the same in both the elliptic and the hyperbolic case. In the hyperbolic case, the subRiemannian geodesics generated by (a, a3 ) with a23 > 1 are of the same form as the geodesics in the elliptic case. For any T > 0, v = 0 for T2 α = nπ , n = 1, 2, . . . . For such values of α, the endpoint of the geodesic g(T) generated ' by A 0 any a, |a| = 1, reaches the isotropy group K = : |A| = 1 . In 0 A¯ fact, if αn is defined by T 2

+

αn2 + = nπ , = ±1,

then T

An = (−1)n e−iαn 2 . This means that all the geodesicsgenerated by {(a, αn ) : |a| = 1} arrive at the An 0 in K. Therefore gn (T) is the intersection of same point gn (T) = 0 A¯ n several geodesics, and hence gn (t) cannot be optimal for t > T. This implies that for any T > 0, there exist geodesics g(t) that emanate from I at t = 0 and arrive at K in any time less than T. Hence g(T) is in WI (T) but not in SI (T). small T. To see this, Furthermore, WI (T) is not closed for any sufficiently ( T note that limn→∞ (−1)n e−iαn 2 = 1 whenever T2 αn2 + = nπ , = ±1. That means that the sequence of geodesics gn (t) defined by any (a, αn ), subject to |a| = 1, converges to the identity at time T. But I is not reachable by a geodesic for sufficiently small time T. Let us end this discussion with the hyperbolic geodesics in D. An easy calculation shows that if z(t) = a tanh 2t then t = log

1 + |z(t)| . 1 − |z(t)|

To find the hyperbolic distance between any two points z1 and z2 use the z−z1 isometry w = 1−¯ z1 z to transform z1 to the origin. Hence, the distance between z2 −z1 z1 and z2 is the same as the distance from 0 to 1−¯ z1 z2 . It follows that this distance is given by

112

7 Hamiltonian view of classic geometry z2 −z1 1 + 1−¯ z1 z2 . log z2 −z1 1 − 1−¯ z1 z2

Hyperbolic geometry plays an important and profound role in the theory of analytic functions on the unit disk largely due to the following proposition [Os]. Proposition 7.5 The Schwarz–Pick lemma Let d(z1 , z2 ) denote the hyperbolic distance between points z1 and z2 in D. Then d( f (z1 ), f (z2 )) ≤ d(z1 , z2 )

(7.25)

for any analytic function on D. The equality occurs only for the Moebius transformations defined by the action of SU(1, 1). The following lemma is crucial. Lemma 7.6 The Schwarz lemma Let F be an analytic function in D such that F(0) = 0. The |F(z)| ≤ |z| for all z ∈ D. If the equality occurs for some z0 ∈ D then F(z) = αz with |α| = 1. The proof can be found in any book on complex function theory. We will also need the following observation: function f (x) = log 1+x 1−x is monotone increasing in the interval (0, 1). We now turn to the proof of the proposition. Proof

2 2 Suppose that S(z) = az+b ¯ a with |a| − |b| = 1 satisfies S(z1 ) = 0, bz+¯

z−z1 then S(z) = 1−¯ z1 z α for some complex number α with |α| = 1. The inverse z+z S−1 (z) = 1+¯z11z α¯ takes 0 to z1 . The composition F(z) = S1 ◦ f ◦ S−1 (z), where z−f (z1 ) S1 (z) = 1− , is an analytic function in D that satisfies F(0) = 0. By f¯ (z1 )z Schwarz’s lemma, |F(z)| ≤ |z|, which means that |S1 ◦f (z)| ≤ |S(z)|. Therefore, f (z2 ) − f (z1 ) z2 − z1 ≤ 1 − f¯ (z )f (z ) 1 − z¯ z . 1 2 1 2

But then f (z2 )−f (z1 ) z2 −z1 1 + 1− 1 + ¯ 1−¯z z f (z1 )f (z2 ≤ log 1 2 = d(z1 , z2 ). d( f (z1 ), f (z2 )) = log (z2 )−f (z1 ) z2 −z1 1 − f1− 1 − 1−¯ z1 z2 f¯ z f (z ) 1

2

In the case F(z) = αz with |α| = 1, f (z) = αS1−1 ◦ S(z).

7.4 Elastic curves

113

7.4 Elastic curves The Hamiltonian H = 12 h23 + h1 associated with the elastic problem is also completely integrable, because H, the Casimir C = h21 + h22 + h23 , and the Hamiltonian lift of any right-invariant vector field form an involutive family on T ∗ G. In fact, the same can be said for any left-invariant Hamiltonian on T ∗ G. Recall that HX (g, ) = (g−1 Ag) is the Hamiltonian lift of a rightinvariant vector field X(g) = Ag in the left-invariant trivialization of T ∗ G. In this setting, integrability of left-invariant Hamiltonians on T ∗ G coincides with the integrability on coadjoint orbits. Each coadjoint orbit in g∗ is identified with the hypersurface h21 + h22 + h23 = c and the integral curves of H reside in the intersection of the surfaces H = c1 and C = c2 for some constants c1 and c2 . H = c1 is a cylindrical paraboloid in the coordinates h1 , h2 , h3 and C = c2 is a cylinder h21 + h22 = c2 , h3 ∈ R in the Euclidean case, the sphere h21 + h22 + h23 = c2 in the elliptic case and the hyperboloid h21 + h22 − h23 = c2 in the hyperbolic case. ∗ The Hamiltonian equation d dt = − ad dH((t))((t)) means that d dt (X) = − [dH, X] for any X ∈ g , where dH = A1 + h3 A3 . It follows that d dh1 = (A1 ) = −([A1 + h3 A3 , A1 ]) = h3 h2 , dt dt dh2 d = (A2 ) = −[A1 + h3 A3 , A2 ]) = h3 − h1 h3 , dt dt dh3 d = (A3 ) = −[A1 + h3 A3 , A3 ]) = −h2 . dt dt Together with dg dt = g(t)(A1 + h3 (t)A3 ) these equations constitute the extremal curves of the elastic problem. The projections x(t) = g(t)K on the quotient M = G /K are called elastic. It follows from the previous section that h3 (t) is the signed geodesic curvature of x(t). Proposition 7.7 ξ(t) = h23 (t) is a solution of the following equation: 2 dξ = −ξ 3 (t) + 4ξ 2 (t)(H − ) − 4ξ(t)(H 2 − C). (7.26) dt Proof 2 dξ = 4h22 h23 = 4(C − h21 − h23 )h23 dt % % & & 2 1 2 2 1 = 4 C − H − h3 − h23 h23 = 4 C − H − ξ − ξ ξ 2 2 = −ξ 3 + 4ξ 2 (H − ) − 4ξ(H 2 − C)

114

7 Hamiltonian view of classic geometry

Change of variable ξ = −4p − 43 (H − ) transforms equation (7.26) into 2 dp = 4p3 − g2 p − g3 , (7.27) dt where g2 =

4 8 1 (H − )2 + (H 2 − C) and g3 = (H − )3 + (H − )(H 2 − C). 3 27 3

The solutions of this equation are well known in the literature on elliptic functions. They are of the form p(t) = ℘ (t − c) where ℘ is the function of Weierstrass [Ap] (meromorphic and doubly periodic in the complex plane having a double pole at z = 0). There is a natural angle θ that links the above equation to the equation of a mathematical pendulum. The pendulum equation is obtained by introducing an angle θ defined by − h1 (t) = J cos θ (t), h2 (t) = J sin θ (t) along each extremal curve h1 (t), h2 (t), h3 (t). Here J is another constant obtained as follows ( − h1 )2 + h22 = 1 − 2h1 + h21 + h22 = 1 − (2H − h23 ) + C − h23 = J 2 , that is, J 2 = 1 − 2H + C. Then, −J sin θ h3 = −h3 h2 = − hence,

dθ dt

dθ dh1 = −J sin θ , dt dt

= h3 (t). This equation then can be written as

( dθ = ± 2(H − ) + 2J cos θ , (7.28) dt 2 2 2 , and H = 1 dθ since dθ = h + − J cos θ . 3 dt 2 dt Equation (7.28) is the equation is the equation of the mathematical pendulum with energy E equal to H − . According to A. E. Love, G. Kirchhoff was the first to notice the connection between the elastic problem and the pendulum (see [Lv] for further details). Kirchhoff referred to the pendulum as the “kinetic analogue” of the elastic problem. The connection with the pendulum reveals two distinct cases, which Love calls inflectional and non-inflectional. The inflectional case corresponds to E < J. In this case there is a cut-off angle θc and the pendulum oscillates between −θc and θc . The curvature changes sign at the cut-off angle (which accounts for the name). The remaining case E ≥ J corresponds to the noninflectional case, where the curvature does not change sign. The case E = H is

7.5 Complex overview and integration

115

the critical case, where the pendulum has just enough energy to reach the top (at t = ∞). Otherwise, the pendulum has sufficient energy to go around the top (see [Je] for further details). The passage from equation (7.27) to the pendulum equation (7.28) transforms the solutions expressed in terms of the elliptic function of Weierstrass to the solutions expressed in terms of Legendre’s function sn(u, k2 ). This transformation is achieved via the following change of variables: if 2φ = θ , then (7.28) becomes + ( ˙ (7.29) 2φ = 2(E + J) 1 − k2 sin2 φ, where k2 =

or

2J E+J .

Under the substitution x = sin φ, equation (7.29) becomes E + J( x˙ = (1 − x2 )(1 − k2 x2 ), 2 (

dx (1 − x2 )(1 − k2 x2 )

The function z = f (u) defined by u =

z

√

=

E+J t. 2

dz 1−z2 )(1−k2 z2 )

is called Legendre’s + E+J 2 2 function and is denoted by z = sn(u, k ). It follows that x = sn 2 t, k , 0

and therefore the extremal curvature κ(t) is given by 1 % & 2 2 ( E+J 2 3 2 2 t, k . κ(t) = 2(E + J) 1 − k sn 2

7.5 Complex overview and integration We will now show that the remaining equations are solvable by quadrature in terms of κ(t). The integration procedure is essentially the same in both the elliptic and the hyperbolic case. This fact is best demonstrated by passing to complex Lie group G = SL2 (C) and to the real forms of the complex Lie algebra g = sl2 (C). a b The Lie algebra g of SL2 (C) consists of matrices M = with c −a complex entries a, b, c. Any such matrix can be written as M = a1 A1 + a2 A2 + a3 A3 where A1 , A2 , A3 are the Pauli matrices defined by (7.7). In fact, b−c a = − 2i a3 , b+c 2 = ia2 , 2 = a1 . A real Lie algebra h is said to be a real form for a complex Lie algebra g if g = h + ih. Both {g , = ±1} are real

116

7 Hamiltonian view of classic geometry

forms of sl2 (C) because su2 is the linear span of matrices with real coefficients a1 , a2 , a3 and su(1, 1) it the linear span of matrices with coefficients ia1 , ia2 , a3 , as a1 , a2 , a3 range over all real numbers. The real Poisson structure also extends to complex Poisson structure on g∗ with { f , h}() = ([df , dh] for any complex functions f and h on g∗ and any point in g∗ . Then any complex function H on g∗ defines a Hamiltonian vector field H on G × g∗ with integral curves (g(t), (t)) the solutions of dg d = g(t)dH((t)), = −ad∗ dH((t))((t)). dt dt

(7.30)

In particular, H = 12 h23 + h1 , where h1 , h2 , h3 now denote the dual (complex) coordinates of a point ∈ g∗ relative to the dual basis A∗1 , A∗2 , A∗3 in g∗ , may be considered as the complexification of the Hamiltonian generated by the elastic problem. It follows that the Hamiltonian equations of H are given by dh1 dh2 dh3 d = g(A1 + h3 A3 ), = h3 h2 , = h3 − h1 h3 , = −h2 . dt dt dt dt

(7.31)

In this setting angle θ is defined in the same manner as in the real case with 1 − h1 = J cos θ , h2 = J sin θ , C2 and C2 = h21 + h22 + h23 . The corresponding pendulum with J 2 = 1 − 2H √ + √ dθ equation dt = 2 E + J cos θ, with E = H − 1, is now complex valued. We will make use of the fact that ∈ g∗ can be identified with L ∈ g via the formula (X) = L, X for all X ∈ g, where A, B = − 12 Tr(AB). Since the Pauli matrices remain orthonormal relative to this quadratic form, A, B = a1 b1 + a2 b2 + a3 b3 , for any matrices A and B in g. Moreover, the invariance property [A, B], C = A, [B, C] extends to the complex setting in g. Then = h1 A∗1 + h2 A∗ + h3 A∗ corresponds to L = h1 A1 + h2 A2 + h3 A3 . Under this identification equations (7.30) are identified with dg dL = g(t)(A1 + h3 A3 ), = [dH((t)), L(t)] = [A1 + h3 A3 , L(t)], (7.32) dt dt and each coadjoint orbit is identified with an adjoint orbit g(t)L(t)g−1 (t) = for some matrix ∈ g. It follows that C2 = L(t), L(t) is the Casimir on g. On each adjoint orbit, L(t), L(t) = , . Since SL2 (C acts transitively by conjugation on the complex variety {X ∈ g : X, X = , }, there is no loss in generality in assuming that = CA3 . Corresponding to this choice of normalization there is an adapted choice of coordinates θ1 , θ2 , θ3 on G defined by g = eθ1 A3 eθ2 A1 eθ3 A3 .

(7.33)

7.5 Complex overview and integration

117

Then gLg−1 = CA3 implies that eθ3 A3 L(t)e−θ3 A3 = Ce−θ2 A1 A3 eθ2 A1 . Since

% θ3 A3

e

=

&

i

e− 2 θ3 0

0 e

i 2 θ3

, eθ2 A1 =

cos 12 θ2 − sin 12 θ2

sin 12 θ2 cos 12 θ2

,

the above implies that h3 = C cos θ2 and h1 + ih2 = −iC(cos θ3 + i sin θ3 ) sin θ2 , −h1 + ih2 = −iC(cos θ3 − i sin θ3 ) sin θ2 . Therefore, h1 (t) = C sin θ3 sin θ2 , h2 (t) = −C cos θ3 sin θ2 , h3 = C cos θ2 .

(7.34)

The remaining angle θ1 will be now calculated from the equation g−1 dg dt = A1 +h3 A3 and the representation (7.33). Differentiation of (7.33) along a curve g(t) results in g−1

dθ1 −1 dg dθ2 −θ3 A3 dθ3 = g A3 g + e A3 . A1 eθ3 A3 + dt dt dt dt

(7.35)

An easy calculation shows that e−θ3 A3 A1 eθ3 A3 = cos θ3 A1 + sin θ3 A2 . If g(t) is a solution of g−1 dg dt = A1 + h3 A3 , then (7.33) yields h1 dθ1 h2 dθ1 h3 dθ1 dθ3 dθ2 dθ2 + cos θ3 = 1, + sin θ3 = 0, + = h3 . C dt dt C dt dt C dt dt (7.36) Therefore, C sin θ3 dθ1 sinθ3 Ch1 = = = . dt h1 sin θ3 − h2 cos θ3 sin θ2 sin2 θ2

(7.37)

Since h3 = C cos θ2 , sin2 θ2 = C2 − h23 , and hence, H − 12 h23 H − 12 κ 2 (t) dt. dt = C θ1 (t) = C C2 − κ 2 (t) C2 − h23 Let us finally note that z(t) = g(t)(0) = eθ1 A3 eθ2 A1 (0) = e−iθ1 tan θ2 is the projection of an extremal curve all the way down to SL2 (C)/SO2 (C). This projection is called complex elastica in [Jm]. Then elliptic elasticae are the projections of the extremal curves with real valued h1 , h2 , h3 , while hyperbolic elasticae are the projections of the extremal curves in which h1 and h2 are imaginary and h3 is real.

8 Symmetric spaces and sub-Riemannian problems

Remarkably, the group theoretic treatment of the classical geometry described in the previous chapter extends to a much wider class of Riemannian spaces, known as symmetric spaces in the literature on differential geometry [C1; Eb; Hl; Wf]. Our exposition of this subject matter is somewhat original in the sense that it explains much of the basic theory through symplectic formalism associated with canonical sub-Riemannian problems on Lie algebras g that admit a Cartan decomposition g = p ⊕ k. The presentation is motivated by two main goals. Firstly, we want to demonstrate the relevance of the mathematical formalism described earlier in this text for this beautiful area of mathematics. Secondly, we want to present the subject in a self-contained way that lends itself naturally to our study of integrable systems on Lie algebras and the problems of applied mathematics.

8.1 Lie groups with an involutive automorphism We will begin with Lie groups with an involutive automorphism, rather than the symmetric spaces themselves as is often done in the existing literature on this subject [Eb; Hl]. An automorphism σ on a Lie group G that is not equal to the identity which satisfies σ 2 = I is called involutive. More explicitly, an involutive automorphism σ satisfies: 1. 2. 3. 4.

σ : G → G is analytic. σ (gh) = σ (g)σ (h) for all g and h in G. σ (g) = e only for g = e. σ 2 (g) = g for all g in G.

The group of fixed points H0 = {g : σ (g) = g} is a closed subgroup of G, and hence a Lie subgroup of G. We will use H to denote the connected component of H0 that contains the group identity. 118

8.1 Lie groups with an involutive automorphism

119

The tangent map σ∗ at the identity is an automorphism of the Lie algebra g, that is, satisfies σ∗ ([A, B]) = [σ∗ (A), σ∗ (B)] for any A and B in g. It follows that (σ∗ + I)(σ∗ − I) = 0, because σ∗2 = I. Therefore, σ∗ (M) = M, or σ∗ (M) = −M for any M in g. Let p = {M : σ∗ (M) = −M} and h = {M : σ∗ (M) = M}. Since σ∗ ([A, B]) = [σ∗ (A), σ∗ (B)], the following Lie bracket conditions hold: [p, p] ⊆ h, [h, p] ⊆ p, [h, h] = h

(8.1)

We will refer to the above Lie algebra decomposition as the Cartan decomposition induced by σ , or simply the Cartan decomposition. The Lie algebraic conditions (8.1) imply that h is a Lie subalgebra of g. Proposition 8.1

h is the Lie algebra of H.

Proof For any M in the Lie algebra of H, exp tM is in H, and therefore σ (exp tM) = exp tM for all t. It follows that σ∗ (M) = M, and hence M belongs to h. We have shown that h contains the Lie algebra of H. Suppose now that M is any point in h. Then σ (exp tM) = exp t σ (M) = exp tM. The above implies that {exp tM : t ∈ R} is a curve in H. But then M belongs to the Lie algebra of H. Proposition 8.2

Adh (p) is contained in p for any h in H.

Proof Let B denote an arbitrary element of h. Then for any A in p, Adexp tB (A) = exp t(adA(B)). Since adA(B) belongs to p it follows that Adexp tB (A) ∈ p for all t. But then any h in H can be written as the product h = exp tm Bm ◦ exp tm−1 Bm−1 ◦ · · · exp t1 B1 . Then Adh (A) = Adexp tm Bm ◦ · · · Adexp t1 B1 (A), and therefore Adh (A) belongs to p for any A in p. p

p

Definition 8.3 We shall write Adh for the restriction of Adh to p, and AdH will p denote the subgroup of Gl(p) generated by {Adh : h ∈ H}. Definition 8.4 The pair (G, H) will be called a symmetric pair if G is a Lie group with an involutive automorphism and H is a closed and connected p subgroup of fixed points of the automorphism. If in addition AdH is a compact subgroup of Gl(p) then (G, H) will be called a Riemannian symmetric pair.

120

8 Symmetric spaces and sub-Riemannian problems

8.2 Symmetric Riemannian pairs The following proposition marks a point of departure for a kind of variational problems that characterize symmetric spaces. Proposition 8.5 Let (G, H) be a symmetric pair. There exists a positivep p definite quadratic form , on p that is invariant under AdH if and only if AdH is compact in Gl(p). Proof Any positive-definite form , on p turns p into a Euclidean vector space. Then the orthogonal group O(p) is the largest group that leaves , p p invariant. If a positive-definite form , is invariant under AdH then AdH is p p contained in O(p). Since O(p) is compact and AdH is closed, AdH is compact. p Suppose now that AdH is a compact subgroup of Gl(p). Let E(p) denote the p vector space of all symmetric quadratic forms on p. Then AdH acts on E(p) by T(Q)(x, y) = T −1 x, T −1 y p

for each quadratic form Q = , in E(p), and each T in AdH . p p Since AdH is compact, continuous functions on AdH with values in E(p) can be integrated so as to obtain values in E(p). In this situation, integration commutes with linear maps [Ad]. It then follows that the integral g(Q) dg p

AdH p

defines a positive-definite, AdH -invariant quadratic form for any positivedefinite quadratic form Q on E(p). We will now suppose that (G, H) is a Riemannian symmetric pair with a positive-definite, AdH -invariant quadratic form , in p, where invariance means AdH (A), AdH (B) = A, B for any A and B in p. It then follows, by differentiating Adexp tC (A), Adexp tC (B) = A, B at t = 0, that [C, A], B + A, [C, B] = 0 for any element C in h, or that [A, C], B = A, [C, B]

(8.2)

for A and B in p and C in h. Recall now the Killing form Kl(A, B) = Tr(ad(A) ◦ ad(B)) defined in Chapter 5 and its basic invariance property relative to the automorphisms

8.2 Symmetric Riemannian pairs

121

φ on g. This invariance property implies that Kl is AdH -invariant, and secondly it implies that p and h are orthogonal, because Kl(A, B) = Kl(σ∗ (A), σ∗ (B)) = Kl(−A, B) = −Kl(A, B). The following proposition is a variant of a well-known theorem about the simultaneous diagonalization of two quadratic forms, one of which is positivedefinite. Proposition 8.6 Let p0 = {A ∈ p : Kl(A, X) = 0, X ∈ p}. Then p0 is ad(h)-invariant. Moreover, there exist non-zero real numbers λ1 , λ2 . . . , λm and linear subspaces p1 , p2 , . . . , pm such that: (i) p = p0 ⊕ p1 ⊕ · · · ⊕ pm . (ii) p0 , . . . , pm are pairwise orthogonal relative to Kl. (iii) Each pi , i ≥ 1 is ad(h)-invariant and contains no proper ad(h)-invariant subspaces. (iv) For each i = 1, . . . , m Kl(A, B) = λi A, B for all A and B in pi . Proof If p0 = {A ∈ p : Kl(A, X) = 0, X ∈ p} then, Kl(ad(h)(p0 ), X) = Kl(p0 , ad(h)(X)) = 0; hence, ad(h)(p0 ) ⊆ p0 . Let p⊥ 0 denote the orthogonal is also invariant under ad(h). complement in p relative to , . It follows that p⊥ 0 = 0. Therefore, Kl is non-degenerate Since , is positive-definite, p0 ∩ p⊥ 0 ⊥ on p0 . Let p1 denote the set theoretic intersection of all ad(h)-invariant subspaces ⊥ of p⊥ 0 . Then p1 is an ad(h)-invariant subspace of p0 that admits no proper subspace that is invariant under ad(h). Now proceed inductively: if p0 ⊕ p1 is not equal to p relative to , then let p2 be equal to the intersection of all ad(h)-invariant subspaces in p⊥ 1 . Continue until the m-th stage at which = 0. Then, p = p ⊕ p ⊕ · · · ⊕ pm . It is clear from the construction that p⊥ 0 1 m each pi , i = 1, . . . , m does not admit any proper ad(h)-invariant subspaces. Therefore, up to a reordering, this decomposition is unique. If we now repeated this process with the orthogonal complements taken relative to Kl, the resulting decomposition would be the same as the one above, up to a possible renumbering of the spaces. This shows that Kl(pi , pj ) = 0, i = j. It remains to show part (iv). Let A denote the point in pi that yields the maximum of Kl(x, x) over the points in pi that satisfy x, x = 1. According to the Lagrange multiplier rule there exists a number λi such that A is a critical point of F(x) = Kl(x, x) + λi (1 − x, x), that is, Kl(A, X) = λi A, X, for all Xpi . The multiplier λi cannot be equal to zero because pi is orthogonal to p0 .

122

8 Symmetric spaces and sub-Riemannian problems

Then Kl(ad(h)(A), X) = λi ad(h)(A), X. Since pi does not admit any proper ad(h)-invariant subspaces, the action h → ad(h)(A) is irreducible. This implies that ad(h)A contains an open set in pi . Therefore, Kl(Y, X) = λi Y, X for all X and Y in pi . Proposition 8.7 Let H be a Lie group such that AdH is a compact subgroup of Gl(g). Then the Killing form is negative-definite on the Lie algebra h provided that h has zero center. Proof Let , denote any positive-definite AdH invariant quadratic form on h, and let A1 , . . . , An denote any orthonormal basis relative to this form. Then for any B in h the matrix of ad(B) relative to this basis is skew-symmetric. If (nij ) are any skew-symmetric matrices. then the trace of M = (mij ) and N = # n n MN is equal to −2 i=1 j=1 mij nij . Therefore, n n Kl(B, B) = Tr(ad(B) ◦ ad(B)) = −2 m2ij ≤ 0. i=1 j=1

If Kl(B, B) = 0 then adB = 0, and hence B belongs to the center of h. But then B = 0. Corollary 8.8 If (G, H) is a Riemannian symmetric pair and if the Lie algebra h of H has zero center, then Kl is negative-definite on h. Proposition 8.9 Let p = p0 ⊕· · ·⊕pm denote the decomposition described by Proposition 8.6 associated with a Riemannian symmetric pair (G, H) where the Lie algebra h of H has zero center. Let gi = pi + [pi , pi ] for each i = 1, . . . , m. Then, (i) Each gi is a an ideal of g that is also invariant under the involutive automorphism σ∗ . (ii) [gi , gj ] = 0 and Kl(gi , gj ) = 0 for i = j. Proof Let Ai ∈ pi , Aj ∈ pj with i = j. Since ad(h)(pi ) ⊆ pi and pi and pj are orthogonal relative to Kl, Kl(ad(h)(Ai ), Aj ) = Kl(h, [Ai , Aj ]) = 0. But Kl[Ai , Aj ], h) = 0 implies that [Ai , Aj ] = 0 since Kl is negative-definite on h. Then [pi , [pj , pj ]] = 0 by the Jacobi’s identity, and the same goes for [[pi , pi ], [pj , pj ]]. Hence [gi , gj ] = 0. Evidently, Kl(gi , gj ) = 0, i = j. To prove that each gi is an ideal of g, let A + B be an arbitrary element of g with A ∈ p and B ∈ h. Since each pi is ad(h)-invariant, it follows that adB(pi ) ⊆ pi . This fact, together with the Jacobi’s identity implies that adB([pi , pi ]) ⊆ [pi , pi ]. Therefore, gi is ad(h)-invariant. Let A = A1 + · · · + Am with Ai ∈ pi for each i = 1, . . . , m. Then adA(gi ) = adAi (gi ) because [Aj , gi ] = 0. Relations (8.1) together with the

8.2 Symmetric Riemannian pairs

123

Jacobi’s identity imply that [Ai , gi ] ⊆ gi . Since it is evident that each gi is σ∗ invariant, our proof is finished. Proposition 8.10 The fundamental decomposition Let g1 , . . . , gm be as in the previous proposition. Define p0 = {A ∈ p : Kl(A, X) = 0, X ∈ p} and g0 = p0 ⊕ {B ∈ h : adB(p) ⊆ p0 }. Then, (i) [p, p0 ] = 0. (ii) g0 is an ideal in g. (iii) g = g0 ⊕ g+ ⊕ g− where g+ = ∪{gi : Kl|pi = λi , |pi , λi > 0}, and g− = ∪{gi : Kl|pi = λi , |pi , λi < 0}. Proof

Let A ∈ p, A0 ∈ p0 and B ∈ h. Then Kl([A, A0 ], B) = Kl(A0 , adB(A)) = 0.

Therefore, [A, A0 ] = 0 since Kl is negative-definite on h. To prove (ii) take first A ∈ p. Because of (i), it suffices to show that ad[A, B](p) ⊆ p0 for any B ∈ g0 ∩ h. But Kl([A, B], p) = Kl(A, [B, p]) = 0, since [B, p] ⊆ p0 . So[A, B] ∈ g0 . Now take A ∈ h and an element A0 +B0 in g0 with A0 ∈ p0 and B0 ∈ h. Since Kl([A, A0 ], p) = Kl(A, [A0 , p) = 0, [A, A) ] ∈ p0 . To show that [A, B0 ] ∈ g0 , we need to show that Kl([[A, B0 ], X], Y) = 0 for any X and Y in p. But that follows immediately from Jacobi’s identity applied to [[A, B0 ], X]. To show (iii), let h1 = ∪{[pi , pi ] : Kl|pi = λi , |pi , λi = 0}. Note that p0 coincides with the space of eigenvectors corresponding to zero eigenvalue of Kl. It is easy to see that h1 is an ideal in h. Let h0 denote its orthogonal complement relative to the Killing form. Claim that h0 ⊆ {B : Ai and adB(p) ⊆ p0 }. Indeed, if A ∈ p and X ∈ p are written as A = X = Xi with Ai and Xi in pi then for any B ∈ h0 , Kl(adB(A), X) = Kl(B, [A, X]) = Kl B, [Ai , Xi ] = 0. It follows that h ⊆ g0 ⊕ g+ ⊕ g− , and therefore (iii) is true. Recall now that a Lie algebra is said to be semi-simple if the Killing form is non-degenerate. A Lie algebra is said to be simple if it is semi-simple and contains no proper ideals. Proposition 8.11 ideals.

Every semi-simple Lie algebra g is a direct sum of simple

Proof Suppose that g0 is a simple ideal in g not equal to g. Let g⊥ 0 denote the orthogonal complement of g0 relative to Kl. Then, g = g0 ⊕ g⊥ 0 because Kl

124

8 Symmetric spaces and sub-Riemannian problems

is non-degenerate on g. Moreover, g⊥ 0 is an ideal in g. The proof is simple: if and B ∈ g , then X ∈ g, A ∈ g⊥ 0 0 Kl([X, A], B) = Kl(X, [A, B]) = 0. ⊥ Hence, [X, A] ∈ g⊥ 0 . Since Kl is non-degenerate on g0 , the argument can be ⊥ repeated whenever g0 is not a simple ideal.

Corollary 8.12

Every semi-simple Lie algebra has zero center.

Proof If C denotes the center of g then C is an ideal in g on which Kl is zero. By the argument above, g = C ⊕ C⊥ . But then C = 0 would violate the non-degeneracy of Kl. Proposition 8.13 Suppose that (G, H) is a Riemannian symmetric pair where the Lie algebra h of H has a zero center. Let gi = pi + [pi , pi ] be the ideals in g such that Kl restricted to pi is given by λi , with λi = 0. Then each gi is a simple ideal, i.e., it contains no proper non-zero ideals. Proof Any proper ideal of gi would split pi into a direct sum of two orthogonal ad(h)-invariant spaces. Since ad(h) acts irreducibly on pi one of these subspaces must be (0). Corollary 8.14 simple.

