01 Csikos Differential Geometry

DIFFERENTIAL GEOMETRY
Series of Lecture Notes and Workbooks for Teaching

Undergraduate Mathematics
Algoritmuselmelet
Algoritmusok bonyolultsaga
Analitikus modszerek a penzugyekben
Bevezetes az analzisbe
Differential Geometry
Diszkret optimalizalas
Diszkret matematikai feladatok
Geometria
Igazsagos elosztasok
Interaktv analzis feladatgyujtemeny matematika BSc hallgatok szamara
Introductory Course in Analysis
Matematikai penzugy
Mathematical Analysis-Exercises 1-2
Mertekelmelet es dinamikus programozas
Numerikus funkcionalanalzis
Operaciokutatas
Operaciokutatasi peldatar
Optimalis iranytasok
Parcialis differencialegyenletek
Peldatar az analzishez
Szimmetrikus kombinatorikai strukturak
Tobbvaltozos adatelemzes
Balazs Csikos
DIFFERENTIAL
GEOMETRY
Eotvos Lorand University

Faculty of Science
Typotex
2014
20142019, Balazs Csikos,
Eotvos Lorand University, Faculty of Science
Referee: Arpad Kurusa
Creative Commons NonCommercial-NoDerivs 3.0 (CC BY-NC-ND 3.0)

This work can be reproduced, circulated, published and performed for non-
commercial purposes without restriction by indicating the authors name, but
it cannot be modified.
ISBN 978 963 279 221 7

Prepared under the editorship of Typotex Publishing House
(http://www.typotex.hu)
Responsible manager: Zsuzsa Votisky
Technical editor: Jozsef Gerner
Made within the framework of the project Nr. TAMOP-4.1.2-08/2/A/KMR-

2009-0045, entitled Jegyzetek es peldatarak a matematika egyetemi oktata-
sahoz (Lecture Notes and Workbooks for Teaching Undergraduate Mathe-
matics).
KEY WORDS: Curve, Frenet frame, curvature, torsion, hypersurface, funda-

mental forms, principal curvature, Gaussian curvature, Minkowski curvature,
manifold, tensor field, connection, geodesic curve
SUMMARY: The aim of this textbook is to give an introduction to differ-

ential geometry. It is based on the lectures given by the author at Eotvos
Lorand University and at Budapest Semesters in Mathematics. In the first
chapter, some preliminary definitions and facts are collected, that will be
used later. The classical roots of modern differential geometry are presented
in the next two chapters. Chapter 2 is devoted to the theory of curves,
while Chapter 3 deals with hypersurfaces in the Euclidean space. In the last
chapter, differentiable manifolds are introduced and basic tools of analysis
(differentiation and integration) on manifolds are presented. At the end of
Chapter 4, these analytical techniques are applied to study the geometry of
Riemannian manifolds.
Contents
1 Preliminaries 1
1.1 Categories and Functors . . . . . . . . . . . . . . . . . . . . . 1
1.2 Linear Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.1 Linear Spaces and Linear Maps . . . . . . . . . . . . . 3
1.2.2 Determinant of Matrices and Linear Endomorphisms . 7
1.2.3 Orientation of a Linear Space . . . . . . . . . . . . . . 11
1.2.4 Tensor Product . . . . . . . . . . . . . . . . . . . . . . 11
1.2.5 Exterior Powers . . . . . . . . . . . . . . . . . . . . . . 15
1.2.6 Euclidean Linear Spaces . . . . . . . . . . . . . . . . . 21
1.2.7 Hodge Star Operator . . . . . . . . . . . . . . . . . . . 30
1.3 Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
1.3.1 Affine Geometry . . . . . . . . . . . . . . . . . . . . . 33
1.3.2 Euclidean Spaces . . . . . . . . . . . . . . . . . . . . . 40
1.4 Topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
1.4.1 Separation and Countability Axioms . . . . . . . . . . 56
1.4.2 Compactness . . . . . . . . . . . . . . . . . . . . . . . 58
1.4.3 Fundamental Group and Covering Spaces . . . . . . . 60
1.5 Multivariable Calculus . . . . . . . . . . . . . . . . . . . . . . 62
1.6 Measure and Integration . . . . . . . . . . . . . . . . . . . . . 67
1.7 Ordinary Differential Equations . . . . . . . . . . . . . . . . . 75
2 Curves in En 81
2.1 The Notion of a Curve . . . . . . . . . . . . . . . . . . . . . . 81
2.2 The Length of a Curve . . . . . . . . . . . . . . . . . . . . . . 83
2.3 Croftons Formula . . . . . . . . . . . . . . . . . . . . . . . . 86
2.4 The Osculating k-planes . . . . . . . . . . . . . . . . . . . . . 92
2.5 Frenet Frames and Curvatures . . . . . . . . . . . . . . . . . 97
2.6 Limits of Some Geometrical Quantities . . . . . . . . . . . . . 107
2.7 Osculating Spheres . . . . . . . . . . . . . . . . . . . . . . . . 112
2.8 Plane Curves . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
i
2.8.1 Evolute, Involute, Parallel Curves . . . . . . . . . . . 118
2.8.2 The Rotation Number Theorem . . . . . . . . . . . . 124
2.8.3 Convex Curves . . . . . . . . . . . . . . . . . . . . . . 127
2.8.4 The Four Vertex Theorem . . . . . . . . . . . . . . . . 129
2.9 Curves in R3 . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
2.9.1 Orthogonal Projections onto Planes Spanned by Frenet
Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . 133
2.9.2 Fenchels Theorem . . . . . . . . . . . . . . . . . . . . 135
2.9.3 The Fary-Milnor Theorem . . . . . . . . . . . . . . . . 139
3 Hypersurfaces in Rn 141
3.1 General Theory . . . . . . . . . . . . . . . . . . . . . . . . . . 141
3.1.1 Definition of a Parameterized Hypersurface . . . . . . 141
3.1.2 Curvature of Curves on a Hypersurface . . . . . . . . 143
3.1.3 The Weingarten Map and the Fundamental Forms . . 145
3.1.4 Umbilical Points . . . . . . . . . . . . . . . . . . . . . 151
3.1.5 The Fundamental Equations of Hypersurface Theory 152
3.1.6 Surface Volume . . . . . . . . . . . . . . . . . . . . . . 161
3.2 Surfaces in R3 . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
3.2.1 Surfaces of Revolution . . . . . . . . . . . . . . . . . . 162
3.2.2 Lines of Curvature, Triply Orthogonal Systems . . . . 166
3.2.3 Ruled and Developable Surfaces . . . . . . . . . . . . 173
3.2.4 Asymptotic Curves on Negatively Curved Surfaces . . 179
3.2.5 Surfaces of Constant Negative Curvature . . . . . . . 183
3.2.6 Minimal Surfaces . . . . . . . . . . . . . . . . . . . . . 195
4 Manifolds 201
4.1 Topological and Differentiable Manifolds . . . . . . . . . . . . 201
4.1.1 Basic Definitions . . . . . . . . . . . . . . . . . . . . . 201
4.1.2 Configuration Spaces . . . . . . . . . . . . . . . . . . . 205
4.1.3 Submanifolds of Rn . . . . . . . . . . . . . . . . . . . 206
4.1.4 Remarks on the Classification of Manifolds . . . . . . 209
4.2 The Tangent Bundle . . . . . . . . . . . . . . . . . . . . . . . 212
4.2.1 The Tangent Bundle . . . . . . . . . . . . . . . . . . . 217
4.2.2 The Derivative of a Smooth Map . . . . . . . . . . . . 218
4.3 The Lie Algebra of Vector Fields . . . . . . . . . . . . . . . . 219
4.3.1 The Lie Algebra of Lie Groups . . . . . . . . . . . . . 227
4.3.2 Subspace Distributions and the Frobenius Theorem . 229
4.4 Tensor Bundles and Tensor Fields . . . . . . . . . . . . . . . 233
4.5 The Lie Derivative . . . . . . . . . . . . . . . . . . . . . . . . 238
4.6 Differential Forms . . . . . . . . . . . . . . . . . . . . . . . . 243
4.6.1 Interior Product by a Vector Field . . . . . . . . . . . 243
ii
4.6.2 Exterior Differentiation . . . . . . . . . . . . . . . . . 245
4.6.3 De Rham Cohomology . . . . . . . . . . . . . . . . . . 251
4.7 Integration of Differential Forms . . . . . . . . . . . . . . . . 254
4.7.1 Integration on Chains . . . . . . . . . . . . . . . . . . 255
4.7.2 Integration on Regular Domains . . . . . . . . . . . . 260
4.7.3 Integration on Riemannian Manifolds . . . . . . . . . 265
4.8 Differentiation of Vector Fields . . . . . . . . . . . . . . . . . 266
4.9 Curvature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
4.10 Decomposition of Algebraic Curvature Tensors . . . . . . . . 290
4.11 Conformal Invariance of the Weyl Tensor . . . . . . . . . . . 298
4.12 Geodesics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304
4.13 Applications to Hypersurface Theory . . . . . . . . . . . . . . 311
4.13.1 Geodesic Curves on Hypersurfaces . . . . . . . . . . . 311
4.13.2 Clairauts Theorem . . . . . . . . . . . . . . . . . . . . 315
4.13.3 Moving Orthonormal Frames Along a Hypersurface . . 317
4.13.4 Relation to Earlier Formulae for Parameterized Hyper-
surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . 318
4.13.5 The GaussBonnet Formula . . . . . . . . . . . . . . . 320
4.13.6 Steiners Formula . . . . . . . . . . . . . . . . . . . . . 321
4.13.7 Minkowskis Formula . . . . . . . . . . . . . . . . . . . 325
4.13.8 Rigidity of Convex Surfaces . . . . . . . . . . . . . . . 330
Bibliography 336
Index 338
iv
Chapter 1
Preliminaries
In this chapter, we collect some definitions and facts that will be used later
in the text.
1.1 Categories and Functors

We shall often use the term natural map or natural isomorphism between two
sets carrying certain structures. The concept of naturality can be properly
defined within the framework of category theory. Category theory yields a
unified way to look at different areas of mathematics and their constructions.
Definition 1.1.1. A category C consists of
a class ObC of objects;
an assignment of a set MorC (X, Y ) to any pair of objects X, Y ObC ,
the elements of which are called morphisms, arrows or maps from X
to Y ;
a composition operation MorC (X, Y ) MorC (Y, Z) MorC (X, Z),
(f, g) 7 g f for any three objects X, Y, Z.
These should satisfy the following axioms.
(i) For any four objects X, Y, Z, W ObC and any morphisms
h MorC (X, Y ), g MorC (Y, Z) f MorC (Z, W ), we have
(f g) h = f (g h).
(ii) There is a (unique) morphism 1X MorC (X, X) for any object X,

called the identity morphism of X, such that for any morphisms f
1
2 1. Preliminaries
MorC (X, Y ) and g MorC (Y, X), the identities

f 1X = f and 1X g = g
hold.
Definition 1.1.2. A morphism f MorC (X, Y ) is called an isomorphism if
it has a two-sided inverse, that is a morphism g MorC (Y, X) with f g = 1Y
and g f = 1X . Two objects are isomorphic if there is an isomorphism be-
tween them. The morphism f is an endomorphism of X if Y = X. The
set MorC (X, X) of endomorphisms of X is also denoted by EndC (X). The
morphism f is an automorphism if f is both an endomorphism and an isomor-
phism. Automorphisms of an object X form a group AutC (X) with respect
to the composition operation.
A basic example is the category of sets, in which the objects are the sets,
Mor(X, Y ) is the set of all maps from X to Y , is the ordinary composition
of maps, 1X is the identity map of X. Isomorphisms of this category are
the bijective maps. Two sets are isomorphic in this category if and only if
they have the same cardinality. Aut(X) is the group of permutations of the
elements of X.
Definition 1.1.3. A diagram is a directed graph, the vertices of which are
objects of a category and the edges are labeled by morphisms from the initial
point of the edge to the endpoint. Any directed path in a diagram gives rise
to morphism from the initial point of the path to its endpoint obtained as
the composition of morphisms attached to the consecutive edges of the path.
A diagram is said to be commutative when, for each pair of vertices X and
Y and for any two directed paths from X to Y , the compositions of the edge
labels of the paths are equal morphism from X to Y .
Definition 1.1.4. A covariant functor F from a category C to a category D
associates to each object X of the category C an object F (X) ObD ; and to
each morphism f MorC (X, Y ) a morphism F (f ) MorD (F (X), F (Y )) in
such a way that
F (1X ) = 1F (X) for any object X ObC ;
F (g f ) = F (g) F (f ) for any f MorC (X, Y ) and g MorC (Y, Z).
Contravariant functors are defined similarly. The difference is that contravari-
ant functors reverse arrows.
Definition 1.1.5. A contravariant functor F from a category C to a category
D associates to each object X of the category C an object F (X) ObD ; and
to each morphism f MorC (X, Y ) a morphism F (f ) MorD (F (Y ), F (X))
in such a way that
1.2. Linear Algebra 3
F (1X ) = 1F (X) for any object X ObC ;
F (g f ) = F (f ) F (g) for any f MorC (X, Y ) and g MorC (Y, Z).
To each category C, there is an opposite category C op . This category has the

same objects as C but for A, B ObC = ObC op , MorC op (A, B) = MorC (B, A).
The composition rule op in the opposite category is f op g = g f . With this
construction, every contravariant functor from a category C to a category D
can be thought of as a covariant functor form C op to D or from C to Dop .
Definition 1.1.6. Let F and G be two covariant functors from the category
C to the category D. Then a natural transformation from F to G assigns
to every X ObC a morphism X MorD (F (X), G(X)) so that for any
morphism f MorC (X, Y ), the diagram
F (f )
F (X) F (Y )

X y

y Y
G(f )
G(X) G(Y )
is commutative, i.e., Y F (f ) = G(f ) X . A natural isomorphism between

two functors is a natural transformation for which X is an isomorphism
for all X objects of C.
The definition of natural transformations can be extended also for the case,
when one of the functors or both are contravariant. Then the contravariant
functors should be substituted by their covariant counterparts going into the
opposite categories. Natural transformations and natural isomorphisms can
be defined by obvious modifications also for multivariable functors.
There is a principle in mathematics, that if there is an isomorphism between
two objects of a category, then the two objects can be identified, as they
behave in the same way and look like the same way from the view point of
the category.
The importance of naturality is that if there is a natural isomorphism between
two functors, then they can be identified with one another, or considered to
be essentially the same.
1.2 Linear Algebra

1.2.1 Linear Spaces and Linear Maps
Definition 1.2.1. A set V is a linear space, or vector space over R if V is
equipped with a binary operation +, and for each R, the multiplication
4 1. Preliminaries
of elements of V with is defined in such a way that the following identities

are satisfied:
(i) (x + y) + z = x + (y + z) (associativity);
(ii) 0 V such that x + 0 = x for all x V ;
(iii) x V x V such that x + (x) = 0;
(iv) x + y = y + x (commutativity);
(v) (x + y) = x + y;
(vi) ( + )x = x + x;
(vii) ()x = (x);
(viii) 1x = x.
Definition 1.2.2. A linear combination of some vectors x1 , . . . , xk V is a

vector of the form
1 x1 + + k xk , (1.1)
where the i s are real numbers. The vectors x1 , . . . , xk V are linearly in-
dependent if the linear combination (1.1) can be 0 only if all of the coefficients
i vanish. A basis of V is a maximal set of linearly independent vectors.
It is known that any two bases of a linear space have the same cardinality.
Definition 1.2.3. The dimension dim V of the linear space V is the cardi-
nality of a basis.

Definition 1.2.4. A map L : V W between the linear spaces V and W is

said to be linear if
L(x + y) = L(x) + L(y)
for any x, y V and , R. A linear isomorphism is a bijective linear map.
Two linear spaces are isomorphic if there is a linear isomorphism between
them.
Linear spaces as objects and linear transformations as morphisms form a

category. Two linear spaces are isomorphic if and only if they have the same
dimension. The automorphism group of a linear space V is called the general
linear group of V and it is denoted by GL(V ).
If V is an n-dimensional linear space, and (e1 , . . . , en ) is a basis of V , then
any vector x V can be written uniquely as a linear combination x =
x1 e1 + + xn en of the basis vectors. The numbers (x1 , . . . , xn ) are called

the coordinates of x with respect to the basis (e1 , . . . , en ).
The indices of the coordinates are not exponents, they are just upper indices.
The reason why it is practical to use both upper and lower indices is the ob-
servation that if we position the indices properly in a linear algebraic formula,
then usually summations go exactly over those indices that appear twice in
a term, once as a lower index, once as an upper one. Therefore, if we take
care of the right positioning of the indices, summation signs show redundant
information and can be supressed. This leads to Einsteins convention which
suggests us to position the indices properly and omit the summation signs.
It is a rule for correct index positioning that if a single index appears on one
side of an equation, then the same index must appear as a single index at the
same (upper or lower) position on the other side as well. In this book, we
shall pay attention to index positioning, but we shall not omit the summation
signs.
Working with coordinates, linear transformations are represented by matri-
ces. Let L : V W be a linear map from the n-dimensional linear space
V to the m-dimensional linear space W . Choose a basis (e1 , . . . , en ) for
V and a basis (f1 , . . . , fm ) for W . Write L(ei ) as a linear combination
L(ei ) = li1 f1 + + lim fm . Arranging the coefficients lij into an m n matrix
l11 . . . ln1

.. .. .
[L] = . . .. ,
l1m . . . lnm
we obtain the matrix of L with respect to the bases (e1 , . . . , en ) and (f1 , . . . , fm ).
If we arrange the coordinates of x with respect to the basis (e1 , . . . , en ) into
a column vector [x], then the column vector [L(x)] of the coordinates of L(x)
with respect to the basis (f1 , . . . , fm ) can be computed by the matrix mul-
tiplication [L(x)] = [L][x]. For an endomorphism of V , we usually use the
same basis for V and W = V .
Examples.
(1) Recall that Rn denotes the set of n-tuples of real numbers

Rn = {(x1 , . . . , xn )|xi R}.
If x = (x1 , . . . , xn ) and y = (y1 , . . . , yn ) are two elements of Rn and
R is a real number, then we define the sum and difference of x and
y and the scalar multiple of x by
xy = (x1 y1 , . . . , xn yn ),
x = (x1 , . . . , xn ).
6 1. Preliminaries
It is clear that Rn is an n-dimensional linear space over the field of real

numbers with respect to the operations defined above.
i
Let ei denote the vector (0, . . . , 0, 1, 0 . . . 0), the only non-zero coordinate
of which is the ith one, being equal to 1. The n-tuple (e1 , . . . , en ) is a
basis of Rn called the standard basis of Rn .
(2) Let V and W be linear spaces, Hom(V, W ) be the set of all linear maps
from V to W . Hom(V, W ) becomes a linear space if for A, B Hom(V, W )
and R, we define the maps A + B and A by
(A + B)(x) = A(x) + B(x), (A)(x) = (A(x)).
(3) A linear subspace of a linear space V is a nonempty subspace W V ,

which contains all linear combinations of its elements. Linear subspaces
of a linear space are linear spaces themselves. The dimension of a linear
subspace W of V is less than or equal to dim V . If V is finite dimensional,
then dim W = dim V holds only if V = W .
The intersection of an arbitrary family of linear subspaces is a linear
subspace, therefore, for any subset S V there is a unique smallest
linear subspace among all linear subspaces containing S. We shall call
this linear subspace the linear subspace spanned or generated by S, or
simply the linear hull of S. We shall denote the linear hull of S by lin[S].
Definition 1.2.5. The Grassmann manifold of k-dimensional linear sub-
spaces of the linear space V is the set Grk (V ) of all k-dimensional sub-
spaces of V . In the special case k = 1, P(V ) = Gr1 (V ) is also called the
projective space associated to V . Later we shall introduce a topology and
a manifold structure on Grk (V ). Then the name Grassmann manifold
will be justified.
(4) If L Hom(V, W ) is a linear transformation, then the image im L =

{L(v) | v V } of L is a linear subspace in W , and the kernel ker L =
{v V | L(v) = 0} of L is a linear subspace in V . The rank rk L of L
is the dimension of im L.
(5) For an element v V of a linear space V , we define translation Tv by
v as the map Tv : V V , Tv (x) = x + v. Translations by a non-
zero vector are not linear transformations. If W is a linear subspace
of V then two translates of W are either equal or disjoint. The set
V /W = {Tv (W ) | v V } of translates of W carries a linear space
structure defined by Tv1 (W ) + Tv2 (W ) = Tv1 +v2 (W ). (Check that
the definition is correct.) The linear space V /W is called the factor space
of V with respect to the subspace W . The surjective linear map L : V
V /W , v 7 Tv (W ) is the factor map. Every linear map L : V W is a

factor map L : V V /(ker L) = im L onto its image. The dimension of
the factor space or the image of a linear space can be computed by the
formula
dim(V /W ) = dim(V ) dim(W ),
(1.2)
rk L = dim(im L) = dim(V ) dim(ker L).
(6) The linear space V = Hom(V, R) consisting of the linear functions on V

is the dual space of V . Assigning the dual space V to a linear space V
is a contravariant functor of the category of linear spaces into itself. This
functor assigns to a linear map L : V W the adjoint map L : W
V , defined by L (l) = l L, where l V .
If (e1 , . . . , en ) is a basis of V , then the linear functions ei V , (i =
1, . . . , n), defined by the equalities ei (ej ) = ji , where

0, if i 6= j,
ji = ij =
1, if i = j
is the Kronecker delta symbol, form a basis of V . This basis is called the
dual basis of the basis (e1 , . . . , en ).
We remark that the Kronecker delta symbol ij is also denoted by ij
and ij . In formulae involving the Kronecker delta symbol, we always
position the indices as dictated by the Einstein convention.
Though for a finite dimensional linear space V , the dual space V has the
same dimension as V , there is no natural isomorphism between these two
spaces. In other words, the identical functor on the category of finite dimen-
sional linear spaces is not naturally isomorphic to the dual space functor. On
the other hand, there is a natural transformation from the identical functor
to the double dual space functor, given by the embeddings V : V V ,
((v))(l) = l(v), where v V , l V . The restriction of this natural trans-
formation onto the category of finite dimensional linear spaces is a natural
isomorphism.
With the help of this natural isomorphism, elements of a finite dimensional
linear space V can be identified with elements of V . With this identification,
the dual basis of the dual basis of a basis of V will be equal to the original
basis.
1.2.2 Determinant of Matrices and Linear

Endomorphisms
Let us denote by Sn the group of all permutations of the set {1, . . . , n}. For
a permutation Sn , we denote by sgn the sign of the permutation ,
8 1. Preliminaries
which can be defined by the formula

Y (j) (i)
sgn = {1, 1}.
ji
1i<jn
Exercise 1.2.6. Show that the sign of a permutation is always equal to 1.

Prove that sgn(1 2 ) = sgn(1 ) sgn(2 ).
Definition 1.2.7. The determinant of an n n matrix A = (aij )1i,jn is

the number

a11 a1n
det A = det ... .. .. = X sgn a
1(1) an(n) .

. .
S

an1 ann n
Some important properties of the determinant are summarized in the follow-

ing proposition.
Proposition 1.2.8. (1) If all but one columns of a square matrix are
fixed, then the determinant is a linear function of the varying column. This
means that if we denote by (a1 , . . . , an ) the square matrix with column vectors
a1 , . . . , an Rn , then
det(a1 , . . . , aj + aj , . . . , an ) =
det(a1 , . . . , aj , . . . , an ) + det(a1 , . . . , aj , . . . , an ).
(2) When we permute the columns of a square matrix the determinant is

multiplied by the sign of the permutation, i.e.,
det(a(1) , . . . , a(n) ) = sgn det(a1 , . . . , an ) for all Sn .
(3) The value of the determinant does not change if an arbitrary multiple of
a column is added to another column.
(4) The determinant of a matrix vanishes if and only if its columns are lin-
early dependent.
(5) A square matrix A and its transposition A> , that is the reflection of A in
the main diagonal, have the same determinant
det A = det A> .
As a consequence, properties (1)-(4) hold also if the columns are replaced by

rows.
(6) The determinant of an upper or lower triangular matrix is the product of
the diagonal elements.
Recall that the product AB of two n n matrices A and B is also an n n

matrix. The determinants of A, B and AB are related to one another as
follows.
Proposition 1.2.9.
det(AB) = det(A) det(B).
Proposition 1.2.10. Let L : V V be a linear endomorphism of the linear

space V . Choose two bases (e1 , . . . , en ) and (f1 , . . . , fn ) of V and consider
the matrices [L]e = (lij )1i,jn and [L]f = (lij )1i,jn with respect to these
bases respectively. Then
det([L]e ) = det([L]f ).
Proof. Let S : V V be the invertible linear endomorphism which takes the

basis vector ei to the basis vector fi for all i. Denote by [S]e = (sji )1i,jn
the matrix of S with respect to the basis (e1 , . . . , en ). Then
n
X n
X n
X
L(ei ) = lij ej , L(fi ) = lj fj ,
i S(ei ) = fi = sji ej ,
j=1 j=1 j=1
and

n X
X n n
X Xn n X
X n
lj sk ek = lj S(ej ) = L(fi ) = L sj
e = sji ljk ek .
i j i i j
j=1 k=1 j=1 j=1 j=1 k=1
Comparing the coefficients we obtain

X n
n X n X
X n
lj sk = sji ljk for all 1 i, k n, (1.3)
i j
j=1 k=1 j=1 k=1
which means that [L]f [S]e = [S]e [L]e . Taking the determinant of both sides
we obtain
det([L]f ) det([S]e ) = det([S]e ) det([L]e ).
Since the columns of [S]e are linearly independent, as they are the coordinate
vectors of the basis vectors fi with respect to the basis (e1 , . . . , en ), the deter-
minant of [S]e is nonzero. Thus, equation (1.3) implies the proposition.
Definition 1.2.11. The determinant of a linear endomorphism L : V V
is the determinant of the matrix of L with respect to an arbitrary basis of V .

The definition is correct according to the previous proposition.
10 1. Preliminaries
Definition 1.2.12. The complexification C V of a linear space V over R

is the set of formal linear combinations v + iw, where v, w V .
C V is a linear space of dimension 2 dim V , which contains V as a linear
subspace. Elements of the complexification can also be multiplied by complex
numbers as follows
(x + iy)(v + iw) = (xv yw) + i(xw + yv).
A linear endomorphism of L : V V can be extended to the complexification

by the formula L(v + iw) = L(v) + iL(w).
Definition 1.2.13. A non-zero vector z = v+iw 6= 0 of the complexification
of a linear space V is an eigenvector of the linear endomorphism L : V V
if there is a complex number C such that L(z) = z. The number is
called the eigenvalue corresponding to the eigenvector z. A complex number
is an eigenvalue of L if there is an eigenvector to which it corresponds.
Proposition 1.2.14. A complex number is an eigenvalue of the linear
transformation L : V V if and only if det(L idV ) = 0. An eigenvalue
is real if and only if there is an eigenvector in V with eigenvalue .
Definition 1.2.15. The characteristic polynomial of a linear endomorphism
L : V V is the polynomial pL () = det(L idV ). Similarly, the character-
istic polynomial of an nn matrix A is the polynomial pA () = det(AIn ),
where In is the n n unit matrix. The algebraic multiplicity of an eigenvalue
is its multiplicity as a root of the characteristic polynomial.
The coefficients of the characteristic polynomial of a matrix A can be ex-
pressed as polynomials of the matrix elements. They can also be expressed
in terms of the eigenvalues 1 , . . . , n using the factorization
pA () = (1 ) (n ).
Comparing these expressions we can relate elementary symmetric polynomi-

als of the eigenvalues to some matrix invariants. For example, the constant
term of the characteristic polynomial of the matrix A is
det A = pA (0) = 1 n .
The coefficient of ()n1 in pA () is equal to the sum of the diagonal

elements of the matrix and also to the sum of the eigenvalues.
Definition 1.2.16. The trace of a matrix A = (aij )1i,jn is the number
n
X
tr A = aii ,
i=1
which is also equal to the sum of the eigenvalues of A. The trace of a linear
endomorphism is the trace of its matrix with respect to an arbitrary basis.

1.2.3 Orientation of a Linear Space

Definition 1.2.17. Let (v1 , . . . , vk ) and (w1 , . . . , wk ) be two ordered bases
of a linear space V . We say that they have the same orientation or they
define the same orientation of V , if the k k matrix (aji ) defined by the
system of equalities
k
X
vi = aji wj for i = 1, 2, . . . , k
j=1
has positive determinant.

Having the same orientation is an equivalence relation on ordered bases,
and there are two equivalence classes. A choice of one of the equivalence
classes, the elements of which will be called then positively oriented bases, is
an orientation of V .
Definition 1.2.18. The standard orientation of Rn is the orientation defined
i
by the ordered basis e1 , . . . , en , where ei = (0, . . . , 0, 1, 0 . . . 0).
1.2.4 Tensor Product

Definition 1.2.19. Let V , W and Z be linear spaces. A map B : V W Z
is said to be a bilinear map if it is linear in both variables, i.e., if it satisfies
the identities
B(v1 + v2 , w) = B(v1 , w) + B(v2 , w) and
B(v, w1 + w2 ) = B(v, w1 ) + B(v, w2 ).
Definition 1.2.20. The tensor product of the linear spaces V and W is a
linear space V W together with a bilinear map : V W V W ,
(v, w) 7 v w, such that for any bilinear map B : V W Z, there is a
unique linear map L : V W Z which makes the diagram

V W V W

yL
B
V W Z
commutative.
12 1. Preliminaries
One can consider the category of bilinear maps defined on V W , in which

the objects are the bilinear maps, a morphism between the bilinear maps
B1 : V W Z1 and B2 : V W Z2 is a linear map L : Z1 Z2 , for
which the diagram
B1
V W Z1

yL
2 B
V W Z2
is commutative. In general, an object X of a category is called an initial
object, if for any other object Y of the category, there is a unique morphism
form X to Y . Using this terminology, the tensor product of the linear spaces
V and W is the initial object of the category of bilinear maps on V W . It
is a simple exercise playing with arrows, that up to isomorphism, a category
can have at most one initial object. However, initial objects do not exist
in all categories. Existence of initial objects are always shown by explicit
constructions in the given category.
To construct the tensor product explicitly, one first considers the linear space
FV W generated freely by the elements of V W . More explicitly, FV W is
the linear space of all formal linear combinations 1 (v1 , w1 )+ +k (vk , wk )
of some pairs (vi , wi ) V W with real coefficients i R. Then we take
the smallest linear subspace Z of FV W that contains all elements of the
form
(v1 + v2 , w) (v1 , w) (v2 , w), (w, v1 + v2 ) (w, v1 ) (w, v2 ),

(v, w) (v, w), (v, w) (v, w)
for all v, v1 , v2 V , w, w1 , w2 W , R. Set V W = FV W /Z and let

: V W V W be the composition of the embedding V W FV W
and the factor map FV W FV W /Z. It is not difficult to check that the
bilinear map : V W V W is an initial object of the category of
bilinear maps on V W .
It is known that if e1 , . . . , en is a basis of V , f1 , . . . , fm is a basis of W , then
the vectors {ei fj | 1 i n, 1 j m} form a basis of V W . In
particular, dim(V W ) = dim(V ) dim(W ).
The tensor product construction can be thought of as a functor from the
category of pairs of linear spaces to the category of linear spaces In the
category of pairs of linear spaces a morphism from the pair (V1 , V2 ) to the
pair (W1 , W2 ) is a pair (L1 , L2 ) of linear maps, where L1 : V1 W1 and
L2 : V2 W2 . Tensor product as a functor assigns to a pair (V1 , V2 ) the
tensor product V1 V2 and to the pair of linear maps (L1 , L2 ) the linear map
L1 L2 : V1 V2 W1 W2 , where the tensor product L1 L2 of the linear
maps L1 and L2 is defined as the unique linear map for which
L1 L2 (v1 v2 ) = L1 (v1 ) L2 (v2 ).
Definition 1.2.21. Let V be an n dimensional linear space and V its dual

space. The tensor product T (k,l) V = V V V V will be
| {z } | {z }
k times l times
called the linear space of tensors of type (k, l). We agree that T (0,0) V is
equal to the ground field R.
If e1 , . . . , en is a basis of V , e1 , . . . , en is its dual basis, then the type (k, l)

tensors eij11...i i1 ik
...jl = e e ej1 ejl form a basis of T
k (k,l)
V . In the
(0,0)
special caseL k = l = 0 the basis vector e of T is the unit element of R. The

direct sum k,l=0 T (k,l) V can be equipped with a bilinear associative tensor
multiplication, which turns it into an associative algebra. Tensor product is
defined on the basis vectors by the formula
i1 ...ik p1 ...pr
eij11...i p1 ...pr
...jl eq1 ...qs = ej1 ...jl q1 ...qs .
k
Pn
If T = i1 ,...,ik ,j1 ,...,jl =1 Tij11...i
...jl i1 ...ik
e
k j1 ...jl
is a tensor of type (k, l), then the num-
j1 ...jl
bers Ti1 ...ik are called the coordinates or components of the tensor T with
respect to the basis e1 , . . . , en .
Exercise 1.2.22. How are the coordinates of a tensor transformed Pnwhen we

change the basis e1 , . . . , en to another one f1 , . . . , fn , where fi = j=1 aji ej ?
Show that if (bji ) is the inverse matrix of the matrix (aji ), then
n n
X X
fi = bij ej , fji11...j
...ik
l
= aip11 . . . aipkk bqj11 . . . bqjkk eqp11...q
...pk
l
.
j=1 1 k 1 l p ,...,p ,q ,...,q =1
Exercise 1.2.23. Find all type (0, 2) and type (1, 1) tensors, the coordinates
of which do not depend on the choice of the basis.
Exercise 1.2.24. The coordinates of a type (0, 2) or a type (1, 1) tensor can
always be arranged into an n n matrix. Do the trace and determinant of
this matrix depend on the choice of the basis?
Exercise 1.2.25. A type (2, 0) or a type (0, 2) tensor is said to be non-

degenerate if the n n matrix built from its coordinates with respect to a
basis has non-zero determinant. Show that non-degeneracy does not depend
on the choice of the basis.
Exercise 1.2.26. Let be a non-degenerate tensor of type (0, 2). Show

that there is a unique type (2, 0) tensor such that the n n matrices built
from the coordinates of and with respect to any basis are inverses of one
another.
14 1. Preliminaries
Exercise 1.2.27. Construct natural isomorphisms between the following

linear spaces:
(a) (V W )
= V W ;
(b) Hom(V, W )
= V W , in particular, End(V )
= T (1,1) V and
(2,0)
Hom(V, V ) = T V;
(c) (T (k,l) V )
= T (k,l) V
= T (l,k) V ;
(d) {V V V V R (k + l)-linear functions}

= T (k,l) V ;
| {z } | {z }
k times l times
(e) {V V W k-linear maps into W }

= T (k,0) V W ;
| {z }
k times
(f) {V V V k-linear maps into V }

= T (k,1) V ;
| {z }
k times
(g) Hom(T (k,l) V, T (p,q) V )

= T (l+p,k+q) V .
Remark. We explain what naturality of an isomorphism means for case

(b). For other cases a similar definition can be given. Consider the cate-
gory of pairs of linear spaces, in which the objects are pairs (V, W ) of linear
spaces, the morphisms from (V1 , W1 ) to (V2 , W2 ) are pairs (, ) of isomor-
phisms : V1 V2 and : W1 W2 . Both F1 : (V, W ) 7 Hom(V, W )
and F2 : (V, W ) 7 V W are functors from this category to the category
of linear spaces. If (, ) is a morphism from (V1 , W1 ) to (V2 , W2 ), then
the linear map F1 (, ) : Hom(V1 , W1 ) Hom(V2 , W2 ) assigns to the lin-
ear map L : V1 W1 the linear map L 1 Hom(V2 , W2 ), while
F2 (, ) : V1 W1 V2 W2 is the linear map (1 ) . The statement
that there is a natural isomorphism between Hom(V, W ) and V W means
that there is a natural isomorphism between the functors F1 and F2 .
There is a more practical (but less formal) way to characterize natural isomor-
phisms. A natural isomorphism between two linear spaces is an isomorphism
for which the image of an element can be described uniquely by a set of
instructions or formulae. If the definition of an isomorphism involves the
random choice of a basis, for example, then it may not be natural, since the
image of an element may depend on the choice of the basis. On the other
hand, the definition of a natural isomorphism is allowed to contain random
choices, but to prove naturality, we have to check that the image of an element
does not depend on the random variables.
1.2.5 Exterior Powers

Denote by Sk the group of all permutations of the set {1, . . . , k}.
Definition 1.2.28. Let V and W be linear spaces, k N. A k-linear map
K: V k = V V W
| {z }
k times
is said to be alternating if for any permutation Sk and any k vectors

v1 , . . . , vk V , we have
K(v(1) , . . . , v(k) ) = sgn K(v1 , . . . , vk ).
Alternating k-linear maps on a given linear space V form a category. A
morphism from an alternating k-linear map K1 : V k W1 to another one
K2 : V k W2 is a linear map L : W1 W2 such that K2 = L K1 .
Definition 1.2.29. The kth exterior power of a linear space V is an initial
object of the category of alternating k-linear maps on V . In other words, it is a
linear space k V together with an alternating k-linear map k : V k k V ,
(v1 , . . . , vk ) 7 v1 vk , such that for any alternating k-linear map
K : V k W , there is a unique linear map L : k V W for which the
diagram

V k k k V

yL
K
V k W
is commutative.
The element v1 vk k V is called the exterior product or wedge product
of the vectors v1 , . . . , vk . Elements of k V are called k-vectors. The words
bivector and trivector are also used for 2-vectors and 3-vectors respectively.

Uniqueness of the exterior power up to isomorphism follows from uniqueness
of initial objects. Existence is proved by an explicit construction as follows.
Consider the kth tensor power T (0,k) V and the k-linear map k : V k
T (0,k) V , (v1 , . . . , vk ) 7 v1 vk . By the universal property of the
tensor product, this is an initial object of the category of k-linear maps on
V . Denote by Wk < T (0,k) V the linear hull of the set of elements of the form
v1 vk sgn v(1) v(k) , where vi V, Sk .
Let k V be the factor space T (0,k) V /Wk and k : V k k V be the com-

position of k with the factor map k : T (0,k) V T (0,k) V /Wk . It can be
checked that k : V k k V is indeed an initial object of the category of
alternating k-linear maps on V .
16 1. Preliminaries
Proposition 1.2.30. Assume that e1 , . . . , en is a basis of V . Then
{ei1 eik | 1 i1 < < ik n}
is a basis of k V . In particular, dim k V = nk .

The following formula has many applications.

Proposition 1.2.31. Suppose that the vectors v1 , . . . , vk can be expressed
as a linear combination of the vectors w1 , . . . , wk as follows
vi = c1i w1 + + cki wk , (i = 1, . . . , k).
Then
c11 ck1

v1 vk = det ... .. .. w w .

. . 1 k
c1k ckk
The next statement is a corollary of the previous two ones.
Proposition 1.2.32. The vectors v1 , . . . , vk are linearly independent if and
only if v1 vk 6= 0. Two linearly independent k-tuples of vectors
v1 , . . . , vk and w1 , . . . , wk span the same k-dimensional linear subspace if
and only if v1 vk = c w1 wk for some c R \ {0}. In addition,
these two collections are bases of the same orientation in the linear space they
span if and only if c > 0.
Corollary 1.2.33. The Grassmann manifold Grk (V ) can be embedded into
the projective space P(k V ) by assigning to the k-dimensional subspace span-
ned by the linearly independent vectors v1 , . . . , vk the 1-dimensional linear
space spanned by v1 vk . This embedding is called the Plucker embed-
ding.
Corollary 1.2.34. There is a natural one-to-one correspondence between
orientations of a k-dimensional linear subspace W of an n-dimensional lin-
ear space V and orientations of the 1-dimensional linear subspace of k (V )
assigned to W by the Plucker embedding.
Ldim V
The direct sum (V ) = k=0 k (V ) of all the exterior powers of V be-
defined uniquely by the
comes an associative algebra with multiplication
rule
(l (vk+1 , . . . , vk+l )) = k+l (v1 , . . . , vk+l )
(k (v1 , . . . , vk ))
k, l N; v1 , . . . , vk+l V.
As we have
...
k (v1 , . . . , vk ) = v1 vk
for any k vectors v1 , . . . , vk V , we may (and we shall) denote the multi-

plication simply by without causing confusion with the earlier notation
k (v1 , . . . , vk ) = v1 vk .
Definition 1.2.35. The associative algebra ( (V ), ) is the exterior algebra

or Grassmann algebra of V .
Exterior Powers and Alternating Tensors

Let us analyze the factor map k : T (0,k) (V ) k (V ) to obtain another con-
struction of the kth exterior power of V , which is, of course, naturally isomor-
phic to the first construction. The permutation group Sk has a representa-
tion on the tensor space T (0,k) V . The representation : Sk GL(T (0,k) V ),
7 is given on decomposable tensors as follows:
(vi1 vik ) = v(i1 ) v(ik ) .
Definition 1.2.36. A tensor T of type (0, k) is called symmetric if (T ) =

T for any permutation Sk . T is said to be alternating if (T ) =
sgn() T for all Sk . We shall use the notation Sk (V ) T (0,k) (V ) for
the linear space of symmetric tensors, and Ak (V ) T (0,k) (V ) for the linear
space of alternating tensors.
Exercise 1.2.37. Show that T (0,2) (V ) = S2 (V ) A2 (V ).
Exercise 1.2.38. Show that dim Sk (V ) = n+k1

k , where n = dim V .
Hint: Find an isomorphism between Sk (V ) and homogeneous polynomials of

degree k in n variables.
Exercise 1.2.39. Compute the dimension of Ak (V ).
Define the linear map k : T (0,k) V T (0,k) (V ) by the formula

X
k (T ) = sgn() (T ).
Sk
Proposition 1.2.40. The image of the map k is Ak (V ). The kernel of k

is Wk = ker k , where k : T (0,k) V T (0,k) V /Wk = k V is the factor map
we defined in the construction of k V . The map k /k! is a projection onto
Ak (V ).
Proof. It is not difficult to check that im k contains only alternating tensors,

and if T is alternating, then k (T ) = k!T , so im k = Ak (V ) and (k )/k! is
a projection of T (0,k) V onto Ak (V ).
18 1. Preliminaries
If v1 vk sgn (v1 vk ) is a generator of Wk , then denoting

v1 vk by T
X X
k (T sgn (T ))) = sgn 0 0 (T ) sgn 00 00 (T ) = 0,
0 Sk 00 = 0 Sk
thus, ker k Wk . Conversely, if T ker k , then

1 X
T = (T sgn (T )).
k!
Sk
Since the image of the linear map (I sgn ) is in Wk for any Sk

by the definition of Wk , the above expression for T shows that T is also in
Wk .
According to the proposition,
T (0,k) (V ) = Ak (V ) Wk ,
therefore, the factor space k (V ) = T (0,k) (V )/Wk is naturally isomorphic

to Ak (V ). There are two different ways to identify these two linear spaces.
One is to identify them with the restriction k = k |Ak (V ) of the factor map
k
k : T (0,k)(V ) (V ) onto Ak (V ).
We can also define another natural isomorphism using the universal property
of the exterior power. Since the map
k k : V k Ak (V ), (v1 , . . . , vk ) 7 k (v1 vk )
is an alternating k-linear map, by the universal property of the exterior power,

there is an induced linear map k : k (V ) Ak (V ) such that
k (v1 vk ) 7 k (v1 vk ).
Since
X
k (k (v1 vk ) = sgn v(1) v(k) = k! v1 vk ,
Sk
k1 = k /k!.
It is a matter of taste which isomorphism is used to identify wedge products
of vectors with alternating tensors, but to avoid confusion, we should choose
one of the identifications and then insist on that for the rest of the discussion.
Making this decision, from now on we use k to identify k (V ) and Ak (V ).
Nevertheless, one should be aware of the fact that some authors prefer the
identification by k .
Taking the direct sum of the linear isomorphisms k : k (V ) Ak (V ), we

obtain a linear isomorphism : (V ) A (V ) between the direct sum
Ldim V
A (V ) = k=0 Ak (V ) and the Grassmann algebra of V . Using this iso-
morphism, we can define an associative multiplication on A (V ) setting
T1 T2 = (1 (T1 ) 1 (T2 )) for T1 , T2 A (V ). More explicitly, if
T1 Ak (V ), T2 Al (V ), then T1 = k (T1 )/k!, T2 = l (T2 )/l!, and
1 1
T1 T2 = k+l (k (T1 ) l (T2 )) = k+l (k (T1 ) l (T2 ))
k! l! k! l!
1
= k+l (k+l (T1 T2 ))
k! l!
1
= k+l (k+l (k+l (T1 T2 )))
k! l! (k + l)!
1 1
= k+l (k+l (k+l (T1 T2 ))) = k+l (T1 T2 ).
k! l! (k + l)! k! l!
Exercise 1.2.41. How should we modify the formula for T1 T2 if we used
the isomorphisms k to identify A (V ) with (V )?
Exercise 1.2.42. Find a natural isomorphism
Ak (V1 V2 )
M
= Ar (V1 ) As (V2 ).

r+s=k
Exterior Powers of the Dual Space and Alternating Forms

As we saw above, there is a natural identification : A (V ) (V ) for any
finite dimensional linear space V . Let us apply this identification to the dual
space V of V . To simplify notation, set Ak (V ) = Ak (V ), A (V ) = A (V ),
and let k , , k and be the isomorphisms analogous to k , , k and
respectively, obtained when V is replaced by the dual space V .
Proposition 1.2.43. The linear space T (k,0) (V ) is naturally isomorphic to
the linear space K of k-linear functions from V k to R. Under this isomor-
phism Ak (V ) < T (k,0) (V ) corresponds to the linear space of alternating k-
linear functions on V .
Proof. The first part of the statement is a special case of Exercise 1.2.27.
((d)), and can be proved as follows. Assign to (l1 , . . . , lk ) (V )k the k-
linear function k (l1 , . . . , lk ) K given by the equality
k (l1 , . . . , lk )(v1 , . . . , vk ) = l1 (v1 ) lk (vk ).
Since k (l1 , . . . , lk ) depends on each li V linearly, k : (V )k K is k-

linear and induces a unique linear map k : T (k,0) V K such that k =
20 1. Preliminaries
k k . We show that k is an isomorphism. Choose a basis e1 , . . . , en of

V . Then
k (ei1 ...ik )(ej1 , . . . , ejk ) = k (ei1 , . . . , eik )(ej1 , . . . , ejk )

= ei1 (ej1 ) eik (ejk ) = ji11 jikk ,
where ji is the Kronecker delta symbol, thus,

Xn n
X
k Ti1 ...ik ei1 ...ik (ej1 , . . . , ejk ) = Ti1 ...ik ji11 jikk = Tj1 ...jk .
i1 ,...,ik =1 i1 ,...,ik =1
Since a k linear function is uniquely determined by its values on k-tuples

of basis vectors, this equality shows that the unique preimage of a k-linear
function : V k R is the tensor
n
X
(k )1 ( ) = (ei1 , . . . , eik )ei1 ...ik .
i1 ,...,ik =1
If Sk is a permutation, then
n
X
((k )1 ( )) = (ei1 , . . . , eik )e(i1 )...(ik )
i1 ,...,ik =1
X n
= (e1 (i1 ) , . . . , e1 (ik ) )ei1 ...ik ,
i1 ,...,ik =1
which shows that (k )1 ( ) is an alternating tensor if and only if is an

alternating k-linear function.
Definition 1.2.44. Alternating k-linear functions on a linear space V are
called alternating k-forms or shortly k-forms on V .
The composition of the isomorphisms k and k yields a natural isomorphism
between k (V ) and the linear space of alternating k-forms on V . Using this
natural isomorphism we shall identify the elements of the two linear spaces.
If l1 , . . . , lk V are linear functions on V , then l1 lk as a k-form
assigns to the vectors v1 , . . . , vk V the number
l1 lk (v1 , . . . , vk ) = (k (l1 lk ))(v1 , . . . , vk )

l1 (v1 ) l1 (vk )
X .. .. .. .
= sgn l(1) (v1 ) l(k) (vk ) = det . . .
Sk lk (v1 ) lk (vk )
1.2.6 Euclidean Linear Spaces

We know that k-linear functions on a linear space V are naturally identified
with tensors of type (k, 0) (cf. Exercise 1.2.27 ((d))). A k-linear function is
symmetric if and only if the corresponding tensor is symmetric, or, equiv-
alently, if the value of the function does not change when we permute its
variables.
Definition 1.2.45. A symmetric bilinear function h , i : V V R is called
positive definite if hv, vi 0 for all v V and equality occurs only when
v = 0.
Definition 1.2.46. A Euclidean linear space is a finite dimensional linear
space V equipped with a positive definite symmetric bilinear function h , i,
which is usually called the inner product or dot product.
For example, Rn with the standard dot product
h(x1 , . . . , xn ), (y1 , . . . , yn )i = x1 y1 + + xn yn
on it is a Euclidean vector space.

The dot product enables us to define the length of a vector.
Definition 1.2.47. The length or Euclidean norm of a vector v in a Eu-
clidean space V is p
kvk = hv, vi.
Proposition 1.2.48 (CauchySchwarz Inequality). For any two vectors v
and w in a Euclidean space V , we have
|hv, wi| kvk kwk.
Equality holds if and only if v and w are linearly dependent.

Proof. If v = 0, then both sides are equal to 0 and the vectors are linearly
dependent. If v 6= 0, then consider the quadratic polynomial
P (t) = ktv + wk2 = kvk2 t2 + 2hv, wit + kwk2 .
Since P (t) 0 for all t,

2

hv, wi 2 hv, wi hv, wi
P = kvk 2hv, wi + kwk2
kvk2 kvk4 kvk2
hv, wi2
= + kwk2 0,
kvk2
which gives the CauchySchwarz inequality after rearrangement.
22 1. Preliminaries
If the CauchySchwarz inequality holds with equality, then

2
hv, wi hv, wi
P 2
=
2
v + w = 0,
kvk kvk
which implies that w = hv,wi

kvk2 v, so v and w are linearly dependent. It is also
clear that if w = v, then both sides of the CauchySchwarz inequality are
equal to kvk2 .
Corollary 1.2.49. For any two vectors v, w V , we have

kv + wk kvk + kwk and kvk kwk kv wk.
Proof. The first inequality is equivalent to the inequality
kv + wk = kvk2 + 2hv, wi + kwk2 kvk + kwk = kvk2 + 2kvk kwk + kwk2 ,
which follows form the CauchySchwarz inequality. The second inequality is

a corollary of the first one applied to the pairs (v, w v) and (w, v w).
Exercise 1.2.50. When do we have equality in the inequalities of Corollary
1.2.49?
Corollary 1.2.51. If v and w are nonzero vectors, then
hv, wi
1 1.
kvk kwk
Definition 1.2.52. If v and w are nonzero vectors, then there is a uniquely

defined number [0, ] for which
hv, wi
cos = .
kvk kwk
The number is called the angle enclosed by the vectors v and w (measured
in radians).
Definition 1.2.53. Two non-zero vectors are orthogonal if the angle enclosed
by them is /2. Orthogonality of two non-zero vectors is equivalent to the
condition hv, wi = 0. Since the latter equation is automatically fulfilled when
v = 0, we agree, that 0 is said to be orthogonal to every vector.
Definition 1.2.54. A collection of some vectors e1 , . . . , ek of a Euclidean
linear space V is said to be an orthonormal system if hei , ej i = ij for all
1 i, j k, i.e., if the vectors have unit length and are mutually orthogonal
to one another.
Theorem 1.2.55 (GramSchmidt Orthogonalization). Assume that the vec-

tors f1 , . . . , fk of a Euclidean linear space V are linearly independent then
there is a unique orthonormal system of vectors e1 , . . . , ek satisfying the fol-
lowing two properties:
(i) lin[{f1 , . . . , fs }] = lin[{e1 , . . . , es }] for s = 1, . . . , k;
(ii) For s = 1, . . . , k, the ordered s-tuples of vectors (f1 , . . . , fs ) and (e1 , . . . ,
es ) are bases of the same orientation in the linear space they span.
Proof. According to condition (i), each vector es must be a linear combination
of f1 , . . . , fs
e1 = 11 f1 ,
..
.
es = s1 f1 + + ss fs , (1.4)
..
.
ek = k1 f1 + k2 f2 + + kk fk .
Condition (ii) on the orientation is fulfilled if and only if
1
1 0 0 ... 0
21 22 0 ... 0

.. ..

det . .. ..
= 11 ss > 0 for all 1 s k,
. . . 0
1 2 s1
s1 s1 . . . s1 0
1 2
s s . . . . . . ss
(1.5)
which means that all the diagonal elements ss must be positive.
We prove the theorem by induction on k and give an explicit recursive formula
for the computation of the vector es . For k = 1, since e1 must be a unit vector
and 11 must be positive, the only good choice for e1 is
f1
e1 = .
kf1 k
Suppose that the theorem is true for k 1. Then for f1 , . . . , fk1 we can find
an orthonormal system e1 , . . . , ek1 satisfying (i) and (ii) for 1 s k 1.
Then ek must be of the form
ek = (k1 e1 + + kk1 ek1 ) + kk fk . (1.6)
Taking the dot product of both sides of (1.6) with ei (1 i k 1), we

obtain
0 = ki + kk hfk , ei i,
24 1. Preliminaries
consequently,
ek = kk fk (hfk , e1 ie1 + + hfk , ek1 iek1 ) .

The parameter kk must be used to normalize the vector which stands on the
right of it. Thus,
kk = kfk (hfk , e1 ie1 + + hfk , ek1 iek1 )k1 .
Since kk > 0, the only possible choice of ek is
fk (hfk , e1 ie1 + + hfk , ek1 iek1 )
ek = .
kfk (hfk , e1 ie1 + + hfk , ek1 iek1 )k
It is not difficult to check that the vectors e1 , . . . , ek will satisfy the require-
ments.
Applying the Gram-Schmidt orthogonalization to a basis we obtain the fol-
lowing corollary.
Corollary 1.2.56. Every finite dimensional Euclidean linear space has an
orthonormal basis.
Definition 1.2.57. The Gram matrix of a system of vectors f1 , . . . , fk of a
Euclidean linear space V is the matrix

hf1 , f1 i . . . hf1 , fk i
G(f1 , . . . , fk ) = ... .. .. .

. .
hfk , f1 i . . . hfk , fk i
Corollary 1.2.58. Let G be the Gram matrix of the vectors f1 , . . . , fk of a
Euclidean linear space V . Then det G 0, and det G = 0 if and only if the
vectors f1 , . . . , fk are linearly dependent.
Pk
Proof. If there is a non-trivial linear relation i=1 i fi = 0, then the rows
of G are linearly dependent with the same coefficients i therefore det G = 0.
If f1 , . . . , fk are linearly independent, then applying the GramSchmidt or-
thogonalization to them we obtain an orthonormal system e1 , . . . , ek and we
can express the vectors fs as follows
f1 = 11 e1
..
.
fs = s1 e1 + + ss es
..
.
fk = k1 e1 + k2 e2 + + kk ek .
The lower triangular matrix T put together from the coefficients ij is the
inverse of the matrix (1.5) coming from the decompositions (1.4), in particular
s1
1 X
ss = = hf i
s i > 0 for s = 1, . . . , k.
, e (1.7)

s
ss
f
i=1
It is clear that G = T T T , so
det G = det(T T T ) = det(T )2 = (11 kk )2 > 0.
Corollary 1.2.59. With the notation used in Corollary 1.2.58, the identity
s1
X 2
det G(f1 , . . . , fs ) = fs hfs , ei iei det G(f1 , . . . , fs1 )

i=1
holds.
Corollary 1.2.60. A symmetric bilinear function h , i on a linear space V is
positive definite if and only if there is a basis f1 , . . . , fn such that the Gram
matrices G(f1 , . . . , fs ) have positive determinants for s = 1, . . . , n.
Proof. The previous proposition shows that if h , i is positive definite, then
any basis will be good. Conversely, assume that f1 , . . . , fn is a basis such that
G(f1 , . . . , fs ) > 0 for s = 1, . . . , n. Then the recursive formula
Ps1
f1 fs i=1 hfs , ei iei
e1 = p , es = p
det G(f1 ) det G(f1 , . . . , fs )/ det G(f1 , . . . , fs1 )
defines an orthonormal basis for h , i, i.e. hei , ej i = ij . (Check this!) The
existence
Pnof an orthonormal basis Pimplies that h , i is positive definite, because
n
if v = i=1 v i ei , then hv, vi = i=1 (v i )2 0, and equality holds only when
v = 0.
The Principal Axis Theorem

Sometimes a Euclidean linear space (V, h , i) is equipped with a second sym-
metric bilinear function { , } : V V R. In general, a symmetric bilinear
function on a linear space can be defined by its matrix with respect to a basis
or by its quadratic form. Using the inner product of the Euclidean structure
it can also be defined with the help of a self-adjoint map.
If (x1 , . . . , xn ) is a basis of the vector space V and { , } is a bilinear function
on V then the n n matrix (bij )1i,jn with entries bij = {xi , xj } is called
the matrix representation or simply the matrix of { , } with respect to the basis
(x1 , . . . , xn ). Fixing the basis we get a one to one correspondence between
bilinear functions and n n matrices. A bilinear form is symmetric if and
only if its matrix with respect to a basis is symmetric.
26 1. Preliminaries
Definition 1.2.61. The quadratic form of a bilinear function { , } is the

function defined by the equality Q{,} (x) = {x, x}.
Symmetric bilinear functions can be recovered from their quadratic forms

with the help of the identity
1
{x, y} = Q{,} (x + y) Q{,} (x) Q{,} (y) . (1.8)
2
The following proposition establishes a bijection between bilinear functions
on a Euclidean linear space V and linear endomorphisms L : V V .
Proposition 1.2.62. Let (V, h , i) be a Euclidean linear space.
(i) Then for any linear endomorphism L : V V , the map
{ , }L : V V R, {x, y}L = hLx, yi
is a bilinear function on V .
(ii) For any bilinear function { , } on V , there is a unique linear map
L : V V such that { , } = { , }L .
(iii) The bilinear function { , }L is symmetric if and only if L satisfies the
identity hLx, yi = hx, Lyi.
Proof. We prove only (ii), the rest is obvious. Choose an orthonormal basis
(e1 , . . . , en ) in V . Given a bilinear function { , }, the only possible choice for
L(x) is
Xn n
X
L(x) = hLx, ei iei = {x, ei }ei ,
i=1 i=1
which proves uniqueness. On the other hand, if we define L by the last

equality, then
n
X n X n o
hLx, yi = {x, ei }hei , yi = x, hei , yiei = {x, y},
i=1 i=1
so L is a good choice.
Definition 1.2.63. A linear endomorphism L of a Euclidean linear space is

said to be self-adjoint (with respect to the Euclidean structure) if it satisfies
the identity hLx, yi = hx, Lyi.
Lemma 1.2.64. All the eigenvalues of a self-adjoint map L : V V are

real.
Proof. Let x+iy C be an eigenvalue of L, v+iw CV be a corresponding

eigenvector. Then the real and imaginary part of the equation L(v + iw) =
(x + iy)(v + iw) gives Lv = xv yw and Lw = yv + xw. Since L is
self-adjoint, we have
xhv, wi ykwk2 = hLv, wi = hv, Lwi = xhv, wi + ykvk2 ,
which yields y(kvk2 + kwk2 ) = 0. However, v + iw 6= 0 because eigenvectors
are non-zero vectors, therefore y = 0.
Theorem 1.2.65 (Principal axis theorem). Let V be a finite dimensional
Euclidean linear space and let L : V V be a self-adjoint linear transfor-
mation on V . Then there exists an orthonormal basis of V consisting of
eigenvectors of L.
Proof. We prove by induction on the dimension n of V . The base case n = 1
is trivial. Assume that it is true for n = k. Suppose n = k + 1. By Lemma
1.2.64, there exists a unit vector v1 in V which is an eigenvector of L. Let
W = v1 = {w V | v1 w}. Then L(W ) W since we have
hLw, v1 i = hw, Lv1 i = hw, 1 v1 i = 1 hw, v1 i = 0

for any w W , where 1 is the eigenvalue belonging to v1 . Clearly LW
is self-adjoint. Since dim(W ) = dim(V ) 1 = k, the induction assumption
implies that there exists an orthonormal basis (v2 , . . . , vn ) in W consisting of
eigenvectors of LW . But each eigenvector of LW is an eigenvector of L, so
(v1 , . . . , vn ) is an orthonormal basis of V consisting of eigenvectors of L.
Induced Euclidean Structures on Tensor Spaces

Let (V, h , i) be a finite dimensional Euclidean linear space. Our goal now
is to extend the dot product in a natural way to tensor spaces and exterior
power spaces constructed from V . Consider the map
K : V 2k = V k V k R,
((v1 , . . . , vk ), (w1 , . . . , wk )) 7 hv1 , w1 i hvk , wk i.
As this map is 2k-linear, it induces a linear map L : T (0,2k) V R such that

K = L 2k . Composing L with the tensor product operation : T (0,k)
T (0,k) T (0,2k) we obtain a bilinear function on T (0,k) , which we also denote
by h , i.
Proposition 1.2.66. The induced bilinear function h , i on T (0,k) V is sym-
metric and positive definite. If e1 , . . . , en is an orthonormal basis of V , then
ei1 ...ik (1 i1 , . . . , ik n) is an orthonormal basis of T (0,k) V with respect to
h , i.
28 1. Preliminaries
Proof. It is enough to show the second part of the proposition since orthonor-
mal basis exists only for symmetric positive definite bilinear functions. This
follows from
hei1 ...ik , ej1 ...jk i = L(ei1 eik ej1 ejk )
= hei1 , ej1 i heik , ejk i = i1 ,j1 ik ,jk .
The linear space Ak (V ) of alternating tensors is a linear subspace of T (0,k) , so

1
the restriction of k! h , i defines a symmetric positive definite bilinear function
on Ak (V ). Using the natural isomorphism k between Ak (V ) and k (V )
we obtain a positive definite symmetric bilinear function h , i also on k (V )
expressed by
1
h1 , 2 i = hk (1 ), k (2 )i 1 , 2 k (V ).
k!
The reason why we divide by k! is to make the following proposition true.
Proposition 1.2.67. If e1 , . . . , en is an orthonormal basis of V , then
{ei1 eik | 1 i1 < < ik n}
is an orthonormal basis of k (V ).
Proof. If the increasing sequences 1 i1 < < ik n and 1 j1 < <
jk n are not the same, then no permutation of them can coincide, so
hei1 eik , ej1 ejk i
* +
1 X X
= sgn ei(1) ...i(k) , sgn ej(1) ...j(k) = 0.
k!
Sk Sk
When the sequences are equal, then the permuted sequences i(1) , . . . , i(k)
and j(1) , . . . , j(k) coincide if and only if = , therefore
hei1 eik , ei1 eik i
1D X X E
= sgn ei(1) ...i(k) , sgn ei(1) ...i(k)
k!
Sk Sk
1 X k!
= kei(1) ...i(k) k2 = = 1.
k! k!
Sk
Proposition 1.2.68. The following identity holds

hv1 , w1 i hv1 , wk i
hv1 vk , w1 wk i = det .. .. ..
. (1.9)

. . .
hvk , w1 i hvk , wk i
Proof. The left-hand side of (1.9) equals
hv1 vk , w1 wk i
1D X X E
= sgn v(1) v(k) , sgn w(1) . . . w(k)
k!
Sk Sk
1 X X
= sgn sgn hv(1) , w(1) i hv(k) , w(k) i
k!
Sk Sk

hv(1) , w1 i hv(1) , wk i
1 X .. .. ..
= sgn det

k! . . .
Sk hv(k) , w1 i hv(k) , wk i

hv1 , w1 i hv1 , wk i
= det .. .. ..
.

. . .
hvk , w1 i hvk , wk i
Let V be the dual space of V . The inner product is a bilinear function,

therefore it defines a linear map l : V V by l(v)(w) = hv, wi. The linear
map l is injective, since if l(v) = 0, then l(v)(v) = kvk2 = 0, so v = 0.
This implies that l is a linear isomorphism between V and V . The linear
isomorphism l can be used to identify V and its dual space, and gives rise to
a Euclidean linear space structure h , i : V V R on V , for which
hl(v), l(w)i = hv, wi.
Proposition 1.2.69. If (e1 , . . . , en ) is an orthonormal basis of V, then its

dual basis (e1 , . . . , en ) is an orthonormal basis of V .
Proof. Since the basis is orthonormal, we have
l(ei )(ej ) = hei , ej i = ij = ei (ej ),
thus, l(ei ) = ei , and
hei , ej i = hl(ei ), l(ej )i = hei , ej i = ij .
The Euclidean structure on V induces a Euclidean structure on the tensor

spaces T (k,0) V and also on k (V ). Finally, since l identifies V with its dual
space, tensors of type (k, l) can be identified with tensors of type (k + l, 0)
and also with tensors of type (0, k + l). Both identifications induces the
same Euclidean structure on T (k,l) V , with respect to which the basis eji11...i
...jl
k
generated from any orthonormal basis of V will be an orthonormal basis in
T (k,l) V .
30 1. Preliminaries
1.2.7 Hodge Star Operator

Let V be an oriented n-dimensional Euclidean linear space with inner prod-
uct h , i. Since n V is an oriented 1-dimensional Euclidean linear space, it
contains exactly one positively oriented unit vector n V .
Definition 1.2.70. The Hodge star operator : k (V ) nk V is a linear
operator. For k (V ), is defined as the unique element in nk V
satisfying the identity
() = h, i for all k V. (1.10)

The following proposition says that the definition is correct.
Proposition 1.2.71. For each k V, there is exactly one element
nk V which satisfies (1.10).
Proof. The proof will give an explicit method to compute . Choose a
positively oriented orthonormal basis e1 , . . . , en in V . Then = e1 en .
The system
{ei1 eik |1 i1 < < ik n} (1.11)
is an orthonormal basis of k V , so we can write and its unknown Hodge
star as
X X
= i1 ...ik ei1 eik , = ()j1 ...jnk ej1 ejnk
1i1 <<ik n 1j1 <<jnk n
Both sides of (1.10) are linear in , so the equation holds for all k (V ) if
and only if it is true for all vectors of a basis. Substituting the basis vector
= ei1 eik into (1.10) we obtain the equality
()j1 ...jnk ei1 eik ej1 ejnk = i1 ...ik e1 . . . en ,
where j1 < < jnk are the elements of the complementary set {1, 2, . . . , n}\
{i1 , . . . , ik } arranged in increasing order. Thus, denoting by (i1 , . . . , ik ) the
sign of the permutation (i1 , . . . , ik , j1 , . . . , jnk ),
()j1 ...jnk = (i1 , . . . , ik ) i1 ...ik .
Corollary 1.2.72. Since the Hodge operator maps the orthonormal basis
(1.11) into an orthonormal basis of nk V it is an orthogonal transformation.
Corollary 1.2.73. As
sgn(i1 , . . . , ik , j1 , . . . , jnk ) = (1)k(nk) sgn(j1 , . . . , jnk , i1 , . . . , ik ),
we have |k (V ) = (1)k(nk) idk (V ) .

Cross Product
Definition 1.2.74. Let V be a 3-dimensional oriented Euclidean vector field.
Then the binary operation : V V V , (a, b) 7 a b = (a b) is called
the cross product operation.
Proposition 1.2.75. The cross product has the following properties, which
give a geometrical way to construct it.
a b is orthogonal to both a and b.
The length of a b is equal to the area of the parallelogram spanned by
a and b. In particular, a b = 0 if and only if a and b are linearly
dependent. If a and b are nonzero and the angle between them is ,
then ka bk = kak kbk sin(). Geometrically, kak kbk sin() is the
area of the parallelogram spanned by the vectors a and b.
If a and b are linearly independent, then (a, b, a b) is a positively
oriented basis of V .
Proof. Let be the only positively oriented trivector of unit length in 3 V .
Then
0 = a (a b) = ha, a bi and 0 = b (a b) = hb, a bi,
which implies that ha, a bi = hb, a bi = 0.

Since the Hodge star operator preserves length,
kak2 ha, bi

2 2
ka bk = ka bk = det G(a, b) = det .
hb, ai |bk2
The last determinant is 0 if any of the vectors a or b is 0. If none of them is

0 and the angle between them is , then ha, bi = kak kbk cos(), and the
value of the determinant is kak2 kbk2 (1 cos2 ()) = (kak kbk sin())2 .
We also have
a b (a b) = ha b, (a b)i = ha b, a bi = ka bk2 .
If a and b are linearly independent, then this is a positive multiple of ,

which means that (a, b, a b) is a positively oriented basis of V .
Proposition 1.2.76. We have the following identities for the cross product:
a b = b a (skew-commutativity);
(a1 + a2 ) b = a1 b + a2 b
(bilinearity);
a (b1 + b2 ) = a b1 + a b2
32 1. Preliminaries
ha, bci = hb, cai = hc, abi (permutation rule for mixed product);

ha, ci ha, di
ha b, c di = det (Lagrange identity);
hb, ci hb, di
a (b c) = ha, cib ha, bic (triple product expansion formula);
a (b c) + b (c a) + c (a b) = 0 (Jacobi identity).
Proof. Skew-commutativity and bilinearity follow directly from analogous
properties of the wedge product. The permutation rule is a consequence of
the identity a (b c) = ha, b ci and the alternating property a b c =
b c a = c a b of the wedge product.
As the Hodge star operator preserves the inner product, ha b, c di =
ha b, c di, and thus, Lagrange identity is a special case of equation (1.9).
Combining the permutation rule and the Lagrange identity, we obtain that
ha(bc), di = hda, bci = hd, biha, cihd, ciha, bi = hha, cibha, bic, di.
Since the first and last terms are equal for any d, a (b c) must be equal
to ha, cib ha, bic.
Expanding the summands on the left-hand side of the Jacobi identity we
obtain
(ha, cib ha, bic) + (hb, aic hb, cia) + (hc, bia hc, aib) = 0.
Proposition 1.2.77. Let (e1 , e2 , e3 ) be a positively oriented orthonormal

basis of V, and let (a1 , a2 , a3 ) and (b1 , b2 , b3 ) be the coordinates of the vectors
a and b with respect to this basis. Then

e1 e2 e3
a b = det a1 a2 a3 .
b1 b2 b3
Proof. It is clear that e1 e2 = (e1 e2 ) = e3 , and similarly e2 e3 = e1

and e3 e1 = e2 . Thus,
a b = (a1 e1 + a2 e2 + a3 e3 ) (b1 e1 + b2 e2 + b3 e3 )
= (a2 b3 a3 b2 )e1 + (a3 b1 a1 b3 )e2 + (a1 b2 a2 b1 )e3 .
Sometimes we have to find a normal vector of an (n 1)-dimensional linear

subspace W of an n-dimensional linear space V . If a1 , . . . , an1 is a basis of
W , then the vector (a1 an1 ) will be a normal vector of W . The
(n 1)-variable operation (a1 , . . . , an1 ) 7 (a1 an1 ) is clearly a
higher dimensional generalization of the cross product and it has analogous
properties that can be proved in a similar way.
1.3. Geometry 33
Proposition 1.2.78.
The vector c = (a1 an1 ) is an alternating (n1)-linear function
of the vectors a1 , . . . , an1 .
c is orthogonal to ai for 1 i n 1.
c = 0 if and only if the vectors a1 , . . . , an1 are linearly dependent.
If a1 , . . . , an1 are linearly independent, then the length of c is the (n
1)-dimensional volume of the (n1)-dimensional p parallelepiped spanned
by the vectors a1 , . . . , an1 , which is equal to det G(a1 , . . . , an1 ).
(a1 , . . . , an1 , c) is a positively oriented basis of V .
If the coordinates of ai with respect to a positively oriented orthonormal
basis (e1 , . . . , en ) of V are (a1i , . . . , ani ), then
. . . a1n1
1
an1

a1
.. .. .. ..
c = det .
. . .
.
a1 n1 n
n1 . . . a n1 an1

e1 . . . en1 en
The proof is left to the reader.
1.3 Geometry
1.3.1 Affine Geometry
Affine Spaces
In traditional axiomatic treatment of Euclidean geometry, vectors are defined
as equivalence classes of ordered pairs of points (also called directed seg-
ments). If X is the set of points of the space, then an ordered pair of points
is simply an element of the Cartesian product X X. If (A, B) X X is an
ordered pair of points, then A is called the initial point, and B is called the
endpoint of the ordered pair. The ordered pairs of points (A, B) and (C, D)
are said to be equivalent if the midpoint of the segment [A, D] coincides with
the midpoint of the segment [B, C]. It can be shown that this is indeed an
equivalence relation. The equivalence classes are called (free) vectors. The
set V of all vectors is equipped with a linear space structure. The sum of two
vectors is constructed by the triangle or parallelogram rule, multiplication by
scalars is defined in the usual way. The map : X X V, which assigns

to each ordered pair of points (A, B) the vector (A, B) = AB represented
by it, i.e., its equivalence class, satisfies the following properties.
34 1. Preliminaries
(A1) For any A X and v V there is a unique B X such that (A, B) =

v.
(A2) (Triangle rule.) For any three points A, B, C X, we have (A, B) +
(B, C) = (A, C).
Figure 1.1: The construction of the sum of two vectors by the triangle and
the parallelogram rule.
The notion of an affine space generalizes this picture.

Definition 1.3.1. An affine space is a triple A = (X, V, ), where X is a set,
the elements of which are called points, V is a linear space, the elements of
which are called vectors, and : X X V is a map satisfying conditions
(A1) and (A2) above. The dimension of the affine space A is the (linear
algebraic) dimension of the linear space V .

