Shape and Shape Theory

WILEY SERIES IN PROBABILITY AND STATISTICS Established by WALTER A. SHEWHART and SAMUEL S. WILKS Editors: K c Barnett, Noel A. C. Cressie, Nicholas I. Fisher, Iain M. Johnstone, J.B. Kadane, David G. Kendall, David W. Scott, Bernard W. Silverman, Adrian F.M. Smith, Jozef L. Teugels Editors Emeritus: Ralph A. Bradley, J. Stuart Hunter A complete list of the titles in this series appears at the end of this volume

Shape and Shape Theory

D. G. KENDALL
Churchill College, University of Cambridge, UK

D. BARDEN
Girton College, University of Cambridge, UK

T. K. CARNE
King’s College, University of Cambridge, UK

H. LE
University of Nottingham, UK

JOHN WILEY & SONS, LTD
Chichester New York Weinheim Brisbane Singapore Toronto

Copyright 0 1999 by John Wiley & Sons Ltd Baffins Lane, Chichester, West Sussex, PO19 IUD, England

National 01243 119111 International (+44) 1243 119111
e-mail (for orders and customer service enquiries): cs-books@wiley.co.uk Visit our Home Page on http:lwww.wiley.co.uk or http:lwww.wiley.com All Rights Reserved. N o part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, electronic, mechanical, photocopying, recording, scanning or otherwise, except under the terms of the Copyright, Designs and Patents Act 1988 or under the terms of a licence issued by the Copyright Licensing Agency, 90 Tottenham Court Road, London, W I P 9HE, UK, without the permission in writing of the Publisher. Other Wiley Editorial Ofices John Wiley & Sons, Inc., 605 Third Avenue, New York, NY 10158-0012, USA Wiley-VCH Verlag GmbH, Pappelallee 3, D 69469 Weinhelm, Germany Jacaranda Wiley Ltd, 33 Park Road, Milton, Queensland 4064, Austraki John Wiley & Sons (Asia) Pte Ltd, 2 Clementi Loop #02-0 I , Jin Xing Distripark, Singapore 129809 John Wiley & Sons (Canada) Ltd, 22 Worcester Road, Rexdale, Ontario, M9W 1L1, Canada

Library of Congress Cataloging-in-Publication Data
Shape and shape theory / D.G. Kendall . . . [et al.]. em.-(Wiley series in probability and statistics) p. Includes bibliographical references and index. ISBN 0-471-96823-4 (alk. paper) 1. Shape theory (Topology)- Statistical methods I. Kendall, D.G. (David George), 1918- . 11. Series. QA612. 7. S48 1999 99- 12433 5 14’ .24- dc2 1 CIP

British Library Cataloguing in Publication Data
A catalogue record for this book is available from the British Library ISBN 0-471-96823-4 Typeset in 10/12pt Times by Laser Words, Madras, India

Contents
Preface Chapter 1 Shapes and Shape Spaces
1.1 1.2 1.3 1.4 1.5 1.6
Origins Some preliminary observations A matrix representation for the shape of a k-ad ‘Elementary’ shape spaces Ef and E$ The Fubini-Study metric on E$ The proof of Casson’s theorem ix

1 1 2 7 10 17 20 23 23
24

Chapter 2 The Global Structure of Shape Spaces
2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9
The problem When is a space familiar CW complexes A cellular decomposition of the unit sphere The cellular decomposition of shape spaces Inclusions and isometries Simple connectivity and higher homotopy groups The mapping cone decomposition Homotopy type and Casson’s theorem

25 25 27 29 30 33 36

Chapter 3 Computing the Homology of Cell Complexes
3.1 3.2 3.3 3.4
The orientation of certain spaces The orientation of spherical cells The boundary of an oriented cell The chain complex, homology and cohomology groups Reduced homology The homology exact sequence for shape spaces

41 41 42 43 45 47 47

3.5 3.6

5 6.2 7.6 6.8 5.5 5.4 6.8 Applications of the exact sequence Topological invariants that distinguish between shape spaces CONTENTS 50 52 Chapter 4 A Chain Complex for Shape Spaces 4.4 5.2 6.1 4.7 6.3 The Riemannian metric The metric re-expressed through natural local vector fields The Riemannian curvature tensor 135 135 138 149 .3 6.3 4.1 7.vi 3.5 4.7 3.8 The action of SO(m) on the pre-shape sphere Viewing the induced Riemannian metric through horizontal geodesics The singular points and the nesting principle The distance between shapes The set of geodesics between two shapes The non-uniqueness of minimal geodesics The cut locus in shape spaces The distances and projections to lower strata Chapter 7 The Riemannian Structure of Shape Spaces 7.3 5.9 Spaces of shapes in 2-space Spaces of shapes in 3-space Spaces of shapes in 4-space Spaces of unoriented shapes in 2-space Spaces of unoriented shapes in 3-space Spaces of unoriented shapes in 4-space Decomposing the essential complexes Closed formulae for the homology groups Duality in shape spaces 69 70 71 72 78 80 82 84 91 100 105 106 107 111 114 116 120 124 127 Chapter 6 Geodesics in Shape Spaces 6.7 The chain complex The space of unoriented shapes The boundary map in the chain complex Decomposing the chain complex Homology and cohomology of the spaces Connectivity of shape spaces Limits of shape spaces 55 55 57 57 59 62 65 68 Chapter 5 The Homology Groups of Shape Spaces 5.6 5.6 4.4 4.2 4.2 5.1 5.7 5.1 6.

1 11.1 8.1 9.8 190 20 1 20 1 204 206 213 217 226 23 1 23 1 233 237 239 243 244 249 249 254 262 264 Chapter 9 Mean Shapes and The Shape of the\Means 9.4 10.6 Concept of means in non-linear spaces Metrics on shape space Uniqueness of Frkchet means of shape-measures FrCchet means and the shape of the means Uniqueness of FrCchet means of R given shapes Procrustean mean shapes Chapter 10 Visualising The Higher Dimensional Shape Spaces 10.5 10.5 9.2 9.CONTENTS vii 157 158 160 164 168 172 Chapter 8 Induced Shape-Measures 8.7 181 8.6 The two-dimensional representation of Z: The cell-decomposition of SO(3) The action of the group D The geodesics of C: Some distributions on C: A diffusion on C i Chapter 11 General Shape Spaces 11. 11: the explicit formula 176 8.2 10.4 Shape spaces for general manifolds Size-and-shape spaces Size-and-shape spaces for the plane Spheres and hyperbolic spaces .3 11.5 8.3 8.3 9.4 9. I: the singular tessellation The shape-measure on Ci induced by three labelled iid uniform distributions in a convex polygon.4 8.6 Geometric preliminaries The shape-measure on Z i induced by k labelled iid isotropic Gaussian distributions on [w" Shape-measures on Cz+' of Poisson-Delaunay tiles Shape-measures on C: induced by k labelled iid non-isotropic Gaussian distributions on [w2 Shape-measures on C induced by complex normal i distributions The shape-measure on Z: induced by three labelled iid uniform distributions in a compact convex set The shape-measure on C: induced by three labelled iid uniform distributions in a convex polygon.2 8.1 10.2 11.3 10.

4 AS Unary operators on groups Binary operators on groups The Universal Coefficient Theorems Duality in manifolds The splitting of exact sequences Bibliography Index 293 299 .l A. 11..2 A.6 Elliptic functions 11.3 A.7 Shape spaces for triangles in the sphere and the hyperbolic plane CONTENTS 267 273 276 283 283 284 286 287 290 Appendix A..Vlll .5 Relative metrics 11.

it is not a trivial matter to define shape in a manner that is susceptible to mathematical and statistical analysis and it is only over the last two or three decades that appropriate definitions have been developed and studied. and make some initial general deductions about their homology groups. more detail than the specialist in that subject might require. We therefore need to describe the topological and geometric properties of these new spaces in some detail. However. Consequently.Preface Everyone knows what is meant by ‘shape’. classical statistical methods are not always adequate or. We start with an introductory survey of the spaces in which we shall represent shapes and describe some of their more important properties and then. Although this may seem a severe restriction. However. Where possible. we have tried to make it accessible to as wide a range of readers as possible by giving. we carry out much of the fundamental analysis that is likely to lie at the heart of further progress. as result of which this book is multidisciplinary. in Chapter 2. One of the central problems in shape theory is that it is not possible to represent the full range of possible shapes of an object in standard Euclidean coordinates without destroying our intuitive feel for the quantitative differences between them. In the first of these we define homology theory and show how it is calculated in the special context that is adequate for our purposes because. with just a few of the more technical topological concepts and results reserved for the appendix. we investigate their global topological structure. shape spaces are still elementary. we do this within the body of the text itself. In Chapter 4 we examine the necessary chain complex for these computations that arises naturally from the topological structure of the spaces. In this book we assume that the shape of an object is essentially captured by the shape of a finite subset of its points and. for each topic. although they are unfamiliar. Then. at least. after giving a range . for the latter. The next three chapters lead up to a full calculation of the homology and cohomology groups of shape spaces. there is no theoretical limit to the number of points we consider and it has the significant advantage that the dimensions of the resulting shape spaces are always finite and only increase linearly with the number of points. not clearly appropriate for the statistical analysis of shape and it is necessary to adapt them to work on unfamiliar spaces. in Chapter 5.

In particular. In the next two chapters we turn to the probabilistic and statistical topics that were the prime motivation for the introduction of shape spaces. Although there is only one topology. Once again. to which we shall relate any other metrics that we use. However. we calculate all the groups explicitly and also derive some intriguing relationships between them. In classical statistics the mean is well-defined and simple to compute. Here we discuss the Riemannian metric that arises from the theory of submersions. for example. it is by no means a standard one. an explicit description of the density function for the shape of a random triangle whose vertices are uniformly independently distributed in a given convex planar polygon. In Chapter 9 we discuss some of the relations between different possible definitions and also identify circumstances. lowest dimensional. they are still elementary and susceptible to computation. We illustrate this claim by obtaining. fortunately fairly general. In Chapter 7. In Chapter 8 we investigate the distributions that arise on shape space from various standard distributions on the points that determine those shapes. we are able to measure the precise extent to which they are curved. problems can arise both in defining a ‘mean’ shape in theory and also in calculating it in practice. that is. In Chapter 10 we address the problem of visualising the first.X PREFACE of low-dimensional illustrative examples. after introducing a little more differential geometry. to a unique result. This is vital. In Chapter 6 we examine the geodesics. Moreover. do allow . the elementary way in which shape spaces are produced enables us to prove most of the results that we require from that theory directly in our context. between two shapes and find simple expressions for the distance between those shapes. We describe them generally by referring to the volume measure on shape space obtained in Chapter 7. The visualisation uses a carefully selected family of 24 two-dimensional sections that. In Chapters 6 and 7 we study the more subtle and more localised geometric properties of the spaces. rather surprisingly. However. the formulae can become quite intricate. shape space that is not already familiar. there is more than one metric. when assessing the extent to which a local linear approximation to shape space is valid. Since the curvature can be arbitrarily large. that is certainly not always the case. that one can put on shape space. as was the case for the homology groups. Although this is topologically a sphere. as it has a singular subset in the neighbourhood of which the curvature becomes arbitrarily large. in which the results we would like to take for granted are actually true. we are able to obtain explicit expressions for the main geometric invariants of the spaces. in principle. As the initial distributions become more general. That is the five-dimensional space of shapes of tetrahedra in 3-space. the analogues of straight lines in Euclidean space. in the final sections of this chapter. with little reference to the general case. as well as the distance from a shape to certain subsets of practical significance. although. the natural quotient topology. our results are quite general. in each case. the range of shape spaces on which we give explicit formulae for the induced distributions tends to become more restricted. It turns out that various ‘obvious’ approaches do not necessarily lead to the same results or even.

. We illustrate this by describing some typical geodesics. indebted to all who have worked on shape theory. We conclude by putting our work into a broader setting where similar studies may be carried out. The first applications still concern finite sets of points in Euclidean space but here we study their size-and-shape. and we are grateful to all our colleagues. some sample paths for a diffusion and a comparison of two distributions on the space. Much of the material presented is previously unpublished work given in local seminars or work produced explicitly for this book. We are. In particular. particularly to Helen Ramsey for her constant encouragement and patience over our ever-receding deadlines. and also an alternative metric. whether or not it lies in the area that we specifically address. In the final sections we consider the shapes of finite sets of points in the other standard spaces of constant curvature. Thanks are also due to our publishers. on the non-degenerate part of the shape space.PREFACE xi us to follow what is going on in the space. for which size is no longer quotiented out. the sphere and hyperbolic space. for many helpful discussions over the years of gestation of this project. but especially to Marge Batchelor. as well as some connections with the classical theory of elliptic functions. this enables us to look at some shape spaces related to those that have been the subject of the rest of this book. one having negative curvature. of course.

G. Thus. These definitions and the related constructions provide the basis for the present book. the shape is defined as the pre-shape modulo rotations. the centroid of the k points will serve as an origin. Accordingly. say. and any two configurations of k labelled points will be regarded as having the same shape if either of their preshapes can be transformed into the other b y a rotation about the shared centroid. The basic object just described will be called the pre-shape. however. In this present book the theory will be developed largely along the lines initiated by Kendall in his 1977 paper. In a typical case the calculations will be concerned with sets of. In Bookstein’s work. if the object in question is a planar representation of a human hand with fingers out-stretched. A third early contributor was Ziezold (1977). k labelled points in a Euclidean space Rn. K. it is appropriate here to stress the fact that for us the labelled points are basic and determine the object being studied. At about the same time Bookstein (1978a. as far as the wrist. concerned with ‘shape’ in archaeology and astronomy. BARDEN & T. but much new material will be presented. At a later stage we shall define ‘size-and-shape’ in a similar way by omitting the ‘unit-sum-of-squares’ standardisation. the ‘marker points’ are selected from a usually two-dimensional or three-dimensional continuum. KENDALL & D. It should be observed that the k constituent labelled points determine the shape. say.b) began to study shape-theoretic problems in the particular context of zoology. The first approach seems to have been that of Kendall (1977) who was. C A R " & H. LE Copyright 0 1999 by John Wiley & Sons Ltd CHAPTER 1 Shapes and Shape Spaces 1. where k 3 2. but it soon became clear that the subject could profitably be studied from a more general standpoint. The resulting assemblage of all possible shapes will be called the shape space and will be denoted by EL. Already in . and the scale will be such that the sum of the squared distances of the points from that origin will be equal to unity. at that time.1 ORIGINS There have been at least three distinct origins of what we call shape theory. the common roots of each pair of adjacent fingers and a few more points on the planar outline of the hand reaching down. then the markers could be the tips of the fingers.Shape and Shape Theory D. Normally. While we do not wish to go deeply into the details of Bookstein’s parallel work.

but as far as shape alone is concerned it is of no interest. x. A copy of Bower’s book was given by him to Kendall. . which contain the shapes of all possible configurations except those for which all the points coincide. . so we normally shrink or expand the size of the centred k-ad about the new origin so as to make the natural quadratic measure of ‘size’ equal to unity.. This contrasts with Kendall’s spaces. x. This provides a context in which it is possible to measure the statistical significance of apparent collinearities or other degeneracies in archaeological or astronomical data. . is moved to the origin of the coordinates. in R m in such a way that their : centroid. It is appropriate here to associate Thompson’s work with that of Bower (1930) who studied ‘size and form’ in plants. of course. Thus. The ‘size’ of this k-ad is. the only such sized-and-centred configurations when k = 2 and m = 1 are the labelled point-pairs: . and it was this event that led many years later to the formulation of shape-theoretic studies in a general mathematical context. and the markers are chosen sufficiently well spaced to identify those objects.2 SOME PRELIMINARY OBSERVATIONS Consider the shape of a configuration of k 3 2 labelled and not totally coincident points x:. How is the shape of such a configuration to be represented? Since we are not interested in the location of the k-ad. Bookstein’s work includes many delicate and important studies concerning the continuous deformation of biological shapes. important as an aspect of ‘size-and-shape’. 1. and it would be still more difficult if the complete three-dimensional surface were to be the object being studied. consider the problem of coding the shape of a potato! A further difference between the two approaches arises at the next stage. he is not interested in configurations in which the markers all lie in a lower-dimensional subspace or two or more of them coincide. This convention makes sense because we have deliberately excluded the maximally degenerate case in which all the points xy coincide. Bookstein is concerned to represent real objects. this being a topic first studied by Thompson [1917] (1942). To take a still more difficult example.2 SHAPES AND SHAPE SPACES this simple case it is clear that the choice of markers is far from being a simple matter even in the two-dimensional case. . we may start by uniformly translating its component points x. x$ in a Euclidean space having m 3 1 Euclidean dimensions. often biological ones. To take the most trivial example.

the dimension of Zk.1) .?m(m . and we define the shape itself to be viewed modulo the rotations in SO(m) acting from the left. which we shall give in more detail later. If we write for the sized-and-centred m x k coordinate matrix with components x* (i.xk) c x i i=l = 0.1)-dimensional subspace and so its shape lies in However. . then the k individual columns of the matrix can be thought of as column mvectors specifying the positions of the k points x. Thus.x.1 because we want to have the centroid of the k points at the origin of the coordinates. while the final term. is x* z* x* x* + k 1 d. and let us move the centroid of the k-ad to the origin and standardise the size as before. -1. A straightforward generalisation of this argument. now the extra dimensions give us room to rotate the configuration onto its mirror image. x. = k . Thus. Ck. so we see that C: consists of just two shapes. . and the corresponding shape space Zk is S i modulo SO(m) with the rotations acting from the left... we will never wish to distinguish between and T x * where T is in SO(m). .I . . shape space C is a unit-radius ( k . we note that d'.* in R" where.j 1 < i : < m. and we call this SL.SOME PRELIMINARY OBSERVATIONS 3 in R2.2)-sphere Sk-*(l) for all i k 3 2. while there are m rows and k columns in the matrix k is here reduced to k . being the result of identifying the shapes of this x*. takes account of the fact that we wish to ignore scale effects.1) to obtain the shape. In particular. a configuration of k labelled points in Rn lies in a (k . = m(k . Then. as above. That is the pre-shape space. is a 'halved' version of Xt-. Provided that k 3 m I . x2. This is because it is a basic feature of our work that rotations acting on the left of are to be regarded as irrelevant. . tells us that shape space C: consists of all standardised k-ads of the form: { k k ( X I . We therefore call the sized-and-centred configuration described by the pre-shape. 1) of dimension zero. This means that the pre-shape is quotiented out by O ( k .1) rather than SO(k . It is easily checked that the complete set of all such pre-shapes is a unit sphere of dimension m(k . from the shape-theoretic point of view. The second term on the right arises because we must quotient out the effect of rotations of S O ( m ) acting from the left. When k < m. But now suppose that m 3 2 and k 3 2.* is the centroid of the k-ad and 1 < j < k. i= 1 =1 I for k 3 2. and that it can be identified with the two-point unit sphere {-I. In this formula the first term on the right follows from the fact that. 1 < j < k). for k < m.2 agrees with our earlier calculations.I ) - 1.1) .

Before discussing general shape spaces Xk.. but the corresponding situation when m is equal to two or more can be quite complicated.2.2. A striking feature of Table 1. zero-dimensional. is ‘over-dimensioned’ in the sense that dk. spheres. In Table 1. the spaces listed in the first column. is a metric 2-sphere of radius one-half.1)-space that are mirror images of each other.2 and a real dimension 2k . When m = 1 the shape is identical with the pre-shape because there are no non-trivial rotations T in SO(1). are all unit spheres. from the corresponding entry in Table 1. We shall see later that X.where the ‘4’ is the appropriate value for the curvature parameter. the entry at ( k .1 is the shape space for four labelled points in three dimensions and. we see that this is a five-dimensional shape space-actually we shall find that it is a topological 5-sphere that possesses singularities. those for which k 6 m.) = d i p l . metric sphere of radius unity. etc. the quotient operation that yields the shape itself is very far from being so. Obviously. it is desirable that in the diagrams we should be able to recognise those shape spaces that are over-dimensioned. We already know that X is the two-point. More details about this will be given later in Section 1. when k 6 m. > dim(Xk.2 follows the same pattern. This. Here. while the second column also has familiar entries: for each choice of k is the classical complex projective space with a complex dimension k .3. : mainly only topological.which will be used elsewhere. 14. is the 2-sphere S 2 ( i ) of radius one-half. In particular. X in Table 1. that is. These will be useful in reminding the reader of ‘what goes where’ and we here mention some of their most important features.1 the over-dimensioned shape spaces are emphasised by the use of lower case (r instead of upper case C.4. This is its name as a classical object. we introduce two important diagrams shown in Tables 1.1.In particular. where m = 1. in more familiar terms. is CP’(4). Xk. . save in a few trivial cases. where the mth entry is + c+’=+ Z m . For example. while the construction of the pre-shape is entirely elementary. but as a shape space we call it Xi. Note that. The accompanying Table 1. As already mentioned. m ) is the dimension o the corresponding shape space.1 and 1.1 is that the main diagonal consists entirely of. m ) in the table is the name of the shape space associated with k labelled points in m dimensions. and a bold f font is used to indicate the region k 3 m 1 in which shape spaces are not over: dimensioned. 9.4 SHAPES AND SHAPE SPACES x* configurations of k labelled points in ( k . but now the entry in position ( k .2. . indicated by the corresponding entries on the main diagonal in Table 1. while the further entries on the main diagonal hold the topological spheres of the dimensions 5 . To illustrate the non-triviality of shape spaces in general it suffices to remark that when k 3 2 the shape of k labelled points in the plane will turn out to be a point in the classical complex projective space CPkp2(4). . 2m ’ .

We shall also show. In fact. Another important feature arises from the fact that in the kth row of Table 1. a simple modification of Casson's proof would show that they are all topological balls. which is the corresponding diagonal entry.1 the spaces beyond the main diagonal are those Ek for which k m so. zero-dimensional space. that none of these spheres is a metric sphere except for the first two. and this two-point space is followed on the right by an infinite part-row of one-point spaces.. and it plays an important role in many later contexts. its two points being separated by two units. private communication). Casson's proof of this theorem will be given at the end of this chapter and alternative proofs will be given elsewhere in this book.SOME PRELIMINARY OBSERVATIONS 5 Table 1. as we have seen. the row labelled k = 2 starts with a twopoint.1 The array of shape spaces (the a-entries are 'over-dimensioned') ~~ ~ ~ k\m 1 2 3 4 5 6 7 8 9 10 This fact was first observed by Casson (1977. in Chapter 6. < . followed by a 2-sphere of radius one-half. the row k = 3 begins with a one-dimensional circle of radius one. this in its turn being followed by an infinite part-row of metric hemispheres having the same dimension and radius. Similarly. they are all identical to halved versions of Xi-. To take the simplest example. and we shall give an alternative proof in Chapter 2.

those that lie to the right of the column m = 2 and to the left of the spheres on the diagonal of the array. leads to a complete characterisation of the whole family of shape spaces. which are precisely ‘halves’ of the 5-sphere on the diagonal.1 the reader will notice that we have still to describe the shape spaces Xk in the infinite triangular region determined by the inequalities m 3 3 and k 3 m 2. equivalently. After inspecting Table 1. followed by a four-dimensional complex projective space of complex curvature 4. They have not yet been determined up to homeomorphism. and then by a five-dimensional topological sphere.2 Shape space dimensions k\m SHAPES AND SHAPE SPACES 1 2 3 0 2 5 8 11 14 17 20 23 26 29 32 35 38 41 44 47 50 53 4 5 6 0 2 5 9 14 20 26 32 38 44 50 56 62 68 74 80 86 92 98 7 8 0 2 5 9 14 20 27 35 43 51 59 67 75 83 91 99 107 115 123 9 10 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 0 2 5 9 13 17 21 25 29 33 37 41 45 49 53 57 61 65 69 0 2 5 9 14 19 24 29 34 39 44 49 54 59 64 69 74 79 84 0 2 5 9 14 20 27 34 41 48 55 62 69 76 83 90 97 104 111 0 2 5 9 14 20 27 35 44 53 62 71 80 89 98 107 116 125 134 0 2 5 9 14 20 27 35 44 54 64 74 84 94 104 114 124 134 144 However. no mo of the ‘diagonal’ shape spaces are the same.6 Table 1. The situation is highlighted in Table 1. For example. These particular shape spaces are truly peculiar in that they appear not to have occurred in any earlier contexts. topological balls. or. X. this last space having singularities.1 by the use of upper-case and lower-case sigmas. but to the left of the diagonal the dimension alone is not sufficient to distinguish between them. That space lies on the diagonal and is followed on its right by an infinite sequence of identical topological hemispheres. In later chapters we will make use of a ‘topological recurrence’ that provides useful structural information about all of them and. in particular. Since they have different dimensions. but in due course we will present the integral homology for each one and describe its global geodesic geometry as well as the Riemannian metric and associated curvature tensors. that is. It will be shown. and Xio each have dimension 34. + . in principle. when k = 4 the row starts off with a unit metric 2-sphere.

all shape spaces with k 3 m are topologically distinct. As an example we present the matrix . whose coordinates . x.I ) representing the new k-ad is obtained from X* by multiplying on the right by a special k x k matrix Q k . These facts were first established by making use of the exact sequences for shape space homology that we introduce in Chapter 5 and that results from the above-mentioned topological recurrence. . . in particular.x l we shall write as the columns of the matrix X * . . and Ei. j Xj*+l - and a striking feature of this construction is the progressive re-centring of xj.l relative to its predecessors XI*. an important consequence implied by (iii): the shape spaces located in the injinite triangular region of Table 1. where k 3 2.A MATRIX REPRESENTATION FOR THE SHAPE OF A k-ad 7 that the shape spaces in the infinite triangular region mentioned above possess the following interesting properties: (i) no one of these is a sphere. for each j > 0.< k . . rotation and dilatation have been quotiented out. . We can see that the matrix (Ax: XI . and that the shape of the k-ad is what is left when all effects attributable to translation. Note.1. Let us consider a labelled set of k points in OBm. We now orthogonally transform the k-ad X* as follows: and for 1 . even at the Z2-level. We recall that degeneracies are allowed except that we insist that the points are not totally coincident. .1 are all topologically distinct from one another. They are also distinct from those in the first two columns so that apart from E.. .. (ii) (iii) no two of them share the same homology. & . + x.3 A MATRIX REPRESENTATION FOR THE SHAPE OF A k-ad x. . x?.< j . . . The second equation also shows that. all of these spaces have torsion in homology. and indeed no one is even a homotopy sphere or a manifold. 1. which are different sized copies of the 2-sphere. . This follows from the form of Q k and it provides the main justification for its use. Xj is a scalar multiple of XI* + .

x ( k . 30=5~6. If we now examine the product . 20=4~5. For most purposes it is convenient to shift the configuration so that its centroid is moved to the origin of coordinates after which the matrix X * will have all its row sums equal to zero. Indeed. 6 = 2 ~ 3 . are as follows: 2 = 1 ~ 2 . However. For our particular choice it turns out that Q is k a rotation. these entries being repeated in the rows below in the same horizontal locations until just before the main diagonal is reached.8 SHAPES AND SHAPE SPACES Q k in the particular case k = 6 as follows: It should be noted that (i) in each column of the array the squares of the entries sum to unity.. x ( ( k .. . . (iii) the integers. (ii) each column is orthogonal to all of the other columns. it remains for us to show that Qk has determinant +1 rather than -1.1) divided by ((1 x 2) x (2 x 3) x (3 x 4) x . In fact. 12=3~4.1) x k>}"2. the square roots of which appear in the denominators in the second and later positions in row one. the fact that it is orthogonal is immediate from properties (i) and (ii) above. which reduces to + I as required. . In order to do this we start by adding to the top row of the matrix the sum of all the subsequent rows. . It follows that the value of the determinant is d i x 1 x 2 x 3 x . these properties suffice to specify the matrix Qk in the general case up to the sign of each column. . This yields a new top row consisting of followed by zeros.

the shape is to be identified with the equivalence class or ‘orbit’ associated with the left action of SO(m) on the pre-shape. We also note that our normalisation for size. 112 . in ( or O f f f f f f f f f f 0 0 0 + f f f f f f f .. we can if we wish transform the pre-shape matrix x = (XI x 2 . our normalisation implies that this pre-shape will lie on the unit (m(k . and we shall be free to represent each such class by any one of its members.1) real matrices with Euclidean ( m x ( k . 0 0 0 0 f f f f f f f i (ii) an upper semi-diagonal matrix with a strictly positive sign for the first non-zero entry in each of the first j < m . k to dividing by I 1i 1 I’} 2 . The result of this normalisation will be a matrix (0 X I { xfz. is equal to zero. We also note that.1))-space. dividing by now corresponds to dividing by { xf=l I I’} I Ix: x2 112 and. . .A MATRIX REPRESENTATION FOR THE SHAPE OF A k-ad 9 we find that it has the form because the first column of Q is ‘constant’ and each row-sum of the matrix k x. Thus. as in the further example with j = 2: . if we identify the space of m x ( k . all the entries in the last row as being 0 or f. X k . since Q is orthogonal. In particular. if we throw away the zero first column.1 rows.l ) .1) . with ‘zero’ rows below this.1) array modulo SO(m) acting on the left.1)-sphere in that space.. . we can exploit these procedures to yield (i) an upper semi-diagonal matrix that has a strictly positive sign for the first non-zero entry in each of the first m . we can represent the pre-shape by x = (XI x2 and then the shape itself is represented by this m x ( k . + . + x. .2 rows. + x. Thus.‘ xk-1) by using one of the rotations in SO(m) to perform various ‘left-hand’ tidying-up operations. and.. Xk-1) .

. equivalently. V is an element of SO(k . f l } with an even number of minus signs. A2. This allows us to present the preshape in the three-factor form U ( A 0)V. A) . We can re-write this decomposition in either of the forms + U ( A D 0) diag{D-'. . with A1 3 A2 3 .I ) and A is the m x m diagonal matrix diagV-1. In this formula the sum of the squares of the A's is equal to unity. . . .l. .4 'ELEMENTARY' SHAPE SPACES C AND C$ i We have already claimed that . and h.10 SHAPES AND SHAPE SPACES . . . - 3 lA. or. OJV. . Accordingly. O O 0 O + f f f f f O O + f f f f 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 f f 0 0 f f 0 0 f f 0 0 f f 0 0 1 Here. so that in that case we are left with (AD-' o)v. 1. on replacing each f by the actual numerical entry we get a tidied-up version of the original data. xk-1) is based on a 'pseudo-singular values decomposition' of X . In fact. where D is any diagonal (m x m)-matrix of the form diag{fl. O}V. . . f l . These transformations will be useful later. U is an element of SO(m). 3 0 unless k = m 1. Of course. when we come to perform our mathematical computations on these matrices it will sometimes be more convenient to have all the potentially non-zero elements at the beginning of each row. 3 A' . U D is a left-rotation and can be dismissed when we are only interested in the shape. U D ( A 0) diag{D-'. Another presentation of the pre-shape X = (XI x2 . 0 0 0 0 0 0 Such 'tidying-up' operations involving an (m x m)-rotation on the left will often be useful. .

2 = 1.x. . so we can think of the 2 x ( k . We begin with . and after dividing by the size s. of course. quotienting on the left by SO(1) is here irrelevant because SO(1) is the trivial group.XT). Next. We then find. l}. clear that we can divide out the size I.. using Q3. so that c. . For m = 2 it is. we consider a not totally coincident k-ad of points x. We turn next to the identification of the shape spaces Ck. =CPy4). x) and we carry out our standard reduction using Q2 to yield the singleton 21 = ‘2( x * .. and x -1 when xz < xT. Assuming that these are the coordinates after we have moved the centroid to the origin and normalised the size. where the components x1 = Xl/s and x2 = 221s of the shape will satisfy the equation XI 2 +x. and it tells us that E. xT and x must be distinct when k = 2. Here. that XT + 1 * XI = -(x2 * -xl) and 22 45 =-{24 1 & - (x? + . and this is So(l) as already noted.. .. When k = 2 we start with a non-degenerate point-pair (xT.‘ELEMENTARY’ SHAPE SPACES C: AND Z$ 11 and that c. Of course.. Accordingly. we find that Xf= {-1. z j . = Sl(1). This argument extends to general k . = SkP2(1). x i in two dimensions. we construct the pre-shape in the form following the specification of Q k . it is . } x).I/2/2 to get 1 when xz > xT. suppose that k = 3. and in this section we provide the evidence for these assertions.. with our conventions. c. confirming our claim for these spaces.. 2 Because. natural to think of each xi as a complex number.1) matrix X as .

. A particular identification already referred to is x. Then. We shall make frequent use of this both as it stands and also in a flat. . the shape is denoted by [XI rather than n ( X ) . . but for the sake of completeness we will set out the details below. that X. . form a subspace isometric with CPkP3 We have thus (4). Then. . but not isometric. 2n)J of complex numbers of unit modulus. Equivalently. - g*. . Zk-l/zj) 3 is invariant under the above action of SO(2). Ignoring the redundant entry 'l'. The resulting quotient space is known as the complex projective space Cpk-2(4). . in fact. on all the shape space except for the points where z j = 0. Although. if z j # 0 for some particular j .12 SHAPES AND SHAPE SPACES identified with the ordered set of complex numbers z = (ZI. in some papers. . Before entering into a more detailed study of X. Zk-1) by the group {eta : a! E [0. Now the metric identification of CP'(4) with S 2 ( i ) is a classical theorem. which we shall employ from time to time. in particular.. . These excluded points. z2/zj. that is. zj+l/zj. . = CP0(4).x H n ( X ) for the quotient mapping. . version obtained by stereographic projection from a point of the sphere S 2 ( i ) onto the tangent plane at the antipodal or some other point. zz. Note. not all the zj will be zero. Zk-1). To get the shape from this pre-shape we still have to quotient out the action of SO(2) acting on the left. Z Z . It is natural to choose the customary metric topology on SL and the corresponding quotient topology on Xi for which the open sets are the images of the SO(m)saturated open sets in S i . is S i modulo SO(m). . Since we have excluded the totally coincident k-ad. it will be helpful to note a few common features of the general shape spaces Xi. whenever (k. confirmed the identification of C!j in the column for m = 2 in Table 1. we can say that the open sets in Ck are those determined by the quotient metric p defined by the fundamental formula . the mapping n is continuous.For every k and m. which is just scalar multiplication of z = (z1. . this being a one-point space. this provides us with a local coordinate system. with this topology. m ) is not (2.. . where the '4' is the value of the complex curvature constant that is determined by our unit-size convention. we shall use the latter notation throughout this book. = CP'(4) = SZ(i). the ordered set of complex numbers (zl/zj. shape space Xk is compact and that it is connected whenever S i is connected.1. 9 zj-l/zj. where Sk denotes the pre-shape unit sphere Sm(k-')-l (1) and the rotations T in SO(m) act from the left on the matrices X that are the points of s. We write n :s . 1. 1). It follows that.

‘ELEMENTARY’ SHAPE SPACES Cf AND E : 13 in which each of X and Y is a pre-shape and. . is required. Y) = 2 arcsin which is equal to arccos tr(YXf) and so is. then it is convenient to use (21 1 2 2 ) < l 2 for the pre-shape. # 0. When k = 3 and 21 # 0 another useful coding of the shape is .*/& when x. where we always have O < p < n . When (1. 1) where ( = 21/22 and (iI IX - Y I I) . with 0 as the mid-point of xx :. horizontal and of unit length without altering the shape situation.2) are combined we see that for m 3 2 the distance p between two shapes n(X) and n(Y) is given by p(n(X). Alternatively. d(X. f We now return to the special. triangle in R2. if we normalise 0x2 to have length l/& in the previous paragraph. This d-metric on Sk is topologically equivalent to the norm-metric or ‘chordal metric’ inherited from ~ ~ ( ~ p l ) . we may reduce the specification to ( i ? ~ . then the resulting x will have coordinate precisely <. all except the excluded shape are represented uniquely by the resulting position of x.. & ) E R2 x R2 or ( 2 .* is obtained as in the previous paragraph. in fact.. When a more symmetrical parameterisation of the shapes in E.*) . . can be arranged to have the vector Ox. < 21 22 # 0. as before. This important result underlies the whole discussion of the metric geometry of shape spaces for all values o m 3 2. as explained above. and similarly for points in Ei with general k . case Ei and. because of the freedom to rotate and rescale the data we can specify the shape either by (i) ( 1 . In this way. We then have a number of different ways to represent the shapes of such triangles. 22) E C2 in the manner already explained. Then. n(Y)) = Tern&rn) arccos tr(TYX‘). x and$ to be the labelled vertices of a not totally degenerate . we take x:. and x coincide. but particularly interesting.x. when m 3 2. <) where = 22/21 and or by (ii) (<.* in the plane.1) and (1. If for the moment we leave on one side the special case when x.x. the great-circle metric on the unit sphere SL. It may be checked that the coordinate in (i) is x. then the triangle (x. Alternatively.

and this. however. Whichever of the above representations of the shapes is chosen.82)) cos + r1rz sin(& . as i1 6 <I = rle where rl and r 2 are real. modulo SO(2) acting on the left.81) (1 + rf)(l + Y.82) sin a a! JGiKT-3 We now need the maximum value of this ratio for 0 < a! < 2n: namely. say. <2) between these two shapes we calculate the trace of the triple matrix-product ( cosa! sina - s i n a ) ( 1 rlcos81) cosa 0 rl sin81 (r2 cos82 1 r2 sin62 divided by d(1 + r f ) (1 + r.C) that are related to . is the shape point (0.:. equivalently. these being identified.").) The value of the inter-shape distance p(<l. e'9. <2) is then the arc-cosine of the above expression. for general k . that is. This formula for the shape-distance. J 1 + rfr. simplifies if we introduce super-abundant shape variables (&6. we can find out what happens by letting r tend to infinity.which reduces to (1 + rl r2 cos(81 . or. which is the only shape excluded from the coordinate representation (i) above."). coincides with x l . 1) in complex coordinates. There the coded shape reduces to (: :. Now suppose that we are interested in two shapes. we must not lose sight of the fact that the shape really is the ratio z2/zlr and similar remarks apply when working with shape spaces Ck. when x.14 SHAPES AND SHAPE SPACES where r = 122/211. non-negative and finite. To find the shape-theoretic distance p(<l. When z1 = 0. (0. That version is often convenient for calculation and it leads to informative pictorial representations. which is necessarily always finite. + 2rl r2 cos(82 .

there is the group of six isometries induced by the permutations of the labels for the vertices of the original triangles. 0. with the point (0. then we obtain. ~ 1 . the six shapes obtained by labelling the vertices will occur at corresponding points in the six lunes.‘ELEMENTARY’ SHAPE SPACES Zf AND Ck. 6.C2)). Moreover. since 2C0s2p(<i. It is important to note that although this projection is not isometric it i measure-preserving and the metric distortions are s reasonably controlled. we see that the shape space X. Thus.1 = C O s 2 p ( r i . is indeed the sphere S2(i). C) satisfy the equation Thus. for each unlabelled triangle. (a. 0. for a suitable choice of cylinder. As a result. all possible shapes may be represented in such a half-lune. To visualise how the various triangular shapes lie upon this sphere we note first that SO(3) acts naturally on it as a group of isometries. C) is a point on the sphere S 2 ( i ) . replacing the compactification point at infinity. r 2 ) = 4 ( ( i i i . up to labelling and reflection. in the above representation. the outline in Figure 1. 8 ) by the formulae r cos 8 for then it will be seen that the new shape coordinates (& b. corresponds to mapping the ‘upper’ half of each lune onto its ‘lower’ half. This takes care of the otherwise excluded point (0. If we now project this lune onto a right circular cylinder that touches S 2 ( i ) along a great circle and then open out that cylinder onto a plane. The shapes of isosceles triangles appear along the other two i) i) . Accordingly. In particular. on the sphere. equilateral triangles at the upper vertex is maximally remote from them. 15 the variables (r.1 on which a selection of shapes are indicated at their representative points on the projection. C 2 ) . if we are only interested in the shapes of unlabelled triangles. In this diagram the shapes of degenerate triangles. (22. these may be specified in just one of these lunes. which. this one-to-one correspondence between the points of that sphere and the shapes in is an isometric correspondence. that is. In addition. those in which the three vertices are collinear. The result of this is that S 2 ( $ ) is split into six equivalent lunes with their common ‘upper’ and ‘lower’ vertices corresponding to the two possible shapes of equilateral triangles. the mapping of each triangle to its reflection induces an isometry 12 of the shape space. Then. lie along the base curve and the shape of the regular.62.<2).

If (x.1 Reproduced with permission from D. so that the equation of the translated sphere is u2 b2 c2 = c. We label the new coordinates a. Kendall. Copyright John Wiley & Sons Ltd boundary arcs with those whose third angle is less than n/3 on the left-hand arc and the third angle progressively increasing to n as they move down the right-hand arc. Barnett (ed). in V. 1) corresponds to the shape of those triangles that have their first two vertices coincident. 0. passing through the centre of the sphere. Interpreting Multivariate Date. The statistics of shape.O) between the sphere and the plane corresponds to the shape of those triangles whose third vertex lies midway between the first two. < + + .y. and the point of contact (0. We start by translating the above sphere of radius one-half upwards along the Z-axis so that it sits on the plane spanned by 0 6 and Ob with the point of contact at the origin 0 and OZ. is the b. The point N at the top of the sphere with coordinate (0. c. c).b. To compute the metric on the shape space sphere S2(k) in terms of the coordinate in (i) above we look first at the way in which stereographic projection links the points on the shape sphere to the points in the compactified plane. 0) is an arbitrary point of the supporting plane and if (a.16 SHAPES AND SHAPE SPACES Figure 1. where c # 1. 0.G.

Since a ~ 1-c = rcos8 and -.5 THE FUBINI-STUDY METRIC ON Xi As X. . INPI .. 0. b2. INql and so that the chordal distance from P to Q is equal to It2 . we write z = .y . b l . q are also collinear. Q. 1) and radius of inversion equal to unity.(1 I J(1 + I t 1 12)(1+ 1<212> and the corresponding arc-distance on the sphere is then the arc-sine of this last expression. Now let P = ( a l .y2. INPI 1 INPI. P . 1) and (x. It follows from this that __ .INPI . These arguments also show that in this planar representation the shapes of degenerate triangles correspond to the points with y = 0 and the points ( = f i are the shapes of the cyclically and the anticyclically labelled equilateral triangles. as X. then we shall have These formulae will reflect the fact that (x. 1. 0) be their projections in the plane. 0). - b 1-c we see the shape coordinate ( in (i) above is just x iy. c l ) and Q = (Q. 0) in the plane and (a.. b. p are collinear and N . we shall here be concerned with the spaces X!j for k 3 4 .lNPl - PQl IP91 INql INqI. As before.THE FWBINI-STUDY METRIC ON C i 17 other point of the sphere collinear with (0. 0.y . has already been considered.. c ) on the sphere are images of one another under an inversion with pole at N = (0. is a one-point space and. 0) and q = (x2. From the geometry of the inversion process we shall have + so that and thus N P Q and Nqp are similar triangles. so that yl N . c2) be two points on the sphere and let p = (XI.r sin 8.

x j v j ) j=1 j= 1 (1. we can always find an isometry that takes that shape into any standard shape that simplifies our calculations. . There is just one non-trivial element here and we can take it to be the conjugation that sends z to Z = (71. This is equal to a €[O. . . xk-1) Yk-1 ("1 u2 212 ''. We now show that is isometric with the complex projective space CPkP2(4).18 SHAPES AND SHAPE SPACES (z1... for any given shape.1 + iy. in particular. These symmetries preserve the pre-shape sphere and commute with the left action of SO(2) and hence induce isometries of the shape space. ?&I)..1) matrix. for z j and u.. This reduces to k-1 k. As a start we consider cos p ( n ( z ) . j=1 j=l k-1 j=1 k-1 - . . This includes the operators that permute the columns of the matrix. We do this by calculating the Riemannian metric and taking into account the global implications of the symmetries noted above.Wk-1)' C I Z j 1 2 ..2n) { ( cosa sina - sina cosa ) (XI x2 y1 y2 . . . zk-1) E Ck-' and w similarly for the result of our standard normalisation under translation of two k-ads in R2. . + iv.= 1 IWj12 C In more detail this is max tr a€[0.l ) / U ( l ) acting on the right. We begin by noting two relevant symmetries.vj + cosa C y j v j j=l k-1 j= 1 j=1 k.1 co sa C x j v j + sina C yjv. The first of these is 0(2)/S0(2) acting on the left. This means.sina C xjuj + cosa C yju.. that the shape space is homogeneous and so. "' uk-l)'} vk-1 v1 divided bv where we have written xj k.1 + yjw. .n(w )). .) + sina C ( y j u j .=l sina C x. for wj. . z ~ . cosa C xjuj + sina C y j u j j= 1 k.~ ) ( w . In matrix terms this is the operator that switches all the algebraic signs in the second row of the 2 x ( k .. 2n) max tr e"(z1.4) divided bv ..1 max tr . . . The other symmetry is U(k .I.

we write wj = zj + E ~ Then. we may assume zl # 0. n(w )) is Using the fact that.2: . a standard calculation transforms this expression in terms of the z-coordinates into the following in terms of the shape space coordinates <j = zj+l/zl. First. between the shapes n(z ) and n(w ). we find that up to quadratic . in an obvious notation. j = 1 . using the isometries that permute the coordinates zj. It is now a simple matter to calculate the Riemannian metric for Ck. sinp E p.THE FUBINI-STUDY METRIC ON C i 19 We now carry out the maximisation for (1. Then. this gives Then. terms in the ~j the value of sin2 p ( n ( z ). k . . n(w 1) = This gives us the distance p in Xk..4) and so obtain and accordingly we have cos p ( n ( z 1. . . for small p. .

and it follows that the orbits of A under left multiplication by O(m) and SO(m) coincide. if A is non-singular. Hence. if A is singular. which varies continuously with A.Cz+'. We also recall that the dimension d.which we know to be S2(i). since for A E X there exists T E O ( m ) such that TA is positive semi-definite and then TA = f(TA) = f(A).l ) ( m 2)/2. so that we obtain a continuous map + + f :X - Y+. is a sphere of curvature 4.+' of Cz+' is ( m . Let X denote the space of all non-zero m x m real matrices A.Whereas. Moreover. the matrix A'A is symmetric and positive semi-definite and so has a unique square root. X.1. two matrices with the same image under f must lie in the same O(m)-orbit in X. f ( A ) . is homeomorphic with a sphere of dimension Proof. Shape space X : + ' dz+'. this map sends all the points of each O(m)-orbit in X to the same point of Y+. and The value '4' for the curvature constant is not surprising when we recall that Xi. the induced maps continuous. X+ the non-linear subspace consisting of those for which det(A) is non-negative. For any m x m matrix A. 1.Conversely. which is one less than the dimension of Y. they induce bijections - f* : x * / s o ( m > between the spaces of SO(m) orbits in X+ and the points of Y+. then its O(m)-orbit splits into two SO(m)-orbits.20 SHAPES AND SHAPE SPACES This is exactly the classical Fubini-Study metric for the complex projective space CPkP2(4) so E. for any T E O(m). are with respect to the quotient topology on &/SO(m). we are interested in the SO(m)-orbits and. Each of the maps f * = f l X + is clearly surjective since f+(A) = A and f-(TA) = A for any A E Y+ and any T E O ( m ) \ SO(m). However. one in X+ and one in X-. - Y+ 7* .the subspace of those for which det(A) is non-positive and XO the subspace of those for which det(A) is zero. In all cases we give our spaces the topology they inherit as subsets of Rm2 with the standard Euclidean metric. then A = TA where T is the reflection in any hyperplane containing the column space of A. Let us also write Y for the m(m 1)/2-dimensional linear subspace of RmZ corresponding to symmetric matrices. (TA)'TA = A". can be identified with that classical space for all values of k . Y+ for its nonlinear subspace of non-zero positive semi-definite symmetric matrices and YOfor the subspace of singular matrices in Y+. Since.6 THE PROOF OF CASSON'S THEOREM We now present an expanded version of Casson's proof of the 'spherical' character of the shape spaces . in Y+. Theorem 1.

which maps are again continuous and clearly bijective: j : &/similarities * Y+/dilations. further composing with the ‘vertical’ projection of the upper hemisphere %+ onto the lower hemisphere X . is itself compact and the sphere Sd:” is Hausdorff so the continuous bijection g is. Since A has an orthonormal basis of eigenvectors. in fact. we have further induced maps between the induced maps f + and the corresponding quotient spaces. then. obtain continuous bijective maps we g* : X*/similarities --+ ‘H*.hi) < 1. respectively.h)Z hA : 0 < h < 1/(1 . which agree on the common subspace Xolsimilarities. and hence it commutes with Thus. the segment {(l .. We shall show below that there is a continuous bijective mapping from y+/dilations onto the upper or ‘northern’ hemisphere X+ of the standard dE+’dimensional unit sphere taking &/dilations onto the ‘equatorial’ (d. To complete the proof of Casson’s theorem it remains to construct the continuous bijection of Y+/dilations onto the dz+’-dimensional hemisphere W+. Then ef((1 . For any positive definite matrix A let ei be an eigenvector of norm one with eigenvalue hi. . the multiplicative group of positive reals acting by scalar multiplication. being the continuous image of the compact pre-shape sphere Xldilations.= 1 - h +A&. and. in the latter case.Taken together these maps give us a continuous bijection: g : X/similarities --+ sdE+’.THE PROOF OF CASSON’S THEOREM 21 It remains to quotient by the action of the group of dilations. However. provided p < 1. mapping it onto yo/dilations. which is positive if h ( l . X/similarities. taking yo/dilations onto the equator No.h)Z +U J e . Composing this map with j+ and j . Now Y+ is a convex subset of the linear subspace Y of Rm2 and it contains an open neighbourhood in Y of the identity matrix I .. 7. a homeomorphism. if p = p ( A ) is its least eigenvalue. that is.p ) ] . mapping it onto the equator No.+’ 1)-sphere. + which ends at the singular matrix is the complete intersection with Y+ of the ray from Z through A. it follows that. We note that these maps agree on the common subspace Xolsimilarities. This commutes with the action of SO(m) and also with the map f.

is in fact a rn homeomorphism. the segment from I to any singular matrix A in YOlies entirely in Y+. then we have a continuous bijective map of %+ onto Y+/dilations which. However.However. A lies in Y+ and only a finite closed segment of the ray from I through A lies in Y+. Moreover. being determined as above by the least eigenvalue. which we are excluding. has its least eigenvalue less than one and so lies on such a finite segment from Z to a singular positive semi-definite matrix. as %+ is compact and the cone on Nrepresenting Y+/dilations is Hausdorff. these are all represented by the matrix I. Thus. every dilation class in Y+ is represented uniquely on such a segment from I through N. except for that through 0. Note that the length of these finite segments. is a continuous function of the choice of A in N. Conversely. Thus. Its inverse is the map we require. then the entire semi-infinite ray from Z through A lies in Y+. if we map the ‘north pole’ of H+ onto the matrix I and the ‘longitudinal’ arcs from the pole to Ho onto the segments of Y+ from I through N. except for 0 itself. let N be the (dz+’ . . of radius E and lying in the hyperplane through . We may choose E > 0 sufficiently small that for every A E N all the eigenvalues of A lie between 0 and 1.22 SHAPES AND SHAPE SPACES If p 3 1.More precisely.1)dimensional sphere in Y centred on iZ. with only A itself lying in YO.I perpendicular to iZ. in the same dilution class as A. it is clear that for every non-singular matrix A in Y+ there is a real h > 0 such that M . the zero matrix. since the points of any segment from I are dilation-equivalent to those of one through N.

We shall mainly consider the shape spaces introduced in Chapter 1. we may also add the spaces that lie above the diagonal in Table 1. LE Copyright 0 1999 by John Wiley & Sons Ltd CHAPTER 2 The Global Structure of Shape Spaces In the next sequence of chapters. these latter spaces are simple quotients of the former ones and. we may add the ‘diagonal’ spaces Z{+’ to our list of known spaces since. Recall that shape space Z i is the quotient of the unit pre-shape sphere S in A by the appropriate left action of SO(m). that is. which is a continuous bijection with a continuous inverse. In fact. BARDEN & T. G. These are the only shape spaces that we can identify with other metric spaces that have already been sufficiently well studied in their own right to be able to be regarded as ‘familiar’. we shall find it convenient to work with both types of shape space at the same time. Then. since these are all topological hemispheres. these are all homeomorphic to spheres. In Chapter 1 we showed that the spaces Zf were isometric with the unit spheres SkP2(1) the and spaces X$ were isometric with the complex projective spaces CPkP2(4). Z i has the natural induced Hopf metric. identification via a homeomorphism. KENDALL & D. .1. 2 1 THE PROBLEM . K.Shape and Shape Theory D. which are at once those for which most is known as well as those that are involved in the most applications. Chapters 2-5. by Casson’s theorem. the spaces that arise from identifying a shape with its reflection are also involved. since SO(m) is compact and so are its orbits in S$. C A R " & H. which we denote by p. Similarly. Rather than metric equivalence we could ask only for topological equivalence. we turn to the global description of shape spaces so far as we are able to achieve it. where ‘4’ is the complex sectional curvature. particularly when computing homology groups. However. although they are of limited practical interest. for which the distance between two shapes is the minimum distance between any two corresponding pre-shapes. in discussing them. The sphere S$ has a standard metric in which the distance is measured along great circular arcs and.

there is a sense in which our spaces are already as simple as they could be and. and the correct set of building blocks. that is. that is. should be regarded as elementary spaces themselves. the simplest global invariant is the dimension. ‘diagonal’. to identify the spaces up to homeomorphism. subgroup of its isometries. although this description will not be fully detailed in the sense of giving full instructions for anyone to follow in building a copy of the space. we find global topological invariants that will allow the builder to check whether he has assembled the pieces correctly. This is something we may achieve for shape spaces.whereas. First. as was mentioned in Chapter 1 and will be proved below. another aspect of familiarity is the feeling of ‘having been here before’: the standard spaces arise so frequently and in so many different situations that we already have a good deal of information about their nature and behaviour. we have identified shape space with a space having a much simpler description. that does not get us much further since most familiar spaces are manifolds. For example. In fact. no one has ever studied any of the remaining unidentified shape spaces in any other context. Instead. to prove that we had done so. First. Indeed. Each is the orbit space of a sphere under a very important. however assembled.24 THE GLOBAL STRUCTURE OF SHAPE SPACES However. we must somehow render them more familiar ourselves. it will be helpful first to look at the coarser global topological structure. This is something we are very unlikely to be able to achieve for the remaining spaces. to the best of our knowledge. and. The converse. 2. If the invariants are wrong. but it is also true of the complex projective spaces if we think of them as the spaces of complex lines in the next higher dimension complex space. However. This is certainly true of the spheres and hemispheres.2 WHEN IS A SPACE FAMILIAR There are perhaps two aspects of familiarity shared by the above identifications. none of the shape spaces other than those explicitly mentioned above is homeomorphic to a manifold. we would need to show that they could only all be correct if the entire construction were correct. Topologists have certainly studied many spaces that are not manifolds. What then are the properties we should endeavour to elucidate here? As we shall be studying the fine local Riemannian structure of the spaces in Chapter 7. but we shall work towards it from two complementary directions. is not true. will always have the correct dimension. they have a dimension n such that every point has a neighbourhood homeomorphic to an open subset of Rn. . There are just not enough simply described spaces to hand. we must build up our own databank of ‘previous experiences’. having arisen in the natural way we have described. we shall describe how the spaces are built up out of standard elementary building blocks. these are spaces with very particular properties. Thus. he knows that he has not assembled the space he intended. To complete this classification program we would need to find sufficient such invariants and. constructed explicitly for special purposes. except that. However. since they have not been studied before. unfortunately. we shall not strictly achieve that aim.

(Sdw-') is contained in the union of the cells of lower dimension. we only need give the following simplified formal definition. or closed d-cell. automatically satisfied if the complex has only finitely many cells and. since that will be our case. and such that h. The main global invariants we compute will be the homology groups. at the structure and basic properties of CW complexes that were introduced by Whitehead precisely in order to describe efficiently. is called an open d-cell. one or two d. Then. the topological properties of spaces such as ours that were not quite. Moreover. however. in fact. A given Hausdorff topological space P is a finite CW complex. We should look first. This decomposition is derived from the following decomposition of unit spheres. for each cell em! in h. let ei denote the unit vector in the positive direction along the ith coordinate axis. They are. . : B d m +-P.A CELLULAR DECOMPOSITION OF THE UNIT SPHERE 25 2.4 A CELLULAR DECOMPOSITION OF THE UNIT SPHERE Turning our attention to Sd in Rid+'. we shall first undertake a preliminary assembly of the cells into larger groups of cells that render these computations more transparent and. a cell complex is a particular choice of such data and provides a ceZlulur decomposition of P. a CW complex is a space that may be constructed inductively by 'attaching closed cells' of increasing dimensions. which restricts to a homeomorphism of the interior of Bdm with e . as well behaved as manifolds. Then we take zero-cells. Cells may be further decomposed and re-combined so that a cell complex has uncountably many possible decompositions. we mean any homeomorphic image of the closed unit ball Bd in Rid. which we shall apply to the pre-shape spheres.. or $finite cell complex. but almost. its inverse image hi'(e. and deal effectively with. then. For this to work in general.(Sdm-l). 2. Among those for shape spaces we shall choose ones that have the additional special property that each h. each mapped homeomorphically into the shape space by h. More precisely. The union of the cells of dimension less than. Its interior. By a closed cell of dimension d . certain additional conditions have to be observed that account for the adjective 'CW'. or vertices.3 CW COMPLEXES The building blocks we shall use will be topological cells and the building regulations we shall observe are those for CW complexes which are described below. at the same time..-cells in Sdm-'. or equal to. if it may be expressed as the disjoint union of open cells {e.r) will be a finite disjoint union of. : w E Q} in such a way that for each w E Q there is a continuous map h. n is called the n-skeleton of the complex and the idea is to prove results about CW complexes by induction over its skeleta.(Sdm-') is the union of a set of cells of lower dimension. however. so that it is the disjoint union of the corresponding open cells. give some additional insight into the structure of our spaces. homeomorphic with the open unit ball.

1 indicates the decomposition of the upper hemisphere xz)-plane and one. if ~i take the values f l . We shall indicate this by letting + + . = f l and the ij are all distinct. then for each pair (Eiei. Thus. . and that the other conditions for a cellular decomposition are satisfied. to distinguish them. . We shall refer to the cells in the above decomposition of a sphere as basic i +~) cells. . ) . e is~determined as the set of i points on the sphere for which x . together with four more 1-cells and four 2-cells. ~ i ~ + . = 0 otherwise. Four 1-cells lie in the plane of the diagram and four others lie above it. Figure 2. + e j ) of vertices that determine it. ... usually most. projecting onto the obvious coordinate space of dimension n . . Note that the basic cell (&ileil. isometric with each other.1 and xi. we shall refer to as elementary cells. there is a unique n-cell on the sphere having precisely those n 1 zero-cells on its boundary. it. -e3. this 1-cell. E j e j ) there is a unique 1-cell in Sd having these two zero-cells for its boundary: namely. for example. ej = +1.= 1. Next.. > 0 if ~ i . For example. The lower hemisphere has one more vertex. where E . in fact. and which. is the part of that circle for which x < 0 < xj. above of S2 in R3. each i . if E . We shall require cells that are unions of these basic cells. It has five zero-cells. +e3. They are determined similarly but with some. x i . " + . the appropriate part of the unit circle in the ( x . four in the ( X I . All the cells of a given dimension are. . Similarly. There are then four 2-cells that also lie above the plane. It is easily checked that the closure of each of these n-cells is indeed the homeomorphic image of a closed ball by. of the non-zero coordinates unrestricted. .26 THE GLOBAL STRUCTURE OF SHAPE SPACES at each of the points f e i in S d . . x j ) plane. ~ i ~ + ~ e . = . none of which are indicated. then the corresponding elementary cell will have both vertices ei and -ei on its boundary. . if xi is such a coordinate. < 0 if E. for any set of vectors ( ~ i e. = -1. Thus. which we shall denote by the pair (--ei.

then it determines an elementary n-cell lying on that sphere. all preceding rows must still end in rather than ‘&’. We recall that. It is easy to see that such a canonical choice always exists and is unique. and for which. the four upper 1-cells (+el. +e2) is the union of the basic cells +e2) and (-el. *e2. Note that. from among the entire ‘+’ ‘+’ . We shall first use it directly to gain some insight into the global topological structure of shape spaces. Thus. ( . Similarly. +e3) is the union of the basic cells (-el. if all the ~i in the symbol (&il ei. by ‘&’ followed by zeros and all other entries by ‘*’. + 2. the symbols for the basic cells in an elementary cell are formed from the symbol for the elementary cell by allowing the ‘*’s independently to take the values +1 or -1 or the corresponding vertices to be omitted. + q ) The elementary cell (*el. Note that this allows for the possibility that the final row or rows in a pattern matrix consist entirely of zeros. then ) it represents the n-sphere in the relevant (n 1)-subspace. the elementary 1-cell (*el. +e3) and (-el. the final non-zero entry in the last row. entire open upper hemisphere comprising the vertex e3. recalling that we are writing n for the quotient map of the pre-shape sphere onto shape space and n(X) for the shape of X. except possibly in the last row.THE CELLULAR DECOMPOSITION OF SHAPE SPACES ~i = 27 *. Rather than the form that we discussed in Chapter 1. +e3). together with their common vertex +e2. .1) pre-shape matrix X we can obtain a matrix can(X) that is in canonical form and has the same shape as X . the last non-zero entry in each row occurs before that in the preceding one. However. +e2. We then define the pattern patt(X) of the pre-shape X to be the symbolic matrix. such as (. together with the common boundary 1-cell ( . the last non-zero entry in each row is strictly positive.5 THE CELLULAR DECOMPOSITION OF SHAPE SPACES We now study the cellular decomposition of shape spaces that was already implicit in our discussion of the matrix representations of shapes in Chapter 1. using the row operations that are achieved by left multiplication by suitable elements of SO(m). ein+l are ‘*’s. :: : ):> f O O O O O * * * * + o O which is obtained from can(X) by replacing the final positive entry in each row except the last by followed by the appropriate number of zeros. if it exists. .T.e l . if at least one E L is f l .e l . what we have done is to select. +e3) is the . and the four upper 2-cells. In other words. we shall now choose the canonical form that is ‘reversed echelon’: that is. in the following chapters. we shall obtain from it a chain complex with which to compute the homology and cohomology of the spaces. forming the upper open semi-circle in the (XI. -e2. xz)-plane. +e2). . . . However. . Chapters 3-5. from any m x (k . *e2. Thus. Then.

which are -1.28 THE GLOBAL STRUCTURE OF SHAPE SPACES SO(m)-orbit n-l (n(X)) of pre-shapes on the pre-shape sphere that have the same shape as X. in order for the resulting matrix to lie on the pre-shape sphere. however. We refer to the images CJ = T C ( O J ) as cells in shape space. so that n(07)= C J . We can deal with the second type by premultiplying by the diagonal matrix in SO@). is. we shall have to be more precise when we wish to compute the homology groups. that the sum of the squares of all these entries be equal to unity. in fact. can be found from patt(X) by replacing ‘+’ by a strictly positive number. The set of all such pre-shapes Y sharing a common pattern J = patt(X) forms one or. from the number at certain of its points. both can(X) and patt(X) are functions of the shape n(X). where the ‘-’ in 7is in the ith row. or to 7. of course. which shares the same pattern as X. they correspond to a symbolic matrix L that fails to be a pattern because two of its rows are of equal length. all of whose diagonal entries are +1 except for the ith and the last. that not all elementary cells correspond to patterns. negative or zero number with the additional requirement. Since each shape has a unique canonical pre-shape. TC is an open mapping and so the bijection. which is the restriction of n to each relevant elementary cell O J . we find that they are either cells OJ. However. Note. then. If we examine the cells in the boundary of OJ in the pre-shape sphere. Similarly. For the moment. corresponding to other patterns J’ with dim(oJ() = dim(aJ) . is again a lower dimensional cell in shape space. replacing f by a strictly positive or strictly negative number and replacing ‘*’ by a positive. The dimension of the cell OJ is one less than the number of non-zero entries in the pattern J. so that we are making a very restricted selection from the full set of elementary cells available on the pre-shape sphere. since each point of an elementary cell corresponding to an allowable pattern represents an entire SO(m)-orbit. any pre-shape Y in canonical form Y = can(Y). or they are cells that correspond in a similar manner which is a pattern J’ with one of its ‘+’ entries replaced by a ‘-’. which differs. each of these spherical cells is mapped injectively into shape space by the quotient map TC. such a cell OJ must have a codimension in the pre-shape sphere at least equal to the dimension of SO(m). if the pattern has a final ‘&’ entry. no two distinct cells having a common image point and the union of the images of all the cells being determined by the various patterns covering the entire shape space. This simultaneously reduces all the pre-shapes in that cell to the canonical forms that belong to O J ! . a homeomorphism. two of the elementary spherical cells OJ on the pre-shape sphere that we described above. the ambiguity as to whether OJ denotes one or two cells will not matter. the unique one can(X) that has our chosen canonical form and that then determines patt(X). Conversely.1. On the other hand. The first of these types obviously maps to a lower dimensional cell or cells in shape space. Since SO(m) acts as a group of homeomorphisms on the pre-shape sphere. if we reduce the pre-shapes that correspond to L to . Clearly.

Then. This reflects the fact that on shape space ZL we have the isometric involution I. . we may think of EL as being embedded in Ck+‘. Thus.. and so forming a ‘tower’ of spaces as k varies in a manner analogous to that of the well-known special case of the projective spaces C. we note that. since this leaves fixed all the cells determined by patterns with zero last row and interchanges the two cells determined by each of the remaining patterns. that. Let T be any element of O(m)\SO(m). We shall denote it by 1. Each cell of Ek maps bijectively to a similarly named cell in EL+.. although what we get is an elementary cell on the pre-shape sphere. if we map Sk into S. together with others that are obtained by replacing the new ‘*’ by ‘0’ and repeating the procedure and. in particular. induced by appending n 1 rows ‘*’ + . It is the union of two such cells. -1) to induce i . which arises as follows. whose final entry is ‘*’. Similarly. in the pattern by ‘0’ to reveal a final ‘*’. with all the image cells being distinct. this commutes with the left action of SO(m) and so induces a map of Ek into EL+’ with each cell of the former mapping bijectively onto a similarly named cell of the latter. we shall have expressed the boundary of CJ = ~ ( C J J as a union ) of cells of lower dimension. which. .. We note.INCLUSIONS AND ISOMETRIES 29 canonical form./tm. is precisely Z. That is. corresponding to the patterns in which the new ‘*’ is replaced by ‘+’ and ‘-’. . If we use the element diag{1. this isometry is independent of the choice of T and is an involution. Finally. it is clear that. at the end of the day. and so we have established all the requirements to express shape space as a finite cell complex.. Its inverse image in S:. induced by appending a row of zeros at the bottom of each pre-shape matrix.. is the set of matrices of rank at most m.+’ by appending a column of zeros on the right-hand side. when determining the cells in the boundary of CJ and replacing an entry ‘+’. the image of X k .. if necessary. it is not one that corresponds to a pattern. then we find that their images comprise a union of cells of lower dimension. except that the two cells in Zk determined by a pattern with final entry ‘*’ in the final row both map to the single cell in corresponding to the pattern in which that is replaced by ‘+’ before appending the row of zeros. since shape space has a metric compatible with its topology. respectively. see that the image of EL in EL+. 1.n in Xi+’. : Moreover. maps each SO(m)orbit onto another such orbit. since SO(m) is a normal subgroup of O(m). Nevertheless. or ‘&’. reducing to canonical form again. it induces an isometry of X onto itself.. On the other hand.6 INCLUSIONS AND ISOMETRIES This cellular decomposition of shape spaces already reveals or confirms some of their important properties. under the mapping we induced by appending a row of zeros at the bottom of each pre-shape matrix. 2. the mapping of Ek to Ck. it is certainly Hausdorff. behaves slightly differently. Note that. . left multiplication by T is an isometry of the pre-shape sphere.

acting on CE+' can be represented by an involution that changes the sign of only one coordinate. since it preserves rank. xk-1) : CiZl2 = 1} xi and 11 changes the sign of all the coordinates simultaneously. since T acts as column operations. where n > 0. . . we note that the image of E+ in C::. We therefore refer to the latter sequence of subspaces as a strat$cation of Xk.1) is an isometry that commutes with the left action of SO(m) and so induces a mapping T . multiplication on the right of the pre-shape sphere by an element T of O(k . if X = U ( A 0 ) V is a pseudo-singular values decomposition of X E S$.. the subspaces C in X. On the other hand. is C. itself. where k > m I . in order to compute homology.As a special case. + + 2. an isometry with respect to the induced metric p. we identify each configuration with any of its reflections.-n.30 THE GLOBAL STRUCTURE OF SHAPE SPACES of zeros. At the other extreme.. as there are no cells :' in the decomposition of ZETi that are not images of those in C + . this is the :' entire image. As for the two-dimensional case described in Chapter 1. : are not geometrically determined in the way that the subspaces Ek-n/~m-n are. is isometric with the quotient EL-n/i. equally. Although this evidence falls well short of a proof. For example. its induced action T . re-scaling and rotation. we shall need to define orientations of the cells and just how 1. must restrict to an isometry of each subspace Ck.7 SIMPLE CONNECTIVITY AND HIGHER HOMOTOPY GROUPS A space is said to be simply-connected if every loop. 11 is the antipodal map of the sphere and the resulting quotient is well-known to be the real projective space RPk-2. In the case of X : / L ~we can identify the quotient explicitly. and T is V-' composed with the permutation that reverses the order of the first m columns. which is the image of the set of matrices of rank at most m . in it may be shrunk to a point. We shall be meeting the involution 1."++'. These quotients are. ( n ( X ) ) lies in C ' and : + further isometries.. which is. that is.-n onto k. Thus. shape spaces in their own right. . then T . Then. ."+'/L. . or. but. They are the spaces of shapes of configurations for which. of Ek onto itself. Czzi E X:+'/L. Thus. : X being a topological ball. 11 -+ Q such that h)pxlO)= f . affects those orientations will be important for the homology calculations. induced by swapping the mth column for the (n . k-1 Ct is identical with its pre-shape sphere Sk-2 = { ( X I . we say that two continuous maps f and g between topological spaces P and Q are homotopic if there is a continuous map h : P x [O. Since the group SO(1) is trivial.pn/~. We shall meet this again when we study the Riemannian structure of shape spaces. on X. .l)st. 1. in addition to translation. again. therefore. More precisely. In other words. will take it into Ck for any n such that m 1 < n < k . There is another significant difference between the above two types of inclusion. by which we mean the continuous image of a circle. of course. it is certainly consistent with the quotient of the action of 1.n in S$+l. on the topological sphere E.

it suffices for simple connectivity that the 2-skeleton contains a simply-connected subcomplex that contains the 1-skeleton. has only one zero-cell: namely. If we consider the points on the pre-shape sphere whose non-zero coordinates are in the top left-hand corner. and so to give an elementary self-contained proof of simple connectivity. is that represented by the pattern whose only non-zero entries are (* * +) in the first row. our cell decomposition of shape spaces Xk. a loop in a CW complex is homotopic to one that lies in the 1-skeleton. we know to be homeomorphic with a 2ball and so simply-connected. we already have sufficiently detailed information on the structure of shape spaces to prove the above special case of Casson’s theorem directly. These three cells form the entire cell decomposition of X i . if to a simply-connected CW complex we attach further cells of dimension greater than one. that a pattern represents is one less than the number of its non-zero entries. radially. A similar argument shows that. It is a typical theorem for CW complexes that. then it remains simply-connected. for k 2 4. Such a map h is called a homotopy between the maps f and g. or cells.SIMPLE CONNECTIVITY AND HIGHER HOMOTOPY GROUPS 31 and h ( P x { l ~ g. in any case. the 1-cell is the image of the cell comprising the points where b > 0 and c = 0. if m and k are each at least three. c > 0. and the 2-cell is the image of the cell comprising the points where b > 0 and c > 0. which. In fact. off the interior of the cell altogether. then the zero-cell is the image of the point where a = 1 and so b = c = 0. corresponding to the pattern with the entries ( * ) in the first row and zeros everywhere else. that which corresponds in the top left-hand corner. since any segment of a loop that lies in such a cell can be deformed by a homotopy to avoid the centre and then moved. every Xk with m 3 3 and k 2 3 must be simply-connected: it is formed from this ball in its 2-skeleton by adding further cells of dimension at least two. to the pattern with the single non-zero entry Similarly. Since the dimension of the cell. Thus. Then. The part of its boundary given by c = 0 maps bijectively onto the 1-cell in shape + + . there is also the 2-cell represented by the pattern whose only non-zero entries are ‘+’ + in the top left-hand corner. are simply-connected. We can now deduce quite easily that all shape spaces. there is just one 1-cell. a space is simply-connected if every loop in it is homotopic to a constant loop. where m is at least two. c 2 0 and a2 b2 c2 = 1. In fact. Hence. except for the circle C:. = Then. where b 3 0. This 2-cell on the pre-shape sphere is the quarter-sphere lying in the quadrant b > 0. the only further two-dimensional cell. by Casson’s theorem.