If g0 in Proposition 7.10 is equal to zero, then g is semi-

Remark 8.15 It is not necessary for g0 to be equal to zero for g to be semi-simple. It may happen that g0 is a subalgebra of h, in which case g is semi-simple, even though g0 = 0. Definition 8.16 Let g = g0 ⊕ g+ ⊕ g− denote the fundamental decomposition of g. Then M = G/H is said to be of Euclidean type if g0 = g, it is said to be of compact type if g0 = g+ = 0, and is said to be of non-compact type if g0 = g− = 0. The simplest type is the Euclidean type. It is characterized by the decomposition g = p ⊕ h where p is a commutative algebra. In such a case, G is isomorphic to the semi-direct product p AdH and G/H p. Let us now consider some particular cases covered by the above general framework. Example 8.17 Euclidean spaces Let K denote a Lie group that acts linearly on a finite-dimensional vector space V and let G denote the semi-direct product V K. Then G admits an involutive automorphism σ (x, k) = (−x, k) for each

8.2 Symmetric Riemannian pairs

125

(x, k) in G. It follows that H = {0} K is the group of fixed points of σ , and that the corresponding Lie algebra decomposition is given by p = V {0} and h = {0} k, where k denotes the Lie algebra of K. For (G, K) to be a symmetric Riemannian pair it is necessary and sufficient that AdH |p be a compact subgroup of Gl(p). It is easy to check that Adk (A, 0) = (k(A), 0) for each (A, 0) in p and each k ∈ K. It follows that K = AdH |p . Hence, for (G, H) to be a symmetric Riemannian pair K needs to be a compact subgroup of Gl(V). Any positive-definite AdH -invariant quadratic form on p turns V into a Euclidean space En and identifies K with a closed subgroup of the rotation group O(En ). Evidently, G/H = En . Example 8.18 (SLn (R), SOn (R)) and positive-definite matrices Let σ be the automorphism on G = SLn (R) defined by σ (g) = (gT )−1 , g ∈ G, where gT denotes the matrix transpose of g. Then σ (g) = g if and only if gT = g−1 , that is, whenever g belongs to SOn (R). Thus, the isotropy group H is equal to SOn (R). The Lie algebra g of G consists of n×n matrices with real entries having zero trace. Then σ∗ decomposes g into the sum p ⊕ h with p the space of symmetric matrices in g, and h the algebra of skew-symmetric matrices in g. This Cartan decomposition recovers the well-known facts that every matrix in g is the sum of a symmetric and a skew-symmetric matrix while Cartan relations reaffirm that the Lie bracket of two symmetric matrices is skew-symmetric. The Killing form on sln (R) is given by Kl(A, B) = 2nTr(AB). The form 1 Kl(A, B) is more convenient, since then A, B = A, B = 12 Tr(AB) = 4n 1 n a b for any matrices A and B in sln (R). In particular, A, A = i, j ij ji 2 n 2 n 2 i≥j aij for any symmetric matrix A, and A, A = − i>j aij for any skewsymmetric matrix A. It follows that , is a positive-definite, SOn (R)-invariant quadratic form on the space of symmetric matrices. This inner product turns (SLn (R), SOn (R)) into a symmetric Riemannian pair. It is easy to verify that the ideal g0 in Proposition 8.10 is equal to zero, which then implies that [p, p] = k. Since Kl is a positive multiple of , , (SLn (R), SOn (R)) is a symmetric Riemannian pair of non-compact type. The quotient space G/H can be identified with the space of positive-definite n × n matrices P(n, R) via the action (g, P) → gPgT , P ∈ P(n, Rn ), g ∈ SLn (R). For then, SOn (R) is the isotropy group of the orbit through the identity matrix, and therefore, P(n, Rn ) can be identified with the quotient SLn (R)/SOn (R).

126

8 Symmetric spaces and sub-Riemannian problems

Example 8.19 Self-adjoint subgroups of SLn (R) A subgroup G of SLn (R) is said to be self-adjoint if gT is in G for each g in G. It can be easily verified that G = SO(p, q) and Spn are self adjoint subgroups of SLn (R). Recall that p SO(p, q) is the group that leaves the Lorentzian form x, yp,q = − i=1 xi yi + n y invariant, while Spn is the group that leaves the symplectic form i=1 xi+n i x, y = ni=1 xi yn+i − ni=1 xn+i yi invariant. For self-adjoint subgroups G, the restriction of σ (g) = (gT )−1 to G can be taken as an involutive automorphism on G, in which case H = G ∩ SOn (R) becomes the group of fixed points. Moreover, p, the subspace of symmetric matrices in the Lie algebra g of G, and h, the subalgebra of skew-symmetric matrices in g, become the Cartan factors induced by this automorphism. Then the restriction of the trace metric A, B = 12 Tr(AB) to p defines an ad(h)-invariant metric on p. With this metric (G, H) becomes a symmetric Riemannian pair. When G = SO(p, q), the Lie algebra gconsists ofall n × n, n = p + q A B with A and C skewmatrices M having the block form M = BT C symmetric p×p and q×q matrices, and B an arbitrary p×q matrix. The Cartan 0 B , and h is the complementary space p consists of matrices M = BT 0 A 0 space M = . The isotropy group H is the group of matrices 0 C g1 0 0 g2 with g1 ∈ Op (R), g2 ∈ Oq (R) such that Det (g1 )Det(g2 ) = 1. We will denote this group by S(Op (R) × Oq (R)). The symmetric pair (G, H) together with the quadratic form A, B = 12 Tr(AB), A, B ∈ p becomes a Riemannian symmetric pair (of non-compact type). The homogeneous space M = SO(p, q)/S(Op (R)×Oq (R)) can be identified with the open subset of the Grassmannians consisting of all q-dimensional p subspaces in Rp+q on which the quadratic form x, xp,q = − i=1 xi2 + p+q 2 i=p+1 xi is positive-definite. Example 8.20 (SO(1, n), 1 × SOn (R)) and the hyperboloid Hn When p = 1, q = n, the isotropy group S(O1 × On ) is equal to {1} × SOn (R) and the quotient SO(1, n)/{1} × SOn (R) can be identified with the hyperboloid Hn via the following realization. Let x0 , x1 , . . . , xn denote the coordinates of a point x ∈ Rn+1 relative to the standard basis of column vectors e0 , e1 , . . . , en in Rn+1 . The group SO(1, n)

8.2 Symmetric Riemannian pairs

127

n n+1 : x2 − (x2 + · · · + x2 ) = acts on the n 0 1 points of the hyperboloid H = x ∈ R 1, x0 > 0 by the matrix multiplication. The action is transitive, and {1} × SO)n(R) is the isotropy group of the orbit through e0 . Hence, Hn can be identified with the orbit through the point e0 . In this identification, curves x(t) on Hn are identified with curves g(t) in SO(1, n) that satisfy x(t) = g(t)e0 for all t. Then, dg dx 0 B (t) = (t)e0 = g(t) e0 , BT C dt dt where B = (b1 , b2 . . . , bn ) and C is an n × n skew-symmetric matrix. It follows that dx bi g(t)ei , (t) = dt n

i=1

and therefore, . n n dx0 2 dxi 2 2 1 dx dx , =− + = bi (t) = TrM 2 , dt dt 1,n dt dt 2 i=1

i=1

0 B in p. BT 0 The above development shows that the canonical hyperbolic metric on Hn can be extracted from the trace metric on the Cartan space p in g. where M denotes the matrix

Example 8.21 (Spn , SUn ) and the generalized Poincar´e plane Matrices g 0 −I in Spn are defined by gT Jg = J, where J = . Since J T = −J, I 0 Spn is a self-adjoint subgroup of SL2n (R). The isotropy group H is equal to SO2n (R) ∩ Spn , which is equal to SUn . Therefore, (Spn , SUn ) is a symmetric Riemannian pair. We will presently show that the quotient space Spn /SUn can be identified with Pn , the set of n × n complex matrices Z of the form Z = X + iY with X, and Y symmetric matrices with real entries and Y positive-definite. This space is called the generalized Poincar´e plane. A B Let us first note that each g in Spn can be written in block form C D satisfying ABT = BAT , CDT = DCT , ADT − BCT = I. For n = 2, matrices A, B, C, D are numbers, and the preceeding conditions reduce to AD − BC = 1, from which it follows that SL2 (R) = Sp1 , and

128

8 Symmetric spaces and sub-Riemannian problems

SU1 = SO2 (R). Therefore, the generalized Poincar´e plane coincides with the hyperbolic upper half plane discussed in the preceding chapter. For a general n, matrices I B 0 I A 0 T , A ∈ GL (R), g = , g = , B = B g= n 0 I −I 0 0 (AT )−1 (8.3)

0 −I I C 0 I generate Spn . To see this, first note that = I 0 0 I −I 0 I 0 I 0 , therefore, every matrix with C an arbitrary symmetric C I C I matrix is generated by the matrices in (8.3). But, then, the exponentials & % 0 etA I tB I 0 , , T 0 I Ct I 0 e−tA are the generators of Spn , by the Orbit theorem (Proposition 1.6). We shall presently show that the fractional transformations W = (AZ + B)(CZ + D)−1

A B defined by the matrices g = in Spn define an action on Pn . It C D suffices to restrict our proof to matrices in (8.3) since they are the generators of Spn . Let Z = X + iY with X symmetric and Y positive-definite. Then, A 0 (Z) = AZAT = AXAT + iAYAT , 0 (AT )−1 I B (Z) = Z + B = (X + B) + iY. 0 I T Evidently, bothAXAT + iAYA , and (X + B) + iY belong to Pn . 0 I When g = then G(Z) = −Z −1 , so it remains to show that −I 0 −Z −1 belongs to Pn for any Z in Pn . Let V be a matrix in GLn (R) such that VYV T = I. Then

VZV T = VXV T + iI = S + iI. It follows that −Z −1 = −(X + iY)−1 = −((V −1 (S + iI)(V T )−1 )−1 = −V T (S + iI)−1 V.

8.2 Symmetric Riemannian pairs

129

Since S2 + I = (S + iI)(S − iI), −Z −1 = −V T (S + iI)−1 V = V T ((S2 + I)(S − iI)−1 )−1 )V = −V T (S − iI)(S2 + I)−1 V = P + iQ, with P = −V T S(S2 + I)−1 V and Q = V T (S2 + I)−1 V. Evidently S2 + I is positive-definite. Since the inverse of a positive-definite matrix is positivedefinite, Q is a positive-definite matrix. Matrix S is symmetric and commutes with S2 + I, and therefore S commutes with (S2 + I)−1 . Hence P is symmetric. This argument shows that −Z −1 is in Pn . We have now shown that Spn acts on Pn . Any point on the orbit through iI is of the form g(iI) = (iA + B)(iC + D)−1 . If Z = X + iY, then % & A 0 I B A AB (iI) (iI) = T o I 0 (AT )−1 0 A−1 = iAAT + ABAT = X + iY if A and B conform to Y = AAT and X = ABAT . Hence, Spn acts transitively on Pn . The reader can easily verify that the isotropy group consists of matrices I 0 A B T g = , and therefore g belongs to . But then gg = 0 I −B A SO2n (R) ∩ Spn . It follows that Spn /SUn = {X + iY X T = X, Y > 0}. Example 8.22 (SO(p, q), S(Op (R) × SOq (R)) and oriented Grassmann manifolds Let p and q be positive integers such that p + q = n and let D denote the diagonal matrix with its diagonal entries equal to −1 in the first p rows, and equal to 1 in the remaining q rows. Then D−1 = D. Let σ : SOn (R) → SOn (R) denote the mapping σ (g) = DgD−1 = DgD. It is easy to verify that σ is an involutive automorphism on G. The set of matrices A 0 g in G that are fixed by σ are of the form g = where A is a p × p 0 D matrix and D a q × q matrix. Denote this group by S(Op (R) × Oq (R)). It consists of pairs of matrices (A, D) with A ∈ Op (R) and D ∈ Oq (R) such that Det(A)Det(D) = 1. Then σ∗ splits the Lie algebra g = son (R) into the Cartan factors p ⊕h, with 0 B , and h p equal to the vector space of matrices P of the form P = −BT 0

130

8 Symmetric spaces and sub-Riemannian problems

A 0 , with both A and D antisymmetric; 0 D h is the Lie algebra of the isotropy group of H = S(Op (R) × Oq (R)). The pair (G, H) is a symmetric Riemannian pair with the metric on p defined by the Lie algebra of matrices Q =

1 P1 , P2 = − Tr(P1 P2 ). 2 We leave it to the reader to show that the Killing form on SOn (R) is given by Kl(A, B) = (n − 2)Tr(AB). Thus the Riemannian metric on p is a scalar multiple of the restriction of the Killing form. Since the Killing form is negative-definite on g, the pair (G, H) is of compact type. The homogeneous space Grp = SOn /SOp (R) × SOq (R) is the space of all oriented p- dimensional linear subspaces in Rp +q . So the symmetric space SOn /S(Op (R) × SOq (R)) is a double cover of Grp . Example 8.23 (SO1+n (R), {1} × SOn (R)) and the sphere Sn When p = 1 and q = n, then SOp (R) = 1 and the isotropy group H is equal to {1} × SOn (R). The set of oriented lines in Rn+1 is identified with the n-dimensional sphere Sn . More explicitly, the sphere is identified with the orbit of SOn+1 through the point e0 = (1, 0, . . . , 0) of Rn+1 under the action of matrices in SOn+1 on the column vectors of Rn+1 . The Riemannian metric induced by the Killing form on p in g coincides with the standard metric on Sn inherited from the Euclidean metric on Rn+1 . Indeed, any curve x(t) on Sn is identified with a curve g(t) in SOn+1 (R) via the formula x(t) = g(t)e0 . Let dg dt (t) = g(t)(P(t) + Q(t)) with ⎛ ⎜ ⎜ ⎜ P(t) = ⎜ ⎜ ⎝

in p and Q(t) ∈ h. Then, and therefore,

dx dt (t)

n dxi 2 0=1

dt

0 p1 p2 .. .

−p1 0 0 .. .

pn

0

=

(t) =

dg dt (t)e0 n i=1

... 0... 0... ... 0...

−pn 0 0 .. .

⎞ ⎟ ⎟ ⎟ ⎟ ⎟ ⎠

0

= g(t)(P + Q)e0 =

pi (t)2 = P(t), P(t).

n

i=1 pi (t)g(t)ei ,

8.3 The sub-Riemannian problem

131

8.3 The sub-Riemannian problem Let us now return the symmetric pairs (G, H) and the associated Cartan decomposition g = p ⊕ h. Definition 8.24 A curve g(t) in G defined over an open interval (a, b) will be called horizontal if g−1 (t) dg dt (t) belongs to p for all t in (a, b). Horizontal curves can also be considered as the integral curves of a leftinvariant distribution P defined by P(g) = {gA : A ∈ p}, g ∈ G, and as such they conform to the accessibility theory of families of analytic vector fields discussed in the previous chapters. In particular, the orbit of P through the group identity is a connected Lie subgroup K of G, and the orbit through any other point g ∈ G is the left coset gK. Let k denote the Lie algebra generated by p. The Herman–Nagano theorem implies that k is the Lie algebra of K. It follows that the group identity can be connected to any point g1 in K by a horizontal curve. More precisely, for any T > 0 there exists a horizontal curve g(t) defined on the interval [0, T] such that g(0) = e and g(T) = g1 . In the language of bundles, G is a principal H bundle over G/H, with H acting on G by the right action (h, g) → gh. The distribution P is called a connection. We will use π to denote the natural projection from G onto G/H. We will now show that for any curve x(t) in M and any g0 ∈ π −1 (x(0)) there is a unique horizontal curve g(t) such that g(0) = g0 and π(g(t) = x(t) for all t. We shall call such a curve the horizontal lift to g0 of a base curve x(t). Let x(t) = y(t)H for some representative curve y(t) ∈ G. Then π −1 (x(0)) = y(0)H and g0 = y(0)h0 for some h0 ∈ H. Every curve y(t) ∈ G is a solution of dy (t) = y(t)(A(t) + B(t)) dt for some elements A(t) ∈ p and B(t) ∈ h. In our case take g(t) = y(t)h(t), where h(t) is the solution of dh dt (t) = −B(t)h(t) with h(0) = h0 . Evidently π(g(t)) = x(t) and g(0) = y(0)h0 = g0 . To show that g(t) is a horizontal curve we need to differentiate: dg (t) = y(t)(A(t) + B(t))h(t) − y(t)B(t)h(t) dt = g(t)(h−1 (t)A(t)h(t) + h−1 (t)B(t)h(t)) − g(t)(h−1 (t)B(t)h(t)) = g(t)(h−1 (t)A(t)h(t)). Since h−1 (t)A(t)h(t) belongs to p, g(t) is horizontal. It is clear from the construction that such a lift is unique. If g1 (t) and g2 (t) are two horizontal lifts of the same base curve x(t), then g2 (0) = g1 (0)h for some h ∈ H. Then g(t) = g1 (t)h is a horizontal curve that

132

8 Symmetric spaces and sub-Riemannian problems

projects onto x(t) and satisfies g(0) = g2 (0). Therefore, g2 (t) = g1 (t)h, by the uniqueness of lifts. Horizontal curves that differ by a right multiple in H are called parallel. It follows that the horizontal lifts of a curve x(t) are parallel translates of a horizontal lift to any point g0 ∈ π −1 (x(0)). Definition 8.25 Hol(P) = K ∩ H is called the holonomy group of P. Proposition 8.26 The holonomy group is isomorphic to the loop group at any point x in the base space G/H. Proof Let x0 = g0 H be any base point, and let x(t) be any curve such that x(0) = x0 = x(T). Let g(t) denote the horizontal lift of x(t) to a point g0 h0 in the fiber over x0 . Since x(T) = x0 , the terminal point g(T) = g0 h0 h for some h ∈ H. The correspondence x(t) → h is one to one. In this correspondence, h−1 corresponds to the lift g(T − t)h−1 of the reversed curve x(T − t). Moreover, if x1 (t), t ∈ [0, T1 ] and x2 (t), t ∈ [0, T2 ] are any two loops, their concatenation x(t) = x1 (t), t ∈ [0, T1 ], x((t) = x2 (t − T1 ), t ∈ [T1 , T1 + T2 ] is a loop at x0 . If g1 (t) and g2 (t) denote the horizontal lifts of x1 (t) and x2 (t) then g1 (T1 ) = g0 h0 h1 and g2 (T2 ) = g0 h0 h2 , for some h1 and h2 in H. It follows that g(t) = g1 (t), t ∈ [0, T1 ], g(t) = g2 (t − T1 )(g0 h0 )−1 g1 (T1 ) is the horizontal lift of x(t) to g0 h0 that satisfies g(T2 + T1 ) = g0 h0 h1 (g0 h0 )−1 (g0 ho )h1 = g0 h0 h2 h1 . If g(t) is a horizontal curve that originates at g0 h0 , then g(t) = g0 h0 g0 (t), where g0 (t) is a horizontal curve that originates at e for t = 0, since P is leftinvariant. Then g(T) = g0 h0 g0 (T) = g0 h0 h reduces to h = g0 (T). Hence, h ∈ K ∩ H. Conversely, any point h ∈ K ∩ H corresponds to the terminal point g0 (T) of some horizontal curve that satisfies g0 (0) = e. Then g(t) = g0 h0 g0 (t) is the horizontal curve at g0 h0 whose projected loop corresponds to h. The previous discussion shows that there is no loss in generality in assuming that a symmetric Riemannian pair (G, H) satisfies the condition that any pair of points in G can be connected by a horizontal curve. In such a situation H is equal to the holonomy group of the connection P. Definition 8.27 A symmetric Riemannian pair (G, H) satisfies the controllability condition if any two points in G can be connected by a horizontal curve. Proposition 8.28 Suppose that (G, H) is a Riemannian symmetric pair such that the Lie algebra h of H has zero center. Then any two points of G can be connected by a horizontal curve if and only if [p, p] = h. Proof Any two points of G can be connected by a horizontal curve if and only if Lie(p) = g (the Orbit theorem). According to the fundamental

8.3 The sub-Riemannian problem

133

decomposition of g (Proposition 7), Lie(p) = g if and only if {B ∈ h : adB(p) ⊆ p0 } = 0. But {B ∈ h : adB(p) ⊆ p0 } = 0 is a necessary and sufficient condition that h = ∪{[pi , pi ] : Kl|pi = λi , |pi , λi = 0}. Corollary 8.29 Suppose that (G, H) is controllable such that h has zero center. If g has also zero center, then g is semi-simple. Proof We will continue with the notations in Proposition 8.10. If (G, H) is controllable, then g0 = p0 . Since g0 is an ideal in g, adh(p0 ) ⊆ p0 . But, [p, p] = h and [p, p0 ] = 0. Therefore, adh(p0 ) = 0 by Jacobi’s identity. Hence, p0 is in the center of g and by our assumption must be zero. But then the Killing form is non-degenerate on g. When the pair (G, H) is Riemannian and controllable, then there is a canonical sub-Riemannian problem on G induced by a positive-definite AdH invariant quadratic form , on p. Any horizontal curve g(t) in G can be assigned the length in an interval [0, T] given by T 1 (U(t), U(t)) 2 dt, (8.4) LT (g(t)) = 0

g−1 (t) dg dt (t).

Assuming that Lie(p) = g, any two points can where U(t) = be connected by a horizontal curve. Therefore the problem of finding a horizontal curve of minimal length that connects two given points in G is well defined. This minimal length is called the sub-Riemannian distance. The subRiemannian problem is left-invariant, so the solutions through any point will be the left-translates of the solution through the group identity e. The sub-Riemannian metric on G induces a Riemannian metric on the base curves in G/H with the Riemannian length of a curve x(t) in M equal to the sub-Riemannian length of a horizontal curve g(t) that projects onto x(t). The length of x(t) is well defined because the sub-Riennian lengths of parallel horizontal curves are all equal, since , is AdH -invariant. That is, . dx dx , = A(t), A(t), dt dt where A(t) = g−1 (t) dg dt defined by a horizontal curve that projects onto x(t). As we have already stated, it suffices to consider curves of minimal length from a fixed initial point x0 , since the curves of minimal length from any other point are the translates by an element of G under the left action. In what follows we will consider the horizontal curves g(t) that satisfy g(0) = e. Then their projections x(t) = g(t)H satisfy x(0) = H. A horizontal lift of a curve x(t) that is of minimal length in an interval [0, T] is also of minimal sub-Riemannian length on [0, T], but the converse may not

134

8 Symmetric spaces and sub-Riemannian problems

be true. To get the converse, we need to consider the totality of horizontal curves that connect H to a given left coset g1 H. Then the projection x(t) of the curve of minimal sub-Riemannian length in this class is of minimal length in M. We shall now recast each of the above variational problems as optimal control problems in G and obtain the extremal curves through the Maximum Principle. For simplicity of exposition we will assume that g has zero center, which in turn implies that g is semi-simple. It will be advantageous to adapt the problem to the decomposition g = g1 ⊕ · · · ⊕ gm with gi = pi + [pi , pi ] explained in Propositions 8.9 and 8.10. Then let (1)

(2)

(m)

(2) (m) A1 , . . . , A(1) m1 , A1 , . . . , Am2 , . . . , A1 , . . . , Amm , m1 + m2 + · · · + mm = n, (k)

(k)

be an orthonormal basis in p such that A1 , . . . , Amk is an orthonormal basis for pk , k = 1, . . . , m. Then horizontal curves g(t) are the solutions of the control problem k dg (k) (t) = g(t) u(k) i (t)Ai dt

m

m

(8.5)

k=1 i=1

(1) (m) with control functions u(t) = u1 (t), . . . , umm (t) measurable and bounded on compact intervals [0, T]. It will be convenient to work with the energy functional & % m mk # (k) $2 1 T 1 T u(t), u(t) dt = ui (t) dt, (8.6) E(g(t)) = 2 0 2 0 k=1 i=1

instead of the length functional. Our variational problems are reformulated as follows: 1. The sub-Riemannian problem. Let T > 0 and g1 in G be fixed. Find a trajectory (ˆg(t), uˆ (t)) of the control system (8.5) in the interval [0, T] that satisfies gˆ (0) = e, gˆ (T) = g1 such that E(ˆg(t)) ≤ E(g(t)) for any trajectory (g(t), u(t)) in [0, T] that satisfies the same boundary conditions g(0) = e and g(T) = g1 . 2. The Riemannian problem. Let T > 0 and g1 in G be fixed. Find a trajectory (ˆg(t), uˆ (t)) of the control system (8.5) in the interval [0, T] that satisfies gˆ (0) ∈ H, gˆ (T) ∈ g1 H such that E(ˆg(t)) ≤ E(g(t))

8.4 Sub-Riemannian and Riemannian geodesics

135

for any trajectory (g(t), u(t)) in [0, T] that satisfies g(0) ∈ H and g(T) ∈ g1 H. The only difference between the two problems is that the terminal points e, g1 in the sub-Riemannian problem are replaced by the terminal manifolds H and g1 H in the Riemannian problem.

8.4 Sub-Riemannian and Riemannian geodesics To take advantage of the left-invariant symmetries, the cotangent bundle T ∗ G will be realized as the product G × g∗ . Recall that a linear function ξ at the tangent space Tg G is identified with (g, ) in G × g∗ via the formula (A) = ξ(Lg )( A) for all A ∈ g. In this representation the Hamiltonians of the leftinvariant vector fields X(g) = gA are linear functions (A) on g∗ . (k) (k) Let h1 , . . . , hmk denote the Hamiltonians of the left-invariant vector fields (k) (k) (k) (k) X1 (g) = gA1 , . . . , Xmk (g) = gAmk , k = 1, . . . , m. According to the Maximum Principle, optimal trajectories (g(t), u(t)) of the preceeding variational problems are the projections of the extremal curves (g(t), (t)) in T ∗ G, which can be of two kinds, normal and abnormal. The fact that our distribution is of contact type, namely that p + [p, p] = g, implies that the abnormal extremals can be ignored, because every optimal trajectory is the projection of a normal extremal curve (the Goh condition [AS, p. 319]). Normal extremals are the integral curves of a single Hamiltonian k 1 (k) 2 hi (). H() = 2

m

m

(8.7)

k=1 i=1

Its integral curves are the solutions of the following differential system: k d dg (k) (k) (t) = g(t)dH((t)), (t) = −(ad∗ dH((t))((t)), dH = hi Ai . dt dt

m

m

k=1 i=1

(8.8) Sub-Riemannian geodesics are the projections on G of the extremal curves on energy level H = 12 . The projections to G/H of the sub-Riemannian geodesics that satisfy the transversality condition (T)(h) = 0 are called Riemannian geodesics.

136

8 Symmetric spaces and sub-Riemannian problems

Proposition 8.30 the form

Each sub-Riemannian geodesic g(t), g(0) = e, is of g(t) = e(P−Q)t eQt ,

(8.9)

for some matrices P ∈ p and Q ∈ h with P, P = 1. Each Riemannian geodesic that originates at π(e) is the projection of a curve g(t) = ePt for some matrix P ∈ p with P, P = 1. Proof Let g∗ = g∗1 ⊕· · ·⊕g∗m , where each g∗k is identified with the annihilator of the complementary space g⊥ k , and then let = 1 ⊕ 2 · · · ⊕ m denote the corresponding decomposition of ∈ g∗ into the factors g∗i . Also, let dH = mk (k) (k) m k=1 dHk , where dHk = i=1 hi Ai , k = 1, . . . , m. Then for any X = X1 + X2 · · · + Xm in g, 4 m 5 m m d (X) = − dHk , X = − [dHk , Xk ] = − k [dHk , Xk ], dt k=1

k=1

k=1

because [gi , gj ] = 0 for i = j. Hence, equation (8.8) decomposes into a system of equation dk = −ad∗ dHk ()(k (t)), k = 1, . . . , mm . dt

(8.10)

Since g is semi-simple, ∈ g∗ can be identified with L ∈ g via the formula (X) = Kl(L, X) for all X ∈ g. If (t) is a solution of (8.8), then L(t) is a solution of dL = [dH, L(t)], dt

(8.11)

as can be verified easily: d dL ,X = (X) = −[dH, X] = Kl([X, dH], L) = Kl(X, [dH, L]). Kl dt dt Since X is arbitrary, equation (8.11) follows. Let L1 , . . . , Lm denote the projections of L ∈ g onto the factors g, . . . , gm . Since the factors g1 , g2 , . . . , gm are orthogonal relative to the Killing form, each element k ∈ g∗k corresponds to Lk , in the sense that k (Xk ) = Kl(Lk , Xk ), Xk ∈ gk .

(8.12)

Therefore, equation (8.11) breaks up into the invariant factors dLk = [dHk , Lk ], k = 1, . . . , m. dt

(8.13)

8.4 Sub-Riemannian and Riemannian geodesics

137

Let now Lk = Pk + Qk with Pk ∈ pk and Qk in hk = [pk , pk ]. Recall that pk and hk are orthogonal relative to the Killing form. Then, dPk dQk = [dHk , Qk ], = [dHk , Pk ]. (8.14) dt dt On pk the Killing form is a constant multiple of , , that is, Kl(A, B) = mk (k) (k) li Ai , then λk A, B. If we now write Pk = i=1 1 1 (k) 1 (k) (k) (k) (k) = lk Ai = hi (). li = Pk , Ai = Kl Pk , Ai λk λk λk Hence, Pk = λ1k dHk , k = 1, . . . , m, and therefore [dHk , Pk ] = 0. It follows that the solutions of (8.14) are given by dQk dPk = λk [Pk , Qk ], = 0. (8.15) dt dt Equation (8.15) is readily solvable: Qk is a constant matrix, and Pk (t) = e−λk Qk t) Pk (0)eλk Qk t . Then dHk = e−λk Qk t λk Pk (0)eλk Qk t , and dH = e−Qt PeQt , where Q = λ1 Q1 ⊕ · · · ⊕ λm Qm , P = λ1 Pi (0) ⊕ · · · ⊕ λm Pm (0). The sub-Riemannian geodesics are the solutions of dg = g(t)(dH(t)) = g(t) e−Qt PeQt . (8.16) dt # −Qt Qt $ 0 Let h(t) = e−Qt and g0 = gh. Then, dg Pe h−g0 hQ = g0 (P−Q). dt = g e Hence, g(t) = e(P−Q)t eQt . If (t) is to satisfy the transversality condition (T)(h) = 0), then Qk (T) = 0 for each k. Since Qk (t) is constant, Qk = 0 and g(t) = ePt . The sub-Riemannian sphere of radius r centered at the identity consists of all points in G which are a distance r away from e. The set WT (e) = {g : g = e(P−Q)T eQT , P ∈ p, Q ∈ h}, P, P = 1 is called the wave front at T or the exponential mapping at T. These concepts are natural extensions of their counterparts in Riemannian geometry. In particular, in this situation, the exponential mapping at π(e) is given by P → ePT (mod H), P, P = 1, and the exponential mapping at any other point x = g (mod H) is equal to the left-translate by g of the exponential mapping at x = I (mod H). It follows from the above proposition that each Riemannian symmetric space M = G/H is complete, in the sense that the exponential map through any point x is defined for all times T. For small T, the

138

8 Symmetric spaces and sub-Riemannian problems

exponential mapping is surjective and coincides with the Riemannian sphere of radius T. That is not the case for sub-Riemannian problems: each wave front WT (e) contains points whose distance from e is greater than T, as we have already seen in the preceding chapter. Definition 8.31 A diffeomorphism on a Riemannian manifold M with its metric , is called an isometry if ∗ v1 , ∗ v2 (x) = v1 , v2 x for all x in M and all tangent vectors v1 and v2 in Tx M. Each left translation by an element in G and each right translation by an element in H is an isometry for the symmetric Riemannian space G/H, and so is the restriction of the involutive automorphism σ . Therefore, Sp defined by Sp (gH) = (g0 (σ (g0 ))−1 σ (g)H is an isometry for each point p = g0 H. It follows that Sp (p) = g0 H = p. Let γ (t) = g0 (exp tA)H be any geodesic from p. An easy calculation shows that σ (exp tA) = exp tσ∗ (A) = exp −tA, and therefore, Sp (γ (t)) = g0 (σ (g0 ))−1 σ (g0 exp tA)H) = g0 σ (exp −tA)σ (H) = Sp (γ (−t)). (8.17) The symmetry Sp is called the geodesic symmetry at p. Every Riemannian manifold that admit a geodesic symmetry at each of its points is called a symmetric space [Eb; Hl]. The literature on symmetric spaces begins with the geodesic symmetries as the point of departure and then arrives at the symmetric Riemannian pairs (G, H) via the isometry arguments ([Eb] or [Hl]). We have shown that one can take the opposite path, and start with the symmetric pair (G, H) and finish with the geodesic symmetry at the end.

8.5 Jacobi curves and the curvature Consider now the second-order conditions associated with the above geodesic problems. Any parametrized curve x(t) in G/H is the projection π(g) of a A(t) = g−1 (t) dg horizontal curve g(t) such that dx dt = π∗ (g(t)A(t)),where dt . 2 dx dx If dt denotes the Riemannian norm of dx dt , then dt = A(t), A(t). To every curve of tangent vectors v(t) defined along a curve x(t) there corresponds a curve B(t) in p such that X(t) = π∗ (g(t)B(t)). Definition 8.32 The covariant derivative Dx (v(t)) of a curve of tangent vectors v(t) along x(t) is given by π∗ (g(t) dB dt (t)).