When it leads to no confusion, (A, B) is also denoted by AB.
Definition 1.3.2. Let A = (X, V, ) and B = (Y, W, ) be two affine spaces.
Then an affine transformation from A to B consists of a map T : X Y and
a linear map L : V W such that
L((A, B)) = (T (A), T (B)) for all A, B X. (1.12)
Observe that since is surjective, T determines L uniquely by (1.12).

Affine spaces form a category in which the morphisms are the affine trans-
formations. An affine transformation is an isomorphism if and only if T is a
bijection.
Example. Let V be an arbitrary linear space, X = V , and define : X
X V by (p, q) = q p. Then AV = (X, V, ) is an affine space.
Thus, every linear space carries an affine space structure. If V and W are
1.3. Geometry 35
linear spaces, then an affine transformation from AV to AW has the form

T (v) = L(v) + w0 , where L : V W is a linear map, w0 W is a fixed
vector.
Linear spaces and affine spaces are very similar objects. The main difference
is that in a linear space, we always have a distinguished point, the origin 0,
whereas in an affine space, none of the points of X is distinguished. This is
essentially the only difference, because if we choose any of the points of X
for the origin, we can turn X into a linear space isomorphic to V .
Indeed, choose a point O X. By the first axiom (A1) of an affine space,
the map O : X V , O (P ) = (O, P ) is a bijection between X and V ,
and we have
(A, B) = O (B) O (A)
by (A2). This identity means that O together with the linear map L = idV
is an isomorphism between the affine space A = (X, V, ) and the affine space
AV .
The vector O (P ) is called the position vector of A from the base point O.
Identification of points of an affine space with vectors with the help of a fixed
base point O is called vectorization of the affine space.
Affine Subspaces
Definition 1.3.3. A nonempty subset Y X of the point set of an affine
space A = (X, V, ) is an affine subspace of A if the image (Y Y ) = W
of Y Y under is a linear subspace of V and (Y, W, |Y Y ) is an affine
space. W is called the direction space of the affine subspace, its elements are
the direction vectors of Y , or the vectors parallel to Y . Since affine subspaces
are affine spaces themselves, their dimension is properly defined.
The 0-dimensional affine subspaces of an affine space A = (X, V, ) are the

points of X. The 1-dimensional affine subspaces are called straight lines.
The 2-dimensional affine subspaces are the ordinary planes of A. In gen-
eral, k-dimensional affine subspaces of an affine space will be called shortly
k-planes. The (n 1)-planes of an n-dimensional affine space are called hy-
perplanes. The set of all k-dimensional affine subspaces of an affine space A
is the affine Grassmann manifold AGrk (A). At this moment we defined the
affine Grassmann manifold just as a set, but this set can be endowed with a
(k + 1)(n k)-dimensional manifold structure, which justifies its name.
The following proposition gives a more explicit description of affine subspaces
in vectorized affine spaces.
Proposition 1.3.4. A subset Y V of the affine space AV is an affine

subspace with direction space W < V if and only if Y is a translate of W .
36 1. Preliminaries
Proof. If Y is an affine subspace with direction space W , then according to

axiom (A1), for any a Y , the map Y W , b 7 b a is a bijection
between Y and W , which means that Y is the translate of W by the vector
a.
Assume now that Y = Ta (W ) is a translate of the linear subspace W of V ,
and show that Y is an affine subspace of AV with direction space W . Two
typical elements of Y have the form y1 = a + w1 and y2 = a + w2 , where
w1 , w2 W . Then
(y1 , y2 ) = y2 y1 = w2 w1 ,
which shows that (Y Y ) = W .

To check axiom (A1) for (Y, W, |Y Y ), observe that for y0 = a + w0 Y ,
the map Y W , y 7 (y0 , y) = y y0 is a translation by y0 restricted
onto Y . Since Ty0 = Tw0 Ta and Ta maps Y onto W bijectively, and
Tw0 maps W onto itself bijectively, Ty0 maps Y onto W bijectively.
Since the triangle rule (A2) is fulfilled for any triple in X, it is true for any
three points in Y as well.
Affine and 0-weight Linear Combinations

When we vectorize an affine space, points are identified with vectors and,
therefore, we can take linear combinations of points. The result can be con-
sidered both a point and a vector. However, the resulting point, or vector
may depend on the choice of the base point of the vectorization. From the
viewpoint of affine geometry, linear combinations the result of which does not
depend on the choice of the vectorization are of special importance.
Let us compare vectorizations with base points O and O0 . If (O, O0 ) = a,
and a point P corresponds to the vectors O (P ) = p and O0 (P ) = p0 under
the vectorizations with these base points, then we have p = p0 + a by the
triangle rule (A2).
If the position vectors of the points P1 , . . . , Pk X from the base point O
are p1 , . . . , pk , then the linear combination 1 p1 + + k pk as a vector in
V does not depend on the vectorization if and only if
1 p1 + + k pk = 1 (p1 a) + + k (pk a)
for any a V , or, equivalently, if 1 + + k = 0.
Definition 1.3.5. A linear combination 1 p1 + +k pk is a 0-weight linear

combination if the sum 1 + + k of the coefficients is equal to 0.
According to the previous computation, we obtain the following statement.

1.3. Geometry 37
Proposition 1.3.6. For 0-weight linear combinations the vector 1 p1 + +

k pk V does not depend on the vectorization, thus it can be denoted by
1 P1 + + k Pk V as well.
We get a different condition on the coefficients if we want that the point with
position vector 1 p1 + + k pk with base point O be the same as the point
with position vector 1 p01 + + k p0k with base point O0 . For this we need
that the equation
1 p1 + + k pk a = 1 (p1 a) + + k (pk a)
should hold for any choice of a. This condition is clearly equivalent to the
condition that 1 + + k = 1.
Definition 1.3.7. A linear combination 1 p1 + + k pk is an affine com-
bination if the sum 1 + + k of the coefficients is equal to 1.
The importance of affine combinations is summarized in the following propo-
sition.
Proposition 1.3.8. The point represented by an affine combination 1 p1 +
+ k pk of the position vectors of some points P1 , . . . , Pk does not depend
on the choice of the base point. This way, it makes sense to denote it by
1 P1 + + k Pk X.
Affine subspaces can be characterized in terms of affine combinations.
Proposition 1.3.9. A subset Y of an affine space is an affine subspace if
and only if it is nonempty and it contains all affine combinations of its points.
Proof. Identify the space with AV by vectorization.
If Y is an affine subspace, then it is a translation Ta W of its direction space.
Since 0 W , a Y is not empty. Furthermore, if yi = a + wi Y ,
(i = 1, . . . , k) are some points, then an affine combination of the has the form
1 y1 + + k yk = (1 + + k )a + (1 w1 + + k wk )
= a + (1 w1 + + k wk ).
Since W is a linear space, w = 1 w1 + + k wk W , therefore, 1 y1 +

+ k yk = a + w Y .
To show the other direction, assume now that Y is a nonempty subset of V
containing all affine combinations of its points.
Consider the set W of all 0-weight linear combinations of elements of Y . Since
the sum of two 0-weight linear combinations and any multiple of a 0-weight
linear combination are 0-weight linear combinations, W is a linear subspace
of V .
38 1. Preliminaries
Choose a vector a Y , and let us show that Y = Ta (W ). Adding a = 1a to a

0-weight linear combination of elements of Y we obtain an affine combination
of some elements of Y , which belongs to Y by our assumptions on Y . Thus,
Y Ta (W ). On the other hand, every y element of Y can be written as
y = a + (y a) Ta (W ), which completes the proof.
As a byproduct of the proof we obtain
Corollary 1.3.10. The direction space of an affine subspace is the set of all
0-weight linear combinations of its elements.
Corollary 1.3.11. The intersection of affine subspaces of an affine space is
either empty or it is also an affine subspace.
Corollary 1.3.12. For a nonempty subset S X there is a smallest affine
subspace among the affine subspaces containing S.
Definition 1.3.13. The smallest affine subspace containing the nonempty
subset S is called the affine subspace spanned by S and it is denoted by aff[S].
Proposition 1.3.14. The affine subspace spanned by the nonempty subset

S consists of all affine combinations of the elements of S.
Proof. Denote by YS the set of all affine combinations of elements of S. By
Proposition 1.3.9, aff[S] YS . Since for any s S, 1 s is an affine combi-
nation of s, therefore S YS . It is easy to check that affine combinations
of affine combinations of elements of S is again an affine combination of ele-
ments of S, so YS is an affine subspace containing S. However, among such
affine subspaces aff[S] is the smallest one, so aff[S] YS .
Affine Independence
A direction space of the affine subspace spanned by a system of k + 1 points
P0 , . . . , Pk consists of 0-weight linear combinations of these points. This
linear space is generated by the differences P1 P0 , . . . , Pk P0 , therefore its
dimension is at most k.
Definition 1.3.15. The points P0 , . . . , Pk are called affinely independent if
any of the following equivalent conditions is fulfilled.
P0 , . . . , Pk span a k-dimensional affine subspace.
The vectors P1 P0 , . . . , Pk P0 are linearly independent.
A 0-weight linear combination of P0 , . . . , Pk equals 0 only if all the
coefficients are equal to 0.
1.3. Geometry 39
Affine Coordinate Systems

An affine coordinate system on an affine space A = (X, V, ) is given by the
following data:
a point O X, which will be the origin of the coordinate system;
and a basis e1 , . . . , en of V .
Given an affine coordinate system, the coordinates of a point P are the real
numbers (x1 , . . . , xn ) for which
(O, P ) = x1 e1 + + xn en .
Assigning to each point of X its coordinate vector gives an affine isomorphism
between the affine space A and the affine space ARn .
Tangent Vectors at a Point, and the Tangent Bundle

A free vector v V of an affine space (X, V, ) has no given initial point
or endpoint. It can be represented by any ordered pair of points (P, Q) for
which (P, Q) = v. Sometimes we have to consider vectors with a given base
point instead of free vectors. For example, forces in Newtonian mechanics are
vector like quantities. However, to describe a diagram of forces properly, it is
not enough to know the directions and the magnitudes of the acting forces,
we have to know also the (base) points at which the forces act on a given
body. For instance, if we push a wardrobe by a horizontal force, sufficiently
large to overcome friction, then the wardrobe will slide horizontally if we push
it close to the floor, but it may fall over if it is pushed at the top. The reason
of this fact is that the torque of a force depends on the point at which the
force acts on a body.
Figure 1.2: Pushing a body with equal forces at different base points.
40 1. Preliminaries
Definition 1.3.16. Let P X be a point of an affine space (X, V, ). A

vector based at P or a tangent vector at P is a pair (P, v), where v V.
The tangent space TP A (or TP X) of the affine space A at P is the set of all
tangent vectors at P , i.e., TP A = {(P, v) | v V }.
Tangent vectors at P form a linear space with the operations
(P, v) + (P, w) = (P, v + w).
Forgetting the base point is a linear isomorphism P : TP A V , P (P, v) =

v between TP A and V . If P and Q are two different points in X, then the
linear isomorphism 1
Q P : TP A TQ A is called parallel transport between
tangent spaces at P and Q.
Definition 1.3.17. The disjoint union of S all tangent spaces of an affine
space is called its tangent bundle T A = P X TP A = X V . The map
: T A X, which assigns to each tangent vector (P, v) its base point P is
called the projection of the tangent bundle.
1.3.2 Euclidean Spaces

Definition 1.3.18. An affine space (X, V, ) is a Euclidean space if the
linear space V is endowed with positive definite symmetric bilinear function
h , i, making it a Euclidean linear space.
Since every linear subspace W of V inherits a Euclidean structure by restrict-
ing the inner product h , i onto W W , affine subspaces of a Euclidean space
are Euclidean spaces as well.
The Euclidean structure enables us to introduce metric notions like distance,
angle, area, volume etc.
Distance and the Isometry Group

Definition 1.3.19. The distance between two points P and Q of a Euclidean
space (X, V, ) is d(P, Q) = k(P, Q)k.
Definition 1.3.20. A metric space is a pair (X, d), where X is a set, d : X
X R is a functions, called distance function, satisfying the following axioms
d(P, Q) 0 for all P, Q X and d(P, Q) = 0 if and only if P = Q.
d(P, Q) = d(Q, P ) for all P, Q X (symmetry).
d(P, Q) + d(Q, R) d(P, R) for all P, Q, R X (triangle inequality).
Proposition 1.3.21. Every Euclidean space is a metric space with its dis-
tance function.
1.3. Geometry 41
Proof. The triangle inequality follows from Corollary 1.2.49

d(P, Q) + d(Q, R) = k(P, Q)k + k(Q, R)k k(P, Q) + (Q, R)k
= k(P, R)k = d(P, R).
The rest is trivial.
Definition 1.3.22. A map : X Y between the metric spaces (X, dX )
and (Y, dY ) is said to be a distance preserving map or isometry if dY ((P ),
(Q)) = dX (P, Q) for all P, Q X.
An isometry is always injective but not necessarily surjective.
Exercise 1.3.23. Give an example of an isometry : X X which maps
a metric space (X, d) onto one of its proper subsets.
Definition 1.3.24. Bijective isometries of a metric space form a group with
respect to composition. The group of bijective isometries of the metric space
X is called the isometry group of X, and is denoted by Iso(X).
Theorem 1.3.25. Every isometry of a Euclidean space En = (X, V, ) into
itself is bijective. Isometries of En are affine transformations. An affine
transformation T : X X is an isometry if and only if the corresponding

linear map L : V V , for which L(AB) = T (A)T (B), preserves the norm
of vectors.
A linear transformation of the Euclidean linear space V preserves the norm
of vectors if and only if it preserves the dot product of vectors. If M is the
matrix of L with respect to an orthonormal basis, then this property of L is
also equivalent to the matrix equation
M M T = M T M = I. (1.13)
Definition 1.3.26. Linear transformations of a Euclidean space preserving
the dot product are called orthogonal linear transformations. Matrices satis-
fying equation (1.13) are called orthogonal matrices.
Definition 1.3.27. An orientation preserving isometry or motion of a Eu-
clidean space is an isometry, the associated linear transformation of which
has positive determinant.
Orientation preserving isometries form a normal subgroup Iso+ (En ) of index
2 in the isometry group Iso(En ) of the Euclidean space.
The Angle Between Affine Subspaces

Angle between affine subspaces of the same dimension will be defined as the
angle between their direction spaces. Direction spaces are linear subspaces,
so we first define the angle between linear subspaces.
42 1. Preliminaries
Definition 1.3.28. The angle (e, f ) between two 1-dimensional linear sub-
spaces e = lin[v] and f = lin[w] of a Euclidean linear space V , is the smaller
of and , where is the angle between v and w.
Clearly, (e, f ) is the unique angle in [0, /2] for which
|hv, wi|
cos((e, f )) = .
kvk2 kwk2
Proposition 1.3.29. is a metric on the projective space P(V ).
Proof. We prove only the triangle inequality, the rest is trivial. Let e = lin[v],
f = lin[w], g = lin[z] be three 1-dimensional subspaces and let us show that
(e, f ) + (f, g) (e, g).
Assume that the vectors v, w and z have unit length. Denote by , and
the angles between (v, w), (w, z) and (z, v) respectively. Changing the
direction of the vectors w and z if necessary, we may assume that ,
[0, /2]. By Corollary 1.2.58, the Gram matrix of the vectors v, w, z has
nonnegative determinant, thus

1 cos() cos()
det cos() 1 cos()
cos() cos() 1
= 1 + 2 cos() cos() cos() cos2 () cos2 () cos2 () 0.
This is a quadratic inequality for cos(), which is fulfilled if and only if cos()
is in the closed interval the endpoints of which are the roots of the polynomial
P (t) = (1 cos2 () cos2 ()) + 2 cos() cos()t t2 .
The roots of P are

p
cos() cos() cos2 () cos2 () + 1 cos2 () cos2 () =
p
= cos() cos() (1 cos2 ())(1 cos2 ()) = cos( ).
Since the cosine function is strictly decreasing on the interval [0, ], this
implies
| | + ,
in particular,
(e, f ) + (f, g) = + min{, } = (e, g).

1.3. Geometry 43
We define the angle between k-dimensional linear subspaces using the Plucker
embedding Grk (V ) P(k V ) (see Corollary 1.2.33).
Definition 1.3.30. The angle between two k-dimensional linear subspaces

W1 , W2 of a linear space V is the angle between the 1-dimensional linear
subspaces lin[w11 w1k ] and lin[w21 w2k ], where w11 , . . . , w1k is a
basis of W1 , w21 , . . . , w2k is a basis of W2 .
By Proposition 1.2.67, the angle between W1 and W2 is the unique angle

in the interval [0, /2] for which
1 1
hw1 , w2 i . . . hw11 , w2k i

.. .. ..
det

. . .

hw1k , w21 i . . . hw1k , w2k i
cos() = p p .
det G(w11 , . . . , w1k ) det G(w21 , . . . , w2k )
According to Proposition 1.3.29, the angle between subspaces is a metric on

the Grassmann manifold Grk (V ).
Definition 1.3.31. The angle between two affine subspaces of a Euclidean

space is the angle between their direction spaces.
Since two different affine subspaces can have the same direction space, the
angle function is not a metric on the affine Grassmann manifold AGrk (A).
Cartesian Coordinate Systems

Definition 1.3.32. An affine coordinate system given by the origin O and
the basis vectors (e1 , . . . , en ) is a Cartesian coordinate system, if the basis
(e1 , . . . , en ) is orthonormal.
A Cartesian coordinate system defines an isometric isomorphism between

any n-dimensional Euclidean space and the standard model ARn of the n-
dimensional Euclidean space, where the inner product on Rn is the standard
dot product.
Equations and Parameterizations of Affine Subspaces

Linear subspaces
Let us first deal with equations and parameterizations of linear subspaces of
a linear space V . A k-dimensional linear subspace W of V can be given by
k linearly independent vectors a1 , . . . , ak spanning W . Then a vector x V
belongs to W if and only if x can be written as a linear combination of the
vectors a1 , . . . , ak .
44 1. Preliminaries
This way, the mapping
r : Rk V, r(x1 , . . . , xk ) = x1 a1 + + xk ak
maps Rk bijectively onto W . The map r is called a (linear) parameterization

of W . A linear subspace has many linear parameterizations, since it has many
different bases.
We can also define a linear subspace as the kernel of a linear map, i.e., as the
set of solutions of a system of linear equations.
For this purpose, observe that x is in W if and only if a1 , . . . , ak , x are
linearly dependent. By Proposition 1.2.32, linear dependence is equivalent to
the equation
x a1 ak = 0.
This is called the equation of the linear subspace W . This equation requires
the vanishing of a (k + 1)-vector. If dim V = n, then the space of k+1 V
n
has dimension k+1 . If we fix a basis in V , then the equation of W becomes
n

equivalent to a system of k+1 equations requiring the vanishing of all the
coordinates of the (k + 1)-vector on the left-hand side. Since the coordinates
are linear functions of x, all the equations in the system are linear. Obviously,
these equations are not linearly independent in general. By the dimension
formula (1.2) any maximal linearly independent subsystem of this system of
equations contains exactly (n k) equations. Thus there are many different
ways to convert the equation of a linear subspace to an independent system
of (n k) linear equations.
In the special case k = n 1 of linear hyperplanes, however, we have only one
linear equation on the coordinates, and it is uniquely determined by W up to
a scalar multiplier. If we introduce an orientation and a Euclidean structure
on V , then the equation of W can be written also as
hx, (a1 an1 )i = 0.
This way, W contains all vectors orthogonal to (a1 an1 ). This last
vector is a normal vector of W .
Affine subspaces
Assume that we want to parameterize or write the equation of a k-dimensional
affine subspace W spanned by k + 1 affinely independent points a0 , . . . , ak .
It is not difficult to parameterize W , that is to obtain it as the image of a
bijection r : Rk W . We know that a point belongs to W if and only if it is
an affine combination of the points a0 , . . . , ak , consequently, the mapping
r(x1 , . . . , xk ) = (1 (x1 + + xk ))a0 + x1 a1 + + xk ak
is a parameterization of W .
1.3. Geometry 45
To write the equation of an affine subspace we use a trick rooted in projective

geometry. Vectorize our affine space by choosing an origin. This identifies
the space with AV , where V is a linear space. Consider the linear space
hV i = RV and embed V into hV i by the mapping v 7 v, where v = (1, v).
The image V of V is the translate of the linear subspace {0}V by the vector
(1, 0), thus, it is an affine subspace of hV i, just as the image W of W . Denote
by hW i the linear subspace of hV i spanned by W
hW i = lin[W ] = aff[W {(0, 0)}].
hW i is a (k + 1)-dimensional linear subspace and uniquely determines W

as the projection of W = hW i V onto V . This way, the map W 7 hW i
gives an embedding of the affine Grassmann manifold AGrk (V ) into the linear
Grassmann manifold Grk+1 (hV i).
Figure 1.3: The correspondence between affine subspaces of V and linear

subspaces of hV i.
Using this picture, x belongs to the affine subspace W if and only if x

hW i. Since the linear subspace hW i is spanned by the vectors a0 , . . . , ak , the
equation of the affine subspace W is
x a0 ak = 0.
Introducing a basis in V , this equation can be converted into a system of

equations, in which each equation requires the vanishing of a linear combi-
nation of the coordinates of x. This means that every affine subspace can be
defined by a system of inhomogeneous linear equations of the form
c0 + c1 x1 + + cn xn = 0. (1.14)
The number of independent equations that define W is (n+1)(k+1) = nk.

In the case of a hyperplane, we need only one equation of the form (1.14).
46 1. Preliminaries
In a Cartesian coordinate system, the coefficients describe the hyperplane

as follows. N = (c1 , . . . , cn ) is a normal vector of the hyperplane, and the
hyperplane goes through the point c0 N/kNk2 . The constant c0 can be
computed from any point p0 lying in the hyperplane as c0 = hN, p0 i.
Equations of Spheres
Definition 1.3.33. Let be a (k+1)-dimensional subspace of an n-dimensional
Euclidean space, O be a point, r > 0 be a positive number. Then the
k-dimensional sphere or shortly the k-sphere in centered at O with radius
r is the set of those points in the distance of which from O is r. A hyper-
sphere is an (n 1)-dimensional sphere. The case r = 0 is considered to be
a degenerate case, when the sphere degenerates to a point.
In the non-degenerate case, a k-sphere determines its (k + 1)-plane, its center

and its radius uniquely.
To write the equations of k-spheres introduce a Cartesian coordinate system
on the space. This identifies our space with Rn .
Consider first hyperspheres ( = Rn ). If o is the center of the sphere, then
the equation of the sphere is
kx ok2 r2 = kxk2 + h2o, xi + (kok2 r2 ) = 0.
This proves that every hypersphere can be defined by an equation of the form
akxk2 + hb, xi + c = 0, (1.15)
where a, c R, b Rn .
Definition 1.3.34. Equations of the form (1.15) will be called hypersphere

equations or shortly sphere equations.
The solution set of a sphere equation is described in the following proposition.
Proposition 1.3.35. Let S = {x | akxk2 + hb, xi + c = 0} be the set of

solutions of the sphere equation (1.15). Then we have the following cases.
If the equation is trivial, i.e., a = c = 0, b = 0, then the solution set

is the whole space Rn .
If the equation is contradictory, i.e., a = 0, b = 0, but c 6= 0, then

S = .
If the equation is inhomogeneous linear, that is a = 0, but not trivial

and not contradictory, that is b 6= 0, then S is a hyperplane.
1.3. Geometry 47
If the equation is quadratic, i.e., a 6= 0, then S depends on the sign of

the discriminant d = kbk2 4ac.
d > 0, then S is a hypersphere centered at b/(2a) with radius

If
d/|2a|.
If d = 0, then S is a degenerated hypersphere of radius 0 which
contains a single point b/(2a).
If d < 0, then S = .
Proof. The linear case is simple, the quadratic case follows from the equiva-
lent rearrangement
2 2
x b = kbk 4ac

2a 4a2
of equation (1.15).
Proposition 1.3.36. The intersection S of a hypersphere of radius r centered

at O and a hyperplane in an n-dimensional Euclidean space is described as
follows. Let the orthogonal projection of O onto be O0 and denote by d the
distance
OO0 . Then if r > d, then S is an (n 2)-sphere in with radius
r d , and center at O0 . S consists of the single point O0 when r = d and
2 2
S = when r > d.
The proof is a corollary of the Pythagorean theorem, since for any P in ,

the triangle P OO0 4 is a right triangle.
Figure 1.4: The intersection of a sphere and a hyperplane.
Corollary 1.3.37. The intersection of a finite system of hyperspheres and

hyperplanes is either empty or an affine subspace or a k-sphere.
48 1. Preliminaries
Proof. From an algebraic viewpoint, the intersection is the set of solutions of

a finite number of sphere equations. If there is a quadratic sphere equation in
the system, then subtracting a suitable multiple of it from the other equations,
we can make all the other equations linear, so we may assume that there is
at most one hypersphere in the family.
If there is no hypersphere at all, the statement follows from Corollary 1.3.11.
If there is a hypersphere in the family, then we can prove the statement
inductively, using the previous proposition.
To write the equation of a k-sphere, we use an embedding of Rn into a larger
linear, space analogous to the affine embedding V hV i = R V that was
used when we wrote equations of affine subspaces.
To simplify notation, denote the Euclidean linear space Rn by V . Set hhV ii =
R V R. Embed V into hhV ii with the map
v 7 v, where v = (1, v, kvk2 ). (1.16)
Because of the quadratic term in the last coordinate, this is not an affine
embedding. Its image V is a paraboloid. The advantage of this embedding
is that a linear equation on the coordinates of v is a hypersphere equation
on v.
Proposition 1.3.38. Let a0 , . . . , ak be (k + 1) 2 affinely independent
points. Then there is a unique (k 1)-sphere through these points and it can
be defined by the equation
x a0 ak = 0. (1.17)
Proof. The (k 1)-sphere must be in the unique k-plane spanned by the

points. If the points were contained in two different (k 1)-spheres of , then
they would be contained in their intersection, which is a (k 2)-sphere. This
would contradict affine independence of the points, since every (k 2)-sphere
is contained in a (k 1)-plane. Thus, uniqueness is proved.
To complete the proof, we should check that the set given by the equation
defines a (k 1)-sphere passing through the given points. Let S be the set
of solutions of (1.17).
Since (1.17) is equivalent to a system of linear equations on the coordinates
of x, which is a system of sphere equations on x. Thus, by Corollary 1.3.37,
S is either empty, or an affine subspace, or a sphere.
The set S cannot be empty, since it contains the points a0 , . . . , ak as
ai a0 ak = 0.
This also gives a lower bound on the dimension of S. If it is an m-sphere,

then m k 1, if it is an m-plane, then m k.
1.4. Topology 49
If x S, then x is a linear combination of the vectors a0 , . . . , ak , hence
(1, x, kxk2 ) = 0 (1, a0 , ka0 k2 ) + + k (1, ak , kak k2 )
for some 0 , . . . , k R. This equation splits into three components
1 = 0 + + k ,
x = 0 a0 + + k ak ,
kxk2 = 0 ka0 k2 + + k kak k2 .
The first two equations show that x must be an affine combination of the
points a0 , . . . , ak . Thus, S , and we have only two possibilities left.
Either S is a (k 1)-sphere, in which case we are done, or S = . The latter
case can be excluded by showing that the midpoint (a0 + a1 )/2 is not in S. If
it were in S, then by affine independence of the ai s, the only possible choice
for the i s would be 0 = 1 = 1/2 and i = 0 for 2 i k, however, the
third equation is not fulfilled with these coefficients as
a0 + a1 2 a0 a1 2

1 1
ka0 k2 + ka1 k2
=
> 0.
2 2 2 2
This proves the proposition.
Denote by SPHk (V ) the set of all k-spheres in the Euclidean linear space
V . As a corollary of the previous proposition, we can construct an embed-
ding SPHk (V ) Gr1 (k+1 hhV ii) in the following way. Given a k-sphere
S, choose k affinely independent points a0 , . . . , ak from it, and assign to S
the 1-dimensional linear space spanned by the (k + 1)-vector a0 ak
k+1 hhV ii.
Exercise 1.3.39. Prove that the center of the sphere defined by (1.17) is the
point 0 a0 + + k ak , where the coefficients are obtained as the solution
of the following system of linear equations
0 + + k = 1
0 hai a0 , a0 i + + k hai a0 , ak i = kai k2 ka0 k2 for i = 1, . . . , k.
1.4 Topology
Definition 1.4.1. A pair (X, ) is said to be a topological space if X is a set,
is a collection of subsets of X, which we call the open subsets of X, such
that
50 1. Preliminaries
(i) the empty set and X are in ;

(ii) the intersection of any two open subsets is also open;
(iii) the union of an arbitrary family of open subsets is open.
The family of open subsets is called the topology on X.
Definition 1.4.2. We say that a subset U of X is a neighborhood of a point
x X, if there is an open subset V such that x V U .
Definition 1.4.3. A subset Y of a topological space (X, ) is said to be
closed if X \ Y is open.
Warning. Most of the subsets of a topological space are neither open nor
closed.
Definition 1.4.4. If A X is a subset of a topological space (X, ), then a
point p X is an interior point of A if A is a neighborhood of p. The point
p is an exterior point of A if X \ A is a neighborhood of p. When all open
sets containing p intersect both A and X \ A, p is called a boundary point
of A. The sets of interior, exterior and boundary points of A are denoted
by int A, ext A and A respectively. The closure of the set A is the union
A = A A.
Definition 1.4.5. A subset A X a topological space (X, ) is said to be
dense in X if its closure A is equal to the whole space X. The subset A is
called nowhere dense if its closure has no interior points at all.
Examples.
Let X be an arbitrary set. The discrete topology on X is the maximal
topology on X, in which every subset is open.
The anti-discrete topology on X is the minimal topology on X, in
which only the empty set and X are open.
Metric Topology
The metric topology of a metric space (X, d) is introduced as follows.
Definition 1.4.6. The open ball in X with center x X and radius > 0 (or
an -ball centered at x) is the set B (x) = {y X | d(x, y) < }. Similarly,
the closed ball in X with center x X and radius > 0 (or a closed -ball
centered at x) is defined as the set B (x) = {y X | d(x, y) }.
Definition 1.4.7. A subset U of a metric space X is called open if for each
x U there is a positive such that the ball B (x) is contained in U .
1.4. Topology 51
U e
x
Figure 1.5: Open subsets of a metric space.
Proposition 1.4.8. The family of open subsets is a topology on X.
Proof. Obviously, the empty set and X are open. If U and V are open
subsets, and x is a common point of them, then there exist positive numbers
1 , 2 such that B1 (x) U and B2 (x) V . Let be the smaller of 1
and 2 . Then B (x) U V , showing that the intersection U V is open.
Finally, let {Ui : i I} be an arbitrary family of open sets, and x be an
element of their union. S S j I and a positive
Then we can find an index
such that B (x) Uj iI Ui , thus the union iI Ui is open.
This topology is referred to as the metric topology of X. A topology is called

metrizable, if it can be derived from a metric. Characterization of metrizable
topologies is a nontrivial problem of general topology, answered partially by
Urysohns metrization theorem and completely by the theorems of Nagata,
Smirnov and Bing (see [9], Chapter 6).
Examples.
Every affine space (over R) can be turned into a Euclidean space by

introducing an inner product on the linear space of its vectors. It can
be checked that though the metric of the space depends on the choice
of the inner product, the topology does not. Thus, every affine space
has a standard topology metrizable by any of the Euclidean structures
on it.
Introducing a positive definite inner product on a linear space V , the

angle between k-dimensional linear subspaces is a metric on the Grass-
mann manifold Grk (V ). As in the previous example, the metric itself
depends on the Euclidean structure on V , but the metric topology in-
duced by it does not. This topology is the standard topology on Grass-
mann manifolds.
52 1. Preliminaries
Subspace Topology
The topology of a topological space defines a topology on every of its subsets
by the following construction.
Figure 1.6: Open subsets in the subspace topology of Y X.
Proposition 1.4.9. Let Y be a subset of a topological space (X, ). Then

the family |Y = {U Y |U } is a topology on Y .
Proof. The proof of the proposition is straightforward from the following

identities:
(i) Y = , X Y =Y;
(ii) (U Y ) (V Y ) = (U V ) Y ;
S S
(iii) iI (Ui Y ) = ( iI Ui ) Y .
Definition 1.4.10. The topology |Y is called the subspace topology or the

topology induced on Y by .
Examples.
As a special case of the subspace topology, all subsets of an affine space,

in particular, all spheres of a Euclidean space inherit a subspace topol-
ogy from the ambient space.
At the end of Section 1.3.2 the affine Grassmann manifolds and the
set of k-spheres of a Euclidean linear space were embedded into a pro-
jective space, which has a standard topology. Through these embed-
dings, affine Grassmann manifolds AGrk (V ) and the sets SPHk (V ) of
k-spheres also inherit a subspace topology.
1.4. Topology 53
Factor Topology
The topology of the Grassmann manifolds of all k-dimensional affine or linear
subspaces of an n-dimensional linear space can also be described with the help
of the factor space topology construction.
Assume that a topological space (X, ) is divided into a disjoint union of its
subsets. Such a subdivision can always be thought of as a splitting of X
into the equivalence classes of an equivalence relation on X. Denoting by
Y = X/ the set of equivalence classes we have a natural mapping : X Y
assigning to an element x X its equivalence class [x] Y .
Proposition 1.4.11. The set
0 = {U Y | 1 (U ) }
is a topology on Y .
Proof. The proof follows from the following set theoretical identities:
(i) 1 () = , 1 (Y ) = X;
(ii) 1 (U V ) = 1 (U ) 1 (V );
(iii) 1 ( iI Ui ) = iI 1 (Ui ).
S S
Definition 1.4.12. The family 0 is called the factor space topology on Y .
As an example, consider the set

V (n, k) = {(x1 , . . . , xk ) (Rn )k | x1 , . . . , xk are linearly independent}.
V (n, k) is an open subset in (Rn )k = Rnk hence it inherits a subspace topol-
ogy from the standard topology of Rnk . If we say that two elements of
V (n, k) are equivalent if they span the same k-dimensional linear subspace
of Rn , then the set V (n, k)/ of equivalence classes is essentially the same as
the set Grk (Rn ) of all k-dimensional linear subspaces of Rn . This set becomes
a topological space with the factor space topology.
We can define a factor topology on the affine Grassmann manifolds similarly.
We set
V (n, k) = {(x0 , . . . , xk ) (Rn )k+1 | x0 , . . . , xk are not in a (k 1)-plane},
furnish V (n, k) with the subspace topology inherited from Rn(k+1) and de-
fine an equivalence relation on V (n, k) by (x0 , . . . , xk ) (y0 , . . . , yk )
x0 , . . . , xk and y0 , . . . , yk span the same k-plane.
V (n, k)/ is essentially the set of AGrk (Rn ) affine k-dimensional subspaces
Rn and it is equipped with the factor space topology.
Exercise 1.4.13. Show that the factor space topologies on Grk (Rn ) and
AGrk (Rn ) coincide with the metric topologies introduced above.
54 1. Preliminaries
Product Topology
Definition 1.4.14. If (X1 , 1 ), . . . , (Xn , n ) are topological spaces, then the
product topology on the Cartesian product X1 Xn is defined as follows.
A subset U X1 Xn is open with respect to the product topology
if and only if for each point (x1 , . . . , xn ) U , we can find open subsets
U1 1 , . . . , Un n such that (x1 , . . . , xn ) U1 Un U .
Exercise 1.4.15. Show that the standard topology on Rn coincides with the
product topology on n copies of R.
Convergence, Continuity
Definition 1.4.16. We say that a sequence x1 , x2 , . . . of points of a topo-
logical space (X, ) converges to a point x X, if for any neighborhood U
of x there is a natural number N such that for n > N , xn U .
Figure 1.7: Continuity at a point.
Definition 1.4.17. Let (X, ) and (Y, 0 ) be topological spaces. A mapping

f : X Y is said to be continuous at the point x (with respect to the given
topologies) if for each neighborhood U of f (x), we can find an open set
V such that x V and f (V ) U . The mapping f is continuous if
it is continuous at each point or, equivalently, if for each U 0 , we have
f 1 (U ) .
Exercise 1.4.18. Show that for Rn the above definition is equivalent to the
definition of continuity at a point: f : Rn Rm is continuous at
x Rn with respect to the standard topologies if and only if for any > 0
one can find a positive such that |x x0 | < implies |f (x) f (x0 )| < .
The map f is a homeomorphism, if it is a bijection such that both f and f 1
are continuous.
1.4. Topology 55
We say that two topological spaces are homeomorphic or have the same topo-
logical type, if there is a homeomorphism between them.
Homeomorphic topological spaces are considered to be the same from the
viewpoint of topology.
Figure 1.8: Intuitive description of a homeomorphism.
Intuitively, two spaces are homeomorphic if a rubber model of one of them

can be deformed into that of the other. We are allowed to stretch and shrink
the model but not allowed to cut the model or glue pieces together. More
exactly, we may cut the model somewhere only if later on we glue together the
parts we get in the same way as they were joined. Of course, this description
of homeomorphism is applicable only for nice spaces such as surfaces, curves
etc. and by no means substitutes the precise definition.
For example, the circle, the perimeter of a square, and the trefoil knot are
homeomorphic, so are a solid disc and a solid square, however a circle is not
homeomorphic to a solid disc. In most cases it is easy to show that two
homeomorphic spaces are indeed homeomorphic: we only have to present
a homeomorphism. However, to show that two topological spaces are not
homeomorphic, we have to find a topological property, which is possessed
only by one of the spaces. For example, the fact that Rn is not homeomor-
phic to Rm if n 6= m, is a non-trivial theorem of topology (the dimension
invariance theorem), the proof of which uses techniques of algebraic topology
or dimension theory.
Sometimes it is difficult to prove continuity of a map using the definition of
continuity directly. However, there are some theorems, that make such proofs
easier. For example, a mapping h : X Rn from a topological space (X, )
into Rn is continuous if and only if all of its coordinate functions hi : X R
are continuous. If f, g : X R are real valued continuous functions on X
then their real valued linear combinations f + g and their product f g are
also continuous. The quotient f /g is also continuous everywhere, where g
is not 0. Thus, for example, if p X Rm and f : X Rn , then if p
has a neighborhood U such that the coordinates of f (q) can be expressed
56 1. Preliminaries
not
homeomorphic homeomorphic homeomorphic
Figure 1.9: Examples of homeomorphic and non-homeomorphic topological

spaces.
as rational functions (quotients of polynomials) of the coordinates of q U ,

then f is surely continuous at p.
Exercise 1.4.19. Show that Rn and the open balls in Rn (with the subspace
topology) are homeomorphic.
Exercise 1.4.20. Show that the punctured sphere Sn \ {p}, where
Sn = {x Rn+1 | |x| = 1} and p Sn ,
is homeomorphic to the plane Rn .

Exercise 1.4.21. Show that the inverting map GL(V ) GL(V ), M 7
M 1 is a homeomorphisms.
Exercise 1.4.22. Show that Grk (Rn ) is homeomorphic to Grnk (Rn ).
Exercise 1.4.23. A k-sphere in a Euclidean linear space V is uniquely deter-
mined by the affine subspace spanned by it, its center, and its radius. Thus,
there is a bijection between SPHk (V ) and the set
g k (V ) = {(, O, r) AGrk+1 (V ) V R | O , r > 0}.

SPH
Show that SPHk (V ) is homeomorphic to SPHg k (V ) endowed with the sub-

space topology inherited from the product topology of AGrk+1 (V ) V R.
1.4.1 Separation and Countability Axioms

Sometimes to prove substantial theorems on a topological space, we have to
assume that it satisfies some simple properties that do not follow from the
1.4. Topology 57
axioms of a topological space. Some of the frequently used properties can

be formulated as axioms that can be added to the system of axioms of a
topological space or as definitions of certain types of topological spaces that
satisfy the additional axioms. We list here some of these axioms that will be
used later. The axioms can be grouped according to their nature. First we
start with the group of separation axioms.
Definition 1.4.24. A topological space (X, ) is said to be T0 , if for any two
distinct points p 6= q X, there is an open subset U , which contains
exactly one of the points.
Discrete topology on a set of at least two points is not T0 .
Definition 1.4.25. A topological space (X, ) is said to be T1 , if for any two
distinct points p 6= q X, there is an open subset U , which contains p
but does not contain q.
If X = {p, q} is a two point set, and the topology on X consists of the sets
X, {q}, and , then (X, ) is T0 but not T1 . T1 -spaces are T0 .
Definition 1.4.26. A topological space (X, ) is said to be T2 , or Hausdorff,
if for any two distinct points p 6= q X, there exist open subsets U, V
such that p U , q V and U V = .
Every T2 -space is T1 . An example of a topological space which is T1 but not
T2 is an infinite set X equipped with the cofinite topology . By definition,
tau contains exactly the emptyset and those subsets of X, the complements
of which are finite.
Definition 1.4.27. A topological space (X, ) is regular if for any point
p X and any closed set F X \ p, there are open subsets U, V such
that p U , F V and U V = . A regular T0 -space is called a T3 -space.

A set with at least two points is regular but not T0 . On the other hand,
every T3 -space is T2 . A simple example of a Hausdorff space (X, ) which is
not regular was constructed by J.R. Munkres. In the example of Munkres,
the set X is the set of real numbers R. Elements of are sets of the form
U \ L, where U is an open subset of R with respect to the standard (metric)
topology, L is a subset of the set K = {1/n | 0 < n N}. (X, ) is not
regular because the point 0 cannot be separated from the closed subset K
with disjoint open subsets.
Definition 1.4.28. A topological space (X, ) is called normal if for any
two disjoint closed set F1 and F2 , one can find disjoint open subsets U and
V such that U F1 and V F2 . A topological space is a T4 -space if it is
normal and T1 .
58 1. Preliminaries
The two point set X = {p, q} with topology = {X, {q}, } is normal and
T0 , but not T1 , hence not T4 . Every T4 space is T3 . The Sorgenfrey plane is an
example of a T3 space which is not normal. The Sorgenfrey plane is obtained
as the product space (R,S ) (R, ), where the (nonstandard) topology on
R consists of all unions iI [ai , bi ) of left closed right open intervals. (See
[9] Example 3. in 31, Ch. 4. for details.)
Countability axioms are of different character. They typically require the
existence of a countable family of subsets with some properties. Recall that
two sets have the same cardinality if and only if there is a bijection between
them. There is an ordering on the family of cardinalities. The cardinality
of a set A is less then or equal to the cardinality of the set B if there is an
injective map from A into B. A set is called countable if it is either finite
or has the same cardinality as the set of natural numbers. Equivalently, a
set is countable if there is a sequence p1 , p2 , . . . listing all elements of the
set. Every infinite set contains a countable infinite set, therefore a countable
infinite sets have the smallest cardinality among all infinite sets.
Definition 1.4.29. A topological space (X, ) is first-countable if for any
point p X, there is a countable family of open sets U1 , U2 , . . . such that
for any U containing p, there is an element Ui of the family for which
p Ui U .
Metric topologies are always first-countable. Choosing for Ui the open ball
of radius 1/i centered at p, we obtain a countable family of neighborhoods
of p satisfying the conditions.
Definition 1.4.30. A topological space (X, ) is second-countable if there is
a countable family of open sets U1 , U2 , . . . such that every openSset U is
the union of those elements Ui that are covered by U , i.e., U = Ui U Ui .
Every second-countable space is first-countable. The discrete topology on R
is metrizable but not second-countable.
Definition 1.4.31. A topological space (X, ) is said to be separable if con-
tains a countable dense subset.
Proposition 1.4.32. Every first-countable separable space is second-countable.
1.4.2 Compactness
Definition 1.4.33. A topological
S space (X, ) is said to be compact if from
any open covering X = iI Ui , Ui of X, we can choose a finite subcov-
ering X = Ui1 Uik , where i1 , . . . , ik I.
A related notion is sequential compactness.
1.4. Topology 59
Definition 1.4.34. A topological space (X, ) is said to be sequentially

compact if every sequence x1 , x2 , . . . in X has a convergent subsequence
xi1 , xi2 , . . . , (1 i1 < i2 < ).
We emphasize that the limit of the subsequence must be in X even if X
is contained in a larger space as a subspace. Compactness and sequential
compactness are independent properties, none of them implies the other in
general. However, for metric spaces they are equivalent.
Proposition 1.4.35.
A first-countable compact space is sequentially compact.
A second-countable sequentially compact space is compact.
A sequentially compact metric space is separable, hence second-countable
and compact.
Compact metric spaces can also be characterized as follows.
Definition 1.4.36. A sequence of points x1 , x2 , . . . in a metric space (X, d)
is a Cauchy sequence if for any > 0, there exists a number N such that we
have d(xm , xn ) < whenever i > N and j > N . A metric space (X, d) is
said to be complete if all of its Cauchy sequences are convergent to a point
of X.
Definition 1.4.37. A metric space (X, d) is totally bounded if and only if
for any > 0, X can be covered by a finite collection of open balls of radius
.
Proposition 1.4.38. A metric space is compact if and only if it is complete
and totally bounded.
Corollary 1.4.39 (HeineBorel theorem). A subset K of Rn with the sub-
space topology is compact if and only K is bounded and closed.
Proposition 1.4.40. The image f (X) of a compact space X under a con-
tinuous map f : X Y is a compact subspace of Y .
In particular, if f : X R is a real function on a compact space, then f (X)
is bounded and closed in R, therefore, it has a maximal and minimal element.
This gives us the extreme value theorem due to Weierstrass.
Proposition 1.4.41. A continuous function f : X R on a compact space
(X, ) is bounded and attains its maximum and minimum values.
Definition 1.4.42. A function f : X R on a metric space (X, d) is uni-
formly continuous, if for any > 0, one can find a positive such that
x, y X, d(x, y) < implies |f (x) f (y)| < .
60 1. Preliminaries
It is clear that every uniformly continuous function is continuous, but the

converse is not true. For example, the function f : R R, f (x) = x2 is
continuous, but not uniformly continuous, since no matter how small |x y|
is, |x2 y 2 | = |x y||x + y| can be arbitrarily large. However, if the domain
of the function is compact, then continuity implies uniform continuity.
Proposition 1.4.43. A continuous function f : X R defined on a compact
metric space (X, d) is uniformly continuous.
Proof. Suppose to the contrary, that there is an > 0 for which we can
not find a suitable . Then there exists a sequence of pairs of real numbers
xn , yn such that xn , yn X, d(xn , yn ) < 1/n, but |f (xn ) f (yn )| > . By
compactness of X, we can select a convergent subsequence xin x of the
sequence (xn ). Condition d(xn , yn ) < 1/n ensures that yin x as well,
and so, by the continuity of f at x we have |f (xin ) f (yin )| 0. But this
contradicts the condition |f (xn ) f (yn )| > for all n.
1.4.3 Fundamental Group and Covering Spaces

Definition 1.4.44. A pointed topological space is a pair (X, x0 ) consisting
of a topological space X and a distinguished base point x0 X.
Pointed topological spaces form a category. A morphism f : (X, x0 ) (Y, y0 )
of the category is a continuous map f : X Y such that f (x0 ) = y0 .
Definition 1.4.45. A loop in the topological space X with base point x0 X
is a continuous map : [0, 1] X such that (0) = (1) = x0 .
Definition 1.4.46. The concatenation of the loops and in X with a
common base point x0 is the loop defined by equipped with the group
multiplication defined by
(
(2t), if 0 t 21 ,
( )(t) =
(2t 1), if 12 t 1.
Definition 1.4.47. The loops and in X with a common base point x0
X are said to be homotopic if there is a continuous map H : [0, 1] [0, 1] X
such that H(0, t) = (t), H(1, t) = (t), and H(t, 0) = H(t, 1) = x0 for all
t [0, 1].
Being homotopic is an equivalence relation on the set of loops with a given
base point. The equivalence class of a loop is also called its homotopy class
and is denoted by [].
Definition 1.4.48. The fundamental group 1 (X, x0 ) of a pointed topological
space (X, x0 ) is the set of all homotopy classes of loops in X with base point
x0 together with the multiplication induced by concatenation as follows
[] [] = [ ].
1.4. Topology 61
Proposition 1.4.49. The product of the elements of the fundamental group

is properly defined, and multiplication satisfies the group axioms.
The fundamental group construction gives a functor from the category of
pointed topological spaces to the category of groups. Indeed, every continuous
map f : (X, x0 ) (Y, y0 ) induces a group homomorphism f : 1 (X, x0 )
1 (Y, y0 ) by the formula f ([]) = [f ].
Definition 1.4.50. A topological space X is connected if X cannot be
decomposed into the disjoint union of two nonempty open subsets. X is
path-connected if for any two points x, y X, there is a continuous map
: [0, 1] X such that (0) = x and (1) = y. A topological space is locally
path-connected if for any point x X and any neighborhood U of x, there is
a path-connected neighborhood V of x which is contained in U .
Every path-connected space is connected. The union of the y-axis and the
graph of the function f : (0, 1) R, f (x) = sin(1/x) is connected but not
path-connected with its subspace topology inherited from R2 . However, a
connected and locally path-connected topological space is path-connected.
The set {(x, y) R | y = 0 or x Q} is path-connected but not locally
path-connected.
The isomorphism class of the fundamental group of a path-connected space
does not depend on the choice of the base point.
Definition 1.4.51. A topological space is said to be simply connected if it is
path-connected and its fundamental group is trivial with respect to any base
point. A topological space X is semilocally simply-connected if each point
x X has a neighborhood for which the the embedding : U X induces a
trivial homomorphism : 1 (U, x) 1 (X, x).
The latter condition means geometrically that any loop in U based at x can
be contracted to a point in X.
The fundamental group has many applications in topology. The main reason,
why we introduced it here is its role played in the classification of covering
spaces of a given space.
Definition 1.4.52. A continuous map p : X X is called a covering map
if each point x X has an open neighborhoodS U X such that p1 (U ) can
1
be decomposed as a disjoint union p (U ) = iI Ui of open subsets Ui of
X, such that the restriction p|Ui : Ui U is a homeomorphism for all i I.
Neighborhoods U with this property are called evenly-covered, the sets Ui are
called the sheets/slices/layers over U .
The cardinality of p1 (x) is locally constant on X. Thus, if X is connected,
then it is constant. Whenever the cardinality of p1 (x) is constant on X, it
62 1. Preliminaries
is called the degree of the covering. A k-fold covering is a covering of degree

k.
Covering maps have the following lifting property for cubes.
Proposition 1.4.53. Let p : (X, x0 ) (X, x0 ) be a covering map between
two pointed spaces, : ([0, 1]k , {0}) (X, x0 ) be a continuous map from a
pointed k-dimensional cube to (X, x0 ). Then there is a unique continuous
map : ([0, 1]k , {0}) (X, x0 ) such that p = .
The map is called the lift of .
Corollary 1.4.54. If p : (X, x0 ) (X, x0 ) be a covering map between two
pointed spaces,then the induced map p : 1 (X, x0 ) 1 (X, x0 ) is injective.
Thus, we can assign to each pointed covering map the subgroup im p of
the fundamental group of 1 (X, x0 ). It turns out that for nice spaces this
subgroup determines the covering uniquely up to a natural notion of isomor-
phism.
Definition 1.4.55. Two covering maps p1 : (X1 , x1 ) (X, x0 ) and
p2 : (X2 , x2 ) (X, x0 ) are isomorphic if there is a base point preserving
homeomorphism h : (X1 , x1 ) (X2 , x2 ) such that p2 h = p1 . We define
isomorphism of (non-pointed) coverings in a similar way, ignoring the base
points.
Theorem 1.4.56. Let (X, x0 ) be a path-connected, locally path-connected,
and semilocally simply-connected topological space. Then there is a bijection
between the set of base point preserving isomorphism classes of path-connected
covering spaces p : (X, x0 ) (X, x0 ) and the set of subgroups of 1 (X, x0 ),
obtained by assigning the subgroup p (1 (X, x0 )) 1 (X, x0 ) to the covering
space (X, x0 ).
If base points are ignored, this correspondence gives a bijection between iso-
morphism classes of path-connected covering spaces p : X X and conjugacy
classes of subgroups of 1 (X, x0 ).
In terms of this correspondence, the degree of the covering is the index
|1 (X, x0 ) : p (1 (X, x0 ))| of the subgroup p (1 (X, x0 )) in 1 (X, x0 ). Spaces
satisfying the conditions of the theorem have a unique simply connected cov-
ering space up to isomorphism. It corresponds to the trivial subgroup of the
fundamental group of X. This covering space is the universal covering space
of X.
1.5 Multivariable Calculus

Definition 1.5.1. A map F : U Rn defined on a neighborhood of U Rm
x0 Rm is differentiable at x0 and its derivative at x0 is the linear map
1.5. Multivariable Calculus 63
A : Rm Rn if
kF (x) F (x0 ) A(x x0 )k
lim = 0.
xx0 kx x0 k
The derivative of F at x0 and also its n m matrix is denoted by F 0 (x0 ).

The derivative (matrix) F 0 (x0 ) is also called the Jacobian (matrix) of F at
x0 .
Exercise 1.5.2. Prove that the derivative of a map at a point is uniquely
determined by the condition defining it.
Exercise 1.5.3. Show that if F is differentiable at a point x0 , then it is also
continuous at x0 .
The derivative map is the linear part of the affine transformation x 7 A(x)+
(F (x0 ) A(x0 )) giving the best approximation of F around x0 .
Proposition 1.5.4 (Chain Rule). If U Rm is a neighborhood of x0 U ,
and the function F : U Rn is differentiable at x0 , furthermore, the function
G : V Rp is defined on a neighborhood of F (x0 ) and it is also differentiable
at F (x0 ), then the composition G F : U Rp is differentiable at x0 and
(G F )0 (x0 ) = G0 (F (x0 )) F 0 (x0 ).
The sign on the right-hand side denotes composition of linear maps or

matrix multiplication depending on whether the derivative maps are thought
of as linear maps or the matrices of these linear maps.
Definition 1.5.5. Let F : U Rn be differentiable at x0 U , where U
Rm is a neighborhood of x0 . The directional derivative of F along a tangent
vector (x0 , v) Tx0 Rm is the vector
d
v F (x0 ) = F (x0 + tv)|t=0 Rn .
dt
Recall that the tangent vector (x0 , v) consists of the vector v, and a base point
x0 . There are several other notations used to denote directional derivatives,
e.g., v F |x0 , v F (x)|x=x0 , (v |x0 )(F ), sometimes is replaced by the letter
D. We shall use usually the sign , but otherwise any of these notations,
always the one which seems to be most convenient typographically.
Remark also that though we defined v F (x0 ) as a vector, there is a natural
choice of a base point for this vector, namely F (x0 ), so it would have been
also a possibility to define v F (x0 ) as a tangent vector of Rn at F (x0 ).
Directional derivatives can be computed easily from the chain rule
v F (x0 ) = F 0 (x0 )(v).

64 1. Preliminaries
Definition 1.5.6. Let F : U Rn be as above. The partial derivative

i F (x0 ) of F with respect to its ith variable at x0 is its directional derivative
i F (x0 ) = ei F (x0 )
along (x0 , ei ), where ei is the ith vector of the standard basis of Rm .

Partial derivatives are also denoted in many different ways. In some cases,
the variables of F are denoted by some fixed symbols like (x1 , . . . , xm ). Then
F
we may also use the notations i F (x0 ) = x i
F (x0 ) = x i
(x0 ) = xi F (x0 ).
Some of these notations can also be used when the variables do not carry nu-
merical indices just denoted by different letters, like (u, v, . . . ). For example,
in the latter case, 1 F (u0 , v0 , . . . ) can also be denoted by u F (u0 , v0 , . . . ),
F
u F (u0 , v0 , . . . ), u (u0 , v0 , . . . ).
Warning. When the variables are denoted by fixed symbols, the variable
symbols should not denote anything else, otherwise one can run into am-
biguous expressions leading to confusion. For example, if F (u, v) = uv, then
u F (u, v) = v. However, the expression u F (u2 , v 2 ) = u u2 v 2 can mean
both v 2 or 2uv 2 . The confusion is caused by the circumstance that instead
of introducing two new variables, say (x, y) and making the substitution
(u, v) = (x2 , y 2 ), we denoted the new variables also by u and v, and made
the strange substitution (u, v) = (u2 , v 2 ). Thus, the different expressions
u F (x2 , y 2 ) = y 2 and x F (x2 , y 2 ) = 2xy 2 collapsed.
We can express the matrix of the derivative of F with the help of its partial
derivatives. Indeed, the columns of the matrix of F 0 (x0 ) are the images of the
standard basis vectors that is the partial derivatives 1 F (x0 ), . . . , m F (x0 ).
To write the matrix more explicitly, suppose the the coordinate functions of
F are (F 1 , . . . , F n ). The F i s are real valued functions on U related to F by
F (x) = (F 1 (x), . . . , F n (x)). Then
1 F 1 (x0 ) . . . m F 1 (x0 )

F 0 (x0 ) =
.. ..
. .
1 F n (x0 ) . . . m F n (x0 )
Definition 1.5.7. If the real valued function F : U R defined in a neigh-
borhood U of x0 Rm is differentiable at x0 , then its derivative matrix is a
row vector
grad F (x0 ) = (1 F (x0 ), . . . , m F (x0 )),
which we call the gradient vector of F at x0 .
According to this definition, for a vector valued function F , the rows of the
derivative matrix of F are the gradient vectors of the coordinates functions
of F .
1.5. Multivariable Calculus 65
Using the explicit form of the derivatives matrices and the multiplication
rule for matrices, we can reformulate the chain rule equating the matrix
coefficients.
(F i G)
Proposition 1.5.8 (Chain Rule | a reformulation). If U Rm is a neighbor-
hood of x0 U , and the function F = (F 1 , . . . , F n ) : U Rn is differentiable
at x0 , furthermore, the function G = (G1 , . . . , Gp ) : V Rp is defined on a
neighborhood of F (x0 ) and it is also differentiable at F (x0 ), then
n
X
i (G F )j (x0 ) = i (Gj F )(x0 ) = k Gj (F (x0 )) i F k (x0 ).
k=1
The existence of all directional derivatives of F at x0 is necessary for the

differentiability of F at x0 but not sufficient. However we have the following
theorem.
Theorem 1.5.9. If F : U Rn is defined on an open subset U of Rm and
all the partial derivatives 1 F (x), . . . , m F (x) exist at each point x U , and
all depend continuously on x, then F is differentiable at each point of U and
the derivative of F is a continuous function of x U .
Definition 1.5.10. A function F : U Rn defined on an open subset U of
Rm is said to be of k times continuously differentiable of a function of class
C k , if all the k-th order partial derivatives i1 . . . ik F (x) exist and depend
continuously on x U .
Functions which are of class C k for all k are said to be function of class C
and called infinitely many times differentiable or smooth functions.
There is also a recursive definition of this notion. A function F is of class C 0
if and only if F is continuous. For k 1, F is of class C k if and only if it
is differentiable at each point of U and the map F 0 : U Rnm is of class
C k1 .
There is important theorem which is referred to by many different names in
the literature: Youngs Theorem, Clairauts Theorem, Schwarzs Theo-
rem, Symmetry of Second Derivatives, Equality of Mixed Partials.
Theorem 1.5.11 (Youngs Theorem). If F : U Rn is defined on the open
set U Rm and is of class C 2 , then i j F = j i F for any 1 i, j m.
Another fundamental theorem of multivariable calculus is the inverse function
theorem
Theorem 1.5.12 (Inverse Function Theorem). Suppose that U is an open
subset of Rn and F : U Rn is a function of class C 1 with invertible deriva-
tive F 0 (x0 ) at x0 . Then x0 has an open neighborhood V U such that
66 1. Preliminaries
W = F (V ) is an open neighborhood of F (x0 ), F |V : V W is a bijection,

and its inverse (F |V )1 : W V is also of class C 1 .
If, in addition, F is of class C k , then (F |V )1 : W V is also of class C k .
The derivative of the inverse map G = (F |V )1 can be computed by differ-
entiating the identity
(F G)(x) x
by the chain rule. Rearranging we obtain
G0 (x) = (F 0 (G(x)))1 .
Sometimes we want to speak about the differentiability of maps the domains
of which is not an open subset (e.g. a closed interval). For this purpose, we
say that a map F : A Rn defined on an arbitrary set A Rm is of class C k
or k times differentiable if there exists an open set U Rm and a mapping
F : U Rn of class C k such that A U and F = F |A .
We can define the partial derivatives i1 . . . ir F (x) of order r k of F
as i1 . . . ir F (x), but it should be kept in mind that these derivatives are
defined properly only on the closure of the interior of A. At other points, the
derivatives usually depend on the choice of F .
Bump Functions
Proposition 1.5.13. If K Rn is a compact subset of Rn and > 0, then
there exists a smooth function h : Rn [0, 1] such that h(x) is equal to 1 if
x K and h(x) = 0 if d(x, K) .
Proof. Define the function h1 : R [0, 1] by the formula
( 1
h1 (t) = e t2 1 , if t (1, 1),

0, if t
/ (1, 1).
Exercise 1.5.14. Prove that h1 is a smooth function on R.

We can extend h1 to Rn in a rotationally symmetric way by setting hn : Rn
R, hn (x) = h1 (kxk).
Exercise 1.5.15. Prove that hn is a smooth function on Rn . (Caution: The
norm function x kxk is not smooth at the origin.)
n+1
R
The graph of hn is a bell shaped hypersurface in R . Let cn = Rn hn (x)dx
denote the volume of the domain under the bell.
Denote by the indicator function of the closed neighborhood of radius /2
of K, that is, : R R is the function defined by

1, if d(x, K) /2,
(x) =
0, otherwise.
1.6. Measure and Integration 67
Figure 1.10: The graphs of the functions h1 and h2 .
Then the function
2n

xy
Z
h(x) = (y)hn dy
cn n Rn /2
has the desired properties.
Exercise 1.5.16. Show that h is indeed smooth, h|K 1, and that h is

vanishing outside the -neighborhood of K.
1.6 Measure and Integration

Extending the classical geometrical treatment of the notion of area and vol-
ume leads to the notion of Jordan content. Roughly speaking, assuming that
we already know how to define the volume of a compact polyhedron, that is a
finite union of n-dimensional simplices by dissecting it into non-overlapping
simplices and adding the volumes of the simplices, a bounded subset A of Rn
is said to have Jordan content if the supremum of the volumes of polytopes
contained in A is equal to the infimum of the volumes of compact polytopes
containing A. In this case, the common value of the supremum and the
infimum is the Jordan content Jn (A) of A.
Instead of Jordan content, we shall rather use Lebesgue measure as the vol-
ume of sets. The reason for this is that whenever a set has Jordan content,
it is also Lebesgue measurable and its Lebesgue measure equals its Jordan
68 1. Preliminaries
content. However, having Lebesgue measure is a much less restrictive prop-

erty then the property of having Jordan content. Lebesgue measure is also
more convenient to work with due to its -additivity.
Measure Spaces
Definition 1.6.1. A -algebra of subsets of a set X is a family of subsets
of X satisfying the following axioms.
X .
If A , then X \ A .
S
If A1 , A2 , . . . is a sequence of elements of , then i=1 Ai , i.e.,
is closed under countable unions.
Definition 1.6.2. The Borel algebra of a topological space is the smallest

-algebra containing its open subsets. Borel sets are elements of the Borel
algebra.
Exercise 1.6.3. Show that closed subsets are Borel sets.
Exercise 1.6.4. Show that if a topological space is Hausdorff, which means

that for any two points p 6= q there exist disjoint open set U and V such
that p U and q V , then compact subsets of the space are closed, hence
Borel sets.
Definition 1.6.5. If is a -algebra of subsets of X, then a function :

[0, +] is a measure if it satisfies the following axioms.
() = 0.
If S
A1 , A2 , . . . isPa sequence of pairwise disjoint elements of , then

( i=1 Ai ) = i=1 (Ai ). (-additivity.)
The triple (X, , ) is called a measure space. Elements of are the -

measurable sets or shortly measurable sets.
Definition 1.6.6. A measure space (X, , ) is complete if any subset B

which is contained in a measurable set A of measure (A) = 0 is also mea-
surable.
Exercise 1.6.7. Show that given a measure space (X, , ), there is a unique
complete measure space (X, , ) such that A if and only if there exist
-measurable sets A, B such that (A \ A) (A \ A) B and (B) =
0, furthermore, if this is the case then (A) = (A). The measure space
(X, , ) is called the completion of the measure space (X, , ).
Definition 1.6.8. A Borel measure on a topological space is a measure on

its Borel sets. A Borel measure is called outer regular if for any Borel set B,
(B) = inf{(U ) | B U and U is open}.
If compact sets of the space are Borel sets, for example, if the space is Haus-
dorff, then is said to be inner regular if for any Borel set B, we have
(B) = inf{(K) | K B and K is compact}.
The Borel measure is regular if it is both inner and outer regular.
Lebesgue Measure
There are several methods to construct the Lebesgue measure n on Rn .
We can start with the construction of 1 as the completion of the unique
regular Borel measure 1 the value of which on open intervals is given by
1 ((a, b)) = b a. Uniqueness of 1 is clear since any open subset of R is a
countable union of disjoint open intervals, so the 1 measure of open subsets
is uniquely prescribed by -additivity. Then the measure of other Borel sets
is uniquely determined by outer regularity of 1 . The proof of existence
requires more work.
To proceed from 1 to n , we can use the product measure construction.
Theorem 1.6.9. Let (X1 , 1 , 1 ) and (X2 , 2 , 2 ) be two measure spaces.
Denote by 1 2 the smallest -algebras of subsets of X1 X2 containing
the family of subsets {A1 A2 | A1 1 , A2 2 }. Then there exists a
measure 1 2 on 2 such that
1 2 (A1 A2 ) = 1 (A1 ) 2 (A2 )
for all A1 1 and A2 2 . If both measure spaces are -finite, that is, if
they can be presented as a countable union of sets of finite measure, then the
measure 1 2 is unique.
Definition 1.6.10. If 1 and 2 are -finite, then the unique measure 1 2
is called the product measure of 1 and 2 .
Exercise 1.6.11. Prove that if X1 = Rk , X2 = Rl and i is the Borel
algebra of Xi , then 1 2 is the Borel algebra of Rk+l .
Starting from the Borel measure 1 on R, the product measure construction
gives a Borel measure n = 1 n1 on Rn = R Rn1 recursively for all
n. Completion of the Borel measure n is the Lebesgue measure n .
Exercise 1.6.12. Show that k l = k+l .
70 1. Preliminaries
Integral of Measurable Functions

Definition 1.6.13. A function f : X [, ] on a measure space (X, , )
is called measurable if the level sets At = {x | f (x) < t} are in for all
t R.
Exercise 1.6.14. Prove that for a measurable function f on the measure
space (X, , ), the preimage f 1 (B) of any Borel subset of R is also in .
The integral of measurable functions with respect to the measure is intro-

duced in some steps.
Definition 1.6.15. The indicator function A of a subset A if X is the
function A : X R such that A (x) = 1 if and only if x A, other-
wise A (x) = 0. A step function is a finite linear combination of indicator
functions.
Exercise 1.6.16. Prove that a step function is measurable if and only if it
can be written as a linear combination of indicator functions of measurable
sets. If, in addition, f is non-negative, then it is possible to write it as a
linear combination with non-negative coefficients.
Definition
Pk 1.6.17. The integral of a non-negative measurable step function
f = i=1 ai Ai , where ai 0 and Ai is
Z k
X
f d = ai (Ai ).
X i=1
As usual in integration theory, we use the convention 0 = 0.

Definition 1.6.18. The integral of a non-negative measurable function
f : X [0, +] on the measure space (X, , ) is
Z Z
f d = sup sd | 0 s f and s is a measurable step function .
X X
As for the integral of an arbitrary measurable function f decompose f into

a non-negative and a non-positive part by setting
f+ (x) = max{f (x), 0} and f (x) = max{f (x), 0}.
Definition 1.6.19. We say that the Rintegral of theR measurable function f
exists if at least one of the integrals X f+ d and X f d is finite and in
that case we define the integral as
Z Z Z
f d = f+ d f d.
X X X
Sometimes we want to restrict the domain of integration to a measurable set

D. This can be done easily with the help of indicator functions.
Definition 1.6.20. The integral of a measurable function f on the measur-
able set D is defined to be the integral
Z Z
f d = D f d,
D X
provided that the second integral exists.

The integral of a Lebesgue measurable function f on Rn will also be denoted
by Z Z
f (x)dx = f dn .
Rn Rn
We have several important and useful tools for Lebesgue integrals. An im-
portant integration tool is Fubinis theorem.
Theorem 1.6.21 (Fubinis Theorem). If (X1 , 1 , 1 ) and (X2 , 2 , 2 ) are
finite measure spaces and f : X1 X2 [, ] is 1 2 measurable,
then xR1 7 f (x1 , x2 ) is 1 measurable for any fixed x2 X2 , the function
x2 7 X1 f (x1 , x2 )d1 (x1 ) is 2 measurable, and
Z Z Z
f (x1 , x2 )d1 (x1 ) d2 (x2 ) = f d(1 2 )
X2 X1 X1 X2
Z Z
= f (x1 , x2 )d2 (x2 ) d1 (x1 ).
X1 X2
On Rk+l = Rk Rl , Fubinis theorem gives

Z Z Z Z Z
f (x, y)dx dy = f (z)dz = f (x, y)dy dx,
Rl Rk Rk+l Rk Rl
where z = (x, y). However, one has to modify slightly the statement for
Lebesgue measures.
R For example, in the case of the Lebesgue measure, the
partial integral Rk f (x, y)dx may not exist for all x. Nevertheless, the set
of those points, for which the integral is not defined has Lebesgue measure
0. Since Lebesgue integrability and the Lebesgue integral of a function is
not affected by the modification of the function on a set of measure zero,
integration of functions whose values are not defined on a set of measure 0
makes sense.
Theorem 1.6.22 (Integration by Change of Variables). Let U and V be

two open subsets of Rn , h : U V be a diffeomorphism of class C 1 , i.e., a
72 1. Preliminaries
bijection between U and V such that both h and h1 are of class C 1 . Then for
any Lebesgue measurable function f : V [ , ], the Lebesgue integral
of f over V exists if and only if the Lebesgue integral of (f h) | det h0 | over
U exists and if the integrals exist, then they are equal
Z Z
f (v)dv = f (h(u)) | det h0 (u)|du.
V U
Theorem 1.6.23 (Sards Lemma). Let U Rm be an open subset. For a

C 1 -map f : U Rn , define the set f of singular points of f by
f = {x U | rk(f 0 (x)) = dim(im f 0 (x)) < n}.
Then the set f (f ) of singular values has Lebesgue measure 0, i.e., n (f (f ))
=0.
Combining the last two theorems we obtain a generalization of the first one.
Theorem 1.6.24 (Integration by Non-Bijective Change of Variables). Sup-
pose that U and V are open subsets of Rn and h : U V is a map of class
C 1 . For a point v V , denote by #h1 (v) the number of h-preimages of v.
Then for any Lebesgue measurable function f : V [, ] we have
Z Z
f (h(u))| det h0 (u)|du = f (v)#h1 (v)dv,
U V
provided that both integrals exist. The integrals may not exist, but if any of
them exists the other exists as well.
Measure and Integration on the Sphere

One can introduce a measure on the unit sphere Sn1 = {x Rn | kxk = 1}
using the Lebesgue measure on Rn . For a set A Sn1 , define the spherical
cone c(A) over A as the set c(A) = {tu | u Sn1 , t (0, 1]}. The set A
will be -measurable if and only if its cone c(A) is Lebesgue measurable,
and if this happens to be the case, then we define the -measure of A as
(A) = n(c(A)).
The factor n is motivated by the heuristic picture that any set can be split
into very small ones. When A is very small, A is almost flat, and the spherical
cone over it is almost like a pyramid over A of height 1, and in Rn , the volume
of a pyramid is 1/n times the volume of the base times the height.
If f : Sn1 R is a -measurable function, then its integral can also be
expressed as an integral with respect to the Lebesgue measure. Extend the
function f to a function f c on the unit ball B n = c(Sn1 ) by

x
f , if x 6= 0,
f c (x) = kxk
0, if x = 0.

Then Z Z
f d = n f c d. (1.18)
Sn1 Bn
Equation (1.18) can be checked easily for step functions. The general case
can be proved by approximating measurable functions with step functions.
A Lebesgue integral over a ball, can be obtained by integrating the function
over concentric spheres with respect to a suitably scaled measure and then
integrating these spherical integrals with respect to the radius.
n
Theorem 1.6.25 (Integration in Spherical Coordinates). Let f : BR R
n
be a Lebesgue measurable function on the n-dimensional ball BR of radius R
centered at the origin. Then
Z Z R Z
f d = rn1 f (ru)dudr,
n
BR 0 Sn1
where integrals over Sn1 are taken with respect to the measure , all the
other integrals are computed with respect to the Lebesgue measure.
Proof. Using equation (1.18), the right-hand side can be transferred to an
integral on B n [1, R]
Z R Z Z RZ
x
rn1 f (ru)dudr = n rn1 f r dxdr.
0 Sn1 0 B n \{0} kxk
Compute the last integral by the substitution
y
n
h : (BR \ {0}) (0, 1] (B n \ {0}) (0, R], h(y, s) = s , kyk .
kyk
The integrand composed with h is
y/kyk
kykn1 f kyk
y/kyk = kyk
n1
f (y).
Thinking of y as a column vector, the determinant of the derivative matrix

of h is

sIn syyT y
0

det(h (y, s)) = det
kyk kyk3 kyk

yT /kyk 0

sIn y

= det
kyk kyk ,

yT /kyk 0
74 1. Preliminaries
where In is the n n unit matrix. The second equality holds because sub-
tracting suitable multiples of the last column from the preceding ones we can
eliminate the term syyT /kyk3 from the upper left corner.
Exercise 1.6.26. Show that for any column vectors a, b Rn , we have

In a
= ha, bi.
det

T
b 0
Applying the result of the exercise,

sIn y

det
kyk kyk

yT /kyk 0

n+1 y n1
s In s
= det s = .

kyk kyk
yT /s 0
Therefore, the substitution gives
Z RZ Z 1Z n1
n1 x n1 s
n r f r dxdr = n kyk f (y) dyds
0 B n \{0} kxk 0 BR n kyk
Z 1 Z Z
=n sn1 ds f (y)dy = f (y)dy,
0 n
BR n
BR
so the theorem is proved.

