Upload
others
View
5
Download
0
Embed Size (px)
Citation preview
Grassmannian Geometry
of Framed Curve Spaces
by
Thomas Richard Needham
(Under the direction of Jason Cantarella)
Abstract
We develop a general framework for solving a variety of variational and computer vision prob-
lems involving framed space curves. Our approach is to study the global Riemannian and
symplectic geometry of the moduli space of similarity classes of framed loops in R3. We show
that this space is an infinite-dimensional Frechet manifold with a natural Kahler structure.
The proof uses novel coordinates on the space of framed paths, which are used to locally
identify the moduli space with the Grassmannian of 2-planes in an infinite-dimensional com-
plex vector space. Results on the geometry of framed loop space are obtained, including
a characterization of its sectional curvatures and an in-depth description of some natural
Hamiltonian group actions on the space. We give a variety of applications of this structure,
such as a classification of critical points of a generalization of the Kirchhoff elastic energy
functional and a shape recognition algorithm for, e.g., protein backbones. We show connec-
tions between previous results by various authors on infinite-dimensional Kahler geometry,
fluid dynamics and moduli spaces of linkages.
Index words: Infinite-dimensional geometry, symplectic geometry, Riemanniangeometry, elastic shape matching
Grassmannian Geometry
of Framed Curve Spaces
by
Thomas Richard Needham
B.S., University of Wisconsin-Milwaukee, 2007
M.S., University of Wisconsin-Milwaukee, 2009
A Dissertation Submitted to the Graduate Faculty
of The University of Georgia in Partial Fulfillment
of the
Requirements for the Degree
Doctor of Philosophy
Athens, Georgia
2016
c©2016
Thomas Richard Needham
All Rights Reserved
Grassmannian Geometry
of Framed Curve Spaces
by
Thomas Richard Needham
Approved:
Major Professor: Jason Cantarella
Committee: Joseph H.G. FuDavid T. GayWilliam H. KazezMichael Usher
Electronic Version Approved:
Suzanne BarbourDean of the Graduate SchoolThe University of GeorgiaMay 2016
Acknowledgments
This thesis is dedicated to my wonderful wife Chelsie. She provided unwavering support
throughout my work in graduate school; in particular, she was completely indispensable in
keeping me sane throughout the stressful process of writing a dissertation and venturing into
the academic job market. I would never have achieved these goals without her, and I can’t
thank her enough for her patience and encouragement.
I am also, of course, deeply indebted to my advisor Jason Cantarella. He spent countless
hours discussing the material in this work with me and was consistently supportive, while
never failing to push me to work just beyond my abilities. I hope that I have, at least in
some small part, internalized some of his amazing mathematical intuition. Overall, I can’t
imagine a better advisor.
There are many, many more people who deserve thanks, and I will mention a few of them
here. I would like to thank my committee members Joseph Fu, David Gay, William Kazez
and Michael Usher, each of whom has taught me interesting mathematics at some point in
my time at UGA. I would also like to thank Simon Foucart, who has graciously mentored me
in the field of compressive sensing. Countless other mathematicians at UGA and elsewhere
have helped me throughout my time in graduate school, and special thanks goes to Sybilla
Beckmann, Sa’ar Hersonsky, Ken Millett, Clayton Shonkwiler and Dennis Sullivan. I would
also like to thank the office staff at the UGA math department, especially Laura Ackerley.
iv
Finally, I would like to thank the many close friends that I have made in Athens, both inside
and outside of the math department.
v
Contents
1 Introduction 1
1.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Main Results and Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2 Preliminaries and Notation 15
2.1 Classical Geometry of Framed Curves . . . . . . . . . . . . . . . . . . . . . . 15
2.2 Frechet Spaces and Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.3 Preliminaries from Differential Geometry . . . . . . . . . . . . . . . . . . . . 35
2.4 Quaternions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3 Complex Coordinates for Framed Paths and Loops 44
3.1 Framed Path Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.2 Framed Loop Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4 The Geometry of Spaces of Framed Paths and Loops 82
4.1 The Symplectic Structure of Framed Loop Space . . . . . . . . . . . . . . . . 82
4.2 Principal Bundle Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
4.3 Symplectic Reduction by the action of LS1{S1 . . . . . . . . . . . . . . . . . 111
4.4 The actions of Diff`pr0, 2sq and Diff`pS1q . . . . . . . . . . . . . . . . . . . . 130
4.5 Riemannian Geometry of Framed Path and Loop Spaces . . . . . . . . . . . 143
vi
5 Application: Critical Points of Energy Functionals 159
5.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
5.2 Weighted Total Twist . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
5.3 Total Elastic Energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
6 Application: Shape Recognition 196
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
6.2 Shape Matching for Framed Paths . . . . . . . . . . . . . . . . . . . . . . . . 199
Bibliography 217
vii
List of Figures
3.1 An element of AC. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.2 A knot diagram with Wr “ 3´ 5 “ ´2 . . . . . . . . . . . . . . . . . . . . . . . 69
3.3 Links determined by (a) pγ0, V0q, (b) A perturbation of pγ0, V0q along the homotopy,
and (c) pγ1, V1q. In each image, the blue loop is the image of γ and the red loop is
the image of γ ` εV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.1 The complex structure ofM. . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
4.2 Regions described in the proof of Lemma 4.3.3. . . . . . . . . . . . . . . . . 116
4.3 Geodesic in M joining a p2, 3q-torus knot to a p2, 5q-torus knot. . . . . . . . 146
5.1 A torus knot realized as a pushoff of a critical point of ĂTw and the corre-
sponding Clifford torus knot as a stereographic projection of Φ2,3. . . . . . . 169
5.2 The 1-parameter family of critical points with h “ 2, k “ 1. . . . . . . . . . 194
5.3 The 1-parameter family of critical points with h “ 3, k “ 2. . . . . . . . . . 195
6.1 Geodesic before reparameterization. . . . . . . . . . . . . . . . . . . . . . . . 215
6.2 Geodesic after reparameterization. . . . . . . . . . . . . . . . . . . . . . . . . 215
6.3 Geodesic distances. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
viii
Chapter 1
Introduction
This chapter introduces the results of the thesis. We begin with a general discussion of the
motivations for this work. We then give a more specific outline of the structure of the paper
and highlight the main results.
1.1 Overview
1.1.1 Motivating Questions
The overarching goal of this thesis is to develop a convenient framework to answer a diverse
collection of questions such as:
1. What is the natural notion of shape similarity for protein backbones?
2. What are the equilibrium shapes of an extensible elastic rod?
3. How should one efficiently compute the gradient flow of a knot energy functional such
as Mobius Energy (see [61])? What does “gradient flow” mean in this setting?
The common thread between these questions is that they can be rephrased as questions
about the space of shapes of framed curves in R3. By a framed curve in R3, we mean a
1
pair pγ, V q consisting of a parameterized open path or closed loop γ in R3 together with a
unit normal vector field V along γ. The shape of a framed curve refers to an equivalence
class with respect to some “shape-preserving” group action—e.g., the Euclidean group of
rotations and translations, or the diffeomorphism group of reparameterizations.
Thus it remains to explain what is meant by the space of shapes of framed curves. We will
show in Chapter 3 that many natural collections of shapes of framed paths and loops admit
manifold structures. In this setting, manifold means a topological space which is locally
modelled on a function space such that the transition functions are smooth in an appropriate
sense. The function space formalism that we adopt here is that of (tame) Frechet spaces.
These are infinite-dimensional vector spaces which are flexible enough to describe spaces of
smooth maps (as opposed to Hilbert or Banach spaces), but which have enough structure to
admit an inverse function theorem (as opposed to more general topological vector spaces).
We work with spaces of smooth framed curves with the viewpoint that spaces of framed
curves of lower regularity are realized as completions of the smooth spaces. Thus any results
obtained in the smooth category should generalize without much work to, say, the L2 category
or a Sobolev regularity category. For example, the results in this thesis can be generalized
to spaces of absolutely continuous framed curves almost immediately. Another reason for
working in the smooth category is that we wish to study actions of diffeomorphism groups,
and topological groups of Ck-diffeomorphisms do not admit Lie group structures [62].
From this perspective we are able to outline solution strategies for the above questions
as follows:
1. The space of protein backbone shapes can be viewed as the infinite-dimensional mani-
fold of equivalence classes of framed paths, where equivalence is with respect to transla-
tion, rotation, scaling and reparameterization. This manifold has a natural Riemannian
metric, so shape similarity of protein backbones can be measured by geodesic distance.
2
2. An extensible elastic rod in equilibrium can be viewed as a critical point of a certain
functional on the manifold of Euclidean similarity classes of framed paths or loops.
Thus the goal is to classify the critical point set of this functional.
3. A knot energy functional can be viewed as a function on the manifold of Euclidean
similarity classes of framed loops. This manifold also admits a natural metric, so that
gradient is well-defined. Then the task is to determine the existence of the gradient
flow and to develop effective methods of computing it.
Now that we have strategies for attacking the questions, our concern is whether these
strategies are actually tractable. For example, the usefulness of the solution strategy for
Question 1 depends on our ability to compute geodesic distances efficiently. While the
(short time) existence of geodesics in a finite-dimensional Riemannian manifold is trivial,
the geodesic equation on an infinite-dimensional manifold becomes a PDE—thus even short
time solutions are not guaranteed to exist. Moreover, even if solutions do exist, actually
computing geodesic distance could conceivably be quite difficult. We contend that the main
results of this thesis also make these solution strategies tractable. The reason is that we do
not only show that many shape spaces of framed paths and loops admit manifold structures
and natural Riemannian metrics; we also show that the shape spaces are isometric to infinite-
dimensional versions of classical manifolds such as spheres and Grassmannians. Despite
their infinite-dimensionality, these classical manifolds are remarkably easy to work with.
For example, the infinite-dimensional Grassmannians that appear admit completely explicit
geodesics.
1.1.2 Complex Coordinates for Framed Curves
As is frequently the case in differential geometry, a key step in unlocking the structure of the
manifolds of interest is to find an appropriate coordinate system. Our approach to studying
3
manifolds of shapes of framed curves is to develop a convenient coordinate system where
a pair of complex-valued functions represents a framed path or loop. The idea of using
functional coordinates for shape spaces of curves is actually quite classical, and we now draw
parallels with the classical approach.
It is a basic fact from elementary differential geometry that the curvature κ and torsion τ
of a space curve completely determine the curve up to Euclidean isometry, provided κ ą 0. A
natural question, attributed independently to Fenchel [18] and to Chern [34] is the following:
What are necessary and sufficient conditions on κ and τ that ensure that the corresponding
space curve is periodic? Certainly a necessary condition is that κ and τ are themselves
periodic, but periodicity of κ and τ is far from sufficient (e.g. constant κ and τ produce a
non-periodic helix). Hwang Cheng Chung gave a complete answer to this question in [34],
but the sufficient conditions include that an infinite series of integrals of functions of κ and
τ vanish. Various other answers exist (e.g., [22, 60]), but the sufficient conditions given are
similarly impractical. The general consensus is that the question has no practical answer.
This situation can be interpreted as follows. One can view pκ, τq as a global system of
coordinates on the infinite-dimensional manifold of isometry classes of arclength parameter-
ized paths in R3 with nonvanishing curvature. This has the benefit of endowing an seemingly
complicated infinite-dimensional manifold with the differential structure of an open subset
of a vector space; that is, we identify the space
tγ P C8pr0, 1s,R3q | }γ1} ” 1, κ ą 0u{ttranslation, rotationu
with
tpκ, τq P C8pr0, 1s,Rą0q ˆ C8pr0, 1s,Rq | κ ą 0u Ă C8pr0, 1s,R2
q.
4
However, the lack of reasonable conditions on κ and τ which guarantee periodicity of a curve
means that there are no constraint equations in κ and τ that can be used to identify the
submanifold consisting of closed loops.
In this thesis we introduce a global coordinate system of pairs of complex-valued func-
tions, denoted pφ, ψq, on the space of framed paths. We show that pφ, ψq determine a
smoothly closed framed loop if and only if they satisfy two simple constraint equations, in
stark contrast with the classical pκ, τq-coordinates. The simple constraint equations turn
the moduli space of smoothly closed framed loops into a finite-codimension submanifold of
a vector space, and this allows us to work with it in an extraordinarily hands-on way.
1.1.3 Why Framed Curves?
Framed paths and loops are ubiquitous in biology and physics. They are frequently used to
model protein backbones [28, 33, 72], circular DNA [13, 19, 31], magnetic field lines [58] and
oriented trajectories [25], and they are the central objects of study in the theory of Kirchhoff
elastic rods [66]. The main purpose of this paper is to develop a natural framework for
studying the dynamics of these physical objects in terms of the geometry and topology of
moduli spaces of framed paths and loops.
Framed Curves are also interesting from a pure mathematics perspective; e.g. it is well-
known that every orientable 3-manifold can be obtained by surgery on a framed link in the
3-sphere [46]. The geometric structures of a variety of interesting spaces of moduli spaces of
loops in 3-manifolds have been studied extensively in, e.g., [8, 49, 55]. We show in this thesis
that, in the case of R3, geometric structures on loop spaces can be seen as arising from more
complicated structures on the moduli space of framed loops. We conjecture that this is also
true in the general 3-manifold setting, and this will be the subject of future work.
5
1.2 Main Results and Outline
1.2.1 Some Definitions
Before giving a detailed outline of the thesis, some more precise definitions are in order. We
define a framed path to be a pair pγ, V q, where γ is a parameterized immersed path in R3
and V is a unit normal vector field along γ. For concreteness, we take the domain of any
parameterized path to be r0, 2s—this convention is a normalization whose convenience will
be made apparent later on. A relatively framed path is an equivalence class of a framed path
with respect to the circle-action of rotating all frame vectors simultaneously around γ by
the same angle. A framed loop (respectively, relatively framed loop) is an ordered pair of
maps pγ, V q : S1 Ñ R3 ˆ R3, such that γ is an immersion and V is a normal vector field
along γ (respectively, an equivalence class of normal vector fields up to the frame rotation
circle-action). We identify the circle S1 with the smooth quotient r0, 2s{p0 „ 2q so that the
collection of framed loops can be considered as a subset of the collection of framed paths;
i.e., a framed loop is a framed path which happens to smoothly close. To distinguish this
realization of S1 from the more standard case, will use the notation Sn to denote the standard
radius-1 n sphere embedded in Rn`1.
Let M denote a finite-dimensional manifold. Throughout the rest of the thesis, we will
use the notation
PM :“ C8pr0, 2s,Mq and LM :“ C8pS1,Mq
for the path space and loop space of M , respectively.
1.2.2 Chapter 2
Chapter 2 collects some background material. We summarize some results from the classical
theory of the geometry of framed curves in R3. We then give some background on the theory
6
of Frechet spaces, paying special attention to the most relevant cases of path and loop spaces
of a finite-dimensional manifold. Next we collect some results on Riemannian and symplectic
geometry, both in the classical setting and in the infinite-dimensional setting. Finally, we
give a brief description of the geometry of the quaternions. This chapter also serves to set
much of the basic notation that will be used throughout the thesis.
1.2.3 Chapter 3
The first original results of the thesis appear in this chapter. We define the moduli space of
framed paths
S :“ tframed pathsu{ttranslation, scalingu,
where scaling refers to scaling the length of the base curve (for concreteness, we take all
framed loops to have fixed length 2). The space S is shown to have the structure of a
manifold modelled on Frechet spaces. It is also shown to admit a natural Riemannian metric.
This is constructed by first defining a 4-parameter family of metrics ga,b,c,d on S called the
framed curve elastic metrics. This definition is a direct generalization of the elastic metrics
ga,b defined on spaces of plane curves by Mio, Srivastava and Joshi [57]. It is then shown that
g1,1,1,1 arises as a pullback metric from a natural embedding of S into an infinite-dimensional
Lie group with a left-invariant metric (Proposition 3.1.5).
Next we utilize the well-known Hopf map, which gives an isometric double cover SUp2q Ñ
SOp3q, to prove that, with respect to g1,1,1,1, S has the geometry of an infinite-dimensional
sphere.
Theorem (Theorem 3.1.12). The moduli space of framed paths S is isometrically double
covered by an open subset of the L2 sphere in PC2.
7
Next we move on to the moduli space of framed loops
M :“ trelatively framed loopsu{ttranslation, scaling, rotationu
consisting of Euclidean similarity classes of relatively framed loops. This space has two
connected components, which are indexed by mod-2 self-linking number (Corollary 3.2.5).
The moduli space admits a natural metric induced by g1,1,1,1, and our next main theorem is
roughly stated as:
Theorem (Theorem 3.2.8). Each component of the moduli space of relatively framed loops
M is isometric to an open subset of the Grassmann manifold of complex 2-planes in LC.
Theorems 3.1.12 and 3.2.8 serve as the basis for the rest of our results. We recall that the
basic premise of the thesis is that many diverse questions can be phrased as questions about
the geometry of moduli spaces such as S and M. These are, on first glance, quite abstract
manifolds, but the results of this chapter show that (with respect to natural Riemannian
metrics), these spaces actually have surprisingly well-behaved geometry.
We also mention here that these theorems generalize some remarkable work of Younes,
Michor, Shah and Mumford [76]. It was shown in [76] that various moduli spaces of immersed
plane curves admit a natural metric (an example of one of the elastic metrics of Mio et. al.)
with respect to which they are isometric to classical manifolds (e.g. real infinite-dimensional
Grassmannians). The authors were primarily interested in the topic for its application to
shape recognition. We are also interested in this application and it is treated in Chapter
6. Theorems 3.1.12 and 3.2.8 generalize the work of Younes et. al. in two ways. First,
the shape spaces they consider embed as totally geodesic submanifolds of the shape spaces
considered here. Second, the proofs give some extra insight into why the results of [76] are
true, and demonstrate the dimensions where such strategies could potentially work—these
are the dimensions where Hopf fibrations exist.
8
1.2.4 Chapter 4
In this chapter we more thoroughly investigate the geometry of S and M. We begin by
noting that
Gr2pLCq :“ tcomplex 2-planes in LCu
is a complex Kahler manifold. We show that the Kahler structure is obtained by viewing
Gr2pLCq as a Kahler reduction of LC2 with its natural L2 Kahler structure. From this we
are able to describe an interesting complex structure on the moduli space of framed loops
M (Corollary 4.1.8).
Next we study the action of the Lie group LS1{S1 onM. Recall that we use the notation
S1 to specifically denote the standard unit circle embedded in C « R2. The action of this
Lie group is by adjusting the framing of (an equivalence class of) a framed loop. Let
B :“ t(unframed) immersed loops in R3u{ttranslation, rotation, scalingu.
We show that B is a Frechet manifold (Lemma 4.2.1) and then prove the following theorem.
Theorem (Theorem 4.2.5). Framed loop space M has the structure of a principal bundle
over B with structure group LS1{S1.
This theorem has a variety of applications. An immediate corollary (Corollary 4.2.6)
roughly states that any curve framing algorithm (see, e.g., [26]) satisfying some reasonable
conditions is bound to fail on some subset of the space of immersed curves.
Since M has a Kahler structure, a natural question to ask is whether the action of the
Lie group LS1{S1 is Hamiltonian. We show that it is by explicitly defining its momentum
map. Moreover, we show that symplectic reduction by this group produces a space which
9
has been previously studied. This is the moduli space of Millson and Zombro [55] defined by
MMZ :“ tarclength-parameterized loops in R3u{ttranslation, rotationu.
Millson and Zombro showed that MMZ admits a symplectic structure by showing that it is
realized as a symplectic reduction of the loop space of the 2-sphere by the action of SOp3q.
We give an alternate description of the symplectic structure.
Theorem (Theorem 4.3.7). The space of Millson and ZombroMMZ is realized as symplectic
reduction of M by the action of LS1{S1.
Next we examine the action of the group Diff`pS1q of orientation-preserving diffeomor-
phisms onM by reparameterization. The action is not free, and because of this we focus on
subgroup Diff`0 pS1q of basepoint-preserving diffeomorphisms, which does have a free action
onM. We show that the spaceM{Diff`0 pS1q is a manifold (Proposition 4.4.3) and that the
tangent spaces toM split orthogonally with respect to the action of Diff`0 pS1q (Proposition
4.4.4).
We conclude this chapter by studying the Riemannian geometry ofM and S. In particu-
lar, we are able to explicitly describe the geodesics of these moduli spaces using the fact that
they are isometric to classical manifolds. Such a description is extremely useful for shape
recognition applications, as we shall see in Chapter 6. This description is also used to prove
a result on the sectional curvatures of M and its quotient by Diff`pS1q.
Theorem (Theorem 4.5.8). The spaces M and M{Diff`0 pS1q both have non-negative sec-
tional curvatures with respect to their natural metrics.
In shape recognition applications, one typically wants to optimize geodesic distance over
reparameterizations, and for this reason it is useful to know that the quotient is nonnegatively
curved. Furthermore, nonnegative sectional curvature is important for numerical stability
10
if one were to use a Newton’s method-type approximation for the gradient flow of some
functional onM. The proof of this theorem uses a variety of results of independent interest.
These include that the exponential map of M is well-defined (Proposition 4.5.6) and that
for any collection of 2-planes in Gr2pLCq, there exists a totally geodesic copy of a finite-
dimensional complex Grassmannian containing the collection (Proposition 4.5.2). The latter
result somewhat formalizes the notion that the geometry of Gr2pLCq can be understood from
a finite-dimensional perspective.
1.2.5 Chapter 5
Chapter 5 turns towards applications to energy functionals on the moduli space M. First
we study the weighted total twist functional, ĂTw : M Ñ R, which measures the total
accumulated twisting of the normal vector field along its base curve, weighted by the pa-
rameterization speed of the curve. We show that ĂTw is a moment map for a natural circle
action onM (Proposition 5.2.2). The circle which acts is a subgroup of Diff`pS1q, acting by
rotating the base point of a based framed loop. We use this fact to characterize the critical
points of ĂTw:
Theorem (Theorem 5.2.8). The critical points of ĂTw : M Ñ R are equivalence classes
of framed loops pγ, V q such that γ is an arclength parameterized, length-2, multiply covered
round circle and V has constant twist rate. Thus a small pushoff γ ` εV forms a multiply-
covered torus knot and every torus knot type is realized as such a pushoff for some critical
point. Each critical framed loop has a complex-coordinate representative as a torus knot on
the standard Clifford torus in S3.
11
Next we introduce the total elastic energy functional, E :MÑ R, which is a straightfor-
ward generalization of classical Kirchhoff elastic energy
EKir “
ż
S1
κ2` tw2 ds
of a framed loop pγ, V q. Here κ is the curvature of γ, tw is the twist rate of V around γ
and ds is measure with respect to arclength of γ (see Section 2.1.2 for the definition of twist
rate and Section 5.1.1 for a short history of the study of Kirchhoff elastic energy and related
functionals). Kirchhoff elastic energy should be viewed as the potential energy of a uniform
inextensible elastic rod, whereas our generalization can be viewed as the potential energy of
an extensible rod whose bending and twisting tension depends on how it is stretched.
Total elastic energy has an incredibly natural representation in our coordinate system.
We are able to completely classify the critical points of E, and in fact we can give explicit
parameterizations of all critical points in our coordinate system. This is due to the following
theorem, which is stated here in a nontechnical form.
Theorem (Theorem 5.3.10). Each critical point of E :MÑ R has a unique representation
in complex coordinates as a pair of eigenfunctions of the operator φ ÞÑ φ2.
To get a more geometric picture of the types of critical points of E, we show:
Theorem (Theorem 5.3.16). Let h, k be integers such that gcdph, h`kq “ gcdpk, h`kq “ 1.
The critical point set of E :MÑ R contains a 1-parameter family of equivalence classes of
framed loops pγu, Vuq such that
(i) pγ0, V0q is an arclength parameterized h-times-covered round circle linked k-times,
(ii) γε is an ph, h` kq-torus knot for sufficiently small ε ą 0,
(iii) γ1´ε is a p´k, h` kq torus knot for sufficiently small ε ą 0,
12
(iv) pγ1, V1q is an arclength parameterized k-times-covered round circle linked ´h-times.
Moreover, there exists a unique u P p0, 1q such that γu is not embedded.
This result can be seen as an analog of Ivey and Singer’s classification of critical points
of the classical Kirchhoff elastic energy functional [35]. The Ivey-Singer result shows that
critical points of the classical elastic energy come in 1-parameter families. Their result was
concerned with arclength parameterized framed loops, and we see here that allowing for
arbitrary parameterizations has produced a richer critical point set, as one might expect.
1.2.6 Chapter 6
In this chapter we address the application of this framework to shape recognition problems.
We first address the geodesic distance problem in the quotient space S{SOp3q, which can be
interpreted as optimizing geodesic distance in the L2-sphere of PC2. Optimizing registration
over rotations is an important part of any shape recognition algorithm, and is typically
treated computationally via matrix decompositions (e.g., [68]). In our setup, the problem
has a pleasing closed form solution, which is most naturally stated by identifying PC2 with
the path space of the quaternions H.
Theorem (Theorem 6.2.2). Geodesic distance in S{SOp3q has an explicit closed form. In
quaternionic coordinates q0 and q1, the distance is given by
min
distpq0, q1 ¨ pq | p P S3Ă H
(
“?
2 arccos1
2
›
›
›
›
ż 2
0
q1 ¨ q0 dt
›
›
›
›
H,
where dist denotes geodesic distance in the radius-?
2 L2-sphere of PH.
We then turn to computing distance in S{Diff`pr0, 2sq; i.e., we wish to optimize geodesic
distance over reparameterizations. Here a technical issue arises in that the full L2-sphere
of PC2 modulo the induced action of Diff`pr0, 2sq is a non-Hausdorff space. This is treated
13
by replacing this quotient space by the space of L2-closures of Diff`pr0, 2sq-orbits. We show
that optimizing over reparameterizations in this new space produces a well-defined distance
metric (Proposition 6.2.5).
The chapter concludes with a desription of a computational algorithm for optimizing
over reparameterizations. This optimization is approximate (by necessity), as we replace the
group of diffeomorphisms with the set of increasing piecewise-linear homeomorphisms with
vertices on a fixed grid.
14
Chapter 2
Preliminaries and Notation
2.1 Classical Geometry of Framed Curves
In this section we recall basic facts about the differential geometry of framed space curves,
following [15] and [27]. We also set some of the notation to be used throughout the rest of
the thesis.
2.1.1 Definitions
A smooth space curve is a C8 map γ : Σ Ñ R3, where Σ is a one-dimensional manifold,
perhaps with boundary. We will almost exclusively consider the case where either Σ “ r0, 2s
(this particular closed interval is convenient for normalization purposes), or Σ “ S1, where
S1 denotes the round circle of length 2. We will identify S1 with the smooth quotient
S1« r0, 2s{p0 „ 2q « R{2Z.
Those space curves with domain r0, 2s will be called paths and those with domain S1 will be
called loops.
15
Our primary interest will be immersed space curves, where γ1ptq ‰ 0 for all t P Σ. An
immersed space curve has a well-defined rank-2 normal bundle, which is necessarily trivial.
A trivialization of the unit normal bundle of a particular curve γ is called a framing. For
concreteness, we realize a framing as a smooth map V : Σ Ñ R3 satisfying
〈γ1ptq, V ptq〉 “ 0 and }V ptq} “ 1 @ t P Σ.
In the above and throughout the thesis, 〈¨, ¨〉 refers to the standard Euclidean inner product
on R3—the various other inner products which will arise will be given specialized notation.
Likewise, } ¨ } will be reserved for Euclidean norm in R3. A pair pγ, V q will be called a framed
curve (respectively, framed path if the domain is r0, 2s or framed loop if the domain is S1).
Although it is implicit in the definition, we emphasize that in the case that γ is a loop, we
require that V is also a closed loop.
We note that by our conventions, a pair of framed curves pγ1, V1q and pγ2, V2q which
differ only by a reparameterization—that is, pγ2ptq, V2ptqq “ pγ1 ˝ ρptq, V1 ˝ ρptqq for some
orientation-preserving diffeomorphism ρ of Σ—are considered to be distinct framed curves.
It should also be noted that other authors use other definitions of a curve framing. For
example, [27] takes a framed curve to be a triple pγ, V,W q, where γ and V fill the same
role as our definition and W “ γ1{}γ1} ˆ V , while [67] defines a framing of γ to be a one-
parameter family of planes Π such that γ1ptq P Πptq for all t. Each of these definitions is
clearly equivalent to the one that we will use, perhaps up to choice of orientation.
2.1.2 Geometric Invariants and Examples
Before giving some fundamental examples of framed space curves, it will be useful to define
the various geometric invariants of a framed space curve—these are geometric invariants in
the sense that they do not depend on a particular parameterization. First, we introduce the
16
invariants of a smooth immersed space curve γ. The curvature of γ, denoted κ, is defined
by the formula
κptq :“
›
›
›
›
d
dsT ptq
›
›
›
›
,
where
d
ds:“
1
}γ1ptq}
d
dt
denotes derivative with respect to arclength and
T ptq :“γ1ptq
}γ1ptq}
denotes the unit tangent vector to γptq. Working this out, κ can be written more explicitly
as
κ2 :“
›
›
›
›
γ2
}γ1}2
›
›
›
›
2
´
⟨γ2
}γ1}2, T
⟩2
.
In the above, we supress explicit dependence on t P Σ. This convention will be adopted
frequently. Note that if γ is arclength-parameterized—i.e., }γ1} ” 1—then this formula
reduces to κ “ }γ2}.
Example 2.1.1. For a curve γ with nonvanishing curvature, we have our first example of a
framing. Define the principal normal vector to γ to be
N :“1
κ
d
dsT “
1
κ
ˆ
γ2
}γ1}2´
⟨γ2
}γ1}2, T
⟩T
˙
.
The principal normal N is unit length, by design. Then pγ,Nq defines a framed curve, called
the Frenet frame of γ. Since N is defined pointwise by a geometric invariant of γ, it is clear
that if γ is a closed loop, then N is as well, so that pγ,Nq is well-defined according to our
conventions.
17
One should observe that if we treat N as the fundamental object, then κ is recovered by
the formula
κ “
⟨d
dsT,N
⟩.
Similarly, another geometric invariant of a space curve γ with nonvanishing curvature is its
torsion τ , defined by
τ :“
⟨d
dsN,B
⟩,
where the vector B :“ T ˆ N is called the binormal vector of γ. We have now defined the
ingredients necessary to state the fundamental theorem of space curves.
Theorem 2.1.1 (Fundamental Theorem of Space Curves, [15], Section 1-5). Let κ, τ :
r0, 2s Ñ R be arbitrary smooth functions with κ ą 0. There is an arclength-parameterized
space curve γ : r0, 2s Ñ R3 with curvature κ and torsion τ which is unique up to translations
and rigid rotations.
Proof. By the above definitions, the problem of recovering γ can be rephrased as solving the
vector ODE
d
ds
¨
˚
˚
˚
˚
˝
T
N
B
˛
‹
‹
‹
‹
‚
“
¨
˚
˚
˚
˚
˝
0 κ 0
´κ 0 τ
0 ´τ 0
˛
‹
‹
‹
‹
‚
¨
˚
˚
˚
˚
˝
T
N
B
˛
‹
‹
‹
‹
‚
,
and then defining γ by the formula
γptq “
ż t
0
T`
t˘
dt.
By existence and uniqueness of solutions of ODEs, the first step is possible up to choice of
initial position pT p0q, Np0q, Bp0qq, hence we have uniqueness only up to rotations. Moreover,
implicit in our definition of γ was the choice of starting point at the origin, hence uniqueness
up to translations.
18
Thus pκ, τq can be seen as global coordinates on the space
tarclength parameterized paths with nonzero curvatureu{ttranslation, rotationu.
However, these coordinates are not suitable to describe the subspace of closed loops and this
was a major motivation for the new coordinate system developed in this thesis (see Section
3.1 for more details).
Now we consider a general framed curve pγ, V q. The definitions of κ and τ given above
immediately generalize to this setting. It will be convenient to use the notation W :“ T ˆV .
Then we define the Darboux curvatures κ1 and κ2 of pγ, V q by the formulas
κ1 “ κ1pγ, V q :“
⟨d
dsT, V
⟩and κ2 “ κ2pγ, V q :“
⟨d
dsT,W
⟩
and the twist rate tw of pγ, V q by
tw “ twpγ, V q :“
⟨d
dsV,W
⟩.
Note that in the case that V “ N , κ1 “ κ, κ2 “ 0 and tw “ τ .
Example 2.1.2. Using the definition of tw, we obtain a second example of a framing for an
open curve γ, called the Bishop framing [4]. This is the framing V obtained by demanding
that the twist rate of pγ, V q is identically zero. This amounts to solving a simple ODE, and
“the” Bishop framing is only well-defined up to an initial choice of V p0q. In general, we will
use the term relative framing to refer to an equivalence class of framings up to choice of initial
frame vector. Note that for a closed loop the Bishop framing will not necessarily be closed,
so the Bishop framing is not always well-defined for loops. Indeed, it follows easily from the
Calugareanu-White-Fuller Theorem [20] that a closed loop admitting a Bishop framing is
highly non-generic.
19
Relatively framed curves can be defined more geometrically as follows. For any given
framed curve pγ, V q, we have a circle action defined by
θ ¨ pγ, V q :“ pγ, cos θV ` sin θW q, θ P r0, 2πs{p0 „ 2πq.
A relatively framed curve is an equivalence class of a framed curve under this action. For a
framed curve pγ, V q, we denote the corresponding relatively framed curve by pγ, rV sq. The
geometric invariants of θ ¨ pγ, V q are easily computed:
κ1pθ ¨ pγ, V qq “
⟨d
dsT, cos θV ` sin θW
⟩“ cos θκ1pγ, V q ` sin θκ2pγ, V q.
Similarly,
κ2pθ ¨ pγ, V qq “ ´ sin θκ1pγ, V q ` cos θκ2pγ, V q
and
twpθ ¨ pγ, V qq “ twpγ, V q.
Thus twist rate tw is well-defined even for relatively framed curves, whereas the Darboux cur-
vatures κ1, κ2 are well-defined only up to phase. We define the projective Darboux curvatures
to be the equivalence class of function pairs
rκ1, κ2s :“ tpcos θκ1 ` sin θκ2,´ sin θκ1 ` cos θκ2q | θ P r0, 2πqu. (2.1)
Theorem 2.1.2 (Fundamental Theorem for Relatively Framed Curves). Let r, κ1, κ2, tw :
r0, 2s Ñ R be arbitrary smooth functions with r ą 0. There is a relatively framed space curve
pγ, rV sq : r0, 2s Ñ R3ˆR3 satisfying }γ1} “ r with projective Darboux curvatures rκ1, κ2s and
twist rate tw which is unique up to translations and rigid rotations.
20
Proof. Choosing a representative of rκ1, κ2s, say pκ1, κ2q, we solve the system of ODEs
d
dt
¨
˚
˚
˚
˚
˝
T
V
W
˛
‹
‹
‹
‹
‚
“ r
¨
˚
˚
˚
˚
˝
0 κ1 κ2
´κ1 0 tw
´κ2 ´tw 0
˛
‹
‹
‹
‹
‚
¨
˚
˚
˚
˚
˝
T
V
W
˛
‹
‹
‹
‹
‚
,
then define γ by
γptq “
ż t
0
r`
t˘
T`
t˘
dt.
The ODE solving step is possible by existence of solutions of ODE’s and is unique up to
choice of initial conditions pT p0q, V p0q,W p0qq, hence we have uniqueness up to rotations.
We have also implicitly chosen γp0q “ 0, hence uniqueness up to translations. Finally, a
different choice of representative of rκ1, κ2s would have produced a different framing in the
class rV s, thus the solution is only well-defined as a relatively framed curve.
Moreover, if we considered the r in the theorem up to constant multiplication, i.e.,
r1 „ r2 ô Dx ą 0 s.t. r1ptq “ x ¨ r2ptq @ t,
then the resulting framed curve would be well-defined up to scaling. This conclusion moti-
vates the definition of the moduli space that we will study for the majority of this thesis,
the moduli space of framed loops :
M :“ trelatively framed loopsu{ttranslation, scaling, and rotationu.
See Section 3.2.3 for a description of the geometry of M.
21
2.2 Frechet Spaces and Manifolds
As opposed to the finite-dimensional setting, there are a variety of flavors of model spaces for
infinite-dimensional manifolds with distinct topological properties. The infinite-dimensional
manifolds that we use in this work will be modeled on Frechet spaces. The following sections
outline some of the basic background material on Frechet spaces and manifolds, following
the standard references [24] and [39].
The theory of Frechet spaces is vast and sometimes forbiddingly technical. To keep the
conversation grounded, we will frequently refer back to a pair of relevant (and more down-
to-Earth) classes of examples: path and loop spaces of Rn or a finite-dimensional manifold
M .
2.2.1 Tame Frechet Spaces
Let V denote a topological vector space over R or C. By convention, we will also assume
that the topology on V is Hausdorff.
A seminorm on V is a map V Ñ Rě0 which has all of the properties of a norm except
nonzero vectors may take the value zero (in particular, any norm is also a seminorm). We say
that V is locally convex if there exists a family of seminorms t}¨}kukPK (with K not necessarily
countable) on V such that the topology on V is generated by the family of seminorms. This
means that the topology of V is the coarsest one in which, for any fixed v0 P V , every map
of the form
V Ñ Rě0 : v ÞÑ }v ´ v0}k
is continuous. A base for this topology is given by sets of the form
UH,εpv0q “ tv P V | }v ´ v0}h ă ε @h P Hu,
22
where H runs over all finite subsets of K and ε ą 0. An equivalent characterization is that
a sequence vn P V converges to v P V if and only if
}vn ´ v}k Ñ 0 as nÑ 8 @ k.
A simple exercise shows that our assumption that V is Hausdorff implies that any family
of seminorms generating the topology of V must satisfy the property
}v}k “ 0 @ k P K ñ v “ 0.
We can characterize the metrizability of a locally convex space as follows.
Lemma 2.2.1. The topology of V can be generated by a countable family of seminorms
t} ¨ }kukPN if and only if V is metrizable.
Proof. If the topology of V can be generated by a countable family of seminorms, we define
a metric by the formula
dpv, wq :“ÿ
kPN
2´k}v ´ w}k
1` }v ´ w}k. (2.2)
Conversely, suppose that V is metrizable with metric d. Then we define a family of seminorms
} ¨ }k on V by
}v}k “ inftλ P Rą0 | v P Bp0, λ{kqu,
where Bp0, λ{kq is the d-ball around the origin of radius λ{k.
If V is a Hausdorff, locally convex, metrizable vector space which is complete as a metric
space, then it is called a Frechet space. The space is complete as a metric space if and only
if every Cauchy sequence of points vn P V satisfying
}vi ´ vj}k Ñ 0 as i, j Ñ 8 @ k
23
converges in V .
Example 2.2.1 (Trivial Example: Banach spaces.). Any Banach space is trivially a Frechet
space with countable family of seminorms consisting of only the Banach norm.
Example 2.2.2 (Main Example: The loop and path spaces of Rn.). The space C8pS1,Rnq is
a vector space with pointwise addition and scalar multiplication. The Whitney C8 topology
is generated by the family of seminorms
}γ}k :“kÿ
j“0
maxt}γpjqptq}Rn | t P S1u, k “ 0, 1, 2, . . . ,
where } ¨ }Rn is the Euclidean norm, γpjq denotes the j-th derivative of γ, and γp0q “ γ. Note
that each } ¨ }k is actually a norm, and is in particular the standard norm on the Banach
space CkpS1,Rnq.
We claim that this space is Frechet. To see this, let γn be a Cauchy sequence in
C8pS1,Rnq. Using the filtration
C8pS1,Rnq Ă Ck
pS1, Rnq Ă Ck´1
pS1,Rnq Ă ¨ ¨ ¨ Ă C1
pS1,Rnq Ă C0
pS1,Rnq
and the fact that } ¨ }k is the standard Banach norm, we conclude that γn is Cauchy in
CkpS1,Rnq for all k. Thus γn converges in each CkpS1,Rnq and these limits must agree, so
that γn converges to a smooth function.
We can similarly define a Frechet structure on the path space of Rn, C8pr0, 2s,Rnq. We
use the special notations LRn for C8pS1,Rnq and PRn for C8pr0, 2s,Rnq with this topology.