:) Thus.1)-connected. the first nonzero homology group is isomorphic with the first non-zero homotopy group. Thus. Xk is (d:+' . we shall show in Chapter 6 is certainly not a metric result. simple connectivity implies the vanishing of the first homotopy group and. It follows that the image. it follows that the space C:. In this case. Since the image of X. where b 3 0 and both a and c are unrestricted except that a2 b2 c2 = 1. showing that Ez 2 B2. we show that Ek has the same connectivity as that which we already know for C:+' by Casson's theorem. in later chapters we shall compute the homology groups of our spaces and. On the other hand. is equivalent to it. all these points map to the unique zero-cell in shape space. usually taken to be connected. may be constructed from C: by attaching the additional 2-cell determined by the pattern (1 :). we shall be showing that. is defined to be the set of homotopy classes of maps of S" into P fixing some chosen base point. Thus. As far as shape spaces are concerned. We shall first give a geometric proof of this.32 THE GLOBAL STRUCTURE OF SHAPE SPACES space. for a connected space. There are similar elementary proofs. with the quotient topology. is a closed ball with a closed circular arc on its boundary collapsed to a point. for any point with b = 0. This is easily seen to be still homeomorphic with the ball. we shall be able to extend these methods to obtain an alternative proof of Casson's theorem. a space that is connected and in which the first n homotopy groups vanish is called nconnected. for k > m. However. . which is also the 2-skeleton of Xi for all k > 2. where we first met Ef and Et. we have a hemisphere in the pre-shape sphere with its entire boundary identified with a single point in shape space. The nth homotopy group of a topological space P. Simple connectivity has a natural generalisation that arises as follows. where we recall that d:+' is our notation for the dimension of EE+'. for a connected and simply-connected CW complex P. 2 s2. Then. in Xi is homeomorphic with the quotient of X by the : involution 12. which interchanges the two top dimensional open cells and fixes the rest of the space. of the homeomorphisms Cf Z S k P 2 . except for the first two cases. we should note that our current methods have lost the metric precision of those of Chapter 1. directly from the cell decompositions. which. That is. However. X. confirming the homeomorphism + + X. the corresponding canonical matrix is (. Thus. However. analogous to the above direct proof of simple connectivity. is the image of the set of points on the pre-shape sphere with coordinates (: :).

P * Q.instead of simply n.For example. 11 x Q under the equivalence relation: ( p . (dE+' . As just mentioned. 11) U Q under the equivalence relation: ( p . 1) = ( p ' . the mapping cone contains a subspace homeomorphic with Q . although it does identify each set f . P x {0}. This is the quotient space of P x [0. q ) E ( p ' . 1. q ) for all p . It is the quotient space of ( P x [0. Q ) .S ( P ) = So * P. q ) E ( p . It is important to observe that this equivalence relation never identifies two separate points of Q. Equivalently. p' in P. of the cone is identified with its image under f . When Q is a single point the join P* Q is called the cone on P. The former we shall find is closely related to the Riemannian structure that we are able to put on shape spaces. starting from its subspace Xi-'. we see that P * S"-' is the nth iterated suspension S n ( P )of P. The essence of our homology computations is to identify which cells in the above decomposition actually affect the homology.(P. and these are indeed all of dimension at least LIE+'.in which we may identify P with the copy P x (0). 0. The complement C(P)\P is the open cone Co(P). S(S"-l) = S" and. S(So) = S'. even when using it to compute homology groups we shall find it convenient to concentrate just on those cells that are essential to express the homology and to discard a large subcomplex that makes no contribution to it. in fact.The other topological construct that we need is a variant of this cone. p' in P and q. denoted by C f (P. There is. we obtain the suspension of P. for example the zero-sphere. 1. 0) = f(p) for all p . where each point in the base. then P * S"-' is a (d n)-dimensional ball. q' in Q. which we shall usually identify with Q itself.for the + Uf . Thus.The first k' of these constructs that we require is the join. In the special case that Q is two points. We can now show how to express Xk as a mapping cone. Since two other shape spaces will be involved we shall write nk. associated with a continuous map f : P + Q.1)-connectivity would also follow from the vanishing of all the homology groups in degrees less than dz+'. q') and ( p . The latter is more closely tied in with the topological structure and we shall now examine it in more detail. C. both of which have their uses. if P is In a d-dimensional ball.-' to X . the so-called mapping cone.l ( q ) in P to the single point q in Q. of two topological spaces P and Q.denoted by C ( P ) .THE MAPPING CONE DECOMPOSITION 33 granted that the spaces are simply-connected. a simple topological recipe for building shape space Xk.3)-sphere in addition to two standard topological constructs. particular.8 THE MAPPING CONE DECOMPOSITION For many purposes the decomposition of shape space into cells is too fine. Q ) is C ( P ) Q .. 2. stratum by stratum. more generally. We have already described the much coarser stratification of Xk and also its tower of subspaces Xi. since the operation * is both associative and commutative. 1) and ( p . and it is glued on using the quotient map from s". without going back to the individual cells. 0. The additional part can be constructed from CkI'l using only a ( k .

we shall illustrate the theorem. and thus W through all possible canonical pre-shapes. shape space Eh is homeomorphic with the mapping cone where f . we have established the T following result.(id * T C ~ : ' ~ ) is the restriction of to the subset of all the ): .. r f m 3 2 and k 3 3. Thus. which it will generally be very difficult to describe explicitly. in SiI:. However. b and c are non-negative real numbers such that a2 b2 c2 = 1. We observe first that Ei is homeomorphic with Xi. by looking at the special case where it is possible to identify f t explicitly. the row matrix U is such that U' E Sk-3 and W is a canonical pre-shape in S I with Li pattern J' derived from J by removing the first row and last column.1. together with the image of the point for which b = 1 . However. Thus. . which b = 0. which we have already shown to be homeomorphic with a 2-ball B2. + + Theorem 2. the image that we get in Ck is a copy of Sk-3 * Ekyll. J' will run through all possible patterns. and also provide the start of our inductive topological proof of Casson's theorem. forms * ~ The base of the cone is the image of points for the open cone C O ( S ~ -Ck:'l). 1 and allowing the other variables to take all possible values. Thus.( embedded in SL-' and where the inclusion of the base space Xh-' in the mapping cone is the natural inclusion of Xi-' in Xi. The quotient map j-ck is injective on the entire set for which b # 0.34 THE GLOBAL STRUCTURE OF SHAPE SPACES quotient map from SL to Ek. fixing b # 0. where I: ceases to be injective. we can decompose X in block form as . 1 and so its image. any applications of it will tend to be those that only require the existence of such a continuous map. the weak link in this result is the map f k. Of course.:( g) where a. If X = can(X) is any canonical pre-shape matrix with pattern J having a full first row. as J and X vary.

(* + * 0 ').b2. 4 and 5 .. where all the letters represent real numbers with a. (: 0 * + o ') and (:+ o oi). forms the vertex of the open cone Co(B4). which is the image in X of all the points except those g for which c = 0. b.determined by : together with the four other cells determined by (* * +). When we include these last points the result is the full cone C(B4) with its base B4 collapsed onto B2 2 Xi by the map which it remains to identify. Then. Thus. This is easily seen to be homeomorphic with a 4-ball B4: at least S' * B' Z B3 can be visualised in R3 and the case we want is analogous in 0 8 ~ . Similarly to the case for the : boundary component ' b = 0' in Xi.where a semi-circle collapsed onto a single .f i) + for a general point on the pre-shape sphere that has such a pattern.THE MAPPING CONE DECOMPOSITION 35 Then. We write : (. We recall that f describes the way in which the points + + + + + fz. as a and c vary subject to a2 c2 = I . c. the decomposition of X contains the three cells of Xi. u2 v2 = 1 and w 2 x2 y2 = 1. 0 0 collapse onto X under the left action of SO(4). as in the theorem. We note that the last three of these patterns may be formed by placing the first above each of those for the three cells of X.x and y . which have dimensions 2. x and y non-negative subject to the relations a2 b2 c2 = 1. When b = 1 we get a single point that. of which the former is mapped injectively into Z and the latter is collapsed : onto a 2-ball in exactly the same way as the cells that produced Xi. b and c is fixed and neither of a or c is zero is homeomorphic with the product of the circle parameterised by the variables u and 'u and the quarter-sphere determined by the variables w .When a = 1 we just get a copy of the circle and when c = 1 we get the 2-ball. respectively. 3. for a fixed choice of b # 1 still. we see that the set of points for which each of a. we get a copy of the join S' * B 2 .

In each case. the necessary visualisation is already getting tricky and.36 THE GLOBAL STRUCTURE OF SHAPE SPACES point on the boundary of X.2. for n 3 1. Theorem 2. Hence. Shape space X : has the same homotopy type as a sphere of + ' dimension dG+' and XE:fi. they form a two-dimensional hemispherical cell as required. This can be seen directly for the points (i H H) 0 0 0 which collapse onto the vertex (+). has the same homotopy type as a ball of dimension dz+'. . For the remaining points we think of the left action of SO(4) rotating the non-zero column vectors of the above matrix until the second vector lies along the first axis and then rotating in the space orthogonal to this axis until the first vector has only its first two coordinates non-zero. We could now show that X S S5 in a similar manner to that in which i we showed X. each of the maps f and g is called a homotopy equivalence. S S2 and continue to give similar direct proofs of the higher dimensional cases of Casson's theorem. since we may rotate in the plane spanned by this and the. Obviously. Two spaces P and Q are said to have the same homotopy type if there are continuous maps f : P + Q and g : Q + P such that the composition f o g is homotopic to the identity map of Q and go f is homotopic to the identity map of P. in the next section. We can now recover Casson's theorem in two stages. homeomorphic spaces have the same homotopy type but the converse is far from true. 2. As a basic reference for the results that we shall need to quote here and in later sections see Spanier (1966). so far unused. here a hemispherical ball collapses onto a single : point. It follows from this description of the collapsing map f t of the base of C(B4) onto the 2-ball X that the quotient space X is still homeomorphic with B5 = : : C(B4). However. However. It follows that z the matrices of the above form that project to a given one in X are obtained by rotating the standard one through an angular interval of size n in two orthogonal directions. spaces of the same homotopy type do have the same homotopy groups and a necesssary and sufficient condition for the map f : P + Q to be a homotopy equivalence is that it induces isomorphisms of all homotopy groups. we shall instead employ more powerful topological techniques. fourth dimension to change the sign to the required positive value. the resulting direction along the axis may be positive or negative.9 HOMOTOPY TYPE AND CASSON'S THEOREM A weaker classification than the topological one is that of homotopy type.Then.

may be slid by a homotopy into the base X. we obtain the required 1 homotopy equivalence between a d i + ' -dimensional ball and + + XzI.+k-' of the mapping cone. which we are assuming by induction to have the homotopy type of a ball and so also that of a point.rn+l. its dimension.HOMOTOPY TYPE AND CASSON'S THEOREM 37 Proof. for the spheres C l 2 Sk-2 much more is true since Sk-2 is (k . at least for a topologist. in general. that is. Since. First.' at least is ) l+(rn+k-3)+d. it is a straightforward matter.-' -l>m+d. Xk is (dE+' . + Theorem 2. We may develop this theorem in two ways. given a homotopy equivalence between a sphere of dimension dEp1 and XZPl.-l . but not including. We work by induction from the cases m = 1 and m = 2 using either the metric results given in Chapter 1 or.+'-l connected if k 3 2. The previous theorem gives us homeomorphisms where in each case the relevant map has its image in a shape space. Proof. by induction on m or from our knowledge of when m = 2.3. Since. remain zero on inclusion in Zk for k > m 1. This follows immediately from the previous theorem for k = m and k = m + 1. those that we proved or. Of course. indicated above.-l and hence between a sphere of dimension dz+' = d"m-l m and Xz+'.3 * x ~ + ~ . for d < dz+'. . We complete the proof by induction on m and then on k.. XE or XE+'. In that case. our mapping cones have the same homotopy type as the corresponding suspensions S(SrnP2* XG-.1)-connected for all k 3 m 3 2. We also use the homeomorphisms m+l 2: p + l Ern+.) and S(S"-* * CE). the homeomorphism shows us that the image of any map of a d-sphere into Cz+k.3)connected. this result is the best possible. since the suspension of a ball is again a ball. since the relevant homotopy groups of the ball and sphere vanish. we show that the initial sequence of homotopy groups that vanish for Xz+'. . all those up to. the cone q ~ m + k . to construct one between a sphere of dimension dG-. Similarly.1 =d..Then. m . for n 3 1.1 and S"-2 * E. However. for m = 1. for a purely topological proof.

certainly all the points of this open cone have the required ball neighbourhood. by showing that the 2-skeleton is a ball. of course. when the dimension of the boundary is four. we shall show later that that part of the algebraic analogue of the cellular decomposition that carries all the homology has no 'algebraic cells' in the dimensions below d. That is. which we restate in full.+' and E :. This theorem.+' . Shape space C + is homeomorphic with a sphere of dimension .' d. The resulting shape then lies in the above open cone and its ball neighbourhood is carried back by the inverse of the isometry to a neighbourhood of the given shape. to ensure that the first row has a non-zero entry and then permute its columns. Then. it suffices now to show that shape space E. Similarly. with an analogous result for balls. may be more satisfying. However. for any pre-shape matrix with a given shape. this gives us a direct explanation of the result in the theorem. Our second extension of Theorem 2. though our earlier proof. although for odd m it then has order two rather than infinite order. we recall Casson's theorem (Theorem 1. assuming that E is known to be homeomorphic to a ball. Similarly. then it is homeomorphic to that sphere. we may assume by induction that we already know this for smaller m. preserving the shape.1) + dZp1= d.4.+'. is homeomorphic with a bull of dimension d:+' for all n 3 1.2 and the fact that d. which states that. again.1)-connected. in view of Newman's theorem. Then.+' 3 5 for m 3 3. corresponding to an isometry of shape space. since : then CO(S"-~ C z ) is a cone on a ball.+' remains non-zero on inclusion in C:+k. The top dimensional homotopy group in C. First. Theorem 2. We shall also establish there that the theorem is the best possible. To achieve this we shall use a result of Newman (1966). for the relative version of Newman's theorem. we have shown above that + . Theorem 2. Proof.-l) in C * : + ' is an open ball of dimension 1 ( m . to give a topological proof of Casson's theorem.+' is a topological manifold.38 THE GLOBAL STRUCTURE OF SHAPE SPACES by induction on k . Then. we can permute its rows. in other words. We take the cases m = 1 and m = 2 as known. we may then further contract the image to a point.+'. this base is already known to be (d.1). the open cone CO(S"-~ C. So. we find that * is a manifold with boundary: each point has a neighbourhood homeomorphic with that of a point either in the interior or on the boundary of the standard ball. to ensure that the final entry in the first row is non-zero.2 is to go from homotopy type to homeomorphism. For C:. Since the first non-zero homotopy and homology groups coincide. includes a proof that the shape spaces in this range are simply-connected. we need either to know that boundary is simplyconnected and of dimension at least five or to know already that it is a sphere. if a topological manifold of dimension at least five is homotopy equivalent to a sphere. we must show every point has a neighbourhood homeomorphic with a ball and.

+'/L~.) the space ELI'. For Xi to be a topological manifold and also to be homeomorphic with the mapping i'. Then. By that time we shall be able to give an alternative proof that the other shape spaces are not topological manifolds since their homology groups do not satisfy the 'duality' relations that the homology groups of a manifold would have to satisfy. indeed a ball of dimension d{+'.HOMOTOPY TYPE AND CASSON'S THEOREM 39 this boundary is a sphere. there are no more. so too : : is X : for n > 1 since that is also isometric with X { + ' / L ~ . would have at least to have its homology identical to that of a sphere. However. and in higher dimensions simple connectivity follows from the fact that the boundary is composed of the cone on a sphere together with a ball. One proof again follows from the mapping cone decomposition of shape spaces. is Thus. the answer is no. X . which we recall is E. with identifications on the latter induced by a map into another ball of lower dimension. W Since all shape spaces we have looked at in detail so far are topological manifolds. it is natural to enquire whether there are any other such manifolds among shape spaces. the base ball of the mapping cone. . cone Cfh(Sk-3* XkI'l. X:zf. We defer the proof of this statement until we are also able to show that the relevant shape spaces do not have the homology of a sphere.

which in our case is also the highest degree in which the homology is non-zero. especially those of an inductive nature. In Section 3. we conclude the chapter with a warning that we still have more to do to distinguish between various shape spaces. For example. 3. dimension. is far weaker than might appear from Table 1. In particular. it will be sufficient to discuss orientations in that context. of cells on the pre-shape sphere. BARDEN & T. LE Copyright 0 1999 by John Wiley & Sons Ltd CHAPTER 3 Computing the Homology of Cell Complexes In this chapter we introduce homology groups and make some initial deductions about them for shape spaces. C A R " & H. as we shall see. the Eilenberg-Steenrod axioms that the homology groups satisfy and that suffice to determine them.6 we give sufficient topological definitions to be able to state. under the quotient map. KENDALL & D. Such exact sequences are a standard tool of topologists and do indeed enable us to draw some initial inferences about the homology groups of shape spaces without actually calculating them. However. we emphasise the use of reduced homology to simplify theorems and proofs. We first show how these groups are defined and calculated for a finite cell complex. K. G.Shape and Shape Theory D.1 THE ORIENTATION OF CERTAIN SPACES For most of our homological calculations it will be necessary for our cells to be oriented. These axioms enable us to derive. and for the cells on their boundary to be oriented ‘coherently’. the obvious topological invariant. Since the cells in the decomposition of shape space that we described in Chapter 2 are all homeomorphic images. For. without proof. a .2: there are arbitrarily large sets of shape spaces for which all members of a set have the same dimension. restricting the technicalities of orientations and the computation of boundaries to the special case of spherical cells. from the mapping cone decomposition of the previous chapter. a ‘long exact sequence’ involving the homology groups of three closely related shape spaces. which is all that we shall require in the context of shape spaces.

our cells are obtained by restricting certain of the coordinates to be zero. The natural home of an orientation is in Euclidean space. This last is the approach that generalises most readily. Note that this quality cannot vary on a connected set. In fact.2 THE ORIENTATION OF SPHERICAL CELLS As described above. We shall give an account of the orientation of such cells. we may define a diffeomorphism. x4. it may be viewed from a number of different points of view. at least for our purposes. We recall that for Sd in Rd+’. XI. We need to define a standard orientation for such cells with which to compare those orientations that occur in the course of our calculations. the cells in which we are interested lie on the sphere. fore finger and middle finger. 278). the composition f w r o ( f w ) . based on that in Hilton and Wylie (1960. an orientation of a space P diffeomorphic with a connected open subset U of R” is a class of diffeomorphisms { f w : w E Q) of P onto U such that. the usual choice of a ‘positive’ orientation being that which aligns the positive x-. for any two members f o .and z-axes in the directions of the thumb. and Orientation reversing otherwise. usually indicated by an arrow. where. the even permutations of the natural ordering. within the intersection of certain of the coordinate hyperplanes and delineated by others. In R3 an orientation is most commonly interpreted as a choice of ‘handed-ness’ for the axes. possibly one to be negative and making no restriction on the remaining coordinates.42 COMPUTING THE HOMOLOGY OF CELL COMPLEXES homeomorphism of an oriented cell onto another cell determines an orientation of the latter. often indicated by a circular arc with an arrow at one end. in R4. respectively. We recall that the particular cells on the sphere that we use are unions of contiguous sets of the cells into which the sphere is divided by the coordinate hyperplanes. determined by any particular diffeomorphism in the class. y. a . Now. writing ei for the unit vector along the ith coordinate axis. respectively. of orderings of the axes. x3 is a member of the positive orientation. Finally. In effect. Thus. we may declare a linear isomorphism to be ‘orientation preserving’ if its determinant is positive and ‘orientation reversing’ otherwise. 3. In R’ an orientation is simply a choice of direction. a differentiable homeomorphism with a differentiable inverse. p. In Rn we may define an orientation to be a permutation class. of the right hand. f u r of the class. depending on whether its determinant is + I or -1. the sequence x2. particularly in low dimensions. Such an orientation is. even or odd. In R2 it is a direction of rotation of the plane. of course. that is.’ is an orientationpreserving diffeomorphism of U onto itself. others to be positive. between connected open subsets of Rn to be orientution preserving if its Jacobian matrix is orientation preserving. To orient more general spaces we first note that an orthogonal transformation either preserves the orientation classes or interchanges them. this just extends each orientation class of axes to include those obtainable from any particular member by a linear isomorphism of positive determinant. Thus.

) . that is. We shall next show how to use an ordering of the vertices of a spherical cell to determine an orientation of the cell itself. Thus. 3. the orientation of the n-cell determined by the composite of these two diffeomorphisms is well-defined.. but a linear map inducing an even permutation of the vertices of the standard simplex is orientation preserving. = f l and the i j are all distinct. This means that homeomorphisms determining the orientation of the basic cells can be chosen to fit together to provide a homeomorphism of the entire elementary cell with an open subset of Euclidean space. However. We note that the standard orientations of the basic cells of. &in+. A diffeomorphism of a basic n-cell on the sphere with this standard n-simplex. In describing these cells.THE BOUNDARY OF AN ORIENTED CELL 43 basic cell. and hence an orientation of that n-cell.). First. where each 8. We also note. We may also take its vertices in the order of the origin followed by the unit vectors in their natural order. which is also given by radial projection.ein+. . This is a diffeomorphism on the open cell and extends to a homeomorphism on the closed cell. is uniquely determined by a set of vectors (&ilei1. say. equal to -1. .1 there is an obvious way in which the orientation of a 2-cell induces an orientation of each of the 1-cells on its boundary. choosing a linear map that preserves the chosen orderings of their two sets of vertices. following the order of the axes along which they lie. so it makes sense to define the standard orientation of an elementary cell as that which restricts to the standard Orientation of each of its component basic cells. . This will then give us a standard orientation for our particular spherical cells. which is determined as the convex linear hull of the origin and the positive unit coordinate vectors and has a natural orientation as an open subset of R". It is not difficult to see that this is true in general. the order in which we wrote the vertices in the symbol did not matter.3 THE BOUNDARY OF AN ORIENTED CELL In Figure 3. is then determined as follows. In R" we have the standard linear cell or n-simplex. irrespective of whether that is in the positive or negative direction. that the standard orientations of the two basic cells that comprise the elementary . . equal to . ( w 2 .1 in the usual manner for one-dimensional and two-dimensional cells. In this way. one whose boundary contains at most one member from each pair ( + e . +e3) are compatible with an orientation of the full elementary cell. We then map this simplex linearly onto the standard n-simplex. the basic 1-cells and 2-cells in the upper hemisphere receive the orientations that are indicated in Figure 3. and we may think of a standardly oriented elementary cell as the union of its standardly oriented basic cells. when there is an even number of the ~ i . we project the cell radially onto the simplex determined by the vertices that define it. . for example. then we take an odd permutation of this natural ordering. If there is an odd number of the &i. now we need to insist that. we write the vertices in the symbol in their natural order. The ordering of the vertices of the cell is only determined up to an even permutation.1.e . or an even permutation of this order.

as well as to proceed to the computation of the boundary formula in the chain complex for shape spaces. we precede the entire symbol by a minus sign so that. respectively. with each transposition that is necessary reversing the . + This is equivalent to requiring that the vertex to be deleted first be moved to the front of the sequence. -e2) would now be denoted by (+el. just mentioned. . -(-el. it will be convenient to introduce an alternative notation for our oriented cells. . +e3. Instead of changing the order of the vertices in a symbol for a basic cell..n+l ). the following rule. then a boundary n-cell is determined by the same sequence with just one vertex. . If the (n 1)-cell is determined by the sequence of vertices ( E .. el"+. to extend to cells of arbitrary dimension the definition of an induced orientation of the boundary cells. (-el -e2 +e3) and -(+el. This reflects the fact. +e3). (--el. Then. when there is an odd number of the &i equal to . in this representation. -e2. in order to describe them. +e3). +e2. say the j t h vertex e. +e3). e. -2. In other words. 3 . we apply.1. that the elementary cell has an orientation compatible with the standard orientations of its basic cells. Thus.. (--el.. . the four 2-cells that we would previously have denoted by (+el. a change of orientation is indicated by formal multiplicaiton by . +e3). +e2).1 Standard orientations of the cells of the upper hemisphere in R3 cell ( + e l .1.. +e2. +e3) induce opposite orientations on the common boundary cell (+el. e. +e2. the orientation chosen as the induced orientation of that boundary cell is (-l)'-' times the standard one determined by the remaining sequence of vertices. +e3) and ( + e l . These phenomena are completely general and. *e2. the total number of minus signs appearing is always even. omitted. +e3.e. Then. first for basic cells. .44 COMPUTING THE HOMOLOGY OF CELL COMPLEXES Figure 3. we always write them in the natural order but. +e3).

( S d . -e2. If we write a for the operation of taking the union of the oriented n-cells in the boundary of an oriented ( n 1)-cell. On the other hand. +e3) = (+el. Once again. consider the free abelian group C. each of the cells in the boundary of the standardly oriented basic cell again has the standard orientation. which may easily be checked for the examples given above. more precisely. then we have. writing the same symbol for the generator of C . +e3) that appear with opposite signs. in the notation introduced above: + Here. for example. because the total number of minus signs involved in each term on the right-hand side is even. Such a sequence of groups and homomorphisms satisfying a2 = 0 is called a chain complex. that is. a cell complex in which every cell has a chosen orientation. +es). HOMOLOGY AND COHOMOLOGY GROUPS 45 orientation. +e3) their standard orientation.THE CHAIN COMPLEX. before it is removed. the union of their oriented boundaries is just the set of cells in the boundary of the elementary cell. Z) with one generator for each oriented n-cell. where the image is now regarded as an element of C1 ( S d . we may rewrite the above equation for the boundary of a 2-cell as a(+el. each with its standard orientation.Z). however. Note. HOMOLOGY AND . for example. of the boundaries of the oriented basic cells that comprise it. +e2. the nth chain group of P with integer coeflcients. 3 4 THE CHAIN COMPLEX. we may. +e3). +e3) + (+e2. COHOMOLOGY GROUPS Given an oriented cell complex P. with appropriate cancellations. this phenomenon is quite general and enables us to define the oriented boundary of an oriented elementary cell as the union. elements in the kernel of a are referred to as cycles and . In general. +ez) - (+el.Z) as for the cell to which it corresponds. having only one minus sign in its symbol. However. if we consider then the second of these boundary cells. provided we agree to extend our formalism to allow the cancellation of the two occurrences of (+el. for each natural number n up to the dimension of the complex. Then. +e2. +e3) and -(+el. has the opposite of the standard orientation chosen for it. so their union is the standardly oriented elementary cell (+el. the boundary map extends to a homomorphism which has the property that a2 = 0.(P. This is called the nth chain group of P or. that we have given each of the contiguous cells ( + e l . *e2.

of P is defined to be the quotient of cycles modulo boundaries. Q). the homology group ker(a. and a is a sequence of homomorphisms a. : D. To obtain homology groups with coefficients in the abelian group Q we form the chain group C. there are 'Universal Coefficient Theorems' (see the Appendix) that spell out how to compute the homology or cohomology of a space. 8 = 0 so that we have a chain complex (C. or just by H. which is a direct sum of copies of G with one summand for each oriented n-cell. of P with coefficients in Q. = 0 and.. a) where D. when computing the homology of shape space. from a specific topological context.. but the converse is false and the nth homology group. we can still define the nth cohomology group.46 COMPUTING THE HOMOLOGY OF CELL COMPLEXES elements in the image as boundaries.Q). we shall look for a subcomplex D. As before. just as before with -(gei) being interpreted as (-g)ei. For example. However. (P. with integer coeficients.(P. of its chain complex whose homology is the same as that of the space.(D. + D. does a) not arise. where we denote by gei the element of the chain group that is the element g of Q in the copy that corresponds to the cell ei.(P.Q ) to Q and the coboundary homomorphism SQ is G . When (D. ) homology groups to be the homology groups. in this book we shall mainly be concerned with the integers Z.-1 such that a.(P. Z)for the direct sum of all the indicated chain groups and H . We then define the boundary operator a.. Q). = 0 implies 6.the group Z z of two elements or the rationals Q as coefficients.). ) induced by a ~ Then. although 66 raises rather than lowers degree. starting from its homology with integer coefficients and we could proceed in that way. The nth cochain group C"(P. or of its elements. a). It follows that boundaries are always cycles. For a cycle c the homology class.(D.G ) to Cn+'(P.Q ) is the group of homomorphisms from C. To define cohomology groups we work with the cochain complex derived from the chain complex as follows. with coeflcients in G . the equivalence class modulo boundaries that it determines is usually denoted [c]. 8 ~ and we define its . 8.)/image(a. that from C"(P. The suffix n here is called the degree of the homology group. Z ) for the direct sum of the corresponding homology groups. (P. as above. we may define the homology groups of any chain complex (D.+l) would be denoted by H.It is also standard notation to write C.+l = 0 for all n . where -g is the inverse of g in the additively notated group Q. of P to be In fact.d. is the direct sum of abelian groups D. H. Similarly. and the chain complex from . ( P .

P O)are the corresponding homology groups. observe that the 1-skeleton P”) of P. which we denote by &.G ) we may find a chain c1 = E~C.Po). we shall obtain the reduced homology by simply omitting the corresponding generator from the chain complex. G ) . including. we have acl = co . We check that this does form a chain complex and then the H . First.G ) and Co(P. when the cofficients are unspecified in any reference to homology or cohomology.(P.(P. When P has r components. When we need to make the distinction. (P) in which the coefficient of every generator corresponding to a cell in Po is zero and.THE HOMOLOGY EXACT SEQUENCE FOR SHAPE SPACES 47 which we compute the integral homology is so simple that it is just as easy to compute those other homology and cohomology groups directly.Q ) . one G for each component. (P . If (TI is an oriented 1-cell with bounding zerocells a and 0 and with a oriented from a to a then. is just ordinary homology with one summand Q removed from Ho(P. the case of the In cellular homology. (P) as absolute homology groups.G ) such that ac. G ) and we shall be working with reduced homology since that tends to simplify many statements and proofs. PO). The linear combination of generators that . (P) by taking PO= 0. all zero-cells determine the same homology class [co]. of course. union of its zero-cells the and 1-cells. then. which therefore generates Ho(P. the so-called cellular homology. Note that we recover H . we refer to the groups H .(P. 3. if PO is a subspace of P we need to define the relative homology group H. all such generators are similarly ignored. POwill need to be a cellular subcomplex and then H. on taking the boundary map. ) . must be connected.6 THE HOMOLOGY EXACT SEQUENCE FOR SHAPE SPACES Most of our discussion and calculations will concern integer coefficients and. Reduced homology.cb.5 REDUCED HOMOLOGY For a connected complex P we always have Ho(P. or the corresponding chain or cochain groups. First. G ) will be the direct sum of r copies of G . for the corresponding 1 0 & generators of C l ( P . that which we have just defined.(P. It follows that for any two generators co and cb of Co(P.cb. similarly.Po) is the subgroup of C. they are to be understood to be the integers. = co .G ) 2 Q.(P. Even before we settle down to compute the homology and cohomology groups of shape spaces. PO) is defined using the chain complex C. Hence. For that we shall need to recall the Eilenberg-Steenrod axioms that are satisfied by any homology theory. Since the cell complex that we shall use for shape spaces has a unique zero-cell. H o ( P . in C1 (P. where each C. which we see as ‘ follows. 06 xi 3. we can obtain some information from standard results in homology theory.

we obtain the following exact sequence for shape spaces. Qo) if. (v) Dimension. more generally.1. which is. Qo) of pairs induces a homomorphism f * : H. = id (fog)* = f*%*. (ii) Homotopy. (i) Functoriality. Po)-+ 0.48 COMPUTING THE HOMOLOGY OF CELL COMPLEXES we ignored on taking the boundary map in the chain complex forms an ( n . also has all its reduced homology groups zero.(P) are replaced by the corresponding reduced homology groups.then H . a cycle and may be shown to induce a homomorphism. into the pair (Q. PO)-+ (Q. that is. any space that is homotopically equivalent to a point. p ) for any point p in P or. If PIis an open subset whose closure is a subcomplex of P contained in the interior of the subcomplex PO. . The reduced homology of a point is zero in every degree. then f is called a relative mapping o the pair (P. If POis a cellular subcomplex of the cellular complex P. (P. ( Q . This.(P.-l(Po). We may now state the axioms. The axioms also imply that H . If PO# 0 it is also valid when the groups H. it is a relative mapping: that is. at every stage. from Hn(P. Qo) such that (id). then f* = g*. o ) and a homotopy from f to g is called f PO) Q f PO) a relative hornotopy o the pair (P.(Po) and H. it also maps POinto Qo. PO )-+ H . then there is an exact sequence of groups and group homomorphisms To say that the sequence is exact means that the image. it ends with Ho(P. in fact.PO) into the pair (Q. If POis a subspace of P and QO is a subspace of Q and if f and g are continuous maps from P to Q each mapping POinto Qo. From the exact sequence (iii) and the mapping cone decomposition given in Theorem 2.-l(Po). implies that any contractible space. of each homomorphism is the kernel of the next. Each map f : (’P. completely determined by these axioms. in fact. (P) 2 H . (iii) Exactness. also denoted by a. Po\%). If the maps f and g of the pair (P. for p any contractible subcomplex of P. 2 PO) H.1)chain in C. The homology theory is. (’P. (iv) Excision. PO)to H. Q o ) are relatively homotopic. In either case. An immediate consequence of axioms (i) and (ii) is that homotopically equivalent pairs or spaces have isomorphic homology groups.(’P\Pl. together with axiom (v). into the pair ( Q . necessarily a subgroup.

1.I]$ C[a. . tfor the part of the mapping cone that is the image. 0 ) all y1 since the cone C [ u .C [ u .b] onto extends to a ~ ~) relative homotopy equivalence between the pairs ( C I 11. ~ ] . shows that is ~ This last group may be re-interpreted using the fact that for any space P the suspension S ( P ) = So * P is. On C[o. OThis.OI Ck-'. in ]which f i . along the [0.I].1. so that ffn(C[a. Then. or of ((Sk-3 * CiZ'l) x [ s .b]) ffn(C[a..t ] . = we have in which the absolute groups are already as required by the theorem. and so has the same 11 = homology groups as. ( C [ a=~ ~for ~ ~ ).JI the suspension of Z. taking the exact sequence (iii) with P = C[0. Xk.L ~ . writing S[O. We write C [ s . . the exact sequence for the pair ( C [ u . under the quotient mapping. ~ ] Ck-' are I. the mapping cone CIO. of (SkP3* x [ s.THE HOMOLOGY EXACT SEQUENCE FOR SHAPE SPACES 49 Theorem 3.b]there is a homotopy. . ' : i X x [ s . Then. T . b and ( C [ u . from the identity to the obvious projection. ~ I . by Theorem 2.onto ~I C [ O ... To re-interpret the relative group in the sequence we choose a with 0 < a < b and use the excision axiom to conclude that We next observe that the deformation retraction of CcU.1] factor.t ] ) U Xk-' if s = 0. of . and Xi-' indirectly using the mapping cone decom~ position of Xk. implies that C [ O . c Then. and pick b with 0 < b < 1. in fact. Then. u ] ) .l for that part for of the suspension that is the image. lC .Zl1 and S[. ~contractible.[ u .11and PO C [ O . t ] if s > 0. We represent Ck. called a deformation retraction. under the quotient mapping. the mapping cone of the mapping of P to a single point. and homotopically equivalent and so have the same homology and reduced homology groups. For m > 2 and k > 2 we have the long exact sequence Proof. is homeomorphic with. fixing C[O. of C [ O .

1 of their homology groups that we do not yet know are isomorphic.+') for all n such that n .l]> 2: ~ n ( ~ [ o .m . ~ S. H Stating the theorem in terms of reduced homology groups has the advantage that. it is always a homomorphic image of it and the sequence + + < + . we have fi. and from the homology exact sequence together with the fact that S L ~ isJcontractible. this means that the first k .-' .+' . for all n dz+' k .1 = d. 0 1 2: Hn(S[o.2.m . I Hn_1(Sk-3 * x.:ll) so that from our original sequence we obtain the one stated in the theorem. b j ) 2 H n ( S [ a . as in the above proof. However. are known to be (LIZ+' . f i n ( ~ ( ~ k ~ ' 1 ) ) fin(S[o. + 2 ) is isomorphic with the group Z = H n ( X $ + ' ) .4 all three spaces are topological balls. 1~1[ 0 .m . Since C and . When k > m. that is. l l . and the dimension dz+' = (m . the five non-identity isomorphisms follow respectively from the fact that SI0. from the homotopy axiom again. is d-connected. similar to those that we have just used. for n = dz+'.) 2 H. Similarly.7 APPLICATIONS OF THE EXACT SEQUENCE When k < m. b ] ) . 3. Since where d = dEPl .ol is a single point.1)-connected. ) g H n ( S [ u . H n ( X .50 COMPUTING THE HOMOLOGY OF CELL COMPLEXES we have the following sequence of isomorphisms.(X. the excision axiom. where again we take 0 < a < b < 1. from the homotopy axiom together with the fact that there onto from is a deformation retraction of S[O.1. the reduced and usual homology groups are identical anyway. with the . Cz+' of :l of dimension d. Iterating this result we see that . we do get new information.k 1 < d . since by Theorem 2. i' isomorphism induced by the natural inclusion of X in X + . the entire sequence is a trivially exact sequence of zeros. we do not have to keep treating the case n = 0 as special. however.[ a .OI. Note that this just fails to tell us whether.(X. for n > 0. ~ [ o .a]) f f i d (XkLIJ . S [ u . Here. when k = m.l)(m 2 ) / 2 and Zonly non-trivial part of the long exact sequence is the suspension isomorphism between their respective unique non-zero reduced homology groups. a ] ) l g fin-l(S[a.~I S[O. the space Xz is a ball and the other two spaces are topological spheres. It has no real disadvantage since.ll.:lI) 2: Hn-k+l(X.m.

' is induced by mapping the cell in Xi determined by a pattern J to the cell in Xi+' determined by the same pattern. are isomorphic with the kth suspension of those of X i .+' . In fact. The homomorphism 2. which is derived from J by removing its first row and last column.+! (C2+2) is finite. falls little short of complete knowledge of the boundary homomorphism and. then. . the homomorphism J* also factors through the relative group . The information needed for this approach. the corresponding group is Z.ranging up to the dimension of the space. However. We shall then be able to see how to choose a new basis in the chain complex. we state without detailed proof. ( At the other end of the sequence we find that. however.d i p 1 = k . it is possible to gain sufficient knowledge of the boundary i' homomorphism in the chain complex to identify the cycles that generate the homology groups.(Xi) -+ fin(C:+') is induced by the natural inclusion z of Xi in C + and so. The full integral homology can then be read off immediately and the homology groups with other coefficients can be deduced using the Universal Coefficient Theorem or again read off directly. . In particular.(Ck+') -+ fin-k(X. adapted to that homology. Since d.. at cellular level. we shall indeed compute it explicitly.+~ (XE+k) Z Z2 for k 3 2 and so for m even HdE+1 X z + k ) Z Z for k 1. the exact sequence will still be of value when discussing and interpreting our results. this is how most of the homology computations were first carried out. iterating the isomorphism between the groups in the top degree we find that In the intermediate degrees it is less obvious how the homology groups of combine with the suspended homology groups of X i p l to form those of X + .APPLICATIONS OF THE EXACT SEQUENCE 51 tells us that. Indeed. this means that the 'top' m potentially non-zero homology groups of Xi+'. and hence certainly no homology in those degrees.k ( C i .+' .d i = m and d. since the re-based chain complex is particularly simple. in the next chapter. This is achieved at cellular level by mapping the cell corresponding to a pattern J to the cell corresponding to the pattern J'. which is the morphism 1% H. we shall find that for m odd fi. To do so we shall require the following identification of the cellular maps that induce the homomorphisms in the exact sequence and which. there is an isomorphism between fin(C:+') and f i n . for the next m.. since they are inherently plausible and no longer required for the calculation.l ) for n > d i . and also sufficient insight into the homomorphisms to be able to use the exact sequence as a basis for an inductive computation of all the homology groups starting from the known special cases. since Xi has no cells in dimensions greater than d. However.-l) inverse of the suspension isomorphism. The main component of the homo: is the 'de-suspension'. : H. so that all the generators and relators appear in the basis. Nevertheless. if for any m the group Ed.

any two different shape spaces have very different sets of homology groups and so cannot be homeomorphic. that the first non-zero homology group of Ek is in degree d:+' for all k > m. 3 Even if we allow a space on the 'main diagonal'. is isomorphic with 72. which. although the dimension d:+' (k . k 3 m + 2 corresponds to + 12 d3 = d . the first set of three spaces of the same dimension lying in the triangle m 3 3. EL) and so is zero on all cells. but not yet proved. or generators. Even if we make restrictions such as m > 2 or k > m + 2. To complete the characterisation we would need.2. We then take the boundary components of this cell that lie in Ck.8 TOPOLOGICAL INVARIANTS THAT DISTINGUISH BETWEEN SHAPE SPACES In the next two chapters. Thus. . 3.52 COMPUTING THE HOMOLOGY OF CELL COMPLEXES ffa(Ek+'. when we calculate the full homology groups of shape spaces.k ( C i P l )+ f i n . 10 -41. which correspond to patterns without a full first row of k . for which all homotopy invariants must vanish. The first set of four such spaces is 14 12 d 23 = d 6 = d I13 = d 9 = 6 2 . Ck with k 6 m. we shall see that.1 '*'s and a '+'.10 = dz = 29. The fact that dimension alone will not suffice is clear from Table 1. However. This would determine m and then k could be deduced from the above expression for the dimension. there is then the earlier set of four: dl' = d l * = d" = d10 = 44. as we have seen. the repetition of dimensions is about as bad as it could be and we may show the following. and the next is 16 13 d 3 = d 4 = d . provided we exclude the contractible spaces.l (Ck) is essentially suspension followed by the boundary homomorphism of the homology exact sequence. 3 5 6 9 In fact. This also ensures that the decrease in degree is indeed k as stated in the theorem.l ) m of Ek is determined by the degree of its top non-zero homology group. some dimensions occur more than once. It is induced by first mapping the cell corresponding to a pattern J' to that corresponding to the pattern J formed from J' by adding a column of zeros at the right and then adding a full first row. : f i n . Finally.m . The limited topological information we already have is not quite sufficient to characterise the shape spaces among themselves. there is no earlier set of three. the fact referred to above. at the top. the homomorphism d. for example. of the new length.

then d i = d. then this would require 2" = 0 mod m. the dimensions of Xk. Similarly. there are other sequences of dimensions that are not realisable under appropriate restrictions on m and k . which again leads to a contradiction. similar or distinct. which is congruent to + + { -1 modm ( m / 2 .1. On the other hand. we may easily check that.for various k .1. if m is an odd multiple of n . Thus.m(m 1112.1) mod m if m is odd. for some integer 1.There are also arbitrarily large integers that are not the dimension o any shape f space Xk with k > m > 1. It is clear from the examples above that there are other ways of producing sets of spaces of the same dimension as that used in the proof. then it would require 2" = m / 2 mod m. if m is even.1. while if m were even. odd integers to obtain m'. multiplying m by a sequence of r . are congruent modulo m to dz+' = m(m 1)/2 . every one of the spaces X k J for k > m' will lie in a set of r spaces with the same dimension as one of the Xk's. Suppose then that Xk had dimension 2" . .1 . Proof. = mk . Since Xk has dimension d. If m were odd.

it does allow us to compute any particular group and also to prove general results. C A R " & H. We are also able to identify the non-zero reduced homology group of lowest degree and in the final section we calculate the homology of the space of 'shapes of infinite sequences of points' in R". with respect to the natural inclusions Eft. In Section 4. leaving this subset fixed. G. This enables us in Section 4. such as the connectivity of the spaces and the fact that they are not topological manifolds. We shall find it convenient throughout this chapter to work with the spaces C k / i m of unoriented shapes at the same time as the spaces X i . K. KENDALL & D. LE Copyright 0 1999 by John Wiley & Sons Ltd CHAPTER 4 A Chain Complex for Shape Spaces We are now in a position to make our initial computations of the homology of shape spaces.Shape and Shape Theory D. and hence the homology and cohomology of shape spaces using the cellular decomposition obtained in . Eft. formed by taking the direct limit of the spaces Eft.4 to identify a maximal sub-chain-complex that is generated by a subset of the basis elements and makes no contribution to the homology. Although this does not yet give us a closed form for the homology groups. We start by labelling the generators or basis elements in the chain complexes determined by our cell decomposition of shape spaces and then give an explicit formula for the boundary map in terms of these generators. 4. L.1 THE CHAIN COMPLEX To compute the boundary map in the chain complex. This complement we call the essential chain complex since it determines the homology of the shape space and no smaller sub-chain-complex could do so. the remaining basis elements generate a direct complement to the first subcomplex. BARDEN & T.+'. After a simple change of basis.5 we make a slight modification to the basis of the essential complex as a result of which each new basis element carries either a generator or a necessary relation for the homology of shape space. The changes of basis in this and the preceding section are sufficiently modest that these generators and relations can be associated with particular patterns for the cells in the original decomposition of shape spaces.

We shall need to distinguish the two cases (i) k > j l > j 2 > . or (ii) k > j l > j 2 > . and 7 the corresponding generator ~ ( 1 . . or Z-module. = 0 = j n + l = . 0 . It will be convenient also to refer to this sequence as a pattern. and by .. .(Ci).(Ck) is the free abelian group. as explained above. Thus. . . corresponding to J has dimension IJI = ji) . when the final entry in the last row is restricted to be negative. for some n < m. corresponding to each of the above oriented cells of dimension n in Xi. which we denote by O J .).(Xk) has two generators CJ and CJ for each J E .. referring to the pattern matrix or pattern sequence when we need to clarify the distinction. Denote by 3 = 3 1 1 the set of pattern sequences of type (i) and similarly define 3 2 and write =31 U3 2 For J E 3 1 there are two elementary cells on the pre-shape sphere SL. . . The cell OJ or.0. These form a sequence of integers J = ( j l . C.We shall give each of these generators the same name as the cell to which it corresponds. since we are only concerned with reduced homology we shall throughout adopt the following convention. where. We denote the images of these cells in with their induced orientations by cJ and CJ. > j . which are positive or negative unit vectors along various coordinate axes. The standard orientation is then that determined by the natural order of the axes or its opposite according to whether an even or odd number of negative unit vectors is involved. row by row. . = j . . ) . .1. or basis element. > 0. with the coordinates on the pre-shape sphere taken in the order in which they appear. respectively.7.71(i) and one generator CJ for each J E 3 2 ( . . . which are strictly decreasing until they reach zero. . recall that we require the chain complex C. where C. For J E 3 2 there is a unique elementary cell in S. . Recall that the orientation of an elementary spherical cell is compatible with that of each component basic cell and that that is determined by its sequence of vertices. in each case. which we denote by OJ when the final entry in the last row is restricted to be positive. (cy!n=l We exclude the shape with pattern sequence (1. We note that a pre-shape pattern is determined by the lengths of the rows of non-zero symbols. . For homology computations with integer coefficients. we shall introduce some terminology that we shall use throughout this and the next chapters. . j. . Note however that. 0) from all the sets .7. > j .0) from Co(Ck). when relevant.56 A CHAIN COMPLEX FOR SHAPE SPACES the previous chapters. These cells are oriented as elementary spherical cells in the manner described in Chapter 3. so that we have Co(Ck)= 0. IJI = n . J(b) (i) (i) (i). . with one generator. in the representative matrices.

1 when i < m and either j i = 1 or j > ji+l i otherwise. since each of the final j .6. {@ Note that aiJ is non-empty if and only if the sequence formed by lowering the ith entry in the pattern sequence J by one. negative unit vectors. define when i = m and J E 3 .l . . .(Xk) the involution 1.) < m. . which maps CJ to (-l)jmcJ for J E 3 1 and fixes CJ for J E Jz. j i + l . We also see that our cell decomposition fits in nicely 1 with the stratification of shape spaces described in Section 2. since UJEJ. if instead we quotient out by O ( m ) . . . for J E 31. It is also clear that each cell CTJ for J in 3 2 is fixed by 1. It follows that the space of unoriented shapes C k / i m has a cell decomposition with no longer any distincwith just one cell for each pattern sequence J in tion between 3 and Jz. Then. as in Chapter 2. which we denote by (i.. which is a finite formed by the union of all the cell decomposition for the subspace XL-l/i. we shall henceforth use H instead of fi to denote them. . induces the isomorphism. .1) ( j 1 . j m . J(i) uJEJ2 43 THE BOUNDARY MAP IN THE CHAIN COMPLEX . . CJ forms the topmost stratum of Xk and CJ is a subcomplex. we use the .1)Im times the standard orientation. 1. we get the space of unoriented shapes. since all our spaces are connected and so have reduced homology zero in degree zero and since otherwise the reduced and standard homology groups are the same. j i . .-l lower strata.1. For each J E Jand each i such that 1 < i ( j 1 . aiJ = + 1. j i . . . and leaving all the other elements of J unchanged. . -1) to induce i then it is clear that. Thus.)*. coordinate axes is reversed. and that. we have the following result on which all our computations of homology and cohomology groups will be based. in C. Since the order of the axes is preserved the image has (. is still a member of J.THE BOUNDARY MAP IN THE CHAIN COMPLEX 57 In addition. j. . j . the basic cell of GJ that is determined by positive unit vectors along each relevant coordinate axis is mapped isometrically onto a basic spherical cell of J with j . 4. It will be convenient to let cg = 0.1 . element diag{1. obtained as the quotient of the pre-shape sphere by the left action of SO(m). . which is the quotient of X i by the isometric involution 1. . If.2 THE SPACE OF UNORIENTED SHAPES Recall that the shape space EL is the space of oriented shapes. .. . .

1) is equal to 1 if j . j . . for J E 3 . > 1 and. coordinate axis corresponding to an '*' or together with the positive unit vector along the axis corresponding to the final f entry when dealing with OJ and the negative unit vector for aJ.. we must use the action of SO(m) to identify each elementary cell in the boundary with one that corresponds to a pattern. the boundary formulae for the chain complex given in the theorem are correct up to the various signs and it remains to check that their orientations are as stated. Since the elementary cells in SL corresponding to the chosen patterns : are mapped homeomorphically and disjointly into Ez and since the latter is the result of quotienting out the action of SO(m) on the former.(Ek) may be computed in C.J since we took the unit vectors in + '+' + + '+' .58 A CHAIN COMPLEX FOR SHAPE SPACES Proof. Moreover. it follows that the boundary map for C. That is.(SL) modulo the action of SO(m).+l + 1 the elementary cell C T ~ Jappears on the boundary of OJ and that aa.J = 0 and so ca. So. .J in fact that the unit vector corresponding to the final in the ith row has to be moved past j l . = 0. When i = m this is just aa. We note first that for i = m and for each i < m with either j .we take that determined by positive unit vectors along each 1 in the pattern determined by J . . In the latter case we must take an odd permutation of the natural order of the vectors since just is then determined by the one of them is negative. We may determine the orientations of the elementary cells appearing in the boundary of an elementary cell by considering those for representative basic cells. these are the only cells of next lower dimension appearing on the boundary. Hence.J = 0. although it is also equal to 1 if j . Note that the factor (1 . since UJ is open. > j.+l 1. > 1 we also have to consider the boundary cell determined by the negative unit vector along the axis corresponding to the final '*' entry in the ith row of the pattern.1 previous unit vectors before being deleted to obtain a similar correctly oriented basic cell of the corresponding elementary boundary cell. > j. The coefficient of aa. so that its boundary is closure(aJ)\aJ. when taking the boundary of an elementary cell corresponding to a pattern.J also occurs there if i = m and j m > 1 or if i < m and j .6. we then have a. When j . = 1 or with j .

( C ~ / L . As a subcomplex of Ck. However.J = (-l)j.(Ck. as do those for which i < rn and j + j . as we may. (iii) in C . .1. > 1. ) for J such that j . ) for the corresponding reduced homology groups. ( X ~ / L . after multiplying on the left.ca.That for a q is similar with the differences accounted for by the facts that for J E 3 and i < m we must 1 have ji > 1 and that a. ca.2 j .(Ci).4 DECOMPOSING THE CHAIN COMPLEX It will be convenient now to write C. 3 1.( . > 1. our homology computations will be based on the following special cases of the theorem. the corresponding chain complex C*(". The subcomplex C. + Corollary 4. of C ~ / Lis. of them are negative. in the chain complex. Ci-' comprises just those cells CJ corresponding to J for which j l < k . no j i can equal one. This completes the computation for aaJ and hence for acJ. in each case. Z ) .. we do obtain a basic cell that belongs to but with (. for J such that j . The boundary maps have the following properties: (i) in C. characterised similarly. 4.J On the other hand. Hence. if j .') is . so we shall denote it by c.l)j~+jm+'times its standard orientation since the unit vectors that span it are in an odd permutation of their natural order and j i . the second i statement follows immediately from the theorem and the facts that.~ for i < m and ca.(L) ambiguously for the chain modules C. removes the vertex that is a negative unit vector.(:) for the corresponding : ( chain complexes and H .1.J = .Z) or C .DECOMPOSING THE CHAIN COMPLEX 59 an odd permutation of their natural order as required for aa. is even. cancel. + It is clear from this theorem that the sum of basis elements CJ CJ will play a special role. Then. when i < m it is not the chosen form for a basic cell in any of the elementary cells corresponding to a pattern. by the diagonal matrix all of whose diagonal entries are + I except for the ith and the mth which are -1. We shall also write C. respectively. The first statement holds since the last summands in acJ and aG.1 ) j m q m ~ the third stateand ment is an immediate consequence of the second since. .-'/L. Proof.

U 3.).l . for each n and both cases. .. . (E. C.(D. that is.(:) and do so as efficiently as possible. We shall thereafter work directly with the essential complexes. m. saying. K . ) 2: D. the boundary a maps C. . . IJI = n ) . for example. . . we subdivide the sets 3 and 3 2 of 1 patterns so that 3 = 3. we propose first to find decompositions. . . n . j . Let K . j m . s induces a bijection between 3 and 3. ( k . be the submodule generated by { c ~ l j= k . just zero. . such : ( that D. so that E .l . in each degree. 1.1. j n + l = 31 o for some n < m). define sJ = ( j l . . as above. . For J E 3:. . = @.J # @ and. Thus. for the latter. in this decomposition H.0. and between . excluded. . is a subcomplex of C. will be the direct sum H. l ) :namely.. since the pattern sequence J = (1. Indeed. . then the homology of C. so that (k): that is.(D.) will be zero and D.J or ca. define SJ = ( j 1 . = @ . will be maximal for this property. for each relevant J . although with it is very close to being one since. j i > 1 as a component o the boundary f of C J . will carry all the homology H. . a subcomplex. one or two components of acJ lie in C . and E . were also a subcomplex. We shall also need to subdivide > 1. (kil) that of K . . U $. into . we introduce some additional notation.D. To this end. = @. for J E 3 with j .$and 3. @ E n . Then. would be the direct sum. In view of the difference in behaviour when the last nonzero term of the pattern sequence J is one. where 1) and.) @ H . We shall refer to each ca. 0) has been . + < . 0. 77) when k is even. those corresponding to a l J .J where a. (kil) itself. the first zero term. that it is non-zero when it has a non-zero coefficient as a summand of acJ. K . as 3.En are both subcomplexes and hence.(Zk/i. that If K . SJ is always a pattern and. .60 A CHAIN COMPLEX FOR SHAPE SPACES a sub-chain-complex of C. of the relevant shape space.U $. which we shall refer to as the essential chain complex. for these J . However. in general. H. . of that of C. . where (i) (i) and 3 = {J : j . U 3. 0). and 3 2 = 3. Our aim will be to decompose our chain complexes as direct sums of subcomplexes in a manner that is valid for all k and m and that facilitates the computation of homology and the inductive checking of related hypotheses. j . K . . Moreover. is not.

We also note that the coefficient of the component CJ in ~ C . Taking the case Z. @ E . Proof. = @.It is now clear that C. = @. In each case. All the sequences of changes for the various J may be carried out independently since the other components of acsJ are all in 3. For Ek we let D.. cJ occurs with coefficient f l in ac.. the inverse to the operation of talung for the maximum i for which this is non-empty. although the moves are not independent. The changes we shall use will be a sequence of elementary changes where an elementary change of basis leaves all but one basis element alone and either changes the sign of that one or adds to it one of the other basis elements. Each C.}. for J E 3. Finally. C. we let D. in each case.(Ek/lm) be the submodule spanned by and E.. ~other than CJ and CJ all have j ..(Ek). J is f l in both cases for J E $. as noted above. and E . comes equipped with a natural basis: the one used to define it. We then perform elementary changes to replace CJ by ac. we make the J obvious elementary changes to replace CJ by c. ( Z k ) be the submodule spanned by {c. for J E 3.. @ E n . and.Again.En are .J for all J E 3. since the boundary components of ~ C .(C:) = D. we replace the natural basis of C. ( subcomplexes and H . increased by one. and so are left unchanged by all moves of each sequence. We may now describe the first decompositions as follows. and in the case of Zk this is also true for J E 3.DECOMPOSING THE CHAIN COMPLEX 61 being. (D. Our method of proof will be to change that basis until the results we require are self-evident. ) k = D.}.1. @ En since D.) be the submodule spanned by {CJ :J E J. and En are spanned by complementary unions of the new basis elements of C.J by a sequence of elementary changes since.It is possible to replace CJ by ac.. Lemma 4. a) = 0. the required moves are possible since CJ has coefficient f l as a component of ac. by first replacing CJ by ac. : J € 3.(Ck) be the submodule spanned by and E . first.U $. It is clear that such a change does indeed produce a new basis. Since (A) . when J is in 3. for Z k / l .(CL) by that required for D.D.(Zk/i. for J E $. D. they may be carried out in increasing order of j .

is a subcomplex follows from the boundary formulae since. since D. That E . and so all the non-zero components of ac.U3. Thus. will be the integers. or only. As a result of this lemma. i[. We now know that the homology of Xk is determined by a complex with one basis element in degree n for each pattern J with 1JI = n and jm-l . CJ occurs with coefficient f l in ac.i. member is congruent to j . by part (i) of Corollary 4. also have their patterns in J. Then. modulo 2.. we pair the entries of a pattern J from the right-hand end and call a pair nice if it is of the form ( n 1.1 = j .1 = j . We further call a pair even. ( C i ) is repeated throughout the pattern. The new bases mentioned above for the spaces of oriented and unoriented shapes are rather different. We recall that homology refers to reduced homology and that coefficients.(C:/i. it is already clear that C. Turning now to the case X / .) = D.1. and E n are spanned by complementary unions of the natural basis elements. for J E J:. with one member of each pair being the boundary of the other. respectively. for J E J:. are paired under a so that D . odd. It remains to show that D. When m is odd we dub the left-hand singleton an honorary pair and call it nice if it is equal to k . so we shall deal with them separately and it also makes sense to consider integer coefficients first since then.J and they may be carried out independently since all the other components are in 3. the component cimJof ac. n ) . not congruent to j . j. Once again. these changes will be sufficiently simple that we can still describe the homology generators in terms of the patterns that originally determined the cells in our decomposition of shape space. respectively. is zero. To describe the generators for H. Once again. the relevant patterns are those with IJI = n and j .) we need to examine the extent to which the behaviour. has trivial homology. a. In the next section we shall perform changes of bases to render the homology calculations transparent. these changes are possible since.-l .62 A CHAIN COMPLEX FOR SHAPE SPACES D . the picture is considerably simplified. > 0._lJ = 0 and. has trivial homology. these new basis elements for D. of the last two members of each pattern determining a basis element of E . by a sequence of elementary moves to obtain the new basis comprising As before.(Xk. the maximum that it could be. For C". is generated by pairs of generators. B E . and so are left unchanged. To achieve that we change the basis of D. which shows both that D . as we shall see. > 1. + .1. when unspecified. 45 HOMOLOGY AND COHOMOLOGY OF THE SPACES . modulo 2. > 0. rather than the much larger complex we started with. if its right-hand. the other cases are obvious. is a subcomplex and that it has trivial homology.

generators of infinite summands can occur only in the complement of E*(Z. in J' would be even again and so give rise to no boundary component. nor any multiple of it. that is. except that the right-hand member of the first pair from the right that is not both nice and even is odd in the type (ii) pattern and one less in the corresponding pattern of type (iii).1. Patterns J of type (i) correspond to infinite summands because each corresponding basis element has zero boundary by part (i) of Corollary 4. which is precisely the criterion for the i' corresponding cell to lie in Z i and not in X . in the former case. modulo 2. Since all boundary coefficients are f 2 . Proof. (ii) the first pair from the right that is not both nice and even is odd.. So neither it itself. There is a one-to-one correspondence between the sets of basis elements of the remaining two types in which the corresponding patterns are identical. j i + l is odd and the corresponding entry j:+. equal to k . For m odd the initial singleton must be nice.1. For such an element c. For the first member of a nice pair can never give rise to a non-empty boundary component and aiJ is also empty when the right-hand member. Note that the pair in (ii) that is the first to fail to be both nice and even may or may not be nice.HOMOLOGY AND COHOMOLOGY OF THE SPACES 63 each basis element of the essential complex corresponds to a pattern of one and only one of the following types: (i) all pairs are nice and even.. However. to be involved in a boundary ac?. that is. There is a direct summand in homology isomorphic with Z for each pattern of z type (iii) and a similar summand in cohomology for each pattern of type (ii). can possibly be a boundary.-') in E. or the initial singleton when m is odd. the first entry in the pattern sequence must equal k . we may make a change of basis replacing each element of type (iii) by half the boundary . Thus. (iii) the first pair from the right that is not both nice and even is even. j i is congruent to j .. which means that it generates a direct summand isomorphic with the integers Z. Theorem 4. each basis element determined by a pattern of type (iii) occurs with coefficient f 2 in the boundary of the element determined by the corresponding pattern of type (ii). whereas that in (iii) cannot be nice. It follows that. The basis elements in the essential complex for Ek corresponding to patterns of type (i). dejned as above.It does exist in Xi+' but there J no longer has type (i). in J . carry the generators of injnite cyclic direct summands in both the homology and cohomology with integer coeflcients.(Ek). For then. when m is odd. since j l + l is even.1. to be nice. the basis element cJ* corresponding to a pattern J of type (i) is a cycle and it is not involved in any boundaries at all.2. it would have to correspond to a. so the J' for which it would be &J' does not exist in Xi. of the same panty as j.' where ji is the first member of a pair. Thus. or for m odd the initial singleton.

. respectively. > 1.. then the three types of pattern we identified for Ck. carry the generators of infinite cyclic direct summands in both the homology and cohomology with integer coefficients. J may occur as a boundary component of C J in Ckli. There is a direct summand in homology isomorphic with Z for each pattern of type (iii) and a similar summand in coho2 mology for each pattern of type (ii). is now the spanned by {cJ : J E Jf). In addition. Theorem 4. when the coefficients are Z2. cells corresponding to patterns in which j . C ~ ./i. for Ck/L.64 A CHAIN COMPLEX FOR SHAPE SPACES of the corresponding element of type (ii). or only. This gives a somewhat larger set of patterns than that for Ck.1 shows that the boundary formulae differ only in that. the essential complexes are so simple that it is just as easy to compute these groups directly. each generator of the essential complex carries the generator of a summand Z for both the homology and cohomology. note that the maximal right-hand block of nice and even pairs of an element is shared by all its boundary components. with this modified definition.. Thus. for either type of shape space. Then. whereas. However. Having computed the homology with integer coefficients. for EL.2. It follows that if. is a cycle and there are no boundaries. However. we have the following analogous result to Theorem 4. odd.. we may carry out the changes of basis in increasing order of the right-hand member of the next pair. for Ck.. J occurs with coefficient f 2 in the boundary of C J whenever j i is odd. 2 2 . the dual of the map n-+n is the same map in the reverse direction. examination of parts (i) and (iii) of Corollary 4. if its right-hand. although it does not include the latter set. a topologist would usually use the Universal Coefficient Theorems to deduce the homology and cohomology groups for other coefficients.. for the cohomology. sharing a common right-hand block of nice and even pairs. serve the same function for Ck/im. odd. will. C ~ . as advertised. member is even. To see that it is possible to carry out all these basis changes. boundary map Z+ Z from generators of type (ii) to those of type (iii) is zero in each case and so every basis element of the essential complex. respectively. among such basis elements. the For example. The stated results now follow since. the essential complex for Ck/i. with the previous definition modiJed as above. it required j i to have opposite parity to j. we now define a pair in a pattern J to be even. since a nice and even pair contributes nothing to the boundary and an additional even pair cannot be created in the components corresponding to the first pair that is either non-nice or odd or both. Thus.3. H 2 Turning to spaces of unoriented shapes. The basis elements in the essential complex for Ck/im corresponding to patterns of type (i).

4.. . Since the shape spaces are simply-connected and so the homotopy and homology groups agree up to and including the first no-zero one. when m is odd.5. H*(Ei. . . .H2) each have one generator of a direct summand in the appropriate dimension for each of the patterns of the relevant types (i). we may now compute their first non-zero homology groups and so complete the characterisation of shape spaces among themselves as indicated at the beginning of Section 3.. the constraints on the cells that carry generators of the homology of C:+" are such that the one of lowest dimension corresponds to the pattern (m. Theorem 4. That cell occurs in all the spaces C z f n . For example. provided m is even. Thus. (ii) and (iii). the first. . In fact.. Thus. Indeed. . it is of type (i) and this remains true in Cz+n when n > 1. with such coefficients and it follows. m. such as the rationals Q. this is of type (i). So neither type of generator of the essential complex contributes to the homology. irrespective of n > 0. .1. and it has the same dimension m ( m + 1)/2 . entry in the pattern ceases to be nice so the pattern is now of type (iii) and it determines a summand of order two in the homology.1)-connected with H. 1).CONNECTIVITY OF SHAPE SPACES 65 When the coefficients are a field of characteristic other than two. except that the constraints on members of the essential complex mean that the lowest dimensional cell corresponds to the pattern ( m 1. Similar considerations apply in Ez+"/i.8.1 = dz+' as Ez+'. we have established the following theorem. H. or cohomology. All the spaces Ez+". singleton. unless m is odd and n > 2 when it becomes type (iii). m . 3. Once again.4.272) and H * ( E k / i m . that the rational homology and cohomology has a summand Q for each pattern of type (i).(EL. For the latter space it is the only cell of the essential complex and so necessarily determines an infinite cyclic summand of the homology. On the other hand.:+I (i" X') = {Z 2 i f m is even or n = 1.Q ) each have one generator of a direct summand in the appropriate dimension for each of the patterns of the relevant type (i). for example. + + Theorem 4.6 CONNECTIVITY OF SHAPE SPACES Before computing the homology groups of shape spaces in detail. for n 3 1. . f m is oddand n > 1. Q ) and H * ( E k / i m .2) and so has dimension dz+' m. 2 . are (dz+' . the boundary map from generators of type (ii) to those of type (iii) is an isomorphism. we close this chapter by making some specific observations that follow directly from these theorems. we have established the following extension of our earlier theorem. but none for the other patterns. the fact that the essential complexes carry all the reduced homology and yet have no cells in low dimensions confirms our previous result that the shape spaces are highly connected..

11 and that K’ K” is homeomorphic with P x [0. n 1/21 onto Kk-l K . respectively. and write Q x [ n . despite being so well-known. that stage of the composition is indeed well-defined and a homeomorphism. it is not easy to locate a proof in the literature. which restricts to the natural identification of the ‘lower’ boundary Q x (0) of K with that Q x { l} of K”. together with the existence of a homeomorphism h of K onto K”. finally translate Kb UK1 onto KLPl UK. l). all the spaces ZE+”/im. Such a homeomorphism will be proper if for all n the projection of the inverse image of Q x [-n. as the composition of a sequence of homeomorphisms: first identify P x [n . since the maps of K’ and K” onto Kb and K1. ifm is odd and n > 2. on the ‘upper’ boundary.1)-connected with H.1/2. of P x [n .1/2. we may describe a homeomorphism g. is a ball of the same dimension. Then. on their common boundary Q x ( I } are both the identity. where the two Q copies of Q x ( n } are identified naturally. K” of the components of f ( P x [1/2. n ] onto R is bounded. We are also now in a position to prove that none of the shape spaces are topological manifolds other than those we already know explicitly to be so. UKL.respectively.66 A CHAIN COMPLEX FOR SHAPE SPACES Similarly. of P x [O 21 in Q x [O 21 such that f ( P x {1/2]) C Q x (0. f ( P x { 3 / 2 } )5 Q x ( 1 .. Two things need to be checked. 11. and which. Proof. 3 / 2 ] .3/2])\Q x ( 1 ) that contain f ( P x ( 1 / 2 } )and f(P x { 3 / 2 } ) . 2 )and Q x ( 1 ) f ( P x [ 1 / 2 . 3 / 2 ] ) . because of our u u + + + u + . We recall that a map is called proper if the inverse image of every compact set is itself compact. We could similarly have shown that CEf2/i.. n 11 = K . We regard Q x R as UnEZ x [ n . Firstly. n 1/21 naturally with P x [ 1 / 2 . since. by Casson’s theorem. + Lemma 4 2 Let P and Q be compact topological spaces and f be an embedding .then map by f onto K’UK”. we already knew that these particular spaces would have the homology stated in this theorem.:+I+. Z. is homeomorphic to a sphere of dimension dE+’ m.+’ is a sphere of dimension dE+’ and C : + ’ / i . (CZ’”/lm) = U 222 i f n = 2 or m is even and n > 2. 2 ] \ f ( P x [ 1 / 2 . Of course.Then P x R isproperly homeomorphic with Q x 08. The required homeomorphism of P x 08 onto Q x R then follows by the socalled ‘Mazur trick’. 2 K and KL Z K‘ for each n. Secondly. then identify K’ with Kb and map K” onto K1 by the inverse of h followed by the natural identification of K with X I . where K . The proof rests on the facts that K K’ is equal to Q x [0.n l ] . and the closures K’. n 3 1. We use the following result from topological ‘folklore’ for which we shall include a proof since. 3 / 2 ] )that contains Q x (0}. thereby explaining its homology groups. are (dE+’ for + m . Consider the closure K of the component of Q x [ 0 . restricts to the natural identification of f ( P x {1/2}) with f ( P x { 3 / 2 } ) .

21 onto itself. We do this in two stages by means of the product structures on P x [0. + 1. This then implies that Sk-3 * has the same homology as a sphere of dimension d. fixing the vertex until it lies in the interior of the other as required by Corollary 4.CONNECTIVITY OF SHAPE SPACES 67 choice of h between the ‘upper’ boundaries. In fact. fixing the boundaries and mapping P x {1/2} onto P x {3/2}. for example. The resulting proper homeomorphism of P x R onto Q x R induces the required homeomorphism between the open cones since these are the one-point compactifications at one end of each space. respecting the cone points. fixed on Q x [ 1 E .3)-fold suspension since for any suspension S‘ (P) of a space P the homology group H . It remains to construct the homeomorphism h. This uses the fact that in a topological manifold every point must have a neighbourhood that is a topological ball. The composition of the two homeomorphisms.2]. gnlPx{n+l/21 = gn+llPx(n+l/Z) and so the homeomorphisms g. Then. is the required homeomorphism h. Proof. together to form a homeomorphism. the only information we use from this corollary is that P and Q must be homotopically equivalent and so have the same homology groups. . respectively. with rn > 2 and k > rn . using compactness of P and Q. is a topolog- Proof. between P x R and Q x R. a homeomorphism of P x [0. which we can also think of as a cone on a sphere with the given point at its vertex.1. 1 E] _C IC’’ and we can produce a homeomorphism of Q x [0. n E Z.2 are easily achieved. First. will induce a homeomorphism of Q x [0. . respecting cone points. fit which by the nature of this construction is necessarily proper.2. + .I (P). we already know from the mapping cone decomposition that there are points in Xi that are at the vertex of a neighbourhood homeomorphic with the cone C(Sk-3* X:IJl). fixing the boundaries and taking f ( P x {1/2}) onto f ( P x {3/2}). 21 onto itself. provided k > rn 1 and rn 3 3. then P x R 2: Q x R and Co(P) 2: Co(Q). Hence neither does its ( k .2] onto Q x [1. I P and Q are compact spaces and the open cone Co(P) embeds f in the open cone Co(Q).2. 21. when restricted to K .6. ical manifold. (Sl (P))is isomorphic with H . Theorem 4. be induced by a suitable increasing function of the factor [0.. However. we can find E > 0 such that Q x [1.2] and Q x [0. However. The hypotheses of Lemma 4. Theorem 4. extending the natural identification of the ‘lower’ boundaries. No shape space X. which could.5 shows that ll:: X has non-zero homology groups that are not shared by a sphere of the same dimension. We can obviously shrink one of these neighbourhoods. 21. + + Corollary 4. by compactness.2] onto itself.