8.5 Jacobi curves and the curvature

139

The geodesic curvature κ(t) along along a curve x(t) parametrized by arc dx length is equal to κ(t) = Dx(t) dt . If g(t) is a horizontal curve that projects onto a curve x(t) parametrized by arc length then ||A(t)|| = 1, dg −1 where A(t) is a curve in p defined by g (t) dt = A(t). It then follows that κ(t) = dA dt . Proposition 8.33 The projections of sub-Riemannian geodesics on G/H have constant geodesic curvature. Riemannian geodesics are the projections of subRiemannian geodesics having zero curvature. Proof The sub-Riemannian geodesics are the horizontal curves that are the −Qt PeQt for some constant matrices solutions of dg dt = g(t)A(t), where A(t) = e P ∈ p with ||P|| = 1 and Q ∈ h. the geodesic curvature of the projected Then dA curve in G/H is given by κ = dt = ||[Q, P]||. The Riemannian geodesics are the projections of sub-Riemannian geodesics with Q = 0. ˆ

With each geodesic curve g(t) = etA consider now the variational curve d −tAˆ t(A+A) ˆ e e vAˆ (t)(A) = π∗ |=0 , vAˆ (0)(A) = 0, (8.18) d ˆ

ˆ

generated by a fixed direction A in p. Let g(t, ) = e−tA et(A+A) and x(t, ) = π(g(t, )). Since g(t, 0) = e, ∂g ∂ |=0 is a curve in g. The following proposition is fundamental. Proposition 8.34

Let JAˆ (t)(A) denote the projection of JAˆ (t)(A) =

∞ k=0

∂g ∂ |=0

on p. Then,

t2k+1 ˆ ad2k A(A). (2k + 1)!

Proof ∂g ˆ ˆ (A+A)t ˆ ˆ ˆ (t, ) = −e−At Ae + e−At (Aˆ + A)e(A+A) ∂t ˆ

ˆ

= g(t, ) e−t(A+A) Aet(A+A) .

ˆ

ˆ

Then, e−t(A+A) Aet(A+A) ) = (P(t, )(A) + Q(t, )(A)), where P(t, ) =

∞ ∞ t2k t2k+1 ad2k (Aˆ + A), Q(t, ) = ad2k+1 (Aˆ + A). (2k)! (2k + 1)! k=0

k=0

Since ad2k (Aˆ + A)(A) ∈ p and ad2k+1 (Aˆ + A)(A) ∈ h, it follows that P(t, )(A) belongs to p and Q(t, )(A) belongs to h.

140

8 Symmetric spaces and sub-Riemannian problems

ˆ ), where h(t, ˆ ) denotes the field of solution curves Let gˆ (t, ) = g(t, )h(t, ∂ hˆ ˆ ˆ 0) = I. It then follows that in H defined by ∂t (t, ) = −h(t, )Q(t, ), h(t, ∂ ˆ )). gˆ (t, ) = gˆ (t, )( hˆ−1 (t, )P(t, )(A)h(t, ∂t Since π(g(t, )) = π(ˆg(t, )), gˆ is a horizontal field of curves that projects onto ˆ )) and x(t, ). Hence, ∂x g(t, )( hˆ−1 (t, )P(t, )h(t, ∂t = π∗ (ˆ ∂ ˆ−1 D ∂x ˆ (t, ) = π∗ gˆ (t, ) h P(t, )(A)h . ∂ ∂t ∂ = g(t, )((V(t, ) + W(t, )) with V(t, ) ∈ p and ¯ ) with h(t, ¯ ) the solution of ∂ h¯ (t, ) = W(t, ) ∈ h. If g¯ (t, ) = g(t, )h(t, ∂ ¯ )W(t, ), h(t, ¯ 0) = I then, −h(t, Let now

∂ ∂ g(t, )

∂ g¯ ¯ = g¯ (t, )(h¯ −1 V h). ∂ It follows that g¯ (t, ) is a horizontal field of curves over the field of curves ∂x ¯ Then, (t, ) = π∗ (¯g(t, )(h¯ −1 V h)). x(t, ) in G/H, and therefore, ∂ ∂ ¯ −1 D ∂x ˆ (t, ) = π∗ g¯ (t, ) (h (t, )V(t, )h)(t, ) . ∂t ∂ ∂t # $ # $ D ∂x D ∂x Since ∂t ∂ (t, ) = ∂ ∂t (t, ), ∂ ¯ −1 ∂ ˆ−1 ˆ ˆ h (t, )V(t, )h(t, ) = π∗ gˆ (t, ) h P(t, )(A)h . π∗ g¯ (t, ) ∂t ∂ dJ At = 0, the left-hand side of this equation is equal to π∗ dtAˆ (t)(A) , while the right-hand side is equal to π∗ (P(t, 0)(A)). But then, ∞ dJAˆ t2k ˆ (t)(A) = P(t, 0)(A) = ad2k A(A), dt (2k)! k=0

hence, JAˆ (t) = Corollary 8.35

∞

t2k+1

k=0 (2k+1)! ad

2k A. ˆ

JAˆ (t) is the solution of the equation $ # d2 JAˆ (t) = ad2 Aˆ JAˆ (t) , JAˆ (0) = 0. 2 dt

(8.19)

Equation (8.19) is called Jacobi’s equation, and its solutions are called Jacobi’s curves.

8.5 Jacobi curves and the curvature

141

Corollary 8.36 1 ˆ JAˆ (t)(A)2 = t2 A2 + t4 A, ad2 A(A) + R(t), 3 where the remainder term R(t) satisfies limt→0

R(t) t4

(8.20)

= 0.

ˆ in the Taylor series expansion of JAˆ (t)(A) is The second term A, ad2 A(A) the negative of the Riemannian curvature of the underlying manifold G/H, a fact true on any Riemannian manifold M [DC]. Let us now consider the curvature in its own right and investigate the implications on the structure of the symmetric space. Definition 8.37 The bilinear form κ : p × p → R defined by κ(A, B) = [[A, B], A], B for each A and B in p is called the Riemannian curvature of the Riemannian space G/H. It follows that κ(A, B) = −ad2 A(B), B and that κ(A, A) = 0. The fact that adA(h), B = −A, adB(h) easily implies that κ is symmetric, in the sense that κ(A, B) = κ(B, A). Definition 8.38 The sectional curvature of a two-dimensional linear subspace P in p is equal to κ(A, B), where A and B are any orthonormal vectors in P. We leave it to the reader to show that the sectional curvature is well defined, in the sense that it is independent of the choice of a basis for P. Definition 8.39 A symmetric Riemannian space G/H is said to be of constant curvature if the sectional curvatures of any pair of two-dimensional linear subspaces of p are equal. Proposition 8.40 Let G/H be a symmetric Riemannian space and let KA denote the restriction of ad2 A to p for A ∈ p, ||A|| = 1. Then, 1. G/H is of zero constant curvature if and only if KA = 0 for each A. This happens if and only if p is a commutative Lie subalgebra of g. 2. G/H is a space of non-zero constant curvature if and only if the restriction of KA to the orthogonal complement of A is a scalar multiple of the identity independent of A. Proof We will first show that KA is a symmetric linear operator on p. Note that for any X, Y, Z, W in p [[X, Y], Z], W = −[Z, [X, Y]], W = −Z, [[X, Y], W] = −[[X, Y], W], Z,

142

8 Symmetric spaces and sub-Riemannian problems

from which it follows that [[X, Y], Z], Z = 0.

(8.21)

Then for arbitrary elements X, Y of p, KA (X), Y = [A, [A, X]], Y = A, [[A, X], Y] = A, [[A, Y], X] + A, [[Y, X], A] = [A, [A, Y]], X + [[X, Y]], A], A = ad2 A(Y), X = KA (Y), X. Let now E1 = A, and lef E2 , . . . , En be an orthonormal basis in the orthogonal complement of A in p with respect to which KA is diagonal, that is, KA (E1 ) = 0, and KA (Ej ) = λj Ej , j = 2, . . . , n. Then the sectional curvature of the plane spanned by E1 , Ej is given by λj . The space G/H is of constant curvature if and only if λ2 = · · · = λn = λ, that is, if and only if KA = λI on the orthogonal complement to A. The curvature is zero if and only if λ = 0 for each A. In such a case p must be a commutative algebra because the Killing form is negative-definite on h. In fact, Kl([A, B], [A, B]) = Kl([[A, B], A], B) = −Kl ad2 A(B), B = 0 for any A, B in p, Therefore, [A, B] = 0.

8.6 Spaces of constant curvature Proposition 8.41 Suppose that (G, H) denotes a Riemannian symmetric pair such that G is connected and the Lie algebra h of H has zero center. Then the quotient space G/H has zero curvature if and only if G is isomorphic to En K, where K is a closed subgroup of the rotation group O(En ). Proof We have already seen that the semi-direct products G = En K with K a closed subgroup of the rotation group O(En ) realizes En as a symmetric space G/({I} × K). In such a situation, p = En × {0} and h = {0} × k, where k denotes the Lie algebra of K. The Lie brackets in g are of the form [(A1 , B1 ), (A2 , B2 )] = ([B2 A1 ] − [B1 A2 ], [B1 , B2 ]) for all (A1 , B1 ) and (A2 , B2 ) in g. Evidently the Lie brackets of elements in p are zero, and hence G/H has curvature equal to zero.

8.6 Spaces of constant curvature

143

Conversely, suppose that (G, H) denotes a Riemannian symmetric pair such that the Lie algebra h of H has zero center, and suppose that G/H is a space of zero curvature. It follows from Proposition 8.34 that p is a commutative subalgebra of g, hence a Euclidean space En . The mapping B → ad(B) is a Lie algebra representation of h into the Lie algebra of skew-symmetric matrices son (R). Let k = {ad(B) : B ∈ h}. The mapping φ : g → p×k defined by φ(A+ B) = (A, −ad(B)), A ∈ p, B ∈ h is a Lie algebra isomorphism onto the semidirect product En k. It is easy to see that k is the Lie algebra of AdH , and that AdH is a subgroup of O(p) since it leaves the Euclidean norm on p invariant. ˆ denote the semi-direct product En AdH . It remains to show that Let G ˆ Let F denote the family of right-invariant vector fields G is isomorphic to G. X(g) = gA such that A ∈ p. If P denotes the orbit of F through the identity of G, then P is an Abelian subgroup of G. It can be proved that g ∈ G can be written as g = ph for some p ∈ P and h ∈ H, as a consequence of the fact that ˆ defined by φ(ph) = (p, Adh ) is the desired ad(h)(p) = p. Then, φ : G → G isomorphism. Consider now the spaces of constant non-zero curvature. There is no loss in generality in assuming that λ = ±1, since this can be accomplished by rescaling the metric on p. We will show that there are only two simply connected cases: the hyperboloid Hn = SO(1, n)/{1} × SOn (R) when λ = −1, and the sphere Sn = SOn+1 (R)/SOn (R) when λ = 1. We will consider both cases simultaneously in terms of a parameter . Let SO = SOn+1 (R) when = 1 and SO = SO(1, n) when = −1. In addition, let S = Sn when = 1 and S = Hn when = −1. So S = G /H, where H = {1} × SOn (R). Let us first show that S is a space of constant curvature with the metric induced by A, B = − 2 Tr(AB) on the Lie algebra so of SO . As we have 0 −aT , seen earlier in this chapter, so consists of matrices M = a A where a is a column vector in Rn , aT is the corresponding row vector, and A a skew-symmetric n × n matrix. The Cartan space p consists of all matrices T 0 −a , where a is a column vector in Rn and 0n the n × n matrix A= a 0n with zero entries. With each such matrix A, Aˆ will denote the vector a. In this notation, A, B = Aˆ · Bˆ where Aˆ · Bˆ denotes the Euclidean product in Rn . The Lie algebra h of H = {1} × SOn (R) consists of matrices 0 ⊕ B = 0 0T with B in son (R). The reader can easily verify the following 0 B formula:

144

8 Symmetric spaces and sub-Riemannian problems ˆ [A1 , A2 ] = 0 ⊕ (Aˆ 1 ∧ Aˆ 2 ), and [A, 0 ⊕ B] = C if and only if Cˆ = BA. (8.22)

for any A1 and A2 in p and any B in son (R). Then, −ad2 A(X) = X for ˆ = 1, and any X ∈ p . Therefore, S has constant curvature any A ∈ p , ||A|| equal to . Let us now realize S as the orbit of SO through e0 , the column vector in n+1 with the first coordinate equal to 1 and all other coordinates equal to R = g0 e0 in S are given by zero. The geodesic curves x(t) through a point xT0 0 −a with ||a|| = 1. An easy the formula x(t) = g0 etA e0 for A = a 0n calculation yields a2 0 2 A = , and A3 = −A. (8.23) 0 a ⊗ aT Therefore,

& % ∞ k t4 t6 t k t2 A e0 = e0 1 − + − + · · · e e0 = k! 2! 4! 6! k=0 % & t3 t5 t7 + a t − + − + ··· , 3! 5! 7! tA

with the understanding that a is now embedded in Rn+1 with its first coordinate equal to zero. It follows that the geodesic curves which originate at e0 in Hn are given by x(t) = e0 cosh t + a sinh t, while the geodesic curves which originate at e0 in Sn are of the form x(t) = e0 cos t + a sin t. Evidently, these geodesics are great circles on the sphere, and “great” hyperbolas on the hyperboloid. Proposition 8.42 Suppose that G/H is a symmetric space with curvature = ±1. Then the Lie algebra g of G is isomorphic to so . If φ denotes this isomorphism, and if , denotes the sub-Riemannian metric on p then A1 , A2 = (φ(A1 ), φ(A2 ) = − 2 Tr(φ(A1 )φ(A2 )) for any A1 and A2 in p . Proof For simplicity of notation we will omit the dependence on the sign of the curvature, except in the situations where it really matters. The lemma below contains the essential ingredients required for the proof. Lemma 8.43

Let A1 , A2 , . . . , An denote an orthonormal basis in p. Then,

1. [Ai , [Aj , Ak ]] = 0 for distinct indices i, j, k. 2. {[Ai , Aj ], 1 ≤ j < i ≤ n} is a basis for h.

8.6 Spaces of constant curvature

145

Proof It follows from (8.21) that [Ai , [Aj , Ak ]] is orthogonal to Ai . Since Ai , Aj , Ak are mutually orthogonal, ad2 (Ak )(Ai ) = −Ai and ad2 (Ak )(Aj ) = −Aj . Then, [Ai , [Aj , Ak ]], Ak = [Ai , [[Aj , Ak ], Ak ] = −Ai , Ak = 0.

(8.24)

Therefore, [Ai , [Aj , Ak ]] is orthogonal to Ak . An identical argument shows that [Ai , [Aj , Ak ]] is orthogonal to Aj . Additionally, ad2 Ai ([Aj , [Ak , Ai ]]) = [Ai , [Ai , [Aj , [Ak , Ai ]]]] = −[Ai , [[Ak , Ai ], [Ai , Aj ]]] − [Ai , [Aj , [[Ak , Ai ], Ai ]]] = [[Ai , Aj ], [Ai , [Ak , Ai ]] + [[Ak , Ai ], [[Ai , Aj ], Ai ]] − [Ai , [Aj , [Ak , Ai ], Ai ]]] = [[Ai , Aj ], Ak ] + [[Ak , Ai ], Aj ] + [Ai , [Aj , Ak ]] = 2[Ai , [Aj , Ak ]].

(8.25)

This yields ad2 Ai ([Ai , [Aj , Ak ]]) = −2ad2 Ai ([Aj , [Ak , Ai ]]) = −4[Ai , [Aj , Ak ]]. After permuting indices in (8.25), first with respect to i and j, and then in respect to j and k, we get ad2 (Aj ([Ai , [Ak , Aj ]]) = 2[Aj , [Ai , Ak ]], ad2 Ak ([Ai , [Aj , Ak ]]) = 2[Ak , [Ai , Aj ]].

(8.26)

These relations imply ad2 Aj ([Ai , [Aj , Ak ]]) = −2[Aj , [Ai , Ak ]] and ad2 Ak ([Ai [Aj , Ak ]) = 2[Ak , [Ai , Aj ]].

(8.27)

Since the curvature is constant, ad2 Aj ([Ai , [Aj , Ak ]]) = −4[Ai , [Aj , Ak ]], ad2 Ak ([Ai , [Aj , Ak ]]) = −4[Ai , [Aj , Ak ]]. Equations (8.27) and (8.28) yield 2[Ai [Aj , Ak ]] = [Aj , [Ai , Ak ]] = [Ak , [Aj , Ai ]], which further implies that 4[Ai , [Aj , Ak ]] = [Aj , [Ai , Ak ]] + [Ak , [Aj , Ai ]] = [Ai , [Aj , Ak ]], and therefore, [Ai , [Aj , Ak ]] = 0. To show the second part, assume that ni>j aij [Ai , Aj ] = 0. Then, ⎛ ⎞ n 0= aij [Ak , [Ai , Aj ]] = ⎝ −akj Aj + aik Ai ⎠ , i>j

k>j

i>k

(8.28)

146

8 Symmetric spaces and sub-Riemannian problems

by the previous part. Therefore, akj = 0, k > j and aik = 0, i > k for any k = 1, . . . , n. But, this implies that Aij = 0, for all i, j. The fact that [p, p] = h implies that {[Ai , Aj ], i > j} is a basis for h. To show that g is isomorphic to so we will introduce the coordinates relative to the basis A1 , . . . , An , [Ai , Aj ], i > j in g . Any A ∈ p can be written as A = ni=1 ai Ai , and any B ∈ h can be written as B = ni>j aij [Ai , Aj ]. These 0 −aT coordinates define a matrix φ(A + B) = , A = (aij ) in so . a A The correspondence A + B → φ(A + B) is one to one and onto so . We leave it to the reader to verify that φ is a Lie algebra automorphism. 0 −aT is isomorphic to p . The Cartan space consisting of matrices a 0n Evidently, A1 , A2 = 2 Tr(φ(A1 ), φ(A2 )) for any elements A1 and A2 in p . Therefore, (G , H ) and (SO {1}, SOn (R)) are isometric.

9 Affine-quadratic problem

Let us now return to the symmetric Riemannian pairs (G, K) and the Cartan decompositions g = p⊕k. In the previous chapter we investigated the relevance of the left-invariant distributions with values in p for the structure of G and the associated quotient space G/H. In particular, we showed that the controllability assumption singled out Lie group pairs (G, K) in which the Lie algebraic conditions of Cartan took the strong form, namely, [p, p] = k, [p, k] = p.

(9.1)

In this chapter we will consider complementary variational problems on G defined by a positive-definite quadratic form Q(u, v) in k and an element A ∈ p. More precisely, we will consider the left-invariant affine distributions D(g) = {g(A + X) : X ∈ k} defined by an element A ∈ p. Each affine distribution D defines a natural control problem in G, dg = g(t)(A + u(t)), dt

(9.2)

with control functions u(t) taking values in k. We will be interested in the conditions on A that guarantee that any two points of G can be connected by a solution of (9.1), and secondly, we will be interested in the solutions of (9.1) which transfer an initial point g0 to a T given terminal point g1 for which the energy functional 12 0 Q(u(t), u(t)) dt is minimal. We will refer to this problem as the affine-quadratic problem. In what follows, we will use u, v to denote the negative of the Killing form Kl(u, v) = Tr(ad(u) ◦ ad(v)) for any u and v in g. Since the Killing form is negative-definite on k, the restriction of , to k is positive-definite, and can be used to define a bi-invariant metric on k. This metric will be used as a bench mark for the affine-quadratic problems. For that reason we will express the quadratic form Q(u, v) as Q(u), v for some self-adjoint linear mapping 147

148

9 Affine-quadratic problem

Q on k which satisfies Q(u), u > 0 for all u = 0 in k. Then, Q = I yields the negative of the Killing form, i.e., the bi-invariant metric on k. The transition from the quadratic form to the linear mapping can be justified formally as follows: any non-degenerate and symmetric quadratic form on k ˜ ˜ ˜ : k → k∗ defined by Q(u)(v) = (Q(u), v) = Q(u, v), v ∈ k. induces a mapping Q Here, (, v) denotes the natural pairing (v) between ∈ g∗ and v ∈ g. Then ˜ the mapping Q is defined by Q(u), v = (Q(u), v) for all v ∈ k. Let us also note that any non-degenerate, symmetric form Q on k has a dual form Q∗ on k∗ defined by ˜ −1 (k1 )), k1 ∈ k∗ , k2 ∈ k∗ . Q∗ (k1 , k2 ) = (k2 , Q

(9.3)

˜ ˜ and k2 = Q(v). It is easy to verify that Q∗ (k1 , k2 ) = Q(u, v) when k1 = Q(u) ∗ −1 In terms of the mapping Q, Q (k1 , k2 ) = L1 , Q (v), after the identification of k1 ∈ k∗ with L1 ∈ k. In particular, Q∗ (k, k) = L, Q−1 (L), for all k ∈ k∗ with (k, X) = L, X, X ∈ k. Before going into further details associated with the above optimal problem, let us first comment on the degenerate case A = 0. When A = 0, the reachable set of (9.1) from the group identity is equal to K and the reachable set from any other point g0 is equal to the right coset g0 K. It also follows that any curve in K is a solution of (9.1). Hence the preceding problem essentially reduces to a left-invariant geodesic problem onK of finding the curves of minimal length T√ in K where the length is given by 0 Q(u(t), u(t)) dt. Then the Maximum Principle identifies H(k) =

1 ∗ Q (k, k) 2

as the appropriate Hamiltonian. To recapitulate briefly, each extremal control u(t) must satisfy 1 1 hu(t) (k(t)) = − Q(u(t), u(t)) + (k(t), u(t) ≥ − Q(v, v) + (k(t), v), v ∈ k 2 2 ˜ −1 (k). Therealong each extremal curve k(t) ∈ k∗ . This implies that u(t) = Q 1 ∗ fore, hu(t) = 2 Q (k(t), k(t)) = H. The Hamiltonian equations are then given by dg ˜ −1 (k(t)), dk = −ad∗ dH(k(t))(k(t)) = −ad∗ Q ˜ −1 (k(t))(k(t)), = g(t)Q dt dt or in equivalent form on K × g as dg dL = g(t)Q−1 (L(t)), = [dH(L), L(t)] = [Q−1 (L(t)), L(t)]. dt dt

9 Affine-quadratic problem

149

In the canonical case Q = I the preceding equations reduce to dg dt = dl = 0. In this situation, the geodesics are the left-translates of the g(t)L(t), dt one-parameter groups {etL : t ∈ R} for L ∈ k. Remark 9.1 The above shows that the exponential map is surjective on a compact Lie group K, because the identity can be connected to any other terminal point by a curve of minimal length. But then, the curve of minimal length is the projection of an extremal curve, hence it is geodesic. Let us now return to the general case, under the assumption that G is semisimple and that the strong Cartan conditions (9.1) hold. Condition [p, p] = k implies that {p ∈ p : adk(p) = 0, k ∈ k} = 0. This fact, in turn, implies that AdH admits no fixed non-zero points in p. Additionally, the strong Cartan conditions imply that g = g1 ⊕ g2 · · · ⊕ gm ,

(9.4)

where each factor gi is a simple ideal of the form gi = pi + [pi , pi ] as described by Proposition 8.10 of the previous chapter. Definition 9.2 An element A in p is called regular if {X ∈ p : [A, X] = 0} is an abelian subalgebra in p. It is easy to prove that the projection of a regular element A on each factor gi in (4) is non-zero. Proposition 9.3 The affine system (9.2) is controllable whenever A is an element in p such that its projection Ai on gi is not zero for each i = 1, . . . , m. For the proof we shall borrow some concepts from differential geometry. Let , denote the AdK invariant Euclidean inner product on p and let Sn denote the unit sphere in p. A subset S of Sn is said to be convex if the angle between any two points of S is less than π and if any two points of S can be connected by an arc of a great circle in Sn that is entirely contained in S [Eb, p. 65]. The dimension of a convex subset S of Sn is the largest integer k such that a k-dimensional disk can be smoothly embedded in S (relative to the canonical metric on Sn ). The interior int(S) of S is the union of all smoothly imbedded k-disks smoothly imbedded in S. The boundary S of ∂(S) is equal to S¯ − int(S), where S¯ denotes the topological closure of S. If x1 and x2 are any points of Sn let d(x1 , x2 ) denote the spherical distance between them and let d(x, X) = inf {d(x, y) : y ∈ X} for any closed subset X of Sn . If S is a convex subset of Sn with a non-empty boundary then there exists a unique point s(S) in int(S) that yields the maximum of the function

150

9 Affine-quadratic problem

f (x) = d(x, ∂(S)). This point is called the soul of S. Finally, if φ is an isometry of Sn that leaves S invariant, then s(S) is a fixed point of φ, i.e., φ(s(S)) = s(S) [Ch; Eb]. With these notions at our disposal we turn to the proof of the proposition. Proof We will first show that the Lie algebra generated by = {A+X; X ∈ h} is equal to g. There is no loss in generality in assuming that g is simple to begin with, since the argument could be reduced to each simple factor by showing that the Lie algebra generated by the projection of on each factor gi is equal to gi . Evidently h ∈ Lie(). Therefore, it suffices to show that p ⊂ Lie(). Let V *∞ * j j denote the linear subspace spanned by ∞ j=0 ad k(A) = j=0 ad B(A) : B ∈ k}, and let V ⊥ denote its orthogonal complement in p relative to the Killing form. Both V and V ⊥ are ad(k) invariant. If X ∈ V and Y ∈ V ⊥ then Kl(k, [X, Y]) = Kl([k, X], Y) = 0 because [k, X] ⊆ V. Hence [X, Y] = 0. It follows that [V, V ⊥ ] = 0. Therefore, V + [V, V] is an ideal in g and a subset of Lie(). But then it must be equal to g since V = 0. The rest of the proof uses the notion of the Lie saturate LS() introduced in the previous chapters. Let C denote the closure of the positive, convex hull generated by all elements of the form {AdXk (A) : X ∈ k, k = 0, 1, . . . }. Then C is an AdK invariant subset of LS(). The fact that Lie() = g implies that C has a non-empty interior in g. Suppose that C is not equal to p. Then there exists a vector v ∈ p such that v, int(C) > 0, i.e., C lies on one side of the hyperplane v, x = 0. Let Sn denote the unit sphere in p relative to the AdK invariant metric , and let S be equal to Sn ∩ int(C). Then S is a convex subset of Sn with a non-empty interior in Sn that is invariant under AdK . If s(S) denotes the soul of S then it follows from above that s(S) is a fixed point of AdK , which is not possible since AdK acts irreducibly on p.1 To each affine problem there is a “shadow affine problem” on the semi-direct product Gs = p H because the Lie algebra g, as a vector space, admits two Lie bracket structures: the original in g and the other induced by the semidirect product. To be more explicit, note that the Lie algebra gs of Gs consists of pairs (A, B) in p × h with the Lie bracket (A1 , B1 ), (A2 , B2 )]s = (adB1 (A2 ) − adB2 (A1 ), [B1 , B2 ]). 1 I am grateful to P. Eberlein for suggesting a proof of Proposition 9.4 based on the concept of the

soul of a convex subset of the sphere.

9 Affine-quadratic problem

151

If (A, B) in p h is identified with A + B in p + k then (A1 + B1 ), (A2 + B2 )]s = [B1 , A2 ] − [B2 , A1 ] + [B1 , B2 ]. Hence, [p, p]s = 0, [p, h]s = [p, h], [h, h]s = [h, h]. Thus vector space g is the underlying vector space for both Lie algebras g and gs . Proposition 9.4 The shadow system dg dt = g(t)(A + U(t)), U(t) ∈ h is controllable in Gs whenever AdH acts irreducibly of p. Proof Since AdH acts irreducibly on p, {Adh (A) : h ∈ H} has a non-empty interior in p. Therefore, p ⊂ Lie(), where = {A + X : X ∈ h} and consequently, Lie() = gs . For the rest of the proof we can either mimic the proof in Proposition 9.3, or use the result in [BJ], which says that the semigroup S generated by {etX : t ≥ 0, X ∈ } is equal to Gs whenever Lie() = gs . Proposition 9.5 Suppose that system (9.2) is controllable. Then both an affine-quadratic problem on G and its shadow problem on p K admit optimal solutions for each pair of boundary points g(0) = g0 and g(T) = g1 , that is, for each pair of points g0 and g1 there exists an interval [0, T] and a control u(t) in L2 ([0, T]) that generates a trajectory g(t) of minimal energy T 1 2 0 Q(u(t), u(t) dt among all other trajectories that satisfy g(0) = g0 and g(T) = g1 . Proof Let T (g0 , g1 , T) denote the set of trajectories g(t) of (9.2) generated by the controls in L∞ ([0, T]) that satisfy g(0) = g0 and g(T) = g1 . Let T be Tbig enough that T (g0 , g1 , T) is not empty. Let∞α denote the infimum of 0 Q(u(t)), u(t) dt over all controls u(t) in L ([0, T]) that generate trajectories in T (g0 , g1 , T). Let L2 ([0, T]) denote the Hilbert space of measurable curves u(t) in k that T satisfy 0 Q(u(t)), u(t) dt < ∞. Then any control u0 (t) that generate a trajecT tory in T (g0 , g1 , T) defines a closed ball B = {u ∈ L2 ([0, T]) : 0 Q(u(t)), u(t) T dt ≤ 0 Q(u0 (t)), u0 (t) dt} in L2 ([0, T]). T Let {un } denote a sequence in B such that α = lim 0 Q(un (t)), un (t) dt. Since closed balls in a Hilbert space are weakly compact, {un } contains a weakly convergent subsequence. For simplicity of notation, we will assume that {un } itself is weakly convergent. Let u∞ denote the weak limit of {un }. But then the trajectories gn (t) of (9.2) that are generated by the controls un converge uniformly to the trajectory g∞ that is generated by u∞ [Jc, p. 118]. Since the convergence is uniform, g∞ belongs to T (g0 , g1 , T). We will complete the proof by showing that g∞ is the optimal trajecT tory, i.e., we will show that α = 0 Q(u∞ (t)), u∞ (t) dt. To begin with,

152

9 Affine-quadratic problem

weak convergence implies that un (t) dt. Secondly, T Q(u∞ (t)), un (t) dt ≤ 0

T 0

T 0

Q(u∞ (t)), u∞ (t) dt = lim

Q(u∞ (t)), u∞ (t) dt)

12

T

T 0

Q(u∞ (t)),

Q(un (t)), un (t) dt)

12 .

0

T T Together, they imply that 0 Q(u∞ (t)), u∞ (t) dt ≤ ( 0 Q(u∞ (t)), u∞ (t) T 1 1 dt)) 2 α 2 , or that 0 Q(u∞ (t)), u∞ (t) dt ≤ α. Remark 9.6 In general the maximum principle is not valid for L2 controls. Fortunately, that is not the case here. In the class of affine-quadratic systems, optimal trajectories generated by L2 controls do satisfy the Maximum Principle [Sg; Tr].

9.1 Affine-quadratic Hamiltonians Let us now turn to the Maximum Principle for the appropriate Hamiltonians associated with the affine problems. To preserve the left-invariant symmetries, the cotangent bundle T ∗ G will be trivialized by the left-translations and considered as the product G × g∗ . Then g∗ = p∗ ⊕ k∗ , where p∗ is identified with the annihilator k0 = { ∈ g : (X) = 0, X ∈ k} and k∗ with the annihilator p0 = { ∈ g : (X) = 0, X ∈ p}. Furthermore, each ∈ g∗ will be identified with L ∈ g via L , X = (X) for all X ∈ g. Then p∗ is identified with p and k∗ is identified with k. The above implies that = p + k , p ∈ p∗ and k ∈ k∗ is identified with L = Lp + Lk where Lp ∈ p and Lk ∈ k. Since p and k are orthogonal relative the Killing form, Lp and Lk are orthogonal relative to , . Under these identifications the Hamiltonian lift of the cost-extended system 1 dg dx = Q(u), u, = g(t)(A + u(t)), dt 2 dt is given by 1 hu (t)(L) = λ Q(u), u + Lp , A + Lh , u(t), λ = 0, −1. 2 The Maximum Principle then yields the Hamiltonian H=

1 −1 Q (Lk ), Lk + A, Lp 2

(9.5)

9.1 Affine-quadratic Hamiltonians

153

defined by u = Q−1 (Lk ) for λ = −1. The abnormal extremals (λ = 0) are the integral curves of hu(t) (L) = A, Lp + Lk , u(t),

(9.6)

subject to the constraint Lk = 0. Therefore, the normal extremals are the solutions of dg dL = g(A + Q−1 (Lk )), = [Q−1 (Lk ) + A, L]. dt dt

(9.7)

The projection on g then can be written in expanded form as dLp dLk = [Q−1 (Lk ), Lk ] + [A, Lp ], = [Q−1 (Lk ), Lp ] + [A, Lk ]. dt dt

(9.8)

The abnormal extremals are the solutions of dL dg = g(t)(A + u(t)), = [u(t), L(t)]], Lk = 0. dt dt

(9.9)

On the semi-direct product gs = p k, the passage from g∗s to its equivalent representation on gs leads to slightly different equations because the quadratic form , is not invariant relative to the semi-direct bracket. On semi-direct dl (X) = −ad∗ (dh(l)(l(t)(X) corresponds products the Hamiltonian equation dt to dL dt , X = − L, [dh, X]s , for any left-invariant Hamiltonian h. This equa dL k = −Lp , [dhp , Xk ] + [dhk , Xp ] + tion implies that dtp , Xp + dL , X k dt Lk , [dhk , Xk ], or . . dLp dLk , Xp + , Xk = [dhk , Lk ] + [dhp , Lp ], Xk + [dhk , Lp ], Xp . dt dt Therefore, dLh dLp dg = g(t)dh(L), = [dhh , Lh ] + [dhp , Lp ], = [dhh , Lp ]. dt dt dt It follows that both the semi-simple and semi-direct Hamiltonian equations can be amalgamated into a single equation dLh dLp = [dhh , Lh ] + [dhp , Lp ], = [dhh , Lp ] + s[dhp , Lh ], dt dt

(9.10)

with the understanding that the parameter s is equal to 0 in the semi-direct case and equal to 1 in the semi-simple case. Of course, it is also understood that g(t) evolves in the appropriate group according to the equation dg dt = g(dh).