Example 2.2.3 (Important Example: Fast-falling sequences.). Let pV , } ¨ }Vq be a Banach
space. The collection of all sequences tvnu in V such that the sum
}tvnu}k :“ÿ
n
ekn}vn}V
24
converges for all k forms a Frechet space with seminorms } ¨ }k. This is known as the space
of fast-falling sequences in V—see [24], Section II, Example 1.1.2. We will denote the space
of fast falling sequences ΣpVq. Note that the seminorms of ΣpVq have the property that
}tvnu}1 ď }tvnu}2 ď }tvnu}3 ď ¨ ¨ ¨
for all tvnu P ΣpVq.
Remark 2.2.2. We will later define a tame Frechet space to be a Frechet space which
is “close enough” (in a precise sense) to a space of fast-falling sequences for some Banach
space. The category of tame Frechet spaces is convenient in that it admits an inverse function
theorem. The heuristic reason is that spaces of fast-falling sequences are close enough to
Banach spaces that they admit an inverse function theorem, and tame Frechet spaces come
with built-in estimates which allow the inversion result to translate.
A graded Frechet space is a Frechet space V together with a choice of countable family
of seminorms t} ¨ }ku satisfying
}v}1 ď }v}2 ď }v}3 ď ¨ ¨ ¨
for all v P V .
Example 2.2.4. The loop space LRn with the family of seminorms of 2.2.2 is graded.
We note that any Frechet space admits a grading. Indeed, for any Frechet space, choose
an arbitrary countable family of seminorms t} ¨ }ku generating its topology. We then define
a new family of seminorms t} ¨ }1ku by the formula
} ¨ }1k “
kÿ
j“1
} ¨ }j.
25
This new family of seminorms is graded and generates the same topology as the original
family.
A linear map L between graded Frechet spaces pV , t} ¨ }kuq and pV 1, t} ¨ }1kuq is called tame
if there exists r P N and a sequence of real numbers ck P R such that
}Lv}1k ď ck}v}k`r
for all v P V and all sufficiently large k. We say that a graded Frechet space V is a tame direct
summand of another graded Frechet space V 1 if there exist tame linear maps L : V Ñ V 1
and L1 : V 1 Ñ V satisfying L1 ˝ L “ IdV .
Example 2.2.5. The linear map L : LRn Ñ LRn defined by Lpγq “ γphq is tame for any h.
It satisfies
}Lpγq}k ď }γ}k`h
for all k.
A graded Frechet space V is said to be tame if there exists a Banach space W such
that V is a tame direct summand of ΣpWq. Tame Frechet spaces form a category of infinite-
dimensional topological vector spaces which are simultaneously flexible enough to treat many
interesting spaces of smooth maps and well-behaved enough to satisfy an inverse function
theorem, which we state in Section 2.2.4.
Example 2.2.6. Let M be a compact (hence finite-dimensional) manifold and E a vector
bundle over M . Consider the space of smooth sections ΓpM,Eq. Choosing a Riemannian
metric gM on M and a metric gE and connection ∇ on E, this space admits a family of
seminorms t} ¨ }ku defined by
}σ}k :“kÿ
j“1
suppPM
›
›
›
`
∇jσ˘
p
›
›
›
gEbjgM,
26
where ∇j denotes the j-th covariant derivative,
gE bj gM “ gE b gM b ¨ ¨ ¨ b gM
is the induced metric on the vector bundle E bj TM defined on pure tensors by
gE bj gMpσ1 bX11 b ¨ ¨ ¨ bX
j1 , σ2 bX
12 b ¨ ¨ ¨X
j2q
“ gEpσ1, σ2qgMpX1
1 , X12 q ¨ ¨ ¨ g
MpXj
1 , Xj2q
and extended linearly, and } ¨ }gEbjgM is the induced norm. With this family of seminorms,
ΓpM,Eq is a tame Frechet space—see [24], Section II, Corollary 1.3.9 for details.
2.2.2 Calculus in Frechet Spaces and Frechet Manifolds
Unlike the Banach space setting, the total derivative of a map between Frechet spaces is not
well-defined. Going back to a more basic concept, we can define the derivative of a path into
a Frechet space γ : RÑ V in the usual way:
γ1ptq :“ limεÑ0
γpt` εq ´ γptq
ε,
provided the limit exists. We can similarly define higher-order derivatives, and we define a
path to be smooth if all of its derivatives exist. We now define a map P : O Ă V Ñ V 1 from
an open subset of a Frechet space into another Frechet space to be smooth if for every path
γ : RÑ O, the composition P ˝ γ : RÑ V 1 is smooth.
Remark 2.2.3. In the case that V and V 1 are finite-dimensional, this definition of smooth-
ness agrees with the usual one. Surprisingly, this fact is nontrivial and was not proved until
1967 [5].
27
Although total derivatives are not defined for maps between Frechet spaces, it still makes
sense to define directional derivatives. For a smooth map P : O Ă V Ñ V 1 and a pair of
vectors v1 P O, v2 P V we define the derivative of P at v1 in the v2-direction in the expected
way:
DP pv1qpv2q :“d
dε
ˇ
ˇ
ˇ
ˇ
ε“0
P ˝ γpεq P V 1,
where γpεq is the path in V defined by v1` εv2. This is sometimes referred to as the Gateaux
derivative.
A (not necessarily linear) map P from an open subset U of a graded Frechet space V into
another graded Frechet space V 1 satisfies a tame estimate on U if
}P pvq}1k ď ckp1` }v}k`rq,
for some r P N and some sequence of real numbers ck for all sufficiently large k and for all
v P U . A map is said to be tame if it satisfies a tame estimate near each point.
Example 2.2.7. Any continuous map P from Rn with its usual norm into a graded Frechet
space V is tame. Indeed, let x0 P Rn and pick a compact set X containing a neighborhood
of x0. Since P is continuous, }P p¨q}k is a continuous map X Ñ R for all k. Thus there exist
ck such that
}P pxq}k ď ck ď ckp1` }x}q
for all x in X.
We define a Frechet manifold to be a Hausdorff topological space equipped with an open
cover by charts such that each chart is homeomorphic to an open subset of a Frechet space
and such that all transition maps (defined in the usual way) are smooth. The Frechet spaces
in the definition are called model spaces for the manifold. A tame Frechet manifold is a
28
Frechet manifold whose model spaces are tame and whose transition maps are both smooth
and tame.
Example 2.2.8 ([24], Chapter II, Corollary 2.3.2). Every space of smooth maps from a
compact manifold into another finite-dimensional manifold is a tame Frechet manifold.
Let F be a Frechet manifold with charts Pα : Oα Ă F Ñ Vα. A submanifold of F is a
subset G such that for each Oα with Oα X G ‰ H, the model space Vα splits as
Vα “ V1 ‘ V2
and Pα|OαXG has image in V1 ‘ t0u.
Let F be a Frechet manifold and let f P F . The tangent space to F at f is defined in
the same way as the finite-dimensional case. That is, the tangent space is the collection of
derivatives
d
dε
ˇ
ˇ
ˇ
ˇ
ε“0
γpεq,
where γpεq is a path in F satisfying γp0q “ f . The derivative is calculated in local coordi-
nates, and is well-defined by smooth compatibility of the charts of F . Each tangent space
to F of course forms a vector space, denoted TfF , and the collection of all tangent spaces
forms a vector bundle, denoted TF .
Finally, a map P : F Ñ G between Frechet manifolds is called smooth if it is smooth in
local coordinates. Smooth maps induce linear maps on tangent spaces called derivatives as
follows. For f P F , let δf P TfF (we will frequently adopt variational notation for tangent
vectors). Then
δf “d
dε
ˇ
ˇ
ˇ
ˇ
ε“0
γpεq
29
for some path γ. Then we define the directional derivative DP pfqpδfq by
DP pfqpδfq :“d
dε
ˇ
ˇ
ˇ
ˇ
ε“0
P ˝ γpεq P TP pfqG.
2.2.3 Extended Examples: Path and Loop Spaces
The main Frechet spaces and manifolds which appear in this thesis are path and loop spaces
or submanifolds thereof. Thus we take extra care to explore these examples and to set
notation.
Recall that for a finite-dimensional vector space V, we define
PV :“ C8pr0, 2s,Vq, and
LV :“ C8pS1,Vq
and that these are tame Frechet spaces.
Similarly, for a finite-dimensional manifold M , we define the path space of M and loop
space of M to be the spaces
PM :“ C8pr0, 2s,Mq, and
LM :“ C8pS1,Mq,
respectively. These spaces are tame Frechet manifolds. Indeed, this follows from the much
more general theorem of Example 2.2.8:
Theorem 2.2.4 ([24], Section II, Corollary 2.3.2). Let M and M 1 be smooth finite-dimensional
manifolds with M compact. Then the space of smooth maps C8pM,M 1q is a tame Frechet
manifold.
30
We will give explicit manifold charts for the specific case of path spaces, as the ideas in
the proof will be useful later on. The proof sketched here follows [70], and we skip technical
details.
Proposition 2.2.5. The path space PM is a Frechet manifold for any finite-dimensional
manifold M . Moreover, if M is n-dimensional then PM is locally modelled on its tangent
spaces, which take the form
TγPM “ Γpr0, 2s, γ˚TMq,
where γ P PM and γ˚TM Ñ r0, 2s is the pullback bundle. Each tangent space is isomorphic
to PRdimpMq.
Similarly, the loop space LM is a Frechet manifold, locally modelled on its tangent spaces
TγLM “ ΓpS1, γ˚TMq « LRdimpMq.
Proof sketch. We prove the statement regarding path spaces; the loop space proof is essen-
tially the same. We define a local addition on M to be a smooth map ν : TM Ñ M such
that
(i) The composition of ν with the zero section is the identity on M , and
(ii) there exists an open neighborhood Udiag of the diagonal in M ˆM such that the map
π ˆ ν : TM Ñ M ˆM is a diffeomorphism onto Udiag, where π : TM Ñ M is the
natural projection.
We will denote the restriction of ν to the fiber over p PM by νp : TpM ÑM . A local addition
is essentially a generalization of the exponential mapping on a Riemannian manifold. It is a
fact that any M admits a local addition (see [39], Section 42.4).
We will use a fixed local addition ν : TM Ñ M to define charts pPγ,Oγq on PM . Let
Udiag Ă M ˆM be the image of the diffeomorphism from the definition of local addition.
31
For fixed γ P PM , define Oγ Ă PM by
Oγ :“ tβ P PM | pγ, βq P PUdiagu.
Let γ˚TM Ñ r0, 2s denote the pullback bundle of TM by γ : r0, 2s Ñ M . We define
Pγ : Γpr0, 2s, γ˚TMq Ñ Oγ as follows: For α P Γpr0, 2s, γ˚TMq, define
Pγpαqptq :“ νγptqpαptqq.
Choosing a global trivialization of the vector bundle α˚TM Ñ r0, 2s, we have Γpr0, 2s, α˚TMq «
PRdimpMq, and we have an atlas of charts on PM .
Next we need to show that the transition charts for this atlas are smooth. Let β, γ P PM ,
and consider the map
τβγ :“ P´1β ˝ Pγ : P´1
γ pOβ XOγq Ă Γpr0, 2s, γ˚TMq Ñ P´1β pOβ XOγq Ă Γpr0, 2s, β˚TMq.
Let α P P´1γ pOβ XOγq. Then
τβγpαqptq “ ν´1βptq ˝ νγptqpαptqq.
Let αpuq denote a smooth path in Γpr0, 2s, γ˚TMq with parameter u. Then
τβγpαpuqqptq “ ν´1βptq ˝ νγptqpαpuqptqq,
which can be considered as a smooth map from R ˆ r0, 2s into M . Smoothness in the
u-parameter means that our transition functions are smooth, by definition.
32
Our next task is to explore the functorial properties of P and L. For example, given a
smooth map of finite-dimensional manifolds f : M Ñ M 1, we obtain new smooth maps of
Frechet manifolds denoted
Pf : PM Ñ PM 1 and
Lf : LM Ñ LM 1.
The path space version is defined by the formula
Pfpαqptq :“ fpαptqq
and the loop space version is defined similarly. These new maps behave well with respect to
differentiation in the case that M is an open subset of a vector space.
Lemma 2.2.6. Let f : M Ñ M 1 be a smooth map between (finite-dimensional) manifolds,
where M is an open submanifold of a vector space. Let α P PM and δα P TαPM . Then the
derivative of Pf at α in the direction δα satisfies
DPfpαqpδαqptq “ Dfpαptqqpδαptqq.
A similar formula holds for Lf : LM Ñ LM 1.
Proof. This is a simple calculation using the vector space structure of M :
DPfpαqpδαqptq “ d
dε
ˇ
ˇ
ˇ
ˇ
ε“0
fpαptq ` εδαptqq
“ Dfpαptqqpδαptqq.
33
These functors also respect principal bundle structures. We first note that for a smooth
finite-dimensional Lie group G, we obtain new infinite-dimensional Lie groups PG and LG,
where group composition is applied pointwise. Then we have the following proposition.
Proposition 2.2.7 (See [70], Section 4.3). Let G ãÑ F Ñ M be a principal fiber bundle.
Then applying the loop space functor to these maps produces a new principal fiber bundle
LG ãÑ LF Ñ LM . A similar statement holds for the path space functor.
We close this subsection by noting that our identification of S1 with r0, 2s{p0 „ 2q has a
special relevance. This is predicated upon our choice of defining paths to have domain r0, 2s,
which is merely a convenient normalization for the results of this thesis. The relevance is
that it will be convenient to think of the loop space LM as the embedded submanifold of the
path space PM consisting of paths which happen to smoothly close. We record this obvious
but useful fact below.
Proposition 2.2.8. The inclusion map LM ãÑ PM is a smooth embedding of Frechet
manifolds.
2.2.4 The Nash-Moser Inverse Function Theorem
We are now prepared to state the Nash-Moser Inverse Function Theorem. This is a hard
theorem whose proof is well beyond the scope of this thesis. Hamilton’s treatise [24] is
devoted to this theorem and its many applications, while a condensed version of the proof
can be found [39], Chapter 10, Section 51.
Theorem 2.2.9 (Nash-Moser Inverse Function Theorem, [24], Section III, Theorem 1.1.1).
Let V and V 1 be tame Frechet spaces and P : V Ñ V 1 a smooth tame map. Let O Ă V be an
open set such that for each v P O, the derivative
DP pvq : V Ñ V 1
34
has a unique inverse, denoted
pDP q´1pvq : V 1 Ñ V .
If
pDP q´1 : O ˆ V 1 Ñ V
is a smooth tame map, then P is locally invertible and its inverse is smooth and tame.
We will frequently make use of a simpler version of the theorem, which assumes that
the range of the map is finite-dimensional. This simpler version more closely resembles the
classical inverse function theorem which holds in the Banach space category.
Theorem 2.2.10 ([24], Section III, Theorem 2.3.1). Let V be a tame Frechet space and
P : O Ă V Ñ V be a smooth map of an open subset of V into a finite-dimensional vector
space. If the derivative DP pv0q is surjective for some v0 P O, then the level set
F “ tv P O | P pvq “ P pv0qu
is a smooth tame submanifold in a neighborhood of f0 with tangent space
TvF “ kernelpDP pvqq.
2.3 Preliminaries from Differential Geometry
In this section we define symplectic and Riemannian structures for Frechet manifolds. Note
that these definitions are more-or-less formal, and we do not claim that any general the-
orems from finite-dimensional differential geometry automatically apply (in fact, we will
point out some finite-dimensional theorems which are known to fail in infinite-dimensions).
35
Nevertheless, these geometric structures frequently provide intuition for specific cases of
infinite-dimensional manifolds and we will find them quite useful throughout the thesis.
2.3.1 Symplectic and Riemannian Structures
We define a weak symplectic structure on a Frechet space V to be a 2-form ω on V satisfying:
1. ω is closed. This means dω “ 0, where d is the exterior derivative defined by the
formula
dωvpδv1, δv2, δv3q “
Dω‚pδv2, δv3qpfqpδv1q ´Dω‚pδv1, δv3qpfqpδv2q `Dω‚pδv1, δv2qpfqpδv3q
for any triple of vectors δvj P TvV « V . In the formula we are treating the δvj as
constant vector fields near v. The expression ω‚pδv2, δv3q (for example) is considered
as a map from V to R and
Dω‚pδv2, δv3qpvqpδv1q
as the derivative of this map at v in the δv1 direction.
2. ω is nondegenerate. That is, ω induces an injection from TvV into T ˚v V for each v P V
via the formula
δv ÞÑ ωvpδv, ¨q. (2.3)
This definition also works for a Frechet submanifold F of a Frechet space V , where
tangent spaces TfF are naturally identified with linear subspaces of TfV « V . The definition
similarly extends to the base space of certain submersions with total space a submanifold
of a Frechet space and with finite-dimensional fiber. These are the only cases of symplectic
Frechet manifolds which appear in this thesis, so we will not worry about extending the
36
definition to more general situations where technicalities involving the existence of a Lie
bracket arise. For a more general treatment, see [39], Section 48.
In the case that V is finite-dimensional, this agrees with the standard definition of a
symplectic structure. Moreover, the injection defined by (2.3) will be an isomorphism in
this case. A symplectic structure on a general Frechet manifold is called strong if (2.3) is an
isomorphism—this situation will not come up again in this thesis.
An almost complex structure on a Frechet manifold F is a smooth choice of involution
J : TfF Ñ TfF , J2“ ´Id.
Such an involution is called a complex structure if it arises via holomorphic charts on F
taking values in a complex Frechet space. In this case the almost complex structure is also
frequently referred to as integrable. We note that in finite-dimensions the integrability of an
almost complex structure J is equivalent to the vanishing of a certain tensor defined in terms
of J , called the Nijenhuis tensor. The definition of the Nijenhuis tensor extends to infinite-
dimensional manifolds, but a vanishing Nijenhuis tensor no longer implies the integrability
of the almost complex structure (see [43]).
A Riemannian metric on F is a smoothly-varying choice g of inner product on each
tangent space
gf p¨, ¨q : TfF ˆ TfF Ñ R.
The Riemannian metrics used in this thesis will always be weak in the sense that they induce
injections
TfF Ñ T ˚f F : δf ÞÑ gf pδf, ¨q,
but we do not requre them to induce bijections (in the case that the induced map is a
bijection, the metric is called strong). The definition of a Levi-Civita connection for a
37
metric g extends without alteration to the infinite-dimensional setting, although we make
no claim of existence or uniqueness for such a connection. We define a path γ in F to be a
geodesic with respect to g if
∇γ1ptqγ1ptq “ 0
for all t, where ∇ denotes a Levi-Civita connection for g. We once again do not claim
the existence or uniquenss of geodesics (such a claim can frequently be translated into a
statement about the existence and uniqueness of a solution to a PDE).
Finally, we define a Kahler structure on F to be a triple pg, ω, Jq, where g is a weak
Riemannian metric, ω is a weak sympectic structure, and J is a complex structure such that
the three structures are related by the formula
gp¨, ¨q “ ωp¨, J ¨q.
2.3.2 Marsden-Weinstein Reduction
In this section we describe the process of reducing a finite-dimensional symplectic manifold
by the Hamitonian action of a Lie group, introduced by Marsden and Weinstein [48]. The
definition given here is somewhat nonstandard, but it is easy to see that it is equivalent to
the usual definition in the finite-dimensional case. The reason for the departure is that it
will allow us to avoid working with the dual to a Frechet vector space later on in the thesis.
LetG be a finite-dimensional Lie group that freely and properly acts on a finite-dimensional
symplectic manifold pM,ωq. Let g denote the Lie algebra of G. For each ξ P g, there is an
associated vector field Xξ on M describing the infinitesimal action of ξ. Let 〈¨, ¨〉g be any
choice of inner product on g. A moment map for the action of G is a smooth map
µ : M Ñ g,
38
which satisfies
Dp 〈µppq, ξ〉g pY |pq “ ωppY |p, Xξ|pq (2.4)
for all p P Σ, ξ P g and vector fields Y on Σ. The left side of the equation should be
understood as the derivative of the function
p ÞÑ 〈µppq, ξ〉g (2.5)
at the point p in the direction Y |p.
Remark 2.3.1. In the standard definition of a moment map, µ is g˚-valued, and the unnat-
ural introduction of an auxilliary inner product on g is avoided. As previously mentioned,
our goal is to avoid working wth dual spaces. The Lie groups that we will work with come
endowed with preferred metrics so that this definition will be natural enough in practice.
In the case that a moment map exists, it is a fact that for any regular value ξ P g, µ´1pξq
is a smooth submanifold of Σ such that the restriction of ω degenerates along exactly the
G-orbits. Therefore one obtains a new manifold, called the symplectic reduction of Σ by G
and denoted
Σ �G :“ µ´1pξq{G,
which has a canonical symplectic form. If pM,ωq is a Kahler manifold and G acts by
isometries with respect to the Riemannian structure and preserves the complex structure,
then the symplectic reduction is also a Kahler manifold and we refer to it as the Kahler
reduction of M by G ([54], Chapter 8).
The Kahler reduction construction extends without issue to the case that the original
manifold and the Lie group are infinite dimensional. However, the general finite-dimensional
theorems (or general theorems in the Banach space category as in, e.g., [73]) no longer apply,
and it must be checked by hand that the resulting space is indeed a Kahler manifold.
39
2.4 Quaternions
Throughout the thesis, it will be convenient for calculations to identify C2 with the algebra
of quaternions. The quaternions are elements of the real vector space
H :“ spanRt1, i, j,ku,
which is endowed with an algebra structure by extending the relations
i2 “ j2“ k2
“ ijk “ ´1
linearly. Elements of the quaternions will typically be denoted by q “ Q0`Q1i`Q2j`Q3k.
In this notation, the real part of q is
Repqq “ Q0 P R
and the imaginary part of q is
Impqq “ Q1i`Q2j`Q3k.
We will quite frequently identify the purely imaginary quaternions with R3 so that Impqq P
R3. The quaternionic conjugate of q P H is the quaternion
q “ Q0 ´Q1i´Q2j´Q3k.
40
We identify C2 with H via the map
C2Ø H
pz, wq Ø z ` wj (2.6)
where we also identify the complex i with the quaternionic i.
Let 〈¨, ¨〉C2 denote the standard Hermitian inner product on C2 and let } ¨ }C2 denote
the corresponding norm. From the identification (2.6), this endows H with a Hermitian
inner product and norm, denoted 〈¨, ¨〉H and } ¨ }H, respectively. The real part of 〈¨, ¨〉C2 is a
Euclidean inner product on C2, and by extension Re 〈¨, ¨〉H is a Euclidean inner product on
H. A simple calculation shows that the Euclidean product on H is given by
Re 〈p, q〉H “ Reppqq.
However, we should be careful to note that in general 〈p, q〉H ‰ pq (the latter is not even
complex-valued).
The identification (2.6) restricts to an identification of the unit 3-sphere S3 Ă C2 with
the unit quaternions
tq P H | }q}H “ 1u.
The product of a pair of unit quanterions is still a unit quaternion, thus this endows S3 with
a Lie group structure. We endow S3 with its standard, constant curvature-1 Riemannian
metric by restricting the real part of the Hermitian inner product on C2. We denote this
metric gS3
and note that it is left-invariant with respect to the Lie group structure of S3.
We also endow the unit quaternions with the Riemannian metric obtained by restricting the
real part of 〈¨, ¨〉H. By the above discussion, it is clear that, with respect to these metrics,
S3 and the unit quaternions are isometric.
41
We conclude this section by identifying S3 and the unit quaternions with a third well-
known Lie group, the special unitary group SUp2q. Elements of SUp2q are of the form
¨
˚
˝
z w
´w z
˛
‹
‚
,
where }pz, wq}C2 “ 1. The mapping
S3Ñ SUp2q
pz, wq ÞÑ
¨
˚
˝
z w
´w z
˛
‹
‚
(2.7)
is a diffeomorphism. Moreover, we claim that it is an isometry. That is, the standard
left-invariant metric on the Lie group SUp2q is
gSUp2qpξ1, ξ2q :“
1
2tracepξ1ξ
˚2 q, for ξ1, ξ2 P TASUp2q,
where TASUp2q is isomorphic to the Lie algebra of skew-Hermitian 2 ˆ 2 complex matrices
sup2q via the identification
ξ P TASUp2q Ø A´1ξ P sup2q.
Then we have the following.
Lemma 2.4.1. The map (2.7) is an isometry with respect to gS3
and gSUp2q.
Proof. For the remainder of this proof, we denote the map (2.7) by f . Since each metric is
left-invariant and f maps p1, 0q P S3 to the idenity matrix in SUp2q, it suffices to prove the
claim at the identity element. The first step is to compute the derivative of f at p1, 0q P S3
42
in the direction of pa, bq P T1S3:
Dfpp1, 0qqpa, bq “d
dε
ˇ
ˇ
ˇ
ˇ
ε“0
fp1` εa, εbq
“d
dε
ˇ
ˇ
ˇ
ˇ
ε“0
¨
˚
˝
1` εa εb
´εb 1` εa
˛
‹
‚
“
¨
˚
˝
a b
´b a
˛
‹
‚
.
Next we compute the pullback of gSUp2q by f :
f˚gSUp2q1 ppa1, b1q, pa2, b2qq “ g
SUp2qId pDfpp1, 0qqpa1, b1q, Dfpp1, 0qqpa2, b2qq
“ gSUp2q
¨
˚
˝
¨
˚
˝
a1 b1
´b1 a1
˛
‹
‚
,
¨
˚
˝
a2 b2
´b2 a2
˛
‹
‚
˛
‹
‚
“1
2trace
¨
˚
˝
¨
˚
˝
a1 b1
´b1 a1
˛
‹
‚
¨
˚
˝
a2 b2
´b2 a2
˛
‹
‚
˚˛
‹
‚
“1
2
`
a1a2 ` b1b2 ` b1b2 ` a1a2
˘
“ Re`
a1a2 ` b1b2
˘
“ gS3
ppa1, b1q, pa2, b2qq.
43
Chapter 3
Complex Coordinates for Framed
Paths and Loops
This chapter describes the complex coordinate system for framed paths and loops which
will be used throughout the rest of the thesis. We will first describe the moduli spaces of
framed paths and loops of interest. These moduli spaces are endowed with Riemannian
metrics, which are natural from a variety of perspectives. Finally, we will establish our first
main results, which isometrically parameterize moduli spaces of framed paths and loops by
infinite-dimensional versions of classical Riemannian manifolds.
3.1 Framed Path Space
As described in Section 1.1.2, our goal is to develop a coordinate system for spaces of sim-
ilarity classes of framed curves in analogy with the classical approach of using curvature
κ and torsion τ as a global coordinate system for the space of arclength parameterized,
nonvanishing curvature curves in R3 up to ambient Euclidean isometry. In Section 3.1.3 we
describe how to associate a framed path to a pair of smooth complex-valued functions, which
44
we denote
pφ, ψq : r0, 2s Ñ C2zt0u.
Such a pair will also frequently be denoted Φ “ pφ, ψq. In contrast with pκ, τq-coordinates,
the conditions on pφ, ψq that ensure periodicity of the resulting framed path are easy to
check, and this makes the submanifold of framed loops quite convenient to work with. We
note that the construction of a framed curve from pφ, ψq given here is not new—see, e.g.,
[22, 27]. The novel contributions of this thesis are the interpretation of pφ, ψq as global
coordinates on the moduli space of framed paths and the resulting description of the space
of framed loops as a submanifold.
In this section, we introduce the space of framed paths. We represent a framed path
pγ, V q as an element of PpSOp3q ˆ R`q via the map
pγ, V q ÞÑ ppT, V, T ˆ V q, }γ1}q , where T :“γ1
}γ1}. (3.1)
Up to translation, this map is a bijection with inverse
ppU, V,W q, rq ÞÑ
ˆ
t ÞÑ
ż t
0
rU dt, V
˙
, (3.2)
and we identify
rS :“ tpγ, V q | γp0q “ 0u
« tframed pathsu{ttranslationu
« PpSOp3q ˆ R`q. (3.3)
From the discussion in Section 2.2.3, we conclude that rS naturally has the structure of a
tame Frechet manifold. We claim that it also admits a natural Riemannian metric, which
45
we describe in the following section. Since we wish to define a Riemannian metric, it will be
useful to have an explicit description of the tangent spaces of rS. Throughout the thesis we
denote tangent vectors to an infinite-dimensional manifold using variational notation; e.g.,
a tangent vector to pγ, V q will be denoted pδγ, δV q.
Lemma 3.1.1. The tangent space at pγ, V q P rS is the space of vectors
pδγ, δV q P C8pr0, 2s,R3ˆ R3
q
satisfying the constraints
δγp0q “ 0,⟨d
dtδγ, V
⟩`
⟨d
dtγ, δV
⟩“ 0, and
〈δV, V 〉 “ 0.
Proof. We first note that we can consider elements of rS as paths
pγ, V q : r0, 2s Ñ R3ˆ R3
satisfying the constraint equations
γp0q “ 0, (3.4)
〈γ1, V 〉 “ 0, and (3.5)
}V }2 “ 1. (3.6)
46
A tangent vector at pγ, V q is thus an equivalence class of derivatives of the form
pδγ, δV q “d
dε
ˇ
ˇ
ˇ
ˇ
ε“0
pγε, Vεq,
where pγε, Vεq is a one-parameter family of paths r0, 2s Ñ R3 ˆ R3 satisfying constraint
equations (3.4)-(3.6), as well as pγ0, V0q “ pγ, V q. From (3.4), we conclude
δγp0q “d
dε
ˇ
ˇ
ˇ
ˇ
ε“0
γεp0q “ 0.
From (3.5), we obtain
d
dε
ˇ
ˇ
ˇ
ˇ
ε“0
〈γ1ε, Vε〉 “ 0,
which implies
〈δγ1, V 〉` 〈γ1, δV 〉 “⟨d
dε
ˇ
ˇ
ˇ
ˇ
ε“0
γ1ε, V0
⟩`
⟨γ10,
d
dε
ˇ
ˇ
ˇ
ˇ
ε“0
Vε
⟩“ 0.
Similarly, the third constraint defining rS implies the third constraint defining the tangent
space.
The framed path space rS admits an R`-action by scaling of the base curve: for x P R`
and pγ, V q P rS, the action is defined by
x ¨ pγ, V q :“ pxγ, V q (3.7)
We are particularly interested in the quotient space
S :“ tframed pathsu{ttranslation, scalingu,
47
which we refer to as the moduli space of framed paths. This space is of interest because
it is more relevant for shape recognition applications and has especially nice geometry. In
particular, we will see in Section 3.1.4 that it is isometric to an L2 sphere. The R`-action
admits a global slice chart, and we will identify S with the concrete model space of framed
paths of length 2,
S «!
pγ, V q P rS | lengthpγq “ 2)
.
Using this identification, we show that S has the structure of a manifold.
Proposition 3.1.2. The moduli space of framed paths S is a tame Frechet submanifold of
rS with tangent spaces of the form
Tpγ,V qS “"
pδγ, δV q P Tpγ,V q rS |ż 2
0
〈δγ1, T 〉 dt “ 0
*
, (3.8)
where T “ γ1{}γ1}.
Proof. By transfer of structure via the identification (3.3), it suffices to show that the space
"
pA, rq P PpSOp3q ˆ R`q |ż 2
0
r dt “ 2
*
(3.9)
is a tame Frechet manifold. As previously noted, PpSOp3qˆR`q is a tame Frechet manifold.
The set (3.9) is the inverse image of 2 of the map
` : PpSOp3q ˆ R`q Ñ R
pA, rq ÞÑ
ż 2
0
rptq dt.
48
This map is smooth and has 2 as a regular value. Indeed, taking a variation pδA, δrq of pA, rq
where δA is arbitrary and δr is constantly 1, we have
D`pA, rqpδA, δrq “d
dε
ˇ
ˇ
ˇ
ˇ
0
`pA` εδA, r ` εδrq “d
dε
ˇ
ˇ
ˇ
ˇ
0
ż 2
0
prptq ` εδrq dt “
ż 2
0
δr dt “ 2 ‰ 0.
Thus S is a submanifold by Corollary 2.2.10. Moreover, in pγ, V q-coordinates δr “ 〈δγ1, T 〉,
whence
D`pγ, V qpδγ, δV q “
ż 2
0
〈δγ1, T 〉 dt
and we conclude that the space (3.8) is the kernel of D`pγ, V q.
3.1.1 The Natural Metric on Framed Path Space
In this section we define a Riemannian metric on
rS “ tframed pathsu{ttranslationu
which is invariant under scaling, rotations and reparameterizations. Invariance under these
transformations is essential for the metric to be useful for shape recognition algorithms.
Using the identification (3.3), we first define the metric on PpSOp3q ˆ R`q.
Recall that the standard left-invariant metric on the Lie group SOp3q is
pξ1, ξ2q ÞÑ1
2tracepξ1ξ
T2 q, for ξ1, ξ2 P TASOp3q,
where TASOp3q is isomorphic to the Lie algebra of skew-symmetric 3ˆ 3 matrices sop3q via
the identification
ξ P TASOp3q Ø A´1ξ P sop3q.
We denote this metric by gSOp3q.
49
Similarly, the standard left-invariant metric on R` (a Lie group under multiplication) is
given by
px1, x2q ÞÑx1 ¨ x2
r2, for x1, x2 P TrR`,
where TrR` is isomorphic to the abelian Lie algebra R via the identification
x P TrR` Øx
rP R.
We endow the product Lie group SOp3q ˆ R` with the product metric
gSOp3qˆR`pA,rq ppξ1, x1q, pξ2, x2qq :“
1
2tracepξ1ξ
T2 q `
x1 ¨ x2
r2.
Applying the isomorphisms
TASOp3q « sop3q and TrR` « R
pointwise produces an isomorphism
TpA,rqPpSOp3q ˆ R`q « Ppsop3q ˆ Rq
for each pA, rq P PpSOp3q ˆ R`q. The natural reparametrization-invariant L2-type metric
on PpSOp3q ˆ R`q is given by
gPpSOp3qˆR`qpA,rq ppδA1, δr1q, pδA2, δr2qq :“
ż 2
0
gSOp3qˆR`pAptq,rptqq ppδA1ptq, δr1ptqq, pδA2ptq, δr2ptqqq ds,
where
ds “ dsprq :“ rptqdt
50
corresponds to measure with respect to arclength in pγ, V q-coordinates. Integrating against
this measure is necessary to ensure that the metric is invariant under reparametrizations.
By construction, we also have that gPpSOp3qˆR`q is invariant with respect to the action of R`
defined by
x ¨ pAptq, rptqq :“ pAptq, x ¨ rptqq, x P R`.
In pγ, V q-coordinates, this action corresponds to scaling the base curve γ, as described in
(3.7). Therefore the metric restricts to a well-defined metric on the moduli space of framed
paths
S “ tframed pathsu{ttranslation, scalingu.
To simplify notation, we use the notation
gS :“1
4gPpSOp3qˆR`q
ˇ
ˇ
ˇ
rSor
1
4gPpSOp3qˆR`q
ˇ
ˇ
ˇ
S.
There are two important remarks to be made here:
1. The slight abuse of notation employed here by using the same notation for either
restricted metric has the benefit of avoiding an unwieldy list of specialized notations.
Such abuses will occur frequently throughout the remainder of the thesis.
2. The extra factor of 1{4 in this definition is a normalization in anticipation of cleaning
up notation later on, and has no deep conceptual signifcance.
The metric gS has an interesting form when expressed in pγ, V q-coordinates. It is related
to a well-studied family of metrics called the elastic metrics [57] on the Frechet space
Nplanar :“ tImmersions r0, 2s Ñ R2u{ttranslation, scalingu.
51
This family of metrics ga,b, a, b ě 0 are defined as follows. Let γ be an element of the space
and let δγ1 and δγ2 be tangent vectors. Then ga,b is defined by the formula
ga,bγ pδγ1, δγ2q “1
2`pγq
ż 2
0
a
⟨d
dsδγ1, N
⟩⟨d
dsδγ2, N
⟩` b
⟨d
dsδγ1, T
⟩⟨d
dsδγ2, T
⟩ds,
where d{ds is derivative with respect to arclength, N is the unit normal to γ, T is the unit
tangent to γ, and `pγq is the length of γ. The first term in the integrand compares bending
of the variations and the second term compares stretching. We note that this metric is in
fact defined on the space of immersions r0, 2s Ñ R2, but that it is invariant to translation
and scaling (invariance to scaling is facilitated by the `pγq´1 factor before the integral)
so that it induces a well-defined metric on Nplanar. Moreover, the metrics are invariant to
rotations of R2 and reparameterization—such symmetries are necessary for shape recognition
applications.
This definition cannot be extended to the full space of immersed curves in R3, since
the normal vector to a parameterized space curve γ is not well-defined when γ2 vanishes.
However, the analogues of elastic metrics can be defined for framed space curves. Let pγ, V q
be a framed curve, let T denote the unit tangent to γ, and let W “ T ˆ V . Let pδγj, δVjq,
j “ 1, 2, be tangent vectors to pγ, V q. We define the family of framed curve elastic metrics
ga,b,c,d, a, b, c, d ě 0, on S by the formula
ga,b,c,dpγ,V q ppδγ1, δV1q, pδγ2, δV2qq
“1
2`pγq
ż 2
0
a
⟨d
dsδγ1, V
⟩⟨d
dsδγ2, V
⟩` b
⟨d
dsδγ1,W
⟩⟨d
dsδγ2,W
⟩` c
⟨d
dsδγ1, T
⟩⟨d
dsδγ2, T
⟩` d 〈δV1,W 〉 〈δV2,W 〉 ds.
The first and second terms in ga,b,c,d compare the bending deformation of γ under the vari-
ations δγj, the third term compares the stretching deformation of γ under the δγj and the
52
last term compares the rotational deformation of the framing V under the variations δVj.
As in the planar case, we note that these metrics are actually defined on the space of framed
paths, but that they are invariant to translation and scaling so that they each induce a
well-defined metric on rS and S. Each metric is also invariant under rotations of R3 and
reparameterizations.
Remark 3.1.3. The somewhat unnatural-looking factor of 1{2 in front of ga,b and ga,b,c,d is
a convenient normalization whose purpose will become apparent in Section 3.1.4.
The next proposition shows that the natural metric metric gS from the perspective of the
Lie group embedding is also a particularly nice example of this generalized family of elastic
metrics; namely it is equal to g1,1,1,1. First, we prove a lemma.
Lemma 3.1.4. The formula for the metric g1,1,1,1 restricted to S reduces to
g1,1,1,1pγ,V q ppδγ1, δV1q, pδγ2, δV2qq “
1
4
ż 2
0
⟨d
dsδγ1,
d
dsδγ2
⟩` 〈δV1,W 〉 〈δV2,W 〉 ds.
Proof. First note that by restricting to S, we have `pγq “ 2, so the 1{2`pγq factor in front of
the integral sign reduces to a constant factor of 1{4. Then note that, since pT ptq, V ptq,W ptqq
forms an orthonormal frame for all t, we can express the arclength derivative of any curve
variation δγ as
d
dsδγ “
⟨d
dsδγ, T
⟩T `
⟨d
dsδγ, V
⟩V `
⟨d
dsδγ,W
⟩W.
The result then follows by an elementary computation using the orthonormality of pT, V,W q
and the normalization a “ b “ c “ d “ 1.
Proposition 3.1.5. In pγ, V q-coordinates, the natural metric gS restricted to S is equal to
the framed curve elastic metric g1,1,1,1.
53
Proof. For the remainder of the proof, let f denote the identification (3.2); i.e.,
f : PpSOp3q ˆ R`q Ñ rS
ppU, V,W q, rq ÞÑ
ˆ
t ÞÑ
ż t
0
rU dt, V
˙
.
Our goal is to compute the pullback of g1,1,1,1 by f .
Let A “ pU, V,W q P PSOp3q and r P PR`. By the discussion above, tangent vectors to
PpSOp3q ˆ R`q at pA, rq take the form pAξ, xq, where ξ P Psop3q and x P PR. It will be
useful to explicitly write ξ as the skew-symmetric matrix
ξ “
¨
˚
˚
˚
˚
˝
0 ´ξ3 ξ2
ξ3 0 ´ξ1
´ξ2 ξ1 0
˛
‹
‹
‹
‹
‚
,
where ξj P PR. Then
Aξ “ pξ3V ´ ξ2W,´ξ3U ` ξ1W, ξ2U ´ ξ1V q.