1. (XE/i.7 LIMITS OF SHAPE SPACES A further direct consequence of our basic homology theorems concerns the limit of the shape spaces for fixed m and varying k . and for k > n . with the identical proviso for change of pattern type. except that when m is odd cells corresponding to patterns of type (i) in which j l = k . which may be thought of as the space of shapes of infinite sequences of points in 08" and arises as follows.(Xk. and for k > n . ..W = lim C k. Similarly. m 2. is a subspace of Ck+' for each k we may define the limit C. we have shown the following result. + + + Theorem 47 H . m . . ( X L ) for k 3 n . when m is odd.d:+' 2m. . In addition.. .(Ck+').) determine cells of the same type in E.(Ck+'). ( Ck. when m is even. + + + When m = 2 the shape space C z is well-known to topologists as CP". .m.d:+' . There is a corresponding inclusion of the essential complex E . / i .68 A CHAIN COMPLEX FOR SHAPE SPACES 4. ( C F ) Z H . ( Z . Thus. H .) 2 H .dE+' + 2m. Since Ck. + we could observe that the involution 1.1 become type (iii) in E. when m is odd.dz+' m. it is well-defined on X F and consequently the notation we have used for the limit is not misleading. Alternatively. 1) and so has dimension d:+' k . and plays a fundamental role in both algebraic and geometric topology. rn . We may define C z / i m= lim Xk/im and its topology similarly. is open. ) for k 3 n .2.+' ) and the lowest dimensional cell of the latter that does not appear in the former corresponds to the pattern ( k . = m 1. . when m is even. infinite dimensional complex projective space. . with the topology determined k by requiring a set in ZE to be open if and only if its intersection with every Xk. Hence. Similar remarks apply to the homology of unoriented shape spaces. the cells of E. . 4 to be the union of the Xi. m. ( C k ) in E . has the same effect on a cell to whichever space it is regarded as belonging.

Where.2. these are not particularly transparent or easy to use so. C A R " & H.7 we study a number of ways in which the essential complexes may be decomposed as a direct sum of subcomplexes. we shall illustrate the computations in spaces of shapes of configurations in two. KENDALL & D. which could. allow an inductive computation of all the groups. before giving the general results. which is the number of partitions of a set of size i into j parts each of which may be empty but has at most 1 elements. Thus. Since the generators were identified in the previous chapter. this counting will require the partition function. In Section 5. 2. In Section 5. These examples are varied enough to display the different types of behaviour that occur in general and to check our formulae. Note that in each table the value of m. The results are displayed in a sequence of tables. l ) . which we call P(i. groups of shape spaces. contrary to the arrangement in Tables 1. 1) would now be written J = 4321. Some of the formulae are rather too intricate to give directly in terms of these functions. while the value of k .8 we give the explicit formulae for the homology groups. the dimensions of the cells or. A generator of a chain complex is indicated just by the pattern that determines it. or of the resulting homology groups. increases along the horizontal axis. the degrees of the corresponding generators of the chain complex. all the members of the pattern sequence J are known explicitly we shall write it more concisely without intervening commas or enclosing brackets. LE Copyright 0 1999 by John Wiley & Sons Ltd CHAPTER 5 The Homology Groups of Shape Spaces In this chapter we obtain closed formulae for homology. and hence also cohomology. J = (4. in theory at least. all that remains is to count those generators. In addition to binomial coefficients. For the spaces C k / i mthis could be self-contained but any inductive calculation of the homology of the spaces Zk seems to require prior knowledge of that for the spaces of unoriented shapes. but our results do allow an explicit formula to be obtained for . is fixed. the dimension of the space in which the shapes lie. increase down the vertical axis and remain constant along each row. BARDEN & T. as will now frequently be the case.1 and 1. K. Then. This leads to recurrence relations among the various homology groups.j . three and four dimensions that have a sufficiently small number of vertices.3. the number of points defining the shape.Shape and Shape Theory D. However. G. equivalently. This also applies when J is used as a suffix to name a cell or generator.

comprising the leftmost one in each non-empty row. for all n > k . 5.it also appears in F.B.1 shows the essential complexes for spaces of oriented shapes in 2-space. In all subsequent diagrams we shall use this feature . In the final section we show that the notable feature. The cumulative nature of the essential complexes as k increases is clear. it is already obvious that all boundary maps Table 5. by displaying each cell only once: in the complex in which it first appears.C..1 SPACES OF SHAPES IN 2-SPACE Table 5. once a cell occurs in Xi.. . In this case.1 The essential complexes for spaces of shapes in 2-space 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 21 21 32 21 32 43 21 32 43 54 21 32 43 54 65 21 32 43 54 65 76 21 32 43 54 65 76 87 21 32 43 54 65 76 87 98 21 32 43 54 65 76 87 98 A9 21 32 43 54 65 76 87 98 A9 BA 21 32 43 54 65 76 87 98 A9 BA CB 21 32 43 54 65 76 87 98 A9 BA CB DC 21 32 43 54 65 76 87 98 A9 BA CB DC ED 21 32 43 54 65 76 87 98 A9 BA CB DC ED FE . represent the numbers 10. .11. . akin to duality.70 THE HOMOLOGY GROUPS OF SHAPE SPACES any particular homology group or family of such groups. The corresponding version of this diagram would just be a single diagonal of entries.. that appears in our first examples does indeed hold in general.12. . Each complex is a subcomplex of its right-hand neighbour and. to simplify the picture. granted that the only essential cells are the ones displayed. where the letters A. In these complexes all the cells are of type (i) and hence carry generators of infinite cyclic summands of the homology.

As before. there is no need to re-base the essential complex in order to read off the homology. each entry Z2 in the diagram exists not only at its location but also. there are also generators of infinite cyclic summands. for the only cells that appear. unlike the case for the homology. as indicated in the theorems.(Xi+2") and H . in every space to the right of that in which it appears. ( X i + 2 n ) each have n summands equal to Z. we find that H 1 l ( X i o ) = Z2 @ Z and H I I ( C ~ ) Z2 @ Z2 = for k > 10. these the first always has type (iii) and the second type (ii). in the same degree. using the Universal Coefficient Theorem. Similarly. they follow directly from the homology computations. in particular.(C$) is never a direct summand of H.(X:+2"). We note also that. These results may be obtained.(E:'2n) and H. respectively. Each arrow in the diagram indicates. as we showed directly from the homology exact sequence. that the homology exact sequence of Chapter 3 does not break up into a set of short exact sequences as we might have hoped. a cell with non-zero boundary and. This is the case. at its tail. In Table 5. in each shape space to the right of its first occurrence. in the same degree. t ' 5. in H 11 (Xio) and H 11 (Xi1) calculated above. for example. by reversing the arrows in the essential complex.(X. This shows. the essential complex for Cg for k 3 10 has three generators in degree .2 SPACES OF SHAPES IN 3-SPACE In Table 5. In Table 5. In these complexes there is only one such cell in each non-zero boundary. which for X 2: CPkP21s 2k . the summand represented by '00' corresponds . In this and all subsequent diagrams the complexes are cumulative: each space is a subcomplex of its right-hand neighbour. Alternatively. each shape space has its non-zero homology group of highest degree isomorphic to Z and its degree is the same as the dimension of the space itself.SPACES OF SHAPES IN 3-SPACE 71 must be zero since. that H. and we have adopted the convention of displaying each generator only once: in the complex in which it first appears. However c9*21 of type (i) in Xio but becomes type (iii) on inclusion in is X: for k > 10. we see that H. However. This diagram is also mainly cumulative in that each summand Z2 also exists. there is never any essential cell in the next lower dimension. Thus. which do not persist but on inclusion in the next largest space generate the summand Z 2 noted there. a cell that appears in that boundary with coefficient f 2 .2 we show the essential complexes for spaces of shapes in 3-space.3 we show the resulting homology groups of these shape spaces. hence.4 we display the cohomology groups of the spaces of shapes in 3space. Thus.+') and. However. Hence. all of which become Z2 on inclusion in H. Of eleven: c & ~~7*32and cg*21. We recover the well-known homology groups of the complex projective spaces: Z in each even dimension up to the real dimension of the space. while the entries '00' are only relevant to the locations in which they appear.4. at its head. indicated by 00.

2 The essential complexes for spaces of shapes in 3-space deg Z : Z: C : C : c: c c:” xi‘ xi2 y xi4 cis 5 6 7 8 9 I0 11 32 1 -421 52 1 432 -621 -532 72 1 632 -821 543 -732 92 1 643 832 -A21 743 -932 B2 1 654 -843 A32 KC21 -754 943 XB32 0 21 854 -A43 C32 E21 765 -954 B43 -032 -865 A54 -C43 E32 965 \B54 043 E43 876 -A65 c54 976 865 -D54 A76 -C65 E54 987 -B76 065 -A87 C76 E65 B87 -076 A98 KC87 E76 B98 087 C98 E87 BA9 “098 CA9 E98 DA9 CBA -EA9 DBA EBA DCB ECB 12 13 14 IS 16 - 17 18 I9 20 21 22 23 24 25 26 21 28 29 30 31 32 33 34 35 36 37 38 39 40 41 ---. The essential complexes for these spaces are indicated in Table 5. As before.72 THE HOMOLOGY GROUPS OF SHAPE SPACES Table 5.-EDC F21 F32 F43 F54 F65 F76 F87 F98 FA9 FBA FCB FDC FED to a summand Z in the next higher degree cohomology group of the next space 2 in the sequence. now there are cells with more .. each arrow goes from a cell to one of its non-zero boundary components and every such boundary component is indicated.5. However. 5 3 SPACES OF SHAPES IN 4-SPACE .

.2cq. which first appears in Zf.6. This we do in the next table.. Table 5. 7 indicates . .3 The homology groups of spaces of shapes in 3-space deg 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 C i C: Cg C: C! Cg Cko Ckl CY Xi3 Ci4 Xi5 C' i 33 34 35 36 37 38 39 40 41 than one non-zero boundary component. All boundary coefficients are f 2 and the sign does not affect the calculation.SPACES OF SHAPES IN 4-SPACE 73 Table 5.. where we show the same essential complexes for spaces of shapes in 4-space but with the new choice of basis. has boundary 2c&. so the diagram contains all the information we need to read off the homology of the spaces. The symbols that are not overlined represent the same basis elements as before. For example. cq532. However.2. however.. the structure is clearer if we first carry out the change of basis used in the proof of Theorem 4.

which. For such J . for .. some symbol J ” . Thus.cT432. Then. 2cf is one component of the boundary of a cell c.5. which fact is indicated by an arrow from the symbol J’ to the symbol J in Table 5. the symbol 7 in Table 5.6 represents half of the full boundary of c.74 THE HOMOLOGY GROUPS OF SHAPE SPACES Table 5.twice which is the boundary of remains in the new basis.The element c... since all boundary components have coefficient f 2 and in all cases here there are just two of them. previously represented by J .4 The cohomology groups of spaces of shapes in 3-space S 0 0 z 2 6 7 8 9 10 11 12 13 14 1s 0 0 0 0 722 722 00 0 0 00 722 722 00 722 0 0 0 0 z 2 z 2 z 2 0 0 0 0 z 2 z 2 z 2 00 00 00 0 0 z 2 0 0 0 0 722 722 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 z 2 z 2 00 00 0 0 0 0 0 0 722 722 722 00 z 2 z 2 z 2 0 0 0 0 z 2 z 2 0 0 z2 0 0 0 0 z 2 z 2 0 0 0 0 722 00 0 0 00 772 0 0 z 2 z 2 z 2 00 722 00 0 0 0 0 00 0 0 z 2 0 0 z 2 0 0 0 0 z 2 z 2 z 2 00 00 00 0 0 00 z 2 00 722 722 772 0 0 0 0 722 722 0 0 0 0 0 0 a new basis element to replace the basis element c...432.. the entry 7432 represents and which replaces the sum c& . the old basis element c.1532 . is c f c.

5 The essential complexes for spaces of shapes in 4-space deg 9 10 11 E : 4321 Et x : c:: c.’ 5321 5421 6321 6?2 1 7321 7421 8321 7521 8421 4z 12 13 14 15 16 5432 6521 6432 ‘ 4z 6532 7432 7621 8521 2 3‘ ! 7 6543 7632 7543 1 2‘ ! 8 8432 4z 4z z4 9321 9421 A321 9521 A421 A521 A432 A532 A632 17 18 8721 8532 9432 9621 4f 4z 19 20 21 22 23 7643 8543 8732 9632 9821 A721 ‘8k3 7654 8743 8654 29$32 94‘ !1 53 2 A 47 4 9832 9643 A543 A921 A732 9743 A643 A832 A932 A654 A943 A754 8754 9654 9843 A743 3 4 A ! ‘ 8765 9854 9765 4z 4z z4 4z 4z 4z z4 4z 4z 4 4 4 24 25 26 27 28 29 30 31 32 33 9876 A854 A954 9865 A765 ‘Ak5 A965 4 A876 A976 4 A987 .SPACES OF SHAPES IN 4-SPACE 75 Table 5.o c.

4 8643 8632 9532 9721 A62 1 4 4 A432 4 7643 8543 8732 9632 9821 A 7 2 1 4 A532 A632 9732 9543 A821 7654 8743 8654 9832 9643 A543 A921 A732 4 4 4 9743 A643 4 8754 9654 9754 4 A832 ~ 4 4 9843 A743 A843 4 A932 A654 A943 A754 8765 9854 9765 9865 A765 A865 9876 4 4 A854 4 4 A954 A965 A876 A976 4 A987 .4 9521 A42 1 6543 7632 7543 8721 8532 9432 9621 A521 4 .76 THE HOMOLOGY GROUPS OF SHAPE SPACES Table 5.6 The re-based essential complexes for spaces of shapes in 4-space 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 4321 5321 4 5421 6321 642 1 4 7321 4 ~ 5432 6521 6432 7421 8321 7521 842 1 4 4 4 6532 7432 7532 4 9321 ~ 4 4 7621 8521 8621 8432 942 1 A 3 2 1 4 .4 4 4 .

then move on to 7432 and finally 6321. in E. As usual.(C:) is a direct summand of H .”( ac. and that we can always take the sum or difference of two existing basis elements to replace one of them in a new basis. plus or minus.” will have at least one component c. Corresponding to the direct sum decomposition of the essential complexes. of E. so are denoted by 72. This suspension isomorphism onto the complementary summand is perhaps even clearer at the level of the essential complex where it is . qua chain complex. then 8521 and 8321.6 are vertical.7. the diagram is cumulative. since all components c. The homology groups of spaces of shapes in 4-space are shown in Table 5. For example. corresponding to a basis element of the essential complex. although never for odd values as we have already seen for the case m = 3. ( C : ) lies in the column under the symbol Cz. together with.” have J’ < J . Then. The reason is that the boundary formula ensures that. In this case. to signify this fact. for m even and any J with leading entry k . The proof that the new elements do form a basis follows from the facts that each new element is the sum of an old one.6. It includes most of those for the next five spaces whose chain complexes are not displayed in Tables 5. all the revised boundary maps in Table 5. This fact is visible in the table. so this choice of bases makes it clear that E . ( C $ + ’ ) . For example.5 and 5.3 we see that this is identical with the homology of Xi. forming c. This phenomenon also occurs for all other even values of m. we take as ‘foundation’ for the new basis element that component of the boundary that is determined by the largest symbol with respect to their natural ordering. the boundary ac. ( C $ ) is a direct summand. when there may be more than one component of the boundary that does not appear in a previous space of the sequence. for which this is not the case. ( C i ) that is not the image of H .From Table 5.(Cz) ___ we can form 8543. in decreasing order of of J will not lead us into any trouble. In particular.except that all the degrees have been increased by eight. one from a previous space in the sequence. that part of H . rather than 00. because of the choices of new basis elements that we have made. the infinite cyclic groups also persist in the next space. ~ ~ ~ ~ The last three spaces also have non-zero groups in degree greater than 41. ( X $ ) suspended k times. More generally for larger m..(C$’’). which are not shown in the table. For example. Working thus in decreasing degree of the homology and then decreasing value of k we ensure that we never change an old basis element that we may need later to add to another to form a new basis element. Moreover. we see that the homology H. whereas for m odd there will always be J . for which the leading entry of J‘ is also k .SPACES OF SHAPES IN 4-SPACE 77 This choice ensures that all the boundary maps with respect to the new basis are either zero or map to twice another new basis element. the type of the generator does not change on inclusion in the next larger shape space. It follows that the exact homology sequence of Chapter 3 must split (see the Appendix) and that the complementary summand must be isomorphic with H .

7 The homology groups of spaces of shapes in 4-space z z z 2 2 €3 z2 7 achieved by prefixing a symbol that represents an essential cell in X3 by an ‘8’ .and so we have not shown them. whereas that in Xi would not have been. the shapes of configurations in 2space already produce cells with more than one non-zero boundary component.8. Unlike the oriented case. in appropriate cases. 5.However.8 to produce the symbol for the corresponding cell in Xi. The cohomology groups differ from the homology groups only in that each summand isomorphic with Z appears in the next higher degree precisely as was 2 the case for the cohomology of the spaces Xi.4 SPACES OF UNORIENTED SHAPES IN 2-SPACE The essential complexes for spaces of unoriented shapes in 2-space are shown in Table 5. . the new symbol is an overlined one.78 THE HOMOLOGY GROUPS OF SHAPE SPACES Table 5.

SPACES OF UNORIENTED SHAPES IN 2-SPACE 79 Table 5.8 The essential comolexes for maces of unoriented shaoes in 2-soace 4 32 42 !3\52 3 $ ‘ 54 62 t37 !\2 3! ‘4 6 82 5 6 7 8 9 I0 II 5 !\ 3!\ !4 8 A2 A3 ‘B2 12 13 14 15 16 4\ A4 \ ‘83 a 5 C2 23\D2 c 4 v 3 CiD !z 4 kzB4 A6 b 98 B 6 ‘: 7 7! ‘8 A 17 I8 19 A9 \ 8 B ‘ 5C L !‘ 6 i7\ 6 !D 7‘ L !8 C B 7:9! A LC CA 9! ‘ L 20 21 22 CB 4z D9 DA 23 24 25 26 27 28 FE .

-'/12) with more than one boundary component./12)\E. as in C!.(C:+2"+1/~. for J = ( j l . The homology of these spaces of unoriented shapes is displayed in Table 5. Note also that here again the infinite cyclic groups are indicated by Z since they do persist as infinite groups in the next space. Z2 indicated there. Thus.(Ck. that is. the spaces with top homology group zero have top cohomology group.(Ci+2n/13) cannot be a direct summand of H.). As usual. j3 1) so that such replacements can only occur when j 2 > j 3 1. On the other hand.(Ck. E. equal to Z2.(Ci+2"/13) has n summands equal to Z. the overlined basis elements represent generators CJ that have been replaced by ( 1 / 2 ) k s J . or one of the groups.80 THE HOMOLOGY GROUPS OF SHAPE SPACES However. H18(Ci2/13)is a direct sum of six copies of Z 2 .1 1.(Ck./l.(Ck. The zeros indicate homology groups that are zero and whose degree is equal to the dimension of the space: a phenomenon that did not occur for spaces of oriented shapes. are indicated by an overline. Table 5.5 SPACES OF UNORIENTED SHAPES IN 3-SPACE We move straight to the re-based essential complexes for spaces of unoriented shapes in 3-space.(X$+2"/13). becomes the group.)\E. Similarly.(CF-'/13) cannot be a direct summand of E. j 3 ) .(C?/13) since the homology is not a direct summand. Note. ( X y / 1 3 ) is a direct summand of E*(Cp+'/13). whereas the other spaces.(C:+2n/in) will not be a direct summand of H. As before. as was the case for all spaces of oriented shapes.10 shows the resulting homology groups of spaces of unoriented shapes in 2-space.Here. on inclusion in the next space. Since the cohomology is similar. Thus. H.+'/12) and so the corresponding homology group H. All the revised boundaries are vertical. These re-based essential complexes for spaces of unoriented shapes in 2-space are shown in Table 5. that in the dimension of the shape space itself. have their top cohomology group equal to Z. for each cell in E. j 2 . As above. As before. however. (Ck. we can see + + . which.(C./12) will be a direct summand of H. The fact that no 'diagonal' boundaries are now required means that the essential complex E.-'/12) and so as before we can choose the new basis elements so that all the nonzero boundary maps become vertical. with each 222 appearing in the next higher degree. ) for all even m and all k . for all odd m.all of which become Z2 on inclusion in H. so that H.(Xt/12) is a direct summand of E. 5. On the other hand. ( C ~ / L . the generator at the head of an arrow is half the boundary of that at the tail. E .(C:+2n/13).9.k. sJ will be ( j l . which are shown in Table 5. j 2 . Thus. H. at least one of those components is also in E. the new basis elements.12. the change of basis will remove the diagonal components since. In general. where different from the old. This feature will occur in H .(C:+'/12). that the change of basis does not remove all the 'diagonal' boundaries. the diagram is cumulative. the summands Z2 persist in each space to the right and '00' represents an infinite cyclic summand 22.

9 The re-based essential complexes for spaces of unoriented shapes in 2-space 4 5 32 42 4 6 43 sz 53 54 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 t 62 4 4 63 64 65 73 4 82 74 75 76 4 4 83 84 92 93 4 t 4 A2 85 86 % 95 t - A3 A4 B2 B3 B4 B5 4 4 C2 4 87 96 97 98 4 A5 A6 4 C3 C4 - 4 02 03 t 4 t A7 A8 A9 B6 B7 B8 - 4 E2 t C5 C6 C7 C8 04 D5 t 4 E3 E4 F2 F3 - 4 t 4 - B9 BA 4 06 07 4 4 E5 E6 F4 4 4 C9 CA CB 08 D9 4 F5 4 El F6 F7 E8 4 4 E9 DB DC 4 F8 F9 - EA 4 t t EB EC FA 4 FB - ED FC FD FE 4 .SPACES OF UNORIENTED SHAPES IN 3-SPACE 81 Table 5.

(Cy2"/13). is precisely what prevents H. . in fact.10 The homologv of spaces of unoriented shapes in 2-space 27 28 z that each homology group H . Again.(X:+2"/i3) being a direct summand of H . ( C : + 2 " / ~ 3 )Similar results hold for all odd m . ( X i + 2 " / ~ 3 )is a direct summand of H . which map to isomorphic summands of the latter. which .6 SPACES OF UNORIENTED SHAPES IN 4-SPACE For our final examples we calculate the homology of spaces of unoriented shapes in 4-space. 5. ( X : + 2 n / 1 2 ) does not persist with infinite order in H. that the suspension of an element of infinite order in H . This. however. we start from the re-based essential complexes. Note.(Ei+2n/~2). As before.82 THE HOMOLOGY GROUPS OF SHAPE SPACES Table 5. and easily derived from.(C2f2n/12). ( X : + 2 " / ~ 3 ) since all summands of the former are Z2. There is a homology exact sequence for the homology groups of spaces of unoriented shapes analogous to. that for spaces of oriented shapes and we see here that. although the original element did so persist into H. the zeros indicate 'missing' groups in degree equal to the dimension of the space and the corresponding cohomology groups are all equal to Z2. once again. the complementary homology groups are the 2nth suspension of those of H.

(C:/14) is a direct summand of H . as for Ck.f. This can easily be seen from Tables 5.11 The re-based essential comolexes for maces of unoriented shaoes in 3-soace 8 9 43 2 ‘532 542 543 642 632 ‘I32 10 II 12 13 14 15 4 643 652 4 - 142 832 842 ‘ 9 3 2 843 653 752743 654 1 5 3 162 2154 1!3 862 164 765 4 4 .The complementary summand is equal to Hn-k(C. The . (C$+’/14). 4 852 853 942 A32 A42 A43 A52 f 16 17 18 19 863 872 854 953 962 864 873 865 874 875 876 4- 4 f 4 ‘ 952 943 4 4 - 4 4 954 963 972 A62 A53 4- 4 - 964 982 973 A63 A54 965 983 4 A72 4 f 4 974 A64 A82 A73 4 20 984 915 A65 A83 f 4 .12 and 5. In this case. all the boundary maps are vertical in the re-based essential complex and so each H .14. The resulting homology groups are in Table 5.13.SPACES OF UNORIENTED SHAPES IN 4-SPACE 83 Table 5. as well as those for Ci1/14 whose essenfor tial complex there was insufficient space in Table 5. for spaces up to and including Ci0/14./13). in Table 5./i*.t A92 A74 21 ‘976 2 5 986 987 A84 A93 A15 A76A85m A86 A95 A87 22 23 24 + + 4 4 4 4 -t A96 A91 25 26 A98 are displayed.13: it has six cells in each of the dimensions twenty-four to twenty-seven.14.

the correspondence being achieved by preceding the pattern for a generator in Ci/ig by a ‘9’ to obtain the pattern for the corresponding generator in Ci0/14.13. are in bijective correspondence with the entire set of generators in the essential complex for Ei/t3. 5.7 DECOMPOSING THE ESSENTIAL COMPLEXES In this section we continue our analysis of the chain complexes for shape spaces by further decomposing the essential subcomplexes in ways that lead to relationships between the homology groups of two or more different spaces. the list of additional generators in the essential complex for C . which were not already in the essential complex for Cz/14. we .84 THE HOMOLOGY GROUPS OF SHAPE SPACES 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 0 0 z 2 n 2 0 0 z 2 z 2 0 z2 z 2 z 2 00 zze3z2 772 0 0 0 z 2 0 0 0 27 28 29 z 2 0 reason is clear from a comparison of Tables 5. the stated isomorphism of homology groups follows. i 0 / 1 4 . This is a special case of the first theorem in the next section. This correspondence respects boundary maps and. First. For example.11 and 5. since those resulting in Ci0/14 are vertical.

.* I . which are the result of the splitting. of the homology exact sequence for shape spaces: H n .k ( . In fact.. when it does not split and .13 The re-based essential complexes for spaces of unoriented shapes in 4-space deg 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 C$/L~C:/L~ C:li4 C:lL4 ci0/L4 5432 6432 4 6532 6542 ~ 7432 4 7532 ~ 8432 4 6543 4 ~ 75427632 8532 85428632 9432 4 75437642 f 9532 4 76527643 f ~~ f 95429632 8543 8732 8642 4 7653 7654 4 f f 8752 ~~ f 4 ~ 8652 8742 8643 8653 8743 9543 9732 9642 f 965297429832 9643 f 8654 8762 8753 4 ~~~ f ~ 9653 9743 9752 9842 f 4 f 4 8754 8763 8764 965497629852 97539843 4 4 ~~ 4 4 f 975497639853 9862 976498729854 9863 8765 4 ~ 4 4 f 9875 9876 9874 9865 9765 9873 9864 look at the isomorphisms. I( ! .DECOMPOSING THE ESSENTIAL COMPLEXES 85 Table 5. -+ H n . in special cases.d. t* Hn (i) H n (kL') d.k + l ). m odd for Ek and m and k both odd for Ek/im. k l ) . the sequence splits in nearly all cases: there are just two circumstances..

86 THE HOMOLOGY GROUPS OF SHAPE SPACES Table 5. To write them uniformly for all n we take the reduced homology groups in negative degrees to be zero. It also splits in all cases when we take the homology with Z2-coefficients.1. induce . Theorem 5. For all m 3 1. from which it is also clear that they are the best possible.('L'). The results for integral homology are illustrated in the special cases we discussed in the previous sections. The inclusions a : Ek homomorphisms a* : En ~f Ek+' and a : C k / i m 9E. k and n 3 0 there is an isomorphism of reduced homology groups Proof.+l/Lm En (k) .14 The homology of spaces of unoriented shapes in 4-space 13 14 z z 2 15 16 17 18 19 20 21 z 2 z 2 z€3772 z 2 z 2 z 22 23 24 25 26 27 28 29 30 31 32 so prevents the homology groups being particularly simple to compute.

(L) (kL') Note that. and we have the result. Then. (L). in the chain complex for computing &-homology where all coefficients are evaluated in 272. and the rest of the proof is as before.Similarly. injects. the homomorphism 2. that is the only possible obstruction to the direct sum decomposition. ( k ) is the same as H. So.1 and part (i) of this theorem are valid for all n . is zero and we have the short exact sequence: (ki').DECOMPOSING THE ESSENTIAL COMPLEXES 87 which provide an isomorphism of the essential complex E. + Theorem 5. the complex E . since the isomorphisms in Theorem 5. in both cases. (k) (L). The above result only has content. we used above to elicit the homology and cohomology generators. For all m 3 2. if k 3 2m for the case of Xi and k 3 2m 1 for C L / i m .-k(mkl).(L) onto a direct This follows from the bases of the essential complexes that summand of E . However. induces. part (ii) of . This shows that the homology of the direct complement to E . we have the following results.j component of the boundary is zero and. the similarly labelled homomorphism that occurs in the exact homology sequence. In all cases. for each value of n . of course. (kL'). which again have content only for the analogous values of k . the Universal Coefficient Theorems ensure that the isomorphisms hold for reduced homology groups with any coefficients. those basis elements not in the image of the complex E.2. Thus. maps to a direct summand of E . in each case.(L) form a sub-chain-complex that is a direct complement of it in E . In the case of C$+'/im. all the boundary maps are zero for either type of space. the ca. and the correspondence between those of types (ii) and (iii) preserves the first entry in the pattern and only i31J can produce components in En-' Thus. since 2. k and n 3 0 there are isomorphisms of reduced homology groups (ii) Proof. once again. the boundary components have coefficients f 2 so.1. as explained in the proof of Theorem 5. each d. since each of the new basis elements that does not lie in En has no boundary component in En-l For. the elements of type (i) have no boundary.

this correspondence respects boundaries since the ith boundary component is determined by the parity of j i . is generated by the basis elements corresponding to those pattern sequences J . in the excluded cases. . For all m 3 3. There is no similar extension for homology with coefficients in Q. unclear. except for homology with coefficients in 222. .(Xk+') of F . if any..-2 . . E E n . is generated by the remaining basis elements. Then. where F .' is non-zero if and only if j . . That these are subcomplexes follows easily from the boundary formulae. j . 2. . whose last two entries are '21'. Historically.3. The first result involves spaces of both oriented and unoriented shapes. where J' is the pattern sequence ( j l . this correspondence respects boundaries since the ith boundary component in Xi. Then. for J = ( j l . . in Jy. . where J' is the pattern sequence ( j l . Thus.(C. the above isomorphisms were expected since they arose from the known homology exact sequence and were involved in the initial computation of the homology groups. Thus. that underly them are. are precisely those that prevent the direct sum decomposition by changing their nature on inclusion in the next space.1). and the inductive computation of the homology of the spaces X. E . starting from the known ones. .88 THE HOMOLOGY GROUPS OF SHAPE SPACES this theorem is really only filling in the cases for Z2-coefficients that those theorems omitted. 1) the basis element CJ E E. since the summands there are generated by elements of type (i). Again. This results from a decomposition of the essential complex.k2/i2 in the special case of Theorem 5. k and n 3 0 Proof. together with the known homology of the spheres X and the t complex projective spaces Xi. So.1.enables the inductive calculation of the homology . as yet. which. corresponds to a basis element C J J E En-m-l(Xk-2/im-2).j . . starting from the known homology of the real projective spaces X?/i. the correspondence induces an isomorphism of homology groups. ./im for all m > 2. However./im-2 is non-zero if and only if j i is odd. The topological identities.m ( X k ) .1). j m ._. Theorem 5. @ G.3 gives the inductive computation of the homology of X.. . they just fail to give a simple inductive computation of all the groups.+') of G. . Moreover. . = F .2 . . Similarly. Corollary 5. corresponds to a basis element C J . which is unchanged. and G. j. of X for m > 2. In the next two theorems we give isomorphisms that only became apparent to us on studying the special cases that we computed in the first sections of this chapter. each basis element CJ E E. while that in C.4. these isomorphisms do have the advantage of allowing an inductive computation of all the homology groups. this information. We note that. it is a chain isomorphism and hence induces an isomorphism of homology groups. is i odd.j. again.1.

and only then.2(k . Corollary 5. d:+' = d:+' (k . k . This is obviously the best possible whenever H d a (Ek-2/im-z) # 0 . When m and k have opposite parity this top group is zero. where each summand corresponds to a particular choice of the last. Corollary 5. .m . . so that the suspension isomorphism holds in just the one extra degree.1.:.+'/i.+'/im) is isomorphic with . + Theorem 5. + + + + + + In other words.1 > d k P 2 . but then the group in the next lower degree is non-zero. the decomposition of Theorem 5. that the generator of the essential complex for C.m l ) ( m . pair of elements of the patterns.However. k . this isomorphism follows from Theorem 5. for m-2 example. the suspension isomorphism holds in the top 2(k . However. 1 3 1 and n 3 0 the homology group H. Since C. For m 3 3 and n 3 0 the reduced homology group H. In fact. since these homology groups are already known we shall not spell it out.DECOMPOSING THE ESSENTIAL COMPLEXES 89 A fairly obvious modification of this theorem holds for m = 2.(C.1. dm-1 m--2 (k .m ) we have Proof. that corresponding to the pattern sequence (k .m).+' = d.3 when iterated gives the direct sum decomposition stated.2 ) imply that d k P 2 m 1 = dL+' .m)m and dk-2 = the facts that d. For each m 3 3. necessarily nice. Recalling our notation d k for the dimension of and so also of Ek/im.2.m ) degrees.(X.+') may be expressed as the direct sum Proof.. For n > dk+' - 2(k . 2m . for it is then.1. for m = 4 and k even or m = 5 and k odd. is of type (i). is contractible.4. the general necessary and sufficient condition is that m and k have the same parity. .3 when n .2. .-Jim-2 of highest degree. Our next result gives an isomorphism that allows an inductive calculation of the homology of spaces of unoriented shapes.m 2).

-2 @ ffn-4m ( mfl-4 Em-l /h-l. 2 2 ) m+l-2 CB ~ n .-l = 3 and j .@. from the correspondence In the case of G. . Again.+'/i. is generated by the basis elements corresponding to pattern sequences J in 3.-1 gn+/-2 /lm-2) . .2) and for K .-l (CET.. .) is isomorphic with the direct sum ~ n . 13 2 and n 3 0 the reduced homology group . (XE-t. . . The other subcomplexes may be interpreted as before.-l 2 4 there is one pattern in G. 3 4. from the correspondence ( j 1 .. . ( xm-l /1m-i.) - ( j l -2.. .-l > 3. where F . and K . by half the boundary of the former. Hence. = 2 or 3 and j. for each pattern in the essential complex of ELI'. (X.2 . z2) ( m+l-4 Em-2 /lm-2) $.2. is generated by the basis elements corresponding to pattern sequences J with j ...j .. corresponding to the pattern '32'..-2/1m-2) i f 1 is even. When m = 2 the subcomplex K . j. then G.1 ( @ Hn-4../im-2) i f 1 is odd.2 m . . for the re-based .. ffn-lm+m Hn-1.90 THE HOMOLOGY GROUPS OF SHAPE SPACES H4(Ci+'/12) When m = 2 we have S 2 and. = F . The corresponding full decomposition of the essential complex is not quite as simple as in the case of spaces of oriented shapes. for n > 4. j. 2 Proof. in the change of basis. E . the stated result. H n (X. is generated by the remaining basis elements corresponding to pattern sequences J with j. = 2. This gives the stated homology group in degree four.with j . these are subcomplexes. . . For example. z2) @ ffn-h+. Corollary 5 3 For m 3 3. the latter being replaced..-. has just the single generator./lm-l. $ G . Again.. . and the stated homology groups are evident for F .. we may iterate this formula to express the homology in terms of that of spaces of unoriented shapes in R"-' and Rn-2. /i. with j . = 3 and one with j . It follows that we get a Z2-summand for each pattern sequence ( j l . of degree four. This results from a decomposition of the essential complex. @ K.-l .2): that is. = 2. for each pattern sequence J with j.

j ./i3. To state our results we shall require. with that degree increased by 19. those whose patterns end in ‘54’ form the subcomplex that gives the homology of X . those pairs whose patterns are identical. we now give closed formulae for the various homology groups. ‘total ranks’: that is. In addition to computing these ranks for the homology groups in each degree we shall give the.2-m.m . although the latter terminology is not standard. k .m - (m Proof. generate the homology of Cl/12 with coefficients in 772 but with degrees increased by six. for each space. the function P(i. Each of our groups is a direct sum of a number of summands isomorphic with Z and a number of summands isomorphic with 772. and the generator with pattern ‘876’ gives the homology of E.j . Z 2 ) with degrees increased by 12. ( C i . / i l with degrees increased by 13. those generators whose patterns end in ‘32’ form the subcomplex that gives the homology of Ey/il. Theorem 5. m 2 ( m + 4 ) ( m. I) = 1. For consistency we take P ( i . Theorem 4. rather simpler.j . In other words. We shall call these numbers of summands the Z-rank and the Z2-rank. the total number of such summands in the homology groups of all degrees. We look first at homology with coefficients in Zz.1) . but with all degrees increased by seven. 5.CLOSED FORMULAE FOR THE HOMOLOGY GROUPS 91 essential complex of E.l ./il. 1) that is the number of partitions of a set of size i into j parts each of which may be empty but has at most 1 elements. respectively.5. 1) = 0 when i is not a non-negative integer. and the three similar pairs of generators with patterns ending in ‘4’ or ‘5’ produce H*(C:/i2. Of the remaining generators. k . The homology H * ( E i .4 shows that the total number of Z2-summands in all degrees is equal to the number of generators of the essential complex. Z )has total Z2-rank 2 The Z2-rank of H . Z2) is P (“ + 2)(m . except that one ends in ‘2’ and the other in ‘3’.1 ) -P(n2 . we .8 CLOSED FORMULAE FOR THE HOMOLOGY GROUPS Having obtained some feeling for the general nature of the results. which is just Z in degree one. Note that this implies that P(0. in addition to the common combinatorial functions. as well as various relationships between them.

m ways of achieving this. The remaining non-zero entries must be placed along the remaining k . . > j m .1.1 numbers j l > j 2 > . To count the number of these generators in each degree we note that all such patterns contain the top left-hand triangle of m(m 1)/2 non-zero symbolic entries.1 = (rn 2)(m . To correspond to a generator of degree n the remaining n . giving a total Z2-rank of Then.(rn 2)(m . be greater than or equal to two. Z2) is /k-2\ . These are determined by them .2 . Ignoring the requirement that in the resulting pattern the last pair be nice there are + + + + P (n (m - + 2)(m.1 and jm-l 3 2.92 THE HOMOLOGY GROUPS OF SHAPE SPACES need to find the number of patterns for which the final pair is nice: jm-l .1 = j . k 2 1 -m.1)/2.m counter-diagonals of length at most m.2 possibilities from which the m distinct row lengths j i may be chosen. Thus. of the n 1 non-zero entries in the pattern for a cell of dimension n . The Z2-rank of H*(E:/im.1)/2 nonzero entries must be arranged in initial segments of weakly decreasing length along successive counter-diagonals. Those whose last pair is not nice can be seen to be precisely those for which at least one part has size m .l where j l 6 k . Thus.6. > 0. there are of them. accounting for degree m(m 1)/2 . a total of m(m + 1)/2 + m must lie in the upper right-hand triangle and the first counter-diagonal adjacent to it. The number of these is r ( n . we have the following result. The only constraint now on the pattern for a generator of the essential complex is that its final entry j . There are thus k . . + Theorem 5.(m + 4)(m 2 - 1) Looking now at spaces of unoriented shapes we again have a relatively straightforward computation for the homology with &-coefficients.1) .

The Z-rank of Hn(Ek.1 ) 8 2(n-m)-(m+2)(m-l) 8 when m is even and. The total number o summands isomorphic with Z is the number of generf ators of type (i): that is. Turning to the integral homology.2(k . m . when m is odd.2(k .1) . Theorem 5.( m when m is odd.1 . For the number of Z-summands we have the following.CLOSED FORMULAE FOR THE HOMOLOGY GROUPS 93 with m(m 3 ) . We consider the cases m even and m odd.(m + 2 ) ( m .77) is + 2)(m .2 . the corresponding formulae are somewhat more complicated.1 8 2 ’ 2 i f k is odd. of the even rows. each being at least two shorter than its predecessor and all differences being even.m.1 2 ’ 2 2n . ( C k . z for summands in H n ( E k / i m Z ) each n. When m is even the allowable patterns are determined by the lengths: that is. There are . Proof. the number o non-zero entries f with which they start.m 2 + .m ) k 8 - 1. those in whose patterns all pairs are nice and even.7. The homology H . separately.2 ) k .1) . in which [a]denotes the integer part of a. especially those counting the Zz-summands. it is 2n i f k is even and - ( m + 2)(m .k. Z)has total Z-rank i f m is even and (13’) 2n .m m .2-m.

1 8 ’ 2 > 2 if k is even and n . To identifi the degrees in which these summands lie we note that again the contribution ( m 2)(m . fi there are to be any such generators. the remaining n .2 .1.1) 8 2(n . since the first row must be full and the remaining m .( m 8 + 2)(rn .1 ’ 2 ’ 2 if k is odd. for m even. is congruent to k - THE HOMOLOGY GROUPS OF SHAPE SPACES 1 and in which j . odd and the second counting those with j .2(k .1 .2) k .m m . is congruent to k . . This time.j . even.1)/2 to the degree is accounted for automatically by the top left-hand triangle of the pattern.1 . This means. that n .( m 2)(m . When m is odd the number o such generators is f 2n .( m + 2)(m . the first of these partition functions f counting the relevant patterns with j .1) 4 - (k .m ) .1)/2 non-Zero entries have to be arranged in pairs of initial segments of equal even length along the counter-diagonals with the lengths of successive pairs weakly decreasing.94 in which j .1) generators o Z-summands in this degree.(rn + 2)(m . and so of length k . however. When m is odd the first row must be full.1 rows are chosen as before. and the final row must have the same parity.(rn + 2)(m .( m 2)(m .1) . + + + 2n . This allows choices.m ) k . 1 ) is zero when i is not a non-negative integer.2.1)/2 must be a multiple of four and there are then.m m . This expression remains valid for all n on account of our convention that P(i.

(EL.--2.1 = 41 with 0 < 1 < k / 2 . which was computed in Theorem 5.k 2 ) 13 for k = 5 . expressions. 7 .2 for k even and if and only if n .. 1. 3 + 1. 5 . if it exists. The computation of the Z2-rank of H .2) in degree n. Z 2 ) . the total Z-ranks are 1 +0. for the case m = 4. ( C L . where. since the inJinite cyclic summands there do not persist on inclusion in the next spaces to the right. if somewhat unwieldy. This is evident in Table 5. 6 . Of these generators. 1 summands in degrees 9. etc. we get an infinite cyclic summand in H. 3 . 6 + 3. 7 with the distribution of summands determined by the partition function . 2.Z) ij' and only i f n . For m = 3 we have Z-ranks 1. 3 + 3 . Theorem 5. 6 . 9 and so on.3 = 41 with 0 < 1 < (k . + P ( n7. two are required for each Z2-summand. of the rest.l 2 p(n-:-l ) 2. These details may be checked against the data in Table 5. Z)is . of course. for k = 4 . Thus. giving us totals of 1 . Z)are distributed with P (?.21 and 25. for example. and. We look next at the Z2-rank of the homology with integer coeficients. To describe it as concisely as possible we shall use the Z2-rank of H .8. by which we recall we mean the number of summands isomorphic with Z2. 2 .7. The total &-rank of H . ( C . is unique. respectively. since all non-empty subsets in the partition are singletons. (EL.CLOSED FORMULAE FOR THE HOMOLOGY GROUPS 95 Thus.(C:. Z z ) . 8 .6. a known number are required for the Z-summands.2. one generator for each summand. 2 . 17.k . 1 . be substituted in the formulae given here to convert them intofully explicit. which is also the same as the Z 2 rank of H . the partition. and the Z-rank of H .7.3. ( C L .k . .2 for k odd. The total Z 2 rank is easily found since the essential complex has the minimum number o f generators consistent with the Z-homology. Z . Z requires a ) little more effort.. The closed expressions found there can. ) which was computed in Theorem 5. the nine infinite summands in H * ( C i .for k even and p( n-k-3 4 k . 1 + 1 .2. In this case.1)/2 .1 ' 2 - for k odd. 13.

(mod 2 ) but jm-2r-3 > jm-lr-2 1 and j h arbitrary. the type (iii) generators are those for which the pattern J = ( j l . + + + + + + + . . (mod 2) and jm-zr-2 = j . ~ 2 ) ) ( H n 2 (&+I. that is. These are those whose pattern is such that the first pair of entries. and so not nice. together with the right-hand. The constraints on r in the above block decomposition are that r may be zero but 2r must be less than or equal to m . (Hn. j .) generator. j m . As indicated before the statement of the theorem.+(m-2r-I)L=n-l x c rankz. the total Z2-rank o f H . The top left-hand block of this decomposition is full. (X:?2r-l / ~ .Z)is [(m-2)/21 r=O cc 1 nl+n. ..2 ~ . ) Proof. i = 1 . .2 to allow for a non-nice even pair.The fact that 2r 1 is odd here guarantees that. Z is ) so the stated result follows from Theorems 5. . for h < m . those that form the nice even pairs. a pair is even if its right-hand member has the same parity as the last entry in the pattern. are even. . and the right-hand member of each intervening nice pair have the same parity as 1 = j.96 i f m is even and k-2 THE HOMOLOGY GROUPS OF SHAPE SPACES i f m is odd. r with jm-zr = j . ( C k r + IZ .4. In the top right-hand block is a pattern in which the last row has length at least two since jm-zr-3 3 jm-zr-2 2 = 1 2. To calculate how many of these summands lie in each homology group we recall that the &-generators are of type (iii). . j. counting from the right.1 . but not nice. or only. The Z2-rank of H n ( C k . which are not both even and nice. Since. ( C i . We look at the block decomposition of such a pattern determined by the first I = jm-2r-2 columns and by the last 2r 1 rows. subject to the monotonic decreasing rule. pair. for each such . Z) . .7. and the bottom right-hand block is empty. we can interpret the lower left-hand block as representing a generator of a Z-summand of H n . Z ) .5 and 5. Zz). . it determines a Zz-summand of H . j z . all its entries are non-zero. Since there is no other restriction. member of the first even. where y11 1 is the total number of non-zero entries in that block. ( C k . rankz (xkY2r-. Similarly.2r .-zr-2. . for H . all its entries are zero. ./lm-2r-l. that is.) ends with a number of pairs jm-21-1 = jm-2[ 1.l .

2 . being zero for the given choice of r and 1.2r .. .CLOSED FORMULAE FOR THE HOMOLOGY GROUPS 97 This is clear when m is even. also on that of k .l)Z = n + 1. However. when m is even we similarly have generators corresponding to + + + ( m + l . . Although again the total Z-rank and &-rank have simple expressions. if then 2r = m . 1) and ( m + 1 .1.-1/~m-2r-1 and n2 to be at least that of X i : . 3 . 1).. as before. m .. that the ranks. otherwise we shall be counting patterns that do not fit properly into the appropriate blocks. the bottom left-hand block would be the entire non-zero part of the pattern and would represent a generator of a Z-summand. Z)) For the constraints on n1 and n2 we note that there will be n1 1 non-zero entries in the top right-hand block and 112 1 in the lower left-hand block. 3 . Since the top left-hand block has all of its (m . Then we can easily check that Theorem 5. . the integral homology of unoriented shape spaces is rather more complicated than that with Z2-coefficients. As was the case for oriented shape spaces.1 . 2 . 2 . .. .2 to identify such a generator: when m is odd the generator corresponding to the pattern (m 1. . these bounds need not be mentioned explicitly in the statement of the theorem since they will be automatically enforced by the appropriate values of rankz. once again. 1) has a boundary twice that corresponding to (m. then there will be n 1 together.m is necessary to allow a pattern representing a generator of the relevant homology group to fit in the top right-hand block. ( H n l(Xk?2r-1/~m-2r-1. but also holds for m odd since. m . when m is odd. necessarily of order two.2 is specified by k > m 1 and m > 2. the results depend on the parity of m and. We deal first with the Z-ranks. we get + + + + ni + 1 + n2 + 1 + (m . 2. . With s and 1 already chosen. we could use Theorem 4.2r .l)Z entries non-zero.. of the homology group in a particular degree are given by more complex formulae. If the result is to be a generator of degree n .1 and 1. . . m . . m . Once again. so that these rn constraints also do not need to be explicitly mentioned. However.1 > 1. this is automatically accounted for by the fact that such patterns do not represent homology generators. we must also restrict nl to be at least the dimension of ZE1. we find. . .2 .(X~r+l.2 > m . m .. Since the ‘triangular region’ in Tables 1.8 confirms our assertion that the homology groups of such spaces always have torsion. A lower bound 2r for Z is necessary to ensure that a type (i) pattern will fit in the lower left-hand block and an upper bound 2r k . as given in the theorem. m . for any space Xi in that region we have k . .1. In fact. especially the &-rank. 1) and both lie in for k 3 m 2.1. 2. Z2)) or rank~(H.

We can count those of a given degree n as before: of the n 1 non-zero entries in the pattern matrix. a total of m m(m 1 ) / 2 are accounted for in the top left-hand triangle. When m is even that specifies the patterns J that determine infinite cyclic summands.l .) of row lengths in which j . . .Z. the remainder are then shared among pairs of counter-diagonals of equal length. .m(m+ 3 ) 8 summands in H . . We recall from Theorem 4. ( X k / i m .1 ’ 2 summands in degree n Proof. this requires the degree n to be such that the number of these remaining entries is divisible by four and then the number of patterns in such a degree is given by the stated partition function.2 i . j. with 2(n + 1) . Z) has pattern J = ( j l . The total Z-rank of H * ( X i / i m . which may take any decreasing sequence of even values in the range [2.k . . that the generator of an infinite summand of H * ( X k / i . and the definitions that preceded it. When m is odd the unpaired first entry j l must be even and also equal to k ..3. When m is even such a pattern is determined by the m / 2 entries j 2 . j 4 .When m is odd the total Z-rank is zero unless k is ) also odd. with the pair lengths being weakly decreasing. is even and all pairs (j m . . jm-2i-l = jm-2i 1. together with the first counter-diagonal.21.9.98 THE HOMOLOGY GROUPS OF SHAPE SPACES Theorem 5. .1. There are thus + ( ‘TI 1 + such patterns. so k must be odd. with all right-hand members jrn-2i even. . As in the case of oriented shape spaces. . + + . j m P z i )are nice: that is. Then it is with 2(n+3)-m(m+1)-2k 8 k-2-m ’ 2 m. j.Z)is when m is even. .

1 nonzero entries has even length. just Z2 is.6 and 5. Z ) . The remainder of the pattern is then determined exactly as it would be for m .Once again. Z)is L{(k. .2)-( 2 for m even. By Theorems 5.8 that gives the stated expressions.1 and k . the total Z2-rank of all the homology groups of a particular such space is easily found as half of the difference of the total number of generators. I?[ 5 )} for m odd and k even and '{( k .Z)is Proof. The total Z2-runk of H * ( X k / i m . Theorem 5. as before.10. We turn finally to the &-rank of the groups H n ( Z k / i m . The expression for the Zz-rank of H n ( X k / ~ m . Zis./im. as for oriented shape spaces. which is the same as the total Z2-rank of H * ( X k / i m .1.CLOSED FORMULAE FOR THE HOMOLOGY GROUPS 99 When m is odd we need k to be odd also so that the first row with k . Z 2 ) . The &-rank of Hn(Zk. and those required for each ) Z-summand. rather less elegant and more conveniently expressed in terms of other previously computed ranks. This gives the stated results on allowing for the shift in degree caused by the entries in the first row.2 )-(T)} 2 for m and k both odd. The total number of summands of H * ( E ~ / Z ~isomorphic with L) .

1. in the essential chain complexes. for infinite summands. or + + .r (cf. 5.8. d:+' + m is mirrored by a unique generator in the top degree dE+' + ( k . however. 0 < r < (k . . . except for the more restricted . the calculation of the Z2-rank of H. Given a pattern J = ( j l . In this section we show that these isomorphisms can be explained. Tables 5. In particular. by + + + J = (k+ 2 - - j. . and hence also at the cellular level. determining a generator in X:+'/i. . j.1.9 DUALITY IN SHAPE SPACES There may be observed in the tables of homology and cohomology groups of shape spaces de facto isomorphisms that are similar to those induced in manifolds by intersection.8. we shall establish that these isomorphisms are valid for all appropriate shape spaces.' ( k . Then. The only difference is that the lower left-hand block here necessarily determines a generator of a Z-summand in the integral homology of a space of unoriented shapes.9. in this case. at least at the chain level..100 THE HOMOLOGY GROUPS OF SHAPE SPACES Since the characterisation of a type (iii) pattern that determines a 272-summand of H * ( X ~ / L . we further called a pair even or odd according to the parity of its right-hand. > 1 and L~ this is satisfied for 3 if and only if it is satisfied for J . that.).. except for a shift in the degrees involved.. Z) follows from a similar block decomposition of the appropriate pattern matrices to that used in the proof of Theorem 5. and corresponding generator. is d:+' = dE+' ( k .m)m . Then. for EL+' /L. having paired the entries of a pattern J from the right-hand end. there are as many generators of degree dE+' + m + r as there are of degree d". This is explicable as follows./i. Recall that this has one Z2-summand for each generator of the essential complex and that the dimension of or Eh+'/i. > j ..11 and 5.m)m/2..m)m and for each r . 5.(Ck. Z) is the same as for H . we may define a 'dual' pattern. meaning of a pair being even. . where dE+' = (m2 m ) / 2 . every generator of the essential complex determines a generator of the homology. This duality restricts to an involution between the generators of the essential complex. . n ) and. when there remains a left-hand singleton.. . this induces a corresponding duality in homology and cohomology with 272coefficients since. Z). . ( X k .1 Spaces of Unoriented Shapes The most complete example of such 'duality' occurs in the homology of spaces of unoriented shapes with Z2-coefficients.m)m.. and linking for finite summands. since the necessary and sufficient condition for J to determine a generator of the essential complex for C ~ + ' /is that~k 1 > j l > . 5. we called it nice if it were equal to k . We omit the details. In the first instance. not just the ones in which they were observed.k + 2 - j1). in the unique generator of the lowest degree the essential complex for E:+'/i.13). we called a pair nice if it were of the form ( n 1. . Recall.

the generator corresponding to J carries an infinite cyclic summand if all pairs are nice and even. . in the boundary of CJ if and only if j . we have chosen one particular component to represent the entire boundary in such a way that the rightmost pair that is not both nice and even determines the character of the corresponding basis element: that is. is odd and. member. for elements of order two we have made an asymmetric choice..j i . These facts ensure that.ji. since j i P l > ji.. For a generator of an element of infinite order all pairs of integers in the sequence J had to be nice and even and. G. So.+l 1. whenever 8. we see that these are indeed the only cases where the duality exists: when m is even the infinite cyclic summands in the homology of C i + l / ~ r are n isomorphic with those in C z / i m .1 tells us that the basis element ca.so cannot satisfy the appropriate duality.DUALITY IN SHAPE SPACES 101 only. Clearly. for m even when at least the pairs are preserved.ji 1 and k 2 . the former has the same parity as j i . which gives rise to the usual duality isomorphisms in manifolds + + + + + + + - + + + + + + + G. this translates to a different choice of new basis for the essential complex such that it is now the leftmost pair that is not both nice and even.J determines a basis element that has coefficient f 2 in the boundary of the basis element determined by J .1.J be non-zero. Recall that 8. to preserve both niceness and parity of the pairs of integers jm-2. We also note that. which determines the character of the corresponding basis element. the same is true of the element determined by in the boundary of that determined by However. when m is odd there is either no infinite cyclic summand. a sequence with j . > 1. as above. do not fit the above intersection-duality pattern.respectively.12 and 5.10. when k 1 is odd. When a generator in the essential complex has several boundary components.-l in the sequence J . or those that occur. However. which is congruent to k 1 jm-2i-l modulo 2. then the corresponding entries in 8J and are k 2 . provided m is even. in order that ca. whether it carries a generator for the homology or whether its boundary carries twice such a generator.jm-2i. in the above duality the preservation of nice pairs was clear by symmetry. that is precisely the relationship between a homomorphism from one free abelian group to another and the dual homomorphism between the dual groups.j i 1 in differs from it by at least two.1 .14 and remembering that they are cumulative. since the righthand member of the corresponding pair in J is k 2 . with coefficient f 2 .J differs from J only in that j . when k 1 is odd. and needs only. On dualising. the ‘parity’ of the pair is preserved if and only if k 1 is even. However. Scanning Tables 5. is replaced by ji . j i > j. and. Then. For such a basis element c. part (iii) of Corollary 4. and it carries a &summand in cohomology if the first pair that is not both nice and even is odd. the entry k 2 . it carries a ?&-summand in homology if the first pair from the right that is not both nice and even is even. For this ‘suspended intersection’ type of duality to preserve the set of generators of infinite cyclic summands it needs.J appears. jrn-2. niceness of all pairs is preserved if and only if m is even and. 5. which follows k 2 . that is. for the duality between elements of order two we shall work in the essential complex with its original basis. the elements corresponding to the generators determined by a sequence J in J. when k 1 is even.

j . Tables 5.14. . This arises from the fact that to be 'nice' the first singleton must be equal to k and remain . . In the case of even dimensional spaces this involves summands of both finite and infinite orders. the free quotients.+' times.+' a manifold of dimension m(k . if m is odd. so there is perfect duality. duality fails dismally on the torsion subgroups: there is no linking duality.1). j.j l ) . . since j l and j . . we have established that for all m. between all such summands. have opposite parities. Tables 5. However. Thus. or. . It follows that. 5. then the dual determines a generator of degree x + + + + + + Since d. . . . 3-space and 4-space (cf. satisfy all the duality isomorphisms that m)th suspension of would be exhibited by the homology groups of the (d. In particular.2.. + + + 5. Combining the above results.10. except for the shift in the degrees of the groups between which it occurs. this is clearly visible in the homology of spaces of unoriented shapes in 2-space. . we have similar isomorphisms here. there is a suspended form of Hodge Duality for the real cohomology of these spaces. . That the intersection duality holds for all even m follows from the bijection betweenJ = ( j .+' and dim(Ck+') are the extreme degrees in which infinite summands occur.102 THE HOMOLOGY GROUPS OF SHAPE SPACES (see the Appendix). When m = 2 the spaces are manifolds.1. j 2 i = j 2 i . this induces a bijection between infinite summands. Then. also on the parity of k . . since when m is odd and k 1 even there are no infinite summands.l 1 for i = 1. . for all even m there is a suspended form of the intersection duality between the free quotients of the homology groups.12 and 5. In this case. Once again. When m is odd (cf. the full homology groups of E ~ + ' / L . more naturally. when k 1 is even. Then. m/2 and j 2 i = jzn(mod 2) for all i and n . . although in the case of odd dimensional spaces the relevant shape spaces only have torsion groups on their homology. both properties are preserved under the duality. this is indeed a complete duality. as required. respectively). if the pattern J determines a generator of degree dE+' 1 = j1 . but now the amount by which the degree is suspended depends on the parity of k. .9.j. .) and J = ( k 1 . Recall that a generator of an infinite cyclic summand in the homology corresponds to a pattern J in which all pairs are nice and even: that is.2 and 5. .3) there is again an 'intersection' duality between the free summands.2 Spaces of Oriented Shapes In the case of spaces of oriented shapes the observed duality is restricted to the infinite summands and its precise form depends on the parity of m and.m . k 1 . suspended d.

. . .j. . the correspondence between ( k . j 2 . In the case of the spaces Xk. in order for the right-hand members of the pairs in the dual sequence to have the same parity as k .+jm)-1=d~+1+m(k-m)-1=dim(X~')-1 . when k 1 is even.' k . This does indeed give rise to a duality. . . since the least degree in which an infinite summand occurs is d".) and (k.m ) 1 = k j 2 .we wonder what it can mean for the rational cohomology to be the suspension of that of a manifold when the integral cohomology is not. . Then.) and (k. . . . However.j. k 1 . . j.. j . . so too will their duals. should have homology groups consistent with their being suspensions of manifolds seems unlikely to be fortuitous. . which involves all the infinite summands. If the former has degree (d:+' k . ..1.j 2 ) has the required properties: since k is odd.k . j. the correspondence must be between (k.DUALITY IN SHAPE SPACES I 03 so under the duality. suspended d:+' k . . . then the dual has degree + + + + + + mk-(j2+. . as the sequences (k. For example. j 2 is even and so k . k .j 2 is again odd as required. So.m. that all the spaces X:/i. .. any explanation would need to be quite subtle in view of the limited extent to which this is true in the other cases. . When k 1 is odd. . j 2 . j. It is difficult to imagine that the phenomena we have described in this section do not have an underlying topological explanation..) run through all the sequences J that determine generators of infinite summands. + + + + + .j 2 ) . . . So. j 2 . . k 1 . the right-hand member of each succeeding pair must have the same parity as k .1 times.

elaborated in Section 6.3 we revisit the stratification of shape space that we first described in Chapter 2. being as elementary and straightforward to compute as the usual examples but much less trivial. there may be many geodesics between two given points of shape space. On a sphere the geodesics are arcs of great circles. The projection of the pre-shape sphere onto shape space is such that certain geodesics on the pre-shape sphere. As for antipodal points on the sphere. in this chapter we shall concentrate on results that may be expressed in terms of geodesics and we shall give the necessary formal definitions only in the context of shape spaces where they are usually much simpler and more transparent than in the general case.5 and 6. those cells are not unique and may be moved around the space. We now begin the study of the geometric structure of shape spaces. although we may. to join arbitrary pairs of points in shape space. a certain homology group is a global and not a local invariant we mean that. gives us a firm grasp of the geodesics in shape space and enables us to achieve the following results. G. is the analogue in a curved space of a straight line in flat Euclidean space. Indeed.2. In Sections 6. However. BARDEN & T. which was. LE Copyright 0 1999 by John Wiley & Sons Ltd CHAPTER 6 Geodesics in Shape Spaces In studying the topology of shape spaces we discovered the extent to which shape spaces differ ‘globally’ from familiar Euclidean spaces. only between two points of the upper two strata. which provides some measure of the locd differences between shape spaces and more familiar standard spaces. This. for example. which is already a fairly sophisticated concept. then. For instance. project isometrically onto geodesics on the shape space. C A R " & H.2. The full range of geometric properties is determined by a Riemannian metric. KENDALL & D. The successive strata fit together so smoothly that it is possible to extend the definition of shape space geodesics from Section 6.6 we . By saying that. shape spaces provide an excellent domain in which to illustrate these basic geometric structures. A geodesic. the ‘horizontal’ ones. and did. any circle S’ x {x} in the torus S’ x S’ carries the same generator of the first homology group. in fact.Shape and Shape Theory D. K. In Section 6. identify cells that determine it. and our account will require further technical ideas.

which we think of as lying vertically above n ( X ) . also open.1. This fibre. which fixes the subspace spanned by the first j rows. without loss of generality. We may also observe at this point that. the quotient map n is both continuous and. 6.= SO(m)/SO(m. an important subgroup of SO(m) is the isotropy subgroup of SO (m) at X or. since SO(m) acts as a group of homeomorphisms. Let Dj denote the subset of S i comprising all pre-shapes whose representing matrices have rank less than or equal to j .7 we identify it for each shape point. which is 2 = {T E SO(m) : T X = X } . .8 we compute the nearest degenerate shape to a given one and the distance between these two shapes.1) pre-shape matrix X E M ( m . isomorphic with SO(m . the stabiliser of X . An important application of these shape spaces is in the evaluation of the statistical significance of degeneracies and in Section 6..1) denotes the space of m x ( k . the expressions that arise lend themselves well to computation.l). Then T X = X implies that T lies in the subgroup. k . consists of all the pre-shapes that have the same shape as X . which is the orbit of X under the action of SO(m). as we move out along each geodesic we arrive at a point immediately beyond which it fails to be minimal. is the quotient space SO(m)/Tx and. It plays an important role in stochastic analysis and in Section 6. in Rm is via the sized-and-centred m x ( k . The locus of such points along various geodesics from the fixed point is called its cut locus. where M ( m . The isotropy subgroup at each point in the orbit of X is conjugate in SO(m) to 1 and hence is isomorphic with it. as it is also known. the fibre of n above n ( X ) is homeomorphic with SO(m). is also referred to as the Jibre of n over n ( X ) . Because of its size-normalisation. and with the Stiefel manifold Vm. X. For each pre-shape X . n-'(n(X)) homeomorphic with .' .1 THE ACTION OF SO(m) ON THE PRE-SHAPE SPHERE We recall from Chapter 1 that our standard procedure for representing by a matrix the shape of a configuration of k labelled points.j rows zero.j ) . if j 3 m . The space E is then the quotient space of Sk by SO(m) and. we assume that X has rank j and. As for the distance between shapes. to identify this quotient.1. also that its representing matrix has its last m .106 GEODESICS IN SHAPE SPACES find the full set of geodesics of minimal length between any two shapes and also examine when that set has more than one element. so that n(Dj)consists of the shapes . For any given X in the pre-shape sphere the subset n-'(n(X)) = {TX : T E SO(m)}. Moreover. is Hausdorff. which are not all identical. Starting from a fixed shape point. if it is given the : quotient topology. since these fibres are compact.1) real-valued matrices. the degree of degeneracy being measured by the stratum in which it lies. Therefore. k .j ) of orthonormal j frames in R"'. if j < m . X is a point on the unit pre-shape sphere S i = S m ( k p l ) .

1) : tr(X2‘) = 0) of M ( m . an example of a Riemannian submersion. Under the above assumption that k 3 m 1. is empty. no element except the identity has any fixed points. the curves that may be characterised as takmg the shortest path between any two sufficiently close points upon them. then the tangent vectors to the sphere at that point correspond to the vectors orthogonal to X . the non-singular part of shape space also inherits a unique Riemannian metric that is compatible with its differential structure and with respect to which the quotient map n is particularly well-behaved. if X is the ‘position vector’ of a point on the sphere.2 VIEWING THE INDUCED RIEMANNIAN METRIC THROUGH HORIZONTAL GEODESICS A Riemannian metric on a differential manifold is a smooth choice of inner product on its various tangent spaces.VIEWING THE INDUCED RIEMANNIAN METRIC 107 of those configurations of k labelled points for which there is a j-dimensional affine hyperplane in Rm that contains all k points. Thus.1) matrix. and also to make a modest extension of the general results in our context.) as the singularity set and its complement as the non-singular part of shape space. in fact. and basic theorems from differential geometry tell us that the corresponding part of the shape space inherits a differential structure. Since SO(m) acts as isometries of the sphere. in due course. Outside Dm-z the action of SO(m)is free.which is. A Riemannian metric on a manifold enables us to define geodesics. since otherwise we would need continually to make exceptions and special statements for the other cases. the singularity set is non-empty and is an important feature of shape space. rather than quote the general theorems. segments of straight lines and on a . Since all matrices in the pre-shape sphere have rank at least one.k . We shall conversely start by defining the geodesics and use them to define and explore the Riemannian structure of shape spaces. It is. In Euclidean space geodesics are. the singular part of Ck. of course. with the coordinates arranged in an m x ( k . In particular. We therefore refer to n(Dm-. as usual for the latter. + + 6. we shall.k . thought of as naturally embedded as the unit sphere in this is just the usual inner product between vectors in a Euclidean space. we are able to give a direct proof of all the properties we require.) the to pre-shape sphere at X is the subspace (2 E M ( m . thought of as being ‘parallel translated’ to act at X . In the case of the pre-shape sphere. a concept introduced and studied by O’Neill (1966. Before proceeding further we recall from Chapter 1 that it is sufficient to consider the case k 3 m 1 and we shall indeed make that assumption for the remainder of this chapter. in our notation.1) Z Rm(k-’). compatible with its topology. However. it is sufficiently elementary that. return to the case k < m to flesh out the reasons that it is already covered by the general case. Dm-2 is the subset of S: of points at which the isotropy subgroup is non-trivial. However. However. 1967). from that on the sphere. the tangent space Ix(Sk. that is. for all m > 2.

a unique such path-length parameterisation and we shall always assume that is the parameterisation chosen. between X and rZ(s). We note that it restricts to a diffeomorphism of {Z E Ix(S”. This is indeed the origin of the term ‘exponential’ map for a general manifold and it is easily checked at least that. . we see that there is a natural bijection between directed geodesics through X and tangent vectors of unit length at X.1) joining them. = = z. there is the geodesic 0 < s < n. since the space of m x m skew-symmetric matrices has the same dimension. is a curve in SO(m) starting at I so that its tangent vector at s = 0. The map %s) A. Any geodesic has. the length of the shortest great circular arc. or orbit. For a point X on the pre-shape sphere and any other point 2 orthogonal to it.1)/2. We refer to the subspace of tangent vectors . For exp(sA) E SO(m) ++ exp(sA) exp(sA)t = 1 *exp(s(A+A‘))=ZtiA+A‘=O. for each direction. when 0 < s < n. for suitable matrices A. except for the point antipodal to X . When SO(m) is similarly regarded as a Riemannian submanifold of R”’. (6. Thus.) : llZll < n} onto the whole of the sphere. where exp(sA) = Z + SA + -A2 + . it is the distance travelled along Tz from r Z ( 0 ) = X to rz(s) and so. the exponential map at the identity is FA($)= exp(sA). The parameter s here is so chosen that. when it is positive. it follows that i/A(s) = exp(sA)X lies in the fibre. through X . m(m . z * ~Z/llZl~(IlZII) ds ls=o is tangent to SO(m) at the identity. S2 - s . which converges for all s and A. is known as the exponential map.1) at X is rz(s) Xcoss + Z sins. arcs of great circles. it is the entire tangent space to SO(m) at 1. as SO(m). any such skew-symmetric matrix A represents a vector tangent to SO(m) at Z and. Since the geodesic exp(sA) lies in SO(m) whenever At = -A. 2! is the obvious series of powers of m x m matrices. . it is the geodesic distance. Since the tangent vector to the geodesic (6.108 GEODESICS IN SHAPE SPACES sphere.

VIEWING THE INDUCED RIEMANNIAN METRIC

109

to such curves at X as the vertical tangent subspace at X , and to its orthogonal complement in %(S;) as the horizontal tangent subspace of '&(S;). Thus, the vertical subspace at X is given by Vx = {AX : A' = -A}. When n(X) is a non-singular point, that is, X @ Dm-2, this vertical subspace is isomorphic with the tangent space to SO(m) at I . However, at a singular point the vectors A that are tangent to the isotropy subgroup give rise to the zero tangent vector AX since the corresponding curves initially remain stationary at X. Thus, at points X of Dj \ 23-1, V X ,specified as above, is indeed isomorphic to the tangent space to V m , jat the frame corresponding to the first j rows of Z. It follows that the horizontal subspace at X is
= { Z E M ( m , k - 1) : tr(XZ') = 0 and

tr(AXZ') = 0 VA s.t. A' = -A}

= (2 E M ( m , k - 1) : tr(X2') = 0 and XZ' = ZX'}.

Again, the specification is independent of the rank of X. When X has rank less than m - 1, so that dim(Vx) < dim(SO(m)), then the space of matrices Z that are tangent to S; and such that XZ' = ZXt is correspondingly larger. We are now in a position to make two crucial observations.

Proposition 6.1. (i) I f a geodesic in S; starts out in a horizontal direction, then its tangent vectors remain horizontal throughout its length. (ii) I f r z (s) is the distance-parameterisedgeodesic through X with initial direction Z, then, for any T in SO(m), TrZ(s) is the distance-parameterised geodesic rTz(s) through TX with initial tangent vector TZ. Moreover, rz(s) i horizontal s ifand only ifTTz(s) is horizontal. Proof. (i) The geodesic rz(s) is horizontal at s = 0 if and only if XZ' = ZX'. Then, for all s,

That is, each tangent vector %(s) is horizontal at rz(s). (ii) That TrZ(s) = TX coss TZ sins is a distance-parameterised geodesic through TX follows from Equation (6.1) since tr{TX(TZ)'} = tr(TXZtTt) = tr(XZ') = 0. Then, rz(s) is horizontal if and only if XZ' = ZX' and so if and only if TX(TZ)' = TXZ'T' = TZX'T' = TZ(TX)',

+

which is if and only if T r Z ( s ) is horizontal.