154

9 Affine-quadratic problem

It follows that the integral curves of the affine-quadratic Hamiltonian H are the solutions of the following system of equations: dg = g(t)(A + Q−1 (Lk )), dt (9.11) dLp dLk −1 −1 = [Q (Lk ), Lk ] + [A, Lp ], = [Q (Lk ), Lp ] + [A, Lk ]. dt dt Let us now address the abnormal extremals (equation (9.9). But first, we will need to introduce additional notations: kA will denote the subalgebra of k consisting of X ∈ k such that [A, X] = 0, and KA will denote the subgroup in K generated by {etX : X ∈ kA , t ∈ R}. Then, Proposition 9.7 Suppose that A is a regular element in p. Then abnormal extremal curves (g(t), Lk , Lp ) are the solutions of dg dt = g(t)(A + u(t)) subject to the constraints Lk = 0, [A, Lp ] = 0, [Lp , u(t)] = 0. Consequently, Lp is constant. If Lp is regular, then g(t) = g(0)eAt h(t) , where h(t) denotes the solution of dh dt (t) = −h(t)u(t), h(0) = I. If g(t) is optimal, then h(t) is the curve of minimal length relative to the metric Q(u), u in KA that connects I to h(T). Proof Suppose that (g(t), Lp (t), Lh (t)) is an abnormal extremal curve generated by a control u(t) in k. Then equations (9.9) imply that Lk = 0 and dLp dt = [u, Lp ], [A, Lp ] = 0. This means that Lp (t) belongs to the maximal dL

abelian subalgebra A in p that contains A. Therefore, dtp also belongs to A. If B is an arbitrary element of A, then dLp , B = Kl([u(t), Lp ] , B) = Kl(u(t), [Lp , B]) = 0. Kl dt dL

Since the Killing form is non-degenerate on A, dtp = 0, and therefore Lp (t) is constant. This proves the first part of the proposition. To prove the second part assume that Lp is regular. Then [A, Lp ] = 0 implies that [Lp , [A, u(t)]] = 0. Since Lp is regular and belongs to A, [A, u(t)] also belongs to A. It then follows that [A, u(t)] = 0 by the argument identical to the one used in the preceding paragraph. But then g(t) = g(0)eAt h(t), where h(t) is the solution of dh dt (t) = h(t)u(t) with h(0) = I. Since [A, u(t)] = 0, . If g(t) is to be optimal, then h(t) is a curve in KA of shortest length h(t) ∈ K A T √ Q(u(t), u(t) dt. 0

9.2 Isospectral representations

155

Corollary 9.8 If g(t) is an optimal trajectory that is a projection of an abnormal extremal generated by a control u(t) that satisfies [A, u(t)] = 0, then g(t) is also the projection of a normal extremal curve associated with Q. Proof The projection g(t) of an abnormal extremal curve that projects onto an optimal trajectory is of the form g(t) = g0 eAt h(t) with h(t) the curve of shortest length in KA that satisfies the given boundary conditions. Hence h(t) is the projection of an extremal curve in T ∗ KA = KA × k∗A . That is, there exists a curve Lk (t) in kA such that h(t) satisfies dh dLk = h(t)(Q−1 (Lk ), = [Q−1 (Lk ), Lk ]. dt dt Let Lp (t) denote the solution of Then,

dLp dt

= [Q−1 (Lk ), Lp ] such that [Lp (0), A] = 0.

d ([A, Lp (t)]) = [A, [u(t), Lp ]] = [[A, Lp (t)], u(t)]. dt It follows that [A, Lp (t)] is the solution of a linear equation that is equal to zero at t = 0. Hence [A, Lp (t)] = 0 for all t. But then, g(t), L(t) = Lk (t) + Lp (t) are the solutions of equations (9.8) and (9.11), hence g(t) is the projection of a normal extremal curve. Remark 9.9 The above proposition raises an interesting question: is every optimal trajectory of an arbitrary affine-quadratic problem the projection of a normal extremal curve? It seems that G = SLn (R) is a good testing ground for this question. In this situation there are plenty of abnormal extremal curves but it is not clear exactly how they relate to optimality. My own guess is that every optimal solution is the projection of a normal extremal curve.

9.2 Isospectral representations Let us now return to the normal extremal curves equations (9.8) and (9.11). The projection of these equations on either g or gs is given by dLp dLk = [Q−1 (Lk ), Lk ] + [A, Lp ], = [Q−1 (Lk ), Lp ] + s[A, Lk ], s = 0, 1 dt dt (9.12) Equations (9.12) can be regarded also as the Hamiltonian equation on coadjoint orbits in g (resp. gs ) associated with the Hamiltonian H = 12 Q−1 (Lk ), Lk + A, Lp .

156

9 Affine-quadratic problem

Definition 9.10 An affine-quadratic Hamiltonian will be called isospectral if its Hamiltonian system (9.12) admits a spectral representation of the form dLλ = [Mλ , Lλ ], dt

(9.13)

Mλ = Q−1 (Lk ) − λA, Lλ = Lp − λLk + (λ2 − s)B, for some matrix B that commutes with A. Proposition 9.11 Suppose that L(λ) is a parametrized curve in a Lie algebra g that is a solution of dL dt (λ) = [M(λ), L(λ)] for another curve M(λ) in g. Then each eigenvalue of L(λ) is a constant of motion. Proof

Let g(t) be a solution of

dg dt

= g(t)Mλ in G. Then,

d −1 dLλ g (t)Lλ (t)g(t) = g−1 (t)([Lλ , Mλ ] + )g(t) = 0, dt dt along the solutions of constant.

dL dt (λ)

= [M(λ), L(λ)]. Hence, the spectrum of Lλ (t) is

Functions of the eigenvalues of L(λ) are called spectral invariants of L(λ). It follows that each spectral invariant of L(λ) is constant along the solutions of the differential equation for L(λ). In particular, the spectral invariants of Lλ = Lp − λLk + (λ2 − s)B are constant along the solutions of (9.12). Proposition 9.12 The spectral invariants of Lλ = Lp − λLk + (λ2 − 1)B Poisson commute with each other relative to the semi-simple Lie algebra structure, while the spectral invariants of Lλ = Lp − λLk + λ2 B Poisson commute relative to the semi-direct product structure. To address the proof, it will be advantageous to work in g∗ rather than in g. Recall that the splitting g = p ⊕ k induces the splitting of g∗ as the sum of the annihilators p0 ⊕ k0 , in which case p0 is identified with k∗ and k0 is identified with p∗ . The affine Hamiltonian H = 12 Q−1 (Lk ), Lk + A, Lp is given by H(p + k) = Q∗ (k, k) + (p, A) with d = −ad∗ dH()() dt equal to its Hamiltonian equations on each coadjoint orbit in g∗ (respectively, on each coadjoint orbit of g∗s ). Definition 9.13 Functions f and h on g∗ which Poisson commute, i.e., { f , h} = 0, are said to be in involution. Functions on g∗ that are in involution with any other function on g∗ are called invariant or Casimirs.

9.2 Isospectral representations

157

Invariant functions can be defined alternatively as the functions f that satisfy ad∗ df () = 0 for any ∈ g∗ , which is the same as saying that they are constant on each coadjoint orbit. On semi-simple Lie algebras, coadjoint orbits are identified with the adjoint orbits and the above is the same as saying that [df , L] = 0 for each L ∈ g. Hence the spectral invariants of L correspond to invariant functions on g∗ The proof below is a minor adaptation of the arguments presented in [Pr] and [Rm]. Proof Let T : g∗ → g∗ be defined by T(p+k) = λ1 p−k+μb for p ∈ k∗ , k ∈ h∗ . Here, b is a fixed element of p∗ and λ and μ are parameters. Thus T is induced 2 by the spectral matrix Lλ = λ1 Lp − Lk + μB with μ = λ λ−s . It follows that T −1 = λp − k − λμb, and hence, T is a diffeomorphism. This diffeomorphism extends to the Poisson form { , } on g∗ according to the formula { f , g}λ,μ (ξ ) = (T{ f , g})(ξ ) = {T −1 f , T −1 g}(T(ξ )). A simple calculation shows that { f , g}λ,μ = −λ2 { f , g} − λμ{ f , g}b − (1 − λ2 ){ f , g}s ,

(9.14)

where { f , g}b = { f , g}(b) and { f , g}s is the Poisson bracket relative to the semi-direct product Lie bracket structure. For the semi-direct Poisson bracket { , }s the analogous Poisson bracket { f , g}s λ,μ (ξ ) = (T ◦ { f , g}s )(ξ ) takes on a slightly different form: { f , g}sλ,μ = −{ f , g}s − λμ{ f , g}s b .

(9.15)

If f is any invariant function relative to { , }, then fλ,μ = T ◦ f satisfies { fλ,μ , g}λ,μ = 0 for any function g on g∗ and any parameters λ and μ. In the case that g is another invariant function, then fλ,μ and gλ,μ = T ◦ g satisfy fλ1 ,μ1 , gλ2 ,μ2

λ1 ,μ1

= fλ1 ,μ1 , gλ2 ,μ2

λ2 ,μ2

= 0.

for all λ1 , μ1 , λ2 , μ2 . The same applies to the semi-direct Poisson bracket { , }s and its invariant functions. 2 Suppose now that μ = λ λ−1 . It follows from (9.14) that { f , g}λ,μ = −λ2 { f , g} − (1 − λ2 )({ f , g}s − { f , g}b ).

158

9 Affine-quadratic problem

Therefore, 0=

1 fλ1 ,μ1 , gλ2 ,μ2 λ21 − 1

λ1 ,μ1

−

1 fλ1 ,μ1 , gλ2 ,μ2 λ22 − 1

λ2 ,μ2

λ21 − λ22 $ # $ fλ1 ,μ1 , gλ2 ,μ2 . =# 1 − λ21 1 − λ22 Since λ1 , and λ2 are arbitrary fλ1 ,μ1 , gλ2 ,μ2 = 0. This argument proves the first part of the proposition because λT = Lp − λLk + (λ2 − 1)B, after the identifications p → Lp , k → Lk and b → B, 2 and μ = λ λ−1 . To prove the second part let λ = μ. Then λT = Lp − λLk + λ2 B , after the above identifications. Relative to the semi-direct structure, 1 1 f fλ1 ,μ1 , gλ2 ,μ2 , g − λ ,μ λ2 ,μ2 1 1 s λ1 μ1 λ21 λ22 % & 1 1 fλ1 ,μ1 , gλ2 ,μ2 . = − 2 s λ22 λ1

0=

Therefore, fλ1 ,μ1 , gλ2 ,μ2

s

s λ2 ,μ2

= 0.

Evidently the isospectral representation yields a number of constants of motion for the Hamiltonian H all in involution with each other, and this begs the question whether these integrals of motion are sufficient in number to guarantee complete integrability of H. Before addressing this question in some detail, however, let us first address the question of the existence of isospectral representations. In this regard we then have the following: Proposition 9.14 An affine Hamiltonian H = 12 Q−1 (Lk ), Lk + A, Lp is isospectral if and only if [Q−1 (Lk ), B] = [Lk , A] for some matrix B ∈ p that commutes with A. Proof

If [Q−1 (Lk ), B] = [Lk , A], then

dLp dLλ dLk = −λ = [Q−1 (Lk ), Lp ] + s[A, Lk ] − λ([Q−1 (Lk ), Lk ] + [A, Lp ]) dt dt dt = [Q−1 (Lk ), Lp − λLk ] + s[A, Lk ] − λ[A, Lp ] = [Q−1 (Lk ) − λA, Lp − λLk ] − λ2 [A, Lk ] + s[A, Lk ] = [Mλ , Lλ ] − (λ2 − s)[Q−1 (Lk ), B] + (λ2 − s)[Lk , A] = [Mλ , Lλ ].

9.3 Integrability

Conversely, if

dLλ dt

159

= [Mλ , Lλ ], then

dLk = [Q−1 (Lk ), Lk ] + [A, Lp ] − λ(λ2 − s)[A, B], dt dLp = [Q−1 (Lk ), Lp ] + (λ2 − s)([Q−1 (Lk ), B] + λ2 [A, Lk ]). dt Therefore, [A, B] = 0 and [Q−1 (Lk ), B] = [Lk , A]. Corollary 9.15 The canonical case Q = I is isospectral. The spectral matrix Lλ is given by Lλ = Lp − λLk + (λ2 − s)A, i.e., B = A. Proposition 9.16 Let H = 12 Q−1 (Lk ), Lk +A, Lp be an isospectral Hamiltonian. Then every solution of the corresponding homogeneous Hamiltonian −1 k system dL dt = [Q (Lk ), Lk ] is the projection of a solution of the affinequadratic system (9.12). Proof Let B be a matrix that satisfies [Q−1 (Lk ), B] = [Lk , A] with [A, B] = 0 and let Lk (t) denote a solution of the homogeneous system on k. Then L(t) = Lk (t) + Lp with Lp = sB is a solution of (9.12). Corollary 9.17 The projections Lk (t) of the solutions of L = Lp + Lk of an isospectral system (9.12) with Lp = sB are the solutions of the following spectral equation: dLk = [Q−1 (Lk ) − λA, Lk − λB]. dt

(9.16)

Conversely, every Hamiltonian system on k that admits a spectral representation (9.16) gives rise to an isospectral affine Hamiltonian H = 1 −1 2 Q (Lk ), Lk + A, Lp . Corollary 9.18 The spectral invariants of Lk −λB are in involution with each other. Proof Lλ = Lp − λLk + (λ2 − s)B = −λLk + λ2 B = λ(Lk − λB). Hence the spectral invariants of Lk − λB are a subset of the ones generated by the spectral matrix Lλ for the affine system.

9.3 Integrability Our presentation of isospectral representations makes an original contact with the existing theory of integrable systems on Lie algebras inspired by the seminal work of S. V. Manakov on the integrability of an n-dimensional

160

9 Affine-quadratic problem

Euler’s top, a left-invariant quadratic Hamiltonian on the space of n × n skew-symmetric matrices subject to some restrictions (which we will explain in complete detail later on in the text) [Mn]. This study originated an interest in the spectral matrix L − λB. Motivated by this work of Manakov, A. S. Mishchenko and A. T. Fomenko [FM] considered quadratic Hamiltonians H on a complex semi-simple Lie group G that admit a spectral representation dL dt = [dH − λA, L − λB] for some elements A and B in the Lie algebra g of G. More specifically, they considered the set of functions F that are functionally dependent on the shifts f ( + λb), where functions f range over the functions that are constant on the coadjoint orbits in g∗ , and they asked the question, when is a given quadratic Hamiltonian H in F? Here b is a fixed element in g∗ and λ is an arbitrary real number. When H is an element of F, then each element of F is an integral of motion for the Hamiltonian flow of H on the coadjoint orbits in g∗ , since the elements of F are in involution with each other relative to the Poisson bracket { , } on g∗ . In this remarkable paper [FM], the authors showed first that H is in F if and only if there exists a vector A ∈ g such that b ◦ adA = 0 and b ◦ ad dH() = ◦ adA for all ∈ g∗ .

(9.17)

If the dual g∗ is identified g via the Killing form , b, then is identified with L and b is identified with B ∈ g, and the preceding condition is equivalent to A, [B, X] = 0 and B, [dH(L), X] = L, [A, X], X ∈ g, which is the same as [dH, B] = [L, A] and [A, B] = 0. Secondly, they showed that F is an involutive family and that it contains n functionally independent functions where 2n is the dimension of the orbit, i.e., they showed that H is integrable, in the sense of Liouville, on each generic coadjoint orbit. Having in mind applications to an n-dimensional Euler’s top, they applied these results to compact real forms to show that Manakov’s top is a particular case of integrable tops defined by sectional operators. The following case is a prototype of the general situation; not only does it illustrate the importance of this study for the theory of tops, but it also provides a constructive procedure for generating affine-quadratic Hamiltonians that admit a spectral matrix Lλ = Lp − λLk + (λ2 − s)B, for some B.

9.3.1 SLn (C) and its real forms Let g = sln (C) and let h denote the commutative sub-algebra consisting of all diagonal matrices in g. We will adopt the following notations: e1 , . . . , en will denote the standard basis in Rn , and ei ⊗ ej will denote the matrix given by

9.3 Integrability

161

(ei ⊗ ej )ek = δjk ei , k = 1, . . . , n. In addition, , will denote the trace form A, B = − 12 Tr(AB). It follows that {ei ⊗ ej : i = j} is a basis for the matrices in g whose diagonal entries are zero. If A is a matrix in h with its diagonal entries a1 , . . . , an then adA(ei ⊗ ej ) = (aj − ai )ei ⊗ ej . In the language of Lie groups, α : h → C given by α(A) = (aj − ai ) is called a root and the complex line through Xα = ei ⊗ ej is called the root space corresponding to α, that is, Xα satisfies adA(Xα ) = α(A)Xα for each A ∈ h. Then −α is also a root with X−α = ej ⊗ ei . If Hα = [Xα , X−α ], then [Xα , X−α ] = ej ⊗ ej − ei ⊗ ei , and therefore Hα ∈ h. More generally, a Cartan subalgebra of a semi-simple Lie algebra g is a maximal abelian subalgebra h of g such that ad(h) is semi-simple for each h ∈ h. It is known that every semi-simple complex Lie algebra g contains a Cartan subalgebra h and that g=h⊕

gα , α ∈ ,

α

where denotes the set of roots generated by h [Hl]. On sln (C) every matrix with distinct eigenvalues λ1 , . . . , λn is called regular. It is well known that the set of matrices which commute with a regular matrix A is a Cartan subalgebra h. In the basis of eigenvectors of A, h consists of all diagonal matrices having zero trace, and that brings us to the situation above. It follows that {Hα , Xα , X−α , α ∈ } is a basis for sln (C). It is called a Weyl basis. It is known that every semi-simple Lie algebra admits a Weyl basis [Hl]. With this terminology at our disposal let us now return to the work of Mishchenko, Trofimov, and Fomenko, [FM; FT; FT1]. Let g0 denote the real vector space spanned by {iHα , i(Xα + X−α ), (Xα − X−α )}, α ∈ } Evidently, g0 consists of matrices A + iB with A and B matrices having real entries, A skew-symmetric and B symmetric. Therefore, g0 is a Lie algebra equal to sun . It follows that sln (C) = g0 + ig0 , that is, g0 is a real form for g. It is called a compact normal form for g (since SUn is compact). Let us note that g0 = k ⊕ p, with k = son (R), and p = {iX : X ∈ sln (R) : X T = X}, is the Cartan decomposition induced by the automorphism σ (g) = (g−1 )T , g ∈ SUn . Suppose now that A is a regular diagonal matrix in p, that is, suppose that the diagonal entries ia1 , . . . , ian of A are distinct. Then,

162

9 Affine-quadratic problem adA(i(ei ⊗ ej + ej ⊗ ei ) = −(aj − ai )(ei ⊗ ej − ej ⊗ ei ), adA(ei ⊗ ej − ej ⊗ ei ) = (aj − ai )i(ei ⊗ ej + ej ⊗ ei ).

Therefore, adA(i(Xα + X−α ) = iα(A)(Xα − X−α ), adA(Xα − X−α ) = iα(A)i(Xα + X−α ), for α ∈ . It follows that α(B) (i(Xα + X−α ), α(A) α(B) ((Xα − X−α ), ad−1 A ◦ adB((Xα − X−α ) = α(A)

ad−1 A ◦ adB(i(Xα + X−α ) =

for any other diagonal element B in p. Let Q = ad−1 A ◦ adB restricted to k = son (R). If X ∈ k, then X = n i, j xij (ei ∧ ej ) = α xα (Xα − X−α ). Therefore, Q(X) =

α(B) α(A)

α

xα =

bj − bi i, j

aj − ai

xij ,

hence Q is a linear mapping on k. Relative to the trace form, X, Q(X) = n bj −bi 2 i, j aj −ai xij . Moreover, [B, Q−1 (X)] = adB ◦ ad−1 B ◦ adA(X) = adA(X), for all X. Therefore, [Q−1 (X), B] = [X, A], X ∈ k and hence the affinequadratic Hamiltonian H = 12 Q−1 (Lk ), Lk + A, Lp is isospectral. The quadratic form Q(X, X is positive-definite if and only if all the b −b quotients ajj −aii are positive. For this reason, we need to relax the condition that the matrices A and B have zero trace. This change is inessential for the bracket condition [Q−1 (X), B] = [X, A], but it allows for greater flexibility b −b in choosing the ratios ajj −aii . For if A and B are diagonal matrices in un instead of sun that satisfy the Tr(B above bracket condition, then A0 = A − Tr(A) n I and B0 = b − n I are diagonal −1 matrices with zero trace that satisfy [Q (X), B0 ] = [X, A0 ]. There are many choices of A and B in un that produce positive quadratic forms. One such choice, which makes a contact with the top of Manakov, is b −b given by bi = a2i , i = 1, . . . , n. For then, ajj −aii = ai +aj . Another choice, which relates to Jacobi’s geodesic problem on an ellipsoid (as will be shown later on in the text), occurs when A is a matrix with positive entries and B = −A−1 . Then

bj −bi aj −ai

=

− a1 + a1 j

aj −ai

i

=

also take bi = a3i . Then,

1 ai aj .

bj −bi aj −ai

Such a case will be called elliptic. One could

=

1 1 aj − ai

aj −ai

= a2j + ai aj + a2i .

9.3 Integrability

163

Let us conclude this section by saying that sln (R) is another real form for sln (C). When realized as the real vector space spanned by {Hα , (Xα + X−α ), (Xα − X−α , α ∈ }, sln (R) lends itself to the same procedure as in the case of sun . The sectional operators in this setting induce isospectral affinequadratic Hamiltonians under the same conditions as in the previous case, except that the matrices A and B are diagonal with real entries, rather than matrices with imaginary diagonal entries. Both of these two real forms figure prominently in the theory of integrable systems as will be demonstrated in the subsequent chapters. The above is a special case of the following general situation. Proposition 9.19 Suppose that H = 12 Q−1(Lk ), Lk + A, Lp is an affinequadratic Hamiltonian that admits a spectral representation dLλ = [Mλ , Lλ ], Lλ = Lp − Lk + (λ2 − s)B, [A, B] = 0, dt where B is a regular element. Let kA = {X ∈ k : [X, A] = 0}. Then kA is a Lie subalgebra of k that is invariant under Q, and Q restricted to the orthogonal −1 complement k⊥ A is given by Q = ad A ◦ adB. It will be convenient first to prove the following lemma. Lemma 9.20 Proof

{X ∈ k : [X, A] = 0} = {X ∈ k : [X, B] = 0}.

Let h = {X ∈ p : [X, A] = 0} and let kB = {X ∈ k : [X, B] = 0}. Then, X, [A, kB ] = [X, A], kB = 0, X ∈ h.

Therefore, [A, kB ] belongs to the orthogonal complement of h in p. But, [B, [A, kB ]] = −[kB , [B], A]] − [[A, [kB , B]] = 0. Regularity of B implies that the set of elements in p that commute with B coincides with the set of elements in p that commute with A, since [A, B] = 0. Therefore, [B, [A, kB ]] = 0 implies that [A, kB ] ⊂ h. It follows that [A, kB ] = 0. Similar argument shows that [B, kA ] = 0. Let us now come to the proof of Proposition 9.19. Proof The isospectral condition [Q−1 (Lk ), B] = [Lk , A] is the same as [Lk , B] = [Q(Lk ), A]. Hence adB(Lk ) = adA(Q(Lk )) for all Lk ∈ k. If Lk ∈ kA then Lk ∈ kB , therefore adB(Lk ) = 0, and hence Q(Lk ) ∈ kA .This argument shows that kA is invariant for Q.

164

9 Affine-quadratic problem

When Lk is an element of k⊥ A then adB(Lk ] = 0 for Lk = 0, Since adB(Lk ) ∈ h⊥ and adA is injective on h⊥ , ad−1 A ◦ adB(Lk ) = Q(Lk ), Lk ∈ k⊥ A.

Corollary 9.21 If A and B are regular elements in a Cartan subalgebra in p and if kA = 0 then H = 12 Q−1 (Lk ), Lk + A, Lp with Q = ad−1 A ◦ adB is isospectral. Let us now return to the main theme, the existence of integrals of motion associated with the affine-quadratic Hamiltonians on semi-simple Lie that admit isospectral representations. To include the shadow Hamiltonians on the semi-direct product Gs = p K, it will be necessary to think of g as a double Lie algebra equipped with two sets of Lie brackets, the semi-simple [ , ] and the semi-direct [ , ]s . As already remarked earlier, the two types of Lie brackets give rise to two Poisson structures on the dual of g. The following definition applies to both situations. Definition 9.22 A family of functions F on the dual of a Lie algebra g of a Lie group G is said to be complete if it contains an involutive subfamily F0 that is Liouville integrable on each coadjoint orbit of G as well. In regard to complete integrability of isospectral affine-quadratic Hamiltonians, let F denote the class of functions functionally dependent on the shifts f (p − λk + (λ2 − 1)b) of invariant functions f on g∗ , where p, k and b denote the corresponding elements of Lp , Lk and B in g. Analogously, Fs will denote the shifts f (p − λk + λ2 b) of invariant functions f on g∗s . To each of the above families of functions we will adjoin the Hamiltonians generated by the left-invariant vector fields that take values in kA . These functions are of the form f (p, q) = q(X), X ∈ kA . We will use FA to denote their functional span. Then we have the following proposition. Proposition 9.23 If g is a function in FA , then { f , g} = 0 for any function f in F. The same applies to functions in Fs , i.e., { f , g}s = 0. Proof Let us adopt the notations used in the proof of Proposition 9.20. If f 1 ∗ is an invariant function on g , let fλ,μ (p, k) = f λ − k + μb . If { , }λ,μ is as defined in Proposition 9.12 then 0 = { fλ,μ , g}λ,μ = −λ2 { fλ,μ , g} − λμ{ fλ,μ , g}b − (1 − λ2 ){ fλ,μ , g}s

9.3 Integrability

165

for any function g on g∗ . If g is a function in FB then λμ{ fλ,μ , g}b = 0. It follows that λ2 { fλ,μ , g}s − { fλ,μ , g}) − { fλ,μ , g}s . Since λ is arbitrary, { fλ,μ , g}s = { fλ,μ , g} = 0. The methods of Mischenko and Fomenko could be extended to the affine Hamiltonians to show that F ∪ FA is complete on the generic orbits in g∗ . Apparently, Fs ∪ FA is complete relative to the semi-direct Poisson structure. [Bv; Pr], although the proofs do not seem to be well documented in the existing literature. So every isospectral affine-quadratic Hamiltonian H that belongs to F ∪ FA is integrable in the Liouville sense on each coadjoint orbit in g∗ provided that H is in involution with the functions in FA , and the same applies to the corresponding shadow Hamiltonian on g∗s . This means that every isospectral affine-quadratic Hamiltonian is integrable when kA = 0. In the general case when kA = 0, H will be in involution with the Hamiltonians in FA whenever the left-invariant fields with values in kA are symmetries for H. If X is a symmetry for the affine-quadratic problem then its Hamiltonian lift hX (g, L) = L, X(g). Poisson commutes with the affine Hamiltonian H as a consequence of Noether’s theorem (Proposition 6.25 in Chapter 6). Proposition 9.24 Each left-invariant vector field X(g) = gX with X ∈ kA is a symmetry for the canonical affine-quadratic problem dg dt = g(t)(A + u(t)), u(t) ∈ k , Q = I Proof Let g(t) be a solution of = exp λXg(t) = g(t)eλX . Then,

dg dt

= g(t)(A + u(t)) and let hλ (t)

dhλ = hλ (t)(e−λX A(eλX + eλX u(t)(eλX ) dt = hλ (t)(A + vλ (t)), with vλ (t) = eλX u(t)(eλX . Since the Killing form is AdK invariant, vλ , vλ = u, u. Corollary 9.25 Every canonical affine Hamiltonian is completely integrable on the coadjoint orbits of G in g and the same applies to its shadow problem on the coadjoint orbits of Gs in gs . In contrast to the existing literature on integrable systems on Lie algebras, in which functions on g∗ are implicitly regarded as functions on the coadjoint

166

9 Affine-quadratic problem

orbits in g∗s and their integrability is interpreted accordingly, the Hamiltonians in the present context are associated with left-invariant variational problems over differential systems in G, and as such, they are regarded as functions on the entire cotangent bundle. Consequently, they will be integrable in the sense of Liouville only when they are a part of an involutive family of functions on T ∗ G that contains n independent functions, where n is equal to the dimension of G. Evidently, this notion of integrability is different from integrability on coadjoint orbits. For Hamiltonian systems on coadjoint orbits, the Casimirs and the Hamiltonians of the right-invariant vector fields are irrelevant, apart the fact that they figure in the dimension of the coadjoint orbit, but that is not the case for the Hamiltonians on the entire cotangent bundle of the group. At this point we remind the reader that the Hamiltonian lift of any rightinvariant vector field X(g) = Xg, in the left trivialization of the cotangent bundle, is given by hX ((, g) = (g−1 Xg). Saying that each such lift hX is constant along the flow of a left-invariant Hamiltonian H is equivalent to saying that the projection (t) of the integral curve of H on g∗ evolves on a coadjoint orbit. Since {hX , hY } = h[X,Y] , the maximum number of independent integrals of motion hX defined by right-invariant vector fields X in involution with each other is independent of H. This observation brings us to the Cartan algebras. Recall that a Cartan subalgebra of a semi-simple Lie algebra g is a maximal abelian subalgebra h of g such that ad(h) is semi-simple for each h ∈ h. Every complex semi-simple Lie algebra has a Cartan subalgebra and any two Cartan subalgebras are isomorphic [Hl, p. 163]. For real Lie algebras there may be several non-conjugate Cartan subalgebras but they all have the same dimension. This dimension is called the rank of g. An element A of g is called regular if the dimension of ker(adA) is equal to the rank of g. On a semi-simple Lie algebra each coadjoint orbit that contains a regular element has dimension equal to dim(g) − rank(g). These orbits are generic. In fact, every element of the Cartan algebra h that contains A is orthogonal to the tangent space of the coadjoint orbit at A relative to the Killing form. So elements in the Cartan algebra do not figure in the integrability on coadjoint orbits, but they do in the integrability on the cotangent bundle of G. The same applies to the invariant functions on g. It is important to note that integrability on the level of the Lie algebra implies solvability by quadrature on G, because once the extremal control u(t) is known, the remaining equation dg dt = g(t)(A+u(t)) reduces to a time-varying equation on G, which, at least in principle, is solvable in terms of a suitable system of coordinates on G. We will get a chance to see in Chapter 14 that the “good” coordinates on G make use of vector fields in the Cartan algebra in g defined by A.

10 Cotangent bundles of homogeneous spaces as coadjoint orbits

This chapter provides an important passageway from Lie algebras to cotangent bundles that connects isospectral Hamiltonians with integrable Hamiltonians systems on these cotangent bundles. We will presently show that the cotangent bundles of certain manifolds M, such as the spheres, the hyperboloids, and their isometry groups SOn (R) and SO(1, n), can be represented as the coadjoint orbits on sln (R), in which case the restriction of the isospectral integrals of H to these coadjoint orbits provides a complete family of integrals of motion for the underlying variational problem on the base manifold. Recall the classification of semi-simple Lie algebras into compact, noncompact, and Euclidean types. On semi-simple Lie groups of compact type each coadjoint orbit is compact, and cannot be the cotangent bundle of any manifold. In fact, the coadjoint orbits of compact simple Lie groups have been classified by A. Borel and are known to be compact K¨ahler manifolds [Bo]. Remarkably, the coadjoint orbits of Lie algebras of non-compact type are all cotangent bundles of flag manifolds [GM]. We will show below that the same is true for the coadjoint orbits relative to the semi-direct products. We will exploit the fact that the dual of any Lie algebra g that admits a Cartan decomposition carries several Poisson structures which account for a greater variety of coadjoint orbits. Our discussion, motivated by some immediate applications, will be confined to the vector space Vn+1 of (n + 1) × (n + 1) matrices with real entries of zero trace, but a similar investigation could be carried out on any semi-simple Lie algebra with a Cartan decomposition g = p ⊕ k.