Treating PpSOp3q ˆ R`q as an embedded subset of PppR3q3 ˆ R`q in the obvious way,
we are abe to compute the derivative of f as
DfpA, rqpAξ, xq “d
dε
ˇ
ˇ
ˇ
ˇ
ε“0
fpA` εAξ, r ` εxq
“d
dε
ˇ
ˇ
ˇ
ˇ
ε“0
ˆ
t ÞÑ
ż t
0
pr ` εxqpU ` εpξ3V ´ ξ2W qq dt, V ` εp´ξ3U ` ξ1W q
˙
“
ˆ
t ÞÑ
ż t
0
xU ` rpξ3V ´ ξ2W q dt,´ξ3U ` ξ1W
˙
.
54
If we denote the last line by pδγ, δV q and the image fpA, rq by pγ, V q, then we have
d
dsδγ “
1
}γ1}
d
dt
ˆ
t ÞÑ
ż t
0
xU ` rpξ3V ´ ξ2W q dt
˙
“1
rpxU ` rpξ3V ´ ξ2W qq
“x
rU ` ξ3V ´ ξ2W.
We can now compute f˚g1,1,1,1. Let pA, rq P PpSOp3q ˆ R`q and denote the image of
pA, rq under f by pγ, V q and let pAξj, xjq, j “ 1, 2, denote elements of TpA,rqPpSOp3q ˆ R`q
with images pδγj, δVjq under DfpA, rq. The entries of ξj will be denoted ξkj , k “ 1, 2, 3.
Applying Lemma 3.1.4 and the previous calculation, we obtain
pf˚g1,1,1,1qpA,rqppAξ1, x1q, pAξ2, x2qq “ g1,1,1,1
pγ,V q ppδγ1, δV1q, pδγ2, δV2qq
“1
4
ż 2
0
⟨d
dsδγ1,
d
dsδγ2
⟩` 〈δV1,W 〉 〈δV2,W 〉 ds
“1
4
ż 2
0
⟨x1
rU ` ξ3
1V ´ ξ11W,
x2
rU ` ξ3
2V ´ ξ22W⟩
`⟨´ξ3
1U ` ξ11W,W
⟩ ⟨´ξ3
2U ` ξ12W,W
⟩ds
“1
4
ż 2
0
x1x2
r2` ξ3
1ξ32 ` ξ
21ξ
22 ` ξ
11ξ
12 ds
“1
4
ż 2
0
1
2tracepξ1ξ
T2 q `
x1x2
r2ds
“ gSpA,rqppAξ1, x1q, pAξ2, x2qq.
3.1.2 The Hopf Map
To obtain complex coordinates on framed path space which were promised in Section 3.1,
we recall the definition of the frame-Hopf map FrameHopf : C2 Ñ R3ˆ3 (e.g., [27]), defined
55
by the formula
FrameHopfpz, wq :“
¨
˚
˚
˚
˚
˝
|z|2 ´ |w|2 2Impzwq ´2Repzwq
2Impzwq Repz2 ` w2q Impz2 ` w2q
2Repzwq Imp´z2 ` w2q Repz2 ´ w2q
˛
‹
‹
‹
‹
‚
.
The first part of this section outlines some of the important properties of the frame-Hopf
map.
The frame-Hopf map has a natural representation when written in quaternionic coordi-
nates. Under the identification C2 « H described in Section 2.4, the frame-Hopf map takes
the form
FrameHopfpqq “ pqiq, qjq, qkqq P timaginary quaternionsu3 « pR3q3.
Another important fact about the frame-Hopf map is that the first column of the frame-
Hopf map gives an explicit form of the Hopf fibration S1 ãÑ S3 Ñ S2 when restricted to
S3 Ă C2 (actually, each column gives a form of the Hopf fibration, but we only treat the first
column for the sake of simplicity). This fact and can be deduced from the following pair of
lemmas. Each lemma can be verified by an elementary computation, which we omit.
Lemma 3.1.6. The first column of the frame-Hopf map, FrameHopf1, is invariant under
the diagonal Up1q-action on C2 by multiplication; that is, for all eiθ P Up1q,
FrameHopf1peiθz, eiθwq “ FrameHopf1pz, wq.
56
Lemma 3.1.7. The columns of the frame-Hopf map square norms. That is, for any pz, wq P
C2,›
›FrameHopfjpz, wq›
› “ }pz, wq}2C2 ,
where FrameHopfj, j “ 1, 2, 3,, denotes the j-th column of the frame Hopf map.
Thus the image of the restriction of FrameHopf1 to S3 is S2 and the map has circle fibers,
thereby establishing our earlier claim.
Recall from Section 2.4 that S3 inherits a Lie group structure via identification with the
Lie group of unit quaternions. Also recall that S3 is naturally identified with the Lie group
SUp2q. The following lemma is well-known (e.g., [21] Section I.1.4) and very useful. It may
be verified by an elementary computation.
Lemma 3.1.8. The restriction of the frame-Hopf to S3 Ă C2 gives a Lie group homomor-
phism onto SOp3q Ă R3ˆ3 which satisfies
FrameHopfpz, wq “ FrameHopfpz1, w1q ô pz1, w1q “ ˘pz, wq.
Thus the frame-Hopf map induces a homomorphic double covering SUp2q Ñ SOp3q.
Moreover, the frame-Hopf map has the following local (almost) isometry property.
Lemma 3.1.9. The restriction of the frame Hopf map
FrameHopf : S3Ñ SOp3q
satisfies
FrameHopf˚gSOp3q“ 4gS
3
.
57
Thus the induced frame-Hopf homomorphism SUp2q Ñ SOp3q is also a local isometry up to
a factor of 4.
Proof. We work in quaternionic coordinates. For the remainder of the proof we use the
notation
Impqq :“ pq1, q2, q3q P R3
for q “ Q0 `Q1i`Q2j`Q3k P H.
Since FrameHopf is known to be a homomorphism and the metrics involved are left-
invariant, it suffices to prove the claim at 1 P S3 Ă H. The derivative of FrameHopf at 1 in
the direction q is given by
DFrameHopfp1qpqq “ 2pImpiqq, Impjqq, Impkqqq.
Note that in quaternionic coordinates, the tangent space of S3 at 1 is the space
T1S3“ tq P H | Re 〈1, q〉H “ 0u “ tq “ Q0 `Q1i`Q2j`Q3k P H | Repqq “ Q0 “ 0u.
58
For p, q P T1S3 the pullback metric is given by
FrameHopf˚gSOp3q1 pq, pq “ g
SOp3qId pDFrameHopfp1qpqq,DFrameHopfp1qppqq
“ gSOp3qId p2pImpiqq, Impjqq, Impkqqq, 2pImpipq, Impjpq, Impkpqqq
“ 4 ¨1
2trace
¨
˚
˚
˚
˚
˝
¨
˚
˚
˚
˚
˝
0 Q3 ´Q2
´Q3 0 Q1
Q2 ´Q1 0
˛
‹
‹
‹
‹
‚
¨
˚
˚
˚
˚
˝
0 P3 ´P2
´P3 0 P1
P2 ´P1 0
˛
‹
‹
‹
‹
‚
T˛
‹
‹
‹
‹
‚
“ 4 ¨1
2¨ 2pQ1P1 `Q2P2 `Q3P3q
“ 4Repqpq
“ 4gS3
1 pp, qq.
Thus the first claim holds. The second follows immediately from Lemma 2.4.1, which says
that SUp2q and S3 are isometric with respect to their standard metrics.
3.1.3 Complex Coordinates for Framed Paths
In this section we finally describe our complex coordinate system for the moduli space of
framed paths. It follows from Lemma 3.1.8 that the frame-Hopf map induces a smooth
double-cover
H : C2zt0u Ñ SOp3q ˆ R`
defined by
Hpz, wq :“
ˆ
1
|z|2 ` |w|2FrameHopfpz, wq, |z|2 ` |w|2
˙
,
where Hpz, wq “ Hpz1, w1q if and only if pz, wq “ ˘pz1, w1q. Indeed,
H “ H2 ˝ H1,
59
where H1 is the diffeomorphism
H1 : C2zt0u Ñ S3
ˆ R`
pz, wq ÞÑ
˜
1a
|z|2 ` |w|2pz, wq,
a
|z|2 ` |w|2
¸
,
and H2 is the smooth double cover
H2 : S3ˆ R` Ñ SOp3q ˆ R`
ppz, wq, rq ÞÑ pFrameHopfpz, wq, r2q.
It follows from Proposition 2.2.7 that applying the path functor to H produces a smooth
double-cover
H : PpC2zt0uq Ñ PpSOp3q ˆ R`q,
which we still denote H by an abuse of notation. From the above, we see that Φ1 “
pφ1, ψ1q, Φ2 “ pφ2, ψ2q P PpC2zt0uq correspond to the same framed path if and only if
Φ1 ” ˘Φ2. By composing H with the identification (3.3) of PpSOp3q ˆ R`q with rS, we
obtain a smooth double cover of framed path space, denoted
pH : PpC2zt0uq Ñ rS “ tframed pathsu{ttranslationu.
Thus a pair pφ, ψq of complex valued functions determines a unique framed path, up to trans-
lation. In the following section, we show that we restrict pH to the L2-sphere in PpC2zt0uq,
then the image is the moduli space of framed paths S. Moreover, we show that this double
covering is an isometry.
60
3.1.4 The L2-Sphere
We endow PC2 with the L2 Hermitian inner product
〈Φ1,Φ2〉L2 :“
ż 2
0
〈Φ1ptq,Φ2ptq〉C2 dt,
where 〈¨, ¨〉C2 denotes the standard Hermitian inner product on C2. The real part of the L2
Hermitian inner product determines a real-valued inner product on PC2. Since this real part
will frequently be of central interest, we introduce the notation
gL2
:“ Re 〈¨, ¨〉L2 ,
which we call the L2 Riemannian metric. We refer to it as a metric since it will soon be
restricted to various submanifolds of PC2, where it will induce a weak Riemannian metric.
We will also use gL2
to denote the L2 Riemannian metric on PH
gL2
‚ pq, pq :“ Re
ż 2
0
qp dt.
These metrics agree under the identification C2 « H described in Section 2.4, so no confusion
should arise from the conflation of notations.
Our first goal is to show that the map pH defined in the previous section is an isometry.
Proposition 3.1.10. The map
pH : PpC2zt0uq Ñ rS
is a local isometry with respect to gL2
and gS “ g1,1,1,1.
61
Proof. It suffices to show that the map
H : PpC2zt0uq Ñ PpSOp3q ˆ R`q
is a local isometry, as the result then follows immediately from Proposition 3.1.5.
Throughout the proof, it will be convenient for calculations to use quaternionic notation.
We write H “ PH2 ˝ PH1, where
PH1 : PpHzt0uq Ñ PpS3ˆ R`q
q ÞÑ
ˆ
1
}q}Hq, }q}H
˙
,
and
PH2 : PpS3ˆ R`q Ñ PpSOp3q ˆ R`q
pq, rq ÞÑ pPFrameHopfpqq, r2q.
are path space versions of the maps defined in the previous section, written in quaternionic
coordinates.
Our first goal is to compute the pullback PH˚2gS . Lemma 2.2.6 implies
DPFrameHopfpqqpδqqptq “ DFrameHopfpqptqqpδqptqq
for any q P PS3 Ă PH and any δq P TqPS3. Similarly, if s : R` Ñ R` denotes the squaring
map r ÞÑ r2, then
DPsprqpδrqptq “ 2rptqδrptq.
62
Thus the variation of PH2 at pq, rq in the direction pδq, δrq is given by the formula
DPH2pq, rqpδq, δrqptq “ pDFrameHopfpqptqqpδqptqq, 2rptqδrptqq .
We will for the moment shorten our notation FrameHopf to F in order to avoid an unwieldy
calculation of the pullback metric:
`
PH˚2gS˘
pq,rqppδq1, δr1q, pδq2, δr2qq
“1
4
ż 2
0
gSOp3qFpqptqq pDFpqptqqpδq1ptqq, DFpqptqqpδq2ptqqq `
4rptq2δr1ptqδr2ptq
prptq2q2ds
“1
4
ż 2
0
ˆ
4gS3
qptq pδq1ptq, δq2ptqq `4δr1ptqδr2ptq
rptq2
˙
rptq2dt (3.10)
“
ż 2
0
Repδq1δq2q ¨ r2` δr1δr2 dt,
where (3.10) follows by Lemma 3.1.9 and the definition of ds.
Next we use Lemma 2.2.6 once again to compute the variation of PH1 at q in the direction
δq to be
DH1pqqpδqq “
ˆ
}q}δq ´ qRepqδqq{}q}
}q}2,Repqδqq
}q}
˙
.
63
In the above and for the remainder of this proof, we will denote the quaternionic norm by
} ¨ } in order to simplify notation. We conclude the proof by calculating the pullback metric:
`
H˚gS˘
qpδq1, δq2q
“`
H˚1`
H˚2gS˘˘
qpδq1, δq2q
“`
H˚2gS˘
pq{}q},}q}qpDH1pqqpδq1q, DH1pqqpδq2qq
“
ż 2
0
Re
ˆ
}q}δq1 ´ qRepqδq1q{}q}
}q}2¨}q}δq2 ´ qRepqδq2q{}q}
}q}2
˙
}q}2
`Repqδq1q
}q}¨
Repqδq2q
}q}dt
“
ż 2
0
Repδq1δq2q `Repqδq1qRepqδq2q
}q}2
´ Re
ˆ
δq1qRepqδq2q ` δq2qRepqδq2q
}q}2
˙
`Repqδq1qRepqδq2q
}q}2dt
“
ż 2
0
Re`
δq1δq2
˘
dt
“ gL2
q pδq1, δq2q.
Going through the calculations in the proof of Proposition 3.1.10, we obtain the following
useful corollary. This will later be used in Section 4.2 to describe the framed curve geometric
invariants of Section 2.1.2 in our complex coordinate system.
64
Corollary 3.1.11. Let Φ P PpC2zt0uq correspond to q P PpHzt0uq under the identification
C2zH and to the framed path pγ, V q under pH. Let the variations δΦj, j “ 1, 2, of Φ correspond
to the variations δqj of q and pδγj, δVjq of pγ, V q. Then we have the following equality as
functions:
1
4
ˆ⟨d
dsδγ1,
d
dsδγ2
⟩` 〈δV1,W 〉 〈δV2,W 〉
˙
}γ1} “ Re 〈δΦ1, δΦ2〉C2 “ Re 〈δq1, δq2〉C2 .
Now we arrive at our first main result, which says that the moduli space of framed paths
S is isometrically double covered by an open subset of an L2 sphere. We use the notation
SpPC2q :“
!
Φ P PpC2zt0uq | }Φ}L2 “
?2)
,
for the radius-?
2 L2 sphere of PC2. We are particularly interested in a dense open subset,
denoted
S˝pPC2q :“ tΦ P SpPC2
q | Φptq ‰ 0 @ tu.
Theorem 3.1.12. The map pH restricts to an isometric double cover
pH|S˝pPC2q : S˝pPC2q Ñ S
with respect to the metrics gL2
and gS .
Proof. Recall that the moduli space S is concretely realized as
S :“ tpγ, V q P rS | lengthpγq “ 2u.
65
By Proposition 3.1.10, we only need to show that
pH´1pSq “ S˝pPC2
q Ă PC2.
Indeed, let pγ, V q P S and Φ P PpC2zt0uq such that pHpΦq “ pγ, V q. Then
2 “
ż 2
0
}γ1} dt “
ż 2
0
}FrameHopf1pΦq} dt “
ż 2
0
}Φ}2C2 dt (3.11)
“ }Φ}2L2 ,
where the last equality in (3.11) follows from Lemma 3.1.7.
This theorem can be phrased differently as follows.
Corollary 3.1.13. The moduli space of framed paths S is isometric to
S˝pPC2q{pΦ „ ´Φq,
an open subset of the projective space of real lines in PC2 « PR4 with its induced Riemannian
metric.
3.2 Framed Loop Space
As stated in Section 3.1, the benefit of the complex coordinates pφ, ψq for framed paths is
that the necessary and sufficient conditions for the periodicity of the resulting framed path
are surprisingly nice. Before describing the conditions, we require a new definition: a path
66
-1.0 -0.5 0.5 1.0
-1.0
-0.5
0.5
Figure 3.1: An element of AC.
Φ P PC2 is called smoothly antiperiodic if
dk
dtk
ˇ
ˇ
ˇ
ˇ
t“2
Φ “ ´dk
dtk
ˇ
ˇ
ˇ
ˇ
t“0
Φ for all k “ 0, 1, 2, . . .
(see Figure 3.1). The space of antiperiodic paths in C2, or the antiloop space of C2, is denoted
AC2.The antiloop space is a complex tame Frechet vector space. We define ApC2zt0uq and
AC similarly.
Lemma 3.2.1. A path Φ “ pφ, ψq P PpC2zt0uq corresponds to a framed loop under pH if and
only if
(1) The path Φ is either smoothly closed or smoothly antiperiodic, and
(2) the maps φ and ψ have the same L2 norm and are L2-orthogonal:
ż 2
0
|φ|2 dt “
ż 2
0
|ψ|2 dt and
ż 2
0
φψ dt “ 0.
Proof. Let Φ “ pφ, ψq be either a loop or an antiloop. It follows easily from the definition
of H that this is a necessary and sufficient condition for HpΦq to be smoothly periodic; i.e.,
67
that
ppT, V,W q, rq :“ HpΦq P LpSOp3q ˆ R`q.
Then pHpΦq is a smoothly periodic framed loop if and only if
ż 2
0
rT dt “ 0. (3.12)
The key point here is that T and r are smooth loops, so the only obstruction to the curve as-
sociated to pT, rq being smoothly closed is that its endpoints match up—hence the vanishing
integral condition. Equation (3.12) is written in pφ, ψq-coordinates as
ż 2
0
|φ|2 ´ |ψ|2 dt “
ż 2
0
2Impφψq dt “
ż 2
0
2Repφψq dt “ 0,
which are exacly the conditions given in (2).
Since LpC2zt0uq and ApC2zt0uq are disjoint, the lemma implies that the space
tframed loopsu{ttranslationu
is disconnected. Indeed, the space of framed loops has two components, which are indexed
by mod-2 self-linking number. This fact is well-known, but a proof is seemingly absent from
the literature. The details of the proof are given in the next section.
3.2.1 Connected Components of Framed Loop Space
We use the classical concepts of writhe, twist, and linking number. Basic definitions are
given here for the convenience of the reader and to set notation (see, e.g., [14] for more
details).
68
+
-‐
-‐
+
-‐
-‐+-‐
+ -‐
Figure 3.2: A knot diagram with Wr “ 3´ 5 “ ´2
Recall that the writhe of an oriented knot diagram is a signed count of its crossings—see
Figure 3.2. We define the writhe of a (parameterized) knot γ, denoted Wrpγq, to be the
average writhe of all knot diagrams for γ which are obtained by projecting to a plane.
The total twist Twpγ, V q of pγ, V q is defined by the formula
Twpγ, V q :“1
2π
ż 2
0
twpγ, V q ds
“1
2π
ż 2
0
⟨d
dsV, T ˆ V
⟩ds,
where
d
ds“
1
}γ1}, T “
γ1
}γ1}, ds :“ }γ1}dt
and twpγ, V q is twist rate, as defined in Section 2.1.2.
Finally, we define the self-linking number of a framed knot pγ, V q, denoted Lkpγ, V q, to
be the topological linking number of the link formed by the disjoint knots γ and γ ` εV ,
where ε ą 0 is sufficiently small.
69
Note that Wr and Lk are only defined in the case that γ is embedded, while Tw makes
sense for general pγ, V q. In fact, one can check using the explicit formula for Tw that it is a
continuous functional on framed loop space—this observation will be useful in a moment.
We will employ the Calugareanu-White-Fuller Theorem [20], which says that for a framed
loop pγ, V q with embedded base curve,
Lkpγ, V q “ Twpγ, V q `Wrpγq.
We use the notation
Lk2pγ, V q :“ Lkpγ, V q mod 2
for the mod-2 linking number of pγ, V q. We will show that this quantity is well-defined even
for framed loops with nonembedded base curves.
The following lemma is well well known, but we include the proof for completeness.
Lemma 3.2.2. Let pγ0, V0q and pγ1, V1q be a pair of homotopic framed loops such that γ0
and γ1 are embedded loops. Then Lk2pγ0, V0q “ Lk2pγ1, V1q.
Proof. By standard genericity arguments and induction, it suffices to prove the lemma in
the case that there is a homotopy pγu, Vuq of framed loops such that γu is embedded for all
u ‰ 1{2, and that γ1{2 has a single transverse self-intersection.
By the isotopy invariance of Lk, the continuity of Tw, and the Calugareanu-White-Fuller
Theorem, a discontinuity in Lk can only occur due to a jump in Wr at u “ 1{2. By taking
ε sufficently small, we ensure that for almost every projection direction, the projection of
γ1{2´ε differs from the projection of γ1{2`ε by switching exactly one crossing. This has the
effect of changing Wr by ˘2, and we conclude that Lk changes by ˘2 at exactly u “ 1{2,
i.e.
Lk2pγ0, V0q “ Lk2pγ1, V1q.
70
Remark 3.2.3. It follows immediately from the lemma that Lk2pγ, V q determines the ho-
motopy class of pT, V,W q in SOp3q, where T “ γ1{}γ1} and W “ T ˆ V . Computing any
example shows that
Lk2pγ, V q “ 0 ô pT, V,W q generates π1pSOp3qq.
Thus the space of framed loops has at least two connected components. We now prove the
claim stated above: the space of framed loops has exactly two components. It is convenient
to identify
TS2zS2
0 « SOp3q ˆ R`
pW, ξq Ø
ˆˆ
ξ
}ξ},W,
ξ
}ξ}ˆW
˙
, }ξ}
˙
,
where S20 is the image of the zero section in TS2. Then the map (3.1), restricted to loops,
becomes
tframed loopsu Ñ LpTS2zS2
0q
pγ, V q ÞÑ pV, γ1q.
Note that the map is not onto—a general loop pW, ξq P LpTS2zS20q does not satisfy
ş2
0ξ dt “ 0,
hence does not correspond to a framed loop. The fact that a general homotopy in TS2zS20
can always be adjusted to lie in the image of this map is the content of the proposition:
Proposition 3.2.4. Let pγj, Vjq for j “ 0, 1 be framed loops such that there exists a ho-
motopy pWu, ξuq between their images pVj, γ1jq in LpTS2zS2
0q. Then pγ0, V0q and pγ1, V1q are
homotopic.
71
This proposition has the flavor of an h-principle [17, 23] (in particular, it is reminiscent of
the Smale-Hirsch theorem), but we are unaware of any general result that implies it. In any
case, we give a simple proof based on the classical proof of the Whitney-Graustein theorem
[75].
Proof. Without loss of generality, we assume that γjp0q “ 0 and }γ1j} ” 1 for j “ 0, 1.
Let pWu, ξuq be a homotopy of loops in TS2zS20 such that pWj, ξjq “ pVj, γ
1jq for j “ 0, 1.
Since TS2zS20 is homotopy equivalent to the unit tangent bundle of S2, we may assume that
}ξu} ” 1 for all u.
We construct a homotopy of framed loops pγu, Vuq by first defining
γuptq :“
ż t
0
ξuptq dt´t
2
ż 2
0
ξuptq dt.
Then γup0q “ γup2q “ 0 for all u and
γ1uptq “ ξuptq ´1
2
ż 2
0
ξuptq dt.
Assuming for now that ξu is nonconstant for all u, we conclude that γ1uptq ‰ 0 for all u, t.
This follows since the average value of a nonconstant spherical loop lies in the interior of the
sphere.
Now we note that Wuptq and γ1uptq are linearly indpendent for all u, t. Indeed,
Wu ˆ γ1u “ Wu ˆ ξu ´W ˆ
1
2
ż 2
0
ξuptq dt,
which is nonzero since the first term is a unit vector and the second has strictly less than
unit length (once again assuming that ξu is nonconstant for all u). Thus we can define Vu
72
by
Vu :“Wu ´ 〈Wu, γ
1u〉
γ1u}γ1u}
›
›
›Wu ´ 〈Wu, γ1u〉
γ1u}γ1u}
›
›
›
.
Then Vu is a normal vector field along the loop γu and we have produced a homotopy pγu, Vuq
between our original framed loops.
Finally, we note that ξu may be assumed to be nonconstant for all u without loss of
generality, perhaps after a small perturbation of our initial homotopy relative to its endpoints.
Corollary 3.2.5. Framed loop space has two connected components. The component in
which a framed curve pγ, V q with embedded base curve γ lies is determined by its mod-2
linking number. It follows that the definition of Lk2 extends to nonembedded framed loops.
Proof. The first statement follows from the proposition, as
LpTS2zS2
0q « LpSOp3q ˆ R`q
has two components—one for each element of
π1pSOp3q ˆ R`q « π1pSOp3qq « Z2.
The second statement follows immediately from Lemma 3.2.2. The third statement follows
by defining the mod-2 linking number of a nonembedded framed loop pγ, V q to be the mod-2
linking number of any embedded framed loop which is homotopic to pγ, V q.
Example 3.2.1. This definition of Lk2 is sometimes counterintuitive. For example, let
γ0ptq “1
2πp0,´ sinp2πtq, cosp2πtq ´ 1q, V0ptq “ p1, 0, 0q.
73
The image of γ0 is a doubly-covered circle, and any pushoff in the V0-direction is disjoint
from γ0 — see Fig. 3.3(a). Thus, for this particular nonembedded framed curve, the usual
self-linking number is well-defined; it is Lkpγ0, V0q “ 0.
On the other hand, pγ0, V0q is homotopic to pγ1, V1q, where
γ1ptq “1
πp0, cospπtq ´ 1, sinpπtqq, V1ptq “ p´ sinpπtq, cos2
pπtq, cospπtq sinpπtqq,
which has as its image an embedded framed curve with Lkpγ1, V1q “ 1 — see Fig. 3.3(c).
An explicit homotopy is given by
¨
˚
˚
˚
˚
˚
˚
˚
˚
˚
˚
˚
˚
˝
¨
˚
˚
˚
˚
˚
˚
˚
˚
˚
˚
˚
˚
˝
pu´ 1qu cospπtq
πp1´ u` u2q
cospπtqpu` pu´ 1q sinpπtqq
π?
1´ u` u2
sinpπtqpu` pu´ 1q sinpπtqq
π?
1´ u` u2
˛
‹
‹
‹
‹
‹
‹
‹
‹
‹
‹
‹
‹
‚
,
¨
˚
˚
˚
˚
˚
˚
˚
˚
˚
˚
˚
˚
˝
´´1` u` u sinpπtq?
1´ 2u` 2u2
upu` u cosp2πtq ` 2pu´ 1q sinpπtqq
2´ 4u` 4u2
u cospπtqp1´ u` u sinpπtqq
1´ 2u` 2u2
˛
‹
‹
‹
‹
‹
‹
‹
‹
‹
‹
‹
‹
‚
˛
‹
‹
‹
‹
‹
‹
‹
‹
‹
‹
‹
‹
‚
.
We conclude that Lk2pγ0, V0q “ 1 according to our definition, even though Lkpγ0, V0q “ 0.
3.2.2 The Stiefel Manifolds
For V “ LC or AC, we denote the Stiefel manifold of L2-orthonormal 2-frames in V by
St2pVq :“
"
pφ, ψq P V2|
ż 2
0
|φ|2dt “
ż 2
0
|ψ|2dt “ 1 and
ż 2
0
φψ dt “ 0
*
.
74
(a) (b) (c)
Figure 3.3: Links determined by (a) pγ0, V0q, (b) A perturbation of pγ0, V0q along the homotopy,and (c) pγ1, V1q. In each image, the blue loop is the image of γ and the red loop is the image ofγ ` εV .
This space is indeed a manifold—it is the level set of a regular value of the smooth map
LC2Ñ R4
pφ, ψq ÞÑ
ˆż 2
0
|φ|2 dt,
ż 2
0
|ψ|2 dt,Re
ż 2
0
φψ dt, Im
ż 2
0
φψ dt
˙
, (3.13)
hence it is a tame Frechet manifold by Corollary 2.2.10.
Recall that we denote the L2 Hermitian inner product on PC2 by 〈¨, ¨〉L2 . Likewise, we
denote the L2 Hermitian inner product on PC by
〈φ, ψ〉L2 :“
ż 2
0
φψ dt,
and the induced norm denoted } ¨ }L2 . Moreover, the restrictions of these inner products and
norms to LC, AC, LC2 and AC2 will still be denoted 〈¨, ¨〉L2 and } ¨ }L2 . The meaning of the
notation should always be clear from context.
75
We are interested in a particular open subset of the Stiefel manifold, called the stable
Stiefel manifold of V and denoted
St˝2pVq :“ tpφ, ψq P St2pVq | pφptq, ψptqq ‰ p0, 0q @ tu.
The use of the word stable refers to the fact that this is the stable submanifold (in the sense
of GIT quotients [73]) with respect to the action of the Lie group LS1 given by poinwise
multiplication. This fact is further explored in Section 4.3 below.
A simple but important fact is that the Stiefel manifolds are diffeomorphic via the map
St2pLCq Ñ St2pACq
pφptq, ψptqq ÞÑ expp´iπt{2qpφptq, ψptqq. (3.14)
This diffeomorphism restricts to a diffeomorphism of the stable Stiefel manifolds.
It is clear that the Stiefel manifold St2pVq is connected. Indeed, let Φ0 “ pφ0, ψ0q and
Φ1 “ pφ1, ψ1q be elements of St2pVq and assume for now that
spantφ0, ψ0, φ1, ψ1u “:W
is 4-dimensional. Any isometry of C4 withW (with respect to the standard Hermitian inner
product on C4 and the restriction of the L2 Hermitian inner product to W) induces an
embedding of the standard complex Stiefel manifold
St2pC4q “ tpz, wq P pC4
q2| z and w are Hermitian orthonormalu
into St2pVq. Moreover, we can consider Φ0 and Φ1 as elements of the embedded St2pC4q.
Thus we can find a homotopy Φu between Φ0 and Φ1 which lies in the connected submanifold
76
St2pC4q, as this classical manifold is well-known to be connected. The same argument works
if dimpWq “ 2 or 3. A less obvious fact is that the stable Stiefel manifolds are connected.
Lemma 3.2.6. The stable Stiefel manifold St˝2pVq is connected.
Proof. In light of the diffeomorphism (3.14), it suffices to prove the claim for V “ LC .Let
Φ0,Φ1 P St˝2pLCq. In particular, Φ0,Φ1 P St2pLCq Ă LC2. By the above discussion, there
exists a homotopy rΦu in St2pLCq which joins Φ0 and Φ1, but such a homotopy may not stay
in StS2 pLCq. To remedy this, we consider rΦu as a homotopy of loops in C2. We perturb rΦu,
relative to Φ0 and Φ1, to obtain a new homotopy pΦu “ ppφu, pψuq which satisfies
(i) pΦuptq ‰ 0 for all t, u and
(ii) pφu and pψu are linearly independent as elements of LC for all u,
where we appeal to the fact that a small perturbation will preserve the stable condition that
the coordinates of pΦu are linearly independent.
Now we can orthonormalize the coordinate functions of pΦu with respect to 〈¨, ¨〉L2 :
φu :“pφu
}pφu}L2
, ψu :“
pψu ´⟨pψu, φu
⟩L2φu
›
›
›
pψu ´⟨pψu, φu
⟩L2φu
›
›
›
L2
.
The orthonormalization process does not introduce new simultaneous zeroes for φu and ψu,
thus the new homotopy Φu :“ pφu, ψuq lies in StS2 pLCq for all u and it still joins Φ0 and
Φ1.
Next we note that St˝2pVq is a subset of
S˝pPC2q “ tΦ P PC2
| }Φ}L2 “?
2 and Φptq ‰ 0 @ tu.
By Theorem 3.1.12, S˝pPC2q covers the space S of framed paths of length 2 via the restriction
of pH. From Lemma 3.2.1 we conclude that the orthonormal pairs pφ, ψq P LpC2zt0uq Y
77
ApC2zt0uq (i.e., pairs which lie in a Stiefel manifold) correspond exactly (under a further
restriction of pH to framed loops of length 2. By considering only framed loops of length 2,
we have chosen a horizontal slice with respect to the R` action on framed loop space by
scaling of the base curve.
Finally, we note that St˝2pVq Ă V2 Ă PC2, so that the Stiefel manifold has a natural
metric obtained by restricting the real part of the L2 Hermitian inner product—we will
continue to denote this restricted Riemannian metric by gL2. The restricted map
pH : S˝pPC2q Ñ S
was shown to be a local isometry with respect to gL2
and gS , thus its restriction to the Stiefel
manifolds is still a local isometry. Putting all of this together, we have:
Proposition 3.2.7. The restricted map
pH : St˝2pLCq \ St˝2pACq Ñ tframed loopsu{ttranslation, scalingu
is a smooth double covering and a local isometry with respect to the restricted metrics gL2
and gS .
The Lk2 “ 1 component of framed loop space is covered by St˝2pLCq and the Lk2 “ 0
component is covered by St˝2pACq.
Proof. Combining the preceding discussion with Corollary 3.2.5 and Lemma 3.2.6, the only
thing we need to check is that an element of St˝2pLCq maps under pH to a framed loop with
Lk2 “ 1. Indeed, from
Φptq “1?
2pcospπtq ´ i sinpπtq, 1q P St˝2pLCq
78
we obtain pHpΦq “ pγ1, V1q, where
γ1ptq “1
πp0, cospπtq ´ 1, sinpπtqq, V1ptq “ p´ sinpπtq, cos2
pπtq, cospπtq sinpπtqq.
From Example 3.2.1, we see that Lk2pγ1, V1q “ 1.
3.2.3 The Grassmann Manifolds
From the Stiefel manifold St2pVq, V “ LC or AC, we obtain a Grassmannian in the usual
way. We define
Gr2pVq :“ St2pVq{Up2q,
where A P Up2q acts on St2pVq by pointwise multiplication from the right:
ppφ, ψq ¨ Aq ptq :“ pφptq, ψptqq ¨ A for A P Up2q.
The Up2q-orbit of a point Φ “ pφ, ψq P St2pVq is the complex span of tφ, ψu and will be
denoted rΦs “ rφ, ψs. Thus the Grassmannian Gr2pVq is the space of 2-dimensional complex
subspaces of V . It is a compex manifold and we sketch the construction of holomorphic
charts for Gr2pVq in Proposition 4.1.3.
We define the stable Grassmannian of V to be the open submanifold
Gr˝2pVq :“ St˝2pVq{Up2q Ă Gr2pVq.
Recall from Lemma 3.1.8 that FrameHopf restricts to a homomorphism
FrameHopf : SUp2q Ñ SOp3q.
79
From this fact and the definition of H, we deduce that
Hppφ, ψq ¨ Aq “ Hpφ, ψq ¨ FrameHopfpAq for A P SUp2q, pφ, ψq P St˝2pVq. (3.15)
Therefore modding out by the action of the SUp2q term of the product
SUp2q ˆ Up1q “ Up2q
has the effect of modding out by the action of SOp3q by rotations on framed loop space. The
Up1q term of Up2q acts on framed loop space by global frame twisting—i.e. twisting each
frame vector V ptq around γ by the same angle. Explicitly, let eiθ P Up1q, let Hpφ, ψq “ pγ, V q
and let Hppφ, ψq ¨ eiθq “ pγ, V q. Then γ “ γ and
V “ cosp2θqV ` sinp2θqpT ˆ V q.
Recall from Section 2.1.2 that an equivalence class of a framed loop with respect to this
Up1q-action is called a relatively framed loop.
Thus the map pH induces a map on Gr˝2pLCq\Gr˝2pACq with image the moduli spaceM,
where we recall that
M “ trelatively framed loopsu{ttranslation, scaling, rotationu.
The induced map will still be denoted pH, and by further abuse of notation we will denote
pHprΦsq by pHpΦq.
The Grassmannians inherit natural Riemannian metrics from the Stiefel manifolds. This
follows immediately from checking that the restricted metric gL2
on St2pVq is invariant with
80
respect to the Up2q-action. Moreover, the restricted metric gS on
tframed loopsu{ttranslation, scalingu Ă S
is invariant with respect to rotations. Thus gS induces a well-defined Riemannian metric on
M. By abuse of notation, we will continue to denote these induced metrics by gL2
and gS ,
respectively. By the equivariance property (3.15) of H and by Proposition 3.2.7, we conclude
that the induced map pH on the Grassmannians is a local isometry with respect to gL2
and
gS .
Finally, we note that since Lk2 is invariant with respect to translation, scaling, and
rotation,M also has two connected components, indexed by Lk2. Thus we have proved our
second main result for this chapter.
Theorem 3.2.8. The induced map
pH : Gr˝2pLCq \Gr˝2pACq ÑM
is an isometry with respect to the induced metrics gL2
and gS . The Lk2 “ 1 component M
is isometric to Gr˝2pLCq and the Lk2 “ 0 component is isometric to Gr˝2pACq.
Proof. This follows from the discussion above. We should point out that this map is an
honest isometry, as opposed to the local isometry of Proposition 3.2.7. This follows since
Φ,´Φ P St2pVq differ by right multiplication by the negative 2ˆ2 identity matrix, an element
of Up2q. Thus rΦs “ r´Φs as elements of Gr2pVq.
81
Chapter 4
The Geometry of Spaces of Framed
Paths and Loops
The Grassmannian Gr2pVq (for V “ LC or AC) has a variety of natural geometric structures
which are inherited by the moduli space of framed loops M. In particular, Gr2pVq is a
complex Kahler manifold. It is complex in the strong sense that it admits holomorphic
charts (recall that there are weaker notions of complex manifolds in the Frechet category
[43]). The goal of this chapter is to describe the geometric structures on Gr2pVq and the
corresponding structures on M.
4.1 The Symplectic Structure of Framed Loop Space
4.1.1 The Kahler Structure of the Grassmannian
In this section, we describe the natural symplectic structure of the Grassmannians and the
induced symplectic structure of the moduli space of framed loopsM. We begin by describing
the tangent spaces of the Stiefel manifolds.
82
Lemma 4.1.1. The tangent space to St2pVq at Φ “ pφ, ψq is given by the following codimension-
4 subspace of V2:
TΦSt2pVq “"
δΦ “ pδφ, δψq P V2| Re
ż 2
0
φδφ dt “ Re
ż 2
0
ψδψ dt “
ż 2
0
φδψ ` δφψ dt “ 0
*
.
Proof. To derive this description of the tangent space, one simply notes that this vector space
is the kernel of the derivative at Φ of the map (3.13) used to define the manifold structure
of St2pVq, and the description follows by Corollary 2.2.10.
Recall that we denote the Up2q-orbit of Φ “ pφ, ψq P St2pVq by rΦs “ rφ, ψs P Gr2pVq
and that
rΦs “ rφ, ψs “ spanCtφ, ψu Ă V .
From Lemma 4.1.1, we will obtain a description of the tangent spaces to the Grassmannians.
We define the vertical tangent space to St2pVq at Φ to be the space of directions along
Up2q-orbit of Φ; that is,
T vertΦ St2pVq :“ tΦ ¨ ξ | ξ P up2qu.
The horizontal tangent space to St2pVq at Φ is then the orthogonal complement with respect
to the L2-metric
T horΦ St2pVq :“ tδΦ P TΦSt2pVq | gL2
Φ pδΦ, δΨq “ 0 @ δΨ P T vertΦ St2pVqu.
By the construction of Gr2pVq as the image of the submersion
St2pVq Ñ St2pVq{Up2q “ Gr2pVq,
83
the tangent space to Gr2pVq at rΦs can be identified with the horizontal tangent space to
St2pVq at Φ. Moreover, since the vertical space is finite-dimensional, this horizontal space
can be explicitly described.