As a corollary we obtain other ways to express spherical integral with Lebesgue
integral.
Corollary 1.6.27. Let f : Sn1 R be a -measurable function, g : [0, 1]
R1
R be a continuous function for which mg = 0 g(r)rn1 dr 6= 0. Then
Z Z
1 x
f d = g(kxk)f dx.
Sn1 mg B n kxk
For example, choosing g 1 we obtain (1.18), g(r) = r yields mg = 1/(n + 1)
and Z Z
x
f d = (n + 1) kxkf dx. (1.19)
Sn1 Bn kxk
1.7. Ordinary Differential Equations 75
1.7 Ordinary Differential Equations

Definition 1.7.1. Let U Rn be an open subset of Rn . A vector field over
U is a map F : U T Rn such that F(p) Tp Rn for all p U .
A tangent vector based at p is a pair (p, v), where v Rn , therefore, F(p)
must have the form (p, F (p)). As the function F : U Rn obtained from
F by ignoring the base points determines F uniquely, we can define a vector
field uniquely by a mapping from U to Rn .
Definition 1.7.2. Let U Rn be an open subset of Rn , F : U Rn be a
vector field on U . A (first order autonomous vector valued) ordinary differ-
ential equation with right-hand side F is the problem of finding differentiable
parameterized curves, i.e., differentiable maps : I U satisfying the equa-
tion
0 (t) = F ((t)) for all t I,
where I is a finite or infinite interval in R. Solutions of the problem are called
the integral curves of the differential equation or that of the vector field F .
F is also called the right-hand side of the differential equation.
More generally a kth order vector valued ordinary differential equation is
given by a map F : U (Rn )k1 Rn and is posing the problem to find
k-times differentiable curves : I U satisfying
(k) (t) = F ((t), 0 (t), . . . , (k1) (t)) for all t I.
Despite its more general form, every kth order differential equation is equiva-
lent to a first order one. The equivalent problem is to find a curve (, 1 , . . . ,
k1 ) : I U (Rn )k1 which satisfies the first order differential equation
(, 1 , . . . , k1 )0 (t) = (1 (t), . . . , k1 (t), F ((t), 1 (t), . . . , k1 (t))).
The adjective autonomous refers to the circumstance that the right-hand

side F depends only on the position (t) but not on other parameters, say
t. For example, a time dependent non-autonomous differential equation has
the form
0 (t) = F (t, (t)), (1.20)
where F : R U Rn is the time dependent right-hand side. However, such
a differential equation can also be rephrased as an autonomous differential
equation. Indeed, it is equivalent to finding curves (, ) : I RU satisfying
(, )0 (t) = (1, F ( (t), (t))).
For these reasons, we shall summarize here the fundamental theorems of

ordinary differential equations only for autonomous first order systems.
76 1. Preliminaries
Definition 1.7.3. A maximal integral curve of a differential equation is a

solution which cannot be extended to a larger interval as a solution of the
differential equation.
Theorem 1.7.4 (Existence and Uniqueness of Solutions). If the right-hand
side F : U Rn of a first order differential equation is continuously dif-
ferentiable, then for any p U , there is a unique maximal integral curve
p : (ap , bp ) U such that ap < 0 < bp + and p (0) = p.
Remark that for continuous right-hand side only the existence part of the
theorem is true, uniqueness may fail.
Theorem 1.7.5 (Smooth Dependence on the Initial Condition). Suppose
that the right-hand side of a first order ordinary differential equation is smooth.
Then the set W = {(p, t) | p U, t (ap , bp ) is an open subset of U R,
and the map : W U , (p, t) = p (t) is a smooth map.
In other words, the point at which we arrive at after traveling for time t along
an integral curve starting at p depends smoothly on p and t. As a corollary,
for any t R, the set Wt = {p | (p, t) W } is an open subset of U , and the
map t : Wt U , t (p) = (p, t) is smooth. If (p, t0 ) W and q = p (t0 ),
then the map (ap t0 , bp t0 ) U , t 7 p (t + t0 ) is a maximal integral
curve starting at q, therefore, q (t) = p (t + t0 ), q Wt0 and t0 (q) = p.
In conclusion, t : Wt Wt is a diffeomorphism for all t R.
Definition 1.7.6. The family {t }tR is called the flow or one-parameter
family of diffeomorphisms of the ordinary differential equation or the vector
field F .
The flow satisfies the following group property. If (p, t) W and (t (p), s)
W , then (p, t + s W ) and
t+s (p) = s (t (p)).
The typical reason why an integral curve cannot be extended to [0, ) is
that it is running out to infinity or to the boundary of U within a finite time.
Theorem 1.7.7 (Unboundedness of Maximal Solutions in Time or Space).
If for a maximal integral curve p : (ap , bp ) U , ap 6= (or bp 6= +),
then the trace ((ap , 0]) (or ([0, bp )), respectively,) cannot be covered by a
bounded closed subset of U .
Linear Differential Equations

Denote by Rnm the linear space of n m matrices and identify Rn with the
space Rn1 of column vectors. A linear differential equation is an equation
of the form
x0 = A x, (1.21)
where x : I Rn is an unknown column vector valued function defined on the

interval I, A : I Rnn is a given matrix valued function, denotes matrix
multiplication. As the matrix A typically depends on t, linear differential
equations are usually non-autonomous.
Of course, the fundamental theorems on ordinary differential equations are
true also for linear differential equations, but linearity has some additional
consequences not true in the general case. These are sum up in the following
theorem.
Theorem 1.7.8.
If A : I Rnn is smooth, t0 I is a given initial point and x0 Rn
is an arbitrary initial value, then there is a unique solution x : I Rn
of (1.21) for which x(t0 ) = x0 .
Recall that for a general differential equation, solution exist only in a
certain, maybe small neighborhood of the initial point t0 . In the linear
case, however, solution exists on the whole interval I.
Solutions of the linear differential equation (1.21) form an n-dimensional
linear space with respect to pointwise addition and multiplication by real
numbers. This property is a characterization of linearity of a differen-
tial equation.
In the special case when A is a constant matrix, solution of (1.21) with
initial condition x(t0 ) = x0 can be written explicitly as
x(t) = eA(tt0 ) x0 ,
Mk
P
where the exponential of a matrix M is defined as eM = k=0 k! .
Systems of Total Differential Equations

Systems of total differential equations are multivariable generalizations of the
non-autonomous differential equation (1.20).
Definition 1.7.9. A system of total differential equations is an equation of
the form
G0 (u) = F (u, G(u)) (1.22)
for an unknown multivariable function G, where F : Rnm is a given
matrix valued function on an open subset of Rm Rn . G : U Rn is a
solution of the system, if it is defined on an open subset U of Rm , its graph
{(u, G(u) | u U } is contained in , and (1.22) is satisfied by all u U .
A system of total differential equations is integrable if for any (u0 , v0 ) ,
there is a solution G of the system satisfying the initial condition G(u0 ) = v0 .
78 1. Preliminaries
If G is a solution with given initial value G(u0 ) = v0 , (e1 , . . . , em ) is the

standard basis of Rm , then for any 1 i m, the curve i (t) = G(u0 + tei )
satisfies the ordinary differential equation
i0 (t) = F (u0 + tei , i (t)) ei
with the initial condition i (0) = u0 . (In this equation, vectors are column
vectors, denotes matrix multiplication.) Solving these differential equa-
tions we can compute the values of G in a neighborhood of u0 along segments
through u0 parallel to one of the coordinate axes. As any point of Rm can
be connected to u0 by a broken line the segments of which are parallel to
one of the coordinate axes, iterating this process we can compute the val-
ues of G in a small neighborhood of u0 . This proves that the solution of a
system of total differential equations with a given initial condition is unique
in a neighborhood of the initial point. However, the solution may not exist.
When m 2, there are many ways to choose the broken line connecting u0
to a point u1 nearby, and it can happen that computing G(u1 ) with the help
of different broken lines we get different values. In such a way the system has
no solution with the given initial value and the system is not integrable.
Frobenius Theorem gives a necessary and sufficient condition for the in-
tegrability of a system of total differential equations. The easiest way to
paraphrase the condition is that the system is integrable if and only if it does
not contradict to Youngs Theorem (Theorem 1.5.11). Compute what this
means in terms of formulae. Let Fji (u, v) be the matrix element of F (u, v)
in the ith row and jth column. If G(u) = (G1 (u), . . . , Gn (u))> is a solution
of the system, then
j Gi (u) = Fji (u, G(u)) for all 1 i n, 1 j m.
Differentiating this equality with respect to the kth variable we obtain
n
X
k j Gi (u) = k Fji (u, G(u)) + m+s Fji (u, G(u))k Gs (u)
s=1
n
X
= k Fji (u, G(u)) + m+s Fji (u, G(u))Fks (u, G(u)).
s=1
By Youngs Theorem, the right hand side of this equality should not change
if we flip the role of j and k.
Theorem 1.7.10 (Frobenius Theorem). The system of total differential
equations (1.22) is integrable if and only if
n
X n
X
k Fji + m+s Fji Fks = j Fki + m+s Fki Fjs
s=1 s=1
holds on for all 1 i n, and 1 j, k m.

A geometrical version of Frobenius Theorem will be proved in Section 4.3.2.
Chapter 2
Curves in En
2.1 The Notion of a Curve

In elementary geometry, one meets a lot of examples of curves: straight lines,
circles, conic sections, cubic curves, graphs of functions defined on an interval
or the whole real line, intersections of surfaces etc. Based on these examples
everyone gets the feeling of what a curve is, however, it is not easy to give an
exact definition of a curve which is satisfactory in all respect. To illustrate
this, we give some commonly used definitions of certain classes of curves.
Definition 2.1.1. A simple arc in a topological space is a subset homeo-

morphic to a closed interval [a, b] of R. A parameterization of a simple arc is
a homeomorphism : [a, b] .
Probably anyone agrees that simple arcs are curves, but since open segments,
straight lines, conic sections and many other important examples of curves
are not simple arcs, this class of curves is too narrow.
We could define curves as finite unions of simple arcs. This wider class
includes circles, ellipses, but still excludes non-compact examples like straight
lines, hyperbolae, parabolas. Non-compact examples would be included if we
considered countable unions of simple arcs. This class of curves seems to
be wide enough, but maybe too wide. For example, it contains the set of
all those points in Rn which have at least one rational coordinate and it is
questionable whether we could call this set a curve.
Definition 2.1.2. A 1-dimensional topological manifold with boundary is a

second countable Hausdorff topological space, in which each point has an
open neighborhood homeomorphic either to an open interval or to a left-
closed, right-open interval of R.
81
82 2. Curves in En
This is a technical definition, but there is a simple description of 1-dimensional

topological manifolds with boundary. They have a finite or countable number
of open connected components and each connected component is homeomor-
phic either to an open, closed, or half-closed interval, or to a circle.
The class of 1-dimensional manifolds with boundary is wider than the class
of simple arcs, it includes much more important examples of curves, but as
it fixes the local structure of a curve quite strictly, it excludes examples of
curves having certain kind of singularities. For example, figure-eight shaped
curves, like Bernoullis lemniscate shown in Figure 2.13, are not a topological
manifolds, because the self-intersection point in the middle does not have a
neighborhood with the required property.
Definition 2.1.3. An algebraic plane curve in R2 is the set of solutions of

a polynomial equation P (x, y) = 0, where P 6= 0 is a polynomial in two
variables with real coefficients.
Algebraic plane curves may have a finite number of singular points, for exam-
ple self intersections, so they are not necessarily 1-dimensional manifolds, but
removing the singular points, the remaining set is a 1-dimensional manifold,
maybe empty. On the other hand, algebraic curves are very specific curves.
For example, if a straight line intersects an algebraic curve in an infinite
number of points, then it is contained in the curve. In particular the graphs
periodic non-constant functions (like the sine function) are not an algebraic
curves.
One can also define curves as 1-dimensional topological or metric spaces.
For such a definition one must have a proper notion of dimension. Possible
definitions of dimension for a topological or metric space are discussed in
a branch of topology called dimension theory, and within the framework of
geometric measure theory. These theories are out of the main focus of this
textbook.
All the above definitions define curves as topological spaces or subsets of
topological spaces having a certain property satisfied by a sufficiently large
family of known examples of curves.
The second approach, which will be more suitable for our purposes, derives
curves from the motion of a point. This view is reflected in the definition of
a continuous curve:
Definition 2.1.4. A continuous parameterized curve in a topological space

is a continuous map of an interval I into the space. The interval I can have
any of the forms (a, b), (a, b], [a, b), [a, b], (, b), (, b], (a, +), [a, +),
(, +) = R, where a, b R. If I = [a, b], then the images of a and b are
the initial and terminal points of the curve respectively. The path is said to
connect the initial point to the terminal point.
2.2. The Length of a Curve 83
We stress that according to this definition, a parameterized curve is a map

and not a set of points as in the earlier definitions. However, we can associate
to any parameterized curve a subset of the ambient space, the set of points
traced out by the moving point.
Definition 2.1.5. The trace or trajectory of a parameterized curve : I X
is the image of the map . We say that is a parameterization of the subset
A of X when A is the trace of .
In many cases, the trace of a continuous parameterized curve is a curve in
one of the above sense, but there are examples, when the image is not a curve
at all. The Italian mathematician Giuseppe Peano (1858-1932) constructed
a continuous parameterized curve that passes through each point of a square.
Such a pathology can not occur if we restrict ourselves to smooth curves.
Definition 2.1.6. A smooth parameterized curve in the Euclidean space En
is a smooth map : I En from an interval I into En .
Definition 2.1.7. We say that the continuous curve 1 : I1 Rn is obtained
from the curve 2 : I2 Rn by a reparameterization if there is a homeomor-
phism : I1 I2 such that 1 = 2 . We say that the reparameterization
preserves orientation if is increasing. A reparameterization is called regular
if is smooth and 0 (t) 6= 0 for all t I1 .
Intuitively, orientation preserving reparameterizations describe motions along
the same route with different timings.
2.2 The Length of a Curve

Definition 2.2.1. The length of a continuous curve : [a, b] Rn is the limit
of the lengths of inscribed broken lines with consecutive vertices (t0 ), (t1 ),
. . . , (tN ), where a = t0 < t1 < < tN = b and the limit is taken as
max1iN |ti ti1 | tends to zero. Provided that this limit is finite, the
curve is called rectifiable.
Exercise 2.2.2. Show that the limit of the lengths of the inscribed broken
lines always exists and it is equal to the supremum of the lengths. Construct
a continuous curve : [a, b] R2 having infinite length.
The following theorem yields a formula that can be used in practice to com-
pute the length of curves.
Theorem 2.2.3. A smooth curve : [a, b] Rn is always rectifiable and its
length is equal to the integral
Z b
l() = | 0 (t)|dt.
a
84 2. Curves in En
Proof. Denoting by x1 , x2 , . . . , xn the coordinate functions of x the length of

the broken line considered in Definition 2.2.1 is equal to
v
N uX
X u n
= t (xj (ti ) xj (ti1 ))2 .
i=1 j=1
By the Lagrange mean value theorem we can find real numbers ij such that
xj (ti ) xj (ti1 ) = x0j (ij )(ti ti1 ), ti1 < ij < ti .
Using these equalities we get

v
N
u n 0
X uX
= (ti ti1 )t xj (ij )2 .
i=1 j=1
Fix a positive . By Proposition 1.4.43, for each > 0, we can find a positive
such that t, t [a, b] and |t t | < imply |x0j (t) x0j (t )| < for all
1 j n.
Suppose that the approximating broken line is fine enough in the sense that
|ti ti1 | < for all 1 i N . Then we have by the triangle inequality
v v v
u n
u n 0
u n

uX
uX uX
x 0 ( )2 t xj (ti )2 t (x0 (ij ) x0 (ti ))2 n.
ij
t
j j j
j=1 j=1 j=1
Making use of this estimation we see that

v
N
u n 0
uX
X
(ti ti1 )t 2
xj (ti ) n(b a)
(2.1)

i=1 j=1
In this formula
v
N
u n 0 N
X uX X
(ti ti1 )t xj (ti )2 = (ti ti1 )| 0 (ti )|
i=1 j=1 i=1
Rb
is just an integral sum which converges to the integral a | 0 (t)|dt when
maxi |ti ti1 | tends to zero. Taking into account inequality (2.1) we see
that in this case the length of the inscribed broken lines also tends to this
integral.
2.2. The Length of a Curve 85
Let : I Rn be a smooth curve, a I be a given point. Consider the

function s : I R Z t
s(t) = | 0 ( )|d.
a
s(t) is the signed length of the arc of the curve between (a) and (t). It is
a monotone but not necessarily a strictly monotone function of t in general.
This fact motivates the following definition.
Definition 2.2.4. A smooth curve is said to be regular if 0 (t) 6= 0 for all
t I.
If is a regular curve, then s defines a regular reparameterization of . The
map s1 : s(I) Rn is referred to as a natural or unit speed parameteri-
zation of the curve or as a parameterization of by arc length. The second
name is justified by the fact that the speed vector
1 0 (s1 (t))
( s1 )0 (t) = 0 (s1 (t)) (s1 )0 (t) = 0 (s1 (t)) =
s0 (s1 (t)) | 0 (s1 (t))|
of this parameterization has unit length at each point.
Exercise 2.2.5. The curve cycloid is the trajectory of a peripheral point of
a circle that rolls along a straight line. Find a parameterization of the cycloid
and compute the length of one of its arcs.
Figure 2.1: Generation of the cycloid.
Sometimes it is more convenient to use the polar coordinate system in the

plane.
Definition 2.2.6. The polarpcoordinates (r, ) of a point (x, y) R2 in the
plane are the distance r = x2 + y 2 of the point from the origin and the
direction angle of the vector (x, y). The direction angle is not defined at
the origin (0, 0), and it is defined only modulo 2 at other points. There
is no continuous choice of for the whole punctured plane R2 \ {(0, 0)}.
However, one can choose continuously on the complement of any closed
half-line starting at the origin. E.g., on the complement of the half-line
{(x, y) | x 0, y =0}, can be defined continuously by the formula =
2 arctan y 2 (, ).
x+ x +y 2
86 2. Curves in En
Cartesian coordinates can be expressed in terms of the polar coordinates as
x = r cos(), y = r sin().
Exercise 2.2.7. Let : [a, b] R2 \ {0} be a smooth curve and denote by

(r(t), (t)) the polar coordinates of (t), where (t) is chosen to be a smooth
function of t. Prove that the length of is equal to the integral
b

Z q
r2 (t)0 2 (t) + r0 2 (t) dt.
a
Figure 2.2: A helix.
Exercise 2.2.8. Find a natural reparameterization of the helix (t) = (a cos t,

a sin t, bt).
2.3 Croftons Formula

There are more sophisticated integral formulas for the length of a curve. In
this section, we discuss some of them.
Topology and measure on the set of straight lines

Let E = AGr1 (R2 ) be the set of straight lines in the Euclidean plane R2 .
For (, p) R2 , denote by e,p E the straight line defined by the equation
x cos() + y sin() = p. As every straight line can be defined by such an
equation, the map : R2 E, (, p) = e,p is surjective. However, is not
injective since (, p) = (, p) if and only if = k and p = (1)k p for
some integer k Z. In particular, the band [0, ) R, closed from the left
and open from the right, is mapped onto E bijectively, and a boundary point
(, p) on its right side corresponds to the same line as the point (0, p). This
implies easily that if we equip the set E with the factor topology induced by
2.3. Croftons Formula 87
the surjective map , then E will be a Mobius band without its boundary
circle, and becomes a covering map, the universal covering map of E.
The Lebesgue measure on R2 defines a measure on E with the help
of the covering map as follows. Let a subset A E be -measurable if
and only if 1 (A) ([0, ] R) Lebesgue measurable and then let (A) =
(1 (A) ([0, ] R)).
Proposition 2.3.1. The measure is invariant under the isometry group
of the plane, that is, if Iso(R2 ) is an arbitrary isometry, A E is
a -measurable set of straight lines, then (A) is also -measurable, and
(A) = ((A)).
Proof. The isometry group of the plane is generated by the following three
types of transformations:
Rotations R by angle about the origin;
Translations Ta by a vector (a, 0) parallel to the x axis;
Reflection M in the x axis.
Hence it is enough to check the invariance of the measure under these trans-
formations.
It is clear that R (e,p ) = e+,p . Since the translation (, p) 7 ( + , p)
preserves Lebesgue measure, the statement is true for the rotations R .
The action of the translation Ta on the straight line parameters is not so sim-
ple as Ta (e,p ) = e,p+a cos() , nevertheless, as the determinant of the deriva-
tive of the transformation (, p) 7 (, p + a cos()) is 1, this transformation
also preserves Lebesgue measure.
Finally, the action of M on the line parameters is given by the map (, p) 7
(, p). The latter transformation is a reflection, therefore preserves the
Lebesgue measure.
The Planar Crofton Formula

Theorem 2.3.2. Let : [a, b] R2 is a C 1 curve, m : E N {} is
defined by m(e) = #{t [a, b] | (t) e}. In other words, m(e) is the
number of intersection points of the curve and the straight line e counted
with multiplicity. Then the length of is
Z
1
l = m d.
2 E
Proof. Consider the C 1 -map h : [0, ] [a, b] [0, ] R defined by
h(, t) = (, x(t) cos + y(t) sin ),
where x(t) and y(t) are the coordinates of (t), and apply Theorem 1.6.24
to h and the constant 1 function on [0, ] R. The number of preimages
88 2. Curves in En
of a point (, p) [0, ] R is the number m(e,p ) of intersection points

of the straight line e,p with the curve counted with multiplicities. The
determinant of the derivative matrix of h at (, t) is

1 0
det = x0 (t) cos +y 0 (t) sin
x(t) sin + y(t) cos x0 (t) cos + y 0 (t) sin
Thus, we get
Z Z Z b Z
m d = m(e,p )ddp = |x0 (t) cos + y 0 (t) sin |ddt.
E [0,]R a 0
R
To compute the integral 0 |x0 (t) cos + y 0 (t) sin |d, fix the value of t and
write the speed vector 0 (t) = (x0 (t), y 0 (t)) as v(t)(cos , sin ), where v(t) =
k 0 (t)k, is a direction angle of the speed vector or any angle if v(t) = 0.
Then Z Z
0 0
|x (t) cos + y (t) sin |d = v(t) |(cos( )|d.
0 0
Since the absolute value of the cosine function is periodic with period , the
last integral does not depend on and its value is
Z Z Z /2
| cos( )|d = | cos()|d = 2 cos()d = 2.
0 0 0
Combining these equations we get

Z Z bZ Z b
m d = |x0 (t) cos + y 0 (t) sin |ddt = 2v(t)dt = 2l ,
E a 0 a
as we wanted to show.
Topology and Measure on the Set of Hyperplanes

One can generalize the planar Croftons formula for curves in Rn . In the
higher dimensional version, we have to count the number of intersection
points of the curve with hyperplanes. Let H = AGrn1 (Rn ) denote the set
of hyperplanes in Rn . Let B n = {x Rn | kxk 1} be the unit ball centered
at the origin, Sn1 = {x Rn | kxk = 1} be its boundary sphere. Assign to
each pair (u, p) Sn1 R the hyperplane Hu,p = {x Rn | hu, xi = p}.
The map : Sn1 R H, (u, p) 7 Hu,p is surjective and since Hu,p = Hu,p
if and only if (u, p) = (u, p) or (u, p) = (u, p), is a double covering.
The map induces a factor topology and also a measure on H. A subset
A H is -measurable if and only if 1 (A) is ( 1 )-measurable, where
is the surface measure on Sn1 , 1 is the Lebesgue measure on the real line,
and if A is -measurable, then its measure is (A) = 21 ( )1 (1 (A)).
Crofton Formula in Rn
Theorem 2.3.3. Let : [a, b] Rn be a C 1 curve, and m : H N {}
be the map assigning to a hyperplane H the number m(H) = #{t [a, b] |
(t) H} of intersection points of the curve and the hyperplane H counted
with multiplicities. Then the length of is
Z
1
l = md.
n1 H
where n1 is the volume of the (n 1)-dimensional unit ball.
Proof. The proof is analogous to the planar case. Consider the C 1 -map
h : B n [a, b] B n R, h(x, t) = (x, hx, (t)i),
and apply Theorem 1.6.24 to it and the constant 1 function on B n R, For

x 6= 0, the number of h-preimages of (x, p) is the number of intersection
points of the hyperplane H kxk
x p .
, kxk
The number of h-preimages of (0, 0) is while for p 6= 0 h1 (0, p) = .
However, these values can be ignored since the set {0} R has measure 0 in
B n R.
The determinant of the derivative matrix of h at (x, t) is

1 0 ... 0 0
0 1 ... 0 0
.. . .

0
det(h (x, t)) = det . .. .. .. = hx, 0 (t)i.

.
0 0 ... 1 0
(t) hx, 0 (t)i
Thus, Theorem 1.6.24 yields

Z Z
|hx, 0 (t)i|dxdt = m(H kxk
x p )dxdp
, kxk
B n [a,b] B n R
Substituting p = kxkp in the second integral we see that

Z Z

m H kxk , kxk dxdp =
x p kxkm H kxkx
,p dxdp
n
B R n
B R
Z Z Z
1 2
= m(Hu,p )dudp = md.
n + 1 R Sn1 n+1 H
To compute the integral B n |hx, 0 (t)i|dx for a fixed t, write 0 (t) as v(t)u,
R
where v(t) = k 0 (t)k, u is a unit vector. Slice the ball with the hyperplanes
90 2. Curves in En
n
Hu, orthogonal to u. If [1, 1], then Hu, 0B is an (n 1)-dimensional
2
ball of radius 1 . Since the function hx, (t)i is equal to the constant
v(t) on this ball,
Z
|hx, 0 (t)i|dx = n1 v(t)| |(1 2 )(n1)/2 .
Hu, B n
Integrating the integrals over the slices

Z Z 1
|hx, 0 (t)i|dx = n1 v(t) | |(1 2 )(n1)/2 d
Bn 1
Z 1
= 2n1 v(t) (1 2 )(n1)/2 d
0
1
1 2n1
= 2n1 v(t) (1 2 )(n+1)/2 = v(t)
n+1 0 n+1
by Fubinis theorem. Integrating with respect to t we get
Z Z b
2n1 2n1
|hx, 0 (t)i|dxdt = v(t)dt = l (2.2)
n
B [a,b] a n + 1 n+1
which completes the proof.
Crofton Formula for Spherical Curves

To finish this section, we compute yet another version of the Crofton formula
for curves lying on the unit sphere Sn1 Rn . This formula will express the
length of a curve using only the number of intersection points with hyper-
planes passing through the origin. The scheme of producing the formula is
the same as before.
Denote by H0 = Grn1 (Rn ) the set of all hyperplanes passing through the
origin. For u Sn1 , let Hu H0 be the hyperplane orthogonal to u. The
map : Sn1 H0 , u 7 Hu is a double cover of H0 . The spherical measure
on Sn1 induces a measure on H0 for which A H0 is -measurable if
and only if 1 (A) is -measurable and if this is the case, then (A) = 21 (A).
Theorem 2.3.4. Let : [a, b] Sn1 be a spherical curve. For H H0 ,
denote by m(H) the number of intersection points of H with counted with
multiplicity, i.e.,
m(H) = #{t [a, b] | (t) H}.
Then the length l of can be obtained as the integral
Z Z
2
l = md = m(Hu )du,
nn H0 nn S n1
where n is the volume of the n-dimensional unit ball.
Proof. For a fixed t [a, b], any vector w Rn can be decomposed uniquely
into the sum of two vectors so that the first vector is parallel to (t), the
second one is orthogonal to it. The decomposition is
w = hw, (t)i(t) + (w hw, (t)i(t)).
Consider the map
h : B n [a, b] Rn R, (w, t) (w hw, (t)i(t), hw, (t)i)
encoding the two components of this decomposition.

If (x, s) Rn R and x 6= 0, then (w, t) h1 (x, s) if and only if t [a, b]
is a point such that (t) w and w = x + s(t). Since kwk2 = kxk2 + s2 ,
w B n if and only if (x, s) B n+1 . Thus, the image of h is the unit ball
B n+1 Rn R, and the number of h-preimages of (x, s) B n+1 is m H kxk x .
The determinant of the derivative matrix of h at (w, t) is

In (t) T (t) (t)
det(h0 (w, t)) = det

.

T
hw, 0 (t)i T (t) hw, (t)i 0 (t) hw, 0 (t)i
Adding suitable multiples of the last column to the previous ones we can
eliminate the (t) T (t) term from the upper left corner. Then applying
Exercise 1.6.26 we get

In (t)
det(h0 (w, t)) = det

hw, (t)i 0 (t) hw, 0 (t)i
0 0
= hw, (t)i + hw, (t)ih(t), (t)i.
This reduces to det(h0 (w, t)) = hw, 0 (t)i as is on the unit sphere, therefore
(kk2 )0 = 2h, 0 i 0.
Applying Theorem 1.6.24 for h and the constant 1 function on B n+1 yields
Z bZ Z
0
|hw, (t)i|dwdt = m(H kxk
x )dxds
a Bn (x,s)B n+1
Z p
= 2 1 kxk2 m(H kxk x )dx.
Bn
The second equation comes from Fubinis theorem by slicing the ball B n+1
by straight lines orthogonal to the hyperplane Rn {0}. By equation (2.2),
we have
92 2. Curves in En
Z b Z
2n1
|hw, 0 (t)i|dwdt = l .
a Bn n+1
R1
According to Corollary 1.6.27, if we denote the integral 0 rn1 1 r2 dr by
an , then
Z p Z
x )
2m(H kxk 1 kxk2 dx = 2an m(Hu ) du.
Bn Sn1
The last three equations show that

Z
l = cn m(Hu ) du, (2.3)
S n1
where cn = (n+1)a n
n1 . The value of cn can be obtained by computing an , but
it can also be obtained by evaluating (2.3) for a great circle. The length of a
great circle on Sn1 is 2 and almost all hyperplanes through the origin cut
a great circle in exactly 2 points, that is the set of exceptional
R hyperplanes,
that contain the great circle has measure 0. This way, Sn1 m(Hu )du for a
great circle is twice the surface measure (Sn1 ) = nn of the sphere. This
gives that cn = /(nn ).
Exercise 2.3.5. Find a direct proof of the equations
Z 1 p n1 n+1
an = rn1 1 r2 dr = =
0 n(n + 1)n 2nn
2k1 (k1)!
(2k+1)!! , if n = 2k,

=
(2k1)!!

(2k+2)!! 2 , if n = 2k + 1.
2.4 The Osculating k-planes

One of the most important tools of analysis is linearization, or more gener-
ally, the approximation of general objects with more easily treatable ones.
E.g. the derivative of a function is the best linear approximation, Taylor
polynomials are the best polynomial approximations of the function around
a point. Adapting this idea to the theory of curves, the following questions
arise naturally. Given a curve : [a, b] Rn and a point (t) on it, how can
we find the straight line, circle, conic or polynomial curve of degree less than
or equal to n etc. that approximates the curve around (t) the best or how
can we find the k-plane, k-sphere, quadric surface etc. that is tangent to the
curve at (t) with the highest possible order?
2.4. The Osculating k-planes 93
We shall deal now with the problem of finding the k-plane which is tangent
to a curve at a given point with the highest possible order. The classical
approach to this problem is the following. A k-plane is determined uniquely
by k + 1 points of it not lying in a (k 1)-plane. Let us take k + 1 points
(t0 ), (t1 ), . . . , (tk ) on the curve. If is a curve of general type then
these points span a unique k-plane, which will be denoted by A(t0 , . . . , tk ).
The k-plane we look for is the limit position of the k-planes A(t0 , . . . , tk ) as
t0 , . . . , tk tend to t.
To properly understand the last sentence, we need a definition of convergence
of k-planes. In section 1.4, we defined a topology on the affine Grassmann
manifold AGrk (Rn ) in different but equivalent ways. Constructing the topol-
ogy as a factor topology gives the following notion of convergence of k-planes.
Proposition 2.4.1. The sequence of k-planes X1 , X2 , . . . tends to the k-
plane X with respect to the topology of the affine Grassmann manifold
AGrk (Rn ) if and only if one can find points pj Xj , p X and linearly
independent direction vectors v1j , . . . , vkj of Xj and v1 , . . . , vk of X such that
limj pj = p and limj vij = vi for i = 1, . . . , k.
2
v1
vv
1 1
1
p
p1 p2 v2
2
v21 v2
Figure 2.3: Convergence of k-planes
Exercise 2.4.2. Show that convergence of k-planes as defined above is the

same as convergence with respect to the topology constructed in section 1.4.
The topology of each Grassmann manifold is metrizable, hence Hausdorff,

which implies that a convergent sequence in a Grassman manifold can have
at most one limit point. We give a direct proof of this fact, relying only on
Proposition 2.4.1.
Proposition 2.4.3. A sequence of k-planes can have at most one limit.
94 2. Curves in En
Proof. Suppose that the sequence X1 , X2 , . . . has two limits, say X and Y .
Then by the definition, one can find points pj , qj Xj , p X, q Y and
linearly independent direction vectors {v1j , . . . , vkj } and {w1j , . . . , wkj } of Xj ,
{v1 , . . . , vk } of X and {w1 , . . . , wk } of Y , such that lim pj = p, lim qj = q,
lim vij = vi and lim wij = wi for i = 1, . . . , k as j tends to infinity. Since
{v1j , . . . , vkj } and {w1j , . . . , wkj } span the same linear space, which contains
the direction vector pj qj , there exist a unique k k matrix (ajrs )1r,sk
and a vector (bj1 , . . . , bjk ) such that
k
X
vij = ajis wsj for i = 1, . . . , k;
s=1
(2.4)
k
X
pj qj = bjs wsj
s=1
The components of this matrix and the numbers bjs can be determined
ajrs
by solving the system (2.4) of linear equations, thus by Cramers rule, they
are rational functions (quotients of polynomials) of the components of the
vectors vij and wij . Using the fact that rational functions are continuous at
each point where they are defined one can show that the limits lim ajrs = ars ,
lim bjr = br exist as j . Taking j in (2.4) we obtain
k
X
vi = ais ws for i = 1, . . . , k;
s=1
k
X
pq= bs ws ;
s=1
from which follows that the vectors v1 , . . . , vk and w1 , . . . , wk span the same
linear space, i.e. the k-planes X and Y must be parallel and that the point
q is a common point of them, consequently X = Y .
Definition 2.4.4. Let : [a, b] Rn be a curve, t (a, b), 1 k n. If the

k-planes A(t0 , . . . , tk ) are defined for parameters close enough to t and their
limit exists as t0 , . . . , tk t, then the limit is called the osculating k-plane of
the curve at t. The osculating 1-plane of a curve is just the tangent of the
curve. (The word osculate comes from the Latin osculatio, which means
kissing or kiss).
The definition of the osculating k-plane is justified by intuition. It is the
k-plane that passes through k + 1 points infinitely close to a given point.
However, it is usually not convenient to determine the osculating k-plane
using its definition directly. Fortunately, we have the following theorem.
2.4. The Osculating k-planes 95
Theorem 2.4.5. Let : [a, b] Rn be a smooth curve, t (a, b), 1 k n.

If the derivatives 0 (t), 00 (t), . . . , (k) (t) are linearly independent, then the
osculating k-plane of is defined at t and it is the k-plane that passes through
(t) with direction vectors 0 (t), 00 (t), . . . , (k) (t).
To prove the theorem we need some preparation.
Definition 2.4.6. Let f : [a, b] Rn be a vector valued function, t0 , t1 , . . . .

[a, b] are different numbers. The higher order divided differences or difference
quotients are defined recursively
f0 (t0 ) := f (t0 ),
f (t1 ) f (t0 )
f1 (t0 , t1 ) := ,
t1 t0
..
.
fk1 (t1 , . . . , tk ) fk1 (t0 , . . . , tk1 )
fk (t0 , t1 , . . . , tk ) := .
tk t0
Exercise 2.4.7. Show that the k-th order divided difference is a symmetric
function of the variables t0 , . . . , tk and has the following explicit form
k
X 1
fk (t0 , t1 , . . . , tk ) = f (ti ) ,
i=0
0 (ti )
where (t) is the polynomial (t t0 )(t t1 ) . . . .(t tk ), hence

0 (ti ) = (ti t0 )(ti t1 ) . . . (ti ti1 )(ti ti+1 ) . . . (ti tk ).
Lemma 2.4.8. If f : [a, b] R is a smooth function, then there exists a
number (a, b) such that
f (k) ()
fk (t0 , t1 , . . . , tk ) = .
k!
Proof. Let P (t) be the polynomial of degree k for which f (ti ) = P (ti ) for
i = 0, 1, . . . , k. Such a polynomial exists and is unique. P is unique, since
if Q is also a polynomial of degree k such that f (ti ) = P (ti ) = Q(ti ) for
i = 0, . . . , k, then the polynomial P Q is of degree k, and P Q has k + 1
roots, which is possible only in the case when P Q = 0, i.e., P = Q. We
show the existence of P by an explicit construction. Set
(t t0 )(t t1 ) . . . (t ti1 )(t ti+1 ) . . . (t tk )

Pi (t) = .
(ti t0 )(ti t1 ) . . . (ti ti1 )(ti ti+1 ) . . . (ti tk )
96 2. Curves in En
Obviously, Pi is a polynomial of degree k such that Pi (tj ) = ij . (The

Kronecker symbol denotes ij = 1 if i = j and ij = 0 if i 6= j.) Thus, the
polynomial
Xk
P (t) = f (ti )Pi (t)
i=0
is good for our purposes.
The difference f P has k + 1 roots. Since by the mean value theorem the
interval between two zeros of a smooth function contains an interior point
at which the derivative vanishes, the derivative f 0 P 0 has at least k roots.
Similarly, f 00 P 00 has at least k 1 roots, etc. f (k) P (k) vanishes at a
certain point [a, b]. The k-th derivative of a polynomial of degree k is k!
times the coefficient of the highest power, from which
k
X 1
f (k) () = P (k) () = k! f (ti ) = k!fk (t0 , t1 , . . . , tk ).
i=0
0 (t i)
Corollary 2.4.9. Since can be chosen from the interval spanned by the
points t0 , . . . , tk , if these points tend to t [a, b], then also tends to t,
f (k) () f (k) (t)
consequently fk (t0 , t1 , . . . , tk ) = tends to .
k! k!
Corollary 2.4.10. Applying the previous corollary to the components of a
vector valued function f : [a, b] Rn , we obtain that fk (t0 , t1 , . . . , tk ) tends
f (k) (t)
to as t0 , . . . , tk tend to t.
k!
Proof of Theorem 2.4.5. Let us recall that if p0 , . . . , pk are position vectors
of k + 1 points in Rn , then the affine plane spanned by them consists of linear
combinations the coefficients in which have sum equal to 1
A(p0 , . . . , pk ) = {0 p0 + + k pk | 0 + + k = 1}.
The direction vectors of this affine plane are linear combinations 0 p0 + +
k pk such that 0 + + k = 0.
We claim that if : [a, b] Rn is a curve in Rn , then k (t0 , . . . , tk ), (k 1) is
a direction vector of the affine subspace spanned by the points (t0 ), . . . (tk ).
Pk
To see this, it suffices to show that i=0 1/ 0 (ti ) = 0. Consider the function
f = 1 and construct the polynomial P of degree k which coincides with
f at t0 , . . . , tk using the general formulae. By the above proposition, there
exists a number such that
k
X 1
0 = f (k) () = P (k) () = ,
i=0
0 (ti )
2.5. Frenet Frames and Curvatures 97
This way, (t0 ) is a point and 1 (t1 , t0 ), . . . , k (tk , . . . , t0 ) are direction vectors
of the affine subspace spanned by the points (t0 ), . . . , (tk ). If is smooth,
then (t0 ) tends to (t), 1 (t1 , t0 ) tends to 0 (t), and so on, k (tk , . . . , t0 )
tends to (k) (t)/k! as the points t0 , . . . , tk tend to t [a, b]. Since by our
assumption the first k derivatives of are linearly independent at t, so are
the vectors 1 (t1 , t0 ), . . . , k (tk , . . . , t0 ) if t0 , . . . , tk are in a sufficiently small
neighborhood of t, and in this case the k-plane A(t0 , . . . , tk ) tends to the
k-plane that passes through (t) with direction vectors 0 (t), 00 (t)/2, . . . ,
(k) (t)/k!.
Exercise 2.4.11. Prove by induction on k that k (t0 , . . . , tk ), (k 1) is a
direction vector of the affine subspace spanned by the points (t0 ), . . . , (tk ).
Definition 2.4.12. Let V be an n-dimensional vector space. A flag in V

is a sequence of linear subspaces {0} = V0 V1 Vn = V such that
dim Vi = i. An affine flag is a sequence of affine subspaces A0 A1
An = V such that dim Ai = i.
Definition 2.4.13. A curve : I Rn is called a curve of general type in Rn
if the first n 1 derivatives 0 (t), 00 (t), . . . , (n1) (t) are linearly independent
for all t I.
Definition 2.4.14. The osculating flag of a curve of general type at a given
point is the affine flag consisting of the osculating k-planes for k = 0, 1, . . . , n
1 and the whole space.
2.5 Frenet Frames and Curvatures, the Fun-

damental Theorem of Curve Theory
Our plan is the following. A curve of general type in Rn is not contained in
any affine subspace of dimension k < n 1 (prove this!), so we may pose the
question how far it is from being contained in a k-plane. In other words, we
want to measure the deviation of the curve from its osculating k-plane. One
way to do this is that we measure how quickly the osculating flag rotates
as we travel along the curve. Since the faster we travel along the curve the
faster change we observe, it is natural to consider the speed of rotation of the
osculating flag with respect to the unit speed parameterization of the curve.
This will lead us to quantities that describe the way a curve is winding in
space. These quantities will be called the curvatures of the curve. There is
one question of technical character left: how can we measure the speed of
rotation of an affine subspace? This problem can be solved by introducing an
orthonormal basis at each point in such a way that the first k basis vectors
98 2. Curves in En
span the osculating k-plane at the point in question, then measuring the
speed of change of this basis.
Definition 2.5.1. A (smooth) vector field along a curve : I Rn is a
smooth mapping v : I Rn .
Remark. There is no formal difference between a curve and a vector field
along a curve. The difference is only in the interpretation. When we think of
a map v : I Rn as a vector field along the curve : I Rn we represent
(depict) v(t) by a directed segment starting from (t).
Definition 2.5.2. A moving (orthonormal) frame along a curve : I Rn
is a collection of n vector fields t1 , . . . , tn along such that hti (t), tj (t)i = ij
for all t I.
There are many moving frames along a curve and most of them have nothing
to do with the geometry of the curve. This is not the case for Frenet frames.
Definition 2.5.3. A moving frame t1 , . . . , tn along a curve is called a
Frenet frame if for all k, 1 k n, (k) (t) is contained in the linear span of
t1 (t), . . . , tk (t).
t2(t) t2(t)
v(t) t1(t)
g(t) g(t) g(t)
t1(t)
Figure 2.4: A vector field, a moving frame and the distinguished Frenet frame
along a curve.
Exercise 2.5.4. Construct a curve which has no Frenet frame and one with
infinitely many Frenet frames. Show that a curve of general type in Rn has
exactly 2n Frenet frames.
According to the exercise, a Frenet frame along a curve of general type is
almost unique. To select a distinguished Frenet frame from among all of
them, we use orientation.
Definition 2.5.5. A Frenet frame t1 , . . . , tn of a curve of general type in
Rn is called a distinguished Frenet frame if for all k, 1 k n 1, the
vectors t1 (t), . . . , tk (t) have the same orientation in their linear span as the
vectors 0 (t), . . . , (k) (t), and the basis t1 (t), . . . , tn (t) is positively oriented
with respect to the standard orientation of Rn .
Proposition 2.5.6. A curve of general type possesses a unique distinguished

Frenet frame.
Proof. We can determine the first n 1 vector fields of the distinguished
Frenet frame by application of the GramSchmidt orthogonalization process
pointwise to the first n 1 derivatives of . According to this recursive
procedure, explained in Theorem 1.2.55, we start with setting
0
t1 = .
k 0 k
If t1 , . . . , tk1 have already been defined, where k n 1, then we compute
the vector
fk = (k) (h (k) , t1 it1 + + h (k) , tk1 itk1 ),
and then set
fk (k) (h (k) , t1 it1 + + h (k) , tk1 itk1 )
tk = = (k) .
kfk k k (h (k) , t1 it1 + + h (k) , tk1 itk1 )k
To finish the proof, we have to show that given n 1 mutually orthogonal
unit vectors t1 , . . . , tn1 in Rn , there is a unique vector tn for which the
vectors t1 , . . . , tn form a positively oriented orthonormal basis of Rn . The
condition that a vector is perpendicular to t1 , . . . , tn1 is equivalent to a
system of n 1 linearly independent linear equation, the solutions of which
form a 1-dimensional linear subspace (a straight line). There are exactly two
opposite unit vectors parallel to a given straight line, and exactly one of them
will fulfill the orientation condition. (Replacing a vector of an ordered basis
by its opposite changes the orientation.)
Exercise 2.5.7. Show that if e1 , . . . , en is the standard basis of Rn and
ti = i1 e1 + + in en for i = 1, . . . , n 1,
then tn can be obtained as the formal determinant of the matrix

11 ... 1n
.. ..
. . .

(n1)1 . . . (n1)n

e1 ... en
Proposition 2.5.8. Let be a curve of general type in Rn . Denote by
t1 , . . . , tn its distinguished Frenet frame and set v = k 0 k. Let the matrix
(ij )1i,jn be such that
n
1 0 X
ti = ij ti . (2.5)
v j=1
Then
100 2. Curves in En
(i) ij = 0 provided that j > i + 1, and
(ii) the matrix (ij )1i,jn is skew-symmetric, i.e. ij = ji .
Proof. (i) Since ti , 1 i n 1, is a linear combination of the vectors

0 , . . . , (i) , t0i is a linear combination of the vectors 0 , . . . , (i+1) . As the
latter vectors are linear combinations of the vectors t1 , . . . , t(i+1) , the first
statement is proved.
(ii) Since hti , tj i ij is a constant function, we get
1 0
ij + ji = (ht , tj i + hti , t0j i) = 0
v i
by differentiation. The proposition is proved.
According to Propositions 2.5.8, only the entries i,i+1 = i+1,i of the

matrix (ij ) may differ from zero. Setting
1 = 12 , 2 = 23 , . . . ., n1 = n1,n ,
we see that equations (2.5) collapse to the following form
1 0
t = 1 t2
v 1
1 0
t = 1 t1 + 2 t3
v 2
..
.
1 0
t = n2 tn2 + n1 tn
v n1
1 0
t = n1 tn1 .
v n
These formulae are called the Frenet formulae for a curve of general type
in Rn . The functions 1 , . . . , n1 are called the curvature functions of the
curve.
We formulate two invariance theorems concerning the curvatures of a curve.
They are intuitively clear and their proof is straightforward.
Proposition 2.5.9 (Invariance under isometries). Let be a curve of general

type in Rn , : Rn Rn be an isometry (distance preserving bijection).
Then the curvature functions 1 , . . . , n2 of the curves and are the
same. The last curvatures n1 of these curves coincide if is orientation
preserving and they differ (only) in sign if is orientation reversing.
Proposition 2.5.10 (Invariance under reparameterization). If is a regular

reparameterization of the curve i.e. = h for some smooth function
h with property h0 > 0 or h0 < 0, then the curvature functions of and
are related to one another by i = i h for 1 i n 2 and n1 =
n(n+1)
(sgn(h0 )) 2 n1 h.
Exercise 2.5.11. Assume a curve of general type in Rn lies in Rn1 Rn .
Then we can compute the curvatures 1 , . . . , n1 of this curve considering
a curve in Rn and also we may compute the curvatures 1 , . . . , n2 of this
curve considering a curve in Rn1 . What is the relationship between these
two sets of numbers?
Computation of the Curvature Functions

Our goal now is to express the curvatures of a curve of general type in Rn
in terms of the derivatives of . For this purpose write the derivatives of
as linear combinations of the distinguished Frenet frame. By the definition
of the Frenet frame, the kth derivative must be the linear combination of the
first k Frenet vector fields
0 = a11 t1 ,
00 = a12 t1 + a22 t2 ,
..
.
(2.6)
(k) = a1k t1 + + akk tk ,
..
.
(n) = a1n t1 + a2n t2 + + ann tn .
The coefficients aji can be expressed recursively by the speed length function
v = k 0 k and the curvature functions. To illustrate this, let us compute the
first three decompositions assuming n 3.
The speed vector is simply
0 = vt1 . (2.7)
Differentiating and applying the first Frenet equation
00 = v 0 t1 + vt01 = v 0 t1 + v 2 1 t2 . (2.8)
Differentiating again and using the first two Frenet equations we get
000 = v 00 t1 + v 0 t01 + (v 2 1 )0 t2 + v 2 1 t02
= v 00 t1 + v 0 v1 t2 + (v 2 1 )0 t2 + v 2 1 v(2 t3 1 t1 ) (2.9)
00 0 0
= (v v 3 21 )t1 2 3
+ (v v1 + (v 1 ) )t2 + v 1 2 t3 .
102 2. Curves in En
In principle, iterating this method we can compute all the coefficients aji but
already the third derivative shows that the coefficients become complicated
and do not show any pattern except for the last coefficients.
Proposition 2.5.12.
akk = v k 1 k1 for 1 k n.
Proof. We prove the equation by induction on k. The base cases are verified
above for k 3. Assume now that the statement holds for k 1 < n. Then
differentiating
(k1) = a1k1 t1 + + ak2 k1

k1 tk2 + ak1 tk1 .
and applying the Frenet formulae we obtain

k1
X
(k) = ((ajk1 )0 tj + ajk1 t0j )
j=1
k1
X
= ((a1k1 )0 t1 + a1k1 v1 t2 ) + ((ajk1 )0 tj + ajk1 v(j tj+1 j1 tj1 )).
j=2
We see that tk appears only in the last summand, when j = k 1 and its
coefficient is
akk = ak1
k1 vk1 = (v
k1
1 k2 )vk1 = v k 1 k1 .
Proposition 2.5.13. The curvature functions 1 , . . . , n2 of a curve of

general type in Rn are positive. (However, there is not any restriction on the
sign of n1 .)
Proof. By the orientation condition on the distinguished Frenet frame, for
k n 1, 0 , . . . , (k) has the same orientation as t1 , . . . , tk in the linear
space they span. According to the definition of having the same orientation,
this means that the determinant
1
a1 . . . 0 . . . 0
.. . . . .. .
. . .. . ..
1 k
1 i

k = det ai . . . ai . . . 0 = a1 ak (2.10)

. . . . .
. . . . .

. . . . .
a1k . . . aik . . . akk
of the lower triangular matrix of coefficients is positive. Since the quotients of

positive numbers is positive, we obtain also that a11 = 1 and akk = k /k1
for 2 k n 1 are positive. However, then the quotients akk /ak1

k1 = vk1
are also positive for 2 k n 1. Since v = k 0 k > 0,this means that all
the curvatures 1 , . . . , n2 are positive.
The above computation shows also that
sgn(n1 ) = sgn(ann /an1 n
n1 ) = sgn(an ) = sgn(n /n1 ) = sgn(n ).
Consequently, the sign of the last curvature coincides with the the sign of
n , which is positive (or negative) if and only if ( 0 , . . . , (n) ) is a positively
(or negatively) oriented basis of Rn . It is 0 if and only if 0 , . . . , (n) are
linearly dependent. The remark at the end of the proposition follows from
the Fundamental Theorem of Curve Theory, which will be proved below.
This theorem that for any smooth function n1 on an interval, there is a
curve of general type in Rn the last curvature of which is n1 .
As a byproduct of the proof, we obtain the following expressions for the
curvatures
a2 2 /1 2
1 = 21 = = 3
va1 v1 v
and
ak+1
k+1 k+1 /k k+1 k1
k = = =
vakk vk /k1 v2k
for 2 k n 1. Thus, to find formulae for the curvatures it is enough
to express the determinants k or the diagonal elements akk in terms of the
derivatives of .
If we take the wedge product of the first k equations (2.6) and apply Propo-
sition 1.2.31, we get
0 (k) = k t1 tk . (2.11)
As t1 , (t) . . . , tn (t) an orthonormal basis for each parameter t,
k 0 (k) k = |k |. (2.12)
Since the determinants k are positive for k n 2, the absolute value can
be omitted for all but the last k. The length of the k-vector 0 (k) can
be expressed using (1.9) with the help of the Gram matrix G( 0 , . . . , (k) ).
Thus for k n 1,
q
k = k 0 (k) k = G( 0 , . . . , (k) )
v 0 0
h , i . . . h 0 , (k) i
u
u
=u
u .. .. ..
tdet

. . .
h (k) , 0 i . . . h (k) , (k) i
104 2. Curves in En
We lost information on the sign of n when we took the absolute value of

the sides of (2.11), so to obtain n together with its sign, return to equation
(2.11) and consider it for k = n. As the Frenet frame assigns a positively
oriented orthonormal basis to each parameter, denoting by e1 , . . . , en the
standard basis of Rn
t1 tn = e1 en ,
consequently
0 (n) = n e1 en .
By the last equation we can obtain n by writing the derivatives of as
linear combinations of the standard basis and computing the determinant of
the matrix of coefficients. If we decompose as a linear combination
= (1 , . . . , n ) = 1 e1 + + n en
of the standard basis, then
(k) (k)
(k) = (1 , . . . , n(k) ) = 1 e1 + + n(k) en ,
and therefore
10 n0 0

...
100 ... n00 00
n = det .. .. .. = det ..

. . . .
(n) (n)
1 ... n (n)
We summarize our computation in the following theorem.
Theorem 2.5.14. Let : I Rn be a curve of general type in Rn , 1 k
n 1. Then the curvature functions of can be computed by the equations
2 k+1 k1
1 = and k = for k 2,
v3 v2k
where v = k 0 k, and the numbers k are given by
v 0 0
h , i . . . h 0 , (k) i
u
u
k = u
u .. .. ..
tdet

. . .
(k) 0 (k) (k)
h , i . . . h , i
0

00
for k < n and n = det .. .

.
(n)
When we want to compute the curvatures of a curve directly, without com-

puting the Frenet frame, the quickest method is probably the evaluation of
the formulae in the theorem. However, if we need to compute the Frenet
frame as well, and we compute them following the proof of Proposition 2.5.6,
then curvatures pop up simply during the GramSchmidt orthogonalization
process. Indeed, in the kth step of the procedure we compute the vector
k1
X
fk = (k) h (k) , ti iti .
i=1
Taking the decompositions (2.6) into account, we see that
h (k) , ti i = aik ,
and
k1
X
(k)
fk = aik ti = akk tk .
i=1
This way, the length of fk , which we also have to compute for evaluating the
equation
fk
tk = ,
kfk k
is equal to
kfk k = |akk | = v k |1 k1 |,
and therefore
kfk+1 k
|k | = .
vkfk k
This formula allows us to compute also |n1 |, but to determine the sign
of n1 , one needs further consideration of the orientation of the first n
derivatives of .
Exercise 2.5.15. Compute the curvatures of the moment curve (t) =
(t, t2 , . . . , tn ) at t = 0.
Fundamental Theorem of Curve Theory

Theorem 2.5.16 (Fundamental Theorem of Curve Theory). Given n 1
positive smooth functions v, 1 , . . . , n2 and a smooth function n1 on an
interval I, there exists a curve : I Rn of general type in Rn such that
k 0 k = v and the curvatures of are the prescribed functions 1 , . . . , n1 .
This curve is unique up to orientation preserving isometries of the space, that
is, if is another curve with the same properties, then there is an orientation
preserving isometry Iso+ (Rn ) such that = .
106 2. Curves in En
Proof. We can eliminate the freedom given by orientation preserving isome-

tries if we fix a parameter t0 I and restrict our attention to curves : I
Rn such that (t0 ) = 0 and the Frenet basis of at t0 coincides with the
standard basis of Rn . We are to show that within this family of curves, each
collection of allowed curvature functions corresponds to a unique curve. The
proof below describes a way to reconstruct the curve from its curvatures by
solving some differential equations.
Let T denote the n n matrix, whose rows are the Frenet vector fields of
the unknown curve : I Rn . If C = (cij ) : I Rnn is the matrix valued
function the only non-zero entries of which are ci,i+1 = ci+1,i = vi for
1 i n 1, then T must satisfy the linear differential equation
T0 = C T (2.13)
Thus, prescribing the Frenet basis at a given point T(t0 ), we can obtain
the whole moving Frenet frame as a unique solution of (2.13). For each
t I, T(t) must be an orthogonal matrix, that is a matrix the rows of which
form an orthonormal basis of Rn . Orthogonality is equivalent to the equation
T TT = TT T = In , where TT is the transposition of T, In is the n n unit
matrix. Furthermore, since the Frenet frame must be a positively oriented
basis of Rn , T(t) must have positive determinant for all t I.
Recall that orthogonal matrices with positive determinant are called special
orthogonal matrices. The set (in fact group) of n n special orthogonal
matrices is usually denoted by SO(n).
We claim that if T is a solution of (2.13) and there is a number t0 , for which
T(t0 ) SO(n), then T(t) SO(n) for all t I. This means that if we take
care of the restrictions on T when we choose the initial matrix T0 , then we
do not have to worry about the other values of T.
To show the statement, set M = TT T. Then
M0 = (T0 )T T + TT T0 = (C T)T T + TT C T = TT (C T + C) T = 0,
(2.14)
as C is skew symmetric. (2.14) shows that M is constant, therefore M
M(t0 ) = In . This proves that T(t) is an orthogonal matrix for all t I.
As for the determinant of T(t), observe that the determinant of an orthogonal
matrix A Rnn satisfies
(det A)2 = det A det AT = det(A AT ) = det In = 1,
hence det A = 1. Since det(T(t)) is a continuous function of t, which takes

only the values 1, it must be constant. We have assumed that T(t0 ) > 0,
thus T(t) = 1 > 0 for all t I.
2.6. Limits of Some Geometrical Quantities 107
The derivative of is vt1 , thus we can obtain from the Frenet frame by
integrating vt1
Z t
(t) = v( )t1 ( )d. (2.15)
t0
Solving (2.13) we get the only possible Frenet frame, then formula (2.15)
defines the only curve which can satisfy the conditions. This proves unique-
ness. To show the existence part of the theorem we have to check that the
curve we have obtained has the prescribed curvature and length of speed
functions. The lengthy but straightforward verification of this fact is left to
the reader.
Exercise 2.5.17. Plot the curve, called astroid, given by the parameteriza-
tion : [0, 2] R2 , (t) = (cos3 (t), sin3 (t)). Is the astroid a smooth curve?
Is it regular? Is it a curve of general type? If the answer is no for a property,
characterize those arcs of the astroid which have the property. Compute the
length of the astroid. Show that the segment of a tangent lying between the
axis intercepts has the same length for all tangents.
Exercise 2.5.18. Find the distinguished Frenet frame and the equation of
the osculating 2-plane of the elliptical helix t 7 (a cos t, b sin t, ct) at the point
(a, 0, 0) (a, b and c are given positive numbers).
Exercise 2.5.19. Show that a curve of general type in Rn is contained in

an (n 1)-dimensional affine subspace if and only if n1 0.
Exercise 2.5.20. Prove that if a curve of general type in R3 has constant

curvatures then it is either a circle or a helix.
Exercise 2.5.21. Describe those curves of general type in Rn which have

constant curvatures.
2.6 Obtaining Curvatures

as Limits of Some Geometrical Quantities
In this section, we compute the limit of some geometrical quantities depending
on a collection of points on the curve as the points tend simultaneously to a
given curve point. The limits will be expressed by the curvatures at the given
point, thus, each of these results gives rise to a geometrical interpretation of
curvatures.
108 2. Curves in En
The angle between osculating k-planes

Let : I Rn be a curve of general type in Rn parameterized by arc length.
Fix a point t0 I and for t0 + s I, denote by k (s) the angle between the
osculating k-planes of at t0 and at t0 + s for every k < n.
Proposition 2.6.1. As the parameter t0 + s I tends to t0 , the quotient

k (s)/|s| tends to |k (t0 )|.
Proof. The direction space of the osculating k-plane at t I is spanned by

the first k Frenet vectors t1 (t), . . . , tk (t). Thus, the angle k (s) is the angle
between the unit k-vectors (t1 tk )(t0 ) and (t1 tk )(t0 + s). By
elementary geometry, if an isosceles triangle has two equal sides of unit length
enclosing an angle , then the length of the third side is 2 sin(/2). Applying
this theorem gives

k (s) k (s)/2 (t1 tk )(t0 + s) (t1 tk )(t0 )
= .
|s| sin(k (s)/2) s
When s tends to 0, the angle k (s) also tends to 0, so
k (s)/2
lim = 1.
s0 sin(k (s)/2)
As a corollary, we get

k (s) (t1 tk )(t0 + s) (t1 tk )(t0 )
lim = lim

s0 |s| s0 s
= k(t1 tk )0 (t0 )k.
Computing the derivative with the Leibniz rule and applying the Frenet for-
mulae
k
X
(t1 tk )0 = t1 t0i tk
i=1
k
X
= (1 t2 t2 tk + t1 (i ti+1 i1 ti1 ) tk .
i=2
As the wedge product of vectors vanishes if two of the vectors are equal, the
only nonzero summand on the right-hand side occurs at i = k
(t1 tk )0 = t1 tk1 (k tk+1 ).

Combining all we have

k (s)
lim = k(t1 tk )0 (t0 )k
s0 |s|
= |k (t0 )| k(t1 tk1 tk+1 )(t0 )k = |k (t0 )|,

as it was to be proven.
Volume and Dihedral Angles of Inscribed Simplices

Let : I Rn be a curve of general type in Rn , k n. Fix a point t I
and choose k + 1 pairwise distinct parameters t0 , . . . , tk I. Then the curve
points (t0 ), . . . , (tk ) span a k-dimensional simplex, the k-dimensional vol-
ume of which we denote by Vk (t0 , . . . , tk ). It is clear that if all the parameters
t0 , . . . , tk tend to t simultaneously, then this volume tends to 0.
Proposition 2.6.2. The limit of the volume Vk (t0 , . . . , tk ) divided by the
product of the edge lengths of the simplex spanned by the points (t0 ), . . . , (tk )
as all the parameters ti tend to t, is equal to
k
!
Vk (t0 , . . . , tk ) 1 Y 1
lim Q = |k1
1 (t)k2
2 (t) k1 (t)|.
i:ti t 0i<jk k(tj ) (ti )k k! m=1 m!
Proof. The volume of a simplex can be obtained from the wedge product of
the edge vectors running out from one of the vertices
1
k((t1 ) (t0 )) ((tk ) (t0 ))k.
Vk (t0 , . . . , tk ) =
k!
We know that the mth order divided difference of has the following explicit
form
m
X 1
m (t0 , . . . , tm ) = Qm (ti ).
j=0 (ti tj )
i=0 j6=i
During the proof of Theorem 2.4.5 we showed that right-hand side is a 0

weight linear combination of the points (ti ). Consequently, equation
m
X 1
m (t0 , . . . , tm ) = Qm ((ti ) (t0 )) (2.16)
i=1
j=0 (ti tj )
j6=i
also holds. Consider equation (2.16) for m = 1, . . . , k and take the wedge
product of these k equations. Applying Proposition 1.2.31 we get
1 (t0 , t1 ) k (t0 , . . . , tk )
1
= Qk Qm1 ((t1 ) (t0 )) ((tk ) (t0 )).
m=1 j=0 (tm tj )
(2.17)
110 2. Curves in En
Taking the length of both sides and rearranging the equation gives
Vk (t0 , . . . , tk ) 1 k1 (t0 , t1 ) k (t0 , . . . , tk )k

Q = Q .
0i<jk k(tj ) (ti )k k! 0i<jk k1 (ti , tj )k
As we know the limits of the higher order divided differences as all the ti s
tend to t, we can compute the limit of the right-hand side easily
k
!
Vk (t0 , . . . , tk ) 1 Y 1 k( 0 (k) )(t)k
lim Q = .
i:ti t 0i<jk k(tj ) (ti )k k! m=1
m! k 0 (t)k(k+1)k/2
Making use of (2.12), (2.10) and Proposition 2.5.12, we obtain
k 0 (k) k = |k | = |a11 akk | = v (k+1)k/2 |k1

1 k2
2 k1 |,

By elementary geometry, if the area of a triangle is A and its side length are
a, b, c, then the radius R of the circumcircle of the triangle is expressed by
the formula
abc
R= .
4A
Thus, our theorem for k = 2 has the following corollary.
Corollary 2.6.3. For a curve : I Rn of general type in Rn , and for
any three distinct points t0 , t1 , t2 in I, denote by R(t0 , t1 , t2 ) [0, ] the
circumradius of the triangle (t0 )(t1 )(t2 ). Then
1
lim R(t0 , t1 , t2 ) = .
t0 ,t1 ,t2 t |1 (t)|
Though for k > 2 the limit in Proposition 2.6.2 involves several curvatures,
it is possible to construct expressions of the volumes of faces of various di-
mensions of a simplex whose limit gives one single curvature. An example of
such a formula is Darbouxs formula for the second curvature.
Corollary 2.6.4 (Darboux). Using the notations of Proposition 2.6.2,
9 V3 (t0 , t1 , t2 , t3 )
lim p = |2 (t)|.
t0 ,t1 ,t2 ,t3 t 2 V2 (t1 , t2 , t3 )V2 (t0 , t2 , t3 )V2 (t0 , t1 , t3 )V2 (t0 , t1 , t2 )
Consider again a curve : I Rn of general type in Rn , and let 2 k n

be an integer. Fix a point t I and if k = n, then suppose that n1 (t) 6= 0.
By Proposition 2.6.2, when the points t0 , . . . , tk I are sufficiently close to t,
the simplex S spanned by (t0 ), . . . , (tk ) will be non-degenerate. Consider
two facets of S, say the facet F spanned by the points (t0 ), . . . , (tk1 ) and
F spanned by (t1 ), . . . , (tk ). These two facets intersect along a (k 2)-
dimensional face L = F F . Let = (t0 , . . . , tk ) be the angle of the
(k 1)-planes spanned by the facets F and F .
We remark that if is the angle between the outer unit normals of the facets
F and F , then the dihedral angle of the simplex at the (k 2)-dimensional
face L is defined to be the complementary angle . The angle of
the (k 1)-planes is the smaller of and the dihedral angle , that is,
= min{, } [0, /2].
Proposition 2.6.5. If the distinct parameters t0 , . . . , tk I tend to t simul-

taneously, then
(t0 , . . . , tk ) k1 (t)
lim = .
i:ti t k(tk ) (t0 )k k
Proof. Denote by P the vertex (tk ), by P 0 is orthogonal projection onto

the (k 1)-plane containing F , and by P 00 its orthogonal projection onto the
(k 2)-plane spanned by L. The triangle P P 0 P 00 4 is a right triangle with
0
right angle at P 0 and angle at P 00 . For this reason, sin() = PPPP00 . P P 0 is
the height of the original simplex corresponding to the facet F , P P 00 is the
height of the simplex F corresponding to its face L. Thus, we can compute
the lengths of these segments from the volume formula for simplices:
Vk (t0 , . . . , tk ) Vk1 (t1 , . . . , tk )

PP0 = k , and P P 00 = (k 1) .
Vk1 (t0 , . . . , tk1 ) Vk2 (t1 , . . . , tk1 )
For simplicity, denote by ES , EF , EF , and EL the product of the edge lengths

of the simplices S, F , F , and L respectively. It is clear that ES EL =
EF EF k(tk ) (t0 )k. Consequently
sin() P P 0 EF EF
=
k(tk ) (t0 )k P P 00 ES EL
Vk (t0 , . . . , tk ) Vk2 (t1 , . . . , tk1 )

k ES EL
= .
k 1 Vk1 (t0 , . . . , tk1 ) Vk1 (t1 , . . . , tk )

EF EF
Computing the limit of the right-hand side applying Proposition 2.6.2 we

obtain
112 2. Curves in En
sin((t0 , . . . , tk ))
lim
i:ti t k(tk ) (t0 )k

k1 (t) k3 (t)
k1 (t) (t)
1 1 k3

1!2! k! k! 1!2! (k 2)! (k 2)!

k k1 (t)
= 2 = .
k1 k

k2 (t)
k2 (t)

1
1!2! (k 1)! (k 1)!

The above equation shows that tends to 0 as all the ti s tend to t, therefore
/ sin() tends to 1, so the equation implies the statement.
2.7 Osculating Spheres

According to Corollary 2.6.3, the circumradius of a triangle (t0 )(t1 )(t2 )4
inscribed in a curve of general type tends to 1/|1 (t)| as the parameters
t0 , t1 , t2 tend to t. In this section we show that under some conditions not
only the radii, but also the circles themselves have a limit position. In fact
we prove such a statement for the circumspheres of inscribed simplices of any
given dimension. Then we investigate how to compute centers and radii of
the limit spheres.
The set SPH(Rn ) of all k-spheres has a natural topology induced from the
embedding into Gr1 (hhV ii). This embedding was described in the subsection
Equations of Spheres of Section 1.3.2. According to Exercise 1.4.23, conver-
gence of a sequence of k-spheres can be split into three conditions. Namely,
a sequence S1 , S2 , . . . of k-spheres tends to a ksphere S if and only if tending
with i to infinity the affine subspace aff[Si ] tends to the affine subspace aff[S],
the center of Si tends to the center of S and the radius of Si tends to the
radius of S.
Theorem 2.7.1. Assume that for a smooth parameterized curve : I Rn
and a parameter t I, the first (k+1) derivatives of at t, 0 (t), . . . , (k+1) (t)
are linearly independent. Then t has a neighborhood U such that for any k +2
distinct parameters t0 , . . . , tk+1 from U I, the points (t0 ), . . . , (tk+1 ) are
affinely independent, therefore define a unique k-sphere S(t0 , . . . , tk+1 ).
Furthermore, the limit of the k-sphere S(t0 , . . . , tk+1 ) exists as the parameters
t0 , . . . , tk+1 tend to t.
Proof. The existence of the open neighborhood U was proven as part of
Theorem 2.4.5.
To show convergence, consider the embedding V hhV ii, v 7 v defined by
equation (1.16). Then by Proposition 1.3.38, the equation of S(t0 , . . . , tk+1 )
2.7. Osculating Spheres 113
is
x (t0 ) (tk+1 ) = x (t0 )((t1 ) (t0 )) ((tk+1 ) (t0 )) = 0.
If we apply formula (2.17) for , the equation of the sphere can be rewritten
as
x (t0 ) 1 (t0 , t1 ) k+1 (t0 , . . . , tk+1 ) = 0.
As the limit of higher rth order divided difference of as all the parameters
tend to t exists and equals (r) (t)/r!, we have
lim (t0 ) 1 (t0 , t1 ) k+1 (t0 , . . . , tk+1 )

i:ti t
1
= (t) 0 (t) (k+1) (t) 6= 0.
1! (k + 1)!
Hence the limit of the k-sphere S(t0 , . . . , tk+1 ) exists as the parameters t0 , . . . ,
tk+1 tend to t, and the equation of the limit sphere is
x (t) 0 (t) (k+1) (t) = 0. (2.18)
Definition 2.7.2. The limit of the k-spheres S(t0 , . . . , tk+1 ) as each param-
eter ti tends to t is called the osculating k-sphere of at t.
The equation of the osculating k-sphere at t is (2.18). The position of the

center is not seen immediately from this equation, but it can be obtained
by solving some linear equations (see Exercise 1.3.39), the coefficients in
which are combinations of the dot products of the derivatives of . Thus,
if is smooth, then the center ok (t) of the osculating k-sphere at t is a
smooth function of t on intervals, where the first k + 1 derivatives are linearly
independent.
Assume that : I Rn is a curve of general type in Rn . For k < n 1,
denote by ok : I Rn the center of the osculating k-sphere at t, as a smooth
function of t, and by Rk : I R the radius function of the osculating k-
sphere. We define on1 and Rn1 similarly, but these maps are defined only
on the open subset J of I, where the last curvature function n1 is not 0.
Our goal now is to find some recursive formulae to compute these functions.
We start with some simple geometrical facts.
Since the osculating spheres at t go through the curve point (t),
Rk2 = kok k2 . (2.19)
Proposition 2.7.3. The difference ok ok1 is parallel to tk+1 .

114 2. Curves in En
Proof. Let t I be a fixed point and take k + 2 distinct points t0 , . . . , tk+1

in I. For m = k 1 and k, denote by Sm = Sm (t0 , . . . , tm+1 ) the m-
sphere through the curve points (t0 ), . . . , (tm+1 ), and denote by Om =
Om (t0 , . . . , tm+1 ) its center. Sk1 is a subsphere of Sk , therefore, its center
Ok1 is the orthogonal projection of the center of Sk onto the affine subspace
spanned by Sk1 . This orthogonality must be preserved when we take the
limit as all parameters tend to t. So in the limit we obtain that ok ok1 must
be orthogonal to the osculating k-plane. On the other hand, the difference
vector ok ok1 is a direction vector of the osculating (k+1)-plane. However,
among direction vectors of the osculating (k + 1)-plane only vectors parallel
to tk+1 are orthogonal to the osculating k-plane.
Corollary 2.7.4. There exists smooth functions k : I R for k n 1

and n : J R such that
ok = + 1 t1 + + k+1 tk+1 for k = 0, . . . , n 1. (2.20)
The corollary is proved by a trivial induction on k, starting from the base

case o0 = . Using this form of ok , and equation (2.19) we can express the
radius Rk with the functions i as
Rk2 = 12 + + k+1
2
(2.21)
Proposition 2.7.5. The derivative o0k is orthogonal to the osculating k-

plane.
Proof. Use the same notation as in the proof of the previous proposition.
Keeping the parameters t0 , . . . , tk+1 fixed for a moment, consider the function
F : I R given by
F (t) = k(t) Ok (t0 , . . . , tk+1 )k2 Rk (t0 , . . . , tk+1 )2 .
F is a smooth function on I vanishing at the k +2 distinct points t0 , . . . , tk+1 .