110

GEODESICS IN SHAPE SPACES

It follows that exp lRx maps the subset of vectors of length less than n onto a submanifold WX of SJ all of whose tangent vectors at X are horizontal. Now, if X lies outside Dm-2, then so do all sufficiently close points and, since the tangent spaces to the fibre and to the WX at X are orthogonal, there is a neighbourhood U of X in WX such that, for all Y in UX, tangent spaces to the fibre and x the to WX at Y remain transverse. Thus, the fibre at Y meets UX only at Y and we have established the following picture. Given X outside Dm-2, through each point TX of the fibre at X there is a submanifold U T X ,traced out by local horizontal geodesics through T X , such that
X (i) the submanifolds UT,Xand U T ~ are disjoint if T I# T2, (ii) each submanifold UTX is mapped by the quotient map bijectively, and hence homeomorphically with respect to the quotient topology, onto a neighbourhood of n ( X ) in Z i ,

~ (iii) the action of each S E SO(m) restricts to a diffeomorphism of U T onto UsTX that also preserves the Riemannian metric, that is, it maps geodesics to geodesics of the same length and its derivative maps horizontal tangent vectors at TX to horizontal tangent vectors of the same length at STX. It follows, first, that we may use UXand nlux to determine a differential structure on the non-singular part of shape space Xi \ n(Dm-2), since for any other choice ( U T X nlu,,) the composition ( j ~ / u ~ ~ ) - ' o ( is Ijust )the diffeomorphism TI,. , n u~ We may also define a Riemannian structure there by defining the inner product of two tangent vectors at n(X) to be the inner product of the corresponding horizontal vectors at X . Again, by (iii) above, it does not matter at which point in the fibre n-'(n(X)) we do this. This Riemannian metric on the non-singular part of shape space will be investigated in more detail in Chapter 7. The way we have defined the induced metric means, of course, that the projection n maps the horizontal subspace of the tangent space to the pre-shape sphere at X isometrically onto the tangent space to the shape space at n(X),making it a so-called Riemannian submersion. It is usual to deduce from this that geodesics that start out in a horizontal direction remain horizontal and that they are mapped isometrically onto geodesics in the quotient manifold. However, in our case, we have already proved the former property and used it to derive the Riemannian submersion itself, and the definition ensures that horizontal geodesics are mapped onto curves of the same length. That they are also geodesics follows from the fact that any curve in shape space may be lifted to one of the same length in Sk whose tangents are everywhere horizontal. If, instead of SO(m), we consider the action of O(m) on the pre-shape sphere, similar considerations to the above apply outside D,,-l, which is the set of points with non-trivial isotropy subgroup in O(m). However, as O(m) = SO(m)USO(m)T, where T is any matrix in O(m) of determinant -1, each O(m)-orbit may break up into two SO(m)-orbits and Sk/O(m) = Zk/im, where 1, maps each shape to the shape of the reflected configuration. Since 1, is induced by multiplication by any T E O(m) of determinant -1 on the pre-shape sphere,

THE SINGULAR POINTS AND THE NESTING PRINCIPLE

111

which is an isometry mapping fibres to fibres and hence horizontal subspaces to horizontal subspaces, this confirms that 1, is an isometric involution of Xk and so induces a metric on the quotient space C k / i m of unoriented shapes. The O(m) and SO(m) orbits of X coincide if and only if X lies in Vm-l, since there we may find T E SO(m) such that n ( T X ) is the reflection of n(X) and this is, of course, impossible if X has rank m.Hence 1, leaves fixed all points of n(D m -l ) but no others.

6.3 THE SINGULAR POINTS AND THE NESTING PRINCIPLE
To understand the singular part of shape space we observe first that each Vj, 1 < j < m - 1, is a submanifold of Sk. The following elementary proof of this fact is due to Milnor. Without loss of generality, we may write any matrix X E S; of rank j in block form as

(: $);
where A is j x j and non-singular. Then, left multiplication by

(-&' :)
shows that the rank of X is j if and only if D = CA-'B. So the map is a diffeomorphism of a neighbourhood

of

(: :)
onto a neighbourhood of X in M ( m ,k - l), taking matrices with D = 0 onto 1 matrices of rank j . This is precisely the form of chart that is needed to show that the latter matrices form a submanifold of dimension

Since rank(X) = rank(hX) for h # 0, it follows that Sk meets this submanifold transversely in a submanifold Vj of one lower dimension. We also have a natural embedding & of the pre-shape sphere S; in Sk: that f is, a diffeomorphism with a submanifold &Sf, mapping a matrix XI E S to the

112

GEODESICS IN SHAPE SPACES

same matrix extended by m - j rows of zeros,

x=

) ( ;

€a;.

Then, €Sf is clearly the submanifold of V j given in the above chart by C1 = D1 = 0. We now consider the action of SO(m) on these submanifolds. The whole of SO(m) maps V j into itself and, since each SO(m) orbit in D j contains points of &Sf, it follows that n ( D j ) = n(&Sf). However, the subgroup mapping I S f into itself is isomorphic with O ( j ) and acts precisely as it does on SJ” itself, and so n(&Sf)is isometric with Cg/ij. Thus, the nested sequence of submanifolds Dm 3
Dm-l 3

. . ’ 3 Dj

3

... 3 Dl

gives rise to the nested sequence of subspaces in Ck, or stratification,
k C 3 Cm-l/Lm-l 3 . . . 3 :
q / L j

3

. .. 3

q/11,

which we first met in Chapter 2. Each stratum, the difference between two consecutive subspaces ( C g / i j ) \ ( C5-1 / i j - 1 ) , has a differential and Riemannian structure induced on it in the same manner as for Xk \ n(DmP2), with the restriction of the projection n to Dj \ Dj-1 being a Riemannian submersion. This is a convenient point to look at the special case k 6 m that we have been excluding. Since the pre-shape matrices are m x ( k - 1 ) and k < m + 1, we now have S i = 2)k-1 and = n(Q-1) = n(&S,k_,),which, as above, is isometric with Ci-l/ik-l. Thus, the special case is indeed covered by knowledge of the which lie on the diagonal in Table 1.1, and of the isometric involutions
lk-1.

Returning to the general case k 3 m 1, we have already seen that the first two strata together form a non-singular manifold and, in fact, the second stratum ( C ; - , / L ~ - I\) ( C k - 2 / ~ m - 2 is a submanifold of this since SO(m) acts freely on ) the invariant submanifold Dm-l \ Dm-2 of SL. This is the only case where adjacent strata combine in this way. Nevertheless, consecutive strata are embedded in each other in a regular manner, which, like the metric structure itself, is best explored through the geodesics.

+

Proposition 6.2. All geodesics through points of Dj that start out in horizontal directions tangent to Dj remain in D,. Those that start out in directions tangent to D, \ Dj-1 will only meet the submanifolds D for i < j in isolated points. i Proof. The horizontal subspace at

x=

) ( ;

THE SINGULAR POINTS AND THE NESTING PRINCIPLE

113

for the action of the subgroup, isomorphic with O(j ) , mapping ES; into itself is
: tr(X1Z;) = 0

and X1Z: = ZlXi

Since n ( D j ) = n(&Sj), this must also be the full horizontal subspace at X for the action of SO(m) on D,. It follows that a horizontal direction Z tangent to Dj at X E &S; is, in fact, a horizontal vector

Z=

(2)

tangent to ES;. Thus, the horizontal geodesic rz through X is the image &rz, in & S j of the horizontal geodesic rz,through X 1 in Sf. In particular, rz remains in Dj. However, the general point of Dj is TX for some T E SO(m) and some X E &S;, and a horizontal direction tangent to Dj at TX is TZ for some Z, which is horizontal and tangent to ES; at X. Then, the corresponding geodesic r T Z at

TX is TrZ by Proposition 6.1, and this lies in Dj since rz does. As horizontal geodesics are reversible, if T Z is tangent to Dj \ Dj-l and r T Z meets Diwith i < j , then it must do so transversely and hence in isolated points, since if it were tangent to D iit would have to remain in it and hence be tangent to D; throughout its length and, in particular, at TX.
Note that, if j < m, there will always be horizontal geodesics meeting V j transversely, since the fibre through X in D, is the full fibre through X in S:, and so the horizontal subspace at X for the action of SO(m) on S: must have a further direct summand: namely, the orthogonal complement to the subspace tangent to Dj. Now, horizontality is definable everywhere in S i and a geodesic in each stratum is the image of a horizontal geodesic. Thus, we may extend the usual definition of a geodesic in a Riemannian manifold to shape spaces as follows.

Definition 6.1. A geodesic in Ek is the image of any horizontal geodesic in S i .
This concept of a geodesic is well defined by Proposition 6.1, and it applies whether or not the image remains in a single stratum. It is a geodesic in the usual sense, except at isolated points where it meets lower strata, and even there it is locally a limit of geodesics in an obvious manner. This definition makes each subspace C?/ij totally geodesic in each higher one in the usual sense that a geodesic in one subspace is still a geodesic when it is regarded as a subset of a larger one. In particular, \ n(Dm-2))/~m-1 a totally geodesic submanifold is ; of z \ n(Dm-2).

114

GEODESICS IN SHAPE SPACES

6.4 THE DISTANCE BETWEEN SHAPES A Riemannian metric on a manifold determines a distance function that is the minimum length of the geodesics between two points and this definition extends naturally to the whole of shape space using our extended definition of geodesics. Since a geodesic between two shapes n(X)and n(Y) is the image of a horizontal geodesic r from X to some point TY in the fibre over n(Y) and since r meets the fibres through X and TY orthogonally at those points, it follows that the induced distance between any two shapes x ( X ) and n(Y) is given by
p ( n ( X ) , n(Y)) =

rr&~,)d(X, TY) = arccos TtSO(m) tr(TYX'), max

(6.2)

where d is the shortest great circular arc distance on the pre-shape sphere. This is precisely the formula (1.3) already given in Chapter 1. Another commonly used distance ; the shape space, sometimes called the on procrustean distance, is the one induced from the chordal distance 2 on the preshape sphere. Here, 2 is the distance measured directly in the Euclidean space in which the pre-shape sphere lies, so that

&X, Y) = IIX - YIJ
and

Although the distance j3 does not give the Riemannian metric on the shape space, these two induced distances p and 6 are topologically equivalent and there is even a simple functional relation between them. This relation is obtained by noting that 2 ( ~Y)' = tr{(x - Y ) ( X - Y)') = 2 - 2 t r ( x ~ ' ) ) , , which leads to

= 211 - cos(p(n(X>, n(Y)>>).

(6.3)

To simplify the expression on the right-hand side of (6.2) we shall use pseudosingular values decompositions of the m x m square matrix YX'. We recall, from Chapter 1, that this is the expression YX' = U A V , where U and V are in SO(m) and A = diag(h1, . . . , h,} with the diagonal elements hi satisfying h , 3 . . . 3 A- 3 lhml and sign(h,) = sign(det(YX')). Then, the hi are called the pseudo,, singular values of YX', or of XY'. Two points should be emphasised. Firstly, we should distinguish the above from the analogous, more familiar, singular values decomposition for which the matrices U and V may be chosen in O ( m ) rather than SO(m). Consequently, for uniqueness, we require h, 3 0 and we now call the hi singular values. We note that, for both singular and

where R = UV and Q = V'AV. the rows of V are unit length eigenvectors of X Y ' Y X ' and the columns of U are unit length eigenvectors of Y X ' X Y ' . although the matrices U and V are not unique. = 0 if and only if at least one of n ( X ) and n ( Y ) is in n(Dm-l). However. If at least one of X and Y is in Dm-l. We also note that. if we denote by y the matrix obtained from Y by changing the signs of the entries of the last row so that n(7)is the reflection i m ( n ( Y ) ) the of shape n ( Y ) . we refer to such a decomposition as a pseudo-polar decomposition. Then. Thus. it should be noted that we use the pseudo-singular values of the square matrix Y X ' to compute the distance between the shapes n ( X ) and n ( Y ) and later we shall use the full decomposition to describe the geodesics between these shapes. In general. introduced in Chapter 1. Again by analogy with the classical case. when T = (UV)'. the corresponding pre-shapes are square matrices and hence h. that is. when k = m 1. R is in SO(m)and Q is symmetric and. Kendall. perhaps different.. so Q is positive semi-definite and we have a polar decomposition. since Q = V f V is a pseudo-singular A values decomposition. In particular. = 0. these rows and columns being arranged in the same order as the corresponding eigenvalues A?.1) matrix X as U ( A 0 ) V. then Y X ' and r X t have the same pseudo-singular values save that the last one has a different sign if it is not zero. if k > m and neither of X and Y is in Dm-. + m T&O(m) max tr(TYX') = max tr(VTUA) = c h i 3 0. then Q is not necessarily positive semi-definite. psuedo-singular values decompositions of the m x ( k .8 to investigate the projections onto lower strata and also in Lemma 7. will be used in Section 6. The matrix A also remains unaltered if X and Y are each pre-multiplied by. A pseudo-singular values decomposition of Y X ' gives rise to an alternative decomposition Y X ' = RQ. for the pseudo-singular values decomposition. Now. rotations in SO(m).THE DISTANCE BETWEEN SHAPES 115 pseudo-singular values. then we have h. 1984). T&O(m) i= 1 the supremum being attained when V T U = I (cf. so it depends only on the pair of shapes n ( X ) and n(Y). there is not a simple relationship between the pseudo-singular values decomposition of Y X ' and those of the individual pre-shapes X and Y themselves. the sign of det(YX'). its pseudo-singular values coincide with its eigenvalues. and hence we have . Secondly. their squares are the eigenvalues of X Y ' Y X ' .2 to obtain local coordinates on shape space in the neighbourhood of the shape n ( X ) . the diagonal matrix A is uniquely determined by X Y ' Y X ' and. Whereas.

so that such pairs always exist. Y)T'. T2Y) = T 2 R ( X . except that the summands hi are now the singular values. we replace SO(m) by O ( m ) and pseudo-singular values by singular values. if T I . there are natural bijections between R(TIX. R ( X . Y ) since R(TIX. In particular. we can always find a non-zero Y such that YX' = 0. which means that if p(n(X). the existence of maximally remote pairs n(X) and n(Y) is equivalent to the condition that the pseudo-singular values of YX' satisfy hl = -hZ. if X lies in V.). throughout the above. then both n(X) and n(Y)must lie in n(Dm-l). and the diameter of shape space is n/2. / i m is the minimum of the distances p(n(X). f (iii) I f k > 2m. YX' = 0 means that the row space of Y is orthogonal to the row space of X. : R E R(X. That is. Then. n(Y)) Zk. Y) to be the set of all R E SO(m) such that Q = R'YX' is symmetric and its pseudo-singular values coincide with its eigenvalues. Note that this statement is false when m = 2. Y)} and. (ii) If m 1 < k < 2m and n(X) and n(Y)are maximally remote. we deduce that the diameter of shape space is bounded above by n/2. to YX' = 0. n(Y)) = n / 2 we must have xyZl rank(X) + rank(Y) < k - 1.5 THE SET OF GEODESICS BETWEEN TWO SHAPES To study the geodesics in shape space between any two given shapes we define R(X. Hence. Y). n(Y)) and p(n(X). R(X. Y). and again n(X) and n(Y) will be maximally remote. Y ) is the set of all rotations R that can occur in a pseudo-polar decomposition of YX'.116 GEODESICS IN SHAPE SPACES Thus.-. Then. we obtain analogous results for C k / i m . in Since 0 < Cy!l h. Y) = R(X. n More generally. T2Y) and R ( X . as each of rank(X) and rank(Y) is not less than one.X) = (R' : R E R(X. = (RT. + + 6. (i) If k = m 1 > 3 and n(X) and n(Y) are maximally remote. Theorem 6. For m = 2. . maximally remote pairs n(X) and n(Y) exist i all cases.T2 E SO(m).1. we obtain the following result. then there exist maximally remote n(X)and n(Y) with neither ofn(X) and n(Y) in n(Dm-l). and we shall call shapes that are n/2 apart maximally remote. hi = 0 is equivalent to A = 0: that is. Thus. it is clear that the calculation required to find the Riemannian distance between any two shapes is elementary and simple to program. then at least one o n(X) and n(Y) must lie in n(D..4). the distance has the same expression (6. If. we see that the distance between n(X) and n(Y) in X . Y ) } and R(X. < 1. However. When m 3 3. R(Y.-I.

So. R'Y) = p ( n ( X ) . hi = hi for all 1 < i < m. where 0 < SO < n/2 assuming n ( X ) # n ( Y ) .2. = -A. we have. and (ii) the non-uniqueness of Q in the pseudo-polar decomposition YX' = RQ when A. Choosing Z E I x ( S . Proof. Y ) fails to be unique will be confirmed later. in fact. Y ) if and only if the complex number corresponding to R'YX' is. d ( X . ) to be Z=- 1 {R'Y -XCOSSO). Proposition 6. Y ) we know that R'YX' is symmetric with trace Er!l hi and that p ( n ( X ) .3.n ( Y ) ) is SO = arccos(tr(R'YX')). hi. R'Y). Suppose that R E SO(m) and that R'YX' is symmetric with eigenvalues hi. The claim that these features create a situation in which R E R ( X . I xy=l It follows from Proposition 6. Y ) symmetric and p ( n ( X ) . For R E R ( X . if and only if R'YX' is Proof.-* < 0.1)-dimensional complex vectors. then the corresponding YX' will be a complex number. real and equal to the norm of the complex number corresponding to YX'. sin SO . if we use ( k . R'Y) = arccosxr!.THE SET OF GEODESICS BETWEEN TWO SHAPES 117 In the case that m = 2. Y ) to characterise the horizontal lifts of minimal geodesics between n ( X ) and n Y ) ( as follows. after relabelling to make {hf : 1 < i < m } non-increasing. Theorem 6. and eie = R E SO(2) is in R ( X . such a point R'Y on the fibre will not always be unique. However.1) matrices. the equivalence between the two statements is established. d ( X . that each hi is f l h i l . We first use the set R(X. Since the pseudo-singular values of R'YX' are the same as those hi of YX'. ( Y ) ) = SO with 0 < SO < n/2. Let R E SO(m).3 that the point R'Y for R E R ( X . Then R E R(X. However. instead of 2 x ( k . Y ) . Then.2. which implies that the great circle through X and R'Y on the pre-shape sphere is orthogonal to the fibre above n ( Y ) . to represent the pre-shapes and identify SO(2) with S1. This lack of uniqueness is associated with (i) the fact that U and V can be variedindependently in the decomposition YX' = UAV when rank(YX') < m . Y ) is a point on the fibre above n ( Y ) whose great circular distance from X is the minimum possible for all points on the fibre above n ( Y ) . {h. n ( Y ) ) = d ( X . n ( Y ) ) if and only if hi = Cy=l and it easily follows that this holds if and only if hi. : 1 < i < m } is the set of eigenvalues of XY'YX'. So. then a curve in the pren shape sphere is a horizontal l$t starting at X of a minimal geodesic from n ( X ) to n ( Y ) if and only if it can be expressed in the form with R E R(X. r f p(n(X).

3 we showed that geodesics meet the lower strata of Zk in isolated points. using their identification in Theorem 6. we must have R E R(X.-2). if p(n(X). It follows that. ~ ( S O )= R'Y and by Proposition 6. : R E W X . However. SO)) = SO. positive semi-definite and has the same rank as X and that. which can only occur if rank(X) = rank(Y) = m and is equivalent to . for such s. 3 0.2. n(Y)) and R(X.3. Proof. when h. together with Proposition 6. Nevertheless. i .s) v'xxtv > 0.3. The relation between G(n(X). n(Y)) is not necessarily in one-to-one correspondence with R(X. RtYXt is also positive semi-definite. 3 0 and R E R(X. Conversely. S O ] . the curves y(s) corresponding to two different R in R(X._l \ D. Y ) such that the corresponding curves y(s) give the same great circular arcs in the pre-shape sphere. n o y ~ minimal . Note. rank(Y)}. In other words.1) for s E [0. G(n(X). Theorem 6. sin SO provided 0 < s < SO. since d(y(O). When h. Thus.S O ] ) is a horizontal lift starting at X of a minimal geodesic from n(X) to n(Y). n(Y)) > p ( n ( X ) .2. rank(yR(s)) 3 rank(yR(s)X') 3 rank(XX') = rank(X). there may exist R1 # R2 E R(X. is since its length is SO. that G(n(X). SO] is y ~ Then. Then.s]) remains a minimal geodesic for s > so ifand only if the eigenvalues of y(s)X' are the same as its pseudo-singular values. Corollary 6.118 GEODESICS IN SHAPE SPACES we have tr(XZ') = 0 and XZ' = ZX' so that Z determines a horizontal great circle through X of which the restriction to [0.1. all minimal geodesics in Ek arise in this way. Y) either coincide for all s such that 0 < s < SO or never meet for any such s.XCOSSO)/ sinso as required.3. If y([O. In Section 6. Whence Z = (R'Y . A minimal geodesic between n(X) and n(Y) in Zk lies entirely in a single stratum except possibly for its end points being in lower strata and. y will be some Tz as in (6. we have V"R(S)X'V 3 sin(s0 . however. n(Y)). then noy([O. if h. rank(yR(s)) 3 max(rank(X). Y)l sol) gives all possible minimal geodesics between n(X) and n(Y) in Zk. where SO = p(n(X). Y). The following result follows immediately from Theorem 6. For minimal geodesics we can say more. if v is a non-zero eigenvector corresponding to a non-zero eigenvalue of XX'. Y). < 0. at most one point ofn(D. Y) will be investigated later. For suppose a minimal geodesic from n(X)to n Y) lifts to a horizontal great circular arc y starting ( at X. n(Y)) = { n o y ~ ( [ O . Thus. the distance between n(X) and n(Y). ( n ( Y ) ) ) . We observe that XX' is symmetric. Y). So.

y ~ ( [ 0 . with X’ and Y’ of rank rn.-l for 0 < s < so and the Note that. ) proof of the theorem is complete. Hence. n(Y)) > p(n(X).(n(Y)) = n ( Y ) .-l) and a point n ( y ~ ( s 2 )in n(Dm_l) meets n(D. i. n(i. if y ~ ( s 1 )and y ~ ( s 2 )were in D. if ‘u is an eigenvector of P Y X t with eigenvalue h.\ Di-1 arbitrarily close to X. However. there must be points X’ = y ~ ( s ’ ) D. and only meets the lower strata in so]) 3 in isolated points.Then.2 that it can only do so in isolated points.. when k = rn 1..-l) again ) that contrary to what we have just proved. then.THE SET OF GEODESICS BETWEEN TWO SHAPES 119 the condition p(n(X). if p(n(X). then n o y ~L ~ . where (Y > 0.. If j = r > max{rank(X). rank(Y)} the results stated in the theorem are already established. in this case.2. i.(n(Y’))) for such X’ and Y’ sufficiently close to X and Y. by the above.1 on which R‘YX‘ is positive definite.would be a minimal geodesic between a point I ~~] n(X)not in n(D.-l. and lying between X and Y along Y R . then V‘ y~(s)X‘v= 0 when tans = tan so 1 . n(Y)) > p(n(X). However. respectively. XX‘ is positive definite and there is a subspace of dimension rn . and hence precisely once..(Y))). n(Y’)) > p(n(X’). Since the restriction of a minimal geodesic is necessarily still a minimal geodesic. a similar argument to the above shows that YR can only meet D. For. at n ( y ~ ( s l ) ) . ) ) except possibly when j = rn. by Proposition 6. rank(Y)}. i. lies in 2 . y ~ ( s E D. Suppose that the tangent vector to the curve yR at s = 0 is tangent to 2 . we must have p(n(X’). where 0 < s1 < s2 < so. and we already know from Proposition 6. and both n(X’) and n(Y’) lie along a minimal geodesic from n(X) to n ( Y ) . When j = rn = max{rank(X). the above result shows that y ~ ( s E 2 . i is an isometry.(n(Y))) stated in the theorem. then YR must meet D. we can find points X’ n arbitrarily close to X and Y’ arbitrarily close to Y. \ Vj-1 for 0 < s < SO. and not to Dj-1. j 3 rank(X) ) and. < 0 and if dXX’w = adv. \ V. so this would imply that which is impossible.-1.-I at least once. If y~ meets Dm-l. . So.( & / a ) sec so + . Then.

a 2the sphere with radius one-half. there may then be more than one minimal geodesic of length SO from n(X) to n(Y). Thus. where so in [0.2 indicates that.120 GEODESICS IN SHAPE SPACES and so 0 < s < so. -1/(22/2)} so that A2 < 0 and we are in the exceptional case. the non-uniqueness of geodesics between two shapes is no longer as obvious as (i). from any point on the fibre above n ( X ) . 6. For then Theorem 6. in shape space. we also have R ( X . The following example illustrates that this situation can occur in C:. Thus. n(Y)) > p(n(X). we shall be able to identify precisely when this does occur. as we pointed out earlier. This phenomenon of the non-uniqueness of minimal geodesics between two shapes is associated with the fact that R(X. We shall also see in the next section that when max{rank(X). n(i. Although this is a topological sphere. although distinct members of R ( X . set G ( n ( X ) . we may be able to find more than one minimal horizontal great circular path to the fibre above n ( Y ) . Let + Then. n/2] is determined by cosso = 1/2 . Y) = { I } and the horizontal lift starting at X of the unique minimal geodesic between n ( X ) and n(Y) is for 0 < s < so. rank(Y)} = j < m it may be possible for a minimal geodesic from n(X) to n(Y) to lie in a higher stratum than ( C g / i j )\ (Cg-l/ij-l).1/(22/2). However. YX' = diag{l/2. so that n(y([O. Y) do not necessarily give rise to distinct geodesics. when k > m 1 and p ( n ( X ) . y ~ ( s ) X is singular at this point and so y ~ ( s must be ' ) singular since X t is a non-singular square matrix. the following shows that. Thus. Indeed. SO] is two. it is not a metric one. for any point n(X) E Xi. as we shall see in Section 6. Y) may contain more than one element.(Y))). It is clear that the rank of y(s) for all s E [0.6 THE NON-UNIQUENESS OF MINIMAL GEODESICS Recall that the shape space E: of three labelled points in the plane is S2($). On the other hand. the geodesic may avoid n(Dm-l) altogether. n(Y)) of all minimal geodesics joining n(X)and the antipodal point n ( Y ) at a distance which we could take to be the equator n / 2 from it can be identified with S' associated with n(X).8. Correspondingly.so])) avoids nm>. . each being of length SO.

if h.THE NON-UNIQUENESS OF MINIMAL GEODESICS 121 it was in the metric sphere C. where 1 is the multiplicity of the eigenvalue h i of XY'YX'. we have Q = V ' A V . Y). these minimal geodesics lie in the top stratum. that is.oss &c 0 a sins b sins c sins 0 where the only restriction is that a2 b2 c2 = 1. We note also that. Q is the unique non-negative square root of Q2 = V t A 2 V = XYtYXt and so is uniquely determined by X and Y. < 0. ( Y ) ) of minimal n geodesics between two arbitrary points of shape space. Y ) is unique unless h. which means that it is an eigenvector .-l h. using our extended definition of geodesics. where. Since YXt = 0. Then. the horizontal geodesics are -toss Y(S>= 0 1 . where 1 = rank(YXt).l ) .v~w. (i) When rank(YX') = m. (ii) When rank(YX') 6 m . Y ) is unique if and only if h. Thus.cos2 s L b sinscoss Jz 0 L a sins cos s Jz sin2 s which shows that the shape n ( y ( s ) ) depends on a and b for 0 < s < SO = n/2. when a2 b2 = 1.1. despite their endpoints being in lower strata.2h. Then. is the unique positive of XY'YXt with eigenvalue h i . Y) is in one-to-one correspondence with RPl-'. let v be the final row of V .-l + h. Proof. . the set R(X.. 3 0. R(X.1. > 0. = Q . for some pseudo-singular values decomposition YXt = U A V . + Lemma 6. R E R(X.cos2s Y(S>'Y(S) = ( h a si:scoss . then R E R(X. 0 0 0 so that rank(X) = 2 and rank(Y) = 1. Let us take X = ( T 5 0 0 0 0) 0 and 0 0 1 Y=(o o 0). in which case the set R ( X . Y) is in one-to-one correspondence with SO(m . Y) = SO(3) so that the choice of R is unrestricted. If h. We examine first the indeterminacy in the factor Q in the pseudo-polar decomposition YX' = RQ. We first examine the lack of uniqueness of R E R(X. + Thus. = 0. except when c = 0. In this section we shall determine the size of the set B(n(X). This two-parameter family of horizontal geodesics in the pre-shape sphere does map to a two-parameter family of minimal geodesics from n(X) to n(Y) since + + .

4. the sign of v. geodesics from n(X) to n Y) coincide ( if and only if they have the same tangent vector at n(X). Finally. the quotient map n is a Riemannian submersion and so maps the horizontal tangent space at X isomorphically onto the tangent space at n(X). G(n(X). Thus. we assume that rank(Y) = ro and we may choose Y within its orbit to lie in f S k . Q = VbAVo. rank(YX') = m implies that rank(Y) = m and so the minimal geodesics coincide if and only if R = k. Thus. > 0. This implies that VoRF'R2Vb = diag(Zl.where r = min{rank(X). In case (ii). where is A with the final diagonal entry h. = 0. = 0. as we have seen. Suppose then that R1 and R2 are any two possible elements of R(X.122 GEODESICS IN SHAPE SPACES square root of XY'YX' since Q* = V ' A V . path-length parameterised. is unique. In case (i).ro rows of YXl must also . T } with T E SO(m . Two.dyk(s) I S=O ds that is. (i) When rank(YX') = m.1(i). so the resolution of the indeterminacy in Q is equivalent to choosing a point in the real projective space RPI-'.1) and thus. and then it is in one-toone correspondence with RPl-'.. B ( n ( X ) . T}Vo. if Ro = UOVOis any element of R(X. if and only if R'Y = k'Y. Y). the result follows from Lemma 6.ro-l. Then. where T is an arbitrary element of SO(m . replaced by -A.-l + h. say. the factor Q. which may be any unit vector in the I-dimensional eigenspace in Rm associated with the eigenvalue h i of Q2.1). We next use our knowledge of R(X. if h. where 1 3 2 is the multiplicity of the eigenvalue h i of XY'YX'. Proof. if h. the others are given by R = U Odiag{Zl. so that VoRF'R2VbA = A . the last m . without loss of generality. Hence. then YX' is non-singular and there is a one-to-one correspondence between the R's and the Q's that occur in the pseudo-polar decomposition YX' = RQ since then Q = R-' YX' is non-singular and R = YX'Q-' . However. So. is irrelevant. (ii) When rank(YX') < m. Now. Y). However. rank(Y)} 3 1 and O 1 = rank(YX'). if YR and yjj are the horizontal lifts starting at X of the two minimal geodesics. n(Y)) is in one-to-one correspondence with the Stiefel manifold Vm-l. the indeterminacy of R coincides with that of Q. Q is determined by v. # 0. RL'R2Q = Q. Then. n(Y)) consists o a single f geodesic save in the special case when h. Y ) to identify G'(n(X). n(Y)). the two geodesics in shape space coincide if and only if dYR(S) ds I S=O . Theorem 6.

and (2) to lie in ES.3 we allowed for the possibility that a minimal geodesic may move into a higher stratum than that in which it starts and. in particular. This completes the proof. So. if R = UV.r . n(X) could be a non-singular point and n(Y) a singular point. Y) can be uniquely written in the form = U diag(Zl. and if 0 ) Vl is a singular values decomposition of and V = diag{det(Vl)VI. The following corollary follows from Theorem 6. ] . and that each T generates an element of R(X. then there are minimal geodesics between n(X) and n(Y) that do not lie in n ( D r . Proof. T ) V . Then.3. R'Y = Z'Y e U'Y = diag{Zl. When k = rn 1 and rank(X) = rn we always have rank(YX') = rank(Y).-l + Note.Z m . For example.We know from the proof of Lemma 6. Zm-ro] . T'}U'Y T = diag(Zro-l.THE NON-UNIQUENESS OF MINIMAL GEODESICS 123 vanish and so the matrix U E SO(m) in the pseudo-singular values decomposition where U1 E UAV of YX' may be chosen to have the form U = diag{Ul. SO(r0).l(ii) that each element of R(X.1 ) has the above form with T2 E SO(m . I f max{rank(X). T2}. we gave an example of such behaviour in Xi.2. Y) in this way. )ifand only ifthere is more than one minimal geodesic between them. ( In Theorem 6. rank(Y)} < m. . Without loss of generality we may assume that 1 = rank(YXt) 6 ro = rank(Y) 6 rl = rank(X). that is. U = diag{det(Ul)Ul. We may also choose X= to lie in ES. then rank(Y) < m is sufficient for the minimal geodesic between n(X) and n Y) to be unique. if U I ( A1 YIX. if and only if T E SO(m ..ro). at the beginning of this section. Thus. only ifeither h. that there is a unique minimal geodesic between n(X) and n ( Y ) when rank(YX') = rank(Y) < rank(X).We are now in a position to say precisely when it can occur. Corollary 6. for T E SO(m . + Corollary 6.4. There is a unique minimal geodesic joining n ( X ) and n ( Y ) if and A > 0 or rank(YX') = min(rank(X). rank(Y)} = i-1 < m.l ) .

We shall denote by Co(n(X))the set {n(Y) : 16(n(X). in shape space we shall see that.5 to a neat characterisation of the cut locus. Since h.7 THE CUT LOCUS IN SHAPE SPACES We are now ready to determine the cut locus associated with any point n(X) in EL. by the cut locus C(n(X)) of n(X).1). However. and when rank(X) < m Co(n(X)) n ( D l . as the set of points to which there is more than one minimal geodesic. there is at least one alternative and shorter geodesic back to n(X) that will not pass through n(Y). Y ) if and only if it can be expressed as R = U diag{Zl. when rank(X) = m. 6. U where 1 = rank(X).-l = 0 is equivalent to h. there are minimal geodesics between n(X) and n(Y) that do not lie in n(D.4 we see that.-1 = 01.. where T I E O(r0 . = -h. The precise situation is reminiscent of that on a sphere. From Theorem 6. : + A.124 GEODESICS IN SHAPE SPACES then UV will be in R(X. Here.rl rows of R'Y are all zero: that is. That is. if and only if T = diag(T1.-l 3 0.l ) . Similarly. beyond which that geodesic ceases to be minimal. outside a set of codimension two. h. Lemma 6. apart from the singularity set. Thus.Y O ) and det(Tl)det(Tz) = 1. if and only if rank(Y) > rank(YX').I ) = {n(Y) : rank(YX') < rank(X)). one on each geodesic through n(X). Proof. Co(n(X))is the union of {n(Y) : h. + h.-1 < 0.2 this last condition holds if and only if there is more than one minimal geodesic between n(X)and n(Y). When rank(X) = m Co(n(X)) u n ( R . By Corollary 6. n(Y))I > 1) of shapes to which there is more than one minimal geodesic from n ( X ) . Y).2 )= {n(Y> h. since we have extended the definition of geodesics into the singular part of shape space. for each point n ( Z ) beyond n(Y) along the geodesic.) if and only if YO > 1: that is. and this is no accident: it can be traced back to the fact that the pre-shape space is a sphere and the restriction of the quotient map to each stratum is a Riemannian submersion.-1 < 0) with { n ( Y ): rank(YX') < rank(Y) < m ) U{n(Y) : rank(YX') + 1 < rank(Y) = m}. .2. TJV with T E SO(m . We first note a rather surprising lemma that ties together the geometry and the algebra and leads in Theorem 6. there is an equivalence between nonextensibility and the existence of a multiplicity of minimal geodesics from n ( X ) to n(Y).. the whole of the corresponding YR lies in &S$ if and only if the last m . = -h. this automatically extends the definition of the cut locus also. However. we mean the locus of cutpoints n(Y). So R is in R ( X . This is a standard definition of the cut locus for a manifold. T2 E O ( m . In a manifold the cut locus is characterised. T21.

That occurs if and only if there is an E > 0 such that the eigenvalues &(so + E ) coincide with the pseudo-singular values of y ~ ( s 0 &)Xt:that is._. Hence. so that i i ( s 0 ) = hi.(so &)I. Let h l ( s ) 3 i 2 ( s ) 3 .n / 2 ] )is horizontal everywhere. Proof. Magnus and Heudecker. If we write w > 0 for the smallest eigenvalue of XX'.2 to give a complete characterisation of the cut locus of an arbitrary shape n(X). if and only if i r n . vf R (s)X'v y VIZl Bv=O ' where B is an (m . = -h. Since i m . n ( Y ) contributes to the cut locus precisely when there exists EO > 0 such that.-l = 0 is equivalent to rank(YX') < m . Z rank(X) = m. the identity follows from Co(n(X)) = {n(Y) : rank(YX') < min{rank(X).5. n(Y) $ C(n(X)) if and only if there is an E > 0 such that SO E is the distance between n ( X ) and n ( y ~ ( s 0 E ) ) .THE CUT LOCUS IN SHAPE SPACES 125 from which the stated result follows since h.i) x m matrix (cf.(s) be the eigenvalues of y~(s)X'for s 3 so. we have hm-l(s0 E ) SO E ) < 0.-l = 0) if rank(X) = m. then for v # 0 we have 0 < adw 6 v'XX'V. so that it is a symmetric matrix.l ( s o E ) 3 li. We now employ Lemma 6. + h. . By Lemma 6. for 0 < E < EO. Y).so > 0. + + + + + + + + + + (i) The case that rank(X) = m.1 ) of Z. 3 h. v'yR(s)x'v 6 from which it follows that I ~ sin s sin E vtR'YX'v wv'v. . the cut locus C(n(X)) of n(X) is the union of the f singularity set n(D. 1988). rank(Y)]].2 it is sufficient to prove that {n(Y) : h. with the set Co(n(X)).-2) of E with the set Co(n(X)) and. for any R E R(X. i f 1 = rank(X) < m. sin SO sin SO ~ hj(S) < sin s ~ sin SO hi. when rank(X) < m and so rank(YX') < m. 16iGm. it : is the union ofthe subspace n ( D l .Since yR([O. Theorem 6. for any %r # 0 and E = s . if rank(X) < m. . {n(Y) : rank(YX') < rank(X)] Note that. Similarly. The eigenvalues of ~R(s)X' be expressed as can i i ( s > = max min BB'=I.l (so E ) 3 im(so E ) .2.

if w is an eigenvector of hi for rank(X) < i 6 m. = 0. However. n(Y)) = n / 2 . This is not suprising i if we notice that the spaces = CPkp2(4)are complete Riemannian manifolds for all k > 2. have the same null space of dimension at least one since it is always true that ker(YX') 3 ker(X') = ker(XX').-l h.-. that is.1 positive eigenvalues + + + + + + &(so + E) 3 . It follows that y ~ ( s 0 E)X' has rank(X) positive eigenvalues and m . + + + When m = 2. so that y ~ ( [ 0SO E ] ) is a horizontal lift of a minimal geodesic. Therefore. E ] ) is still a horizontal lift of a minimal geodesic and n(Y) $ C(n(X)). n(X) and n ( Y )are a maximally remote pair. when h. as in manifolds. Finally. when m = 2. + u'yR((S0 + E)X'u >0 =0 if w is an eigenvector of hi for 1 < i < rank(X).-l = -A.2. the cut locus of n(X) consists precisely of all shapes that are most remote from it. > 0. however. E) > 0. > 0. so that n(Y)E C(n(X)). for E > 0 small enough._l (so E ) < 0 for any small enough E > 0. although y~((so E)X' still has m . so that n ( Y ) E C(n(X)) in either case. . i m ( ~ ~ o e ) 5 h.rank(X) > 0 zero eigenvalues. its pseudo-singular values cannot equal these eigenvalues and therefore n(Y) is in the cut locus of n(X). that -I. So. When h. however. If rank(YX') < rank(X). then if m = 2 we must have the two pseudo-singular values of YX' satisfying hi = -h2 and if m > 2 we must have YX' = 0. If p(n(X). this is no longer the case when rn > 2: then the cut locus of n(X) contains more shapes than just those that are most remote from it.: (so E ) h. . The situation for shape space is described in the following corollary. . If rank(YX') = rank(X). ensures that each point is in the cut locus of the other. which. then XX' and YX'. which follows from Theorem 6. Thus. by continuity.126 GEODESICS IN SHAPE SPACES When h. then for small enough E > 0 we have. (ii) The case that rank(X) < m. + where a = sin($" E ) / sinso.-l(so hm-1 + -h. In a manifold the relationship of one point lying in the cut locus of another is symmetric. y ~ ( s 0 E)X' will have one negative eigenvalue and one zero eigenvalue for any E > 0. which therefore agree with its pseudo-singular values. and hence again n(Y) will lie in the cut locus of n(X). o + + E) < u h m = -a < + E ) < 0. 3 h. that is not possible when there is more than one minimal geodesic between the two points: a symmetric relationship. n(Dm-2) = ll and so C(n(X)) = Co(n(X)). . and so also R'YX'. ~ R ( [ OSO . for any small enough E > 0.5 and Corollary 6. However. . Hence. Hence. (SO we have s (: . For our extended definition in shape space this is no longer the case: it is possible for a minimal geodesic to remain minimal when extended beyond one end point but not when extended beyond the other. there is a non-zero w1 E ker(XXt) & ker(R'YX') and there is also a non-zero u2 E ker(R'YX') \ ker(XX') such that uiXX'u2 > 0.-1 = -A.(sg E ) > 0.

tr(T0YSoX') = h. i= 1 S€O(k-1) and. we shall. 6.(Y) 3 0. respectively. Note that the zero m x ( k . In particular. E SO(m) and SO = V". Suppose that X and Y are two m x ( k . + Lemma 6. for this section.1) matrix X to vary.1) matrix that borders Ax is only required if k > m 1 and in this case we may. . = hm(X) = 0. these basic properties turn out to have simple expressions involving pseudo-singular values. on the other. to Ekl Proof. then. for To = U x U . VX E SO(k .i(X)hj(Y). is the collinearity set of shapes of k labelled points in in R2 and n(Dm-l) in is its generalisation to 08". However. and this maximal value is equal hi(X)hi(Y).X E SO(k . insist that h.THE DISTANCES AND PROJECTIONS TO LOWER STRATA 127 Corollary 6. if U X ( A x 0 ) VX and U y ( Ay 0 ) V y give pseudo-singular values decompositions of X and Y. then n(X) E C(n(Y)) implies that n(Y) E C(n(X)).3. Ckl .l).8 THE DISTANCES AND PROJECTIONS TO LOWER STRATA In this section we calculate the distance from a given shape n(X) to the various strata n(Dj \ Dj-l). (cf. write its pseudo-singular values decomposition as where we recall that U X E SO(m). it is the pseudosingular values of X itself that we need. since rank(X) = rank(Ax). We note first that. 1974) max tr(TYSX') T€SO(rn) S€SO(k-I) < m TcO(m) max tr(YSX'T) 6 C A ~ ( X ) A ~ ( Y ) . we see that. r f rank(X) 3 rank(Y). as mentioned above. and find the nearest points to the shape on these strata.I).S E SO(k . 1 6 j 6 m . the converse implication holds if and only if there is more than one minimal geodesic between n(X)and n(Y). which is if and only if there exists more than one minimal geodesic between them.m .1)) is attained for some To E SO(m). Then. and do.1.4.1) and Ax is the diagonal matrix with diagonal entries the pseudo-singular values &(X) of X.1 since n(D1) Ck. Once again. SO E SO(k . except that now. .(X)h. if n(X) and n(Y) are in the same stratum or one each in the two uppermost strata. On the one hand. Since we shall need to allow the m x ( k . if and it lies in the jth stratum if and only if in addition hj(X) # 0.1) matrices such that h. Ben-Israel and Greville. The most interesting and important case in applications will be when j = m .(X) 3 0. max{tr(TYSX') : T E SO(m). then n(X) E C(n(Y)) if and only if n(Y) E C(n(X)). n ( X ) is in n(Dj) and only if hj+l (X) = .

.(X) # 0.(X)’.(Y) the square root of C-(=. > h. )’l Moreover. the maximum of the function E:=l h. cos(p(n(X).128 GEODESICS IN SHAPE SPACES Now.(n(X)) the distance p(n(X). . if h. j + 1 < i < rn. m .. denote by s.)) any point at which a geodesic of length s j ( n ( X ) ) and from n(X) hits n(Dj). Proof. since h L ( Y ) 0 for i > J .) n ( X ) .1 < j = i=l . . is h.). and we may take X(j) = (AX’ 0 ) V X . Moreover.. <m . (ii) Ifhl ( X ) > . ) ) are not necessarily unique. ( Y ) }subject to the condition h.where A(f“ = diag(hy).3. then both n(X(. which is achieved when h . by Lemma 6. ifX $ D~i-1. We emphasise that the n ( X ( . V y to be equal to V X .(Y>*= 1.6. then the above to argument shows that the diagonal matrix of the pseudo-singular values of p must be A$’. i ( ~ ) ~ .)) and the corresponding geodesic will be unique. (i) For Y in Vj let hi. and x:=l I . . but in any case it is natural to regard them as the projections of n(X) onto n(D. h i ) }and hy’ = hi(X) chr(X)2 (r:l . as we may.+l ( X ) . Then.) by n(X(. Theorem 6. 1<i <j . any choice of X(J) will have its jth pseudo-singular value A:”’ # 0 and so n ( X ( J ) )will not lie in n ( D j . Thus. and hi” = 0 . . If we next vary the { h . then. n ( Y ) ) )= c h i = max tr(TYX‘) T€SO(rn) i=l < . n ( D j ) ) from n(X) E Xi to the closed set n(D. for given {hl(X)}. 1 < i < rn.l ) .(X)h.-l). then cos2(s.(n(~))) C i . (ii) If n(p)is one of the nearest points in ~(23. ( Y ) is equal to the A stated for 1 < i < j.I. necessarily doing so perpendicularly. the : ) given choice of X ( J ) is valid and. be the pseudo-singular values of YX‘. the last maximum is = achieved if we choose. (i) I f U X ( A x 0 ) V X is a pseudo-singular values decomposition of X . . S€SO(k-1) TESO(m) max tr(TYSX‘) = i= 1 hi(X)hi(Y).then any choice for n(X(J))lies in the stratum n(Dj\V.

each 6. . from each other and from the remaining entries. . the first j rows of Vp and V x are identical. 1937) that AB and BA are both symmetric so that there is an S1 E SO(k . for n ( X ) E n(D. which in this case are the eigenvalues of X‘X. n ( X ( j ) )is unique. Now. In particular. hj(X)’ is the square of the sine of the distance between n ( X ) and n(DJ-l).1 .) and. hence.j pseudo-singular values of Y are all : zero. It follows that SOhas the form diag{Zj. and the geometric meaning of the eigenvalues h i ( X ) 2 of X l X is now obvious: h i ( X ) 2 = c o s 2 { s i ( n ( x ) ) ]. AB = S~BASI. SO chosen in SO(k . > h . the fact that hl ( X ) > . Then.1) such that tr(AB) = h i ( X ) h y ) . That involve the pseudo-singular values of Y‘X. the smallest eigenvalue h. ..). ( X ) > 0 implies that = h y ) for 1 < i < j . the first j rows of V X are uniquely determined where the submatrix 30 up to signs and SO must have the form VxVZ diag{Zj. the fact that A is a diagonal matrix with the first j diagonal entries distinct. must be fhj!’ or zero. The uniqueness W of the geodesic between n ( X ) and n ( X ( j ) )follows from Corollary 6. j l } .(X)2 is the square of the sine of the distance from n ( X ) to n(D.j ) . Hence z!=l B= However.1) such that AS1 and S i B are both symmetric and. On the other hand.is interesting to compare It this with the expression we obtained in Section 6.m ) > = 2s.-l(n(X)) = 2Sm-l (h(n(x>>> . Since hl ( X ) > . von Neumann. the diagonal matrix Ax supplies us with information on the position of n ( X ) relative to the various strata in Zk.1).C O S 2 { S i & l ( n ( X ) ) ) . Thus. &}: that is. > h j ( X ) > hj+l ( X ) . are also eigenvectors for B. This implies (cf. each generating one-dimensional eigenspaces of A. It then follows that the first j basis vectors. .THE DISTANCES AND PROJECTIONS TO LOWER STRATA 129 with To. we have n(p) = n( ( A$) 0 ) V X ) that is. since B‘B only has eigenvalues zero or (hjJ’y. whereas here we have expressed the distance between n ( X ) and n(Dj)in terms of the squares of the pseudo-singular values of X .2.-.4 for the distance between any two shapes n ( X ) and n(Y). It is also worth noting that p ( n ( X ) . SO} is in SO(k . means that S1 must take the form diag{Zj.1 < i < j . Since the last rn . tr(AB) 3 tr(TAB) and tr(BA) 3 tr(TBA) for all T E SO(k . b n ( j .

Each of (i) and the second result of (ii) is obvious. and the shapes n(X) that are maximally remote from n(Dm-l) have the pseudosingular values hl ( X ) = . Also.) n(Dj. ) is the composite of the projection of n ( X ) onto n ( D j 2 ) with that of n(Dj.n(D.6. . when sjl ( n ( X ) ) is small enough. ( n ( X ) ) 2 2 5 Sj.(X>*). .). = Ihm(X)I = l / f i and are also maximally remote from n(D.1 . Proof. for any 1 6 j l < j 2 6 m ..(n(X>). ( n ( X ) ) and n(X(’)). the projection of n ( X ) onto n ( D . From Theorem 6. The second statement in (ii) says that. 1 < j l < j26rn-1.6 we can deduce the following properties of s .130 GEODESICS IN SHAPE SPACES since the first term is arccos(1 . Corollary 6. (n(X(j2)))2. i f n ( X ) $ n(D.))) i= 1 9 hi ( X ) 2 - 5hi( X ) 2 i= 1 . and the first result of (ii) follows from cos2 (q( n ( X ( j 2 ) ) ) )= cos2 ( p ( n ( X ( j ’ ) ) . for < . This implies that any shortest geodesic from n ( X ) to n(Dm-1) extends to pass through 1.2h.(T(X))2 + sj. (i) For n ( X ) E Zk.1 we shall have sj. Corollary 6.5. ! then (ii) For 1 < j l < j 2 6 m .). . 1 j 6 m .2. The maximum possible value of the distance of shapes from n(Dj) is arccos E.1. and. it follows onto ). from (ii) that whenever 1 < j l < j 2 6 m .

However.) . . such . the fact that n ( X + ) and n ( X . as f is bijective. In Cz+' the configurations whose shapes are maximally remote from n(D. 1. However. However. + { & z!=l Theorem 6. .) and.l )) = S. which is the greatest distance between n(X*) and points of n ( D j ) for any j .) are equi-distant from all points of n(D. we have h. we recall that..= 2 diag(1. n(Y)cannot map into Sand. increases to n/2. These observations may be expanded to give an alternative proof of Casson's theorem and. the distance to YO).. for m 3 3 this is no longer the case and. = h. then it would not be possible for each of f ( n ( X + ) ) and f ( n ( X . Le.-l) means that f ( n ( D m ..1.-l) are characterised by their pre-shapes X satisfying X ' X = AZ. 1991). . .THE DISTANCES AND PROJECTIONS TO LOWER STRATA 131 This corollary allows us to take a more detailed look at the special case k = m 1.) ) to be maximally remote from f ( n ( D . 1 < j < m .(Y) i=l fi # P ( n ( X + ) > n(Y>>. and X .1. since we have already given more than one proof of Casson's theorem.) are both maximally remote from n(D. when n(Y) is not in n(D.(Yo) = 0. J. = h. 1 < j < m .bn(n(Y>>> ) = arccos (-f i c 1 lr-l 1 h i ( Y >. then the geodesic distance between n ( X * ) and n ( Y 0 ) is arccos(l/fi). In particular. Then. C:+' is homeomorphic with a sphere. Proof.--hh. for there to be points f ( n ( Y 1 ) ) and f ( n ( Y 2 ) ) in f ( n ( D . as m increases.-l) as the corresponding 'equator'. .)for all j . at the same time. the isometry 1. when m 3 3. However. . if Y O is any matrix with pseudosingular values h l ( Y 0 ) = 1 and h2(Y0) = . then J.) decreases to zero and the longest distance to its points.(Y) # 0 so that P(n(X-). for any point n ( Y ) in n(D. Thus. If we take X + = I Z . Cz+' is not isometric with Sdz+' i f m 3 3. Suppose it were possible to find an isometry f : E:+' Sd:.7. at the same time. First. Thus.-l).(Y) = 0 and the pseudo-singular values of Y(X*)' are the ( h i ( Y ) / f i : 1 < i < m ) . we must have f ( n ( D m . with n(D. the distance of n ( X * ) from n(D. n ( Y > )= p ( n ( X f > . interchanges n ( X + ) and n ( X . . .) distances the p ( n ( X + ) .l > ) at different distances from f ( n ( X + ) ) . if f were an isometry.l ) > and. . by Casson's theorem. n ( Y ) ) and p ( n ( X .. all points of n(D1) are of the form n ( Y 0 ) so that the distances from n(X*) to all points of n(D1)are the same. Geometrically. n ( Y ) ) are both equal to arccos hi(Y)} since hj+l ( Y ) = . which would be required. and behave very like antipodal poles of the sphere. to show that C:+' is not isometric with a sphere (cf.+' of the shape space with a standard sphere in RdE+'+l.l ) ) would have to be contained in a totally geodesic sphere S of codimension one. we shall now assume it and deduce the latter result. When m = 2. -11. . n ( X + ) and n ( X . .

we may write X = (c. Then. Clearly. A pre-shape for the shape (x. then. When m = 2 the points X.y ) is and a pre-shape for its projection onto the collinearity set is determined by a unit eigenvector (a.where X E SmP2. 0.X>). where X E S"-' and the ith coordinate of the corresponding pre-shape is T + xi = c . b ) corresponding to the larger eigenvalue h: of X t X . We then have. are all equal to c / ( 1 + c ) since the other solution X>) * xj -* c = (Xi. the eigenvalues of X t X are . T i for those i.c 2 ) ( q .5) must be -1/2. .132 GEODESICS IN SHAPE SPACES a configuration is a regular simplex. However. and X of So must be f l and then the constant c appearing in ' i (6. 0) and its second vertex to (l/&. rotating and scaling a labelled triangle in the plane to move its first vertex to (-l/&.5) to show by induction that. ) = c2 c = 1 of (6. y ) of the third vertex of this resultant triangle as the coordinate of the shape of the original triangle. Chapter 1). x = X T : g EFT. we T must have the first coordinate of X equal to the common inner product c so that. for i > 1. To prove it in general we write T for the circumradius and X . + . 1 < j < m. (6. j > 1 and i # j .) = . 0). 0). . the collinearity set is then represented by the x-axis. . If we rotate the configuration so that X. As a further application of the results in this section.1. we calculate the xcoordinate of the orthogonal projection of a shape onto the collinearity set n(D1) in the planar representation of the punctured version of X (cf. We may then use the non-trivial solution of (6.5) corresponds to the totally degenerate configuration.l / m for all 1 6 i # j < m + 1. .q+.for the circumcentre of the simplex with vertices xi*. +XT) . equivalent to the original configuration X * being a regular simplex. d n X i ) . This result has already been seen for the case m = 2 in Figure 1. The : : planar representation of X to which we refer is obtained by translating. Using the increasing sequence of submatrices of X t X formed by the first j rows and columns. ( : X. This is X. .(q d T where c' is a scaling constant.Xg) is constant for all i # j .5) which implies that the (Xi. we may show by induction that X'X = $Z if and only if (XT. for all rn. + + (1 . We then use the coordinate (x. for 1 < i < m 1. = (1. .

x2 . 0) so that its x-coordinate is 1 .THE DISTANCES AND PROJECTIONS TO LOWER STRATA 133 so we obtain a=- X J W + + - and b = 4 W ' 1 -h: where h: = h:( 1 x2 y2).J ( 1 + x2 $. . which was to be expected since we know that the geometry of the planar representation of C: differs from the standard Euclidean geometry of R2. The planar representation of the projection is thus given.4y2 - 1-x2- y2 + 2x 2x J(1 x2 + + y 2 ) 2 . when x # 0.y 2 .y2)2 . by (b/a.4y2' This is different from the x-coordinate of the given shape unless y = 0.

G. It is an important feature of these results that the geometric invariants are all expressed in terms of the pseudo-singular values of the pre-shape matrices that we use to represent the configurations that determine particular shapes. Since. inner product on the Euclidean space [ W m ( k . BARDEN & T. C A R " & H. as is often convenient. In that chapter we represented the tangent space to SL at X as the set of matrices Z such that tr(XZ') = 0. 7. to ignore the Riemannian structure and treat shape coordinates as if they were standard Euclidean coordinates. 1 6 j < k < - 1 of the tangent space to Rm(k-') . in terms of which it is natural to describe probability measures on shape space. and that these pseudo-singular values are easily computed in practice. We now turn to the problem of quantifying those aspects of the metric that are necessary for statistical purposes. the entries of Z are the coordinates of a tangent vector .Shape and Shape Theory D.1 and since the geometric quantities with which we are concerned are all local in nature.1 THE RIEMANNIAN METRIC We recall from Chapter 6 that a Riemannian metric is a choice of inner product on each tangent space. restrict our attention to the case k 3 m + 1 throughout this chapter. which is the infinitesimal generator of Brownian motion. it follows that we may.with respect to the & : 1 < i < rn. Z is orthogonal to X with respect to the standard . K. More precisely. as usual. we shall compute the volume element. In particular. In the final section we calculate various standard measures of the curvature of shape space. LE Copyright 0 1999 by John Wiley & Sons Ltd CHAPTER 7 The Riemannian Structure of Shape Spaces In Chapter 6 we introduced the Riemannian metric and made an initial qualitative study of it through its geodesics. That is. KENDALL & D. which varies smoothly as we move around the manifold.l ) in which SL lies. Xk is isometric with the quotient of Xipl by the involution l k . when k < m. These quantify the extent to which it is feasible. and the Laplace-Beltrami operator.

Hence.-. .t ~contains the subspace specified in (7. k - # j}. where the relevant choice of n and p will be determined by the context.1). 0 < i < 2 . .1) + m(k - 1. A multiplies each of the first m . = 0) We may now use these spaces of matrices to identify the horizontal subspace at X as follows. then Z E M ( m .m) . Recall that we are writing M ( n . since hl 3 . ifrank(X) 3 m . We further define matrix spaces WO { ( w .I. p ) for the space of all n x p real-valued matrices and we shall write E. > 0. Accordingly. Since we shall need to change basis from time to time in this chapter it will be necessary to make this distinction between a tangent vector itself and its matrix of coordinates with respect to a particular basis. tr(XZ') = tr(AW0) = x h i w i i = 0 i= 1 and XZ' = U(AW1A + AWoJU' = U(AW1A + WoA}U' = ZX'.t ~ the space of standard coordinate matrices be o horizontal vectors at X. j = wji and 1 . m ) : w .1 = m + i m ( m . m ) : w . the Riemannian structure at a shape n(X) is directly inherited from that of the horizontal subspace of the tangent space to the pre-shape sphere S i at any pre-shape X in the fibre above n(X). Then.m). To specify that metric we shall require a more precise analysis of the horizontal subspaces above non-singular points of shape space. If Z = U {A ( W I 0 ) + (Wo W z ) }V . f f Proof.136 THE RIEMANNIAN STRUCTURE OF SHAPE SPACES at X.1) is dim(W0) + dim(W1) + dim(W2) .j for the matrix whose ( i . the dimension of the subspace (7. j = 0 if i = and W2 = M ( m . are the diagonal elements o WO .1. 7 . Thus. wii W1 = {(wij) E M ( m . k m - I). However. k . Suppose that the pre-shape matrix X has a pseudo-singular values decomposition U ( A 0)V and let 7. x h i w i i = 0 : i= 1 o M ( m . j)th entry is one and all of whose other entries are zero. we shall refer to Z as the standard coordinate matrix of <.1.l). 3 h. in the non-singular part Zk\n(Dm-2) of shape space. where the w. .t ~ the subspace f is m U{A(W1 O)+(WO W 2 ) } V W i ~ W i . As explained in Chapter 6. j ) E M ( m . 7 . Lemma 7. m).1 rows of W I E W1 by a nonzero multiple and so acts as a monomorphism of W1 into M ( m ..

l<i<m<j<k-l. the inner products of the vectors n*((. 1 <i<m.2) where W 11 . where if i # j .1. 0 = 1 if i = j and 0 . .E. However. i < j < k . in T & ~ ( Z ~ \ n ( V + ~ )of the corresponding vectors . ].l ~is specified by (7. : 1 :1 ( . 1 < i < j < m.Eizj2 = S i l j 2 S . Proof. .l ~ .It follows that 7 . hl 11. J . . with respect to which we may calculate the Riemannian metric on shape space. otherwise. the tangent space to Zi\n(Vm-2) at n(X)has a basis {n*((. Wij = A{Eij W .)< i < m } u { n *--'J ( .1) (7.) --rJ --'J is thus sufficient to compute the inner products of the Zij to obtain the required result. ifl<il=i2<m< jl=jz<k-l. is the derivative at X of the projection n of Sk onto Xk. ..1. This description of 7 . whose standard coordinate --I J matrices Zij in 7-t~are given by zij = uwjjv. l j z . It . .1) as required. Hence. Ifrank(X) 3 m . the matrices Zjj given by (7.. . are the same as those of (. since n is an isometry when restricted to 'Hx. (zilj l 9 z Z 2 = tr i j ) pi1 j l ~ : ~ =~tr) (witj t j Wf2j2 Sjj 1 H and since.3) [I+$ ifl<il=iz=jl=j2<m.2) form a basis of E X . By Lemma 7. tr EiIj. form a basis of that tangent space and.1.) ) --1J . the stated results follow.11 h . I:.l ~ gives rise to a simple choice of basis vectors involving the components of a pseudo-singular values decomposition U ( A 0 ) V of X . . the images n*((..E .. where n. for example. (7. Theorem 7. Suppose that X has a pseudo-singular values decomposition U ( A 0 ) V .) -11 such that the inner product is given by < i < m . l + Eji}.>E . This basis is the set of horizontal tangent vectors (.THE RIEMANNIAN METRIC 137 which is the dimension of 7 .

2 THE METRIC RE-EXPRESSED THROUGH NATURAL LOCAL VECTOR FIELDS In the previous section we found. although the matrices Zij given by (7. Note. A) determined by A lies on the unit sphere S"-' in R". For example. when it is embedded in the space of m x m real-valued matrices that is identified with Rm2 with coordinates xi. We note that.2) are defined in terms of globally defined vectors Wij E M ( m .m . then the tangent vector itself that a skew-symmetric matrix A determines at the identity is the linear combination tr(A ax:) = tr(A' ax. and the local embeddings enable us to transfer naturally defined vector fields on SO(m). the corresponding shape. To describe the required vector fields on SO(m) we recall that. dense subset of shape space.l). so -'I that their images under n form a basis for the tangent space to shape space at . . = + + St-' St-' St-' &. may also make general computations difficult. we may still use the same notation since Vm. .' of S"-' and FLV]of Vk-1. the presence of the m x ( k . we are able to describe open subsets S t .. which is a local phenomenon. that span the horizontal subspace. however.1) matrix X. The nature of this involvement of the pseudosingular values decomposition in these fields.1) block of zeros in a pseudo-singular values decomposition means that the last k .m . since A D = D A for any diagonal matrix D. the tangent space to SO(m) at the identity I is the set of vectors whose standard coordinates form m x m skew-symmetric matrices." such that the product S O ( m ) x x Fp.). k . so that we are really concerned with the Stiefel manifold Vk-1. the Zi. . it is not always easy to find the expression for the Poisson bracket between two corresponding tangent vectors. since the way in which they are expressed depends on a particular choice of pseudo-singular values decomposition of X.such factors may be transferred from the Stiefel manifold component on the right into SO(m) on the left. themselves are only defined locally. X lies on the pre-shape sphere S: if and . In all cases. U of Rm2 at which we are taking that basis. When k > m 1. and F[VI into local vector fields on the image neighbourhoods. and hence x F ~ V I mapped diffeomorphically onto a neighbourhood of is n(X) in shape space."is SO(m) itself.. Thus. that. varying as X varies. it will be convenient to turn to the study of another family of local vector fields that arise naturally via a pseudo-singular values decomposition X = U ( A 0 ) V of a fixed m x ( k . Left multiplication by an element U of SO(m) is a linear transformation of the matrix space sending SO(m) onto itself and . The union of these neighbourhoods will be an open. . a set of independent horizontal tangent vectors {. When k = rn 1 and there is no such block of zeros. if we denote the standard basis of the tangent space to Rm2 formally and by axu when we wish to specify the point by ax. . However.138 THE RIEMANNIAN STRUCTURE OF SHAPE SPACES 7. only if the point ( h l . for each pre-shape outside Vm-2. where (ax).1 rows of V are not involved in the determination of X." of orthonormal rn frames [V] in Rk-'. Hence. since tr(X'X) = tr(A2).] is mapped diffeomorphically onto a neighbourhood of X in the pre-shape sphere.

as the quotient of SO(k .l ) .1) unchanged. : 1 i < j < m} are those for a basis {qi. to UbUt. it will frequently be convenient to use the full matrix U'aX. Then.2 when i l < j l and i2 < j 2 . T z ) .. similar to the above left-invariant fields on SO(m). more precisely. since det(D) = 1.m ) in SO(k .l)). by linearity..1 .} is in D if and only if each di = f l and.1) matrices to those similarly named above. we note first that. these basis vectors are all orthogonal to each other and each has norm a.. are now the analogous ( k .(V)') = 2&li28jl. which is necessarily tangent to SO(m) at U . Hence.1) submatrix of V . the tangent vector tr(A ax:) maps x to tr{A(U'aXu)'} = tr(UAaXb).1) x ( k . It follows that the space of standard coordinate matrices of the tangent vectors to SO(m) at U is {UA : A' = -A}. .aXuIu0 to = ( U l Uo)tL3Xu. To describe the subsets of Vk-l. so that we may regard the skew-symmetric matrix A as the matrix of coordinates with respect to this 'basis' of our tangent space at U . However. u0. Since tr{tiIjl (V)&. If Sij is the m x m skew-symmetric matrix Eij . Since left multiplication by U1 sends axuo U'.. containing [ V ] on which the restriction of the quotient map is injective. The Stiefel manifold component of the pseudo-singular values decomposition is not mapped injectively into shape space. is a metric space. since tr(UA dXb) = tr{A(U'aXU)'}. if [ V ]is any point of Vk-l. and D i ( [ V ] ) 1 < i < io.uo and so UbaXu. which we shall use.Eji. so that we may represent Vk-l.. .. we shall also use [ V ] to denote the corresponding m x (k . that we require we embed the product SO(m) x SO(k . the Stiefel manifold of orthonormal m-frames in R i p ' .1) in the usual manner by mapping (TI. T z ) to diag(T1. d.1) of l v ( S O ( k .m ) acting on the left. there is a matrix aXvV' of right-invariant vector fields on SO(k . To define the vector fields on Vk-l. These basis vectors are double those that lie above the diagonal in the skew-symmetric matrix U'aXu of tangent vectors at U .THE METRIC =-EXPRESSED THROUGH NATURAL LOCAL VECTOR FIELDS 139 mapping 8 1 to U'aXu. then {Sij : 1 < i < j < m ) are the standard coordinate matrices for a basis of T(SO(m))and so {USi. The fields above the diagonal in this matrix provide a basis { & j ( V ): 1 < i < j < k . . we assume such a choice made. are its images . . . whose standard coordinate matrices are {SijV : 1 < i < j < k . We have to consider instead its quotient by the subgroup D of diagonal matrices in SO(m).aXu. where D = diag{dl. or.j. these vectors form global left-invariant fields on SO(m). Without this convention we should need to take double the entries appearing above the diagonal in A as our coordinates with respect to the true basis. where the Si..1 . under the elements of D and if we choose a set of disjoint balls Bi with each Bi containing D ( [ V ] ) then the intersection i . < io F [ = n~ . the number of minus signs is even. since D is a finite group of isometries and Vk-l.. For each given [V]. Since that subgroup leaves the first m rows of V E SO(k . the frame [ V ] is that determined by the first m rows of V . However. on each FLVI.(U) : 1 < i < j < m } of 7u(SO(m)). l ( B i ) ~ D i=l is an open set in Vk-l.1) by the subgroup { I m } x SO(k .1)..

it follows. Using our 'doubled basis' convention.2. Thus.1) will therefore be a skew-symmetric matrix of the form Finally."] to 2: dejined by (S.>. as open subsets of their respective spaces.) x SO(k . .[TI)H S(diag(x1. We are now ready to state the lemma that will enable us to transfer our vector fields onto shape space.. into which we shall be mapping. It induces a diffeomorphism ~ [ V o ST-' x [F[vl onto an open neighbourhood of n(X).1 . If .'(no(V)) of n above [V] is given by o ({I.l .m ) ) V . that V k . in the same manner as was explained in Chapter 6 for the non-singular part of shape space. so the vertical tangent subspace at V is given. Lemma 7.1) x ( k .l . at [V].l). . Since the vector fields of E(V) are everywhere orthogonal and the horizontal subspace is the orthogonal direct complement to the vertical subspace. A} : A' = -A}.1) coordinate matrices {diag(O. . m inherits a Riemannian metric with respect to which the projection n : V H [V] of SO(k . is o a Riemannian submersion. we define the open subset of the pre-shape sphere. the elements of Q(V) together with those above the diagonal in E l (V) will form a basis of the horizontal subspace and so be mapped isometrically onto a basis of the tangent space to V k .m ) act as isometries of SO(k .3(V). if we partition our matrix of vector fields on SO(k .. . The fibre n.} x SO(k .1) in the obvious way as then the vertical tangent space at V is spanned by the elements above the diagonal in E.l ) . For any X E X with a pseudo-singular values decomposition : U ( A 0 ) V. the map @["I from SO(m) x St-' x [F.. with respect to the basis {(ij(V) : 1 < i < j < k .x . X and SrP1 both inherit natural differential : structures and also that n(X:) is an open and dense submanifold of the nonsingular part of Xk. x .1) onto Vk-l. . .140 THE RIEMANNIAN STRUCTURE OF SHAPE SPACES Since the elements of the group {I. } O)T is a diffeomorphism onto an open subset of Xi containing X .1 .. by the set of ( k .. the coordinate matrix with respect to the E(V) of a horizontal vector on SO(k . Note that.(XI.

~ path esBV is the horizontal lift of the path [e"V] = no(e"'V) in Vk-l. . .vl)* is an isomorphism at each point of the domain. the first m rows of 5. . x m } 0 ) T2 E X.x } . by the Inverse Function Theorem. the diagonal For a general element Y = S (diag{xl. . . However. .. It is clearly continuous.5) . However. ."]..m). then we have S2 = SlD and diag{D.I .1 . . 0 ) T in Xk.1 .This implies that @["I is injective. to prove that (@rvl)* is an isomorphism it is enough to show that the entire tangent space to 9 at . I ~ . r(s) is the composition of @[vl with a path in SO(m) x St-' x IF["] and so its derivative at s = 0 is a tangent vector in the image of (@["I)*..the . for any skew-symmetric matrix A determining a vector in 7 u ( S O ( m ) ) . x } . It now suffices. . . 0 ) diag[I. where w = (A-IAA + A-'D + B~ B ~ ) . .l)). S is determined up to a right diagonal factor D E D and the first m rows of T are determined up to a same left diagonal factor D. . . .. : 0 ) TI = S2 ( diag{xl. . Since S (diag{xl. Hence. we may re-write this coordinate matrix as UAWV. .m)."I is well-defined. any diagonal matrix D with tr(AD) = 0 determining a vector in 7~(s"-'). then D = I so that Sl = S2 and [ T I ]= [Tz]. ?}T = S ( diag{xl. and any skew-symmetric matrix determining a horizontal tangent vector in l"(SO(k .~ } T ~ = diag{l. . For that purpose we shall realise elements of %(Si) as the tangent vectors at X to curves F(s) in Xi starting at X and lying in the image of @["I. Since A # 0 and so 12-' exists. we consider the curve r ( s ) on Sk given by F(s) = UeSA(cos s A + sins D - 0 ) esBV. . X lies in its image. to show that the derivative (@. it follows that @. 0)T for any ? in SO(k . So. x } . . ? } T I for some ? in SO(k . .4) Since 5 is horizontal with respect to the projection of SO(k . (7.THE METRIC RE-EXPRESSED THROUGH NATURAL LOCAL VECTOR FIELDS 141 Proof. x} entries xi are uniquely determined.I . if we know further that [ T I ] and [T2]are both in IF. (7. and its image is open and contains X . Thus. Since Xk has the same dimension as the domain of @["I. that derivative has a standard coordinate matrix where B = (B1 B z ) . . if Sl (diag(x1.1) onto V ~ . . since they are the pseudo-singular values of Y .