Corollary 4.1.2. The tangent space to Gr2pVq at rΦs may be identified with the codimension-
4 Up2q-horizontal subspace of TΦSt2pVq given by
TrΦsGr2pVq « T horΦ St2pVq
“
"
δΦ P V2|
ż 2
0
φδφ dt “
ż 2
0
ψδψ dt “
ż 2
0
φδψ dt “
ż 2
0
ψδφ dt “ 0
*
. (4.1)
Note that if Φ1 and Φ2 lie in the same Up2q-orbit, then
T horΦ1St2pVq “ T horΦ2
St2pVq,
so that this description of the tangent spaces of Gr2pVq does not depend on the choice of
representative of rΦs.
Proof. Elements of up2q can be written in the form
A “
¨
˚
˝
ai z
´z bi
˛
‹
‚
,
where a, b P R and z P C. Thus the defining condition of the horizontal subspace can be
expressed as
gL2
pδΦ,Φ ¨ ξq “ gL2
ppδφ, δψq, pφai´ ψz, φz ` ψbiqq “ 0 @ a, b P R and z P C,
84
where δΦ “ pδφ, δψq and Φ “ pφ, ψq. Simplifying, we have
gL2
ppδφ, δψq, pφai´ ψz, φz ` ψbiqq “ Re 〈pδφ, δψq, pφai´ ψz, φz ` ψbiq〉L2
“ Re
ż
δφpφai´ ψzq ` δψpφz ` ψbiq dt,
and we conclude
a ¨ Im
ż
φδφ dt` b ¨ Im
ż
ψδψ dt` Re
ˆ
z
ż
φδψ ` δφψ dt
˙
“ 0
must hold for all a, b P R and z P C. By taking various choices of a, b and z and combining
with conditions which are already assumed on δΦ P TΦSt2pVq, we are able to derive the
conditions in (4.1). For example, by taking a “ 1 and b “ z “ 0 we obtain
Im
ż
φδφ dt “ 0.
Since
Re
ż
φδφ dt “ 0
was already a condition required to be tangent to St2pVq, we conclude
ż
φδφ dt “ 0.
85
The remaining conditions can be derived by other choices of parameters. We leave it to the
reader to check that the following choices work:
b “ 1, a “ z “ 0 ñ
ż
ψδψ dt “ 0
z “ 1, a “ b “ 0 ñ
ż
φδψ dt “ 0
z “ i, a “ b “ 0 ñ
ż
ψδφ dt “ 0.
We note that the tangent spaces to the Grassmannian are complex subspaces of V2, so
each tangent space inherits the complex structure
TrΦsGr2pVq Ñ TrΦsGr2pVq
δΦ ÞÑ iδΦ. (4.2)
Thus Gr2pVq has an almost complex structure, and we show it is in fact an honest complex
structure on Gr2pVq, as described in Section 2.3.1. To do so, we explicitly construct holomor-
phic charts on Gr2pVq. This construction uses the usual charts for a real finite-dimensional
Grassmannian, adapted to the complex infinite-dimensional setting (e.g., [56], Section 5).
We outline the proof here for the convenience of the reader.
Proposition 4.1.3. The Grassmannian Gr2pVq is a complex manifold with complex structure
(4.2).
Proof. Fix rΦ0s “ rφ0, ψ0s P Gr2pVq and let rΦ0sK denote the codimension-2 subspace of V
consisting of elements which are orthogonal to the 2-dimensional subspace rΦ0s Ă V with
respect to gL2.
86
We will model Gr2pVq near rΦ0s on the complex vector space TrΦ0sGr2pVq, described in
(4.1). We define a map
frΦ0s : TrΦ0sGr2pVq Ñ Gr2pVq
pδφ, δψq ÞÑ spanCtφ0 ` δφ, ψ0 ` δψu.
The image of frΦ0s is the open subset U Ă Gr2pVq consisting of 2-dimensional complex
subspaces of V which intersect rΦ0sK in exactly 0. Moreover, frΦ0s is a smooth bijection onto
its image with smooth inverse defined on S P U by
f´1rΦ0spSq :“
´
p´1S,rΦ0s
pφ0q ´ φ0, p´1S,rΦ0s
pψ0q ´ ψ0
¯
,
where pS,rΦ0s is orthogonal projection from S to rΦ0s—this is a well-defined isomorphism,
since S and rΦ0s are both 2-dimensional and S P U .
We leave it to the reader to check that transition maps take the form
f´1rΦ1s
˝ frΦ0spδφ, δψq “ p〈φ1, φ0〉L2 pφ0 ` δφq ` 〈φ1, ψ0〉L2 pψ0 ` δψq ´ φ1,
〈ψ1, φ0〉L2 pψ0 ` δφq ` 〈φ1, ψ0〉L2 pψ0 ` δψq ´ ψ1q
and that the derivative of this transition map at pδφ, δψq in the direction´
Ăδφ,Ăδψ¯
is given
by
D´
f´1rΦ1s
˝ frΦ0s
¯
pδφ, δψq´
Ăδφ,Ăδψ¯
“
´
〈φ1, φ0〉L2Ăδφ` 〈φ1, ψ0〉L2
Ăδψ,
〈ψ1, φ0〉L2Ăδφ` 〈ψ1, ψ0〉L2
Ăδψ¯
.
Thus the transition maps are clearly holomorphic.
87
4.1.2 The Grassmannian as a Kahler Reduction
We explained in Section 3.2.3 that the metric gL2
on V2 obtained by taking the real part
of the natural L2 Hermitian metric restricts to a metric on St2pVq and descends to a well-
defined metric on Gr2pVq, each still denoted gL2. By the same reasoning, the entire Hermitian
structure descends to Gr2pVq as well. In particular,
ωL2
:“ ´Im 〈¨, ¨〉L2
is a symplectic form on V2 which we claim descends to a well-defined symplectic form on
Gr2pVq—we will abuse notation and continue to denote the induced form by ωL2. Indeed,
the closure property dωL2“ 0 follows from the closure of ωL
2as a form on the vector space
V2, which is clear since ωL2
doesn’t depend on its basepoint. The nondegeneracy of the
induced form follows easily because Gr2pVq is a complex manifold. In particular, for any
nonzero δΦ P TrΦsGr2pVq, iδΦ P TrΦsGr2pVq, and
ωL2
rΦspδΦ, iδΦq “ ´Im
ż
〈δΦ, iδΦ〉C2 dt “ Re
ż
}δΦ}2C2 dt ‰ 0.
This induced symplectic form can be seen as arising from a symplectic reduction, as
defined in Section 2.3.2. We claim that the map
V2Ñ R4
pφ, ψq ÞÑ
ˆż 2
0
|φ|2 dt,
ż 2
0
|ψ|2 dt,Re
ż 2
0
φψ dt, Im
ż 2
0
φψ dt
˙
“`
}φ}2L2 , }ψ}2L2 ,Re 〈φ, ψ〉L2 , Im 〈φ, ψ〉L2
˘
88
introduced in Section 3.2.2 to show that St2pVq is a Frechet manifold can be viewed as the
moment map for the Up2q-action on V2 by right multiplication. More precisely, we have the
following proposition.
Proposition 4.1.4. The map
µUp2q : V2Ñ up2q
pφ, ψq ÞÑ i
¨
˚
˝
〈φ, φ〉L2 ´ 〈φ, ψ〉L2
´ 〈ψ, φ〉L2 〈ψ, ψ〉L2
˛
‹
‚
is a moment map for the Up2q-action on V2 by pointwise right multiplication. We conclude
that
Gr2pVq “ V2 � Up2q.
Remark 4.1.5. The moment map is closely related to the frame-Hopf map (see Section
3.1.2). The difference of the diagonal coordinates of the moment map is
i p〈φ, φ〉L2 ´ 〈ψ, ψ〉L2q “
ż 2
0
i`
|φ|2 ´ |ψ|2˘
dt.
The integrand is i times the first entry of FrameHopf1. The upper right entry of the moment
map is
´i 〈φ, ψ〉L2 “
ż 2
0
´iφψ dt.
The real part of the integrad is the second entry of FrameHopf1, while the imaginary part is
the negation of the remaining entry of FrameHopf1. Thus the closed framed loops correspond
to the level set of¨
˚
˝
i 0
0 i
˛
‹
‚
P up2q.
89
Proof. Let 〈¨, ¨〉up2q denote the standard inner product on up2q (an extension of the standard
inner product on sup2q), and fix an arbitrary matrix
ξ “
¨
˚
˝
ia z
´z ib
˛
‹
‚
P up2q “ tskew-Hermitian matricesu.
For the remainder of this proof, let F denote the map
F : V2Ñ R
Φ ÞÑ⟨µUp2qpΦq, ξ
⟩up2q
.
More explicitly,
F pΦq “1
2trace
`
µUp2qpΦq ¨ ξ˚˘
“1
2trace
¨
˚
˝
i
¨
˚
˝
〈φ, φ〉L2 ´ 〈φ, ψ〉L2
´ 〈ψ, φ〉L2 〈ψ, ψ〉L2
˛
‹
‚
¨
˚
˝
´ia ´z
z ´ib
˛
‹
‚
˛
‹
‚
“1
2
`
a}φ}2L2 ´ iz 〈φ, ψ〉L2 ` iz 〈ψ, φ〉L2 ` b}ψ}2L2
˘
“1
2
`
a}φ}2L2 ` b}ψ}2L2
˘
` Im pz 〈φ, ψ〉L2q
Our goal is to show that
DF pΦqpδΦq “ ωL2
pδΦ,Φ ¨ ξq, (4.3)
90
where Φ ¨ ξ is the infinitesimal vector field associated to ξ. This can be established directly:
DF pΦqpδΦq “d
dε
ˇ
ˇ
ˇ
ˇ
ε“0
1
2
`
a}φ` εδφ}2L2 ` b}ψ ` εδψ}2L2
˘
` Im pz 〈φ` εδφ, ψ ` εδψ〉L2q
“ aRe 〈φ, δφ〉L2 ` bRe 〈ψ, δψ〉L2 ` Im pz p〈φ, δψ〉L2 ` 〈δφ, ψ〉L2qq
“ ´Im 〈δφ, iaφ〉L2 ´ Im 〈δψ, ibψ〉L2 ´ Im 〈δφ,´zψ〉L2 ´ Im 〈δψ, zφ〉L2
“ ´Im 〈pδφ, δψq, piaφ´ zψ, zφ` ibψq〉L2
“ ωL2
pδΦ,Φ ¨ ξq.
It is easy to check that
ξ0 :“
¨
˚
˝
i 0
0 i
˛
‹
‚
is a regular value of µUp2q—for example, the collection
tpφ, 0q, p0, ψq, pψ, φq, piψ,´iφqu
maps to a spanning set for up2q under DµUp2qpφ, ψq. Recalling that St2pVq “ µ´1Up2qpξ0q and
Gr2pVq “ St2pVq{Up2q, we have shown that
Gr2pVq “ V2 � Up2q,
i.e., the Grassmannians are obtained as symplectic reductions.
A primary feature of the Marsden-Weinstein reduction process is that the restriction of
the symplectic form to the level set submanifold degenerates exactly along the group orbit
fibers—i.e. the vertical space of the submersion. We record a useful version of this fact in
the special case of interest in the following lemma.
91
Lemma 4.1.6. Let rΦs P Gr2pVq, δΦ1 P ThorΦ St2pVq « TrΦsGr2pVq and δΦ2 P TΦSt2pVq and
let proj : TΦSt2pVq Ñ T horΦ St2pVq denote orthogonal projection. Then
ωL2
Φ pδΦ1, δΦ2q “ ωL2
Φ pδΦ1, projpδΦ2qq,
where ωL2
is the symplectic form of V2 restricted to St2pVq.
Proof. This statement is equivalent to the statement that
ωL2
Φ pδΦ1, δΦvert2 q “ 0
for all δΦ1 P ThorΦ St2pVq and δΦ2 P TΦSt2pVq, where Φvert
2 is the vertical component of the
tangent vector. Indeed, the vertical component can be written as δΦvert2 “ Φ ¨ ξ for some
ξ P up2q, and
ωL2
Φ pδΦ1,Φ ¨ ξq “⟨DµUp2qpΦqpδΦq, ξ
⟩up2q
“ 〈0, ξ〉up2q “ 0.
By Theorem 3.2.8, the various geometric structures of Gr2pVq are passed on to the moduli
space of framed loopsM. We have already seen (as part of the theorem) that pGr2pVq, gL2q
is locally isometric to pM, gSq, where gS is the metric induced from the natural metric on
framed path space constructed in Section 3.1.1. We now describe the induced complex and
symplectic structures on M.
4.1.3 The Induced Kahler Structure of M
To describe the complex structure of M, we note that it is an easy corollary of Proposition
3.1.10 that the map pH restricts to a locally isometric double-cover from
W :“ LpC2zt0uq \ApC2
zt0uq Ă PC2,
92
a disjoint union of open subsets of complex vector spaces, onto the space of frame-periodic
paths
rSper “ tframe-periodic framed pathsu{ttranslationu Ă rS.
An element of rSper is a framed path pγ, V q with γp0q “ 0 such that V is periodic and γ is
almost periodic in the sense that
γpkqp0q “ γpkqp2q @ k “ 1, 2, . . . ,
but γp0q is not necessarily equal to γp2q. Thus such a pγ, V q determines a loop in SOp3qˆR`.
By the previous discussion, M can be viewed as a symplectic reduction of rSper by the
action of SOp3qˆS1, where SOp3q acts by rigid rotations and S1 acts by global frame twists.
This is summarized by the commutative diagram:
W Gr˝2pLCq \Gr˝2pACq
rSper M
�Up2q
pHpH
�pSOp3q ˆ S1q
Therefore, it suffices to describe the complex structure on the simpler space rSper, as this
structure descends to M.
Let pγ, V q P rSper and let T “ γ1{}γ1} and W “ T ˆ V . Any variation pδγ, δV q can be
recovered from pδγ1, δV q, and this can be expressed uniquely as
pδγ1, δV q “ λstretchXstretch ` λtwistXtwist ` λbend1Xbend1 ` λbend2Xbend2 , (4.4)
93
T
V
WJ
-‐Jγ
J
-‐J
Xstretch Xtwist Xbend1 Xbend2
Figure 4.1: The complex structure ofM.
where the λ’s are R-valued functions and the X’s are the following four basic variations :
Xstretch “ p}γ1}T, 0q “ stretch tangent vector, leave frame vector unchanged
Xtwist “ p0,´W q “ twist the frame in the negative direction around γ
Xbend1 “ p}γ1}W, 0q “ bend the curve in the negative direction around V
Xbend2 “ p´}γ1}V, T q “ bend the curve in the negative direction around W.
Then we define an almost complex structure, denoted J , by extending the rule
J : Xstretch ÞÑ Xtwist, J : Xbend1 ÞÑ Xbend2 , and J2“ ´Id. (4.5)
over linear combinations. The complex structure’s pointwise action is depicted in Figure 4.1.
We claim that this almost complex structure is, in fact, an honest complex structure and
that it is the induced structure coming from our identification with rS. To show this, we first
express the basic variations in quaternionic coordinates.
94
Lemma 4.1.7. Let pγ, V q P rSper. Using quaternionic coordinates, let q PW satisfy
pHpqq “ pγ, V q.
Then, under the map DpHpqq, we have the correspondences
q
2Ø Xstretch
iq
2Ø Xtwist
jq
2Ø Xbend1
kq
2Ø Xbend2 .
Proof. In quaternionic coordinates, pH takes the form
pHpqq “
ˆ
t ÞÑ
ż t
0
Impqiqq dt, Im
ˆ
qjq
}q}2H
˙˙
,
so for any variation δq, we have
DpHpqqpδqq “d
dε
ˇ
ˇ
ˇ
ˇ
ε“0
˜
t ÞÑ
ż t
0
Imppq ` εδqqipq ` εδqqq dt, Im
˜
pq ` εδqqjpq ` εδqq
}q ` εδq}2H
¸¸
“
ˆ
t ÞÑ
ż t
0
2Impqiδqq dt,2
}q}2H
`
Impqjδqq ´ RepqδqqV˘
˙
“: pδγ, δV q,
where we once again use Im and Re in this context to denote the maps
Impq0 ` iq1 ` jq2 ` kq3q “ pq1, q2, q3q P R3, Repq0 ` iq1 ` jq2 ` kq3q “ q0 P R.
Thus
δγ1 “ 2Impqiδqq.
95
Now the claim follows by simply calculating explicitly. For δq “ q{2 we have
δγ1 “ 2Im´
qiq
2
¯
“ Impqiqq “ γ1
and
δV “2
}q}2H
ˆ
Im´
qjq
2
¯
´ 2Re
ˆ
q´q
2
¯
˙
V
˙
“1
}q}2H
`
Impqjqq ´ Re`
}q}2H˘
V˘
“ V ´ V “ 0.
Thus δq “ q{2 corresponds to Xstretch, as claimed.
Similarly, for δq “ iq{2,
δγ1 “ 2Im
ˆ
qiiq
2
˙
“ Imp´}q}2Hq “ 0
and
δV “2
}q}2H
˜
Im
ˆ
qjiq
2
˙
´ Re
˜
q
ˆ
iq
2
˙
¸
V
¸
“1
}q}2HpImp´qkqq ´ Rep´i}q}2HqV q
“ ´Impqkqq
}q}2H“ ´W,
so δq “ iq{2 corresponds to Xtwist.
The remaining computations are similar. For δq “ jq{2,
δγ1 “ 2Im´
qkq
2
¯
“ }γ1}W and δV “2
}q}2H
ˆ
Im´
2
¯
´ Re
ˆ
2j
˙
V
˙
“ 0.
96
and for δq “ kq{2,
δγ1 “ 2Im´
qp´jqq
2
¯
“ ´}γ1}V and δV “2
}q}2H
ˆ
Im´
qiq
2
¯
´ Re
ˆ
2p´kq
˙˙
“ T.
Then we have the immediate corollaries:
Corollary 4.1.8. The complex structure defined in (4.5) on rSper is induced from the complex
structure of W via the map pH
Proof. In quaternionic coordinates, the usual complex structure is replaced by multiplication
by the quaternionic i. Thus Lemma 4.1.7 gives correspondences
i ¨q
2“
iq
2Ø J ¨Xstretch “ Xtwist
i ¨jq
2“
kq
2Ø J ¨Xbend1 “ Xbend2
i2 “ ´Id Ø J2“ ´Id.
Remark 4.1.9. It follows from Lemma 4.1.7 that rSper actually has a quaternionic structure.
Corollary 4.1.10. Let Φ “ pφ, ψq P W correspond to pγ, V q P rSper under pH. The basic
variations are represented in complex coordinates as
1
2Φ Ø Xstretch
i
2Φ Ø Xtwist
1
2p´ψ, φq Ø Xbend1
1
2p´iψ, iφq Ø Xbend2 .
Proof. Recall that the identification of C2 with H is given by pz, wq Ø z`wj. The corollary
follows by rewriting each quaternionic variation using this identification.
97
This complex structure transfers from rSper to M, although this requires some clarifica-
tion. Any functional linear combination of the form
λstretch1
2Φ` λtwist
i
2Φ` λbend1
1
2p´ψ, φq ` λbend2
1
2p´iψ, iφq (4.6)
determines an admissible variation of Φ “ pφ, ψq PW , since W is a union of open subsets of
vector spaces. Thus any variation of the form (4.4) of pγ, V q P rSper is admissible. However,
when we restrict to Gr2pVq, we impose conditions on which variations of rΦs P Gr2pVq of
the form (4.6) are admissible—this is because the tangent spaces to the Grassmannians have
positive codimension in W . Nonetheless, we have shown in the previous section that the
complex structure of Gr2pVq is well-defined. Similarly, when we restrict to M we impose
conditions on admissible variations of rγ, V s P M of the form (4.4). The fact that the
complex structure of M described in this section is well-defined is a corollary of Theorem
3.2.8, Proposition 4.1.4 and the discussion above.
ThusM has a well-defined complex structure. Readers familiar with the work of Millson
and Zombro will note that it is no coincidence that J closely resembles the complex structure
constructed on a related moduli space of loops in [55]. This will be treated later in Section
4.3.
Finally, the symplectic form on rSper is defined by
ωSp¨, ¨q :“ gSpJ ¨, ¨q.
As usual, we will not use special notation to distinguish this form from the induced form on
M.
98
4.2 Principal Bundle Structure
4.2.1 Statement of the Result
As we have so far identified S1 with r0, 2s{p0 „ 2q, we introduce the special notation S1
for the length-2π circle, considered as a subset of C in the natural way. The space LS1 is
a tame Frechet Lie group with the obvious pointwise multiplication. It will frequently be
convenient to denote points of LS1 by eiα, where α P PR is 2π ¨ k-periodic for some k P Z.
By identifying S1 with the set of constant loops in LS1, S1 embeds as a Lie subgroup. The
quotient LS1{S1 is once again a tame Frechet Lie group and it has a natural free action on
Gr2pVq by pointwise multiplication. Explicitly, the action is
preiαs ¨ rφ, ψsqptq :“ reiαptq ¨ φptq, eiαptq ¨ ψptqs,
where reiαs P LS1{S1 denotes the equivalence class of eiα P LS1.
According to Theorem 3.2.8, this action corresponds to an action of LS1{S1 on the moduli
space of relatively framed loops, M. Indeed, the action is by adjusting the framing: an
element eiα P LS1 acts on a framed loop pγ, V q by fixing γ and adjusting V according to the
formula
V ÞÑ cosp2αqV ` sinp2αq
ˆ
γ1
}γ1}ˆ V
˙
.
This action induces a well-defined action on equivalence classes of framed loops up to trans-
lation, scaling and rotation, and descends to a well-defined action of LS1{S1 on equivalence
classes up to relative framing. From this interpretation, we see that the action of LS1{S1 is
free—if
reiαs ¨ rγ, V s “ rγ, V s,
99
then the framed loopˆ
γ, cosp2αqV ` sinp2αq
ˆ
γ1
}γ1}ˆ V
˙˙
differs from pγ, V q by a rigid rotation. This is only possible if α is constant, and thus reiαs
is the identity class in LS1{S1.
The main result of this section is that this action gives M the structure of a principal
bundle over the moduli space of unframed loops. To be precise, let B denote the moduli
space of unframed loops
B :“ timmersed loops in R3u{ttranslation, scaling, rotationu.
Lemma 4.2.1. The moduli space B is a Frechet manifold.
Proof. We model B as
B « tγ P Imm0pS1,R3
q | lengthpγq “ 2u{SOp3q,
where we let Imm0pS1,R3q denote the space of immersed loops based at zero—an open
submanifold of a tame Frechet vector space.
The space
tγ P Imm0pS1,R3
q | lengthpγq “ 2u
is a tame Frechet manifold as it is the level set of a regular value of the length functional and
we once again invoke Corollary 2.2.10. We obtain smooth cross-sections for the SOp3q-action
by following an argument of Millson and Zombro ([55], Lemma 1.6). Let Ut0 denote the open
set of length-2 based immersed curves containing the curves with nonvanishing curvature at
t0 P r0, 2s. Every such curve γ has a principal normal vector Nγpt0q at γpt0q, and we take
the SOp3q-cross-section through Ut0 to be the inverse image of the identity matrix under the
100
map
Ut0 Ñ SOp3q
γ ÞÑ
ˆ
γ1pt0q
}γ1pt0q}, Nγpt0q,
γ1pt0q
}γ1pt0q}ˆNγpt0q
˙
.
It is easy to see that these sets do define cross-sections for the SOp3q-action and that
transition maps are smooth, thus proving the claim (see [55] for more details).
Thus our goal for this section is to show that LS1{S1 ãÑM Ñ B is a Frechet principal
bundle.
4.2.2 The Failure to Close Map
The obvious projection M Ñ B is the “forget framing” map rγ, V s ÞÑ rγs. It will be useful
to express this projection map in terms of the Grassmannian formalism. Let pHp1q denote the
first coordinate of pH; that is,
pHp1qprΦsq “ rγs P B ô pHprΦsq “ rγ, V s PM.
Then, by Theorem 3.2.8, the forget framing map can be expressed as
pHp1q : Gr2pLCq \Gr2pACq Ñ B
In this form it is easy to check from the explicit formula for pH that each fiber of the projection
is diffeomorphic to LS1{S1.
101
Thus it remains to show that M is locally diffeomorphic to B ˆ LS1{S1. We begin by
recalling that the twist rate of a framed path pγ, V q is defined to be
twpγ, V q :“
⟨d
dsV,W
⟩,
where d{ds is derivative with respect to arclength of γ and W “ γ1{}γ1}. The twist rate
is well-defined for relatively framed curves and is invariant under translation, scaling, and
rotation. We also recall that the total twist of pγ, V q is the integral of the twist rate with
respect to arclength; i.e.,
Twpγ, V q “1
2π
ż 2
0
twpγ, V q }γ1ptq}dt.
It will be useful to express the twist rate in pφ, ψq-coordinates. We first prove an auxilliary
lemma.
Lemma 4.2.2. Let q P PpHzt0uq correspond to the framed path pγ, V q in quaternionic
coordinates. Then the variation δq “ q1{}q}2H of q corresponds to the variation
pδγ1, δV q “
ˆ
d
dsγ1,
d
dsV
˙
of pγ, V q under DpHpqq.
Proof. Recall from the proof of Lemma 4.1.7 that the variation corresponding to δq is
pδγ, δV q “
ˆ
t ÞÑ
ż t
0
2Impqiδqq dt,2
}q}2H
`
Impqjδqq ´ RepqδqqV˘
˙
102
Thus for δq “ q1{}q}2H, we have
δγ1 “ 2Im
ˆ
qiq1
}q}2H
˙
“1
}q}2H2Impqiq1q
“1
}q}2H
d
dtImpqiqq
“1
}γ1}
d
dtγ1 “
d
dsγ1.
Moreover,
δV “2
}q}2H
˜
Im
ˆ
qjq1
}q}2H
˙
´ Re
˜
q
ˆ
q1
}q}2H
˙
¸
V
¸
“1
}q}4H
ˆ
d
dtImpqjqq ´ V
d
dt}q}2H
˙
“1
}q}2H
}q}2Hddt
Impqjqq ´ Impqjqq ddt}q}2H
}q}4H
“1
}γ1}
d
dtV “
d
dsV.
Lemma 4.2.3. Let Φ “ pφ, ψq P St˝2pVq correspond to a framed loop pγ, V q. Then
twpγ, V q “ ´2Im
`
φ1φ` ψ1ψ˘
p|φ|2 ` |ψ|2q2. (4.7)
Proof. Rearranging the expression on the right, we obtain
´2Im
`
φ1φ` ψ1ψ˘
p|φ|2 ` |ψ|2q2“ ´
2
}Φ}4C2
Im 〈Φ1,Φ〉C2
“ ´4
}Φ}2C2
Re
⟨Φ1
}Φ}2C2
,iΦ
2
⟩C2
.
103
In quaternionic coordinates, the last expression is equal to
´4
}q}2HRe
⟨q1
}q}2H,iq
2
⟩H.
Applying Lemma 4.2.2, Lemma 4.1.7 and Corollary 3.1.11, this is equal to
´4
}γ1}¨
1
4
ˆ⟨d
dsδγ, 0
⟩`
⟨d
dsV,W
⟩〈´W,W 〉
˙
}γ1} “
⟨d
dsV,W
⟩“ twpγ, V q,
where W “ γ1{}γ1} and δγ satisfies δγ1 “ ddsγ1.
The quantity on the right hand side of (4.7) will also be denoted by twpΦq. We may also
express the total twist in pφ, ψq-coordinates as
TwpΦq “1
2π
ż 2
0
twpΦq`
|φ|2 ` |ψ|2˘
dt.
Moreover, these maps induce well-defined maps on equivalence classes rΦs P Gr˝2pVq.
We now introduce the map
ĆFTC :MÑ S1Ă C
rγ, V s ÞÑ exppiπTwpγ, V qq
(“FTC” stands for “failure to close”—this will be explained momentarily). It is clear that
this map is well-defined, as Tw is well-defined for relative framings and is invariant under
translation, scaling and rotation. Perhaps less obvious is the fact that the value of ĆFTC does
not depend on the framing.
Lemma 4.2.4. The map ĆFTC :MÑ S1 induces a well-defined map on B.
104
Proof. We need to show that ĆFTC is LS1{S1-invariant. Let rγ, V s PM have Grassmannian
representation rΦs “ rφ, ψs. For reiαs P LS1{S1. It follows easily from Lemma 4.2.3 that
twpeiαΦq “ ´2α1
|φ|2 ` |ψ|2` twpΦq.
Thus
TwpeiαΦq “1
2π
ż 2
0
ˆ
´2α1
|φ|2 ` |ψ|2` twpΦq
˙
p|φ|2 ` |ψ|2qdt
“1
πpαp0q ´ αp2qq ` TwpΦq.
Since eiα is a loop, αp0q ´ αp2q is a multiple of 2π, and we have
exppiπTwpeiαΦqq “ exppiπTwpΦqq,
as claimed.
We denote the induced map by FTC : B Ñ S1. This is the failure to close map; it
measures the angle by which the Bishop frame [4] of a given loop fails to close. Recall from
Example 2.1.2 that the Bishop frame of a space curve γ is the relative framing obtained by
choosing a normal vector V p0q and parallel transporting it in the normal bundle of γ, and
that this framing does not necessarily close up even if the base curve γ is closed. With FTC
defined, we are equipped to prove the main theorem of this section.
4.2.3 The Principal Bundle Theorem
We are prepared to prove the main result of this section, which is stated formally as follows.
105
Theorem 4.2.5. The “forget framing” map
MÑ B
rγ, V s ÞÑ rγs
gives M the structure of a principal bundle over B with structure group LS1{S1.
The proof strategy is stated informally as follows. We divide the space of unframed loops
into two open sets using the map FTC. Roughly, there is an open set containing loops
whose Bishop frames fail to close by a small amount and an open set containing loops whose
Bishop frames fail to close by a large amount. For each open set, we assign a preferred
relative framing to each loop by “smoothly closing the Bishop framing”—more precisely, we
increase or decrease the twist rate by the smallest amount possible so that the resulting frame
smoothly closes. These frames are essentially the rotation minimizing frames of, e.g., [74].
The amount by which to increase the twist rate is not canonical over the space of all loops,
which is why the principal bundle is nontrivial. To show that M is locally diffeomorphic to
B ˆ LS1{S1, we define a map which takes rγ, V s to prγs, reiαsq, where reiαs is the loop class
measuring the difference between the relative framing rV s and the preferred relative framing.
We now make this precise.
Proof. We will construct explicit local trivializations ofM. First define local charts pfj, Ujq
for j “ ˘1 on S1 Ă C by
Uj “ S1ztju, j “ ˘1
and
fj : Uj Ñ p0, 2πq
f´1peiθq “ θ ` π for θ P p´π, πq and f1pe
iθq “ θ for θ P p0, 2πq.
106
Let
Uj :“ FTC´1pUjq and rUj :“ ĆFTC
´1pUjq.
Then Uj is an open subset of B, rUj is an open subset of Gr˝2pLCq \ Gr˝2pACq « M, and
pHp1qp rUjq “ Uj. Referring back to the informal discussion above, the open subset U´1 corre-
sponds to those loops whose Bishop framing “fails to close by a small amount”.
We claim that
rUj « Uj ˆ LS1{S1, j “ ˘1.
We will define an explicit diffeomorphism by first defining maps
Gj : rUj Ñ PR j “ ˘1,
where Gj is defined on Φ “ pφ, ψq by the formula
GjpΦqptq :“1
2π
ż t
0
twpΦqp|φ|2 ` |ψ|2q dt´t
2πfj ˝ĆFTCpΦq
and we note that this descends to a well defined map on equivalence classes rΦs. We also
note that
GjpΦqp0q “ 0,
and
GjpΦqp2q “ TwpΦq ´ pTwpΦq mod 2πq “ 2πk for some k P Z.
Therefore
rexppiGjpΦqqs P LS1{S1,
107
and we define our local trivializations by
rUj Ñ Uj ˆ LS1{S1
rΦs ÞÑ´
pHp1qpΦq, rexppiGjpΦqqs¯
j “ ˘1. (4.8)
Informally, the equivalence class of loops rexppiGjpΦqqs records the difference between the
original framing of the framed loop corresponding to Φ and the modified Bishop framing of
its base loop.
To show that these maps are diffeomorphisms, we construct their inverses. Let rΦs be
any element of the fiber of rγs P Uj. We claim that rexppiGjpΦqqΦs only depends on rγs and
not on the choice of rΦs; i.e., we wish to show that for any reiαs P LS1{S1,
rexppiGjpeiαΦqqeiαΦs “ rexppiGjpΦqqΦs.
From the definition of Gj, we easily deduce that
GjpeiαΦqptq “ αp0q ´ αptq `GjpΦqptq.
Up to S1-equivalence, we have
rexppipαp0q ´ αptq `GjpΦqptqqqs “ rexppiGjpΦptqqq expp´iαptqqs,
We conclude
rexppiGjpeiαΦqqeiαΦs “ rexppiGjpΦqqe
´iαeiαΦs “ rexppiGjpΦqqΦs.
108
Now we define the inverse map of (4.8) by
Uj ˆ LS1{S1
Ñ rUj
prγs, reiθsq ÞÑ rexppipGjpΦq ´ θqqΦs j “ ˘1,
where rΦs is any element of the fiber of rγs. Informally, this map produces of a framing of
a loop γ by taking its modified (closed) Bishop framing and adjusting it according to reiθs.
We leave it to the reader to check that these are our inverse maps and this completes the
proof.
4.2.4 Non-Existence of a Global Framing Algorithm
This theorem has a corollary which follows trivially but is of practical interest. A well-
studied problem in applied differential geometry is to algorithmically assign a framing or
relative framing to a given parameterized space curve [9, 14, 26, 74]. This has applications
to computer graphics, where one uses the framing to construct a tube around a given curve
for visual clarity [26], as well as animation, motion planning and camera tracking.
To make the problem precise, we define a curve framing algorithm to be a continuous
map from any subset of the space of immersed paths, Immpr0, 2s,R3q, to the space of framed
paths that is of the form γ ÞÑ pγ, V q. Similarly, we define a relative framing algorithm to
be a continuous map which assigns a relative framing to each immersed path lying in some
subset of Immpr0, 2s,R3q. If γ is a loop, then we require a (relative) framing algorithm to
produce a smoothly closed (relative) framing.
For example, the most classical curve framing algorithm assigns to an immersed path
γ its Frenet framing, where the frame vector is the principal normal vector to γ. The
principal normal vector is not well-defined any point where the curvature of γ vanishes. In
109
our language, the curve framing algorithm is defined on the subset
tγ P Immpr0, 2s,R3q | κpγq ą 0u.
A classical relative framing algorithm assigns to each path its Bishop framing. This
algorithm is defined on the full space of immersed paths. However, if γ is a loop then the
Bishop framing may fail to close and this is a problem in many applications (e.g., for the
application of providing a tubing of the curve, the smooth closure of the framing is required
if the tube is to be textured). Therefore, the restriction of this curve framing algorithm to
loops is only defined on the subset
tγ P ImmpS1,R3q | Wrpγq P Zu,
(by the Calugareanu-White-Fuller Theorem [20]) . In fact, every framing algorithm that has
been introduced fails on some subset of loop space. From our perspective, the reason for this
phenomenon is obvious: there is no continuous global section B ÑM.
Corollary 4.2.6. There is no Euclidean similarity-invariant curve framing algorithm or
relative framing algorithm with domain ImmpS1,R3q.
Proof. Any such framing algorithm or relative framing algorithm would induce a continuous
global section of the principal bundle MÑ B. This would imply that M is homeomorphic
to B ˆ LS1{S1, but M has two components and B ˆ LS1{S1 has countably infinitely many
componenents, so this is impossible.
Remark 4.2.7. Using similar methods to this section, one is able to show that
PS1{S1 ãÑ trel. framed pathsu {ttransl., rot., scal.u Ñ Immpr0, 2s,R3
q{ttransl., rot., scal.u
110
is a principal bundle. This bundle does admit a global section—e.g., the Bishop framing.
There is no contradiction here, as PS1{S1 is contractible.
4.3 Symplectic Reduction by the action of LS1{S1
Our goal in this section is to show that the symplectic reduction of the moduli space of
framed loopsM by the action of LS1{S1 described above is isomorphic to the moduli space
of Millson-Zombro, denoted MMZ . We recall that the Millson-Zombro moduli space is
MMZ :“ tarclength-parameterized (unframed) loopsu{ttranslation, scaling, rotationu.
In [55] it was shown that MMZ is a Frechet manifold. Moreover, it was shown to admit
a complex Kahler structure. This was accompished by showing that MMZ is obtained as
the symplectic reduction of LS2 by the action of SOp3q by rigid rotations. The work in
this section provides a new interpretation of the Kahler structure of MMZ and moreover
demonstrates an infinite-dimensional Gel’fand-Macphereson Correspondence.
4.3.1 The Momentum Map
We note thatMMZ is connected, whileM has two connected components. Indeed, perform-
ing the Kahler reduction process on M will produce two copies of MMZ . For this reason,
we will actually work with only one of the components of M—say the component contain-
ing framed loops with odd linking number. For this reason, we will work specifically with
St2pLCq and Gr2pLCq in this section (ignoring their anti-loop counterparts).
111
We begin by identifying the relevant pieces in the construction. The Lie algebra of LS1{S1
is LR{R. We define an inner product on LR{R by the formula
〈rαs, rβs〉LR :“1
2
ż 2
0
αβ dt´1
4
ż 2
0
α dt
ż 2
0
β dt, α, β P LR.
This is (up to a constant multiple) the inner product induced by the standard L2 inner
product on LR.
Now we note that the LS1{S1-action on Gr2pLCq is induced from the action of LS1 on
St2pLCq, defined by
peiα ¨ Φqptq :“ eiαptqΦptq.
Let α P LR, the Lie algebra of LS1. The exponential map LR Ñ LS1 is α ÞÑ eiα, so the
infinitesimal vector field associated to α is given by
Xα|Φ “d
dε
ˇ
ˇ
ˇ
ˇ
ε“0
eiεαΦ “ iαΦ.
Thus a point rαs P LR{R induces a vector field Xrαs on Gr2pLCq defined by orthogonally
projecting; i.e.,
Xrαs|rΦs “ projpiα ¨ Φq,
where proj is orthogonal projection
proj : TΦSt2pLCq Ñ T horΦ St2pLCq « TrΦsGr2pLCq
with respect to gL2. This is well-defined, since the horizontal space is finite codimension in
the full tangent space.
112
We define our candidate for the moment map for the LS1{S1-action to be
µ “ µLS1{S1 : Gr2pLCq Ñ LR{R
rφ, ψs ÞÑ r|φ|2 ` |ψ|2s.
The rest of this section is devoted to proving that the Kahler reduction process is well-
defined in this setup and that it produces the space MMZ .
Lemma 4.3.1. The map µ is a moment map for the action of LS1{S1.
Proof. We first compute the derivative of the map
rΦs “ rφ, ψs ÞÑ 〈µprΦsq, rαs〉LR “1
2
ż 2
0
p|φ|2 ` |ψ|2qα dt´1
2
ż 2
0
α dt
at the point rΦs in the direction δΦ “ pδφ, δψq P T horΦ St2pLCq for fixed rαs P LR{R. This is
given by
d
dε
ˇ
ˇ
ˇ
ˇ
ε“0
1
2
ż 2
0
p|φ` εδφ|2 ` |ψ ` εδψ|2qα dt´1
2
ż 2
0
α dt “
ż 2
0
Repφδφ` ψδψqα dt.
On the other hand, we need check that this is equal to
ωL2
rΦspδΦ, Xrαs|rΦsq “ ´Im 〈δΦ, projpiαΦq〉L2 .
To see this, we write
Xrαs|rΦs “ projpiαΦq “ iαΦ´ piαΦqK,
113
where piαΦqK is perpendicular to T horΦ St2pLCq with respect to gL2
(i.e., piαΦqK is tangent to
the Up2q-orbit of Φ). Then
´Im 〈δΦ, projpiαΦq〉L2 “ ´Im⟨δΦ, iαΦ´ piαΦqK
⟩L2
“ ´Im 〈δΦ, iαΦ〉L2 ` Im⟨δΦ, piαΦqK
⟩L2
“ Re 〈αΦ, δΦ〉L2 ` Re⟨´iδΦ, piαΦqK
⟩L2
“
ż 2
0
Re`
φδφ` ψδψ˘
α dt` 0,
and this proves the claim.