167

168 10 Cotangent bundles of homogeneous spaces as coadjoint orbits

10.1 Spheres, hyperboloids, Stiefel and Grassmannian manifolds We will let , denote the quadratic form A, B = − 12 Tr(AB), A, B in Vn+1 . Since Vn+1 carries the canonical Lie algebra sln+1 (R) with its Killing form equal to 2(n + 1)Tr(AB), , is a negative scalar multiple of the Killing form on sln+1 (R) and hence, inherits all of its essential properties. In particular, , is non-degenerate and invariant, in the sense that A, [B, C] = [A, B], C for any matrices A, B, C in sln+1 (R). It is also positive-definite on k = son+1 (R). As remarked before, Vn+1 has two distinct Poisson structures: the first associated with the dual of sln+1 (R) with its standard Lie–Poisson bracket and the second associated with the dual of the semi-direct product Sn+1 ⊕s son+1 (R), where Sn+1 denotes the spaces of symmetric matrices in Vn+1 . In addition to these two structures, we will also include the pseudo-Riemannian decompositions of Cartan type in Vn+1 and the induced Poisson structures. The pseudo-Riemannian decompositions are due to the automorphism σ (g) = D(gT )−1 D−1 , g ∈ SLp+q (R), defined by a diagonal matrix D with its first p diagonal entries equal to 1 and the remaining q diagonal entries equal to −1. It is easy to check that σ is involutive and that the subgroup K of fixed points of σ is equal to SO(p, q), p+q = n+1, the group that leaves the quadratic form (x, y)p,q =

p i=1

xi yi −

n+1

xi yi

(10.1)

n−p+1

invariant. The decomposition Vn+1 = Sp,q ⊕ k induced by the tangent map σ∗ consists of the Lie algebra k = so(p, q) and the Cartan space ' A B T T (10.2) ,A = A ,C = C . Sp,q = P ∈ Vn+1 : P = −BT C The Cartan space Sp,q can be also described as the set of matrices P in Vp+q which are symmetric relative to the quadratic form (x, y)p,q , i.e., (Px, y)p,q = (x, Py)p,q . In what follows it will be convenient to regard the Riemannian metric (x, y) = n+1 i=1 xi yi as the limiting case of the pseudo-Riemannian case with q = 0. The symmetric pair (SLp+q (R), SO(p, q)) for q = 0 is strictly pseudoRiemannian because the subgroup of the restrictions of AdSO(p,q) to Sp+q is not a compact subgroup of Gl(Sp+q ) (since the trace form is indefinite on Sp+q ). Nevertheless, this automorphism endows Vn+1 with another semidirect product Lie algebra structure, namely Sp,q ⊕s so(p, q).

10.1 Spheres, hyperboloids, Stiefel and Grassmannian manifolds

169

So Vn+1 has three distinct Poisson structures, the first induced by the semisimple Lie brackets, and the other two induced by the semi-direct products corresponding to the Riemannian and pseudo-Riemannian cases. In each case, Vn+1 is foliated by the coadjoint orbits. We will presently show that the cotangent bundles of many symmetric spaces can be realized as the coadjoint orbits on Vn+1 , and that will make a link with mechanical and geometric systems of the underlying spaces. In the semi-simple case, the coadjoint orbits can be identified with the adjoint orbits via the Killing form. As a consequence, each adjoint orbit is symplectic and hence is even dimensional. The adjoint orbit through any matrix P0 can be identified with the quotient SLn+1 (R)/St(P0 ), where St(P0 ) denotes the stationary group of P0 , consisting of the matrices g such that gP0 g−1 = P0 . The Lie algebra of St(P0 ) consists of the matrices X ∈ sln+1 (R) which commute with P0 . The orbits of maximal dimension, called generic, correspond to matrices P0 whose stationary group is of minimal dimension. It can be proved that each generic orbit is generated by a matrix P0 that belongs to some Cartan subalgebra of sln+1 (R). Since Cartan algebras in sln+1 (R) are n-dimensional, generic orbits in sln+1 (R) are n(n − 1)-dimensional. A more detailed study of adjoint orbits in sln+1 (R) is sufficiently complicated, and would take us away from our main objectives, so will not pursue it here except in the simplest cases. Instead, we will turn attention to the coadjoint orbits relative to the semi-direct products. The following lemma describes the coadjoint orbits on any semi-simple Lie algebra g that admits a Cartan decomposition p ⊕ k. Lemma 10.1 Let Gs denote the semi-direct product p K and let , denote ∗ (l ), for some any scalar multiple of the Killing form. Suppose that l = Adg−1 0 ∗ 0 ∈ g and some g = (X, h) ∈ Gs . Suppose further that l0 −→ L0 = P0 + Q0 , and l −→ L = P + Q are the correspondences defined by the quadratic form , with P0 and P in p and Q0 and Q in k. Then P = Adh (P0 ), and Q = [Adh (P0 ), X] + Adh (Q0 ). Proof

(10.3)

If Z = U + V is an arbitrary point of g with U ∈ p and V ∈ h, then d −1 (g (sU, esV )g)|s=0 ds d = (−Adh−1 (X) + Adh−1 (sU + esV (X)e−sV ), h−1 esV h)|s=0 ds = Adh−1 (U + [X, V]) + Adh−1 (V).

Adg−1 (Z) =

170 10 Cotangent bundles of homogeneous spaces as coadjoint orbits

Hence, l(Z) = l0 (Adg−1 (Z) = P0 , Adh−1 (U + [X, V]) + Q0 , Adh−1 (V) = Adh (P0 ), U + Adh (P0 ), [X, V] + Adh (Q0 ), V = Adh (P0 ), U + [Adh (P0 ), X], V + Adh (Q0 ), V = P, U + Q, V. Since U and V are arbitrary, P = Adh (P0 ), Q = [Adh (P0 ), X] + Adh (Q0 ), and (10.3) follows. Consider now the coadjoint orbit through a symmetric matrix P0 in the semidirect product Sn+1 ⊕s son+1 (R), where Sn+1 denotes Sp+q with p = n + 1, q = 0. For simplicity, we will dispense with the constraint that P0 has trace 1 Tr(P)I, where zero. Any symmetric matrix P can be written as P = P0 + n+1 P0 has zero trace. Then the coadjoint orbit through P0 is essentially the same as the coadjoint orbit through P, they differ by a constant multiple of the identity. If a and b are any points in Rn+1 then P = a ⊗ b will denote the rank one matrix defined by Px = (x, b)a for any x ∈ Rn+1 . It is easy to verify that a ⊗ b + b ⊗ a is symmetric while a ∧ b = a ⊗ b − b ⊗ a is skew-symmetric. In particular, a ⊗ a is symmetric for any a ∈ Rn+1 . Proposition 10.2 Let a1 , . . . , ar , r > 1 be any orthonormal vectors in Rn+1 . Then the coadjoint orbit through P0 = ri=1 ai ⊗ ai relative to the semi-direct bracket in gs is the tangent bundle of the oriented Grassmannian Gr(n+1, r) = SOn+1 (R)/SOr (R) × SOn+1−r (R). The coadjoint orbit through P0 = ri=1 λi (ai ⊗ ai ) with distinct numbers λ1 , . . . , λr is the tangent bundle of the positively oriented Stiefel manifold St(n + 1, r). Proof Let V0 denote the linear span of a1 , . . . , ar oriented so that a1 , . . . , ar is a positively oriented frame. With each matrix Ph in (10.3) given by Ph = r i=1 h(ai ) ⊗ h(ai ), with h ∈ SOn+1 (R), we will associate the vector space Vh spanned by h(a1 ), . . . , h(ar ) and then identify Ph with the reflection Rh around Vh . This reflection is defined by Rh x = x for x in Vh and Rh x = −x for x in the positively oriented orthogonal complement of Vh . Every r-dimensional vector subspace V of Rn+1 admits a positively oriented orthonormal frame b1 , . . . , br which can be identified with Vh for some h ∈ SOn+1 (R), because SOn+1 (R) acts transitively on the space of r-dimensional spaces that are positively oriented relative to V0 .

10.1 Spheres, hyperboloids, Stiefel and Grassmannian manifolds

171

In the identification with the reflections Rh , the action of SOn+1 (R) on the vector spaces translates into the adjoint action Rh = h−1 R0 h, h ∈ SOn+1 (R). The isotropy group of R0 in SOn+1 (R) is equal to SOr (R) × SOn+1−r (R). Hence, the orbit through R0 is diffeomorphic with G(n + 1, r), which in turn is diffeomorphic with the matrices P in expression (10.3). The tangent space at Vh can be realized as the vector spaces of skewsymmetric matrices obtained by differentiating curves R() in the space of reflections of Vh . Each such curve R() is a curve in SOn+1 (R) that satisfies dR dR R2 () = I. Therefore, R dR d + d R = 0. If R d = R() (), then the preceding yields R R = − .

(10.4)

R (x) = −x, x ∈ Vh , and R x = x, x ∈ Vh⊥ .

(10.5)

This means that

Hence, (Vh ) ⊆ Vh⊥ and (Vh⊥ ) ⊆ Vh . If c1 , . . . , cn+1−r denotes any orthonormal basis in Vh⊥ , then

=

r n+1−r

jk cj ∧ bk where jk = (cj , bk ).

(10.6)

j=1 k=1

It follows that the tangent space at Vh and the linear span of matrices cj ∧ bk are isomorphic But this linear span is also generated by the matrices Q in equation (10.3) since [P, X] =

r

[bi ⊗ bi , X] =

i=1

=

r

[bi ∧ Xbi ]

i=1

Xbi , bj bi ∧ bj +

i, j=1

=

r

n+1−r r j=1

(cj , Xbk )cj ∧ bk

k

n+1−r r

(cj , Xbk )cj ∧ bk ,

j=1

k

because Xbi , bj bi ∧ bj + Xbj , bi bj ∧ bi = 0 for any symmetric matrix X. This proves the first part of the proposition. Suppose now that P0 = ri=1 λi (ai ⊗ ai ) with λi = λj , i = j. The correspondence between points p = (b1 , . . . , br ) of the Steifel manifold St(n + 1, r) and matrices P = ri=1 λi (bi ⊗bi ) is one to one. Then Adh (P0 ) = ri=1 λi (h(ai )⊗ h(ai )) corresponds to the point hp0 = h(a1 , . . . , ar ). That is, the adjoint orbit {Adh (P0 ); h ∈ SOn+1 (R)} corresponds to the orbit {hp0 : h ∈ SOn+1 (R)}. The

172 10 Cotangent bundles of homogeneous spaces as coadjoint orbits action of SOn+1 (R) on St(n + 1, r) by the matrix multiplications from the right identifies St(n + 1, r) as the homogenous space SOn+1 (R)/SOn+1−r (R). The tangent space at (b1 , . . . , br ) = h(a1 , . . . , ar ) consists of vectors of the form h (a1 , . . . , ar ) for some skew-symmetric matrix in the linear span of bj ∧ bk , 1 ≤ j, k ≤ r and bj ∧ ck , j = 1, . . . , r, k = 1, . . . , n + 1 − r, where c1 , . . . , cn+1−r stands for any orthonormal basis in the orthogonal complement of the vector space spanned by b1 , . . . , br . As X varies over all symmetric matrices having zero trace, matrices Q = [P, X] span the same space as the ’s in the above paragraph. The following argument demonstrates the validity of this statement. Matrices (bi ⊗ bj ) + (bj ⊗ bi ), (ci ⊗ bj ) + (bi ⊗ cj ), (ci ⊗ cj ) + (cj ⊗ ci ) form a basis for the space of symmetric matrices, and hence, any symmetric matrix X can be written as the sum X=

r Xbi , bj ((bi ⊗ bj ) + (bj ⊗ bi )) i, j

+

r n+1−r i=1

+

Xbi , cj ((bi ⊗ cj ) + (cj ⊗ bi ))

j=1

n+1−r

Xci , cj ((ci ⊗ cj ) + (cj ⊗ ci )).

i, j

Then, r [λi (bi ⊗ bi ), (bj ⊗ bk ) + bk ⊗ bj )] i=1

=

r

λi (δij (bk ⊗ bi ) + δik (bj ⊗ bi ) − δij (bi ⊗ bk ) − δik (bi ⊗ bj ))

i=1

= λj (bk ∧ bj ) + λk (bj ∧ bk ) = (λj − λk )(bk ∧ bj ), r

[λi (bi ⊗ bi ), (bj ⊗ ck ) + (ck ⊗ bj )]

i=1

=

r

λi (δij (ck ⊗ bi ) − (bi ⊗ ck )) = λj (ck ⊗ bj ), and

i=1 r [λi (bi ⊗ bi ), (cj ⊗ ck ) + (ck ⊗ cj )] = 0 i=1

10.1 Spheres, hyperboloids, Stiefel and Grassmannian manifolds yields Q = [P, X] = λj (ck ∧ bj ).

r

j,k (bj , Xbk )(λj

− λk )(bk ∧ bj ) +

r j=1

n+1−r k=1

173

(ck , Xbj )

Proposition 10.3 The canonical symplectic form on each of St(n + 1, r) and Gr(n + 1, r) agrees with the symplectic form of the corresponding coadjoint orbits. Proof The symplectic form on coadjoint orbits is given by ωL (V1 , V2 ) = L, [V2 , V1 ], where V1 and V2 are tangent vectors at L. Every tangent vector V at a point L = P + Q is of the form V = [P, A] + [Q, A] + [P, B],

(10.7)

for some skew-symmetric matrix A and some symmetric matrix B. The argument is simple: if (X(t), h(t)) denotes any curve in Gs such that dh (0) = A and dX dt dt (0) = B0 , then P(t) = Adh (P), and Q = [Adh (P), X]+Adh (Q0 ) is the corresponding curve on the coadjoint orbit through P + Q. Therefore, dQ dP dt (0) = [P, A], dt (0) = [[P, A], X] + [P, B0 ] = [Q, A] + [P, [A, X] + B0 ] with the aid of Jacobi’s identity. This yields (10.7) with B = [A, X] + B0 . r r = [P, X] = Suppose now that P = i=1 xi ⊗ xi and Q i=1 xi ∧ yi r is a point of the coadjoint orbit through P0 = a ⊗ a corresponding i i i=1 to some orthonormal vectors a1 , . . . , ar in Rn+1 . That is, xi = h(ai ) for n+1−r Xxi , cj cj for each i = 1, . . . , r, where some h ∈ SOn+1 (R) and yi = j=1 c1 , . . . , cn+1−r is any orthonormal extension of x1 , . . . , xr . Then P˙ =

r

xi ⊗ x˙i + x˙i ⊗ xi , Q˙ =

i=1

r

x˙i ∧ yi + xi ∧ y˙i

(10.8)

i=1

are tangent vectors at P, Q in the tangent bundle of Gr(n + 1, r) for any vectors x˙1 , . . . , x˙r and y˙1 , . . . , y˙r that satisfy (xi , x˙j ) + (x˙i , xj ) = 0 and (xi , y˙j ) + (x˙i , yj ) = 0. The canonical symplectic form on the cotangent (tangent) bundle of Gr(n + 1, r) is given by r

(z˙i , y˙i ) − (x˙i , w˙ i )

(10.9)

i=1

for any tangent vectors (x˙1 , . . . , x˙r , y˙1 , . . . , y˙r ) and (z˙1 , . . . , z˙r , w˙ 1 , . . . , w˙ r ) at (P, Q). The coadjoint symplectic form is given by ωP+Q (V1 , V2 ) = L, [B2 , A1 ] + [A2 , B1 ] + [A2 , A1 ]

(10.10)

for any tangent vectors Vi = [L, Ai + Bi ], i = 1, 2. To equate these two expressions, let P˙ = [P, A] and Q˙ = [Q, A] + [P, B] for some matrices A and B with A skew-symmetric and B symmetric. Then,

174 10 Cotangent bundles of homogeneous spaces as coadjoint orbits

xi ⊗ x˙i + x˙i ⊗ xi =

i=1

r

[xi ⊗ xi , A] =

i=1

r

xi ∧ Axi + Axi ∧ xi

i=1

implies that x˙i = Axi −

r (Axj , xi )xj .

(10.11)

j=1

Similar calculation yields y˙i = Ayi − Bxi +

r (Bxi xj )xj .

(10.12)

j=1

The following lemma will be useful for the calculations below. Lemma 10.4 Suppose that B is a symmetric and A a skew-symmetric matrix. Then, x ⊗ x, [A, B] = (Bx, Ax), x ∧ y, [A2 , A1 ] = (A2 x, A1 y) − (A1 x, A2 y). Proof

First note that Tr(x ⊗ y) =

n

i=0 xi (y, ei )

= (x, y). Then,

1 Tr(x ⊗ [A, B]x) 2 1 1 = (Tr(x ⊗ BAx − x ⊗ ABx)) = ((x, BAx) − (x, ABx)) 2 2 1 = ((Bx, Ax) + (Ax, Bx)) = (Bx, Ax). 2

x ⊗ x, [A, B] =

The second formula follows by an analogous argument. With these notations behind us, let x˙i = Axi −

r r (A1 xj , xi )xj , y˙i = A1 yi − B1 xi + (Bxi xj )xj , j=1

z˙i = A2 xi −

j=1

r r (A2 xj , xi )xj , w˙ i = A2 yi − B2 xi + (Bxi xj )xj j=1

j=1

denote the tangent vectors at P = ri=1 xi ⊗ xi , Q = ri=1 xi ∧ yi defined by the matrices B1 + A1 and B2 + A2 in Sn+1 ⊕ son+1 (R). Then,

10.1 Spheres, hyperboloids, Stiefel and Grassmannian manifolds

175

ωP+Q (V1 , V2 ) = L, [B2 , A1 ] + [A2 , B1 ] + [A2 , A1 ] = = =

r xi ⊗ xi , [B2 , A1 ] + [A2 , B1 ] + xi ∧ yi , [A2 , A1 ] i=1 r

(A1 xi , B2 xi ) − (A2 xi , B1 xi ) + (A2 xi , A1 yi ) − (A1 xi , A2 yi )

i=1 r

r (A1 xi , −w˙ i + A2 yi + (B2 xi , xj )xj ) − (A2 xi , − y˙i + A1 yi

i=1

j=1 r + (B1 xi , xj )xj ) + (A2 xi , A1 yi ) − (A1 xi , A2 yi ) j=1

r r = (A1 xi , −w˙ i ) + (A2 xi , y˙i ) = (z˙i , y˙i ) − (x˙i , w˙ i ), i=1

r

i=1

r

because i, j=1 (B1 xi , xj )(A2 xi , xj ) = i, j=1 (B2 xi , xj )(A1 xi , xj ) = 0. The calculation on the Stiefel manifold is quite analogous and will be omitted. Corollary 10.5 The generic coadjoint orbit of Gs through a non-singular symmetric matrix P0 is equal to the tangent bundle of SOn+1 (R) realized as the Stiefel manifold St(n + 1, n + 1) while the coadjoint orbit through rank one symmetric matrix P0 = a ⊗ a is equal to the tangent bundle of the sphere Sn = {x : ||x|| = ||a||}. Next consider analogous orbits through a point P in Sp+q defined by the pseudo Riemannian symmetric pair (SLp+q (R), SO(p, q)). In contrast to the Riemannian case, where each symmetric matrix admits an orthonormal basis of eigenvectors, matrices P in Sp,q may have complex eigenvalues, which accounts for different types. The hyperbolic rank one matrices are matrices of the form x ⊗ Dx for some vector x ∈ Rn+1 , where D is the diagonal matrix with its first p diagonal entries equal to 1, and the remaining q diagonal entries equal to −1. These matrices will be denoted by x ⊗p,q x, while x ∧p,q y will denote the matrix x ⊗p,q y − y ⊗p,q x. It follows that (x ⊗ x)p,q u = (x, u)p,q x, and therefore, (x ⊗p,q x)u, v)p,q = (x, u)p,q (x, v)p,q = (u, (x ⊗p,q x)v)p,q . 1 (x, x)p,q I Since the trace of (x ⊗p,q x) is equal to (x, x)p,q , P = (x ⊗p,q x) − n+1 is in Sp,q . A similar calculation shows that Q = x ∧p,q y belongs to so(p, q).

176 10 Cotangent bundles of homogeneous spaces as coadjoint orbits The coadjoint orbits relative to Gp,q = Sp,q SO(p, q) can be identified as the tangent bundles of the “pseudo-Stiefel” and “pseudo-Grassmannians” in a manner similar to the Riemannian setting. We will not pursue these investigations in any detail here, except to note the extreme cases, the coadjoint orbit through P0 = a ⊗p,q a and the coadjoint orbit through P0 = n+1 i=1 λi ai ⊗p,q ai , λi = λj , i = j. As in the Riemannian case, we will omit the zero-trace requirement. Proposition 10.6 The coadjoint orbit of the semi-direct product Gp,q = Sp,q SO(p, q) through P0 = a ⊗p,q a is equal to {x ⊗p,q x + x ∧p,q y : (x, x)p,q = (a, a)p,q , (x, y)p,q = 0}. The latter is symplectomorphic to the cotangent bundle of the hyperboloid Hp,q = {(x, y) : (x, x)p,q = (a, a)p,q , (x, y)p,q = 0}. The symplectic form on the coadjoint orbit coincides with the canonical symplectic form ω on Hp,q given by ωx,y ((x˙1 , y˙p,q ), (x˙2 , y˙2 )) = (x˙1 , y˙2 )p,q − (x˙2 , y˙1 )p,q , where (x˙1 , y˙1 ) and (x˙2 , y˙2 ) denote tangent vectors at (x, y). p q Proposition 10.7 Let P0 = i=1 λi (ai ⊗p,q ai ) + i=1 μi (bi ⊗p,q bi ) for vectors a1 , . . . , ap , b1 , . . . , bq that satisfy (ai , aj )p,q = δij , (ai , bj )p,q = 0, (bi , bj )p,q = −δij and some distinct numbers λ1 , . . . , λp and μ1 , . . . , μq . The coadjoint orbit of the semi-direct product Gp,q = Sp,q SO(p, q) through P0 is symplectomorphic to the cotangent bundle of SO(p, q). The proofs are similar to the proofs in Propositions 10.2 and 10.3 and will be omitted. The case p = 1, q = n is particularly interesting for the applications, for then the hyperboloid H1,n defined by a vector a that satisfies (a, a)1,n = 1 coincides with the standard hyperboloid Hn = x ∈ Rn+1 : 2 x12 − n+1 i=2 xi = 1 . Its realization as a coadjoint orbit allows for conceptually clear explanations of the mysterious connections between the problem of Kepler and the geodesics of space forms [Ms1; Os]. This connection brings us to an important general remark about the modifications of the affine problem in the case that the quadratic form , is indefinite on k, as is on k = so(p, q) where the TKilling form is indefinite. In these situations the problem of minimizing 12 0 u, u dt over the trajectories of dg dt = g(t)((A + u(t)) in Gp,q with the controls u(t) in so(p, q) is not well defined since the cost is not bounded below. Of course, the affine system is still controllable for any regular drift A. So any pair of points in Gp,q can be connected by a trajectory in some finite time T, but a priori there is no guarantee that the infimum of the cost over such trajectories will be finite.

10.2 Canonical affine Hamiltonians on rank one orbits

177

Nevertheless, the affine Hamiltonian H = 12 Lk , Lk +A, Lp is well defined and corresponds to the critical value ofthe cost, ratherthan to its minimum. T That is to say, the endpoint map u → 12 0 u(t), u(t) dt, g(T)) is singular at the contol u(t) = Lk (t) associated with an integral curve (g(t), L(t)) of the Hamiltonian vector field H. As we have already remarked above, the semi-simple adjoint orbits are also the cotangent bundles. For the sake of comparison with the sphere, the proposition below describes such an orbit through rank one matrices. 1 I, (a, a) = 1, Proposition 10.8 The adjoint orbit through P0 = a ⊗ a − n+1 is symplectomorphic to the cotangent bundle of the real projective space Pn+1 .

Proof Let S denote the orbit gP0 g−1 , g ∈ SLn+1 (R). If R ∈ SOn+1 (R) then 1 I. Since SOn+1 (R) acts transitively on the unit sphere RP0 R−1 = Ra⊗Ra− n+1 in Rn+1 , a can be replaced by e0 . 1 1 −1 −1 It follows that g(e0 ⊗ e0 − n+1 I)g =−1g(e0 ⊗ e0 )g − n+1 I, hence S is diffeomorphic to the orbit g(e0 ⊗ e0 ))g : g ∈ SLn+1 (R) . The reader can readily verify that g(e0 ⊗ e0 )g−1 = ge0 ⊗ (gT )−1 e0 . This relation shows that the the Euclidean inner product to 1 x · y is equal T −1 −1 wheneverx = ge0 and y = (g ) e0 . Hence the orbit g(e0 ⊗ e0 ))g : g ∈ SLn+1 (R) is confined the set x ⊗ y : x ∈ Rn+1 , y ∈ Rn+1 , (x · y) = 1 . But these sets are equal since every pair of points x and y such that x · y = 1 can −1 be realized as x = ge0 and y = gT e0 for some matrix g ∈ SLn+1 (R). Then the set of matrices {x ⊗ y : x · y = 1} can be identified with the set of lines {αx, α1 y), x · y = 1}, and the later set is diffeomorphic to the cotangent bundle of the real projective space Pn+1 .

10.2 Canonical affine Hamiltonians on rank one orbits: Kepler and Newmann We will now return to the canonical affine Hamiltonian H = 12 Lh , Lh + A, Lp and consider its restrictions to the coadjoint orbits that correspond to the cotangent bundles of spheres and hyperboloids. This means that Lp is ||x||2

restricted to matrices x⊗ x− n+1 I, where ||x||2 = (x, x) , and Lh is restricted to matrices x ∧ y, (x, y) = 0. Then,

178 10 Cotangent bundles of homogeneous spaces as coadjoint orbits 1 1 ||x||2 1 I = − Tr(A(x ⊗ x)) = − (Ax, x) A, Lp = − Tr A x ⊗ x − 2 n+1 2 2 1 Lh , Lh = − Tr(x ∧ y)2 = Tr(||x||2 y ⊗ y + ||y||2 x ⊗ x) = ||x||2 ||y||2 . 2 Hence H reduces to 1 1 H = ||x||2 ||y||2 − (Ax, x) . (10.13) 2 2 The Hamiltonian equations dLh dLp = [A, Lp ], = [Lh , Lp ] dt dt reduce to ||x||2 d d x ⊗ x − I = ||x||2 (x ⊗ y + y ⊗ x), (x ∧ y) = x ∧ Ax. dt n+1 dt (10.14) It follows that x˙ ⊗ x + x ⊗ x˙ −

˙ 2||x|| (x, x) I = ||x||2 (x ⊗ y + y ⊗ x), x˙ ∧ y + x ∧ y˙ n+1 = x ∧ Ax.

On the “sphere” ||x||2 = h2 , (x, x) ˙ = 0, and (x, y) ˙ = −(x, ˙ y) , and ˙ 2||x|| (x, x) I x = (||x||2 (x ⊗ y + y ⊗ x))x. x˙ ⊗ x + x ⊗ x˙ − n+1 ˙ = (x ∧ Ax)x Therefore x˙ = h2 y. The remaining equation (x˙ ∧ y + x ∧ y)x yields y˙ = Ax − h12 ((Ax, x) + h2 ||y||2 )x. It follows that equations 10.14 reduce to 1 x˙ = h2 y, y˙ = Ax − 2 ((Ax, x) + h2 ||y||2 )x. (10.15) h If we replace A by −A and take h = 1 then equations (10.15) become x˙ = y, y˙ = −Ax + ((Ax, x) − ||y||2 )x.

(10.16)

Equations (10.16) with = 1 correspond to the Hamiltonian equations for the motions of a particle on the sphere Sn under a quadratic potential 1 2 (Ax, x) [Ms4]. This system originated with C. Newmann in 1856 [Nm] and has been known ever since as the Newmann system on the sphere. It is a point of departure for J. Moser in his book on integrable Hamiltonian systems [Ms2]. We will presently show that all integrals of motion found in Moser’s book are simple consequences of the spectral representation discussed in the previous chapter. For = −1 equations (10.16) coincide with the Hamiltonian

10.3 Degenerate case and Kepler’s problem

179

equations on the hyperboloid Hn for the motion of a particle on Hn in the presence of a quadratic potential (Ax, x)−1 and the same can be said for the mechanical system on the projective space on the semi-simple orbit. Of course, the integration procedure in all cases is essentially the same, as will soon be made clear. Remarkably, even the degenerate case A = 0 sheds light on mechanical systems; we will presently show that it provides natural connections between Kepler’s system and the Riemannian geodesics on spaces of constant curvature [Ms2; O2].

10.3 Degenerate case and Kepler’s problem When A = 0, the Hamiltonian H reduces to H = equations (10.15) reduce to x˙ = ||x||2 y, y˙ = −||y||2 x.

1 2 2 2 ||x|| ||y|| ,

and

(10.17)

It will be convenient to revert to the notations employed earlier, and use SO to denote the group SOn+1 (R) when = 1, and SO(1, n) when = −1. It is clear that our Hamiltonian H is invariant under SO . The associated moment map is given by = x ∧ y.

(10.18)

Evidently, is constant along the solutions of (10.17). The solutions of (10.17) satisfy x¨ + ||x||2 ||y||2 x = 0. On any sphere ||x||2 = ||a||2 and energy level H = h2 , the solutions are given by x(t) = a cos th + b sin th, with a and b constant vectors that satisfy ||a||2= ||b||2 = h2 , (a, b) = 0. Dx(t) dx 2 = 0, therefore x(t) is a When h2 ||a||2 = 1, then || dx dt || = 1 and dt dt geodesic on the sphere ||x||2 = ||a||2 . It follows that each solution curve traces a geodesic circle with speed h on either the sphere ||x||2 = ||a||2 when = 1 or the hyperboloid ||x||2−1 = ||a||2−1 when = −1. In the latter case the geodesic is confined to the region n dx0 2 dxi 2 − > 0. dt dt i=1

180 10 Cotangent bundles of homogeneous spaces as coadjoint orbits In the complementary open region, H = −h2 , the solutions are given by x(t) = a cosh ht + b sinh ht. These solutions are also geodesics on each hyperboloid ||x||2 = ||a||2 , but now they trace hyperbolas with speed h. The remaining case, H = 0, generates horocycles on each hyperboloid ||x||2 = ||a||2 . These solutions are given by x(t) = at + b.

10.3.1 Kepler’s problem Let us now digress briefly into Kepler’s problem. Kepler’s problem concerns the motion of a planet around an immovable planet in the presence of the gravitational force. On the basis of the empirical evidence, Kepler made two fundamental observations about the motion of the moon around the earth, known today as Kepler’s laws. The first law of Kepler states that the moon moves on an elliptic orbit around the earth, and the second law of Kepler states that the position vector of the moving planet relative to the stationary planet sweeps out equal areas in equal time intervals as the planet orbits around the stationary planet. The main motivation for this diversion, however, is not so much driven by this classic knowledge but rather by a remarkable discovery that the solutions of Kepler’s problem are intimately related to the geometry of spaces of constant curvature. This discovery goes back to A. V. Fock’s paper of 1935 [Fk] in which he reported that the symmetry group for the motions of the hydrogen atom is O4 (R) for negative energy, E3 O3 (R) for zero energy, and O(1, 3) for positive energy. It is then not altogether surprising that similar results apply to the problem of Kepler, since the energy function for Kepler’s problem is formally the same as the energy function for the hydrogen atom. This connection between the problem of Kepler and the geodesics on the sphere was reported by J. Moser in 1970 [Ms1], and even earlier, by G. Gy¨orgyi in 1968 [Gy], while similar results for positive energies and the geodesics on spaces of negative constant curvature were reported later by Y. Osipov [O1; O2]. As brilliant as these contributions were, they, nevertheless, did not offer any explanations for these enigmatic connections between planetary motions and geodesics on space forms. This enigma later inspired V. Guillemin and S. Sternberg to take up the problem of Kepler in a larger geometric context with Moser’s observation as the focal point for this work [GS], but this attempt did not give any clues for the basic mystery. Our aim is to show that the affine Hamiltonian lifts this mystery behind the Kepler’s problem.