1 < i < m.1) comprising those W for which tr(A2Wo) = 0 has the same dimension as G(Sk). k . B I . this completes the proof that (@["I)* is surjective and hence an isomorphism. k . W1 = K ' A A B1. we define the skewsymmetric matrices A = ( a i j ) and B I = ( b i j )by Then. E W2 and W I is m x m with all W2 its diagonal entries zero. we have + tr(X2') = tr(U ( A 0 ) VV'W'AU'} = tr(( A2 0 ) W'} = tr(A2Wo). . . a second diagonal matrix D such . For. . we have tr(AD) = 0 as specified for (7.4) and an arbitrary choice of W2 is obviously realised by the same choice of B2. If we then separate out the diagonal part of W by writing W = ( W O W . We now claim that. if we define D = AWO. so that the condition for Z to represent a tangent vector at X becomes tr(A2WO) = 0. W 2 ). the resulting matrix W given by (7.1) satisfying this tangency condition. For Lemma 7. Finally. where W OE WO. B2 and D vary. for our next result we need to identify a specific basis of 5 ( S m p ' ) for 5 A E St-'. so that all choices of W1 can be realised by the appropriate choice of A and B1 in (7. as A .5) ranges over the subset of M ( m .5).6) . Since the dimension of the subspace of M ( m . the standard coordinate matrix 2 of an arbitrary tangent vector to Sk at X may also be expressed in the form Z = U A W V by taking W = A-'U'ZV'. x in Rm. and so completes the proof of the lemma. regarding the diagonal entries in A as standard coordinates (XI.142 THE RIEMANNIAN STRUCTURE OF SHAPE SPACES On the other hand. ) that tr(AD) = 0 would define a tangent vector to Sr-' at A in the usual way.2 it was sufficient to observe that. However. . To that end we change to the coordinates + From the inverse transformation we find i a a x i a (7. for any m x m matrix W1 = ( w i j ) with wii = 0.

6) implies that the diagonal matrix D in (7.THE METRIC RE-EXPRESSED THROUGH NATURAL LOCAL VECTOR FIELDS 143 which."] with the direct sum of the three tangent spaces and then Lemma 7.at A by aYL &.=-. Lemma 7 3 For X in XL. the standard coordinate matrix of the vector a.. for any X = U ( A 0 ) V in X k . We now look at the relation between this basis and the basis {<.2). -1 of the horizontal subspace at X . of course. in (7. Proof. As we shall only use the coordinates hi when they are the diagonal entries of A E S Z . whose standard coordinate matrices are the corresponding Zii and Zij given by (7. We note first that Equation (7. since the corresponding integral curves with y1 = 1 lie on Sm-'. so that h = Cii as stated.Henceforth. We may. : 1 < i < m. it will be convenient to denote these tangent vectors 2.2). the images under (@["I)* of the tangent vectors form a basis of G(S2).2 implies that.1} . . we may take as a basis for 7*(Sm-').. i < j < k . we shall follow the usual convention and omit explicit mention of (@["I)* and (no)* when the context makes it clear which particular vectors are involved.5) corresponding to for 1 < i < m has diagonal coordinates & Hence. identify the tangent space to the product SO(m) x Sr-' x IF. <..' . & will be the matrix Zii . --I1 a ahi 1<i<m. : 1 < i -11 < m } U { <J .

2) as U A { A . since for 1 < i < m < j < k .1. For simplicity. so that A is invertible. any vertical vectors at X correspond to those in Tu(SO(m)).4) and (7. 0 ) V are the standard coordinate matrices of the horizontal and vertical components of the tangent vector with the standard coordinate matrix U A ( C 0 ) V . for 1 6 i < j 6 m we need to consider the special case -1 J D = 0 and B2 = 0 of (7. Recall that A and B1 are both skewsymmetric and. it follows that for 1 6 i 6 m < j < k .~ } .' A A B1 = C = (cij) and write cfj and cyj for the components of the matrices Ch and C. since for any X in Xk we have h. ( V ) . corresponding to A = 0.1 the matrix Zij gives the standard coordinates of tijlki.A ~ A A . since c7j = csi from Lemma 7. . respectively. determine tangent vectors in 'Tu(SO(m))and 'Tv(SO(k. its projection onto the horizontal subspace at X . this may be further rewritten as + ux= {un(A 0 ) v : At = . which. the standard coordinate matrix of the horizontal component of & j . 1 6 i < j < m.4). so that the <.cyi and. is that which corresponds to Since the corresponding cyj is 2h. respectively. rn .so the vertical subspace at X is VX = { U A ( A 0 ) V : A ' = -A}. we have h?cyj = -h.5). since U A = (UAUt)U and UAU' is also skew-symmetric.1 we may re-write Zij in (7. -'J Finally. does agree with the expression for the vertical subspace used in Chapter 6.144 THE RIEMANNIAN STRUCTURE OF SHAPE SPACES To determine the <. the standard coordinate matrix of its vertical component is that corresponding to + Thus. via the fields qij(U) and t i . Clearly. Thus.' E i j } V and since the vector field t i j has integral curve F(s) at X .. is the basic skewsymmetric matrix S i j defined above. we write A . given by (7. we deduce that Since & j corresponds to the case in which A = 0 and B . we note that. . Moreover.l))./(h. # 0. where U A ( C h 0 ) V and U A ( C.. D = 0 and B = E i j . hs)Sij. are as stated.

This component is -1 <. since their horizontal components together 1 < i < m.2 also provides a diffeomorphism "VI of x F[VI onto a neighbourhood of n ( X ) .. we could and shall denote by 1 . which. However. n * ( ( i z j z ) ) j ) are equal to (Ziljl. and the image under (*["])* of the natural vector fields : 1 < i < m and { & j : 1 < i < m.(cij) for 1 < i < j < in remain orthogonal to those in the rest of the basis and to each other.7) whose standard coordinate matrix is Since. which can then be seen by continuity to hold throughout the non-singular part of shape space.i< j<k- form a basis of T&. Zizj2). for X in X. are also not vertical. Nevertheless.1 are horizontal tangent vectors and so are mapped isometrically onto their images in the tangent space to shape space at n ( X ) . and then ( i j will be vertical at X and hence n*(&j) will be zero. although the vectors n. for 1 < i < j < m.(x) (n(X:)). for example. Firstly. it is not available outside n(X:). l<i<rn. form a basis of with the vectors the horizontal subspace at X . the image vectors &. since for X in the remainder of S$\D. hi and hj are never zero. which. ) . and & j . h i and h. we see are given by Lemma 7. i < j < k . Lemma 7.1. it is the horizontal component of & j that is mapped isometrically onto its image. from the proof of Theorem 7. 1 < i < m < j < k . .(n(X:)).3 tells us that the & sij (7. be indices i and j such that hi = h. it follows that ( i j is never horizontal : : for i and j in the latter range.-z there could. 1 are horizontal and the qij are vertical. their norms differ from those of t i j since their inner products ( x * ( &.1) on S t .~ . Secondly. with our previous convention of omitting explicit reference to the relevant linear map.THE METRIC RE-EXPRESSED THROUGH NATURAL LOCAL VECTOR FIELDS 145 Since the for 1 < i < m and & j for 1 < i < m < j < k . We should note two points about this choice of basis. However.1. are unequal for X in X and so the ( i . Then. it will often be possible to use this basis to obtain results on n ( X .' x FLV]gives us St-' {& another basis of I.

Re-interpreting the basis (7.). where @. +A. for the remaining & j . have the following relation between our two bases of the tangent space to shape space at n(X).1. a ahi 1 < i < m. for i > j in expression (7.( --rJ for 1 < i < j < m and h. for 1 < i = j . for 1 < i < m < j < k .. 1 < i m.' ) @ ? v l ( [ F .j. This throws further . This will (&) cause no confusion as far as the A. .10) Note that. and n . ( & j ( V ) ) in (7..10) for the metric on n(Xl).@. i=l j=m+l m k-l (7.. n.J (&. then obtain exactly the correwe sponding expression for the Riemannian metric on n(X. the Riemannian metric can be expressed as dp2 = C dhf + C -dhi i=2 i=2 m @. Since the Riemannian metric on a Riemannian manifold with local coordinates x. e). that is. dht are concerned. where g. On n(X. < m. For n(X) in n(Xi). 1 6 i m < j < k .146 THE RIEMANNIAN STRUCTURE OF SHAPE SPACES the same symbols.1).1. Since n. m@.@. .9) for that basis of z ( ~ l ( n ( X iLemma 7. these are precisely the vectors n .8). or the &. = the next theorem follows from Theorem 7.. if we choose j < m and put hi = 0.dx.o(@[v])* restricted to 7 ~ ( S r . writing (7.3 and the above remarks imply that we )). Theorem 7. their inner products will depend on whether they represent tangent vectors on Xl or on n(Xi). ~ l ) is just ( ~ [ v I ) .1. It follows that the dual 1-forms of n( ) are dh.J are the dual 1-forms of ClJ on SO(k . < < Corollary 7.j + C C a.dx. However.1 . and hence dhi = 0. may be expressed in terms of those coordinates as ds2 = C g.(gii)= -.2.8) in this way.").

". What it means for the lower stratum to be isometric with X : / L ~ so locally isometric and with C: is that. so the vectors {& : 1 < i <m do not form part of an orthogonal basis of I. . For that we require the determinant of g in (7.+h. which is equal to l/A:. we may combine these two approaches to obtain the following result.10) to obtain the geometric invariants of shape space..-. Alternatively. so the metric of Ck approaches that of C:.3). dAj+l) and is equal to 1 -A. A +A: : ~ 2 ~. The Laplace-Beltrami operator on a Riemannian manifold may be given locally in one of two ways. we shall be using formula (7. 3 (7.j Since the subspace spanned by : 1 < i < m is orthogonal to that spanned by { & j: 1 < i < m. as we approach that stratum by letting hi + 0 for i > j .+l is.).3.>) at n ( X ) in ~(2.11).11) Azh.. 1 - vvivi} 9 where V is the Levi-Civith connection determined by the metric.1) matrix of the inner products g 1 J . i < j < k .. Nevertheless. (gij) = -h2Am -A3Am -a2k3 . ( ") ah.THE METRIC RE-EXPRESSED THROUGH NATURAL LOCAL VECTOR FIELDS 147 light on the stratification of shape space that we described in Section 6. h.9) and (7. A3Am . we use it to find the Laplace-Beltrami operator. on shape space. which will enable us to study Brownian motion. it is X ( d e t ( g ) ) ..(x.. because of the simple relation between the bases (7. First. The ( m .101.' / axi 2a i.1) x (m. which gives the corresponding inner products g'j = (dAi+l.-3. For an orthonormal basis of vector fields 2ri it may be expressed as C {u. for local coordinates xi.l}.(Xk\n(D.+l by (7.. and the inverse of g. -h2h.ah. {& } .

the Laplace-Beltrami operator can be expressed as Similarly. nr=ln51i+l ny=l Corollary 7.i j defines a constant multiple of the usual @ invariant measure on V ~ . depends only on [ V ] and is invariant relative to the transformations [ V ]H T [ V ]for T in SO(m) and [ V ]H [V]Tfor T in S O ( k . which we shall use to describe probability measures on shape space.4.I(cf. I f n ( X ) is a non-singular point of shape space. we may compute the volume element. for local coordinates xi.2. it follows that the volume element of the fibre n-l ( n ( X ) ) for any n ( X ) in n(X:) is . .3. Hence. Muirhead. n?Ii+.1). the volume element on a Riemannian manifold may be given either as the product of the forms dual to an orthonormal basis of vector fields or. then the volume of for some positive constant c i .~ We may also find the volume element and hence the volume of the fibre on the pre-shape sphere above any non-singular point of shape space.148 THE RIEMANNIAN STRUCTURE OF SHAPE SPACES Corollary 7. The volume element of Xi\n(Dm-2) on n(Xi)is Note that the product &. Corollary 7. On n(X:). as (det(g))’I2 times the product of forms d x i . Thus. 1982).

This apparent plethora of information was encapsulated by Riemann in a single tensor having four. Riemannian metric can also be expressed as the A useful observation is that (7. the situation is more complex but still well-structured. when convenient. (7. but with the factor (1 replaced by ( 1 &I2. vary from point to point on the surface. On n(Xk).5. by continuity. Then. lead to the following result. its curvature at that point will be equal to this sectional curvature. if it is none of these. with the remarkable property that. + Corollary 7. Then. on substitution into (7.THE RIEMANNIAN CURVATURE TENSOR 149 where the $ i j are the dual I-forms of ylij on SO(m).3 THE RIEMANNIAN CURVATURE TENSOR The curvature of a two-dimensional surface is intimately bound up with its shape. if necessary. measuring whether.13) can be replaced by A . for any surface embedded in the manifold with its tangent space spanned by those two vectors. saddle-shaped or. rather than just two. . Note that when m = 2 we have = 1. it is convex. and to what extent. It is an intrinsic property of the surface that does not depend on how it may be embedded isometrically in any Euclidean space or higher dimensional manifold. i < I . This information will. flat. the sectional curvature. concave. Although we shall find it convenient to work with the non-orthogonal tangent vectors on Sr-'. and in the general case the metric can be expressed as in Corollary 7. For that higher dimensional manifold itself. of course. and so the expression for the volume of the fibre holds for n ( X ) in n & .hi .10). it is possible. to obtain an orthogonal basis by using the customary stereographic coordinates ui on the unit sphere. '-1+hl ~ and hi = A ~ hi+] . 1 +hl l<i<m-I.5.13) is the case 1 = 1. These are given by the formulae (7. . + + 7.13) which. it is also () true for any non-singular point. so that in this case the fibre has constant volume everywhere. For each two linearly independent tangent vectors at a point there is an invariant.

Rather than working directly from the definition of shape space itself. it follows that all the sectional curvatures of each shape space are greater than. To state O’Neill’s theorem (cf.G2Gl f However. G21 f = 6162f . 6 2 1 denotes the bracket of the vector fields 5. V l ] and 621 = 0. the sectional curvatures never decrease. In our case the ‘top’ space is the pre-shape sphere of radius one.and 6 that acts on a smooth function f as 2 [ G I . the 6 are i i i horizontal and n. We shall also find that some of the sectional curvatures. while.150 THE RIEMANNIAN STRUCTURE OF SHAPE SPACES arguments.9) form a basis of the tangent space to ~ ( 2 ’ at) n ( X ) and that the horizontal lifts of these vectors are those of the same name for : 1 < i < m and for {&. &] tij (7. Recall that the vectors in (7. and so also the scalar curvature of shape space. we also know that . we shall in ti. and are determined by. So. The full set of sectional curvatures determine. unity. since S i has all its sectional curvatures everywhere equal to unity. as the measure of scalar curvature is normalised so that a unit sphere in any Euclidean space has all its sectional curvatures equal to unity. respectively. In particular.(6i) = u . or equal to. the non-zero brackets of the pairs of horizontal vector fields from our chosen basis are determined by [&. gi.I}. & j or 1 Since [61. = -[62. for the remaining {& vectors. O’Neill.7). . which in turn determines other standard measures of curvature.-2) and their horizonal lifts 6 on S i . we shall employ a famous theorem of O’Neill that expresses how the curvature changes when passing from the ‘top’ space to the ‘bottom’ space of a Riemannian submersion.Then O’Neill’s theorem tells us that the Riemannian curvature tensors Rsh and R c t on the pre-shape sphere and on shape space. 1966) we consider vector fields u . : 1 < i < m < j < k .. For simplicity. on C:\n(D. this Riemannian curvature tensor. they are the horizontal components by denote all the horizontal lifts. such as the Ricci and scalar curvatures. i 1 < i < 4. Thus.. are related by where uv denotes the vertical component of the vector u and [GI. tend to infinity as we approach a singularity.

.)2 I if 1 < i and 1 = j . + h. Thus.A? (a. otherwise. the non-zero vertical components of the brackets of the pairs of horizontal vector fields from the basis are determined by if i = 1 and 1 < 1 = j < m .THE RIEMANNIAN CURVATURE TENSOR 151 ifi=landl=j. . A.

In what follows we compute the sectional curvatures and the Ricci curvature tensor. by continuity. J I< . 16 i l < j 6 m and 16 i 2 # il.which. .152 [<.3 and the fact that --I J the curvature tensor is linear in all four arguments. we obtain expressions for the curvature on ~ ( 2 .. remain valid throughout Ek\n(Dm-2) ). (7.-~ .]" - 1J2 Jzl l<il <j<rnandl<i2#j<m.. I THE RIEMANNIAN STRUCTURE OF SHAPE SPACES lJ1 3 t. terms of the above choice of basis vectors.15). using the relations given in Lemma 7.-2. recalling that the former determine the Riemannian curvature tensor completely.. j Gm. .\DD.14) and (7.-2) associated with the . together with Theorem 7. Theorem 73 All the sectional curvatures K of C. distinct pairs of vectorsfrom the basis (7. form a basis of the horizontal subspace throughout --rJ k S.3)are unity save for 3 i f i = 1 an d 1 < 1 = j < m.]" = -[<... allow us to compute the Riemannian curvature tensor on ~ ( 2 in) . . the (. If we then re-interpret this in terms of the alternative basis of the image of the horizontal fields <. unlike the t i j .2.15) Formulae (7.. since.\n(D.

1 of Ck.. n*(<.. j l 6 m.1.12).. + Theorem 7 4 The components Ric(n. while the remaining non-negative term represents the effect of the submersion. For those. where we recall that we write dk.(<. n*(<. By definition. . where we always assume that 1 < s for -1 s when it is not specijied. . ).m(m 1)/2 . are all zero tensor on C:\n(D. the pre-shape sphere.. given by . In the case of the Ricci curvature tensor we obtain the following result.9). the theorem is true on n(Xk)and so on E:\n(Dm-2) by continuity. the sectional curvature K ( q . n.)) of the Ricci curvature .(<.. Hence. by the corresponding members of the basis (7.3).-2). except for the following: Ric(n.THE RIEMANNIAN CURVATURE TENSOR 153 1 6 i < j l 6 m and 1 < j 2 # i . < Proof. ).). Thus.3 represents the contribution from the top space. we may compute K(n. the results stated in the theorem follow from (7. the sectional curvature of the tangent 2-plane spanned by vectors v1 and 212 is which by (7. for the dimension mk .) -13 Corollary 7.(<.(<.)) 1 < i 6 m is given by for -1 I -I1 . Note that the summand ‘1’ in each of the formulae in Theorem 7.15) and the equality (7.(<. plane spanned by v1 and v2. . v2) is independent of the basis chosen for the .14) simplifies to Now. -I131 --12/2 associated with the vectorjelds in (7.)) by --1111 -1232 replacing the n.

.) --‘J -lJ <m < j <k - 1 is given b y Proof.14).) I < m is given b y Ric(n. Then.(<. ) ) f o r 1 < i < j --‘J --rJ < m is given b y Ric(n. . that the non-zero components of the Ricci curvature tensor on n(Xi)associated with this basis are . at each point n(X) of Ek\n(Drn-2).154 -1 THE RIEMANNIAN STRUCTURE OF SHAPE SPACES -JJ Ric(n.).)).. to be the symmetric (0. we find. ) f o r 1 < i # j . replacing V I and 712 by various members of our basis. e.9). n.).(<. On n(XL)we choose { e i } to be the basis (7.2)-tensor given by where the ei are d: independent vectors on ‘T&(C~\n(Drn-2)) and (g:j) is the inverse matrix to the matrix of inner products ((ei. the above formula becomes where g‘j = (dhi+l.).(<.(<. T*(<.) f o r 1 < i . n. .d h j + l ) . using (7. The Ricci curvature tensor Ric is defined.(<. Now.

. .. 1 <i <m < j <k - 1.are given by . the result follows from Theorem 7. The scalar curvature Scal o Xh\n(DmP2) at n(X) is f Scal(n(X>> d k ( d k . 1 Ric(n*(<.).12). that is. the stated result is true on n(2:) by Equations (7. = hi Ric([.THE RIEMANNIAN CURVATURE TENSOR 155 +4 3 C l<l<s<m ~ ~ t l s ~ ~ .(<.15) and (7. .1) = Proof. j<m.). t i s l v ) 9 < i < m < j < k .1. Ric(tij9 t i j ) = (d: .4 after some computation. n*(<.+l). dh. n(<. However. ) non-zero components of the Ricci curvature the tensor associated with n*(<.)) --2J -1J &j). tisIv3[ t i j . That the scalar curvature is given by (7. t1slv)3 1 ~ 1<i < j < m.16) follows immediately from the fact that the scalar curvature is the metric contraction of the Ricci curvature tensor. . Corollary 7.i 7j1s 9 [ t i j .l>IItijIt2 +4 3 I<l<m<s<k-l C 1 11tls11-* x ( [ t i j .* ( [ t t .6. We obtain the theorem by continuity.j..)) -11 -JJ 1<i. the linearity of the Ricci curvature tensor with respect to each of its variables implies that on ~ ( 2 .) -1 I Ric(n. = Ric . it is equal to where gfJ = (dh.+l. Then. Thus.

~ )as n(X)approaches the singularity set n(Dm-2). 3 h.). Scal(n(X))/dk (dk .)@. certain sectional curvatures and certain components of the Ricci curvature tensor all tend to infinity as 0 (p(n(X). we have average sectional curvature at n(X) = 1 3 1 + 2(h. + : + A. and in the components of the Ricci and sectional curvature tensors that tend to infinity fastest.2 ) ) ) . + 10 (hi + A. ~ ( D .Thus.h: + h:h. However. we see that scalar curvature.n ( ~ .’ .l h i } . when m = 3 and k = 4.I. The remaining sectional curvatures and components of the Ricci curvature tensor either remain bounded or tend to infinity more slowly than this. we may now be more precise. as we saw in Chapter 6. + hm-l 2 + 1 = sin2(p(n(X). However. In the simplest case. where p is the Riemannian distance on Xk. Since the pseudo-singular values satisfy hl 3 .z ) ) . that the scalar curvature and certain of the sectional and Ricci curvatures tend to infinity as we approach a singularity._l 3 Ih.156 THE RIEMANNIAN STRUCTURE OF SHAPE SPACES The scalar curvature is equal to the sum of the sectional curvatures associated with distinct pairs of vectors from an orthogonal basis.h. Ricci and scalar curvatures that we have obtained above confirm our claim. . + h. . the dominant term in the scalar curvature. is a multiple of ( h i . .1) is a measure of the ‘average’ sectional curvature at n(X). at the beginning of this section. . .) @ ) A - The explicit expressions for the sectional. . Thus.

we now turn our attention to probabilistic topics that arise in the study of shape and naturally involve the structure of shape spaces. . G. . [ p ] can be uniquely determined. .). LE Copyright 0 1999 by John Wiley & Sons Ltd CHAPTER 8 Induced Shape-Measures Having described the topological and geometric properties of shape spaces. .Shape and Shape Theory D. . x. We then remove the effect of scaling by dividing {tr(. then the shape of the configuration with labelled random vertices x:. . .*) that satisfies the above condition.* is the by centroid of the xi*. k labelled random variables such that x). k ) E Rmk represents a configuration in Rm with k labelled vertices. we have that d[pI(n(X)) = 1 11 SO(") R+ IWm dP(x.. KENDALL & D. is continuous. P(X* 1 = . .?z)}1/2 to obtain the pre-shape X. . In this chapter we shall discuss probability measures on shape spaces induced by various types of distribution of the vertices that determine the shape. we write f for any bounded measurable function from Ek to R..x. . . Suppose now that p is a joint distribution of (x. To obtain the induced measure [ p ] on EL.with the quotient topology. = xk*) = 0. C A R " & H. Since the composite map from Rmk onto the shape space Xk. Thus. if X* = (x: . . . up to a set of measure zero. . . by Hence. . then we first remove the effects of translation via our standard orthogonal projection Q k E SO(k) such that X*Qk = ( a x . x. Then. . . . . * where x. it is measurable with respect to the corresponding Bore1 a-fields. . xk*h where the integration over R" is related to the procedure for quotienting out translations and that over R' is related to the procedure for normalising the scale. BARDEN & T. K.) E M ( m . We shall follow our usual notation for pre-shape and shape coordinates. . . x in Rm are . if xT. Thus.>. x$ will be a random variable in and distributed there according to the measure induced by the joint distribution of (x:.

158 INDUCED SHAPE-MEASURES However. the different expressions are always equivalent. . For different p we usually need to apply different methods to quotient out these transformations or to use different coordinates on Ek to perform the integrations. in general. of course. provides a coordinate system on the corresponding subset of Z:. If we write z1 = roe"o and zJ+l = r. . ~ by the first m rows of V. .' . although. As usual. when m = 2. We recall that. z z ) . and z. . m(3)). if r is induced by the transposition that interchanges the first two labels and fixes the third.3) . . we have the alternative represention of configurations of k labelled vertices in R2 by k-dimensional row vectors z* = (zT. . < < + This planar parameterisation of Z: also allows us to give a simple description of the isometry m induced by a permutation. 2 . = (<I.yJel(oo+ol). (8.1 GEOMETRIC PRELIMINARIES As in Chapter 7. we obtain a ( k . whose complement has volume measure zero. to measure its rotation from a standard position with z1 along the positive real axis. and [ V ]for the projection of V E SO(k . zk-1) in Q k . . m is obviously given by In particular. instead of by 2 x k real matrices. In terms of the coordinate (. . equivalently. We also write (@ij) for the matrix V dXb of the l-forms dual to the right-invariant vector fields VaXb on SO(k . we denote the shape of z* by n(z).I)-dimensional row vector z = (z1.2 ) . we write U ( A 0)V for a pseudo-singular values decomposition of X . on the subset of Ck-' determined by ZI# 0. 3 ) ++ (m(l). 1 6 j 6k . then. after the standard real-orthogonal transformation to quotient out translations. . in Q k . (1. .1) onto V ~ . where = r. ro can be used to represent the scale factor of the labelled configuration in R2 and 0. [V]) when the hi are all different. of the labels or. by a simultaneous re-ordering of the vertices of the triangles. then = -4. When m = 2 and k = 3 and when the first two vertices z. the shape coordinate = z2/z1 = x i y is given by <.2. these integrations are not easy to carry out. in which case the resulting expression for [ p ] will depend on that choice of coordinates.Ithat is represented .1). Thus. Then. < k . . 8.eLOJ. m(2). of a labelled triangle are distinct. so that n ( X ) has local coordinates (A.

Returning to the general case. the corresponding shape-measure is k! times the shape-measure induced by the same k labelled iid random variables on R". However. dvs. Thus. then the two corresponding shape-measures would be the same. we have.(X>.3 gives us an explicit local expression for dv% that is valid away from a set of measure zero.(X). in the case m = 2 and k = 3 . a Riemannian submersion. we write d w for the volume measure of a Riemannian manifold M. Then Corollary 7. if we choose with equal probability l / k ! a label for each set of k unlabelled points in R". part. on the non-singular : . . E Sk Hence. Finally. since each unlabelled k-ad determines k! shapes of labelled vertices. at X \ Dm-2. nor is it invariant under isornetries with respect to the metric on the plane induced from the Riemannian metric on Xi. Since the projection n from s".4 implies that.. whose coordinates are 00. Corollary 7. in the case that a random shape on Xi is generated by k unlabelled iid random variables on R". if we are using the planar representation of the punctured version of X. to Xi is. (XI = (n(X>> dun-1 (. is the induced volume measure duck normalised so that the total volume is unity. we see that the induced isometry 5 is given by the Mobius transformation The full group of six permutations of the vertex labels induces an isomorphic group Q of isometries.GEOMETRIC PRELIMINARIES 159 Since the cyclic permutation of the labels permutes the shapes of the three degenerate triangles with two coincident vertices. we note that... away from the singularities. We shall denote by do& the uniform measure on E which. on Sk \ Dm-2. is invariant with respect to any isometric transformations on Xi and we shall normally consider Radon-Nikodpm derivatives of induced shape-measures with respect to do:. do. respectively. l/& and -l/a. However. This is a natural generalisation of the Euclidean case.we shall also consider Radon-Nikodgm derivatives with respect to d< = dx dy.The precise relation between dwi and d< will be discussed in the next section. generated by 5 and t in the usual way. but d{ is not a constant multiple of the volume measure in the plane induced by du.

Z2)random variables when the components z* of z* are iid N(p*.1. the joint 1 J distribution of zJ can be expressed in terms of z1 = roe'Hnand TJ = zJ+l/zl = rJelsi as . The set of squares of the pseudo-singular values of X is invariant under the group O(k .' )ds dusn..5). the result follows from (8. This corollary suggests a method for obtaining an expression for the uniform measure on Ck.1 iid Gaussian N(0.1) acting on the right of X. dfk-1 = s ~ ( ~ . Proof. x$ are k labelled iid Gaussian N(p*. Corollary 8. Since EL. on the set z1 # 0. with respect to the measure dfl . However. .a2Z. If x:. using the complex representation of configurations on R2. and hence the shape-measure on Zk. up to a normalising constant. The shape-measure on induced by k labelled iid isotropic Gaussian distributions on Rim coincides with its uniform measure if and only if m = 2... a2Z. . ) in Rim.I I l2 + tr(X'X). . induced by k labelled iid Gaussian distributions on R" is invariant under the group of isometries induced by O(k . 2 ) random variables.1 columns Xi of 2 are k . . . Thus. The Radon-Nikodjm derivative at n(X). we have the following corollary. the k . -' k ' the density of the joint distribution of 2 is a function of the squared size s2 = s ( X * ) ~ tr@Z). Thus. then since random variables tr{( X * ) ' X * } = tr{(X*Qk)'X*Qk} = *1.1) acting on the right of elements of s.1. = 1.1.) random variables.160 INDUCED SHAPE-MEASURES 8. and so the induced measure of X on S is the normalised = : volume measure of Sk. of the shape-measure on CL induced by k labelled iid isotropic Gaussian distributions on Rim is. .2 THE SHAPE-MEASURE ON Z i INDUCED BY k LABELLED iid ISOTROPIC GAUSSIAN DISTRIBUTIONS ON Rim Theorem 8.with respect to d w i . The real-orthogonality of our standard transformation from z * to z implies that the components zJ of z are iid N(O. given by Theorem 8.

is isometric with the quotient of Cf by the reflection. which. (ii) its projection on the collinearity set is uniformly distributed there. only differs from dv.2. n(Dm-l )). use the normalised Riemannian volume element to give a measure over ~ k . Proof. the local expression of the uniform measure on the collinearity set of Ck. < Theorem 8.kzinduced by k labelled iid isotropic Gaussian random variables in R2. Thus. l n(X) E C t is generated by k labelled iid isotropic Gaussian f random variables in R2. . up to a normalising constant. However.7) + 1<12)2 dw. and with n(X("-')). We shall consider the case when n(X)E C i is generated by k labelled iid isotropic Gaussian points.: by a constant.SHAPE-MEASURES VIA ISOTROPIC GAUSSIAN DISTRIBUTIONS ON [Wm 161 By integrating out z1 we obtain. (iii) its distance to. It follows from Corollary 7. since we get the same measure on C. the projection of n ( X ) onto n ( D m . 11.6. (8. The next two theorems involve the set n(Dm-l) of shapes lying in hyperplanes. we see that the induced joint probability measure of the distance of n(X)from the collinearity set and its projection onto it can be obtained by integrating out . the Riemannian distance from the shape point n ( X ) to n(Dm-l). = d<. is the collinearity set. when m = 2. the expression for the uniform measure on the subset of E!j determined by z1 # 0: < In particular. and its projection on. can be expressed via a pseudo-singular values decomposition of X as By Corollary 8. . in terms of the coordinates = (<I. . this is also the shape-measure on C. alternatively. where = 22/21 as before. Using the expression for the projection of n ( X ) on the collinearity set in terms of a pseudo-singular values decomposition of X obtained in Theorem 6.l ) . certainly when m = 2. They are concerned with the induced joint distribution of sm-l ( n ( X ) ) = p ( n ( X ) . the collinearity set are independent.kz whether we use that Gaussian model or.3 that the volume measure dux. . .~ ((2 s( lX ) ) )is uniformly distributed in [0. of C. <k-2). then n (i) ~ i n ~ .1. This is less artificial than might at first be supposed. q 5 l j gives the volume measure of Cf = Sk-2 and the collinearity set of Ck. on the subset of C: determined by n(1 z1 # 0.

by n:il It follows that the distance of n ( X ) from the collinearity set and its projection onto it are independent. away from the singularity set of Zk. up to another constant.162 INDUCED SHAPE-MEASURES q52j in the above expression for dv k Thus. and the result of Theorem 8. ) points can be expressed locally as where Z k is a constant. Recall that defines an invariant measure on V k . So. given in Corollary 7.l ) ) ~tan2(sm-l1) 12=1 (Sm-I ) COSk(m-l'-l (Sm-lh where. the induced joint k Radon-Nikodjm derivative.l ) . the A?-') are theJirst m .l ) ) onto it is m-l I m-1 x JJ { ( ~ j r . as before. and that the distribution of the projection is duel.3. Now n(Dm-l)is isometric with Z i . Proof.3.l)}-'. with respect to the product measure ds. up to a normalising constant. that. by Theorem 6.1 pseudo-singular values of X ( m . the distribution of SI( n ( X ) ) is given by up to the constant {2k-2(k .l ) and its projection n ( X ( m . since.-l(n(X>) of n ( X ) from r ( D m . We now return to the case of general dimension m. then. the local .~ .6.1 imply that the shape-measure on Xk induced by k labelled independent N(0. I f n ( X ) E Zk for m > 2 is generated by k labelled iid isotropic Gaussian random variables. q5ij Theorem 8. is given. s l ( n ( X ) ) = arcsin(lh21). n::. this joint probability measure c2. m .l / ~ m . The expression for duck.1 .-l x of the distance s. where [. up to some constant.-I denotes our usual isometric involution on ELpl.I .

l ) lie. when m = 2.m-l.l) is locally diffeomorphic to the product of Sm-'. so that.SHAPE-MEASURES VIA ISOTROPIC GAUSSIAN DISTRIBUTIONS ON [Wm 163 expression for the volume element of n(Dm-. ) n d h i 1=2 m nn k-1 @n. Theorem 8. for m > 2. l )) joint probability measure becomes On the other hand. the single factor (hi")' .l .3 with m replaced by m . we can find the induced Radon-Nikodym .) differs from only by a constant. the form nrIi+l i j on @ Vk-l. the space where the frame determined by the first m .1 rows of the matrix V occurring in a pseudo-singular values decomposition of X ( m .l ) lies. sm-l ( n ( X ) ) and n(X("-')) are no longer independent and also that the shape-measure of n(X("-')) is no longer uniform on n(Dm-l).?i is another constant. up to a constant since the latter is given by the formula of Corollary 7.? n . Hence. with Vk-l. this expression divided by that in the statement of the theorem is dsm-l x dwk-.+.m)-sphere bundle.'given in that theorem.m can be regarded as the product of the form @ i j on Vk-l.tan2(sl) and so this only involves the first component.1 .m * Vk-1. the induced . In particular. On substituting in this formula the expression for s ( n ( X ) ) and the coordinates of ~ T ( X ( ~ . :n m-1 ny=l .n. it follows from Theorem 6. By integrating the expression for the induced joint Radon-Nikodym derivative in Theorem 8.3 shows that.' ) ) is nr:f.k n '1 n m Ihilk-l-m i=l n I < j l <jz<m ( ~ -5h ~? .1 .m-l and the form @mj on Hence.m-1 is a ( k .1 i=l of the induced joint Radon-Nikodym derivative involve both components in a non-trivial manner. the space where the pseudo-singular values of X ( m . n( Dm.6 that the induced joint probability measure of sm-l ( n ( X ) ) and n ( X ( " . = l nz=nl+l where . locally. Also Vk-l.3 with respect to s m .tan2(sl) that occurs reduces to 1 . However. the theorem is established. The independence breaks down because the factors m. ': nrz:+.

Then. . Since the condition (a) above is satisfied with probability one. (ii) the numbers of points of the process in n disjoint Borel sets form n independent random variables. 8. . A homogeneous Poisson point process is characterised by the following two fundamental properties: + + + + (i) the number of points of the process in a bounded Borel set U has a Poisson distribution with mean equal to the volume of U multiplied by a constant i . To describe the shape-measure of tiles in a Poisson-Delaunay tessellation we need an ancillary shape variable. where the constant k is called the intensity of the process. the length of the perpendicular. even if we only know the position of the foot of the geodesic perpendicular roughly. and such tiles will be non-overlapping and will yield a complete tessellation of EX". in the light of that._.1)-dimensional sphere. we draw the simplex determined by these m 1 points if and only if their circumsphere contains no point of the set in its interior. Hence.1 This amounts to treating the position of the foot of the perpendicular as an ancillary statistic and assessing the degree of degeneracy. This distribution is smooth as a function of n(X("-') ). this enables us to get a good approximation to the conditional law of s. . From these two properties we can deduce that a realisation of an m-dimensional homogeneous Poisson point process satisfies the conditions in the previous paragraph with probability one and so we may apply the above construction to obtain the Delaunay tessellation of a realisation of an m-dimensional homogeneous Poisson point process of intensity C and study the shapes of the tiles of its .*.3 SHAPE-MEASURES ON C:+' OF POISSON-DELAUNAY TILES If we have a countably infinite discrete set of distinct points in R" such that (a) no m 1 points lie on an (rn . we know the conditional distribution of s. .8. This yields a Delaunay tile. as at the end of Section 6. then we can construct the Delaunay tessellation of such a set of points as follows. . If f: is the centroid off:.. : where E S"-'. simplexes. so that. we may. then the shape variable Q that we require is given by fr + . use the circumradius F of a simplex and its circumcentre F to write its vertices x t as xi* = F. .X.164 INDUCED SHAPE-MEASURES derivative for n(X("-')). We look at each subset of m 1 points in this set in turn.1)-dimensional hyperplane and (b) no m 2 points lie on an (m .(n(X)) given n(X("-')). We shall refer to this tessellation as a Poisson-Delaunay tessellation of intensity k. Xk+l.

m 1 < i < m. with respect to the shapemeasure induced by m 1 labelled iid isotropic Gaussian distributions on Rim. Theorem 8. . l<i<m. Let i i + = -x. i= 1 Since i=l i= 1 i=l . xk + Ax:).~ ~ m ~1mi i)l n a . of the shape-measure of tiles of a Poisson-Delaunay tessellation of intensity k on 08" is cmem2. m ~ [ i ~ ~ [ i ~1 ~i < m ~ ~x + ~ i ~ ) . where Em is the volume of the unit ball B" in 08".\ : 1 < i < m 1) + 0. and the circumsphere of x. If x:.x < . where \Ax. 1<i < m + 11 = km+l i= 1 exp(-ClAxjCJ}JAxjCI m+ 1 i= 1 as max{lax. 1987) that. the Palm distribution of homogeneous Poisson point processes implies (cf. is a constant depending only on m. The Radon-Nikodjm derivative at n ( X ) . . Thus.*I\' = tr(x'?& then Q is equal to Y I S . we have + P[ij E [iio. .x.SHAPE-MEASURES ON Cg+' OF POISSON-DELAUNAY TILES 165 Note that..+~ are m 1 Poisson points of intensity k on Rm. .".". . where c. + + Px [: E [xk. i ~ ~ ~ i=l ~ ~ However.] = i m e x p { . .x. we have. the number of Poisson points that are contained in the circumsphere of these m 1 points has a Poisson distribution with intensity kEmYm. i i o + Aii). We may now describe the shape-measure of tiles in a Poisson-Delaunay tessellation in terms of the basic shape-measure we obtained in Theorem 8. so that s2 = s ( X * ) ~ C:=. = 11x7 ..1.4.. + Proof. .\ here denotes the product of the absolute values of the components of the vector Ax:. contains no other Poisson points I x. for any fixed m 1 distinct points x$ in 08" and incremental vectors Ax. Then. Stoyan et al. for any fixed m distinct points i i o in Rm. if s = s ( X * ) is the standard measurement that we are using of the size of a simplex.

we see that. . (8. On the other hand. we use our usual notation = 23/21 = reie. = du.Theorem 8.{cos28. of the above probability is nyZl ny=ld. to . (8. . up to a constant.4. we find that < {sin201 + sin2 8 2 + sin2 1 ~ 3 1 ~ and d81 d82 = 6& cos2 4 cos2 8. after some elementary calculation. .e-msm}. as we showed in the proof of Theorem 8.?i = km exp{-2C.z. the square of the expression (8. x. + When m = 2. x. . .11).3r2 cos2 0)2 2 r sin0 -_ d<. is dug+' up to a normalising constant. then. .(x1.Tm} = Zn exp{-kC. if. . . together with (8. ./(2n2) and 22 = n. we have @ 2 = 2 . Then.121z.z2 1 . (8.11) i where the B are the three interior angles of the triangle. = s ( X * ) . with respect to dwi.11) is.). . r sin 0 d< (1 .10) q+' induced The result of the theorem follows by comparing this with the distribution on by m 1 labelled iid isotropic Gaussian distributions on R".7). . gives an alternative expression for the shape-measure of tiles of a . where (XI.166 INDUCED SHAPE-MEASURES the Radon-Nikodym derivative. the Radon-Nikodym derivative. . . on the subset determined by z1 # 0.10) and (8. in fact.1. we integrate the measure corresponding to (8. since the induced shape-measure of three labelled iid isotropic Gaussian distributions on R2 is the uniform measure on Xi. of the shape-measure of tiles of a Poisson-Delaunay tessellation on R2. (XI. lz. with respect to the measure dii..9) Now.9) with respect to s ( X * ) to get (8. . . = :{sin2 $1 + cos 282 + cos 283 J + sin2 8 2 + sin2 031. which.) E %+'. Since dw. X) quotient out the scaling effects.and so.

2. Similarly. Having chosen a representative region. This follows from the fact that e 6 (rn 1)1/2 with equality if and only if f: = 0.4 is a decreasing function of /IF:// that tends to zero when I I I I tends to one. Kendall.1 Reproduced with permission from D. as they are in the present context. the Radon-Nikodfm derivative of Theorem 8. 1983 . Studies in Probability and Related Topics. I IXz I I = 1 would require all X : the Xf to be coincident on the sphere. In Figure 8. Montreal. we illustrate.1 we illustrate in that manner the shape points resulting from a simulation of 5000 Poisson-Delaunay triangles. This need not alarm us since we have only defined the Radon-Nikodym derivative on the. which corresponds to the totally degenerate shape. complement of ~ T ( D ~ . in M. As explained in Chapter 1. However. pp 321-330.On the other hand.8). when labelling and sensing are of minor importance. which is precisely the formula that Miles (1970) first derived. as shown in (8. the shape space $ can be divided into twelve isometric regions such that. any one of them will suffice for illustrative purposes. points most remote from x(D. in Figure 8. However. it has been shown. Nagard.SHAPE-MEASURES ON Z ' : + OF POISSON-DELAUNAY TILES 167 Poisson-Delaunay tessellation on R2 as & sin 61 sin 62sin 83d61d02. Demetrescu and M. non-compact. the maximum is achieved at the ).-l). Losifescu (eds).G.1. as in Figure 1. Since Q is a decreasing function of l[f:ll.C. it may then be represented on a 'bell-diagram' in a way that preserves the uniform measure. at the end of + Figure 8. and so this minimum cannot be achieved. the contours for the Radon-Nikodfm derivative of the shape measure of Poisson-Delaunay triangles with respect to the uniform measure that we obtained above.. The shape of PoissonDelaunay triangles..

maximal remoteness is also necessary for X. this is not true. ) if and only if the corresponding simplex X * is regular. Studies in Probability and Related Topics.8. For example.G. that n(X) E Xt+' is maximally remote from z ( D m .l ) . The shape of PoissonDelaunay triangles. Nagard. . Losifescu (eds). and this implies that X. pp 321-330. = When m = 2. Demetrescu and M.* 0. = when m > 2.. the X corresponding to z each of the following two configurations is zero. Kendall. although not in z ( D m .1. INDUCED BY k LABELLED iid NON-ISOTROPIC GAUSSIAN DISTRIBUTIONS ON R2 For the remainder of this chapter we shall restrict ourselves to the case that m = 2 and use the complex representation of configurations that we discussed in Section 8. in M.2 Reproduced with permission from D.4 SHAPE-MEASURES ON X. However. but the shape of the first one is in n(Dm-l) that of the second one.C. for m = 3.168 INDUCED SHAPE-MEASURES Figure 8. 1983 Section 6.* 0. Montreal. is not maximally and remote from n(Dm-l): and 8.

along similar lines to the proof of Theorem 8. Tzk-1) has the same shape as z . .12) where s (n(z)) the distance from n ( z )to n(D1) and P.~ } T for some T E SO(2). The covariance matrix C can be expressed as C = T‘diag{a2.diag(a2.SHAPE-MEASURES VIA NON-ISOTROPIC GAUSSIAN DISTRIBUTIONS ON R2 169 Theorem 8. Since Tz = (Tzl. Thus. a . we see that the induced shape-measure is . Then. . a-2}. induced by k labelled iid bivariate Gaussian distributions on R2 with principal component variances a2> 1 and 1/a2is given by (8. T z j are iid N(0. that the components z j of z are iid N(0. on the set z1 # 0. (x) is the Legendre polynomial 1 is for which the Laplace ‘second’ integral representation is Proof..1. . C) random variables. o . . The Radon-Nikodjm derivative at n(z).I and so. of the shape-measure on Ck.5. a . with respect to dwk. by integrating out the redundant coordinates ro and 00 from the above. we may without loss of generality assume that the covariance matrix of z. Because of the real-orthogonality of our standard transformation from z * to z . itself has the form diag{02.2 ) )random variables. it follows. the joint distribution of z j can be written in terms of z1 and < j as where gl(a. where C is the covariance matrix of zT. b ) = a sin2(a + 6 ) + a-l cos2(a + b ) The complement of the set determined by z1 # 0 has Lebesgue measure zero on c k .

<) = -(a . sin2 <) and da h2(a.I 70 where INDUCED SHAPE-MEASURES + hl (ap'. :r: ) 1 .0-1) 3 1 2 ( + -y<.

5 achieves its minimum when f o ( n ( z ) )= 1.6) gives the stated result. the Radon-Nikodym derivative (8. = the Radon-Nikodfm derivative has its maximal value Pk-z((a a-')/2) on the collinearity set. Then.1 and h.hl(o-'.12) is Coo everywhere on X. ( n ( z ) ) = 1 if and only if s1 ( n ( z ) ) n/4: that is. ( x ) is an increasing function of x 3 1 and so the Radon-Nikodym derivative (8. if and only if n ( z ) is maximally distant from the = collinearity set. (fu It follows that gz(a. it is : asymptotically equal to P k . n/2] is such that .21 for the two pseudo-singular values of the real-valued 2 x ( k .~ ( ( a a-')/2) on any small neighbourhood of the + + + .12) in Theorem 8. which is achieved when $1 ( n ( z ) ) 0: that is. To identify this set we write h.(O ' + Equation (8. From the last equation it follows that Then.13) implies that f . = 1 and lA. since a > 1. = J(hl(o.<) h3(19<) a 0-1 2f. Thus.2 k () 0. the maximal value of f . when n(z)is in the collinearity set.SHAPE-MEASURES VIA NON-ISOTROPIC GAUSSIAN DISTRIBUTIONS ON 08' 171 and q E [0.12) has the minimal value + crzf { &T-l.1) matrix corresponding to the pre-shape z /I )zI 1. since h3(a. The Legendre polynomial P . the set consisting of all shapes z! = 0. we have where the xj and y j are the real and imaginary parts of z j .21 = sin(sl(n(z))).2 with h.1 3 lh. ( n ( z ) ) is (a a-')/2. Finally. on this set. Equation (8. -. 2h2(0. When a > 1. since h: A. . r> o > ~ <)2 + 4h2(a. Since the function given by (8. and maximally distant from the collinearity set is characterised by so. Similarly. sina. <) = h3(ar { ) P ' ) P .

. where C and C.1) acting on the right of the elements of Sk and the next result follows from Theorems 8. . - @ given in the corolWhen k = 3 . then the Riemannian distance S I(n(z)) of n(z) to the collinearity set and the projection of n(z)into it are independent a n d i ( n ( z ) )= sin(2sl (n(z))) [O. .2. is a complex normal CN((z&. . if C* = C symmetric and C .. then C is positive-definite T T is.2 ) { 2 ~ } ~ . .. Thus. . where U is a unitary k x k matrix and D is diagonal with all its diagonal elements positive. the Radon-Nikodym derivative (8.. Zfn(z) E Zt is generated by k labelled iid non-isotropic Gaussian random variables with principal component variances a2 > 1 and 1/02. . .2 and 8. .’ i{~ .12) is invariant under the group of isometries induced by SO ( k . .1)2i2)3/* + + Thus. y z ) is multivariate normal with a covariance matrix of the form + The assumption that C* is positive-definite Hermitian implies that it can be decomposed as C* = U‘DU. det(C*) = . this value gives the factor by which a degree of ellipticity 0 increases the chance of an approximate multiple alignment. xz. 11 has distribution E ( k . . Since it depends on the shape only via its Riemannian distance to the collinearity set. .” 2 ) d i . z c ) E C k ./(det(C*)) > 0.l ) 2 P . Thus.5 SHAPE-MEASURES ON Zf INDUCED BY COMPLEX NORMAL DISTRIBUTIONS Let C* be a positive definite k x k complex-valued Hermitian matrix. Pk-2(x) = x and so the distribution of i(n(z)) lary becomes 4o2(a2 1) d? . Then. z&). (402 (02 .~ 4a2 + (a2 1 ) 2 i 2 } x P k 4 ( ( a 2 + 1){402+ (a2.5. . That T iC. is skew-symmetric. where Z*J = x* + iy?. J able if (xT. Corollary 8. . . yT.172 INDUCED SHAPE-MEASURES collinearity set. are real. we have and E[i(n(z))] = 2 0 / ( 0 + 8. we say that (z. C*) random vari.

Zk-1) is still a complex normal distribution.2 ) are our usual coordinates on the subset of shape space .. . - zlzoc-l(l. . C).Zl(1.. after some calculation.with respect to dwi. j=l where (tl. . is identical with the induced Riemannian metric on Xi and K.. Zk-1) is 1 exp{-i(T-Fo)C-'(z (2~)~-ldet( C) -zo)' j=l k-2 . The Radon-Nikodjm derivative at n(z).*. say. * ) is C (8.6. . .14) where I F1 is Kummer's confluent hypergeometricfunction dejined by and r ( n ) denotes the Gummu-function. Clearly. 12(k-2)dZi. is the squared size of the configuration determined by the means. Zk-1) C N ( z 0 . = ZoC-'z.. Proof. determined by Z I # 0. We note that pc is invariant under scalar multiplication of C . and so the induced shape-measure is the product of exp { -: x (lzl12(1. t k . as the C-based squared size of the configuration determined by the means. t>'. . z$). .. . we can see. T)C-'z. C*) implies that the joint distribution of (ZI. but that KC is not.~ i ( 1 TIC-'^^) . The joint distribution of (ZI. . Theorem 8. +zoc-lz. . zl) is a complex normal distribution CN((zG. .5. . . that the assumption that the joint distribution of (z. IZI I (8.SHAPE-MEASURES VIA COMPLEX NORMAL DISTRIBUTIONS 173 Following similar arguments to those at the beginning of the proof of Theorem 8. T>C-'CI. . . . . z = (ZI. We introduce the C-based metric pc on X$ defined by - and write K. . on of the shape-measure induced by CN(z. .)}Iz112(k-2)dZ1 n dtj.15) . .

(8. for any n # 0: Hence. we can rewrite However.15) as E Q.174 INDUCED SHAPE-MEASURES with Write a = (1. and. <)' > 0 and b = (1. (8. r)C-'z. ?)C-' (1. Then.16) is equal to .

14) can be simplified accordingly. as already noted. is not defined and the Radon-Nikodfm derivative the shape of the means (8. If. given by (8. then the z. then K$ = {Fozi}/a2 and. in addition. . .*. i=l When C = 0 2 Z and the z:j are not all identical. is C*) The stated result then follows from the expression for the uniform measure dw. all the zij are identical. the zoj are all zero and so KC = 0. the Radon-Nikodfm derivative (8. are independent.1 . j 1 )distribution and ZC-'wt = a22 a j 2 Z j w j . .14) becomes cy=i ' k-1 i=l k. then (8.l ~ ' } k . It follows that the Radon-Nikodym derivative (8. pc coincides with the Riemannian metric p induced on X i from the standard metric on Si.l ' When C = diag{af.SHAPE-MEASURES VIA COMPLEX NORMAL DISTRIBUTIONS 175 we deduce that the shape-measure induced by CN(z. a:-l}.6). In this case. Thus.14) becomes . When all the zGj are identical.14) simplifies to 1 {zzf}k-l det(C) { T C . each having isotropic Gaussian N(zGj. .

) be the polar coordinates.of the projection from o of the ith vertex onto 8K. then. Z 2. the labelled triangle (Z. 11 and. of course.The method also extends naturally to the case when the number of vertices of the configuration is more than three.0.) has the same shape as that of (z?. with respect to 0.. i = max{125 I/r.j : 1 6 j 6 3 ) takes values in the interval (0. Then. but we shall not discuss that here. given that the ith vertex is on aK. 2 .z. ) . Similarly. of the projection from that vertex onto 8K of the jth vertex. 8i. z. if Zg = z ? / i .6 THE SHAPE-MEASURE ON X.*. we let (r.. respect to dw. First. z. .*. z. (roi. of three labelled iid uniform random variables (27..7. z. of K with the induced shape-measure on X. Our proof proceeds by rescaling and translating a random triangle to successively project two of its vertices onto aK in such a way that the relevant conditional distributions may be calculated..) are three labelled iid uniform random variables in K.) at z and be ’ (To.z.. we let 0. if (z:. 2 : ) in a compact convex set K in R2 can be expressed as Proof. The Radon-Nikodjm derivative m at ~ ( z ) . the interior angle of the triangle (z:. In the general case. INDUCED BY THREE LABELLED iid UNIFORM DISTRIBUTIONS IN A COMPACT CONVEX SET The method that we shall use to obtain such measures is distantly related to what is sometimes called ‘Crofton’s boundary formula’.) are its polar coordinates. Q. this gives a 2-fold integral for the Radon-NikodLm derivative of the induced shapemeasure with respect to the uniform measure on Xi.) be the polar coordinates.*. we may without loss of generality suppose that the origin is inside K... with respect to the ith vertex. Theorem 8.176 INDUCED SHAPE-MEASURES This expression was first obtained by Mardia and Dryden (1989a) using a different method. To describe our result we let aK denote the boundary and IKI the area of a compact convex subset K of R2 with non-empty interior and we let o be any fixed point in K\aK. 8. that o is at the origin and that IKI = 1. Since we are working modulo translations and scaling... When the ith vertex is already on aK.

) dZ. of (i. given that the ilth vertex lies on aK. so that each summand of the right-hand side of the above becomes f rifi. then. z.. are two iid uniform random variables in K. and almost surely distinct. j # i l } . Z. = Zi*.2.*. for given. z. z. = max(lZ. that lies on i3K is and. as and T before. of (zT.. $... If Pi.*. of the labelled triangle (Zr .=l j#i1 Since i is a scale factor. may be integrated out. the 25.ZJI/ri... i.j. Again. the joint distribution fi i= I 7 1K(Z. given that the i2th vertex is on i3K. and ? on i3K there will be : a one-to-one correspondence between iQ and the shape of the labelled triangle . the probability of the vertex that lies on i3K being the ilth one is 1/3 and. j # il.) has the same shape as that of (z.SHAPE-MEASURES VIA UNIFORM DISTRIBUTIONS 177 and lies in K with at least one vertex on i3K. the labelled triangle i. where the sum. .) and it also lies in K with at least two vertices on i3K. we may integrate it out so that each summand of the right-hand side of the above equation becomes where the point ( r o i l . Our second projection is from the vertex Z z .ddei i l l K (2: ) d.) on i3K is now the location of the vertex Zi*.. Moreover. The probability that it is the i2th vertex. ranging over all permutations of the labels. Therefore.*.).). z. Thus. in addition to the ilth vertex. deoi. Now. the remaining vertex is uniformly distributed in K. Hence (?T. r. the scale factor &. which is now known to lie on i3K. we i=l J . runs through all possible choices of two ordered vertices on i3K and the third one in K.) reduces to the joint distribution i. ?i*. Z. after two such succesive changes of scaling. i = Z5/Fil for j # il. .

the uniform measure on Ei is invariant with respect to such a transformation. and it. 3n A* -? . with respect to d<. z.18) Finally. from (8. that the Radon-Nikodym derivative GZK.2) and (8. of the shape-measure associated with (8. . we have This Radon-Nikodym derivative has exactly the same singularities as the corresponding Radon-Nikodgm derivative with respect to do: ..: (8.19) It is clear from the proof of Theorem 8. 3) H ( m ( l )u ( 2 ) . by (8. we have Thus. as stated in the theorem follows from -- 2 sin2 01 3 + sin2 02 + sin2 03 sin2 0i3 (8. the expression for the shape-measure on E.3) and (8.178 INDUCED SHAPE-MEASURES (z?. for fixed d?:3 = l2 d(m(<)) = 7Izi2 . etry.. Since E.18) and (8.19).7. 2. u ( 3 ) ) is an isom. by (8. is a homogeneous space and the transformation u on induced by the relabelling mapping (1. zi).4).7).17) at can be expressed as < where. in particular.12{ 1 + IU(<)I~}~do: 2 2 3 = ?r:iz{l + Iu(<)I 1 do.

We also have r. This result was first obtained by Small (1982). the fact that zi*.. (8. To close this section we show that the double integral in (8. the remaining two double integrals over i3K are independent of the position of n(z) within that component.i. Theorem 8. when both zi*. which is independent of the position of zrI on i3K.8. The expression (8. = 1. we have = n/2 6 i l i .17)kills four of the six terms.8 sin(20. = z has been mapped to 00.SHAPE-MEASURES VIA UNIFORM DISTRIBUTIONS 179 When n(z) is in the collinearity set and is not any one of the three special shape points corresponding to triangles with two of their three vertices coincident. passes through any one of the three special shape points. since it is the radius of the disk.) 1=1 + sin(40i) + sin4 i Proof. if we write 6 i l j 2 for the angle between the directed line from zi*. by the comments in the is previous paragraph.* = 2 sin 6 i l i 2 .17) is discontinuous as n ( z ) . that the first derivative of #ZK given by (8. In addition. The Radon-Nikod$m derivative. travelling along the collinearity set. l/&) of x-axis. Then. so that.20)also makes it clear that the value of #ZK is a constant in the interval (-l/&.20) discontinuous at the two special points rl/& on the x-axis since the special shape point . The complement of the three special shape points in the collinearity set has three components and the terms to be killed depend on which of these components contains n(z). This implies. However. to zT2 and a directed tangent at zi*.19). with respect to dw. of the shapemeasure on induced by three labelled iid uniform distributions in a disc is given by 120i . the factor 1K(Zr3(n(z))) in (8. by (8. and ri.17)is amenable to computation by evaluating the formula for the case when K is a disc.17)is proportional to It is easy to verify from this formula that the first derivative of mK in (8.. zr2 and are on aK. When K is a circular disc. and so + The condition zi*. it is convenient to take the fixed point o in K at its centre and its radius to be unity. coil) onto i3K is uniformly distributed there. assumed to be is uniformly distributed inside K implies that its projection (roil. corresponding to z. E K is equivalent to requiring that .

3 Reproduced from D. two minor local maxima at the equilateral shapes. we have the result. and six saddle points at a special isosceles shape and its images under permutations of the vertices. Figure 8.8 are illustrated in Figure 8. Advances in Applied Probability 17 (1985) 308-29. There are three major local maxima at the shapes of triangles with two coincident points.3 on a projection of the sphere S 2 ( i ) = Xi onto a central plane chosen. so that the central horizontal line is the visible part of the collinearity set with the midpoint being the shape of the degenerate isosceles triangle having its third vertex at the midpoint of the line joining the other two. Thus. - 8 sin(28il) + sin(48i3)}. The contours for the Radon-Nikodym derivative of the shape-measure obtained in Theorem 8.G.180 INDUCED SHAPE-MEASURES so that the above integral is equal to 64n Jd $8 7 sin4 8d8 = 2x1 128i. three local minima at the three degenerate triangles with one vertex at the point mid-way between the other two. Kendall. by permission of the Applied Probability Trust . Exact distribution for shapes of random trangles in convex sets.

However. For each term (8. en ordered in the counter-clockwise direction of aK and. for the first time. by Li and denote by p i j the point of intersection of Li with Lj. We also write ci for the length of the perpendicular from the fixed point o in K to ei. We first observe that each of the six terms of GZK in (8.. Those that we noted were on the collinearity set. they could appear anywhere and. oriented in the counter-clockwise sense. . at o with a positive vl-axis parallel to the direction of ei and replace the polar coordinates (q2.20) has a further decomposition into terms of the form + The arguments in the proof of Theorem 8. u1 is equal to the constant ci.7 by Cartesian coordinates ( 4 . w = pi+l .) that lie on aK are on different edges. j ) . In addition to the notation we used there. makes with the u-axis.1) of these terms for which the two vertices zT and z. u2 gives the . It will also be convenient to have the associated complex coordinate w = u iw and to write pi for the complex coordinate of the vertex common to i the edges ei-1 and ei of K. discontinuities appearing in the derivatives of the Radon-Nikodgm derivatives of shape-measures. The shape-measures concerned are a special case of those of the previous section. we now suppose that K is a convex polygon with n edges el..pi is the complex representation i of the edge ei with modulus a equal to the length of ei and argument pi equal to the angle that ei. v l ) 1. v) in the plane of K. w2) at zT with a positive v2-axis in the direction of e j . Then. Since the induced shapemeasure depends only on the shape of K itself. we study a class of shape-measures whose Radon-NikodLm derivatives have singularities along the edges of an appropriate tessellation of the planar representation of the punctured version of E. of the triangle (zT. in this and the next sections.7 THE SHAPE-MEASURE ON X. . z. so that K lies in the upper half of the (u. where i # j . v)-plane.21) associated with ( i . will be non-zero. Then. I: THE SINGULAR TESSELLATION In the previous section we observed. we make the convention that i l and i2 determine the same edge when il = i2 modulo n .7 show that only the n ( n . . . we set up Cartesian coordinates (u. which we shall use throughout. we replace the polar coordinates ( ~ ~&I ) at the fixed point o in K by Cartesian coordinates (u1. INDUCED BY THREE LABELLED iid UNIFORM DISTRIBUTIONS IN A CONVEX POLYGON. by taking the positive u-axis to start at the midpoint of el and then to run along its second half in the counter-clockwise direction of aK. d12) at z. for general integer suffixes. denote the affine line that contains e. 181 8. z.A SHAPE-MEASURE ON Z I: THE SINGULAR TESSELLATION .

the Jacobian < + < factor I 1 2 in xij(m(<)) is a real analytic function of (x. t) = . f To understand how such singularities can occur. have their third vertex + + . at p i twj. t)-coordinates o the vertices of K+(i.22) can be rewritten as + z E ei to T where f ( s . If we take the vertices z at pi swi and z.20) as = x i y varies in C = R2. t)-plane to the (u. z.182 INDUCED SHAPE-MEASURES E length of the perpendicular from z. The reason for its importance is that a necessary condition for to be a singularity of i j Z ~ ( < ) is that at least one of its summands in (8. (8. and examine the nature of the set K+(i. t) given by t qij(<)(s. to e. and so one of the summands (8.. the point z along the oriented edge ei may be represented by the complex coordinate pi + swi and z. j .(m(<)) to i j Z ~ singular a t ifand only if is the same can be said of the (s.z.) of shape <.y).< n . the triangle (z. <I = (qij(<))-'(K) of K under qij(<) is the set of those (s.with vertices zT at pi swi and z. j . the set K+(i. m(<)). However.) with shape will have its third vertex + < < < + at the point qij(<)(s.. m(<)) the intersection with the unit square of is a. and the inverse image ~ o + ( ji .20). Hence.{swi(l .23). ) away from the y special shape points ~ l / & . the integrand is a real analytic function of (s. ei onto e. Then. possibly degenerate. lie For any given convex polygon K. j . t)-values for which triangles (zT. <) and.< i # j . g(s) is a linear polymonial giving the length of the perpendicular from z. should have a singularity at <. the term (8. z. in particular. and (8. t)-plane associated with the requirement that zi(m(<)) in K.24) Clearly.21) takes the form T Now. v)-plane in which K lies. polygon whose edges vary smoothly with (x. we choose E C and i. z ( ) will play a central r<) role in our description of the singularities of i j Z ~ ( < ) in (8. twj. 2. qij(<) is an affine transformation from the (s. it will follow that the contribution xi. in terms of real coordinates. and K+(i. j . along e j by p j twj. 13. where s and t lie in the real interval [0. m(<)) the subset of the unit square in the is (s. t) is a quadratic polynomial giving the square of the distance from E ej.h<)w j ( l + &<) + pi(1- + + p j ( l + &<)I. how it varies with <.t) and we shall show that K+(i. at T p. j such that 1 . j ..

n B. for &+(i.B.. j . (cf. Being affine.(<) is an affine line. However. since K + ( i . However. and then it is the whole of [0. . and zg lie on the arc of the circle through z. j by the initial points pi and p . lies between zT and z.4(a)). <) will be the linear segment joining the vertex (0.. Figure 8. that is. K $ ( i . and the image of pi.(<) is defined for all s and t. t)-plane and. We must also have zg = p. The set K. (cf.(i. which. the angle at the third vertex z. we choose the arguments of w . j .In.B. z. Since we already know that the special shape points T l / A are singularities of &. . t)-plane with non-empty interior. <) will be a strip in the (s. j . we require z.z. respectively.1 and so p . between the arguments of wj and w.4(c)). <) = K i G .4(b)). then. It is also clear that the intersection < + < + < < . it will strictly contain the unit square. j ( < ) is singular if and only if the quotient is real or infinite. In that case. equally. j . Figure 8.l/& 1/&]. <) has a non-empty interior and the image of pi. Figure 8. of a triangle (zT.. and p.) of shape 4 . and e. <) is a compact convex polygon in the (s.. q . Now. Then. j = p . that line must meet K in an affine segment. so we have K+G. . K + ( i . after e j in the cyclic order around aK. l/&). we shall not consider them further. (<) is non-singular.. arg <+I/& wi < + where I is an integer. we see that. we must have j = i . The alternative condition corresponds to the angle subtended at by the real segment [. and L. of e. as illustrated in Figure 8. or. not on their prolongations L. 0n [O.Bi is the angle subtended at p ..z5.A SHAPE-MEASURE ON C: I: THE SINGULAR TESSELLATION 183 in K. and zg to lie on the chosen edges of K.and wj such that Bi > B j and 0 < B. . <) is non-empty.. if a triangle of shape has its third angle equal to n B.. provided the affine transformation p.5(a). (i) The case that 0 < Bi . by z. the following cases arise.. Since B.-l(<) will be an affine line meeting K only at p . if and only if E (-l/&.B j < 2n. the set K+(i. differing by a multiple of n from the difference B.. j . . In the latter case. l]’. <) to be non-empty. = arg . # B. for each in the given range. qi.is if and only if either is at one of the special shape points 1/& or F l / & w. and zi..B j < n. j .Taking e. Then. since a complex number is real if and only if its argument is an integral multiple of n. 11’.B j = n. j . j . on which p. 1 ) of the unit square with a point on one of its opposite sides. which lie on these edges (cf.. Since K is convex and the points at which this angle is subtended by z. (ii) The case that .

Then. with the edges ei and e. since there are appropriate triangles of shape arbitrarily close to the former point.1 and z.(c) all triangles shown have the same shape.. is in K if and only if j = i . In either of the above general cases. let qij be any point concyclic with z. Figures 8. j . so that KJ(i. j .5 s << 0 requires f < < 0 for z. Its image must be the affine line joining pi.1 (cf.6(a) and 8.6(b)). no longer necessarily contiguous. + + < < .. if is the shape of the triangle (z. Thus. (b) z. except that now we require i = rz + j ..4 (a) All angles are measured anti-clockwise. then vij(<) will be singular. in K of KJ(i. Figure 8.8). <) will be a strip in the (s. = pi. the unit square.B i . and zz is 3n B j . and qi. z z ) with z. but now need not contain.7 and 8. <) with the second and fourth quadrants in the (s. (iii) The case that n < . Figures 8. zz.B j < 2n. t)-plane that meets. = q i j .Bi in each case (cf. and similar results to those in the first case apply. (i) or (iii). t)-plane must be bounded (cf. the angle subtended at T q i j by zy and zz is 2n Bj . the angle subtended at p i j by z. s + 0 if and only iff+ 1 Figure 8. Then.5(b)). lies between p i j and qij. zz and pij such that each of z and z.184 INDUCED SHAPE-MEASURES (4 (b) (C) Figure 8.

= pij (a) (b) Figure 8. II.p l ] / w l is real. t ) . The convex polygon K is bounded by the n affine lines Li = { P ~ + u w u E EX]. ( < ) ( s . t ) lies on Ll if and only if { q i j ( ( ) ( s .. and q ~ j . z t .. I: THE SINGULAR TESSELLATION 185 \ lei Figure 8.7 In (b) all the dashed triangles have the same shape as (z.6 zg is in K if and only if i = j +n - 1 and z. That is. + where .A SHAPE-MEASURE ON C.. j . < ) . zz) We now examine the boundary of KJ(i.25) U i l ( 0 s + ujl ((1 t w i j l ( 0 = 0. . if and only if (8. ~ i z= 1..

m(()). is also conceivable. j . functions of ( and it follows that K:(i. Otherwise. the inverse image q L . indeed linear. m(<)) varies smoothly with <. m(())having a vertex on [0. 112 on an edge of K:(i..(<) is singular and p. ( m ( ( )will also be a smooth function of ( at (0. which.l(<) = 0 = v. j .J(m(())(R2) meets a vertex of K. j .8 and This is a line in the (s.. m(<o)). j .. Hence.' ( L lis either empty or an affine ) line whose coordinates are smooth. m(<o)) a is line meeting a vertex of [0. when all the inverse images of Ll are parallel and distinct. Thus. the domain of integration Kf(i. It when qLj(m(<O)) singular and so K:(i. / ( ( ) is also zero. u(<)) will be far away from the origin and certainly not on the boundary of [0. since then two of them coincide. ( ( ) .(<) shows that this cannot occur.(m(<)) TO. at then all the non-empty inverse images of the bounding lines Ll are parallel since the image q L j ( m ( < ) ) ( R 2 ) cannot coincide with any of them. in all cases. j .186 INDUCED SHAPE-MEASURES Figure 8. for ( sufficiently close to (0 the vertices of K:(i. that is there are points arbitrarily close to (0 with q+(i. However. 11' or a vertex of [0. the last case can only occur if cp. and it is then empty unless w . 11'. Thus./(().. . t)-plane unless u.(<)(R2) = L. m((0))is a line or a strip. j . j . which again could give rise to a singularity of x. except ) that a continuous but non-smooth variation of the domain may be caused by a vertex of K l ( i . corresponds to case (i) or (iii) above when K:(i. j . m((0))on an edge of [0. j . m(())of x L . t)-plane. we are led to the following definition.. 112. l]'. They are also distinct unless q. moving off in different directions as varies and so giving rise to differently shaped domains K+(i. < < . However. in which case it is the entire (s. and the above analysis of the singular cpo.

112. then the two complex analytic functions Wi{l + 1/55(m(O)) 2wj and ~ l {. (oij(m(<))(s. K:(i. m(<)) varies. j . where. 112 lies on an edge of K:(i. z:) of shape m(<) z. if K:(i. j . The integral x. Then.~ p j 2 2wj would have the same imaginary part for all and so themselves would differ by a real constant independent of <. < H t ( < ) is an open mapping. However. Now. at pi sw. then. j . the imaginary parts of complex analytic functions of <. we need to study the movement of the vertices of K:(i. This occurs if and only if there is a triangle (2T. 112. by (8. we have < < + + + + vij(dm(t))) s(<> + w / i j ( d a ( < ) >= 0 > < and so the first coordinate s(<) of ( o . 1]*. when both of the following conditions hold: (i) no vertex of [0. u(<)). for fixed i.we find (P < and. z l at p j twj and z i at p i . 2.1.fig(m(<))~ l + pi{l + f i < ( m ( < > >. is if Proof. again. by continuity.1. To establish the converse. Non-constant harmonic functions are open mappings and so the mapping H s(<) is also open unless it is constant. for the second coordinate t(<)of (oij(m(<))-' I ) . that xiJ(m(<)>singular at <O if K i ( i . j . j and 1. this is easily seen to be impossible. 112. is the vertex pl of K if and only as t) if there is a triangle (zT. with zT at p i swi. j ( m ( < ) ) . this remains the case on some neighbourhood of and so no singularity can be introduced. 2. and 2: at p j t w j . However. We show next that general position is equivalent to non-singularity of the corresponding contribution to &. The discussion motivating the definition has already shown that general position is a sufficient condition for xij(m(<)) be non-singular at (0 when to K f ( i . we regard s and t as functions of <. <) is said to be in general position with respect to [0.y).25). Lemma 8. g(m(<)) with 2: at p l . ~ j i ( < ~ ( m ( t )t(<>+ wjli(q2(m(<>>) 0 )) = . m ( < ~ ) ) is fails to be in general position.. m(<)) on an edge of [0. lies (ii) no vertex of Kof(i. j . Z. j . Similarly. u(<)) general position is in because it fails to meet [O.A SHAPE-MEASURE ON C: 1: THE SINGULAR TESSELLATION 187 Definition 8. j .) of shape ..' ( p l ) is the quotient of two real harmonic functions of (x. m(<o)) in general position with respect to [0.(m(<)) non-singular at (0 if and only is K:(i. However. j . ~ ( ( 0 ) ) is non-empty.