4.3.2 Manifold Structure of the Space of Arclength-Parameterized
Framed Loops
Next we choose a suitable level set of the moment map to construct an LS1{S1-invariant
submanifold of Gr2pLCq. Let 1 P LR{R denote equivalence class of the loop that takes
the value 1 identically. We wish to show that µ´1p1q is a submanifold with the desired
properties. From Lemma 3.1.7 we deduce that µ´1p1q corresponds under pH to the subset of
M consisting of relatively framed loops with arclength parameterized base curve:
µ´1p1q «Marc :“ trγ, V s PM | }γ1} ” 1u.
Proposition 4.3.2. The space µ´1p1q is an LS1{S1-invariant smooth Frechet manifold with
tangent space
TrΦsµ´1p1q «
δΦ P T horΦ St2pLCq | Re 〈Φ, δΦ〉C2 ” 0(
.
114
The strategy we will use to prove this theorem is to note that there is a natural identifi-
cation
µ´1p1q « pLS3
X St2pLCqq{Up2q,
where S3 Ă C2 denotes the standard 3-sphere and we consider LS3 as a submanifold of LC2.
Thus we first aim to show that LS3X St2pLCq is a Frechet manifold. This follows by noting
that it is a level set of the Frechet manifold LS3 ˆ R` with respect to the map
µUp2q : LS3ˆ R` Ñ up2q
pφ, ψq ÞÑ i
¨
˚
˝
〈φ, φ〉L2 ´ 〈φ, ψ〉L2
´ 〈ψ, φ〉L2 〈ψ, ψ〉L2
˛
‹
‚
Note that this is a restriction of the momentum map introduced in Proposition 4.1.4 for the
Up2q-action on LpC2zt0uq. We are considering LS3 ˆR` as embedded in LpC2zt0uq via the
embedding
LS3ˆ R` ãÑ LpC2
zt0uq
pΦ, rq ÞÑ rΦ.
Thus our goal is to show that elements of
LS3X St2pLCq « pLS3
ˆ R`q X St2pLCq
are always regular points for µUp2q, and then the proposition follows by Corollary 2.2.10. It
is not so easy to show explicitly that elements of LS3 X St2pLCq are regular points, so we
introduce the following technical lemmas. Their purpose is to allow us to produce enough
115
Figure 4.2: Regions described in the proof of Lemma 4.3.3.
tangent vectors δΦ P TΦLS3 to map to a spanning set for up2q under DµUp2qpΦq for any
Φ P LS3 X St2pLCq.
Lemma 4.3.3. Let eiα P LS1 such that
ż
S1
eiα dt “ 0.
Then there exist functions B1, B2 : S1 Ñ Rě0 such that
ż
S1
B1 ¨ eiα dt and
ż
S1
B2 ¨ eiαdt
are linearly independent over R as elements of R2 “ C.
Proof. The assumption that
0 “
ż
S1
eiαdt “
ż
S1
cosα dt` i
ż
S1
sinα dt
116
says that the average value of the loop eiα lies at the origin in C “ R2. We claim that there
exists an open subarc I1 Ă S1 such that | cosα| ą | sinα| on I1. This claim means that the
image of eiα on I1 lies in one of the highlighted regions in Figure 4.3.2 intersected with S1.
From this interpretation, it is clear that such a subarc must exist due to the average value
assumption. Similarly, there must exist some open subarc I2 Ă S1 such that | cosα| ă | sinα|
on I2.
Now we define Bj : S1 Ñ Rě0 to be a nonzero bump function supported on Ij. Then
the image of B1eiα lies entirely in one of the shaded regions of Figure 4.3.2 intersected with
S1, and this implies that the average value of B1eiα is nonzero and lies in the same shaded
region (though not necessarily on S1). Similarly, the average value of the image of B2eiα
is nonzero and lies in an unshaded region. Multiplication by a real number will not take a
point in a shaded region to a point in an unshaded region, so the proof is complete.
Before moving on to the next lemma, we note that the tangent space to LS3 at Φ is
concretely realized as
TΦLS3“ tδΦ P LC2
| Re 〈Φ, δΦ〉C2 ” 0u.
Lemma 4.3.4. Let Φ “ pφ, ψq P LS3 X St2pLCq. Then there exists pδφ, δψq P TΦLS3 such
that
Im⟨pφ, ψq, pδψ, δφq
⟩L2 ‰ 0.
Proof. Consider the tangent vector
pδφ, δψq “ p´iψ, iφq ´ Re 〈p´iψ, iφq, pφ, ψq〉C2 pφ, ψq P Tpφ,ψqLS3.
We claim that
Im⟨pφ, ψq, pδψ, δφq
⟩C2 ı 0
117
for this choice of δΦ. Indeed, if this is not the case, then
0 ” Im⟨pφ, ψq, piφ´ Re 〈p´iψ, iφq, pφ, ψq〉C2 ψ, ψ ´ Re 〈pψ, φq, pφ, ψq〉C2 φq
⟩C2
“ Im`
´i|φ|2 ´ i|ψ|2 ´ 2Re 〈p´iψ, iφq, pφ, ψq〉C2 φψ˘
“ ´1´ 2Impψφ` φψqImpφψq
“ ´1` 4Impψφq2.
This implies that
Impψφq ”1
2or Impψφq ” ´
1
2,
contradicting pφ, ψq P St2pLCq as it implies
ż
Impψφq dt ‰ 0.
Now assume without loss of generality that Im⟨pφ, ψq, pδψ, δφq
⟩C2 takes some positive
value. Then there exists an open subarc I Ă S1 so that Im⟨pφ, ψq, pδψ, δφq
⟩C2 ą 0 on I.
Let I 1 Ă I be a proper open subset, and let B : S1 Ñ Rě0 be a smooth bump function with
B|I 1 ” 1, and supportpBq Ă I. Then
Im⟨pφ, ψq, pB ¨ δψ,B ¨ δφq
⟩L2 “
ż
S1
B ¨ Im⟨pφ, ψq, pδψ, δφq
⟩C2 dt
ě
ż
I 1Im⟨pφ, ψq, pδψ, δφq
⟩C2 dt
ą 0.
Since pB ¨ δφ,B ¨ δψq P TΦLS3, this proves the claim.
118
Proof of Proposition 4.3.2. Fix Φ “ pφ, ψq P LS3 X St2pLCq. Our goal is to show that Φ is
a regular point of the restricted map µUp2q. That is, we wish to find four tangent vectors in
TΦpLS3ˆ R`q “ tδΦ “ pδφ, δψq P LC2
| Re 〈Φ, δΦ〉C2 ” 0u ‘ tλΦ | λ P Ru
whose images are linearly independent under
DµUp2qpΦqpδΦq “ i
¨
˚
˝
2Re 〈φ, δφ〉L2 ´ 〈φ, δψ〉L2 ´ 〈δφ, ψ〉L2
´ 〈ψ, δφ〉L2 ´ 〈δψ, φ〉L2 2Re 〈ψ, δψ〉L2
˛
‹
‚
.
The first obvious choice is δΦ “ Φ, so that
DµUp2qpΦqpΦq “ i
¨
˚
˝
2 0
0 2
˛
‹
‚
“: A1.
To find the second tangent vector, we first observe that either
ż
Repφψq2 dt or
ż
Impφψq2 dt
must be positive. This follows since the assumptions |φ|2 ` |ψ|2 ” 1 and }φ}L2 “ }ψ}L2 “ 1
imply that |φ|2|ψ|2 ı 0, which in turn implies
0 ă
ż
|φ|2|ψ|2 dt
“
ż
|φψ|2 dt
“
ż
Repφψq2 `
ż
Impφψq2 dt.
119
Assume without loss of generality thatş
Repφψq2 dt ą 0 and consider the tangent vector
pδφ, δψq “ p´Repφψqψ,Repφψqφq P TΦLS3.
We have
Dµpφ, ψqp´Repφψqψ,Repφψqφq “ i
¨
˚
˝
´2ş
Repφψq2 dt 0
0 2ş
Repφψq2 dt
˛
‹
‚
“: A2.
Then A2 is linearly independent from A1, as the diagonal entries of A2 differ by a sign while
the diagonal entries of A1 are the same. In the case thatş
Impφψq2 dt ą 0, a similar choice
of tangent vector could have been made.
Thus it remains to find two more tangent vectors which map to complete a spanning set
for up2q. Since A1 and A2 are diagonal matrices, it suffices to show the set
t´ 〈φ, δψ〉L2 ´ 〈δφ, ψ〉L2 | pδφ, δψq P TΦLS3u.
of upper-right entries of DµUp2qpΦqpδΦq is two-dimensional (over R).
We proceed by checking two cases: either |ϕ|2 ” |ψ|2, or not. In the case that the
functions are not identically equal,
ż
|φ|2 ´ |ψ|2 dt “
ż
|φ|2 dt´
ż
|ψ|2 dt “ 0
implies that the function |φ|2´|ψ|2 must take both positive and negative values. Let I Ă S1
be an open subarc such that |ϕ0|2 ´ |ψ0|
2 ą 0 on I and let I 1 Ă I be a proper open subarc.
Let B : S1 Ñ Rě0 be a smooth bump function with supportpBq Ă I and B ” 1 on I 1. Now
we consider the tangent vector δΦ “ pB ¨ψ,´B ¨φq. The upper-right entry of DµUp2qpΦqpδΦq
120
is
〈φ,B ¨ φ〉L2 ´ 〈B ¨ ψ, ψ〉L2 “
ż
B ¨ p|φ|2 ´ |ψ|2q dt.
This is a positive real number; i.e.,
ż
B ¨ p|φ|2 ´ |ψ|2q dt ě
ż
I 1|φ|2 ´ |ψ|2 dt ą 0.
Finally we note that Lemma 4.3.4, implies that there exists a tangent vector δΦ such that
the upper-right entry of DµUp2qpΦqpδΦq has non-zero imaginary part, and this completes the
proof of the claim for this case.
Now assume that |φ|2 ” |ψ|2. As |φ|2 ` |ψ|2 ” 1, this implies that |φ|2 ” |ψ|2 ”?
2{2
and we write
φ “
?2
2eiθ1 and ψ “
?2
2eiθ2
for some smooth eiθ1 , eiθ2 P LS1. Consider the tangent vector
δΦ “ pireiθ1 ,´ireiθ2q P TΦLS3,
where r P LRě0 is any smooth function. We have that the upper-right entry ofDµUp2qpΦqpδΦq
is?
2i
ż
reipθ1´θ2q dt.
Since
0 “
ż
φψ dt “1
2
ż
eipθ1´θ2q dt,
Lemma 4.3.3 implies that there exist functions B1, B2 : S1 Ñ Rě0 such that choosing r “ B1
and r “ B2 for the tangent vector above will yield linearly independent elements of R2 after
applying DµUp2qpΦq. This completes the proof that LS3 X St2pLCq is a Frechet manifold.
121
We immediately have a description of the tangent spaces of LS3 X St2pLCq. These are
given by
TΦLS3X St2pLCq “ tδΦ P TΦLS3
| δΦ P kerDµUp2qpΦqu
“ tδΦ P TΦSt2pLCq | Re 〈Φ, δΦ〉C2 ” 0u.
It remains to show that
µ´1p1q “ pLS3
X St2pLCqq{Up2q
is a Frechet manifold and to describe its tangent spaces. This follows easily from the principal
bundle structure
St2pLCq Ñ St2pLCq{Up2q « Gr2pLCq.
Indeed, any choices of slice charts for this principal bundle will restrict to slice charts for
LS3X St2pLCq Ñ pLS3
X St2pLCqq{Up2q « µ´1p1q.
Going through the proof of Corollary 4.1.2, one sees that the tangent spaces to µ´1p1q are
described by
TrΦsµ´1p1q « tδΦ P T horΦ St2pLCq | Re 〈Φ, δΦ〉C2 ” 0u.
Finally, the fact that µ´1p1q is LS1{S1-invariant is obvious from the definition of the
action on Gr2pLCq.
122
4.3.3 Proof of the Main Result
Now our goal is to show that µ´1p1q{pLS1{S1q may be identified with MMZ , that it has
a natural symplectic structure, and that this structure agrees with the one constructed by
Millson and Zombro. We first give a more careful description of the tangent spaces to µ´1p1q.
Throughout the following lemmas, we consider µ´1p1q as a Riemannian manifold with metric
defined by restricting the metric gL2
of Gr2pLCq. Similarly, the intersection LS3 X St2pLCq
inherits a Riemannian metric, and with these metrics
LS3X St2pLCq Ñ pLS3
X St2pLCqq{Up2q “ µ´1p1q
is a Riemannian submersion.
Lemma 4.3.5. The subspace of the tangent space to µ´1p1q that is horizontal with respect
to the LS1{S1-action is
T horrΦs µ´1p1q “
δΦ P T horΦ St2pLCq | 〈Φ, δΦ〉C2 ” 0(
.
The tangent space splits orthogonally (with respect to Re 〈¨, ¨〉PC2) as
TrΦsµ´1p1q “ T horrΦs µ
´1p1q ‘ T vertrΦs µ
´1p1q,
where T vertrΦs µ
´1p1q consists of tangent vectors along the LS1{S1-orbits through rΦs.
Proof. We begin by working in LS3XSt2pLCq, which we will denote by F for the remainder
of this proof. Fix Φ P F . The vertical space T vertΦ F consists of tangent vectors of the form
iαΦ P TΦF , where α P LR. Thus the horizontal subspace with respect to the LS1{S1-action
is the set
T horΦ F :“ tδΦ P TΦF | Re 〈iαΦ, δΦ〉L2 “ 0 for all α P LRu.
123
Examining the defining condition more closely, we see
Re 〈iαΦ, δΦ〉L2 “ ´Im 〈αΦ, δΦ〉L2
“ ´Im
ż 2
0
α 〈Φ, δΦ〉C2 dt
“ ´
ż 2
0
αIm 〈Φ, δΦ〉C2 dt
is required to be zero for all α P LR. This is only possible if Im 〈Φ, δΦ〉C2 ” 0. On the other
hand, combining this with the description of TΦP obtained in the proof of Proposition 4.3.2,
we obtain
T horΦ P “ tδΦ P TΦSt2pLCq | 〈Φ, δΦ〉C2 ” 0u.
Projecting via the submersion P Ñ P{Up2q “ µ´1p1q, we have
T horrΦs µ´1p1q “ tδΦ P T horΦ St2pLCq | 〈Φ, δΦ〉C2 ” 0u.
Thus we have proved the first claim.
We now explicitly show that the tangent space splits orthogonally. We once again begin
by working with P . For an arbitrary δΦ P TΦP , we write
δΦ “ pδΦ´ 〈Φ, δΦ〉C2 Φq ` p〈Φ, δΦ〉C2 Φq .
Then the second term in this decomposition is an element of the vertical space. Indeed,
〈Φ, δΦ〉C2 is purely imaginary by the assumption that δΦ P TΦP , thus
〈Φ, δΦ〉Φ “ iαΦ
124
for α “ ´i 〈Φ, δΦ〉C2 P LR. It remains to show that the first term lies in the horizontal
space. Indeed,
〈Φ, δΦ´ 〈Φ, δΦ〉C2 Φ〉C2 “ 〈Φ, δΦ〉C2 ´ 〈Φ, δΦ〉C2 }Φ}2C2 “ 0.
Finally, we note that this property is preserved by the Riemannian submersion P Ñ
µ´1p1q.
Lemma 4.3.6. The restriction of the symplectic form ωL2
to µ´1p1q degenerates exactly
along directions tangent to LS1{S1-orbits.
Proof. Let Xrαs|rΦs denote an element of T vertrΦs µ
´1p1q. Then for any δΦ P TrΦsµ´1p1q, Lemma
4.3.1 implies
ωL2
pδΦ, Xrαs|rΦsq “
ż
Re 〈Φ, δΦ〉C2 dt “ 0.
Thus the symplectic form degenerates along LS1{S1-orbits (i.e. the vertical tangent spaces).
By the orthogonal splitting given in Lemma 4.3.5, it suffices to show that the symplectic
form is non-degenerate when restricted to T horΦ µ´1p1q. Indeed, this follows immediately since
the horizontal subspace is closed under complex multiplication.
Finally we come to the main result of this section.
Theorem 4.3.7. The “forget framing map”
µ´1p1q ÑMMZ
rγ, V s ÞÑ rγs
gives µ´1p1q the structure of a principal bundle over MMZ with fibers diffeomorphic to
LS1{S1. Thus
Gr2pLCq � pLS1{S1q “ µ´1
p1q{pLS1{S1q «MMZ
125
inherits a symplectic structure, and this structure is isomorphic up to a constant factor to
the one constructed by Millson and Zombro.
Proof. Going through the proof of the first principal bundle theorem, Theorem 4.2.5, we see
that it can be directly adapted to treat the first claim. Thus Gr2pLCq� pLS1{S1q andMMZ
are diffeomorphic as Frechet manifolds.
We endow Gr2pLCq � pLS1{S1q with a symplectic structure by restricting ωL2
to the
horizontal tangent spaces of µ´1p1q, where it was shown in Lemma 4.3.6 to be nondegenerate.
We likewise endow Gr2pLCq � pLS1{S1q with a complex structure.
To see that the induced structures agree with those of Millson-Zombro, we recall from
Section 4.1.3 that the admissible variations of rγ, V s PM can be written pointwise as linear
combinations of the basic variations Xstretch, Xtwist, Xbend1 and Xbend2 . From Lemma 4.1.7,
we see that variations which are tangent to M � pLS1{S1q « Gr2pLCq � pLS1{S1q can no
longer have any Xstretch or Xtwist component. Thus elements pδγ, δV q of the tangent space
toM� pLS1{S1q at rγ, V s must satisfy 〈δγ, γ1〉 ” 0 and 〈δV,W 〉 ” 0. Moreover, the induced
complex structure takes Xbend1 to Xbend2 . This can be succinctly rewritten as
J ¨ δγ “ T ˆ δγ,
and this agrees with the complex structure constructed in [55]. Moreover, the induced
Riemannian metric reduces to
g1,1,1,1ppδγ1, δV1q, pδγ2, δV2qq “
1
4
ż 2
0
⟨d
dsδγ1,
d
dsδγ2
⟩ds “
1
4
ż 2
0
〈δγ11, δγ12〉 dt
since 〈δV,W 〉 ” 0 for any admissible variation. This metric agrees with the Millson-Zombro
metric up to a constant. Thus the symplectic structure agrees up to a constant as well.
126
Remark 4.3.8. The complexification of LS1{S1 is LC˚{C˚. This space is a Frechet Lie
group with an obvious action on Gr2pLCq that extends the action of LS1{S1. Using the
terminology of GIT quotients [73], the stable submanifold of Gr2pLCq with respect to the
action of LS1{S1 is the set of points rΦs such that there exists an element rreiαs P LC˚{C˚
with
rreiαs ¨ rΦs P µ´1´”
~1ı¯
.
It is easy to see that the stable submanifold is Gr˝2pVq, justifying our choice of terminology.
4.3.4 Discussion of the Main Result
It is interesting to more thoroughly compare the approach given here with the original
approach of Millson and Zombro. They showed that MMZ is realized as the symplectic
reduction of LS2 by the rotation action of SOp3q. Note that LS2 is the base space of the
principal bundle
LS1 ãÑ LS3Ñ LS2
obtained by applying the loop functor to the Hopf fibration. The action of LS1 on LS3 which
appears in this fibration is a restriction of the same LS1-action on LC2 that we have been
considering in this section. Moreover, LS3 can be seen as a level set of the momentum map
of LS1.
Proposition 4.3.9. The momentum map for the natural action of LS1 on LC2 is
LC2Ñ LR
Φ “ pφ, ψq ÞÑ |φ|2 ` |ψ|2.
127
The loop space LS3 is the level set of this map for the loop constantly taking the value 1. We
conclude that
LS2“ LC2 � LS1.
Proof. The first statement can be proved by following the proof of Lemma 4.3.1 almost
exactly. The second and third statements are obvious.
Thus we have the following commutative diagram which summarizes the two approaches
to obtaining the symplectic structure of MMZ .
LC2
St2pLCq LS3
Gr2pLCq LS2
µ´1Up2qp1q µ´1
SOp3qp0q
MMZ
�LS1�Up2q{Up2q {LS1
�pLS1{S1q �SOp3q
{pLS1{S1q {SOp3q
We have thus demonstrated an infinite-dimensional Gel’fand-Macphereson correpson-
dence: the space LC2 has Hamiltonian actions by the groups Up2q and LS1, and symplecti-
cally reducing in either order produces the same space.
The content of this section is expressed informally in terms of framed paths and loops in
the next diagram. Every space in the diagram is also up to translations and the left side is
modulo scaling.
128
frame-periodicframed paths
framed loopsarclength param.f-p framed paths
relativelyframed loopsmod rotations
arclength param.tangent-periodicunframed paths
arclength param.rel. framed loops
mod rotations
arclength param.unframed loops
arclength param.unframed loopsmod rotations
arclength param.
�LS1�Up2q
closure
mod rotations& rel. framing mod framings
�pLS1{S1q �SOp3qarclength param.
mod framings
closure
mod rotations
Finally, we remark here that a similarly pleasing picture holds in the finite-dimensional
world of moduli spaces of linkages. For fixed ~r “ pr1, . . . , rnq P Rně0, let
Pol~r :“ tpolygons in R3 with edgelengths determined by ~ru{ttransl., rot., scal.u.
Kapovich and Millson showed that (for generic ~r) this space is a symplectic manifold, ob-
tained as a symplectic reduction of pS2qn by the diagonal action of SOp3q [37]. Hausmann
and Knutson showed that Pol~r is obtained as a symplectic reduction of Gr2pCnq [30]—this
viewpoint has subsequently been used quite successfully in the theory of sampling off-lattice
random walks (e.g., [10, 11]). It was shown by Howard, Manon and Millson that the full
129
Grassmannian Gr2pCnq also has a polygon-theoretic interpretation, similar to the interpre-
tations of the Grassmannians Gr2pLCq and Gr2pACq given here [32]. It will be interesting to
study whether the bending flows introduced in [37] for polygon spaces have a smooth analog.
This will be the subject of future work.
4.4 The actions of Diff`pr0, 2sq and Diff`pS1q
Let Diff`pΣq denote the group of orientation-preserving diffeomorphisms of Σ, for Σ “ r0, 2s
or S1. This is a Frechet Lie group which acts by reparameterization on the space of framed
paths when Σ “ r0, 2s and the space of framed loops when Σ “ S1. More precisely, for
ρ P Diff`pΣq, the action on a framed curve pγ, V q is defined by
pρ ¨ pγ, V qqptq “ pγpρptqq, V pρptqqq.
The action of Diff`pr0, 2sq descends to a well-defined action on S, and the action of a subgroup
of Diff`pS1q descends to M (see Section 4.4.1 for details).
We are interested in the quotient spaces with respect to these group actions. For example,
the quotient space S{Diff`pr0, 2sq can be considered as the space of unparameterized framed
paths, or the space of shapes of framed paths. For some shape recognition applications, one
wants to match unparameterized framed paths and this can be formulated as determining
geodesic distance in the quotient space. See Chapter 6 for a more thorough explanation of
the use of these spaces in the shape recognition setting.
130
4.4.1 Manifold Structures of the Quotient Spaces
Our first goal for this section is to show that the spaces
S{Diff`pr0, 2sq and M{Diff`pS1q
are smooth Frechet manifolds. We will see below that there is a technical issue in the case
of the moduli space of framed loops. However, the claim does hold for framed path space.
Proposition 4.4.1. The space S{Diff`pr0, 2sq admits the structure of a Frechet manifold.
Proof. The action of Diff`pr0, 2sq is free on S. We can take a global slice of the action by
identifying
Sarc :“ tpγ, V q P S | }γ1} ” 1u « S{Diff`pr0, 2sq
via the bijection which takes pγ, V q to its Diff`pr0, 2sq-equivalence class. This map has
an inverse obtained by taking the Diff`pr0, 2sq-equivalence class of a framed loop pγ, V q
to pγ ˝ ρ, V ˝ ρq where ρ P Diff`pr0, 2sq is the unique reparameterization so that γ ˝ ρ is
arclength-parameterized.
It follows from Lemma 3.1.7 that Sarc is double-covered via pH by PS3, which is a Frechet
manifold. Thus
Sarc « PS3{pΦ „ ´Φq,
and we claim that the latter space has a manifold structure. Indeed, this amounts to choosing
smooth cross-sections to the Z2-action Φ ÞÑ ´Φ and this can be accomplished near Φ by using
the usual manifold charts for PS3 chosen small enough to not contain antipodal points.
We now turn to the Diff`pS1q action onM, where we immediately run into two problems.
The first is that with our usual identification,
M « tframed loops pγ, V q | γp0q “ 0, lengthpγq “ 2u{trelative framing, rotationu, (4.9)
131
the natural action is not well-defined. That is, the reparameterization γ ˝ ρ does not neces-
sarily satisfy γ ˝ρp0q “ 0. This problem may be remedied by instead considering the induced
action on the image of M under the map
MÑ LpSOp3q ˆ R`q{pSOp3q ˆ R` ˆ S1q
rγ, V s ÞÑ rpT, V, T ˆW q, }γ1}s .
In the above, the circle factor in the quotient on the right corresponds to relative framing.
The induced action is
ρ ¨ rpT, V, T ˆW q, }γ1}s “ rρ1pT ˝ ρ, V ˝ ρ, T ˝ ρˆW ˝ ρq, ρ1}γ1 ˝ ρ}s .
The above approach leads to a much more serious problem since the action is not free.
As a simple example, take
pγ, V q “ ppcospπtq, sinpπtq, 0q, p0, 0, 1qq
and
ρptq “ t` θ mod 2
for any θ P p0, 2q. Then
rpT, V, T ˆW q, }γ1}s “
»
—
—
—
—
–
¨
˚
˚
˚
˚
˝
´ sinpπtq 0 cospπtq
cospπtq 0 sinpπtq
0 1 0
˛
‹
‹
‹
‹
‚
, 1
fi
ffi
ffi
ffi
ffi
fl
(4.10)
132
and
ρ ¨ rpT, V, T ˆW q, }γ1}s “
»
—
—
—
—
–
¨
˚
˚
˚
˚
˝
´ sinpπpt` θqq 0 cospπpt` θqq
cospπpt` θqq 0 sinpπpt` θqq
0 1 0
˛
‹
‹
‹
‹
‚
, 1
fi
ffi
ffi
ffi
ffi
fl
. (4.11)
After applying trigonometric identities, we see that the matrix in (4.11) differs from the one
in (4.10) by right multiplication by
¨
˚
˚
˚
˚
˝
cospπθq 0 sinpπθq
0 1 0
´ sinpπθq 0 cospπθq
˛
‹
‹
‹
‹
‚
P SOp3q.
Thus our example is stablilized in (the image of) M by the Diff`pS1q-action.
To circumvent each of these problems, we restrict the action to the reparameterization
action by the group of based diffeomorphisms
Diff`0 pS1q :“ tρ P Diff`pS1
q | ρp0q “ 0u.
Remark 4.4.2. Diffeomorphisms of S1 of the form
t ÞÑ t` θ mod 2 θ P r0, 2q
form a subgroup of Diff`pS1q which is obviously diffeomorphic to S1 and we can identify the
based diffeomorphism group with the quotient
Diff`0 pS1q « Diff`pS1
q{S1.
133
Clearly the action by the subgroup S1 on M is not free. We will see later in Theorem 5.2.8
that the points stabilized by this S1 subgroup action are exactly the critical points of a certain
natural functional on M.
The natural action of Diff`0 pS1q is well-defined on the usual model ofM (4.9). Moreover,
the action is now clearly free. We are easily able to show that the quotient space is a manifold
based on previous work.
Proposition 4.4.3. The spaceM{Diff`0 pS1q has the structure of a smooth Frechet manifold.
Proof. We concretely represent M as
M « tpγ, V q | γp0q “ 0, lengthpγq “ 2u{trelative framing, rotationu.
By Proposition 4.3.2, the space
Marc :“ tframed loops pγ, V q | γp0q “ 0, }γ1} ” 1u{trelative framing, rotationu
is a smooth Frechet manifold. We claim that M{Diff`0 pS1q can be identified with Marc.
Indeed, we define a bijection from Marc to M{Diff`0 pS1q by
rγ, V s ÞÑ rrγ, V ss,
where the double brackets on the right denote the Diff`0 pS1q orbit of rγ, V s. The inverse of
this map is
rrγ, V ss ÞÑ rγ ˝ ρ, V ˝ ρs,
where ρ is the unique based reparameterization such that γ˝ρ is arclength parameterized.
134
We note that all Riemannian metrics defined in the previous chapter are invariant un-
der the Diff`-actions—indeed, this was by construction! Thus the metrics descend to the
quotient spaces.
4.4.2 Splitting the Tangent Space
It will be useful to decompose the tangent space ofM into the subspace tangent to Diff`0 pS1q-
orbits (the vertical tangent space) and its orthogonal complement (the horizontal tangent
space). For example, we will use this decomposition in Section 4.5.4 to study the sectional
curvatures of the quotient M{Diff`0 pS1q. This task is nontrivial since M and Diff`0 pS
1q are
both infinite-dimensional. We will first decompose the tangent spaces of the simpler space
ĂM :“ tframed loops pγ, V q | γp0q “ 0u.
Since the moduli space of interest is obtained from ĂM by quotienting by a finite-dimensional
group action, a decomposition of the tangent spaces to ĂM will easily yield a decomposition
for M.
By arguments similar to those of the previous chapter, ĂM is a tame Frechet manifold
(this can be seen by, e.g., viewing ĂM as a codimension-3 submanifold of LpSOp3q ˆ R`q)
with tangent spaces
Tpγ,V q ĂM “ tpδγ, δV q P LpR3ˆ R3
q | 〈δV, V 〉 “ 〈δγ1, V 〉` 〈γ1, δV 〉 ” 0, δγp0q “ 0u.
The diffeomorphism group Diff`0 pS1q acts on ĂM by reparameterization. The framed curve
elastic metric g1,1,1,1 is well-defined on ĂM, and going through the proof of Lemma 3.1.4, one
135
sees that it reduces to
g1,1,1,1pγ,V q ppδγ1, δV1q, pδγ2, δV2qq “
1
2`pγq
ż 2
0
⟨d
dsδγ1,
d
dsδγ2
⟩` 〈δV1,W 〉 〈δV2,W 〉 ds,
where `pγq is the length of γ.
We are now prepared to state the splitting result for ĂM. We use the notation
L0R :“ tξ P LR | ξp0q “ 0u
for the based loop space of R.
Proposition 4.4.4. The subspace of Tpγ,V q ĂM which is tangent to the Diff`0 pS1q-orbits is
T vertpγ,V qĂM :“ tpξT, ξp´κ1T ` twW qq | ξ P L0Ru,
where T is the unit tangent to γ, W “ T ˆ V , κ1 is the first Darboux curvature of pγ, V q
and tw is the twist rate of pγ, V q.
The orthogonal complement with respect to g1,1,1,1pγ,V q is
T horpγ,V qĂM :“
"
pδγ, δV q P Tpγ,V q ĂM |
⟨d2
ds2δγ, T
⟩´ 〈δV,W 〉 tw ” 0
*
.
The tangent space Tpγ,V q ĂM splits as
Tpγ,V q ĂM “ T vertpγ,V qĂM‘ T horpγ,V q
ĂM.
The proof of the proposition requires a technical lemma.
Lemma 4.4.5. . Let pγ, V q be a framed loop, let d{ds denote derivative with respect to
arclength of γ, let κ and tw denote the curvature and twist rate of γ, respectively. The
136
operator L defined by
L : L0RÑ LR
ξ ÞÑd2
ds2ξ ´ pκ2
` tw2q ¨ ξ
is invertible.
Proof. We first show that L is injective. This follows because the unique solution to the
second-order linear ODE
d2
ds2ξ ´ pκ2
` tw2q ¨ ξ “ 0
with boundary conditions ξp0q “ ξp2q “ 0 is ξ “ 0.
Now we wish to show that L is onto. We first note that L is a degree-2 elliptic operator.
Indeed,
Lpξq “d2
ds2ξ ´ pκ2
` tw2qξ
“1
}γ1}
d
dt
1
}γ1}
d
dtξ ´ pκ2
` tw2qξ
“1
}γ1}2ξ2 ´
〈γ2, γ1〉}γ1}4
ξ1 ´ pκ2` tw2
qξ.
Following an argument in the proof of Lemma 4.5 in [53], we claim that the index of L,
indexpLq “ dim kerL´ dim cokerL,
is zero. Since L is injective, this implies that its cokernel is zero, hence L is onto. This
follows because L is homotopic to the invertible operator
ξ ÞÑ ξ2 ´ ξ, (4.12)
137
through perturbations by elliptic operators and index is invariant under such homotopies.
The index of (4.12) is zero (see [53]), thus the proof is complete.
We now proceed with the proof of Proposition 4.4.4.
Proof of Proposition 4.4.4. Let ρε be a path in Diff`0 pS1q with parameter ε such that ρ0 is
the identity and let ξ “ ddε
ˇ
ˇ
ε“0ρε so that
ξ P L0R,
which is the Lie algebra of Diff`0 pS1q (strictly speaking, the Lie algebra of Diff`0 pS
1q is the
space of vector fields on S1 with the zero vector at 0, but this is identified with LR by
comparing to the unit vector field). Then tangent vectors along the Diff`0 pS1q-orbit of pγ, V q
take the form
d
dε
ˇ
ˇ
ˇ
ˇ
ε“0
pγpρεq, V pρεqq “ pγ1pρ0qξ, V
1pρ0qξq
“
ˆ
ξ}γ1}d
dsγ, ξ}γ1}
d
dsV
˙
“ pξ}γ1}T, ξ}γ1}p´κ1T ` twW qq ,
thus
T vertpγ,V qĂM “ tpξ}γ1}T, ξ}γ1}p´κ1T ` twW qq | ξ P L0Ru “ tpξT, ξp´κ1T ` twW qq | ξ P L0Ru.
To determine the horizontal subspace, we note that pδγ, δV q is horizontal if and only if
g1,1,1,1pγ,V q ppδγ, δV q, pξT, ξp´κ1T ` twW qqq “ 0
138
for all ξ P L0R. Writing this out, this means that for all ξ we have
0 “1
2`pγq
ż 2
0
⟨d
dsδγ,
d
dsξT
⟩` 〈δV,W 〉 〈ξp´κ1T ` twW q,W 〉 ds
“1
2`pγq
ż 2
0
⟨d
dsδγ,
ˆ
d
dsξ
˙
T ` ξpκ1V ` κ2W q
⟩` 〈δV,W 〉 ξtw ds
“1
2`pγq
ż 2
0
d
dsξ
⟨d
dsδγ, T
⟩` ξ
ˆ⟨d
dsδγ, κ1V ` κ2W
⟩` 〈δV,W 〉 tw
˙
ds
“1
2`pγq
ż 2
0
ξ
ˆ
´
⟨d2
ds2δγ, T
⟩` 〈δV,W 〉 tw
˙
ds,
where we have integrated by parts to obtain the last line. This condition holds for every
ξ P L0R if and only if ⟨d2
ds2δγ, T
⟩´ 〈δV,W 〉 tw ” 0,
and we have established the claimed characterization of the horizontal tangent space.
Finally, we wish to show that the tangent space splits. Assuming that we have such a
projection, we can express an arbitrary tangent vector as
pδγ, δV q “ pξT, ξp´κ1V ` twW qq ` pδγhor, δV horq (4.13)
where
pδγhor, δV horq P T horpγ,V q
ĂM
and ξ P L0R. From (4.13) we conclude
⟨d2
ds2δγ, T
⟩“
d2
ds2ξ ´ ξκ2
`⟨δγhor, T
⟩, (4.14)
where κ2 “ κ21 ` κ
22 is the squared curvature of γ. On the other hand,
〈δV,W 〉 “ ξtw `⟨δV hor,W
⟩.
139
Multiplying this expression by tw and subtracting the result from (4.14) yields
⟨d2
ds2δγ, T
⟩´ 〈δV,W 〉 tw “ d2
ds2ξ ´ ξκ2
`⟨δγhor, T
⟩´ ξtw2
´⟨δV hor,W
⟩tw.
Utilizing the horizontality of pδγhor, δV horq, this reduces to
⟨d2
ds2δγ, T
⟩´ 〈δV,W 〉 tw “ d2
ds2ξ ´ pκ2
` tw2qξ “ Lpξq,
where L is the invertible linear operator from Lemma 4.4.5. Therefore we define
ξ :“ L´1
ˆ⟨d2
ds2δγ, T
⟩´ 〈δV,W 〉 tw
˙
.
This gives us a well-defined projection onto T vertpγ,V q
ĂM, thus proving that the tangent spaces
split.
Corollary 4.4.6. There is a well-defined projection
Trγ,V sMÑ T vertrγ,V sM,
where T vertrγ,V sM is the space of tangent vectors to Diff`0 pS
1q-orbits.
Proof. Since M “ ĂM{pSOp3q ˆ R`q is the image of a submersion with finite-dimensional
fibers, we can identify Trγ,V sM with a finite codimension subspace of Tpγ,V q ĂM. Thus we can
first project onto T vertpγ,V q
ĂM using Proposition 4.4.4, then project onto the finite codimension
subspace.
140
4.4.3 Diff`-Actions in Complex Coordinates
We conclude this section with a description of the Diff`-actions in complex coordinates. We
will once again focus on the case of loops, so that we will work in the Stiefel and Grassmann
manifolds. The case of framed paths is similar.
Lemma 4.4.7. The action of Diff`0 pS1q on St2pVq corresponding to the one described above
is
ρ ¨ Φ :“a
ρ1Φpρq.
This induces a well-defined action on Gr2pVq.
Proof. Let pHpΦq “ pγ, V q. We wish to show that?ρ1Φpρq maps under pH to pγpρq, V pρqq.
Since we are working up to translations, it suffices to show that this holds at the level of
LpSOp3q ˆ R`q; i.e. that
Hpa
ρ1Φpρqq “ pρ1γ1pρq, V pρqq
This is easy to see using quaternionic coordinates Φ Ø q. We have
Hpa
ρ1qpρqq “
˜
Impa
ρ1qpρqia
ρ1qpρqq,Imp
?ρ1qpρqj
?ρ1qpρqq
}?ρ1qpρq}2H
¸
“
˜
Impqpρqiqpρqqρ1,Impqpρqjqpρqq
}qpρq}2H
¸
“ pρ1γ1pρq, V pρqq.
The second claim follows by noting that if Φ0 and Φ1 differ by right multiplication by an
element of Up2q then ρ ¨ Φ0 and ρ ¨ Φ1 do as well, so that rρ ¨ Φs is well-defined.
We now describe the horizontal and vertical tangent spaces in these coordinates. Since
we have already used T horSt2pVq to denote the horizontal space with respect to the Up2q-
141
action, we will use T hor1
St2pVq and T vert1
St2pVq to denote the horizontal and vertical spaces
with respect to the Diff`0 pS1q-action.
Corollary 4.4.8. The tangent space of St2pVq decomposes as
TΦSt2pVq “ T hor1
Φ St2pVq ‘ T vert1
Φ St2pVq,
where T vert1
Φ St2pVq is the space of tangents to Diff`0 pS1q orbits, described explicitly as
T vert1
Φ St2pVq :“ tp1{2qξ1Φ` ξΦ1 | ξ P LRu,
and T hor1
Φ St2pVq is its orthogonal complement with respect to gL2, described explicitly as
T hor1
Φ St2pVq :“ tδΦ P TΦSt2pVq | Rep〈δΦ,Φ1〉C2 ´ 〈δΦ1,Φ〉C2q ” 0u.
Similarly,
TrΦsGr2pVq “ T hor1
rΦs Gr2pVq ‘ T vert1
rΦs Gr2pVq,
where T hor1
rΦs Gr2pVq (respectively, T vert1
rΦs Gr2pVq) is the orthogonal projection of T hor1
Φ St2pVq
(respectively, T vert1
Φ St2pVq) to TrΦsGr2pVq.
Proof. The statements that the tangent spaces split follows from Proposition 4.4.4, thus we
only need to show that the vertical and horizontal spaces are as described. Let ρε be a path
in Diff`0 pS1q such that ρ0 is the identity, and let ξ denote d
dε
ˇ
ˇ
ε“0ρε. Then
d
dε
ˇ
ˇ
ˇ
ˇ
ε“0
?ρεΦpρεq “
1
2?ρ0
ξΦpρ0q `?ρ0Φ1pρ0qξ “
1
2ξΦ` ξΦ1.