10.3 Degenerate case and Kepler’s problem

181

Let us first recall the essential facts. It was not until Newton that a mathematical foundation was laid out from which Kepler’s observations could be deduced mathematically. According to Newton, the gravitational force of attraction exerted on a planet of mass m by another planet of mass M is given by q , F = −kMm q3 where k denotes the gravitational constant and q denotes the position vector of the first planet relative to the second planet. This formulation of the force presupposes that the planets are immersed in a three-dimensional Euclidean space E3 with an origin O so that the position vector q is equal to the difference 1 , and that the distance between the planets 2 − OP of the position vectors OP is expressed by the usual formula 2 − OP 1 2 . q||2 = OP For the problem of Kepler one of the planets is assumed stationary, in which where OP denotes case the origin of E3 is placed at its center. Then q = OP, the position vector of the moving planet, and the movements of the planet are governed by the equation q d2 q = −kmM , 2 dt q3

(10.19)

according to Newton’s second law of motion. Equation (10.19) can be expressed in Hamiltonian form dp q dq = mp, = −kM . dt dt q3

(10.20)

In the Hamiltonian formalism the gravitational force is replaced by the potenq and the equations of motion are generated by the tial energy V(q) = −kM q

energy function E(q, p) = 12 m2 p2 + V(q), in the sense that equation (10.2) coincides with ∂E dp ∂E dq = , =− . dt ∂p dt ∂q

The solutions of the above equations are easily described in terms of the energy E and the angular momentum L = q×p, both of which are constant along each solution of equation (10.2). The simplest case occurs when L = 0. Then the initial position q(0) is colinear with the initial momentum p(0). In such a case the solutions of (10.2) remain on the line defined by these initial conditions. The angular momentum is not equal to zero whenever the initial conditions are not colinear. Then each solution (q(t), p(t)) of (10.2) is confined to the

182 10 Cotangent bundles of homogeneous spaces as coadjoint orbits

plane spanned by the initial vectors q(0) and p(0), a consequence of the conservation of the angular momentum, and traces a conic in this plane; a hyperbola when E > 0, a parabola when E = 0, and an ellipse when E < 0. To make our earlier claim more compelling, we will consider the n-dimensional Kepler problem given by the Hamiltonian E=

1 1 p2 − 2 q

in the phase space Rn /0 × Rn corresonding to the normalized constants m = kM = 1, and the associated differential equations dp 1 dq = p, =− q. dt dt q3 Lemma 10.9 Each of L = q ∧ p and F = Lp − solution of (10.21).

(10.21) q q

is constant along any

The proof is simple and will be omitted. It follows that L is a direct generalization of the angular momentum since x × (q × p) = (x · p)q − (x · q)p = L(x). The other conserved quantity F is called the Runge–Lenz vector, or the eccentricity vector. The length of F is called the eccentricity for reasons that will become clear below (see also [An] for more historical details). Lemma 10.10 son (R). Then,

Let L2 = − 12 Tr(L2 ) denote the standard trace metric on

(a) L2 = q2 p2 − (q · p)2 . (b) F2 = 2L2 E + 1. Proof

Let e1 , . . . , en denote the standard basis in Rn . Then 1 ei · L2 ei 2 n

L2 = −

i=1

=−

n 1

2

ei · ((pi ((q · p)q − q2 p) − qi (p2 q − (q · p)p))

i=1

= q2 p2 − (q · p)2 . 1 To prove (b), let e = q q. Then F = p2 q − (p · q)p − e = (p2 q − 1)e − (p · q)p. It follows that

F·e=

L2 − 1 and F · p = −e · p. q

10.3 Degenerate case and Kepler’s problem

183

Then, F2 = F · F = (p2 q − 1)F · e − (p · q)F · p L2 2 = (p q − 1) − 1 + (p · q)e · p q 1 L2 + (−L2 + p2 q2 ) − (p2 q − 1) q q L2 + 1 = 2EL2 + 1. = (p2 q − 2) q = (p2 q − 1)

Proposition 10.11 Let (q(t), p(t)) denote the solution of (10.21) that originates at (q0 , p0 ) at t = 0. Then q(t) evolves on a line through the origin if and only if q0 and p0 are colinear, that is, whenever L = 0. In the case that q0 and p0 are not colinear, then q(t) remains in the plane P spanned by q0 and p0 , where it traces an ellipse when E(q0 , p0 ) < 0, a parabola when E(q0 , p0 ) = 0 and a hyperbola when E(q0 , p0 ) > 0. Proof If q0 and p0 are colinear then L = 0 and therefore q(t) and p(t) are colinear. But then F = − e, and hence, e is constant. To prove the converse reverse the steps. Assume now that q0 and p0 are not colinear and let P denote the plane spanned by q0 and p0 . Since L is constant, q(t) and p(t) are not colinear for all t. It follows that for any x in the kernel of L both x · q(t) and x · p(t) are equal to zero. But this implies that both q(t) and p(t) are in P for all t. Evidently the Runge–Lenz vector is in P. Let φ denote the angle between F and q. Then, L2 − q = F · q = Fq cos φ, and therefore q =

L2 . 1 + F cos φ

(10.22)

It follows that q(t) traces an ellipse when F < 1, a parabola when F = 1, and a hyperbola when F > 1. Since F2 = 2L2 E + 1, E < 0 if and only if F < 1, E = 0 if and only if F = 1, and E > 1 if and only if F > 1. With this background at our disposal let us now connect Kepler to the affine Hamiltonians. Let us first establish some additional notations that will facilitate the transition to equations (10.21). We will use S (h) to denote the Euclidean

184 10 Cotangent bundles of homogeneous spaces as coadjoint orbits sphere ||x||2 = h2 for > 0 and the hyperboloid (x, x) = h2 , x0 > 0 when < 0. The cotangent bundle of S (h) will be identified with the tangent bundle via the quadratic form ( , ) . Points of the tangent bundle will be represented by the pairs (x, y) ∈ Rn+1 × Rn+1 that satisfy ||x|| = h2 and (x, y) = 0. For each x ∈ S (h) the stereographic projection p in Rn is given by λ(x − he0 ) + he0 = (0, p) with λ =

h . h − x0

The inverse map x = (p) is given by x0 =

h(||p||2 − εh2 ) 2h2 , and x ¯ = (x , . . . , x ) = p. 1 n ||p||2 + εh2 ||p||2 + εh2

(10.23)

It follows that Rn ∪ {∞} is the range under the stereographic projection of S (h) in the Euclidean case and the annulus {p ∈ Rn : ||p||2 > h2 } ∪ {∞} in the hyperbolic case. ∂xi denote the Jacobian matrix with the entries ∂p ,i = Let ∂ ∂p j 0, . . . , n, j = 1, . . . , n. Then, ∂ dp = dx = ∂p 4εh3 2h2 4h2 p · dp p · dp, dp − p , (10.24) (||p||2 + h2 )2 ||p||2 + εh2 (||p||2 + h2 )2 that is, dx0 =

4εh3 2h2 4h2 p · dp p · dp and d x ¯ = dp − p. (||p||2 + εh2 )2 ||p||2 + εh2 (||p||2 + εh2 )2

Assume that the cotangent bundle of Rn is identified with its tangent bundle × Rn via the Euclidean inner product ( · ), and let (p, q) denote the points of Rn × Rn . Let q = (x, y) denote the mapping such that

Rn

(dx, y) = (dp · (x, y)),

(10.25)

for all (x, y) with x ∈ S and (x, y) = 0. It follows that % & n n n n n ∂x0 ∂xi ∂x0 ∂xi y0 dpj + yi dpj = + yi y0 dpj ∂pj ∂pj ∂pj ∂pj j=1

i=1 j=1

=

j=1 n j=1

i=1

qj dpj = q · dp.

10.3 Degenerate case and Kepler’s problem

185

Therefore, ∂xi ∂x0 y0 + yi , j = 1, . . . , n. ∂pj ∂pj n

qj =

(10.26)

i=1

Then (10.24) yields 2h2 q= ||p||2 + h2

2hy0 2(¯y · p) p + y¯ − p , y¯ = (y1 , . . . , yn ). ||p||2 + h2 ||p||2 + h2 (10.27)

Hence, ||p||2 + h2 2hy0 ||p||2 − h2 2 (q · p) = ||p|| − (¯y · p). 2h2 ||p||2 || + h2 ||p||2 + h2

(10.28)

y0 Since y is orthogonal to x, (¯y · p) = − 2h (||p||2 − h2 ). Therefore,

||p||2 + εh2 1 q·p q · p, y¯ = q − 2 p. (10.29) 2 h 2h h The transformation (p, q) ∈ Rn × Rn → (x, y) ∈ TS (h) is a symplectomorphism since it pulls back the Liouville form (dx, y) on S (h) onto the Liouville form (dp·q) in Rn (symplectic forms are the exterior derivative of the Liouville forms). It then follows from (10.24) and (10.29) that y0 =

||dx||2ε

4h4 ε (||p||2 + εh2 )2 ||dp||2 , ||y||2ε = ε ||q||2 . 2 2 2 (||p|| + εh ) 4h4

(10.30)

To pass to the problem of Kepler, write the Hamiltonian H = 12 ||x||2ε ||y||2ε in the variables (p, q). It follows that 1 (||p||2 + εh2 )2 1 2 (||p||2 + εh2 )2 2 h ε ε ||q|| = ||q||2 . 2 2 4h4 4h2 The corresponding flow is given by H=

∂H (||p||2 + εh2 )2 dq ∂H ||p||2 + εh2 dp = =ε = − = −ε q, ||q||2 p. ds ∂q ds ∂p 4h2 2h2 ) ||q||2 = 1 and the preceding equations On energy level H = 2hε 2 , (||p|| +εh 4 reduce to q ||q|| dq dp =ε 2 = −ε 2 p. , (10.31) 2 ds h ||q|| ds h s After the reparametrization t = − −ε 2 0 ||q(τ )||dτ , equations (10.31) become h 2

2 2

dp dp ds q dq dq ds = =− = = p. , 3 dt ds dt ds dt ||q|| ds

186 10 Cotangent bundles of homogeneous spaces as coadjoint orbits

Since

(||p||2 +εh2 )2 ||q||2 4

= 1,

1 1 1 ||p||2 − = (||p||2 ||q|| − 2) 2 ||q|| 2||q|| 1 1 (2 − εh2 ||q|| − 2) = − εh2 . = 2||q|| 2

E=

(10.32)

So E < 0 in the spherical case, and E > 0 in the hyperbolic case. The Euclidean case E = 0 can be obtained by a limiting argument in which ε is regarded as a continuous parameter which tends to zero. 1 ¯ It To explain in more detail, let w0 = lim→0 x0 and w = lim→0 2h 2 x. follows that (10.23) and (10.24) have limiting values w0 = h, w =

1 1 p · dp p, dw0 = 0, dw = dp − 2 p. 2 2 ||p|| ||p|| ||p||4

The transformation p → w with w = ||p||2

1 p ||p||2

(10.33)

is the inversion about the circle

= 1 in the affine hyperplane w0 = h, and ||dw||2 =

1 ||dp||2 ||p||4

is the

corresponding transformation of the Euclidean metric ||dp||2 . The Hamiltonian H0 associated with this metric is equal to

1 ||p||4 2 2 4 ||q||

. This Hamiltonian can

) also be obtained as the limit of ( h ) 12 (||p||4h+h ||q||2 when → 0. On energy 2 1 2 level H = 2 , ||p|| ||q|| = 2, and therefore E = 0 by the calculation in (10.32). Of course, the solutions of (10.17) tend to the Euclidean geodesics as tends 2 to zero. Consequently, w(t) = lim→0 2h12 ε (x(t)) ¯ is a solution of ddtw 2 = 0, and hence, is a geodesic corresponding to the standard Euclidean metric. q The angular momentum L = q∧p and the Runge–Lenz vector F = Lp− ||q|| for Kepler’s problem correspond to the moment map (10.18) according to the following proposition. 2

Proposition 10.12 H = 2h 2 ,

2

2 2

Let x = x0 e0 + x¯ and y = y0 e0 + y¯ . On energy level

¯ and F = h(y0 (e0 ∧ x) ¯ ε − x0 (e0 ∧ y¯ )ε )e0 . L = (¯y ∧ x) Proof

If we identify x¯ and y¯ with their projections on Rn , then x¯ =

2h2 ||p||2 + h2 (q · p) p p, y ¯ = q− 2 2 2 h ||p|| + h 2h

according to (10.23) and (10.29). Therefore, 2h2 ||p||2 + h2 (q · p) (¯y ∧ x) p∧ ¯ = q− p = q ∧ p = L. h 2h2 ||p||2 + h2

10.4 Mechanical problem of C. Newmann

187

Then, h(y0 (e0 ∧ x) ¯ ε − x0 (e0 ∧ y¯ )ε )e0 = h(x0 y¯ − y0 x) ¯ 2 2 h (||p||2 − εh2 ) 1 2εh p+ = −h q·p h ||p||2 + εh2 ||p||2 + εh2 2 2 ||p|| + εh q·p × q− 2 p 2 2h h ||p||2 − εh2 ||p||2 − εh2 q = (q ∧ p) p − ||p||2 q + q 2 2 ||p||2 + h2 = (q ∧ p)p − q. 2

= −(q · p)p +

The preceding expression reduces to F on H =

2h2

because

(||p||2 +εh2 ) ||q|| = 1. 2

To make the correspondence complete, let us mention that the stereographic projections of the great circles on the sphere trace ellipses in the plane spanned by the initial data q0 and p0 , while the projections of the hyperbolas on the hyperboloid trace hyperbolas in the same plane. The Euclidean geodesics trace parabolas. We will leave these details to the reader.

10.4 Mechanical problem of C. Newmann Let us now return to the full affine Hamiltonian H = 12 Lh , Lh + A, Lp with A a regular element in p . When = 1 then p is equal to the space symmetric matrices relative to the Euclidean inner product. Every symmetric matrix is conjugate to a diagonal matrix with real entries. Then A is regular if and only if all the diagonal entries are distinct, that is, if and only if A has distinct eigenvalues. For = −1, A is symmetric relative to the Lorentzianquadratic form x0 y0 − n a0 −aT , where i=1 xi yi . Then every matrix A in p is of the form a A0 0 −aT can be a ∈ Rn and A0 is a symmetric n × n matrix. The matrix a 0n written as a ∧ e0 , where now a is embedded in Rn+1 with a0 = 0. Hence, a0 0 , a0 a every matrix in p is of the form A = a ∧ e0 + S with S = 0 A0 real number, and A0 a symmetric n × n matrix. There are two distinct possibiities. Either a is zero, in which case A0 can be diagonalized. In this situation, A is regular if the eigenvalues of A0 are distinct

188 10 Cotangent bundles of homogeneous spaces as coadjoint orbits and different from a0 . Or, a = 0. In this situation A cannot be conjugated to a diagonal matrix by the elements of SO(1, n), because g(a ∧ e0 + S)g−1 = g(a) ∧ g(e0 ) + gSg−1 = b(g) ∧ e0 + S(g) with b(g) = (g(a) ∧ g(e0 ))e0 . Hence, b(g) = 0 for any g ∈ SO(1, n). The above implies that regular elements to different conjugacy may belong D0 0 , where D0 is a 2 × 2 classes. For instance, the matrix A = 0 D α −β matrix with α and β are real numbers and D a diagonal β α (n − 1) × (n − 1) matrix with non-zero diagonal real entries, is regular whenever α 2 + β 2 = 0 and the remaining diagonal entries are all distinct, but is not conjugate to any diagonal matrix. Remarkably on the coadjoint orbits through rank one matrices, the restriction of H is completely integrable, and the required integrals of motion are easily obtained from the spectral representation Lλ = Lp − λLk + λ2 A. The text below contains the necessary details. Consider first the coadjoint orbits in g . It follows from Section 10.2 that ' ||x||2 2 2 I, ||x|| = ||x0 || , Lh = {x ∧ y : (x, y) = 0}. Lp = x ⊗ x − n+1 The zero trace requirement is inessential for the calculations below and will be disregarded. Additionally, A will be replaced by −A and Lλ will be rescaled by dividing by −λ2 to read Lλ = − λ12 Lp + λ1 Lh + A. The spectrum of Lλ is then given by 1 1 −1 − 2 Lp + Lh . 0 = Det(zI − Lλ ) = Det(zI − A)Det I − (zI − A) λ λ It follows that on Det(zI − A) = 0, 0 = Det(zI − Lλ ) whenever 1 1 . 0 = Det I − (zI − A)−1 − 2 Lp + Lk λ λ Matrix M = I − (zI − A)−1 − λ12 Lp + λ1 Lk is of the form M=I+

1 1 Rz x ⊗ x − (Rz x ⊗ y − Rz y ⊗ x), Rz = (zI − A)−1 . λ λ2

Lemma 10.13 Det(M) =

1 ((Rz x, x) + (Rz x, x) (Rz y, y) − (Rz x, y)2 ) + 1. λ2

10.4 Mechanical problem of C. Newmann

189

Proof Let V be the linear span of {x, y, Rz x, Rz y} and let V ⊥ denote the orthogonal complement relative to ( , ) . Since M(V) ⊆ V and M is equal to the identity on V ⊥ , the determinant of M is equal to the determinant of the restriction of M to V. Then 1 1 1 ||x||2 Rz x + ||x||2 Rz y, My = y − ||y||2 Rz x, λ λ λ2 1 1 1 MRz x = Rz x + 2 (Rz x, x) Rz x − (Rz x, y) Rz x + (Rz x, x) Rz y, λ λ λ 1 1 1 MRz y = Rz y + 2 (Rz x, y) Rz x − (Rz x, y) Rz x + (Rz x, x) Rz y. λ λ λ

Mx = x +

The corresponding matrix is given by ⎛ 1 0 0 ⎜ 0 1 0 ⎜1 ⎝ 2 ||x||2 − 1 ||y||2 1 + 12 (Rz x, x) − 1 (Rz x, y) λ λ λ λ 1 1 2 0 λ ||x|| λ (Rz x, x)

⎞ 0 ⎟ 0 ⎟. 1 1 (Rz x, y) − λ (Rz y, y)⎠ λ2 1 + λ1 (Rz x, y)

The determinant of this matrix is equal to 1 1 1 1 + 2 (Rz x, x) − (Rz x, y) (Rz x, y) + 1 λ λ λ 1 1 1 (Rz x, y) − (Rz y, y) − (Rz x, x) λ λ λ2 1 = 2 ((Rz x, x) + (Rz x, x) (Rz y, y) − (Rz x, y)2 ) + 1. λ Lemma 10.14 Function F(z) = (Rz x, x) + (Rz x, x) (Rz y, y) − (Rz x, y)2 , z ∈ R is an integral of motion for H. It follows from above that Det(I − Wz ) = 0 if and only if F(z) = −λ2 . −1 Since z(zI − A)−1 = I − 1z A , limz→±∞ (z(zI − A)−1 = I. Therefore, Proof

lim zF(z) = ||x||2 ||y||2 + ||x||2 .

z→±∞

This argument shows that F(z) takes negative values on some open interval. Hence, F(z) is an integral of motion on that interval. But then F(z) is an integral of motion for all z since F is an analytic function of z on Det(zI − A) = 0. Function F is a rational function with poles at the eigenvalues of the matrix A. Hence, F(z) is an integral of motion for H if and only if the residues of F are constants of motion for H.

190 10 Cotangent bundles of homogeneous spaces as coadjoint orbits

In the Euclidean case, the eigenvalues of A are real and distinct, since A is symmetric and regular. Hence, there is no loss in generality in assuming that A is diagonal. Let α1 , . . . , αn+1 denote its diagonal entries. Then F(z) =

n k=0

Fk , z − αk

where F0,..., Fn denote the residues of F. It follows that F(z) =

n k=0

=

n k=0

−2

&2 % n n n xk yk xk2 y2j xk2 + − z − αk (z − αk )(z − αj ) z − αk k=0 j=0

k=0

n n xk2 y2j xk2 + z − αk (z − αk )(z − αj ) k=0 j=0, j =k

n

n

k=0 j=0, j =k

xk yk xj yj . (z − αk )(z − αj )

(10.34)

Hence, Fk = lim (z − αk )F(z) z→αk

= xk2 +

n xj2 yk + xk2 y2j j=0, j =k

= xk2 +

n j=0, j =k

(αk − αj )

−2

n j=0, j =k

xk yk xj yj (αk − αj )

(xj yk − xk yj )2 , k = 0, . . . , n. (αk − αj )

The preceding calculation yields the following proposition. (x y −x y )2 Proposition 10.15 Each residue Fk = xk2 + nj=0, j =k j(αkk −αk j )j , k = 0, . . . , n is an integral of motion for the Newmann’s spherical system (10.16) with = 1. Moreover, functions F0 , . . . , Fn are in involution. Proof The Poisson bracket relative to the orbit structure coincides with the canonical Poisson bracket on Rn+1 × Rn+1 . These results coincide with the ones reported in [Ms2; Rt2], but our derivation and the connection with the affine-quadratic problem is original. Similar results hold on the hyperboloid. When A is a diagonal matrix with distinct diagonal entries, then (Rz x, y) =

n vi wi , z − αi i=0

10.4 Mechanical problem of C. Newmann

191

where v0 = x0 , w0 = y0 and vi = ixi , wi = iyi , i = 1. . . . , n Therefore, the residues Fk are exactly as in (10.34) with x and y replaced by v and w. Hence, Fk = v2k +

n (vj wk − vk wj )2 , k = 0, . . . , n (αk − αj )

j=0, j =k

are the integrals of motion Newmann’s problem on the for the hyperboloid. D0 0 α −β In the case that A = with D0 = and D a diagonal 0 D β α with diagonal matrix with diagonal entries α2 , . . . , αn , then define 1 1 v0 = √ (x0 + ix1 ), v1 = √ (x0 − ix1 ), vi = ixi , i = 2 . . . , n, 2 2 1 1 w0 = √ (y0 + iy1 ), w1 = √ (y0 − iy1 ), wi = iyi , i = 2, . . . , n. 2 2 Then,

(Rz x, y)−1 = (z − A)−1 x, y

−1

1 = (z − α)(x0 y0− x1 y1 ) − β(x0 y1+ x1 y0 )) (z − α 2 ) + β 2 n 1 − xj yj z − αj j=2

1 1 1 v0 w0 + v1 w1 + vj wj z − (α + iβ) z − (α − iβ) z − αj n

=

j=2

=

n j=0

1 vj wj , z − αj

provided that we identify α0 = α + iβ and α1 = α − iβ. Therefore, the spectral function F(z) has the same form as in the spherical case: % n &2 n n n vk wk v2k w2j v2k − + . F(z) = z − αk (z − αk )(z − αj ) z − αk k=0

k=0 j=0

k=0

It follows that F(z) =

n k=0

n (vj wk − vk wj )2 Fk , k = 0, . . . , n . , Fk = v2k + z − αk (αk − αj ) j=0, j =k

192 10 Cotangent bundles of homogeneous spaces as coadjoint orbits An easy calculation shows that F¯ 0 = F1 and that each Fk , k ≥ 2 is real valued. Therefore, Re(F0 ), Im(F0 ), F2 , . . . , Fn

(10.35)

are n independent integrals of motion. The integration technique based on the use of elliptic coordinates is intimately tied to another famous problem in the theory of integrable systems, Jacobi’s geodesic problem on the ellipsoid. We will defer this connection to the next chapter. In the meantime we would like to contrast this class of integrable systems with another famous integrable system, the Toda system.

10.5 The group of upper triangular matrices and Toda lattices A Toda lattice is a system of n particles on the line in motion under an exponential interaction between its nearest neighbors. This mechanical system is at the core of the literature on integrable systems, partly for historical reasons, as the first mechanical system whose integrals of motion were found on a coadjoint orbit of a Lie algebra, but mostly because of the methodology which seemed to carry over to other Lie algebras with remarkable success [Pr; RS]. Its magical power led to a new paradigm in the theory of integrable systems based on the use of R-matrices. This paradigm provides an abstract procedure to construct large families of functions whose members are in involution with each other, and is looked upon in much of the current literature as an indispensable tool for discovering integrable systems [Pr; RT]. Our approach to the Toda system is completely different and makes no use of double Lie algebras. Rather than starting with a particular Hamiltonian on the space of symmetric matrices, as is commonly done in the literature on Toda systems, we will start with a geodesic problem on the group G of upper triangular matrices, and along the way discover the integrals of motion for a Toda lattice. The basic setting is as follows: G will denote the subgroup of matrices in SLn (R) consisting of matrices g = (gij ) such that gij = 0 for i < j, and g will denote its Lie algebra, i.e., the algebra of all upper triangular matrices of trace zero. Then A, B = Tr(ABT ) is positive-definite on g, since A, A = ni≥j a2ij for any matrix A = (aij ) in g. We will use g0 to denote the space of upper triangular matrices having zero diagonal part. Then g0 is an ideal in g, and its orthogonal complement g⊥ 0 in g consists of all diagonal matrices having zero trace.

10.5 Upper triangular matrices and Toda lattices

193

In order to make contact with Toda lattices, we will introduce the weighted quadratic form (A, B) = DA , DB + 12 A0 , B0 , where DA and DB are the diagonal parts of A and B, and A0 and B0 are the projections of A and B onto g0 . Then ||A||2 will denote the induced norm (A, A). This T norm can be used T to define the length 0 ||U(t)|| dt and the energy E = 12 0 ||U(t)||2 dt of any curve g(t) in G where U(t) = g−1 (t) dg dt (t). The associated Riemannian problem is left-invariant, hence the corresponding Hamiltonian is a function on g∗ when T ∗ G is represented as G × g∗ . A distinctive feature of this set-up, however, is that g∗ can be identified with the space of symmetric matrices in sln (R). The identification depends on the choice of a complementary subalgebra in sln (R) and goes as follows. Any upper triangular matrix U can be expressed as the sum of a symmetric and a skew-symmetric matrix. In fact, U=

$ 1# $ 1# U + UT + U − UT . 2 2

Conversely, any symmetric matrix S defines an uper triangular matrix S+ with T = its entries equal to sij for i ≤ j, and zero otherwise. It follows that S+ + S+ 1 T S−Sd , where Sd denotes the diagonal part of S. Therefore, S+ = 2 (S+ +S+ )+ 1 T 2 (S+ − S+ ) can be written as 2S+ = S − Sd + (S+ − S− ),

(10.36)

T . These decompositions show that sl (R) = so (R)⊕g, because where S− = S+ n n any matrix in sln (R) is the sum of a symmetric and a skew-symmetric matrix. The decomposition sln (R) = son (R)⊕g induces a decomposition of the dual sln (R)∗ in terms of the annihilators so0n (R) = { ∈ sln (R)∗ : |son (R) = 0} and g0 = { ∈ sln (R)∗ : |g = 0}. Then g∗ will be identified with so0n (R), and so∗n (R) with g0 . When sln∗ (R) is identified with sln (R) via the trace form , , then so0n (R) is identified with matrices L ∈ sln (R) such that L, X = 0 for all X ∈ son (R). That is, g∗ is identified with the orthogonal complement of sln (R) relative to the trace form, which is exactly the space of symmetric matrices having zero trace.

Remark 10.16 The above reasoning shows that the dual of any Lie algebra which is transversal to son (R), such as the Lie algebra of lower triangular matrices, can be represented by the symmetric matrices. Hence the identification of g∗ with the symmetric matrices is not canonical. In this identification the Hamiltonian lift of any left-invariant vector field XA (g) = gA is identified with the symmetric matrix SA = 12 (A + AT ) via the

194 10 Cotangent bundles of homogeneous spaces as coadjoint orbits formula hA (ξ ) = S, 12 (A + AT ). In particular, hij = hEij is identified with the the matrix Sij = 12 (Eij + Eji ) and hD is identified with D, where D is a diagonal matrix and Eij = ei ⊗ ej , i < j, i = 1, . . . , n, j = 1, . . . , n. Matrices Eij , i < j forms an orthonormal basis for g0 relative to , . Let D1 , . . . , Dn−1 denote any ⊥ orthonormal basis in g⊥ 0 (the two forms agree on g0 ). Then any curve g(t) ∈ G is a solution of ⎛ ⎞ n−1 n dg = g(t) ⎝ ui (t)Di + uij (t)Eij ⎠ (10.37) dt i = gF?for some Hermitian 1 L dV ds for an arbitrary matrix F(s) that satisfies df (V) = i 0 (s), dF ds , ds tangential direction V(s). The above is equivalent to L -

/ 0 . d2 1 dF dV (s), , ds = 0. + i ds ds ds2

(17.21)

0

Since the mapping U → i[, U] is bijective on the space of Hermitian matrices orthogonal to (s), there exists a Hermitian matrix U(s) such that U(0) = 0 and dV ds = i[(s), U(s)] Then equation (17.21) becomes 0 0 . 0 // L - / 2 d dF i , (s) , U(s) ds = 0. , (s) + (s), ds ds2 0

It follows that /

0 0 0 // d2 dF , (s) =0 , (s) + (s), ds ds2 > > ?? because U(s) is arbitrary. Since (s), (s), dF = dF ds ds , the above becomes i

/ i

0 d2 dF = 0. , (s) − ds ds2

It follows that Xf (g) = g(s)F(s) with F(s) = i

s/ 0

0 d2 (x), (x) dx. dx2

The integral curves t → g(s, t) of Xf are the solutions of the following equations: 0 s/ 2 d ∂g ∂g (s, t) = g(s, t)i (s, t) = g(s, t)(s, t). (x, t), (x, t) dx, 2 ∂t ∂s dx 0 (17.22)

388 17 Non-linear Schroedinger’s equation and Heisenberg’s equation D The equality of mixed partial derivatives dsg ∂g ∂t = the matrices (s, t) evolve according to / 2 0 ∂ ∂ (s, t) = i , (s, t) . ∂t ∂s2

Dg dt

∂g ∂s

implies that

To prove the analogous formula in the spherical case, let Y(s, t) denote a family of anchored horizontal-Darboux curves such that Y(s, 0) = X(s) and v(s) = ∂Y ∂t (s, t)t=0 = X(s)V(s). Denote by Z(s, t) the matrices defined by ∂Y ∂s (s, t) = Y(s, t)Z(s, t). It follows that (s) = Z(s, 0), and that V(s) is the solution of dV ds = [(s), V(s)] + U(s) (s, 0). with U(s) = ∂Z ∂t Then, . L∂Z ∂Z 1∂ (s, t), (s, t) ds|t=0 df (V) = 2 ∂t 0 ∂s ∂s . Ld dU (s), (s) ds = ds ds 0 .s=L . L- 2 d d (s), U(s) =− (s), U(s) ds + . ds ds2 0 s=0 Analogous to the hyperbolic case, the boundary terms vanish, and therefore . L- 2 d dfX (V) = − , U(s) ds. ds2 0 The Hamiltonian vector field Xf that corresponds to f is of the form Xf (X)(s) = X(s)F(s) for some curve F(s) ∈ h. Since Xf (X) ∈ TX PHDs (L), F(s) is the solution of dF (s) = [(s), F(s)] + Uf (s), F(0) = 0 ds for some curve Uf (s) ∈ h that satisfies Uf (0) = 0 and (s), Uf (s) = 0. The curve Uf (s) is determined by the symplectic form ω: L df (U) = − (s), [Uf (s), U(s)] ds,

(17.23)

0

where U(s) is an arbitrary curve in h that satisfies U(0) = 0 and (s), U(s) = 0. Equation (17.23) yields . L- 2 d − [(s), Uf (s)], U(s) ds = 0. (17.24) ds2 0

17.3 Geometric invariants of curves

389

If we write U(s) as U(s) = [(s), C(s)] for some curve C(s) that satisfies C(0) = 0, equation (17.24) becomes 0 . L -/ 2 d (s), (s) − [[(s), U (s)], ], C(s) ds = 0. f ds2 0 Since C(s) is arbitrary, /

0 d2 , − [[, Uf ], ] = 0. ds2 ? > 2 . The integral But then [[, Uf ], ] = Uf , and therefore Uf = − , dds 2 curves t → X(s, t) of Xf are the solutions of ∂X ∂X (s, t) = X(s, t)F(s, t), and = X(s, t)(s, t), ∂t ∂s > ? ∂2 where F(s, t) is the solution of ∂F (s, t) = [(s, t), F(s, t)]− (s, t), (s, t) . 2 ∂s ∂s Matrices F(s, t) and (s, t) satisfy the zero-curvature equation [, F] = 0. Combined with the above, this equation yields / 2 0 ∂ ∂ (s, t) = (s, t), (s, t) . ∂t ∂s2