1. we make the following definition. ( i .188 INDUCED SHAPE-MEASURES It follows that t+ ~ p ~ ~ ( m p l<is also ~ open mapping unless there is ( ( ) ) ) ~ an a constant c such that s(<) = ct(<). z. the union of all the forms its inverse image under m. z.1. position. after the full analysis of the relevant contributions to h~in the next section. j . m(<o)) on the boundary of the unit square.. both algebraically and geometrically. a(<)) to be in general fails which there exist i. So. at Thus. j and m such that &+(i.(i. <) is in general position. m(<)) lies outside the unit square and for which that vertex lies inside the square. and similar arguments show that such a singularity also arises when a vertex [0. so that this necessary condition is also a sufficient condition. in all cases.25) implies that there is a triangle (z?. in view of Lemma 8. respectively. the locus of inverse images of p1 is a line transverse to the boundary of the unit square. forms a copy of aK. then the union of the & j l . both for which the corresponding vertex of K. at p j and z. z:) of shape with its first vertex zT at the vertex pi of K. it will be clear that they do not cancel each other’s singularities. 1 or t(<) = 0. j. which we shall denote by aKt. there are. < < Clearly..(L j . m((o))lies on the boundary of the unit square. similarly.0 ) ) . when a lies vertex of Kof(i. since we have already excluded the possibility that s(<) = 0. and. Thus. However. That set will turn out to be a tessellation of the <-plane by segments of lines and arcs of circles. w . < Definition 8.l(<) = 0. it is necessary for to be in J(K) for it to be a singularity of h ~ . Then. j . To describe Z(K) we first note that Equation (8. arbitrarily close to <o. 1 < I 6 n . rn < < < Rather than specifying when a certain K . j . the set of points for j . Thus. shifted. we shall characterise.2. 112 lies on an edge of K. we do indeed get a singularity of xij(u(<)) (0 when a vertex of K. on the extended edge Ll if and only if w. Ztl . The set J(K) of points at which there exist 1 6 i # j 6 n and a E d such that the contribution xij(m(<)) h~ not smooth at is called ! to is the singular tessellation of h ~ . rotated and rescaled so that the ith and jth vertices are located at the special shape points ~l/a. if we write Z i j l for the Zth edge of the polygon in the <-plane bounded by the n lines < < and write q/ {< : m(<)E = &j/}.(i.

5 2.5 -3..0 1.9. an extremity of the edge.0-1.G.0-2.0 2.A SHAPE-MEASURE ON E. zg) of shape m(<) z. z of shape u(<) each of z. at a vertex of K and with one of zT and z at zz. u(<))to be in general position for some i and fails j when there exists a triangle of shape m(<) with its vertices on dK and at least two of them coinciding with the vertices of K itself. Le. given m and <. 1 { K: # j}. of K.5 4. the failure of condition (ii) is equivalent to there being a triangle (zf.0 -3. on which it lies. by permission of the Applied Probability Trust . 896-916.9 Reproduced from D.0 0. Then.5 3. j . The singular tessellation J(K) o f h ~ is J(K) = U{Z$[u E (1. that occurs if and only if there is a triangle A of shape ( with the same property.5 0. 1 < i. 1 < n : < i. Advances in Applied Probability 19 (1987). . Similarly.5 1. 5. The failure of condition (i) in the definition of general position is equivalent to the existence of a triangle (zf. either A -4.5 -2. g2}. ei or e j . and z. Thus.0 Figure 8.0 3.0-0. j . I: THE SINGULAR TESSELLATION 189 Theorem 8. Kendall and H.52). the polygon Kof(i. However. respectively. j < n and i # j } and i = U d $ m E (1. with . : ) with z at an extremity of the edge of K on which it lies and with z also on an edge .5-1. Proof. The structure and explicit determination of convex-polygonally generated shape-densities. q.

8 THE SHAPE-MEASURE ON X i INDUCED BY THREE LABELLED iid UNIFORM DISTRIBUTIONS IN A CONVEX POLYGON. The reason for introducing this method of description is that. since we are excluding the special shape points. 8.6 that G K is a constant on the real segment [-1/2/5.Zij[. j and 1 in each case. the subset &(K) of J(K). so that it is prohibitively laborious to work out the shape-measure in every tile from first principles for a given K.8. j and 1.26) Jo(K) = e i .1 The Expression for ~FZK the Basic Tile in We know from Section 8. 8. for an appropriate choice of i. the number of tiles increases rapidly with the number of vertices of K and the change in the singular tessellation is generally abrupt when K changes.I90 INDUCED SHAPE-MEASURES itself. we obtain the stated results. and observing that the two vertices common to the triangle and to K cannot coincide. <) and K+(i. which we shall call the basic segment. i . finally. . 1/2/51 of the <-plane. t(<))are [0. we shall first identify a ‘basic’ reference tile and compute the function GK there. 1 6 i 6 n ) and consists of some half lines in the <-plane with finite endpoints at one of the special points 7 1/2/5 and some circular arcs with endpoints at both of the special points. We then find the ‘jump functions’ that express the difference between its expression on pairs of ajacent tiles and.Zfjl or .[Z . j . when is on this segment. 11: THE EXPLICIT FORMULA We close this chapter with a description of a so-called stepping-stone method to obtain an explicit formula for the Radon-Nikodqm derivative (8. 11’ for all i # j and the polygons K+ corresponding to the < . m). < Figure 8. i + l . The argument there also implies that. j . This subset is is given by -m (8.17) of the shapemeasure of random triangles whose labelled vertices are iid uniformly distributed in a compact convex polygon K. <(A) or <’(A) has its first two vertices coinciding with vertices of K and so lies on . ‘step’ from the basic tile to any given tile to find its expression there. Moreover.. consisting of those points at which $rij(m(<)) singular for some triple (i.9 shows the singular tessellation J(K) for the case when K is a certain irregular pentagon. will be important.l: m E Q. When we come to discuss the differential order rather than the position of the singularities. K+(i. and it is impossible to get a general formula for the shape-measure that is suitable for all convex polygons. So. as we saw in the previous section. Taking the union over all possible i. j . the singularities of the Radon-Nikodym derivative of this shape-measure split the whole space into irregular tiles on which it has entirely different expressions.

t ) at one corner of [0. m K+(i. the set K+(i. except that then the roles of and t(<)are interchanged. as moves away from the basic segment into upper half-plane. Kf(i. However. On the other hand. for sufficiently close to the basic segment. m(<)) remains empty. This discussion shows that f i can be expressed in the basic tile as ~ < < < < < + + < kK(<) = fig'(<> + fig)(t(<)). where K. t(<)) remains equal to [0. then. <) loses an increasing triangular area near the (0. Since f i is ~ real-analytic in each tile. This will be the case until the choice of ( forces the third vertex to lie on aK. So.O)corner and (1. it will be enough = for us to consider the expression for f i in the one. in this case. 112 and. necessarily for (s.(i. if ei and ej are neither contiguous nor have a single edge of K between them. t ) at some point of a([O. j . for m = g. when j = i . m(<)) remains as it was for in the basic segment and so xij(zu(<))will remain constant. # 1. respectively. Again. we are able to obtain its expression in these two tiles by examining it at any sufficiently close to the basic segment and. m(())remains empty. each such summand xij(m(<)) remains constant until meets fails the appropriate edge of J(K). K+(i. Kf(i. When j = i . j . j . s2t(<)) begin to grow out from the (0.1. However. t(<)) remains equal to [0.2.A SHAPE-MEASURE ON C: 11: THE EXPLICIT FORMULA 191 other relabelling transformations are all void or degenerate. Hence. both the previously empty sets K+(i. where . In all that follows this tile will be called the basic tile. 112. that lies ~ in the upper half-plane. Similar results hold when j = i 1. 112 and. it is clear that. 112). 1)-comer. for m E a. <) loses two increasing congruent triangular areas containing the (0. j . t. j .l)-corner. these summands certainly remain constant throughout the basic tile. for (s. since the relabelling operation t yields the symmetry f i ~ ( < ) %K(-<). we note that it is a property of the singular tessellation that the basic segment is always an edge of exactly two tiles and that these tiles are congruent triangles. j . Looking from within K. st(<))and K+(i. the constant value CK of f i on the basic segment can be expressed as ~ To describe our stepping-stone method we need at least to find the expression ~ for f i in one of the tiles of the singular tessellation. K+(i. m(<)) to be in general position. To do so. j . j . j . g2. Thus. i 2. the set K+(i.l)-corner that maps to the common vertex. of these two tiles. j . or the third vertex to coincide with a vertex of K.

Theorem 8. K. However.<) by substituting i for i + 1.+la. The analysis summarised above also tells us that the domains of the corresponding K+ are determined by the intersection with the unit square of at most two bounding lines of K:. the resultant expression for “. t)g(s) ds dt K+(i+l.+l J’ J’ f ( s .t M s >ds dt may be found from that for C. the non-constant contributions to f i ~ ( < ) the in expression for fig)(t(<)) may be found directly from appropriate ones in fig)(<). we get the following result.192 INDUCED SHAPE-MEASURES and KZ denotes the complement of K+ in [0.r. .+( I . Clearly. E e.aIal+l J’ J’ a. i + 1 for i and t(<) for <. with these comments kept in mind. After some computation. for each fixed i there are eight non-constant contributions to f i ~ ( < ) associated with z.10. m) f (s. I +1 . l]*. For example. The expressionfor f i in the basic tile is given by ~ n where and where .

Thus. and so xl. Explicit formulae for polygonally generated shape-densities in the basic tile.a 2 is elzJzlz an arc y of positive length. 101 (1987).o 1. each edge ZEl accounted for several summands of GZK.(Zl. j ) .1).5 -1 .o -1. in which the contours for &K are drawn on the basic tile and its five equivalents. .o 0. if the intersection of Zyjlll and . That will cause K. to ZYJll1 y .5 Figure 8. j ) . The ‘primary’ edges Z F l of J(K) can overlap one another.1 .0 0.10 we show the singular tessellation for K being a rectangle.9.10 Reproduced from H.5 0. and will link the local forms of GZK in the two tiles that lie either side of that edge. (0) not to be in general position.5 -1.)) be singular. 1 2 . z. (i . Le. one on a neighbourhood of each open edge of the singular tessellation J(K).8.2 The Jump-Functions The jump-functions for &K that we shall give will be defined.0 -0. when we discuss the jump of GZK with respect y. We now need to make that correspondence precise. 8.z.. at p L .5 -1. Mathematical Proceedings o the Cambridge f Philosophical Society.5 193 1 .5 0. (i.12((. we temporarily ignore the jump of t i Z ~arising from When identifying the singular tessellation in Theorem 8.A SHAPE-MEASURE ON Ci 11: THE EXPLICIT FORMULA 1. at p J and zz in the interior of el. with z.1. To that end we recall that a point (0 lies on the open edge = Z r J l \ {end points} if and only if there is a triangle (zT. j - 1) or (i .0 -0. j .5 1 .) of shape (0 z. but we shall need to distinguish them in order to describe the jump-functions. whenever to n Zzplz n (11.1 2 ) = (i. 313-21 by permission of Cambridge University Press In Figure 8.

194

INDUCED SHAPE-MEASURES

However, since, for each m E Q, the triangle (z;, z;, z;) may also be regarded as a triangle of shape m(<o) with z & ~ , at p i , at p j and z&) in the interior of e l , we see that the full set of summands of mK that are singular at <o on account of <o lying on Zijl, is 1;j’ = j , j - I}, (8.27) where x12 : N x N x N H N x N is the projection on the first two factors. Suppose now that a part of E J(K) contributes to the boundary y that )) separates tiles TI and T2, so that w i j l ( m ~ ( < = 0 on y. Then, for in the interiors of TI and T2, the function wijl(mo(<))is non-zero and has different signs that are constant throughout each set. Let the sum of all the contributions (8.27) to AK be in the tile T I ,where wijl(mo(<))< 0, and be 7 2 in the tile T2, where wijl(uo(<)) 3 0. Then each of 71 and 7 2 can be continued as real-analytic functions into a two-dimensional neighbourhood Nijl of the open arc .; If these continuations are called and then
-

(xl,12(m(<o)) E Q; (11, I 2 ) = xl2(m(i’, j ’ , I ) ) ; i’ = i, i :m

2 3

<

71

7; 7;,

is defined to be the jump-function for &K with respect to Then, if y is the intersection of i?zjb,lb, 1 < b < bo, the total jump of AK along y will be , given, on the intersection N, of the neighbourhoods N Z j b , l bby the composite jump-function:
J,(<) =

23.

C sign(~ib,j~,lb(mb(<)))z j ~ , l ~ ( < ) , ~ b=l
n

hi

where sign(t) is the function that is 1 if t > 0, -1 if t < 0, and 0 if t = 0. Thus, once we know either one of f l = A K ~ Tor f 2 = AKIT~, , say, f l , we are able to construct the other, f 2, in N, T2 by adding the total jump-function to of AK along ; the continuation f T of f 1 : that is, we have
f2(<)

= fT<<) J,K)

+

in N, T2. Then, since each of the summands of AK has a unique expression in the interior of each tile, the expression found for J , in N, remains valid for its analytic continuation throughout TI UT2. Hence, the expression we found for f 2 in N, T2 actually remains valid throughout T2. The structure of the tessellation also shows that the jump-functions for &K with respect to Z;! and Z$ can be deduced from the jump-functions for AK with respect to the segments i?:jl = Z i j l by making use of the transformation operators 5 and 52 and applying the appropriate absolute Jacobian factors to allow for the difference between the complex differentials d< and d(u(<)). Hence, in what follows it is sufficient to find the jump-function for AK with respect to each segment Z i j l .

n

n

A SHAPE-MEASURE ON C : 11: THE EXPLICIT FORMULA

195

The set (8.27) of contributions to GK that are singular along Z i j l is composed of two distinct subsets. First, there is the subset, in which m = 1 or t,corresponding to the polygons K l ( 1 1 , 1 2 <) and KJ (11 , 1 2 , t(<))that fail to satisfy condition , (i) when lies on Z ; j l . For each such polygon it is its Ith edge that contains a vertex of [0, 112 and so breaks the analyticity of the associated integral terms. Then, there is the complementary subset of eight summands corresponding to the set of polygons that fail to satisfy condition (ii) on .Zijl. In this case, one of the suffices ( Z 1 , 1 2 ) is equal to 1 and it is the other suffix that determines the vertex of the appropriate polygon that lies on an edge of [0, 1]* and so accounts for the lack of analyticity of the corresponding summand of GK.In this second subset, m takes the values 5, 52,t5 or ts2.The terms in the jump-function corresponding to the four members of the first subset that involve the same value of m are all very similar and depend only on the choice of the suffixes (1 1 , 1 z), and the same is true for each pair of members of the second subset that involve the same value of m. In is addition, for a given choice of (ZI, 1 2 ) , the term corresponding to t(u) directly deducible from that corresponding to u in a similar manner to that described in the computation for the formula for &K in the basic tile. Thus, it is sufficient to carry out the explicit computation only in the following three special cases. (i) The Zth edge of K;(i, j , <) passes through (0, 0) in the (s, t)-plane when is on .Zijl. Therefore, the change in the type of shape of K+(i, j , <) is associated with the inequalities

<

<

uil(<) s

+ u j l ( < ) t + wij l ( 0 3 0

and s, t 3 0. It follows that the integral formula for the associated jump term is

1 ciuia, 1921KI3
where
$0

1” JG’”
ds

f ( s , t ) g(s) dt,
wi j l (<I

wij = -~l ( < >

ui/ 0 (

and

to = -

+ uil(<) s

uj/ (<I

(ii) The jth vertex of Ki(1, i, 5(<)) on the edge t = 0 of [0, 112 when is lies on Z i j l . Therefore, the change in the type of shape of K+(I, i, 5(<)) associated is with the inequalities

<

and t 3 0, where we have used the relation

196

INDUCED SHAPE-MEASURES

Because the t-coordinate of the jth vertex of KO+(/, s(<)) u;l(<> = -w;jl(<), i, is to it follows that the integral formula for the associated jump term is

where

(iii) The ith vertex of K;(j, I , s2(<)) on the edge s = 0 of [0, 112 when lies Therefore, the change in the type of shape of K + ( j , I , s2(<)) is associated with the inequalities

< is on Z i j l .

and s > 0, where we have used the relation

Because the s-coordinate of the ith vertex of K;(j, I , q2(<)) is vjl(<) so = -w;jl(<), it follows that the integral formula for the associated jump term is

where

197

After calculating and simplifying the integrals, we obtain an expression for the jump-function for f i as follows. ~

Theorem 8.11. The jump-function Jijl for f i at with respect to & j l is given by: ~

<

and

i-1

198

INDUCED SHAPE-MEASURES

We now examine the order of the first discontinuous derivative at a singularity of f i ~Theorem 8.11 shows that the jump-function Ji;l for f i with respect to . ~ the component Zi;l of J(K) can be expressed as a finite sum of quotients of binary polynomials in (x,y) in which all the denominators of the summands are products of powers of u;r(<), u i - l , l ( < ) , vjr(<) and w,-~,l(<), where = (x,y). Hence, the denominators occurring in the summands of J;l(<), as i, j , 1 and m vary, can be split into powers of factors of the following forms:

<

where the last factor arises from the Jacobian dm0 . The way in which the dC above factors are involved in the summands of the jump-functions means that the type of singularity of a jump-function will be determined by whether or not the relevant edge belongs to the set &(K), defined at the end of Section 8.7, in the following manner.

I I

2

Theorem 8.12. The order of differentiability of the jump of f i with respect to ~

.?El is three when .?$

E

&(K), and otherwise it is at least four.
& [ ;

Proof. The analyticity of the jump-function J ; j l ( ( ) in a neighbourhood of asserts that none of the singularity lines of J ; , l ( < ) ,
uil(<)

= 0,

ui-i,i(<)

= 0,

v;r(<> = 0

and

vj-i,i(<) = 0,

(8.29)

can intersect Z i j l unless & j l is part of one of these lines. The open arc &ijl is part of the line determined by wijl(<) = 0. So, if Z i j l is not part of one of the lines (8.29), it follows from (8.28) that J;,l(<) has a zero of order four and hence that the order of differentiability of the jump of 6~ with respect to Z i j l is at least four. On the other hand, Z i j l is part of one of the lines (8.29) if and only if one of the following conditions holds:
j = i + 1 , 1 = z 1; j = i - 1,1 = i - 2; . . J = z - 1 , l = i; j = i + 1 , l = i - 1.
'

+

Then, the corresponding functions

wijr

(<) can be rewritten as

respectively, and, on these segments, the singularity is removable and then J i j l ( < ) is O(W?~~(<)) is not O(wfjl(<)). Thus, if we substitute m(<) and for in (8.29)

<

A SHAPE-MEASURE ON C; 11: THE EXPLICIT FORMULA

199

and compare it with (8.26), we find that only if E Jo(K), as required.

ZiEl

ZiGl is part of one of loci (8.29) if and

We observe that each jump-function for h~ with respect to a component of

J(K) is real-analytic away from a natural singularity set consisting of either
infinite lines through one of the special shape points ~ l / & complete circles or through both of these special points, or both of such infinite lines and complete circles. We also note that the union of all the singularity sets of the jump-functions for h~ coincides with the union of the x-axis and the set
{< : qij(rn(<)) is a singular mapping,
tir E

2 and 1 !

<i # j <n).

8.8.3 The Stepping-Stone Principle and an Explicit Formula for r i i ~
We are now in a position to describe our stepping-stone method to obtain an explicit formula for h ~ , whatever the convex polygon K may be. This will allow our evaluation of h~ to spread out from the basic tile and eventually to cover the whole of the plane using our knowledge of the singular tessellation J(K), the formula for h~ in the basic tile and the formulae for the jump-functions for h ~ , together with the fact that h ~ ( < ) ~ ( - < ) . =h We proceed as follows. Let Ti, be a tile in the upper half-plane in which we wish to find an explicit formula for h ~ . are many ways of stepping from There the basic tile To to Ti, in the upper half-plane and, from our present point of view, it does not matter which route we choose, although obviously some routes will be more convenient than others. Suppose we have decided on the route {Ti : i = 0, 1, . . . , io}. Let h~= f ; in the tile T; for each j and suppose that we already know the functions fo, f l , . . . , f i . To get a formula for f i + l we recall that, if f is the real-analytic continuation of f i , then

T

J!+1

= f ' 1+1

-

fT

will obviously be real-analytic, at least throughout an open neighbourhood of the open arc of the boundary separating Ti and Ti+l, and at most differ from I= + ' our jump-function J , for &K by a sign: that is, . ; fJ,. The sign is totally + determined by the equations of the bounding lines of & together with the route I is + ' we have chosen, so that . : known to us by Theorem 8.11, and we are then able to express f i+l in Ti+1. Thus, we have established the following result.

Theorem 8.13. For any tile Ti, in the upper half-plane we have
io-1

i=O

whenever the tiles To, T I ,. . . , Ti, form a chain in which each pair of adjacent tiles has a boundary-arc in common.

since.. we can first attempt a direct evaluation at the given point and then. to obtain the explicit form for &K in any named tile by starting with its form in the basic tile and adding to it certain jumpfunctions describing the switches in analytic form when crossing tile boundaries. the theorem would imply that there must be more than one such singular term and the singularities viewed algebraically must cancel out. For example. at least theoretically. when we work numerically. the points on the lines x = & l / & all have this pathological character when K is a square. this sum being formed following any desired stepping-stone route from the basic tile to the target tile via a stepwise continuous sequence of intermediate tiles. which consists of a finite number of lines and circles. this would appear as 0/0 and straightforward numerical computation would fail to evaluate it. with the aid of the formulae for the jump-functions. However. however. it would usually be possible to examine alternative steppingstone routes from the basic tile to the target tile and to try to choose one so that no term in the decomposition formula is singular at a given point. However. attempt one at a slight perturbation of the given point. The direction of the perturbation can always be chosen after inspecting the union of the global singularity curves for all jump-functions in the decomposition formula. there is a problem that might arise in a practical computation based on this approach. Such an approximation technique will also deal with the removable singularities. may not always be possible: sometimes there is no stepping-stone route from the basic tile such that no term in the decomposition formula is singular at every point in the target tile. Then. we know exactly where the singularities of each term lie. when 0 0 there is a removable singularity. On the other hand. say. This. the unwanted terms would appear as 0 . .0 and then the numerical computation will fail. Nevertheless. nine-figure accuracy. Similarly.13 as an algebraic identity can be used.200 INDUCED SHAPE-MEASURES The decomposition formula given in Theorem 8. if that fails. as such a numerical value is usually rarely wanted to more than. The existence of the singularities of the jump-functions makes it possible that some terms in the decomposition formula are singular at some points of the tile Ti.

we are considering the shapes of configurations of random vertices in a Euclidean space. To understand the problems that can arise when we try to define a mean in the absence of a given linear structure. it is quite natural to choose the shape of the means of the vertices to be the mean. Neither is it clear that such an intrinsically defined mean might not be a more appropriate statistic for any particular problem. and discuss the relations both among them and between them and the shape of the means. In the case of the statistical analysis of shape it is not even clear what the mean should mean and different circumstances may require different definitions. LE Copyright 0 1999 by John Wiley & Sons Ltd CHAPTER 9 Mean Shapes and The Shape of the Means In any classical statistical analysis the estimation of the mean of a sample is an important first step. BARDEN & T. The most we can say. it is not obvious. Then. rather artificially and certainly . it is enough to consider the simplest non-linear manifolds. which would be the next best possibility. The result would surely accord with our intuitive preconception of where the mean should lie. C A R " & H. G. that this is the mean of the shapes in any sense that can be defined directly on the appropriate shape space. and indeed may not be the case. What. shape spaces are not linear spaces and most of them are not even manifolds. if the probability measure were supported on an entire circle? There is no way to map that isometrically into the real line.1 CONCEPT OF MEANS IN NON-LINEAR SPACES As we have seen in earlier chapters. perform our calculation of the mean there in the usual way and then transfer it back onto the arc. However. 9. in general. We shall also discuss practical problems that arise in their application. KENDALL & D.Shape and Shape Theory D. such as associated Laws of Large Numbers and algorithmic methods of computation. as is often the case. Further confusion can arise when. We might. K. For a probability measure on an arc of a circle we could map the arc isometrically onto an interval of the real line. In this chapter we shall describe various concepts of mean shape. by which we understand a mean that can be defined directly on shape space. however. is that they are all metric spaces with a natural stratification in which all the strata are Riemannian manifolds.

this approach fails. dist) be a metric space and 4 be an M-valued random variable with probability measure p on M.5"s. for example. almost any approach will lead to the conclusion that every point of the equator is a mean. lie entirely within the space on which the probability measure was originally defined. that would give us a mean that does not lie on the circle. Let (M. the set of means could occupy any submanifold or more general subspace and it makes sense to ask for topological or geometric invariants of the set of means. consider a distribution with equal weights at + I and -1. The problems that can arise when similar embedding techniques are applied to probability measures in higher dimensional non-linear spaces are even more intractable. However.5"N and 23. between latitudes 23. a uniform distribution on the part of the sphere lying in the 'tropics'.202 MEAN SHAPES AND THE SHAPE OF THE MEANS non-intrinsically. probability measures whose sets of means had different dimensions could be expected to be significantly different in other ways. Three of the most successful ones are the notions of FrCchet means of random variables (cf. at least if we want to obtain a unique well-defined mean. However. However. Where is its mean? Any reasonable approach will give +i and -i as equally valid answers. 1977). choose an isometric embedding of the circle in the plane and carry out the classical computation for a planar distribution. Then.1. of barycentres of measures (cf. Emery. In general. which is surely a necessary pre-requisite for a sensible theory. Provided the planar mean is not at the centre of the circle. where the calculations. there are various non-linear operators we may use to replace the usual linear one. 1977). as we shall see below. If this were a shape calculation. that is. 1989) and of Riemannian centres of mass (cf. that will give us a well-defined mean on the circle itself. It could also fail more fundamentally if the mean falls outside the curve or if the interior of the curve has no point from which the projection onto the curve is continuous. . At the very crudest level. dist) and it will be clear from the definitions that follow that none of them is necessarily unique. These means are all defined on a metric space (M. In fact. then we would have a mean shape that is not a shape! We could recover the situation by projecting the planar mean from the centre back onto the circle. simple examples will show that we should not expect them to do so. not only when the planar mean is at the centre of the circle but also for less regular simple closed curves for which there is no obvious preferred projection of the interior back onto the curve. the potential non-uniqueness of means is a feature we shall have to accept in general since. To obtain an intrinsic definition for the mean of random variables or probability measures on general metric spaces or on manifolds where there are no linear structures. Karcher. it is easy to construct examples where the set of means has dimension greater than zero. and the solution. the intrinsic methods of calculation that do exist do not guarantee to give a unique answer. Ziezold. Once again. Regarding the unit circle S' as the set of points of unit modulus in the complex plane. so clearly there is a need for an intrinsic definition of a mean. Definition 9. Consider.

dist(x.LL therein is not very severe since. on a sphere. Le.1. a regular geodesic ball could be almost a hemisphere and. However. that to obtain results such as the Law of Large Numbers in Proposition 9. 1998). a non-trivial matter to establish that a FrCchet mean for U.CONCEPT OF MEANS IN NON-LINEAR SPACES 203 (i) The point xo is said to be a FrCchet mean of function F. It is.2 below we need to use a metric. are related to random variables concentrated on domains that can be described geometrically in terms of convex geometry. f dLL(x) for all bounded convex functions f : U H R. on a space of non-positive curvature. Note. (iii) The point xo is said to be a Cartan mean of with respect to ‘dist’ if xo < is a local minimum of the function F . is no longer a FrCchet mean for F . it could be the entire space. since this involves taking into account the global topological nature of the space M. s. The sufficient conditions for uniqueness of Cartan means referred to above generally involve the curvature of a manifold. defined on the entire space M. as is also the case for barycentres. The FrCchet mean could be generalised by replacing the metric in the expression (9. which a priori is only a Cartan mean for M. defined as minimum for the function F . while for Proposition 9. To specify such conditions we follow Hildebrandt (1985) in calling a geodesic ball of radius a regular if the supremum of the sectional curvatures upon it is less than ( ~ / 2 a )This bound ~. Many results concerning Cartan means. Proposition 9. by a member of a wider class of functions on M x M (cf. Note that the restriction on . Suppose further that (M.1) . where even a FrCchet mean. is indeed a FrCchet mean for M.1 we need to have the specific metric induced by the Riemannian metric. (ii) The point xo is said to be the barycentre of 6 if f (xo) < Jc. . ( x ) < 00 for some x E M. Suppose that . provided F . xo E U and range(e) c U a s . dist) is a Riemannian manifold.1).LL is a probability measure on a Riemannian manifold (M.(x) = 6 with respect to ‘dist’ if the (9. it is easy to construct examples similar to those described above. in (9. however. YI2 W Y ) attains its global minimum at xo. rather than the entire manifold.1) for F . where U c M is open and convex. is required for the following results. The corresponding uniqueness results that can be obtained relate also to those domains. in general. restricted to an appropriate convex domain U. dist) and that &(xo) is a regular geodesic ball contained in the manifold with centre xo and radius a. for example.

To state his result we introduce the analogue of the FrCchet mean for a sample of n points. Then be 9. . Karcher’s hypothesis is. we shall now primarily be concerned with Frtchet means. then F . we shall also need to consider the more .2.. The first of these results was obtained by Kendall (1990b) and the second by Karcher (1977). a FrCchet mean of such a probability measure is a global minimum of the function dist ( x . of course. . necessarily a Cartan mean. However. As already noted. F. then there is one and only one Cartan mean of P on Ba (xo1. we can define a discrete uniform probability measure po supported on these points. dist) and such that E[dist(x. determinate. .204 MEAN SHAPES AND THE SHAPE OF THE MEANS (i) I f p is supported on Ba(xg). 61)2] < 00 for some x E M. Let M..x ~ ) ~ . . However. Proposition 9. For a given. be iid random variables with common probability measure p. strictly stronger. . . be the set of Fre‘chet means of p with respect to ‘dist’ and M(t1. (ii) I f p is supported on the concentric ball Bap(xg)of halfthe radius. an appropriate generalisation to random variables on metric spaces of the classical Strong Law of Large Numbers has been obtained by Ziezold (1977). Let 61. if we assume that the n given points are iid random variables on M with common probability measure p on M. Then. is convex and has one and only one critical point. . to estimate such means in practice we need a related Law of Large Numbers and. the relation that Ziezold obtained between the set of FrCchet means of samples and the set of FrCchet means of p is the following.2 METRICS ON SHAPE SPACE Of the means introduced in the previous section. having values in a separable metric space (M. in fact. on Ba/2 ( X O 1.62. set of n points ( x j E M : 1 < j < n } . then we may define the set of FrCchet means of samples of size n from p to be I Then.(x) = ’ j=1 2 Analogously. tn) dejined as above. the conclusion is more precise and it remains important as a practical method of ‘finding the mean.

We recall that.the natural way to induce a metric on shape space from one on the pre-shape sphere is to define dist(n(X). W ) l l = (9.and the induced distance between the corresponding shapes by = cosp(n(z). ) = 1 . ) = 1 . most of our results will make use of three particular metrics.3) on shape space. which we have already met in Section 6.' ( n ( ~ ) ) . Both concepts require that we select metrics. n(Y)) = min dist(X'.4.) = sind(X. (9. i ( z . Y ) . on shape space we may define . then the distance between two points z and w on the pre-shape sphere is given by cosd(z. w ) = z W'. ~ a>O As before. we obtained the metric p given by Equation (6. or possibly more general functions.{tr(XYf>}2 ~ d ( ~Y. Y ' ) . Applying this procedure to the standard Riemannian metric or great circular arc distance d on the pre-shape sphere. In fact.METRICS ON SHAPE SPACE 205 general Cartan means at some stages of our discussion.3). n ( w ) > 2 2{1 .a Y ) ' } . from the chordal metric 2 on the pre-shape sphere. The first two of these. where (z . Y ~ on the pre-shape sphere.a Y ) ( X . W ) l .2) Similarly. For the complex representation of the case m = 2. Thus. A useful variant of the second of these metrics arises by defining 2 ( ~Y.l(z. Y >= mintr{(X . In the special case m = 2. w ) is given by I Iz . if we use the complex row vectors instead of matrices to represent the pre-shapes. Its square may also be expressed as J(x. on shape space. where the minimum is taken over all X' in the orbit n-'(n(X)) X and all Y' of in n .(X. we obtained the procrustean metric j on Xk. W ) = %((z. since shape space is the orbit space of the pre-shape sphere under the action of a compact group with quotient map n.given by Equation (6.w I I on the pre-shape sphere and Xn(z). which is the length of the perpendicular from Y onto OX in the Euclidean space containing the pre-shape sphere. are induced from metrics on the pre-shape sphere.4). n(w>> l(z. w)).

However.n ( Z ) ) and p ( n ( Z ) .s 2 ) + sins2 = coss2 + sins2 3 1 3 G ( ~ ( x ) n ( ~ ) > . Similarly.1 gives what are.1. writing n n $1 and sz. . w ) l2 = sin2 p ( n ( z >. this is also the most important case.n ( Y ) > . note that p takes values in the interval [0. as we made clear in the first section of this chapter. we cannot. the corresponding formulae are on the pre-shape sphere and = 1- I ( z . and throughout this chapter. generally via the minimising procedures that give rise to them.n ( Y ) > . ( Y > )= 0 and.4) on shape space.X > = 0. from a practical point of view. is most easily applied when the probability measure itself has a unique FrCchet mean and. in general. ( Z ) )+ p ( n ( Z ) . ( Y > )= 0 if and only if p(n(X). for which the above formulae would not remain valid. we have p(n(X). expect there to be a unique FrCchet mean. which we mentioned in Section 9. n / 2 ] on which sine is an increasing function. To see that. This contrasts with our previous use of such symbols to denote general points in Ck-'. whether they remain FrCchet means on the entire space. although a regular geodesic ball can be arbitrarily . in fact. we use z and w to denote points on the pre-shape sphere Si. . ( Y ) >6 sin(s1 + s2) = sins1 cossz + cossl sins2 n 6 sins1 sins2 = F(n(X). nevertheless 5 defined on shape space is a metric. n + provided sins1 s 1 + s2 6 n / 2 . that is. When s1 + s2 > n / 2 . 9. we have + sins2 > sin ($ . p(n(X). So. unique Frkchet means as far as the relevant geodesic balls are concerned. those results do not address the question as to whether they are global minima. All three of these metrics have been used in various statistical analyses of shape. for p(n(X).206 MEAN SHAPES AND THE SHAPE OF THE MEANS For the complex representation when m = 2. respectively. Although 2 is not a metric on the pre-shape sphere since J ( X . Although Proposition 9.n(w )) (9. We should emphasise that here.3 UNIQUENESS OF FRECHET MEANS OF SHAPE-MEASURES The Strong Law of Large Numbers due to Ziezold.

. . whichever of p. 0. in particular. To describe these results we shall now. all the Fre'chet means of p with respect to p . . cannot be applied to most shape-measures induced from distributions of vertices which. it is sufficient to prove that. given two shapes whose Riemannian distance apart is s. which is clear from the expression (9. is a shape-measure on Z supported on a geodesic ball of radius . In this section we shall use our explicit knowledge of shape space firstly to show that.1) of isometries. Writing &/8(n(zo)) for the given geodesic ball. . which is even unbounded above for m > 2. that is not the case for shape spaces since they have positive curvature. Z i is a Riemannian manifold. . We shall use several of the properties of this metric starting with the fact. 0. the integrand in F . p ( n ( z ) . . we may map them to the standard shapes n (I . Then.1) induces an isometric involution of shape space that interchanges the latter shapes and hence determines an isometric involution that interchanges the two given shapes. to . and throughout the remainder of this chapter. for shape-measures supported on suitable geodesic balls. We consider first the question of the location of FrCchet means. for n(w) in &/s(n(zo)).22 are all increasing functions of p. Secondly. To establish this we use the fact that Z i is homogeneous to take zo in the standard position (1. stabiliser in U(k .(n(zo)) for n/8 < s < n/4. n ( w ) ) is increasing along such normals. This will enable us to translate results on the uniqueness of Cartan means for such shape-measures into uniqueness theorems for their FrCchet means.Z2 and . (n(z is increasing as n(z) moves )) along the outward normal to the boundary of B. The basic metric on shape space is p and.2) for it given above. the reflection ( toss sins sins -coss @ lk-3 in U(k . . with this metric.0. are commonly supported by the entire shape space. 0) and to take any given geodesic starting from n ( z 0 ) to be vp . This means that those results. n/8. sins. Moreover. . .2 is involved. . Since p2. Indeed.Z or to . which require the distribution to be supported on a geodesic ball. do not require any restriction on their support. 0)) and ( n (cos s. Zi is homogeneous with respect to p since the action of the unitary group on the pre-shape sphere is transitive and commutes with the action of SO(2).2 lie in that ball. since this induced action of U(k . we shall obtain conditions for the uniqueness of FrCchet means that hold for a large class of shape-measures and that. . 0)). .1) of each shape n(z)is transitive on each set of the shapes at a given distance from n ( z ) . their Frkchet means will also lie in that ball. we show that. as we saw in Chapter 8.1.1) is its standard action on CPkP2. restrict ourselves to the case m = 2. that it is invariant under the group induced by the unitary group U(k . although some of the results we shall obtain can easily be generalised to the case m > 2. Thus. . Lemma 9.UNIQUENESS OF FRBCHET MEANS OF SHAPE-MEASURES 207 large in a Riemannian manifold of non-positive curvature. then Proof. 5 or .

. on Xt. is increasing in all radial directions throughout B.p ( n ( z 0 ) )\ &/8 (n(zo)). and since cos2 s is a strictly decreasing function of s for 0 < s < n / 2 .1) sin 2s - + + tl> 2s) cos (<I < cos2 p(n(zo>.208 MEAN SHAPES AND THE SHAPE OF THE MEANS the image n(z. (9. the metric distance p(n(z. in fact.n(w 1) {(r: < cos2 p(n(zo>. j or 5..cos2 p(n(z.). Thus. for each of p. d ds n(w = cos2 p(n(zo>. Since the subset of Et determined by the pre-shapes with w1 = 0 does not meet &p(n(zO)).. that between n(z. n(w 1) = cos2 p ( n ( z o ) . .n(w { ( r .2. we may represent n(w) in &p(n(zo)) by the standard inhomogeneous coordinates <j = wj+l/wl for 1 6 j < k ./4(n(zo)) must all be taken inside &/8 ( n ( z 0 ) ) . for 0 < s < n / 2 .2s) 6 0. cos2 p ( n ( z o ) . for each n(w) in &/8(n(zo)).n ( w 1) sin 2s - + 2 rl cos 2s) { (tan2 f 1) sin 2s +2 tan f cos 2s) = cos2 p(n(zo). n ( z 0 ) ) is equal to s. n(w)) is increasing as n(z) moves along the outward normal to the boundary of a. 5 It follows that. Then.) of the horizontal geodesic (cos s.the same must be true for F . as required. . ) . . and does so orthogonally. n(w 1) C O S .). 0) through it (cf.n(w )) { (cos s + <I sins) (cos s + t.cos2 p(n(z. since for n/8 < s < n/4 we have .6) Since the geodesic n(z.n(w 1) = Iw1 l2 = 1 + r: + + rZp2 > cos2 f ' ' ' 1 (9. n(w 1) < d ds 0 for n/8 6 s < n/4. the minimum values of F .).~ sin ($ . sins. that implies that these minima must.(n(zo)). it follows that.) and n(w) is given by cos2 p ( n ( z . However.sins)}. since each value of the integrand of F. in order to show that.. p(n(z).) hits the boundary of Bs0(n(zo))when s = so. on the compact set B. . For if n(z)is any .However. 0. for 0 < s < n/2. be global minima of F .it suffices to show that . In terms of these coordinates and writing rj for the Riemannian distance between n ( z 0 ) and n ( w ) is given by Irjl.5) and. itself. Chapter 6).

2. in particular. To establish that we require the following lemma. We show first that the integral on the right-hand side of this inequality is independent of the choice of n(z0). for all n(w>in &/8(n(zo)). with respect to the uniform measure dW5 on Xi.n(w >>.(n(zo)) and so n ( z > cannot be a FrCchet mean of p with respect to any of p. Then. We recall that in Theorem 7. we have p M z 1. Let f 1 and f 2 be two strictly decreasing functions on the nonnegative real line and n(z1) and n(z2) be distinct points in Zk. By transitivity we may choose g in the group = induced by U(k .UNIQUENESS OF FRBCHET MEANS OF SHAPE-MEASURES 209 point outside Dn/4(n(zo)) distant E > 0 from the boundary of &. for any n(zo> in xi. we saw that with the metric p the maximal sectional curvature at any point of Xi is equal to four. 6 or 5. Lemma 9. we have .1 to give the following result. Theorem 9. E + p m z o > .4(n(zo>>.(n(z)) is strictly greater than F.1 (ii) with Lemma 9. then p has a unique Fre'chet mean with respect to the Riemannian metric p and that mean lies in &/8(n(zo)>. F. We first give a sufficient condition for that point to be the unique Fr6chet mean. Proof. $ or 6. Our more general uniqueness results will depend on the Radon-Nikodym derivative of the shape-measure. then.1.1) such that g(n(zo>> n(z1). > Thus. so we may combine Proposition 9. n ( w >> > E + . being a function of the Riemannian distance p from a fixed point. gp is a shape-measure on C that is supported on a geodesic ball i &/8(n(zo)). since g is an isometry. It follows that any geodesic ball of radius n/4 is regular.3 we computed the sectional curvatures of shape spaces and. calculated with respect to any of the metrices p. Then.

2 below. and 5 explicitly mentioned in the theorem. Once we note that Zi has finite diameter with respect to each of p. Then.210 MEAN SHAPES AND THE SHAPE OF THE MEANS However. ( p ( n ( z l ) .1) that interchanges n(z1) and n(z2).2 may obvii ously be weakened: it is sufficient to assume that the functions j in the proof are non-increasing and are strictly decreasing on a subset of A of positive measure. taking n(zo>to be at n(z1) and observing that the fixed point set of j has measure zero. which. Similarly. d&n(w >).=. we see that the difference between the right-hand and left-hand sides of the inequality in the lemma is which is strictly positive since A has positive wi-measure. implies that the uniform measure dwi is invariant with respect to the induced U(k . let A be the subset of Zi that consists of all for points of Xi that are closer to n(z1) than to n ( z 2 ) and write j. Using these two facts. the point n(zo>from which the distance is measured would also be the unique ‘generalised . Now let j be an isometric involution in the group induced by U(k . we see that the final expression above is equal to f 1 (p(n(z1>.1)-action. >>) * ( p ( n ( z 1 > n ( w I>> n(w f . It may be worth noting that our results lend themselves to some obvious generalisations. 1 for a probability measure p on M. the hypotheses of Lemma 9. a We now observe that since d M . For example. in particular.n(w>)>. Zi is metrically homogeneous. jj and 5 and that these metrics are all strictly increasing functions of p. which simultaneously identifies the FrCchet mean and proves its uniqueness.(n(w)) f . Theorem 9.c”.2. Frkchet mean of p with respect to dist is a global maximum of where 6(M) is the diameter of M with respect to ‘dist’. . is a direct consequence of Lemma 9. although our current applications only involve FrCchet means with respect to the metrics p.

However. As in the observation that precedes Theorem 9.(n(w )) is invariant with respect to the induced U(k . it follows from the relation between p and 5 that a point n(z)is a FrCchet mean of the shape-measure f (n(w)>dw:(n(w)) with respect to 5 if and only if it is a global maximum of the function defined by * W z 1) = s.7) Since the uniform measure dw.3. has finite 'diameter'. n(zo) is the unique Frkchet mean of the given shape-measure with respect to p. expressed as a non-increasing function of the distance p from a Jixed point n(zo) that is strictly decreasing on a set of positive measure. the sufficient condition of Theorem 9. cos2 p(n(z )>n(w >> f(n(w >)d&(w >>.2. it follows that \I. then n ( z 0 ) is the unique Fre'chet mean of the given shape-measure with respect to 5 if and only if Proof. Theorem 9.1) is transitive on each set of points at a given distance from n(zo)..Iff is a function of the distance p from a fixed point n(zo). our hypothesis on f implies that for any such induced isometry g that fixes n(zo>. have we Since the stabiliser of n(zo) in U(k . is a function of the distance from . where f can be .2 for that shape to be the unique FrCchet mean is not a necessary condition.1)-action. Suppose that f dw: is a shape-measure on X . is a shape-measure on Xi. Given that the Radon-Nikodfm derivative of the shape-measure is a function of Riemannian distance from a fixed shape. in the case of the metric ? it is possible to give a necessary /I and sufficient condition as follows.UNIQUENESS OF FRECHET MEANS OF SHAPE-MEASURES 21 I FrCchet mean' with respect to any strictly increasing function of p for which Ck. to or to 5. Then. Suppose that f do. (9.2. Theorem 9.

n(zo)is a global maximum of 8 if and only if it is a global maximum along any particular geodesic through n(zo). As in the proof of Lemma 9.9) The metric homogeneity implies that (9. for any s such that 0 < s < n / 2 and any w with w1 # 0.1. 0. + r. (9. = J{ cos2 s cos2 p(n(zo).6). the Riemannian distance p between n(z. Thus.7) and (9./wl r. . ) ) the following homogeneous manner.) and n(w) is given by (9. (9. 0. . the homogeneity of S!j means that we may take 20 in the standard position (1.n(w )> . .>> ( k .-2}k-' j=1 n (2) d d(r. in terms of these coordinates.8) Now. .8) imply that we can express 8 along the above geodesic as 8(n(z. . . . 0) and consider the horizontal geodesic zs = (cos s.9) is equal to the similar expression in with any r.5) and (9. 0) through it. We also know from Chapter 8 that the uniform measure has the expression (k .n ( w )) + k -12 ~ sin2 s sin2 p(n(zo>. We recall that.).6).2 ) ! = (cos s + rl sins) (cos s + sins) dB1 = (k - 2)! s { cos2 s + r: sin2s} (9. for any s such that 0 < s < n / 2 and any w with w1 # 0. in place of rl and hence we can express 8 ( n ( z . .2. . determined by the preshapes with w = 0 has wt-measure zero. As the subset of Ck. .212 MEAN SHAPES AND THE SHAPE OF THE MEANS the given shape n(zo). sins. e''] for 1 < j < k .2 ) ! (1 k-2 dwi(r(w)) = + r: + . it makes no contribution to 8 and again l = we may work with the standard inhomogeneous coordinates (j = wj+. (9.

In this section we shall describe some of these measures and investigate the optimal case in which not only is there a unique FrCchet mean. The case f = 1 of Theorem 9.3. First.a2Z) on R2. also those of Theorem 9. Conversely.2. . but it is clear from the proof that it also applies to such means defined with respect to any strictly increasing function of the Riemannian metric p.2 and so. if and so Q(n(z.3 but not those of Theorem 9. when the given shape-measure on C$ is uniform. the following result follows directly from Theorem 9.a)f a h will satisfy the hypotheses of Theorem 9.2 and 9.2.3 holds and Theorem 9.F~CHET MEANS AND THE SHAPE OF THE MEANS 213 Hence. for any of which we are naturally interested in knowing the relation between the shape of the configuration that is the mean of the joint distribution of the vertices. and the FrCchet mean of the induced shape-measure.2 is not implied by Theorem 9. There is some overlap between our Theorems 9.3 then coincides with the case when the metric is 6 in Theorem 9.3. then the inequality required for Theorem 9. that is. (1 .3. with z{j not all equal. shows that for any fixed point n(z0) E Ck. since the RadonNikod9m derivative of the induced shape-measure of k labelled independent Gaussian distributions N(zGj. we have for all 0 < s < n/2 attains the unique global maximum at s = 0. for sufficiently small a > 0. + 9. since not only does the former apply to FrCchet means defined with respect to any of three metrics rather than just to one. However.)) and only if the inequality in the theorem holds.4 F R ~ H E T MEANS AND THE SHAPE OF THE MEANS The results of the previous section hold for a broad range of shape-measures induced from distributions of vertices. is a strictly decreasing function of distance p from the shape of the means (cf. then. Chapter S).2. that is. but it also coincides with the shape of the means. at n(zo). If h is a function of distance from n(z0) having a positive jump at a distance for which f does not have a negative one and if h is normalised so that hdwi is a shape-measure on Z$. we have From this we can deduce that. if f is a non-increasing function of the distance p of n(z) from a given point n(z0) and f is strictly decreasing on a set of positive measure. that is. the shape determined by the means of the marginal distributions of the vertices.2. Theorem 9. let f satisfy the hypotheses of Theorem 9. as noted above.

. .z & . w . are isotropic on R2k. . using Theorems 9. respectively. With the above considerations.3. like that of Theorem 9. Kent and Mardia. Without loss of generality. That is. . . &)&I. . . which gives us a significant class of induced shape-measures on X that g satisfy the main hypothesis of those two theorems with distance measured to the shape of the means. z k . which are not all equal.1 labelled random variables 21.214 MEAN SHAPES AND THE SHAPE OF THE MEANS Theorem 9. We l may express the joint density as a function f of the squared distance to the origin. 1. . This is equivalent to the density function of x being a function of distance to the origin. which. . and the corresponding points to which they project on the pre-shape sphere by z . . . In the proof of this and later results we shall follow the convention we have used in the general case by denoting points in Ck-' by 2. . These extensions depend on the following result. is the unique Frkchet mean shape of the corresponding induced shape-measure with respect to p. Proposition 9.1 . z i in R2 have means z&. . l). Zk-1 is I . etc. The shape of the means of k labelled independent Gaussian distria21) butions N(z. .3. . which follows a uniform distribution on S2k-1. when written in polar coordinates n = sy. the joint distribution of 21. the shape of the means.G . . . Hence. if k labelled points in R2 jointly follow an isotropic distribution and if we remove the effects of translation by a projection onto to the hyperplane orthogonal to (1.& I . For a random vector x E RZk to follow an isotropic distribution means that. with respect to the uniform measure. To obtain the induced shape-measure on shape space we write 21 = roeiHo we then have and .*~.k is isotropic.1. . .. . on Z of the induced shape-measure is a : function of the distance p of n ( z )from ~ ( z o ) . . . . the Radon-Nikod9m derivative. zz .z. to and to 5. . . then the resulting z" in Ck-' also follows an isotropic distribution. Then. and suppose that the joint distribution on R2k of . . we may assume that Zol = a > 0 and Zoj = 0 for 2 < j < k . Proof. zZk. Suppose that k labelled random variables zT. That. where s > 0 and y E S2kp'. . However. and that the joint density of Z .4. . 00) independently of y . of course.the radial part s follows an arbitrary distribution on [0. This implies that z has an offset uniform distribution on the pre-shape sphere (cf. . is a rather special case. we suppose that k .1 in [w2 have means 201.4.2 and 9. provided those means are not all equal. on R2. we can extend it to other joint distributions of k vertices. . i k . 1997). respectively. .i ( ) k p l is a function of distance to the origin.

n(z)). in is Cos-2(k1) p(=(zo>.11) Although this inequality looks complex. . .2 that the Radon-Nikodym derivative. with respect to the uniform measure. then the condition required for Theorem 9. . the uniform measure on X t has the expression (9.2 ) at n(z) on shape space.n(z)) 00) 3 0 cos . cos d 12" 1 " 2 J'"" .8) and dil . = rikP4dZld(1 .~p(n(zo>.n ( z ) >. when using the inhomogeneous coordinates ((I. To proceed from this result to a proof that the shape of the mean of a joint distribution of the vertices is the unique FrCchet mean of the induced shapemeasure. where ( j = Z j + l / Z l = zj+l / z l .2a i cos p(n(zo).10).Im n o f (a2 + + where i = ro/ cos p ( n ( z o ) . It holds. ~ ( z ) ) 60)} i2k-3i d & . . it is frequently straightforward to verify in practice.2 a i C O S ~ ( T ( ZZ ()Z ) ) C O S ~ O ) } cos6odidQo > 0. =(z >> + i2. with respect to the uniform measure. However.2a ro c o s ~ o . if f is a C' convex function. ( k . there are two routes we may pursue. .n(z)) 6 0 ) cos .( Z I > Z O l ) . .FRECHET MEANS AND THE SHAPE OF THE MEANS 215 and Ilz" -z"01l2 = 11z"1I2 + lZ12. On the one hand.(201.2a i cos p(n(zo).= ( z ) )cos 00) = o +f (a2 i2 2~ i cos p(n(zo). if the function f in the proof of this proposition is differentiable.f ' (a2 + i2 2a i cos p(n(zo). O . n(z)) cos 0. Hence. lOl 21) = a2 + ri C O S . ?2k-2 (9. we need to integrate out ro and 60. scaling effects can be represented by ro and rotation effects can be represented by 60. We recall that.) i2k-3 i do0 d = - { f (a2 + i2. Thus. ' d(k-2. the Radon-Nikodym derivative. For then f' is increasing and so f' (a2 + i2+ 2a i cos p ( n ( z o ) . . of the induced shape-measure be a decreasing function of the distance p is satisfied if and only if -f' ( a 2 + i 2 . to obtain the shapemeasure on shape space induced by (9. This is clearly a function of p as required. for example.

f ( s ) = (1 f (s) = ecCs. is equivalent to b > 0. n/2] and for all cosp. . mean of the induced shape-measure with respect to 5.c < 0 and b > 0.3. with respect to the more general functions mentioned before Theorem 9.216 MEAN SHAPES AND THE SHAPE OF THE MEANS for all i > 0.2 are much weaker than those of Theorems 9. Then. z: in 08' are given as in Proposition 9. Theorem 9. However. which require.2. ~ ( z o ) ) . . E[F'z] must be of the form: Hence. p(n(w ). the shape of the means. H + + We should note that the hypotheses of Theorem 9. By the results of Theorem 9. c < -1. b is greater than zero.(n(w)) = w E[F'zI I t . the shape of the mean of the joint distribution of the vertices is always the FrCchet mean with respect to 5 of the induced shape-measure since the hypotheses of Theorem 9. we have Q(T(w = a )) \.3 and 9.3 are always satisfied. Suppose that k labelled random variables z. so that our necessary and sufficient condition. However. On the other hand. as was shown by Kent and Mardia (1997). for the random variables in Proposition 9. which is now a b > l/(k .3.2. j or 6 or. in particular. that the distributions . ~ ( z o )is the unique Fre'chet .5. if we pull back I the induced shape-measure onto the pre-shape sphere. we have \I.1l). when 25 satisfy the condition given in the theorem. we shall give a summary of Kent and Mardia's proof that ties it in with our present point of view. we only need to show that the induced shapemeasure satisfies the necessary and sufficient condition given in Theorem 9.3. The class of convex C' functions includes the following functions that often arise in practice: (i) (ii) (iii) + s)c. n ( z 0 ) is the global maximum of However.4. f ( s ) = (1 + s)' ecbs.1)a b = 1. that the shape of the mean of the joint distribution of the vertices is the unique FrCchet mean of the induced shape-measure whether this is calculated with respect to p.c > 0. in Kent and Mardia (1997) it is proved that. I if and only if b is greater than zero. from Theorem 9. Whenever we are able to verify the necessary and sufficient condition (9.. which. It is easy to check that E[F'z] is invariant with respect to U(k . in turn. indeed.l).3.3. Proof. by the spectral decomposition.1) and so. Rather than a complete proof. is equivalent to showing that n(zo) is the unique global maximum of the function \.3 and Proposition 9. . defined in the proof of Theorem 9. we deduce. . for all 80 E [O. +b COS* where a > 0 and ( k .

f d ~ ~ ~ k . or 5.5 are unable to tell us whether or not it is the shape of the means. . that satisfies the hypotheses of Theorem 9. Define a function f on the pre-shape sphere S2k-3 by z f ( z >= f((z1.2. attains its global minimum. Then.. and. . n < defined on Ck. f dup-3 is the marginal distribution of d p on S2k-3. for any function h on Xi. or 5. Then.5 UNIQUENESS OF FRECHET MEANS OF ~t GIVEN SHAPES Suppose that {n(zj) E C : 1 j < n } are the n given shapes.1 ) ) = celzl12. 0 . The following is such an example of a non-isotropic distribution on Ck-' whose marginal distribution on S2k-3. induces a shape-measure on Ck.' . 0 ) in s 2 k . . Now. by a Frtchet mean of the n shapes with respect to the metric 'dist' we understand any n(z + ) E X at which the function i .UNIQUENESS OF FRECHET MEANS OF n GIVEN SHAPES 217 on Ck-' be isotropic.3S2k-3 induces the shape-measure on Z. On the other hand. as explained i in Section 9. but for which Theorems 9. therefore. for which.3 . nevertheless. dup-3 is the volume element of S2k-3. Consider the distribution d p on Ck-' defined by where C is another normalising constant. we know that there is a unique Frtchet mean of the induced shape-measures with respect to p . where zo = (1. but not the latter. . There is indeed a large class of distributions that satisfy the former. is Cecos2 a decreasing function of p .n(z)) and.we have on Thus. this shape measure has n(zo> as its unique Fr6chet mean with respect to p. .4 and 9. whose Radonp(n(zo). with respect to dwi. where c > 0 is a constant such that As before. . f (z ) = c ecos2 p(n(zo). 9. so that d p is not an isotropic distribution. so that f is the composite of the projection with a function defined on Xk. it is clear that the density function of d p is not a function of the distance from a fixed point on C k .2.n(z)). z k . . . since this is Nikodfm derivative.

1: if the n given shapes are contained in an open geodesic ball of radius n/8. then they have a unique FrCchet mean with respect to p and this mean lies inside the ball. it follows from (9. we can already see that the uniqueness of the FrCchet mean of the n given shapes with respect to 5 corresponds to the maximum eigenvalue of this matrix having geometric multiplicity one. In view of these results for p and 5. nor that there are no other FrCchet means outside the ball. it will be more convenient to work on the pre-shape sphere. we do not know that it is a FrCchet mean.3 realises the minimum that is. from now on.12) So. if and only if z E S Z k p 3is an eigenvector corresponding to the maximum eigenvalue of the matrix Cs=l T3zj (cf. for the FrCchet means of the n given shapes with respect to the Riemannian metric p we already have the uniqueness result.218 MEAN SHAPES AND THE SHAPE OF THE MEANS To find uniqueness results for such FrCchet means with respect to two of our three metrics. the metric j on shape space can be expressed in terms of pre-shapes by (9. However. whose global minima we shall be seeking on shape space will be given by . Similarly. we shall devote the rest of this section to obtaining a similar condition for the uniqueness of FrCchet means of the n given shapes with respect to j. we see that n(z +) is a FrCchet mean of the IZ given shapes with respect to j?~ if and only if z + E S2k-3 realises the minimum of the function n (9. Kent. Theorem 9. since the set of points on the pre-shape sphere with the same shape as w is {e'"w : 0 < a < 2n}. For example. Hence. in that case. the function F. associated with 5 or p back to the pre-shape sphere to obtain analogous expressions. Similarly. then we still have the weaker result of Proposition 9. we can pull the F..4) that n(z+) is a FrCchet mean of the IZ given shapes with respect to 5 if and only if Zf ~ 2 k . If the n given shapes are only known to lie in a geodesic ball of radius n/4.13) on the pre-shape sphere. So. In particular.. 1992).l(i) that there is a unique Cartan mean on that ball.

. and G. since the partial derivatives of the latter functions along the S0(2)-orbits are always zero. Since the pre-shape sphere lies in Ck-' and the SO(2) action . respectively. in particular.This extension is given by a similar formula to (9.12). where n j= 1 and. taken one from each of the n fibres {ei"lzj : cllj E [ 0 . and the quotient map Ck-' + SC. Since n(z ') is a global minimum of F.is the minimum value of H ( g . that extends the similarly named metric on Ek. g€[0. taken by F. 2 n ) }lying above the shape points n(zj) for 1 6 j 6 n .. for fixed g.are the same as those taken by G. are constant on S0(2)-orbits so that G. and G. we see that - g_E[0.) E [ 0 .' . . and 2 Is. We also write C . = kfi. on the entire space C k . we also have Z c SZ. which will play an important role in our analysis.= n. respectively. We shall denote the size-and-shape space by SCk.. the mean of the corresponding n points in Ck-'.. 2) as Z varies in Ck-'. for g = ( a .. 2 ~ )to . i ) = min h(g). preserves size.' for the set of all such means.2n)" . which. It also follows that the critical points of F. and F.oit and the values.. and G. to Ffingiven by the same expression on the size-and-shape space.... we also extend G.2rr)" TgCk-1 min min H ( g . the minimum values... if and only if z t is a global minimum of G. writing h ( g ) for H ( g . ~ write It will be convenient. in the case m = 2 is the quotient space of Ckp' by SO(2) acting on the left. = F. the norm metric on Ck-' induces a metric 5 on SCk.. but with 2 ( f ) replacing n(z).Moreover. etc. we recall from Chapter 1. to the similarly defined function G. . which. Then..... on the pre-shape sphere. We now extend the above function F... . are the images of those of Gfi. a. . .. by 2..UNIQUENESS OF FReCHET MEANS OF n GIVEN SHAPES 219 For this we need to bring into play the size-and-shape space. and FJL. Note that both G. 2n)")5 ck-' .= { M ( g ) : g E [0. M ( g ) ) .on and G.

2n)".. has a unique global minimum and that it lies in that ball.ll = 1 for each j . in the latter case. those of H . if where g realises the minimum of h ( g ) . ifzt E S2k-3 is a global minimum of G. any minimum point Z" = ~ ( gof )G..14) . Each critical point of G. lies in To obtain our result for FrCchet means with respect to fi we shall first correlate the global minima of G.. Since we are assuming that llz.220 MEAN SHAPES AND THE SHAPE OF THE MEANS Since [0. Then. and the related function h and. Conversely. In particular. the global minima of F. which map under the quotient map n to the FrCchet means that we seek. First. which is essentially the product of n copies of the unit circle. will satisfy this condition to be a local minimum. E S2k-3 : 1 < j < n } be given. M ( g ) ) we find that h ( 4 = H ( c r ... We have already seen that Z" realises the minimum of Gwo and only if = M ( g ) . on expanding H ( g . with those of G.1IM(41I2).. we correlate the global minima of G. Lemma 9. we shall be able to show that F. equivalently. will determine a critical point of h. On the other hand. 2n)". then. so that Z must be a point of maximum norm in Ck-'. Proof. on the pre-shape sphere is (9. then there exists an sli > 0 such that Z"? = stz realises the global minimum of CF0. we obtain a sufficient condition for such a point to be a local minimum. if the n given shapes lie in a sufficiently small geodesic ball. c. then the corresponding Z " is a global minimum of G...-'.. or..3... I f i t E Ck-' is a global minimum of G. with those of GPO.. the minimum value of G... provided the corresponding shape is sufficiently close to the n given shapes... which. We then characterise the critical points of G.. is compact. Similarly. this minimum will be attained in [0. Let {z. M ( g ) )= n(1 .

) . . or h. Lemma 9. E (blezB21. . when subject to c k-1 a. i' It follows from this lemma that the global minima of GPO.. which implies that $(ji(2"). .14) of G.UNIQUENESS OF FRECHET MEANS OF n GIVEN SHAPES 221 However. we shall need to study their general critical points. . . . cos(B1. 3 0 and z = .=I achieves its maximum value realises the maximum value of to solving max 82] = 81.b. as the arguments are allowed to vary. bk-le'B2k-l)with b. . as follows. . 6k-l) in if and only if (2. is now clear..21(2. . for any 2 = (21.=I . . which correspond to those of h.b.. . 3 0. W(Z. However.. since both are determined by points of maximum norm in C . %(*)I2 = 112112 + 151 . A point Zt solution of the equation E Ck-' is a critical point of GPO and only if . As we showed on in Chapter 6 for pre-shapes. we shall need an alternative expression for the metric .M ( g ) ) )as z varies in S2k-3 is equivalent a. bJ20 . $)I. c k. rather than looking directly for global minima of Gw.B 2 . writing M(g) = (ale'B11. eiatG) is real and non-negative. and G.4.. is { The stated relation between the global minima of G. and hence those of F.=1 a. size-and-shape space. ak_le'B1k-l) Ci-'with a. size-and-shape space enables us to find the critical points on of G.b. then. .. that we require.}"2 with the maximal value That is attained at h: = a l / { This implies that the minimum value (9. . M ( d ) )= c k-1 a.. 2k-l) at realises the minimum and I.15) This expression for . For this purpose..G = (GI. Thus. also determine those of G. it follows similarly that. to find z t that % ( ( z .. 112 ifft is a (9.. .1 b: = 1.

. may work on the size-and-shape space SX:. Lemma 9.5. Next. Since f t is a critical point of GPO and only if ?(ft) is a critical if we point of kPo. We differentiate $(it(?). the nature of a critical point of h is determined by the behaviour of h restricted to H. Since h ( g ) = n ( l . a!) is a critical point of h if and only if (9. ?(I?))2 as a function of ? ( i ) SC...4 is that all the critical points of and hence those of FPomust be contained in n(C.11M(g)112) and I IM(g)I I = I IM(a')I I whenever g-0l' has all n of its components equal. . we have the required result.222 MEAN SHAPES AND THE SHAPE OF THE MEANS Proof. In the case on when 21 is real and positive we get Using this to differentiate the expression n j=1 with respect to it(?) in SC. 2 7 ~ of ~ hyperplane perpendicular ) the to L.. then g is a ' strict local minimum of the restriction of h to H. lie in Ck. If we restrict to f E Ck-' having one particular component real and positive. we look at the critical points of h. If these inequalities are all strict.-').. .. rn A straightforward consequence of Lemma 9. . - .-' are corresponding to the fact that GPOand GPO constant along S0(2)-orbits. This gives us local coordinate systems on SC. we see that h is constant along the line segment c. g+ = .16) then it is a local minimum of h. then there is a local one-to-one correspondence between such Z and the corresponding ?(f) in SCk. by using such local coordinate systems. the intersection with [0. Accordingly.

the matrix @ of the Hessian of h with respect to the tangent vectors : 1 < j < n } is {& where and Then. (a1 . if h attains a local minimum at gt. As the product space [0. a n ) @ (u) an (9. 2n)" is flat. a2h . 2 ~ for any ( a l . at the critical point gt E [ 0 ..16) are all non-negative. is equal to it follows that..zJ)}= --gt 2 n (eL(aJ-al) zl))..18) . .z / ) aa aal n + e'(ai-aJ) (zl. For 1 # j . The first statement follows from the fact that Then. at critical points of h.UNIQUENESS OF FRECHET MEANS OF n GIVEN SHAPES 223 Proof. (ZJ. since which._ -1 {e'(aJ-a/) (zJ. then the n real numbers in (9. )~.

it follows that (z+. . for a = arg{(z'. However. eia:zj) real and non-negative. 2 ~ by) ~ g ( f >= (al(f>. I f f t in Ck-' is a critical point of point of h. Thus.i?> = arg{(zt. the expression (9.)}) = Lemma 9.17) holds.4 says that f t is a critical point of GPO and only if M ( g ( i t ) ) . then gt is a local minimum of h. ( f > > (arg{(f. . if the inequalities in (9. I in addition. I IIzjI12 E R.zl)l. and also a sufficient condition for a critical point of G. .17). g(ft) satisfies the necessary and sufficient condition (9. ~ ( ~ + ) z j .: never has all its components equal and so. Assuming now that p(n(z').eia:z. with strict inequalities in (9. .18) is equal to zero if and only if all aj are equal. Lemma 9. .zj)) = . if Proof. In the following lemma we obtain a relation between critical points of GPO and those of h.then gt = g(ft) is a critical P(X(Zt)?X(Zi>>< n/4.6.. Moreover. to determine a local minimum of h. since ft = aj(. . a tangent vector to H. z j ) ) .) Let = cos(p(n(zt>. the restriction of h to Hat will have a strict local minimum at (211.. then this will always be non-negative and hence gt will be a local minimum.n(zj>)> 3. arg{(f.224 MEAN SHAPES AND THE SHAPE OF THE MEANS If condition (9. For that purpose we define a function g from Ck-' to [O. we : have ( f t .~ ). i =) ~ ( ~ t . n ( z j ) ) < n/4.arg{(zj. f + ) ) .For such it.17) are all strict. a . z j ) R + E l and so = ~ I -(7 Z z .. since.16) for it to be a critical point of h. for 1 < j < n j 6. we have ( e ' .". > 1 -t 1 20 = xz 1- . which is strict when h is restricted to Hat.z.

Firstly. l ) .l12 = $<%(z.. Since all the given shapes also lie in &/s(n(zo)). which is strict when h is restricted to Hat./x(n(z. is any global minimum of F..4. Since X i is compact.15).15). %(z. then we would like to know whether any of the FrCchet means also lie in that ball. %(E..>> < 2. .). we work on a geodesic ball or other submanifold containing the given shapes. This is particularly relevant in practice when we have found an explicit mean by algorithms or other methods. Proof. as we commonly do. that n(z t . This Frkchet mean will lie in Bx/8(n(zo)) and may be identijied as the unique point n ( z t) in that ball such that there exists a real positive st for which stzt is a solution of (9. the . We are now ready to discuss the FrCchet means of our n given shapes. FrCchet means exist and by Lemma 9. Secondly. - ZJI + I le"izl I t -2 . by the triangle inequality. Ifthe set o n given shapes. We are now able to answer these questions under sufficient hypotheses on the n given shapes.3 shows that there is st > 0 such that stz is a global minimum and hence a critical point of G.1./s ( n ( z 0 ) ) . which are similar to some of those we have used earlier for general shape-measures.~)). if. . which. and Lemma 9. 1 Thus. However. by Lemma 9.z.. is contained f in a geodesic ball &l~(n(zo))with centre n(z0) and radius n/8... 112. FrCchet means will always exist. z will be a global minimum of G. that it is a solution of (9.To establish uniqueness we show that such FrCchet means are characterised as stated in the theorem. I le'a.e'a:z.15) with n(zt) in U. + I~ > By Lemma 9. e'.6.))2 = 1 + - 1/2cosp(n(zt).5.UNIQUENESS OF FRECHET MEANS OF n GIVEN SHAPES 225 Then. { n(zj) E Eg : 1 6 j 6 n } . II < 1/2 = $(%(z. for each pair ( j . as already mentioned in Section 9. 2(&) + $ ( 2 ( Z l ) .>. then. is a solution of Equation (9. since shape space is compact. this implies that gt is a local minimum of h. Theorem 9.> >5 1 { I I ~ I~I ~I ~ II -~ 1 = 0. given the hypotheses of this theorem.)) and it follows that 8 ((e'":z. z(z. they all lie in B. Conversely.3. Of course.e+zl II < I lela:z. then those shapes will have a unique Frkchet mean with respect to $. to show that any such critical point of GFL0 projects to a FrCchet mean. it is very desirable to know when there is a unique FrCchet mean. we may assume. without loss of generality. Granted.:zl. we have. in practice we need to know rather more than that.

which are generally referred to as the ‘partial’ and ‘full’ procrustean mean shapes. rotation and scale change. Hence.8(n(zo)) must be a strict global minimum. z.. and n(zt) is a strict global minimum of F p o . . of the geodesic ball B.. 1995a) and we shall show that they are. We note that the argument of the previous paragraph in fact shows that any critical point of k. in fact. just the way in which we obtained induced metrics on shape. there must be precisely one of them and so there is also precisely one FrCchet mean on 6 r / 8 M z o 1). which is strict when h is restricted to Hat. The basic principle involved is to use as a measure of the difference between two configurations of k labelled points in Rn the square-root of the sum of the squares of the Euclidean distances between corresponding points after an optimal matching of the two configurations under such operations as translation. .(. as usual. 1991. it is homeomorphic with a cone on B. . or size-and-shape. which. in effect. are isolated and all are global minima. have strict minima at those points. similarly. This is. when restricted to a local codimensi& one submanifold transverse to the S0(2)-orbit through Z t . space from the norm metric on the configuration space Rmk.6 are satisfied and hence gt = g(it)is a local minimum of h.p(n(zo)). . if all the critical points of F. 1992 and Le. with image in Bn. since it must be the image of a critical point of GPO.6 PROCRUSTEAN MEAN SHAPES In the final section of this chapter we show that so-called procrustean means are special cases of FrCchet means and use our results to answer some questions that naturally arise in procrustean analysis. Specialising to the case m = 2 and writing. Hence. It follows that G. Procrustean methods are a much used practical means for comparing two configurations in terms of their vertices. the partial and full procrustean mean shapes implicitly involve the metrics on shape space. will satisfy Equation (9.226 MEAN SHAPES AND THE SHAPE OF THE MEANS hypotheses of Lemma 9.. 9. which we have labelled fi and 5.. We shall continue. G.15). (cf.. to work with the case of configurations in R2. z* for the complex k-ad representing the vertices of a configuration. Kent. in turn. FrCchet means with respect to these metrics..In particular. respectively. ?(Zt) is. in fact. However. a strict global minimum of p. a partial procrustean mean shape of the n points z. and. Goodall.* in the configuration space Qk is the shape of a configuration (z*)? of norm one that realises . respectively z t. as for most of this chapter.. the pre-image in SC.(n(zo)) is connected: in fact. where we shall mainly discuss the two most commonly used means.