This proves that the vertical space is as described.
142
A tangent vector δΦ “ pδφ, δψq lies in the horizontal space if and only if
0 “ gL2
ˆ
δΦ,1
2ξ1Φ` ξΦ1
˙
“ Re
ż 2
0
⟨δΦ,
1
2ξ1Φ` ξΦ1
⟩C2
dt
“ Re
ż 2
0
´1
2ξ p〈δΦ1,Φ〉C2 ` 〈δΦ,Φ1〉C2q ` ξ 〈δΦ,Φ1〉C2 dt
“ Re
ż 2
0
1
2ξ p〈δΦ,Φ1〉C2 ´ 〈δΦ1,Φ〉C2q dt
holds for all ξ P LR. This is true if and only if
Re p〈δΦ,Φ1〉C2 ´ 〈δΦ1,Φ〉C2q ” 0.
The claims that the tangent spaces split follow as corollaries of Proposition 4.4.4.
4.5 Riemannian Geometry of Framed Path and Loop
Spaces
4.5.1 Explicit Geodesics in Framed Path Space
A major motivation for the framework developed in the previous chapter is for applications to
elastic shape matching. This is a vibrant subfield of computer vision where shapes are treated
as points in an infinite-dimensional “shape manifold”, and shape similarity is measured by
geodesic distance in the shape manifold. In the following sections, we describe the geodesics
of the shape spaces S andM. The details of their application to shape recognition are treated
later in Chapter 6. For a more general overview of the field of elastic shape matching, see,
e.g., the survey articles [69, 77] and the references therein.
143
The geodesics of
S “ tframed pathsu{ttranslation, scalingu
are particularly easy to describe by virtue of Theorem 3.1.12. We recall that the theorem
says that S is locally isometric to an L2 sphere—in particular, it is isometrically double
covered by
S˝pPC2q “ tΦ P PpC2
zt0uq | 〈Φ,Φ〉L2 “?
2u Ă SpPC2q,
which is an open subset of the radius-?
2 L2-sphere in PC2, denoted SpPC2q. It follows that
the geodesics of S are exactly the geodesics of the sphere SpPC2q. There is a technical issue
in that geodesics joining points of S˝pPC2q do not necessarily stay in S˝pPC2q. However,
geodesic distance in the full L2-sphere is still a well-defined, geometrically-motivated distance
measure for S. Moreover, it is the case that geodesics of SpPC2q between elements of S˝pPC2q
tend to stay in S˝pPC2q. Heuristically, this is because a geodesic of SpPC2q can be thought
of as special homotopies between paths in C2. Then the geodesic staying in S˝pPC2q means
that the special homotopy never passes through 0 P C2, and this is a highly generic condition.
As SpPC2q is a round sphere in a vector space, its geodesics are great circles. These are
parameterized explicitly as follows. Let Φ0,Φ1 P SpPC2q and define
θ “ θpΦ0,Φ1q “ arccos1
2
ż 2
0
Re 〈Φ0,Φ1〉C2 dt.
Then geodesic distance in SpPC2q is given by?
2θ and the geodesic segment connecting Φ0
to Φ1 in SpPC2q is
Φu “sinpp1´ uqθq
sinpθqΦ0 `
sinpuθq
sinpθqΦ1.
144
4.5.2 Explicit Geodesics in Framed Loop Space
Now we turn to the slightly more involved task of explicitly describing the geodesics of M
with respect to its natural metric. By Theorem 3.2.8, these are the geodesics of Gr2pVq for
V “ LC2 orAC2. We once again run into the technical issue here thatM is actually isometric
to the disjoint union of stable Grassmannians, so that geodesics in the Grassmanians can
potentially correspond to paths in framed loop space which pass through singular framed
loops. We once again note that geodesic distance in the full Grassmannian gives a well-
defined metric onM and, by the same heuristic argument of the previous section, geodesics
between points of the stable Grassmannian tend to stay in the stable Grassmannian.
The description of the geodesics in Gr2pVq given here can be traced back to Neretin [59],
who gave a description of the geodesics of the finite-dimensional Grassmannians GrkpRnq
which is amenable to adaptation to the infinite-dimensional setting. The Neretin geodesics
were already adapted to describe the geodesics in the space of immersed plane curves in [76].
Let rΦ0s, rΦ1s P Gr2pVq—recall that these equivalence classes represent complex 2-planes
in V , where we have chosen particular orthonormal bases Φj for each subspace. The basic
description of the geodesic rΦus joining these points is to find the optimal bases for each
2-plane and then to simply interpolate. This process is described completely explicitly by
the following steps.
1. Write the projection map π : rΦ0s Ñ rΦ1s as a 2 ˆ 2 matrix in terms of the chosen
orthonormal bases.
2. Compute the singular value decomposition of the matrix π. This produces new or-
thonormal bases rΦ0 “
´
φ0, ψ0
¯
for rΦ0s and rΦ1 “
´
φ1, ψ1
¯
for rΦ1s such that πpφ0q “
λφφ1 and πpψ0q “ λψψ1, where 0 ď λφ, λψ ď 1.
3. Let θφ “ arccospλφq and θψ “ arccospλψq. These are the Jordan angles of rΦ0s and
rΦ1s.
145
Figure 4.3: Geodesic in M joining a p2, 3q-torus knot to a p2, 5q-torus knot.
4. If θφ, θψ ‰ 0, then the geodesic joining the subspaces is given by rΦus, where Φu “
pφu, ψuq is described by the formulas
φuptq “sinpp1´ uqθφqφ0ptq ` sinpuθφqφ1ptq
sin θφ
ψuptq “sinpp1´ uqθψqψ0ptq ` sinpuθψqψ1ptq
sin θψ.
If θφ “ 0 (i.e., πpφ0q “ φ1, which implies φ0 “ φ1), then the geodesic is simply given
by φuptq “ φ0 and likewise for ψu.
We also note that the geodesic distance between rΦ0s and rΦ1s is given explicitly by
dist “b
θ2φ ` θ
2ψ.
An example geodesic is shown in Figure 4.5.2.
146
This description of the geodesics of Gr2pVq has a useful corollary. For any isometric
embedding L : Cn Ñ V (with respect to the standard Hermitian metric on Cn and the
Hermitian L2-metric 〈¨, ¨〉L2 on V), we obtain an induced map
L˚ : Gr2pCnq Ñ Gr2pVq
spantw1, w2u ÞÑ spantLpw1q, Lpw2qu
for any linearly independent pair w1, w2 P Cn. In other words, Gr2pCnq embeds as Gr2pimagepLqq.
The following corollary shows that this is a totally geodesic embedding and follows essen-
tially immediately from our description of geodesics above. We note that a similar result
was shown to hold for Stiefel manifolds of real Hilbert spaces in [29].
Corollary 4.5.1. Let L : Cn ãÑ V be an isometric embedding as described above. Then
the induced map L˚ : Gr2pCnq ãÑ Gr2pVq is an isometric embedding as a totally geodesic
submanifold with respect to the standard induced metric on Gr2pCnq and the induced metric
gL2
on Gr2pVq.
Proof. One can easily show that the induced map L˚ is an embedding by representing L
explicitly in terms of a choice of orthonormal basis. The embedding takes the horizontal
tangent spaces of St2pCnq to the tangent spaces to the embedded Grassmannian, so the
assumption that L was an isometry implies that that L˚ is an isometric embedding. Thus
it remains to show that Gr2pCnq embeds as a totally geodesic submanifold.
Let rΦ0s and rΦ1s be elements of the image of L˚ and let rΦus “ rφu, ψus denote the
geodesic joining them in Gr2pVq, as described above. With the interpretation that the
embedded Grassmannian is Gr2pLpCnqq with metric induced by Gr2pVq, we see that the
geodesic in L˚pGr2pCnqq is obtained by following the same procedure above. Thus the image
of Gr2pCnq is totally geodesic.
147
As a sanity check, we will explicitly show that φu and ψu each lie in the image of L for all
u. First note that φ0, ψ0, φ1 and ψ1 all lie in the image of L. Indeed, since rΦ0s P imagepL˚q,
φ0 and ψ0 lie in imagepLq, hence φ0 does as well, and the same argument holds for the other
vectors. Then φu lies in the image of L for all u, as φu P spantφ0, φ1u. Therefore rΦus lies in
the image of L˚ for all u.
We can take this line of thought further and show that every collection of planes is
contained in a totally geodesic copy of some finite-dimensional Grassmannian. This aligns
with our point of view that much of the geometry of Gr2pVq has a finite-dimensional flavor.
Proposition 4.5.2. Let rΦ1s, . . . , rΦks be distinct elements of Gr2pVq. There exists an em-
bedded finite-dimensional Grassmannian which contains every rΦjs.
Proof. Let V denote the linear subspace of LC of minimal dimension which contains all planes
rΦ1s, . . . , rΦks and let dimpV q “ n ď 2k. Choose an isometric embedding L : Cn Ñ LC such
that imagepLq “ V . Then L˚pGr2pCnqq is a totally geodesic Grassmannian which contains
every plane rΦjs.
In particular, it follows that every geodesic triangle of Gr2pVq is contained in a finite-
dimensional totally geodesic embedded Grassmannian. The heuristics of Toponogov’s The-
orem (see, e.g., [12] Theorem 2.2) suggest that the sectional curvatures of Gr2pCnq and
Gr2pVq should be comparable. We will use a similar idea to describe the curvature of Gr2pVq
in Section 4.5.4.
To implement this framework as an elastic shape matching algorithm, it remains to com-
pute distance in S{Diff`pr0, 2sq andM{Diff`pS1q—i.e., to optimize over reparameterization.
This is achieved via dynamic programming and is treated in detail in Chapter 6.
148
4.5.3 The Exponential Map
The aim of this section is to describe geodesics of St2pVq and Gr2pVq in terms of initial
position and velocity. We once again obtain explicit geodesics. This is accomplished by
determining the geodesic equation of St2pVq and showing that it is integrable with an explicit
solution. The derivation given here adapts the finite-dimensional real version found in [16].
The results at the end of the section are proved by methods similar to those in [29] for the
real Hilbert space case.
In this section, it will be useful to identify St2pVq with a space of linear maps C2ˆC2 Ñ
V2. We first consider Φ “ pφ, ψq P V2 as a linear map by using matrix multiplication on the
right:
pφ, ψq
¨
˚
˝
z1 w1
z2 w2
˛
‹
‚
“ pφz1 ` ψz2, φw1 ` ψw2q.
The formal adjoint of Φ to be the linear map Φ˚ : V2 Ñ C2 ˆ C2 defined by
Φ˚pφ1, ψ1q “
¨
˚
˝
〈φ1, φ〉L2 〈ψ1, φ〉L2
〈φ1, ψ〉L2 〈ψ1, ψ〉L2
˛
‹
‚
.
This adjoint is formal in the sense that we do not claim that any properties of the adjoint
in a Hilbert space carry over. However, it will be useful later to note that some of the usual
basic properties of an adjoint hold for this definition. In particular, for any
A “
¨
˚
˝
a b
c d
˛
‹
‚
P C2ˆ C2,
149
we have
pΦAq˚pφ1, ψ1q “ paφ` cψ, bφ` dψq˚pφ1, ψ1q
“
¨
˚
˝
〈φ1, aφ` cψ〉L2 〈ψ1, aφ` cψ〉L2
〈φ1, bφ` dψ〉L2 〈ψ1, bφ` dψ〉L2
˛
‹
‚
“
¨
˚
˝
a 〈φ1, φ〉L2 ` c 〈φ1, ψ〉L2 a 〈ψ1, φ〉L2 ` c 〈ψ1, ψ〉L2
b 〈φ1, φ〉L2 ` d 〈φ1, ψ〉L2 b 〈ψ1, φ〉L2 ` d 〈ψ1, ψ〉L2
˛
‹
‚
“ A˚Φ˚pφ1, ψ1q,
so that pΦAq˚ “ A˚Φ˚ as linear maps V2 Ñ C2 ˆ C2.
With this formalism, the Stiefel manifold can be expressed as
St2pVq “ tΦ P V2| Φ˚Φ “ Id2ˆ2u (4.15)
and the tangent space to St2pVq at Φ is
TΦSt2pVq “ tδΦ P V2| Φ˚δΦ P up2qu.
Indeed, this is a direct translation of the description of TΦSt2pVq given in Lemma 4.1.1. For
points Φ1 P St2pVq, Φ2 P V2 and A P C2 ˆ C2, it therefore holds that
Φ1A “ Φ2 ô A “ Φ˚1Φ2.
Now we note that since St2pVq is a Riemannian submanifold of V2 with a flat metric, the
geodesics of St2pVq are paths Φu such that B2
Bu2Φu lies in the normal direction to St2pVq for all
u. Thus it will be useful to describe the normal spaces of St2pVq explicitly. In the following
150
lemma, we also give a description of the normal space to the Up2q-horizontal tangent space
of St2pVq, as this will be useful later on.
Lemma 4.5.3. Let Φ P St2pVq. The normal space to TΦSt2pVq Ă TΦV2 has L2-orthonormal
basis
pφ, 0q, p0, ψq,1?
2pψ, φq,
1?
2p´iψ, iφq. (4.16)
Let rΦs P Gr2pVq. The normal space to TrΦsGr2pVq « T horΦ St2pVq Ă TΦSt2pVq has orthonor-
mal basis
piφ, 0q, p0, iψq,1?
2piψ, iφq,
1?
2pψ,´φq. (4.17)
Proof. One simply needs to check that the vectors in the first list are orthonormal and that
they are each orthogonal to TΦSt2pVq. This is sufficient, since St2pVq is codimension-4 inW .
We leave it to the reader to check the first point and note that the second point is also easy
by our characterization of the tangent spaces of the Stiefel manifolds given in Lemma 4.1.1.
For example, let pδφ, δψq P TΦSt2pVq. Then
Re 〈pφ, 0q, pδφ, δψq〉PC2 “ Re
ż 2
0
φδφ dt “ 0.
The remaining vectors are also easy to check.
To prove the second statement, we note that orthonormality follows by the first statement
as each vector in the second list is obtained by applying the complex structure to a vector in
the first list. Moreover, the fact that the Grassmannian is a Kahler reduction of the Stiefel
manifold implies that the vectors in the seond list are normal to the horizontal tangent space
of the Stiefel manifold. Alternatively, this can be proved by explicit compuatations using
the characterization of the tangent spaces to the Grassmannian given in Corollary 4.1.2.
151
It is straightforward to check that the normal space to TΦSt2pVq can then be written as
tδΦ P V2| Im 〈φ, δφ〉L2 “ Im 〈ψ, δψ〉L2 “ 〈φ, δψ〉L2 ´ 〈δφ, ψ〉L2 “ 0u. (4.18)
Indeed, this set is real codimension-4 in V2 and it contains each vector in the list (4.16).
Using the description (4.15) of St2pVq, the set (4.18) translates to
tδΦ P V2| Φ˚δΦ is Hermitianu.
We are now prepared to describe the geodesics of St2pVq. For a path Φu in St2pVq, we
will use dots to denote differentiation with respect to u; that is,
9Φu :“B
BuΦu.
Proposition 4.5.4. The geodesic equation for St2pVq with its L2 metric is
:Φu ` Φu9Φ˚u 9Φu “ 0.
Thus the geodesic Φu with Φ0 “ Φ and 9Φu “ δΦ, say with }δΦ}L2 “ 1, is given by
Φu “ pΦ, δΦq expu
¨
˚
˝
Φ˚δΦ ´δΦ˚δΦ
Id2ˆ2 Φ˚δΦ
˛
‹
‚
Id4ˆ2 expp´uΦ˚δΦq, (4.19)
where pΦ, δΦq is treated as a map C4 ˆ C4 Ñ V2. Therefore the exponential map of St2pVq
is well-defined.
152
Proof. As stated above, the geodesics of St2pVq are those paths Φu such that :Φu lies in the
normal space to Φu for all u. Taking two u-derivatives of the defining of St2pVq yields
B2
Bu2Φ˚uΦu “
B2
Bu2Id2ˆ2,
so that
:Φ˚uΦ` 2 9Φ˚u 9Φu ` Φ˚u:Φ “ 0. (4.20)
If :Φu lies the normal space to Φu, then
Φ˚u:Φu “ A
for some Hermitian 2ˆ 2 matrix A. We claim that this can be rewritten as
:Φu “ ΦuA. (4.21)
This holds because ΦuΦ˚u can be viewed as the linear map V2 Ñ V2 which linearly projects
each coordinate of an element of V2 to the span of φu and ψu in V . Since :Φu is normal to
the Stiefel manifold, the explicit desription of the normal spaces in Lemma 4.5.3 shows that
each coordinate of :Φu already lies in the span of φu and ψu. Thus ΦuΦ˚u
:Φu “ :Φu.
Replacing (4.21) into (4.20) and multiplying on the left by Φu, we have
0 “ ΦupΦuAq˚Φu ` 2Φu
9Φ˚u 9Φu ` ΦuΦ˚uΦuA
“ ΦuA˚Φ˚uΦu ` 2Φu
9Φ˚u 9Φu ` ΦuA
“ ΦuA` 2Φu9Φ˚u 9Φu ` ΦuA
“ 2:Φu ` 2Φu9Φ˚u 9Φu.
153
Now we wish to show that (4.19) is the solution to the geodesic equation with the given
initial conditions. Let Su “ 9Φ˚u 9Φu, interpreted as a path of 2ˆ 2 complex matrices. Then
9Su “ :Φu9Φu ` 9Φ˚u:Φu
“
´
´Φu9Φ˚u 9Φu
¯˚9Φu ´ 9Φ˚u
´
Φu9Φ˚u 9Φu
¯
“ ´SuΦ˚u
9Φu ´ 9Φ˚uΦuSu.
Now define Tu “ Φ˚u 9Φu. Then Tu is a skew symmetric matrix, since 9Φu is a tangent vector
to St2pVq. Moreover, it is straightforward to check that
T ˚u “9Φ˚uΦu.
Thus we have
9Su “ TuSu ´ SuTu. (4.22)
Next we note that
9Tu “ 9Φ˚u 9Φu ` Φ˚u:Φu
“ 9Φ˚u 9Φu ´ Φ˚uΦu9Φ˚u 9Φu
“ 9Φ˚u 9Φu ´ 9Φ˚u 9Φu
“ 0.
Thus Tu is constant and we let T :“ T0. Then (4.22) becomes the ODE
9Su “ TSu ´ SuT,
154
which has the unique solution
Su “ exppuT qS0 expp´uT q.
Therefore the geodesic equation becomes
:Φu “ ´Φu exppuT qS0 expp´uT q.
This can be rewritten as the integrable equation
d
du
´
Φu exppuT q, 9Φu exppuT q¯
“
´
Φu exppuT q, 9Φu exppuT q¯
¨
˚
˝
T ´S0
Id2ˆ2 T
˛
‹
‚
.
It is easy to see that (4.19) is the unique solution with the given initial conditions.
Lemma 4.5.5. For any path Φu in St2pVq, there exists a path Au in Up2q so that ΦuAu is
horizontal.
Proof. From the description of TrΦsGr2pVq « T horΦ St2pVq in Corollary 4.1.2, we rewrite the
horizontal space condition in the language of this section as
T horΦ St2pVq “ tδΦ P V2| Φ˚δΦ “ 0u.
Thus ΦuAu is horizontal if and only if
0 “ Φ˚B
BuΦuAu “ Φ˚ 9ΦuAu ` Φ˚Φu
9Au.
We then solve the problem by taking Au as a solution of the ODE 9Au “ ´Φ˚u 9ΦuAu.
155
Now we note that if a geodesic in St2pVq has initial data pΦ, δΦq with δΦ P T horΦ St2pVq,
then the geodesic stays horizontal. Otherwise it could be shortened by applying the pro-
jection of Lemma 4.5.5. We conclude that a geodesic in St2pVq with horizontal initial data
represents a geodesic in Gr2pVq. The next proposition follows immediately.
Proposition 4.5.6. The exponential map exp : TrΦsGr2pVq Ñ Gr2pVq is well-defined.
4.5.4 Sectional Curvatures
To get an idea of how the geodesics inM andM{Diff`0 pS1q behave, it is useful to characterize
the sectional curvatures of each space with respect to our metric; this is the content of our
next main theorem, which states that each space is nonnegatively curved. This result is
similar to Theorem 5.3 in [76], where the sectional curvature for the shape space of planar
loops is shown to be nonnegative. The proof strategy used here is different than that of [76].
The proof relies on O’Neill’s formula, which requires some setup to state. Let M be a
Riemannian manifold with a smooth free action by a Lie group G, so that π : M Ñ M{G
is a submersion. Then we note that any tangent vector to M{G at rps can be expressed as
DπppXq for some X P TpM . The tangent space TpM decomposes as
TpM “ T horp M ‘ T vertp M,
where T vertp M is the kernel of Dπp and T horp M is its orthogonal complement. For X P TpM ,
we denote its decomposition as
X “ Xhor`Xvert.
We are now prepared to state O’Neill’s theorem.
156
Theorem 4.5.7 (O’Neill [63]). Let π : M Ñ M{G be a Riemannian submersion, as above.
Let DπppXq, DπppY q P TrpsM{G be orthonormal with respect to the induced metric, where
X, Y P TpM . Then the sectional curvature of M{G at rps for the plane spanned by X and
Y is given by
secpXhor, Y horq `
3
4
›
›
›
“
Xhor, Y hor‰vert
›
›
›
2
g,
where sec is sectional curvature computed in M and r¨, ¨s is the Lie bracket of M .
Going through the proof of O’Neill’s formula, we see that it extends to infinite-dimensional
Riemannian submersions, provided all terms are well-defined (also see [76]). We are in par-
ticular considering the submersion M ÑM{Diff`0 pS1q. By Corollary 4.4.8, the horizontal
and vertical projections are well-defined. Moreover, we can compute the Lie bracket in M
easily via its correspondence with the Grassmannian. Given vector fields X, Y on Gr2pVq,
we can locally think of these as horizontal (with respect to the Up2q-action) vector fields on
LC2. Thus we can consider X and Y as smooth functions from LC2 to itself. Thus their
bracket in LC2 is given by
rX, Y sΦ :“ DXpΦqpY |Φq ´DY pΦqpX|Φq.
This bracket can then be projected to a Up2q-horizontal vector field, thus giving us the
bracket on Gr2pVq.
Theorem 4.5.8. The spacesM andM{Diff`0 pS1q both have non-negative sectional curva-
ture.
Proof. By Theorem 3.2.8, M is locally isometric to Gr2pVq, and it suffices to compute the
curvature of the Grassmannian to prove the first part of the theorem. Assuming that M
has non-negative sectional curvature, it follows by O’Neill’s formula thatM{Diff`0 pS1q does
as well, since MÑM{Diff`0 pS1q is a Riemannian submersion by construction.
157
We proceed with proving that Gr2pVq is nonnegatively curved. Let rΦs P Gr2pVq and let
δΦ1, δΦ2 P TrΦsGr2pVq be linearly independent and assume without loss of generality that
each vector is L2-normalized. Using the results of the previous section, and in particular
Proposition 4.5.6, choose a point rΦjs along the geodesic through rΦs with velocity vector
δΦj for j “ 1, 2. These points can be chosen to be distinct, by the explicit description
of geodesics given in Proposition 4.5.4. Next we use Proposition 4.5.2 to choose a totally
geodesic isometrically embedded finite-dimensional Grassmannian containing rΦs, rΦ1s and
rΦ2s. In particular, this Grassmannian contains rΦs and contains each δΦj in its tangent
space. Thus the sectional curvatures of Gr2pVq and the finite dimensional Grassmannian
agree at that point and plane. Finite-dimensional Grassmannians are well known to be
nonnegatively curved as they are compact symmetric spaces, thus the theorem follows.
158
Chapter 5
Application: Critical Points of Energy
Functionals
5.1 Background
In this chapter we consider a pair of natural energy functionals on the moduli space M of
framed loops. Before introducing the functionals, we give some historical context for our
results.
5.1.1 Elastic Energies
The study of critical points of energy functionals on curve spaces dates back to James
Bernoulli in the 17th century. Bernoulli sought to describe the equillibrium shape taken
by a springy wire with fixed boundary conditions. The problem was later solved by Euler,
who, following a suggestion of David Bernoulli, reformulated the problem as determining the
critical points of the energy functional
γ ÞÑ
ż
Σ
κpγq2 ds,
159
where γ : Σ Ñ R2 is an arclength-parameterized plane curve with fixed boundary conditions
(Σ is S1 or a closed interval). Accordingly, critical points of this functional are referred to
as Euler elastica.
Euler utilized variational techniques to give a complete characterization of the elastica. Of
particular interest are the Euler elastica with periodic boundary conditions; i.e., γ is a smooth
closed curve. By the Whitney-Graustein theorem [75], the space of arclength-parameterized
closed planar curves has countably many components—immersed plane curves are homotopic
through immersions if and only if they have the same winding number. For each nonzero
winding number component, the total squared curvature functional has a multiply-covered
oriented round circle as its only critical point, up to rotations and translations. On the
other hand, the zero winding number component has countably many critical points up to
rotation and translation. These are the multiply-covered “Bournoulli 8-curves”. See [44]
for an in-depth historical account of the elasticae and [66] for a modern derivation of the
periodic critical points.
An obvious generalization of the Euler elastica is to consider critical points of the total
squared curvature functional on arclength-parameterized space curves, which we continue to
refer to as elastica. A characterization of the closed elastica in R3 is given by Langer and
Singer [42]. Surprisingly, they showed that the knotted elastica are exactly the ph, kq-torus
knots satisfying h ą 2k.
The Euler elastica problem treats the original physical problem by considering the springy
wire as a 1-dimensional object. A more natural physical model is to treat the wire as having
a thickness. In this case, the wire is more readily modelled by a framed curve pγ, V q, where
γ is an arclength-parameterized curve. A relevant energy functional in this setup is the
Kirchhoff elastic energy
pγ, V q ÞÑ
ż
Σ
κpγq2 ` twpγ, V q2 ds.
160
Once again, the aim is to find critical points of this functional amongst framed curves
with fixed boundary conditions and such critical points are referred to as elastic rods. The
Kirchhoff elastic energy functional penalizes both bending and twisting of the rod. The
study of elastic rods is a vast topic, frequently covering more general forms of Kirchhoff
elastic energy than the one given here (see e.g. [45] and references therein).
The characterization of closed elastic rods was achieved by Ivey and Singer in [35], where
they showed that every torus knot type is realized by the base curve of an elastic rod. More-
over, they showed that for any relatively prime h and k, there exists a homotopy of elastic
rods from the h-times-covered circle to the k-times-covered circle possessing certain inter-
esting features (e.g., there is exactly one self-intersecting elastic rod between the beginning
and end of the homotopy).
5.1.2 Vortex Filament Equation
Another interesting and relevant energy functional is
γ ÞÑ1
2
ż
Σ
}γ1}2 dt, (5.1)
where γ is an arclength-parameterized space curve. As first noticed by Marsden and We-
instein [49], this energy functional is related to the evolution of a vortex filament. We now
demonstrate this fact somewhat informally. Let δγ be a variation of γ. Then the derivative
of the functional at γ in the δγ-direction is
d
dε
ˇ
ˇ
ˇ
ˇ
ε“0
1
2
ż
Σ
}γ1 ` εδγ1}2 dt “
ż
Σ
〈γ1, δγ1〉 dt “ ´
ż
Σ
〈γ2, δγ〉 dt.
Using the Millson-Zombro Kahler structure on the space of (isometry classes of) arclength-
parameterized loops, we see that if γ P LR3, then the gradient of this functional is ´γ2 “
161
´κN , where N is the principal normal to γ. Applying the negative almost complex structure
to the gradient, we see that the Hamiltonian vector field associated to the energy functional
is ´T ˆ ´κN “ κB, where B is the binormal vector field to γ. Therefore the Hamiltonian
flow of the energy functional, γpt, uq, satisfies the vortex filament equation:
B
Buγ “ κB.
Solutions of this PDE are well known to approximate the evolution of a vortex filament in
an incompressible fluid.
5.1.3 Description of Results
We now outline the results of this chapter. In Section 5.2, we study the weighted total twist
functional of a framed loop defined by
ĂTw :“1
4
ż
S1
tw}γ1} ds.
It turns out that the weighted total twist functional is the momentum map for the action
of a subgroup of Diff`pS1q which is diffeomorphic to S1 (the basepoint action). This is the
subgroup containing the stabilizers of the Diff`pS1q action, as described in Section 4.4.1. We
use this fact to give a characterization of the critical points of ĂTw.
In Section 5.3 we introduce the total elastic energy functional, denoted E, which is a
generalization of Kirchhoff elastic energy to not necessarily arclength-parameterized curves.
This functional takes the form
Epγ, V q :“1
4
ż
S1
pκ2` tw2
` st2q}γ1}2 ds,
162
where st is a term penalizing change in velocity per unit arclength, defined by
st :“1
}γ1}
d
ds}γ1}.
This closely resembles the strain energy of an extensible elastic rod, as in [6], Section 10.2.11.
The extra factor of 1{}γ1} must be included in the stretch term so that the functional scales
appropriately with length, and thus descends to a well-defined functional on our moduli
spaces. Restricting to arclength parameterized curves, the total elastic energy functional
reduces to exactly Kirchhoff elastic energy.
We give a description of the critical points to E. The critical points have very simple
descriptions, while the set of all critical points has a rich structure. The key to achieving
this characterization is to show that E is represented in complex coordinates as
EpΦq “
ż
S1
}Φ1}2C2 dt.
One immediately notices that this is strikingly similar to the energy functional 5.1. It turns
out that critical points of the complex energy functional which correspond to closed framed
curves satisfy a simple second-order ODE whose solutions can be written out explicitly, thus
the critical points of E have very simple representations. On the other hand, to compare to
the results of Ivey and Singer, we find a family of homotopies of critical points which take
an h-covered circle to a k-covered circle and which pass through torus knots in between.
163
5.2 Weighted Total Twist
5.2.1 The Basepoint Action
Recall from Remark 4.4.2 that there is a Lie subgroup of Diff`pS1q consisting of rigid ro-
tations of S1, and that this Lie subgroup is naturally diffeomorphic to S1. The reparame-
terization action of Diff`pS1q on M thus restricts to a (non-free) action of S1 on M. This
action is given explicitly for u P S1 « r0, 2s{p0 „ 2q by the forumula
u ¨ rγptq, V ptqs :“ rγpt´ uq, V pt´ uqs.
It follows from Lemma 4.4.7 that this action is given explicitly in Gr2pVq-coordinates by
u ¨ rΦptqs :“ rΦpt´ uqs,
and that this action is induced from the S1-action on St2pVq defined by
pu ¨ Φqptq :“ Φpt´ uq.
Each of these closely related actions of S1 will be referred to as the basepoint action. The
basepoint action can be interpreted as changing the basepoint of a based framed curve.
We claim that the basepoint action is Hamiltonian and that its momentum map has a
natural curve-theoretic interpretation. We will see that its momentum map is weighted twist
rate, defined in Section 5.1.3 by the formula
ĂTwpγ, V q “1
4
ż
S1
twpγ, V q}γ1} ds.
164
Note that this functional is invariant under translation and rotation. It could be made
invariant with respect to scaling, but we instead opt to restrict to framed curves with base
curves of length-2 (see the following remark). Thus ĂTw induces a well-defined functional on
M.
Remark 5.2.1. The factor of 1{4 in the definition of ĂTw is convenient for our purposes.
As defined, ĂTw is invariant under rotations, translations and global frame twists, but it is
not invariant under scaling of the base curve. Thus ĂTw is not well-defined on M. The most
natural remedy for this problem is to instead normalize by 1{p2lengthpγqq. Since we typically
realize M by fixing lengths of base curves at 2, this reduces to our choice of normalization.
Therefore we will take the convention in this section that M is always realized by restricting
to equivalence classes of length-2 curves, so that ĂTw is a well-defined energy functional on
M.
Another natural choice of normalization would be 1{p2πq. With this choice, ĂTw agrees
with the classical total twist functional Tw when restricted to the submanifold of framed loops
with arclength-parameterized base curve. We will stick with the 1{4 normalization only for
the sake of simplifying notation later on.
Proposition 5.2.2. The moment map for the basepoint action of S1 on M is ĂTw.
Before proving the proposition, we prove some preliminary lemmas.
Lemma 5.2.3. Let rΦs P Gr2pVq correspond to rγ, V s under pH. Then
ĂTwpγ, V q “ ´1
2Im 〈Φ,Φ1〉L2 .
Remark 5.2.4. The second statement of the lemma can be written as
ĂTwpγ, V q “1
2ωL
2
Φ pΦ,Φ1q,
165
where the right side is interpreted as the symplectic form on the complex vector space V2.
However, neither Φ nor Φ1 lie in T horΦ St2pVq « TrΦsGr2pVq, so this notation would be con-
fusing in the context of the lemma.
Proof. By Lemma 4.2.3, twist rate is given in complex coordinates by the formula
twpγ, V q “ ´2Im 〈Φ,Φ1〉C2
}Φ}4C2
.
Thus the claimed formula holds by the fact that }γ1} “ }Φ}2C2 (see Lemma 3.1.7), whence
ĂTwpγ, V q “1
4
ż 2
0
twpγ, V q}γ1} ds “1
4
ż 2
0
´2Im 〈Φ,Φ1〉C2
}Φ}4C2
}Φ}2C2 ¨ }Φ}2C2 dt “ ´1
2Im 〈Φ,Φ1〉L2 .
Lemma 5.2.5. For r P R « LiepS1q, the corresponding infinitesimal vector field for the
basepoint action on St2pVq is given by
Xr|Φ “ rΦ1.
Thus the infinitesimal vector field on Gr2pVq is the orthogonal projection of Xr onto the
horizontal tangent space of St2pVq, denoted projpXrq.
Proof. Let uε be a path in S1 « r0, 2s{p0 „ 2q with u0 “ 0 and let ddε
ˇ
ˇ
ε“0uε P LiepS1q « R
be denoted r. Then
d
dε
ˇ
ˇ
ˇ
ˇ
ε“0
Φpt´ uεq “ rΦ1.
It is straightforward to check that this vector lies in TΦSt2pVq. Thus its projection to
T horΦ St2pVq gives the infinitesimal vector field on Gr2pVq.
166
We now proceed with the proof of Proposition 5.2.2.
Proof of Proposition 5.2.2. This follows by straightforward computations in complex co-
ordinates. We first calculate the derivative of ĂTw. Let Φ “ pφ, ψq P LpC2zt0uq and
δΦ “ pδφ, δψq P TΦLpC2zt0uq « LC2. The derivative of ĂTw at Φ in the direction δΦ is
given by
d
dε
ˇ
ˇ
ˇ
ˇ
ε“0
ĂTwpΦ` εδΦq “d
dε
ˇ
ˇ
ˇ
ˇ
0
´1
2Im
ż 2
0
pφ1 ` εδφ1qpφ` εδφq ` pψ1 ` εδψ1qpψ ` εδψq dt
“ ´1
2Im
ż 2
0
δφ1φ` φ1δφ` δψ1ψ ` ψ1δψ dt
“ ´1
2Im
ż 2
0
´δφφ1 ` φ1δφ´ δψψ1 ` ψ1δψ dt (5.2)
“ ´Im 〈δΦ,´Φ1〉L2 “ ωL2
Φ pδΦ,´Φ1q,
where (5.2) follows after integrating by parts and uses the assumption that all of the functions
involved are periodic. If we had instead taken Φ P ApC2zt0uq then it is easy to check that
the same calculation holds. Moreover, the same calculation holds in local coordinates on
Gr2pVq, where we apply Lemma 4.1.6 to obtain
DTwpΦqpδΦq “ ωL2
rΦspδΦ, projp´Φ1qq (5.3)
for δΦ P T horΦ St2pVq « TrΦsGr2pVq. An implicit requirement for this claim to be well-defined
is that ´Φ1 P TΦSt2pVq. Indeed, we can check that the constraint equations of TΦSt2pVq are
satisfied; e.g.,
Re
ż
φp´φ1q dt “ ´
ż
Repφφ1q dt “ ´1
2
ż
pφφq1 dt “ 0.
167
Let 〈¨, ¨〉R denote the standard inner product on R “ LiepS1q (i.e. multiplication). Then
for any r P R
〈DTwpΦqpδΦq, r〉R “ ωL2
rΦspδΦ, projp´rΦ1qq “ ωL2
rΦspδΦ, Xr|rΦsq.
From Equation (5.3) of the proof of the proposition, we immediately obtain the following
corollary.
Corollary 5.2.6. The Hamiltonian vector field of ĂTw : Gr2pVq Ñ R is given at rΦs by
projp´Φ1q, where proj is orthogonal projection from TΦSt2pVq to T horΦ St2pVq « TrΦsGr2pVq.
Thus the gradient of ĂTw at rΦs is projp´iΦ1q, where proj is orthogonal projection from
TΦSt2pVq to T horΦ St2pVq « TrΦsGr2pVq.
Remark 5.2.7. The gradient vector field of ĂTw is a projection of the field ´iΦ1. We will
ignore the projection for a moment and consider the flow of ´iΦ1 in PC2. Denote this flow by
Φpu, tq “ pφpu, tq, ψpu, tqq, where u is the flow parameter. We can interpret each coordinate
function as a map from C to itself, where the domain has coordinates u ` it. It is not hard
to see that this implies that each coordinate function is holomorphic. It would be interesting
to study the implications of this observation for the gradient flow of ĂTw.
5.2.2 Critical Points of Total Twist
By the discussion above, ĂTw is a natural energy functional to study on M. We conclude
this section by characterizing its critical points.
Theorem 5.2.8. The critical points of ĂTw :MÑ R are equivalence classes rγ, V s such that
γ is an arclength parameterized, length-2, multiply covered round circle and V has constant
168
Figure 5.1: A torus knot realized as a pushoff of a critical point of ĂTw and the correspondingClifford torus knot as a stereographic projection of Φ2,3.
twist rate. Thus a small pushoff γ ` εV forms a multiply-covered torus knot and every torus
knot type is realized as such a pushoff for some critical point. All critical points are unstable.
Each critical framed loop has a complex-coordinate representative as a torus knot on the
standard Clifford torus in S3.
Proof. It is a general fact that if µ is the momentum map of a circle action on a finite-
dimensional Kahler manifold, then the points of the manifold where the gradient vanishes
are exactly the fixed points of the circle action. We see this by identifying LiepS1q with R
and considering the momentum map as a Hamiltonian. Then the associated Hamiltonian
vector field is tangent to the group orbits. Moreover, the gradient vanishes exactly when the
Hamiltonian vector field vanishes since the Riemannian metric and symplectic structure are
related by the complex structure. The points where the Hamiltonian vector field vanishes
are the fixed point of the circle action.
The reasoning of the previous paragraph applies to an infinite-dimensional Kahler man-
ifold, provided all of the pieces are well-defined. By Corollary 5.2.6, this is the case for the
169
situation at hand. Thus we see that the points of the Grassmannian where the gradient of
the twist functional is zero correspond exactly to the points fixed by the reparameterization
action of S1. A point rΦs is fixed in the Grassmannian under the S1 action if and only if
rΦpt´ uqs “ rΦptqs for all u P S1« r0, 2s{p0 „ 2q.
Since the equivalence class rΦs is up to Up2q-action, this means that there is an Apuq P Up2q
for each u P S1 such that
Φpt´ uq “ Φptq ¨ Apuq. (5.4)
Setting t “ 0 in (5.4), we obtain
Φp´uq “ Φp0q ¨ Apuq for all u P S1. (5.5)
An element Φ P St2pVq satisfies 5.5 if and only if its corresponding framed loop has constant
parameterization speed as well as constant (classical) curvature, torsion and twist rate.
Indeed, multiplication by an element of Up2q in complex coordinates corresponds to a rotation
together with a global frame twist of the corresponding framed loop. These actions do not
effect any of the geometric invariants. Thus (5.5) implies that the values of the geometric
quantities of such a corresponding framed loop at any time must agree with the values at
time 0. The framed loops with constant curvature, torsion, parameterization speed and twist
rate are exactly those described by the theorem.