∂ ∂t

−

∂F ∂s

+

Thus in both the hyperbolic and the spherical case (s, t) evolves according to the same equation: in the hyperbolic case is Hermitian, while in the spherical case is skew-Hermitian. To pass from the hyperbolic case to the spherical case multiply (s, t) in equation (17.19) by i. We leave it to the reader to show that in the Euclidean case the integral curves g(s, t) of the Hamiltonian flow Xf evolve according to / 2 0 ∂ ∂ ∂g (s, t) = g(s, t)(s, t), (s, t) = i , (s, t) , ∂s ∂t ∂s2 exactly as in the hyperbolic case, except that the curves g(s, t) now evolve in the semi-direct product SH (p). Equations (17.19) and (17.20) will be referred to as Heisenberg’s magnetic equation (HME) (17.1). Equation (17.20), when expressed in terms of the coordinates λ(s, t) of (s, t) relative to the basis of Hermitian Pauli matrices, takes on the following form: ∂ 2λ ∂λ (s, t) = λ(s, t) × 2 (s, t), ∂t ∂s

(17.25)

390 17 Non-linear Schroedinger’s equation and Heisenberg’s equation

while (17.19) becomes ∂ 2λ ∂λ (s, t) = 2 (s, t) × λ(s, t) ∂t ∂s when is expressed in the coordinates relative to the skew Hermitian basis A1 , A2 , A3 . The reason is simple: if X = x1 B1 +x2 B2 +x3 B3 and Y = y1 B1 +y2 B2 +y3 B3 then i[X, Y] = z1 B1 + z2 B2 + z3 B3 with z = y × x. But if X = x1 A1 + x2 A +2 +x3 A3 and Y = y1 A1 + y2 A2 + y3 A3 then [X, Y] = z1 A1 + z2 A2 + z3 A3 with z = x × y. Both of these assertions can be verified through the relations in Table 17.1. Equation (17.22) is referred to as the continuous isotropic Heisenberg ferromagnetic model in [Fa, Part II, Ch.1]. This equation is related to the filament equation [AKh] ∂γ (s, t) = κ(s, t)B(s, t). (17.26) ∂t For, when the solution curves of the filament equation are restricted tocurves ∂γ parametrized by arc-length, i.e., to curves γ (s, t) such that ∂s (s, t) = 1, then ∂γ ∂T ∂ 2γ (t, s), and (s, t) = κ(s, t)N(s, t) = 2 (s, t). (17.27) ∂s ∂s ∂s It then follows that in the space of arc-length parametrized curves the solutions of the filament equation coincide with the solutions of T(t, s) =

∂γ ∂ 2γ ∂γ = × 2, (17.28) ∂t ∂s ∂s because B(s, t) = T(s, t) × N(s, t). For each solution curve γ (s, t) of (17.28) the tangent vector T(s, t) satisfies ∂ 2T ∂T =T× 2 ∂t ∂s as can be easily verified by differentiating with respect to s. But then T(s, t) may be interpreted as the coordinate vector of (s, t) relative to an orthonormal basis in either h or p, which brings us back to the Heisenberg’s magnetic equation. To link the HME (17.1) to the non-linear Schroedinger’s equation, we will now turn our attention to the matrices R(s, t) in SU2 defined by the relations

17.3 Geometric invariants of curves

391

(s, t) = R(s, t)B1 R∗ (s, t) in the hyperbolic, and (s, t) = R(s, t)A1 R∗ (s, t) in the spherical case. The field of curves R(s, t) then evolve according to ∂R ∂R (s, t) = R(s, t)U(s, t), and = R(s, t)V(s, t) ∂s ∂t for some matrices U(s, t) and V(s, t) in h. Since the matrices U(s, t) and V(s, t) are generators of the same field of curves, they must satisfy the zero-curvature equation ∂V ∂U (s, t) − (s, t) + [U(s, t), V(s, t)] = 0, ∂t ∂s

(17.29)

as stated by Lemma 17.9 following Proposition 17.8. Moreover, V(0, t) = 0 for all t, because the horizontal-Darboux curves are anchored at s = 0. Proposition 17.16 Let R−1 (s, t) ∂R uj Aj be such that ∂s (s, t) = U(s, t) = either (s, t)) = R∗ (s, t)B1 R(s, t) or (s, t) = R∗ (s, t)A1 R(s, t) is a solution of the approprite HME. If u(s, t) = u2 (s, t) + iu3 (s, t), then, s u1 (x, t) dx ψ(s, t) = u(s, t) exp i 0

is a solution of the non-linear Schroedinger’s equation ∂ 2ψ ∂ 1 1 |ψ(s, t)|2 + c ψ(s, t) with c(t) = − |u(0, t)|2 . ψ(s, t) = i 2 (s, t) + i ∂t 2 2 ∂s (17.30) In the proof below we will make use of the following formulas: [A, [A, B]] = A, BA − A, AB, A ∈ p, B ∈ p, [[A, B], B] = B, BA − A, BB, A ∈ h, B ∈ p.

(17.31)

They can be easily verified by direct calculation. Proof The proof of the theorem will be done for the hyperbolic case first. Then we will point to the modifications required for the proof in the spherical case. Since (s, t) = R(s, t)B1 R∗ (s, t), then ∂ ∂ ∂ = (R(s, t)B1 R∗ (s, t)) = R[B1 , V]R∗ , = R[B1 , U]R∗ ∂t ∂t ∂s and

0 / ∂ 2 ∂U R∗ . = R [[B1 , U], U] + B1 , ∂s ∂s2

392 17 Non-linear Schroedinger’s equation and Heisenberg’s equation

The fact that (s, t) evolves according to Heisenberg’s magnetic equation implies that 0 0 // ∂U (17.32) , B1 [B1 , V] = i [[[B1 , U], U], B1 ] + B1 , ∂s Relations (17.31) imply that [[B1 , U], U] = U, B1 U − U, UB1 = −(u22 + u23 )B1 + u1 u2 B2 + B3 u1 u3 , hence [[[B1 , U], U], B1 ] = u1 u3 A2 − u1 u2 A3 . Similarly, // 0 0 0 / ∂U ∂u2 ∂U ∂u2 ∂u3 ∂u3 B1 , = B2 − B3 , and , B1 = − A3 − A2 . B1 , ∂s ∂s ∂s ∂s ∂s ∂s Therefore equation (17.32) reduces to ∂u3 ∂u2 A3 − A2 [B1 , V] = i u1 (u3 A2 − u2 A3 ) − ∂s ∂s ∂u3 ∂u2 B3 + B2 . = −u1 (u3 B2 − u2 B3 ) + ∂s ∂s If V = v1 A1 +v2 A2 +v3 A3 , then [B1 , V] = v3 B2 −v2 B3 , which, when combined ∂u2 3 with the above, yields v2 = −u1 u2 − ∂u ∂s , and v3 = −u1 u3 + ∂s . These relations can be written more succinctly as v(s, t) = −u1 (s, t)u(s, t) + i

∂u (s, t), ∂s

where u = u2 + iu3 and v = v2 + iv3 . The zero curvature equation implies that ∂v1 1 ∂ 2 ∂u1 = + u2 + u23 , ∂t ∂s 2 ∂s 2 ∂ u ∂u ∂u ∂u1 = i 2 − 2u1 − u − i(v1 + u21 )u. ∂t ∂s ∂s ∂s The first equation in (17.33) implies that ∂ ∂t

s

(17.33)

1 u1 (x, t)dx = v1 (s, t) + (u22 (s, t) + u23 (s, t)) + c(t), 2

0

where c(t) = −v1 (0, t) − 12 (u22 (0, t) + u23 (0, t)) = − 12 (u22 (0, t) + u23 (0, t)), since s ∂ u1 (x, t)dx − 12 |u(s, t)|2 − c into the V(0, t) = 0. Substituting v1 (s, t) = ∂t 0

second equation in (17.33) leads to

17.3 Geometric invariants of curves

∂u ∂ + iu ∂t ∂t

s

393

∂u ∂ 2u ∂u1 1 2 2 u1 (t, x) dx = i 2 − 2u1 −u − i − |u| − c + u1 u. ∂s ∂s 2 ∂s (17.34)

After multiplying by exp (i

s 0

u1 (x, t) dx), equation (17.34) can be expressed as

2 ∂ ∂ u ∂u1 ∂u 1 ψ(s, t) = i 2 − 2u1 −u − i u21 − u|2 − c u ∂t ∂s ∂s 2 ∂s s u1 (x, t) dx , (17.35) × exp i 0

s where ψ(s, t) = u(s, t) exp (i 0 u1 (x, t) dx). s $ # ∂u It follows that ∂ψ ∂s = ∂s + iuu1 exp (i 0 u1 (x, t dx), and therefore, ∂ 2ψ = ∂s2

s ∂u ∂ 2u ∂u1 2 + 2iu u exp i u (x, t) dx , + iu − u 1 1 1 ∂s ∂s ∂s2 0

or i

2 s ∂u1 ∂u ∂ u ∂ 2ψ 2 − u − iu = i − 2u u exp i u (x, t dx , 1 1 1 ∂s ∂s ∂s2 ∂s2 0

(17.36)

Combined with (17.35), equation (17.36) yields ∂ ∂ 2ψ 1 2 |ψ| + c(t) ψ. ψ(t, s) = i 2 + i ∂t 2 ∂s In the spherical case equation (17.32) is replaced by 0 0 // ∂U [A1 , V] = [[[A1 , U], U], A1 ] + A1 , , A1 . ∂s The substitution A1 = iB1 brings us back to equation (17.32). Therefore the calculations that led to the non-linear Schroedinger equation in the hyperbolic case are equally valid in the spherical case, with the same end result. The steps taken in the passage from the Heisenberg equation to the Schroedinger equation are reversible. Any solution ψ(s, t) of (17.2) generates matrices ⎞ ⎛ & % 0 ψ i ∂ψ 1 −i( 12 |ψ|2 + c(t)) 1⎝ ∂s ⎠ and V = U= ¯ 2 2 i ∂∂sψ i( 12 |ψ|2 + c(t)) −ψ¯ 0

394 17 Non-linear Schroedinger’s equation and Heisenberg’s equation

that satisfy the zero-curvature equation. Therefore, there exist unique curves R(s, t) in SU2 with boundary conditions R(0, t) = I that evolve according to the differential equations: ∂R ∂R (s, t) = R(s, t)U(s, t), (s, t) = R(s, t)V(s, t). ∂s ∂t Such curves define (s, t) through the familiar formulas (s, t) = R(s, t)B1 R∗ (s, t) or (s, t) = R(s, t)A1 R∗ (s, t) depending on the case. In the first case, / 0 00 / / ∂ 2 ∂U i , 2 = R B1 , [[B1 , U], U] + B1 , R∗ = ∂s ∂s 00 / / ∂U 1 0 ius R∗ = R( R∗ . R B1 , B1 , iu¯s 0 ∂s 2 > ? 2 Therefore i , ∂∂s is equal to 2

∂ ∂t

= R[B1 , V]R∗ =

v = ius . The spherical case follows along similar lines.

1 2

0 −¯v

v 0

because

z 0

0 z¯

Remark 17.17 The above proposition reveals that SO2 = , ' |z| = 1 is a symmetry group for the non-linear Schroedinger equation, $ # s since ψ(s, t) = u(s, t) exp i 0 u1 (x, t) dx remains a solution for any u1 , independently of the symmetric space. This observation removes some mystery an ingenious observation # behind $ by H. Hasimoto [H1] that ψ = κ exp i τ dx of a curve γ (s, t) that satisfies the filament equation is a solution of the non-linear Schroedinger equation, for when R(s, t) is a Serret–Frenet frame then u1 = τ , u2 = 0, u3 = κ in the Euclidean and the hyperbolic case, while u1 = τ + 12 , u2 = 0, u3 = κ in the spherical case. s Hasimoto’s function κ(s, t) exp i 0 τ (x, t) dx) coincides with u(s, t) $ # s exp i 0 u1 (x, t) dx in the hyperbolic and the Euclidean case, but not in the spherical case. Of course, the most natural frame is the reduced frame u1 = 0 which bypasses these inessential connections with the torsion. The correspondence between the HME (17.1) and NLS (17.2) described by the above proposition strongly suggests that (17.1) is integrable, in the sense that there exist infinitely many functions in involution with each other that are constant along the solutions of (17.1), for it is well known that (17.2) is integrable (see, for instance [Fa; Mg; LP]). Its conserved quantities can be obtained either by an iterative procedure based on isospectral methods due to

17.3 Geometric invariants of curves

395

Shabat and Zakharov [ShZ], or alternatively, through a procedure known as the Magri scheme. The Magri scheme is applicable in situations in which there are two compatible Poisson structures { , }1 and { , }2 relative to which a given vector field X is Hamiltonian. This means that there exist two functions f and g such that X = f1 = g2 , where f1 is the Hamiltonian vector field of f relative to the first Poisson structure and g2 is the Hamiltonian vector field of g relative to the second Poisson structure. Then {f , g}1 and {f , g}2 are the constants of motion for X. Then the Hamiltonian vector field induced by these two function is also bi-Hamiltonian and generates two new integrals of motion. The iteration of this process produces a sequence of conserved quantities in involution with each other for the flow of X. It is entirely possible that Magri’s scheme could be used to prove the integrability of (17.1) by introducing another symplectic form on the space of anchored curves given by L , [V1 , V2 ] ds.

(V1 , V2 ) = 0

Such a form is mentioned elsewhere in the literature (see, for instance [AKh; Br; LP]) but never used for this specific purpose. We will not undertake this study here. Instead, we will just mention some functions that might be in this hierarchy of conserved quantities. ? 2 L > d Proposition 17.18 The Hamiltonian flow of f = −i 0 , d ds , ds2 ds is given by . 3 - 3 . ∂ ∂ ∂ ∂ 2 ∂ =2 . (17.37) − , − 3 , ∂t ∂s3 ∂s3 ∂s2 ∂s L Function f Poisson commutes with f0 = 12 0 κ 2 (s) ds. Proof Let V(s) be an arbitrary tangent vector at a frame-periodic horizontalDarboux curve g(s). Then the directional derivative of f in direction V is given by the following expression: 0 2 . L -/ ∂Z ∂ Z ∂ Z(s, t), (s, t) , 2 (s, t) ds|t=0 , df (V) = −i ∂t 0 ∂s ∂s where Z(s, t) denotes a field of Hermitian matrices such that Z(s, 0) = (s) and

∂Z dV (s, 0) = (s). ∂t ds

396 17 Non-linear Schroedinger’s equation and Heisenberg’s equation

It follows that

L

df (V) = −i

... ˙ + , ¨ [V, ˙ + , ¨ [, V] ˙ ] ¨ ds V , [, ]

0 L

¨ ], V ¨ ], ˙ V ¨ − [, ˙ ds 2[, . d ¨ ¨ ˙ 2 ([, ]) + [, ] , V˙ ds =i ds 0 L ... ¨ ], ˙ V ˙ ds, =i 2[, ]) + 3[, = −i

0 L-

0

where the dots indicate derivatives with respect to s. In the preceding calculations periodicity of is implicitly assumed to eliminate the boundary terms in the integration by parts. Let now V1 (s) denote the Hermitian matrix such that df (V) = ω (V1 , V) for all tangent vectors V. Then, L ... 1 L ¨ ], ˙ V ˙ ds = ˙ ds, 2[, ] + 3[, [, V˙1 ], V i i 0 0 which implies

L

... ¨ ], ˙ V ˙ ds = 0. [, V˙1 ] + 2[, ] + 3[,

0

When V˙ = [, C] the above becomes L ... ¨ ], ˙ ], C(s) ds = 0. [[, V˙1 ], ] + 2[[, ], ] + 3[[, 0

Since C(s) is an arbitrary curve with C(0) = 0 the preceeding integral equality reduces to ... ¨ ], ˙ ] = 0. [[, V˙1 ], ] + 2[[, ], ] + 3[[, The Lie bracket relations (17.31) imply that ... ... ¨ ˙ = 0. V˙1 + 2(, − ) + 3, Now it follows by the arguments used earlier in the paper that the Hamiltonian flow Xf satisfies ... ... ∂ ¨ . ˙ = 2( − , ) − 3, ∂t

17.3 Geometric invariants of curves

397

To prove the second part, we need to show that Lthe Poisson bracket of {f0 , f1 }, given by the formula ω (V0 (), V1 ()) = 1i 0 (s), [V˙ 0 (s), V˙1 (s)] ds with ... ... ¨ ] and V˙1 = −(2(, − ) + 3, ¨ ), ˙ is equal to 0. V˙ 0 () = i[, An easy calculation shows that ... ... ¨ − , ¨ , ) − 3, ¨ , ¨ . ˙ [V˙0 , V˙1 ] = i(2(, ) Hence,

{f0 , f } =

L

... ... ¨ − 2, , ¨ − 3, ¨ , ˙ ) ¨ ds. (2,

0

... ¨ = d , ¨ . ¨ The integral of the first term is zero because 2, ds Since ... d ¨ , ˙ , ¨ ¨ ¨ 2 − 2, 2, , = , ds ¨ , ˙ . ¨ But then the remaining integrand reduces to one term −, 1 d ˙ ˙ 2 ¨ , ˙ ¨ because , ˙ ˙ = −, . ¨ , = , 4 ds

The above functional L can be restated in terms of the curvature and the torsion of the base curve as 0 κ 2 (s)τ (s) ds for the following reasons. If T(s) = (s), then g(s)T(s) projects onto the tangent vector of a base curve γ ∈ H3 . Then N(s) and B(s), the matrices that correspond to the normal and the binormal vectors, are given by / 0 1 i d 1 d and B(s) = [T(s), N(s)] = − , . N= κ ds i κ ds According to the Serret–Frenet equations dN ds = −k + τ B. Therefore, . / 0. dN 1 dκ d 1 d2 1 d τ= , B = −i − 2 + , , ds κ ds2 κ ds κ ds ds -/ 0 2 . d d 1 , 2 , = −i 2 , ds κ ds ? 2 > d and hence, κ 2 τ = −i , d ds , ds2 . Proposition 17.19 Suppose that (s, t) = R(s, t)B1 R∗ (s, t) evolves according to the equation (17.37), where R(s, t) is the solution of ∂R 0 u(s, t) (s, t) = R(s, t) , R(0, t) = I. −u(s, ¯ t) 0 ∂s Then, u(s, t) is a solution of ∂u ∂ 3u ∂u − 3 |u|2 − 2 3 = 0. ∂t ∂s ∂s

(17.38)

398 17 Non-linear Schroedinger’s equation and Heisenberg’s equation

This proposition is proved by a calculation similar to the one used in the proof of Proposition 17.16, the details of which will be omitted. Equation (17.38) isknown as the modified L Korteweg-de Vries equation [Mg]. L Functions f0 = 0 k2 ds and f1 = 0 k2 τ ds also appear in the paper of C. Shabat and V. Zacharov [ShZ], but in a completely different context. The first two integrals of motion in the paper of Shabat and Zacharov are up to constant factors given by the following integrals: ∞ ∞ ˙¯ t)) − u(s, |u(s, t)|2 ds, C2 = (u(s, t))u(s, ¯ t)˙u(s, t)) ds, C1 = −∞

−∞

where C1 interpreted as the number of particles in the wave packet and C2 is interpreted as their momentum. To see that C1 and C2 are in exact correspondence with functions f0 and f1 assume that the Darboux curves are expressed by the reduced frames R(s), i.e., as the solutions of dR (s) = R(s)U(s) with U(s) = u2 (s)A2 + u3 (s)A3 . ds Then, f0 =

1 2

L

2 ˙ ||(s)|| ds =

0

Hence, C1 corresponds to 0

L

L 0

1 2

L

||[B1 , U(s)]||2 ds =

0

1 2

L

|u(s)|2 ds.

0

κ 2 (s) ds. Furthermore,

1 κ τ ds = 2i

2

L

˙¯ − u(s)˙ (u(s)u(s) ¯ u(s)) ds,

0

because ˙ ] ¨ = [[B1 , U], [B1 , U]], ˙ B1 = Im(¯uu˙ ), i, [, where Im(z) denotes the imaginary part of a complex number z. Therefore f1 corresponds to C2 . L In the language of mathematical physics, vector 0 (s) ds is called the total spin [Fa]. Here it appears as the moment map associated with the adjoint action of SU2 . It is a conserved quantity since the Hamiltonian is invariant under this action. This fact can be verified directly: 0 L/ 2 L L ∂ ∂ ∂ (t, s) ds = i (t, s) ds = , ds ∂t 0 ∂s2 0 ∂t 0 L ∂ ˙ ds = 0. [, ] =i 0 ∂s

17.4 Affine Hamiltonians and solitons

399

The third integral of motion C3 in [ShZ], called the energy, is given by & 2 ∞ % ∂u 1 4 (s, t) − |u(s, t)| ds. C3 = ∂s 4 −∞ It corresponds to the function L 5 (s) ¨ 2 − (s) ˙ 4 ds f2 = 4 0 L ∂κ 2 1 (s) + κ 2 (s)τ 2 (s) − κ 4 (s) ds. = ∂s 4 0 Functions C0 , f1 , f2 are in involution relative to the Poisson bracket induced by the symplectic form. There seems to be a hierarchy of functions that contains f0 , f1 , f2 such that any two functions in the hierarchy Poisson commute. For L instance, it can be also shown that f3 = 0 τ (s) ds is in this hierarchy and that its Hamiltonian vector field generates the curve shortening equation [Ep] ∂ ∂ (s, t) = (s, t) = κ(s, t)N(s, t). ∂t ∂s These findings are in exact correspondence with the ones discovered by J. Langer and R. Perline over the rapidly decreasing functions in R3 [LP].

17.4 Affine Hamiltonians and solitons For mechanical systems the Hamiltonian function represents the total energy of the system and its critical points correspond to the equilibrium configurations. In an infinite-dimensional setting the behavior of a Hamiltonian system at a critical point of a Hamiltonian does not lend itself to such simple characterizations. L We will now show that the critical points of the Hamiltonian f = 12 0 k2 ds over the horizontal Darboux curves are naturally identified with the affine Hamiltonian H=

1 2 (H + H32 ) + h1 2 2

(17.39)

generated by dg = g(s)(B1 + u2 (s)A2 + u3 (s)A3 ) ds L and the cost functional 12 0 (u22 (s) + u23 (s)) ds.

(17.40)

400 17 Non-linear Schroedinger’s equation and Heisenberg’s equation

To explain, we will need to go back to the parallel frames discussed earlier in the text. For simplicity of exposition we will confine our discussion to the hyperbolic Darboux curves. 0 Each anchored Darboux curve g0 (s) is a solution of dg ds = g0 (s)(s)), where (s) is a curve of Hermitian matrices such that |(s)|| = 1 and (0) = B1 . Every such matrix (s) can be represented by a curve R(s) in SU2 that satisfies (s) = R(s)B1 R∗ (s) and R(0) = I. Then g(s) = g0 R(s) is a solution of dg = g(s)((B1 + u1 (s)A1 + u2 A2 + u3 (s)3 )A3 ds in SL2 (C), for some real functions u1 (s), u2 (s), u3 (s), that satisfies g(0) = I. It projects onto the same base curve x(s) = g(s)g∗ (s) = g0 (s)g∗0 s). Each reduced Darboux curve g0 (s) defines a parallel frame v1 (s) = 2g0 (s)B1 g∗0 (s), v2 (s) = g0 (s)B2 g∗0 s, v3 (s) = g0 (s)B3 g∗0 (s) over the base curve x(s) = g0 (s)g∗0 (s), since v1 (s) =

dx ds

and

Dx (v2 ) = g0 (s)[B2 , u2 (s)A2 + u3 (s)A3 ]g∗0 (s) = u3 (s)v1 (s), ds Dx (v3 ) = g0 (s)[B3 , u2 (s)A2 + u3 (s)A3 ]g∗0 (s) = −u2 (s)v1 (s). ds L Therefore, the cost functional 12 0 (u22 (s) + u∗3 (s)) ds associated with (17.40) is L equal to 12 0 κ 2 (s) ds where κ(s) denotes the curvature of the base curve x(s). Therefore the affine Hamiltonian (17.39) is the Hamiltonian associated L with the optimal control problem of finding the minimum of 12 0 (u22 (s) + u23 (s)) ds over the trajectories in (34) that conform to the given boundary conditions (the Euler–Griffiths problem in Chapter 13). The variables h1 , h2 , h3 , H1 , H2 , H3 are the coordinates of an element ∈ g∗ relative to the dual basis B∗1 , B∗2 , B∗3 , A∗1 , A∗2 , A∗3 . Then, dL dg = g(s)(B1 + H2 (s)A2 + H3 (s)A3 ), = [dH(s), L(s)] ds ds

(17.41)

are the equations of the integral curves of H having identified g∗ with g via the trace form. Again, we will take advantage of the Cartan decomposition 0 u(s) 1 , where u(s) = H2 (s) + iH3 (s) g = p ⊕ h and write U = 2 −u(s) ¯ 0

17.4 Affine Hamiltonians and solitons

401

3 3 so that dH = B1 + U. Then P = i=1 hi B1 and Q = i=1 Hi Ai are the Hermitian and skew-Hermitian parts of L. It follows that 1 −iH1 −u 1 iw h1 ,Q = , P= u¯ iH1 2 −iw¯ −h1 2 where w = h2 + ih3 . Then equations (17.41) take on the familiar form dP dQ = [U(s), Q(s)] + [B1 , P(s)], = [U(s), P(s)] + [B1 , Q(s)], = −1. ds ds (17.42) This equation is formally the same as equation (8.7) in Chapter 8 and can be written in expanded form as dh1 dH1 du dw = 0, = iH1 u(s) − iw(s), (s) = iRe(w(s)u(s)), ds ds ds ds = i(h1 − )u(s). (17.43) Except for some minor notational details this equation is valid for the spherical and the Euclidean cases as well. We are now in the position to relate these this material to the solutions of (17.2). Proposition 17.20 Let u(s) and H1 be as in equation (37) and let H = 1 2 + ξ t) is a solution of the non-linear 2 ||u(s)|| + h1 . Then ψ(s, t) = u(s 2 ∂ψ ∂ ψ Schroedinger’s equation ∂t = −i ∂s2 + 12 |ψ|2 ψ precisely when ξ = −H1 and H = . Proof

If ψ(s, t) = u(s + ξ t) then ∂ 2ψ ∂ψ = iξ(H1 ψ − w) and = −H12 ψ + H1 w + (h1 − )ψ. ∂t ∂s2

Therefore,

∂ 2ψ 1 2 |ψ| + ψ 2 ∂s2 1 = ξ(H1 ψ − w) − −H12 ψ + H1 w + (h1 − )ψ + |ψ|2 ψ 2

0 = −i

∂ψ − ∂t

= ξ(H1 ψ − w) − (−H12 ψ + H1 w + (h1 − )ψ + ψ(H − h1 )) = −(ξ + H1 )w + (ξ H1 + H12 + − H)ψ, whenever ξ = −H1 and H = .

402 17 Non-linear Schroedinger’s equation and Heisenberg’s equation Thus the extremals which reside on energy level H = generate soliton solutions of the non-linear Schroedinger’s equation traveling with speed equal to the level surface H1 = − ξ . These soliton solutions degenerate to the stationary solution when H1 = 0, that is, when the projections of the extremals on the base curve are elastic. To show that periodic solutions exist on energy level H = requires an explicit formula for u(s), which is tantamount to solving the extremal equations explicitly. The integration procedure is essentially the same as in the Kirchhoff– Lagrange equation. To begin note that (H2 h3 − H3 h2 )2 + (H2 h2 + H3 h3 )2 = (H22 + H32 )(h22 + h23 ). Then,

d h1 ds

2 = (H2 h3 − H3 h2 )2 = (H22 + H32 )(h22 + h23 ) − (H2 h2 + H3 h3 )2 = (H22 + H32 )(I1 − (H12 + H22 + H32 ) − h21 ) − (I2 − h1 H1 )2 = 2(H − h1 )(I1 − H12 − 2(H − h1 ) − h21 ) − (I2 − h1 H1 )2 = 2h31 + c1 h21 + c2 h1 + c3 ,

where I1 = h21 + h22 + h23 + (H12 + H22 + H32 ) and I2 = h1 H1 + h2 H2 + h3 H3 denote the universal integrals on g , and c1 , c2 , c3 are the constants given by the following expressions: c1 = − H12 − 2H − 4 , c2 = 2I2 H1 − 2H12 + 4H − 2I1 , c3 = 2H I1 − H12 − 2H − I22 . It follows that h1 (s) is solvable in terms of elliptic functions, and since k2 = + H32 = 2(H − h1 ), the same can be said for the curvature of the projected elastic curve. The remaining variables u = H2 + iH3 and w = h2 + ih3 can be integrated in terms of two angles θ and φ on the sphere H22

(h1 − )2 + |w|2 = J 2 , where J 2 = I1 − H12 − 2H + 2 .

(17.44)

This new constant J is obtained as follows: I1 = h21 + |w|2 + H12 + |u|2 = h21 + |w|2 + H12 + 2(H − h1 ) = (h1 − )2 + |w|2 + H12 + 2H − . Then angles θ and φ correspond to the spherical coordinates on the above sphere. They are defined by (h1 (s) − ) = J cos θ (s) and w(s) = J sin θ (s)eiφ(s) .

(17.45)

17.4 Affine Hamiltonians and solitons

403

It follows that

dθ dw cosθ dθ dφ dh1 = −J sin θ , and =w +i . ds ds ds sin θ ds ds

Furthermore, u uw¯ H2 h2 + H3 h3 + i(H3 h2 − H2 h3 ) = = w |w|2 J 2 − (h1 − )2 =

1 I2 − h1 H1 + i dh ds

J 2 sin2 θ I2 − h1 H1 i dθ . = − 2 2 J sin θ ds J sin θ This formula shows that I2 − h1 H1 i dθ u(s) = w(s) − J sin θ ds J 2 sin2 θ

(17.46)

Equations (17.43) combined with (17.45) yield cos θ dθ dφ dw w +i = = i(h1 − )u sin θ ds ds ds (I2 − h1 H1 ) cos θ dθ w. + = iJ cos θ sin θ ds J 2 sin2 θ Hence, dφ (I2 − h1 H1 ) J cos θ (I2 − H1 − H1 J cos θ ) = J cos θ = . 2 2 ds J sin θ J 2 sin2 θ

(17.47)

Now, the equation dh1 2 = (H2 h3 − H3 h2 )2 ds = (H22 + H32 )(h22 + h23 ) − (H2 h2 + H3 h3 )2 = 2(H − h1 )|w|2 − (I2 − H1 h1 )2 is the same as 2 dθ (I2 − H1 ( + J cos θ ))2 = 2 (H − − J cos θ ) − , ds J 2 sin2 θ

(17.48)

after the substitutions h1 − − J cos θ and |w|2 = J 2 sin2 θ . The solutions of (17.48) parametize the extremal curves: for then φ is given by equation (17.47) and u and w by equations (17.45) and (17.46). We now return to the question of periodicity. Evidently, both u and w are periodic whenever φ(0) = φ(L) and θ (0) = θ (L). Soliton solutions propagate

404 17 Non-linear Schroedinger’s equation and Heisenberg’s equation with speed −ξ = H1 on energy level H = . On this energy level, φ(0) = φ(L) and θ (0) = θ (L), if and only if L J cos θ (I2 + H1 − H1 J cos θ ) ds = 0, J 2 sin2 θ 0 where θ denotes a closed solution of the equation 2 dθ (I2 − H1 ( + J cos θ ))2 = −2J cos θ − . ds sin2 θ The paper of T. Ivey and D. Singer demonstrates that there are infinitely many closed solutions generated by the constants I1 , I2 , H1 [IvS]. Each such solution generates a solution of the non-linear Schroedinger equation that travels with speed equal to the corresponding value of the integral −I1 .