PROCRUSTEAN MEAN SHAPES 22 7 where 25. the partial procrustean mean shapes of the n given configurations are. 2n). To compute such means we denote the composition M o by g.12) and (9. As pointed out at the beginning of the previous section. from the expression (9. denotes the centroid of the components of z. We consider first the practical computation of procrustean means. Thus.. ZSzj. However.21) we minimise over the a. Similarly... a FrCchet mean of the n given shapes { n ( z j ): 1 < j < n } with respect to the metric 2 is the shape of an eigenvector corresponding to the maximum eigenvalue of Cy=. and l k denotes the row vector in Ck with all its components equal to one. where u j ( z ) E S' are given by I if ( z . is the shape of a configuration ( z * ) of norm one that realises ~ min (9. we may now use our work on the latter to answer questions about the former. . in fact. > 0 . defined there on the pre-shape sphere.11z*11 = 1 . Similarly. ~ that is. there is generally no closed form for FrCchet means of the n given shapes with respect to fi.4) for 2 we see that their full procrustean means are the corresponding FrCchet means with respect to 5. a full procrustean mean shape. first.13) that z is the minimum of G. If in (9. z j )=o. the FrCchet means of the corresponding shapes defined with respect to fi.20) It follows that the partial and full procrustean means may be calculated on the pre-shape sphere as the shape of any z t that realises r C llz* . Having identified the various procrustean means as special cases of FrCchet means. 1 and the shape of any z that realises respectively.cj1k j=1 - a j e ' " ~ z ~ / 1 cj E C. aj :2 aj E [o. we see from (9. defined on the configuration space.

Ifthe shapes ~ ( z jof the n configurations z s all lie in a geodesic ) ball & / ~ ( n ( z o > ) .4 using the definition of Q".a 2 Z 2 k ) for. as was - .the generalised then procrustean algorithm has a f i e d p o i n t f t such that n ( z t. we tend to take the shape of any stable point of this algorithm to be 'the' FrCchet mean of the n given shapes with respect to . is a non-increasing function of the distance p from n(zo). which shows consistency of both estimators under certain conditions on one type of simple statistical model. { s j Jare iid errors in R2k. we see that it is just the set of critical points of Gfio.8.2 and Ziezold's Strong Law of Large Numbers. for any given W O . We turn finally to the question of the consistency of the procrustean mean shapes defined by (9. in that case. n(Ck. of then the quotient set of Q".228 MEAN SHAPES AND THE SHAPE OF THE MEANS We denote the set of fixed points of g by Q".2n) and aj > 0 are unknown nuisance parameters. we have the following theorem. the shape of the means.-' by SO(2) acting on the left contains exactly one element. Theorem 9.6 gives us the following uniqueness result.-') is the set of critical points of Ffio and. We can now establish conditions under which this assumption is justified. I$ f o r any given starting point W O . + An important special case of this theorem is the consistency of these means when 81 N(0.Thus.19) as estimators of the shape of the means under various statistical models. By Theorem 9. the intersection of that ball with the projection onto Ck. On the other hand.18) or (9. of z$ ~1 in Ck. If the induced RadonNikodjm derivative. we get the following result on fixed points of GPA. the hypotheses are satisfied. In practice.-'. by Lemma 9. in particular. with respect to do:. Suppose that {zs : 1 6 j 6 n } is a set of independent configurations of k labelled points in R2 each obtained from the model: where z$ is a fixed configuration. n(z+)is the unique partial procrustean mean shape of the n given conjigurations. Theorem 9. and that element is the unique partial procrustean mean shape of the n given configurations. their . Suppose that the shapes n ( z j ) of the n configurations z.4. Qi-' Proposition 9. Reinterpreting Proposition 9. which is strictly decreasing on a set ofpositive measure.7. lies in Br~8(n(zo)). and cj E Q. Theorem 9. the sequence {w.4.+1 = g(wn) n 3 0) is actually the same as the : sequence obtained by the well-known generalised procrustean algorithm or GPA. then both their partial and full procrustean mean shapes will tend to n(z0) as n tends to infinity. all lie in a geodesic ball &/s(n(zo)). that is. aj E [0.' so that Q".' is necessarily contained in the image of M and. procrustean mean. contains all its local minima.

when its hypotheses are satisfied. we can in practice use either the partial or full procrustean mean shapes to estimate the shape of the means. Further results may be deduced similarly from Theorems 9. the shape of the means. we can use the partial procrustean mean shapes to approximate the full procrustean mean shapes or vice versa. Proposition 9. while the former gives necessary and sufficient conditions on the induced shape-measure for consistency of the full procrustean mean. as n tends to infinity. In addition.5. More generally.4. then the partial and full procrustean mean shapes will both tend to the shape of z. in those circumstances and for large sample size.3 and 9.8 the errors ~j are isotropic and satisfy the conditions mentioned in that discussion. The latter shows that. if in Theorem 9. it is sufficient that the ~j in the model considered in Theorem 9.3 and the ensuing discussion imply that.PROCRUSTEAN MEAN SHAPES 229 remarked before the statement of Theorem 9.8 that. . as sample size tends to infinity. For then both the procrustean mean shapes tend to the same limit.8 be isotropic. It also follows from Theorem 9. for the consistency of the full procrustean mean.

We shall illustrate the problem. For example. this may be considered the first non-standard space in the family of shape spaces. as the dimension and the topological complexity increase. Of course. and our approach to its solution. G.Shape and Shape Theory D. Obviously. BARDEN & T. 10. even more crucially. C A R " & H. to represent the eight-dimensional shape space E: in a similar manner to that which we propose for X i would require 672 diagrams. in fact. Since the projective spaces Ei may be considered well known and since Ei is a five-dimensional manifold with singularities homeomorphic. KENDALL & D. corresponding to that number of eight-dimensional cells. there is no perfect or final solution to the problem. the problem rapidly. becomes more complicated and difficult. is also non-Euclidean. our technique for retaining as much fidelity as possible will be to break the space up into a number of sizeable chunks and represent each chunk on a two-dimensional diagram. in some sense exponentially. K. we shall attempt to describe how we may ‘look at’ stochastic phenomena in Ei sufficiently clearly to get some feeling for what is happening. They arise as follows.1 THE TWO-DIMENSIONAL REPRESENTATION OF Xi Since a two-dimensional representation of a five-dimensional space certainly cannot be faithful. together with an indication in terms of these diagrams of how the corresponding chunks fit together to form the shape space. LE Copyright 0 1999 by John Wiley & Sons Ltd CHAPTER 10 Visualising The Higher Dimensional Shape Spaces In this chapter we discuss the problem of trying to visualise what is happening in a shape space that is not only of a higher dimension than those with which we are familiar but. . We hope thereby to persuade the reader that ‘even when the desirable is unattainable. 1995). Our ‘chunks’ will. a second best can be quite startlingly informative’. with S5. but not isometric. However. be topological cells and there will be 24 of them.b. by considering the representation of the space Ei that was developed by Kendall (1994a.

and A are uniquely determined by the shape up to multiplication by matrices in D. x2. h2. of which an even number are -1. the corresponding shape lies in the non-singular part of the shape space and we can make a unique choice by restricting A to lie 1 in Si = {(XI.x3) E S2 : x > x2 > Ix31}. representation of the factor Si. conformal. the passage from pre-shape to shape corresponds to omitting the special orthogonal matrix U .Then. starting at Figure 10.1 The projection of S i with closure points: the indicated sets of coordinates are ( h l . though not isometric. hZ>A31 .5.=l The set Si is an open spherical triangle that we can most faithfully project onto a plane by rotating until its centroid is on a horizontal axis. then the diagonal matrix A has entries h l . which lie in S0(3). which represent one twenty-fourth of the surface of the unit sphere are rather different from those we Figure 10. The result is a bellshaped triangle of the form displayed in Figure 10. The matrices V.232 VISUALISING THE HIGHER DIMENSIONAL SHAPE SPACES We recall that the pre-shape of four points in a 3-space is represented by a 3 x 3 matrix of norm one. This group D is comprised of the diagonal matrices with entries f l . z. When the hi are distinct. the central vertical axis represents points where 1 3 = 0 and along which hl ranges from 1/& to 1. which can also be seen in the figures that appear later in the chapter. Thus. projecting it horizontally onto an enveloping cylinder with a vertical axis and then flattening out that cylinder after cutting it along a suitable generator. If it has a pseudo-singular values decomposition U A V . such that h? = 1. Here. which is a one-codimensional subset of the non-singular part. the part of C i where the corresponding pre-shape matrices have distinct pseudo-singular values. This gives us a bijective. We note that these bells.1. is parameterised by Si x S0(3)/D. the centre of SO(3).

Except. To represent shapes where some of the hi coincide we need the boundary of S.THE CELL-DECOMPOSITION OF SO(3) 233 used in Chapter 1 to represent one twelfth of the surface of the sphere of radius one-half. -l}. . after which the third is uniquely determined. these extra quotients do not correspond to singularities of the shape space. We have 24 vertices corresponding to matrices in which two entries are chosen to be f l . In particular. 0 +1 0 0 0 0). where the additional lower suffix indicates the column permutation needed to obtain the matrix from the corresponding diagonal one. each of which is contiguous with others across the three 2-cells that form its boundary. represented by the topmost vertices of our bell-diagrams. Since we shall need to refer to these explicitly later.+ and V-. V + . The extra information contained in the factor S0(3)/D is only utilised discretely. Then. The singularities of Ei occur at the points where the rank of UAV is one. we may record its two-dimensional Si-coordinate in the copy that corresponds to the 3cell in which its remaining coordinates lie. +1}. Then. 10. there are six possible choices for the first row and then four for the second. the cells in our decomposition are determined by the various special cases that can arise. Our representation of C i then consists of 24 belldiagrams. with V . Since these cannot lie in the same row or column. however. In the next section we shall describe a cellular decomposition of SO(3) which has 96 3-cells that are freely permuted by D.2 THE CELL-DECOMPOSITION OF SO(3) We represent an element of SO(3) by a 3 x 3 matrix with orthonormal rows and columns. it will appear in more than one bell-diagram because of the necessary extra quotienting of SO(3) mentioned above. when two of the pseudo-singular values h. +1.i coincide. this gives D us 24 quotient 3-cells. -1. h.1 = 1. we adopt the notation V + + for the matrix diag{+l.2 = h.defined similarly in terms of the first two diagonal entries. there is less distortion of the metric in these bells. that is. Since this extra quotient will not be explicitly included in our representation. we may define v++ . together with an indication of which cells are contiguous.. At these boundary points the second coordinate factor SO(3) has to be quotiented by a larger group than D.1 with the corresponding coincidences.. Then. one for each cell of S0(3)/D.for diag{+l. which is marked in Figure 10. Since I 1 = 4. where h. for a general point in the non-singular part of Ei.(1 (12) . since now the matrix A also commutes with all rotations in the plane corresponding to the coincident hi. that. it needs to be kept in mind when interpreting the pictures we obtain.3 = 0. 1 v&= 0 +1 (0 0 1 0 0 etc. rather they are singularities of the representation. In particular.

(n/2. we have 72 edges or 1-cells lying. a compartment is determined by the sign pattern of the entries in the matrices that comprise it.. for our purposes it is more convenient to note that. Each a-circle meets each j3-circle in a unique vertex and so the induced decomposition of the torus has 16 vertices. It contains eight of the 18 circles: namely. although. {( cos a cos j3 sin j3 O . A circle here is the set of all matrices (of SO(3)) in which a particular entry is 5 1 . -} and is then formed of all the . the signs of all the entries in the corresponding matrices of SO (3) retain a pattern characteristic of that 2-cell.2. S:. the set.2n). since this is not so immediately obvious. At the two-dimensional level we have 144 faces or 2-cells lying. Thus. where a torus is formed by the set of all matrices with a zero entry in a given position. For example 0 1 0 -sina and 0 cosa -1 0 0 { 0 (cosa sina 0 sina . matrices with the sign pattern (10.(n. For example. which. respectively. /3 E [0. Thus. of four disjoint B-circles and the set. The four open 1-cells. ~ 2 ~. However. n). 3 ~4 independently in {+. by continuity. of four disjoint a-circles. and (3n/2.2n). the 16 2-cells on this torus are determined by requiring each of a and B to lie in one of the four ranges (0. each of the 16 2-cells above is determined by a choice of & I . . : 2 which we shall denote by S and S. as a and j3 vary in any such pair of ranges. and Stl. n/2). we shall refer to them as ‘compartments’. 3n/2) and (3n/2.cos a cos B ) : a. together with the relevant vertices. which we shall denote by T11. Once again. Once again. n/2). (n/2. four at a time. n).234 VISUALISING THE HIGHER DIMENSIONAL SHAPE SPACES Next.1) The three-dimensional components of our decomposition will also turn out to be 3-cells. with the lower suffix indicating the position of the special entry and the upper one its sign.co s a ) : a E [0.2n)}. on 18 circles. and S f 3 .sina sin j3 sin a cos j 3 sin a cos a sin j3 . form a particular circle are determined by restricting a to lie in the four ranges (0. 32 edges and 16 faces arranged as in Figure 10. on nine tori. 3n/2) (IT.2n) is clearly a torus. 16 at a time.

-}. Thus.2) determined by 61 = ~2 = ~3 = ~4 = the 2-cell boundary component + above.2 Decomposition of T I 1 each 2-cell (10. edges. G. :!) on the torus T11 that we have been considering.1) above will be contiguous with the two compartments having sign patterns ( where again 60 E Eo E3 El -E2&4 E2E3 E2 E1E4 -EIE3 ). at the same time. change the sign of a row or a column. circles. 2-cells and tori. We consider the subgroup. Accordingly. This group G is easily seen to be transitive on each of the sets of vertices. To count the compartments we need to know how many such 2-cells lie on the boundary of each compartment. as we have just seen. E4 [+. of isometries of SO(3) generated by those that interchange two rows or two columns and. which. with sign pattern (10. has (. it suffices for us to examine the faces of one particular compartment adjacent to one particular 2-cell. . we look at the 3-cell.THE CELL-DECOMPOSITION OF SO(3) 235 Figure 10.

form the entire boundary of the compartment. By this criterion we see that the compartment with sign pattern (10.3(a). . 1 0 0 0 V. Let us now look at the vertices.i2. We find that they each have both of the vertices 0 1 V&i.3(b) with the vertices labelled. !l). there must be 9 x 16 x 2/3 = 96 compartments altogether. where the relevant parts of the standard picture.2) are shown in Figure 10. of these 2-cells. Since no other entry in the 3-cell matrix satisfies the necessary criterion. with no two tori ever sharing a common 2-cell. As each compartment has three 2-cell faces. Hence.236 VISUALISING THE HIGHER DIMENSIONAL SHAPE SPACES To find other 2-cell boundary components we note that the special form of sign pattern (10.= ( 1 0 0 0 (e 0 0 -1 0 1 0).2) above will also have the boundary 2-cells on the torus T23 and (::!) 0 0 ) -1 and V&3)= on the torus T32. and each of these 2-cells is one face of two different compartments.= (i 0 l . The number is determined by the facts that each of the nine tori has 16 2-cells. as indicated in Figure 10. = ( 0 0 I ) .2) for a 3-cell by a zero.1) in such 2-cells means that we cannot replace an arbitrary sign in the pattern (10. This is the classic picture of the double suspension of a triangle with each pair of triangles suspending the same edge of the basic triangle amalgamated to form a creased rectangle. each of which is a 3-cell. that appear on the boundary of the cell with sign pattern (10. and their bounding edges and vertices.3. 0 together with two of the three vertices 0 1 0 V&13. It is now clear that the decomposition is completed by 96 compartments. the boundary of the compartment is homeomorphic to a 2-sphere and it follows from the topological fact that SO(3) Z RP3 is indecomposable that the compartment itself must be a 3-cell. Figure 10. = ( 1 0 0 1 0 0) 1 0 and V f . and hence the edges. It is not difficult to check that we can do so if and only if the complementary 2 x 2 submatrix has determinant zero when the signs f are replaced by the numbers f l . these three 2-cells.

faces and 3-cells are all seen to have four elements and hence the induced decomposition of SO(3)D has six vertices. (b) Double suspension with amalgamations: comprising three creased rectangles AUBV. (1 1 I). write X j for the matrix obtained from X by changing all the signs in its i ith and jth columns. BVC. + + (: I T)? (. we write A . CVA. BUCK CUAV 10. and.i).4 we represent each quotient 3-cell in SO(3)D by the sign pattern of the member of the orbit that has the signs '+' at the beginning of each of the first two rows. edges. j)th entry remains unchanged. . C . 18 edges.+ 1 T)? (++ + + . F for the basic sign patterns (+ (.3 THE ACTION OF THE GROUP D We recall that D. F } . + +) - + + + + . E . D . Since . x where X E { A . The orbits of vertices.X23. Then. so that the (i. the complete set of 3-cells of S0(3)/D is represented by the 24 sign patterns X. B . B. similarly.3 (a) Double suspension of the triangle ABC: comprising six triangles AUB. each of the four elements of D simultanously changes all the signs in two of the rows of V.THE ACTION OF THE GROUP D 237 V v=v+13 Figure 10. 36 faces and 24 3-cells. Each such operation clearly preserves the above decomposition of SO(3) and has no fixed points. the centre of S0(3). In Figure 10. Thus. 3 2 and 3x3. E . C. BUC. + ) + + - + + + + + + and then write i X j for the matrix derived from X by changing all the signs in the ith row and jth column.I. acts on the left of V E SO(3) in a psuedo-singular values decomposition. CUA and AVB. D . Then.

However. It will not always be possible for the sign difference to occur between our chosen representatives. in Figure 10.3 0 3 . From our previous discussion it will be clear that two 3-cells of S0(3)/D are contiguous if two of their representative 3-cells in SO(3) differ in their sign pattern at just one location that satisfies the criterion for it to determine a 2-cell when those differing signs are replaced by zero.4 we display our 24 chosen patterns in such a way that adjacent members of each circle of 12 patterns represent contiguous 3-cells as well as the four pairs A . We need to identify and record which three.238 VISUALISING THE HIGHER DIMENSIONAL SHAPE SPACES Figure 10.4 The complete set of adjacencies each 3-cell has three facial 2-cells. 3D2 . For example.A23 and 3A2 . the leading diagonal entry in B=(+ + + + + + - +) . The remaining eight pairs of contiguous cells do not lie close together in the diagram and are indicated by the eight curved lines.0 2 3 . D .3A3. it is contiguous with three other 3-cells.

we get (: + + . It will be observed that.THE GEODESICS OF C! 239 satisfies the criterion for an adjacency to occur. When we change it. including such pairs as 3C3 .5 we see the projection of a typical geodesic with ‘jumps’ between 3-cells of S0(3)/D indicated by dotted lines. On closer inspection we see that these jumps in fact take place between points on the left-hand and right-hand boundaries of the relevant bells.3 0 3 . is our representation of X In each of the 24 circles we shall : . possible.E23 and 3F2 . but there is no way to arrange that all contiguous pairs of cells be placed adjacent to each other in the diagram.( 2 2 3 . ) which is not one of our representative patterns. together with the sign pattern that labels the corresponding 3-cell in S0(3)/D.4 THE GEODESICS OF X ! As a first application of our visualisation technique for Xi. of course.F . etc. : + 10. but the reader needs to bear their existence in mind when interpreting the results that we shall display. This.4 the eight adjacencies between cells B and 3E2. These are points where the pseudo-singular values satisfy h.3B3. it is in the same Dorbit as (t I + . A . for each given 3-cell. At the cost of increasing the number of extra lines to 12 we could arrange for all cells contiguous to the cells in the outer ring to be adjacent. If we flip the inner circle of cells about its vertical axis. Other arrangements of the cells are. incribe a copy of our projection of S. as we would expect.3 = f h . using the results produced by Kendall (1995). Note that the Si-factor means that each k-cell in S0(3)/D represents a ( k 2)-ce11 in c. the complete set of three 3-cells with which it shares a face. However. n/3.- I). In Figure 10. In Figure 10. 2 . etc. then. we obtain a diagram in which alternate pairs. these jumps take place between contiguous 3-cells with a common face. are indicated by the longer curved lines to show. In the interests of clarity we shall not include the lines indicating the non-adjacent pairs that represent contiguous 3-cells. We also see. However. which are not adjacent in the diagram. which is the cell 3E2. 2n/3 to the x-axis. E23 . C and 3F3. 3B3 . after changing the signs in the second and third rows of . there are also a number of jumps between E23 and 3C2 that do not share a common face.i. There are then just six pairs of non-adjacent contiguous cells lying on the three diameters at angles 0.we will ‘look at’ some geodesics.. have all the 3-cells that are contiguous with them in S0(3)/D placed adjacent to them in the diagram.

by permission of the Applied Probability Trust 3C2 that both the pattern E23 and 3C2 have on their boundary the 1-cell of ST3 specified by (I 4 i). E 2 3 . 3 c 2 . How to look at objects in a five-dimensional shape space: looking at geodesics. Advances in Applied Probability 27 (1995). that these eight cells are c 2 3 .240 VISUALISING THE HIGHER DIMENSIONAL SHAPE SPACES Figure 10. 0 2 3 . with a maximum angle of 3n/4 between the curves that . each edge must lie on eight 3-cells. Clearly. Kendall.5 Reproduced with permission from D.G. 35-43. as there are 96 3-cells and 72 edges. 3 E 2 and 3 F 2 in the cyclic order in which successive pairs share a common face. the two faces of each of these 3-cells that meet along the common edge must do so at an angle of n/4. To understand what is happening here we should recall that each 3-cell has six edges in our decomposition and. by symmetry. on changing all the signs in the last two rows of the patterns 3 x 2 . so that corresponding faces of E 2 3 and 3 C 2 meet at right angles. In the case of the edge of SC3 specified by the above pattern we may easily check. F 2 3 . 3 0 2 .

h2.Advances in Applied Probability 27 (1995). when we superimpose all the geodesic segments in the various bells onto a single bell. larger bell in Figure 10.G. by permission of the Applied Probability Trust .5. Kendall. for it to be possible for the geodesic to emerge from a different 3-cell. h3 are continuous functions of shape.THE GEODESICS OF C: 241 meet that edge from the two 3-cells. How to look at objects in a five-dimensional shape space: looking at geodesics. this is true of the eight corresponding 5-cells in Ci since they are identified along their common three-dimensional face. However. since the geodesic is ‘straight’ in terms of the metric structure of X:. this projection of the geodesic is also smooth.6 Reproduced with permission from D. Since we are in the non-singular part of X i and geodesics cannot turn corners and since its full projection on the bell-diagram is not tangent to the edge at the point where it moves from one cell to the other. these eight 3-cells must be twisted since they are identified along their common edge. Moreover. as we have done in the central. The parameters h l . 35-43. we must conclude that the A-coordinates of the geodesic remain constant for a while before the geodesic emerges in the different 3-cell.6 Figure 10. Generally. More precisely. we get a continuous image. so. in Figure 10.

Another possible source of non-smoothness is illustrated by the geodesic in Figure 10. the geodesic passes through the singularity set of the shape space.2 = h. the geodesic has moved from 3A3 to 3F2 across an edge of their common face. Advances in Applied Probability 27 (1995). Here. by permission of the Applied Probability Trust we show the portrait of a geodesic with a pair of cusps caused by the Sicoordinates along the geodesic retracing their path while the S0(3)/D-coordinates remain constant. resulting in a discontinuity of the tangents to the full projection of the geodesic in the central bell representing the Si-coordinates. . from 3Ag via E23 to 3F2. h. In other words. As it does so. This geodesic also displays two double ‘edge jumps’ similar to those in Figure 10.5. which does have a face in common with 3A3.3 = 0. for example. beyond that which is explicitly contained in our representation. and even the structure of shape space itself.7. this time it goes.7 Reproduced with permission from D.G. How to look at objects in a five-dimensional shape space: looking at geodesics.3-coordinate changes sign. the h. These examples illustrate how it is possible to make inferences about the behaviour of the geodesics.242 VISUALISING THE HIGHER DIMENSIONAL SHAPE SPACES Figure 10.1 = 1. 35-43. However. which in C$ is characterised by h. Kendall.

242-247 .8 Reproduced with permission from D. Figure 10. How to look at objects in a 5-dimensional shape space I: Looking at distributions. Teoriya Veroyatnostel (1994).9.8. The first of these. Despite the apparent crudity of our representation. Figure 10. 39. plots the shape points of tetrahedra associated with 5000 independent 3 x 4 Gaussian matrices. In the next diagram.SOME DISTRIBUTIONS ON C. 243 10. Figure 10. Kendall. we display the shape points associated with the independent random tiles of Poisson-Delaunay tessellations.G. the different nature of these two distributions is clearly visible: the striking empty regions lying at the top of each bell and around its median vertical line in the Poisson-Delaunay case corresponds to the fact that a Delaunay tetrahedron is most unlikely to be nearly flat. we display the projections of two of the distributions of random shape points that we discussed in Chapter 8.5 SOME DISTRIBUTIONS ON As a second example of our visualisation process for Xi. in contrast to the Gaussian case where there is no bias against flatness.

244

VISUALISING THE HIGHER DIMENSIONAL SHAPE SPACES

Figure 10.9 Reproduced with permission from D.G. Kendall, How to look at objects in a 5-dimensional shape space I: Looking at distributions, Teoriya Veroyatnostel (1994), 39, 242-247

10.6 A DIFFUSION ON C i
We look now at the progress of a diffusion on Xi. We shall use the process described by Kendall (1994b), where he constructs a discrete model analogous to the way in which a random walk can be thought of as a discrete model for Brownian motion. The straightforward example he works with is sufficiently representative and yet enables us to forget about most of the technicalities. In contrast to the case of a geodesic, where its path to its present position determines the direction in which it must continue, successive steps in our process are independent and we shall observe much freer action around the edges of the 3-cells, and even across the faces, than we did for geodesics. We start by choosing four labelled iid Gaussian points in 08’. We then displace these so that the centroid of the tetrad lies at the origin and then scale the

A DIFFUSION ON C i

245

configuration to have unit size in the usual sense of shape theory. We then record the shape of this standardised tetrad on the appropriate bell-diagram. At the next stage, we displace each of the four points of the standardised tetrad by independent three-dimensional Gaussian perturbations of linear standard deviation E , re-standardise the resulting tetrad and plot the corresponding shape point. Continuing inductively in this manner, we generate a stochastic process whose only parameters are the value E and the initial randomising seed. The start of the resulting process is displayed in Figure 10.10. We see the process starting in the bell-diagram that is labelled 3E3 in Figure 10.4. At first, it travels about that bell in a more or less Brownian manner, allowing for the discrete nature of our construction. However, after a while there is a jump, indicated by a dotted line, into the bell labelled B23 and, after some Brownian-like

Figure 10.10 Reproduced with permission from D.G. Kendall, How to look at objects in a 5-dimensional shape space 1 : Looking at diffusions, in F.P. Kelly (ed), Probability, 1 Statistics and Optimization, Wiley (1994), pp. 315-24. Copyright John Wiley & Sons Ltd

246

VISUALISING THE HIGHER DIMENSIONAL SHAPE SPACES

Figure 10.11 Reproduced with permission from D.G. Kendall, How to look at objects in a 5-dimensional shape space 11: Looking at diffusions, in F.P. Kelly (ed), Probability, Statistics and Optimization, Wiley (1994),pp. 315-24. Copyright John Wiley & Sons Ltd

motion there, into the bell labelled 3E2. These are, of course, only apparent jumps, not true ones, caused by our suppression of three of the five dimensions in which the process is taking place. However, it is possible to infer something of what is happening in these remaining dimensions. For example, the 3-cells 3Eg and B23 share the common face formed by matrices with the sign pattern (10.1) in which ~1 = ~2 = - and ~3 = ~4 = f. Obviously, the jump took place across that face between the two 3-cells. On the other hand, the 3-cells B23 and 3E2 do not share a commom face. However, we notice that they do both have the same I-cell of the circle Sll on their boundary, namely, the 1-cell specified by the pattern

A DIFFUSION ON C!

24 7

In fact, looking more closely, we see that the jump did not occur directly between and 3E2 but paused briefly at the 3-cell 3 0 2 , which also has that 1-cell on its boundary. So, it is clear that in the unrecorded dimensions the coordinates remain on or close to that 1-cell as the process moves between the three 3-cells. In Figure 10.11, which gives a later stage of the diffusion, as well as the further diagrams in Kendall (1994b), we see that these observations remain true: the jumps take place between 3-cells with a common face or a common edge. Moreover, it is clear from the later development that, as we would expect, most jumps take place between two 3-cells with a common face and relatively few between those with just a common edge. We also note that, again as might be expected, after a jump across a common face, the process will either remain in the new cell or jump back to the one whence it came, whereas, after moving to a new 3-cell sharing an edge, it is equally likely, as in the example above, to pass on to a third 3-cell sharing that edge.
B23

Shape and Shape Theory
D. G. KENDALL & D. BARDEN & T. K. C A R " & H. LE Copyright 0 1999 by John Wiley & Sons Ltd

CHAPTER 11

General Shape Spaces
11.1 SHAPE SPACES FOR GENERAL MANIFOLDS
In Chapter 1 we considered the shapes formed by a k-ad of vectors in the Euclidean space R”. Two k-ads have the same shape if there is a Euclidean similarity that maps one k-ad on to the other. The set of Euclidean similarities forms a Lie group, which acts on the set of k-ads. The orbits of this group action are the shapes in the space Xk.Later chapters showed how we could study the geometry of the shape space by using the simpler properties of the space of k-ads. In this chapter we wish to examine a more general situation that will allow us to consider shapes for k-ads of points in other manifolds, such as spheres or hyperbolic spaces. The general situation we wish to consider is where a Lie group Q acts isometrically on a manifold E. Thus, E is a smooth manifold with a metric that we will denote by d. Each g in the group gives an isometry

We will say that two points x and x‘ are the same modulo Q if there is an isometry g in the group G with gx = x’. This is an equivalence relation on E and the () equivalence class containing x is the orbit { ( ) : g E Q}. The set of equivalence gx classes will be denoted by E/Q and is called the quotient of E by Q. The quotient map n :E E/Q; x H {g(x) : g E Q } sends each point of E to its equivalence class. We wish to study the geometry that the quotient E/Q inherits from E. We can define the distance in E/C from one point n ( x ) to another n(x’) as the shortest distance between the orbits of x and x’in E. More precisely, this is
p ( n ( x > ,n(x’>> inf{d(g(x),g’(x’>>: g, g’ =
E Q).

E

E; x H g(x).

Since each g so

E

Q is an isometry, we have d(g(x), g(’) = d(x,g-’g’(x’)) and ’x)

p ( n ( x ) ,n(x’))= inf{d(x, g(x’)) : g

E

Q}.

250

GENERAL SHAPE SPACES

It is clear that p is symmetric and satisfies the triangle inequality, so it is always a pseudo-metric on E/Q. Furthermore, the distance is zero precisely when x is in the closure of the orbit of x. Consequently, p is a metric if and only if all the ' orbits are closed. We will only consider cases where this is true. The situation described above does include the shape spaces Xi for k-ads of points in the Euclidean space R". By translating and enlarging we can map any k-ad of points in R", which are not all equal, to a k-ad in the set

E=

{

k

k

(x,,x2,. . . , X k ) : c x j = 0
j=l

and
j=l

(Ixj1l2= 1 .

1

This set E is the pre-shape sphere and, being a subset of metric:

has the Euclidean

Two k-ads (XI, x2, . . . ,xk) and 0 1 1 , y 2 , . . . ,y k ) in E have the same shape if there is a Euclidean similarity S : R" + R" that has S ( x j ) = y j for j = 1, 2, . . . , k. k This similarity must send the centroid C;=, j to the centroid Cj,lyj so x S fixes the origin. Similarly, S must preserve length because C:=,Ixj I l 2 = 1 = I l!yj1I2. Therefore, S is in the group SO(m) of rotations of R" about the origin. The quotient E/SO(rn) is then the shape space that we called Xi.The metric defined on the quotient

i

i

x;=,

is the procrustean metric. In a similar way, we can consider the shape spaces for k-ads of points in any Riemannian manifold M. Let Q be a group of isometries of M and let E be the set of all k-ads from M. Then, the product metric

is a metric on E and the group Q acts isometrically on E by

Since y is the shortest path. the implicit function theorem shows that. We can also reverse the above argument. suppose that d is the distance on E derived from a complete Riemannian metric. If this is the case. Hence. (. the orbits ( ( ) : g E Q } must be submanifolds of E with constant dimension equal gx to dim(E) . . 2 .Then. be the tangent space to the orbit through x.)(E/Q) - is a linear projection onto the tangent space at n(x). in a neighbourhood U of each point x E E. . so V. we can go further and make the quotient E/Q into a Riemannian manifold also. we have constructed a Riemannian metric on the quotient E/Q so that n : E + E/Q is a Riemannian submersion (see Chapter 6 and O’Neill. y k ) have the same x2.SHAPE SPACES FOR GENERAL MANIFOLDS 251 We will say that two k-ads (XI. onto H . k. . . the quotient E/Q has one point for each shape of k-ads modulo Q. Projection onto the first coordinate then gives a chart for E/G at n ( x ) and this shows that E/Q is a manifold. we would like to make the quotient E/G into a manifold so that the quotient map n : E + E/Q is a submersion.x’ E E. shape modulo Q if there is an isometry g E Q with gx) = y. The derivative n.n(x’))is the length gx of the shortest path y in E from the orbit ( ( ) : g E Q} to the point x’. onto 7r(x1(E/Q). the gx . 1966. .dim(E/Q). bijectively onto 7&)(E/Q). onto V . 1983 for further details). for j = 1 . We call it the verticd tangent space at x. . isometrically. it must be orthogonal to { ( ) : g E Q} at x. We have already seen in the case of Euclidean shape spaces that there can be singular points where E/Q is not a manifold.(E) + I. let V. The Riemannian metric on E gives us an inner product on each tangent space. For each x E E.(x)has V. . To include all of these cases we will study the general quotient and try to understand its geometry in terms of the geometry of E. This Riemmanian metric is closely related to the distance p on E/Q defined earlier. We can then define an inner product on 7&1(E/Q) so that the bijection is an isometry from H . When E is a Riemannian manifold. In the final part of this chapter we will consider such shape spaces when M is the sphere or the hyperbolic plane. For any two points x.For. in Z ( E ) . we can find a chart 4 : U + R” x R”’ that maps Q-orbits into the sets ( u } x R”’ for u E Rn. It will also be useful to consider the situation where E is a proper subset of Mk and the metric on E is not the product metric. since they are the inverse images of points under the submersion. the derivative g * ( x ) : I. . . as its kernel and maps 3-1. Suppose that each of the orbits is a submanifold of E with fixed dimension IZ’. Thus. Since g is an isometry. so we can define the horizontal tangent space H . . Where possible.(E) maps V. This inner product does not depend on the choice of x. then y = g ( x ) for some g E Q. at x to be the orthogonal complement of V. Then. and H . the distance p(n(x). if n(x) = n ( y ) . If this occurs we can restrict our attention to a smaller subset of E that excludes these singularities. is a vector subspace of the tangent space q ( E ) . . . We are assuming that the orbit is a submanifold of E. y2. For. This means that the derivative n*(X>: Z ( E ) Tr(. xk) and ( y l .

u2).. u2). consider a tangent vector u ( n ( x ) )at a point n ( x ) of E/Q. where and V are the covariant derivatives on E and E/Q.n(x')). It will be useful later to consider symmetries of the quotient E/Q. For vector fields u1 and u2 we denote by (V1. From this we can calculate the Riemannian curvature tensor R. so we find that 62 and (Vu.The Koszul formula allows us to calculate the covariant derivative of these horizontal lifts in terms of inner products and Lie brackets. Two elements f and f ' have = f l if and only if f ' = f g for some g E Q.Gz) function x I+ (211 (x). This will induce an isometry - f : E/Q - E/Q. v2)on. respectively. This shows how the curvature Ruiu2v3 the shape space can be found in terms of in the curvature R~. &]7. G 2 ) is equal to (v1. Suppose that f : E + E is an isometry. that n*(x) . it is straightforward to calculate the metric and other geometric invariants in terms of those of E.1. Therefore. Set 7 7 vcl V(Vl. we have uf = G ( f o n ) .differ by a vector field that is vertical. Also. A similar argument applies at each point of y . This implies that the Riemannian length of n o y is the same as the Riemannian length of y.] ~ 2 ~ 3 + 2(v(u43 u3)3 v(V1. So the group N/Q acts isometrically on E/Q. G2(x)).If 71 is a vector field on E/Q. the distance from n ( x ) to n(x')for the Riemannian metric we constructed on E/Q is the same as the distance p ( n ( x ) . R.. the way the Riemannian the metric on E/Q was defined ensures that (61.. for any smooth function f : E/Q + R. V ( V > u3)). the quotient (E/G)/(N/Q) is naturally identified with E/N.252 GENERAL SHAPE SPACES derivative of y at x is a tangent vector orthogonal to the orbit (g(x) : g E Q}. The definition of the metric shows that there is a unique horizontal tangent vector G(x) in ?. the results can be quite complicated.2'3) = ('4. First.(V. Once we have defined a Riemannian metric on the shape space E/Q in this way.I = Q. so y is a geodesic in E with its derivative always horizontal. f Q f . 2 1 1 ) > V('U2. as was the case for the Euclidean shape space. 213)) + (V(u49 u2).' E Q for each g E Q: that is.u2 u3 : ('4. provided that f g f . Moreover. So. It is easy to check that V is a skew-symmetric tensor field. as we saw in Chapter 7. and so is a horizontal tangent vector at x. This is defined bv . I maps onto v(n(x)). These in formulae simplify when we consider the sectional curvature K ( q .( V ( V 4 . Then. . although. if Q is a normal subgroup of a larger group N of isometries. w2)-.~213E.G* . then these horizontal tangent vectors give a vector field V on E that is everywhere horizontal.212) = f77. We call this the horizontal lift of v. n ( x >H n ( f ( x > ) . then each f E N induces an isometry : E/Q + E/G.. v2)) ) ~ .

~on E to be a Markov process with O infinitesimal generator .o. which passes through x. For suppose that f is a smooth function on E/Q. The image of this (n(x. remains over y and.SHAPE SPACES FOR GENERAL MANIFOLDS 253 So. It is. there is a differential operator C on E/Q with A ( f 0 n ) = ( C f ) o n . C and A are not usually the . the tangent vector is then a horizontal lift of Now. x2. at each point. Since each g E Q acts isometrically on E. iC is the infinitesimal generator of the process (n(x.o then gives a random process on the quotient E/G.A.. the properties of Riemannian submersions ensure that the integral curve r of this horizontal vector field. a Markov process. x in Rm follow independent Brownian motions. and V is skew-symmetric. in fact.. Finally. % 2. we wish to consider probability distributions on the quotient E/Q related to the metric. Thus. r is a geodesic in E. The Riemannian metric on E gives rise to a Laplacian A and then we can define Brownian motion ( X ~ ) . The geodesics in the shape space E/Q are curves y with tangent vectors that are parallel along y : that is. . it preserves the Laplacian. Since E/Q is already a Riemannian manifold. it has a Laplacian A and $ A generates a Brownian motion on E/G. we obtain For further details see O'Neill (1966). their size-and-shape will give a Markov process in the size-and-shape space SC. Then. Then.)). However. suppose that the k vectors XI. .E/Q of a geodesic in E that has all its tangent vectors horizontal. so Thus. For example. . This shows that every geodesic in the shape space E/Q is the image under the quotient map n : E +. We will concentrate on the distributions associated with Brownian motion. As for the special case we described in Chapter 6. . If y is a geodesic through n(x).. so A ( f 0 n ) is constant on each orbit.)). the vector field along y extends to a vector field on E/Q that lifts to a horizontal vector field. which is therefore a Markov k process.

say. Theorem 3. The orientation-preserving isometries are the maps g:R"---+R". x ~ R ( x ) + b . Consider two smooth functions f and g on E/Q with compact support. 2. then the divergence theorem gives However. . H 11. . . . Let u ( u ) be the d-dimensional volume of n-'( )and H the vectorjield Vlog v. y z . g . d. .xk) and 6 . If we denote the volume elements on E and E/Q by dUE and duE/Q. v duE/Q. Then. Proof. a Cf =Af +Hf. Vf) = A f +H f . .254 GENERAL SHAPE SPACES same and consequently the process ( n ( x r ) ) t g not usually a Brownian motion.7 and Equation (33').respectively. Two k-ads (XI. so Cf = v-'div(u. . for example. the first integral can also be written as JEIQ Cf .y k ) of points in R" will have the same size-andxz. k . (See. . Helgason. 1 shape if there is an orientation-preserving isometry g : 08" + R" with g ( x j ) = y j for j = 1. the a orbit z-' ( )is a compact submanifold of E withjixed dimension. 1984. For each (T E E/Q. They have similar properties to the shape spaces but have a simpler geometry. .1.2 SIZE-AND-SHAPE SPACES We will now show how to apply the results of the previous section to shape spaces by considering the size-and-shape space for k points in Euclidean space R". . O f ) = A f This is the desired formula. where A is the Laplacian on E/Q.) Proposition 11. is The following well-known proposition explains the connection between C and A. These size-and-shape spaces have arisen earlier in Chapter 9. + u-'(Vv. . . Let G be a compact Lie group of isometries of the Riemannian manifold E with n : E 3 E/Q a Riemannian submersion.

= e j . in Rk-'. . . a translation vector. = 0. the orthoge onal complement V of 1 is a (k . . The corresponding inner product is ( T .C&. . This will have its centroid u.x. We now need to be more precise in defining the standard k-ad. the m x k matrix X with columns xl. uk) of points in Rk-l. The restriction of X to V is then a linear map T : V -+ R" with T(uj) = x j . . the standard k-ad (u1. Two such maps T and T' give k-ads with the same size and the same shape when TI = RT for some rotation R E SO(m). These form ei. . . . . . Let u. Thus. .1)-dimensional vector space and so it is isometric to Rk-' as we may see by choosing an orthonormal basis for it.. = C:=. We can translate any kad (XI. u2. For any k-ad (XI. the k-ad is regular. In the size-and-shape spaces we distinguish between k-ads that have the same shape but different sizes. xk) with k I Ix. . Then. any linear map T : Rk-' -+ 08" will map ( U I . so we can apply the results of the previous section. x Therefore.] 1' = 1. R") of linear maps from Rk-' to R" by the group SO(m). . xk) in 08" with x. . R") a Euclidean norm by setting IlTll = {tr(T'T))'I2. . . uk) to u2. u2. . q. TI) = tr(TfT'). . . j is moved to the origin. a rotation. . and b . xk) so that its centroid x. . . so u. Let e l . .SIZE-AND-SHAPE SPACES 255 for R E SO(m). X2. : There is a slightly different way of thinking about k-ads of points in R" that will prove useful later. at the origin and will be specified more precisely in a moment. . If we restrict attention to k-ads (XI.. . we obtain . . uk) in V or. in Rk and 1 the vector C. Note that So. . . T(u2). j with all components equal to 1. ) a k-ad (T(ul). be the orthogonal projection of e j onto V. . .=. . the shape space C as a subset of the size-and-shape space SC. ek be the standard unit vectors. = 0) by the group SO(m). . R m ) / S O ( m ) the space Horn(@'. . . x2.xk ) E ( R")k : X. fix a standard k-ad (u1. can be thought of as the quotient of i E = {(XI. . T ( u ~ )of points in R" with its centroid at the origin. xk gives a linear map X : Rk + R" that maps e j to x. . Then. . the size-and-shape space SCk. equivalently. we can identify the size-and-shape of space with the quotient Hom(Rk-'. e3. . First. We can give the vector space Hom(V. . e2. x2. x2.Each rotation acts isometrically on E for the product metric. . .

So. = 0 and X the corresponding m x k matrix. . Let (XI. ( f ' ) ) ~ ~ ~ = s ( T ) 2 s(T')2 This is the usual relationship between the metric on a cone and the metric on its cross-section at unit radius from the vertex.X(f'))* . It is clear that the set ( n ( T ): s ( T ) = I } is the shape space Zk and that the metric is the same as the procrustean metric considered in Chapter 6. . x k ) be a k-ad in R" with centroid x. = ~ s RESO(") + + + s ( T ) s(T') { p ( n ( f ) . This metric is We can also relate this to the metric on the shape space EL.. . Consequently. Therefore. xk) be a k-ad of points in R" with x.. Any linear map T : V -+ R" can be written as s ( T ) . The size of the k-ad (x1.x j ) and these inner products are clearly unaltered when we replace (XI. the metric on the size-and-shape space is identical to that on the quotient Hom(V. Let (xi. x2.s ( T ' ) >+ s ( T ) s(T')p(n(f). . x2. . x k ) is {tr(TfT))I" = J I T J I . . . corresponding to ( s ( T ) . The metric satisfies p ( n ( ~ )n(T)12 s ( T ) ~ s ( T ' ) . This corresponds to the linear map X : Rk + R" with Xej = x j . Let P = X ' X . .. . where p is denoted by 6. . . x. . .x.n(f)) in R x with all the points (0. x ' j .2} = ( s ( T ). the size-and-shape space S E i can be regarded as a cone on the shape space E i with n ( T ) E S C .x2. . We now wish to describe coordinates for the size-and-shape space and calculate the Riemannian metric relative to these. .2 s ( ~ ) ( ~ ' ) sup tr(f'R?') .. R")/SO(rn). k ( T . f for some map f of unit norm. .x.256 GENERAL SHAPE SPACES Let (XI. so we will be brief. This is very similar to the calculations for the shape space in Chapter 6. n(f)) identified together to represent the k-ad with all points equal. .x'. X'X' agrees with T'T' on V and is zero on R1.) be another k-ad with X' and T' the corresponding maps.). . The restriction of X to V is the linear map T : V -+ 08". xk) by . T ' ) = tr(T'T') = tr(X'X') = )Jxj j=1 . Then. x. x2.. . The entries in P are pij = (xi.

Since 4 : SZt-l -+ P(k) is two-to-one on the shapes X with rank(X) = m. we see that S1 = 0.1 < m. nonnegative k x k matrices P with P1 = 0 and rank(P) 5 m. As in (ii) we see that SCL consists of two copies of Pm(k) joined along the boundary. The set P ( k ) is readily seen to be a cone with dimension k(k . This means that the map 4 : SC. R E O(m) with det(R) = -1 is equal to UJ with U E SO(m). (iii) k . . k). .We thus obtain a well-defined HP 4 : SCk. . X = U S : Rk + R" has X1 = 0. non-negative k x k matrices P with Pl = 0. the matrices X and X' may have different shapes.1)/2 homeomorphic to a closed half-space of the same dimension. which we regard as a subspace of R k . R x z . = X1 = 0. However. depending on the relative sizes of k and m. Now. R x k ) for any of rotation R map SO(m). Consequently. In this case. we see that SCi-. . This shows that 4 is surjective. the rank of P is rank(P) = rank(X).1)-dimensional subspace of 08". . This identifies SZk up to homeomorphism.1 > m. This happens precisely when all the points XI. The orbit of T is the image of the smooth map p~ : S O ( m ) - Hom(V. Then. (ii) k . consists of two copies of the cone P(k) joined along their boundary.1 = m. Moreover. . X and X' = RX have the same size-and-shape if and only if X and J X do. It is therefore homeomorphic to 08k(k-')/2. Choose a linear map T : V -+ 08" and let (XI. This should be compared with Casson's theorem identifying the shape space Xi-l with S'kfl)(k-2)/2. Since P1 = 0. which is at most m. Thus. . if and only if rank(X) < m.SIZE-AND-SHAPE SPACES 257 E (Rxl. X k ) be the image of the standard k-ad in V under this map. Let J be a fixed reflection of 08". when det(R) = -1. X and RX have the same shape. Rk). so Pm(k)= P(k) and the map 4 : SZk + Pm(k)= P( k ) is a bijection. . Also. non-negative and has P1 = 0. . R H RT.xk) = X'X on the size-and-shape space. . we see that X' = RX for some orthogonal matrix R E O ( m ) . xk lie in an (m . Therefore. --+P E Pm(k) sends two distinct points to each P E Pm(k) with rank(P) = m and maps only one point to each P E Pm(k)with rank(P) < m. For.If det(R) = +1. . so X gives a point in SC. 4 maps into the set Pm(k) of all symmetric. rank(S) = rank(P) 5 min(m. There are three different cases to consider. let P(k) be the set of all symmetric.1 < m. x2. X' have the same size-and-shape. We will show that 4 maps onto this set. The matrix P = XtX is symmetric. . with XtX = P. . xz. n(x1. Moreover. because x. Every such matrix has rank(P) < k . if X and X' satisfy XtX = X"X/ = P.k). (i) k . 08"). . . for R E O ( m )\ SO(m). . so we can choose an orthogonal map U that sends the image of S into 08m'n(m.x2. --+ Hom(Rk. Let us now investigate the Riemannian metric on the size-and-shape space. then R E SO(m) andX. which consists of matrices P with rank(P) < m. each P E Pm(k) has a symmetric square root S with S'S = S2 = P. .

where r is the rank of T: that is. The corresponding change of bases transforms T E Hom(V. . q . more exactly it is convenient to change coordinates in V and [w" so that T has a simple form. the x2.f m ) f of R" and scalars t 3 t2 3 t3 3 .f basis (f~ 2 . . Vk-I) of V. ( H . Note that we may assume that the basis ( q v2. v2. This should be compared with the pseudo-singular values decomposition in Chapters 1 and 6. . vk-1) is positively . 3 0 with l m j=1 So..1 ) / 2 . The horizontal tangent space at T is the orthogonal complement of VT. . R")/SO(m).. we obtain an orthonormal basis (vl. the vertical tangent space is Vr = {AT E Hom(V. .AT) = 0 for all A with A f+ A = 0). m ) . 7 .AT) = tr(H'AT) = tr(ATH') and this is zero for all A with A' + A = 0 precisely when TH' is symmetric. since T and RT have the same size-and-shape for each R E SO(m). . t2. the . . . If we have two orthonormal bases for V. are called the l approximation numbers of T. . we see that T maps the orthogonal complement of v1 into the orthogonal complement of f i . . R") : ( H . R") : A' + A = 0). Since IITvllI is maximal.1. then there is an orthogonal linear map U : V -+ V that maps one to the other. So. . When k . an orthonormal basis (fi. R") : TH' = HT'} To describe the Riemannian metric on SC.The tangent space to SO(m) at I is the Lie algebra so(m) = {A E Hom(R". because we may replace vk-1 by -0k-l if necessary. . T is diagonal relative to these bases. oriented. . However. Then. R") : At + A = 0). Changing orthonormal bases in V or in R" does not change the metric in the size-and-shape space Hom(V.1 3 m. So.1 and singular points otherwise. this means we have a manifold when rank(T) is m or m .258 GENERAL SHAPE SPACES So.f m ) may then not be positively oriented. . . The numbers t . Note that t j must be 0 for j > min(k . . . . dimension of the space spanned by (XI. Thus. Tvl = t l f l for t 3 0 and a l unit vector f1 E R". This is clear for R". 2 . Xk). Repeating this process for the restriction of T to the orthogonal complement of v1.. R") to . . This has dimension r(2m . so it is %T = {H E Hom(V. the derivative of pr at the identity I will map the tangent space to SO(m) at I onto the vertical tangent space V T .l ~= {H E Hom(V. . Now. We can choose a unit vector v1 E V with IITv1I) maximal.r . . . we see that we obtain a manifold structure on the size-and-shape space at those points where this dimension is locally constant.

this transformation commutes with T ++ RT for any R E SO(m). This is Interchanging the roles of i and j and averaging gives . For T = r jf j $ we have cy=. we may assume that T is diagonal in our calculations.(S) equal to I IH I I. S = CTj=. Rm) with H with V vertical and H horizontal. there is an isometry Consequently. Now. sijfivs. the orthogonal projection of S onto the horizontal tangent vectors is The Riemannian metric for SC: is given by making the length of n. we can write S = V Consider any tangent vector at T represented by a matrix S E Hom(V. so. + We can solve these equations to find H : Therefore. Thus. Then.SIZE-AND-SHAPE SPACES 259 T U . as we saw in the previous section.

For each symmetric linear map S : V + V. To calculate the curvature and other geometric invariants of SCk.V. V 2 : T H V.. The vector space Hom(V. we will concentrate on the case where the approximation numbers are nonzero. v2 on S . so that HT' = T H t and A + A ' = 0.w2). f j and [Sl. we can find these by talung the horizontal and vertical parts of T H TSlS2. Therefore. TSlS2T' = HT' +ATT' + and transposing gives TS2SlT' = TH' - TT'A.v2). where [Sl. For. C . = However. we know from the previous section that V.(T){Vl(T)} TSlS2. the composite H = TS has H T t = TST' symmetric.7 ~ 2 )is a vertical vector field.S2S1.260 GENERAL SHAPE SPACES Note that this is well-defined unless two of the approximation numbers si are zero.. S2lij = (T. = (V. if A = CTj=. R") is flat.v2 of vector fields v1.vz)" V ( q . T H TS is the horizontal lift of a vector field v on SC. . We can explicitly determine the sectional curvature K ( q . S21 = ~j a Cf." + T+j. Let V j : T H T S j be two such vector fields.. Subtracting these gives T[S1.jL1 S 2 ] i j v i ~ > . then TJj[S..S2] is the commutator SlS2 . i j f . so the covariant derivative is given by V . Hence. 712) by . V : Hom(V. so it is a horizontal tangent vector at T . where (V. Then. Therefore. For simplicity. when we are at a singular point of the size-and-shape space. We can solve this by using the diagonal form of T = Cy=l f j v f . This determines the curvature of the size-and-shape space at T . R").is the horizontal lift of the vector field Vvlv2and V(v1. [Sl. R") --+ Hom(V. we need to find the covariant derivative Vv.S21T' = ATT' + TT'A.. Suppose that TSlS2 = H +AT with H horizontal and AT vertical.

. x2.1 to find the operator C on S . We will find its volume by computing how much the map J alters the volume of SO(m). PropoC . the vectors i x. For simplicity.b)(x. j=1 which is the condition that the angular momentum of the particles in the plane spanned by a and b is zero. sition 11. .SIZE-AND-SHAPE SPACES 261 A straightforward calculation gives This certainly shows that the curvature is non-negative everywhere on the sizeand-shape space.a)} ) =o. we will assume that k > m. This orbit is the image of the map J : SO(m) --+Hom(V. . x 2 . At each time t . b E Rm. x 2 . a. t i j form the columns of a m x k matrix X t X . we may apply Proposition 1 1. we will insist that the initial centroid x. j is also zero.(ij. we we obtain xi=. . R H RT. . In this expression. Expanding this. This is equivalent to X X ' = X X ' . R"). . . Having calculated the covariant derivative. must have a'(XXf .XX')b = 0. b ) ( i j .) trace out a geodesic in the size-and-shape space when their angular momentum about their centroid is zero. positioned at the points X I . . = is there and that the momentum C. Thus. x k ) . u ( n ( T ) )is the volume of the orbit {RT : R E SO(m)] + + in Hom(V. . each of unit mass. Xk and with velocities X l . .1 shows that C = A H . R"). for any two vectors a . it is easy to see that the geodesics through a shape n ( T ) are given by t t+ n(T + t H ) for a horizontal tangent vector H . . Since we want the centroid xj always to be at the origin. We can also interpret this in terms of the k-ad (XI.. i k . Thus. Finally. Now. Rm). Consider k particles.where A is the Laplacian induced by the Riemannian metric and H is the gradient of log u. . . the particles ( x j ti.=. + + + k x { ( x j . H is horizontal at T when HT' = TH'. The restriction of this to the subspace V of Rk gives a linear map T tH E Hom(V.

R") -+ so(rn). R"). A I--) $ (ATT' + TT'A) then changes volumes by a factor 6'. S H $(STt .so the irn(rn . . . This is a linear map that maps so(rn) onto a vector subspace of Hom(V. ( / ) is J*(/)' : Hom(V. First. similar argument applies to A the tangent space at any other point R E SO(rn). it changes volumes by niij(r. we are concerned with the restriction of this map to the ( k . observe that the transpose of J . Suppose that (z1. . z k ) is a k-ad of points in C. . A H AT. Since the plane is an important source of examples.3 SIZE-AND-SHAPE SPACES FOR THE PLANE When m = 2 we may identify R" with the complex plane C. and the derivative of J at the identity is J.(Z) : so(rn) --+Hom(V. R"). The composite J*(Z)'J*(Z) : so(rn) --+ so(m). z 2 .T S t ) .1)-dimensional subspace V. By choosing an orthonormal basis for V . the vector field H is the gradient of 11. we see that r j fjv> J*(I)'J*(z) : f i f : -fjfi' H (2 2 +$ ) ~ f : -fjfi'>s Therefore. For T = Cy='=.e j H zj.262 GENERAL SHAPE SPACES The tangent space to SO(m) at the identity is the set so(rn) of skew-symmetric m x rn matrices.1)-dimensional volume of the orbit J(SO(rn))is + x Volume(SO(m>).' ti"). This significantly simplifies the arguments of the previous section in certain cases. Consequently. We think of this vector as defining a real linear map x : R~ ---+C. We need to find the factor 6 by which it changes the volume. As in the previous section. we will describe the situation briefly.

Each w2. So the size-and-shape space SC: is the E Q quotient Ck-’/{o : 101 = 1). . Consequently. If the mean of w is WO.SIZE-AND-SHAPE SPACES FOR THE PLANE 263 we see that this corresponds to a vector w = (w1. dw ) = 0 and SO lly(t>1l2 I(w 112 t211dw 112. . . w’) = Cri:%jwj’. Finally.Two vectors w and w’ give the same size-and-shape if and only if w’ = o w for some w E Q with unit modulus. For the Riemannian metric on SC: we note that at any non-zero point w E Qk-I a tangent vector dw is vertical when dw = ihw for some real number h.. each with an isotropic normal distribution with variance a2. It follows that a horizontal geodesic in Qk-’ is of the form y :t Hw + tdw where (w. d w ) +t211dw112. More generally. lly(t)ll2 = I I w ~+ 2 t~( w . Then. A rotation in SO(2) acts on Q by multiplication by a complex number w with modulus 1. the distance in the size-and-shape space is where ( . component wj is a linear combination of the points z. .then the components of the corresponding vector w E Qk-’ will also be independent isotropic normal distributions with variance IS’.IIW -wall ’}. Z k ) are independently distributed in Q. then (w . 2a2 To compute the corresponding density on the size-and-shape space we need to average this over all rotations. dw) E R. . then the density for w is 1 (2nIS2)k-1 exp{ . If we choose w as this point. dw is horizontal when (w . sothereisauniquepointonthe ~ geodesic with minimal norm. . . wk-1) E Ck-’. . dw) is real. This is the metric on size-and-shape space that we used in Chapter 9. let us consider the probability distribution on SC: given by a k-ad of points that are independently normally distributed in Q. An isotropic normal random variable with mean zo and variance IS^ has density + __ 1 exp 2nd { ->}. z 2 . We will denote the quotient map by n. 202 Iz - ZOI If the points ( z l . This describes the size of k-ads along the = geodesic. the shapes trace out a geodesic in the shape space. It is simpler to work directly with the vector w. so we obtain . ) is the complex inner product: ( w .

264

GENERAL SHAPE SPACES

where

~ o ( x= )

1

2rr

d0 exp(x cos 0)2n

is the Bessel function of order 0. This should be compared with the results obtained by Mardia and Dryden in (1989a) and Section 8.5. Closely related to this is the Markov process on SC: obtained when the k points (z1, z2, . . . , z k ) execute independent Brownian motion in C. Then, the vector w will follow a Brownian motion in C k - ’ . Since the size-and-shape space is a cone, we obtain a skew-product process on it. The size I I I I is clearly a Bessel process w of dimension 2(k - l), while the shape follows a time-changed Brownian motion on CPk-2. We can prove this analytically by applying Proposition 11.1.
11.4 SPHERES AND HYPERBOLIC SPACES

In the next three sections we will use the hyperbolic plane and in the final section we will need hyperbolic 3-space. It is usually simplest to think of the hyperbolic plane as the upper half-plane w: = { z E CC : S(z) > 0) with the Riemannian [ metric
ds =
I2 d1
~

3(2)

or the unit disc B = { z E C : IzI < 1) with the Riemannian metric
ds =
~

2ldzl 1 - 1212’
1 + IPI
~

The hyperbolic distance between two points z and z’ is then
p(z, z’) = log

1 - IPI ’

(11.1)
- z’

where

z

for B.

Every Mobius transformation z H for a, b, c, d E C with ad - bc # 0 that maps R to R or B to B, is an isometry for the hyperbolic metric. All of : ,:

SPHERES AND HYPERBOLIC SPACES

265

the results we require in Sections 11.5 and 11.6 follow readily for these models. However, in studying hyperbolic 3-space in Section 11.7 it will be easiest to use the hyperboloid model for hyperbolic space, which views it as one sheet of a hyperboloid in a space with an indefinite inner product. Therefore, we will develop the properties of the hyperbolic spaces in this context. To show the analogy between hyperbolic spaces and the spheres we will recall the properties of the spheres. The sphere S" is the subset
((XO,Xl,.

. . , x) ,

E

R"+I : x;

+Xf

+ . . . + x ,2 = 1)

of OB"+'.

We will write ( , ) for the standard inner product
(X,Y)

= xoyo

+ xlyl + . . . + x , y , .

So, S" = (x E Rn+' : ( x , x ) = 1). This is a submanifold of R"+l and inherits a Riemannian metric d with

d ( x , y ) = arccos ( x , y ) , which measures distances on the sphere rather than the chordal distance through R"+'. The group of rotations SO(n 1) acts as a group of isometries of S" and acts transitively. We can map S" onto R", plus an extra point 00, by stereographic projection. This was described in Chapter 1, but it will be more convenient for us to project from the south pole rather than the north. Thus, the point x E S" is projected radially outwards from (- 1, 0, 0, . . . , 0) until it hits the set ((0, y l , y2, . . . , y n ) E R"+l}, which we identify with R". It hits at the point

+

l+xo ' l+xo

' . . . ' A). metric on S" is then transformed to the metric I+xo The

sin(Zd(y,y')> =

1

JU

IlY -Y'll + IIY 112> (1

+ llV'1I2>

fory,y'

E [Wn

u (00).

We can also consider the real projective space RPn, which consists of all the one-dimensional vector subspaces of R"+l. Each such one-dimensional subspace cuts S" at two antipodal points. So we may think of RP" as the space obtained from S" by identifying together every pair of antipodal points. The hyperbolic spaces W" are defined in a similar way to the spheres, or, more accurately, to the projective spaces, by replacing the inner product ( , ) on R"+' by an indefinite bilinear form. Let R'," be the vector space Fin+' with the indefinite, symmetric bilinear form

266

GENERAL SHAPE SPACES

Then. W" is the set

W" = (x

E R",' : ( x , x ) = -1 andxo > 0).

This is one of the two sheets of a hyperboloid. It is a manifold of dimension n. The hyperbolic metric d is given by d ( x , y ) = arccosh ( - ( x , y ) ) . It is often more convenient to use different models for hyperbolic space due to Poincare. These allow us to identify the hyperbolic space W" with the unit ball Bn = Cv E R" : lbll < 1 ) or with the upper half-space R? = Cv E R" : yn > 0). Stereographic projection gives us a map from W n to the ball Bn.We map each point x E W n to the pointy E Rn = ((0, y1, y2, . . . , y n ) : y j E R}, where the straight . . . , A) . line through (-I, 0, 0, . . . , 0) andx cuts R". This point is

(&,&,W n gives a The image pointy lies in the unit ball B". The hyperbolic metric on
Riemannian metric on B" with

s i n h ( i d 6 , y ') ) =

Ib - Y 7 l
J(1- llv1I2>(1 llv'1I2) -

fory,y'

E

Bn.

In a similar way, the mapping

enables us to identify the hyperbolic space W" with the upper half-space R l . The corresponding Riemannian metric on R? is

IVY I1 d s = -,
Yn

s i n h ( i d 6 , y ') ) =

-y'JI m

fory,y' E R .;

The hyperbolic spaces are highly symmetric with a group of isometries that maps any point to any other. To see this we will consider 'reflections' in hyperplanes in W". Let u be a vector in R'," with ( u , u ) = +I and set

P, = {x

E (HI" :

(x, u ) = 0).

This is a subset that is isometric to the ( n - I)-dimensional hyperbolic space Wn-1 . Reflection in this subspace is given by

RELATIVE METRICS

267

It is easy to check that R, is a linear map that sends W" into itself. Also, (R,x ,R,x ') = (x , x '), so R, is an isometry of the hyperbolic space. These reflections generate the group O'(1, n ) = { T E Hom(R"", R1,") : ( T x , Tx') = (x,x') and

T ( W n )= W"}.

Within this group the maps have determinant f1, depending on whether they are the product of an even number or an odd number of reflections. Those with determinant +1 form the subgroup SO'(1, n ) , which consists of all of the orientation-preserving isometries of hyperbolic space. Those with determinant - 1 preserve distances but reverse orientation. In the ball or the upper half-space models, the hyperbolic hyperplane P, corresponds to a spherical cap orthogonal to the boundary. Reflections correspond to inversions in such spheres orthogonal to the boundary. These generate the group of Mobius transformations of R" U (00) that generalise the Mobius transformafrom c u {oo]to itself. tions z H We will also need to consider the boundary of hyperbolic space. For the ball B" the boundary is the sphere S"-' = : 1 I I = l}, and for the upper halfspace 08; the boundary is Cy E R" : yn = 0) U (00). However, the boundary of the hyperboloid model W" is a little more difficult to describe. Let x be a point of W" and y the point of B" that corresponds to it under stereographic projection. As y tends to a boundary point in Sn-', so the direction vector x converges to a direction in the cone

2

K" = {u E R""

: ( u , u ) = 0 and uo > 0},

the vectors of which we call null vectors. So, we will consider the half-lines {hu : h > 0) in K" as the boundary points of W".

11.5 RELATIVE METRICS
In this section we wish to consider a different metric on the shape space that measures the relative displacement between triangles. This is the metric considered in much more detail by Bookstein (1978a, 1991). As we shall see, this gives no useful information for the shapes of triangles with collinear or coincident vertices. Indeed, it will have the property that all other triangles are infinitely far from such degenerate ones. Thus, this model of shape space is only apposite for a context in which it is natural to place such a constraint on the shapes we consider. It is convenient to identify R2 with the complex plane C. We will translate all triangles so that their centroid is at the origin 0. Fix a standard triangle A in the plane with vertices u1, uz, 113 that are not collinear and with its centroid at the origin. For example, we might take u1, u2, ug to be the three cube roots of 1.

268

GENERAL SHAPE SPACES

Then, any triangle with its centroid at 0 is the image of A by a real linear map T : R2 -+R2. Any such map is of the form

T :x

+ i y H w1x + w2y

for two complex numbers W I and w2. We will think of the set of all triangles with centroid at the origin as the set Hom(R2, R2) of real linear maps. The shape of the triangle is unchanged when we multiply w and w2 by the same non-zero l complex number. So, the ratio t = w2/w1 E C U {GO} determines the shape of the triangle. This is the way we identified the shape space with the extended complex plane in Chapter 1. (See also Section 11.3.) The triangle is flat, with collinear vertices, when T is singular or, equivalently, when t E 08 U {GO}. If the triangle is oriented in the same way as A, then t is in the upper half-plane R = {t E Q : S ( t ) > 0} and, if it is oppositely oriented, then S ( t ) < 0. In the : previous chapters we have used the procrustean metric on the shape space to measure the distance ;(t, t’)between two shapes t and t’ as

This metric depends on the choice of the standard triangle A . For suppose that we chose a different triangle, say, S(A), where S is the linear map S : R2

t=

R2;

(;)

H

(: :) (5)
-b

with a, b, c, d E R.

For S ( A ) not to be flat we need S to be non-singular, so the determinant 6 = ad - be is non-zero. For this new standard triangle, the map T is replaced by TS-’. Therefore, the shape t = T(i)/T(l) is replaced by

-

T ( ( d -bi)/S) - dwl T ( ( - c a i ) / 6 ) -cwI

~ 2

+

+ aw2

-

-bt+d at -c

Thus, 5 is the image of

t

under the Mobius transformation
tw

-bt

+d

at - c

Since a, b, c, d E R, all of these transformations map R U {GO} into itself and, consequently, map the complement Q \ 08 into itself. We now wish to consider a different metric p that is unaltered when we change the standard triangle. Thus, we should have

dt,t’>= P

-bt d -bt’ d at-c at’-c

+

+

(11.2)

RELATIVE METRICS

269

for any a , b, c , d E R with ad - bc # 0. It will measure the relative displacement of one triangle shape from another rather than the absolute displacement. Such a metric would be uninteresting on R U (oo}, which represents the flat triangles. For we would have

So, all distances between distinct points of R U (00) are equal and we have a discrete metric. Also, the metric on the upper half-plane 0: determines the metric 8 on the lower half-plane since
p ( t , t’)= p(t-1,

P).