The fact that pushoffs produce torus knots for these types of framed loops is obvious, as
is the fact that all torus knot types are obtained.
We observe that the critical points are unstable because for any critical point rγ0, V0s,
there is a one parameter family rγτ , Vτ s such that Twpγτ , Vτ q “ Twpγ0, V0q for all τ . For
170
example, such a one-parameter family can be obtained by keeping γτ “ γ0 and altering Vτ
by increasing twist rate at one point and decreasing it elsewhere to keep total twist fixed.
Finally, we wish to show that a critical rγ, V s has a complex representation as a knot
on the Clifford torus in S3. We do so by providing such representations explicitly. We
assume that the (perhaps multiply-covered) torus knot corresponding to rγ, V s is of type
ph, kq P Z ˆ Z. In other words, γ is an h-times-covered round circle, and V has constant
twist rate π ¨ k. We claim that the Clifford torus knot
Φh,k “ pφh,k, ψh,kq :“1?
2
ˆ
exp
ˆ
i
2pk ` hqπt
˙
, exp
ˆ
i
2pk ´ hqπt
˙˙
corresponds to rγ, V s under pH.
Indeed, applying pH to Φh,k, we get the base curve
γ “1
πhp0,´ cosphπtq, sinphπtqq .
This curve has been translated to take its average value at 0 (rather than our usual convention
of taking its basepoint at 0). Then γ is clearly an h-times-covered round circle. Moreover,
we apply Lemma 4.2.3 to calculate the twist rate of Φh,k:
twpΦh,kq “ ´2Im
`
φh,k ¨ φh,k 1 ` ψh,k ¨ ψh,k 1˘
}Φh,k}4C2
“ ´2Im
ˆ
1
2exp
ˆ
i
2pk ` hqπt
˙
¨ exp
ˆ
´i
2pk ` hqπt
˙
¨´i
2pk ` hqπ
`1
2exp
ˆ
i
2pk ´ hqπt
˙
¨ exp
ˆ
´i
2pk ´ hqπt
˙
¨´i
2pk ´ hqπ
˙
“ ´2Im
ˆ
´i
4pk ` hqπ `
´i
4pk ´ hqπ
˙
“ kπ,
as claimed.
171
5.3 Total Elastic Energy
5.3.1 Framed Curve Invariants
Before moving on to study the total elastic energy functional defined in Section 5.1.3, we
derive the complex-coordinate representations of the remaining framed curve invariants.
Lemma 5.3.1. Let Φ be a complex-coordinate representation of a framed loop pγ, V q. Then
the Darboux curvatures of pγ, V q are given by
κ1 “2Impφ1ψ ´ ψ1φq
}Φ}4C2
and κ2 “ ´2Repφ1ψ ´ ψ1φq
}Φ}4C2
.
Proof. Let Φ “ pφ, ψq correspond to the quaternion q under the identification C2 Ø H and
to pγ, V q under pH. Recall from Lemma 4.2.2 that the quaternionic variation δq1 “ q1{}q}2H
corresponds to the framed curve variation
pδγ11, δV1q “
ˆ
d
dsγ1,
d
dsV
˙
under DpHpqq. Also recall from Lemma 4.1.7 that the quaternionic variation δq2 “ kq{2
corresponds to the framed curve variation
pδγ12, δV2q “ p´}γ1}V, T q.
172
Now we note that under the identification C2 Ø H, kq corresponds to p´iψ, iφq. Putting
this together, we conclude
2Impφ1ψ ´ ψ1φq
}Φ}4C2
“2
}Φ}4C2
Im⟨pφ1, ψ1q, pψ,´φq
⟩C2
“2
}Φ}4C2
Re⟨pφ1, ψ1q, piψ,´φq
⟩C2
“4
}q}2HRe
⟨q1
}q}2H,´
kq
2
⟩H
“4
}γ1}
1
4
ˆ⟨1
}γ1}
d
dsγ1, V
⟩`
⟨d
dsV,W
⟩〈T,W 〉
˙
}γ1} (5.6)
“
⟨1
}γ1}
d
dsγ1, V
⟩(5.7)
Line (5.6) follows from Corollary 3.1.11 and Lemma 3.1.7 and by noting that δγ11 “ddsγ1
implies that ddsδγ1 “ p1{}γ1}q d
dsγ1 and likewise for δγ12 “ ´}γ1}V . Now we notice that for
T “ γ1{}γ1}
⟨1
}γ1}
d
dsγ1, V
⟩“
1
}γ1}
⟨d
ds}γ1}T, V
⟩“
1
}γ1}
⟨ˆd
ds}γ1}
˙
T ` }γ1}d
dsT, V
⟩“
⟨d
dsT, V
⟩“ κ1.
This completes the proof for κ1. To derive the formula for κ2, one repeats the argument
above using the basic variation jq{2.
Recall that the stretch rate of a parameterized framed curve pγ, V q is given by
st “1
}γ1}
d
ds}γ1}
173
and represents the relative change per unit arclength of the parameterization speed of γ. We
now derive an expression for stretch rate in complex coordinates.
Lemma 5.3.2. Let Φ be a complex-coordinate representation of a framed loop pγ, V q. Then
the stretch rate of pγ, V q is given by
st “2Repφφ1 ` ψψ1q
}Φ}4C2
.
Proof. The derivation is easier in this case:
1
}γ1}
d
ds}γ1} “
1
}Φ}4C2
d
dt}Φ}2C2
“1
}Φ}4C2
d
dtpφφ` ψψq
“2Repφφ1 ` ψψ1q
}Φ}4C2
.
Remark 5.3.3. A simple calculation shows
st “1
}γ1}
d
ds}γ1} “
1
}γ1}2d
dt〈γ1, γ1〉1{2 “ 〈γ
2, γ1〉}γ1}3
.
Recall from Section 2.1.2 that the normal vector of an arbitrarily parameterized space curve
γ is given by the formula
κN “γ2
}γ1}2´〈γ2, γ1〉}γ1}3
T “γ2
}γ1}2´ st ¨ T,
and we see that stretch rate has already appeared naturally in the context of curve theory.
174
5.3.2 Total Elastic Energy
We now return to the total elastic energy functional on defined in Section 5.1.3. This
functional is defined on a framed path pγ, V q by the formula
Epγ, V q “1
4
ż 2
0
`
κ2` tw2
` st2˘
}γ1}2 ds,
We claim that the functional E is natural from a variety of perspectives. First, the total
energy functional is a straightforward generalization of the Kirchhoff elastic energy func-
tional described in Section 5.1—in particular, restricting to the submanifold of arclength-
parameterized framed curves produces Kirchhoff energy exactly. Secondly, applications
sometimes require elasticity in the stretch term, for example in the case of polymers where
the distance between the atoms is able to vary [38]. Another feature of the functional E
is that it has an incredibly nice representation in complex coordinates. Finally, E induces
a well-defined functional on the moduli space M, and the critical point set of the induced
functional on M is quite rich.
Remark 5.3.4. The statement that E is well-defined on M is still under the convention
that M is realized as a restriction to equivalence classes of length-2 curves. As in the case
of ĂTw, we could replace the 1{4 normalization factor by 1{p2lengthpγqq to achieve an energy
functional which is invariant under scaling of the base curve. The choice of 1{4 is convenient
for calculations to follow.
We begin our study of E by describing it in complex coordinates.
Proposition 5.3.5. In complex coordinates for framed paths, the total elastic energy func-
tional takes the form
EpΦq “
ż 2
0
}Φ1}2C2 dt.
175
Proof. Appealing to Lemmas 4.2.3, 5.3.1 and 5.3.2, we write
EpΦq “1
4
ż 2
0
4
}Φ}4C2
`
}φψ1 ´ φ1ψ}2C ` }φφ1 ` ψψ1}2C
˘
}Φ}2C2dt
“
ż 2
0
1
}Φ}2C2
`
}φψ1 ´ φ1ψ}2C ` }φφ1 ` ψψ1}2C
˘
dt. (5.8)
Simplifying }φψ1 ´ φ1ψ}2C ` }φφ1 ` ψψ1}2C yields
}φ1ψ ´ ψ1φ}2C ` }φφ1 ` ψψ1}2C “ }φ
1ψ}2C ` }ψ1φ}2C ` }φφ
1}
2C ` }ψψ
1}
2C
´ 2Re`
φ1ψψ1φ˘
` 2Re´
φφ1`
ψψ1˘
¯
“ p}φ}2C ` }ψ}2Cqp}φ
1}
2C ` }ψ
1}
2Cq “ }Φ}
2C2}Φ1}2C2 .
Replacing this into (5.8) yields the result.
For a moment, we consider E as a functional on the space W “ LpC2zt0uq \ApC2zt0uq,
which was previously introduced in Section 4.1.3. Recall that under pH, W double-covers the
space of frame-periodic framed paths rSper.
Proposition 5.3.6. The gradient of E :W Ñ R with respect to the natural L2 metric takes
the form
gradpEq|Φ “ ´2Φ2.
Proof. We calculate the derivative of E at Φ PW in the direction δΦ P TΦW «W :
DEpΦqpδΦq “d
dε
ˇ
ˇ
ˇ
ˇ
ε“0
ż 2
0
}Φ1 ` εδΦ1}C2 dt
“
ż 2
0
φ1δφ1 ` δφ1φ` ψ1δψ1 ` δψ1ψ1 dt
“ ´2Re
ż 2
0
φ2δφ` ψ2δψ dt
“ Re 〈´2Φ2, δΦ〉L2 ,
176
where we have once again employed integration by parts using the assumption that either
all functions involved are periodic or all functions are antiperiodic.
5.3.3 Critical Points of E
Noting that E is invariant under rigid rotations, global frame twists, translations and scaling
(see Remark 5.3.4), we conclude that it induces a well-defined functional on M, which we
continue to denote by E. Our goal for this section is to determine the critcal points of
E on M. Locally, M can be considered as a codimension-8 submanifold of rSper. Indeed,
M is identified with the disjoint union of Grassmannians, each Grassmannian is locally a
codimension-4 submanifold of the corresponding Stiefel manifold (by virtue of a slice chart
to the Up2q-action), and each Stiefel manifold is an honest codimension-4 submanifold ofW .
Thus we wish to find the points where the gradient of E points in the 8-dimensional normal
direction toM, which we recall was described explicitly in Lemma 4.5.3. Before stating the
main theorem of this subsection (Theorem 5.3.10), we introduce notation and prove some
preliminary lemmas.
For the remainder of this section, we consider complex paths of the form
Φptq “
¨
˚
˝
φptq
ψptq
˛
‹
‚
“
¨
˚
˝
z1 exppic1πt{2q ` z2 expp´ic1πt{2q
w1 exppic2πt{2q ` w2 expp´ic2πt{2q
˛
‹
‚
, (5.9)
where c1, c2 P Z satisfy c1 “ c2 mod 2 and z1, z2, w1, w2 P C. The collection of all such Φ
will be denoted Sol. We will see in the following lemma that these are certain solutions to a
family of second-order ODEs, and we will see in the proof of Theorem 5.3.10 that this family
of ODEs is relevant to the critical point set of E.
177
Lemma 5.3.7. Consider the system of second order ODEs
¨
˚
˝
φ2
ψ2
˛
‹
‚
“
¨
˚
˝
λ1 0
0 λ2
˛
‹
‚
¨
˚
˝
φ
ψ
˛
‹
‚
(5.10)
where Φ “ pφ, ψq : RÑ C2 and λ1, λ2 P R. The set Sol consists of solutions lying in W.
Proof. General solutions φ : RÑ C of the equation
φ2 “ λφ, λ P R (5.11)
take the form
φptq “ z1e?λt` z2e
´?λt
for z1, z2 P C. Indeed, such functions satisfy (5.11) and solutions of (5.11) are unique up
to choice of initial conditions. There are four (real) degrees of freedom for choosing initial
conditions, hence the (complex) parameters z1 and z2 appearing in the solution. We conclude
that general solutions Φ : RÑ C of (5.10) take the form
Φptq “
¨
˚
˝
φptq
ψptq
˛
‹
‚
“
¨
˚
˝
z1 expp?λ1tq ` z2 expp´
?λ1tq
w1 expp?λ2tq ` w2 expp´
?λ2tq
˛
‹
‚
(5.12)
for z1, z2, w1, w2 P C.
Of course, not all paths Φ of the form (5.12) lie inW . The path Φ is a loop (respectively an
anti-loop) if and only if exppa
λjtq is a complex loop (anti-loop) for j “ 1, 2. We conclude
thata
λj is a complex integer multiple of π (respectively, half-integer multiple of π) for
j “ 1, 2. Thus we have shown that a necessary and sufficient condition for Φ to lie in W is
that it Φ P Sol.
178
Thus Sol is a subset of W—if the cj are even, then Φ P LC2 and if the cj are odd then
Φ P AC2. Next we wish to characterize which elements Φ P Sol lie in St2pVq.
Lemma 5.3.8. Let Φ P Sol. Then
Φ P St2pVq ô
$
’
&
’
%
|z1|2 ` |z2|
2 “ |w1|2 ` |w2|
2 “ 1 c1 ‰ c2
pz1, z2q and pw1, w2q are orthonormal in C2 c1 “ c2 ‰ 0
Proof. It is straightforward to conclude that Φ P Sol is never an element of St2pVq if c1 “
c2 “ 0. Indeed, if Φ “ pφ, ψq “ pz, wq for some z, w P C, then
ż
φψ dt “ 2zw,
which is zero if and only if z or w is zero, and this violates normality.
Next we note that the set texppcπt{2q | c P 2Zu is L2-orthonormal as a subset of LC.
Similarly, texppcπt{2q | c P Z is oddu is an L2-orthonormal subset of AC. Let Φ P Sol with
c1 ‰ c2. Then
〈φ, φ〉L2 “ |z1|2 〈exppic1πt{2q, exppic1πt{2q〉L2 ` |z2|
2 〈expp´ic1πt{2q, expp´ic1πt{2q〉L2
` Re pz1z2 〈exppic1πt{2q, expp´ic1πt{2q〉L2q
“ |z1|2` |z2|
2.
Similarly 〈ψ, ψ〉 “ |w1|2 ` |w2|
2.
179
Finally,
〈φ, ψ〉L2 “ z1w1 〈exppic1πt{2q, exppic2πt{2q〉L2 ` z1w2 〈exppic1πt{2q, expp´ic2πt{2q〉L2
` z2w1 〈expp´ic1πt{2q, exppic2πt{2q〉L2 ` z2w2 〈expp´ic1πt{2q, expp´ic2πt{2q〉L2
“ 0
This establishes the first set of necessary and sufficient conditions.
In the case that Φ P Sol with c1 “ c2 “ c ‰ 0, the same calculations show 〈φ, φ〉L2 “
|z1|2 ` |z2|
2 and 〈ψ, ψ〉L2 “ |w1|2 ` |w2|
2. However, in this case
〈φ, ψ〉L2 “ z1w1 〈exppicπt{2q, exppicπt{2q〉L2 ` z1w2 〈exppicπt{2q, expp´icπt{2q〉L2
` z2w1 〈expp´icπt{2q, exppicπt{2q〉L2 ` z2w2 〈expp´icπt{2q, expp´icπt{2q〉L2
“ z1w1 ` z2w2,
and this completes the proof.
We will show in the following theorem that critical points of E : M Ñ R can be repre-
sented as elements of Sol X pSt˝2pLCq \ St˝2pACqq. These representations are not unique, so
we impose some normalizing conditions. We define the set
Crit :“ tΦc | c P Zu Y tΦ P SolX pSt˝2pLCq \ St˝2pACqq | c1 ą c2 ě 0, z1 P Rě0u
“: Crit1 Y Crit2,
where
Φcptq “
¨
˚
˝
φcptq
ψcptq
˛
‹
‚
:“
¨
˚
˝
exppicπt{2q
expp´icπt{2q
˛
‹
‚
. (5.13)
We will see in the following lemma the reason for the distinction of the set Crit1
180
Lemma 5.3.9. The action of Up2q on St2pVq restricts to a free transitive action on each set
Solc :“ tΦ P SolX pSt˝2pLCq \ St˝2pACqq | c1 “ c2 “ c ‰ 0u.
Thus Crit1 is a Up2q-cross-section ofď
cPZ
Solc.
Proof. Let
A “ eiθ
¨
˚
˝
u v
´v u
˛
‹
‚
denote an arbitrary element of Up2q and let Φ P Solc. Then
Φ ¨ A “
¨
˚
˝
eiθpuz1 ´ vw1q exppicπt{2q ` eiθpuz2 ´ vw2q expp´icπt{2q
eiθpvz1 ´ vw1q exppicπt{2q ` eiθpvz2 ´ vw2q expp´icπt{2q
˛
‹
‚
.
By a straightforward calculation, we see that the assumption that the parameters zj, wj
satisfy the conditions of Lemma 5.3.8 implies that the new parameters do as well. We
conclude that Up2q acts on Solc, and the fact that the action is free follows from the freeness
of the action on St2pVq.
Then it remains to show that the action is transitive. Let
Φptq “
¨
˚
˝
φptq
ψptq
˛
‹
‚
“
¨
˚
˝
z1 exppicπt{2q ` z2 expp´icπt{2q
w1 exppicπt{2q ` w2 expp´icπt{2q
˛
‹
‚
P Solc.
The claim follows by showing that there exists A P Up2q such that Φc ¨ A “ Φ. Indeed, this
is satisfied by the matrix
A “
¨
˚
˝
z1 w1
z2 w2
˛
‹
‚
,
181
which is an element of Up2q since the parameters zj, wj must satisfy the conditions of Lemma
5.3.8 and this implies that the columns of A are Hermitian orthonormal.
Since the Up2q-action is free, transitivity implies that Crit1 is a Up2q-cross-section ofŤ
cPZ Solc.
We are now prepared to state our main result.
Theorem 5.3.10. The critical points of E :MÑ R are in one-to-one correspondence with
the complex parameterizations in Crit.
Proof. Locally we consider M as a codimension-8 submanifold of rSper. Critical points of E
restricted toM occur when gradpEq lies entirely in the normal space toM. We now work in
complex coordinates, where the previous statements are translated into the statement that
critical points of E restricted to Gr2pVq occur when gradpEq lies in the 8-dimensional normal
space to Gr2pVq, considered locally as a submanifold of W . For rΦs P Gr2pVq, the normal
space to rΦs is spanned by the eight vectors listed in Lemma 4.5.3. Easy calculations show
that the gradpEq|Φ is orthogonal to each of the four vectors in the second list (4.17); e.g.
integrating by parts shows
Rep〈pφ2, ψ2q, piφ, 0q〉L2 “ Re
ż
φ2piφq dt
“ Re
ż
i|φ1|2 dt “ 0.
Thus we are looking for Φ P St2pVq such that the gradpEq|Φ lies in the 4-dimensional
(over R) normal space to the Stiefel manifold at Φ. That is,
Φ2 “ λ1pφ, 0q ` λ2p0, ψq ` λ3pψ, φq ` λ4p´iψ, iφq, for some λj P R.
182
We can rewrite this as
¨
˚
˝
φ2
ψ2
˛
‹
‚
“
¨
˚
˝
λ1 λ3 ´ iλ4
λ3 ` iλ4 λ2
˛
‹
‚
¨
˚
˝
φ
ψ
˛
‹
‚
.
The matrix is Hermitian, hence diagonalizable by unitary matrices by the spectral theorem.
Since we only care about solutions up to multiplication by unitary matrices, we can rename
parameters and replace our system by
¨
˚
˝
φ2
ψ2
˛
‹
‚
“
¨
˚
˝
λ1 0
0 λ2
˛
‹
‚
¨
˚
˝
φ
ψ
˛
‹
‚
. (5.14)
From Lemmas 5.3.7 and 5.3.8 we see that the critical points of E can be written as elements of
SolXpSt2pLCq\St2pACqq, and that this is described by explicit conditions on the parameters.
We conclude the proof by showing that for any Φ P Sol X pSt2pLCq \ St2pACqq, there
exists A P Up2q such that Φ ¨ A P Crit, and then that such an A is unique. This will be
accomplished in two cases. We first assume that c1 “ c2 “ c, so that Φ P Solc. By Lemma
5.3.9, there is a unique A P Up2q such that Φ ¨ A “ Φc P Crit1. Moreover, Φ ¨ A R Crit2 for
any A P Up2q. Thus the proof is complete in this case.
We now move on to the second case and assume that Φ P SolXpSt2pLCq\St2pACqq with
c1 ‰ c2. Our goal is to find a matrix A P Up2q such that Φ P Crit2. We first wish to show
that it can be arranged that c1 ą c2 ě 0. Indeed, by renaming parameters we can ensure
that c1 ě ´c1 and c2 ě ´c2. Since c1 ‰ c2, we finish rearranging by multiplying by
¨
˚
˝
0 1
1 0
˛
‹
‚
P Up2q
183
if necessary, thus ensuring that c1 ą c2. We finish this part of the argument by writing
z1 “ r1eiθ1 and multiplying by the matrix
A “
¨
˚
˝
e´iθ1 0
0 eiθ1
˛
‹
‚
P Up2q.
Clearly, Φ ¨A P Crit2, and we have proved the existence part of our claim in the case c1 ‰ c2.
Now we prove that such an A is unique. We will prove the equivalent statement that for
Φ, rΦ P Crit2, if Φ ¨A “ rΦ, then A is the identity matrix. Let Φ and rΦ be written in the form
(5.9), with the parameters of rΦ given by rzj, rwj, and rcj, and let
A “ eiθ
¨
˚
˝
u v
´v u
˛
‹
‚
be an arbitrary element of Up2q. Then Φ¨A “ rΦ implies that their first coordinates cooincide;
i.e.,
eiθ puz1 exppipc1qπt{2q ` uz2 expp´ipc1qπt{2q
´vw1 exppipc2qπt{2q ´ vw2 expp´ipc2qπt{2qq
“ rz1 exp pi prc1q πt{2q ` rz2 exp p´i prc1q πt{2q .
From the L2-orthonormality of the exponential functions and the conditions c1 ą c2 ě 0 and
rc1 ą rc2 ě 0, we deduce that there are two cases: either rc1 “ c1, or rc1 “ c2. In the first case,
we once again appeal to L2-orthonormality to conclude that
eiθuz1 “ rz1 and v “ 0.
184
The first equality implies that eiθu is a positive real number, as z1, rz1 P Rě0. The second
equality implies that |eiθu| “ 1, therefore A is the identity matrix.
Now we assume that rc1 “ c2. This implies
rz1 “ ´eiθvw1, rz2 “ ´e
iθvw2, u “ 0.
Thus the equality of the second coordinates of Φ ¨ A and rΦ reads
eiθv pz1 exppipc1qπt{2q ` z2 expp´ipc1qπt{2qq
“ rw1 exp pi prc2q πt{2q ` rw2 exp p´i prc2q πt{2q ,
whence we conclude that rc2 “ c1. However, this is a contradiction, as we have
rc2 “ c1 ą c2 “ rc1,
violating rΦ P Crit2.
Using our convenient complex form of E and the previous theorem, we are able to calculate
the possible energy levels that a framed curve can realize. Let Φ “ pφ, ψq P Crit. Then
φ1ptq “iπc1
2z1 exppiπc1t{2q `
´iπc1
2z2 expp´iπc1t{2q,
so
|φ1ptq|2 “
ˇ
ˇ
ˇ
ˇ
iπc1
2z1 exppiπc1t{2q
ˇ
ˇ
ˇ
ˇ
2
`
ˇ
ˇ
ˇ
ˇ
´iπc1
2z2 expp´iπc1t{2q
ˇ
ˇ
ˇ
ˇ
2
` 2Re
˜
iπc1
2z1 exppiπc1t{2q
ˆ
´iπc1
2z2 expp´iπc1t{2q
˙
¸
“
´πc1
2
¯2
´ 2´πc1
2
¯2
Re pz1z2 exppiπc1tqq .
185
Similarly,
|ψ1ptq|2 “´πc2
2
¯2
´ 2´πc2
2
¯2
Re pw1w2 exppiπc2tqq .
Thus
EpΦq “
ż 2
0
}Φ}2C2 dt
“
ż 2
0
´πc1
2
¯2
´ 2´πc1
2
¯2
Re pz1z2 exppiπc1tqq `´πc2
2
¯2
´ 2´πc2
2
¯2
Re pw1w2 exppiπc2tqq dt
“π2pc2
1 ` c22q
2´ 2
´πc1
2
¯2
Re
ˆ
z1z2
ż 2
0
exppiπc1tq dt
˙
´ 2´πc2
2
¯2
Re
ˆ
w1w2
ż 2
0
exppiπc2tq dt
˙
“π2pc2
1 ` c22q
2.
Therefore we have proved the following corollary.
Corollary 5.3.11. The critical energy levels of E :M Ñ R are quantized and the possible
critical energy levels are
π2pc21 ` c
22q
2
for any integers c1, c2 which are not both zero.
5.3.4 One-Parameter Families of Critical Points
We now move on to understanding these critical points of E in framed loop coordinates—e.g.,
what is the knot type of the base curve of a critical rγ, V s? Towards this goal, we examine
a particular collection of 1-parameter families of critical points.
186
Let c1 ą c2 ě 0 be integers. We will consider the simple 1-parameter family Φu “
pφu, ψuq, u P r0, 1s, where
φuptq :“ u exppic1πt{2q `?
1´ u2 expp´ic1πt{2q, ψuptq :“ exppic2πt{2q.
Note that each Φu in this family is an element of the set Crit defined in the previous sub-
section. It will be convenient to introduce the change of variables
h :“c1 ` c2
2and k :“
c1 ´ c2
2.
Then, up to a translation, Φu maps under pH to the framed loop pγu, Vuq, where
γuptq “2
π
¨
˚
˚
˚
˚
˚
˚
˚
˚
˚
˚
˚
˚
˝
u?
1´ u2
h` ksinpph` kqπtq
´u
kcospkπtq `
?1´ u2
hcosphπtq
u
ksinpkπtq `
?1´ u2
hsinphπtq
˛
‹
‹
‹
‹
‹
‹
‹
‹
‹
‹
‹
‹
‚
,
and
Vuptq “2
}Φu}2C2
¨
˚
˚
˚
˚
˝
u sinphπtq ´?
1´ u2 sinpkπtq
cosphπtq cospkπtq ` u?
1´ u2
p1´ u2q sinphπtq cospkπtq ´ u2 cosphπtq sinpkπtq
˛
‹
‹
‹
‹
‚
,
with
}Φu}2C2 “ 2` 2u
?1´ u2 cospph` kqπtq.
187
Then Φ0 corresponds under pH to the framed loop pγ0, V0q, where
γ0ptq “2
πh
¨
˚
˚
˚
˚
˝
0
cosphπtq
sinphπtq
˛
‹
‹
‹
‹
‚
and
V0ptq “
¨
˚
˚
˚
˚
˝
´ sinpkπtq
cosphπtq cospkπtq
sinphπtq cospkπtq
˛
‹
‹
‹
‹
‚
“ cospkπtq
¨
˚
˚
˚
˚
˝
0
cosphπtq
sinphπtq
˛
‹
‹
‹
‹
‚
` sinpkπtq
¨
˚
˚
˚
˚
˝
1
0
0
˛
‹
‹
‹
‹
‚
Clearly γ0 is an arclength-parameterized h-times-covered round circle. Moreover, it is clear
from the second form of V0ptq that the linking number of V0 and the image of γ0 is k.
Similarly, Φ1 corresponds to
γ1ptq “2
πk
¨
˚
˚
˚
˚
˝
0
´ cospkπtq
sinpkπtq
˛
‹
‹
‹
‹
‚
and
V1ptq “
¨
˚
˚
˚
˚
˝
sinphπtq
cosphπtq cospkπtq
´ cosphπtq sinpkπtq
˛
‹
‹
‹
‹
‚
“ .´ cosphπtq
¨
˚
˚
˚
˚
˝
0
´ cospkπtq
sinpkπtq
˛
‹
‹
‹
‹
‚
` sinphπtq
¨
˚
˚
˚
˚
˝
1
0
0
˛
‹
‹
‹
‹
‚
which is an arclength parameterized k-covered round circle whose image is linked ´h times
by V1.
Thus it remains to determine what happens between the endpoints of the one-parameter
family. Our first goal is to show that the base curve γu is nonembedded for exactly one value
188
of u P p0, 1q. This can be achieved by directly analyzing the explicit formula for γu. The
proof follows slightly more easily by utilizing the following generalization of Lemma 3.2.1,
which has essentially the same proof.
Lemma 5.3.12. Let Φ “ pφ, ψq P St2pVq correspond to the framed loop pγ, V q under pH.
Then γ has a self-intersection γpt0q “ γpt1q for 0 ă a ă b ă 2 if and only if Φ satisfies
ż b
a
|φ|2 ´ |ψ|2 dt “
ż b
a
φψ dt “ 0.
Corollary 5.3.13. For the 1-parameter family of critical points described above, the only
u P p0, 1q for which the base curve γu is nonembedded is
u “
c
k2
h2 ` k2.
Proof. The 1-parameter family is given in complex coordinates by
pφuptq, ψptqq “ pu exppic1πt{2q `?
1´ u2 expp´ic1πt{2q, exppic2πt{2qq,
thus the conditionż b
a
φuψu dt “ 0
reads
0 “
ż b
a
u exppic1πt{2q expp´ic2πt{2q `?
1´ u2 expp´ic1πt{2q expp´ic2πt{2q dt
“u
ikπexppikπbq `
?1´ u2
´ihπexpp´ihπbq ´
u
ikexppikπaq ´
?1´ u2
´ihπexpp´ihπaq,
189
or, equivalently,
hupexppikπaq ´ exppikπbqq “ k?
1´ u2pexpp´ihaq ´ expp´ihbqq.
Taking the squared absolute value of each side yields
h2u2p2´ 2Re exppikπpa´ bqqq “ k2
p1´ u2qp2´ 2Re expp´ihpa´ bqqq.
Now we claim that Re exppikπpa ´ bqq “ Re expp´ihpa ´ bqq. If this is the case then we
immediately obtain our result, as this implies
h2u2“ k2
p1´ u2q, or u “
c
k2
h2 ` k2.
Thus it remains to show that Re exppikπpa´bqq “ Re expp´ihpa´bqq, which is equivalent
to cosppa ´ bqkπq “ cosppa ´ bqhπq. From the explicit description of γu, we see that for γu
to have a self-intersection at times a ă b it must be that sinpph ` kqbq “ sinpph ` kqaq.
Therefore it is necessary that
b “
$
’
&
’
%
a` 2jh`k
or
2j`1h`k
´ a,j “ 0, 1, . . . , h` k ´ 1.
Taking b “ a` 2j{ph` kq, we have
cosppa´ bqkπq “ cos
ˆ
2j
h` kkπ
˙
“ cos
ˆ
2j
h` kpph` kq ´ hqπ
˙
“ cos
ˆ
2jπ ´2j
h` khπ
˙
“ cos
ˆ
2j
h` khπ
˙
“ cosppa´ bqhπq.
190
The case b “ p2j ` 1q{ph` kq ´ a follows similarly.
We now describe the knot types which appear in the 1-parameter family γu.
Lemma 5.3.14. Assume that h and h ` k are relatively prime. Then for 0 ă u ă
a
k2{ph2 ` k2q, γu parameterizes an ph, h` kq-torus knot.
Proof. Since γu has the same knot type for 0 ă u ăa
k2{ph2 ` k2q, it suffices to prove that
γu is an ph, h` kq-torus knot for sufficiently small u. Consider the parameterized surface
pt, sq ÞÑ2
π
¨
˚
˚
˚
˚
˚
˚
˚
˚
˚
˚
˚
˚
˚
˝
u?
1´ u2
h` ksinpπsq
cospπtq
ˆ
´u
kcospπsq `
?1´ u2
h
˙
sinpπtq
ˆ
´u
kcospπsq `
?1´ u2
h
˙
˛
‹
‹
‹
‹
‹
‹
‹
‹
‹
‹
‹
‹
‹
‚
.
In this form, it is easy to see that for u ă 1{2, the surface is an embedded torus of revolution
with elliptical cross sections. Making the change of variables s ÞÑ ph ` kqt and t ÞÑ ht, we
obtain a loop on the torus
rγuptq :“2
π
¨
˚
˚
˚
˚
˚
˚
˚
˚
˚
˚
˚
˚
˚
˝
u?
1´ u2
h` ksinpph` kqπtq
cosphπtq
ˆ
´u
kcospph` kqπtq `
?1´ u2
h
˙
sinphπtq
ˆ
´u
kcospph` kqπtq `
?1´ u2
h
˙
˛
‹
‹
‹
‹
‹
‹
‹
‹
‹
‹
‹
‹
‹
‚
which is clearly an ph, h` kq-torus knot.
191
Noting that
cospph` kqπtq cosphπtq “ cospkπtq ´ sinphπtq sinpph` kqπtq
and
cospph` kqπtq sinphπtq “ ´ sinpkπtq ` cosphπtq sinpph` kqπtq,
we conclude that
rγuptq “ γuptq `u
k
¨
˚
˚
˚
˚
˝
0
sinphπtq sinpph` kqπtq
´ cosphπtq sinpph` kqπtq
˛
‹
‹
‹
‹
‚
Therefore γu and rγu are isotopic for sufficiently small u ą 0, and γu is an ph, h ` kq-torus
knot.
Lemma 5.3.15. Assume that k and h` k are relatively prime. Then fora
k2{ph2 ` k2q ă
u ă 1, γu parameterizes a p´k, h` kq-torus knot.
Proof. The proof is the same as that of Lemma 5.3.14, except we start with the parameterized
surface
pt, sq ÞÑ2
π
¨
˚
˚
˚
˚
˚
˚
˚
˚
˚
˚
˚
˚
˚
˝
u?
1´ u2
h` ksinpπsq
cospπtq
ˆ
´u
k`
?1´ u2
hcospπsq
˙
sinpπtq
ˆ
´u
k`
?1´ u2
hcospπsq
˙
˛
‹
‹
‹
‹
‹
‹
‹
‹
‹
‹
‹
‹
‹
‚
,
192
which is an embedded torus for u ą 1{2, and we make the change of variables s ÞÑ ph` kqt
and t ÞÑ ´kt. This produces a p´k, h` kq-torus knot rγu which can be written as
rγuptq “ γuptq ´
?1´ u2
h
¨
˚
˚
˚
˚
˝
0
sinpkπtq sinpph` kqπtq
cospkπtq sinpph` kqπtq
˛
‹
‹
‹
‹
‚
and we conclude that γu is isotopic to rγu for sufficiently large u ă 1, so γu is a p´k, h`kq-torus
knot.
The results of this subsection are summarized as:
Theorem 5.3.16. Let h, k be integers such that gcdph, h ` kq “ gcdpk, h ` kq “ 1. The
critical point set of E :M Ñ R contains a 1-parameter family rγu, Vus, u P r0, 1s, such that
there exists a unique u P p0, 1q where γu is nonembedded, and:
(i) pγ0, V0q is an arclength parameterized h-times-covered round circle linked k-times,
(ii) γε is a ph, h` kq-torus knot for sufficiently small ε ą 0,
(iii) γ1´ε is a p´k, h` kq torus knot for sufficiently small ε ą 0,
(iv) pγ1, V1q is an arclength parameterized k-times-covered round circle linked ´h-times.
An example is shown in Figure 5.3.4 of a 1-parameter family of critical points for the
h “ 2, k “ 1 case. Each shape in the figure is a ribbon parameterized according to pγu, Vuq
with centerline γu (in black). Note that the times are not equally separated and that the
curve is nonembedded ata
1{5 “a
k2{ph2 ` k2q. Also note that the u “ 0 ribbon appears
as a Mobius band because the centerline is actually double-covered.
193
0 0.1 0.25 0.4 1/5
0.5 0.75 0.99 1
Figure 5.2: The 1-parameter family of critical points with h “ 2, k “ 1.
Another example is shown in Figure 5.3.4 with h “ 3 and k “ 1. In this example, only
the base curve γu is shown for clarity. The starting and ending points are omitted from the
figure, as they are multiply covered circles.
194
0.1 0.4 4/13 0.6 0.99
Figure 5.3: The 1-parameter family of critical points with h “ 3, k “ 2.
195
Chapter 6
Application: Shape Recognition
6.1 Introduction
6.1.1 Background
The Riemannian geometry of infinite-dimensional shape manifolds of paths or loops in Rn has
become an active field of research over the last decade. Interest in these spaces is driven by
their applications to computer vision (see the survey papers [3, 69]). Such a shape manifold
is typically realized as a quotient of the space of immersions
ImmpΣ,Rnq “ tγ P C8pΣ,Rn
q | γ1ptq ‰ 0 for all t P Σu ,
where Σ is an interval or S1. To form the shape space, we quotient by “shape-preserving”
group actions—usually isometries and homotheties of the ambient Rn and diffeomorphisms
of Σ, which act by reparameterizing the immersed manifold. The points (equivalence classes
of curves) in the quotient manifold are “shapes” of immersed copies of Σ. By defining
a Riemannian metric on the shape space, we can (in theory) measure geodesic distance
196
between shapes, thereby defining a natural, geometrically-motivated distance function on
the space of shapes.
Example 6.1.1. In the case n “ 2 and Σ “ S1, the elements of the shape space
ImmpS1,R2q{ttranslation, rotation, scaling, reparameterizationu
are viewed as outlines of images of objects. Geodesic distance between points of the shape
space is a measure of distance between shapes. This yields an object classification algorithm.
Example 6.1.2. As a more relevant example, we consider facial recognition. The facial
recognition problem is mathematically realized as defining a distance metric on the space of
immersed surfaces in R3, up to natural equivalences. A computationally effective approach
to elastic shape recognition for surfaces is to compare collections of closed curves immersed
in the surfaces [65]. Curves on an immersed surface in R3 have a natural framing, thus
the surface shape matching problem can be formulated as shape matching of collections of
framed loops. Therefore shape similarity for faces can be measured by geodesic distance in
M or M{Diff`0 pS1q.
To study the Riemannian geometries of shape spaces of paths or loops in Rn, one begins by
choosing a convenient metric on the space of immersions ImmpΣ,Rnq, where Σ is an interval
or S1. Techniques from Riemannian geometry are then applied to study the geodesics in the
more complicated shape space.
The simplest choice of reparameterization-invariant metric on the space of immersions is
the L2 metric: for γ P ImmpΣ,Rnq, let δγ1 and δγ2 denote variations of γ and define the L2
metric by
gL2
γ pδγ1, δγ2q “
ż
Σ
〈δγ1, δγ2〉 ds,
197
where ds is measure with respect to arclength. It was shown in [52] that geodesic distance
with respect to the L2 metric vanishes, so this metric is not suitable for shape recognition
applications.
A wide variety of more complicated Riemannian metrics have been introduced in the
literature, largely produced by adjusting the L2 metric in some manner. The Sobolev-type
metrics [53, 71], which include inner products of higher-order derivatives of the variations
δγj in the integral, are particularly well-suited to shape recognition problems. For example,
it was shown in [7] that ImmpS1,R2q is geodesically complete with respect to a family of
Sobolev-type metrics.
The elastic metrics of [57], which we discussed previously in Section 3.1.1, are a partic-
ularly well-studied family of Sobolev-type metrics on ImmpΣ,R2q of the form
ga,bγ pδγ1, δγ2q “
ż
Σ
a
⟨d
dsδγ1, T
⟩⟨d
dsδγ2, T
⟩` b
⟨d
dsδγ1, N
⟩⟨d
dsδγ2, N
⟩ds, (6.1)
where a, b ą 0, d{ds is derivative with respect to arclength, N is the unit normal to γ and
T is the unit tangent to γ. The first term compares the bending deformation of γ induced
by the variations δγ1 and δγ2, while the second term compares the stretching deformation.
Remarkably, it was shown in [76] that geodesics may be described explicitly with respect to
this metric in the case a “ b. In Section 3.1.1 we introduced an analogous family of metrics
on the space of framed curves and we showed in Section 4.5.1 that geodesics in S and M
are explicitly computable with respect to this metric.