Concluding remarks Well into the writing of this book I became increasingly aware that some of the topics that I initially planned to include in the text had to be left out. To begin with, I realized that integrable systems, even though linked through common symmetries, somehow stood apart from each other, and each required more space and attention than I initially had imagined. Secondly, as the number of pages grew, I was constantly reminded of that old saying that a big book is a big nuisance. So, some inital expectations had to be curtailed. Initially, I wanted to include a chapter on optimal control of finite dimensional quantum systems. This class of systems fits perfectly our theoretic framework since it is described by left-invariant affine control systems of G = SUn subordinate to the natural Cartan decomposition on the Lie algebra sun [BGN]. In that context, it would have been natural to extend the isospectral representations to optimal problems on SUn and investigate the relevance of the associated integrals of motion for the problems of magnetic resonance imaging. Secondly, I wanted to make a more extended study of isospectral systems on manifolds whose contangent bundles can be represented as coadjoint orbits on Lie algebras and compute some of the spectral invariants associated with the spectral matrix −Lp + λLh + (λ2 − s)B. I had also wanted to include a study of the elastic problem on the Heisenberg group and compare its integrability properties with the elastic problem on the space forms.

Concluding remarks

405

Along with the above, I would have liked to present a more detailed study of infinite dimensional Hamiltonian systems associated with the symplectic form described in the last chapter and investigate its possible relevance for the Korteweg–de Vries equation. The end of the last chapter also calls for a more detailed investigation of the hierarchy of the Poisson commuting functions and of its relation to the results found in [LP] and [ShZ]. Having said this, let me add that these topics would make worthwhile research projects, and I invite the interested reader to carry out the required details.

References

[AM] R. Abraham and J. Marsden, Foundations of Mechanics, Benjamin-Cummings, Reading, MA, 1978. [Ad] J. F. Adams, Lectures on Lie Groups, Mathematics Lecture Notes Series, W. A. Benjamin, New York, 1969. [Al] M. Adler, On a trace functional for formal pseudo-differential operators and symplectic structure of the Korteweg-de Vries equation, Invent. Math. 50 (1979), 219–249. [AB] A. Agrachev, B. Bonnard, M. Chyba and I. Kupka, Sub-Riemannian sphere in Martinet flat case, ESAIM: Control, Optimization and Calculus of Variations 2 (1997), 377–448. [AS] A. Agrachev and Y. Sachkov, Control Theory from the Geometric Point of View, Encyclopedia of Mathematical Sciences, vol. 87, Springer-Verlag, New York, 2004. [Ap] R. Appell E. and Lacour, Principes de la theorie des fonctions elliptiques et applications, Gauthier-Villars, Paris, 1922. [Ar] V. I. Arnold, Mathematical Methods of Classical Mechanics, Graduate Texts in Mathematics, vol. 60, Springer-Verlag, New York, 1978. [AKh] V. I. Arnold and B. Khesin, Topological Methods in Hydrodynamics, App. Math. Sci. (125), Springer-Verlag, New York, 1998. [AKN] V. I. Arnold, V. V. Kozlov, and A. I. Neistadt, Mathematical aspects of classical and celestial mechanics, In Encyclopedia of Mathematical Sciences, Vol 3, Springer-Verlag, Berlin-Heidelberg, 1988. [Bl] A. G. Bliss, Calculus of Variations, American Mathematical Society, LaSalle, IL, 1925. [BR] A. I Bobenko, A. G. Reyman and M. A. Semenov-Tian Shansky, The Kowalewski top 99 years later: a Lax pair, generalizations and explicit solutions, Comm. Math. Phys 122 (1989), 321–354. [Bg1] O. Bogoyavlenski, New integrable problem of classical mechanics, Comm. Math. Phys. 94 (1984), 255–269. [Bg2] O. Bogoyavlenski, Integrable Euler equations on SO4 and their physical applications, Comm. Math. Phys. 93 (1984), 417–436. [Bv] A.V. Bolsinov, A completeness criterion for a family of functions in involution obtained by the shift method, Soviet Math. Dokl. 38 (1989), 161–165. [BJ] B. Bonnard, V. Jurdjevic, I. Kupka and G. Sallet, Transitivity of families of invariant vector fields on semi-direct products of Lie groups, Trans. Amer. Math. Soc. 271 (1982), 525–535.

406

References

407

[BC] B. Bonnard and M. Chyba, Sub-Riemannian geometry: the Martinet case (Conference on Geometric Control and Non-Holonomic Mechanics, Mexico City, 1996, edited by V. Jurdjevic and R.W. Sharpe), Canad. Math. Soc. Conference Proceedings, 25 (1998), 79–100. [Bo] A. Borel, K¨ahlerian coset spaces of semi-simple Lie groups, Proc. Nat. Acad. USA, 40 (1954), 1147–1151. [BT] R. Bott and W. Tu, Differential Forms in Algebraic Topology, Graduate Texts in Mathematics, Springer-Verlag, New York, (1982). [Br1] R. Brockett, Control theory and singular Riemannian geometry, in New Directions in Applied Mathematics (P. Hilton and G. Young, eds.), SpringerVerlag, New York, 1981, 11–27. [Br2] R. Brockett, Nonlinear control theory and differential geometry, Proceedings of the International Congress of Mathematicians, Aug. 16–24 (1983), Warszawa, 1357–1368. [BGN] R. Brockett, S. Glazer, and N. Khaneja, Sub-Riemannian geometry and time optimal control of three spin systems: quantum gates and coherence transfer, Phys. Rev. A 65(3), 032301, (2002). [BG] R.Bryant and P. Griffiths, Reductions for constrained variational problems and 1 κ 2 ds, Amer. Jour. Math. 108 (1986), 525–570. 2 [Br] J.P Brylinski, Loop Spaces, Characteristic Classes and Geometric Quantization, Progress in Math. (108), Birkhauser, Boston, Mass, 1993. [Cr] C. Carath´eodory, Calculus of Variations, reprinted by Chelsea in 1982, Teubner, 1935. [C1] E. Cartan, Sur une classe remarquable d’espaces de Riemann, Bull. Soc. Math. France 54 (1927), 114–134. [C2] E. Cartan, Lec¸ons sur les invariants integraux, Herman, Paris, 1958. [CK] A. l. Castro and J. Koiller, On the dynamic Markov–Dubins problem: from path planning in robotics and biolocomotion to computational anatomy, Regular and Chaotic Dynamics 18, No. 1–2 (2013), 1–20. [Ch] J. Cheeger and D. Ebin, Comparison Theorems in Reimannian Geometry, North-Hollond, Amsterdam, 1975. [CL] E. A. Coddington and N. Levinson, Theory of Ordinary Differential Equations, McGraw Hill, New York, 1955. [Dr] G. Darboux, La th´eory g´en´eral des surfaces, Vol III, Translation of the Amer. Math. Soc. original publication Paris 1894, Chelsea, New York, 1972. [D] P. A. M. Dirac, Generalized Hamiltonian Dynamics, Can. J. Math. 1 (1950), 129–148. [DC] M. P. DoCarmo, Riemannian Geometry, Birkh¨auser, Boston, MA, 1992. [Drg] V. Dragovic, Geometrization and generalization of the Kowdewski top, Commun. Math. Phys. 298 (2010), 37–64. [Db] L. E. Dubins, On curves of minimal ength with a constraint on the average curvature and with prescribed initial positions and tangents, Amer. J. Math. 79 (1957), 497–516. [DKN] B. A. Dubrovin, I. M Kirchever and S. P. Novikov, Integrable systems I, In Dynamical systems IV (V. I. Arnold and S. P. Novikov, eds.), Encylopaedia of Mathematical Sciences, Vol 4. (1990), Springer-Verlag, 174–271. [Eb] P. B. Eberlein, Geometry of Nonpositively Curved Manifolds, Chicago Lectures in Mathematics, The University of Chicago Press, Chicago, IL, 1997. [Ep] C. K. Epstein and M. I. Weinstein, A stable manifold theorem for the curve shortening equation, Comm. Pure Appl. Math. XL (1987), 119–139.

408

References

[E] L. Euler, Methodus inveniendi lineas curvas maximi minimive proprieatate gaudentes, sive solutio, problematis isoperimetrici lattisimo sensu accepti, Opera Omnia Ser. Ia Lausannae 24 (1744). [Eu] L. Euler, Evolutio generalior formularum comparationi curvarum inserventium, Opera Omnia Ser Ia 20 (E347/1765), 318–354. [Fa] L. Faddeev L. and L.Takhtajan, Hamiltonian Methods in the Theory of Solitons, Springer-Verlag, Berlin, 1980. [FJ] Y. Fedorov and B. Jovanovic, Geodesic flows and Newmann systems on Steifel varieties: geometry and integrability, Math. Z. 270, (3–4) (2012), 659–698. [Fl1] H. Flaschka, Toda latice, I: Existence of integrals, Phys. Rev. B3(9) (1974), 1924–1925. [Fl2] H. Flaschka, Toda latice, II: Inverse scattering solution, Progr. Theor. Phys. 51 (1974), 703–716. [Fk] V. A. Fock, The hydrogen atom and non-Euclidean geometry, Izv. Akad. Nauk SSSR, Ser. Fizika 8 (1935). [FM] A. T. Fomenko and A. S. Mischenko, Euler equation on finite-dimensional Lie groups, Math USSR Izv. 12 (1978), 371–389. [FT] A. T. Fomenko and V. V. Trofimov, Integrability in the sense of Louville of Hamiltonian systems on Lie algebras, (in Russian), Uspekhi Mat. Nauk 2(236) (1984), 3–56. [FT1] A. T. Fomenko and V. V. Trofimov, Integrable systems on Lie algebras and symmetric spaces, in Advanced Studies in Contemporary Mathematics, Vol. 2, Gordon and Breach, 1988. [GM] E. Gasparim, L. Grama and A.B. San Martin, Adjoint orbits of semisimple Lie groups and Lagrangian submanifolds, to appear, Proc. of Edinburgh Mat. Soc. [GF] I. M. Gelfand and S. V. Fomin, Calculus of Variations, Prentice-Hall, Englewood Cliffs, NJ, (1963). [Gb] V. V. Golubev, Lectures on the Integration of the Equations of Motion of a Heavy Body Around a Fixed Point, (in Russian), Gostekhizdat, Moscow, 1977. [Gr] P. Griffiths, Exterior Differential Systems and the Calculus of Variations, Birkh¨auser, Boston, MA, 1983. [GH] P. Griffiths and J. Harris, Principles of Algebraic Geometry, WileyInterscience, New York, 1978. [GS] V. Guillemin and S. Sternberg, Variations on a Theme by Kepler, vol. 42 Amer. Math. Soc., 1990. [Gy] G. Gy¨orgi, Kepler’s equation, Fock Variables, Baery’s equation and Dirac. brackets, Nyo Vo cimertio A 53 (1968), 717–736. [Hm] R. S. Hamilton, The inverse function theorem of Nash and Moser, Bull. Amer. Math. Soc. 7 (1982), 65–221. [H1] H. Hasimoto, Motion of a vortex filament and its relation to elastica, J. Phys. Soc. Japan (1971), 293–294. [H2] H. Hasimoto, A soliton on the vortex filament, J. Fluid Mech. 51 (1972), 477–485. [Hl] S. Helgason, Differential Geometry, Lie Groups and Symmetric Spaces, Academic Press, New York, 1978. [HV] D. Hilbert and S. Cohn-Vossen, Geometry and Imagination, Amer. Math. Soc., 1991. [Hg] J. Hilgert, K. H. Hofmann and J. D. Lawson, Lie Groups, Convex Cones ans Semigroups, Oxford Univesity Press, Oxford, 1989. [HM] E. Horozov and P. Van Moerbeke, The full geometry of Kowalewski’s top and (1, 2) Abelian surfaces, Comm. Pure and App. Math. XLII (1989), 357-407.

References

409

[IS] C. Ivanescu and A. Savu, The Kowalewski top as a reduction of a Hamiltonian system on Sp(4, R), Proc. Amer. Math. Soc. (to appear). [IvS] T. Ivey T and D.A. Singer, Knot types, homotopies and stability of closed elastic curves, Proc. London Math. Soc 79 (1999), 429–450. [Jb] C. G. J. Jacobi, Vorlesungen uber Dynamik, Druck und Verlag von G. Reimer, Berlin, 1884. [J1] V. Jurdjevic, The Delauney-Dubins Problem, in Geometric Control Theory and Sub-Riemannian Geometry. (U. Boscain, J. P. Gauthier, A. Sarychev and M. Sigalotti, eds.) INDAM Series 5, Springer Intl. Publishing 2014, 219–239. [Ja] V. Jurdjevic, The symplectic structure of curves in three dimensional spaces of constant curvature and the equations of mathematical physics, Ann. I. H. Poincar´e (2009), 1843–1515. [Jr] V. Jurdjevic, Optimal control on Lie groups and integrable Hamiltonian systems, Regular and Chaotic Dyn. (2011), 514–535. [Je] V. Jurdjevic, Non-Euclidean elasticae, Amer. J. Math. 117 (1995), 93–125. [Jc] V. Jurdjevic, Geometric Control Theory, Cambridge Studues in Advanced Mathematics, vol. 52, Cambridge University Press, New York, 1997. [JA] V. Jurdjevic, Integrable Hamiltonian systems on Lie groups: Kowalewski type, Annals Math. 150 (1999), 605–644. [Jm] V. Jurdjevic, Hamiltonian systems on complex Lie groups and their Homogeneous spaces, Memoirs AMS, 178, (838) (2005). [JM] V. Jurdjevic and F. Perez-Monroy, Variational problems on Lie groups and their homogeneous spaces: elastic curves, tops and constrained geodesic problems, in Non-linear Control Theory and its Applications (B. Bonnard et al., eds.) (2002), World Scientific Press, 2002, 3–51. [JZ] V. Jurdjevic and J. Zimmerman, Rolling sphere problems on spaces of constant curvature, Math. Proc. Camb. Phil. Soc. 144 (2008), 729–747. [Kl] R. E. Kalman, Y. C. Ho and K. S. Narendra, Controllability of linear dynamical systems, Contrib. Diff. Equations, 1 (1962), 186–213. [Kn] H. Kn¨orrer, Geodesics on quadrics and a mechanical problem of C. Newmann, J. Riene Angew. Math. 334 (1982) 69–78. [Ki] A. A. Kirillov, Elements of the Theory of Representations, Springer, Berlin, 1976. [KN] S. Kobayashi and K. Nomizu, Foundations of Differential Geometry, Vol I, Interscience Publishers, John Wiley and Sons, New York, 1963. [Km] I. V. Komarov, Kovalewski top for the hydrogen atom, Theor. Math. Phys 47(1) (1981), 67–72. [KK] I. V. Komarov and V. B. Kuznetsov, Kowalewski top on the Lie algebras o(4), e(3) and o(3, 1), J. Phys. A 23(6) (1990), 841–846. [Ks] B. Kostant, The solution of a generalized Toda latrtice and representation theory, Adv. Math. 39 (1979), 195–338. [Kw] S. Kowalewski, Sur le probl`eme de la rotation d’un corps, solide autor d’un point fix´e Acta Math. 12 (1889), 177–232. [LP] J. Langer and R. Perline, Poisson geometry of the filament equation, J. Nonlinear Sci., 1 (1978), 71–93. [LS] J. Langer and D. Singer, The total squared curvature of closed curves, J. Diff. Geometry, 20 (1984), 1–22. [LS1] J. Langer and D. Singer, Knotted elastic curves in R3 , J. London Math. Soc. 2(30) (1984), 512–534.

410

References

[Lt] F. Silva Leite, M. Camarinha and P. Crouch, Elastic curves as solutions of Riemannian and sub-Riemannian control problems, Math. Control Signals Sys. 13 (2000), 140–155. [LL] L. D. Landau and E. M. Lifshitz, Mechanics, 3rd edition, Pergamon Press, Oxford, 1976. [Li] R. Liouville, Sur le movement d’un corps solide pesant suspendue autour d’un point fix´e, Acta Math. 20 (1896), 81–93. [LSu] W. Liu and H. J. Sussmann, Shortest Paths for Sub-Riemannian Metrics on Rank 2 Distributions, Amer. Math. Soc. Memoirs 564, Providence, RI, (1995). [Lv] A. E. Love, A Treatise on the Mathematical Theory of Elasticity, 4th edition, Dover, New York, 1927. [Lya] A. M. Lyapunov, On a certain property of the differential, equations of a heavy rigid body with a fixed point Soobshch. Kharkov. Mat. Obshch. Ser. 2 4 (1894), 123–140. [Mg] F. A. Magri, A simple model for the integrable Hamiltonian equation, J. Math. Phys. 19 (1978), 1156–1162. [Mn] S. V. Manakov, Note on the integration of Euler’s equations of the dynamics of an n dimensional rigid body, Funct. Anal. Appl. 10 (1976), 328–329. [MK] A. A. Markov, Some examples of the solution of a special kind of problem in greatest and lowest quantities (in Russian), Soobsch. Karkovsk. Mat. Obshch. (1887), 250–276. [MZ] J. Millson and B. A. Zambro, A K¨ahter structure on the moduli spaces of isometric maps of a circle into Euclidean spaces, Invert. Math. 123(1) (1996), 35–59. [Ml] J. Milnor, Curvature of left invariant metrics on Lie groups, Adv. Math. 21(3), (1976), 293–329. [Mi1] D. Mittenhuber, Dubins’ problem in hyperbolic spaces, in Geometric Control and Non-holonomic Mechanics (V. Jurdjevic and R. W. Sharpe, eds.), vol. 25, CMS Conference Proceedings American Mathematical Society Providence, RI, 1998, 115–152. [Mi2] D. Mittenhuber, Dubins’ Problem is Intrinsically Three Dimensional, ESAIM: Contrˆole, Optimisiation et Calcul des Variations, vol. 3 (1998), 1–22. [Mo] F. Monroy-P´erez, Three dimensional non-Euclidean Dubins’ problem, in Geometric Control and Non-Holonomic Mechanics (V. Jurdjevic and R. W. Sharpe, eds.), vol. 25, CMS Conference Proceedings Canadian Math. Soc., 1998, 153–181. [Mt] R. Montgomery, Abnormal minimizers, SIAM J. Control Opt. 32(6) (1994), 1605-1620. [Ms1] J. Moser, Regularization of Kepler’s problem and the averaging method on a manifold, Comm. Pure Appl. Math. 23(4) (1970), 609–623. [Ms2] J. Moser, Integrable Hamiltonian Systems and Spectral Theory, Lezioni Fermiane, Academia Nazionale dei Lincei, Scuola Normale, Superiore Pisa, 1981. [Ms4] J. Moser, Finitely Many Mass Points on the Line Under the Influence of an Exponential Potential-an Integrable System, in Lecture notes in Physics, vol 38 Springer, Berlin, 1975, 97–101. [Ms3] J. Moser, Various aspects of integrable Hamiltonian systems, in Dynamical Systems, C.I.M.E. Lectures, Bressanone, Italy, June, 1978, Progress in Mathematics, vol 8, Birkhauser, Boston, MA, 1980, 233–290. [Na] I. P. Natanson, Theory of Function of a Real Variable, Urgar, 1995. [Nm] C. Newmann, De probleme quodam mechanico, quod ad primam, integralium ultra-ellipticoram classem revocatum J. Reine Angew. Math. 56 (1856).

References

411

[NP] L. Noakes L. and T. Popiel, Elastica on SO(3), J. Australian Math. Soc. 83(1) (2007), 105–125. [NHP] L. Noakes, G. Heizinger and B. Paden, Cubic spines of curved spaces, IMA J. Math. Control Info. 6 (1989), 465–473. [O1] Y. Osipov, Geometric interpretation of Kepler’s problem, Usp. Mat. Nauk 24(2), (1972), 161. [O2] Y. Osipov, The Kepler problem and geodesic flows in spaces of constant curvature, Celestial Mechanics 16 (1977), 191–208. [Os] R. Osserman, From Schwarz to Pick to Ahlfors and beyond, Notices of AMS 46(8) (1997), 868–873. [Pr] A. M. Perelomov, Integrable Systems of Classical Mechanics and Lie Algebras, vol. I, Birkhauser Verlag, Basel, 1990. [Pc] E. Picard, Oeuvres de Charles Hermite, Publi´ees sous les auspices de l’Academie des sciences, Tome I, Gauthier-Villars, Paris, 1917. [Po] H. Poincar´e, Les m´ethodes nouvelles de la mechanique celeste, Tome I, Gauthier-Villars, Paris, 1892. [Pt] L. S. Pontryagin, V. G. Boltyanski, R. V. Gamkrelidze and E. F. Mishchenko, The Mathematical Theory of Optimal Processes, Translated from Russian by D. E. Brown, Pergamon Press, Oxford, 1964. [Rt1] T. Ratiu, The motion of the free n-dimensional Rigid body, Indiana Univ. Math. J. 29(4), (1980), 602–629. [Rt2] T. Ratiu, The C. Newmann problem as a completely integrable system on a coadjoint orbit, Trans. Amer. Mat. Soc. 264(2) (1981), 321–329. [Rn] C. B. Rauner, The exponential map for the Lagrange problem on differentiable manifolds, Phil. Trans. Royal Soc. London, Ser. A 262 (1967), 299–344. [Rm] A. G. Reyman, Integrable Hamiltonian systems connected with graded Lie algebras, J. Sov. Math. 19 (1982), 1507–1545. [RS] A. G. Reyman and M. A. Semenov-Tian Shansky, Reduction of Hamiltonian systems, affine Lie algebras and Lax equations I., Invent. Math. 54 (1979), 81–100. [RT] A. G. Reyman and M. A. Semenov-Tian Shansky, Group-theoretic methods in the theory of finite-dimensional integrable systems, Encyclopaedia of Mathematical Sciences (V. I. Arnold and S. P. Novikov, eds.), Part 2, Chapter 2, Springer-Verlag, Berlin Heidelberg, 1994. [Ro] H. L. Royden, Real Analysis, 3rd Edition, Prentice Hall, Englewood Cliffs, 1988. [Sc] J. Schwarz, von, Das Delaunaysche Problem der Variationsrechung in kanonischen Koordinaten, Math. Ann. 110 (1934), 357–389. [ShZ] C. Shabat and V. Zakharov, Exact theory of two dimensional self-focusing, and one dimensional self-modulation of waves in non-linear media, Sov. Phys. JETP 34 (1972), 62–69. [Sh] R. W. Sharpe, Differential Geometry: Cartan’s Generalization of Klein’s Erlangen Program, Graduate Texts in Mathematics, Springer, New York, 1997. [Sg1] C. L. Siegel, Symplectic Geometry, Amer. J. Math. 65 (1943), 1–86. [Sg2] C. L. Siegel, Topics in Complex Function Theory, vol. 1: Elliptic Functions and Uniformization Theory, Tracts in Pure and Appl. Math. vol. 25, WileyInterscience, John Wiley and Sons, New York, 1969. [Sg] E. D. Sontag, Mathematical Control Theory, Deterministic Finite-Dimensional Systems, Springer-Verlag, Berlin, 1990. [So] J. M. Souriau, Structure des Systemes Dynamiques, Dunod, Paris, 1970.

412

References

[Sf] P. Stefan, Accessible sets, orbits and foliations with singularities, London Math. Soc. Proc., Ser. 3, 29 (1974), 699–713. [St] S. Sternberg, Lectures on Differential Geometry, Prentice-Hall Inc, Englewood Cliffs, NJ, 1964. [Sz] R. Strichartz, Sub-Riemannian Geometry, J. Diff. Geometry 24 (1986), 221–263. [Sg] M. Sugiura, Conjugate classes of Cartan subalgebras in real semisimple Lie algebras, J. Math. Soc. Japan, 11(4) (1959). [Ss] H. J. Sussmann, Orbits of families of vector fields and integrability of distributions, Trans. Amer. Math. Soc. 180 (1973), 171–188. [Ss1] H. J. Sussmann, A cornucopia of four-dimensional abnormal sub-Riemannian minimizers, in Sub-Riemannian Geometry (A. Bella¨iche and J. J. Risler, eds), Progress in Mathematics, vol. 144, Birkh¨auser, 1996, 341–364. [Ss3] H. J. Sussmann, Shortest 3-dimensional paths with a prescribed curvature bound, in Proc. 34th Conf. on Decision and Control, New Orleans, 1995, 3306– 3311. [Ss4] H. J. Sussmann, The Markov-Dubins problem with angular acceleration, in Proc. 36th IEEE Conf. on Decision and Control, San Diego, CA, 1997, 2639– 2643. [Sy] W. Symes, Systems of Toda type, inverse spectral problems and representation theory, Invent. Math. 59 (1980), 13–53. [To] M. Toda, Theory of Nonlinear Lattices, Springer, New York, 1981. [Tr] E. Trelat, Some properties of the value function and its level sets for affine control systems with quadratic cost, J. Dyn. Contr. Syst. 6(4) (2000), 511–541. [VS] J. von Schwarz, Das Delaunische Problem der Variationsrechung in kanonischen Koordinaten, Math. Ann. 110 (1934), 357–389. [W] K. Weierstrass, werke, Bd 7, Vorlesungen u¨ ber Variationsrechung, Akadem. Verlagsges, Leipzig, 1927. [Wl] A. Weil, Euler and the Jacobians of elliptic curves, in Arithmetic and Geometry (M. Artin and J. Tate, eds.), Progress in Mathematics, vol. 35, Birkh¨auser, Boston, MA, 353–359. [Wn] A. Weinstein, The local sructure of Poisson manifolds, J. Diff. Geom. 18 (1983), 523–557. [Wh] E. T. Whittaker, A Treatise on the Analytical Dynamics of Particles and Rigid Bodies, 3rd edition, Cambridge Press, Cambridge, 1927. [Wf] J. A. Wolf, Spaces of Constant Curvature, 4th edition, Publish or Perish, Inc, Berkeley, CA, 1977. [Yo] K. Yosida, Functional Analysis, Springer-Verlag, Berlin, 1965. [Yg] L. C. Young, Lectures in the Calculus of Variations and Optimal Control Theory, W.B. Saunders, Philadelphia, PA, 1969. [Zg1] S. L. Ziglin, Branching of solutions and non-existence of first, integrals in Hamiltonian mechanics I, Funkts. Anal. Prilozhen 16(3) (1982), 30–41. [Zg2] S.L. Ziglin, Branching of solutions and non-existence of first, integrals in Hamiltonian mechanics II, Funkts. Anal. Prilozhen 17(1) (1981), 8–23.

Index

absolutely continuous curve, 15 accesibility property, 22 Accessibility theorem, 22 adjoint representation, 60, 101 Affine-quadratic problem, 147 Agrachev, A., 86 angle nutation, 292 precession, 292 angular velocity in absolute frame, 224 in moving frame, 224 Bobenko, I.A., 313 Bogoyavlensky, O., 237 Brockett, R.W., xiv Caratheodory, C., 85, 330 Cartan’s formula, 52 Cartan decomposition, 119 coadjoint representation, 60 complex structure, 46 cone of attainability, 75 connection, 131 control affine system, 21 control drift, 21 control system, 19 controlled vector fields, 21 controllable, 23 controllability criterion, 25 cost-extended control system, 68 covariant derivative, 138 covariant derivative problem, 329 curvature problem, 329

cubic spline, 332 Curve shortening equation, 399 Darboux’s theorem, 51 Darboux curve, 248 anchored, 372 reduced, 373 periodic, 374 framed curves, 372 Delauney–Dubins problem, 331 diffeomorphism, 6 differential form, 3 closed, 5 exact, 5 Dirac, P., 210 distribution, 14 integrable, 15 involutive, 15 Dragovic, V., 313 Dubins, L., 331 elastic problem, 104 elliptic geodesic equations, 202 Elliptic geometry, 101 elementary perturbation vector, 76 escape time positive, 6 negative, 6 exterior derivative, 4 extremal control, 67 extremal curve, 74 normal, 80 abnormal, 80 extremal trajectory, 67 Euler–Griffiths problem, 250

413

414

Euler quadratic form, 225 Euler’s elasticae, 105 Fedorov, Y., 210 Filament equation, 390 flow, 6 Fomenko, A.T., 160 forward-backward principle, 72 frame moving, 222 absolute, 222 Darboux, 248 Frechet manifold, 376 tame, 376 Frobenius theorem, 16 generalized elastic rod, 257 generalized upper half plane, 42, 127 geodesic, 135 geodesic symmetry, 138 Grassmannian, 43 oriented Grassmannian, 129 Griffiths, P., 105 group action, 41 Hamilton, R.S., 376 Hamilton equation, 94 Hamiltonian lift, 71 of a control system, 74 Hamiltonian vector field, 49, 50 Hasimoto, H., 394 Heisenberg’s magnetic equation, 389 Hermann–Nagano theorem, 18 Hermitian matrix, 269 homogeneous space, 41 horocycle, 110 horizontal curve, 131 horizontal lift, 131 holonomy group, 132 heavy top, 220 of Euler, 220 of Kowalewski, 220 of Lagrange, 220 Hyperbolic geometry, 96 inertia tensor, 224 of a body, 225 infinitesimal generator, 6 integral curve, 6 of a distribution, 14 integral manifold, 15 maximal, 15

Index

interior product (contraction) of, 50 invariant functions, 156 invariant quadratic form, 62 involutive automorphism, 118 isometry, 138 isoperimetric problem, 87 isospectral Hamiltonian, 159 isospectral case, 281 isotropic, 45 isotropy group, 41 Jacobi’s curves, 140 Jacobi’s equation, 140 Jovanovich, B., 210 Kepler, J., 180 Killing form, 62 Kirchhoff, G., 114, 221, 256 Kirchhoff–Kowalewski case, 281 Kirchhoff–Lagrange case, 281 Kirillov, A., 63 Knorrer, H., 213 Kowalewski, S., 261, 273 Kowalewski gyrostat in two constant fieldw, 319 Lagrangian, 45 Langer, J., 399 Lebesgue point, 75 left-invariant representation of, 58 Legendre condition, 93 Lie algebra, 31 derived, 65 edge, 40 semi-simple, 63 simple, 123 wedge, 40 Lie algebra generated by vector fields, 13 Lie bracket, 3 Lie derivative, 52 Lie determined system, 17 Lie group, 29 general linear, 29 orthogonal, 36 special, 36 Lie–Poisson bracket, 60 Lie saturate, 25 Lie subgroup, 34 Liouville form, 48 Lorentzian, 36 Semi-direct product, 38 unitary, 37

Index

Love, A.E., 114, 256 Lyapunov, A.M., 274 Magri scheme, 395 Manakov, S.V., 160 Markov, A.A., 274, 331 Martinet plane, 87 Maximum Principle of optimality, 80 Mishchenko, A.S., 160 Modified Korteweg–de Vries equation, 398 Moment map, 84 Monroy’s lemma, 342 Montgomery’s example, 89 Moser, J., 178, 180, 210, 219 Milson, J., 385 n-dimensional Euler top, 231 Newmann, C., 178 Newtonian ptential field, 232 Noether’s theorem, 83 Non-linear Schroedinger’s equation, 391 normalizer, 28 one-parameter group of, 7 optimal control, 67 optimal trajectory, 68 orbit coadjoint, 60 of a family of vector fields, 10 of a vector field, 7 theorem, 10 Osipov, Y., 180 parallel curves, 132 Parallel frame, 249 Perline, R, 399 Poisson bracket, 53 Poisson manifold, 55 principal moment of inertia, 225

second countable, 1 sectional curvature, 141 Shabat, C., 398 shadow problem, 150 solitons, 402 spectral invariants, 156 Steifel manifold, 43 strongly controllable, 23 subordinated to a group action, 84 sub-Riemannian ball, 110 sphere, 110 wave front, 110 sub-Riemannian geodesics, 88, 135 normal, 88 abnormal, 88 strictly abnormal, 88 symmetric pair, 119 symmetric space, 138 of constant curvature, 141 symmetry, of a control system, 82 symplectic form, 50 left-invariant representation, 377 symplectic basis, 45 symplectic leaves, 57 symplectic manifold, 51 symplectic vector space, 44 symplectomorphism, 52 tangency property, 12, 14 Toda lattice, 192 Toda mechanical system, 197 trace of a family of invariant vector fields, 39 transitive action, 41 vector field, 2 complete, 7 left-invariant, 30 right-invariant, 30

Quadratic Newtonian field, 235 real form of a complex Lie algebra, 263 regular point of a curve, 75 Riemannian curvature, 141 Riemannian symmetric pair, 119 Root, 161 Root space, 161 Runge–Lenz (eccentricity) vector, 182 saturated Maximum Principle, 81 Schwarz von J., 330 Schwarz–Pick lemma, 112

415

Weierstrass, C., 69 Weierstrass’ excess function, 91 Weil, A., 307 Weil’s group, 307 Weyl basis, 161 Young, L.C., xiii, 69, 85 Zakharov, V., 398 zero-curvature equation, 378 zero-time ideal, 65 Zombro, B.A., 385