Since by the Lagrange mean value theorem, the derivative of a smooth func-
tion on an interval has a root strictly between any two roots of the function, F 0
has at least k+1 roots inside the interval T = [min0i(k+1) ti , max0i(k+1) ti ].
Going on inductively, the jth derivative F (j) has at least k +2j roots inside
the interval T for j k + 1. Choose a root j T of F (j) for j = 1, . . . , k + 1.
Then
F (j) (j ) = 2h (j) (j ), (j ) Ok (t0 , . . . , tk+1 )i

j1
X j
+ h (s) (j ), (js) (j )i = 0.
s=1
s
2.7. Osculating Spheres 115
Take the limit of these equations as the parameters t0 , . . . , tk+1 tend to t.

Then T shrinks onto t and therefore j tends to t as well, so we get
j1
X j
2h (j) (t), (t) ok (t)i + h (s) (t), (js) (t)i = 0.
s=1
s
This equation is valid for any t. Rearranging the equation we can bring it to
a simpler form
1
h (j) , ok i = (kk2 )(j) for j = 1, . . . , k + 1.
2
Assume that j k end differentiate this equation. We obtain
1
h (j+1) , ok i + h (j) , o0k i = (kk2 )(j+1) .
2
On the other hand, since j + 1 is still less than or equal to k + 1, we have
1
h (j+1) , ok i = (kk2 )(j+1) .
2
The difference of the last two equations gives that o0k is orthogonal to the
first k derivatives of , consequently, it is perpendicular to the osculating
k-plane.
Proposition 2.7.6. The coefficients i can be computed recursively by the
formulae
1 = 0,
1
2 = ,
1
1 k0

k+1 = + k1 k1 .
k v
Proof. A 0-sphere is a pair of points on a straight line, and the center is at the
midpoint. The limit of the center of the 0-sphere {(t0 ), (t1 )} as t0 , t1 t
is (t). Thus
o0 = + 1 t1 = ,
showing 1 = 0.
Differentiating (2.20) for k 1 we obtain
k+1
X
o0k = vt1 + (j0 tj + v j (j tj+1 j1 tj1 )).
j=2
116 2. Curves in En
By the previous proposition, o0k is orthogonal to tk , consequently, the coeffi-

cient of tk on the right-hand side must vanish. When k = 1, the coefficient
of t1 is
v v2 1 = 0,
from which 2 = 1/1 .
If 1 < k n 1 then the coefficient of tk in the decomposition of o0k is
k0 + v (k1 k1 k+1 k ) = 0. (2.22)
Solving this equation for k+1 we obtain the recursive formula of the propo-
sition.
Remark. Observe that equation (2.22) can be rearranged into a form re-
sembling the Frenet equations
1 0
= k k+1 k1 k .
v k
With the help of the proposition, we can compute the location of the centers
of the osculating k-spheres recursively. The initial cases are
1
o1 = + t2 ,
1
1 01
o2 = + t2 t3
1 v 2 21
0 !
01 01

1 2 1
o3 = + t2 t3 + t4 .
1 v 2 21 1 3 v 3 v 2 21
..
.
Proposition 2.7.7. The radius functions Rk of the osculating k-spheres obey
the following recursive rules
R02 = 0,
1
R12 = 2 ,
1
2
R2 Rk0

2
Rk+1 = Rk2 + 2 k2 .
Rk Rk1 v k+1
Proof. The initial cases are obvious. For 1 k n2, differentiate equation
(2.21). Omitting 1 = 0 and dividing by 2 we obtain
k+1
X
Rk Rk0 = j j0 .
j=2
2.8. Plane Curves 117
Express j0 from the recursive equation (2.22). This provides
k+1
X
Rk Rk0 = vj (j+1 j j1 j1 )
j=2
k+1
X
= vj+1 j j vj j1 j1 = vk+2 k+1 k+1 .
j=2
2
If we square this equation and use s+1 = Rs2 Rs1
2
, then we get
(Rk Rk0 )2 = v 2 (Rk+1

2
Rk2 )(Rk2 Rk1
2
)2k+1 .
2
Solving this equation for Rk+1 results the recursive formula of the proposition.
2.8 Plane Curves

This section we deduce some facts on plane curves from the general theory
of curves. A plane curve : I R2 is given by two coordinate functions.
(t) = (x(t), y(t)) t I.
The curve is of general type if the vector 0 is a linearly independent system

of vectors. Since a single vector is linearly independent if and only if it is
non-zero, curves of general type in the plane are the same as regular curves.
From this point on we assume that is regular.
The Frenet vector fields t1 , t2 are denoted by t and n in classical differential
geometry and they are called the (unit) tangent and the (unit) normal vector
fields of the curve. There is only one curvature function of a plane curve
= 1 . The Frenet formulae have the form
t0 = vn,
n0 = vt,
where v = | 0 |.
Let us find explicit formulae for t, n and . Obviously,
1 0 0 1
t= (x , y ) = p 2 (x0 , y 0 ).
v 0
x +y 0 2
The normal vector n is the last vector of the Frenet basis so it is determined
by the condition that (t, n) is a positively oriented orthonormal basis, that is,
in our case, n is obtained from t by a 90 degree rotation in positive direction.
118 2. Curves in En
The right angled rotation in the positive direction takes the vector (a, b) to
the vector (b, a), thus
1 1
n= (y 0 , x0 ) = p 2 (y 0 , x0 ).
v x0 + y 0 2
To express , let us start from the equation
0 = vt.
Differentiating and using the first Frenet formula,
00 = v 0 t + vt0 = v 0 t + v 2 n.
Taking dot product with n and using ht, ni = 0 we get
h 00 , ni = v 2 ,
which gives
x0 y0

det
h 00 , ni x00 y 0 + y 00 x0 x00 y 00
= = = ,
v2 v3 (x0 2 + y 0 2 )3/2
in accordance with Theorem 2.5.14.
2.8.1 Evolute, Involute, Parallel Curves

Since a 1-sphere in a plane is simply a circle, osculating 1-spheres of curve
are called its osculating circles. The osculating circle of a regular plane curve
exists at a point t if and only if (t) 6= 0. Then the center o1 (t) of the
radius R1 (t) of the osculating circle are given by
1 1
o1 (t) = (t) + n(t), R1 (t) = .
(t) |(t)|
Definition 2.8.1. The center of the osculating circle is called the center of
curvature, the radius of the osculating circle is called the radius of curvature
of the given curve at the given point.
The center of the osculating circle can be obtained also as the limit of the
intersection point of two normal lines of .
Theorem 2.8.2. Assume that : I R2 is a smooth regular plane curve,
t I is such that (t) 6= 0. Denote by M (t0 , t1 ) the intersection point of the
normals of taken at t0 and t1 6= t0 . Then M (t0 , t1 ) exists if t0 and t1 are
sufficiently close to t and
lim M (t0 , t1 ) = o1 (t).
t0 ,t1 t
Proof. The intersection point of the normals exist if the corresponding tan-
gents are not parallel. Since t0 (t) = (t)n(t) 6= 0, the unit tangent t rotates
with nonzero angular speed at t, hence no two tangents are parallel in a small
neighborhood of t.
Assume t0 , t1 are close to t and write M (t0 , t1 ) as
M (t0 , t1 ) = (t0 ) + m(t0 , t1 )n(t0 )
using the fact that M (t0 , t1 ) is on the normal at t0 . Since M (t0 , t1 ) is also
on the normal at t1 ,
hM (t0 , t1 ) (t1 ), t(t1 )i = h(t0 ) + m(t0 , t1 )n(t0 ) (t1 ), t(t1 )i = 0.
Solving this equation for m(t0 , t1 ) gives

(t1 ) (t0 )
, t(t1 )
h(t1 ) (t0 ), t(t1 )i t1 t0
m(t0 , t1 ) = = .
hn(t0 ), t(t1 )i n(t0 ) n(t1 )
, t(t1 )
t1 t0
It is easy to calculate the limit of the right-hand side
h 0 (t), t(t)i hv(t)t(t), t(t)i 1
lim m(t0 , t1 ) = 0
= = .
t0 ,t1 t hn (t), t(t)i hv(t)(t)t(t), t(t)i (t)
In conclusion,
1
lim M (t0 , t1 ) = lim (t0 )+m(t0 , t1 )n(t0 ) = (t)+ n(t) = o1 (t).
t0 ,t1 t t0 ,t1 t (t)
Exercise 2.8.3. Let us illuminate a curve by rays of light, parallel to the
normal line of at (t). Denote by F (t1 , t2 ) the intersection point of the
rays reflected from (t1 ) and (t2 ). Show that limt1 ,t2 t F (t1 , t2 ) (t)+
1
t2 (t).
2(t)
Definition 2.8.4. The locus of the centers of curvature of a curve is the

evolute of the curve. The evolute is defined for arcs along which the curvature
is not zero.
If : I R2 is a curve with a nowhere zero curvature function and unit
normal vector field n, then the evolute can be parameterized by the mapping
o1 : I R2 , o1 = + (1/)n.
Exercise 2.8.5. Show that the 2evolute of the ellipse (t) = (a cos t, b sin t)
a b2 2 2
is the affine astroid o1 (t) = a cos t, b a
3
b sin3 t .
120 2. Curves in En
g(t1)
F(t1,t2)
g(t2)
Figure 2.5: Parallel rays reflected from a curve crossing one of the rays or-
thogonally.
evolute
Figure 2.6: Osculating circles and the evolute of a curve
The evolute of a curve was introduced by Christiaan Huygens (1629-1695).

He used the fact that the evolute of a cycloid is a congruent cycloid to modify
the pendulum used in pendulum clocks to increase the accuracy of the clock.
He could also be led to the notion of the evolute in connection with his
investigations on the propagation of wave fronts. If we generate a curvilinear
wave on the surface of calm water (e.g. we drop a wire into it), the wave
starts moving. Geometrically, the shape of the wave front at consecutive
moments of time is described by the parallel curves of the original curve.
Definition 2.8.6. Let be a regular plane curve with normal vector field
n. A parallel curve of is a curve of the form d = + dn, where d R is
a fixed real number.
Experiments on wave fronts show that even if the initial curve is smooth and
regular, singularities may appear on its parallel curves. As the distance d
varies, singularities on d are usually born and disappear in pairs.
Definition 2.8.7. Let be a smooth parameterized curve, t a point of its

parameter interval. We say that (t) (or t) is a singular point(or singular
parameter ) of the curve if 0 (t) = 0.
Exercise 2.8.8. Study the singularities of the parallel curves of an ellipse.

Figure 2.7: Parallel curves of an ellipse and the evolute swept out by the
singular points of them.
Proposition 2.8.9. Singular points of the parallel curves of a regular curve

trace out the evolute of .
Proof. Since d0 = 0 + dn0 = vt vdt = (1 d)vt, singular parameters

are characterized by 1 d(t) = 0. Then the corresponding singular points
d (t) = (t) + (1/(t))n(t) lie on the evolute of the curve. It is also easy to
show that any evolute point is a singular point of a suitable parallel curve.
Proposition 2.8.10. Let : [a, b] R2 be a regular plane curve with mono-

tone non-vanishing curvature , and evolute o1 = + (1/)n.
If 0 (t) 6= 0, then o1 is regular at t and its tangent line at t coincides

with the normal line of at t.
The length of the arc of the evolute between o1 (t1 ) and o1 (t2 ), t1 < t2 ,
is the difference of the radii of curvatures |1/(t1 ) 1/(t2 )|.
Proof. The speed vector of o1 is
o1 0 = 0 + (1/)0 n + (1/)n0 = vt + (1/)0 n (1/)vt = (1/)0 n.
This equation shows that if 0 (t) 6= 0, then o1 is regular at t and its tangent
at t is parallel to the normal n(t) of . This proves the first part of the
122 2. Curves in En
proposition. As for the length of the evolute, it is equal to the integral

Z t2 Z t2 Z t2
|o1 0 ( )|d = |(1/)0 ( )n( )|d = |(1/)0 ( )|d
t1 t t
Z1 t2 1
0

= (1/) ( )d = |1/(t1 ) 1/(t2 )|.
t1
At the third equality we used that 1/ is monotone, thus either (1/)0 0

or (1/)0 0 everywhere.
The above proposition gives a method to reconstruct the curve from its
evolute. Suppose for simplicity that > 0 and 0 0. Take a thread of
length 1/(a) and fix one of its ends to o1 (a). Then pulling the other end
of the thread from the initial position (a) wrap the thread tightly onto the
curve o1 . By Proposition 2.8.10, the moving end of the thread will draw .
involute
evolute
Figure 2.8: Relation of the evolute and involute constructions.
The thread construction gives involutes of a curve.

Definition 2.8.11. Let : I R2 be a regular curve with unit tangent
vector field t, a I. An involute of the curve is a curve
R t of the form
= + (l s)t, where l is a given real number, s(t) = a | 0 ( )|d is the
signed length of the arc of between (a) and (t).
A curve has many involutes corresponding to the different choices of the
initial point a I and the length l of the thread.
Corollary 2.8.12. A curve satisfying the conditions of Proposition 2.8.10 is
an involute of its evolute.
Proposition 2.8.13. Let be a unit speed curve with > 0, (s) = (s) +
(l s)t(s) an involute of it such that l is greater than the length of . Then
the evolute of is .
Proof. We have
0 (s) = t(s) t(s) + (l s)(s)n(s) = (l s)(s)n(s),

00 (s) = [(l s)(s)]0 n(s) (l s)2 (s)t(s).
The first equation implies that the Frenet frame t, n of is related to that
of by t = n, n = t. Computing the curvature of ,
h 00 (s), n(s)i (l s)2 3 (s) 1

(s) = 0 3
= 3 3
= .
| (s)| (l s) (s) (l s)
Thus, the evolute of is + (1/)n = + (l s)t (l s)t = .

We formulate some further results on the evolute and involute as exercises.
Exercise 2.8.14. (a) Suppose that the regular curves 1 and 2 have regular
evolutes. Show that 1 and 2 are parallel if and only if their evolutes are
the same.
(b) Show that if two involutes of a regular curve are regular, then they are
parallel.
Exercise 2.8.15. The curve cardioid is the trajectory of a peripheral point
of a circle rolling about a fixed circle of the same radius.
Find a smooth parameterization of the cardioid.
Compute its length.
Show that its evolute is also a cardioid.
Figure 2.9: Generation of the cardioid.
Exercise 2.8.16. The chain curve is the graph of the hyperbolic cosine
function
ex + ex
cosh(x) = .
2
124 2. Curves in En
Figure 2.10: Construction of the tractrix as an involute of the chain curve.
Determine the involute of the chain curve through the point (0, 1). (This
curve is called tractrix.)
Let the tangent of the tractrix at P intersect the x-axis at Q. Show that
the segment P Q has unit length.
Exercise 2.8.17. Let be a regular plane curve for which the curvature
function and its derivative are positive. Show that for any t1 < t2 from
the parameter domain of the osculating circle of at (t1 ) contains the
osculating circle at (t2 ).
2.8.2 The Rotation Number Theorem

Let : [a, b] R2 be a unit speed curve. The direction angle (s) of the
tangent t(s) is determined only up to an integer multiple of 2, however one
can see easily the existence of a differentiable function : [a, b] R such
that (s) is a direction angle of t(s) for all s [a, b] (see Proposition 2.8.27
below). Then
t(s) = (cos (s), sin (s)).
Differentiating with respect to s gives t0 = 0 ( sin , cos ). If we compare
this equality with Frenet equations, we see immediately that
= 0 ,
i.e. the curvature is the derivative of the direction angle (angular speed) of
the tangent vector with respect to the arc length.
Definition 2.8.18. The total curvature of a curve is the integral of its cur-
vature function with respect to arc length
Z b
(s)ds.
a
By the relation = 0 , the total curvature of a curve is (b) (a), thus it

measures the rotation made by the tangent vector during the motion along
the curve.
Definition 2.8.19. A curve : [a, b] Rn is a smooth
closed curve, if there
exists a smooth mapping : R Rn such that [a,b] = , and is periodic
with period b a, that is (t + b a) (t).
In other words, a smooth curve is smoothly closed if it returns to its initial
point and the periodic repetition of the same motion crosses the initial point
smoothly. Since the tangent vector of a smooth closed planar curve is the
same at the endpoints (a) and (b), the direction angles (b) and (a) differ
in an integer multiple of 2.
Definition 2.8.20. The integer ((b) (a))/(2) is called the rotation
number of the smooth closed planar curve .
Exercise 2.8.21. Construct a smooth closed curve with an arbitrarily given
rotation number k Z.
Solving the exercise we can observe that all our efforts to construct a curve
with rotation number neither 1 nor 1 having no self-intersection are in vain.
The reason for this is the famous Umlaufsatz (rotation number theorem).
Definition 2.8.22. A curve : [a, b] Rn is a simple closed curve if it is
closed and (t) = (t0 ) holds if and only if t = t0 or {t, t0 } = {a, b}.
Theorem 2.8.23 (Umlaufsatz). The rotation number of a planar smooth
simple closed curve in the plane is equal to 1, or equivalently, the total
curvature of such a curve is 2.
The proof of the theorem will use some concepts borrowed from topology.
Definition 2.8.24. Let S 1 denote the unit circle {x R2 : kxk = 1}. The
mapping : R S 1 , (t) = (cos t, sin t) is called the universal covering map
of S 1 . Given a continuous mapping : X S 1 from a topological space X
into S 1 , we say that has a lift to R if there exists a continuous mapping
: X R such that = .
Exercise 2.8.25. Construct a continuous mapping from a topological space
into the circle which has no lift to R.
Lemma 2.8.26. Suppose that the image of the mapping : X S 1 does
not cover the point (cos , sin ) S 1 and that the restriction of onto a
subspace Y X has got a lift such that (Y ) is contained in an interval
of the form ( + 2k, + 2(k + 1)), k Z. Then can be extended to a lift
of . If furthermore Y is non-empty, X is path connected (i.e. any two
points of X can be connected by a continuous curve lying in X) then the lift
is unique, and maps X into the interval ( + 2k, + 2(k + 1)).
126 2. Curves in En
Proof. The restriction of onto ( + 2k, + 2(k + 1)) is a homeomorphism

between ( + 2k, + 2(k + 1)) and S 1 \ {(cos , sin )}. Thus can be
defined as
= ( (+2k,+2(k+1)) )1 .

If X is path connected, then so is its image under a continuous lift . Conse-

quently (X) must be contained in one of the intervals (+2k, +2(k+1)).
If Y
is non-empty then k is uniquely determined, hence must have the form
( (+2k,+2(k+1)) )1 .
Proposition 2.8.27. Any continuous mapping : [a, b] S 1 from an in-

terval into the circle has got a lift.
Proof. Choose a partition a = t0 < t1 < < tk = b of the interval [a, b] fine
enough to insure that the restriction of onto [ti , ti+1 ] does not cover the
whole circle. This is possible because of the uniform continuity of . Then
using Lemma 2.8.26 we can define a lift of step by step, extending
recursively to [a, ti ], i = 1, 2, . . . , k.
Remark. If is differentiable, then so is its lift.
Proposition 2.8.28. Any continuous mapping : T S 1 from a rectangle
T = [a, b] [c, d] into the circle has got a lift to R.
Proof. Divide the rectangle into n n small rectangles so that the image of
any of the small rectangles does not cover the circle. Then we may define
a lift of recursively, applying at each step the lemma above. can be
defined first on the small rectangles of the first row going from left to right
then on the rectangles of the second row, etc.
1 2 3 ... n
n+1 n+2 n+3 ... 2n
Proof of the Umlaufsatz. Let us choose a point p on the regular simple

closed curve in such a way that the curve is lying on one side of the tangent
at p and parameterize the curve by arc length starting from p. Denoting this
parameterization also by : [0, l] R2 , we define a mapping : [0, l][0, l]
S 1 by
0 (0),

if {t1 , t2 } = {l, 0},

(t1 ) (t2 )

if t1 > t2 and (t1 , t2 ) 6= (l, 0),

|(t ) (t )| ,

1 2
(t1 , t2 ) =

0 (t1 ), if t1 = t2 ,

(t2 ) (t1 )
, if t2 > t1 and (t1 , t2 ) 6= (0, l).

|(t2 ) (t1 )|

Figure 2.11: The geometrical meaning of .
It is easy to see that is continuous, so it has a continuous lift . If the

function : [0, l] R is defined by (t) = (t, t), then (t) is a direction
angle of the speed vector t(t) of , thus (l) (0) is 2 times the rotation
number of . Consider the functions (t) = (0, t) and (t) = (t, l). (t) is a
direction angle of the unit vector ((t)p)/|(t)p|, (t) is a direction angle
of its opposite. Thus and differ only in a constant of the form (2k + 1).
Since the vectors ((t) p)/|(t) p| point in a half-plane bounded by the
tangent of at p, the image of is contained in an open interval of length 2
(see Lemma 2.8.26). Thus (l)(0), which has obviously the form (2m+1)
for some m Z, must be equal to . Hence, we conclude that
(l) (0) = (l, l) (0, 0) = ((l, l) (0, l)) + ((0, l) (0, 0))
= (l) (0) + (l) (0) = 2((l) (0)) = 2.
Remark. With some more work but using essentially the same idea, one can
generalize the Umlaufsatz for piecewise smooth closed simple curves. The
generalization says that for a simple closed polygon with smooth curvilinear
edges, the sum of oriented external angles plus the sum of the total curvatures
of the edges equals 2.
Exercise 2.8.29. Let be a simple regular closed curve of length l with
curvature function . Choose a real number d such that 1 d. How long
is the parallel curve d = + dn?
2.8.3 Convex Curves

Definition 2.8.30. A simple closed plane curve is convex, if for any point
p = (t), the curve lies on one side of the tangent line of at p. In other
words the function h(t) p, n(t)i must be 0 or 0 for all t.
Exercise 2.8.31. Show that a simple closed curve is convex if and only
if every arc of the curve lies on one side of the straight line through the
endpoints of the arc.
128 2. Curves in En
Convex curves can be characterized with the help of the curvature function.
Theorem 2.8.32. A simple closed curve is convex if and only if 0 or
0 everywhere along the curve.
Proof. Assume first that is a naturally parameterized convex curve. Let
(t) be a continuous direction angle for the tangent t(t). As we know, 0 = ,
thus, it suffices to show that is a weakly monotone function. This follows if
we show that if takes the same value at two different parameters t1 , t2 , then
is constant on the interval [t1 , t2 ]. The rotation number of a simple curve
is 1, hence the image of t covers the whole unit circle. As a consequence,
we can find a point at which
t(t3 ) = t(t1 ) = t(t2 ).
If the tangent lines at t1 , t2 , t3 were different, then one of them would be

between the others and this tangent would have points of the curve on both
sides. This contradicts convexity, hence two of these tangents say the tangents
at p = (ti ) and q = (tj ) coincide.
e
s
p x q
Figure 2.12
We claim that the segment pq is an arc of . It is enough to prove that this

segment is in the image of . Assume to the contrary that a point x pq is
not covered by . Drawing a line e 6= pq through x, we can find at least two
intersection points r and s of e and the curve, since e separates p and q and
has two essentially disjoint arcs connecting p to q. Since pq is a tangent
of , the points r and s must lie on the same side of it. As a consequence,
we get that one of the triangles pqr and pqs, say the first one is inside the
other. However, this leads to a contradiction, since for such a configuration
the tangent through r necessarily separates two vertices of the triangle pqs,
which lie on the curve.
If is defined on the interval [a, b], then (a) = (b) is either on the segment
pq or not. The first case is not possible, because then would be constant
on the intervals [a, t1 ] and [t2 , b], yielding
(a) = (t1 ) = (t2 ) = (b)
and
rotation number = ((b) (a))/2 = 0.
In the second case is constant on the interval [t1 , t2 ], as we wanted to show.
Now we prove the converse. Assume that is a simple closed curve with
0 everywhere and assume to the contrary that is not convex (the case
0 can be treated analogously). Then we can find a point p = (t1 ),
such that the tangent at p has curve points on both of its sides. Let us find
on each side a curve point, say q = (t2 ) and r = (t3 ) respectively, lying
at maximal distance from the tangent at p. Then the tangents at p, q and
r are different and parallel. Since the unit tangent vectors t(ti ), i = 1, 2, 3
have parallel directions, two of them, say t(ti ) and t(tj ) must be equal. The
points a = (ti ) and b = (tj ) divide the curve into two arcs. Denoting
by K1 and K2 the total curvatures of these arcs, we deduce that these total
curvatures have the form K1 = 2k1 , K2 = 2k2 , where k1 , k2 Z, since
the unit tangents at the ends of the arcs are equal. On the other hand, we
have k1 + k2 = 1 by the Umlaufsatz and k1 0, k2 0 by the assumption
0. This is possible only if one of the total curvatures K1 or K2 is equal
to zero. Since 0, this means that = 0 along one of the arcs between a
and b. But then this arc would be a straight line segment, implying that the
tangents at a and b coincide. The contradiction proves the theorem.
2.8.4 The Four Vertex Theorem

Definition 2.8.33. A point (t) of a regular plane curve is called a vertex
if 0 (t) = 0.
Vertices of a curve correspond to the singular points of the evolute. By
compactness, the curvature function of a closed curve attains its maximum
and minimum somewhere, hence every closed curve has at least two vertices.
Exercise 2.8.34. Find a parameterization of Bernoullis lemniscate
{P R2 | P A P B = 1/4(AB)2 },
where A 6= B are given points in the plane and show that it is a closed curve
with exactly two vertices. Determine the rotation number of the lemniscate.
Exercise 2.8.35. Find the points on the ellipse (t) = (a cos t, b sin t) at
which the curvature is minimal or maximal (a > b > 0).
130 2. Curves in En
Theorem 2.8.36 (Four Vertex Theorem). A convex closed curve has at least
4 vertices.
This result is sharp. For example, an ellipse has exactly four vertices.
Proof. Local maxima and minima of the curvature function yield vertices.
One can always find a local minimum on an arc bounded by two local max-
ima, hence if we have two local maxima or minima of the curvature then we
must have at least four vertices. Thus we have to exclude the case when the
curvature function has one absolute maximum at A and one absolute mini-
mum at B, and strictly monotone on the arcs bounded by A and B. In such
a case, choose a coordinate system with origin at A and x-axis AB.
The arcs of the curve bounded by A and B do not cut the x-axis at points
other than A and B. Indeed, if there were a further intersection point C,
then the curve would be split into three arcs by A,B and C in such a way
that on each arc we could find a point at which the tangent to the curve is
parallel to the straight line ABC. If the three tangents at these points were
different then the one in the middle would cut the curve apart contradicting
convexity, if two of the tangents coincided then we could find a straight line
segment contained in the curve yielding an infinite number of vertices.
If the two arcs bounded by A and B lied on the same side of AB then the
line AB would be a common tangent of the curve at A and B. In this case
the segment AB would be contained in the curve, yielding an infinite number
of vertices as before.
We conclude that for a suitable orientation of the y-axis, y(t)0 (t) 0 for
every t [a, b], where (t) = (x(t), y(t)), t [a, b] is a unit speed parameter-
ization of the curve. Hence we get
Z b
y(t)0 (t)dt > 0.
a
vertices:
A B
Bernoulli's lemniscate
ellipse
Figure 2.13: Vertices of Bernoullis lemniscate and that of an ellipse.

2.9. Curves in R3 131
Integrating by parts,
Z b Z b Z b
b
y(t)0 (t)dt = [y(t)(t)]a y 0 (t)(t)dt = y 0 (t)(t)dt.
a a a
The unit tangent vector field of the curve is t = (x0 , y 0 ), the unit normal
vector field is n = (y 0 , x0 ), hence by the first Frenet formula,
x00 = y 0 .
Integrating,
Z b Z b
0 b
y (t)(t)dt = x00 (t)dt = [x0 (t)]a = 0.
a a
This is a contradiction since a positive number can not be equal to 0.
2.9 Curves in R3
A 3-dimensional curve is a curve of general type if its first two derivatives are
not parallel. From now on we shall suppose that the curves under consider-
ation are all of general type.
The distinguished Frenet frame vector fields t1 , t2 and t3 of a 3-dimensional
curve are denoted in classical differential geometry by t, n and b and they are
called the (unit) tangent, the principal normal and the binormal vector fields
of the curve respectively. These vector fields define a coordinate system at
each point of the curve. The coordinate planes of this coordinate system are
given the following names. We are already familiar with the plane that goes
through a given curve point and spanned by the directions of the tangent and
the principal normal. It is the osculating plane of the curve. The plane that is
spanned by the principal normal and the binormal is the plane that contains
all straight lines orthogonally intersecting the curve at the given point.
For this obvious reason, this plane is called the normal plane of the curve.
The third coordinate plane, that is the plane spanned by the tangent and
the binormal directions is the rectifying plane of the curve. The reason for
this naming will become clear later. As we know from the general theory, a
3-dimensional curve of general type has two curvature functions 1 , which is
always positive and 2 , which may have any sign. The first curvature 1 is
denoted by the classics simply by and it is referred to as the curvature of
the curve while the second curvature 2 is called the torsion of the curve and
is denoted by .
132 2. Curves in En
Using the classical notation, Frenets formulae for a space curve can be writ-
ten as follows.
t0 = vn
n0 = vt + v b
b0 = v n.
Let us find explicit formulae for the computation of these vectors and curva-
tures in an economical way. Assume that : [a, b] R3 is a curve of general
type. The unit tangent vector field t can be obtained by normalizing the
speed vector 0
0
t= 0 .
k k
To obtain the principal normal n we can use the general method based on
the GramSchmidt orthogonalization process
00 h 00 , tit k 0 k2 00 h 00 , 0 i 0
n= = k 0 k2 00 h 00 , 0 i 0 ,

k 00 h 00 , titk
and after this we can compute the binormal as a cross product of t and n
b = t n.
In certain cases, it can be convenient to calculate the binormal first. The

binormal vector is the unit normal vector of the osculating plane, for which
(t, n, b) is positively oriented. The osculating plane is spanned by the first
two derivatives 0 , 00 of , furthermore the pair (t, n) defines the same ori-
entation of the osculating plane as the pair ( 0 , 00 ), thus the basis ( 0 , 00 , b)
is positively oriented. Hence,
0 00
b= .
k 0 00 k
Having computed b, n can be obtained as
( 0 00 ) 0
n=bt= .
k 0 00 )k k 0 k
The formulae obtained for the curvature and the torsion in Theorem 2.5.14
can be rewritten using standard vector operations. The first curvature is
k 0 00 k k 0 00 k
= = ,
v3 v3
while the torsion is

0 0

det 00 k 0 k det 00
000 000
h 0 00 , 000 i
= = = .
vk 0 00 k2 k 0 00 k2 k 0 00 k2
Recall that the numerator of this fraction is the determinant of the matrix
the rows of which are 0 , 00 , 000 and geometrically it is the signed volume
of the parallelepiped spanned by 0 , 00 , 000 , where the sign is positive if and
only if ( 0 , 00 , 000 ) is a positively oriented basis.
2.9.1 Orthogonal Projections onto Planes Spanned by

Frenet Vectors
Study the shape of the orthogonal projections of a curve onto the planes
spanned by the vectors of the distinguished Frenet frame. For simplicity,
suppose that the curve is parameterized by arc length and examine the
curve around (0). Since v 1, equations (2.7)(2.9) can be reduced to the
form
0 = t
00 = t0 = n
000 = 0 n + n0 = 0 n + (t + b) = 2 t + 0 n + b
We can approximate the curve around (0) by its Taylor expansion.

00 (0) 2 000 (0) 3
(t) = (0) + 0 (0)t + t + t + o(t3 ).
2 6
Recall that the little oh notation o(t3 ) is used in the following sense. If f, g,
and h are functions defined around 0, then we write
f (t) = g(t) + o(h(t))

f (t) g(t)
if tends to zero as t tends to 0. Expressing the derivatives of
h(t)
with the help of Frenet vectors we get
(t) (0) =
t3 t2 t3
= t 2 (0) t(0) + (0) + 0 (0) n(0)
6 2 6
t3
+ (0) (0) b(0) + o(t3 ).
6
134 2. Curves in En
b(0) b(0) b(0)
n(0) n(0) n(0)
t(0) t(0) t(0)
Projections of a space curve onto the ...

osculating plane normal plane rectifying plane
n(0) b(0) b(0)
t(0) n(0) t(0)
Figure 2.14
Looking at this expansion we conclude that the projection of the curve on the
osculating plane is well approximated by the parabola tt(0) + ((0)t2 /2)n(0).
Observe that the curvature of this parabola at t = 0 is (0).
The projection onto the normal plane has approximately the same shape as
the semicubical parabola ((0)t2 /2)n(0) + ((0) (0)t3 /6)n(0), in particular,
it has a so called cusp singularity at t = 0.
Finally, the projection onto the rectifying plane has the Taylor expansion
t3
(0) + t + o(t) t(0) + (0) (0) + o(t3 ) b(0).

6
so its shape is like the graph of a cubic function. It is easy to see, that the
curvature of this projection is 0 at t = 0, thus it is almost straight around
the origin. That is the reason why the rectifying plane was given just this
name: projection of the curve onto the rectifying plane straightens the curve
and rectifying means straightening.
Exercise 2.9.1. Given a unit speed curve of general type in R3 with distin-
guished Frenet frame t1 , t2 , t3 , find a vector field along the curve such that
t0i = ti holds for i = 1, 2, 3. ( is called the Darboux vector field of the
curve.)
Exercise 2.9.2. Suppose that the osculating planes of a curve of general type
in R3 have a point in common. Show that the curve is a plane curve. (In
particular, if a point is moving in a central force field, its motion is planar.)
Exercise 2.9.3. Suppose that the normal planes of a regular curve in R3 go

through a fixed point O. Show that the curve lies on a sphere centered at O.
Exercise 2.9.4. Show that the rectifying planes of a curve of general type
in R3 can never share a point in common.
2.9.2 Fenchels Theorem

Though a regular space curve is not necessarily a curve of general type, the
usual formulae allow us to define the unit tangent vector field and the (first)
curvature function for a regular space curve as well.
Definition 2.9.5. The unit tangent vector field of a regular space curve
: I R3 is the vector field t = 0 /v, where v = k 0 k. The curvature of is
the function
kt0 k k 0 00 k
= = .
v v3
The total curvature of is the integral
Z
v(t)(t)dt.
I
We saw that the total curvature of a smooth closed plane curve can take only
isolated values, the integer multiples of 2. This is related to the fact that the
curvature of a planar curve is a signed function. When we try to increase the
total curvature of a plane curve making it wavier by continuously deforming it
within the class of closed regular plane curves, we shall not succeed, because
the total curvature of arcs where the curvature is negative keeps equilibrium
with the total curvature of the arcs where the curvature is negative.
In contrast to the planar case, curvature of a regular space curve is nonnega-
tive even if the curve is contained in a plane. In the latter case, the (spatial)
curvature of the curve is the absolute value of the signed planar curvature.
For this reason, the total curvature of a regular space curve can be increased
continuously and can be made arbitrarily large by making the curve wavier.
Hence for regular closed smooth space curves, the interesting question is, how
small the total curvature can be. This question is answered by a theorem of
Moritz Werner Fenchel (1905-1988).
136 2. Curves in En
Theorem 2.9.6 (Fenchel). The total curvature of a regular smooth closed

space curve is at least 2. Equality holds if and only if the curve is a convex
plane curve.
Remark. We do not assume that is simple.
The proof of the theorem is based on the following Lemma from spherical
geometry.
Lemma 2.9.7. If a closed curve on the unit sphere S2 centered at the origin
0 has length l 2, then the curve is contained in a closed hemisphere. If
the length is strictly less than 2, then the curve is contained in an open
hemisphere.
Proof. Recall that the spherical distance of two points on the sphere S2 is
defined to be the angle between their position vectors from the center. It is a
corollary of the spherical triangle inequality, that the length of any spherical
curve connecting the points P and Q is greater than or equal to the spherical
distance of P and Q, and equality holds if and only if the curve is a shorter
arc of a great circle passing through P and Q. This arc is unique if P and
Q are not antipodal, but when P and Q are antipodal, there are infinitely
many semicircles of minimal length connect them.
Consider a closed spherical curve of length l 2. Choose two points P
and Q on , which divide into two arcs 1 , 2 of equal length l/2 .
If P and Q are antipodal points, then their spherical distance is , so the two
arcs of connecting them must be two semicircles. In this case the length of
is exactly 2 and is contained in one of the closed hemispheres bounded
by the great circle containing the semicircle 1 . We infer that the Lemma
holds in this case.
Figure 2.15
Assume now that P and Q are not antipodal. Then let F be the midpoint of
the shortest great circle arc connecting them. 0F is the bisector of the angle
P 0Q and reflecting P in the line 0F gives Q.
Figure 2.16: The tantrix of a parameterized curve .
Denote by C the great circle, the plane of which is orthogonal to 0F .

We prove that if l < 2, then does not intersect C. Assume to the contrary,
that there is an intersection point R. We can assume without loss of general-
ity, that R is on the arc 1 and it cuts 1 into an arc 01 connecting P to R,
and an arc 001 connecting R to Q. Let denote the reflection in the plane
aff[C]. It is easy to see that (Q) is antipodal to P , and concatenation of the
arcs 01 and (001 ) connect P to (Q). However, this is a contradiction, since
the total length of these two arcs is l/2 < , but the spherical distance of P
and (Q) is . The contradiction proves that if l < 2, then is contained
in one of the open hemispheres bounded by C.
If we allow l = 2, then the arc 1 and 2 may intersect C. Nevertheless, if,
say 1 intersects C in a point R, as above, then 01 and (001 ) must be two
arcs of a semicircle connecting P and (Q). This means that even in this
case the arc 1 is in the closed hemisphere bounded by C and containing P .
The same is true for 2 .
Proof of Fenchels Theorem. Let : [a, b] R3 be a regular smooth closed
curve, t : [a, b] S2 be its unit tangent vector field. t parameterizes a closed
spherical curve . is called the tangent indicatrix or shortly the tantrix of
. The length of equals
Z b Z b
0
kt (t)kdt = v(t)(t)dt,
a a
that is the total curvature of .

Apply the Lemma to the curve . The parameterization t is not necessarily
injective, but this is not important as the proof of the Lemma works also for
curves with a non-injective parameterization.
If we assume to the contrary that the total curvature of is less than 2,
then is a closed spherical curve of length < 2, so it can be covered by an
138 2. Curves in En
open hemisphere. Let u S2 be the spherical center of the open hemisphere

containing . Then u encloses an acute angle with all the vectors t(t), thus,
hu, t(t)i > 0 for all t [a, b]. From this we obtain the inequality
hu, i0 = v hu, ti > 0,
which implies that hu, i is a strictly increasing function. However, this

contradicts the assumption that is closed, as this leads to
hu, (a)i = hu, (b)i.
The contradiction proves that the total curvature of is at least 2.

Consider the case of equality. Applying the lemma for the case when has
total curvature 2, we obtain a unit vector u satisfying u, ti 0. Then the
above argument provides that the function hu, i is weakly increasing and
takes the same values at the endpoints of the interval [a, b]. Thus, hu, i is
constant, which means that lies in a plane orthogonal to u.
We can also see that is simple. Indeed, if had a self-intersection point
(t1 ) = (t2 ), a < t1 < t2 b then we could split into two arcs 1 = |[t1 ,t2 ]
and 2 = |[t2 ,t1 +(ba)] , where : R R3 is the periodic extension of . Then
the points P = t(t1 ) and Q = t(t2 ) would cut into two corresponding arcs
1 and 2 . We show that each of these arcs must have length strictly larger
than . This would contradict the fact that has total length 2.
Assume to the contrary that, say 1 has length . Then the great circle C
that covers 1 has a closed semicircle with midpoint v C which also covers
1 . Following the above line of reasoning, this implies that hv, i must be
weakly increasing on [t1 , t2 ] and take the same values at t1 and t2 , therefore,
hv, i must be constant on [t1 , t2 ]. Differentiating this equation we infer that
t must stay orthogonal to v on [t1 , t2 ]. But t is also orthogonal to the unit
vector u v, consequently, t must be constant on [t1 , t2 ], equal to either
u v or u v. However, in this case, h, u vi is strictly increasing or
decreasing on [t1 , t2 ], which contradict the assumption (t1 ) = (t2 ).
Let us orient the plane of and introduce the signed planar curvature of ,
which we denote by to distinguish it from = | |. Since is simple,
the Umlaufsatz gives
Z
Z b Z b b

v(t)| (t)|dt = v(t)(t)dt = 2 = v(t) (t)dt .

a a a
This equality can be true only if 0 or 0 everywhere, and this is

equivalent to the convexity of the curve.
2.9.3 The Fary-Milnor Theorem

Fenchel proved his theorem in 1929. In 1947 K. Borsuk generalized Fenchels
theorem for curves in the n-dimensional space, and formulated the conjecture
that the total curvature of a nontrivial knot in the three dimensional space
must exceed 4. The conjecture with a weak inequality 4 was proved by
I. Fary in 1949. Independently, in the same year J. W. Milnor published an
analogous, but a bit more technical proof in which the strict inequality was
verified.
A simple closed curve gives an injective continuous map from the circle into
the space. Such a map is also called a knot. A knot is said to be trivial , if
the map from the circle extends to a continuous injective map from the disc
bounded by the circle into R3 .
Theorem 2.9.8 (I. Fary, J.W. Milnor). The total curvature of a nontrivial
knot in R3 is greater than 4.
Below we prove only the week inequality 4.
Proof of the weak form. Assume that the total curvature of a simple regular
closed smooth curve is less than 4. We want to show that the curve is a
trivial knot. The tantrix of is a closed spherical curve of length < 4.
Let us apply the spherical Crofton formula for . For a unit vector u S2 ,
let m(u) denote the number of intersection points of with the great circle
contained in the plane orthogonal to u, counted with multiplicity
m(u) = #{t [a, b] | hu, t(t)i = 0}.
Then by the Crofton formula, the length of can be given as an integral

Z
1
l = m(u)du.
4 S2
The surface area of S2 is 4, and the value of the integral on the right-hand
side must be less than 16, the subset of those points of S2 at which the value
of m is at most 3 must have positive measure. In particular, we can find a
vector u S2 such that the dot product hu, t(t)i vanishes at no more than 3
points.
We can think of u as a vertical vector pointing upward, and hu, (t)i as the
height of the point (t). The height of (t) increases strictly on intervals
where hu, t(t)i > 0 and decreases strictly where hu, t(t)i < 0. The roots of
hu, ti divide the image of into at most 3 simple arcs along which the height
changes in a strictly monotonous way. There must exist both an arc along
which the height increases and another one where it decreases, otherwise the
curve could not return to its initial point. If there are three arcs along which
140 2. Curves in En
the height is a strictly monotone function then two of the arcs can be joined so
that the height will strictly monotone also on the union. We can summarize
the information we got as follows. The curve traced out by can be split
into two simple arcs joining the unique highest and the unique lowest point
of the trace. As we move along any of the two arcs the height changes in a
strictly monotonous way.
Take any horizontal plane between the highest and lowest points of the trace
of . Such a plane intersects both arcs at a unique point. Connect these two
points by a segment. The union of all such horizontal segments connecting
two points of im fill a topologically embedded disc in R3 , the boundary of
which is the trace of . We conclude that is a trivial knot.
Chapter 3
Hypersurfaces in Rn
3.1 General Theory

3.1.1 Definition of a Parameterized Hypersurface
Definition 3.1.1. A smooth parameterized hypersurface in Rn is a differ-
entiable mapping r : Rn from an open domain of Rn1 into the
n-dimensional space.
Smooth curves on a parameterized hypersurface are curves of the form (t) =

r(u(t)), where the mapping t 7 u(t) is a smooth curve lying in the parameter
domain . Curves of the form t 7 r(u1 , . . . , ui1 , t, ui+1 , . . . , un1 ), where
u1 , . . . , ui1 , ui+1 , . . . , un1 are fixed numbers are called the parameter lines
or coordinate lines on the hypersurface. The speed vectors of the parameter
lines t 7 r(u1 , . . . , ui1 , t, ui+1 , . . . , un1 ), which are just the partial deriva-
tives of the mapping r with respect to the i-th variable, will be denoted by
ri (u1 , . . . , ui1 , t, ui+1 , . . . , un1 ).
Since we shall often work with formulae containing partial derivatives of
a function, it is convenient to introduce the shorthand convention that we
shall denote the partial derivative of a multivariable function F with respect
to its i-th variable by Fi . In general, the higher order partial derivative
kF
of F will be denoted by Fi1 ...ik . If there is a danger of confusion
ui1 . . . uik
with ordinary lower indices, the lower indices of the function will be separated
from the indices of variables with respect to which we have to take the partial
k Fj1 ...jr
derivative by a comma. Thus, will be denoted by Fj1 ...jr ,i1 ...ik .
ui1 . . . uik
Similar conventions will be applied to functions the variables of which are
denoted by fixed alphabetical symbols. For example, the partial derivative
141
142 3. Hypersurfaces in Rn
TM
p
r
rn N 1 M
1
p
r
Rn
Rn1
Figure 3.1
of a function F of the variables u, v, w, . . . with respect to the variable u

will be denoted shortly by Fu . We also use this abbreviation for higher order
partial derivatives. E.g., Fuv is a short form of u v F . For conventions about
denoting partial derivatives see also the remarks following Definition 1.5.6.
Definition 3.1.2. A parameterized hypersurface is regular if the vectors
r1 (u), . . . , rn1 (u) are linearly independent for any u . In this case we
also say that r is an immersion of the domain into Rn .
Regularity does not imply that r is injective. However, if r : Rn is a
regular parameterized hypersurface,
then each point u has a neighbor-
hood U such that rU maps U bijectively onto r(U ). Thus, studying a
regular parameterized hypersurface locally (i.e. in a small neighborhood of a
point) we may assume without loss of generality that r is injective. We shall
usually denote the image of r by M .
Definition 3.1.3. The affine tangent space of a regular parameterized hyper-
surface at the point p = r(u) M is the hyperplane through r(u) spanned
by the direction vectors r1 (u), . . . , rn1 (u). The (linear) tangent space of
M at p is the linear space Tp M of the direction vectors of the affine tan-
gent space. The unit normal vector of the hypersurface at the point r(u) is
defined to be the unit normal vector N(u) of the tangent plane, for which
r1 (u), . . . , rn1 (u), N(u) is a positively oriented basis of Rn .
For parameterized surfaces in R3 , the unit normal vector field can be calcu-
lated with the help of cross product
r1 (u1 , u2 ) r2 (u1 , u2 )
N(u1 , u2 ) = .
|r1 (u1 , u2 ) r2 (u1 , u2 )|
3.1. General Theory 143
To get a similar formula in higher dimensions, we need a suitable generaliza-

tion of the cross product.
Let ri = (ri1 , . . . , rin ) Rn , i = 1, 2, . . . , n 1, be n-dimensional vec-
tors, (e1 , . . . , en ) be the standard basis of Rn . Then according to Propo-
sition 1.2.78, the generalized cross product (r1 rn1 ) of the vectors
r1 , . . . , rn1 can be computed by the equality
1
. . . r1n

r1
. . . . . . . . . . . . . . . .
(r1 rn1 ) = det 1
rn1 n .

. . . rn1
e1 . . . en
The vector r1 rn1 is orthogonal to r1 , . . . , rn1 , it is different from 0
if and only if r1 , . . . , rn1 are linearly independent, and finally, r1 , . . . , rn1 ,
(1)n1 r1 rn1 is a positively oriented basis of Rn .
Consequently, for regular hypersurfaces, we have
(r1 rn1 )
N(u) = .
kr1 rn1 k
3.1.2 Curvature of Curves on a Hypersurface

Now we shall study a problem which connects curve theory to surface theory.
If a curve lies on a surface, curvedness of the surface forces the curve to bend.
Thus, curvedness of a surface can be detected by the curvatures of the curves
lying on the surface. It is clear heuristically that the curvature of a curve
on a given surface should be the same as the curvature of the intersection
curve of the surface and the osculating plane of the curve provided that the
osculating plane is not tangent to the surface. This is indeed true and thus
we may pose the question how to compute the curvature of the curve using
only information on the surface and the position of the osculating plane of
the curve. The existence of a formula that answers this question will prove
our heuristics.
Consider the curve (t) = r(u(t)) lying on the regular parameterized hy-
persurface r : Rn , where u = (u1 , . . . , un1 ) is a curve in . Express
the first two derivatives of using the Frenet formulae on one hand and the
special form of on the other. By the chain rule, we get
n1
X
vt1 = 0 = u0i ri (u)
i=1
and
n1
X n1
X n1
X
v 0 t1 + v 2 1 t2 = 00 = u0i u0j rij (u) + u00i ri (u).
i=1 j=1 i=1
Multiplying the last equation by the normal vector of the hypersurface and
using the fact that it is orthogonal to the tangent vectors t1 , r1 , . . . , rn1 , we
obtain
n1
X n1
X
v 2 1 hN(u), t2 i = hN(u), 00 i = hN(u), rij (u)iu0i u0j ,
i=1 j=1
from which
Pn1 Pn1 0 0
1 i=1 j=1 hN(u), rij (u)iui uj
1 = . (3.1)
hN(u), t2 i v2
Let us study this expression. We claim that the right hand side is determined
by the osculating plane of the curve and the hypersurface provided that the
osculating plane is not tangent to the surface.
Let us start with the expression
Pn1 Pn1 0 0
0 i=1 j=1 hN(u), rij (u)iui uj
k( ) = .
v2
Since the quantities hN(u), rij (u)i are determined by the parameterization of
the hypersurface, the functions u01 , . . . , u0n1 are the components of the speed
vector 0 of the curve with respect to the basis r1 , . . . , rn1 of the tangent
space, v is the length of the speed vector 0 , k( 0 ) depends only on the speed
vector 0 of the curve (that justifies the notation k( 0 )).
Definition 3.1.4. Let v 6= 0 be an arbitrary tangent vector of the regular
parameterized hypersurface r : Rn at r(u). The intersection curve of the
hypersurface and the plane through r(u) spanned by the direction vectors
N(u) and v is called the normal section of the hypersurface in the direction
v. Assume that the normal section is parameterized in such a way that the
speed vector of the parameterization at r(u) is v. (Such a parameterization
always exists in a neighborhood of r(u).) Orienting the cutting normal plane
by the ordered basis (v, N(u)), we may consider the signed curvature of the
normal section, which will be called the normal curvature of the hypersurface
in the direction v and will be denoted by k(v).
Applying (3.1) for normal sections one may see easily that the normal cur-
vature of a parameterized hypersurface in the direction v = v1 r1 (u) + +
vn1 rn1 (u) is just
Pn1 Pn1
i=1 j=1 hN(u), rij (u))vi vj
k(v) = , (3.2)
v2
where v = kvk. Since k(v) = k(v) for any 6= 0, the normal curvature
depends only on the direction of v.
Returning to the curve we see that k( 0 ) is determined by the tangent line

of at the given point which is the intersection of the osculating plane of
and the tangent space of the hypersurface.
Since the osculating plane and the tangent line determine the second Frenet
vector t2 uniquely up to sign, we conclude that the curvature 1 =
(1/hN(u), t2 i)k( 0 ) of the curve is determined by the osculating plane up
to sign and since the curvature 1 is positive, if n 3, both t2 and 1 are
determined uniquely (and not only up to sign) by the osculating plane.
To finish this unit with, we formulate an obvious consequence of the formula
expressing the curvature of a curve lying on a hypersurface.
Corollary 3.1.5 (Meusniers theorem). If the osculating plane of a curve

lying on a hypersurface is not contained in the tangent space of the hyper-
surface at a given point (t) = r(u(t)), then the curvature of the curve and
the normal curvature of the surface in the direction 0 (t) are related to one
another by the equation 1 (t) = cos1 k( 0 (t)), where is the angle between
the normal vector N(u(t)) of the hypersurface and the second Frenet vector
t2 (t) of the curve.
3.1.3 The Weingarten Map and the Fundamental Forms

Definition 3.1.6. Let r : Rn be a parameterized hypersurface. A vector
field along the hypersurface is a mapping X : Rn from the domain of
parameters into Rn . X is a tangential vector field, if X(u) is tangent to the
hypersurface at r(u) for all u .
For example, the mappings r1 , . . . , rn1 and N are vector fields along the
hypersurface, and the first n 1 of them are tangential.
Given a vector field along a hypersurface, we would like to express the speed
of change of the vector field as we move along the surface, in terms of the
speed of our motion. This is achieved by the following.
Definition 3.1.7. Let r : Rn be a parameterized hypersurface, X :

Rn be a vector field along it, u0 , v a tangent vector of the hypersurface
at r(u0 ). We define the derivative v X of the vector field X with respect to
the tangent vector v as v X = (X u)0 (0), where u : [1, 1] is a curve
in the parameter domain such that u(0) = u0 and (r u)0 (0) = v.
Since by the chain rule
n1
X
0
(X u) (0) = u0i (0)Xi (u(0)),
i=1
where (u01 (0), . . . , u0n1 (0)) are the components of v in the basis r1 (u0 ), . . . ,
rn1 (u0 ) of the tangent space at r(u0 ), we have the following formula
n1
X
v X = vi Xi (u0 ),
i=1
where v1 , . . . , vn1 are the components of the vector v in the basis r1 (u0 ), . . . ,
rn1 (u0 ). This formula shows that the definition of v X is correct, i.e.
independent of the choice of the curve u(t).
The way in which a hypersurface curves around in Rn is closely related to
the way the normal direction changes as we move from point to point.
Lemma 3.1.8. The derivative v N of the unit normal vector field of a hy-
persurface with respect to a tangent vector v at p = r(u) is tangent to the
hypersurface at r(u).
Proof. We have to show that v N is orthogonal to N(u). Indeed, differenti-
ating the relation 1 hN, Ni, we get
0 = hv N, Ni + hN, v Ni = 2hv N, Ni.
Definition 3.1.9. Let us denote by M the parameterized hypersurface r :
Rn and by Tp M the linear space of its tangent vectors at p = r(u0 ). The
linear map
Lp : Tp M Tp M, Lp (v) = v N
is called the Weingarten map or shape operator of M at p.
We define two bilinear forms on each tangent space of the hypersurface.
Definition 3.1.10. Let M be a parameterized hypersurface r : Rn , u0
, Tp M the linear space of tangent vectors of M at p = r(u), Lp : Tp M
Tp M the Weingarten map. The first fundamental form of the hypersurface
is the bilinear function Ip on Tp M obtained by restriction of the dot product
onto Tp M
Ip (v, w) = hv, wi for v, w Tp M.
The second fundamental form of the hypersurface is the bilinear function IIp
on Tp M defined by the equality

IIp (v, w) = hLp v, wi for v, w Tp M.
The first fundamental form is obviously a positive definite symmetric bilinear

function on the tangent space. Its matrix representation with respect to the
basis (r1 (u0 ), . . . , rn1 (u0 )) has entries hri (u0 ), rj (u0 )i.
An important property of the Weingarten map and the second fundamental
form is stated in the following theorem.
Theorem 3.1.11. The second fundamental form is symmetric, i.e.
hLp v, wi = hv, Lp wi for v, w Tp M,
or in other words, the Weingarten map is self-adjoint (with respect to the

first fundamental form).
Lemma 3.1.12.
IIp (ri (u0 ), rj (u0 )) = hrij (u0 ), N(u0 )i. (3.3)
Proof. We know that the normal vector field N is perpendicular to any tan-
gential vector field, thus
hN, rj i 0.
Differentiating this identity with respect to the i-th parameter we get
hri N, rj i + hN, rji i 0,
from which
hN, rji i hri N, rj i = hLr ri , rj i.
Proof of Theorem 3.1.11. It is enough to prove that the matrix of the second
fundamental form with respect to the basis r1 (u0 ), . . . , rn1 (u0 ) is symmetric.
However, this follows from the lemma, since by Youngs theorem rij = rji .
Comparing the identity (3.3) with (3.2) we see that the normal curvature is
the quotient of the quadratic forms of the second and first fundamental forms
Pn1 Pn1
i=1 j=1 hN(u0 ), rij (u0 )ivi vj IIp (v, v)
k(v) = = ,
|v|2 Ip (v, v)
Pn1
where v = i=1 vi ri (u0 ) is a tangent vector of the hypersurface at p = r(u0 ).
The expression
IIp (v, v)
k(v) = ,
Ip (v, v)
gives rise to a linear algebraic investigation of the normal curvature. It is
natural to ask at which directions the normal curvature attains its extrema.
Since k(v) = k(v) for any 6= 0, it is enough to consider this question for
the restriction of k onto the unit sphere S in the tangent space. The unit
sphere of a Euclidean space is a compact (i.e., closed and bounded) subset,
thus, by Weierstrass theorem, any continuous function defined on it attains
its maximum and minimum.
Definition 3.1.13. Let f be a differentiable function defined on the unit

sphere S of a Euclidean vector space. We say that the vector v S is a
critical point of f if for any curve : [1, 1] S on the sphere such that
(0) = v the derivative of the composite function f vanishes at 0.
Clearly, points at which a function is locally minimal or maximal are critical,

but the converse is not true. The following proposition gives a characteri-
zation of critical points for the restriction of the normal curvature onto the
unit sphere of the tangent space.
Proposition 3.1.14. Let V be a finite dimensional vector space with a pos-

itive definite symmetric bilinear function h, i and let L : V V be a self-
adjoint linear transformation on V . Set S = {x V |hx, xi = 1} and define
f : S R by f (x) = hLx, xi. Then v S is a critical point of f if and only
if v is an eigenvector of L.
Proof. For any curve : [1, 1] S such that (0) = v, we have
d
hL((t)), (t)it=0 = hL( 0 (0)), (0)i + hL((0)), 0 (0)i

dt
= hL 0 (0), vi + hLv, 0 (0)i = 2hLv, 0 (0)i.
This means that v is a critical point of f if and only if Lv is orthogonal

to all vectors of the form 0 (0). Since the speed vectors 0 (0) of spherical
curves through v = (0) range over the tangent space of the sphere S, v is
a critical point of f if and only if Lv is orthogonal to the tangent space of S
at v. However, since the normal vector of this tangent space is v, the latter
condition is satisfied if and only if Lv is a scalar multiple of v, i.e. v is an
eigenvector of L.
Applying Proposition 3.1.14 to the Weingarten map Lp we obtain that the

critical vectors of the normal curvature function kp at p are the eigenvectors of
Lp . According to the Principal Axis Theorem 1.2.65, there is an orthonormal
basis of the tangent space Tp M consisting of eigenvectors of the Weingarten
map.
Definition 3.1.15. For a hypersurface M in Rn parameterized by r, r(u0 ) =

p M , the eigenvalues 1 (p), . . . , n1 (p) of the Weingarten map Lp : Tp M
Tp M are called the principal curvatures of M at p, the unit eigenvectors of
Lp are called principal curvature directions.
If the principal curvatures are ordered so that 1 (p) 2 (p) n1 (p),

the discussion above shows that n1 (p) is the maximal, 1 (p) is the minimal
value of the normal curvature k(v).
Theorem 3.1.16 (Eulers formula). Let v1 , . . . , vn1 be an orthonormal

basis of Tp M consisting of principal curvature directions, 1 (p), . . . , n1 (p)
be the corresponding principal curvatures. Then the normal curvature k(v)
in the direction v Tp M , |v| = 1, is given by
n1
X n1
X
k(v) = i (p)hv, vi i2 = i (p) cos2 (i ),
i=1 i=1
where i = arccos(hv, vi i) is the angle between v and vi .

Proof. Since (v1 , . . . , vn ) is an orthonormal basis, v can be expressed as
n1
X
v= hv, vi ivi .
i=1
Making use of this formula, we obtain

D n1
X n1
X E
k(v) = hLp (v), vi = Lp hv, vi ivi , hv, vi ivi
i=1 i=1
D n1
X n1
X E n1
X
= hv, vi ii (p)vi , hv, vi ivi = i (p)hv, vi i2 .
i=1 i=1 i=1
The determinant and trace of the Weingarten map, that is the product and
sum of the principal curvatures are of particular importance in differential
geometry.
Definition 3.1.17. For M a hypersurface, p M , the determinant K(p) of
the Weingarten map Lp is called the Gaussian or Gauss-Kronecker curvature
of M at p, H(p) = tr(Lp )/(n 1) is called the mean curvature or Minkowski
curvature.
When we want to compute the principal curvatures and directions of a hy-
persurface at a point we generally work with a matrix representation of
the Weingarten map. Dealing with a parameterized hypersurface r :
Rn , it is natural to take matrix representations with respect to the basis
r1 (u), . . . , rn1 (u) of the tangent space at r(u). Let us denote by G =
(gij )1i,jn1 , B = (bij )1i,jn1 and L = (lji )1i,jn1 the matrix rep-
resentations of the first and second fundamental forms and the Weingarten
map respectively, with respect to this basis (gij , bij and lji are functions on the
parameter domain, i is the row index, j is the column index). Components
of G and B can be calculated according to the equations
gij = hri , rj i,
bij = hN, rij i (cf. Lemma 3.1.12).
By definition, the coefficients lji are obtained from the equations
n1
X
Lr (rj ) = lji ri i = 1, 2, . . . , n 1.
i=1
The relationship between the matrices G, B, and L follows from the following
equalities
D n1
X E n1
X n1
X
bji = bij = hLr ri , rj i = lik rk , rj = lik hrk , rj i = gjk lik
k=1 k=1 k=1
expressing that B = GL. As G is the matrix of a positive definite bilin-

ear function, it is invertible (its determinant is positive). Multiplying the
equation B = LG with the inverse of G we get the expression
L = G 1 B
for the matrix of the Weingarten operator.

Corollary 3.1.18. The Gaussian curvature of a hypersurface is equal to
det B
K= .
det G
Recall from linear algebra that in order to determine the eigenvalues of a
linear mapping with matrix representation L one has to find the roots of the
characteristic polynomial pL () = det(L I), where I denotes the identity
matrix.
Having determined the eigenvalues of the linear mapping, components of
eigenvectors with respect to the fixed basis are obtained as non-zero solutions
of the linear system of equations Lv = v, where is an eigenvalue of L.
Exercise 3.1.19. Determine the Weingarten map for a sphere of radius r at
one of its points.
Exercise 3.1.20. Find the normal curvature k(v) for each tangent direction
v, the principal curvatures and the principal curvature directions, and com-
pute the Gaussian and mean curvatures of the following surfaces at the given
point p.
(a) (x21 /a2 ) + (x22 /b2 ) + (x23 /c2 ) = 1, p = (a, 0, 0) (ellipsoid);
(b) (x21 /a2 ) + (x22 /b2 ) (x23 /c2 ) = 1, p = (a, 0, 0) (one-sheeted hyperboloid);
2

p
(c) x21 + x22 + x23 2 = 1, p = (0, 3, 0) or p = (0, 1, 0) (torus).
Exercise 3.1.21. Suppose that the principal curvatures of a connected pa-

rameterized surface in R3 vanish. Show that the surface is a part of a plane.
Exercise 3.1.22. Find the Gaussian curvature K : M R for the following

surfaces
(a) x21 + x22 x23 = 0, x3 > 0 (cone);
(b) (x21 /a2 ) + (x22 /b2 ) (x23 /c2 ) = 1 (hyperboloid);
(c) (x21 /a2 ) + (x22 /b2 ) x3 = 0 (elliptic paraboloid);
(d) (x21 /a2 ) (x22 /b2 ) x3 = 0 (hyperbolic paraboloid).
Exercise 3.1.23. Let M be a (hyper)surface in R3 , p M . Show that for

each v, w Tp M ,
Lp (v) Lp (w) = K(p)v w.
3.1.4 Umbilical Points

Definition 3.1.24. A point p = r(u) of a regular parameterized hypersurface
r : Rn is called an umbilical point or umbilic if the principal curvatures
at p are equal. A umbilical point p is said to be spherical if the common
value (u) of the principal curvatures at p is not equal to zero, and planar if
(u) = 0.
At an umbilical point p, the Weingarten map is a constant multiple of the

identity map of Tp M , where the multiplier is (u). To compute umbili-
cal points in practice, it is more convenient to consider the matrix equation
B(u) = G(u) instead of the equation G 1 (u)B(u) = In1 . By the symme-
try of the fundamental forms, this matrix equation gives n(n + 1)/2 ordinary
equations for the unknown vector u and the scalar .
The following theorem gives a characterization of those hypersurfaces which
have only umbilical points.
Theorem 3.1.25. Suppose that the parameter domain of a regular pa-

rameterized hypersurface r : Rn is connected and that all the points of
the hypersurface are umbilical. If n 3, then the trace M = r() of r is
contained either in a hypersphere or in a hyperplane.
Proof. Denote by (u) the common value of the principal curvatures at r(u).
First we show that is constant. For any pair of indices 1 i < j n 1,
we have
Ni = ri and Nj = rj ,
since ri and rj are principal directions as any tangent vector is. Differentiat-
ing the first equation with respect to the jth variable, and the second with
respect to the ith, we get
Nij = j ri rij and Nji = i rj rji ,
from which j ri = i rj . Since ri and rj are linearly independent, the last

equation can hold only if i = v = j, i.e., if is constant.
Case 1: 0. In this case Ni 0 for all 1 i n 1, therefore the nor-
mal vector is constant along the hypersurface. The derivative of the function
hN, ri with respect to the ith variable is hN, ri i = 0 because N is perpen-
dicular to the tangent vectors ri , hence hN, ri is constant and M = r() is
contained in a hyperplane with equation hN, xi = const.
Case 2: 6= 0. We claim that in this case M is contained in a hypersphere.
The facts we have so far suggest that if the claim is true, then the center of
the hypersphere should be r + (1/)N. Setting p(u) = r(u) + (1/)N(u), we
have to make sure first that p(u) does not depend on (u). Indeed, using the
fact that is constant, differentiation with respect to the ith variable results
pi = ri + (1/)Ni = ri (1/)ri = 0
for any i.
Now to show that the surface lies on a hypersphere centered at p we have to
prove that the function kr pk is constant. This is clear as

1 1
kr(u) pk = N(u)

=

is constant. The theorem is proved.
3.1.5 The Fundamental Equations of Hypersurface The-

ory
Now we derive some formulae for hypersurfaces. Consider a regular param-
eterized hypersurface r : Rn1 . The partial derivatives r1 , . . . , rn1
define a basis of the tangent space of the hypersurface at each point. If we
add to these vectors the normal vector of the hypersurface, we get a basis
of Rn at each point of the hypersurface. The system of the vector fields
r1 , . . . , rn1 , N along r is called the Gauss frame of the hypersurface. Gauss
frame plays similar role in the theory of hypersurfaces as Frenet frame does
in curve theory. Similarity is not complete however, since a Gauss frame is
much more dependent on the parameterization. Nevertheless, in the same
way as for Frenet frames, it is important to know how the derivatives of
the frame vector fields with respect to the parameters can be expressed as a
linear combination of the frame vectors. For this we have to determine the
coefficients kij , ij , jk , j in the expressions
n1
X n1
X
rij = kij rk + ij N, Nj = jk rk + j N. (3.4)
k=1 k=1
Let us begin with the simple observation that since Nj is known to be tan-
gential, and Nj = Lr (rj ), where Lr is the Weingarten map,
j = 0 for all j,
n1
and (jk )j,k=1 is the matrix L = G 1 B of the Weingarten map with respect
to the basis r1 , . . . , rn1 . Denote by gij and bij the entries of the first and
second fundamental forms as usual, and denote by g ij the components of
the inverse matrix of the matrix of the first fundamental form. (Attention!
Entries of G and G 1 are distinguished by the position of indices.) Then
n1
X
jk = ljk = g ki bij .
i=1
Taking the dot product of the first equation of (3.4) with N we gain the
equality hrij , Ni = ij and since hrij , Ni = bij ,
ij = bij for all i, j.
There is only one question left: What are the coefficients kij equal to? Let
us take the dot product of the first equation of (3.4) with rl
n1
X n1
X
hrij , rl i = kij hrk , rl i = kij gkl ,
k=1 k=1
or denoting the dot product hrij , rl i shortly by ijl ,

n1
X
ijl = kij gkl .
k=1
The coefficients kij

and ijk are called the Christoffel symbols of the first and
second type respectively. The last equation shows how to express Christoffel
symbols of the second type with the help of Christoffel symbols of the first
type. It can also be used to express Christoffel symbols of the first type
in terms of secondary Christoffel symbols.
Pn1Indeed, multiplying the equation
with g ls , summing up for l and using l=1 gkl g ls = ks , we get
n1
X n1
X n1
X n1
X
ls
ijl g = kij gkl g ls = kij ks = sij .
l=1 l=1 k=1 k=1
Now let us determine the Christoffel symbols of the second type. Differen-
tiating the equality gij = hri , rj i with respect to the k-th variable and then
permuting the role of the indices i, j, k we get the equalities
gij,k = hrik , rj i + hri , rjk i,

gjk,i = hrji , rk i + hrj , rki i,
gki,j = hrkj , ri i + hrk , rij i.
Solving this linear system of equations for the secondary Christoffel symbols
standing on the right-hand side, we obtain
1
ijk = hrij , rk i = (gik,j + gjk,i gij,k )
2
and
n1 n1
X X 1
kij = ijl g lk
= (gil,j + gjl,i gij,l )g lk . (3.5)
2
l=1 l=1
Observe that the Christoffel symbols depend only on the first fundamental
form of the hypersurface.
Now we ask the following question. Suppose we are given 2(n 1)2 smooth
functions gij , bij i, j = 1, 2, . . . , n 1 on an open domain of Rn1 . When
can we find a regular parameterized hypersurface r : Rn the fundamental
forms of which are represented by the matrices G = (gij ) and B = (bij ). We
have some obvious restrictions on the functions gij and bij . First, gij = gji ,
bij = bji , and since G is the matrix of a positive definite bilinear form, the
determinants of the corner submatrices (gij )ki,j=1 must be positive for k =
1, . . . , n 1. However, several examples show that these conditions are not
enough to guarantee the existence of a hypersurface. For example, if G is
the identity matrix everywhere, while B = f G for some function on , then
the hypersurface (if exists) consists of umbilics. We know however that if
a surface consists of umbilics, then the principal curvatures are constant, so
although our choice of B and G satisfies all the conditions we have listed so
far, it does not correspond to a hypersurface unless f is constant. So there
must be some further relations between the components of B and G. Our
plan to find some of these correlations is the following. Let us express rijk
and rikj as a linear combination of the Gauss frame vectors. The coefficients
we get are functions of the entries of the first and second fundamental forms.
Since rijk = rikj , the corresponding coefficients in the expressions for these
vectors must be equal and it can be hoped that this way we arrive at further
non-trivial relations between G and B. This was the philosophy, and now let
us perform the computation.

n1
! n1
X X
l
rijk = ij rl + bij N = (lij,k rl + lij rlk ) + bij,k N + bij Nk
l=1 ,k l=1
n1 n1
! n1
X X X n1
X
lij,k rl lij slk rs bks g sl rl

= + + blk N + bij,k N bij
l=1 s=1 l=1 s=1
n1
X n1
X n1
X n1
X
= (lij,k + sij lsk bij bks g sl )rl + (bij,k + lij blk )N.
l=1 s=1 s=1 l=1
Comparing the coefficient of rl in rijk and rikj , we obtain

n1
X n1
X
lij,k lik,j + (sij lsk sik lsj ) = (bij bks bik bjs )g sl ,
s=1 s=1
while comparison of the coefficient of N gives

n1
X
bij,k bik,j = (lik blj lij blk ).
l=1
The first (n1)4 equations (we have an equation for all i, j, k, l), are the Gauss
equations for the hypersurface. The second family of (n 1)3 equations are
the CodazziMainardi equations.
Exercise 3.1.26. Express the second order derivatives Nij and Nji as a
linear combination of the Gauss frame vectors. Compare the corresponding
coefficients and prove that their equality follows from the Gauss and Codazzi
Mainardi equations.
The exercise points out that a similar try to derive new relations between G
and B does not lead to really new results. This is no wonder, since the Gauss
and CodazziMainardi equations together with the previously listed obvious
conditions on G and B form a complete system of necessary and sufficient
conditions for the existence of a hypersurface with fundamental forms G and
B.
Theorem 3.1.27 (Fundamental theorem of hypersurfaces). Let Rn be
an open connected and simply connected subset of Rn1 (e.g. an open ball or
cube), and suppose that we are given two smooth (n1)(n1) matrix valued
functions G and B on such that G = (gij ) and B = (bij ) assign to every point
a symmetric matrix, G gives the matrix of a positive definite bilinear form. In
this case, if the functions kij derived from the components of G according to
the formulae (3.5) satisfy the Gauss and CodazziMainardi equations, then
there exists a regular parameterized hypersurface r : Rn for which the
matrix representations of the first and second fundamental forms are G and
B respectively. Furthermore, this hypersurface is unique up to rigid motions
of the whole space. In other words, if r1 and r2 are two such hypersurfaces,
then there exists an orientation preserving isometry : Rn Rn for which
r2 = r1 .
Let us denote the expressions standing on the left-hand sides of the Gauss
equations by
n1
X
l
Rkji := lij,k lik,j + (sij lsk sik lsj ).
s=1
Then Gauss equations can be abbreviated writing

n1
X
l
Rkji = (bij bks bik bjs )g sl .
s=1
Let us multiply this equation by glm and take a summation for l

n1
X n1
X n1
X
l
Rkji glm = (bij bks bik bjs )g sl glm
l=1 l=1 s=1
n1
X
s
= (bij bks bik bjs )m = (bij bkm bik bjm ).
s=1
Pn1 l
Introducing the functions Rkjim := l=1 Rkji glm , we may write
Rkjim = (bij bkm bik bjm ).
Let us observe, that the functions Rimjk can be expressed in terms of the
first fundamental form G.
Corollary 3.1.28 (Theorema Egregium). The Gaussian curvature of a reg-
ular parameterized surface in R3 can be expressed in terms of the first funda-
mental form as follows
R1221
K= .
det G
Theorema Egregium (meaning Remarkable Theorem in Latin) is one of those
theorems of Gauss he was very proud of. The surprising fact is not the actual
form of this formula but the mere existence of a formula that expresses the
Gaussian curvature in terms of the first fundamental form. The geometrical
meaning of the existence of such a formula is that the Gaussian curvature

does not change when we bend the surface.
Bending of a hypersurface is a deformation of it which preserves the lengths
of curves drawn onto the hypersurface. For example, when a flat sheet of
paper is bent into a cylinder, or a cone, it is a bending in the mathematical
sense as well, but blowing up a balloon is not a bending, because it increases
the lengths of curves drawn on the balloon.
If we choose an injective regular parameterization r : Rn of a hypersur-
face M = r() and the mapping : M Rn describes where a point of M
goes when we deform the surface, then r = r : Rn is a parameter-
ization of the image hypersurface (M ). The deformation = r r1 is
uniquely determined by the parameterizations r and r. This motivates the
following formal definition.
Definition 3.1.29. We say that the regular parameterized hypersurfaces

r : Rn and r : Rn are obtained from one another by a bending if
for any smooth curve u : [a, b] , the lengths of the curves = r u and
= r u are equal.
The connection between bendings and the first fundamental form is given in
the following simple proposition.
Proposition 3.1.30. The parameterized hypersurfaces r : Rn and r :

Rn are obtained from one another by a bending if and only if the matrix G of
the first fundamental form of r with respect to the basis (r1 , . . . , rn1 ) is the
same as the matrix G of the first fundamental form of r with respect to the
basis (r1 , . . . , rn1 ).
Proof. Let u : [a, b] , u(t) = (u1 (t), . . . , un1 (t)) be a smooth curve.
Then the length of the curve = r u is equal to
v
Z b uD n1
Z bu X 0
n1
X 0
E
l = k 0 (t)kdt = t ri (u(t))ui (t), rj (u(t))uj (t) dt
a a i=1 j=1
v
u n1
Z bu X 0 0
= t ui (t)uj (t)gij (u(t)) dt,
a i,j=1
where the functions gij = hri , rj i are the entries of the matrix G. Similarly,
the length of = r u can be expressed as
v
u n1
Z bu X 0 0
l = t ui (t)uj (t)gij (u(t)) dt,
a i,j=1
where gij = hri , rj i. Comparing these two equations we see at once that if
G = G, then l = l for all smooth curves u : [a, b] .
Conversely, assume that l = l holds for any smooth curve u : [a, b] .
Then, in particular, this holds also for the restriction of the curve u onto the
interval [a, x] for any x (a, b], consequently
v
Z xu u n1X 0 0
t ui (t)uj (t)gij (u(t)) dt
a i,j=1
v
Z x
u n1
uX 0 0
= t ui (t)uj (t)gij (u(t)) dt for all x (a, b].
a i,j=1
Differentiating this equation with respect to x at x0 (a, b) we obtain

v v
u n1 u n1
uX 0 0
uX 0 0
t ui (x0 )uj (x0 )gij (u(x0 )) = t ui (x0 )uj (x0 )gij (u(x0 )),
i,j=1 i,j=1
or equivalently,
n1 n1
X 0 0 X 0 0
ui (x0 )uj (x0 )gij (u(x0 )) = ui (x0 )uj (x0 )gij (u(x0 ))
i,j=1 i,j=1 (3.6)
for all x0 (a, b).
Let u0 be in arbitrary point of the parameter domain, v = (v 1 , . . . , v n1 )

Rn1 be an arbitrary vector, and consider the curve u : [a, b] Rn1 ,
u(t) = u0 + tv. If the interval [a, b] is a sufficiently small neighborhood of 0,
then the image of u is in and we can apply (3.6) to u at x0 = 0. This gives
n1
X n1
X
v i v j gij (u0 ) = v i v j gij (u0 ) for any u0 and v Rn1 .
i,j=1 i,j=1
For a fixed u0 , both sides of this equation is a quadratic polynomial of the co-
ordinates of v. These two quadratic polynomials are equal, so the coefficients
of each monomial are equal. Comparing the coefficients of the monomial v i v j
we obtain the equality
1 1
(gij (u0 ) + gji (u0 )) = (gij (u0 ) + gji (u0 )).
1 + ij 1 + ij
Since G and G are symmetric matrices, this equation gives G = G.