A common technique used to study geodesics in immersion space with respect to a chosen
metric is to flatten the metric by applying a transform. This entails isometrically mapping
the immersion space into some other Riemannian manifold with a metric which is simpler to
work with. Examples include the Square Root Velocity Function-transform [68], which maps
γ to γ1{a
}γ1}, the Q-transform [41], which maps γ toa
}γ1}γ, and the complex square-root
198
transform studied in [76] (it was recently shown in [2] that, in the planar case, these are all
specific examples of the so-called R-transform). The complex square-root transform is only
defined for planar curves γ, and the transform takes the curve to?γ1, where γ1 is treated as a
complex number. The proof of Theorem 3.2.8 presented in this work made use of a transform
which may be viewed as a quaternionic analogue of the complex square root transform.
6.1.2 Description of Results
In this chapter we are particularly interested in shape matching algorithms for the shape
spaces of open framed paths S{SOp3q and S{pSOp3q ˆ Diff`pr0, 2sqq. The reason for this is
that these shape spaces have the most obvious applications. We are particularly interested
in shape matching for protein backbones, as the shape of a protein backbone is well-known
to correlate to its function. Another potential application of this theory would be oriented
trajectory recognition for, e.g., automated aircraft training or gesture recognition.
We show that the geodesic distance problem for S{SOp3q has an exact closed-form solu-
tion. We then move on to treating the geodesic distance problem for S{pSOp3qˆDiff`pr0, 2sqq.
We will see that geodesic distance degerates on S{pSOp3qˆDiff`pr0, 2sqq and this makes the
space non-Hausdorff. This is remedied by identifying elements of S{SOp3q up to L2 con-
vergence, thereby producing a Hausdorff space. Geodesic distance in this space is treated
computationally using a dynamic programming algorithm.
6.2 Shape Matching for Framed Paths
6.2.1 Lifting a Framed Path
In order to perform shape analysis using our coordinate system, we need a concrete descrip-
tion of how to choose a representative Φ P S˝pPC2q of an arbitrary framed path pγ, V q P S.
199
In this chapter it will frequently be useful to work in quaternionic coordinates. Accordingly,
we define SpPHq to be the L2-sphere in PH (with L2 metric induced by the identification
C2 « H) and
S˝pPHq :“ tq P SpPHq | qptq ‰ 0 @ tu.
In order to lift to the L2-sphere we are assuming that lengthpγq “ 2, which can be easily
arranged as a preprocessing step.
Our approach to lifting a framed path is to use axis-angle coordinates. Any element
of SOp3q besides ˘Id3ˆ3 has a unique representation as p~n, βq up to multiplication by ´1,
where ~n P R3 is a unit vector along the 1-dimensional subspace fixed by the rotation, and
β is the angle rotated around this axis. Since the identity corresponds to a rotation by
β “ 0, there is no well-defined choice of ~n to represent the identity. Note that p~n, βq and
p´~n,´βq clearly correspond to the same rotation. There are easy and numerically stable
algorithms for determining the axis-angle coordinates of rotation matrix—see, e.g., Chapter
16 of [27]. Moreover, it is well-known that the unit quaternion corresponding to SOp3q under
the quaternionic version of FrameHopf is cospβ{2q ` sinpβ{2q~n. In this representation, we
are thinking of ~n as an element of the purely imaginary quaternions under their obvious
identification with R3.
Thus we have a lifting procedure taking a framed path pγ, V q P S to an element of
S˝pPC2q, modulo a technical issue which we will address momentarily. The procedure first
represents pγ, V q as ppT, V,W q, }γ1}q P PpSOp3q ˆ R`q. We then convert to axis-angle co-
ordinates pp~nptq, βptqq, }γ1ptq}q, choosing p~nptq, βptqq to be a smooth path in S2 ˆ R. By
continuity, there is a unique way to make this choice up to global multiplication by ´1.
Finally, we lift to
qptq “a
}γ1} pcospβ{2q ` ~n sinpβ{2qq .
200
The remaining technical issue that this procedure is not well-defined for a framed curve
pγ, V q such that pT ptq, V ptq,W ptqq “ Id3ˆ3 for some t P r0, 2s. We get around this issue
by noting that we are really interested in equivalence classes up to rotations. Thus we can
preprocess by applying some rotation to pγ, V q so that its image curve in SOp3q does not hit
the identity. A generic framed curve has the property that pT ptq, V ptq,W ptqq ‰ Id3ˆ3 for all
t, so such rotations are plentiful.
6.2.2 Optimizing Geodesic Distance Over Rotations
The moduli space of framed paths S admits a free action of SOp3q by rotations. It is easy
to see that the natural metric gS on S is invariant under the SOp3q action, thus gS descends
to a well-defined metric on the quotient space S{SOp3q. We described explicit geodesics in
S in Section 4.5.1 and our goal in this section is to describe geodesics in S{SOp3q. This can
be restated as finding a horizontal geodesic in S with respect to SOp3q. That is, let pγ0, V0q,
pγ1, V1q P S. Then the horizontal geodesic joining the equivalence classes in S{SOp3q is
realized as the geodesic in S between pγ0, V0q and A ¨ pγ1, V1q, where A P SOp3q is given by
A “ argmintdistSppγ0, V0q, B ¨ pγ1, V1qq | B P SOp3qu
and distS is geodesic distance in S. Thus geodesic distance in S{SOp3q between the equiva-
lence classes of framed paths is
mintdistSppγ0, V0q, A ¨ pγ1, V1qq | A P SOp3qu. (6.2)
The goal of this section is to give explicit descriptions of the minimizing rotation A and
geodesic distance distS .
201
We remark here that algorithms for optimizing geodesic distance over rotations are abun-
dant in the elastic shape recognition literature for a variety of shape space regimes. These
algorithms typically involve matrix decompositions. Such decompositions are numerically
efficient, but it is nonetheless useful and aesthetically pleasing to have a closed-form solution
for the rotation registration problem. Our closed-form solution in quaternionic coordinates
also allows us to write out solutions in complex and axis-angle coordinates. This allows us
to compute rotation-optimized geodesic distance without the need to lift data points from
S to S˝pPHq.
Recall from Theorem 3.1.12 that S is isometrically double-covered by the open subset of
the L2 sphere in PC2
S˝pPC2q “ tΦ P PpC2
zt0uq | 〈Φ,Φ〉L2 “?
2u.
We saw in Equation 3.15 that the right action of SUp2q on S˝pPC2q corresponds under pH
to the rotation action of SOp3q on S. The SUp2q-action in the covering space turns out
to be simpler to work with than the SOp3q-action on S. If Φj P S˝pPC2q corresponds to
pγj, Vjq P S for j “ 0, 1, then we can replace the computation of (6.2) with the computation
min tdistpΦ0,Φ1 ¨ Aq | A P SUp2qu , (6.3)
where dist is geodesic distance in SpPHq—for the remainder of this chapter, we slightly abuse
notation and take the convention that dist always refers to geodesic distance in whichever
space makes sense in context. The computation (6.3) can in turn be simplified by applying
the identifications of Section 2.4. Namely, let qj P PH correspond to Φj for j “ 0, 1 under the
identification induced by C2 « H. We also identify SUp2q « S3. The next lemma translates
the SUp2q-action into quaternionic coordinates.
202
Lemma 6.2.1. Let Φ “ pφ, ψq P PC2 correspond to the quaternionic path q P PC2 and let
A “
¨
˚
˝
u v
´v u
˛
‹
‚
P SUp2q
correspond to p “ u` vj P S3. Then the right action of A on Φ corresponds to quaternionic
right multiplication of q by p, q ¨ p.
Proof. This is a straightforward calculation. The group action in complex coordinates is
Φ ¨ A “ puφ´ vψ, vφ` uψq.
On the other hand, p acts on q “ φ` ψj by
q ¨ p “ pφ` ψjq ¨ pu` vjq “ pφu` ψjvjq ` pφvj` ψjuq “ puφ´ vψq ` pvφ` uwqj, (6.4)
where the second equality follows by noting that for any complex number x` iy,
jpx` iyq “ xj´ yk “ px´ iyqj. (6.5)
The right hand side of (6.4) is identified with puφ´ vψ, vφ` uψq.
Then (6.2) becomes
min
distpq0, q1 ¨ pq | p P S3(
. (6.6)
From the discussion in Section 4.5.1 and the fact that
Re 〈Φ0,Φ1〉C2 “ Repq0 ¨ q1q
203
we see that dist is given in quaternionic coordinates by
distpq0, q1q “?
2 arccos1
2Re
ż 2
0
q0q1 dt.
The next theorem shows that (6.6) has a simple closed-form solution.
Theorem 6.2.2. Let q0 and q1 denote points of SpPC2q in quaternionic coordinates. Then
argmin
distpq0, q1 ¨ pq | p P S3(
“pq
}pq}H,
where
pq “
ż 2
0
q1 ¨ q0 dt P H.
We conclude that geodesic distance in SpPC2q{SUp2q is given by
min
distpq0, q1 ¨ pq | p P S3(
“?
2 arccos1
2
›
›
›
›
ż 2
0
q1 ¨ q0 dt
›
›
›
›
H.
Proof. Let q0, q1 P SpPHq. Our goal is to find a solution to
pp :“ argmin
distpq0, q1 ¨ pq | p P S3(
“ argmin
"
?2 arccos
ˆ
1
2Re
ż 1
0
q0pq1 ¨ pq dt
˙
| p P S3
*
.
Any solution pp satisfies
pp “ argmax
"
Re
ż 2
0
q0pq1 ¨ pq dt | p P S3
*
“ argmax
"
Re
ż 2
0
q0 ¨ p ¨ q1 dt | p P S3
*
“ argmax
"
Re
ˆ
p
ż 2
0
q1 ¨ q0 dt
˙
| p P S3
*
.
204
The last line follows by noting that for any quaternions q “ Q0 ` iQ1 ` jQ2 ` kQ3 and
p “ P0 ` iP1 ` jP2 ` kP3,
Repqpq “ Q0P0 ´Q1P1 ´Q2P2 ´Q3P3 “ Reppqq,
and that p is a constant.
Let
pq :“
ż 2
0
q1 ¨ q0 dt P H.
We claim that pp “ pq{}pq}H. Indeed, under the identification H « R4,
Re
ˆ
p
ż 2
0
q1 ¨ q0 dt
˙
“ 〈p, pq〉R4 ,
and this claim is just an application of the obvious fact that for any w P Rn,
argmaxt〈v, w〉Rn | v P Sn´1
Ă Rnu “
w
}w}.
Therefore geodesic distance between the equivalence classes of q0 and q1 in SpPC2q{SUp2q is
given by
min
distpq0, q1 ¨ pq | p P S3(
“?
2 arccos1
2Re
˜
pq
}pq}Hpq
¸
“?
2 arccos1
2}pq}H
“?
2 arccos1
2
›
›
›
›
ż 2
0
q1 ¨ q0 dt
›
›
›
›
H
We easily obtain the geodesic distance formula in complex and axis-angle coordinates.
205
Corollary 6.2.3. Geodesic distance between equivalence classes of points Φ0 and Φ1 in
SpPC2q{SUp2q is given by?
2 arccos 12χ, where
χ2“
›
›
›
›
ż 2
0
φ0φ1 ` ψ0ψ1 dt
›
›
›
›
2
C`
›
›
›
›
ż 2
0
ψ0φ1 ´ φ0ψ1 dt
›
›
›
›
2
C
If
pγ0, V0q “ pp~n0, β0q, }γ10}q, and pγ1, V1q “ pp~n1, β1q, }γ
11}q
are axis-angle representations of the framed curves associated to Φ0 and Φ1, then geodesic
distance between the equivalence classes of these points in S{SOp3q is given by?
2 arccos 12ζ,
where
ζ2“
›
›
›
›
ż 2
0
a
}γ10} ¨ }γ11}
ˆ
cos
ˆ
β0
2
˙
cos
ˆ
β1
2
˙
` sinpβ0{2q sinpβ1{2q 〈~n0, ~n1〉˙
dt
›
›
›
›
2
R
`
›
›
›
›
ż 2
0
a
}γ10} ¨ }γ11}
ˆ
sin
ˆ
β0
2
˙
cos
ˆ
β1
2
˙
~n0 ´ cos
ˆ
β0
2
˙
sin
ˆ
β1
2
˙
~n1
´ sin
ˆ
β0
2
˙
sin
ˆ
β1
2
˙
~n0 ˆ ~n1
˙
dt
›
›
›
›
2
R3
.
Proof. Let points Φj “ pφj, ψjq, for j “ 1, 2, be identified with the quaternionic paths
qj “ φj ` ψjj. Once again using the observation (6.5), we obtain
q0 ¨ q1 “ pφ0 ` ψ0jq ¨ pφ1 ´ ψ1jq
“ pφ0φ1 ` ψ0ψ1q ` pψ0φ1 ´ φ0ψ1qj.
Thus
›
›
›
›
ż 2
0
q0 ¨ q1 dt
›
›
›
›
2
H“
›
›
›
›
ż 2
0
pφ0φ1 ` ψ0ψ1q ` pψ0φ1 ´ φ0ψ1qj dt
›
›
›
›
2
H
“
›
›
›
›
ż 2
0
φ0φ1 ` ψ0ψ1 dt
›
›
›
›
2
C`
›
›
›
›
ż 2
0
ψ0φ1 ´ φ0ψ1 dt
›
›
›
›
2
C.
206
To prove the second claim, write qj “b
}γ1j}pcospβj{2q ` ~nj sinpβj{2qq for j “ 0, 2. Then
q0 ¨ q1 “a
}γ10} ¨ }γ11}pcospβ0{2q ` ~n0 sinpβ0{2qqpcospβ1{2q ´ ~n1 sinpβ1{2qq
“a
}γ10} ¨ }γ11}pcospβ0{2q cospβ1{2q ` sinpβ0{2q sinpβ1{2q 〈~n0, ~n1〉` sinpβ0{2q cospβ1{2q~n0
´ cospβ0{2q sinpβ1{2q~n1 ´ sinpβ0{2q sinpβ1{2q~n0 ˆ ~n1q.
The last line makes use of the following identity for multiplication of purely imaginary
quaternions q “ Q1i`Q2j`Q3k and p “ P1i` P2j` P3k:
q ¨ p “ ´pQ1P1 `Q2P2 `Q3P3q ` pQ2P3 ´Q3P2qi` p´Q1P3 `Q3P1qj` pQ1P2 ´Q2P1qk
“ ´ 〈q, p〉` q ˆ p,
where we are utilizing the identification of purely imaginary quaternions with R3 to write
the last line. Now the result follows by integrating, taking the quaternionic norm and
decomposing accordingly.
6.2.3 Modding out by the Diff`-Action
In this section we aim to describe geodesic distance in the shape space
S{pSOp3q ˆDiff`pr0, 2sqq,
where Diff`pr0, 2sq acts as usual by reparameterizations (see Section 4.4), but we will see
momentarily that there is a problem with this goal. We first record the following lemma,
whose proof is obvious by definition.
Lemma 6.2.4. The actions of SOp3q and Diff`pr0, 2sq on S commute.
207
Thus we can first consider only modding out by Diff`pr0, 2sq. We wish to define a dis-
tance metric on S{Diff`pr0, 2sq by optimizing geodesic distance between points in S over
reparameterizations. Recall that the geodesics of S are the geodesics of S˝pPC2q, but that a
geodesic between points of S˝pPC2q may cross through the set SpPC2qzS˝pPC2q. Nonethe-
less, geodesic distance in SpPC2q is still a well-defined, geometrically-motivated metric on
S. Thus we might try to shift the problem of geodesic distance in S{Diff`pr0, 2sq to that
of geodesic distance in SpPC2q{Diff`pr0, 2sq. Here we run into a major problem in that this
quotient is not Hausdorff. This can be demonstrated by adapting an example from Section
3.5 of [76] as follows.
Example 6.2.1. Let ρ : r0, 2s Ñ r0, 2s be a smooth bijective map which is non-decreasing
and constant on some subinterval I Ă r0, 2s. The sequence of diffeomorphisms tρnu defined
by ρnptq “ p1´ 1{nqρptq ` t{n converges in C8 to ρ. For any Φ P S˝pPC2q, this sequence of
diffeomorphisms acts on Φ to produce a sequence tΦnu in the Diff`pr0, 2sq-orbit of Φ given
by
Φnptq “a
p1´ 1{nqρ1ptqΦpp1´ 1{nqρptq ` t{nq.
This sequence converges in C8 to a smooth path Φ8 P SpPC2q. Moreover, we see that Φ8
lies in SpPC2qzS˝pPC2q, since Φ8 has norm zero on the interval I. Since the Diff`pr0, 2sq-
action fixes S˝pPC2q, it must be that Φ8 is not in the Diff`pr0, 2sq-orbit of Φ. We conclude
that the Diff`pr0, 2sq-orbits of Φ and Φ8 are distinct but arbitrarily close to one another,
thus SpPC2q is not Hausdorff.
The problem boils down to the fact that the Diff`pr0, 2sq-orbits are not closed. The most
obvious remedy for this issue is to partition SpPC2q into slightly larger equivalence classes.
One possibility is to consider elements of SpPC2q up to Frechet equivalence, where we enlarge
the quotient set from Diff`pr0, 2sq to the set of monotone relations (see [76], Section 3.5 for
208
a definition). Another approach, and the one we adopt here, is to partition SpPC2q into
closures of Diff`pr0, 2sq orbits (this is the approach taken in, e.g., [64, 68]).
Let us precisely define the space of interest. We define the shape space of framed paths
pS to be the collection of equivalences classes JΦK, where JΦK is the set
JΦK “ closureL2pta
ρ1Φpρq | ρ P Diff`pr0, 2squq X SpPC2q.
That is, Φ P JΦ0K if and only if Φ P SpPC2q and there exists a sequence of diffeomorphisms
ρn such that Φ0pρnq Ñ Φ in L2. We claim that optimized geodesic distance now gives a
well-defined distance metric on pS. Define
dpSpJΦ0K, JΦ1Kq “ inftdistpΦ0,
a
ρ1Φpρqq | ρ P Diff`pr0, 2squ,
where dist is geodesic distance in SpPC2q. This can be written in quaternionic coordinates
as
dpSpJq0K, Jq1Kq “ inf
"
?2 arccos
1
2Re
ż 2
0
a
ρ1 ¨ q0 ¨ q1pρq dt | ρ P Diff`pr0, 2sq
*
.
We now show that the above definitions are reasonable from two perspectives. First, it
is necessary to show that dpS is actually a distance metric.
Proposition 6.2.5. The map dpS is a distance metric on pS.
Proof. The most important point here is that
dpSpJq0K, Jq1Kq “ 0 ô Jq0K “ Jq1K.
209
The left implication is obvious, so we focus on the right implication. If the distance between
Jq0K and Jq1K is zero, then there exists a sequence of diffeomorphisms ρn such that
Re
ż 2
0
a
ρ1n ¨ q1pρnq ¨ q0 dtÑ 2 as nÑ 8,
which impliesż 2
0
a
ρ1n ¨ q0 ¨ q1pρq dtÑ 2 as nÑ 8.
In other words,?ρ1n ¨ q1pρnq converges weakly in L2 to q0. On the other hand,
?ρ1n ¨ q1pρnq
and q0 are elements of the L2-sphere SpPHq, so their L2-norms agree for all n. Together
these statements imply that?ρ1n ¨ q1pρnq Ñ q0 in L2. But then Jq1K “ Jq0K, and this proves
the claim.
The map dpS is obviously nonnegative-valued and symmetric, and the triangle inequality
follows because it is satisfied by geodesic distance, dist.
Next we wish to show that these equivalence classes are geometrically meaningful in some
sense; i.e., that they are not so large that they are no longer useful for shape recognition
applications.
Proposition 6.2.6. Let q P Jq0K and let pH1pqq “ γ and pH1pq0q “ γ0. Then the images of γ
and γ0 agree.
Proof. Let tρnu be a sequence of diffeomorphisms such that?ρ1nq0pρq Ñ q in L2. Then
there exists a subsequence tρnku such thata
ρ1nkq0pρnkq Ñ q in L2 and pointwise almost
everywhere. Therefore γ0pρnkq Ñ γ pointwise almost everywhere. Let E a subset of r0, 2s
such that γpρnkq converges pointwise on E and such that r0, 2szE has zero measure. We
first claim that the images of γ0 and γ agree on E. Indeed, for t P E, let t0 “ lim ρnkptq.
Then γ0pρnkptqq Ñ γ0pt0q “ γptq. Similarly, for fixed t0 P E, let t “ lim ρ´1nkpt0q. Then
210
γ0pρnkptqq Ñ γptq “ γpt0q. Therefore the images of γ0 and γ agree on E. Since each curve is
smooth, we conclude that the images of γ0 and γ agree on all of r0, 2s.
6.2.4 Computationally Optimizing Over Reparameterizations
We have shown that dpS is a well-defined metric. Now the question becomes whether it
can actually be computed. In this section, we adapt a well-known dynamic programming
algorithm in order to approximate dpS . The approach taken here follows the exposition in
[64], Section 3.4.2.
The task is accomplished by discretizing and approximating diffeomorphisms of r0, 2s
by piecewise linear homeomorphisms of r0, 2s with vertices on a fixed grid such that each
segment has positive slope. The general algorithm begins by choosing partitions 0 “ t11 ă
t12 ă ¨ ¨ ¨ ă t1n1“ 2 of the x-axis and 0 “ t21 ă t22 ă ¨ ¨ ¨ ă t2n2
“ 2 of the y-axis of r0, 2s ˆ r0, 2s.
For the sake of simplicity, let us take the partitions to be the same and evenly spaced, so
that we can use the notation 0 “ t1 ă t2 ă ¨ ¨ ¨ ă tn “ 2. Then gridpoints on r0, 2s ˆ r0, 2s
take the form
a “ aij “ pti, tjq “
ˆ
2pi´ 1q
n´ 1,2pj ´ 1q
n´ 1
˙
.
Each grid point has a collection of potential successors. These are points ak` on the grid such
that the segment connecting ak` to aij has positive slope and does not pass through any other
grid point. The first condition is expressed simply as k ă i and ` ă j. The second condition
is included so that paths can be unambiguously expressed as concatenations of segments
joining vertices, and is expressed as gcdpk ´ i, `´ jq “ 1 (making use of the even spacing of
the partition). Thus admissible paths are represented as sequences pa1, a2, . . . , amq, where
each ai is a gridpoint, and we interpret the path as the concatenation of the edges joining ai
to ai`1.
211
We now introduce a variety of relevant functionals. For fixed q0, q1 P SpPHq and an
increasing linear diffeomorphism ρ : ra, bs Ñ rc, ds between subintervals of r0, 2s, we define
F pρq :“1
2Re
ż b
a
a
ρ1 ¨ q0 ¨ q1pρq dt
“1
2
ˆ
d´ c
b´ a
˙1{2
Re
ż b
a
q0 ¨ q1pρq dt.
This is referred to as the energy of ρ. For a grid point aij and an admissible successor ak` we
will use the notation F pak`, aijq for the energy of the linear diffeomorphism taking rtk, tis to
rt`, tjs. This can also be interpreted as the energy of the edge joining the points ak` and aij.
Remark 6.2.7. Note that the energy functional is derived from the geodesic distance between
q0 and q1 in SpPHq. One might consider using a functional based on the distance formula in
SpPHq{Up2q given by Theorem 6.2.2. We will see that the norm appearing in that formula
would cause problems with the proposed algorithm.
By a slight abuse of notation, we define the energy of an admissible path pa1, . . . , amq to
be the sum
F pa1, . . . , amq :“m´1ÿ
i“1
F pai, ai`1q.
Thus the energy of a path starting at a1 “ pt1, t1q “ p0, 0q and ending at am “ ptn, tnq “ p2, 2q
is, upon applying arccos and multiplying by?
2, an approximation of the geodesic distance
in SpPCq between Φ0 and?ρ1Φ1pρq, where ρ is the diffeomorphism approximated by the PL
path pa1, . . . , amq.
Turning toward optimizing over such PL paths, we define Dpaq to be the energy of the
optimal path joining the origin to the gridpoint a. By optimal we mean of maximum energy.
Large energy will be converted to small geodesic distance after taking arccos. More precisely,
212
we define
Dpaq :“ maxtF pa1, . . . , amq | pa1, . . . , amq is admissible, a1“ p0, 0q, am “ au.
Thus the problem of minimizing geodesic distance over admissible paths is equivalent to
computing Dp2, 2q.
The following lemma is standard (see, e.g., [64]), but it is important in that it shows that
dynamic programming is an appropriate approach to computing Dpaq.
Lemma 6.2.8. If Dpaq is realized by pa1, . . . , am “ aq, then for any grid point ai in this
path, Dpaiq is realized by pa1, . . . , aiq.
Proof. Let pa1, . . . , am “ aq be such an optimal path. If Dpaiq is not realized by pa1, . . . , aiq,
then there exists an admissible path pb1, . . . , bj “ aiq with strictly higher energy. But then
the concatenation pb1, . . . , bj, ai`1, . . . , amq is an admissible path ending at a with higher
energy than the original minimizing path, and we have obtained a contradiction.
Lemma 6.2.8 means that the problem of finding the optimal PL ρ is amenable to the
dynamic programming approach, and we take our functional equation to be
Dpaq “ maxtDpa1q ` F pa1, aq | a1 is an admissible predecessor for au.
The dynamic programming algorithm calculates Dp2, 2q by calculating Dpaq for every grid-
point in row-by-row order. The procedure is outlined in Algorithm 1. The algorithm is
essentially the same as the one employed in, e.g., [57] and the pseudocode is essentially the
same as that of [64], Section 3.4.2. The difference here is that the energy functions used
are specific to our setup, and that the presence of arccos in the geodesic distance formula
causes us to find maximum energy paths (as opposed to minimum energy paths in other
algorithms).
213
Algorithm 1 PL Optimization
Input: Size of grid partition n, Edge energies F pa1, a2q for all grid points a1 and a2
Output: Energy minimizing path Π from p0, 0q to p2, 2q and its energy Dp2, 2qInitialize D: Set Dp0, 0q “ 0, Dpaq “ ´8 for all other gridpoints aInitialize Π: Set Π “ pp2, 2qq
1: for r “ 1 to n do2: for c “ 1 to n do3: for Each predecessor a of acr do4: if Dpaq ` F pa, acrq ą Dpacrq then5: Set Dpacrq “ Dpaq ` F pa, acrq6: Set P pacrq “ a7: end if8: end for9: end for10: end for11: Set a “ p2, 2q12: while a ‰ p0, 0q do13: Set a “ P paq14: Prepend a to Π15: end while
We explore a simple example in Figures 6.1, 6.2 and 6.3. Figures 6.1 and 6.2 show
geodesics between a helix and a semicircle. Each curve has its Frenet framing, which is not
pictured for visual clarity. Figure 6.1 shows the geodesic in SpPHq, where the beginning
and ending curves have arclength parameterizations. The next figure shows the geodesic
after the helix has been optimized over Diff`pr0, 2sq. We can see from the figures that the
geodesics look quite different. Figure 6.3 gives a visualization of the reparameterization of
the helix and shows the corresponding gedesic distances.
A major source of future work will be to rigorously apply this algorithm to perform shape
recognition and modelling for, e.g., protein backbones, immersed surfaces, camera tracking,
root systems, hand gestures and aircraft trajectories.
214
Figure 6.1: Geodesic before reparameterization.
Figure 6.2: Geodesic after reparameterization.
215
~1.55 ~0.98
Figure 6.3: Geodesic distances.
216
Bibliography
[1] Bauer, M., Harms, P. and Michor, P.W. (2012). Almost local metrics on shape space of
hypersurfaces in n-space. SIAM Journal on Imaging Sciences, 5(1), 244-310.
[2] Bauer, M., Bruveris, M., Marsland, S. and Michor, P.W. (2014). Constructing reparam-
eterization invariant metrics on spaces of plane curves. Differential Geometry and its
Applications, 34, 139-165.
[3] Bauer, M., Bruveris, M. and Michor, P.W. (2016). Why use Sobolev metrics on the
space of curves. In Riemannian Computing in Computer Vision ( 233-255). Springer
International Publishing.
[4] Bishop, R.L. (1975). There is more than one way to frame a curve. The American Math-
ematical Monthly, 82(3), 246-251.
[5] Boman, J. (1967). Differentiability of a Function and of its Compositions with Functions
of One Variable. Mathematica Scandinavica, 20, 249-268.
[6] Bower, A.F. (2009). Applied mechanics of solids. CRC press.
[7] Bruveris, M., Michor, P.W. and Mumford, D. (2014). Geodesic completeness for Sobolev
metrics on the space of immersed plane curves. In Forum of Mathematics, Sigma (Vol.
2, p. e19). Cambridge University Press.
217
[8] Brylinski, J.L. (2007). Loop spaces, characteristic classes and geometric quantization
(Vol. 107). Springer Science & Business Media.
[9] Carroll, D., Kse, E. and Sterling, I. (2013). Improving Frenet’s Frame Using Bishop’s
Frame. arXiv preprint arXiv:1311.5857.
[10] Cantarella, J., Deguchi, T. and Shonkwiler, C. (2014). Probability theory of random
polygons from the quaternionic viewpoint. Communications on Pure and Applied Math-
ematics, 67(10), 1658-1699.
[11] Cantarella, J., Grosberg, A. Y., Kusner, R. and Shonkwiler, C. (2015). The expected
total curvature of random polygons. American Journal of Mathematics, 137(2), 411-438.
[12] Cheeger, J. and Ebin, D.G. (2008). Comparison theorems in Riemannian geometry (Vol.
365). American Mathematical Soc..
[13] Chirikjian, G.S. (2013). Framed curves and knotted DNA. Biochemical Society Trans-
actions, 41(2), 635-638.
[14] Dennis, M.R. and Hannay, J.H. (2005). Geometry of Clugreanu’s theorem. In Proceed-
ings of the Royal Society of London A: Mathematical, Physical and Engineering Sciences
(Vol. 461, No. 2062, 3245-3254). The Royal Society.
[15] Do Carmo, M.P. and Do Carmo, M.P. (1976). Differential geometry of curves and sur-
faces (Vol. 2). Englewood Cliffs: Prentice-hall.
[16] Edelman, A., Arias, T.A. and Smith, S.T. (1998). The geometry of algorithms with
orthogonality constraints. SIAM journal on Matrix Analysis and Applications, 20(2),
303-353.
[17] Eliashberg, Y. and Mishachev, N.M. (2002). Introduction to the h-principle. Providence:
American Mathematical Society.
218
[18] Fenchel, W. (1951). On the differential geometry of closed space curves. Bulletin of the
American Mathematical Society, 57(1), 44-54.
[19] Fuller, F.B. (1978). Decomposition of the linking number of a closed ribbon: a problem
from molecular biology. Proceedings of the National Academy of Sciences, 75(8), 3557-
3561.
[20] Fuller, F.B. (1971). The writhing number of a space curve. Proceedings of the National
Academy of Sciences, 68(4), 815-819.
[21] Gel’fand, I. M., Minlos, R. A. F. and Shapiro, Z. Y. (1963). Representations of the
Rotation and Lorentz Groups and their Applications.
[22] Grinevich, P.G. and Schmidt, M.U. (1997). Closed curves in R3: a characterization in
terms of curvature and torsion, the Hasimoto map and periodic solutions of the Filament
Equation. arXiv preprint dg-ga/9703020.
[23] Gromov, M. (2013). Partial differential relations (Vol. 9). Springer Science & Business
Media.
[24] Hamilton, R.S. (1982). The inverse function theorem of Nash and Moser. American
Mathematical Society, 7.
[25] Hannay, J.H. (1998). Cyclic rotations, contractibility and Gauss-Bonnet. Journal of
Physics A: Mathematical and General, 31(17), p.L321.
[26] Hanson, A.J. (1998). Constrained optimal framings of curves and surfaces using quater-
nion gauss maps. In Visualization’98. Proceedings ( 375-382). IEEE.
[27] Hanson, A.J. (2005). Visualizing quaternions. In ACM SIGGRAPH 2005 Courses (p.
1). ACM.
219
[28] Hanson, A.J. and Thakur, S. (2012). Quaternion maps of global protein structure. Jour-
nal of Molecular Graphics and Modelling, 38, 256-278.
[29] Harms, P. and Mennucci, A.C. (2012). Geodesics in infinite dimensional Stiefel and
Grassmann manifolds. Comptes Rendus Mathematique, 350(15), 773-776.
[30] Hausmann, J.C. and Knutson, A. (1997). Polygon spaces and Grassmannians. Enseigne-
ment mathmatique, 43(1/2), 173-198.
[31] Hoffman, K.A., Manning, R.S. and Maddocks, J.H. (2003). Link, twist, energy, and the
stability of DNA minicircles. Biopolymers, 70(2), 145-157.
[32] Howard, B., Manon, C. and Millson, J. (2008). The toric geometry of triangulated
polygons in Euclidean space. arXiv preprint arXiv:0810.1352.
[33] Hu, S., Lundgren, M. and Niemi, A.J. (2011). Discrete Frenet frame, inflection point
solitons, and curve visualization with applications to folded proteins. Physical Review E,
83(6), p.061908.
[34] Hwang, C.C. (1981). A differential-geometric criterion for a space curve to be closed.
Proceedings of the American Mathematical Society, 83(2), 357-361.
[35] Ivey, T.A. and Singer, D.A. (1999). Knot types, homotopies and stability of closed elastic
rods. Proceedings of the London Mathematical Society, 79(02), 429-450.
[36] Jurdjevic, V. (2009). The symplectic structure of curves in three dimensional spaces of
constant curvature and the equations of mathematical physics. In Annales de l’Institut
Henri Poincare (C) Non Linear Analysis (Vol. 26, No. 4, 1483-1515). Elsevier Masson.
[37] Kapovich, M. and Millson, J. (1996). The symplectic geometry of polygons in Euclidean
space. J. Differential Geom, 44(3), 479-513.
220
[38] Kierfeld, J., Niamploy, O., Sa-Yakanit, V. and Lipowsky, R. (2004). Stretching of semi-
flexible polymers with elastic bonds. The European Physical Journal E, 14(1), 17-34.
[39] Kriegl, A. and Michor, P. W. (1997). The convenient setting of global analysis (No. 53).
American Mathematical Soc..
[40] Kuipers, J. B. (1999). Quaternions and rotation sequences (Vol. 66). Princeton: Prince-
ton university press.
[41] Kurtek, S., Klassen, E., Gore, J. C., Ding, Z. and Srivastava, A. (2012). Elastic geodesic
paths in shape space of parameterized surfaces. Pattern Analysis and Machine Intelligence,
IEEE Transactions on, 34(9), 1717-1730.
[42] Langer, J. and Singer, D. A. (1984). Knotted elastic curves in R3. J. London Math.
Soc.(2), 30, 512-520.
[43] Lempert, L. (1993). Loop spaces as complex manifolds. Journal of Differential Geometry,
38(3), 519-543.
[44] Levien, R. (2008). The elastica: a mathematical history. Electrical Engineering and
Computer Sciences University of California at Berkeley.
[45] Li, Y. and Maddocks, J. H. (1996). On the computation of equilibria of elastic rods.
Part I: Integrals, symmetry and a Hamiltonian formulation. Preprint, Department of
Mathematics, University of Maryland, College Park.
[46] Lickorish, W. R. (1962). A representation of orientable combinatorial 3-manifolds. An-
nals of Mathematics, 531-540.
[47] Liu, W., Srivastava, A. and Zhang, J. (2011). A mathematical framework for protein
structure comparison. PLoS Comput Biol, 7(2), e1001075.
221
[48] Marsden, J. and Weinstein, A. (1974). Reduction of symplectic manifolds with symmetry.
Reports on mathematical physics, 5(1), 121-130.
[49] Marsden, J. and Weinstein, A. (1983). Coadjoint orbits, vortices, and Clebsch variables
for incompressible fluids. Physica D: Nonlinear Phenomena, 7(1), 305-323.
[50] McDuff, D. and Salamon, D. (1998). Introduction to symplectic topology. Oxford Uni-
versity Press.
[51] Michor, P. W. and Mumford, D. (2003). Riemannian geometries on spaces of plane
curves. arXiv preprint math/0312384.
[52] Michor, P. W. and Mumford, D. (2005). Vanishing geodesic distance on spaces of sub-
manifolds and diffeomorphisms. Doc. Math, 10, 217-245.
[53] Michor, P. W. and Mumford, D. (2007). An overview of the Riemannian metrics on
spaces of curves using the Hamiltonian approach. Applied and Computational Harmonic
Analysis, 23(1), 74-113.
[54] Mumford, D., Fogarty, J. and Kirwan, F. C. (1994). Geometric invariant theory (Vol.
34). Springer Science & Business Media.
[55] Millson, J. J. and Zombro, B. (1996). A Khler structure on the moduli space of isometric
maps of a circle into Euclidean space. Inventiones mathematicae, 123(1), 35-59.
[56] Milnor, J. W. and Stasheff, J. D. (1974). Characteristic classes (No. 76). Princeton
university press.
[57] Mio, W., Srivastava, A. and Joshi, S. (2007). On shape of plane elastic curves. Interna-
tional Journal of Computer Vision, 73(3), 307-324.
222
[58] Moffatt, H. K. and Ricca, R. L. (1992). Helicity and the Calugareanu invariant. In
Proceedings of the Royal Society of London A: Mathematical, Physical and Engineering
Sciences (Vol. 439, No. 1906, pp. 411-429). The Royal Society.
[59] Neretin, Y. A. (2001). On Jordan angles and the triangle inequality in Grassmann man-
ifolds. Geometriae Dedicata, 86(1-3), 81-91.
[60] Nikolaevskii, Y. A. (1994). On Fenchel’s problem. Mathematical Notes, 56(5), 1158-
1164.
[61] O’Hara, J. (2003). Energy of knots and conformal geometry (Vol. 33). world scientific.
[62] Omori, H. (1978). On Banach-Lie groups acting on finite dimensional manifolds. Tohoku
Mathematical Journal, Second Series, 30(2), 223-250.
[63] O’Neill, B. (1966). The fundamental equations of a submersion. The Michigan Mathe-
matical Journal, 13(4), 459-469.
[64] Robinson, D. T. (2012). Functional data analysis and partial shape matching in the
square root velocity framework. Dissertation. The Florida State University.
[65] Samir, C., Srivastava, A., Daoudi, M. and Klassen, E. (2009). An intrinsic framework
for analysis of facial surfaces. International Journal of Computer Vision, 82(1), 80-95.
[66] Singer, D. A. (2008). Lectures on elastic curves and rods. In AIP Conference Proceedings
(Vol. 1002, No. 1, p. 3).
[67] Slovesnov, A. and Zakalyukin, V. (2013). Applications of Framed Space Curves. Journal
of Mathematical Sciences, 195(3).
[68] Srivastava, A., Klassen, E., Joshi, S. H. and Jermyn, I. H. (2011). Shape analysis of elas-
tic curves in euclidean spaces. Pattern Analysis and Machine Intelligence, IEEE Trans-
actions on, 33(7), 1415-1428.
223
[69] Srivastava, A., Turaga, P. and Kurtek, S. (2012). On advances in differential-geometric
approaches for 2D and 3D shape analyses and activity recognition. Image and Vision
Computing, 30(6), 398-416.
[70] Stacey, A. (2005). The differential topology of loop spaces. arXiv preprint math/0510097.
[71] Sundaramoorthi, G., Yezzi, A. and Mennucci, A. C. (2007). Sobolev active contours.
International Journal of Computer Vision, 73(3), 345-366.
[72] Taylor, W. R. and Aszdi, A. (2004). Protein geometry, classification, topology and sym-
metry: A computational analysis of structure. CRC Press.
[73] Tumpach, A. B. (2007). Hyperkahler structures and infinite-dimensional Grassmanni-
ans. Journal of Functional Analysis, 243(1), 158-206.
[74] Wang, W., Juttler, B., Zheng, D. and Liu, Y. (2008). Computation of rotation minimiz-
ing frames. ACM Transactions on Graphics (TOG), 27(1), 2.
[75] Whitney, H. (1937). On regular closed curves in the plane. Compositio Mathematica, 4,
276-284.
[76] Younes, L., Michor, P. W., Shah, J. M. and Mumford, D. (2008). A metric on shape
space with explicit geodesics. Rendiconti Lincei-Matematica e Applicazioni 19, no. 1, 25-
57.
[77] Younes, L. (2012). Spaces and manifolds of shapes in computer vision: An overview.
Image and Vision Computing, 30(6), 389-397.
224