Consequently, we need only consider the metric on the upper half-plane. This corresponds to restricting our attention to positively oriented triangles. 8 On the upper half-plane 0: we know that the hyperbolic metric does have the invariance property (1 1.2). Indeed, the only Riemannian metrics on R that : satisfy (1 1.2) are constant multiples of the hyperbolic metric. Therefore, we will take p as the hyperbolic metric
p ( t , t’)= log

(

I ’ - t( t

+ It’- t(

I ’ - t - I ’- t t l t l

)

as in (11.1). For certain results about the probability distributions on the shape space, this hyperbolic metric, the Bookstein metric, is more appropriate than the procrustean metric. Consider two probability distributions of triangles in the plane. Let X I , x2, x g have independent normal distributions with means U I u2, u3 and , covariance matrix a multiple a2of the identity. Then, the density for (XI, x2, x3) is

This density depends only on the distance d ( x , u ) =

{ c?=,- u;ll ’)i’* and JIxJ

so is unchanged when we apply any isometry of 08’ to all of the points. However, if we change the shape of the initial triangle A, the density alters. It is clear that this probability distribution is intimately connected with the procrustean metric on the shape space. Now, consider another distribution where we choose XI, x2, xg independently and uniformly from within the triangle A. This distribution has very different symmetries, as we saw in Chapter 8. The differences x j - u j no longer have the same distributions. However, if S : R2 -+ R2 is any non-singular linear map, it will map the triangle A to another triangle S(A) with vertices Sul, Su2, Su3. A point x i uniformly distributed within A is mapped to Sx;, which is uniformly

. t) = 1 I t 1 2 2S(t) ~ + + d i ) 2 + la + ci12 2 Consequently. T ( A )is a random triangle derived from A . A more general setting for this type of distribution is as follows. Then. ( A ) form a Markov chain.. such distributions on the shape space E. T n . . Let { T j : 1 < j < 00) be iid random variables takmg values in SL(2. We wish to examine the shapes of the triangles in this chain.2 70 GENERAL SHAPE SPACES distributed within S ( A ) . and have studied. . We can now construct a Markov chain of triangles. R). so the shape t of T ( A ) is T-( i ) . We can use Furstenberg’s (1993) work on random products of matrices to describe the behaviour of this chain as n -+ 00. So S transforms the uniform distribution on A to the uniform distribution on S(A). . Then. the shape Markov chain satisfies tn of the nth triangle T l T 2 T 3 . R). .5 and 8. Let T be a random variable that takes values in the space SL(2.~ T n ( A ) the in Furstenberg (1963. He considered iid random variables { X j : 1 < j < 00) taking values in SL(2.b T(l) + di a+ci The distance p(i. the triangles TlT2T3 . If the distribution p satisfies E(log tIxjlI> J’loglIxjlldl* < 00 = . T . Theorems 8. The standard triangle A has shape i. it is more natural to consider the hyperbolic metric that has the same symmetries as this distribution.6) proved a Law of Large Numbers for such random products. R) with a common distribution p. S T ( A ) is a random triangle derived from any other triangle S ( A ) . In a similar way. R) of 2 x 2 real matrices with determinant 1. Then.. Although we can study. Let T= (z i) E SL(2. T . t) in the hyperbolic metric is then given by Ib coshp(i.

t. Another application is given in BQrQny et al. X . if a has density f on [-1. as n . Therefore. the real part a' of a ' has density !) () conditional on the value a of % a . Also. . where we choose the vertices of the nth triangle independently and uniformly from within the ( n . we obtain (1). _ ~ . A new triangle A(') is produced by choosing a ' uniformly from within A and () taking A(') to have vertices . so log(2coshp(i. X ~ X I ~ I I a s . . He explicitly computed the value of a for this case in Mannion (1990a). we will assume that -1 5 %(a) 5 1. Thus. t. p(i' n --+ 2a a s . . a ' . the triangles T IT2T3 . This type of argument was used by Mannion (1988) to examine the behaviour of triangles.1.RELATIVE METRICS 2 71 must act irreducibly and if the closed subgroup generated by the support of on R2.Clearly. Suppose that we begin with a triangle A with vertices . We now apply Furstenberg's theorem to the case where X j = TS and u = Provided that the conditions for . ( A ) converge to the set of flat triangles. . Further results on random products of matrices are described in Bongerol and Lacroix (1985) and Ledrappier (1984).)) --+ a 2n a s . By dividing u into real and imaginary parts we see that this also holds for u E Q2.--+ --+a n 0 0 for each non-zero vector u E R2. 1.-lT.). t. as n --+ 00. I]. For simplicity.converge to the boundary R U (00) of 08: as n + 00. then there is a strictly positive constant a with 1 . t. the shapes t.. then a' 1 () 1 :) has density Kf.)) is asymptotic to p ( i . the imaginary part of a(')is no () bigger than the imaginary part of a. So. - 1 log(2coshp(i.1.~ ~ ~ ( I X . A particularly simple application of the above ideas arises when we consider the random triangles where two vertices are fixed and the third is chosen uniformly from within the triangle. are u satisfied. This certainly shows that p ( i .. 1 and a for some a E R. as n --+ oo. T. (1996). In this case we can see explicitly how convergence occurs.where .1)th triangle. Thus.) + 00.

T H RT E by left-multiplication. = SL(m. having integral J d. R). This gives us a right-invariant Riemannian metric d on SL(m. we (RAlS-lR-1.R)*(S)Al. A2)s = (AiS-'.x2).+l) for S E SL(m. R) -+ SL(rn. On this we may use our usual inner product (Al.1. We will call this the relative metric on SL(m. R)/SO(m) be the quotient map that sends each T E SL(m. Von Neumann's ergodic theorem shows that K n f + d. Since K is a well-known integral . = d. Therefore. It is easy to construct such a metric.. R)/SO(m). This gives the '1 l limiting density for the flat triangles to which the Markov process converges.2. R). dn) converges to the boundary of part of d n ) R The real part a' of dn) density K n f . R) to its orbit q ( T ) = {RT : R E S O ( m ) } . We can also find similar results in higher dimensions for the shapes of ( m 1)ads in Rm (cf. SO ((h. R) that is invariant under changes to the standard (m 1)-ad A..(x) = :(I . we obtain a metric p on the quotient SL(m. 1. R). RA2S-'R-')I = tr( (Rt)-' (Sf)-'A~RtRA2S-'II-'} = tr{(S')>-1A". Each rotation R E SO(m) gives a map h .2S-1} = (Al. This is an isometry for d since. almost everywhere with d. + + + This is a right-invariant metric on the group SL(m. Replacing A. A 2 ) S . R)) at the identity and use the right-invariance to define the inner product at S by (Ai. So. Le and Small. choosing the third vertex dn) independently and then we uniformly from within the triangle (. which is sZ(m. R) = {A E Hom(R". we wish to produce a Riemannian metric on E = SL(m. R) transforms the linear map representing a (m 1)-ad from T to TS-'. R). R)/S0(2) by taking the shortest distance between SO(m)-orbits in SL(m. we can evaluate the limit. ) I on the tangent space '&(SL(m..(x)dx = 1. The tangent space at I is the Lie algebra sZ(m.2 72 GENERAL SHAPE SPACES If we repeat this process. R) ~ ---+ SL(m. R). at any S have (AR)*(S): A F-+ RA. see that the imaginary decreases monotonically to 0. So. we can calculate geometric properties of this metric in terms of the metric on SL(m. Let q : SL(m. Solving this we obtain d. d-')). by S(A. R") : tr(A) = O}. 1997 and Small. R). T2S).We constructed the metric . 1996). A2)1 = tr(AiA2).(hR)*(S)A2)RS= ( 1m 2 ) R S m . A2S-')1. we seek a metric d with d(T1. As in Section 11. T2) = d(TiS.. : SL(m. a continuous solution of Kd. For we choose any inner product ( .: ? has operator. First. R)..

in Ahlfors (1996).w2 are the vertices of a plane triangle (w1. triangle is flat with collinear vertices. so we can extend the convergence results to Markov chains of shapes in SL(m. 1963 and Bougerol and Lacroix. This shows. not both 0. w l ] and [0. The simplest connection uses the shape of a fundamental parallelogram.6 ELLIPTIC FUNCTIONS The classical theory of elliptic functions. Let w1 and 0 2 be two complex numbers. which we can do by interchanging w1 and : w2 if necessary. Analogues of Furstenberg’s Law of Large Numbers exist for random products of matrices in SL(m. So. w1 and w2 generate a lattice A = Zwl Zw2. In the all other cases. Then the three points w l . R)/SO(m) (cf. Indeed. w3). so we can think of the quotient C/A as the space we obtain by joining together the opposite sides of P. The line segments [0. and set t = w2/w1 E . t. R). is related in various ways to the shape of triangles in the plane. as described. When t E R U {co}. for example in Helgason (1978). we know that right-multiplication ps : SL(m. for example. R) PS is an isometry. 1985). in particular. R)/SO(m> is a symmetric space and its geometry has been extensively investigated. wg) is positively oriented. Note that the modulus depends on the generators w1. The quotient SL(m. that the curvature tensor at each point is the same. Wm. The set R of possible values for t is called the Teichmuller : space for tori. R). Each point of C is equivalent under A to a point of P. So. for each S E SL(m. 0 2 . R>/sO(m> - + SL(m. R). w2 and w3 = -w1 .ELLIPTIC FUNCTIONS 2 73 d on SL(m. T H TS 11. : SL(m. Furstenberg. A2 5 1). pS is an isometry that maps the orbit q(1) to q(S). . q ( T ) q ( T S ) is well-defined and is an isometry. but there is a more complicated argument that considers double covers of the sphere ramified over infinity and over the vertices of the triangle. w3) with centroid at the origin. This is a torus and t is called its modulus. R)/SO(m) and acts transitively. t classifies tori with marked generators for the fundamental group.t . w2 is + P = {hlwl + h2W2 : 0 5 h l . The fundamental parallelogram for w1 . Each point t corresponds to a conformal equivalence class of tori with marked generators. the induced map R>/sO(m). w2. wz] give closed loops in the torus that generate the fundamental group n (C/A). This corresponds to insisting that the triangle (w1. R) acts as a group of isometries on the quotient SL(m.-1 . l It is usual in the study of elliptic function to insist that t lies in the upper half-plane R = { tE C : S ( t ) > 0}. In particular.C . so the point t determines the shape of (w1. Therefore. 0 2 . R) precisely so that it would be highly symmetric. w2 and not just on the lattice. This triangle has the same shape as the one with vertices 1. the entire group SL(m.

f has critical values z j . This is a quasi-conformal map with complex dilatation p and the hyperbolic metric corresponds to the Teichmuller metric on the upper half-plane. is the Weierstrass @-function for A. we may assume that f ( w j / 2 ) = z . 3. wk. Thus. ) has the same shape as (el. z 3 and infinity. This requires that A@’. provided that f ( w ) / 2 ) = f ( w j / 2 ) for j = 1. Elementary algebra shows that t’. Lehto in (1 994) gives further details on the Teichmuller space for tori. w j ) .t( I ’.5 p=-($)(=) so the hyperbolic distance p ( t . so it induces a mapping C/A -+ C/A’ between the corresponding tori. 0 1 2or. the hyperbolic distance between t and t’ in the upper half-plane is the same as the hyperbolic distance between 0 and p in the disc. the choice of generators ~ 1 . The real linear map T maps the lattice A = Zwl Zw2 onto A‘ = Zwi Zwk. b. 3 . the hyperbolic metric on the shape space and the Teichmuller metric on the Teichmuller space classifying tori are identical. z 2 . B E C. By translating C we may ensure that the poles of f occur at the points of the lattice A and nowhere else. so it is equal to C/A for some lattice A . z 2 . Thus. 0 and equivalent points modulo the lattice. Here. z 3 in the complex plane. t’)is log 1 (-)+ 1 - IPI (PI = log ( I ’. equivalently.: and (w’. Furthermore. consider three distinct points z l . By relabelling the points w1. w2.2 74 GENERAL SHAPE SPACES Now. Then z j = f ( w j / 2 ) = A @ ( w j / 2 ) B for j = 1 . So the triangle ( 2 1 ~ 2 2 . for j = 1 . w3) T :z H k(z + p7) for some k E C \ (0) and p E C. The covering map R + C then gives an elliptic function f : C -+ C that has . The lattice A is the period lattice of z3 f . z ~ . For we could choose a different set of generators w : . so it is uniquely determined by ( z I . the choice of t. but nowhere else. Each of these functions has critical points at w l / 2 . Then. 2 . z ~ ) .t l 1. R is an analytic torus. 2 . For the second connection. we may consider the hyperbolic metric on R Given two triangles .. e2. z 3 and 00.. 2 + + + + . e3). period lattice A and degree two.(z) +B for some A # 0. However.2. Then. there is a unique real linear map T : C -+ C with T ( w j ) = w ) for j = 1 . w2. 2 . J = 2 J h for some h E A . The only elliptic functions of degree two with poles at the points of A are those of the form + f ( z ) = Ab. w3. we may construct a Riemann surface R that consists of two sheets and is ramified of order two at 21. w3/2.m i . is not unique. This map must be of the form (w1. The values @ ( w j / 2 )are usually denoted by ej and satisfy el e2 eg = 0.t t l + It’. if necessary. w2/2. z 2 .tl + I ’ .t t .. 3.

for integers a. Since m i . z 3 ) thus determines a point T in the quotient R$/r(2). b. z2. must also generate A. (: :) is in the principal congruence subgroup The triangle (z1.ELLIPTIC FUNCTIONS 2 75 = awl t bw2. note that the shaded region in Figure 11. w.So. z 2 . we must have ad . The value of t depends only on the shape of (z1. z3). so the modulus t of A and A' are the same. w. the quotient R$/r(2) is obtained by sticking together the sides of the Figure 11.1 The fundamental domain of r(2) . Therefore. any Euclidean similarity S : C + C that sends z j to z ) will induce a conformal map from C/A to C/A'.1 is a fundamental region for r(2).bc = 1. For. = cwl so. d with + dW2 (: i) = (A y ) (mod 2 ) . c. Finally.

we obtain copies of all three geometries. As we shall see. The hyperbolic metric on W3restricts to give a metric on this horosphere and the set h. for a unique vector y with @ . For each vector y E ( 1 8 ' ~ with @.. corresponding to the points -1. with this metric is a Euclidean plane. If we choose any non-zero null-vector u .. However. We will do this geometrically in a way that describes all three similarly. the set h.2. on on the horospheres h. 1 and 0 in R This gives us again the shape space C . u ) = -1) is called a horosphere at u . we get the hyperbolic geometry. so that the same plane with the opposite orientation is given by P. within W3: hyperbolic planes P. x ) = 0 and xo > 0). Euclidean and spherical. In the ball B3 it corresponds to a sphere inside B3 that touches aB3 at the point corresponding to u .: a sphere with the three exceptional points representing triangles with two equal vertices. connected three-dimensional surfaces. = {x divides W3 into two half-spaces: E w : ( x . the Euclidean plane and the hyperbolic plane. and on the boundary sphere aB3 of W3 we have the spherical geometry. To do this we will work in the three-dimensional hyperbolic space W3 and refer to the general discussion and notation introduced at the beginning of Section 11. See Figure 11. the hyperbolic metric on R$/r(2) makes these exceptional points infinitely far away and so differs from the Riemannian metric discussed in previous chapters.2 76 GENERAL SHAPE SPACES shaded region as shown. we get the Euclidean geometry. the Euclidean plane and the hyperbolic plane. W3 contains subsets that are isometric to each of the sphere. 11. y )= 0 ) ' P = (x . Hence. y ) < O } ... the set rL = (y E R193: @ . y ) = 1.. E Py' = {x E W3 : ( x . Thus.Recall that a point of the boundary of the hyperbolic space HI3 is identified with a halfline {hu E R193: h > 0) in the cone K = [x E R'33 : (x. hyperbolic.4. in which the sets K and IL are. This gives a sphere with three punctures. = [X E W3 : ( x . to point outwards into PT. as 0 . y ) > 0) and W3 : ( x . the hyperbolic plane ~ P. y )= 1. y )= l} corresponds to the set of all oriented hyperbolic planes in W3. Every oriented hyperbolic plane in W3 is P.7 SHAPE SPACES FOR TRIANGLES IN THE SPHERE AND THE HYPERBOLIC PLANE We now wish to consider shape spaces for triangles in the sphere. We will orient the planes by choosing the normal to P. We will show that different subsets . in fact.

u3) and (bl. For the Euclidean triangles. points of 8W3= S2 represent the shapes of triangles in the Euclidean plane. we have known from Chapter 1 that the shape space for triangles is the sphere S2. 3) described in Section 1 1. Each isometry in this group maps W3 into itself. So. This consists of the linear maps with determinant 1 in the group O+(1. but also maps [L into itself and the cone K into itself. U Z . 2.. In particular.2 The hyperboloid model for hyperbolic 3-space of R'93 give shape spaces of triangles in each of these three geometries.u2. The points of W3 represent the shapes of triangles in the sphere.. bz. These maps of S2 are the Mobius transformations: if we identify S2 with the extended complex plane QQ3. 3) of orientationpreserving isometries of the hyperbolic 3-space. To prove these results we need to consider the group SO+(1. If Py is a hyperbolic plane in W3 .4. with a ] . b3) in S2.SHAPE SPACES FOR TRIANGLES 2 77 . / . Figure 11. 3. . . a3 distinct and bl . and some of the points in [L represent the shapes of triangles in the hyperbolic plane._. . 6 2 .. . 63 distinct.then they are given by maps + This enables us to identify the isometry group of W3 with the Mobius group. . each isometry gives us a map from S2 into itself. . we know that for any two triads of points (ul. there is a unique Mobius transformation sending u j to bj for j = 1.

Then. and so give an isometry of this hyperbolic plane. This sends e to a boundary point w and x1. 1) in K.0). Note that changing the three points corresponds to applying an isometry For each point w E W3 we can draw hyperbolic geodesics yj from w to of W3. = T ( B j ) are three geodesics from w to u. when we replace the point w E H’ either by a boundary point w E a W 3 or else by a hyperbolic plane Py. . preserving the orientation. Consider first the Euclidean case.. The composite S = (T’)-’T is then another isometry of W’ with S ( x j ) = x i and S(0) = 0.xi) = (S(xl). we might take the points that correspond to (1.x3.0. u3. from the origin to each x. u3. x3) and (xi. 3.x2.0. there is an isometry T’ of W3 with T’(xg) = u j and T’(0) = w .0. xg in he. that is. we draw hyperbolic geodesics ’/I. for j = 1 . Choose any horosphere he and represent the triangle by three points XI. Thus... for j = 1. which is a horosphere at T ( e ) . The same procedure. Conversely. This set is a Euclidean sphere about 0. to u. T ( x 3 ) in T(he). then T will map Py into itself. it must map the set of points at a fixed hyperbolic distance from zero into itself and preserve angles subtended at 0. If h. (1. except u1. 2 . (x1. Then. We will take these three points as the standard points and measure positions relative to them. Since T is an isometry of W’.x2. and extend them until they hit the boundary i3W3 at points [j. u3 in S2. u j . These three points are distinct since xl. For each point w E tWU3. We will use the ball B’ as our model for W3 and think of the sphere S2 as its boundary. so S is a rotation. suppose that we are given a spherical triangle with distinct vertices x1. So. The three tangent vectors to yj at w are then unit vectors and so determine a spherical triangle. Conversely. Consequently. (1.x2. S E SO(3).( x 3 ) )is isometric to the Euclidean triangle (XI 9 x2. 1. u2. suppose that we are given a triangle in the Euclidean plane. x i ) is another triad of distinct points in L IO that yields the same point w in W’. at three distinct points that form a triangle in the Euclidean plane h. uz.T(x2). 2 . y3 from w to u1. Since these are geodesics from the origin. xg). 0. These will cut a horosphere h. x2. into itself and will be an isometry for the Euclidean metric on h. so the triangle ( T ( x l ) .xg to points T ( x l ) .x2. 1. will give the shape space for triads in the Euclidean plane or else the hyperbolic plane.. they are radial straight lines. There is a unique isometry T of W3 that maps x. Draw geodesics p. These are unit vectors giving a triangle in the sphere. 3’ Suppose that (xi.. S(x2). there is an isometry T of W’ with T(6. Draw geodesics p.) = u. S(x3)) have the same spherical shape. Since S is an isometry. Choose three distinct points ul. For example. is a horosphere in W3 and the isometry T of W3 fixes u . because the mapping T is conformal.u2.. x2. x.0).x3 are. . then T will map h. x3).2 78 GENERAL SHAPE SPACES and the isometry T of W3 fixes y . . y2. from e through x. it will preserve distances T on he. y.T ( x ~ ) . and their direction vectors at w clearly determine a spherical triangle with the same shape as (XI. we have shown that HI3 represents the shapes of triangles in the sphere S2 with three distinct vertices. 3 and this sends 0 to a point w E W’.

i3W3\ (u1. we obtain the full sphere a W 3 = S2 as the shape space X i for Euclidean triangles that are not totally degenerate. This shows that the shape space for triangles of distinct points in the hyperbolic plane is part of IL. this gives us the set It is interesting to consider the significance of the remaining part of IL.x2. xi. x2. Since u1. 1. u3 E Py’. u2 = (1. x i ) is another hyperbolic triangle in the plane P.$j)= uj for j = 1. When we take u1 = (1. S = (T’)-’T is an isometry with S ( x j ) = x j for j = 1.0. So. x2. these triangles have the same hyperbolic shape. Consider a point y E IL with u1.0.then the vertex x1 approaches the boundary of the hyperbolic plane. x2. Moreover. E Py that form a triangle in the hyperbolic plane Py. so the triangle has no vertex XI. Then. that = yields the same pointy E IL. 3 and S(z) = z . As the plane Py moves across u1 so that u1 moves into P i . all three points ul. 3 and T ’ ( z ) = y . these two triangles have the same hyperbolic shape. there is an isometry T’ of W 3 with T’(. u3 represent those triangles with two vertices coincident. the are geodesics y j must meet Py at distinct points. u2. It is the part consisting of those y where u1. Finally. Draw geodesics y j normal to this plane from Py to u.0. the triangles (XI. u3 E Py’ and the corresponding hyperbolic triangle ( X I . x2. we consider the hyperbolic case. 2. T ( x 3 ) ) in the hyperbolic plane Py. x i ) is another triad of distinct points in the Euclidean plane he that yields the same point w E aW3. orthogonal to Pz through x j and going in the positive direction into P.2. x3) in Pz isometrically to a triangle ( T ( x l ) . 2 . x i . Conversely. 2 4 3 lie on one side of the plane Py. there is an isometry T of W3 with T(. 3 and T ( z ) is some pointy E L.$.) = B) and S ( e ) = e . Suppose that (xi. Then. every hyperbolic triangle corresponds to a point y E IL. So. Thus. Consequently. This means that S maps the hyperbolic plane Pz into itself.. ~ 2 . Thus.$)) u j for j = 1 . the two sides of 2 . Thus. ( 2 . Then. 6 3 lie in P.. Hence. l). x2. xg E Pz. the triangle. 0. u3 must lie in Pyf. 61. say. If we move y so that the plane Py approaches ul. suppose that we are given any hyperbolic triangle with distinct vertices. When u 1 lies in the plane Py. x1. Then. so : u1. ug lie in P. xi). xi.0). These geodesics will meet i3W3 at three distinct points . x i ) have the same shape. there is an isometry T’ of W3 with T’(B))= yj = T ( B j ) and T ’ ( e )= w . 2 4 3 distinct. T maps the hyperbolic triangle ( X I . xg) and (xi. so it is an isometry from (XI. so x1. the vertex x1 becomes an ‘ideal vertex’ on the boundary of the hyperbolic plane.SHAPE SPACES FOR TRIANGLES 2 79 Suppose that (xi. It is easy to see that the three omitted points u1.u2. It is a . x3).~ 2 .T(x2). x3) to (xi. u3} represents the shapes of Euclidean triangles with distinct vertices. u3 = (1. u2. S must map the horosphere he into another horosphere at e. Suppose that y E IL is a point with u1.. xi. These cut Py at points x.‘. If the horospheres are not equal. Hence. u2. Draw geodesics B. u2. The composite S = (T’)-’T is then another isometry of W3 with S(p.0). u2.Now. x3 are distinct. then S enlarges or contracts Euclidean lengths. ~ 1 x and ~ 1 x 3fail to meet. 1. Such a mapping is a Euclidean similarity.

280 GENERAL SHAPE SPACES Figure 11. For y E [L.3. It is also interesting to consider the metric on the shape spaces that we have found above. This group acts on points in Fils3 and there is a natural bijection from the orbit of any point y E F i l s 3 to the quotient SO'(1. 3) is the isometry group for W3. 3)/Stab@) is the shape space for spherical triangles.3 Hyperbolic trilaterals trilateral rather than a triangle. For y E HI3. So. These metrics are all conformally equivalent and so do give a canonical conformal . where we have used the disc model for the hyperbolic plane. However. In the sphere or the plane. metric on W3 itself is unchanged and gives a canonical metric for spherical triangles. this duality associates any triangle described by its vertices with the dual triangle described by its edges. Thus. These are trilaterals with sides that are hyperbolic geodesics. the points of [L represent triangles that may have 'virtual vertices'. Similar results hold for the other vertices. Finally. This is illustrated in Figure 11. we have a family of metrics on S 2 . The group SO'(1. So. so SO+(l. but those sides may not meet at any point in the hyperbolic plane to give a triangle. which corresponds to a W 3 . The isometry T acts as a Mobius transformation of the sphere a W 3 = S2 and this can indeed change the metric. if we let U be a halfline {hu : h > 0} in K. 3)/Stab@) by the stabiliser Stab@) = (T E S 0 + ( 1 . the orbit is all of L. There are trilaterals described by three sides where the sides do not meet in vertices. SO'(1. The shape spaces for triangles can also be described entirely algebraically. u3 changes the Thus. so SO'(1. 3)/Stab(U) is the shape space for triangles in the Euclidean plane. 3)/Stabb) contains the shape space for hyperbolic triangles. in the hyperbolic plane. 3 ) : T @ ) = y } . then the orbit of U is the set of all half-lines in K. this is no longer true. the hyperbolic point that represents the triangle by some isometry T of W3. Observe that changing the choice of points u1. Note that there is a duality between vertices and edges for triangles. the orbit is all of MI3. u2.

This is invariant under T . However. So. on 1L the indefinite bilinear form ( .SHAPE SPACES FOR TRIANGLES 281 structure on the shape space for triangles in the Euclidean plane. Finally. they do not give the full strength of the metric results described in Chapter 1 and thereafter. we obtain a canonical indefinite metric on the shape space for hyperbolic triangles. . ) gives an indefinite metric that is not positive definite.

Before describing the main results. group of order n. LE Copyright 0 1999 by John Wiley & Sons Ltd Appendix In this appendix we shall outline some of the general results from homological algebra. Similarly. as well as the more specific duality results for manifolds. {g E Q : order of g is finite}. KENDALL & D. ring.The extra generality does not require much extra effort. unless Q is equal to F Q itself. or in some cases non-commutative. Except for nG. If we look at a fixed such direct sum decomposition of Q . for an integer n. abelian group is a direct sum of cyclic groups.Shape and Shape Theory D. C A R " & H. Note that n Q . will be written Q 6 W. we first need to introduce some natural unary and binary operators on groups. We shall continue to write Z for the cyclic 9 . whereas abelian groups are just the modules over Z. so that the identity element is 0 and. ng is the nth 'power'.l UNARY OPERATORS ON GROUPS For any abelian group Q we define nQ = {ng : g tQ = E Q}. which we have taken for granted in the main body of the book. However. there is no canonical choice. although Q has subgroups isomorphic with F Q . A. . FQ = Q/tQ. K. We shall restrict attention to the category of finitely generated abelian groups. then nQ . which we shall write additively. G. "Q and tQ are all subgroups of G. the direct 'product' of two groups. "Q = { g E Q : ng = 0}. Q and W. BARDEN & T. which is fairly common. this is not standard notation for these groups: tQ is more usually denoted tors(Q) and otherwise there is no standard. Most modern texts describe the results for modules over an arbitrary commutative. but it has the effect of making the results look unfamiliar. Recall that every finitely generated. which others might denote by ZlnZ or by C .

if Q = tQand W = FW. W ) = 0. MacLane.f.g €3 hl + g €3 h21 and { g 8 (hl + h2) .2. W ) = ( f : Q --+ W : f is a homomorphism}.1 Horn For two abelian groups Q and W we have (cf. However. it is sometimes denoted Q h W. W ) 2 Hom(Q1. Q*= Hom(Q. 1963. MacLane. @ ) Hom (Q.that is. MacLane. W ) Y " W . 138). 21. p. Q 2 tQ @ F Q . E W} quotiented by the free subgroup generated by all the h elements {(Sl + g2) €9 h .Thus. if all the elements of none of the elements of W are of finite order.c. This is made into an abelian group by defining addition of homomorphisms via addition of the images in the usual way. . Thus. The subgroup tC is the direct sum of all the finite cyclic summands of Q from a given decomposition.)@ Hom(Q. 1963. W ) W . W. h E W] = g €9 hl subject to the relations (gl + g2) €9 h = g 8 hl + g €9 h2 and g 8 (hl + h2) +g 8 h2 (cf. 1963. unlike €9 it is not symmetric. 23) Hom(Z. Q and A. Hom (GI Q2. Z)2 F Q . it is the free abelian group generated by ( g 8 h : g E Q. A.. The subgroup "Q has a cyclic summand of order n for every summand of Q of order a multiple of n . then Hom(Q. In other words. 21) Hom(Q. and the quotient F Q is the direct sum of those that are of infinite order and so each isomorphic to Z.2. W . To make it look more like a binary operation analogous to Q 8 W below. I@ W2) Y Hom(Q.284 APPENDIX has an infinite cyclic summand for every one that G has and. A. W) @ Hom(G2. for example.g 8 hl + g 8 h21. There are isomorphisms (cf. p. of n and n'.2 BINARY OPERATORS ON GROUPS In addition to the direct sum already mentioned. pp. there is one in nC whose order is n' divided by the h. G ) S Q. Horn(&.2 Tensor The tensor product Q 8 W is the abelian group generated by the elements ( g € 3 h :g E Q . for every cyclic summand of Q of order n'. there are four basic operators that produce a new group from two given ones.

Q@WSH@Q. the group that is ‘extended’ is a subgroup of its extension and the quotient is the group ‘by which it has been extended’. Ext(Z:. is isomorphic with the group Bihom(Q. denoted Ext(W.4 Tor Finally. W.not to be confused with the torsion subgroup.Z) of ‘bihomomorphisms’ or bi-additive maps from Q x W to Z. This product is the quotient of the set ((8. n . h ) (8. MacLane. Q@ZZQ. Q ) = 0. Tor(Q. G I ) @ Tor(W. by the fact that the group of homomorphisms from G 8 W to Z. h ) (g. The following facts. Hom(Q @ W. Q@ (Wl @ W2) 2 Q @ Wl @ Q @ W2. GI) Ext(W.. G Tor(&. n’. Q ) 2: Q / n Q . G @ Z” 2 C / n Q . analogous to those just given for Ext. h ) : g E G. n E z. n . 1963..3 Ext An extension of a group Q by a group W is any group K equipped with a homomorphism from K onto W with kernel Q. can be found on p. Q1 @ Q 2 ) 2 Ext(H. . @ Ext (Hi@ W2. n .2. Ext (HI. Q z ) .ng = 0 = nh} h by the equivalence relation generated by (gn. Tor (W. Q ) 2 Tor(Q. h). Ex@. The extensions 06 for a given Q and W can themselves be made into an abelian group. Q): note the rather unnatural order. The basic facts are as follows (cf. Q ) @ Ext(W2. ) A. pp.2. Details of this. Z) 2: tQ. A. n .Q ) 2 Ext(W1. Z 2: Zfl.n . Thus.Z). Q 2 ) . I @ Q 2 ) Z Tor(H. etc. h ) (gl. + . 150 of MacLane (1963). ) Tor(& Q ) = 0. W . Ext(Z.BINARY OPERATORS ON GROUPS 285 The tensor product is also characterised uniquely. together with the following facts. in particular. up to isomorphism. W ) of Q and W. nn’. h ) ( g 2 . we have the ‘torsion product’.- + - Tor(W. 64-71 of MacLane (1963). can be found on pp. n’h) and bi-additivity on Q x W: (gl g2. 139-142). Ext(Q. Q ) . E W. C ) 2 “Q. which is why we are using t for the latter.

is the I . MacLane. H. G ) = Ker(1. those of C. G)/L*(Hom(IF.G) = 0 and Ext(W. .G ) . r @ g * 1 ( r ) @g QHQ01.then the homology and cohomology groups of K are.then there are induced homomorphisms 1.+I h G. G ) 2 Hom(H. . Then.(K. d*).))/d(K. = K h G and . as usual.+I) and H.5 Alternative Definitions of Tor and Ext If we write the group W as the quotient [F/R. G ) = 0. where (K h G). Tor(W. respectively. for example. the Universal Coefficient Theorems are expressed by the following isomorphisms. where. So. 171 and 77. These theorems work at the level of the chain complex. by definition.. which may be found on pp.1 The General Theorems We are now in a position to describe the theorems that relate the homology and cohomology of a space with standard integer coefficients to that with arbitrary coefficients.(K). G ) .286 APPENDIX A. the qth homology group of K. then [F = W and R = 0. of MacLane (1963): H.-l(K).-I (W). G ) and H 4 ( K .(K.: and 1*: R @ G+ F@G. G) 2 H. is the quotient (ker(dlK. For example. [FhG-. Kn +K.(K). 1963. d ' Kl d K O be a chain complex of free abelian groups. When K is the chain complex associated with. Then. d*(Q)= Qod as before. d @ 1 ~ )The cohomology of K with coefficients in G. the degree. 100) Ext(W. A. d ) = . we let (K. where 1 : R + [F is the inclusion.(K) @ Q €3 Tor(H. G ) is the qth homology group of the complex ( W @ G . W(K. mapping K h Q to K. G ) €3 Ext(H. R h G . This gives a method for computing some of the above results. . a cellular decomposition of a space C. p.) and (cf. G ) = Coker(i*) = Hom(R.-1. homology of the cochain complex (K h G . so Hom(R. irrespective of its provenance: it need not arise from a cellular decomposition of a topological space. when W = Z. where IF is the free group generated by the generators of W and R is the free subgroup of [F generated by the relations that determine W. (K h G .2. .3 THE UNIVERSAL COEFFICIENT THEOREMS A. d * ) is called a cochain complex since d* raises .3.G)).

W ) . W ) 2: Ext(Q. more generally.(K) @I Z 2: H 4 ( K )m Z is z 2 a direct sum of r s copies of Z . both for the limited extent to which it does hold for shape spaces and for the more general extent to which it does not. and so direct sums of cyclic groups of various prime power or infinite orders. isomorphisms are special cases of the following. we take coefficients Z2. If Q and W are finitely generated abelian groups. that H 4 ( K ) is the direct sum of r copies of Z and s copies of 722 for suitable r and s.1 PoincarC Duality When M is a compact. then + + Q@I w s Qmw. ) If. A.(M. K will have basis elements ef corresponding to cells Ef of dimension r in E . each of H. L. Z =\ H.4 DUALITY IN MANIFOLDS We have had cause to refer to the duality between the homology and cohomology groups of a manifold. There is a one-to-one correspondence between the cells El of the first decomposition and FL-. determined geometrically by the fact that corresponding cells fi . Tor(Q. Thus.(K). and L will have basis elements corresponding to cells F. where K is chain isomorphic to the dual of L. The proof rests on finding geometrically ‘dual’ cell decompositions E and F . Then. Z )and Hq(K. 2 This agrees with what we obtained by computing directly with the complexes with Z coefficients since in that case all boundary maps are zero. very unnatural. and corresponding chain complexes 06. which may be proved using the listed properties of the various operations. Z )=\ H.-. of the second. of M. ) - z. if it is such a homology manifold.2 The Shape Space Example Suppose. H.(M. 2 The above. 181 of Maunder (1970). then we have the following Poincark duality isomorphisms: D : H’(M. whether it is orientable or not: D : H‘(M. instead of the integers. A. orientable manifold-without-boundary of dimension n . 2 and + Tor(H. as is the case for shape spaces.3. Thus.(K).l . similar isomorphisms are valid for any homology n-manifold that is compact and without boundary.-.(K. 772) E x H q ( K ) 2 Ext(H. These results may be found on p. and W is finite.DUALITY IN MANIFOLDS 287 A. 772) is a direct sum of r s t copies of Z . in F . Z ) 2 is a direct sum of s copies of 772. 2 - Z) 2. if there are t summands isomorphic z 2 with Z in H q ..4. or.

Bo x B" = (point} x B".. . 0 .)'+'A with respect to the dual basis (4:)of Ls. However. Of course. the cohomology of M. it is built up by successively adding r-handles. we are implicitly using here the fact that different cell decompositions of the manifold lead to isomorphic homology groups. The orientability of M is used in the above to relate the orientations of the cells of F with those of the cells of E . which is defined by @ ( f : ) = S i k . Then. the cells of E may be obtained as natural extensions of the cores B' x (0) of these handles. d : Kr+l ?. In other words. Moreover. both K and 11 are chain complexes corresponding to cell decompositions of the manifold M. that of M and the homology of L* is. by definition. fi Z and d* . L. otherwise. with respect to these bases of the chain complexes determined by the respective decompositions of the manifold. we have the required duality isomorphisms.r .-l. Krhas matrix A. If M is not orientable it is not possible to correlate the orientations in this way and then it is only possible to obtain the chain isomorphism when the chain complexes have Z2 coefficients. the relations between the cells of the two decompositions and their chosen orientations are such that the coefficient of e'. and hence correlate the boundary maps in the two complexes. and the cells of F may be obtained similarly from the cocores (0) x Bn-'.([L*) of L*. the diagrams Kr+l 5 11n-r-1 K. a then d* : [ L n P r p 1 -+ L"-' has matrix (-l>'+'(A')' = (-. ( K ) of K and H. starting from a set of 0-handles.288 APPENDIX meet transversally in a single point and. we have the restricted version of the theorem in that case.Ood. again by definition. This can be proved using the Eilenberg-Steenrod axioms for a homology theory given in Chapter 3. That is.Id* L"-' commute for all r and so induce isomorphisms between the cohomology groups H . That is. that cells of complementary dimensions from the two decompositions are disjoint.-.7-1 [L'.). if d(e:+l) = cj d(fX-.f:-. in the boundary d(eL+l)of eL+l is (-l)'+l times ai.l in d(fi-. then the coefficient of f h . On the other hand.> = (-1)'" Ciai. so their homology is.e!. if we construct the dual cochain complex [L* = L m Z of L for which the sth cochain group is Ls = [L. Thus. The easiest way to obtain the geometrically dual cell decompositions E and F of M required for the above proof is to assume that M is presented via a handle decomposition. B' x B"-' attached along S r P 1 x B"-'. So. . Hence.+'A'. the correspondence between the basis elements e i in Krand (-l)r+lq!(pr in L"-' induces a chain isomorphism a : K -+ L*. while d : Ln-r -+ LnPr-l has matrix (-l). Thus..

these points of intersection may be allotted a sign and then their sum is an invariant of the pair h. if ynPs is a cycle representing another torsion element An-.4. The isomorphism ( L ) is obtained similarly. from the above outline proof of duality...(M) is represented by a cycle I . However. An element. yn-. h._l is the boundary of some cycle y. E Z.-1 (M).(M) is represented by a cycle yn-. It is called their intersection number. it is independent of our choice of representative cycles. then ny. similarly. For this. except that in this case the pairing is by linking numbers that are defined in Q/Z.. or hn-q is a torsion element and it determines a non-singular bilinear form on the free quotients.. it induces the required isomorphism.-. . From the isomorphisms H q ( M . Z @ Ext(H. . of H. This defines the linking number h. and provided M is orientable.(M)) @ T(H.The content of this aspect of PoincarC duality is that this pairing H. of H. That is.-. by the order n of the first torsion element is a rational number that is well-defined modulo the integers. ..(M) + Z is zero if either of 8 h. in terms of the natural generators.) and. This is not so easy to see directly from the geometry and the algebraic proof requires the Bockstein boundary homomorphism arising from the exact sequence of coefficients 0 - Z --+ Q --+ Q I Z --+ 0. an element of ker(dIK. Provided these cycles are in general position with respect to each other. h.-l . by means of the Universal Coefficient Theorems we may re-interpret them as isomorphisms between subgroups and quotient groups of homology groups as follows.-. h. is.(M) 1H. and h.2 Duality via Linking and Intersection The natural duality isomorphisms are between the cohomology and homology of the manifold. hn-. Then.. Z) ) ) g F(H. an /that .-.DUALITY IN MANIFOLDS 289 A.(M).-. if a cycle y4-1 represents a homology element h. h.-l of order n. the quotient of the intersection number y. element.-l(M)) The isomorphism (I) may be re-interpreted in terms of intersection numbers and (L) may be re-interpreted in terms of linking numbers. Hence. which in this case means that they meet only in isolated points. in the quotient group Q/Z. Z 2 Hom(H. This should be clear.

Then the short exact sequence 2 0 __f A. Then. which were mentioned at the end of Section A. We say that a sequence of homomorphisms of abelian groups or. z2. obtained from the complex of exterior forms on the manifold and exterior differentiation. z2) =" Hn-9(M. very unnatural. In the case of an orientable differentiable manifold M the cohomology with real coefficients R is the de Rham cohomology.. for which the manifold need not be oriented. If K is a direct complement. isomorphism H9(M. allow us to re-interpret the duality as the. since the cohomology with coefficients in a field has no torsion and so is equal to its free quotient. is exact if the kernel of each is the image of its predecessor in the sequence.(M.3.W IT z 2 and Ext(Hr-l(M). given by where the volume form duM is determined by the orientation and a choice of metric. the exact sequence implies that K is isomorphic with . .1. the isomorphism ( I ) can be re-interpreted as Hq(M. Given such choices. F) Y H"-. each de Rham class is represented by a unique harmonic form and then ( H ) is the Hodge duality between such forms. z2). A S THE SPLITTING OF EXACT SEQUENCES In our calculations of the homology groups of shape spaces we mentioned how very nearly our exact sequence of such groups splits. Then. the isomorphisms H i . the analogue of the intersection pairing is the pairing between q-forms Q. 1 so that A2 = z(A1) @ 06. (H This follows from PoincarC duality and the Universal Coefficient Theorems.q)-forms q&.290 APPENDIX When we take the coefficients in a field F. z2) 2 H r ( W 8 ZZ Tor(ffr-l(M). in a similar manner to ( I ) . There is a similar simplification of the duality theorem with Z2-coefficients.+ A 2 -A3 0 splits if the image of A in A2 is a direct summand. and how much simpler our calculations would have been had it done so. such as the real numbers or the rational numbers. of modules. F). and ( n . together with the Universal Coefficient Theorems. more generally. We recall now what these terms mean and derive the necessary and sufficient condition for a sequence to split to which we referred.

such that p . - A ---+ 1 z K __f 0 where K is the image of z and. o z is the identity of Al. the necessary and sufficient condition for this is that 1 there exists a homomorphism p : A2 + A such that p o z is the identity of Al. This implies that there is a homomorphism.the composition of the projection from A2 onto z(A1) and the identification of z(A1) with A1. Since z is injective. hence also the kernel of J . As for the short exact sequence. More generally.. the existence of such a homomorphism p is also sufficient for the sequence to split: its kernel will be the required direct complement of z(A1) in A2. we say that the long exact sequence splits at A2 if it breaks up into two exact sequences . 1 p : A2 + A.. by exactness..In fact. z A ) is isomorphic with A and so altogether we get (1 1 an isomorphism A2 G A @ A3. .THE SPLITTING OF EXACT SEQUENCES 291 A3.

IEEE Trans. Umea. Sci. Macmillan.140.V. Mathematika 43. 11. Nauk. Bookstein.L. 23. Adv. Bru. F. Comput.R. (1982) Random shapes by factorisation. Lecture Notes in Biomathematics 24. Birkhauser. Ser. 109. McGraw-Hill.IMath. 435-458. Izv. Appl. Broadbent. Bower. (1986) Size and shape spaces for landmark data in two dimensions (with discussion). (1990) The geometry of shape spaces. In: Ranneby. John Wiley & Sons. SOC. S.V. Cambridge University Press. Bookstein. P. LE Copyright 0 1999 by John Wiley & Sons Ltd Bibliography Ahlfors. (1974) Generalized Inverses. F. K.V. L.D. Ambartzumian. (1989) The Statistical Analysis of Shape Data. Bougerol. Machine Intell. T. Proc. 1. New York. 2nd edn. Pattern Anal. . and Carne.L. 165. Res. F. 11-IV. F.F. New York. Cambridge. London Math.L. 284-288. Swedish University of Agricultural Science. Bookstein.L.E. Stat. R. T. (1996) Barycentric subdivision of triangles and semigroups of Mobius maps. F. A. BARDEN & T. K. Cambridge University Press. (1990) Factorization Calculus and Geometric Probability.O. dissertation.F. R. C A R " & H. R. DieudonnC. BLBny. C.N.V. (1989) Processus de Wishart. (1989) Principal warps: thin plate splines and the decomposition of deformations. KENDALL & D. Ben-Israel. (1978b) Linear machinery for morphological distortion. (3) 61. University of Leeds..L. (A) 143.L. Ambartzumian. New York. J.V. Sci. Dryden. Academic Press. Dryden. Ambartzumian. I.L. (1996) Complex Analysis. Came. Cambridge.Shape and Shape Theory D. (1985) Products of Random Matrices with Applications to Schrodinger Operators. and Lacroix. Bookstein. Beardon. (A) 308. and Mardia. Bookstein. 259-276. 181-242. Progress in Probability and Statistics 8. 407-432. T. F. (1985) A Sylvester-type problem for homogeneous Poisson processes. Biomed. London.29-52.R. and Greville. Stat. M. New York. (1991) General shape distributions in a plane. Probab. ed. B. A.171. Ph. Statistics in Theory and Practice. (1970-4) Treatise on Analysis.K.K. (199 1) Morphometric Tools for Landmark Data. Armyaskoi SSR 20. Soc. G. Roy. J.. Springer-Verlag. J. Acad. (1978a) The Measurement of Biological Shape and Shape Change. 11. 567-585. I. (1980) Simulating the ley hunter (with discussion). Paris. (1930) Size and Form in Plants. I.

and Iosifescu. (1987) Riemannian Geometry. In: Demetrescu.1 17. Goodall. Kendall.. (1977) The diffusion of shape. Adv. SOC.G. D. Academic Press. 53. Hermann. K. and Mardia.R. and Mardia.. E. 33-51.1. Lie Groups. C. Appl. pp.R. R.V. (1971a) Construction of maps from ‘odd bits of information’. . Mathematics in the Archaeological and Historical Sciences. H. (1975) Generalized procrustes analysis. C. S. Statistics of 2dimensional Voronoi foams. Berlin. Cambridge. Adv. Stat. J. 21. J. Edinburgh.11. and Lafontaine. (1963) Noncommuting random products. Lecture Notes in Mathematics 1161. M. Nature 231. Chichester. 30. (1 984) Groups and Geometric Analysis. Trans. In: Hodson. Royal SOC.G. and van de Weygaert. D. 445 -457.V. Gallot. Anal. Studies in Probability and Related Topics: Papers in Honour of Octav Onicescu on His 90th Birthday.L. D. (1960a) A sufficient condition that a mapping of Riemannian manifolds be a fibre bundle. Appl. Springer-Verlag. Psychometrika 40. S. 23. I. (1998) StatisticaZ Shape Analysis. H.. Berlin. C. (1991) Procrustes methods in the statistical analysis of shape (with discussion). 108. 9. Invariant Differential Operators and Spherical Functions. Nagard. Multivar. 16-32.R.V.G. 428-430. D. eds. Kendall. (1974) Correspondence analysis: a neglected multivariate method. and Mardia. Furstenberg. Academic Press. Edinburgh University Press. Hilton. (1975) The recovery of structure from fragmentary information. Trans.59. V. Kendall. R. (1993) Multivariate aspects of shape theory. Goodall. (1989) Stochastic Calculus in Manifolds. (1985) Harmonic mappings of Riemannian manifolds. M. S. Stat. 547-582. S. 40. In: Giusti. 23. Helgason. Goodall. K. eds.G. New York. Roy. Astrophysics 184. Proc.J. and Wylie. Kendall. and Mardia. (1977) Riemannian center of mass and mollifier smoothing. Montreal. Kendall... and Tautu. (1971b) Maps from marriages: an application of non-metric multidimensional scaling to parish register data. 285-339. 340-355. 1 . C. 377-428. Hildebrandt. Berlin.G. P. Probab. Appl. Emery.R. Orlando. (1983) The shape of Poisson-Delaunay triangles. Amer. Ann. pp. J. Math. (1991) A geometrical derivation of the shape density. (1992) The non-central Bartlett decomposition. (1960b) On the differential geometry of foliations. Hill. 236-242.496-514. J. (2) 72. F. D. S. Cambridge University Press. Probab. Math. 331 -339. (1960) Homology Theory. Hermann. D. Springer-Verlag. and Symmetric Spaces. Hulin.London (A) 279. and shape densities. M. K. Math. SOC. Math. Integral Geometv.294 BIBLIOGRAPHY Dryden. Amer.C.G. Stat. (B) Goodall. I. 848-866. Commun.R. (1987) Fragmenting the universe. Kendall. Icke. SOC. ed. Springer-Verlag.O. Cower. Astron. Phil. D. Pure Appl. 158. Helgason. Karcher.C. R. M. 509-541. K. 94-108. Ann. 303-3 18. P. pp.V. John Wiley & Sons. (1978) Differential Geometry. Harmonic Mappings and Minimal Immersions.

. Soc. procrustean metrics. In: Grimmett. H. 30. 81-121. Probability. Kendall. Proc. Kendall. Appl. Soc.G. Proc.G.G. and Welsh. 17. Oxford. John Wiley & Sons. D. . J. Amer. 87-120. Kent. Cambridge Philos. Appl.G. 28. pp.G. In: Mardia. Soc. Roy.) 59-72. D. and complex projective spaces. 371 -406. Kendall. John Wiley & Sons. Appl. D.S. Statistics and Optimization. Probab. D. Chichester.V. (1997) Consistency of procrustes estimators. 203-217. W. 115. (1990b) Probability. W. 242-247. (1987a) Explicit formulae for polygonally generated shape densities in the basic tile. K. 225-230. Chichester. B. Oxford University Press. K. Kendall. 317-334. 281-290. 407-412. G. 89-90. Stat. Probab.G. 25. pp. Adv. Disorder in Physical Systems. Lond. and Mardia.S. Probab. (1986b) Further developments and applications of the statistical theory of shape.127. Probab. Appl. D. D. In: Mardia. (1994a) How to look at objects in a 5-dimensional shape space I: Looking at distributions.. Le. D. (1980) Alignments in 2-dimensional random sets of points. Kendall. ed. Conv. 20. (1991b) Contribution to the discussion following Goodall (1991). Kristof. Stat. 31. (1986a) Contribution to the discussion following Bookstein (1986).G. Kendall.G. Appl. (1971) Generalization of the orthogonal Procrustes rotation procedure for more than two matrices.G. Math.V. W. Kendall. and Le. The Art ofstatistical Science. (1990a) The diffusion of Euclidean shape. J. 4.G. and Kendall. D.G.S. Proc. ed. H. 35-43. 12. ed.BIBLIOGRAPHY 295 Kendall. W. 313-321. Math. Kendall.P. eds.V. (1989) A survey of the statistical theory of shape. 775-797.T. (1992) Spherical triangles revisited. Theory Probab. John Wiley & Sons. 102. and Wingersky. Soc. (1998) A diffusion model for Bookstein triangle shape. Adv. and Le. Appl. Teoriya Veroyatnostei 39. H. 225 -234. (1994b) How to look at objects in a 5-dimensional shape space 11: Looking at diffusions. (1987b) Singularities of convex-polygonally generated shape-densities. (1988) Symbolic computation and the diffusion of shapes of triads. Adv. Le. 380-424.S. H. D. 79th Annu. Soc. Adv. In: Kelly.113. London Math. 27. pp. (Suppl. Stat. (1984) Shape manifolds. Kendall. Adv. 587-596. Kent. Kendall. Its Appl. (1986) Exact distributions for shapes of random triangles in convex sets. 61. and harmonic maps with small image I: uniqueness and fine existence. (1990) Random Delaunay simplexes in 08". 19. Probab. Kendall.G. Adv. Chichester. Sci. 16. W. Infer.S. Kendall. 308-329. Kendall. Appl. D. J. Math. Kendall. (1992) New directions in shape analysis. Kendall. (1991a) The Mardia-Dryden shape distribution for triangles: a stochastic calculus approach. Adv. Bull. 896-916. Probab. D. Probab. Kendall.G. Assoc. D. Probab. (1995) How to look at objects in a 5-dimensional shape space 111: Looking at geodesics. The Art o f Statistical Science. Proc. Kendall. Kendall. D. 105..T.G. Plan. D. Psychol. (B) 59. K. Cambridge Philos. J. F. (1987) The structure and explicit determination of convexpolygonally shape-densities. convexity. 315-324. Appl. W..G. 101. (1985) Exact distributions for shapes of random triangles in convex sets. D. J. D.

London (A) 439. pp. (1994) Brownian motions on shape and size-and-shape spaces. H. Shape Anal. (1995b) The mean shape and the shape of the means. Adv. Le. H. Springer-Verlag. Le. Le. I. I. Mardia. Ledrappier. Adv. Appl.R. Ann. Math. C. 22. Probab. Math. Proc. 570-580. 845-865. Con$ Curr. B. (1991) On geodesics in Euclidean shape spaces. Math. Biometrika 76. Berlin. (1963) Homology. . and Heudecker. (1988) A Markov chain of triangle shapes. J. Probab. 20.. Adv.F. K. (1970) Algebraic Topology. (1988) MatriwDiflerential Calculus. 0. D. 83 1-844. (1998) On consistency of procrustean mean shapes. 573-602.D. K. Adv. (1990b) The invariant distribution of a sequence of random collinear triangle shapes. In: Mardia. Leeds. Mardia. Ph. Royal Soc. Appl. H. and Dryden. (1994) Introduction to Teichmiiller Space. Le. 31. London Math. Leeds University Press. and Dryden. Cambridge Phil. Probab. (1995a) Mean size-and-shapes and mean shapes: a geometric point of view. 26. Int. 27. 334-340. H. 21. H. Ecole d ' i t i de Saint-Flour. XII. Springer Graduate Texts in Mathematics. Lecture Notes in Mathematics 1097. K. 11. D. Mannion. H. Soc. Lehto. H. and Gill. Probab. Proc. Probab. Adv. Adv. H. Springer-Verlag. ed. Probab.V.V. van Nostrand Reinhold.G. and Small. and Shape Densities with Unqorm Generators. Geol. Le. (1994) Shape averages and their bias. Adv. and Kendall. 34-39. (1993) The Riemannian structure of Euclidean shape spaces: a novel environment for statistics. (1990) A stochastic calculus approach to the shape distribution induced by a complex normal model. John Wiley & Sons. 25. F. K. dissertation. Putt.L.L. H. 44-55.. J. New York. Probab. Probab. Appl. Mannion. 122. SOC. (1989a) Random spherical triangles. 21. 1225. H.A. Probab. Le. Le. Le. Le. Proc. Le. (1997) On simplifying shapes by subjecting them to collinearity constraints. 21. 32. Appl. In: Hennequin. 271-281. H. Appl. S.142-755.V. 221 -228. Springer-Verlag. C. (1999) Multidimensional scaling of simplex shapes. Issues Stat. (2) 44. Recog. H.R. 101-113. Appl. Proc. 1601-1613. (1992) The shapes of non-generic figures. Mannion. (1984) Quelques propriCtCs des exposants caractkristiques. 22.G.L. Berlin. J. S. Mardia. SOC. Le. (1989a) Shape distributions for landmark data. C. ( I 990a) Convergence to collinearity of a sequence of random triangle shapes. D. Adv.L. Appl. 30. (1988) Shape Theory in Flat and Curved Spaces. Shape densities.53-63. Geometrical background. Adv. Appl. (1993) Euclidean distance matrix analysis (EDMA): estimation of mean form and mean form difference. Magnus. 348-370. Appl. 197-210. H. and Dryden. I. Cambridge Philos. H. Berlin. 360-372. Maunder. (1989b) Random spherical triangles. 109. Probab.1271. and Bhavnagri. 315-323. (1989b) The statistical analysis of shape data. Le. Appl. Stat. D. London. I. University of Cambridge. MacLane.V. 581 -594. eds. 21. and applications to collinearity testing.296 BIBLIOGRAPHY Le. Lele. P.

H. Chichester. Spanier. R.132.G. Math. Cambridge Philos. SOC. Small.G. University of Cambridge. C. B. Small. Z. disserf tation. (1983) Characterisation of distributions from maximal invariant statistics.G. Stat. Muirhead. C. Func. (1977) Orthogonal procrustes rotation for two or more matrices.S. IV. (1994) Mean figures and mean shapes applied to biological figure and shape distributions in the plane. (1981) Distribution o Shape and Maximal Invariant Statistics. 543 -547. SOC. 91. Special publication 2. and Mecke. F. and Williams. Rijksuniversiteit Leiden. Rogers. (Republished in John von Neumann: Collected Works. Proc. J. Biom.555-571. O’Neill. (1967) Submersions and geodesics. Reidel. A. B. New York.H. Reinhart. (1966) Algebraic Topology. (1966) The fundamental equations of a submersion. Math. New York.J. MacMillan (1992). (1990) Proc. Trans. vol. John Wiley & Sons. B. 267-276. 1. Verw. Stat. Rev. J. John Wiley & Sons. Psychometrika 42. (1990) Mean figures and mean shapes in the plane. R. Academic Press. Math. f New York. Ziezold. Duke Math. Workshop. 517-527. H. (1937) Some matrix-inequalities and metrization of matrix-space. (1984) A classification theorem for planar distributions based on the shape statistics of independent tetrads. J. 84. Markov Processes.G. D.D. 243 -257. Wahrsch. J. dissertation. Rev.W. Ten Berge. Math.H. Gebiete 63. (1987) Diffusions.G. O’Neill. 119. University of Michigan Museum of Zoology. Dec.L. Int. Small. (1996) The Statistical Theory of Shape. Springer-Verlag.C. Stoyan. J. Biosci. Originally published in 1917. (1982) Random uniform triangles and the alignment problem.A.G. C. Proc. Newman. Cambridge University Press. E. 96. R. 491-510.F. 13. New York. Dordrecht.363-373. 459-469. J. Chichester. Kendall.L. John Wiley & Sons. von Neumann. (1987) Stochastic Geometry and its Applications. 85-127. Random Processes. Ziezold. (1966) The engulfing theorem for topological manifolds. (2) 69. W. B. 6.) Ziezold. Math. Vol.J. eds.D. London. M. and Bookstein. H. Small. C. (1970) On the homogeneous planar Poisson point process. Ann. van de Weygaert. Ann. (1988) Techniques of shape analysis on sets of points. Schrift.M. Thompson. 286-300. (1991) Voids and the Geometry of Large Scale Structure. Math. D. Kassel. C. 34. D’A. Ph. Small. (1982) Aspects o Multivariate Statistical Theory. 205-219. McGraw-Hill. 7th Prague Con& Inj: Theory.E. Tomsk Univ. (1959) Foliated manifolds with bundle-like metrics.G. Michigan Math. . C. 315-322. Michigan Morphomet. F. 36. (1977) On expected figures and a strong law of large numbers for random elements in quasi-metric spaces. Cambridge Philos. [1917] (1942) On Growth and Form.BIBLIOGRAPHY 297 Miles. 56. O’Neill. Small. 591 -602. and Martingales: It0 Calculus. L.. Rohlf. (1983) Semi-Riemannian Geometry. Ph.

55-68.Shape and Shape Theory D. 31. 56. 286. 64. 260 array of shape spaces 5 barycentres 202. G. 72 of spaces of shapes in 3-space 74 collinearity set 132. 203 basic segment 190. 130 cell boundary 43-5 cell complexes 48 homology of 4 1-53 cell decomposition 32. K. 290 absolute homology groups 47 adjacencies 238 affine transformation 182-3 angular momentum 261 antipodal poles 131 approximation numbers 2. 55. 34. BARDEN & T. LE Copyright 0 1999 by John Wiley & Sons Ltd Index abelian groups 46. 55 of shape spaces 27-9 of SO(3) 233-6 of unit sphere 25-7 cell dimensions 28. 288 cohomology groups 45-7. 101. 52 boundary maps 47.58. 62-5. 36-9.1 basic tile 191-3 basis elements 287 Bessel function 264 Bessel process 264 bi-additive maps 285 bihomomorphisms 285 binary operators on groups 284-6 biological shapes 2 block decomposition 96-7. 286 boundary map in 57-9 decomposing 59-62 chain isomorphism 288 chordal metric 205 closed d-cell 25 closure points 232 cochain complex 46. C A R " & H. 41 -2. 66. 65 cell orientations 30. 264 Cartan mean 203-5 Cartesian coordinates 181 Casson’s theorem 20-3. 288 chain complex 45-7. 253. 286. 100 Bockstein boundary homomorphism 289 Bookstein metric 269 Borel a-fields 157 Borel sets 164 boundaries 46 boundary cells 44-5 boundary coefficient 63 boundary components 35. 77 in chain complex 57-9 properties 59 Brownian motion 135. KENDALL & D. 32. 65. 88 boundary homomorphism 5 1. 254. 57. 283. 35. 51. 284. 162 . 286 cocores 288 coherent orientation 41 -2 cohomology 55.

149. 140. 31 cycles 45-6 deformation retraction 49 degree of homology group 46 Delaunay tessellation 164 de Rham cohomology 290 diagonal coordinates 143 diagonal elements 136. 261 critical points 222. 42. 273 component of boundary 60 configuration space 227 connected shape space 12 connectivity 32-3. 63-5 decomposing 84-91 for spaces of shapes in 2-space 70. 111. 172 diagonal matrix 115. 98 covariance matrix 169. 135. 2. 28. 79 in 3-space 72 in 4-space 75 Euclidean coordinates 135 Euclidean dimensions 2 Euclidean distance 226 Euclidean geometry 133.1 exactness 48 . 65-7 continuous bijective map 21. 58. 129. 205. induced shape-measures 172-6 complex numbers 117.16 distance function 114 distance-parameterised geodesic 109 distances to lower strata 127-33 distributions on C i 243 dual hormomorphism 101 duality in manifolds 287-90 in shape spaces 100-3 isomorphisms 101 suspended intersection type 101 theorem 290 via linking and intersection 289-90 Eilenberg.Steenrod axioms 47 elementary cells 26. 254 Euclidean triangle 277-8 exact sequences applications 50-2 splitting 290. 105. 43. 128. 22.17 elementary spherical cells 56 elliptic functions 273 -6 embedding techniques 202 equilateral triangles 15 equivalence 124 essential complexes 60. 59 elementary shape spaces 10. 276 Euclidean metric 20 Euclidean norm 255 Euclidean plane 276-81 Euclidean shape space 252 Euclidean similarity 249 Euclidean space 1. 9. 141. 110. 172 covariant derivative 260. 201. 224 Crofton’s boundary formula 176 curvature of size-and-shape space 260 curvature of two-dimensional surface 149 cut locus 124-7 CW complexes 25. 143 diffeomorphisms 42. 145 differential manifold 107 differential structure 1 10 diffusion on C 244-7 : dilation class 22 dimension of corresponding shape space 4 distance between shapes 114. 56.300 INDEX compact convex polygon 190-200 compact convex set 176-80 compact spaces 12. 114. 249-50. 94. 67 complex normal distributions. 23 continuous inverse 23 cores 288 counter-diagonals 92.

16 of n given shapes. 87. 253. 29. 62-5 of spaces of unoriented shapes in 2-space 82 in 3-space 84 in 4-space 86 homology groups 23. 140.103 closed formulae 9 1. 144 . 113. 117. 278 between two shapes 116-20 horizontal 113. 66-7 homology 30. 136 volume element 148. 284 homotopy 31.9 finite cell complex 25 FrCchet mean 203. 25. 41.100 of spaces of shapes in 3-space 73 of spaces of shapes in 4-space 77-8 homomorphisms 51. 225 geodesic distance 108 geodesics 105-33. 122. 48 axiom 50 equivalence 37 groups 30-3 type 36-9 horizontal geodesics 113 in pre-shape sphere 121 horizontal great circle 1I8 horizontal lift 252 of minimal geodesics 117. 261.18. 37. 207. 121 of X: 239-42 sphere 105 totally geodesic subspace 113 uniqueness 129 see also minimal geodesics geometric invariants 135. shape spaces for 249-81 general position 289 geodesic ball 203-4. 209. 138. 45-7. 136. 126 of vector field 260 horizontal subspace 113. 57.INDEX 301 explicit formula 190-200 exponential map 108 extended complex plane 277 extension of a group 285 familiarity 24 fibre 106. 85 of cell complexes 41-53 of shape spaces 55. 260 geometric properties 105 geometric structure 105 global homology group 105 global left-invariant fields 139 global structure of shape spaces 23-39 global topological invariants 24 globally defined vectors 138 great circle metric 13 great circular arc distance 205 group D 237-9 group extension 285 group homomorphisms 48 handle decomposition 288 Hausdorff spaces 22. 34. 106 hemispherical ball 36 hemispherical cell 36 higher dimensional manifold 149 higher dimensional shape spaces 23 1-47 Hodge duality 290 homeomorphic image 32. uniqueness 217-26 of random variables 202-3 of shape-measures. 32. 50 homeomorphisms 20-3.13 Fubini-Study metric 17-20 functoriality 48 Furstenberg’s theorem 271 Gamma-function 173 general manifolds. 69. 38. 204 of induced shape-measure 215 . uniqueness 206. 286 class 46 exact sequence 47-50.

270. 204. 28 1 hyperbolic trilaterals 280 hyperboloid 266 hyperboloid model 265 hyperplanes I61 identical topological hemispheres 6 inclusions 29-30 independent Gaussian distribution induced shape-measure 213 shape of the means 214 induced shape-measures 157-200 and shape of the means 214 complex normal distributions 172-6 FrCchet mean 215-16 independent Gaussian distribution 213 isotropic Gaussian distributions 160-4 non-isotropic Gaussian distributions 168-72 Poisson-Delaunay tiles 164-8 uniform distributions in compact convex set 176-80 in convex polygon 181-200 integer coefficients 45-7. 269. 261 Law of Large Numbers 203.8 1 hyperbolic spaces 264-7 hyperbolic 3-space 264. 265. 177. 56. 289 isotropic distribution 216. duality in 287-90 mapping cone 33. 64. 141. 38. 34. 95 intersection number 289 invariance property 269 INDEX invariant submanifold 112 Inverse Function Theorem 141 inverse image 66 inverse transformation 142 involution 30 isometric correspondence 15 isometric embedding 202 isometric involution 57 isometry 29-30. 261 horizontal vector 113. 137. 286. 254. 148 Laplace ‘second’ integral representation 169 Laplacian 253. 253 horosphere 276. 270. 171 Levi-Civitii connection 147 limits of shape spaces 68 linking 289-90 local horizontal geodesics 110 manifolds. 65. 272 . 48 Markov chain of triangles 270 Markov process 253.302 horizontal tangent space 258 horizontal tangent subspace 109 horizontal tangent vector 110. 100. 276 hyperbolic plane 264. 274. 277 hyperbolic triangle 279. 273-6 Lebesgue measure 169 left diagonal factor 141 Legendre polynomial 169. 139. 214 joint probability 161-2. 278. 173. 57. 284. 49 decomposition 33-6. 147. 51.17 isotropic Gaussian distributions 166 induced shape-measures 160-4 isotropy subgroup 106 joint distribution 169. 264. 140 isomorphisms 50. 163 jump-functions 193-9 Koszul formula 252 Kummer’s confluent hypergeometric function 173 Laplace-Beltrami operator 135. 279 hyperbolic distance 274 hyperbolic geometry 276 hyperbolic metric 266. 136. 87-8. 131. 276. 259. 287.

131-2 pre-shape matrix 34. 57. 167. 120 horizontal lift 117-18.18. 205. 277 monomorphism 136 monotonic decreasing rule 96 natural local vector fields 138-49 nesting principle 1 11. 136 pre-shape pattern 56 pre-shape space 3. 122.INDEX 303 matrix representation 7. 30.13 Newman’s theorem 38 non-isotropic distribution 217 non-linear spaces. 102-3 partition function 95. 124 pre-shape sphere 23. 165 Poisson point process 164. 98 pattern sequence 56.10 matrix spaces 136 maximally remote pairs 116 Mazur trick 66 mean in non-linear spaces 201-4 of random variables 202 mean shapes 201-29 metric geometry of shape spaces 13 metric space 120. 56 orientation preserving 42 isometry 254 orientation reversing 42 oriented cell boundary 43-5 oriented shape spaces 57. 32. 121. 135. 264. 206 action of SO(m) 106-7 principal component variances 169 principal consequence subgroup 275 probability distributions 269 probability measures 157 procrustean distance 114 procrustean mean shapes 226-9 procrustean metric 256 projections to lower strata 127-33 proper homeomorphism 66 proper map 66 . means in 201-4 non-singular manifold 112 non-singular part of shape space 107 non-singular square matrix 119 non-uniqueness 117 of minimal geodesics 120-4 normalising constant 160. 202 metric sphere 5 metrics on shape space 204-6 minimal geodesics 117. 268. 165 pre-shape 1.8 Poisson distribution 164. 112. 140. of spherical cells 42-3. 267. 67 open d-cell 25 open mapping 187-8 open neighbourhood 140 orientation. 28. 243 induced shape-measures 164. 105. size-and-shape spaces for 262-4 PoincarC duality 289. 28. 114. 56. 126 non-uniqueness 120-4 unique 123 Mobius transformation 159. 162 n-simplex 43 nth chain group 45 nth cochain group 46 nth cohomology group 46 nth homology group 46 null space 126 null vectors 267 O’Neill’s theorem 150 open cone 33. 290 Poisson-Delaunay tessellations 165. 98 path-length parameterisation 108 pattern matrix 56. 60 planar parameterisation 158 plane.

263-6. 175. 18. 201. 175. 254 Riemannian volume element 161 right diagonal factor 141 right-invariant vector fields 158 scalar curvature 155-6 sectional curvature 149-50. 267-73 Ricci curvature tensor 154-6 Riemann surface 274 Riemannian centre of mass 202 Riemannian curvature tensor 149-56. 172. 50. 110. 146-8. 163. 131. 56-7. 48 relative homotopy 49 relative mapping 48 relative metrics 209. 121. 39. 159. 260 shape configuration 2 coordinates 15. 203. 156. 48. 135. I 26.304 INDEX pseudo-polar decomposition 115-17. 258 decomposition 114-15. 124. 237 quotient groups 289 quotient map 34. 272 re-expressed through natural local vector fields 138-49 Riemannian structure 24. 90 relative homology group 47. 121 pseudo-singular values 126-9. 193 singular values decomposition 114-15. 251. 161. 212 Riemannian length 252 Riemannian manifolds 1 13.13 singular tessellation 181-90. 105. 123 . 159. 65 singular points 111. 269. 256-9. 128. 59. 252 Riemannian distance 116. 272 quotient metric 12 quotient space 106 quotient topology 110 Radon-Nikodgm derivatives 159. 135-7. 87. 171-3. 150. 190. 178-81. 213-17 and induced shape-measure 214 independent Gaussian distribution 214 statistics 16 theory. 251. 112.8 1 global structure 23-39 higher dimensional 23 1-47 in 2-space 70-1 in 3-space 7 1-2 in 4-space 72-8 limits of 68 shape-theoretic distance 14 similar triangles 17 simple connectivity 30-3. 124. 162-7. 49. 122. 252. 159. 38. 169. 123. origins 1 shape spaces 4 array of 5 coordinates 19 dimensions 6 example 287 general manifolds 249 . 163. 176. 107-11. 253. 62. 161. 160. 136-40. 19. 153. 110. 203. 209. 110. 253 Riemannian metric 6. 89. 158. 205. 213-15 random triangles in convex sets 180 re-based essential complexes for spaces of unoriented shapes in 2-space 81 in 3-space 83 in 4-space 76. 17 distance 14 modulo Q 251 of the means 201. 209. 140. 135-56 Riemannian submanifold 108 Riemannian submersion 107. 251-3. 127. 112. 85 reduced homology groups 47.

138 strata 112. 144. 137. 50 torsion product 285. 286-7. 135. 138-42 skew-symmetric tensor field 252 spheres 264-7 geodesics 105 spherical cells 28. 87. 228 sub-cellular complex 47. 262 skew-symmetric matrix 108. 124 distances and projections to lower strata 127-33 stratification 30.1 stabiliser 106 standard coordinate matrix 136. 146. 143. 253 Teichmuller metric 274 tensor product 284-6 topological equivalence 23 topological invariants distinguishing between shape spaces 52-3 topological manifolds 38. 6. 137.12 submodule 61 super-abundant shape variables 14 suspension 33 isomorphism 51. 64. 1 12 Strong Law of Large Numbers 204. 206. 51. 223. 122. 48 sub-chain complex 87 subcomplexes 59-62. 290 unoriented shape spaces 57 duality 100-2 in 2-space 78-80 in 3-space 80-2 in 4-space 82-4 unoriented shapes 64 singularity set 107 size-and-shape 2 size-and-shape spaces 254-62 curvature of 260 for plane 262-4 sized-and-centred configurations 2 skew-symmetric 144. 56 spherical geometry 276 spherical triangles 278. 121. 262 tangent vectors 108. cellular decomposition of 25 -7 unitary operators on groups 283-4 Universal Coefficient Theorem 46. 125 tangency condition 142 tangent space 107. 90 subgroups 289 submanifolds 110. 296 totally geodesic subspace 113 translation vector 255 triangles in the sphere 276-81 triangular shapes 15 trivial homology 62 two-dimensional representation of 231-3 uniform distributions induced shape-measures in compact convex set 176-80 in convex polygon 181-200 uniqueness 1 17 of geodesics 129 unit sphere. 289. 253. 119. 71. 39. 280 splitting of exact sequences 290. 109. 135.INDEX 305 symmetric 117. 138. 89 . 41 orientation 42-3. 122. 141-5 standard coordinates 142 standard integer coefficients 286 standard orientation 45 statistics of shape 16 stepping-stone method 190. 129 symmetric linear map 260 symmetric matrix 20. 67 topological spaces 30 topological 5-sphere 4 topological spheres 4. 199-200 stereographic coordinates 149 Stiefel manifold 106. 149. 66.

140. 262 horizontal lift 260 vector space 260 vector subspace 262 vertical subspace 140 vertical tangent space 258 vertical tangent subspace 109. 151. 140 INDEX vertices 25-6 virtual vertices 280 volume element 148 von Neumann’s ergodic theorem 272 Weierstrass Pfunction 274 zero-cells 25-63 323 47 .306 vector field 139.