According to the proposition, all quantities that can be expressed in terms

of the entries of the matrix of the first fundamental form, in particular the
Gaussian curvature, are invariant under bendings. The bending invariance
of the Gaussian curvature is remarkable because the principal curvatures of
a surface do change in general when the surface is bent.
Definition 3.1.31. Let r : Rn be a hypersurface.
Pn1 Consider the mapping
Pn1
R that assigns four tangential vector fields X = i=1 X i ri , Y = i=1 Y i ri ,
Pn1 Pn1
Z = i=1 Z i ri , W = i=1 W i ri a function according to the formula
X n1
n1 X n1
X n1 X
R(X, Y ; Z, W ) = Rimjk X i Y m Z j W k .
i=1 m=1 j=1 k=1
We shall call R the Riemannian curvature tensor of the hypersurface, the

functions Rimjk the components of the curvature tensor.
Recall that a tensor of type (k, l) over a linear space V can be identified with
a multilinear function
T : V V V V R
| {z } | {z }
k l
defined on the Cartesian product of k copies of V and l copies of V . Mul-

tilinear means that fixing all but one variables, we obtain a linear function
of the free variable. If (e1 , . . . , en ) is a basis of V , (e1 , . . . , en ) is its dual ba-
sis, then every tensor T is uniquely determined by its values on basis vector
combinations, i.e. by the numbers
Tji11...j
...il
k
= T (ej1 , . . . , ejk ; ei1 , . . . , eil ),
which are called the components of the tensor T with respect to the basis
e1 , . . . , en .
Now consider a regular parameterized hypersurface M , r : Rn . A ten-
sor field of type (k, l) over M is a mapping T that assigns to every point
u a tensor of type (k, l) over the tangent space of M at r(u). T (u) is
uniquely determined by its components Tji11...j ...il
k
(u) with respect to the basis
i1 ...il
r1 (u), . . . , rn1 (u). The functions u 7 Tj1 ...jk (u) are called the components
of the tensor field T . T is said to be a smooth tensor field if its components
are smooth.
Examples.
Functions on M are tensor fields of type (0, 0).
Tangential vector fields are tensor fields of type (0, 1) (V is isomorphic
to V in a natural way).
The first and second fundamental forms of a hypersurface are tensor

fields of type (2, 0).
The mapping that assigns to every point of a hypersurface the Wein-
garten map at that point is a tensor of type (1, 1). (The linear space of
V V linear mappings is isomorphic to T (1,1) V in a natural way.)
Let f be a smooth function on M . Consider the tensor field of type
(1, 0) defined on a tangent vector X to be the derivative of f in the
direction X. This tensor field is the differential of f .
The Riemannian curvature tensor is a tensor field of type (4, 0).
Curvature tensor is one of the most basic objects of study in differential
geometry. However, Definition 3.1.31 does not cast any light on its geometric
content. To understand its real meaning, we shall introduce the curvature
tensor in a more natural way on a more general level in the framework of
Riemannian manifolds. For this purpose, we have to get acquainted with
some fundamental definitions and constructions. This will be the goal of the
next chapter.
Exercise 3.1.32. Determine the components of the Riemannian curvature
tensor of the plane with respect to the polar parameterization

r(u, v) = (u cos v, u sin v, 0).
Exercise 3.1.33. Compute the Christoffel symbols of the second kind for
surfaces of revolution with respect to the parameterization
r(u, v) = (x(u) cos v, x(u) sin v, y(u)),
where x(u) and y(u) are given functions of one variable.

Exercise 3.1.34. The components of the first fundamental form of a surface
with respect to a given parameterization are
guu = gvv = 1, guv = 0.
Show that the Gauss curvature of the surface is constant 0.

Exercise 3.1.35. The components of the first fundamental form of a surface
with respect to a given parameterization defined on an open subset of the
upper half-plane {(u, v) | v > 0} are
1
guu = gvv = , guv = 0.
v2
Show that the Gauss curvature of the surface is constant 1. (Poincares
half-plane model of the hyperbolic plane.)
Exercise 3.1.36. Show that the curvature tensor satisfies the following sym-
metry relations:
R(X, Y ; Z, W ) = R(Y, X; Z, W ) = R(X, Y ; W, Z);
R(X, Y ; Z, W ) + R(X, Z; W, Y ) + R(X, W ; Y, Z) = 0 (Bianchi identity).
3.1.6 Surface Volume

In this section we define the surface volume of a domains on a hypersurface.
Let r : Rn be a regular parameterized hypersurface in Rn , denote by
G : R(n1)(n1) the matrix of the first fundamental form with respect
to the basis r1 , . . . , rn1 .
Definition 3.1.37. D be a Lebesgue-measurable subset. We define the

surface volume n1 (r(D)) of r(D) as the integral

Z
n1 (r(D)) = det G dn1 ,
D
where the integral is taken with respect to the Lebesgue measure n1 on

Rn1 . In the case n = 3, 2 (r(D)) is also called the surface area of r(D).

p for choosing the function det G for the integrand is that geo-
The motivation
metrically, det G(u) is the volume of the (n 1)-dimensional parallelepiped
spanned by the partial derivatives r1 (u), . . . , rn1 (u). Since these partial
derivatives are the images of the standard basis vectors of Rn1 under the
derivative map of rpat u, and the standard basis vectors of Rn1 span a unit
cube of volume 1, det G(u) is the factor by which the derivative map r0 (u)
stretches or shrinks (n 1)-dimensional volume.
A rigorous statement, which justifies the definition in another way is the
following one.
Proposition 3.1.38. Suppose that the hypersurfaces r : Rn and r :

Rn are reparameterizations of one another, that is, assume there is a diffeo-
morphism h : such that r = r u. Denote by G and G the matrices of
the first fundamental forms of r and r respectively. Then for any Lebesgue-
measurable subset D ,

Z Z q
det G dn1 = det G dn1 .
D h(D)
For this reason, the surface volume of r(D) = r(h(D)) does not depend on
the way we parameterize the domain.
Proof. We apply the theorem on the change of variables in multivariable

integrals. This gives at once that
Z q Z q
| det h0 | det G h dn1 = det G dn1 ,
D h(D)
p
so we have only to prove | det h0 | det G h = det G.
By the chain rule we have
n1
X
ri (u) = rj (h(u))i hj (u),
j=1
where (h1 , . . . , hn1 ) are the coordinate functions of h. Taking the wedge
product of these equations and applying Proposition 1.2.31 we obtain
r1 (u) rn1 (u) = det h0 (u) r1 (h(u)) rn1 (h(u)).
Computing the length of the (n 1)-vectors on the two sides, with the help
of equation (1.9), we get the required equality
p q
det G(u) = | det h0 (u)| det G(h(u)).
3.2 Surfaces in R3
A regular parameterized surface r : R3 ( is an open subset of the plane)
has two principal curvatures 1 (u, v) and 2 (u, v) at each point p = r(u, v)
of the surface. If 1 (u, v) 2 (u, v) then 1 (u, v) is the minimum of normal
curvatures in different directions at p, while 2 (u, v) is the maximum of them.
If 1 (u, v) < 2 (u, v) then the principal directions corresponding to 1 (u, v)
and 2 (u, v) are uniquely defined, however if 1 (u, v) = 2 (u, v) then the
point r(u, v) is umbilical, the normal curvature is constant in all directions
and every direction is principal.
3.2.1 Surfaces of Revolution

The next example shows how to compute the principal curvatures and di-
rections for a surface of revolution. Consider a regular plane curve =
(x, y) : I R2 , with positive second coordinate function y > 0, and rotate
the curve about the x-axis. This obtained surface of revolution can be pa-
rameterized by the mapping
r(u, v) = (x(u), y(u) cos v, y(u) sin v).

3.2. Surfaces in R3 163
The coordinate lines of this parameterization are the rotations of the gener-
ating curve , and the circles drawn by points of the generator curve as they
rotate about the x-axis. Rotated copies of the generating curve are called
the generators, generatrices or meridians of the surface of revolution, while
the circles are the so-called circles of latitude.
The tangent vectors ru and rv are obtained by partial differentiation
ru (u, v) = (x0 (u), y 0 (u) cos v, y 0 (u) sin v),
rv (u, v) = (0, y(u) sin v, y(u) cos v).
The matrix of the first fundamental form with respect to the basis ru , rv is
02 2

hru , ru i hru , rv i x (u) + y 0 (u) 0
G= = .
hrv , ru i hrv , rv i 0 y 2 (u)
To obtain the matrix of the second fundamental form we need the normal
vector field and the second order partial derivatives of r.

e1 e2 e3
ru rv = det x0 (u) y 0 (u) cos v y 0 (u) sin v
0 y(u) sin v y(u) cos v
= (y 0 (u)y(u), y(u)x0 (u) cos v, y(u)x0 (u) sin v),
ru rv 1
N= =q (y 0 (u), x0 (u) cos v, x0 (u) sin v).
|ru rv |
x0 2 (u) + y 0 2 (u)
The second order partial derivatives of the parameterization are
ruu (u, v) = (x00 (u), y 00 (u) cos v, y 00 (u) sin v),
ruv (u, v) = (0, y 0 (u) sin v, y 0 (u) cos v),
rvv (u, v) = (0, y(u) cos v, y(u) sin v).
The matrix of the second fundamental form is

hN, ruu i hN, ruv i
B=
hN, rvu i hN, rvv i
00
x (u)y 0 (u) y 00 (u)x0 (u)

1 0
=q .
0 x0 (u)y(u)
x0 2 (u) + y 0 2 (u)
The matrix of the Weingarten map with respect to the basis ru , rv is
00
x (u)y 0 (u) y 00 (u)x0 (u)

(x0 2 (u) + y 0 2 (u))3/2 0
L = BG 1 =

0
.
x (u)
0 0 2 0 2 1/2
y(u)(x (u) + y (u))
As we see, the matrix of the Weingarten map is diagonal, consequently ru , rv

are eigenvectors and the diagonal elements of L are eigenvalues of the Wein-
garten map. Thus the principal curvatures of the surface are
x00 (u)y 0 (u) y 00 (u)x0 (u) x0 (u)
1 (u, v) = 2 (u, v) = .
(x0 2 (u) + y 0 2 (u))3/2 y(u)(x (u) + y 0 2 (u))1/2
02
We could have obtained this result in a more geometrical way. For any
point p on the surface, the plane through p and the x-axis is a symmetry
plane of the surface. Thus, reflection of a principal direction of the surface
at p is also a principal direction (with the same principal curvature). The
principal curvatures at p are either equal and then every direction is principal,
or different and then the principal directions are unique. Since a direction is
invariant under a reflection in a plane if and only if it is parallel or orthogonal
to the plane, we may conclude that ru and rv are principal directions of the
surface. Principal curvatures are the curvatures of the normal sections of the
surface in the direction ru , rv .
The normal section of the surface in the direction ru is a meridian of the
surface. Its curvature can be calculated according to the formula known for
plane curves and gives 1 up to sign. The difference in sign is due to the fact
that the unit normal of the surface and the principal normal of the meridian
are opposite to one another.
The plane passing through p perpendicular to the x-axis intersects the surface
in a circle of latitude the tangent of which at p is rv . The curvature of this
1
circle is . The normal curvature 2 = k(rv ) of the surface in the direction
y(u)
rv and the curvature of the circle intersection are related to one another by
Meusniers theorem as follows
1 1
= 2 ,
y(u) cos
where is the angle between the normal of the surface and the principal
normal of the circle. As it is easy to see, is the direction angle of the tangent
to the meridian at p, that is, by elementary calculus tan = y 0 (u)/x0 (u), from
x0 (u)
which cos = q .
x0 2 (u) + y 0 2 (u)
Therefore, we get
x0 (u)
2 (u, v) = q
y(u) x0 2 (u) + y 0 2 (u)
1 1
as before. The equation = 2 has the following consequence.
y(u) cos
1
Corollary 3.2.1. The second principal radius of curvature of a surface
2
of revolution at a given point p is the length of the segment of the normal of
the surface between p and the x-axis intercept.
As an application, let us show that the surface of revolution generated by

the tractrix has constant 1 Gaussian curvature. For this reason the surface
is called pseudosphere. Its intrinsic geometry is locally the same as that of
Bolyais and Lobachevskys hyperbolic plane. This local model of hyperbolic
geometry was discovered by Eugenio Beltrami (1835-1899).
Figure 3.2: The pseudosphere.
The tractrix is defined as the involute of the chain curve (t) = (t, cosh t)
touching the chain curve at (0, 1). The length of the chain curve arc between
(0) and (t) is
Z t Z tp Z t
0 2
k ( )kd = 1 + sinh d = cosh d = sinh t.
0 0 0
This way, the tractrix has the parameterization
0 (t) (1, sinh t)

(t) = (t) sinh t 0
= (t, cosh t) sinh t
k (t)k cosh t

1
= t tanh t, .
cosh t
As we know from the theory of evolutes and involutes, the chain curve is the
evolute of the tractrix, the segment (t)(t) is normal to the tractrix, and its
length is the radius of curvature of the tractrix at (t). This implies from one
1
hand that the first principal curvature of the pseudosphere is 1 = sinh t.
On the other hand, we obtain that the equation of the normal line of the
tractrix at (t) is
y cosh t
= sinh t.
xt
The x-intercept is

cosh t
t ,0 .
sinh t
According to the general results on surfaces of revolution, the second principal
radius of curvature of the pseudosphere is the distance between (t) and
(t coth t, 0), i.e.
2 !1/2
cosh2 t sinh2 t

1
1
2 = k(coth t tanh t, (cosh t)
1
)k = +
cosh t sinh t cosh2 t
1/2
1 1 1
= 2 + = .
sinh t cosh t cosh2 t
2 sinh t
This shows that K = 1 2 1.
3.2.2 Lines of Curvature, Triply Orthogonal Systems

Definition 3.2.2. A regular curve on a surface is said to be a line of curva-
ture if the tangent vectors of the curve are principal directions.
There are many parameterizations of a hypersurface. In applications we

should always try to find a parameterization that makes solving the problem
easier. For theoretical purposes, it is good to know the existence of cer-
tain parameterizations that have nice properties. In what follows, we study
parameterizations, for which coordinate lines are lines of curvature.
Theorem 3.2.3. The coordinate lines of a regular parameterization r :

Rn of a hypersurface are lines of curvature if the matrices of the first and
second fundamental forms are diagonal. The converse is also true if the
principal curvatures of the hypersurface are different at each point.
Proof. The matrix L of the Weingarten map is the quotient BG 1 of the

matrices of the first and second fundamental forms. If these matrices are
diagonal, then so is L. Obviously, the matrix of a linear map with respect to
a basis is diagonal if and only if the basis consists of eigenvectors of the linear
map. In our case, we get that r1 , . . . , rn1 form an eigenvector basis for the
Weingarten map, i.e. these vectors are principal directions. Since ri is the
tangent vector of the i-th family of coordinate lines, the coordinate lines are
lines of curvature.
Now suppose that the coordinate lines are lines of curvature and that the
principal curvatures 1 , . . . , n1 corresponding to the principal directions
r1 , . . . , rn1 are different at every point. In this case,
i hri , rj i = hLri , rj i = hri , Lrj i = j hri , rj i,

(i j )hri , rj i = 0,
and since (i j ) 6= 0 for i 6= j,
gij = hri , rj i = 0 if i 6= j.
Hence, the matrix G of the first fundamental form is diagonal. The matrix L
of the Weingarten map is also diagonal by our assumption, consequently the
matrix B = LG of the second fundamental form is diagonal as well.
Theorem 3.2.4. Every non-umbilical point of a regular parameterized sur-

face in R3 has a neighborhood that admits a reparameterization with respect
to which coordinate lines are lines of curvature.
In view of other applications, we formulate the key part of the proof as a
separate theorem.
Theorem 3.2.5. Let X1 and X2 be two smooth tangential vector fields on
a regular parameterized surface r : R3 . Assume that at a given point
p = r(u0 , v0 ), the vectors X1 (u0 , v0 ) and X2 (u0 , v0 ) are linearly independent.
Then there is a diffeomorphism H : W V between an open neighborhood
W R2 of the origin and an open neighborhood V of (u0 , v0 ) such that
the composition r = r H is a regular reparameterization of the surface in a
neighborhood of p such that the partial derivative vector fields r1 and r2 are
parallel to X1 and X2 respectively.
Proof. Let us decompose X1 and X2 as linear combinations
X1 = X11 r1 + X12 r2 ,
X2 = X21 r1 + X22 r2
of r1 and r2 . The coefficient functions Xij are smooth functions on and

they define two smooth vector fields F1 , F2 : R2 by
Fi (u, v) = (Xi1 (u, v), Xi2 (u, v)) i = 1, 2.
and two ordinary differential equations u0 = Fi u.

Let t and t be the flows generated by F1 and F2 respectively (see Definition
1.7.6) and consider the smooth maps
H1 (s, t) = t (s (u0 , v0 )), H2 (s, t) = s (t (u0 , v0 )).

Since the curve H1 (s, 0) = H2 (s, 0) = s (u0 , v0 ) is an integral curve of F1 ,

1 H1 (0, 0) = 1 H2 (0, 0) = F1 (0, 0),
and similarly
2 H1 (0, 0) = 2 H2 (0, 0) = F2 (0, 0).
Thus, the derivative matrix of H1 and H2 at the origin is
1
X1 (u0 , v0 ) X21 (u0 , v0 )

0 0
H1 (0, 0) = H2 (0, 0) = . (3.7)
X12 (u0 , v0 ) X22 (u0 , v0 )
By our assumptions, X1 (u0 , v0 ) and X2 (u0 , v0 ) are linearly independent vec-

tors, therefore both H1 and H2 has invertible derivative matrix at (0, 0).
Applying the inverse function theorem to both H1 and H2 we conclude that
the origin (0, 0) has open neighborhoods W1 , W2 , and the point (u0 , v0 ) has
open neighborhoods V1 , V2 such that Hi |Wi : Wi Vi is a diffeomor-
phism for i = 1, 2.
Denote by Gi = (G1i , G2i ) : Vi Wi the inverse map of Hi |Wi , and consider
the map G = (G11 , G22 ) : V1 V2 R2 . The derivative matrices of G1 and
G2 at (u0 , v0 ) are equal to the inverse matrix (3.7), consequently, this inverse
matrix is also the derivative matrix of G at (u0 , v0 ). Applying again the
inverse function theorem to G, we obtain that there are open subsets V
V1 V2 around (u0 , v0 ) and W W1 W2 around the origin, such that
G|V : V W is a diffeomorphism. We can assume for simplicity that W
is an open square W = {(s, t) | |s| < , |t| < }. We claim that the inverse
map H = (G|V )1 : W V gives a reparameterization with the required
property.
For this purpose, examine the coordinate lines of the parameterization r H.
Since the first coordinate function of G is the same as the first coordinate
function of G1 ,
G(H1 (s, t)) = (s, (s, t)) for all (s, t) G1 (V ),
where : W R is a smooth function. As G H leaves the first coordinate
invariant, its inverse also must have the form
G1 (H(s, t)) = (s, (s, t)), for all (s, t) W,
where for all |s| < , (s, .) = (s, .)1 . Applying H1 to both sides we get
H(s, t) = H1 (s, (s, t)) = (s,t (s (u0 , v0 )).
For a fixed s, this is a reparameterization by (s, .) of the integral curve of
F2 starting at s (u0 , v0 )). Thus,
rt (s, t) = t (s, t) X2 (H(s, t)).
This means that rt is parallel to the vector field X2 . A similar argument

shows that the vector field rs is parallel to X1 , that is the coordinate lines of
the new parameterization are tangent to X1 and X2 .
Remarks.
The parameterization constructed above has also the property that the
coordinate lines through p are integral curves of X1 and X2 . In par-
ticular, if X1 and X2 are unit vector fields, then these two parameter
lines are parameterized by arc length.
The theorem does not hold for higher dimensions. (Find why the above
proof does not work in higher dimensions.)
Proof of Theorem 3.2.4. Let 1 2 be the two principal curvatures ar-

ranged in increasing order. Since 1 and 2 change continuously and they
are different at p, they are different also in a small neighborhood of p. Re-
stricting our attention to this neighborhood, we may assume that the surface
consists of non-umbilical points. Then at each point q of the surface, there
are unique pairs of opposite unit vectors (X1 )q and (X2 )q pointing in the
principal directions corresponding to 1 and 2 respectively. If we fix the
unit principal directions (X1 )p and (X2 )p arbitrarily and then choose the di-
rection of (X1 )q and (X2 )q at nearby points so that (Xi )q enclosed an acute
angle with (Xi )p then we obtain to smooth vector fields X1 and X2 on a
neighborhood of p. Applying to these two vector fields the above theorem,
we obtain the required reparameterization.
Dupins Theorem and Curvature Lines on Ellipsoids

Now we give a description of the lines of curvature on ellipsoids. Our ap-
proach, which is based on Dupins theorem, works for any surface of second
order. Dupins theorem claims that if we have three families of surfaces such
that the surfaces of any of the families foliate an open domain in R3 , and sur-
faces from different families intersect one another orthogonally, then surfaces
from different families intersect one another in lines of curvature. We can ob-
tain families of surfaces in a natural way considering a curvilinear coordinate
system on R3 .
Definition 3.2.6. A local (curvilinear) coordinate system on an open subset
U of R3 is a smooth mapping : U R3 , which maps U onto its image
= (U ) bijectively and the inverse 1 : R3 of which is also smooth.
The inverse r = 1 is a regular parameterization of U . is covered by
planes parallel to one of the three coordinate planes in R3 . The images of
these families of planes are the coordinate surfaces of the local coordinate
system (or that of the parameterization r). There are three families of
coordinate surfaces. Each family covers the domain U . Coordinate surfaces
from different families intersect one another in coordinate lines. We say that
(or r) defines a triply orthogonal system, if coordinate surfaces from different
families intersect one another orthogonally, or equivalently, if hri , rj i = 0 for
i 6= j.
Theorem 3.2.7 (Dupins theorem). If the curvilinear coordinate system
: U R3 defines a triply orthogonal system on U , then the coordinate
lines are lines of curvature on the coordinate surfaces.
Proof. Let r : = (U ) R3 be the inverse of . We may consider without
loss of generality the coordinate surface parameterized by (u, v) 7 r(u, v, w0 ),
where w0 is constant. It is enough to show that the matrices of the first and
second fundamental forms of the surface with respect to the given parame-
terization are diagonal. The matrix of the first fundamental form is diagonal
by our assumption hr1 , r2 i = 0. The nondiagonal element of the matrix of
the second fundamental form is hr12 , Ni, where N is the unit normal of the
surface. Since r3 is parallel to N, hr12 , Ni = 0 will follow from hr12 , r3 i = 0.
Differentiating the equation hr1 , r3 i = 0 with respect to the second variable
yields
hr12 , r3 i + hr1 , r23 i = 0
and similarly, we also have
hr23 , r1 i + hr2 , r31 i = 0,

hr31 , r2 i + hr3 , r12 i = 0.
Solving this system of linear equations for the unknown quantities hr12 , r3 i,
hr23 , r1 i, hr31 , r2 i, we see that they are all zero.
The canonical equation of an ellipsoid has the form
x2 y2 z2
+ + = 1.
A B C
Suppose A > B > C. We can embed this surface into a triply orthogonal
system of second order surfaces as follows. Consider the surface
x2 y2 z2
F : + + = 1.
A+ B+ C +
F is
an ellipsoid for > C;
a one sheeted hyperboloid for C > > B;
a two sheeted hyperboloid for B > > A.

In accordance with these cases, we get three families of surfaces. Surfaces
obtained by such a perturbation of the equation of a second order surface are
called confocal second order surfaces.
Figure 3.3: One and two sheeted hyperboloids confocal to an ellipsoid.
Proposition 3.2.8. Let (x, y, z) R3 be a point for which xyz 6= 0. Then

there exist exactly three -s, one from each of the intervals (C, +),
(B, C), (A, B) such that (x, y, z) F .
Proof. Condition (x, y, z) F is equivalent to
P () = (A + )(B + )(C + )
x2 (B + )(C + ) + y 2 (A + )(C + ) + z 2 (A + )(B + ) = 0.

This is an equation of degree three for , so the number of solutions is not

more than three. To see that all the three solutions are real and located as
it is stated, compute P at the nodes.
P (A) = x2 (B A)(C A) < 0,

P (B) = y 2 (A B)(C B) > 0,
P (C) = z 2 (A C)(B C) < 0.
Furthermore, F () = 3 + implies lim F () = +. Thus, by

Bolzanos intermediate value theorem, F has at least one root in each of
the intervals (C, +), (B, C), (A, B).
Proposition 3.2.9. If (x, y, z) F F0 , xyz 6= 0, 6= 0 , then F
intersects F0 orthogonally.
Lemma 3.2.10. If a nonempty subset M of R3 is defined by an equation

F (x, y, z) = 0, i.e.,
M = {(x, y, z) R3 | F (x, y, z) = 0},
so that the gradient vector field grad F = (1 F, 2 F, 3 F ) is not zero at points

of M , then every point of M has a neighborhood (in M ) which is the image
of a regular parameterized surface. In this case the tangent plane of M at
pr(u0 ) M is orthogonal to grad F (p).
Proof. The first part of the lemma is a direct application of the inverse func-
tion theorem, we shall prove it later in more generality (see Theorem 4.1.12).
Suppose that M admits a regular parameterization r around p M . Then
F r 0. Differentiating with respect to the i-th variable (i = 1, 2) using
the chain rule we obtain
0 = i (F r) = h(grad F ) r, ri i,
hence grad F (p) is orthogonal to the tangent vectors r1 (u0 ), r2 (u0 ) which
span the tangent plane Tp M .
Figure 3.4: Lines of curvature on an ellipsoid.
Proof of Proposition 3.2.9. We need to show
grad F (x, y, z) grad F0 (x, y, z),
or equivalently
x2 y2 z2
+ + = 0.
(A + )(A + ) (B + )(B + ) (C + )(C + 0 )
0 0
We know that
x2 y2 z2 x2 y2 z2
+ + = 1 and + + = 1.
A+ B+ C + A + 0 B + 0 C + 0
Subtracting these equalities we obtain ( 0 ) times the equality to prove.
Since 6= 0 , we are ready.
3.2.3 Ruled and Developable Surfaces

Regular surfaces swept out by a moving straight line are ruled surfaces. A
bit more generally, we shall call a regular surface ruled if every point of the
surface has a neighborhood with a regular parameterization of the form
r(u, v) = (u) + v(u),
where is a smooth curve, called the directrix, is a nowhere zero vector
field along . The straight lines v 7 (u0 ) + v(u0 ) are the generators or
generatrices of the surface.
Theorem 3.2.11. The following statements are equivalent for ruled surfaces:
(i) the normal vector field N is constant along the generators;
(ii) ruv is tangential for the parameterization r(u) = (u) + v(u);
(iii) the Gaussian curvature K is constant 0.
Proof. (i) = (iii). If N is constant along the generators, then L(rv ) =
Nv = 0 = 0rv , thus the generators are lines of curvature and the cor-
responding principal curvature is 0 everywhere. From this follows that the
Gaussian curvature is 0.
(iii) = (i). The normal section of a ruled surface in the direction of a
generator is the generator itself. Hence, the normal curvature of the surface
in the direction rv of the generators is 0. If rv were not a principal direction,
then 0 would be strictly between the principal curvatures, in which case we
would have K < 0. The contradiction shows that rv is a principal direction
at a given point, and thus Nv = L(rv ) = 0rv = 0, i.e. N is constant along
the generators.
Remark: We have proved here that K 0 for any ruled surface.
(ii) (iii). According to the formula
det B
K= ,
det G
K = 0 if and only if det B = 0. Since rvv = 0,

hruu , Ni hruv , Ni hruu , Ni hruv , Ni
det B = det = det = hruv , Ni2
hrvu , Ni hrvv , Ni hrvu , Ni 0
and thus det B = 0 if and only if ruv is orthogonal to N i.e. if ruv is tangential.
Definition 3.2.12. A ruled surface that satisfies the equivalent conditions

of the previous theorem is called a developable surface.
Examples.
(1) The following surfaces are ruled but not developable:
r(u, v) = (u, v, uv), (hyperbolic paraboloid);
r(u, v) = (a cos u, b sin u, 0) + v(a sin u, b cos u, c), (one sheeted hyper-
boloid);
r(u, v) = (0, 0, u) + v(cos u, sin u, 0) (helicoid).
Figure 3.5: A hyperbolic paraboloid, a one-sheeted hyperboloid, and a heli-

coid.
(2) There are three main types of developable surfaces:
Cylinders over a curve. Let be a regular space curve, v 6= 0 a vector

nowhere tangent to . We define the cylinder over with generators
parallel to v by the parameterization
r(u, v) = (u) + vv.
Since ruv = 0 is tangential, cylinders over a curve are developable.
Cones over a curve. Let be a regular curve, p be the position vector

of a point not lying on any tangent to the curve. The cone over with
vertex p is defined by the parameterization
r(u, v) = v(u) + (1 v)p.

Figure 3.6: Cylinder over a curve with one singular generatrix.
The cone is regular only in the domain v 6= 0. The tangent plane of the
cone is spanned by the vectors
ru (u, v) = v 0 (u) and rv (u, v) = (u) p.
Since
ruv (u, v) = 0 (u) = (1/v)ru (u, v),
cones are developable.
Figure 3.7: Cone over a curve with a singular generatrix.
Tangent surfaces. Let be a curve of general type in R3 . We show

that the regular part of the surface swept out by the tangent lines of
is developable. Indeed, the surface can be parameterized by
r(u, v) = (u) + v 0 (u).
Partial derivatives of r are
ru (u, v) = 0 (u) + v 00 (u) and rv (u, v) = 0 (u).
Since is of general type, 0 (u) and 00 (u) are linearly independent,

hence singularities of the surface of tangent lines are located along the
generating curve .
As ruv (u, v) = 00 (u) = (1/v) ru (u, v) rv (u, v) for v 6= 0, the regular

part of the surface of tangent lines is developable. Surfaces of this type

are called tangent surfaces of regular curves.
Figure 3.8: The tangent surface of a curve.
According to the definition, a developable surface is a ruled surface with

Gaussian curvature equal to zero. The following theorem states that in most
cases the condition of being ruled follows from K 0.
Theorem 3.2.13. If the Gaussian curvature of a surface is 0 everywhere

and the surface contains no planar point, then it is developable.
Proof. Gaussian curvature is positive at spherical points so the surface con-

tains no umbilics. Therefore we may consider a parameterization r around
any point p such that p = r(0, 0), coordinate lines are lines of curvature and
the coordinate lines through p are unit speed curves (see Theorem 3.2.4).
Suppose that ru corresponds to the nonzero principal curvature 1 = 6= 0.
Then
Nu = ru and Nv = 0rv = 0.
The second equation shows that N is constant along v-coordinate lines. What
we have to show is that v-coordinate lines are straight lines. For this purpose,
it is enough to prove that r is linear in v, i.e. rvv = 0. The idea we use is
that the only vector which is perpendicular to each vector of a basis is the
zero vector. According to this observation, it suffices to show that rvv is

orthogonal to the vectors N, ru , rv .
(i) rvv N. This orthogonality follows from
0 = v hrv , Ni = hrvv , Ni + hrv , Nv i = hrvv , Ni.
(ii) rvv ru . Since lines of curvature are orthogonal,
0 = v hru , rv i = hruv , rv i + hru , rvv i.
On the other hand, using the facts that Nu = ru rv and Nuv =

u (Nv ) = 0 we get
hruv , rv i = v Nu / , rv = (v /2 )hNu , rv i (1/)hNuv , rv i = 0.

Combining these two equalities we get hrvv , ru i = 0.

(iii) rvv rv . This will follow from the observation that v-coordinate lines are
all parameterized by arc length, i.e. krv k 1. We know that krv (0, v)k = 1
by the construction of r. By our previous equation we also have
u hrv , rv i = 2hruv , rv i = 0,
showing that hrv , rv i does not depend on u. Thus,
krv (u, v)k = krv (0, v)k = 1 for every u, v.
This completes the proof.

We finish the investigation of developable surfaces with a structure theorem
stating that every developable surface is made up of pieces of cylinders, cones
and tangent surfaces.
Theorem 3.2.14. Let r : [a, b] [c, d] R3 be a developable surface without
planar points and suppose that the parameterization r of the surface is the
one we used in the proof of the previous theorem. Then there exists a nowhere
dense closed subset A of [a, b] the complement of which is a union of open
intervals [a, b] \ A = I1 I2 such that the restriction of r onto In [c, d]
is a part of a cylinder or cone or a tangent surface.
Proof. As it was proved above, r has the form r(u, v) = a(u) + vb(u), where
b(u) is a unit vector field along the curve a(u). We have
ru (u, v) = a0 (u) + vb0 (u), rv (u, v) = b(u), ruv (u, v) = b0 (u).
By the definition of developable surfaces, ruv (u, v) = b0 (u) must be tangential

to the surface, i.e. it lies in the plane spanned by ru and rv . ru is orthogonal
to rv as they are lines of curvature, furthermore, ruv is also orthogonal to rv as

0 = hb(u), b(u)i0 = 2hb(u), b0 (u)i. For there is only one direction in a plane
which is orthogonal to a given nonzero vector, ru and ruv must be parallel:
b0 (u) k a0 (u). Hence, b0 (u) = c(u)a0 (u) for some function c : [a, b] R.
Now let A be the set of those roots of c c0 in [a, b], which do not have a
neighborhood containing only roots of c c0 . Then A is closed and nowhere
dense in [a, b]. If [a, b] \ A is the union of the disjoint open intervals I1 , I2 , . . . ,
then for the restriction of c onto In , we have one of the following possibilities:
(a) the restriction is identically 0;
(b) the restriction is a nonzero constant;
(c) the restriction is strictly monotone and nowhere zero.
In the first case, b0 (u) = 0 and thus b is constant on In , thus the restriction
of r onto In [c, d] is a part of a cylinder.
In the second case, the point p(u) = a(u) (1/c)b(u) does not depend on
u. Indeed, p0 (u) = a0 (u) (1/c)b0 (u) = 0. Furthermore, the point p lies on
every generator of the surface, so this case serves a part of a cone.
Finally, consider the curve (u) = a(u) (1/c(u))b(u) for the last case. As
0 (u) = a0 (u) + (c0 (u)/c2 (u))b(u) b0 (u)/c(u) = (c0 (u)/c2 (u))b(u) k b(u),
the tangent of at (u) coincides with the generator v 7 r(u, v) of the

surface, and is of general type too, as
2
0 (u) 00 (u) = c0 (u)/c2 (u) b(u) b0 (u)
2
= c0 (u)/c2 (u) c(u)b(u) a0 (u) 6= 0.
We conclude that in the third case the restriction of r onto In [c, d] is a

part of the tangent surface of the curve of general type .
Exercise 3.2.15. Prove the inequality
H2 K
for the Minkowski curvature H and the Gauss curvature K of a surface in

R3 . At which points does equality hold?
x2 y2 z2
Exercise 3.2.16. Find umbilical points on the ellipsoid a2 + b2 + c2 = 1,
(a > b > c) parameterized by
r(u, v) = (a cos u cos v, b cos u sin v, c sin u).
Exercise 3.2.17. Show that the Minkowski curvature of the following sur-
faces is 0 everywhere
(a) r(u, v) = (v cos u, v sin u, bu) (helicoid ),

(b) r(u, v) = (cosh v cos u, cosh v sin u, v) (catenoid : surface of revolution
generated by the chain curve).
Exercise 3.2.18. Let M be a surface with a regular parameterization r.
Denote by N the unit normal vector field of M and consider the parallel
surface Md parameterized by
rd (u, v) = r(u, v) + dN(u, v),
where d 6= 0 is a fixed real number.
(a) Show that the singular points of the parameterization rd correspond to
those points of M at which 1/d is a principal curvature.
(b) Show that N(u, v) is a unit normal of Md at rd (u, v).
(c) Express the Gaussian and Minkowski curvature of Md with those of M .

3.2.4 Asymptotic Curves on Negatively Curved Surfaces

Consider a regular parameterized surface in r : R3 , and a point p =
r(u0 ) M = r(), at which the Gaussian curvature Kp = K(u0 ) negative.
Then the product of the two principal curvatures 1 , 2 at p is negative,
thus, the principal curvatures have opposite signs. We may assume without
loss of generality that 1 < 0 < 2 . Let e1 , e2 be unit vectors, pointing in
principal directions corresponding to 1 and 2 respectively. Since 1 and 2
are different eigenvalues of the Weingarten map, e1 is orthogonal to e1 .
The principal curvatures at a point are the extremal values of the normal
curvature function. Since they have opposite sign in our case, there must be
tangent directions at p, in which the normal curvature vanishes.
Definition 3.2.19. If p is a point on a surface, then the tangent directions at
p in which the normal curvature vanishes are called the asymptotic directions
at p.
We can find these directions easily with the help of Eulers formula. If the
direction angle of a unit vector v Tp M relative to orthonormal basis (e1 , e2 )
is , then v = cos()e1 + sin()e2 , and
kp (v) = cos2 () 1 + sin2 ()2 .
Solving the equation kp (v) = 0 for , we obtain
r
1
kp (v) = 0 tan() = .
2
Thus, there are two linearly independent asymptotic directions, arranged

symmetrically in the principal directions. In other words, the principal direc-
tions are the bisectors of the angles determined by the principal directions.
We described above how to find asymptotic direction at a point, where the
Gaussian curvature is negative. Asymptotic directions exist also at points
where the Gaussian curvature is 0, but at such points the picture we obtained
for the negative curvature points degenerates in a way. If 1 = 0 < 2 or
1 < 0 = 2 , then the two asymptotic directions coincide with the principal
direction corresponding to the 0 principal curvature. If 1 = 0 = 2 , then p
is a planar umbilical point and all tangent directions are asymptotic.
There are no asymptotic directions at a point where the Gaussian curvature
is positive.
Definition 3.2.20. A regular curve : I M on a surface M is called an
asymptotic curve if 0 (t) is an asymptotic direction for all t I.
Proposition 3.2.21. A regular curve : I M is an asymptotic curve on
the surface M if and only if 00 (t) is tangent to the surface at (t) for all
t I.
If 0 (t) and 00 (t) are linearly independent, then the second condition is equiv-
alent to the requirement that the osculating plane of at t coincides with the
tangent plane of the surface at (t).
Figure 3.9: Asymptotic curves on the pseudosphere.
Proof. The normal curvature can be expressed with the help of the funda-
mental forms
II(t) ( 0 (t), 0 (t)) h 0 (t), L(t) ( 0 (t))i
k(t) ( 0 (t)) = = .
k 0 (t)k2 k 0 (t)k2
Thus, 0 (t) points in asymptotic direction if and only if
h 0 (t), L(t) ( 0 (t))i = 0.
Denote by Np the unit normal of the surface at p. Differentiating the equation

h 0 (t), N 0 (t) i 0
with respect to t we obtain
h 00 (t), N 0 (t) i + h 0 (t), L(t) ( 0 (t))i = 0,
consequently, 0 (t) is a principal direction if and only 00 (t) is orthogonal to
N(t) .
Theorem 3.2.22 (BeltramiEnneper).
(a) Let : I M be an asymptotic line on a surface M in R3 . Then
h 0 00 , 000 i = K k 0 00 k2 ,
p
(3.8)
where the function K : I R assigns to t I the Gaussian curvature
of M at (t), is the sign of h 0 L ( 0 ), N i.
(b) In particular, if 0 and 00 are linearly independent at t, therefore the
torsion of is defined at t, then
q
(t) = K(t) .
(c) If two asymptotic curves meet at a point p with nonparallel speed vectors
and their torsions are defined at p, then their torsions at p are opposite
to one another.
Proof. (a) Since is an asymptotic curve, both 0 (t) and 00 (t) are in the
tangent plane at (t), consequently, 0 00 = 1 k 0 00 kN and
h 0 00 , 000 i = 1 k 0 00 khN , 000 i.
Where 1 is the sign of h 0 00 , N i.
Differentiating h 00 , N i 0, we obtain
D d E
h 000 , N i = 00 , N = h 00 , L ( 0 )i.
dt
Using the linear algebraic identity
L ( 0 ) L ( 00 ) = (L ( 0 ) L ( 00 )) = (det(L ) 0 00 ) = det(L ) 0 00
and the Lagrange identity we obtain
det(L )k 0 00 k2 = h 0 00 , L ( 0 ) L ( 00 )i
0
h , L ( 0 )i h 00 , L ( 0 )i

= det
h 0 , L ( 00 )i h 00 , L ( 00 )i
= h 00 , L ( 0 )i2 .
In the last step we used that h 0 , L ( 0 )i = 0 and h 00 , L ( 0 )i = h 0 , L ( 00 )i.

The first equation holds since 0 points in asymptotic direction, the second is
true because the Weingarten map is self-adjoint. Taking square root results
in
q
det(L )k 0 00 k = 2 h 00 , L ( 0 )i,
where 2 = sgnh 00 , L ( 0 )i. Combining the equations above we get

q
h 0 00 , 000 i = det(L )k 0 00 k2 ,
where = 1 2 .
If 0 (t) 00 (t) = 0 at a point, then equation (3.8) holds with any choice of
(t), as both sides are equal to 0. Assume that 0 (t) 00 (t) 6= 0. Then 00 (t)
can be decomposed into the orthogonal components
00 (t) = a 0 (t) + b(N(t) 0 (t)), where a, b R, b 6= 0,
hence 1 (t) is the sign of
h 0 (t) 00 (t), N(t) i = h 0 (t) b N(t) 0 (t) , N(t) i = bkN(t) 0 (t)k2 ,

while 2 (t) is the sign of
h 00 (t), L(t) ( 0 (t))i = hb N(t) 0 (t) , L(t) ( 0 (t))i

= bh 0 (t) L(t) ( 0 (t)), N(t) i.
Thus, (t) is the sign of h 0 (t) L(t) ( 0 (t)), N(t) i.

Part (b) of the statement follows directly from (a) and the formula for the
curvature of curves of general type.
As
p for part (c), it is clear from (b) that both torsions have absolute value
Kp , where Kp is the Gaussian curvature at p, so we have to show only
that the signs are opposite. By the description of the sign , we need to
show that if v and w are two nonparallel asymptotic directions at p, then
the vectors v Lp (v) and w Lp (w) have opposite orientation.
Let (e1 , e2 ) be an orthonormal basis of Tp M , consisting of principal directions
ordered so that e1 e2 = Np . Denote by 1 , 2 the corresponding principal
curvatures. If v = v1 e1 + v2 e2 , and w = w1 e1 + w2 e2 then v Lp (v) =
v1 v2 (2 1 )Np and wLp (w) = w1 w2 (2 1 )Np . On the other hand, since
the principal directions bisect the angle between the asymptotic directions,
w must be multiple of v1 e1 v2 e2 , therefore, v1 v2 and w1 w2 have opposite
signs.
3.2.5 Surfaces of Constant Negative Curvature

Surfaces of constant negative curvature are important because their local
geometry coincides with the geometry of the hyperbolic plane. In this section
we discuss some important facts on them.
When we scale a surface by a factor > 0, for example by a central homoth-
ety, the Gaussian curvature K changes to K/2 , so we can focus on surfaces
of constant Gaussian curvature K 1 without loss of generality.
Important results can be obtained by considering special parameterizations of
these surfaces, especially those for which the coordinate lines are asymptotic
lines and those for which the coordinate lines are lines of curvature. Both
types of parameterizations exist locally by Theorems 3.2.4 and 3.2.5.
Consider a regular parameterization r : R3 of a surface with Gaussian
curvature K 1 and assume that the coordinate lines are asymptotic
lines. We shall denote by u and v the first and second variables of the
parameterization respectively. In the standard notations ri , rij , gij , g ij bij
kij , ijk , Rijk
l
of hypersurface theory, the indices i, j, k run over 1, . . . , n 1.
In the case of surfaces in R3 , these indices are equal to 1 or 2. Below we shall
replace in each of these notations all indices by the variable symbols u and
v. Index 1 will be replaced by u, index 2 by v. For example, we shall write
vuv instead of 212 .
By the characterization of asymptotic lines, the condition on r is equivalent to
ruu N and rvv N. Therefore,
the matrix of the second fundamental form
0 C
is of the form B = . Let us introduce also the notation A = kru k,
C 0
B = krv k and denote by (u, v) (0, ) the angle between ru (u, v) and
rv (u, v). Then
A2

AB cos
G= (3.9)
AB cos B2
and by the expression of the Gaussian curvature
C 2 = det(B) = K det(G) = A2 B 2 sin2 (),
so C = AB sin(). Changing the order of the parameters u and v, N and

B changes to N and B, so we may assume without loss of generality that

0 AB sin
B= . (3.10)
AB sin 0
If there is a surface the fundamental forms of which have these special forms,
then it is certain, that the coordinate lines are asymptotic lines and the
surface has constant 1 Gaussian curvature. However, we know from the
fundamental theorem of curve theory, that a surface with given fundamental
forms exist (locally) if and only if the prescribed matrices satisfy the Gauss
and CodazziMainardi equations. These equations are not automatically
satisfied for the above matrices leading to some partial differential equations
that must be satisfied by the functions A, B and . Let us compute these
equations.
First we observe that the inverse of G is
1 cos

A2 sin2
1
AB sin2
G = .

cos 1

AB sin2 B 2 sin2
Lemma 3.2.23. If a regular parameterized surface of negative curvature has
the property that the coordinate lines are asymptotic lines and buv > 0, then
the CodazziMainardi equations reduce to the equations
u (log buv ) = vuv uuu , (3.11)
v (log buv ) = vvv uvu . (3.12)
Proof. The general form of the CodazziMainardi equation is
2
X
bij,k bik,j = (lik blj lij blk ).
l=1
Since both sides are skew-symmetric in the indices j and k, these equations
are true trivially when j = k and the equation for i, j, k is equivalent to the
equation for (i0 , j 0 , k 0 ) = (i, k, j). For this reason, when we have only two
variables u and v then there are only two independent CodazziMainardi
equations for (i, j, k) = (1, 1, 2) and (i, j, k) = (2, 1, 2). The equations for
these indices are
buu,v buv,u = (uuv buu uuu buv ) + (vuv bvu vuu bvv ),
bvu,v bvv,u = (uvv buu uvu buv ) + (vvv bvu vvu bvv ).
Taking into consideration buu = bvv = 0, these equations reduce to
buv,u = uuu buv + vuv bvu ,
bvu,v = uvu buv + vvv bvu .
Dividing these equations by buv = bvu we obtain the required equalities for
the logarithmic derivative of buv
Lemma 3.2.24. The CodazziMainardi equations are satisfied by the matri-
ces (3.9) and (3.10) if and only if
v A = u B = 0.
Proof. First we compute the right-hand side of (3.11)

vuv uuu = uvu g uv + uvv g vv uuu g uu uuv g vu
1 1 1 1
= guu,v g uv + gvv,u g vv guu,u g uu guv,u g vu + guu,v g vu
2 2 2 2
uv 1 vv 1 uu vu
= guu,v g + gvv,u g guu,u g guv,u g
2 2
u B 2v A cos u A
+ u log(AB cos ) cot2 .

= 2 2
B sin A sin
Thus equation (3.11) gives
u A u B
cot u
A B
u B 2v A cos u A
+ u log(AB cos ) cot2 ,

= 2 2
B sin A sin
which is equivalent to
2u B 2v A cos
0= ,
B sin2
after some rearrangement and simplifications. This gives u B = v A cos .
Flipping the role of the two variables, we obtain v A = u B cos similarly.
Since cannot be a multiple of , the last two equations imply that u B =
v A = 0. Conversely, reversing the steps of the computation, if u B = v A =
0, then the CodazziMainardi equations are fulfilled.
Definition 3.2.25. The net of the coordinate lines of a parametrization
r : R3 of a surface is called a Chebyshev net if the opposite sides of any
curvilinear quadrangle bounded by segments of the coordinate lines are of
equal length.
Lemma 3.2.26. The coordinate lines of a regular parametrization r : R3
form a Chebyshev net if and only if v kru k = u krv k 0.
Proof. The parameter lines of r form a Chebyshev net if and only if for any
rectangle [u0 , u1 ] [v0 , v1 ] , we have
Z u1 Z u1
kru k(u, v0 )du = kru k(u, v1 )du
u0 u0
Z v1 Z v1 (3.13)
and krv k(u0 , v)dv = krv k(u1 , v)dv.
v0 v0
The first equation means that the integral

Z u1
I(v) = kru k(u, v)du
u0
is constant on any interval [v0 , v1 ] on which it is defined. The latter condition

holds if and only if
Z u1
I 0 (v) = v kru k(u, v)du 0
u0
for any v u0 , u1 such that [u0 , u1 ] {v} . This is clearly equivalent to

the vanishing of v kru k.
Changing the role of the parameters we can see that the second condition in
(3.13) is equivalent to u krv k 0.
Thus, we obtained as a consequence of Lemma 3.2.24 (the CodazziMainardi

equations) that the asymptotic lines on a surface of constant Gaussian curva-
ture 1 form a Chebyshev net. This means that we can reparameterize the
surface in a way that all coordinate lines are asymptotic curves parameterized
by arc length. Choosing such a parameterization, we shall have A B 1,
and the matrices G and B take the simpler form

1 cos 0 sin
G= , B= .
cos 1 sin 0
Let us compute now what the Gauss equation gives for a surface with these
fundamental form matrices. The inverse of G is
1 cos

sin2
sin2
G 1 = ,

cos 1
2 2
sin sin
the Christoffel symbols of the second kind are
1
uuu = guu,u = 0,
2
1
uuv = (2guv,u guu,v ) = sin u ,
2
1
uvu = (guv,u + guu,v guv,u ) = 0,
2
1
uvv = (guv,v + gvv,u guv,v ) = 0,
2
1
vvu = (2guv,v gvv,u ) = sin v ,
2
1
vvv = gvv,v = 0,
2
and the Christoffel symbols of the first kind are
uuu = uuu g uu + uuv g vu = cot u ,

u
vuu = uuu g uv + uuv g vv = ,
sin
uuv = uvu g uu + uvv g vu = 0,
vuv = uvu g uv + uvv g vv = 0,
v
uvv = vvu g uu + vvv g vu = ,
sin
vvv = vvu g uv + vvv g vv = cot v .
The only nontrivial Gauss equation for this surface is
sin2 = det B = Ruvvu = Ruvv

u v
guu + Ruvv u
gvu = Ruvv v
+ Ruvv cos ,
where the components of the curvature tensor are

X
u
Ruvv = uvv,u uvu,v + (svv usu svu usv ) = uvv,u + uvv uuu
s{u,v}
cos u v uv sin cos u v uv
= = ,
sin2 sin2 sin
and
v
X
v
Ruvv = vvv,u vvu,v + (svv vsu svu vsv ) = vvv,u + uvv vuu
s=u
u v u v
= + uv cot + = uv cot .
sin2 sin2
Substituting these into the Gauss equation leads to
cos2 u
sin2 = uv = uv sin ,
sin
hence satisfies the partial differential equation uv = sin .
Definition 3.2.27. The partial differential equation uv = sin is called
the sine-Gordon equation.
We can sum up our results as follows.
Theorem 3.2.28. A surface of constant Gaussian curvature 1 can be
parameterized in such a way that the coordinate lines are the asymptotic
lines parameterized by arc length. For such a parameterization the angle
: (0, ) between the asymptotic lines gives a solution of the sine-Gordon
equation.
Conversely, if is convex and a function : (0, ) solves the sine-

Gordon equation, then there is a regular parameterized surface r : R3
with fundamental form matrices

1 cos 0 sin
G= and B = .
cos 1 sin 0
This surface has constant Gaussian curvature 1 and the coordinate lines of
the parametrization are unit speed asymptotic curves.
Roughly speaking, surfaces with constant Gaussian curvature 1 are in one
to one correspondence with solutions of the sine-Gordon equation with values
in the interval (0, ).
The integrated form of the sine-Gordon equation is the Hazzidakis formula.
Theorem 3.2.29 (Hazzidakis Formula). Let r : R3 be a surface of
constant Gaussian curvature 1 parameterized as above. Then the surface
area A of the curvilinear quadrangle bounded by asymptotic lines obtained as
the r image of the rectangle [u0 , u1 ] [v0 , v1 ] is
(u1 , v1 ) (u0 , v1 ) (u1 , v0 ) + (u0 , v0 ).
Proof. By the definition of the surface area

Z v1 Z u1 p Z v1 Z u1
A= det(G(u, v)dudv = sin((u, v))dudv.
v0 u0 v0 u0
As solves the sine-Gordon equation,

Z v1 Z u1 Z v1 Z u 1
sin((u, v))dudv = uv (u, v))dudv
v0 u0 v u0
Z v1 0
= (v (u1 , v) v (u0 , v))dv
v0
= (u1 , v1 ) (u0 , v1 ) (u1 , v0 ) + (u0 , v0 ).
Corollary 3.2.30. The area A of any curvilinear quadrangle bounded by

asymptotic lines is less then 2. The function satisfies the inequality
2 > (u1 , v1 ) (u0 , v1 ) (u1 , v0 ) + (u0 , v0 ) > 0. (3.14)
Definition 3.2.31. Let r : R3 be a regular parameterized surface and

define for u0 and R > 0 the set B(u0 , R) as the set of those points
v for which there is a smooth curve u : [0, 1] connecting u0 = u(0)
to v = u(1) so that the length of r u is less then R. The surface r is said
to be complete if for any u0 and any R > 0, the closure of B(u0 , R) is
contained in .
Intuitively, completeness means that moving along a curve of finite length on

the surface we cannot reach (converge to) points that do not belong to the
surface.
Theorem 3.2.32 (Hilberts Theorem). There is no complete regular surface

of constant negative Gaussian curvature in R3 .
Proof. Suppose to the contrary that such a surface exists. Then by com-
pleteness, for any p r() both maximal asymptotic curves starting at
p have a parameterization by arc length defined on the whole real line. Let
p : R r() be a unit speed parameterization of one of the asymptotic lines
starting at (0) = p, and for each t R, denote by p (t) : R r() the unit
speed parameterization of the asymptotic line starting at p (t) (0) = p (t).
Orient the initial speed of p (t) so that 0 (t) 0 p (t) (0) be a positive mul-
tiple of the normal vector Np (t) for all t. Then the map r : R2 R3 ,
r(u, v) = p (u) (v) is a regular parameterized surface with constant Gaussian
curvature 1, for which the coordinate lines are naturally parameterized
asymptotic curves. (We do not claim that r is injective, nor that it has the
same image as r.) This map r would define a solution of the sine-Gordon
equation which would take values in the interval (0, ). However, as the next
lemma states, there are no such solutions.
Lemma 3.2.33. The sine-Gordon equation has no solution on the band

[0, 1] R for which 0 < < .
Proof. Assume to the contrary that such a solution exists. We discuss three
cases separately. In the first case, suppose that (1, 0) > (0, 0). Set c =
((1, 0) (0, 0))/3 and let
u1 = max{s [0, 1] : (s, 0) (0, 0) = c},

u2 = min{s [0, 1] : (1, 0) (s, 0) = c}.
Then it is clear that 0 < u1 < u2 < 1 and for any t [u1 , u2 ], we have
c + (0, 0) (t, 0) (1, 0) c
Let t [u1 , u2 ] and v > 0 be arbitrary. Then applying inequality (3.14), we

obtain
(t, v) > (t, 0) + (0, v) (0, 0) > (t, 0) (0, 0) c,
and
(t, v) < (1, v) + (t, 0) (1, 0) (1, v) c < c.
The last two inequalities imply that sin((t, v))

sin(c) for (t, v) [u1 , u2 ] [0, +). Integrating on
the rectangle [u1 , u2 ] [0, v] gives
Z v Z u2
sin((u, v))dudv v(u2 u1 ) sin(c).
0 u1
On the other hand, the integral is given by the Hazz-

idakis formula and
Z v Z u2
sin((u, v))dudv
0 u1
= (u2 , v) (u1 , v)
(u2 , 0) + (u1 , 0) < 2.
However, this is a contradiction, because the inequal-

ity 2 > v(u2 u1 ) sin(c) cannot be true for any
positive v.
Figure 3.10
If (0, 0) > , then consider the function (u, v) = (1u, v). The function
is also defined on [0, 1] R, it solves the sine-Gordon equation, takes its
values in the interval (0, ), but (1, 0) > (0, 0), so the second case can be
reduced to the first case.
Finally, assume (0, 0) = (1, 0). Then introduce the function (u, v) =
(u, v + 1). The function is again defined on [0, 1] R, it solves the sine-
Gordon equation, takes its values in the interval (0, ), furthermore, using
inequality (3.14)
(1, 0) (0, 0) = (1, 1) (0, 1) > (1, 0) (0, 0) = 0.
Applying the arguments of the first case to we obtain a contradiction.
Backlund transform
A Backlund transform in contemporary mathematics is a method to find
new solutions to certain partial differential equations if one particular solu-
tion is already known. Backlund transforms have their origins in differential
geometry. Namely, in the 1880s L. Bianchi and A.V. Backlund introduced
a nontrivial geometrical construction of new surfaces of constant negative
Gaussian curvature from an initial such surface using a solution of a first or-
der partial differential equation. As surfaces of constant negative curvature
can be described by solutions of the sine-Gordon equation, Backlund trans-
formation of surfaces can be viewed as a transformation of solutions of the

sine-Gordon equation.
Definition 3.2.34. Let r : R3 be a surface of constant Gaussian curva-
ture 1. We say that another regular parameterized hypersurface r : R3
is a Backlund transform of r with parameter if
(1) kr(u, v) r(u, v)k cos ,
(2) r(u, v) r(u, v) is tangent to r at r(u, v) and r at r(u, v),

(3) the angle between N and N is 2 .
Bianchi transform of a surface is a Backlund transform with parameter 0.
To describe Backlund transforms of a surface, it is convenient to reparam-
eterize it with the lines of curvature as the parameter lines. If r is the
parameterization of the surface for which the parameter lines are naturally
parameterized asymptotic curves, then r(u, v) = r( u+v vu
2 , 2 ) is a reparame-
1 u+v vu 1 u+v vu
terization such that ru (u, v) = 2 ru ( 2 , 2 )+ 2 rv ( 2 , 2 ) and rv (u, v) =
12 ru ( u+v vu 1 u+v vu
2 , 2 ) + 2 rv ( 2 , 2 ) point in principal directions. Denote by
the function defined by (u, v) = 12 ( u+v vu
2 , 2 ), where is the angle between
ru and rv . As
1
(uu vv )(u, v) = ([uu + 2uv + vv ]
8
u+v vu
[uu 2uv + vv ]) ,
2 2

u+v vu
= 2uv , ,
2 2
the sine-Gordon equation for can be rephrased in terms of as follows
uu vv = sin cos .
The matrices of the fundamental forms of the parameterization r are
2
cos 0 sin cos 0
G = and B = .
0 sin2 0 sin cos
Computing the Christoffel symbols from G we can express the second order
partial derivatives of r as linear combinations of the Gauss frame. We omit
the details of this mechanical computation, and give just the result
ruu = tan u ru + cot v rv + sin cos N

ruv = tan v ru + cot u rv (3.15)
rvv = tan u ru + cot v rv sin cos N
From the first two conditions, a Backlund transform of r must be of the form

ru rv
r = r + cos cos + sin ,
cos sin
where is an unknown angle function which encodes the direction of the

vector rr. Let us compute the restrictions must obey in order to guarantee
that r is a Backlund transformation of r.
Applying the formulas in (3.15), one can calculate the expressions

sin cos
ru = 1 cos (u + v ) ru + cos (u + v ) rv
cos sin
h i
+ cos sin cos N

sin cos
rv = cos (v + u ) ru + 1 + cos (u + v ) rv
cos sin
h i
cos sin cos N
for the partial derivatives of r.

Since N must be orthogonal to r r and should enclose angle 2 with N,
it must be of the form

ru rv
N = cos sin + cos + sin N
cos sin
The angle function of the transformation is restricted by the conditions

hN, ru i = hN, rv i = 0. Computing these dot products

0 = hN, ru i = cos sin cos + cos (u + v ) + sin sin cos

0 = hN, rv i = cos sin cos + cos (u + v ) sin sin cos
We can assume cos 6= 0 otherwise the Backlund transform coincides with

the original surface and this case is not of interest. Then rearranging the
equations for we obtain
sin cos sin cos sin

u + v =
cos (3.16)
sin cos sin cos sin
u + v =
cos
By Frobenius Theorem (Theorem 1.7.10), this system of partial differential
equations has a unique solution for in the a neighborhood of any point
(u0 , v0 ) with any initial condition (u0 , v0 ) = 0 if and only if the equa-
tions do not contradict the Young theorem uv = vu . As
(u )v = vv
cos cos v sin sin v sin ( sin sin v + cos cos v )
+ ,
cos
(v )u = uu
cos cos u + sin sin u sin ( sin sin u + cos cos u )
+ ,
cos
the integrability condition for is
cos cos sin sin sin

0 = uu vv + (u + v )
cos
sin sin sin cos cos
+ (u + v )
cos
cos cos sin sin sin sin cos sin cos sin
= uu vv +
cos cos
sin sin sin cos cos sin cos sin cos sin
+
cos cos
= uu vv sin cos .
However, this condition is fulfilled, since this is essentially the sine-Gordon

equation written in terms of .
This computation proves the existence theorem for Backlund transforms. Be-
fore we formulate the theorem, observe, that the sign can be incorporated
with as its sign.
Theorem 3.2.35. For any choice of the point (u0 , v0 ) , the parameter
[/2, /2], a vector v of length cos tangent to r at (u0 , v0 ), there is
a unique Backlund transform r of r, defined in a neighborhood of (u0 , v0 ),
which shifts r(u0 , v0 ) to r(u0 , v0 ) = r(u0 , v0 ) + v.
Theorem 3.2.36. Let r be a parameter Backlund transform of r. Then
r is regular at (u, v) if and only if r(u, v) r(u, v) is not a principal

direction of r.
The Backlund transform r has constant Gaussian curvature 1.
the parameter lines of r are the lines of curvature of the Backlund trans-
form.
The angle of the asymptotic curves of r is 2. Parameter lines of the

reparameterization r(u+v, v u) are the asymptotic curves of the Back-
lund transform. That is, the natural correspondence between the points
of the original surface and its Backlund transform moves lines of curva-
ture to lines of curvature, and asymptotic curves to asymptotic curves.
Proof. A lengthy but elementary computation shows that the matrix of the
first fundamental form of r is
2
cos 0
G =
0 sin2
This shows that singular points of the Backlund transform are the points
where is an integer multiple of /2, which means that r r is parallel to
one of the principal directions ru , rv .
Another use of the knowledge of G is that we can compute the Gaussian
curvature from it by the Theorema Egregium. It gives that the Gaussian
curvature of r is
vv uu
K = .
sin cos
K 1 is equivalent to the variant of the sine-Gordon equation uu vv =
sin cos . We saw that the sine-Gordon equation for was the integrability
condition for the partial differential equation (3.16) for the unknown . The
functions and enter into (3.16) in a symmetric way. Since is obtained
as a solution of (3.16) for a given , (3.16) does have a solution in given .
Thus satisfies the integrability condition uu vv = sin cos .
To compute the matrix of the second fundamental form of r we first compute
partial derivative of N with respect to v

cos sin
Nv = cos (v + u )ru (v + u )rv cos cos N
cos sin
+ sin Nv
From this, again some simple but tedious computation shows hru , Nv i = 0,
which means that for the parametrization r of the Backlund transform the
coordinate lines are lines of curvature.
With some more work we can compute also hrv , Nv i = sin cos . Since
the B is symmetric and its determinant is K det G = cos2 sin2 , B must
be of the form
sin cos 0
B = .
0 sin cos
This gives at once that ru rv point in asymptotic directions, from which
the rest of the theorem follows.
3.2.6 Minimal Surfaces

When a wire frame is dipped into a soap solution, and pulled out, in most
cases a soap film is created. When a soap film is stable, its surface area is
locally minimal in the sense that small perturbations of the soap film do not
decrease the surface area. Our goal below is to find a differential geometric
consequence of local minimality. For this purpose, we shall use the notion of
variation of a parameterized hypersurface.
Definition 3.2.37. A smooth variation of a regular parameterized hypersur-
face r : Rn is a smooth map R : (, ) Rn , such that for all
t (, ), the map rt : Rn defined by rt (u) = R(u, t) is a regular
parameterized hypersurface and r0 = r.
A variation is compactly supported if there is a compact set K , such that
R(u, t) = r(u) for all u
/ K and t (, ).
For each point r(u) of the hypersurface, the variation R defines a parameter-
ized curve R(u, .) : (, ) Rn which describes a motion of that point. The
image of the parameterized hypersurface rt consists of those points at which
the points of the original hypersurface r arrive after moving on for time t.
Assigning to each point of the hypersurface the initial speed vector of its mo-
tion, we obtain a vector field along the hypersurface, called the infinitesimal
variation associated to the variation R.
Definition 3.2.38. The infinitesimal variation associated to the variation
R is the vector field X = t R(., 0) : Rn along the hypersurface r, where
t R = n R is the partial derivative of R with respect to the nth variable,
(which we denote by t now).
If the variation R is supported by a compact set K, then X vanishes outside
K, so X has compact support.
Theorem 3.2.39 (First Variation of the Surface Volume). Assume that a
variation R : (, ) Rn of a hypersurface r : Rn is supported by
a compact set K . Denote by X the infinitesimal variation associated to
R, by H the Minkowski curvature of r, and by G the matrix of the first fun-
damental form of r with respect to the basis r1 , . . . , rn1 . Then the derivative
of the surface volume Vn1 (t) of rt (K) is

Z
0
Vn1 (0) = (n 1) hX, NiH det G dn1 .
K
Proof. By the definition of the surface volume,

Z p
Vn1 (t) = det Gt dn1 ,
K
where Gt is the Gram matrix of the partial derivatives rt,1 , . . . , rt,n1 . Dif-
ferentiating with respect to t at t = 0, we obtain
Z Z
0 dp d
Vn1 (0) = det Gt |t=0 dn1 = krt,1 rt,n1 kt=0 dn1 .
K dt K dt
(3.17)
Applying the chain rule, the Leibniz rule for the differentiation of dot and
wedge products, and Youngs theorem, we get

d
d dt rt,1 rt,n1 t=0 , r1 rn1
krt,1 rt,n1 k t=0 =

dt kr1 rn1 k
DP E
n1
i=1 r 1 r i1 t i R(., 0) r i+1 rn1 , r1 rn1
=
kr1 rn1 k
DP E
n1
i=1 r1 ri1 Xi ri+1 rn1 , r1 rn1
= ,
kr1 rn1 k
and
0
Vn1 (0)
DP E
n1
Z
= dn1 .
K kr1 rn1 k
(3.18)
Decompose X into the sum of a vector field X > tangential to the hyper-
surface r and a vector field X orthogonal to it. Obviously, the orthogonal
component is X = hX, NiN. The key observation for the proof of the theo-
rem is that the tangential component X > has no contribution to the integral
in (3.18), that is,
DP E
n1 >
Z
i=1 r1 ri1 X i r i+1 rn1 , r1 rn1
dn1
K kr1 rn1 k
= 0.
To prove this fact, decompose X > P as a linear combination of the vector

n1
fields ri . The coefficients in X > = i=1 F i ri are smooth functions on ,
thus, they yield a smooth vector field F = (F 1 , . . . , F n1 ) : Rn1 on
. F vanishes outside K, so a maximal integral curve of the differential
equation 0 (t) = F ((t)) is constant if it starts at a point (0) / K and it
remains in K if it starts at a point of K. This means that the flow of the
differential equation consists of diffeomorphisms t : for t R such
that t (K) = K.
Consider the variation R : R Rn , R(u, t) = r(t (u)) of the hypersur-

face r. The hypersurfaces rt = R(., t) = rt are simply reparameterizations
of the initial hypersurface r. Thus, by Proposition 3.1.38, the surface volume
Vn1 (t) of rt (K) is constant, in particular, its derivative with respect to t is
0.
The infinitesimal variation of r associated to the variation R is the vector
field
d
t R(., 0) = (r t )|t=0
dt 1
F
0
d
0
. n1
X
= (r 0 ) t |t=0 = r .. = F i ri = X > .
dt n1 i=1
F
Thus, the application of equation (3.18) to the variation R yields

0
0 = Vn1 (0)
DP E
n1
Z
i=1 r1 ri1 Xi> ri+1 rn1 , r1 rn1
=
K kr1 rn1 k
as we claimed. If we omit the vanishing contribution of X > from (3.18), then

0
Vn1 (0)
DP E
n1
Z
= dn1
K kr1 rn1 k
remains. The partial derivatives of X = hX, NiN are Xi = i (hX, Ni)N +

hX, NiNi . As N is orthogonal to all partial derivatives of r,
hr1 ri1 N ri+1 rn1 , r1 rn1 i = 0
by equation (1.9). Write Ni = L(ri ), where L is the Weingarten map of r

as a linear combination
n1
X
Ni = L(ri ) = lij rj
j=1
of the partial derivatives of r. The coefficients lij in the decomposition are

the entries of the matrix of L with respect to the basis r1 , . . . , rn1 . Then
n1 n1
!
X X
r1 ri1 Ni ri+1 rn1 = lii r1 rn1
i=1 i=1
The sum on the right-hand side is tr L = (n 1)H. Substituting back the

results of these partial computations we obtain
(n 1)HhX, Ni hr1 rn1 , r1 rn1 i

Z
0
Vn1 (0) = dn1
K kr1 rn1 k
Z
= (n 1) hX, NiHkr1 rn1 k dn1
ZK

= (n 1) hX, NiH det G dn1 .
K
The theorem gives a necessary condition for a hypersurface to minimize the

surface volume among hypersurfaces with given boundary.
Corollary 3.2.40. Let r : Rn be a regular parameterized hypersurface.
Assume that for any compact subset K , the surface volume of r(K)
is minimal among the surface volumes of r(K), where r : Rn is run-
ning over all regular hypersurfaces coinciding with r outside K. Then r has
constant Minkowski curvature 0.
Proof. Suppose to the contrary that H(u0 ) 6= 0 for a point u0 . As H

is continuous, we can find a closed ball K = B(u0 , r) around u0 such
that H is positive on K. Let h : Rn1 [0, 1] be a smooth function, which
is constant 1 on the ball B(u0 , r/2) and 0 outside K.
Consider the variation R(u, t) = r(u) + t h(u)N(u) of r. It is supported by
the compact set K, and its associated infinitesimal variation is hN. Applying
Theorem 3.2.39 to this variation we obtain

Z
0
Vn1 (0) = (n 1) hH det Gdn1 .
K
The function hH is 0 outside K, non-negative inside K and strictly positive

on a ball around u0 , thus the value of the integral is negative, in particular,
0
Vn1 (0) < 0. This means that if t > 0 is small enough, then the surface
volume of rt (K) will be strictly smaller than that of r(K), which contradicts
the minimality condition on r.
The corollary motivates the following definition.

Definition 3.2.41. A regular parameterized hypersurface is called a minimal
hypersurface if its Minkowski curvature is constant 0.
Exercise 3.2.42. Show that a connected surface of revolution in R3 is a
minimal surface if and only if it is part of a plane orthogonal to the axes of
rotation or a that of a catenoid. (A catenoid is any surface similar to the
catenoid considered in Exercise 3.2.17. (b).)
A union of some disjoint closed curves in R3 can bound several minimal

surfaces having different surface areas. Thus, among surfaces with a given
boundary, not all minimal surfaces minimize the surface area.
Figure 3.11: The catenoid.
Exercise 3.2.43. Assume that C1 and C2 are circles of radius r in parallel

planes and that the straight line through their centers O1 and O2 is orthogonal
to their planes. Show that if d(O1 , O2 )/r is small enough, then there exists
two non-congruent catenoid rings bounded by C1 C2 .
Chapter 4
Manifolds
4.1 Topological and Differentiable Manifolds

4.1.1 Basic Definitions
In the following chain of definitions n is a fixed positive integer.
Definition 4.1.1. Let X be an arbitrary set. A local parameterization of

X is an injective mapping : X from an open subset of Rn onto a
subset of X. The inverse 1 : () of such a parameterization is called
a chart because through 1 the region im X is charted on U Rn ,
just as a region of the earth is charted on a topographic or a political map.
1 is also called a local coordinate system because through 1 each point
p im corresponds to an n-tuple of real numbers, the coordinates of p.
Definition 4.1.2. An atlas on X is a collection of charts AS = {i : i I}

such that every point is represented in at least one chart i.e. iI dom i =
X.
Definition 4.1.3. Two charts 1 : U1 V1 and 2 : U2 V2 are said to

be C r -compatible if the domains 1 (U1 U2 ) and 2 (U1 U2 ) of the transit
mappings 2 1 1 n
1 and 1 2 are open subsets of R , and 2 1 and
1
1
1 2 are r times continuously differentiable. (A mapping is 0 times
continuously differentiable if it is continuous).
Definition 4.1.4. An atlas is C r -compatible if any two charts in the atlas

are C r -compatible.
201
202 4. Manifolds
A C 0 -compatible atlas A on a set X defines a topology on X as follows. Let

U X be open if and only if (U dom ) is open in Rn with respect to any
chart from A.
Proposition 4.1.5. The family of open sets yields a topology on X. The
domains of the charts of A are open subsets of X. Each chart A is a
homeomorphism between its domain and image with respect to these topolo-
gies.
Proof. The first part of the statement follows directly from the following set
theoretical identities.
(i) ( dom ) = ,
(ii) (X dom ) = im ,
(iii) (U V dom ) = (U dom ) (V dom ),

S S
(iv) (( iI Ui ) dom ) = iI (Ui dom ).
The rest of the statement is also a straightforward consequence of the defini-

tions.
Corollary 4.1.6. The topology of a topological space X is induced by a C 0 -

compatible atlas if and only if each point of X has a neighborhood homeomor-
phic to Rn .
Definition 4.1.7. An n-dimensional topological manifold is a pair (X, A)
consisting of a point set X and a C 0 -compatible atlas A on it, such that the
topology induced by the atlas on X is Hausdorff and second countable. The
Figure 4.1: The transit map between two charts.

4.1. Topological and Differentiable Manifolds 203
topological manifolds (X, A1 ) and (X, A2 ) are considered to be the same if

A1 and A2 induce the same topology on X, i.e. if the combined atlas A1 A2
is C 0 -compatible.
Among all atlases defining the same topological manifold structure on a set
X, there is a unique maximal atlas.
Definition 4.1.8. A C r -manifold is a set X equipped with a C r -compatible
atlas A turning X into a topological manifold. Two atlases define the same
C r -manifold structure on X if their union is also C r -compatible.
It is clear that the class of atlases defining the same C r -manifold structure
on X contains a unique maximal atlas.
Definition 4.1.9. A mapping f : X Y from a C r -manifold (X, A) into
the C r -manifold (Y, B) is said to be r times continuously differentiable or of
class C r if for any two charts A and B, the mapping f 1 is r
times continuously differentiable. The map f is a C r -diffeomorphism if it is
a bijection and both f and f 1 are of class C r .
Definition 4.1.10. Two C r -manifolds are C r -diffeomorphic if there is a C r -
diffeomorphism between them.
As before, the adjectives smooth and differentiable will be used in the sense
infinitely many times differentiable or C . Unless otherwise stated, manifolds
and mappings between them are assumed to be smooth.
Examples.
Rn equipped with the atlas consisting of only one chart, the identity mapping
of Rn , is an n-dimensional differentiable manifold.
Open subsets U X of an n-dimensional C r -manifold (X, A) become n-
dimensional C r -manifolds with the atlas {dom()U : A}.
If (X1 , A1 ) and (X2 , A2 ) are two C r -manifolds of dimension n and m respec-

tively, then the product space X1 X2 has a natural (n + m)-dimensional
C r -manifold structure given by the atlas
{(1 2 ) : dom(1 ) dom(2 ) Rn+m : 1 A1 , 2 A2 }.
We have introduced the topology on the Grassmann manifolds Gr(n, k) in

Section 1.4. The topologies of these spaces come from a k(n k)-dimensional
differentiable manifold structure. We construct a chart B on Gr(n, k) to
every ordered basis B = (x1 , . . . , xn ) of Rn . Let us denote by V the subspace
spanned by the first k vectors of B and by W the subspace spanned by the
204 4. Manifolds
last (n k) vectors. It is clear that Rn = V W. Denote by : Rn V the

projection of Rn onto V along W . The chart B will be defined on the set
dom(B ) = {L Gr(n, k) : L W = {0}}.
B assigns to L dom(B ) a k (n k) matrix in the following way.

The restriction of onto L is a linear isomorphism between L and V . The
preimages of the vectors x1 , . . . , xk yield a basis of L which has the form
n
X
( L )1 (xi ) = xi +

ij xj .
j=k+1
It is clear that setting B (L) equal to the matrix of coefficients (ij ), i =

1, . . . , k; j = k + 1, . . . , n, we obtain a bijection between dom(B ) and the
set of all k (n k) matrices. The family of all charts of the form B is a
C -compatible atlas on Gr(n, k).
Gr(n, k) is a compact manifold, it can be covered by a finite number of charts.
Indeed we get a finite atlas on Gr(n, k) if we let B run through different
permutations of the standard basis of Rn . The Grassmann manifold Gr(n +
1, 1) is the n-dimensional projective space. The geometrical way to introduce
projective spaces is the following. We take an n-dimensional Euclidean space
and add to it a collection of extra points, called ideal points or points at
infinity, in such a way, that we attach one point at infinity to each straight
line and two straight lines get the same point at infinity if and only if they are
parallel. If we put the n-dimensional space into the (n + 1)-dimensional one
and fix a point O outside it, then every straight line through O intersects the
projective closure of the n-dimensional Euclidean space in a unique ordinary
or ideal point and this is the natural correspondence between the two ways
of introducing projective spaces.
Definition 4.1.11. A Lie group is a differentiable manifold G with a group

operation such that the mapping
G G G, (x, y) 7 x y 1
is differentiable.
Example. GL(n, R) and GL(n, C) are open subsets in the linear spaces of
all n n real or complex matrices, hence they have a differentiable manifold
structure. They also have a group structure, which is smooth since the entries
of the quotient of two matrices are rational functions of the entries of the
original matrices and rational functions are smooth. This way, general linear
groups are Lie groups.
4.1.2 Configuration Spaces

Many examples of manifolds can be obtained as the configuration space of a
mechanical system.
As the motion of a particle in R3 corresponds to a parameterized space curve,
a motion of a system of n points can be described by n parameterized curves
xi : [a, b] R3 , i = 1, 2, . . . , n.
Putting these mappings together, we obtain a curve
(x1 , . . . , xn ) : [a, b] R3 R3 R3 (n times)
in the direct product of n copies of R3 , the projections of which on the ith

factor of the product is just the curve xi .
For there is a one-to-one correspondence between points of the product R3
R3 R3 (n times) and the possible configurations of n points in the
space, we can call the direct product of n copies of R3 the configuration space
of the system of n points.
In general, the configuration space of a mechanical system is the set of all of
its possible positions, equipped with some natural additional structures such
as topology or the structure of a differentiable manifold.
The advantage of introducing the configuration space is that the motion of
the system can be interpreted as one single curve in the configuration space
instead of a set of space curves.
Non-trivial examples can be obtained by putting some constraints on a system
of n points. For example, some pairs of points can be connected by a rigid
rod, some points can be fixed or forced to move along a line or a surface.
Further constraints can be obtained by specifying the type of joint at the
points where two or more rods meet.
The configuration space of a system of n points with constraints is a sub-
space of R3n and it is quite natural to furnish it with the subspace topology
inherited from R3n .
Remark. In physics, the dimension of the configuration space of a mechan-
ical system (provided that it is a manifold) is called the number of degrees of
freedom.
Examples.
The configuration space of the planar pendulum is the circle S 1 .
The configuration space of the spherical pendulum is the two-dimensional

sphere S 2 .
The configuration space of a planar double pendulum is the direct prod-

uct of two circles, i.e. the torus T 2 = S 1 S 1 .
206 4. Manifolds
The configuration space of a spherical double pendulum is the direct

product of two spheres S 2 S 2 .
A rigid segment in the plane has for its configuration space the direct
product R2 S 1 , which is homeomorphic to the open solid torus.
In the above examples, the configuration space carries a manifold topology.

Theorem 4.1.12 below explains why the configuration space of a system of n
points with constraints so often happens to be a manifold.
4.1.3 Submanifolds of Rn
Theorem 4.1.12. Let F : Rn Rk be a smooth mapping, the image of
which contains 0 Rk . Consider the preimage X = {x Rn : F (x) = 0} of
the point 0.
Let us suppose that the gradient vectors
grad fi (x) = (1 fi (x), 2 fi (x), . . . , n fi (x))
of the coordinate functions of F = (f1 , f2 , . . . , fn ) are linearly independent at

each point x of X.
Then X is an (n k)-dimensional topological manifold, furthermore, there is
a well-defined differentiable manifold structure on X.
Remark. The condition on the independence of the gradient vectors of the
coordinate functions is essential. By a theorem of Whitney, for any closed
set C Rn there exists a smooth function f on Rn such that C = {x Rn :
f (x) = 0}.
Proof. The linear space Rnk can be embedded into Rn by the mapping
: (x1 , x2 , . . . , xnk ) 7 (0, . . . , 0, x1 , x2 , . . . , xnk ).

| {z }
k zeros
Consider the set A of those diffeomorphisms : V U between open subsets

of Rn through which the set X V is mapped onto the intersection (Rnk )
U.
Put
A=
{ : has the form = 1 XV : X V Rnk , where A, dom = V }.

Obviously, elements of A define a homeomorphism between open subsets of

X and that of Rnk . It is also clear from the construction that the mappings
1 , defined on (dom dom ), are smooth for any , A. In such

a way, to prove that A is a C -compatible atlas on X one has only to check
that each point of X is represented in at least one chart belonging to A.
For this purpose, consider an arbitrary point x of X and the gradient vectors
of the coordinate functions of F at x. Since they are linearly independent, we
can obtain a basis by joining further n k vectors a1 , a2 , . . . , ank to them.
The gradient vector of the linear function gi (x) = hai , xi is ai at any point,
consequently, the inverse function theorem (Theorem 1.5.12) can be applied
to the mapping
= (f1 , . . . , fk , g1 , ..., gnk ) : Rn Rn
at the point x. According to the theorem, is a diffeomorphism between

a neighborhood V of x and an open neighborhood of (x). Denote by
the restriction
of onto V . The mapping belongs to A and the chart
1 (V M ) is defined in a neighborhood of x, so the proof is finished.
Exercise 4.1.13. Check that the topology of (X, A) coincides with the sub-
space topology inherited from Rn , consequently, it is Hausdorff and second
countable.
Definition 4.1.14. We say that X Rn is an embedded (n k)-dimensional

submanifold in Rn if in a neighborhood U of every point x X there are func-
tions f1 , f2 , . . . , fk : U R such that the intersection U X is given by the
equations f1 = f2 = = fk = 0 and the vectors grad f1 (x), . . . , grad fk (x)
are linearly independent.
The study of higher dimensional manifolds was launched at the beginning

of the 20-th century by H. Poincare (1854 - 1912). At that time topology
was in its cradle and the abstract definition of a topological space had not
been created. Poincare worked with embedded submanifolds in Rn . This was
not a real loss of generality since by Whitneys theorem every differentiable
manifold of dimension n can be embedded into R2n as a submanifold.
Examples. As an application, consider the set S n1 of points in Rn the

distance of which from the origin is equal to one. They are characterized by
the equality
f (x1 , . . . , xn ) = x21 + + x2n 1 = 0.
The gradient vector of f at the point x Rn is just 2x. It is zero only at the
origin, which does not belong to S n1 , so S n1 is a topological manifold with
a natural differentiable structure on it. S n1 is called the (n1)-dimensional
sphere with the standard differentiable structure.
208 4. Manifolds
Most important Lie groups are obtained as closed subgroups of GL(n, R),
defined by some equalities on the matrix entries. For example, GL(n, C) is
isomorphic
to the subgroup of GL(2n, R) consisting of matrices of the form
A B
, where A, B are nn matrices. The following example illustrates
B A
how we can prove that a closed subgroup of GL(n, R) is a Lie group. (Remark
that according to Cartans theorem, every closed subgroup of a Lie group is
a Lie group, therefore the computation below is an elementary verification of
a special case of Cartans theorem.)
Consider the orthogonal group
O(n) = {A GL(n, R) : AAT = I}.
We claim that it is an n(n1)

2 -dimensional Lie group. If A = (aij ), then the
equation AAT = I is equivalent to the system of equations
n
X
Fij (A) = aik ajk ij = 0, (i, j = 1, . . . , n).
k=1
Since Fij = Fji , these equations are not independent. We get however an
independent system of equations if we restrict ourselves to the equations with
1 i j n. The number of these equations is n(n+1) 2 , so if we show the
independence of the gradients, then we obtain the required expression for the
dimension of O(n).
The gradient of Fij at A O(n) is an n n matrix with entries
n
Fij X aik ajk
= ajk + aik = ajs ir + ais jr .
ars ars ars
k=1
We show that these vectors, are orthogonal with respect to the usual scalar
product on Rn
2
= Mat(n, R). Indeed, taking the scalar product of the gradi-
ent vectors of Fij and Fkl at A we obtain
n
X
(ajs ir + ais jr )(als kr + aks lr )
r,s=1
n
X
= (ajs als ir kr + ajs aks ir lr + ais als jr kr + ais aks jr lr )
r,s=1
X n
= (jl ir kr + jk ir lr + il jr kr + ik jr lr )
r=1
= jl ik + jk il + il jk + ik jl = 2ik jl (1 + ij kl ).
(In the last step, we used that as i j and k l, il jk 6= 0 only if i = j =

k = l.)
We know that the determinant of an orthogonal matrix is 1. Thus O(n) has
two components. The connected component on which determinant is 1, is the
special orthogonal group SO(n), and it has also dimension n(n1) 2 . SO(n) is
the configuration space of a rigid n-dimensional body with one fixed point.
The configuration space of an n-dimensional body without a fix point is the
space SO(n) Rn .
For n = 2, SO(2) is the 1-dimensional group of rotations of the plane. For
n = 3, SO(3) is a 3-dimensional manifold. With the help of quaternions we
can show that this group is homeomorphic to the 3-dimensional projective
space.
Let H denote the 4-dimensional space of quaternions x + yi + zj + wk, and let
us identify R3 with the space of pure imaginary quaternions. For 0 6= q H
let us denote by q the transformation q (h) = q 1 hq. Since |q 1 hq| = |h|,
q is an orthogonal transformation. If h is a real number, then q (h) = h,
thus q maps R H into itself. Consequently, it maps R3 , the orthogonal
complement of R also into itself. The assignment q 7 q R3 is a group
homomorphism from the multiplicative group H\{0} to SO(3). The kernel of
this homomorphism is the center of H\{0}, i.e. R\{0}. This homomorphism
is also surjective as it follows from the following two exercises.
Exercise 4.1.15. Show that every element of SO(3) is a rotation about a
straight line.

Exercise 4.1.16. Show that if a R3 is a pure imaginary quaternion of
length 1, R and q = cos + a sin , then q R3 is a rotation about a with
angle 2.
We conclude that SO(3) is isomorphic to the factor group H \ {0}/R \ {0},
but cosets of R \ {0} in H \ {0} are in one to one correspondence with straight
lines through 0 in H, i.e. there is a natural bijection between SO(3) and the
projective space Gr(4, 1).
4.1.4 Remarks on the Classification of Manifolds

The classification problem of n-dimensional manifolds can be formulated in
different levels. For each r we may consider the category of C r -manifolds and
r-times differentiable mappings. The following two theorems show that the
classification problem of C r -manifolds is the same problem for all 1 r .
Theorem 4.1.17. For r 1, every maximal C r -compatible atlas contains a
C -compatible atlas.
210 4. Manifolds
Theorem 4.1.18. If two C -manifolds are C 1 -diffeomorphic, then they are

C -diffeomorphic.
None of the above theorems can be extended to C 0 -manifolds. There exist
topological manifolds which have no C 1 -compatible atlas, and there exist
homeomorphic but not diffeomorphic differentiable manifolds.
In 1956 J.W. Milnor constructed a differentiable manifold which is homeo-
morphic to the standard 7-dimensional sphere but not diffeomorphic to it.
Such manifolds were given the name exotic spheres. Later on an even more
surprising result was published by Milnor and Kervaire. There are exactly
28 mutually non-diffeomorphic differentiable structures on a topological 7-
sphere. Since then many examples of topological manifolds having many
different differentiable structures have been obtained. One of the most inter-
esting constructions is due to Donaldson, who invented exotic differentiable
structures on R4 .
We do not meet exotic differentiable structures in dimension two. The classi-
fication of compact surfaces is the same up to homeomorphism and diffeomor-
phism. The classification theorem of compact 2-dimensional manifolds gives
a list of non-diffeomorphic compact surfaces and asserts that every compact
surface is diffeomorphic to one of the surfaces in the list. The list of compact
surfaces consists of two lists in fact. The first list contains the orientable
compact surfaces, the second contains the non-orientable ones.
Orientable compact surfaces. The simplest orientable closed surface is the
sphere S 2 = {x R3 : |x| = 1}. The next example is the torus T 2 = S 1 S 1 .
Cutting a small disc out of the torus, we get a surface with boundary, called
a handle. A typical orientable compact surface is a sphere with g handles.
We can obtain this surface if we cut g holes on the surface of the sphere and
glue a handle to each of them.
Non-orientable compact surfaces. The first member of this list is the real
projective plane. To understand the topological structure of the projective
plane, let us cut the projective plane into two parts by a hyperbola. The
interior of the hyperbola has two components in the Euclidean plane, but
these components are glued together along a segment of the line at infinity,
so the interior is a topological disc. The exterior of the hyperbola is a long
infinite band in the Euclidean plane, the ends of which are glued together
along another segment of the line at infinity. One can see that the ends of the
band are glued together by a half twist so what we get is a Mobius band. We
conclude that the projective plane is the union of a disc and a Mobius band
glued together along their boundaries. A typical non-orientable compact
surface is a sphere with g Mobius bands. We obtain this surface cutting g
discs out of the sphere and gluing to the boundary of each hole a Mobius band.
Non-orientable compact surfaces can not be embedded into the 3-dimensional
Euclidean space, so although one can easily construct a 3-dimensional model
Figure 4.2: Orientable closed surfaces.
of a Mobius band and that of a sphere with g holes, it is impossible to glue

the Mobius bands to the sphere in practice. If however we could try this in
a 4-dimensional space there would be no difficulty.
Exercise 4.1.19. The configuration space of the pentagon (closed chain of

five rods in the plane) with one edge fixed is a compact surface (sometimes
with singularities). What kind of surfaces can we obtain?
Exercise 4.1.20. Give an example of a set X with a C 0 -compatible atlas A

on it such that the topology induced on X by A is
(a) second countable but not Hausdorff;
(b) Hausdorff but not second countable.
Exercise 4.1.21. Show that the special unitary group
SU (n) = {A GL(n, C) : AA = I, det A = 1}
is a Lie group and determine its dimension. Prove that SU (2) is diffeomorphic
to the 3-dimensional sphere S 3 .
Exercise 4.1.22. Which surface shall we get from the classification list if
we glue to the sphere k 1 Mobius bands and l handles?
212 4. Manifolds
Figure 4.3: Self-intersecting images of the Mobius band with planar bound-
ary: the cross cap and half of the Klein bottle.
Figure 4.4: A sphere with one and two Mobius bands: the projective plane
and the Klein bottle.
Exercise 4.1.23. Let P be a complex polynomial of degree k having k

different roots. Consider the subset of C2 defined by
M = {(z, w) C2 : z l = P (w)}.
Show that for some natural numbers g and N , M is diffeomorphic to a sphere

with g handles from which N points are omitted. Express g and N in terms
of k and l.
4.2 The Tangent Bundle

The aim of this section is to give and reconcile different commonly used
definitions of a tangent vector to a manifold. Before passing over to the
abstract situation, we shall deal with submanifolds of Rn .
Definition 4.2.1. Smooth mappings : [a, b] M of an interval into a

differentiable manifold (M, A) are called smooth parameterized curves in the
manifold.
4.2. The Tangent Bundle 213
Definition A. Let M be a differentiable manifold embedded in Rn , x0 M .

A vector v is called a tangent vector to M at x0 if there is a smooth curve
: [, ] M passing through x0 = (0) with speed vector v = 0 (0).
The (linear) tangent space Tx0 M of M at x0 M is the set of all tangent
vectors to M at x0 .
The affine tangent space of M at x0 is the translate of Tx0 M with the vector
x0 .
Theorem 4.2.2. Let us suppose that a k-dimensional manifold M embed-

ded in Rn is given in a neighborhood U of x0 M by a system of equal-
ities f1 = = fnk = 0, where f1 , . . . , fnk are smooth functions on
U such that the vectors grad f1 (x0 ), . . . , grad fnk (x0 ) are linearly indepen-
dent. Then the tangent space of M at x0 consists of the vectors orthogonal
to grad f1 (x0 ), . . . , grad fnk (x0 ).
Corollary 4.2.3. The tangent space of a k-dimensional submanifold M of
Rn is a k-dimensional linear subspace of Rn . The affine tangent space of M
at a point is an affine subspace of Rn .
Proof. If x : [, ] M is a smooth curve on M passing through x0 = (0),
and having coordinate functions = (x1 , . . . , xn ), then we have
fi (x1 (t), . . . , xn (t)) = 0 i = 1, . . . , n k
for each t [, ]. Differentiating with respect to t we get
1 fi ((0))x01 (0) + + n fi ((0))x0n (0) = 0,
which means that the vectors grad fi (x0 ) and 0 (0) are orthogonal.
Now we prove that if a vector v is orthogonal to the vectors grad fi (x0 ),
1 i n k, then v is a tangent vector. Let us take a smooth local
parameterization F : Rk M Rn of M around the point x0 . The curve
: t 7 F (F 1 (x0 ) + ty), where y Rk is fixed, is a curve on M passing
through x0 = (0). The speed vector of this curve at t = 0 is
1 F (F 1 (x0 ))y1 + + k F (F 1 (x0 ))yk ,
where y1 , . . . , yk are the coordinates of y. By the construction of local pa-

rameterizations of embedded manifolds, presented in the proof of Theorem
4.1.12, F is a restriction onto Rk of a diffeomorphism between open sub-
sets of Rn , consequently, the vectors i F (F 1 (x0 )) are linearly independent.
We conclude, that the tangent space is contained in the k-dimensional linear
subspace orthogonal to grad f1 (x0 ), . . . , grad fnk (x0 ) and contains the k-
dimensional linear subspace spanned by the vectors i F (F 1 (x0 )) 1 i k,
which means that both linear subspaces coincide with the tangent space.
214 4. Manifolds
The definition of tangent vectors can also be given in intrinsic terms, inde-
pendent of the embedding of M into Rn . Let us define an equivalence relation
on the set
(M, p) = { : [, ] M | (0) = p, is smooth},
consisting of curves passing through p M , by calling two curves 1 , 2
(M, p) equivalent if ( 1 )0 (0) = ( 2 )0 (0) for some chart around p.
Then this condition is true for any chart from the atlas of M . (Prove this!)
Definition B. A tangent vector to a manifold M at the point p M is an
equivalence class of curves in (M, p). The set of equivalence classes is called
the tangent space of M at p and is denoted by Tp M .
In this setting, the speed vector 0 (s) T(s) M of a smooth curve is the
equivalence class of the curve t 7 (s + t) in (M, s).
Given a chart around p, we can establish a one-to-one correspondence
between the equivalence classes and points of Rm , (m = dim M ), assigning
to the equivalence class of a curve (M, p) the vector ( )0 (0) Rm .
With the help of this identification, we can introduce a vector space structure
on the tangent space, not depending on the choice of the chart.
For embedded manifolds, Definition B agrees with Definition A. The advan-
tage of Definition B is that it is applicable also for abstract manifolds, not
embedded anywhere.
Definition 4.2.4. If = (x1 , . . . , xm ) is a chart on the manifold M around
the point p, (M, p), then the coordinates (x1 )0 (0), . . . , (xm )0 (0)
of ( )0 (0) are called the components of the speed vector 0 (0) with respect
to the chart .
The main difficulty of defining tangent vectors to a manifold is due to the
fact that an abstract manifold might not be naturally embedded into a fixed
finite dimensional linear space. Nevertheless, there is a universal embedding
of each differentiable manifold into an infinite dimensional linear space.
Let us denote by F(M ) the linear vector space of smooth functions on M ,
and by F (M ) the dual space of F(M ) that is the space of linear functions
on F(M ), and consider the embedding of M into F (M ) defined by the
formula
[(p)](f ) = f (p), where p M, f F(M ).
Having embedded the manifold M into F (M ), we can define tangent vectors
to M as elements of the linear space F (M ).
Definition 4.2.5. Let M be a differentiable manifold, p M . We say that
a linear function D F (M ) defined on F(M ) is a derivation of F(M ) at p
if the equality
D(f g) = D(f )g(p) + f (p)D(g)
holds for every f, g F(M ).
Derivations at a point p M form a linear subspace Derp M of F (M ).

Each curve (M, p) defines a derivation at the point p by the formula
D 0 (0) (f ) = (f )0 (0), where f F(M ). D 0 (0) is the speed vector of the
curve in F (M ). Since two curves define the same derivation if and
only if they are equivalent, there is a one-to-one correspondence between the
equivalence classes of curves and the derivations obtained as D 0 (0) for some
.
Definition C. A tangent vector to a manifold M at the point p M is a
derivation of the form D 0 (0) , where (M, p).
The tangent space Tp M of M at the point p is the set of derivations D 0 (0)
along curves in M passing through p = (0).
We are going to show that although Definition C defines Tp M just as a
subset of Derp M , we have Tp M = Derp M in fact. To prove this we need
some preliminary steps.
Lemma 4.2.6. If f F(M ) is a constant function and D is a derivation at

a point p M , then D(f ) = 0.
Proof. Because of linearity, it is enough to show that D(1) = 0, where 1 is

the constant 1 function on M . This comes from
D(1) = D(1 1) = D(1)1(p) + 1(p)D(1) = 2D(1).
Lemma 4.2.7. If two functions f, g M coincide on a neighborhood U of

p M and D is a derivation at p then D(f ) = D(g).
Proof. Using the bump function construction of Proposition 1.5.13, we can

define a smooth function h on M which is zero outside U and such that
h(p) = 1. In this case h(f g) is the constant 0 function on M . Thus we
have
0 = D(0) = D(h(f g)) = D(h)(f (p) g(p)) + h(p)D(f g) = D(f ) D(g).
Remarks.
(i) The bump function construction shows that the mapping : M

F (M ) is an inclusion. Indeed, if p 6= q are distinct points of M ,
then there is a smooth function h on M such that
[(p)](h) = h(p) = 1 6= [(q)](h) = h(q) = 0.

216 4. Manifolds
(ii) We can extend a derivation D Derp to functions f defined only in

a neighborhood U of p by taking a smooth function h on M such that
h is zero outside U and constant 1 in a neighborhood of p and putting
D(f ) := D(f), where

f (x)h(x) if x U ,
f(x) =
0 if x
/ U.
By Lemma 4.2.7, this extension of D is correctly defined.
Lemma 4.2.8. Let f : B R be a smooth function defined on an open ball
B Rn around the origin. Then there exist smooth functions gi , 1 i n
on B such that
n
X
f (x) = f (0) + xi gi (x) for x = (x1 , . . . , xn ) B
i=1
and
gi (0) = i f (0).
Proof. Since
Z 1 Z 1X n
d f (tx)
f (x) f (0) = dt = xi i f (tx)dt
0 dt 0 i=1
n
X Z 1
= xi i f (tx)dt,
i=1 0
R1
we may take gi (x) = 0
i f (tx)dt.
Now we are ready to prove the main theorem.
Theorem 4.2.9. The tangent space to a differentiable manifold M at the
point p M coincides with the space of all derivations of F(M ) at p, which
is a linear space having the same dimension as M has.
Proof. Let us take a differentiable manifold (M, A) and a chart = (x1 , . . . ,
xn ) A defined in a neighborhood of p M . Define the derivations i (p)
as follows
[i (p)](f ) := i (f 1 )((p)).
We prove that the derivations i (p) form a basis in the space of derivations
at p. They are linearly independent since if we have
n
X
i i (p) = 0,
i=1
then applying this derivation to the j-th coordinate function xj we get

n
X
i i (xj 1 )((p)) = j = 0.
i=1
On the other hand, if D is an arbitrary derivation at p, then we have D =

P n i
i=1 D(x )i (p). Indeed, let f F(M ) be an arbitrary smooth function on
M and apply Lemma 4.2.8 to the function x 7 f 1 (x + (p)) around 0.
We obtain the existence of functions gi defined around (p) such that
n
X
f = f (p) + (xi xi (p))gi and gi ((p)) = i (f 1 )((p)).
i=1
In this case, however, we have

n
X
D(f ) = D(f (p)) + D((xi xi (p)))gi ((p)) + (xi (p) xi (p))D(gi )
i=1
n
X n
X
= i
D(x )i (f 1
)((p)) = D(xi )[i (p)](f ).
i=1 i=1
To finish the proof, we only have to show that every derivation at the point
p can be obtained as a speed vector of a curve passing through p. Define
the curve : [, ] M by the formula (t) := 1 ((p) + (t1 , . . . , tn )).
Pn
Then obviously the speed vector 0 (0) is just i=1 i i (p).
Definition 4.2.10. The basis (1 (p), . . . , n (p)) of Tp M defined in the proof

of Theorem 4.2.9 is called the basis induced by the chart .
4.2.1 The Tangent Bundle

S
The disjoint union of the tangent spaces of M at the various points, pM Tp M ,
has a natural differentiable manifold structure, the dimension of which is
twice the dimension of M . This manifold is called the tangent bundle of M
and is denoted by T M . A point of this manifold is a vector D, tangent to
M at some point p. Local coordinates on T M are constructed as follows.
Let = (x1 , . . . , xn ) be a chart on M the domain U of which contains p,
and D(x1 ), . . . , D(xn ) be the components of D in the basis i (p). Then the
mapping
D 7 (x1 (p), . . . , xn (p), D(x1 ), . . . , D(xn ))
S
give a local coordinate system on pU Tp M T M . The set of all local
coordinate systems constructed this way forms a C -compatible atlas on
T M , that turns T M into a differentiable manifold.
218 4. Manifolds
Exercise 4.2.11. Check the last statement.
The mapping : T M M which takes a tangent vector D Tp M to the

point p M at which the vector is tangent to M is called the natural pro-
jection of the tangent bundle. The inverse image of a point p M under the
natural projection is the tangent space Tp M . This space is called the fiber of
the tangent bundle over the point p.
4.2.2 The Derivative of a Smooth Map

Definition 4.2.12. Let f : M N be a smooth mapping between the dif-
ferentiable manifolds (M, A) , (N, B), and let p M . The derivative of f
at the point p is the linear map Tp f : Tp M Tf (p) N given in the following
way. Let D Tp M and consider a curve : [, ] M with (0) = p and
speed vector 0 (0) = D. Then Tp f (D) is the speed vector (f )0 (0).
The tangent vector Tp f (D) as a derivation at f (p) assigns to a smooth func-

tion h F(N ) the number
(h (f ))0 (0) = ((h f ) )0 (0) = D(h f ).
Proposition 4.2.13. The derivative Tp f is correctly defined (does not de-

pend on the choice of ) and is linear.
Proof. We derive a formula for Tp f using local coordinates which will show
both parts of the proposition.
Let = (x1 , . . . , xm ) and = (y 1 , . . . , y n ) be local coordinates in a neigh-
borhood of p M and f (p) N respectively.
If the components of D in the basis i (p) corresponding to the chart are
{i : 1 i m} then we have (xi )0 (0) = i . Observe, that i depends
only on D and the fixed chart , but not on the curve . The components
{ j : 1 j n} of Tp f (D) in the basis j (f (p)) induced by the chart
can be computed by the formula j = (y j f )0 (0). Denote by fj the j-th
coordinate function of the mapping f = f 1 , i.e.
fj = y j f 1 .
Then we have
j = (y j f )0 (0) = [(y j f 1 ) ]0 (0) = [fj ( )]0 (0)

Xm m
X
= i fj ((p))(xi )0 (0) = i fj ((p))i ,
i=1 i=1
4.3. The Lie Algebra of Vector Fields 219
which shows that Tp f (D) depends only on D but not on and that Tp f
is a linear mapping the matrix of which in the bases (1 (p), . . . , m

(p)) and

(1 (f (p)), . . . , n (f (p))) is

i fj ((p)) 1im .
1jn
We remark that although the matrix representation of Tp f we have just

computed depends on the choice of the charts and , the linear map Tp f
itself does not depend these choices. This is clear from the definition of Tp F ,
which does not use local coordinates.
Proposition 4.2.14 (Chain Rule). If g : M N and f : N P are smooth

maps, p M , then Tp (f g) = Tg(p) f Tp g.
Proof. If a tangent vector of M at p is represented by the curve , then
Tp (f g)([]) = (f g )0 (0) = Tg(p) f ([g ]) = Tg(p) f (Tp g([])).
4.3 The Lie Algebra of Vector Fields

Definition 4.3.1. A smooth vector field X over a differentiable manifold M
is a smooth mapping of M into its tangent bundle, such that X(p) Tp M
for each p M .
Obviously, smooth vector fields over M form a real vector space with respect
to the operations
(X + Y )(p) := X(p) + Y (p), (X)(p) := X(p),
where X, Y are vector fields, R, p M . We can multiply vector fields

by smooth functions as well, by the rule
(f X)(p) := f (p)X(p).
We denote by X(M ) the vector space of smooth vector fields.

Associated to a local coordinate system = (x1 , . . . , xn ) on M , there is a
basis of Tp M at each p dom , formed by the tangent vectors {i (p) : 1
i n}. The mapping i : p 7 i (p) gives a local smooth vector field in the
domain of the chart for each i. Thus, every smooth vector field X can be
written in the form
Xn
X= X i i ,
i=1
220 4. Manifolds
where the X i -s are smooth functions on the domain of . The functions X i

are called the components of the vector field X with respect to the chart .
Given a smooth vector field X on a manifold M , we may pose the following
problem. Find those smooth curves : (a, b) M in M for which the speed
of at t (a, b) is X((t)). Such curves are called the integral curves of the
vector field. Obviously, a restriction of an integral curve onto a subinterval is
also an integral curve, therefore, it is enough to look for the maximal integral
curves which can not be extended to an integral curve defined on a larger
interval.
The solution of this problem reduces to an ordinary differential equation
of first order. Indeed, is an integral curve if and only if for each chart
= (x1 , . . . , xn ), the vector-valued function
f = : (a, b) Rn
satisfies the differential equation
f 0 (t) = (X 1 1 f (t), . . . , X n 1 f (t)).
Actually, finding integral curves of a vector field is the same problem as

solving an ordinary differential equation, only the language of formulation is
different. Translating the basic results of the theory of ordinary differential
equations into the language of geometry we get the following theorems, we
mention without proof.
Theorem 4.3.2.
(i) (Existence and uniqueness of solutions). Let X be a smooth vector

field on a differentiable manifold M . Then for each point p M , there
exists a unique maximal integral curve p : (a, b) M of the vector
field X such that 0 (a, b) and p (0) = p (a and b are allowed to be
and respectively).
(ii) (Unboundedness of solutions in time or space). If a (or b) is finite

then no compact subset of M contains the image p ((a, 0)) (or p ((0, b))
).
(iii) (Differentiable dependence on the initial point). Let us define the set
Ut M for t R as follows
Ut = {p M : t dom p }.
Then Ut is an open subset of M and the mapping Ht : Ut M defined

by Ht (p) = p (t) is a diffeomorphism between Ut and Ut . If, fur-
thermore, the expression Ht1 (Ht2 (p)) is defined, then so is Ht1 +t2 (p)
and Ht1 (Ht2 (p)) = Ht1 +t2 (p). The family {Ht : t R} is called the
one-parameter family of diffeomorphisms or the flow generated by the
vector field X.
A useful consequence of the fundamental theorems is the following statement
on the straightening of vector fields.
Proposition 4.3.3. Let X be an arbitrary vector field on a manifold M
and p M be a point such that X(p) 6= 0. Then there exists a chart =
(x1 , . . . , xn ) : U Rn around p for which X|U = 1 .
This means that the derivative of the mapping turns the vector field X into
a constant vector field on Rn .
Proof. Let Ht be the flow generated by the vector field X. Choose a chart
: V Rn around p such that (p) = 0. Then the image of contains a
ball Br of radius r about the origin. Decompose X as a linear combination
X = X 1 1 + + X n n . Since X(p) 6= 0, one of the coefficients X i (p)
should also be different from 0. We may assume that X 1 (p) 6= 0 as this can
be reached by permuting the coordinates.
Let Brn1 Br be the (n 1)-dimensional ball consisting of those point of
Br the first coordinate of which is 0, and let j : Brn1 M be the inverse
restricted onto Brn1 . The set W of those points q M for which the
maximal integral curve starting at q is defined on a fixed open interval (, )
is open in M and these open sets for all positive > 0 cover M . Since j(Brn1 )
is compact, it is covered by a finite number of such sets Wi . Thus, taking
for > 0 the smallest of the i s, we obtain a positive number such that all
maximal integral curves of X starting at a point of j(Brn1 ) is defined on the
interval (, ).
Consider the smooth map F : (, ) Brn1 M which is defined by the
formula
F (x1 , . . . , xn ) = Hx1 (j(0, x2 , . . . , xn )).
The partial derivatives of F at the origin are 1 F (0) = X(p), and i F (0) =
i (p) if i 2. These vectors are linearly independent, therefore, by the
inverse function theorem, there is an open neighborhood U of p and W
(, )Brn1 of the origin in Rn such that F |W is a diffeomorphism between
W and U . In other words, F |W is a local parameterization of M and its
inverse is a local coordinate system on M . We claim that has the required
property. Indeed, if q U is an arbitrary point, and (q) = (q1 , . . . , qn ), then
the parameterized coordinate line (, ) M , t 7 F (t, q2 , . . . , qn ) is an
integral curve of X, therefore its speed vector at t = q1 is 1 (q) = X(q).
Since tangent vectors to a manifold at a point are identified with deriva-
tions at the point, vector fields can be regarded as differential operators
222 4. Manifolds
called derivations of smooth functions assigning to a smooth function an-

other smooth function by the formula
[X(f )](p) = [X(p)](f ), where X X(M ), f F(M ), p M.
In this sense, a vector field X is a linear mapping X : F(M ) F(M ),

satisfying the Leibniz rule
X(f g) = X(f )g + f X(g).
Definition 4.3.4. Let A and B be two linear endomorphisms of a vector

space V . Then the linear mapping [A, B] = A B B A is called the
commutator of them.
Proposition 4.3.5. The commutator of linear mappings satisfies the follow-

ing identities (A, B, C End(V ), R)
(i) [A + B, C] = [A, C] + [B, C], [C, A + B] = [C, A] + [C, B]

[A, B] = [A, B] = [A, B] (bilinearity),
(ii) [A, B] = [B, A] (skew-commutativity),
(iii) [A, [B, C]] + [B, [C, A]] + [C, [A, B]] = 0 (Jacobi identity).
Proof. We prove only (iii), the rest is left to the reader.
[A,[B, C]] + [B, [C, A]] + [C, [A, B]]

= [A, (B C C B)] + [B, (C A A C)] + [C, (A B B A)]
= A (B C C B) (B C C B) A + B (C A A C)
(C A A C) B + C (A B B A) (A B B A) C
= A B C A C B B C A + C B A+
+ B C A B A C C A B + A C B+
+C ABC BAABC +BAC
= 0.
Definition 4.3.6. Let us suppose that a linear space L is endowed with a

bilinear operation [ , ] : L L L satisfying conditions (i), (ii), and (iii) of
the above proposition. Then the pair (L, [, ]) is called a Lie algebra.
Proposition 4.3.7. Let X and Y be two smooth vector fields on a manifold

M . Considering them as linear endomorphisms of the vector space of smooth
functions F(M ), the commutator [X, Y ] of them is also a vector field.
Proof. The commutator [X, Y ] is a linear endomorphism of F(M ) so we have

only to check that it satisfies the Leibniz rule. For f, g F(M ) we have
[X, Y ](f g) = (X Y Y X)(f g) = X(Y (f g)) Y (X(f g))

= X(Y (f )g + f Y (g)) + Y (X(f )g + f X(g))
= X Y (f )g + Y (f )X(g) + X(f )Y (g) + f X Y (g) Y X(f )g
X(f )Y (g) Y (f )X(g) f Y X(g)
= [X, Y ](f )g + f [X, Y ](g).
As a corollary, we obtain that he commutator of vector fields, which is gen-

erally called the Lie bracket of them, is a binary operation on X(M ), giving
the space of vector fields a Lie algebra structure.
Proposition 4.3.8. Let us choose a local coordinate system = (x1 , . . . , xn )

on M and denote by 1 , . . . , n the associated coordinate vector fields. Then
we have
(i) [i , j ] = 0;
(ii) [f X, gY ] = f g[X, Y ] + f X(g)Y gY (f )X for each X, Y X(M ),

f, g F(M );
Pn Pn
(iii) if X = i=1 X i i , Y = i=1 Y i i are arbitrary vector fields, then
n
X X n
n X
[X, Y ] = i
(X(Y ) Y (X i
))i = X j j Y i Y j j X i i .
i=1 i=1 j=1
Proof. (i) The first part of the proposition is equivalent to Youngs theorem
since for any smooth function f on the domain of , we have
i (j (f )) = i (j (f 1 ) ) = i (j (f 1 ))
= j (i (f 1 )) = j (i (f )).
(ii) Let h be an arbitrary smooth function on M , and apply the operator

[f X, gY ] to it.
[f X, gY ](h) = f X(gY (h)) gY (f X(h))

= f X(g)Y (h) + f gX(Y (h)) gY (f )X(h) gf Y (X(h))
= (f g[X, Y ] + f X(g)Y gY (f )X)(h).
224 4. Manifolds
(iii) Using (i) and (ii) we get

n
hX n
X i Xn X
n
[X, Y ] = X i i , Y j j = [X i i , Y j j ] =
i=1 j=1 i=1 j=1
n X
X n n
X
= X i i (Y j )j Y j j (X i )i = (X(Y i ) Y (X i ))i .
i=1 j=1 i=1
Suppose that we are given two vector fields X and Y on an open subset
of Rn . The corresponding flows Hs and Gt do not commute in general:
Hs Gt 6= Gt Hs .
To measure the lack of commutation of the flows Hs and Gt , we consider the
difference
(s, t; p) = Gt Hs (p) Hs Gt (p)
for a fixed point p Rn . is a differentiable function of s and t and it is 0 if
t or s is zero. This means, that in the Taylor expansion of around (0, 0; p)

(s, t; p) = (0, 0; p) + s (0, 0; p) + t (0, 0; p) +
s t
2 2
2 t2 2

s
+ (0, 0; p) + st (0, 0; p) + (0, 0; p) + o(s2 + t2 )
2 s2 st 2 t2
2
the only non-zero partial derivative is (0, 0; p).
st
Proposition 4.3.9. Through the natural identification of the tangent space
2
of Rn at p with the vectors of Rn , the vector (0, 0; p) corresponds to the
st
tangent vector [X, Y ](p).
Pn Pn
Proof. Put X = i=1 X i i , Y = i=1 Y i i . Let us compute first the vector
2
st Hs Gt (p) at s = t = 0. Since Hs is the flow of X, for any point q, the
curve q (s) = Hs (q) is an integral curve of X. Thus, we have

Hs Gt (p) (0, t) = G (p) (s) (0, t) = X(Gt (p)).
s s t
Differentiating with respect to t,
n
2 d X
Y (X i )i (p)

Hs Gt (p)s=t=0 = X(Gt (p))t=0 =
st dt i=1
A similar computation shows that

n
2 X
X(Y i )i (p)

Gt Hs (p)s=t=0 =
st i=1
Subtracting these equalities we get

n
2 X
(0, 0, ; p) = (X(Y i ) Y (X i ))i (p) = [X, Y ](p).
st i=1
Now returning to the Taylor expansion of , we see that
2
(s, t; p) = st (0, 0, ; p) + o(s2 + t2 ) = st[X, Y ](p) + o(s2 + t2 )
st
In particular, we obtain the following expression for [X, Y ](p).
Gt Ht (p) Ht Gt (p)
[X, Y ](p) = lim
t0 t2
Definition 4.3.10. We say that two vector fields are commuting if their Lie
bracket is the zero vector field.
Theorem 4.3.11. Let {Ht : t R} and {Gt : t R} be the one-parameter

families of diffeomorphisms generated by the vector fields X and Y respec-
tively and suppose that the vector fields X and Y are commuting. Then the
diffeomorphisms Hs and Gt are commuting as well in the following sense.
For each point p of the manifold there exists a positive (depending on p) such
that for any pair of real numbers s, t satisfying the inequality |s| + |t| the
expressions Hs (Gt (p)) and Gt (Hs (p)) are defined and coincide: Hs (Gt (p)) =
Gt (Hs (p)).
Proof. If both X and Y vanish at p then Hs (p) = Gt (p) = p for any s and t
and thus the assertion holds trivially. We may thus suppose that one of the
vectors X(p), Y (p), say X(p) is not zero. By the theorem on the straightening
of vector fields we may suppose that the manifold is an open subset of Rn ,
with coordinates (x1 , . . . ,P
xn ), and the vector field X coincides with the basis
n
vector field 1 . Let Y = i=1 Y i i be the decomposition of Y into a linear
combination of the basis vector fields i . By the formula for the Lie bracket
of vector fields we have
n
X
0 = [X, Y ] = (1 Y i )i 1 Y i = 0 for each i.
i=1
Consequently, the functions Y i do not depend on x1 , thus the vector field Y

is invariant under translations parallel to the vector e1 = (1, 0, . . . , 0). This
implies that if is an integral curve of the vector field Y , then so is + se1
for any s (the domain of + se1 is an open subset of the domain of ). On
226 4. Manifolds
the other hand, the diffeomorphism Hs is just a translation by the vector

se1 . Hence, for small s and t, we have
Gt (Hs (p)) = Gt (p + se1 ) = (p + se1 )(t)

= p (t) + se1 = Gt (p) + se1 = Hs (Gt (p)).
If f : M N is a smooth map, then the derivative of f can be used to

send forward tangent vectors of M to tangent vectors of N . However this,
in general, cannot be used to send forward vector fields. Taking a smooth
vector field on M , and sending forward its vectors by the derivative map of f
we obtain a bunch of tangent vectors of N which may not be tangent vectors
of a smooth vector field of N . One obstruction of this can be that if f is
not injective, then we can get different tangent vectors at some points of N
coming from different preimages. If we get at most one vector at each point
of N then it is still possible that this partially defined vector field does not
extend to a global smooth vector field. If there is a smooth extension but f
is not surjective, then the smooth extension is definitely not unique. These
problems are eluded by the notion of f -related vector fields which makes the
assumption that the image vectors extend to a smooth vector field a part of
the definition.
Definition 4.3.12. Let f : M N be a smooth map between two smooth
manifolds. We say that the vector fields X X(M ) and Y X(N ) are
f -related if Tp f (Xp ) = Yf (p) for all p M .
Thinking of vector fields as derivations of smooth functions, X and Y are
f -related if and only if for any smooth function h C on N , the equality
Y (h)(f (p)) = Yf (p) h = Xp (h f ) = X(h f )(p) holds for all p M that is,
if Y (h) f = X(h f ). This equivalent version of the definition is convenient
to prove the following proposition.
Proposition 4.3.13. If f : M N is a smooth map, X1 , X2 X(M ) and
Y1 , Y2 X(N ) are smooth vector fields such that X1 is f -related to Y1 and
X2 is f -related to Y2 , then [X1 , X2 ] is f -related to [Y1 , Y2 ].
Proof. Let h C be an arbitrary smooth function on N , then
[Y1 , Y2 ](h) f = Y1 (Y2 (h)) f Y2 (Y1 (h)) f

= X1 (Y2 (h) f ) X2 (Y1 (h) f )
= X1 (X2 (h f )) X2 (X1 (h f )) = [X1 , X2 ](h f ).
Corollary 4.3.14. When f : M N is a diffeomorphism, then for any
smooth vector field X on M , there is a unique smooth vector field f (X)
on N which is f -related to X. The push forward f of vector fields is a

Lie-algebra isomorphism between X(M ) and X(N ).
4.3.1 The Lie Algebra of Lie Groups

Let F : M N be a diffeomorphism between two manifolds. Then F defines
a bijection between X(M ) and X(N ), (which we shall denote also by T F ) and
this bijection is a Lie algebra isomorphism.
Let G be a Lie group, g G. Let Lg denote the left translation by g, i.e.,
Lg : G G, Lg (h) = gh. Lg is a diffeomorphism, its inverse is Lg1 . A vector
field X X(G) is called left invariant if T Lg (X) = X for all g G. Since
T Lg (X) = X and T Lg (Y ) = Y imply T Lg [X, Y ] = [T Lg (X), T Lg (Y )] =
[X, Y ], left invariant vector fields form a Lie subalgebra of X(G).
Definition 4.3.15. The Lie algebra of left invariant vector fields of a Lie
group is called the Lie algebra of the Lie group.
If X X(G) is left invariant, then X(g) = Te Lg (X(e)), thus, a left invari-
ant vector field is uniquely determined by the vector X(e) Te G (e is the
unit element of the group G). Since every vector in Te G extends to a left
invariant vector field this way, the assignment X 7 X(e) yields a linear iso-
morphism between the vector space of left invariant vector fields on G and
Te G. As a consequence, we obtain that the Lie algebra of a Lie group is finite
dimensional and its dimension is the same as that of the Lie group.
As an example, let us determine the Lie algebra of GL(n, R). GL(n, R) is
an open subset in Mat(n, R), so its manifold structure is given by one chart,
the embedding. Tangent spaces at different points can be identified with the
linear space Mat(n, R). For A GL(n, R), the left translation M 7 AM
extends to a linear transformation of the whole linear space Mat(n, R). The
derivative of a linear transformation of a linear space is the linear transfor-
mation itself, if we identify the tangent spaces at different points with the
linear space, so a left invariant vector field X : GL(n, R) Mat(n, R) has
the form X(A) = AM , where M Mat(n, R) is a fixed matrix.
The integral curves of a left invariant vector field on GL(n, R) can be de-
scribed with the help of the exponential function for matrices. If M is an
arbitrary square matrix, then we define eM as the sum of the infinite series

X Mk
.
k!
k=0
If we define the curve A : R GL(n, R) by A (t) = AeM t , then we obtain

an integral curve of the vector field X(A) = AM . Indeed,
0
A (t) = AeM t M = X(AeM t ) = X(A (t)).
228 4. Manifolds
The flow generated by the left invariant vector field X consists of the diffeo-
morphisms
Ht (A) = AeM t ,
that is, Ht is a right translation by eM t .
Now let us take two left invariant vector fields X(A) = AM and Y (A) = AN
and consider the flows Ht and Gt generated by them.
Computing Gt Ht (A) Ht Gt (A) up to o(t2 ), we get
Gt Ht (A) Ht Gt (A) = A(eM t eN t eN t eM t )

1 1
= A I + M t + (M t)2 I + N t + (N t)2
2 2
1 1
A I + N t + (N t)2 I + M t + (M t)2 + o(t2 ) =
2 2
= A(M N N M )t2 + o(t2 ).
We obtain, that the Lie algebra of GL(n, R) is isomorphic to the Lie algebra
of all matrices with Lie bracket [M, N ] = M N N M .
Exercise 4.3.16. Let 1 and 2 be the two coordinate vector fields on R2

determined by the identity mapping. Describe the vector fields
X(x1 , x2 ) = x1 1 + x2 2
Y (x1 , x2 ) = x2 1 x1 2 ,
compute their Lie bracket, and determine the flows generated by them.
Exercise 4.3.17. Show that the Lie algebra of SO(n) is isomorphic to the Lie
algebra of skew-symmetric nn matrices with Lie bracket [X, Y ] = XY Y X.

Exercise 4.3.18. Show that R3 endowed with the cross-product is a 3-

dimensional Lie algebra isomorphic to the Lie algebra of SO(3).
Exercise 4.3.19. For v R3 , let Xv denote the vector field on R3 , defined

by
Xv (x) = v x.
Describe the flow generated by Xv , and prove that [Xv , Xw ] = Xvw .
Exercise 4.3.20. Show that the Lie algebra of left invariant vector fields on
a Lie group is isomorphic to the Lie algebra of right invariant vector fields.

4.3.2 Subspace Distributions and the Frobenius Theo-

rem
Frobenius theorem for subspace distributions is a geometric formulation of
theorem giving necessary and sufficient condition for the solvability of certain
systems of partial differential equations.
Definition 4.3.21. Let M be an n-dimensional manifold, 0 k n a
fixed integer. A k-dimensional smooth subspace distribution on M is a map
D, which assigns to each point p of M a k-dimensional subspace Dp of the
tangent space Tp M in such a way, that each point p M has an open
neighborhood U on which one can define k smooth vector fields X1 , . . . , Xk
with the property that for any q U , the subspace Dq is spanned by the
vectors (X1 )q , . . . , (Xk )q .
The word distribution is used in many different senses in mathematics. It
can mean a generalized function, a probability distribution, a subspace dis-
tribution. In what follows, it will always mean a subspace distribution.
Definition 4.3.22. We say that a vector field X is tangent to the distribution
D if Xp Dp for all p M .
A distribution D is called involutive or completely integrable, if for any two
vector fields that are tangent to D, the Lie bracket of the vector fields is also
tangent to D.
Definition 4.3.23. An injectively immersed submanifold of M is a pair
(N, ), where N is a manifold and : N M is an injective immersion,
that is an injective map the derivative map Tq : Tq N T(q) M of which is
injective for each q N . We say that (N, ) contains, or goes through a point
p of M if p (N ).
Definition 4.3.24. An integral manifold of a distribution D is an injectively
immersed submanifold (N, ) for which T (Tq N ) = D(q) for all q N .
Proposition 4.3.25. If the distribution D on M has the property that each
point of M is contained in an integral manifold of D, then D is involutive.
Proof. Let X and Y be two vector fields tangent to D, p M be an arbitrary
point. To show that [X, Y ]p Dp , consider an integral manifold (N, ) of D
passing through p. From the definition of an integral manifold, there exists
two uniquely defined smooth vector fields X and Y on N which are -related
with X and Y respectively. By Proposition 4.3.13, then [X, Y ] is also -related
to [X, Y ], so if q = 1 (p), then
[X, Y ]p = T ([X, Y ]q ) T (Tq N ) = Dp ,
which was to be proved.
230 4. Manifolds
The theorem of Frobenius claims that the converse of this proposition is also
true.
Theorem 4.3.26 (Frobenius Theorem). Let D be an involutive smooth k-

dimensional subspace distribution on the manifold M . Then every point of
M is contained in an integral manifold of D. Moreover, there is a local
coordinate system = (x1 , . . . , xn ) in an open neighborhood U of any point
p M such that for any q U , the submanifold Nq = {q U | xi (q) =
xi (q), for i = k + 1, . . . , n} of M together with the inclusion map : Nq
M is an integral manifold of D passing through q. Integral manifolds also
have the local uniqueness property that for any connected integral submanifold
(N, ) of D with image in U the functions xi are constant for k +1 i n.
Proof. Observe first that the properties imposed on the chart (x1 , . . . , xn ) are
equivalent to the condition that the subspace Dq is spanned for all q U by
the vectors i |q for i = 1, . . . , k. This means geometrically that any involutive
distribution can be straightened locally, that is locally diffeomorphic to a
subspace distribution on Rn assigning to each point a translate of a given
subspace.
We prove the theorem by induction on the dimension of the distribution.
If k = 0, then the statement is obvious. Assume that k 1 and that the
theorem is proved for (k 1)-dimensional distributions. In order to construct
a suitable chart around p, choose an open set V around p, over which the
distribution can be spanned by the pointwise linearly independent smooth
vector fields X1 , . . . , Xk .
By Proposition 4.3.3, there is a a local coordinate system = (y 1 , . . . , y n )
defined on an open neighborhood W V of p for which X1 |W = 1 . Denote
by W the (n 1)-dimensional submanifold defined by the equation
y 1 = 0.
Define the vector fields Y1 , . . . , Yk on W by the equations Y1 = X1 , Yi =
Xi Xi (y 1 )X1 , (i = 2, . . . , k). It is clear that Y1 , . . . , Yk span the same
subspace at each point of W as X1 , . . . , Xk . For i 2, we also have
Yi (y 1 ) = Xi (y 1 ) Xi (y 1 )X1 (y 1 ) = 0,
which means that the vector field Yi is tangent to the level sets of the first
coordinate function y 1 , in particular, it is tangent to .
Let Zi be the restriction of the vector field Yi onto . We claim that the
distribution D0 spanned by the vector fields Z2 , . . . , Zk on is involutive.
Indeed, as D is involutive, the Lie bracket [Yr , Ys ] of the vector fields Yr and
Ys is tangent to D hence can be written as
k
X
[Yr , Ys ] = cirs Yi (4.1)
i=1
where the coefficients cirs are smooth functions on W . Evaluating both sides of
this equation on the function y 1 for r, s 2 we obtain 0 = c1rs . Thus restrict-
Pk
ing the equation (4.1) onto gives the decomposition [Zr , Zs ] = i=2 cirs Zi
which implies easily that D0 is involutive.
Applying the induction hypothesis for the distribution D0 we obtain a local
coordinate system = (z 2 , . . . , z n ) on defined around p, such that D0 is
spanned by the vector fields i , i = 2, . . . , k. (For convenience, we shifted
the indices by 1. This means that i denotes derivation with respect to
the (i 1)st coordinate z i of .) Suppose that the transit map between the
coordinates (y 2 , . . . , y n )| and (z 2 , . . . , z n ) is given by the smooth functions
f i in such a way that
z i = f i (y 2 , . . . , y n )| , (i = 2, . . . , n). (4.2)
Introduce a new local coordinate system = (x1 , . . . , xn ) on M defined in

some open neighborhood of p with the help of the equations
x1 = y 1 , (4.3)
i i 2 n
x = f (y , . . . , y ), (i = 2, . . . , n). (4.4)
Shrink the open domain on which is defined to an open neighborhood U

of p in such a way that the image of U under the chart be equal to an
open cube K the edges of which are parallel to the coordinate axes. We can
assume without loss of generality that K is centered at the origin, which is
the image (p) of p as well.
We show that the local coordinate system satisfies the requirements of
the theorem. For this purpose, it is enough to check that Yi (xr ) = 0, if
i k and k + 1 r n. Computing the derivative matrix of the transit
map (4.3-4.4) we can see that 1 = 1 = Y1 on U . This implies at once
that Y1 (xr ) = 1 (xr ) = 0 for k + 1 r n. Suppose 2 i k. Using
the decomposition (4.1) for [Y1 , Yi ] and applying this derivation to xr with
(k + 1 r n) we get
k
X
1 (Yi (xr )) = Y1 (Yi (xr )) = [Y1 , Yi ](xr ) = cs1i Ys (xr ). (4.5)
s=1
Let the projection of the cube K onto the first coordinate axis be the interval
(a, a) and denote by K 0 its orthogonal projection onto the hyperplane x1 =
0. Fix an arbitrary point c = (0, c2 , . . . , cn ) K 0 and denote by qt U the
point the coordinates of which are x1 (qt ) = t, x2 (qt ) = c2 , . . . , xn (qt ) = cn .
Consider the functions hir : (a, a) R, hir (t) = (Yi (xr ))(qt ), (2 i k <
r n). According to the equations (4.5), the functions hir (i = 2, . . . , k)
232 4. Manifolds
satisfy the linear ordinary differential equation

k
X
h0ir = cs1i hsr , (i = 2, . . . , k), (4.6)
s=2
where cs1i (t) = cs1i (qt ). We know also that at the point q0 , the vector
Yi (q0 ) Dq0 0 is a linear combination of the vectors s |q0 = s |q0 s = 2, . . . , k.
For this reason, (Yi (xr ))(q0 ) = 0, that is hir (0) = 0 for all 2 i k < r n.
The functions hir 0 also solve the differential equation (4.6) with initial
condition hir (0) = 0, thus, by the uniqueness of solutions with a given initial
condition, hir (t) = (Yi (xr ))(qt ) 0 for all 2 i k < r n. As choosing
the constants c2 , . . . , cn and t (a, a) properly qt can be any point of U ,
Yi (xr ) 0 for all 2 i k < r n, and this is what we wanted to prove.
Definition 4.3.27. A maximal integral manifold of a distribution is a con-

nected integral manifold (N, ) the image (N ) of which is not a proper subset
of the image of any other connected integral manifold.
Theorem 4.3.28. Let D be a k-dimensional involutive distribution on the

manifold M . Then there is a unique maximal integral manifold through any
point p in M . Uniqueness should be understood as follows. If (N1 , 1 ) and
(N2 , 2 ) are two maximal integral manifolds through a given point p, then
there is a diffeomorphism h : N1 N2 such that 1 = 2 h.
Proof. Existence. Let N M be the subset of those points q M that can

be connected to p by a piecewise smooth continuous curve the speed vectors
of which are in the distribution D. It is clear that N contains the image of
any connected integral manifold passing through p, therefore it is sufficient
to create a smooth manifold structure on N , which makes N together with
the inclusion map : N M an integral manifold of D.
By the Frobenius theorem, M can be covered by a countable family of local
coordinate systems
i = (x1i , . . . , xni ) : Ui Rn
in such a way that i (Ui ) is a cube the edges of which are parallel to the
coordinate axes, furthermore, the connected integral manifolds of D lying in
Ui are exactly the connected open subsets of a submanifold of U that can
be defined by a system of equations xri = cr , (k + 1 r n), with some
constants cr .
Call the submanifolds of Ui that can be defined by a system of equations of
the form xri = cr , (k + 1 r n) the slices of Ui .
Choose for all q N an index iq such that q Uiq , and let Siq denote the slice
of Uiq which contains q. Observe that open subsets of the subspace topologies
4.4. Tensor Bundles and Tensor Fields 233
of the slices {Siq | q N } give a base of a locally Euclidean topology on N .

Furthermore, the charts
(x1iq , . . . , xkiq )|Siq : Siq Rk
form a C -atlas on N . It is more or less a routine work to check that this

is a right smooth manifold structure on N . Probably the only point which
requires some idea is the proof of the fact that N is second countable. This is
a corollary of the fact that N contains at most a countable number of slices.
The latter is true, because due to the connectedness of N one can choose for
any two slices S, S 0 contained in N a sequence S = S1 , S2 , . . . , Sm = S 0 of
slices in N such that any two consecutive slices in the sequence have a point
in common. Thus, it is sufficient to prove that each slice intersects at most
countable number of other slices. Since each slice is contained in one of the
countable number of sets Ui , it suffices to show that for any pair of indices
i, j N a slice S of Ui can intersect at most countable number of slices of Uj .
If S cuts the slices {S | I} of Uj , then the intersections S S are disjoint
open subsets in S, however, as S is second countable it can contain at most
a countable number of pairwise disjoint open subsets, hence I is countable.
The proof of uniqueness is simple and left to the reader.
4.4 Tensor Bundles and Tensor Fields

We gave definition of tensor fields on hypersurfaces earlier. In this section,
we are going to extend the definition to manifolds. The general definition will
introduce tensor fields as sections of the tensor bundles over the manifold.
Tensor bundles are some vector bundles produced from the tangent bundle.
Definition 4.4.1. A smooth real k-dimensional vector bundle (E, ) over a
manifold M consists of
a smooth manifold E called the total space of the bundle,
a smooth map : E M , called the projection of the bundle,
a given k-dimensional linear space structure on the preimage Fp =
1 (p) of each point p M .
The linear space Fp is called the fiber over p, M itself is the base space
of the bundle. In order to form a vector bundle, these data should satisfy
the following local triviality condition: for any point p M there is an open
neighborhood U of p and a diffeomorphism h : 1 (U ) U Rk such that for
any point q U , the restriction of h onto the fiber Fq is a linear isomorphism
between Fq and {q} Rk , where the linear space structure of the latter space
is essentially the standard linear structure of Rk .
234 4. Manifolds
Examples.
A trivial vector bundle over M is a product space E = M V with
the projection map = M onto the first component, where V is a
k-dimensional linear space over R.
The tangent bundle : T M M of an n-dimensional manifold M is
an n-dimensional vector bundle. It is usually not a trivial bundle.
There are several methods to construct new vector bundles from given ones.
To construct tensor bundles we can apply the following general construction
for the tangent bundle of a manifold.
Consider the category Vk of k-dimensional linear spaces in which the mor-
phisms between two k-dimensional linear spaces are the linear isomorphisms
between them and let be a covariant functor from Vk to Vl which is smooth
in the sense that for any two k-dimensional linear spaces V W , the map
: Mor(V, W ) Mor((V ), (W )) is smooth. Smoothness means that if
we fix some bases in V ,W , (V ) and (W ) to identify linear maps with their
matrices, then the coefficients of the matrix of (L) are smooth functions of
the matrix coefficients of L Mor(V, W ).
If : E M is a k-dimensional smooth vector bundle over M then applying
to each fiber of E, we can construct a new smooth vector bundle over M
with l-dimensional fibers. The total space of this bundle is the disjoint union

[
E = (Fp ),
pM
where Fp is the fiber over p in E. The projection of E assigns to each

point of (Fp ) the base point p. Thus, the fiber over p in the new bundle is
(Fp ) which is a linear space of dimension l. It remains to define the smooth
structure on the total space E and to show that the new bundle is locally
trivial.
Fix a linear isomorphism between the linear spaces I : (Rk ) Rl .
Choose for any point p in M an open neighborhood which is both the do-
main of a chart : U Rn and a neighborhood over which the bundle E is
trivialized by a diffeomorphism h : 1 (U ) U Rk as in the definition of
local triviality. Then we can define a chart
^
(, h) = ( , I (h)) : ( )1 (U ) Rn Rl
S
from and h on the set ( )1 (U ) = pM (Fp ) in the following way.
The first component of the chart (, ^ h) assigns to each element e
1
in ( ) (U ) the coordinates of the base point (e) with respect to the
chart . These coordinates are the first n coordinates of e. The symbolic
notation I (h)) in the second component of the chart means the following.
Since h is a local trivialization of the original bundle, it defines for each
q U a linear isomorphism hq : Fq Rk . The functor assigns to this
isomorphism an isomorphism (hq ) : (Fq ) (Rk ). Composing (hq )
with the isomorphism I, we obtain a linear isomorphisms (Rk ) Rl . The
^
last l coordinates of (, h)(e) for e (Fq ) are the coordinates of the vector
I((hq )(e)).
The set of all charts constructed this way form an atlas on E . The transit
map between any two charts is defined on an open subset of Rn+l and it is
smooth. Indeed, if and h are another chart and local trivialization over an
1
^
open subset V , then the transit (, ^
h) (, h) map is defined on the open
set (U V ) Rl , and maps (x, v) (U V ) Rl to
1
^
(, ^
h) (, h) (x, v) = (1 (x)), I (h1 (x) )((h1 (x) ))1 I 1 (v) .

This transit map is obviously smooth, since it is built up from compositions

of smooth functions. Thus, this atlas defines a smooth manifold structure on
E . The vector bundle we have constructed is locally trivial, as it is shown
by the diffeomorphisms
( , I ((h))) : ( )1 (U ) U Rl .
Let k, l 0 be two natural numbers. The construction of the tensor space

T (k,l) V from a linear space V can be thought of as a functor from Vn to
Vnk+l . If L : V W is a linear isomorphism, then the associated isomorphism
between T (k,l) V and T (k,l) W is
T (k,l) (L) = (L )1 (L )1 L L : T (k,l) V T (k,l) W.

| {z } | {z }
k times l times
Applying the functor T (k,l) to the tangent bundle of a manifold, we obtain

the the tensor bundle T (k,l) M M or the bundle of tensors of type (k, l)
over M .
There is another important vector bundle over M , the bundle of k-forms.
This is obtained by applying the functor : V 7 k (V ) to the tangent
bundle. The isomorphism corresponding to a linear isomorphism L : V W
under this functor is the isomorphism (L) uniquely characterized by the
identity
(L)(l1 lk ) = (L )1 (l1 ) (L )1 (lk ), for all l1 , . . . , lk V .
The resulting vector bundle of k-forms will be denoted by k T M

236 4. Manifolds
Definition 4.4.2. A smooth section of a vector bundle (E, ) over M is a

smooth map s : M E such that s = idM . In other words, a section
assign to each point of the manifold an element of the fiber Fp over that
point.
Definition 4.4.3. Smooth sections of the tensor bundle T (k,l) M are called
smooth tensor fields of type (k, l) over M .
Definition 4.4.4. A differential k-form over a manifold M is a smooth
section of the bundle k T M . The set of all differential k-forms over M are
denoted by k (M ).
Smooth sections of a vector bundle, in particular, tensor fields and differential
forms can be added and multiplied by a smooth function pointwise. With
these operations they are modules over the ring of smooth functions.
Using the derivative of a smooth map between two manifolds, we can push
forward or pull back certain types of tensors. Let f : M N be a smooth
map. Then for any point p M , the derivative map Tp f is a linear map from
Tp M to Tf (p) N .
If S : M T (0,l) M is a smooth tensor field of type (0, l), then it assigns to
each point p M an element S(p) of the tensor product Tp M Tp M .
Applying the kth tensor power Tp f Tp f of the derivative map Tp f we
can push S(p) forward to Tp f Tp f (S(p)) T (0,l) (Tf (p)N ). Thus, the
tensor assigned to a point by a tensor fields of type (0, l) on M can be pushed
forward to N , but, in general, these pushed forward tensors may not extend
to a smooth tensor field on N or if they extend, the extension may not be
unique. Thus, tensor fields of type (0, l) may not be pushed forward to tensor
fields of the same type on N . The situation is similar to the special case of
vector fields, which was discussed when we introduced f -related vector fields.
Assume now that S : N T (k,0) N is a tensor field of type (k, 0) on the image
manifold N . Then for any q N , S(q) is an element of (Tq N ) (Tq N ) .
For p M , applying the linear transformation (Tp f ) (Tp f ) to S(f (p))
we obtain a tensor of type (k, 0) over the linear space Tp M . In contrast to
type (k, 0) tensors, pulling back a tensor field of type (0, l) on N pointwise,
as described, we obtain a smooth tensor field on M . The pull back of the
tensor field S by the smooth function f will be denoted by f S.
The pull-back of a tensor field of type (k, 0) can also be explained as follows.
The tensor space T (k,0) (Tq M ) is naturally isomorphic to the linear space of
k-linear functions on Tq N . Therefore, we can think of S(q) as a k-linear
function on Tq N . Similarly, for p M f S(p) is a k-linear function on
Tp M . By the definition of f S, the value of f S(p) on the tangent vectors
v1 , . . . , vk Tp M can be computed by the formula
f S(p)(v1 , . . . , vk ) = S(f (p))(Tp f (v1 ), . . . , Tp f (vk )).

Differential forms can be interpreted as alternating tensors of type (k, 0),

therefore, as a special case of the pull back of type (k, 0) tensors, differential
forms on N can be pulled back to a differential form on M . Pull back for
differential forms will also be denoted by f : k N k M .
In general, tensor fields of mixed type (k, l), where k, l 1, can neither be
pushed forward, nor be pulled back by a smooth map. However, there is a
special case, when we can both push forward and pull back tensor fields of
any type. This is the case, when f is a diffeomorphism. The idea is that
a diffeomorphism identifies the two manifolds, so any tensor field on one of
them corresponds to a tensor field on the other.
More formally, if f : M N is a diffeomorphism S is a tensor field of type
(k, l) on M , then S can be pushed forward to a type (k, l) tensor field f S
on N by the formula
(f S)(f (p)) = ((Tp f ) )1 ((Tp f ) )1 Tp f Tp f (S(p)).
| {z } | {z }
k times l times
To pull back a type (k, l) tensor field S on N , we essentially push it forward

by f 1 , that is
(f S)(p) = (Tp f ) (Tp f ) (Tf (p) f 1 ) (Tf (p) f 1 ) (S(f (p))).
| {z } | {z }
k times l times
It is useful to know how one can work with tensor fields in local coordinates.
If = (x1 , . . . , xn ) : U Rn is a local coordinate system on the manifold,
then the vector fields 1 , . . . , n induced by give a basis in the tangent space
Tp M for all p M . Taking the dual basis of the basis at each point we obtain
n tensor fields of type (1, 0), which are usually denoted by dx1 , . . . , dxn . The
tensor fields dxi1 dxik j1 jl , (1 i1 , . . . , ik , j1 . . . .jl n)
yield a basis of the tensor space T (k,l) (Tp M ) at each point p U , therefore,
any smooth tensor field S of type (k, l), can be written uniquely as a linear
combination
X n
S= Sij11...i
...jl
k
dxi1 dxik j1 jl
i1 ,...,ik ,j1 ....jl =1
of these tensor fields, where the coefficients Sij11...i

...jl
k
are smooth functions on
U.
Similarly, any differential k-form on U can be decomposed uniquely as a
linear combination
X
= i1 ...ik dxi1 dxik ,
1i1 <<ik n
where the coefficients i1 ...ik are smooth functions on U .

238 4. Manifolds
4.5 The Lie Derivative

Let M be a smooth manifold and X be a smooth vector field on M . Let
t : Ut Ut , (t R) be the flow generated by the vector field.
Definition 4.5.1. The Lie derivative LX S of a tensor field S of type (k, l)

on M is the type (k, l) tensor field
t S S
LX S = lim .
t0 t
The pulled back tensor field t S is defined only on Ut . Nevertheless, LX S
is properly defined on the whole manifold M , since for each p M , there
is a positive t0 , such that p Ut0 Ut0 , and then t S is defined at p for
|t| t0 , so the limit can be taken at p.
The definition shows exactly the key idea behind the Lie derivative. It is the
initial speed at which the tensor field changes when we start moving it by the
flow of the vector field X. However, the limit on the right-hand side is not
the most practical way to compute the Lie derivative, since to compute the
flow, we have to solve an ordinary differential equation. Our goal now is to
work out formulae for the computation of the Lie derivative without having
to compute the flow of the vector field explicitly. We start with the simplest
types of tensor fields.
Smooth functions on M are the tensor fields of type (0, 0), or differential 0-
forms. The Lie derivative for them is the well known differentiation of smooth
functions with respect to vector fields.
Proposition 4.5.2. For a smooth function f on M , LX f = Xf .
Proof. In general, if F : M N is a smooth map, and f is a smooth function

on N , then the pull back of f with F is the composition F (f ) = f F .
Applying this for M = N , and F = t , we get
f (t (p)) f (p) d
LX f (p) = lim = f (t (p))|t=0 = Xp (f ).
t0 t dt
In the last step, we used the fact that p (t) = t (p) is an integral curve of
X, therefore its speed vector at 0 is 0 (0) = Xp .
Vector fields are tensor fields of type (0, 1). Lie derivation of vector fields is
also a familiar operation.
Proposition 4.5.3. If X and Y are smooth vector fields on M , then LX Y =

[X, Y ].
4.5. The Lie Derivative 239
Proof. Let t and t denote the flows generated by X and Y respectively.

Choose a smooth function h on M and a point p M . Then
((t Y )(h))(p) (Y (h))(p) d
(LX Y (h))(p) = lim = Tt (p) t (Yt (p) )(h)|t=0
t0 t dt
Since p (s) = s (t (p)) is an integral curve of Y , 0 (0) = Yt (p) , and
d
Tt (p) t (Yt (p) ) = ds t s t (p)|s=0 , thus
2
(LX Y (h))(p) = h(t s t (p))|t=s=0 .
ts
Consider the function F defined on a neighborhood of the origin in R3 by the
formula
F (u, v, w) = h(u v w (p)).
Using the chain rule
2
h(t s t (p))|t=s=0
ts
2
= F (t, s, t)|t=s=0
ts
2 2
= F (u, v, 0)|u=v=0 + F (0, v, w)|w=v=0
uv wv
2
= [h(s t (p)) h(t s (p))]|t=s=0
ts
= Xp (Y (h)) Yp (X(h)) = [X, Y ]p (h).
The Lie derivative of a tensor field of type (1, 0), which is the same as a
differential 1-form, it can be obtained from a Leibniz-type differentiation
rule. If 1 (M ) is a differential 1-form, then it assigns to each p M a
linear function p on Tp M . Therefore, if Y is a vector field on M , we can
evaluate on X pointwise and get a smooth function (X), the value of
which at p is p (Xp ). The smooth function (X) depends on and X in a
bilinear way over R. For this reason, the Lie derivative of (X) can be also
computed by the following Leibniz-type formula.
Proposition 4.5.4. For any two vector fields X, Y and any differential
1-form on M , we have
LX ((Y )) = (LX )(Y ) + (LX Y ).
In particular, rearranging, and using the previous two propositions
(LX )(Y ) = X((Y )) ([X, Y ]).
240 4. Manifolds
Proof. We use the standard trick to prove all Leibniz-type rules as follows
t ((Y )) (Y )
LX ((Y )) = lim =
t0 t
(t Y ) (t Y ) + (t Y ) (Y )
lim t
t0 t
t Y Y

= lim (lim t Y ) + lim t
t0 t t0 t0 t
= (LX )(Y ) + (LX Y ).
Tensor fields can always be multiplied by the tensor product, differential
forms can be multiplied by the wedge product pointwise. The Leibniz rule
also holds for these operations.
Proposition 4.5.5. If S1 and S2 are two arbitrary tensor fields, not nec-
essarily of the same type, , are arbitrary differential forms, then for any
smooth vector field X, we have
LX (S1 S2 ) = (LX S1 ) S2 + S1 (LX S2 ),
and
LX ( ) = (LX ) + (LX ).
Both equations can be proved by the standard trick. We leave the details to
the reader.
Proposition 4.5.6. If in addition to the assumptions of the previous propo-
sition S1 and S2 have the same type, and the differential forms and are
of the same degree, then for any two real numbers , R we have
LX (S1 + S2 ) = LX S1 + LX S2 ,
and
LX ( + ) = LX + LX .
Locally, any tensor field can be written as the sum of a finite number of
tensor products of some vector fields and differential one-forms. Similarly,
any differential form can be written locally as a sum of wedge products of
some differential 1-forms. Thus, the above rules are enough to compute the
Lie derivative of any tensor field and differential form.
The following proposition tells us how the Lie derivation operator LX depends
on the vector field X.
Proposition 4.5.7. The operator LX depends R-linearly on X, and satisfies
the commutator relation
LX LY LY LX = L[X,Y ] .
4.5. The Lie Derivative 241
Proof. The proof follows the steps we went through to work out formulae to
compute LX on more and more complicated tensor fields.
If the Lie derivations in the statement are applied to smooth functions, then
linearity in X is obvious, the commutator relation is just a reformulation of
the definition of the Lie bracket.
When Z is a vector field,
LX+Y Z = [X + Y, Z] = [X, Z] + [Y, Z] = (LX + LY )Z,
and
(LX LY LY LX )Z = [X[Y, Z]] [Y, [X, Z]] = [[X, Y ], Z] = L[X,Y ] Z.
by the Jacobi identity.

If is a 1-form, and Z is an arbitrary vector field, then
(LX+Y )(Z) = (X + Y )((Z)) ([X + Y, Z])

= X((Z)) + Y ((Z)) ([X, Z]) ([Y, Z])
= ((LX + LY ))(Z),
furthermore,
(LX LY LY LX )()(Z)
= X((LY )(Z)) (LY )([X, Z]) Y ((LX )(Z)) + (LX )([Y, Z])
= X(Y ((Z))) X(([Y, Z])) Y (([X, Z])) + ([Y, [X, Z]])
Y (X((Z))) + Y (([X, Z])) + X(([Y, Z])) ([X, [Y, Z]])
= [X, Y ]((Z)) ([[X, Y ], Z]) = (L[X,Y ] )(Z).
Next we show that if the statement is true when the Lie derivatives are
evaluated on the tensor fields S1 , S2 , or on the differential forms and ,
then it also holds on S1 S2 or on respectively. We check only for the
tensor product. The wedge product can be treated in the same way.
(LX+Y )(S1 S2 ) = (LX+Y )(S1 ) S2 + S1 (LX+Y )(S2 )

= (LX + LY )(S1 ) S2 + S1 (LX + LY )(S2 )
= (LX S1 ) S2 + (LY S1 ) S2 + S1 (LX S2 )
+ S1 (LY S2 )
= (LX + LY )(S1 S2 ).
242 4. Manifolds
As for the commutator,
(LX LY LY LX )(S1 S2 )
= LX (LY (S1 ) S2 + S1 LY S2 ) LY (LX (S1 ) S2 + S1 LX S2 )
= LX (LY (S1 )) S2 + LY S1 LX S2 + LX S1 LY S2
+ S1 LX (LY (S2 )) LY (LX (S1 )) S2 LX S1 LY S2
LY S1 LX S2 S1 LY (LX (S2 ))
= (LX LY LY LX )(S1 )) S2 + S1 (LX LY LY LX )(S2 ))
= (L[X,Y ] S1 )) S2 + S1 (L[X,Y ] S2 ) = L[X,Y ] (S1 S2 ).
It is easy to see that if the identities we want to show hold on two tensor
fields or differential forms of the same type, then the identity holds also on
their sum. Since locally, e.g. in the domain of any chart, tensor fields can be
written as sums of tensor products of vector fields and 1-forms, and similarly,
differential forms can be written as sums of wedge-products of 1-forms, the
proposition is proved.
Definition 4.5.8. We say that (the flow of ) a vector field X leaves a tensor
field S invariant if t (S) = S|Ut for all t R, where t : Ut Ut (t R)
is the flow generated by X.
Proposition 4.5.9. The flow of a vector field X leaves the tensor field S
invariant if and only if LX S = 0.
Proof. If the flow leaves S invariant, then
t S S SS
LX S = lim = lim = 0.
t0 t t0 t
To prove the converse, assume that LX S = 0. Denote by St the one-
parameter family of tensor fields St = t S. S0 = S and for each p M ,
St (p) is defined on an open interval around 0, so to prove the statement, it
d
suffices to check that dt St 0. However, this is true since

St St0 h S S
lim = lim t0 = t0 (LX S) = 0.
tt0 t t0 h0 h
Corollary 4.5.10. Vector fields leaving a tensor field invariant form a Lie
subalgebra of the Lie algebra of all vector fields.
Proof. Indeed, if X and Y leave S invariant, then LX+Y S = LX S +
LY S = 0 and
L[X,Y ] S = LX (LY S) LY (LX S) = 0.

4.6. Differential Forms 243
4.6 Differential Forms

A differential k-form on a manifold M assigns to each point p M an
element of k (Tp M ) . As elements of k V for a linear space V can be
identified with the linear space Ak (V ) of alternating k-linear functions, we
can evaluate (p) on any k tangent vectors v1 , . . . , vk Tp M . In particular,
if we are given k smooth vector fields X1 , . . . , Xk on M , then we can evaluate
on X1 , . . . , Xk pointwise and get a smooth function on M . This function
will be denoted by (X1 , . . . , Xk ).
The pointwise wedge product of a differential k-form k (M ) and a
differential l-form l (M ) is a differential (k + l)-form. Wedge product of
differential forms is not a commutative operation. Instead of that it satisfies
the supercommutativity relation
= (1)kl .
4.6.1 Interior Product by a Vector Field

Definition 4.6.1. The interior product of a smooth vector field X and a
differential k-form on a differentiable manifold M is a differential (k 1)-
form X . It is defined by the property that
X (X1 , . . . , Xk1 ) = (X, X1 , . . . , Xk1 ).
When k = 0, we set X = 0.
Interior product L
with the vector field X can be thought of as a linear map
dim M k
from (M ) = k=0 (M ) into itself. This is a degree -1 map, as it
decreases the degree of any form by 1.
Definition 4.6.2. In general, a linear map L : (M ) (M ) is a degree

d map, if L(k (M )) k+d (M ) for all k.
It is interesting that although interior product has a linear algebraic charac-

ter, it satisfies a Leibniz-type rule.
Definition 4.6.3. A linear map L : (M ) (M ) is called a superderiva-

tion if for any k (M ) and l (M ), we have
L( ) = L() + (1)k L().
Proposition 4.6.4. Interior product with the vector field X is a superderiva-

tion.
244 4. Manifolds
Proof. Both sides are bilinear functions of and , therefore, it is enough to

check the graded version of the Leibniz rule for forms that generate k (M )
and l (M ) additively. It is also enough to show the graded Leibniz identity
on domains of charts, since any point is contained in the domain of a chart.
In the domain of a chart, any k-form can be written as the sum of wedge
products of k-tuples of 1-forms, so it is enough to show the graded Leibniz
rule for the special case when both and are wedge products of 1-forms.
Let l1 , . . . , ls be 1-forms. Then

l1 (X) l1 (X1 ) . . . l1 (Xs1 )
l2 (X) l2 (X1 ) . . . l2 (Xs1 )
X (l1 ls )(X1 , . . . , Xs1 ) = det . .

.. ..
.. . .
ls (X) ls (X1 ) ... ls (Xs1 )
Computing the determinant by Laplace expansion with respect to the first
column we obtain
X (l1 ls )(X1 , . . . , Xs1 )

X s
= (1)i+1 li (X) (l1 lbi ls )(X1 , . . . , Xs1 ),
i=1
where the hat over li means that it is omitted. This means that
s
X
X (l1 ls ) = (1)i+1 li (X) (l1 lbi ls ).
i=1
Consider the special case of the statement, when = l1 lk and =

lk+1 lk+l . Then
k+l
X
X ( ) = X (l1 lk+l ) = (1)i+1 li (X) (l1 lbi lk+l )
i=1
k
X
= (1)i+1 li (X) (l1 lbi lk+l )+
i=1
l
X
+ (1)k+i+1 lk+i (X) (l1 ld
k+i lk+l )
i=1
= X () + (1)k X ().
By the introductory arguments, this special case implies the general one.
Exercise 4.6.5. Show that for any two vector fields X an Y , we have
X Y = Y X .
Exercise 4.6.6. Show that the Lie derivation of differential forms and the in-
terior product with a vector field satisfies the commutation relation
LX Y Y LX = [X,Y ] .
4.6.2 Exterior Differentiation

We can associate to each smooth function a differential 1-form, its differential.
Definition 4.6.7. The differential df of a smooth function f 0 (M ) is the
differential 1-form df 1 (M ), the evaluation of which on a vector field X
is given by df (X) = X(f ).
Remark. If = (x1 , . . . , xn ) : U Rn is a chart, then the 1-forms deter-
mined by the pointwise dual basis of the basis vector fields 1 , . . . , n were
denoted by dx1 , . . . , dxn . The notation is motivated by the fact that these
1-forms can also be obtained as the differentials of the coordinate functions.
Indeed,
dxi (j ) = j (xi ) = ji .
Our goal is to extend the differential d from functions to all differential forms.
Theorem 4.6.8. There is a unique superderivation d : (M ) (M ) of
degree 1, which has the following properties.
For f 0 (M ), df is the differential of the function f .
d d = 0.
The superderivation d defined uniquely by the theorem is called the exterior
differentiation of differential forms.
Proof. To prove uniqueness, first observe that if two k forms 1 , 1 coincide
on an open neighborhood V of a point p, then d1 and d2 must coincide at
p. Indeed, choose a smooth bump function h 0 (M ) such that supp h V
and h 1 on a neighborhood of p. Then h(1 2 ) = 0, therefore,
0 = d(h(1 2 )) = dh (1 2 ) + h(d1 d2 ).
Evaluating at p gives (d1 )(p) = (d1 )(p).

Let p be an arbitrary point in M and choose a chart = (x1 , . . . , xn ) : U
Rn around p. Fix also a smooth function g 0 (M ) such that supp g U
and g|V 1 on an open neighborhood V U of p. Introduce the notation
that for a smooth function f defined on U , denote by f the smooth function
on M defined by (
f (q)g(q), if q U,
f(q) =
0, otherwise.
246 4. Manifolds
Let be an arbitrary differential k-form. We can write |U uniquely as

X
|U = i1 ...ik dxi1 dxik .
1i1 <<ik n
Consider the differential form k (M ) defined by

X
= i1 d x
i1 ...ik dxf fik .
1i1 <<ik n
The forms and coincide on V and the computation of d on is uniquely

dictated by the properties of d, thus,

X
(d)(p) = (d)(p) = i1 d x
di1 ...ik dxf f ik (p).
1i1 <<ik n
Since by this method we can compute the value of d at any point p, d is

uniquely defined by its properties.
For the existence, we first construct for each chart = (x1 , . . . , xn ) : U Rn
an exterior differentiation d : (U ) (U ) on U . For (U ) write
as X
= i1 ...ik dxi1 dxik ,
1i1 <<ik n
and set X
d = di1 ...ik dxi1 dxik .
1i1 <<ik n
Let us check that d satisfies all the requirements. It is obviously linear and
increases the degree of forms by 1, and coincides with d on smooth functions.
It is also a superderivation. Indeed, if l (M ) has the decomposition
X
= j1 ...jl dxj1 dxjl ,
1j1 <<jl n
then
d ( ) =
X X
d i1 ...ik j1 ...jl dxi1 dxik dxj1 dxjl
1i1 <<ik n 1j1 <<jl n
X X
= d(i1 ...ik j1 ...jl ) dxi1 dxik dxj1 dxjl
1i1 <<ik n 1j1 <<jl n
For smooth function f and g, the Leibniz-type formula d(f g) = df g + f dg

holds, since for any smooth vector field X,
d(f g)(X) = X(f g) = X(f ) g + f X(g) = (df g + f dg)(X).
Applying this special case to the general one, we obtain
d ( ) =
X X
j1 ...jl di1 ...ik dxi1 dxik dxj1 dxjl
1i1 <<ik n 1j1 <<jl n
X X
+ i1 ...ik dj1 ...jl dxi1 dxik dxj1 dxjl
1i1 <<ik n 1j1 <<jl n
= d + (1)k d.
It remains to show that d d = 0. As d is a superderivation,

d (d ()) =
X
d (di1 ...ik ) dxi1 dxik di1 ...ik d (dxi1 dxik )

1i1 <<ik n
X
= d (di1 ...ik ) dxi1 dxik ,
1i1 <<ik n
it is enough to prove that d (d(f )) = 0 for any smooth function f . Since

dx1 , . . . , dxn is the pointwise dual basis of 1 , . . . , n , df can be decomposed
as a linear combination of the 1-forms dxi as
n
X
df = (i f )dxi .
i=1
From this formula,

n
X n X
X n
d (df ) = d(i f ) dxi = (j i f )dxj dxi
i=1 i=1 j=1
X
= (i j f j i f )dxi dxj = 0.
1i<jn
If we consider two charts, : U Rn and : V Rn , then the restrictions

of both and onto U V give an exterior differentiation on U V . But the
uniqueness part of the theorem implies that these exterior derivations must
be the same. This means that if is a globally defined k-form on M , then
the forms d (|U ) taken for all charts : U Rn glue together to a global
(k + 1)-form d on M . The map 7 d is an exterior derivation, since all
the properties can be checked locally, using a local coordinate system.
248 4. Manifolds
In the rest of this section we prove some useful formulae involving the exterior
differentiation.
Proposition 4.6.9. If f : M N is a smooth map, k (N ) is a differ-

ential k-form on N , then d(f ()) = f (d).
Proof. The statement is true for smooth functions h 0 (M ), because for

any smooth vector field X on M and for any p M , we have
d(f (h))(Xp ) = Xp (f (h)) = Xp (h f ) = (Tp f (Xp ))(h)

= dh(Tp f (Xp )) = (f (dh))(Xp ).
It is also clear that a wedge product can be pulled back componentwise, i.e.,
f ( ) = f () f ().
If p M is an arbitrary point, then we can introduce a local coordinate
system = (x1 , . . . , xn ) : U Rn around f (p), where n = dim N (!). By
continuity of f , we can also find an open neighborhood V of p, such that
f (V ) U . Then for any k (N ), we can write
X
|U = i1 ...ik dxi1 dxik ,
1i1 <<ik n
hence
X
(f )|V = f (i1 ...ik )f (dxi1 ) f (dxik )
1i1 <<ik n
X
= (i1 ...ik f )d(xi1 f ) d(xik f ).
1i1 <<ik n
and
X
(f (d))|V = d(i1 ...ik f )d(xi1 f ) d(xik f ) = d(f ()).
1i1 <<ik n
Proposition 4.6.10. Lie derivation and exterior derivation of differential

forms commute, that is
LX d = d LX .
Proof. Using the usual notation, for any differential form , we have
d d
Lx (d) = (t (d))|t=0 = d(t ())|t=0 .
dt dt
In the last expression, we can change the order of differentiation with respect
to the parameter t and the exterior differentiation by Youngs theorem. In-
deed, if we write the one-parameter family of forms t = t using local
coordinates as
X
t (p) = i1 ...ik (p, t)dxi1 dxik ,
1i1 <<ik n
then
n
d X X
d(t )(p) = t j i1 ...ik (p, t)dxj dxi1 dxik
dt
1i1 <<ik n j=1
X Xn
= j t i1 ...ik (p, t)dxj dxi1 dxik
1i1 <<ik n j=1

d
=d t .
dt
d d

Consequently, Lx (d) = dt d(t ())|t=0 =d dt t ()|t=0 = d(Lx ).
Proposition 4.6.11 (Cartans Formula). Exterior differentiation and inte-

rior product with a vector field are related to Lie derivation of differential
forms by the formula
LX = X d + d X .
Proof. The formula is true for functions f 0 (M ), since
LX (f ) = X(f ) = df (X) = X (df ) = X (df ) + d(X f ).
In the last step we used that X is 0 on smooth functions.

If the formula holds for a form , then it holds for d. Indeed,
LX (d) = d(LX ) = d X d() + d d X () = d X d(),
on the other hand,
(X d + d X )(d) = X d d() + d X d() = d X d().
If the identity holds for the forms k (M ) and l (M ), then it holds

for their wedge product as well. Using the Leibniz rule for LX , we get
LX ( ) = LX () + LX ()
= X (d) + d(X ) + X (d) + d(X ).
250 4. Manifolds
Computing the other side
(X d + d X )( )
= X (d + (1)k d) + d(X + (1)k X )
= X (d) + (1)k+1 d X + (1)k X d + (1)2k X (d)+
+ d(X ) + (1)k1 X d + (1)k d X + (1)2k d(X )
= X (d) + d(X ) + X (d) + d(X ).
Since locally, in the domain of a chart, any differential form can be written
as the sum of wedge products of functions and differentials of functions, the
proposition is true.
The formula below can also be used as a coordinate free definition of the
exterior differential of a differential form.
Proposition 4.6.12. If k (M ), and X0 , . . . , Xk are smooth vector fields

on M , then
k
X
d(X0 , . . . , Xk ) = (1)i Xi ((X0 , . . . , Xi , . . . , Xk ))+
i=0
X
+ (1)i+j ([Xi , Xj ], X0 , . . . , Xi , . . . , Xj , . . . , Xk ).
0i<jk
Proof. We prove the formula by induction on k. For a smooth function

f 0 (M ), the second sum is empty, and df (X) = X(f ), so the formula
holds. Assume that the formula is true for (k 1)-forms. Using Cartans
formula
d(X0 , . . . , Xk ) = (X0 d)(X1 . . . , Xk )

= (LX0 )(X1 . . . , Xk ) d(X0 )(X1 . . . , Xk ).
(LX0 )(X1 . . . , Xk ) can be expressed by a Leibniz-type rule as follows
k
X
(LX0 )(X1 . . . , Xk ) = LX0 ((X1 . . . , Xk )) (X1 , . . . , LX0 Xi , . . . , Xk )
i=1
k
X
= X0 ((X1 . . . , Xk )) + (1)k ([X0 , Xi ], X1 , . . . , Xi , . . . , Xk ).
i=1
The induction hypothesis applied to the (k 1)-form X0 provides

k
X
d(X0 )(X1 . . . , Xk ) = (1)i Xi (X0 (X1 , . . . , Xi , . . . , Xk ))+
i=1
X
+ (1)i+j+1 X0 ([Xi , Xj ], X1 , . . . , Xi , . . . , Xj , . . . , Xk )
1i<jk
k
X
= (1)i Xi ((X0 , . . . , Xi , . . . , Xk ))+
i=1
X
+ (1)i+j ([Xi , Xj ], X0 , . . . , Xi , . . . , Xj , . . . , Xk ).
1i<jk
Adding the last two equations we obtain the statement for .
4.6.3 De Rham Cohomology

Definition 4.6.13. A differential form is said to be closed if d = 0. It is
exact if there is a differential form such that = d.
Exact differential forms are closed because of d d = 0, but the converse is
not always true. De Rham cohomology spaces measure in some sense how
much the converse statement fails to be true. Let k 0 be an integer, and
consider the spaces of closed and exact k-forms on M
Z k (M ) = k (M ) (ker d) and B k (M ) = k (M ) (im d).
Definition 4.6.14. The k-th de Rham cohomology space is the factor space
k
HdR (M ) = Z k (M )/B k (M ).
The kth de Rham cohomology space is a linear space. Its elements are equiv-
alence classes of closed differential k-forms, where two forms represent the
same equivalence class, also called cohomology class if and only if their dif-
ference is exact. L
k
It is useful to consider the direct sum HdR (M ) = k=0 HdR (M ) of all de
Rham cohomology spaces, because this linear space has also a multiplicative
structure induced by the wedge product of forms. Let [] and [] be two
cohomology classes represented by the closed forms and . Then is a
closed form and the cohomology class [ ] depends only on the cohomology
classes [] and []. Indeed, if we choose other representatives +d and +d
from these cohomology classes, then their wedge product
( + d) ( + d) = + d ( + d) + (1)deg

is in the same cohomology class as .

252 4. Manifolds

Definition 4.6.15. The linear space HdR (M ) together with its multiplicative
structure defined by
[][] = [ ]
is called the de Rham cohomology algebra of M .
Multiplication of cohomology classes is associative, distributive with respect

to addition. It has a unit element, represented by the constant 1 function
0
in HdR (M ). Instead of commutativity, it is supercommutative, that is, for
k l
[] HdR (M ) and [] HdR (M ), we have
[][] = (1)kl [][].
Every smooth map f : M N between two manifolds induce an algebra

homomorphism f : HdR
(N ) HdR k
(M ), which maps HdR k
(N ) to HdR (M ).
k
First observe, that if Z (N ) is a closed k-form on N , then its pull back
f is also closed, since d(f ) = f (d) = 0. The cohomology class of f
depends only on the cohomology class of since f ( + d) = f + d(f ).
Thus, f : HdR
(N ) HdR
(M ) is properly defined by f [] = [f ] and it
is an algebra homomorphism because f on differential forms is linear and
satisfies f ( ) = f () f ().
k
We can also say that HdR is a contravariant functor from the category of

smooth manifolds to the category of linear spaces, while HdR is a contravari-
ant functor from the category of smooth manifolds to the category of algebras.
Diffeomorphic manifolds have isomorphic de Rham cohomology spaces and
de Rham algebra. In fact, de Rham cohomology spaces are homotopy in-
variant as well. We shall prove here the seemingly weaker smooth homotopy
invariance.
Definition 4.6.16. Two smooth maps f0 : M N and f1 : M N between

the manifolds M and N are said to be piecewise smoothly homotopic if there
is a continuous map H : M [0, 1] N such that f0 (p) = H(p, 0) and
f1 (p) = H(p, 1) for all p M , and the interval has a subdivision 0 = t0 <
t1 < < tN = 1, for which the restriction of H onto M [ti1 , ti ] is smooth
for all 1 i N .
Theorem 4.6.17. If two smooth maps f0 : M N and f1 : M N are

piecewise smoothly homotopic, then they induce the same homomorphism

f0 = f1 : HdR
(N ) HdR (M ).
Proof. We may suppose that the homotopy map H : M [0, 1] N connect-

ing f0 and f1 is smooth, as any piecewise smooth homotopy is a concatenation
of smooth ones. For t [0, 1], define the embedding it : M M [0, 1] by
it (p) = (p, t). Then f0 = H i0 and f1 = H i1 . Set in general ft = H it .
We have to show that if k (N ) is a closed form, then f1 f0 is exact.

For this purpose, set = H . Let X be the vector field on M [0, 1] the
flow of which acts on M [0, 1] by the formula (p, t) = (p, t + ).
Then we have
Z 1
d
f1 f0 = i1 i0 = it dt
0 dt
Z 1 Z 1
= it (LX )dt = it (X (d) + d(X ))dt
0 0
Z 1 Z 1 Z 1
= it (d(X ))dt = d(it (X ))dt = d it (X )dt
0 0 0
Definition 4.6.18. Two smooth manifolds are (piecewise smoothly) homo-

topy equivalent if there are smooth maps f : M N and g : N M such
that f g is piecewise smoothly homotopic to idN and g f is piecewise
smoothly homotopic to idM .
Corollary 4.6.19. Piecewise smoothly homotopy equivalent smooth mani-

folds have isomorphic de Rham cohomology spaces and algebras.
Proof. Let f and g be the maps from the definition of homotopy equiv-
alence. Then the algebra homomorphisms f : HdR
(N ) HdR
(M ) and
g : HdR (M ) HdR (N ) have the property that f g = (g f ) = (idM ) =

(M ) and similarly, g f
idHdR = idHdR
(N ) , which means that f and g are
the inverses of one another, in particular, they are isomorphisms.
Recall that a subset U of Rn is called star-shaped if U has a point p such

that the segment [p, q] = {tp + (1 t)q | t [0, 1]} is contained in U for all
q U.
Corollary 4.6.20 (Poincare Lemma). If U Rn is a star-shaped open subset

of Rn , then every closed k( 1)-form on U is exact.
Proof. In terms of the de Rham cohomology spaces, Poincare Lemma claims

k
that HdR (U ) = 0 for k 1. Choose a point p U from which all other points
of U can be seen, and let V be the 0-dimensional manifold consisting the
single point p. We claim the U and V are smoothly homotopy equivalent.
Indeed, let f : V U be the embedding of V into U , g : U V be the
constant map sending each point to p. Then g f = idV . The composition
f g is not the identity of U but it is smoothly homotopic to it as it is shown
by the smooth homotopy map H : U [0, 1] U , H(q, t) = tp + (1 t)q.
254 4. Manifolds
The smooth homotopy equivalence implies that HdR k

(U ) k
= HdR (V ) for all k.
However, since V is 0-dimensional, there are no non-zero differential k-forms
k
on it for k 1, therefore HdR (V ) = 0 for k 1.
Exercise 4.6.21. Show that a function f is a closed differential 0-form if

and only if it is locally constant. From this fact, prove that the dimension
0
of the cohomology space HdR (M ) is the number of connected components of
M . In particular, for a star-shaped open domain U in Rn , HdR0
(U )
= R.
Remark. Applying some theorems about the approximation of continuous

functions by smooth ones, smoothness conditions in the above theorems can
be weakened. The following two approximation theorems are essential. Any
continuous map is homotopic to a smooth one, where homotopic means that
the homotopy map H connecting the two maps is assumed to be continuous
only. Secondly, if two smooth maps are homotopic, then there is also a smooth
homotopy between them.
This means that we can introduce the homomorphism f : HdR
(N ) HdR
(M )
for any continuous map f : M N taking a smooth map f: M N homo-
topic to f and setting f = f .
Homotopy equivalence of topological spaces is defined as the notion of piece-
wise smooth homotopy equivalence replacing the smoothness assumptions by
continuity. Using the above approximation theorems, we can also show that if
two manifolds are homotopy equivalent then they have isomorphic de Rham
cohomology spaces and algebras.
The latter theorem also enables us to speak about the de Rham cohomology
spaces and algebras of topological spaces homotopy equivalent to a manifold.
4.7 Integration of Differential Forms

To integrate a function, we need a measure. Since manifolds do not come with
a measure, we cannot integrate functions on a manifold, unless there is an
additional structure which produces a measure. On the other hand, differen-
tial k-forms can be integrated on parameterized k-dimensional submanifolds
(more precisely on k-dimensional chains) without any extra structure. We can
also integrate differential n-forms on domains of an n-dimensional manifolds,
but that integration requires an orientation of the manifold.
4.7. Integration of Differential Forms 255
4.7.1 Integration on Chains

Definition 4.7.1. The standard k-dimensional simplex k is the simplex in
Rk defined by the inequalities
k
n X o
k = (x1 , . . . , xk ) Rn xi 0 for i = 1, . . . , k and xi 1 .

i=1
The standard simplex is the convex hull of the affinely independent points
v0k = 0, v1k = e1 , . . . , vkk = ek , where 0 is the origin, e1 , . . . , ek is the standard
basis of Rk . These points are the vertices of k .
Definition 4.7.2. A smooth singular k-dimensional simplex in a manifold
M is a smooth map : k M from the standard k-dimensional simplex
to M .
The word singular does not mean that must be singular. It means that
is allowed to be singular.
Definition 4.7.3. A smooth singular k-dimensional chain in a manifold M is
a formal linear combination 1 1 + +s s of smooth singular k-dimensional
simplices i : k M with real coefficients i R. Smooth singular k-
dimensional chains in M form a linear space which we denote by Ck (M ).

The standard k-dimensional simplex k has (k + 1) facets. Let the i-th
be the convex hull of the vertices v0k , . . . , vc k k
i , . . . , vk . The standard (k 1)-
dimensional simplex can be mapped onto the ith facet of k by a unique
affine transformation lik : k1 k which maps the vertices v0k1 , . . . , vk1 k1
of k1 onto the vertices v k , . . . , vck , . . . , v k of the ith facet of k in the given

0 i k
order. We can give lik explicitly by the formulae
l0k (x1 , . . . , xk1 ) = (1 (x1 + + xk1 ), x1 , . . . , xk1 )
lik (x1 , . . . , xk1 ) = (x1 , . . . , xi1 , 0, xi , . . . , xk1 ) for 1 i k 1.
If 0 i < j k, then both ljk lik1 and lik lj1 k1

are affine maps from
k2 k
onto a (k 2)-dimensional face of , and both maps the vertices
v k2 , . . . , v k2 to the vertices v k , . . . , vck , . . . , vck , . . . , v k in an increasing order,
0 k2 0 i j k
consequently,
ljk lik1 = lik lj1
k1
.
Definition 4.7.4. The boundary of a smooth singular k-dimensional simplex
: k M is the (k 1)-dimensional smooth singular chain
k
X
= (1)i lik .
i=0
256 4. Manifolds
Ps
The boundary of a smooth singular k-dimensional chain j=1 s s is the
(k 1)-dimensional chain

s s

X X
s s = s s .
j=1 j=1
The boundary map is a linear map : Ck (M ) Ck1 (M ) for all k.
Proposition 4.7.5. The boundary of the boundary of a chain is 0, that is
= 0.
Proof. Since the linear space of smooth singular chains is generated by smooth
singular simplices and is linear, it is enough to check the statement for sin-
gular simplices. Let : k M be a smooth singular simplex. Then
k
X k k1
X X
() = (1)j ( ljk ) = (1)i+j ljk lik1 .
i=0 j=0 i=0
For i < j, the compositions ljk lik1 = lik lj1

k1
enter the last sum
with opposite signs, they cancel each other. After these cancellations nothing
remains.
The previous proposition gives rise to the definition of smooth singular ho-
mology groups of a manifold.
Definition 4.7.6. A smooth singular chain is called a smooth singular cycle
if its boundary is 0. A smooth singular chain is called a boundary, if it is the
boundary of a smooth singular chain.
Denote by Zk (M ) and Bk (M ) the linear spaces of k dimensional smooth
singular cycles and boundaries respectively. Then the kth smooth singular ho-
mology space with real coefficients is the factor space Hk (M ) = Zk (M )/Bk (M ).

Now let us define integration of differential forms on singular chains.
Definition 4.7.7. Let k (M ) be a differential k-form on the mani-
fold M , : k M be a smooth singular k-dimensional simplex in M .
Let (x1 , . . . , xk ) : Rk Rk be the identity chart of Rk and let us write the
differential k-form as
= f dx1 dxk ,
where f : k R is a smooth function. Then we define the integral of
over the smooth singular simplex as
Z Z
= f (x)dx,
k
where the second integral is the integral of f over the standard simplex k
with respect to the Lebesgue measure. Ps
The integral of over a k-dimensional smooth singular chain c = i=1 i i
is defined by
Z Xs Z
= i .
c i=1 i
Stokes Theorem is the most fundamental integral formula, a far-reaching

generalization of the NewtonLeibniz formula. Though it is much more gen-
eral then the latter one, we shall see that its proof, in fact, reduces to a finite
number of NewtonLeibniz formulas.
Theorem 4.7.8 (Stokes Theorem for Chains). Let c Ck (M ) be a k-
dimensional smooth singular chain in the manifold M , k1 (M ) be a
differential (k 1)-form. Then
Z Z
= d.
c c
Proof. Since both integrals depend depend linearly on c, it is enough to prove

the theorem for k-dimensional smooth singular simplices. Let : k M be
one. Denote by = the pull-back of onto k . Since is a (k 1)-form,
it can be written as
k
X
= fi dx1 dx
ci dxk .
i=1
The pull-back of d by is
k
X
(d) = d( ) = d = dfi dx1 dx
ci dxk
i=1

k
X k
X
= j fi dxj dx1 dx
ci dxk
i=1 j=1
k
!
X
i1
= (1) i fi dx1 dxk .
i=1
Thus, by the definition of the integral of a form on a smooth singular simplex,

Z k
X Z
i1
d = (1) i fi (x)dx.
i=1 k
258 4. Manifolds
To compute the ith integral on the right let us integrate with respect to
the ith variable first, and then with respect to the others. In the first step
compute the integral using the Newton-Leibniz rule. This gives us
Z
i fi (x)dx
k
Z 1(x1 ++
Z xbi ++xk )

= i fi (x1 , . . . , xk )dxi dx1 . . . dx
ci . . . dxk
(x1 ,...,xbi ,...,xk )k1 0
Z
= fi (x1 , . . . , 1 (x1 + + xbi + + xk ), . . . , xk )dx1 . . . dx
ci . . . dxk
(x1 ,...,xbi ,...,xk )k1
Z
fi (x1 , . . . , xi1 , 0, xi+1 , . . . , xk )dx1 . . . dx
ci . . . dxk .
(x1 ,...,xbi ,...,xk )k1
In the first integral of the last expression, make the substitution
x1 = 1 (y1 + + yk1 ) if i 6= 1,
xj = yj1 for 2 j k and j 6= i,
in the second integral make the substitution
(x1 , . . . , xbi , . . . , xk ) = (y1 , . . . , yk1 ).
The determinants of the derivative matrices are equal to 1 for both substi-
tutions, therefore
Z Z Z
i fi (x)dx = fi (l0k (y))dy fi (lik (y))dy,
k k1 k1
and
Z Z k
! k Z
X X
i1
d = (1) fi (l0k (y)) dy + (1)i fi (lik (y))dy.
k1 i=1 i=1 k1
Now let us compute the integrals of on the facets of . To avoid confusion

with coordinates on Rk and Rk1 , let (y 1 , . . . , y k1 ) be the identity chart on
Rk1 . Assume first that 1 i k. Since xi lik 0 and (lik ) (dx1
ci dxk ) = dy 1 dy k1 ,
dx
( lik ) = (lik ) = (fi lik )dy 1 dy k1

and Z Z
= fi (lik (y))dy.
lik k1
The pull-back of the form dx1 dx

ci dxk by lk is
0
(l0k ) (dx1 dx
ci dxk )
= d(1 y 1 y k1 ) dy 1 dy
\ i1 dy k1
= (1)i1 dy 1 dy k1 ,
consequently,
k
X
( l0k ) = (l0k ) = (1)i (fi l0k )dy 1 dy k1 ,
i=1
and !
Z Z k
X
i1 k
= (1) fi (l0 (y)) dy.
l0k k1 i=1
Adding the integrals of on the facets of with alternating signs we get
Z k
X Z
i
= (1)
i=0 lik
Z k
! k Z
X X
i1
= (1) fi (l0k (y)) dy + (1)i fi (lik (y))dy
k1 i=1 i=1 k1
Z
= d,

Corollary 4.7.9. If c R Zk (M ) is a cycle, k (M ) is a closed differential
form, then the integral c depends only on the homology class [c] Hk (M )
k
of c and the de Rham cohomology [] HdR (M ) of .
Proof. Choose other elements c0 = c + a and 0 = + d from the same
homology and cohomology class respectively. Then
Z Z Z Z Z Z Z Z
0 = + ( +d)+ d = + (d +d(d))+ = .
c0 c c c c c c c
Definition 4.7.10. The integral of the de Rham cohomology class []

k
HdR (M ) over the homology class [c] Hk (M ) is defined by the equation

Z Z
[] = .
[c] c
260 4. Manifolds
4.7.2 Integration on Regular Domains

If M is an n-dimensional oriented manifold, then we can integrate a dif-
ferential n-form on regular domains of M . Below we give details of this
integration.
Definition 4.7.11. An orientation of a 0-dimensional manifold is an assign-
ment of a sign + or to each of its points.
An orientation of a manifold M of dimension n 1 is a choice of an orienta-
tion of each tangent space Tp M of M in a continuous way. Continuity means
that for each point p M , there is a chart : U M around p, such that
the vectors (1 |q , . . . , n |q ) form a positively oriented basis of the tangent
space Tq M for all q U . A chart having this property is called a positive
chart with respect to the orientation. A manifold is orientable if it has an
orientation.
The Mobius band is probably the simplest example of an non-orientable man-
ifold. An orientation of a manifold of dimension 1 can also be given by an
atlas consisting of positive charts.
Definition 4.7.12. A subset D of a manifold M is called a regular domain
if D is closed and for each boundary point p D of D, there is a chart
: U Rn around p such that
(U D) = (U ) {(x1 , . . . , xn ) Rn | xn 0}.
Denote the half-space {(x1 , . . . , xn ) Rn | xn 0} by Rn+ . It is clear that

for the chart in the definition
(U D) = (U ) {(x1 , . . . , xn ) Rn | xn = 0},
Thus, if the coordinate functions of are (x1 , . . . , xn ), then (x1 , . . . , xn1 ) :

U D Rn1 is a chart on D. In particular, the boundary of a regular
domain in M is an (n 1)-dimensional smooth manifold.
A tangent vector v Tp M at a boundary p D is said to point inward
D or, respectively, outward D if the last last coordinate of v in the basis
1 |p , . . . , n |p positive or, respectively, negative, where is the chart form
the definition of regular domains. The notion does not depend on the choice
of the chart as we have the following equivalent characterization. The tangent
vector points inward D (or outward D) if and only if v is not tangent to the
boundary D and for any curve p representing the tangent vector v,
there is a positive > 0 such that ((, 0)) ext D and ((0, )) int D
(or conversely, ((, 0)) int D and ((0, )) ext D).
An orientation of M induces an orientation on the boundary D of any
regular domain as follows. Let p D, and choose a tangent vector v Tp M
pointing outward D. If D is 0-dimensional, dim M = 1, assign to p the sign

+ if and only if v is a positively oriented basis of Tp M . If dim M > 1, then
let the ordered basis (v1 , . . . , vk1 ) of Tp (M ) be positively oriented if and
only if (v, v1 , . . . , vk1 ) is a positively oriented basis of Tp M .
Consider now a regular domain D in an oriented manifold M . To prepare
integration of differential forms on D, we cover D by relatively small open
sets.
Call a smooth singular n-dimensional simplex : n M a positive reg-
ular n-simplex, if extends to an open neighborhood V n of n to a
diffeomorphism : V U the inverse of which is a positive chart of M .
If p int D, then choose a positive chart : U Rn around p, the domain of
which is contained in int D. Composing with a translation, we may assume
that (p) int n . Then composing with a central homothety with center
at p and with sufficiently large ratio, we may also assume that (U ) n .
Then denote by p : n int D M the positive regular n-simplex obtained
from the inverse of . Choose also for p an open neighborhood p Up U
such that (Up ) int n .
If p D, then choose a chart = (x1 , . . . , xn ) : U Rn around p as in
Definition 4.7.12. Restricting to the connected component of U , we may
assume that the vector fields (1 , . . . , n ) define bases with constant positive
or negative orientation on U . If dim M 2, we can make a positive
chart by replacing x1 with x1 . If dim M = 1, then let p denote the sign
corresponding to the orientation of the chart . Observe, that n intersects
the boundary hyperplane of the half-space Rn+ in its nth facet n1 {0}.
Composing with a translation we may assume that (p) is a point of the
relative interior of the nth facet. Then composing with a homothety with
sufficiently large ratio and center at (p) we may assume that (U ) n .
For n 2 let p be the positive regular n-simplex in M obtained from the
inverse of . For n = 1, let p be the chain equal to p times the regular
1-simplex obtained from the inverse of . Choose also an open neighborhood
Up U of p such that (Up D) n and (Up D) intersects only the
nth face of n .
Definition 4.7.13. The support supp of a differential k-form k (M )

is the closure of the set {p M | (p) 6= 0}.
Assume that is a differential n-form on the oriented manifold M the support

of which intersects the regular domain D in a compact set K = supp D.
K is covered by the open subsets Up (p D) constructed above, therefore it is
also covered by a finite number of them. Let Ui = Upi (i = 1, . . . , s) be a finite
subcovering of K and set U0 = M \ K. Denote by i the regular n-simplex
pi . Recall that i is carrying a sign i in the 1-dimensional case, otherwise
it is positive. For the sake of uniform handling of different dimensions, we
262 4. Manifolds
can set i = +1 in dimensions not less than 2. The open sets U0 , U1 , . . . , Us

cover M , therefore, there is a smooth partition of the unity h0 , h1 , . . . , hs
subordinated to this covering. This means Psthat the hi s are smooth function
on M , hi : M [0, 1], supp
R h i Ui and i=0 hi 1.
We are going to define D as
Z Xs Z
= hi . (4.7)
D i=1 i i
However, to justify this definition we have to show that the expression on

the right-hand side does not depend on the many arbitrary choices of the
process. Assume that we repeat the same process but we choose the open
sets Up and the regular simplices p in another way choose another finite
subcovering for K and another partition of unity. Denote all the objects we
choose the second time by the same letters as in the first time, but decorated
with a tilde.
The key lemma is the following.
Lemma 4.7.14. Assume that n (M ) and i , j are regular n simplices
from the above processes. Then
Z Z
hi hj = hi hj .
i i j j
Proof. The regular n-dimensional simplices i and j were obtained as the

inverses of some charts i : U Rn and i : U Rn . Let H = 1 : (U
U ) (U U ) be the transit map between the two charts.
Since supp hi Ui U and supp hj Uj U , supp(hi hj ) Ui Uj
U U . Assume that
i (hi hj ) = f dx1 dxn j (hi hj ) = f dx1 dxn .
Then Z Z Z
hi hj = f (x)dx = f (x)dx
i n (Ui Uj ) Rn
+ (Ui Uj )
and similarly, Z Z
hi hj = f(x)dx
j Rn
+ (Ui Uj )
The transit map H defines a diffeomorphism Rn+ (Ui Uj ) Rn+ (Ui

Uj ), thus, we can compute the integral of f making substitution x = H(u).
This gives
Z Z Z
hi hj = f (x)dx = f (H(u))| det H 0 (u)|du.
i Rn
+ (Ui Uj ) Rn
+ (Ui Uj )
(4.8)
On the other hand, since j = i H.
f dx1 dxn = j (hi hj ) = H (i (hi hj )) = H (f dx1 dxn ).
Denote by (H 1 , . . . , H n ) the coordinate functions H i = xi H of H. Then

H (f dx1 dxn ) = (f H) dH 1 dH n .
As dH i = 1 H i dx1 + + n H i dxn ,
H (f dx1 dxn ) = (f H) (det H 0 ) dx1 dxn ,
which yields f = (f H) (det H 0 ), and
Z
1 n
j (hi hj ) = f dx dx = f (H(x))(det H 0 (x))dx. (4.9)
Rn
+ (Ui Uj )
Comparing equations (4.8) and (4.9) we see that the equation we want to
show holds up to sign. To have the equality without any sign corrections,
it remains to show that i j = sgn det H 0 . If dim M 2, then i = j = 1,
furthermore, and are positive charts, so the derivative of the transit map
between them preserves orientation, that is, it has positive determinant.
For dim M = 1, the derivative of the transit map has positive determinant if
and only if either both and are positive or none of them is positive. This
completes the proof.
Proposition 4.7.15. For any two different choices of the simplices and the
partition of unity
Xs Z Xs Z
hi = hj .
i=1 i i j=1 j j
Proof. The supports of h0 and h0 are in (M \ K) supp ext D, thus,

the integrals of the forms h0 hj and h0 hi on any of the regular n-simplices
k and l are equal to 0. Consequently,

s Z
X s Z
X Xs
hi = hi hj
i=1 i i i=1 i i j=0
s X
X s Z s X
X s Z
= hi hj = hi hj
i=1 j=1 i i i=1 j=1 j j
s Z s
! s Z
X X X
= hi hj = hj .
j=1 j j i=0 j=1 j j
264 4. Manifolds
Definition 4.7.16. Using the notations introduced above, we define the

integral of over the regular domain D by (4.7).
Corollary 4.7.17. Denote by n the union of the interior of n and the
relative interior of its nth facet. This set is what remains if we remove from
n all its facets except for the nth one. R
If supp D is covered by i (n ), then D = i i .
R
Proof. If i (n ) covers K = supp D, then Ui can be chosen so that Ui

covers K alone, so there will be no need for other simplices to compute the
integral. Then hi will be constant 1 on K, so (hi )|D = |D.
Theorem 4.7.18 (Stokes Theorem for Regular Domains). If D is a regular
domain in an oriented n-dimensional manifold M , n1 (M ) is a differ-
ential (n 1)-form such that supp D is compact, and D is endowed with
the induced orientation, then
Z Z
d = .
D D
Ps Ps
Proof. Since i=0 hi 1, i=0 dhi 0, and
s s s
! s
X X X X
d = hi d = d(hi ) dhi = d(hi ).
i=0 i=0 i=0 i=0
The support of d(h0 ) is disjoint from D, so integrating over D we obtain

Z s Z
X s Z
X
d = d(hi ) = d(hi ).
D i=1 D i=1 i i
Applying the Stokes theorem for chains, and the fact that supp(hi ) inter-
sects only the image of the nth facet of i ,
Z s Z
X s Z
X s Z
X
d = d(hi ) = d(hi ) = hi
D i=1 D i=1 i i i=1 i i
Xs Z
n
= i (1) hi .
n
i ln
i=1
The system of regular (n1)-simplices i lnn the relative open subsets Ui D

R D and the partition of unity hi |D form a system to
of compute the integral
D
, however the orientation of the simplices i lnn is not positive. The
orientation of it is given exactly by the sign i (1)n . Indeed, the derivative
of i at the origin takes the standard basis of Rn to the tangent vectors
(v1 , . . . , vn ) at i (0). The sign of the orientation of this basis relative to

the orientation of M is i . The first (n 1) vectors (v1 , . . . , vn1 ) are the
images of the standard basis vectors of Rn1 under the derivative of i lnn
at the origin, vn is pointing inward D. Thus, the sign of the orientation of
(v1 , . . . , vn1 ) relative to the orientation of the boundary is the sign of the
orientation of (vn , v1 , . . . , vn1 ) which is exactly i (1)n . In conclusion we
obtain
Xs Z Z
n
i (1) hi = ,
n
i ln D
i=1
4.7.3 Integration on Riemannian Manifolds

Definition 4.7.19. A Riemannian manifold is a smooth manifold M en-
dowed with a smooth tensor field g of type (2, 0) for which g(p) is a positive
definite symmetric bilinear function on Tp M for each p M . The tensor
field g is called the (Riemannian) metric on M .
Example. A hypersurface of Rn together with the first fundamental form is
a Riemannian manifold.
The bilinear function g(p) is often denoted by h , ip , and if X and Y are two
vector fields on M the function g(X, Y ) is also denoted by hX, Y i.
If M is an n-dimensional oriented Riemannian manifold, then there is a
unique differential n-form on it, called its volume form. The volume form
is defined in the following way.
Let p M be an arbitrary point and choose a positively oriented orthonor-
mal basis e1 , . . . , en of Tp M . Let e1 , . . . , en be the dual basis of the basis
e1 , . . . , en and set (p) = e1 en . The form (p) does not depend
on the choice of the orthonormal basis. Indeed, if f1 , . . . , fn is another pos-
itively oriented orthonormal basis of Tp M with dual basis f 1 , . . . , f n . Then
e1 en and f 1 f n both have unit length with respect to the in-
duced Euclidean structure of the 1-dimensional linear space n (Tp M ) , (see
Proposition 1.2.67), and they must have the same orientation, hence they
ought to be equal.
Denote by Cc (M ) the linear space of continuous functions with compact sup-
port on M . Consider the functional I : Cc (M ) R defined by
Z
I(f ) = f .
M
The right-hand side makes sense, because the integral of compactly supported
forms can be defined in the same manner as we did also in the case when the
266 4. Manifolds
differential form is continuous but not necessarily smooth. The functional I

is a positive linear functional on Cc (M ), which means that it is linear and
assigns positive numbers to nonnegative functions not identically equal to 0.
By the Riesz representation theorem from measure theory, (see Theorem 7.2.8
in [5]), any such positive linear functional
R corresponds to a unique regular
Borel measure on M , for which I(f ) = M f d. The measure obtained this
way is the volume measure of the oriented Riemannian manifold M .
Though the volume form cannot be defined for non-orientable Riemannian
manifolds, there is a volume measure even in that case.
For this purpose, we construct an orientable double covering : M M of
M . M as a set consists of all pairs (p, o), where p is a point of M , o is an
orientation of Tp M . The covering map : M M assigns to the pair (p, o)
the point p. Since a linear space has exactly two orientations, each point
p in M has exactly two preimages. We define a manifold structure on M
with the help of an atlas. Choose a chart : U Rn from the atlas A of
M . We can split 1 (U ) into two disjoint pieces U+ and U as follows. Let
U+ contain those pairs (p, o), for which p U and the basis (1 |p , . . . , n |p )
is positively oriented with respect to o, and let U = 1 (U ) \ U+ be the
remaining subset. The charts |U : U Rn ( A) give an atlas on
M turning it into a differentiable manifold. M has a unique orientation for
which all the charts |U+ are positive. The restriction of onto U is a
diffeomorphism between U and U , in particular, is a covering map. M
can be endowed by the Riemannian metric g, obtained as the pull-back of
the Riemannian metric of M by the covering map .
As M can be oriented, we can produce a volume form and a volume measure
on it as above. Then a subset A of M will be measurable with respect to
the volume measure of M if and only if 1 (A) is -measurable, and if this
is fulfilled, then we set (A) = (1 (A))/2.
4.8 Differentiation of Vector Fields

Although there is a natural way to differentiate a smooth function defined
on a manifold with respect to a tangent vector, there is no analogous natural
way to differentiate vector fields. In fact, there are infinitely many ways to
differentiate vector fields with respect to a tangent vector, and to choose
one of them, (the most appropriate one in some sense), the differentiable
manifold structure alone is not enough. A fixed rule for the differentiation
of vector fields is itself an additional structure on the manifold, called an
affine connection. Later we shall see that on Riemannian manifolds, i.e., on
manifolds the tangent spaces of which are equipped with a dot product we can
introduce a special differentiation of vector fields in a natural way. A precise
4.8. Differentiation of Vector Fields 267
formulation of this statement is the Fundamental Theorem of Riemannian

Geometry.
As far as only vector fields on an open domain of Rn are considered, the rule
Y X = (X )0 (0) (4.10)
seems to be quite natural for the derivation of a vector field X over an

open subset U Rn with respect to the tangent vector Y Tp Rn , where
: [, ] U is any smooth curve that satisfies (0) = p and 0 (0) = Y .
In this formula, using the natural isomorphisms Tp Rn
= Rn , X is thought of
n n
as a map X : R R . This interpretation of X enables us to compute the
derivative of the vector valued function X at 0.
We see that
n
X
Y X = Y (X i )i (p), (4.11)
i=1
where i denotes the i-th coordinate vector field on Rn , X i are the compo-
nents of the vector field X with respect to i . In particular, the value of Y X
does not depend on the choice of .
It is easy to check that this differentiation of vector fields has the properties.
(1) Y1 +Y2 X = Y1 X + Y2 X,
(2) cY X = cY X,
(3) Y (X1 + X2 ) = Y X1 + Y X2 ,
(4) Y (f X) = Y (f )X + f (p)Y X,
(5) X1 X2 X2 X1 = [X1 , X2 ],
(6) Y hX1 , X2 i = hY X1 , X2 i + hX1 , Y X2 i,
where X1 , X2 X(R ), Y Tp Rn , f F(Rn ), c R.
n
Now consider the general case. Let M be a smooth manifold. As we men-

tioned, there is no natural rule for derivation of vector fields on M , so we
introduce such rules axiomatically, as operations satisfying some of the above
properties.
Definition 4.8.1. An affine connection at p M is a mapping which

assigns to each tangent vector Y Tp M and each vector field X X(M ) a
new tangent vector Y X Tp M called the covariant derivative of X with
respect to Y and satisfies the following identities
268 4. Manifolds
(i) Y1 +Y2 X = Y1 X + Y2 X,
(ii) cY X = cY X,
(iii) Y (X1 + X2 ) = Y X1 + Y X2 ,
(iv) Y (f X) = Y (f )X + f (p)Y X,
where X1 , X2 X(M ), Y, Y1 , Y2 Tp M , f F(M ), c R.
Definition 4.8.2. A global affine connection (or briefly a connection) on

M is a mapping which assigns to two smooth vector fields Y and X a new
one Y X called the covariant derivative of the vector field X with respect to
the vector field Y , satisfying the identities (i), (ii), (iii), and
(iv) Y (f X) = Y (f )X + f Y X,
for all X1 , X2 , Y, Y1 , Y2 X(M ), and f, c F(M ).
Lemma 4.8.3. For a global affine connection , X, Y X(M ), p M , the

tangent vector (Y X)(p) depends only on the behavior of X and Y in an
open neighborhood of p.
Proof. Let us suppose that the vector fields X1 and X2 coincide on an open
neighborhood U of p. Choose a smooth function h F(M ) which is zero
outside U and constant 1 on a neighborhood of p. Then we have h(X1 X2 ) =
0, consequently
0 = Y (h(X1 X2 )) = hY (X1 X2 ) + Y (h)(X1 X2 ).
Computing the right-hand side at p we get
0 = (Y X1 )(p) (Y X2 )(p).
Similarly, if the vector fields Y1 and Y2 coincide on an open neighborhood U

of p and h is chosen as above, then we have
0 = h(Y1 Y2 ) X = hY1 X hY2 X,
which yields 0 = Y1 X(p) Y2 X(p).
The lemma implies that an affine connection can be restricted onto any open
subset and can be recovered from its restrictions onto the elements of an open
cover.
Fix a local coordinate system = (x1 , . . . , xn ) on an open subset U of M
and let 1 = 1 ,. . . ,n = n be the basis vector fields on U induced by .
(As we work with a fixed chart , omission of from the notation should not
lead to confusion.) Given an affine connection on U , we can express the

vector field i j as a linear combination of the basis vector fields
n
X
i j = kij k .
i=1
The components kij are smooth functions on U , called the Christoffel symbols
of the covariant derivation with respect to the given chart.
Proposition 4.8.4. The restriction of a global affine connection onto an
open coordinate neighborhood U is uniquely determined by the Christoffel
symbols. Any n3 smooth functions kij on U are the Christoffel symbols of an
appropriate affine connection on U .
Pn Pn
Proof. Let X = i=1 X i i , Y = j=1 Y j j be two smooth vector fields on
U . Then by the properties of affine connections, Y X can be computed as
follows
n
! n n
!
X X X
i j i
Y X = ( n Y j j )
P X i = Y j X i
j=1
i=1 j=1 i=1
n X
X n
Y j j (X i )i + X i j i

=
i=1 j=1
n
n X n
!
X X
j i i
= Y j (X )i + X kji k
i=1 j=1 k=1

n
X n X
X n
= Y (X k ) + X i Y j kji k .
k=1 i=1 j=1
This formula shows that the knowledge of the Christoffel symbols enables us
to compute the covariant derivative of any vector field with respect to any
other one. On the other hand, if kij are an arbitrary smooth functions on U
for 1 i, j, k n, then defining the covariant derivative of a vector field by
the above formula, we obtain an affine connection on U .
Note that the connection defined by (4.11) on Rn has vanishing Christoffel-

symbols with respect to the identity chart of Rn .
Observe also, that in fact, the tangent vector (Y X)(p) depends only on
the vector Y (p), so a global affine connection on a manifold defines an affine
connection at each of its points. Furthermore, we do not need to know the
vector field X everywhere on U to compute (Y X)(p). It is enough to know
X at the points of a curve : [, ] M for which (0) = p, 0 (0) = Y (p).
270 4. Manifolds
Definition 4.8.5. Let : [a, b] M be a smooth curve in M . A smooth

vector field X along the curve is a smooth mapping X : [a, b] T M which
assigns to each t [a, b] a tangent vector X(t) T(t) M .
Now suppose that M is provided with a global affine connection . Then

any vector field X along determines a new vector field t X along called
the covariant derivative of X with respect to the curve parameter t as follows.
For [a, b], choose a local coordinate system around ( ). Let kij be
the Christoffel symbols of with respect to decompose X and 0 as
n
X n
X
X(t) = X i (t)i ((t)) and 0 (t) = Y i (t)i ((t)).
i=1 i=1
Then t X( ) is given by

n n X
n
X 0 X
t X( ) := X k ( ) + X i ( )Y j ( )kji (( )) k (( )).
k=1 i=1 j=1
Remark. Several different notations are used for the covariant derivative of
a vector field X along a curve : t X, X 0
dt , X , X. The most concise
0
0
notation X does not show the dependence of this operation on . Neverthe-
less, it is convenient to use when we work with a fixed connection on M .
Notations t X and X dt can be used only when the parameter of the curve
is denoted by a fixed symbol, in our case by t. Of course, if another symbol
is used for the parameter, then t should be replaced by that. The notation
0 X is motivated by the fact that if X is a vector field defined in an open
neighborhood of ( ) which extends X in the sense that X((t))) = X(t) for
all t close enough to , then t X( ) = 0 ( ) X. In the rest of this book we
shall mainly use the notations t X and 0 X.
Definition 4.8.6. A vector field X along a curve is said to be a parallel

vector field if the covariant derivative 0 X is identically zero.
Proposition 4.8.7. Given a curve and a tangent vector X0 at the point

(0), there is a unique parallel vector field X along which extends X0 .
Proof. The proposition follows from results on ordinary differential equations.

Using local coordinates, condition 0 X = 0 yields a system of ordinary
differential equations for the components of X
n X
n
0 X
Xk + X i Y j kji = 0.
i=1 j=1
Since these equations are linear, the existence and uniqueness theorem for
linear differential equations guaranties that the solutions of this system of
differential equations are uniquely determined by the initial values X k (0)
and can be defined for all values of t.
If X is a parallel vector field along the curve , then the vector X((t)) is
said to be obtained from X0 by parallel transport along .
Definition 4.8.8. A connection is called symmetric or torsion free if it sat-
isfies the identity

X Y Y X = [X, Y ].
Applying this identity to the case X = i , Y = j , from [i , j ] = 0 one
obtains the relation
kij = kji
for the Christoffel symbols of a torsion free connection. Conversely, if kij =
kji then using the expression of covariant derivative with the help of Christof-
fel symbols we get
X Y Y X

Xn n X
X n
= X(Y k ) + Y i X j kji k
k=1 i=1 j=1

n
X n X
X n
Y (X k ) + X i Y j kji k
k=1 i=1 j=1
n
X
X(Y k ) Y (X k ) k = [X, Y ].

=
i=1
There is a useful characterization of symmetry. Consider a parameterized

surface in M that is a smooth mapping s : R M from a rectangular domain
R of the plane R2 into M .
A vector field X along s is a mapping which assigns to each (x, y) R a
tangent vector X(x, y) Ts(x,y) M .
As examples, the two standard vector fields x and y on the plane give rise
to vector fields T s(x ) and T s(y ) along s. These will be denoted briefly by
x s and y s. Here T s : T R2 T M denotes the derivative map of s.
For any smooth vector field X along s the covariant partial derivatives x X
and y X are new vector fields along s constructed as follows. For each fixed
y0 , the map x 7 X(x, y0 ) is a vector field along the curve x 7 s(x, y0 ). The
vector x X(x0 , y0 ) is defined to be its covariant derivative with respect to x
at x0 . This defines x X along the entire parameterized surface s. y X is
defined similarly.
272 4. Manifolds
Proposition 4.8.9. A connection is symmetric if and only if
x (y s) = y (x s)
for any parameterized surface s in M .
Proof. Let us choose a local coordinate system = (x1 , . . . , xn ) on M around

a point s(x0 , y0 ). The mapping s is given locally by the functions si = xi s.
The vector field y s has the form
n
X
y s = y si (i s).
i=1
The partial covariant derivative of this vector field with respect to x is equal
to
n
!
X
i
x (y s) = x y s (i s)
i=1
n
X
= x (y (si )) (i s) + y si x (i s)
i=1

n
X n
X X n
n X
= x (y (si )) (i s) + y si x sj (kji s) k s.
i=1 k=1 i=1 j=1
This formula shows that interchanging the role of x and y we obtain the same
vector field for any s if and only if kij = kji .
Roughly speaking, the torsion free condition halves the degree of freedom
in the choice of Christoffel symbols, a symmetric connection is uniquely de-
2
termined by n (n+1)
2 arbitrarily chosen functions, nevertheless, the space of
symmetric affine connections on a manifold is still infinite dimensional. We
can reduce the degree of freedom further on Riemannian manifolds.
Definition 4.8.10. A connection on a Riemannian manifold M is said to

be compatible with the Riemannian metric if parallel transport along curves
preserves inner products. In other words, for any curve and any pair X, Y
of parallel vector fields along , the inner product hX, Y i is constant.
Proposition 4.8.11. Suppose that the connection is compatible with the met-
ric. Let V, W be any two vector fields along . Then
hV, W i0 = ht V, W i + hV, t W i.
Proof. Choose parallel vector fields X1 , . . . , Xn along which are orthonor-

mal at one point of and hence Pn at every point Pnof . Then the given fields
V and W can be expressed as i=1 vi Xi and i=1 wi Xi respectively Pn(where
vi = hV, Xi i is a real valued function). It follows that hV, W i = i=1 vi wi
and that
Xn Xn
t V = vi0 Xi , t W = wi0 Xi .
i=1 i=1
Therefore
n
X
ht V, W i + hV, t W i = (vi0 wi + vi wi0 ) = hV, W i0 .
i=1
Corollary 4.8.12. An affine connection on a Riemannian manifold is com-

patible with the metric if and only if for any vector fields X1 , X2 on M and
any tangent vector Y Tp M we have
Y (hX1 , X2 i) = hY X1 , X2 i + hX1 , Y X2 i.
Theorem 4.8.13 (Fundamental Theorem of Riemannian Geometry). A Rie-

mannian manifold possesses one and only one symmetric connection which
is compatible with its metric.
Proof. Applying the compatibility condition to the basis vector fields i , j , k
induced by a fixed chart on the manifold and setting hj , k i = gjk one ob-
tains the identity
i gjk = hi j , k i + hj , i k i.
Permuting i, j and k this gives three linear equations relating the three
quantities
hi j , k i, hj k , i i, hk i , j i.
(There are only three such quantities since i j = j i .) These equations
can be solved uniquely; yielding the first Christoffel identity
1
hi j , k i = (i gjk + j gik k gij ).
2
Pn
The left-hand side of this identity is equal to l=1 lij glk . Multiplying by
the inverse (g kl ) of the matrix (glk ) this yields the second Christoffel identity
n
X 1
kij = (i gjl + j gil l gij )g lk .
2
l=1
Thus the connection is uniquely determined by the metric.

274 4. Manifolds
Conversely, defining kij by this formula, one can verify that the resulting
connection is symmetric and compatible with the metric. This completes the
proof.
Definition 4.8.14. The unique symmetric affine connection which is com-

patible with the metric on a Riemannian manifold is called the Levi-Civita
connection.
Since the formula for the Christoffel symbols of the Levi-Civita connection
involves elements of the inverse of the matrix of the metric, it is not a conve-
nient formula to express X Y in a coordinate free way. However, there is an
explicit coordinate free expression of hX Y, Zi for an arbitrary vector field
Z. To obtain this formula, start with the equation
XhY, Zi = hX Y, Zi + hY, X Zi = hX Y, Zi + hY, Z X + [X, Z]i.
Rearranging and permuting the role of X, Y and Z cyclically, we obtain the

following system of equations
XhY, Zi hY, [X, Z]i = hX Y, Zi + hZ X, Y i,

Y hZ, Xi hZ, [Y, X]i = hY Z, Xi + hX Y, Zi,
ZhX, Y i hX, [Z, Y ]i = hZ X, Y i + hY Z, Xi.
Solving this system for the unknown quantities hX Y, Zi, hY Z, Xi and

hZ X, Y i we obtain Koszuls formula
1
hX Y, Zi = XhY, Zi hY, [X, Z]i + Y hZ, Xi
2 (4.12)
hZ, [Y, X]i ZhX, Y i + hX, [Z, Y ]i .
The connection defined by formula (4.11) on open subsets of Rn is just the

Levi-Civita connection of Rn .
Consider now a parameterized hypersurface r : Rn . It is a Riemannian
manifold with the first fundamental form as metric. The basis vector fields
ri through suitable identifications are the same as the basis vector fields i
corresponding to the chart = r1 . Comparing the formulae
n1
X
ri rj = rij = kij rk + bij N (4.13)
k=1
and
n1
X 1
kij = (gjl,i + gil,j gij,l )g lk .
2
l=1
4.9. Curvature 275
proved in Section 3.1.5 with the formulae derived for the Levi-Civita connec-
tion we can conclude that the Christoffel symbols of a hypersurface intro-
duced previously are the Christoffel symbols of the Levi-Civita connection
of the hypersurface. Therefore, denoting by the Levi-Civita connection of
the hypersurface, we can rewrite equation (4.13) as
ri rj = ri rj + II(ri , rj )N.
This formula extends directly to tangential vector fields X, Y , as follows
Y X = Y X + II(X, Y )N = Y X + hY X, NiN. (4.14)
Exercise 4.8.15. Show that if and are two affine connections on a man-
ifold M , then their difference S(X, Y ) = X Y X Y is an F(M )-bilinear
mapping. (In other words, S is a tensor field of type (2, 1)). Conversely, the
sum of a connection and an F(M )-bilinear mapping S : X(M ) X(M )
X(M ) is a connection. (According to these statements, affine global connec-
tions form an affine space over the linear space of tensor fields of type (2, 1).)
Exercise 4.8.16. Show that the torsion T : X(M ) X(M ) X(M ) of a

connection defined by
T (X, Y ) = X Y Y X [X, Y ]
is an F(M )-bilinear mapping (i.e., T is a tensor field).
Exercise 4.8.17. Show that if T is the torsion of an affine connection

then T /2 is a symmetric connection.
Exercise 4.8.18. Check that the connection defined by the Christoffel sym-
bols
n1
1X
kij = (gli,j + glj,i gij,l )g lk
2
l=1
is symmetric and compatible with the metric.
4.9 Curvature
If is an affine connection on a manifold M , then we may consider the
operator
R(X, Y ) = [X , Y ] [X,Y ] : X(M ) X(M ),
276 4. Manifolds
where [X , Y ] = X Y Y X is the usual commutator of opera-

tors. The operator R(X, Y ), depending on X and Y , is called the curvature
operator of the connection. The assignment
X(M ) X(M ) X(M ) X(M )

(X, Y, Z) 7 R(X, Y )(Z)
is called the curvature tensor of the connection. To reduce the number of

brackets, we shall denote R(X, Y )(Z) simply by R(X, Y ; Z). Thus, the letter
R is used in two different meanings, later it will denote also a third mapping,
but the number of arguments of R makes always clear which meaning is
considered.
Proposition 4.9.1. The curvature tensor is linear over the ring of smooth
functions in each of its arguments, and it is skew symmetric in the first two
arguments.
Proof. Skew symmetry in the first two arguments is clear, since
R(X, Y ) = [X , Y ] [X,Y ] = [Y , X ] + [Y,X] = R(Y, X).
According to this, it suffices to check linearity of the curvature tensor in the

first and third arguments.
Linearity in the first argument is proved by the following identities.
R(X1 + X2 , Y ) = [X1 +X2 , Y ] [X1 +X2 ,Y ]

= [X1 + X2 , Y ] [X1 ,Y ]+[X2 ,Y ]
= [X1 , Y ] + [X2 , Y ] [X1 ,Y ] [X2 ,Y ]
= R(X1 , Y ) + R(X2 , Y ).
and
R(f X, Y ; Z) = ([f X , Y ] [f X,Y ] )(Z)

= f X Y Z Y (f X Z) f [X,Y ]Y (f )X (Z)
= f X Y Z f Y X Z Y (f )X Z f [X,Y ] Z + Y (f )X (Z)
= f (X Y Z Y X Z [X,Y ] Z) = f R(X, Y ; Z).
Additivity in the third argument is clear, since R(X, Y ) is built up of the

additive operators X , Y and their compositions. To have linearity, we
4.9. Curvature 277
need
R(X, Y ; f Z) = X Y (f Z) Y X (f Z) [X,Y ] (f Z)
= X (Y (f )Z + f Y Z) Y (X(f )Z + f X Z)
[X, Y ](f )Z f [X,Y ] Z
= XY (f )Z + Y (f )X Z + X(f )Y Z + f X Y Z
Y X(f )Z X(f )Y Z Y (f )X Z f Y X Z
XY (f )Z + Y X(f )Z f [X,Y ] Z
= f (X Y Z Y X Z [X,Y ] Z) = f R(X, Y ; Z).
Proposition 4.9.1 is interesting, because the curvature tensor is built up from
covariant derivations, which are not linear operators over the ring of smooth
functions.
We have already introduced tensor fields over a hypersurface. We can in-
troduce tensor fields over a manifold in the same manner. A tensor field T
of type (k, l) is an assignment to every point p of a manifold M a tensor
T (p) of type (k, l) over the tangent space Tp M . If 1 , . . . , n are the basis
vector fields defined by a chart over the domain of the chart, and we denote
by dx1 (p), . . . , dxn (p) the dual basis of 1 (p), . . . , n (p), then a tensor field is
uniquely determined over the domain of the chart by the components
Tji11...j
...il
k
(p) = T (p)(dxi1 , . . . , dxil ; j1 , . . . , jk ).
We say that the tensor field is smooth, if for any chart from the atlas of M ,
the functions Tji11...j
...il
k
are smooth. We shall consider only smooth tensor fields.
Tensor fields of type (0, 1) are the vector fields, tensor fields of type (1, 0) are
the differential 1-forms. Thus, a differential 1-form assigns to every point of
the manifold a linear function on the tangent space at that point. Differential
1-forms form a module over the ring of smooth functions, which we denote
by 1 (M ).
Every tensor field of type (k, l) defines an F(M )-multilinear mapping
1 (M ) 1 (M ) X(M ) X(M ) F(M )
| {z } | {z }
l times k times
and conversely, every such F(M )-multilinear mapping comes from a tensor
field. (Check this!) Therefore, tensor fields can be identified with F(M )-
multilinear mappings 1 (M ) 1 (M ) X(M ) X(M ) F(M ).
Tensor fields of type (k, 1), that is F(M )-multilinear mappings
1 (M ) X(M ) X(M ) F(M )
can be identified in a natural way with F(M )-multilinear mappings
X(M ) X(M ) X(M ).
278 4. Manifolds
By this identification, R : X(M ) X(M ) X(M ) corresponds to

R : 1 (M ) X(M ) X(M ) F(M ), defined by R(; X1 , . . . , Xk ) =
(R(X1 , . . . , Xk )).
Using these identifications, the curvature tensor is a tensor field of type (3, 1)
by Proposition 4.9.1. It is a remarkable consequence, that although the vec-
tors X Z(p) and Y Z(p) are not determined by the vectors X(p), Y (p), Z(p),
to compute the value of R(X, Y ; Z) at p it suffices to know X(p), Y (p), Z(p).
Beside skew-symmetry in the first two arguments, the curvature tensor has
many other symmetry properties.
Theorem 4.9.2 (First Bianchi Identity). If R is the curvature tensor of a
torsion free connection, then
R(X, Y ; Z) + R(Y, Z; X) + R(Z, X; Y ) = 0
for any three vector fields X, Y, Z.
Proof. Let us introduce the following notation.
P If F (X, Y,PZ) is a function of
the variables X, Y, Z, then denote by F (X, Y, Z) or XY Z F (X, Y, Z)
the sum of the values of F at all cyclic permutations of the variables (X, Y, Z)
X
F (X, Y, Z) = F (X, Y, Z) + F (Y, Z, X) + F (Z, X, Y ).

P
We shall use several times that behind the cyclic summation we may
cyclically rotate X, Y, Z in any expression
X X X
F (X, Y, Z) = F (Y, Z, X) = F (Z, X, Y ).

The theorem claims vanishing of

X
R(X,Y ; Z)

X
= (X Y Z Y X Z [X,Y ] Z)

X
= (X Y Z X Z Y [X,Y ] Z)

X
= (X [Y, Z] [X,Y ] Z)

X X
= (Z [X, Y ] [X,Y ] Z) = [Z, [X, Y ]],

but the last expression is 0 according to the Jacobi identity on the Lie bracket
of vector fields. (At the third and fifth equality we used the torsion free
property of .)
4.9. Curvature 279
The presence of an affine connection on a manifold allows us to differentiate

not only vector fields, but also tensor fields of any type.
Definition 4.9.3. Let (M, ) be a manifold with an affine connection. If

1 (M ) is a differential 1-form, X is a vector field, then we define the
covariant derivative X of with respect to X to be the 1-form
(X )(Y ) = X((Y )) (X Y ), Y X(M ).
In general, the covariant derivative X T of a tensor field
T : 1 (M ) 1 (M ) X(M ) X(M ) F(M )
of type (k, l) with respect to a vector field X is a tensor field of the same
type, defined by
(X T )(1 , . . . , l ; X1 , . . . , Xk ) = X(T (1 , . . . , l ; X1 , . . . , Xk ))
k
X
T (1 , . . . , X i , . . . , l ; X1 , . . . , Xk )
i=1
l
X
T (1 , . . . , l ; X1 , . . . , X Xj , . . . , Xk ).
j=1
For the case of the curvature tensor, this definition gives
(X R)(Y, Z; W ) = X (R(Y, Z; W )) R(X Y, Z; W )

R(Y, X Z; W ) R(Y, Z; X W ).
Theorem 4.9.4 (Second Bianchi Identity). The curvature tensor of a torsion

free connection satisfies
X
(X R)(Y, Z; W ) = 0.
XY Z
Proof. (X R)(Y, Z; W ) is the value of the operator
X R(Y, Z) R(X Y, Z) R(Y, X Z) R(Y, Z) X : X(M ) X(M )
on the vector field W , hence we have to prove vanishing of the operator

X
X R(Y, Z) R(X Y, Z) R(Y, X Z) R(Y, Z) X . (4.15)
XY Z
280 4. Manifolds
First, we have
X
X R(Y, Z) R(Y, Z) X
XY Z
X
= (X Y Z X Z Y X [Y,Z] )
XY Z

(Y Z X Z Y X [Y,Z] X )
X
= (X Y Z X Z Y X [Y,Z] )
XY Z

(X Y Z X Z Y [Y,Z] X )
X
= [Y,Z] X X [Y,Z] .
XY Z
On the other hand,

X
R(X Y, Z) R(Y, X Z)
XY Z
X
= R(X Z, Y ) R(X Y, Z)
XY Z
X
= R(Y X, Z) R(X Y, Z)
XY Z
X X
= R(Y X X Y, Z) = R([Y, X], Z).
XY Z XY Z
Combining these results, operator (4.15) equals

X
[Y,Z] X X [Y,Z] + R([Y, X], Z)
XY Z
X
= [Y,Z] X X [Y,Z] + R[Z, Y ], X)
XY Z
X
= [Y,Z] X X [Y,Z] + [Z,Y ] X X [Z,Y ] [[Z,Y ],X]
XY Z
X
= [[Y,Z],X] = P [[X,Y ],Z] = 0.
XY Z
Curvature of Riemannian Manifolds

In the remaining part of this unit, we shall deal with Riemannian manifolds,
therefore from now on assume that (M, h , i) is a Riemannian manifold with
Levi-Civita connection and R is the curvature tensor of .
4.9. Curvature 281
Introduce the tensor R of type (4, 0), related to R by the equation
R(X, Y ; Z, W ) = hR(X, Y ; Z), W i,
which is called the RiemannChristoffel curvature tensor of the Riemannian

manifold. To simplify notation, we shall denote R also by R. This will
not lead to confusion, since the RiemannChristoffel tensor and the ordinary
curvature tensor have different number of arguments.
Levi-Civita connections are connections of special type, so it is not surprising,
that the curvature tensor of a Riemannian manifold has stronger symmetries
than that of an arbitrary connection. Of course, the general results can be
applied to Riemannian manifolds as well yielding
X
R(X, Y ; Z, W ) = R(Y, X; Z, W ) and R(X, Y ; Z, W ) = 0.
XY Z
In addition to these symmetries, we have the following ones.

Theorem 4.9.5. The RiemannChristoffel curvature tensor is skew-symmetric
in the last two arguments
R(X, Y ; Z, W ) = R(X, Y ; W, Z).
Proof. By the compatibility of the connection with the metric, we have
X(Y (hZ, W i)) = X(hY Z, W i + hZ, Y W i)

= hX Y Z, W i + hY Z, X W i
+ hX Z, Y W i + hZ, X Y W i,
and similarly,
Y (X(hZ, W i)) = hY X Z, W i
+ hX Z, Y W i + hY Z, X W i + hZ, Y X W i.
We also have
[X, Y ](hZ, W i) = h[X,Y ] Z, W i + hZ, [X,Y ] W i.
Subtracting from the first equality the second and the third one and applying
[X, Y ] = X Y Y X, we obtain
0 = hX Y Z Y X Z [X,Y ] Z, W i
+ hZ, X Y W Y X W [X,Y ] W i
= R(X, Y ; Z, W ) + R(X, Y ; W, Z).
282 4. Manifolds
For tensors having the symmetries of a Riemannian curvature tensor, we

introduce the following
Definition 4.9.6. Let V be a finite dimensional linear space over R. An
algebraic curvature tensor or Bianchi tensor over V is a 4-linear map S : V
V V V R satisfying the symmetry relations
S(X, Y, Z, W ) = S(Y, X, W, Z) = S(X, Y, W, Z)
and the Bianchi identity
S(X, Y, Z, W ) + S(Y, Z, X, W ) + S(Z, X, Y, W ) = 0.
An algebraic curvature tensor field or Bianchi tensor field over a manifold M

is a tensor field of type (4, 0), which assigns to each point p of M a Bianchi
tensor over Tp M .
We know from linear algebra (see equation (1.8)) that a symmetric bilinear
form is uniquely determined by its quadratic form. More generally, when a
tensor has some symmetries, it can be reconstructed from its restriction to
a suitable linear subspace of its domain. For algebraic curvature tensors, we
have the following
Proposition 4.9.7. Let S1 and S2 be algebraic curvature tensors (or tensor
fields). If S1 (X, Y ; Y, X) = S2 (X, Y ; Y, X) for every X and Y , then S1 = S2 .
Proof. The difference S = S1 S2 is also an algebraic tensor (field), and

S(X, Y ; Y, X) = 0 for all X, Y . We have to show S = 0.
We have for any X, Y, Z
0 = S(X, Y + Z; Y + Z, X)
= S(X, Y ; Y, X) + S(X, Y ; Z, X) + S(X, Z; Y, X) + S(X, Z; Z, X)
= S(X, Y ; Z, X) + S(X, Z; Y, X) + [S(X, Y ; Z, X) + S(Y, Z; X, X)
| {z }
=0
+ S(Z, X; Y, X)] = 2S(X, Y ; Z, X).
Now taking four arbitrary vectors (vector fields) X, Y, Z, W and using

S(X, Y ; Z, X) 0, we obtain
0 = S(X + W, Y ; Z, X + W )
= S(X, Y ; Z, X) + S(X, Y ; Z, W ) + S(W, Y ; Z, X) + S(W, Y ; Z, W )
= S(X, Y ; Z, W ) + S(W, Y ; Z, X),
4.9. Curvature 283
i.e., S is skew symmetric in the first and fourth variables. Thus,
S(X, Y ; Z, W ) = S(Y, X; W, Z) = S(Z, X; W, Y ) = S(Z, X; Y, W ),
in other words, S is invariant under cyclic permutations of the first three vari-
ables. But the sum of the three equal quantities S(X, Y ; Z, W ), S(Y, Z; X, W )
and S(Z, X; Y, W ) is 0 because of the Bianchi symmetry, thus S(X, Y ; Z, W )
is 0.
Exercise 4.9.8. Let S be an algebraic curvature tensor, and let QS (X, Y ) :=
S(X, Y ; Y, X). Prove that QS (X, Y ) = QS (Y, X) and
6S(X, Y ; Z, W ) =QS (X + W, Y + Z) QS (Y + W, X + Z) + QS (Y + W, X)
QS (X + W, Y ) + QS (Y + W, Z) QS (X + W, Z)+
+ QS (X + Z, Y ) QS (Y + Z, X) + QS (X + Z, W )
QS (Y + Z, W ) + QS (X, Z)

QS (Y, Z) + QS (Y, W ) QS (X, W ).
Theorem 4.9.9. Assume that S is an algebraic curvature tensor. Then
S(X, Y ; Z, W ) = S(Z, W ; X, Y ).
Proof. Label the vertices of an octahedron as shown in the figure. It follows
Figure 4.5
easily from our assumptions that the sum of the labels of the vertices of
any of the shaded faces of the octahedron is 0 by a Bianchi identity, in
which the label of the face stands in the last place and the sum goes over
cyclic permutations of the other three variables. If we add these identities
284 4. Manifolds
for the upper two shaded triangles Z and W and subtract the identities
corresponding to the two lower triangles X and Y , we obtain
2S(X, Y ; Z, W ) 2S(Z, W ; X, Y ) = 0,
as we wanted to prove.
Corollary 4.9.10. For the Riemann-Christoffel tensor, the identity
R(X, Y ; Z, W ) = R(Z, W ; X, Y ) holds.
Definition 4.9.11. Let M be a Riemannian manifold, p a point on M , X
and Y two non-parallel tangent vectors at p. The number
R(X, Y ; Y, X)
K(X, Y ) =
|X|2 |Y |2 hX, Y i2
is called the sectional curvature of M at p, in the direction of the plane
spanned by the vectors X and Y in Tp M .
Proposition 4.9.12. If M is a Riemannian manifold with sectional curva-
ture K, X and Y are linearly independent tangent vectors at p M , and
are non-zero scalars, then
(i) K(X, Y ) = K(X + Y, Y );
(ii) K(X, Y ) = K(X, Y );
(iii) K(X, Y ) = K(Y, X).
Furthermore, K depends only on the plane spanned by its arguments.
Proof. (i) follows from
R(X + Y, Y ; Y, X + Y ) = R(X, Y ; Y, X) + R(X, Y ; Y, Y ) + R(Y, Y ; Y, X)

+ R(Y, Y ; Y, Y ) = R(X, Y ; Y, X)
and
|X + Y |2 |Y |2 hX + Y, Y i2
= (|X|2 + |Y |2 + 2hX, Y i)|Y |2 (hX, Y i2 + 2hX, Y i|Y |2 + |Y |4 )
= |X|2 |Y |2 hX, Y i2 .
(ii) follows from
R(X, Y ; Y, X) = 2 2 R(X, Y ; Y, X)
and
|X|2 |Y |2 hX, Y i2 = 2 2 (|X|2 |Y |2 hX, Y i2 ).
4.9. Curvature 285
(iii) comes from the equalities R(X, Y ; Y, X) = R(Y, X; X, Y ) and |X|2 |Y |2

hX, Y i2 = |Y |2 |X|2 hY, Xi2 .
Finally, the last statement follows by (i), (ii), and (iii) as if x1 , y1 and x2 , y2
are two bases of a 2-dimensional linear space, then we can transform one of
them into the other by a finite number of elementary basis transformations
of the form
(x, y) (x + y, y); (x, y) (x, y), where 6= 0; (x, y) (y, x).
Definition 4.9.13. Riemannian manifolds, the sectional curvature function

of which is constant, called spaces of constant curvature or simply space forms.
A space form is elliptic or spherical if K > 0, it is parabolic or Euclidean if
K = 0 and is hyperbolic if K < 0.
Typical examples are the n-dimensional sphere, Euclidean space and hyper-
bolic space. Further examples can be obtained by factorization with fixed
point free actions of discrete groups.
The following remarkable theorem resembles Theorem 3.1.25, but its proof is
not so simple.
Theorem 4.9.14 (Schur). If M is a connected Riemannian manifold, dim M
3 and the sectional curvature K(Xp , Yp ) (Xp , Yp Tp M ) depends only on p
(and does not depend on the plane spanned by Xp and Yp ), then K is constant,
that is, as a matter of fact, it does not depend on p either.
Proof. By the assumption,
R(X, Y ; Y, X) = f (|X|2 |Y |2 hX, Y i2 )
for some function f . Our goal is to show that f is constant. Consider the
tensor field of type (4, 0) defined by
S(X, Y ; Z, W ) = f (hX, W ihY, Zi hX, ZihY, W i).
It is clear from the definition that S is skew-symmetric in the first and last
two arguments. It has also the Bianchi symmetry indeed,
X X
S(X, Y ; Z, W ) = f (hX, W ihY, Zi hX, ZihY, W i)
XY Z XY Z
X
= f (hY, W ihZ, Xi hX, ZihY, W i) = 0,
XY Z
thus S is an algebraic curvature tensor field. We also have R(X, Y ; Y, X) =

S(X, Y ; Y, X), therefore R = S. Set
S(X, Y ; Z) = f (hY, ZiX hX, ZiY ).
286 4. Manifolds
Then for any vector field W , we have
hR(X, Y ; Z), W i = R(X, Y ; Z, W ) = S(X, Y ; Z, W ) = hS(X, Y ; Z), W i,
that is,
R(X, Y ; Z) = S(X, Y ; Z) for all X, Y, Z.
Differentiating with respect to a vector field U we get
(U R)(X, Y ; Z) = (U S)(X, Y ; Z)
= U (S(X, Y ; Z)) S(U X, Y ; Z)
S(X, U Y ; Z) S(X, Y ; U Z).
Since
U (S(X, Y ; Z)) = U (f )(hY, ZiX hX, ZiY ) + f U (hY, ZiX hX, ZiY )
= U (f )(hY, ZiX hX, ZiY ) + f (U hY, ZiX + hY, ZiU X
U hX, ZiY hX, ZiU Y )
= U (f )(hY, ZiX hX, ZiY ) + f (hU Y, ZiX + hY, U ZiX+
+ hY, ZiU X hU X, ZiY hX, U ZiY hX, ZiU Y ) =
= U (f )(hY, ZiX hX, ZiY ) + S(U X, Y ; Z) + S(X, U Y ; Z)+
+ S(X, Y ; U Z),
we obtain
(U R)(X, Y ; Z) = (U S)(X, Y ; Z) = U (f )(hY, ZiX hX, ZiY ).
Using the second Bianchi identity, this gives us

X X
U (f )(hY, ZiX hX, ZiY ) = (U R)(X, Y ; Z) = 0.
U XY U XY
If X Tp M is an arbitrary tangent vector, then we can find non-zero vectors

Y , Z = U Tp M such that X, Y and U are orthogonal (dim M 3!). Then
X
0= U (f )(hY, ZiX hX, ZiY ) = X(f )hU, U iY Y (f )hU, U iX.
U XY
Since X and Y are linearly independent and hU, U i is positive, X(f ) =

Y (f ) = 0 follows, yielding that the derivative of f with respect to an ar-
bitrary tangent vector X is 0. This means that f is locally constant, and
since M is connected, f is constant.
4.9. Curvature 287
The curvature tensor is a complicated object containing a lot of information

about the geometry of the manifold. There are some obvious ways to derive
simpler tensor fields from the curvature tensor. Of course, simplicity is paid
by losing information.
Definition 4.9.15. Let (M, ) be a manifold with an affine connection, R
be the curvature tensor of . The Ricci tensor Ric of the connection is a
tensor field of type (2, 0) assigning to the vector fields X and Y the function
Ric(X, Y ) the value of which at p M is the trace of the linear mapping
Tp M Tp M
Zp 7 R(Zp , X(p); Y (p)), where Zp Tp M.
Proposition 4.9.16. The Ricci tensor of a Riemannian manifold is a sym-

metric tensor
Ric(X, Y ) = Ric(Y, X).
Proof. Let e1 , . . . , en be an orthonormal basis in Tp M , where p is an arbitrary
point in the Riemannian manifold M . We can compute the trace of a linear
mapping A : Tp M Tp M by the formula
n
X
Tr A = hA(ei ), ei i.
i=1
In particular,
n
X n
X
Ric(X, Y )(p) = hR(ei , X(p); Y (p)), ei i = R(ei , X(p); Y (p), ei )
i=1 i=1
n
X n
X
= R(Y (p), ei ; ei , X(p)) = R(ei , Y (p); X(p), ei )
i=1 i=1
= Ric(Y, X)(p).
Since the Ricci tensor of a Riemannian manifold is symmetric, it is uniquely
determined by its quadratic form X Ric(X, X) (see equation (1.8)).
Definition 4.9.17. Let Xp Tp M be a non-zero tangent vector of a Rie-
mannian manifold M . The Ricci curvature of M at p in the direction Xp is
the number
Ric(Xp , Xp )
r(XXp ) = 2
.
= e1|X
, e2p,| . . . , en we can express the Ricci
p
Fixing an orthonormal basis
|Xp |
curvature as follows
n n
Ric(Xp , Xp ) X R(ei , Xp ; Xp , ei ) X
r(Xp ) = = = K(Xp , ei ).
|Xp |2 i=1
|Xp |2 i=2
288 4. Manifolds
The meaning of this formula is that the Ricci curvature in the direction Xp
is the sum of the sectional curvatures in the directions of the planes spanned
by the vectors Xp and ei , where ei runs over an orthonormal basis of the
orthogonal complement of Xp in Tp M . It is a nice geometrical corollary that
this sum is independent of the choice of the orthogonal basis.
With the help of a scalar product, one can associate to every bilinear function
a linear transformation. For the case of the Riemannian metric and the Ricci
tensor, we can find a unique F(M )-linear transformation Ric : X(M ) X(M )
defined by
Ric(X, Y ) = hX, Ric(Y )i for every X, Y X(M ).
Definition 4.9.18. The scalar curvature s(p) of a Riemannian manifold M

at a point p is the trace of the linear mapping Ric : Tp M Tp M .
Let us find an expression for the scalar curvature in terms of the Ricci cur-
vature and the sectional curvature. Let e1 , . . . , en be an orthonormal basis
in Tp M . Then
n
X n
X n
X
s(p) = Tr Ric = hRic(ei ), ei i = Ric(ei , ei ) = r(ei ),
i=1 i=1 i=1
i.e. s(p) is the sum of Ricci curvatures in the directions of an orthogonal

basis. Furthermore,
n
X X n
n X X
s(p) = r(ei ) = K(ei , ej ) = 2 K(ei , ej ),
i=1 i=1 j=1 1i<jn
j6=i
that is, the scalar curvature is twice the sum of sectional curvatures taken in
the directions of all coordinate planes of an orthonormal coordinate system
in Tp M .
To finish this section with, let us study the curvature tensor of a hyper-
surface M in Rn . According to (4.14), the Levi-Civita connection of a
hypersurface can be expressed as = P , where is the derivation rule
of vector fields along the hypersurface as defined in Definition 3.1.7, P is the
orthogonal projection of a tangent vector of Rn at a hypersurface point onto
the tangent space of the hypersurface at that point. Comparing Definition
3.1.7 to formula (4.10), we see that the derivation of vector fields along a
hypersurface is induced by the Levi-Civita connection of Rn , which we also
denoted by . As the curvature of Rn is 0,
X Y Y X = [X,Y ]
holds for any tangential vector fields X, Y X(M ).

4.9. Curvature 289
We have
X Y Z = P (X Y Z) = P (X (Y Z hY Z, NiN))
= P (X Y Z) P (X(hY Z, Ni)N) P (hY Z, NiX N)
= P (X Y Z) hY Z, NiX N,
where X, Y, Z X(M ). Similarly,
Y X Z = P (Y X Z) hX Z, NiY N.
Combining these equalities with
[X,Y ] Z = P ([X,Y ] Z)
we get the following expression for the curvature tensor R of M
R(X, Y ; Z) = (X Y Z Y X Z) [X,Y ] Z =
= P ((X Y Z Y X Z) [X,Y ] Z)
hY Z, NiX N + hX Z, NiY N
= hX Z, NiY N hY Z, NiX N.
Since hZ, Ni is constant zero,
0 = X(hZ, Ni) = hX Z, Ni + hZ, X Ni
and
0 = Y (hZ, Ni) = hY Z, Ni + hZ, Y Ni.
Putting these equalities together we deduce that
R(X, Y ; Z) = hZ, Y NiX N hZ, X NiY N
= hZ, L(Y )iL(X) hZ, L(X)iL(Y ).
Comparing the formula
R(X, Y ; Z) = hZ, L(Y )iL(X) hZ, L(X)iL(Y )
relating the curvature tensor to the Weingarten map on a hypersurface with
Gauss equations we see that the curvature tensor R coincides with the curva-
ture tensor defined there. This way, the last equation can also be considered
as a coordinate free display of Gauss equations.
Exercise 4.9.19. Prove that if X1 and X2 are two nonparallel principal
directions at a given point p of a hypersurface M , 1 , 2 are the corresponding
principal curvatures, then
K(X1 , X2 ) = 1 2 .
What is the minimum and maximum of K(X, Y ), when X and Y run over
Tp M ?
290 4. Manifolds
Exercise 4.9.20. Express the Ricci curvature of a hypersurface in Rn+1 in

a principal direction in terms of the principal curvatures.
Exercise 4.9.21. Express the scalar curvature of a hypersurface in terms of
the principal curvatures.
4.10 Decomposition of Algebraic Curvature

Tensors
Proposition 4.10.1. Algebraic curvature tensors over an n-dimensional lin-
ear space V form a linear subspace R of the tensor space T (4,0) V , the dimen-
sion of which is (n4 n2 )/12.
Proof. Skew symmetry in the first two and last two variables implies that
every algebraic curvature R tensor can be obtained from a uniquely defined
bilinear function R on 2 V by R(X, Y, Z, W ) = R(XY, ZW ). By Theorem
4.9.9, R must also be a symmetric. bilinear function.
Let R T (4,0) V be the linear space of 4-linear functions S which have the
symmetry properties
S(X, Y, Z, W ) = S(Y, X, Z, W ) = S(X, Y, W, Z) = S(Z, W, X, Y ).
It is clear that R is isomorphic to the space of symmetric bilinear functions

on 2 V . Since dim 2 V = n(n1)2 ,

n(n1) n(n1)
2 2 + 1 n(n 1)(n2 n + 2)
dim R = = .
2 8
By Theorem 4.9.9 R R, and R can be given as the kernel of the map
B : R R defined by
X
B(S)(X, Y, Z, W ) = S(X, Y, Z, W ).
XY Z
The image of B contains only alternating tensors. Indeed, B(S)(X, Y, Z, W )

is clearly invariant under cyclic permutations of (X, Y, Z), furthermore,
B(S)(X, Y, Z, W ) = S(X, Y, Z, W ) + S(Y, Z, X, W ) + S(Z, X, Y, W )

= S(X, Y, W, Z) S(W, X, Y, Z) S(Y, W, X, Z)
= B(X, Y, W, Z),
and the cyclic permutation of (X, Y, Z) and the transposition of (Z, W ) gen-
erates the full permutation group of (X, Y, Z, W ).
4.10. Decomposition of Algebraic Curvature Tensors 291
On the other hand, if S is alternating, then B(S) = 3S, so the image of

B contains all alternating tensors, that is, im B = A4 (V ). Applying the
dimension formula (1.2), we obtain
dim R = dim R dim im B

n4 2n3 + 3n2 2n n(n 1)(n 2)(n 3) n4 n2
= = .
8 24 12
Assume that V is a Euclidean linear space and denote by g the inner product
of V . Then the orthogonal group O(V ) acts on R by
((R))(X, Y ; Z, W ) = R(1 (X), 1 (Y ); 1 (Z), 1 (W )).
We want to study the question how to decompose R into irreducible O(V )-

invariant subspaces. We can find O(V )-invariant subspaces in R with the
help of the KulkarniNomizu product.
Definition 4.10.2. Let V be a finite dimensional linear space, h, k S 2 (V )
be two symmetric bilinear functions on V . Then the KulkarniNomizu prod-
uct of h and k is the 4-linear map h k T (4,0) (V ) given by

h(X, Z) h(X, W ) k(X, Z) k(X, W )
h k(X, Y ; Z, W ) =

+

k(Y, Z) k(Y, W ) h(Y, Z) h(Y, W )
= h(X, Z)k(Y, W ) h(X, W )k(Y, Z) + k(X, Z)h(Y, W ) k(X, W )h(Y, Z).

The KulkarniNomizu product is an R-bilinear commutative operation. Its

relevance to algebraic curvature tensors is that the KulkarniNomizu product
of two symmetric bilinear functions is always an algebraic curvature tensor.
Proposition 4.10.3. S 2 (V ) g = {h g | h S 2 (V )} is an O(V ) invariant
linear subspace of R.
Proof. Since hg is linear in h, S 2 (V )g is a linear subspace. The orthogonal

group O(V ) has a natural action also on S 2 (V ), defined by (h)(X, Y ) =
h(1 X, 1 Y ). If O(V ) is an orthogonal transformation, then (h
k) = (h)(k), and since (g) = g, (hg) = (h)g. As a consequence
(S 2 (V ) g) = S 2 (V ) g.
Now we construct an O(V )-invariant subspace in R complementary to S 2 (V )

g.
292 4. Manifolds
Definition 4.10.4. The trace of a bilinear function h S 2 (V ) defined on a

Euclidean linear space V with inner product g is tr(h) = hh, gi.
Proposition P 4.10.5. If e1 , . . . , en is an orthonormal basis of V , h S 2 (V ),
n
then tr(h) = i=1 h(ei , ei ). In particular, the sum on the right-hand side
does not depend on the choice of the orthonormal basis.
Proof. Using the natural embedding S 2 (V ) T (2,0) V , if e1 , . . . en is the dual
basis of the basis e1 , . . . , en , then
* n n
+ n
X X X
i j k k
hh, gi = h(ei , ej )e e , e e = h(ei , ei ).
i,j=1 k=1 i=1
Definition 4.10.6. Let e1 , . . . , en be an orthonormal basis of V and take

R R. The Ricci tensor of R is the bilinear function Ric(R) T (2,0) (V )
given by
Xn
Ric(R)(X, Y ) = R(X, ei , ei , Y ).
i=1
Proposition 4.10.7.
The Ricci tensor of R does not depend on the choice of the orthonormal
basis.
Ric(R) is symmetric and the linear map Ric : R S 2 (V ) is O(V )
equivariant, that is, Ric((R)) = (Ric R)) for any R R.
Ric(h g) = (2 n)h tr(h)g, in particular, Ric(g g) = 2(1 n)g
and tr(Ric(h g)) = 2(1 n) tr(h).
Proof. Ric(R)(X, Y ) can also be obtained as the trace of the bilinear func-
tion h(Z, W ) = R(X, Z, W, Y ), thus it does not depend on the choice of
orthonormal basis.
Symmetry of Ric(R) follows from
R(X, ei , ei , Y ) = R(ei , Y, X, ei ) = (1)2 R(Y, ei , ei , X) = R(Y, ei , ei , X).
The O(V ) equivariance follows from the first part of the proposition and the
fact that an orthogonal transformation maps an orthonormal basis to an
n
X
Ric((R))(X, Y ) = R(1 (X), 1 (ei ), 1 (ei ), 1 (Y ))
i=1
n
X
= (R(1 (X), ei , ei , 1 (Y )) = (Ric(R))(X, Y ).
i=1
Pn Pn
Finally, using the decompositions X = i g(X, ei )ei and Y = i g(Y, ei )ei ,
we obtain
Ric(hg)(X, Y )
Xn
= (h g)(X, ei , ei , Y )
i
n
X
= h(X, ei )g(ei , Y ) h(X, Y )g(ei , ei )
i
+ g(X, ei )h(ei , Y ) g(X, Y )h(ei , ei )
= (2 n)h(X, Y ) tr(h)g(X, Y ).
Definition 4.10.8. Algebraic curvature tensors with vanishing Ricci curva-

ture are called algebraic Weyl tensors. They form an O(V )-invariant linear
subspace W of R.
Proposition 4.10.9. The space of algebraic curvature tensors decomposes

as
R = (S 2 (V ) g) W.
Proof. If dim V = 1, then there is nothing to prove as R is 0-dimensional.

For dim V = 2, dim R = 1, thus R = Rg g = S 2 (V ) g. Since Ric(g g) =
2g 6= 0, W is 0-dimensional.
Assume now that dim V = n > 2. Then for any h S 2 (V ), we have
Ric(h g) + tr(h)g Ric(h g) tr Ric(h g)g

h= = + ,
2n 2n 2(1 n)(2 n)
therefore the linear map P : R S 2 (V ) g defined by

Ric(R) tr Ric(R)g
P (R) = + g (4.16)
2n 2(1 n)(2 n)
is a projection onto S 2 (V ) g, that is, its restriction onto S 2 (V ) g is the

identity. This implies that R = (S 2 (V ) g) ker P . It is also clear from
the definition of W that ker P W. Conversely, if R is in the kernel of the
projection, then
2(n 1) Ric(R) = tr(Ric R)g.
Taking the trace of both sides 2(n 1) tr(Ric(R)) = n tr(Ric(R)) is obtained,
from which tr(Ric R) = 0 and Ric(R) = 0, thus, R W. As R = (S 2 (V )
g) ker P , equation ker P = W completes the proof.
294 4. Manifolds
Corollary 4.10.10. The components of an algebraic curvature tensor R

R, corresponding to the decomposition R = (S 2 (V ) g) W are P (R)
S 2 (V ) g defined in (4.16), and

Ric(R) tr Ric(R)g
W (R) = R P (R) = R + g W. (4.17)
2n 2(1 n)(2 n)
Definition 4.10.11. We call P (R) the Ricci component of the algebraic cur-
vature tensor R as it depends only on the Ricci tensor Ric(R). The algebraic
Weyl tensor W (R) W given by equation (4.17) is called the Weyl tensor
or Weyl component of R.
The action of O(V ) on S 2 (V ) is not irreducible. Since (g) = g for any
O(V ), the one dimensional subspace spanned by g is an O(V ) invariant
subspace. The trace is an O(V ) invariant linear function on S 2 (V ), so its ker-
nel S02 (V ) = {h S 2 (V ) | tr(h) = 0}, the linear space of 0-trace symmetric
bilinear functions, is an O(V ) invariant linear subspace of codimension 1.
Proposition 4.10.12. If V is an n-dimensional Euclidean linear space with
inner product g, then the linear space S 2 (V ) splits into the direct sum of the
O(V ) invariant subspaces Rg and S02 (V ). The components of h S 2 (V )
corresponding to this direct sum decomposition are (tr(h)/n)g R and h
(tr(h)/n)g S02 (V ).
We leave the proof as an exercise to the reader.
Definition 4.10.13. The scalar curvature of an algebraic curvature tensor
R is the trace s(R) = tr(Ric(R)) of the Ricci tensor of R.
Applying Proposition 4.10.12 to decompose the Ricci component of a curva-
ture tensor, we obtain the following theorem summarizing the above calcula-
tions.
Theorem 4.10.14. Let V be a Euclidean linear space with inner product g.
If dim V = 2, then R = Rg g, therefore, every algebraic curvature

tensor R can be written as R = s(R) 2
4 g g and S0 (V ) g and W are
0-dimensional.
If dim V 3, then the space of algebraic curvature tensors decomposes

into the direct sum of the O(V ) invariant subspaces Rg g, S02 (V ) g
and W. The corresponding decomposition of an algebraic curvature
tensor R is

s(R) 1 s(R)
R= gg+ Ric(R) g g + W (R).
2(1 n)n 2n n
In the special case when dim V = 3, dim R = 6 = dim S 2 (V ), conse-

quently, dim W = 0, so the Weyl component of any algebraic curvature
tensor is 0.
s(R)
Definition 4.10.15. The component 2(1n)n g g Rg g, depending only
on the scalar curvature
of R is called
the scalar component of R, while the
s(R) 2
second component Ric(R) n g g S0 (V ) g is the trace free Ricci
s(R)
component of R as it comes from the trace free part Ric(R) n g of the
Ricci tensor of R.
We mention without proof that the representations of the group O(V ) on the
spaces Rg g, S02 (V ) g and W are irreducible, so we cannot decompose
these spaces into smaller O(V ) invariant subspaces.
Consider a Riemannian manifold (M, g). The Riemannian metric g turns
each tangent space Tp M into a Euclidean linear space and the Riemannian
curvature tensor R yields an algebraic curvature tensor on Tp M for all p M .
Applying the above decomposition of algebraic curvature tensors pointwise,
we obtain a decomposition of the Riemannian curvature tensor into three
components, some of which automatically vanish in dimensions less then or
equal to 3.
In dimension 2, the Riemannian curvature tensor can be written as
s K
R = g g = g g,
4 2
where K assigns to each point p M the sectional curvature in the direction
Tp M . In particular,
s
R(X, Y, Z, W ) = (hX, W ihY, Zi hX, ZihY, W i)
2
= K(hX, W ihY, Zi hX, ZihY, W i).
In dimension 3, the Riemannian curvature tensor is uniquely determined by

the Ricci tensor according to the formula
s s s
R = g g Ric g g = g g + Ric g,
12 3 4
where s = tr Ric is the scalar curvature function.
From dimensions greater than 3, the pointwise Weyl component appears and
defines a tensor field C = W (R) of type (4, 0) on M . This tensor field is called
the type (4, 0) conformal Weyl tensor field. The adjective conformal and
the related standard notation C for the Weyl tensor field refers to the invari-
ance property of this tensor field under conformal changes of the Riemannian
metric. This property will be discussed in the next section.
296 4. Manifolds
Some important classes of Riemannian manifolds can be defined by restric-

tions on the components of the curvature tensor.
For example, as we saw in the proof of Schurs theorem, a space has constant
curvature K R if and only if its curvature tensor has the form K 2 g g.
This means that a constant curvature space has vanishing trace free Ricci
component and Weyl component. Schurs theorem can be interpreted as a
converse of this statement: If the curvature tensor of a connected Riemannian
manifold of dimension at least 3 has vanishing trace free Ricci component and
Weyl component, then it has constant sectional curvature.
Definition 4.10.16. A Riemannian manifold is called an Einstein manifold

if the Ricci curvature r(Xp ) equals the same constant r R for all p M
and Xp Tp M .
For an Einstein manifold with constant Ricci curvature r, the Ricci tensor
Ric is equal to rg, so the trace free Ricci component (Ric rg) is equal to 0.
The following theorem is an analog of Schurs theorem.
Theorem 4.10.17. A connected Riemannian manifold of dimension at least

3 is an Einstein manifold if and only if the trace free Ricci component of its
curvature tensor vanishes.
Proof. The trace free part of the Ricci tensor vanishes if and only if Ric = rg
for a smooth function r on the manifold M . Our goal is to show that in this
case r must be constant.
Since M is connected, it is enough to show that r has vanishing derivative
with respect to any vector field Z. To check Z(r) = 0 at a point p M , choose
a local coordinate system around p and apply the GramSchmidt orthogo-
nalization process to the vector fields 1 , . . . , n to obtain orthonormal vector
. . . , En in a neighborhood of p. Write Z Ei as a linear combination
fields E1 ,P
Z Ei = ij Ej . The coefficients ij can be computed as ij = hZ Ei , Ej i.
Differentiating the equation hEi , Ej i = ij , we obtain that ij = ji .
As a corollary of this skew symmetry, we obtain that for any two vector fields
X, Y , we have
n
X
(R(X, Z Ei , Ei , Y ) + R(X, Ei , Z Ei , Y ))
i=1
n
X
= (R(Z Ei , X, Y, Ei ) + R(Ei , X, Y, Z Ei ))
i=1
Xn
= (R(ij Ej , X, Y, Ei ) + R(Ei , X, Y, ij Ej ))
i,j=1
n
X
= (ij + ji )R(Ej , X, Y, Ei ) = 0. (4.18)
i,j=1
Pn
Differentiating the defining equation s = i,j=1 R(Ei , Ej , Ej , Ei ) of the
scalar curvature with respect to Z using (4.18) and the second Bianchi iden-
tity,
n
X
Z(s) = [(Z R)(Ei , Ej , Ej , Ei ) R(Z Ei , Ej , Ej , Ei )
i,j=1
R(Ei , Z Ej , Ej , Ei ) R(Ei , Ej , Z Ej , Ei )
R(Ei , Ej , Ej , Z Ei )]
n
X
= (Z R)(Ei , Ej , Ej , Ei ) (4.19)
i,j=1
Xn
= [(Ei R)(Ej , Z, Ej , Ei ) + (Ej R)(Z, Ei , Ej , Ei )]
i,j=1
X n
=2 (Ej R)(Z, Ei , Ei , Ej ).
i,j=1
Pn
Differentiate now the equation rhZ, Ej i = Ric(Z, Ej ) = i=1 R(Z, Ei , Ei , Ej )
with respect to Ej to get
Ej (r)hZ, Ej i + rhEj Z, Ej i + rhZ, Ej Ej i
X n

= (Ej R)(Z, Ei , Ei , Ej ) + R(Ej Z, Ei , Ei , Ej )+
i=1

+ R(Z, Ej Ei , Ei , Ej ) + R(Z, Ei , Ej Ei , Ej ) + R(Z, Ei , Ei , Ej Ej )
Xn

= (Ej R)(Z, Ei , Ei , Ej ) + R(Ej Z, Ei , Ei , Ej ) + R(Z, Ei , Ei , Ej Ej ) .
i=1
Since
rhEj Z, Ej i + rhZ, Ej Ej i
= Ric(Ej Z, Ej ) + Ric(Z, Ej Ej )
Xn

= R(Ej Z, Ei , Ei , Ej ) + R(Z, Ei , Ei , Ej Ej ) ,
i=1
we obtain
n
X
Ej (r)hZ, Ej i = (Ej R)(Z, Ei , Ei , Ej ).
i=1
298 4. Manifolds
Summing for j and comparing the result to (4.19) we get

n n
X X Z(s)
Z(r) = (hZ, Ej iEj )(r) = (Ej R)(Z, Ei , Ei , Ej ) = .
j=1 i=1,j
2
On the other hand, taking the trace of the equation Ric = rg, we obtain
s = r dim M , in particular dim M Z(r) = Z(s). As dim M 3, this implies
that Z(r) = Z(s) = 0.
We remark that 3-dimensional Einstein manifolds are spaces of constant cur-

vature, since in dimension 3 the Weyl tensor vanishes automatically.
4.11 Conformal Invariance of the Weyl Tensor

The goal of this section is to compute how the curvature tensor of a Rie-
mannian manifold is transformed when the Riemannian metric is replaced by
another one, proportional to the original metric.
Definition 4.11.1. Two Riemannian metrics g and g are called conformally

equivalent if there is a smooth function on M such that g = g. Since
both g and g are positive definite, the conformal factor must be positive
everywhere. Thus, it can be written as = e2f for some smooth function f .

To understand how the curvature tensors of conformally equivalent metrics

are related to one another, first we compute the transformation rule between
their Levi-Civita connections.
Definition 4.11.2. The gradient vector field grad f of a smooth function f

defined on a Riemannian manifold M is the unique vector field satisfying
Xf = hX, grad f i for any vector field X X(M ).
Let us compute the gradient vector field of a function in local coordinates.

Let = (x1 , . . . , xn ) : U Rn be a chart and 1 = 1 , . . . , n = n be the
basis vector fields induced by the chart. Denote by G = (gij ) = (hi , j i) the
matrix of the Riemannian metric with respect to these basis vectors, and by
G 1 = (g ijP) the inverse matrix. We look for grad f as a linear combination
n
grad fP= i=1 Gi i . By the definition of the gradient, for any vector field
n
X = i=1 X i i we should have
n
X n
X
Xf = X i i (f ) = hX, grad f i = X i Gj gij .
i=1 i,j=1
4.11. Conformal Invariance of the Weyl Tensor 299
Thus, the unknown coefficients Gi can be found by solving the linear system
of equations
n
X
i (f ) = Gj gij , (i = 1, . . . , n).
j=1
Multiplying the ith equation by g ik , and summing up for i we obtain

n
X n
X n
X
i (f )g ik = Gj gij g ik = Gj jk = Gk ,
i=1 i,j=1 j=1
hence
n
X
grad f = i (f )g ij j .
i,j=1
Suppose that the Levi-Civita connection of g is , and that of g = e2f g is

Let us compute the difference tensor X Y
. XY .
Applying the Koszul formula (4.12)
X Y, Z) = X(g(Y, Z)) + Y (g(X, Z)) Z(g(X, Y ))
2g(
+ g([X, Y ], Z) g([X, Z], Y ) g([Y, Z], X),
thus
X Y, Z) = e2f X(g(Y, Z)) + 2X(f )e2f g(Y, Z)+
2e2f g(
+ e2f Y (g(X, Z)) + 2Y (f )e2f g(X, Z)
e2f Z(g(X, Y )) 2Z(f )e2f g(X, Y )+
+ e2f (g([X, Y ], Z) g([X, Z], Y ) g([Y, Z], X)).
Dividing by e2f and taking the Koszul formula
2g(X Y, Z) = X(g(Y, Z)) + Y (g(X, Z)) Z(g(X, Y ))+

+ g([X, Y ], Z) g([X, Z], Y ) g([Y, Z], X).
into consideration we conclude the formula

X Y = X Y + X(f )Y + Y (f )X g(X, Y ) grad(f ).
(4.20)
Definition 4.11.3. Let f be a smooth function on a manifold M endowed

with a connection . The Hesse form of f with respect to the connection
is the (2, 0) type tensor field Hesse(f ) the evaluation of which on the vector
fields X, Y is
Hesse(f )(X, Y ) = X(Y (f )) (X Y )(f ).
300 4. Manifolds
Proposition 4.11.4. The Hesse form is a tensor field indeed, that is,
Hesse(f )(X, Y ) is a bilinear function of X and Y over smooth functions.
If is torsion free, then the Hesse form is symmetric. At a critical point p
of f , i.e., at a point where the differential (df )p vanishes, Hesse(f )p does not
depend on the connection.
Proof. Additivity in X and Y is obvious. If g is a smooth function on M ,

then
Hesse(f )(gX, Y ) = gX(Y (f )) (gX Y )(f ) = gX(Y (f )) g(X Y )(f )

= g Hesse(f )(X, Y ),
and
Hesse(f )(X, gY ) = X(gY (f )) (X gY )(f )

= gX(Y (f )) + X(g)Y (f ) g(X Y )(f ) X(g)Y (f )
= g Hesse(f )(X, Y ).
The equation
Hesse(f )(X, Y ) Hesse(f )(Y, X) = (X Y Y X (X Y Y X))(f )

= T (X, Y )(f )
shows the symmetry for vanishing torsion.

Finally, if p is a critical point of f , then (X Y )p (f ) = (df )p (X Y ) = 0, so
Hesse(f )p (Xp , Yp ) = Xp (Y (f ))
does not depend on . (In the last equation, Xp , Yp Tp M , Y is an arbitrary

vector field the value of which at p is Yp .)
Now we are ready to compute the relation between the curvature tensor R
of the metric g = e2f g and the curvature tensor R of g.
From equation (4.20)
X
Y Z =
X (Y Z + Y (f )Z + Z(f )Y g(Y, Z) grad(f ))
= X (Y Z + Y (f )Z + Z(f )Y g(Y, Z) grad(f )+)
+ X(f )(Y Z + Y (f )Z + Z(f )Y g(Y, Z) grad(f ))+
+ ((Y Z)(f ) + Y (f )Z(f ) + Z(f )Y (f ) g(Y, Z)|| grad(f )||2 )X
(g(X, Y Z) + Y (f )g(X, Z) + Z(f )g(X, Y ) g(Y, Z)X(f )) grad(f )
Expanding the first line of the last expression we get

}
X
Y Z = X Y Z + X(Y (f ))Z + Y (f )X Z + X(Z(f ))Y

+ Z(f )X Y g(X Y, Z) grad(f ) g(Y, X Z) grad(f )
}
g(Y, Z)X (grad(f )) + X(f )Y Z + X(f )Y (f )Z

+ X(f )Z(f )Y X(f )g(Y, Z) grad(f ) +

+ (Y Z)(f )X + Y (f )Z(f )X + Z(f )Y (f )X g(Y, Z)|| grad(f )||2 X

+ ( g(X, Y Z) Y (f )g(X, Z) Z(f )g(X, Y )

+ g(Y, Z)X(f ) ) grad(f )
To work with expressions antisymmetric in two variables, it is convenient to

introduce the antisymmetrizing operator. If F (X, Y, . . . ) is an expression in-
volving the variables X and Y and maybe other ones, then set
XY (F (X, Y, . . . )) = F (X, Y, . . . )F (Y, X, . . . ). The antisymmetrizer X,Y
makes the expression to which it is applied antisymmetric in X and Y .
The terms in the expansion of X Y Z that are in boxes with the same labels
give a symmetric expression in X and Y , therefore, they disappear when we
apply the antisymmetrizer X,Y to both sides, so
X ,
[ Y ]Z =
X
Y Z
Y
X Z = XY (
X
Y Z)
= [X , Y ]Z + [X, Y ](f ))Z + XY (X(Z(f ))Y ) + Z(f )[X, Y ]
g([X, Y ], Z) grad(f ) + XY (g(X, Z)Y (grad(f )))+
+ XY (Y Z)(f )X + Z(f )Y (f )X + || grad(f )||2 g(X, Z)Y +

+ XY (X(f )g(Y, Z) grad(f )).

We know also that
[X,Y ] Z = [X,Y ] Z + [X, Y ](f )Z + Z(f )[X, Y ] g([X, Y ], Z) grad(f ).

Subtracting this equation from the previous one we obtain

R(X, Y )Z = R(X, Y )Z + XY X(Z(f ))Y + g(X, Z)Y (grad(f ))+
+ (Y Z)(f )X + Z(f )Y (f )X + || grad(f )||2 g(X, Z)Y +

+ X(f )g(Y, Z) grad(f ) .
This is the transformation rule of the curvature tensor under conformal trans-
formations of the metric. To produce from this a transformation rule for the
302 4. Manifolds
Riemannian curvature tensor, take the inner product of both sides with a
fourth vector field W with respect to g. This produces
e2f R(X, Y, Z, W ) = R(X, Y, Z, W )+

+ XY X(Z(f ))g(Y, W ) + g(X, Z)g(Y (grad(f )), W )
+ (Y Z)(f )g(X, W ) + Z(f )Y (f )g(X, W )

+ || grad(f )||2 g(X, Z)g(Y, W ) + X(f )W (f )g(Y, Z) .
The expression in the box can be rewritten using the identity
g(Y (grad(f )), W ) = Y g(grad(f ), W ) g(grad(f ), Y W )

= Y (W (f )) (Y W )(f ).
Using this a rearrangement gives
e2f R(X, Y, Z, W )

= R(X, Y, Z, W ) + XY X(Z(f ))g(Y, W ) + g(X, Z)Y (W (f ))
g(X, Z)(Y W )(f )+
+ (Y Z)(f )g(X, W ) + Z(f )Y (f )g(X, W )+
+ || grad(f )||2 g(X, Z)g(Y, W )+

+ X(f )W (f )g(Y, Z) .
Let us flip the role of X and Y in those terms, where there is no X in the
inner product g(., .). Inside the antisymmetrizer XY , this can be done with
a sign change. This results in the formula
e2f R(X, Y, Z, W )

= R(X, Y, Z, W ) + XY Y (Z(f ))g(X, W ) + g(X, Z)Y (W (f ))
g(X, Z)(Y W )(f )+
+ (Y Z)(f )g(X, W ) + Z(f )Y (f )g(X, W )+

g(X, Z)g(Y, W ) g(Y, Z)g(X, W )
+ || grad(f )||2
2
Y (f )W (f )g(X, Z) .
Observe that the expression to which XY is applied is antisymmetric in Z

and W as well, so we can write the right-hand side in a more compressed
form
e2f R(X, Y, Z, W ))
= R(X, Y, Z, W )+

+ ZW XY Y (Z(f ))g(X, W ) + (Y Z)(f )g(X, W )+
|| grad(f )||2
+ Z(f )Y (f )g(X, W ) g(Y, Z)g(X, W )
2
= R(X, Y, Z, W )
!
|| grad(f )||2
+ df df Hesse(f ) g g (X, Y, Z, W ).
2
Thus we obtained the formula

|| grad(f )||2
e2f R = R + df df Hesse(f ) g g.
2
The main corollary of this formula is that the difference e2f R R is the
Kulkarni-Nomizu product of a symmetric tensor of type (2, 0) with g, conse-
quently, its Weyl component is 0. This gives the following theorem.
Theorem 4.11.5. If C and C are the (4, 0) type conformal Weyl tensors of
g and g = e2f g, then C = e2f C.
The decomposition of the Riemannian curvature tensor induces a decompo-
sition of the type (3, 1) curvature tensor as well. In this decomposition, the
Weyl component will be a tensor field C of type (3, 1), which is related to
the type (4, 0) Weyl tensor C by the identity
g(C(X, Y, Z), W ) = C(X, Y, Z, W ).
C is called the type (3, 1) conformal Weyl tensor of the Riemannian manifold.
As a corollary of the previous theorem, we see that the type (3, 1) Weyl tensor
fields of conformally equivalent metrics are equal.
Definition 4.11.6. A Riemannian manifold is called flat if its curvature ten-
sor vanishes. A Riemannian manifold is locally conformally flat if each point
has an open neighborhood over which the metric is conformally equivalent to
a flat one.
It can be proved that
(1) a flat Riemannian manifold is locally isometric to a Euclidean space;
304 4. Manifolds
(2) a space is locally conformally flat if it has an atlas of angle preserving

charts.
Angle preserving charts, i.e., charts with respect to which the matrix of the
Riemannian metric is a multiple of the unit matrix at each point are called
isothermal local coordinate systems.
According to the conformal invariance of the Weyl tensor, if a Riemannian
manifold is locally conformally flat, then its Weyl tensor is 0. It is a remark-
able fact that this condition is not just necessary but also sufficient if the
dimension of the manifold is not 3.
Theorem 4.11.7. Let n be the dimension of a Riemannian manifold M .
If n = 2, then M is locally conformally flat.
If n = 3, then M is locally conformally flat if and only if its Cotton
tensor defined by
C(X, Y, Z) = (Z Ric)(X, Y ) (Y Ric)(X, Z)

Z(s)g(X, Y ) Y (s)g(X, Z)

2(n 1)
vanishes.
If n 4, then M is locally conformally flat if and only if its Weyl tensor
vanishes.
The proof of this theorem goes beyond the scope of this textbook.
4.12 Geodesics
We define the length of a smooth curve : [a, b] M lying on a Riemannian
manifold (M, h , i) to be the integral
Z b p
l() = h 0 (t), 0 (t)idt.
a
It is worth mentioning that the classical definition of length as the limit of

the lengths of inscribed broken lines does not make sense, since the distance
of points is not directly defined. The situation is just the opposite. We can
define first the length of curves as a primary concept and derive from it a so
called intrinsic metric d(p, q), at least for connected Riemannian manifolds
as the infimum of the lengths of all curves joining p to q. The metric enables
us to define the length of broken lines given just by a sequence of vertices
P1 , . . . , PN to be the sum of the distances between consecutive vertices. There
4.12. Geodesics 305
is a theorem saying that the length of a smooth curve : [a, b] M is equal

to the limit of the lengths of inscribed broken lines (t0 ), (t1 ), . . . , (tN ),
a = t0 < t1 < < tN = b as the maximum of the distances |ti ti1 | tends
to zero.
To find the analog of straight lines in the intrinsic geometry of a Riemannian
manifold we have to characterize straight lines in a way that makes sense for
Riemannian manifolds as well. Since the length of curves is one of the most
fundamental concepts of Riemannian geometry, we can take the following
characterization: a curve is a straight line if and only if for any two points
on the curve, the segment of the curve bounded by the points is the shortest
among curves joining the two points. A slight modification of this property
could be used to distinguish a class of curves, but it is not clear at first glance
whether such curves exist at all on a general Riemannian manifold.
For a physicist a straight line is the trajectory of a particle with zero ac-
celeration or that of a light beam. This observation can also give rise to a
definition. We only have to find a proper generalization of acceleration for
curves lying in a Riemannian manifold. It seems quite natural to proceed as
follows. The speed vectors of a curve yield a vector field along the curve. On
the other hand, by the fundamental theorem of Riemannian geometry, the
Riemannian metric determines a unique affine connection on the manifold
which is symmetric and compatible with the metric. In particular, one can
differentiate the speed vector field with respect to the curve parameter and
may call the result 00 = 0 0 the acceleration vector field along the curve.
Definition 4.12.1. Let M be a Riemannian manifold, be a curve on it.
We say that is a geodesic if

0 0 = 0.
Remark. More generally, if (M, ) is a manifold with an affine connection,
then the curves satisfying 0 0 = 0 are said to be autoparallel. Geodesics
are autoparallel curves for the Levi-Civita connection.
Proposition 4.12.2. The length of the speed vector of a geodesic is constant.
Proof. By the compatibility of the connection with the metric, parallel trans-
port preserves length and angles between vectors. The definition of geodesics
implies that the speed vector field is parallel along the curve, consequently
consists of vectors of the same length.
The proposition follows also from the equality
h 0 , 0 i0 = h 0 0 , 0 i + h 0 , 0 0 i = 0.
As a consequence, we get that the property of being geodesic is not invariant

under reparameterization. The parameter t of a regular geodesic is always
306 4. Manifolds
related to the natural parameter s through an affine linear transformation

i.e. t = as + b for some a, b R. This motivates the following definition.
Definition 4.12.3. A regular curve on a Riemannian manifold is a pre-
geodesic if its natural reparameterization is geodesic.
In terms of a local coordinate system with coordinates (x1 , . . . , xn ) a curve
in the domain of the chart determines (and is determined by) n smooth
functions i = xi (1 i n). The equation 0 0 = 0 then takes the
form
n
00 X 0 0
k + kij i j = 0 for all 1 k n.
i,j=1
The existence of geodesics depends, therefore, on the solutions of a certain

system of second order differential equations.
0
Introducing the new functions v i = i this system of n second order differ-
ential equations becomes a system of 2n first order equations
0
k k
= v ,

n
k0
X
k i j
for all 1 k n.

v = ij v v ,
i,j=1
Applying the existence and uniqueness theorem for ordinary differential equa-
tions one obtains the following.
Proposition 4.12.4. For any point p on a Riemannian manifold M and
for any tangent vector X Tp M , there exists a unique maximal geodesic
defined on an interval containing 0 such that (0) = p and 0 (0) = X.
If the maximal geodesic through a point p with initial velocity X is defined
on an interval containing [, ] then there is a neighborhood U of X in the
tangent bundle such that every maximal geodesic started from a point q with
initial velocity Y Tq M is defined on [, ].
Since a geodesic with zero initial speed can be defined on the whole real
straight line, for each point p on the manifold one can find a positive such
that for every tangent vector X Tp M with kXk < , the geodesic defined
by the conditions (0) = p, 0 (0) = X can be extended to the interval [0, 1].
The following notation will be convenient. Let X Tp M be a tangent vector
and suppose that there exists a geodesic : [0, 1] M satisfying the con-
ditions (0) = p, 0 (0) = X. Then the point (1) M will be denoted by
expp (X) and called the exponential of the tangent vector X.
Using the fact that for any c R, the curve t 7 (ct) is also a geodesic we
see that the geodesic is described by the formula
(t) = expp (tX).
4.12. Geodesics 307
As we have observed, expp (X) is defined provided that kXk is small enough.
In general however, expp (X) is not defined for large vectors X. This motivates
the following.
Definition 4.12.5. A Riemannian manifold is geodesically complete if for all
p M , expp (X) is defined for all vectors X Tp M .
This is clearly equivalent to the requirement that maximal geodesics are de-
fined on the whole real line R.
Proposition 4.12.6. For a fixed point p M , the exponential map expp
is a smooth map from an open neighborhood of 0 Tp M into the mani-
fold. Furthermore, the restriction of it onto a (possibly even smaller) open
neighborhood of 0 Tp M is a diffeomorphism.
Proof. Differentiability of the exponential mapping follows from the theorem

on the differentiable dependence on the initial point for solutions of a system
of ordinary differential equations. To show that expp is a local diffeomor-
phism, we only have to show that its derivative at the point 0 Tp M is
a non-singular linear mapping (see Inverse Function Theorem). Since Tp M
is a linear space, its tangent space T0 (Tp M ) at 0 can be identified with the
vector space Tp M itself. Through this identification, the derivative of the
exponential map at 0 maps Tp M = T0 (Tp M ) into Tp M . We show that this
derivative is just the identity map of Tp M , hence non-singular.
Let X be an element of the tangent space Tp M = T0 (Tp M ). To determine
where X is taken by the derivative of the exponential mapping, we represent
X as the speed vector of the curve t 7 (t) = tX at t = 0. The exponential
mapping takes this curve to the geodesic curve = expp , (t) = expp (tX),
the speed vector of which at t = 0 is X, so the derivative of the exponential
map sends X to itself and this is what we claimed.
By the proposition, we can introduce a local coordinate system, based on

geodesics, about each point of the manifold as follows. We fix an orthonormal
basis in the tangent space Tp M , which gives us an isomorphism : Tp M Rn
that assigns to each tangent vector its components with respect to the basis,
and then take exp1p . The map expp
1
is a diffeomorphism between
an open neighborhood of p and that of the origin in Rn , therefore, it is
a smooth chart on M . Coordinate systems obtained this way are called
normal coordinate systems, while we shall call the inverse of them normal
parameterizations.
For a Riemannian manifold M , we can define the sphere of radius r centered
at p M as the set of points q M such that d(p, q) = r, where d(p, q)
denotes the intrinsic distance of p and q. When the radius of the sphere is
increasing, the topological type of the sphere changes at certain critical values
308 4. Manifolds
of the radius. For small radii however, the intrinsic spheres are diffeomorphic
to the ordinary spheres in Rn , and what is more, we have the following.
Theorem 4.12.7. The normal parameterization of a manifold about a point
p maps the sphere about the origin with radius r, provided that it is contained
in the domain of the parameterization, diffeomorphically onto the intrinsic
sphere centered at p with radius r.
We prove this theorem later.
Formula for the First Variation of the Length

Definition 4.12.8. A variation of a smooth curve : [a, b] M is a smooth
mapping from the rectangular domain [, ] [a, b] into M such that
(0, t) = (t) for all t [a, b].
Given a variation of a curve we may introduce a one parameter family of
curves , [, ] by setting (t) = (, t). By our assumption, these
curves yield a deformation of the curve 0 = .
Theorem 4.12.9. Let be a variation of a geodesic . Let l() denote the
length of the curve . Then the following formula holds
0 (b) 0 (a)

l0 (0) = 1 (0, b), 0 1 (0, a), 0 .
k (b)k k (a)k
Proof. By the definition of the length of a curve, one has

Z b Z bp
d d
l0 (0) = k2 (, )k d = h2 (, ), 2 (, )id

d a =0 d a =0
Z b
d p
= h2 (, ), 2 (, )i d

a d =0
b d
h2 (, ), 2 (, )i|=0
Z
d
= p d.
a 2 h2 (0, ), 2 (0, )i
With the help of the covariant differentiation induced by the Levi-Civita

connection this expression can be written as follows.
Z b Z b
0 ( )

h1 2 (0, ), 2 (0, )i
d =
1 2 (0, ), d.
a k 0 ( )k a k 0 ( )k
By the symmetry of the connection, this is equal to

Z b
0 ( )

2 1 (0, ), 0 d
a k ( )k
4.12. Geodesics 309
0
D E
Observe, that the function t 7 1 (0, t), k 0 (t)k
(t)
is a primitive function
(antiderivative) of the function to be integrated. Indeed, the derivative of
this function is
0 (t) 0 (t)

d
1 (0, t), 0 = 2 1 (0, t), 0
dt k (t)k k (t)k
0

+ 1 (0, t), 0 (t) ,
k 0 k
but the second term on the right-hand side is zero since is geodesic. Con-
sequently,
b
0 (t)
Z
0
l (0) = 2 1 (0, t), 0
k (t)k
a
0 (b) 0 (a)

= 1 (0, b), 0 1 (0, a), 0 .
k (b)k k (a)k
Theorem 4.12.10 (Gauss Lemma). Let M be a Riemannian manifold,

p M , and denote by Sr the sphere of radius r in Tp M centered at the
zero tangent vector. Assume r is chosen to be so small that the exponential
mapping is a diffeomorphism on a ball containing Sr and denote the exponen-
tial image of Sr by Sr . Then for any X Sr the radial geodesic t 7 expp (tX)
is perpendicular to Sr .
Proof. Every tangent vector of Sp can be obtained as the speed vector of a

curve expp where is a curve in Sr passing through (0) = X. Given
such a curve, let us define a variation of the geodesic : t 7 expp (tX) in the
following way
(, t) := expp (t()).
For a fixed , the curve is a geodesic of length r so l() is constant. Thus,
the previous theorem implies that
0 (1) 0 (0)

0
0 = l (0) = 1 (0, 1), 0 1 (0, 0), 0 .
k (1)k k (0)k
Since (, 0) = expp (0 ()) = p and (, 1) = expp (()), we have

1 (0, 0) = 0 and 1 (0, 1) = (expp )0 (0), therefore, we get
0 (1)

0
0= (expp ) (0), 0 ,
| (1)|
showing that intersects Sr orthogonally.

310 4. Manifolds
Now we are ready to prove the theorem saying that Sr is a sphere in the
intrinsic geometry of the manifold. It is clear that d(p, q) r for any point
q on Sr , since the radial geodesic from p to r has length r, so all we need is
the following.
Theorem 4.12.11. If : [a, b] M is an arbitrary curve connecting p to a
point of Sr , then its length is r.
Proof. We may suppose without loss of generality that (b) is the only in-
tersection point of the curve with Sr and (t) 6= p for t > a. Then there is
a unique curve in the tangent space Tp M such that = expp . Let N
denote the vector field on Tp M \ {0} that is the gradient vector field of the
function f : X 7 kXk on Tp M , and therefore consists of unit vectors per-
pendicular to the spheres centered at the origin. The theorem above shows
that the derivative of the exponential map takes N into a unit vector field N
on M , perpendicular to the sets St .
We can estimate the length of a curve as follows
Z b Z b
l() = | 0 ( )|d h 0 ( ), N(( ))id.
a a
Since h 0 ( ), N(( )i is the component parallel to N(q) of the speed vector

0 ( ) with respect to the splitting Tq M = RN(q) Tq S at q = ( ), it
is equal to the component parallel to N (X) of the speed vector ( ) with
respect to the splitting
TX (Tp M ) = RN(X) TX S
at X = ( ). Therefore,
h 0 ( ), N(( )i = h 0 ( ), N(( )i = h 0 ( ), grad f (( ))i = (f )0 ( ),
and
Z b Z b
h 0 ( ), N(( )id = (f )0 ( )d = k(b)k k(a)k = r.
a a
The proof also shows that the equality l() = r holds only for curves perpen-
dicular to the spheres S .
Exercise 4.12.12. Show that such curves are pre-geodesics.
Theorem 4.12.13. A smooth curve : [a, b] M parameterized by the
natural parameter in a Riemannian manifold is geodesic if and only if there
is a positive such that for any two values t1 , t2 [a, b] such that |t1 t2 | < ,
the restriction of onto [t1 , t2 ] is a curve of minimal length among curves
joining (t1 ) to (t2 ).
4.13. Applications to Hypersurface Theory 311
Remark. It is not true in general, that a geodesic curve is the a curve of

minimal length among curves joining the same endpoints. To see this, it is
enough to consider a long arc on a great circle on the sphere.
Exercise 4.12.14. Show that a regular curve in a hypersurface M Rn+1

is a geodesic if and only if its ordinary acceleration 00 (t) is perpendicular to
T(t) M for every t. It is a pre-geodesic if and only if 00 (t) is contained in the
plane spanned by 0 (t) and the normal vector of M at (t).
Exercise 4.12.15. Show that great circles on the sphere and helices on a
cylinder are pre-geodesics.
Exercise 4.12.16. Find a regular pre-geodesic on the cone x2 + y 2 = z 2 ,

different from straight lines.
Exercise 4.12.17. Show that straight lines on a hypersurface are pre-geo-

desic curves.
Exercise 4.12.18. Show that symmetry planes of a surface in R3 intersect

the surface in pre-geodesic lines.
Exercise 4.12.19. Write the differential equation of geodesics on a surface

of revolution with respect to the usual parameterization. Derive from the
equations Clairauts theorem: For a pre-geodesic curve on a surface of rev-
olution the quantity d cos is constant, where d denotes the distance of the
curve point from the axis of symmetry, is the angle between the speed vec-
tor of the curve and the circle of rotation passing through the curve point.

4.13 Applications to Hypersurface Theory

4.13.1 Geodesic Curves on Hypersurfaces
The Frenet theory of curves in Euclidean spaces can be extended to curves
lying in oriented Riemannian manifolds. If (M, g) is a Riemannian manifold
and : I M is a smooth parameterized curve, then using the Levi-Civita
connection of M , we can define higher order derivatives of . There is no
need for connection to define the speed vector field 0 . However, since 0 is
a vector field along , to define higher order derivatives of as vector fields
along , we need a connection. Let us define the vector field k recursively
by 1 = 0 and k+1 = 0 (k ).
We say that a curve : I M is of general type in M if the first (n 1)
covariant derivatives 1 (t), . . . , n1 (t) are linearly independent for all
t I, where n = dim M .
312 4. Manifolds
For a curve of general type the distinguished Frenet vector fields ti : I T M

can be defined and computed in the same way as in Rn . The first (n 1)
of them is the result of the GramSchmidt orthogonalization process applied
to 1 , . . . , n1 , the last one is the unique vector which extends the first
(n 1) to a positively oriented orthonormal basis.
Once we defined the distinguished Frenet frame, the curvature functions of
can be defined as the functions 1 , . . . , n1 appearing in the Riemannian
version of the Frenet formulae
1
0 t1 = 1 t2 ,
k 0 k
..
.
1
0 ti = i1 ti1 + i ti+1 ,
k 0 k
..
.
1
0 tn = n1 tn1 .
k 0 k
We can also prove the following extension of Theorem 2.5.14 for curves of
general type in Riemannian manifolds.
Theorem 4.13.1. Let : I M be a curve of general type in M , 1 k
n 1. Then the curvature functions of can be computed by the equations
2 k+1 k1
1 = and k = for k 2,
v3 v2k
where v = k 0 k, and the numbers k are given by
v
u
u
u h1 , 1 i h1 , k i
...
.. ..

k = u
u ..
udet

. . .
t
hk , 1 i k
. . . h , ik

1

2

for k < n and n = det

..
,

.

n
where the last matrix is a matrix whose rows are the coordinates of the deriva-
tives i with respect to an arbitrary positively oriented orthonormal frame.
(The matrix depends on the choice of the frame, but its determinant does
not.)
Although up to the introduction of the Frenet frame, the curvature function,

the Frenet formulae, and the computation of the curvatures everything goes
word by word as in the Euclidean space, it is not true that all the theorems
that were proved for curves in the Euclidean space are valid for curves in
Riemannian manifolds.
Consider now a hypersurface M in Rn , that is an (n 1)-dimensional sub-
manifold of Rn . Then a smooth curve : I M in M lives a double life.
It is both a curve on M and a curve in Rn . If happens to be of general
type in both spaces, then we should distinguish the Frenet frame and curva-
tures coming from the intrinsic geometry of M , from the Frenet frame and
curvatures coming from Rn .
Let us compare the first intrinsic and extrinsic curvatures of a curve in a
hypersurface. Recall that the for a curve in Rn , the first curvature function
can be defined for any regular parameterized curve : I Rn , being not
necessarily of general type in Rn . When n 3, it is simply the non-negative
function
k 0 00 k
1 = ,
k 0 k3
for n = 2, it is the signed function

x0 y0
det
x00 y 00 k 0 00 k
1 = = ,
k 0 k3 k 0 k3
where (x(t), y(t)) are the coordinates of (t) with respect to the standard
basis. The intrinsic counterpart of 1 is what we call geodesic curvature.
Definition 4.13.2. Let : I M be a regular curve in a hypersurface

M Rn (or, more generally, on a Riemannian manifold M ). If dim M 3,
then the geodesic curvature g of is the function
k 0 2 k
g = ,
k 0 k3
where is the Levi-Civita connection of M , 2 = 0 0 . If dim M = 2,

and M is not orientable, or orientable, but not oriented, then we can use the
same formula to define the geodesic curvature of .
314 4. Manifolds
However, if dim M = 2 and M is oriented, then the geodesic curvature of

is given an orientation depending sign as follows. We set
0
k (t) 2 (t)k if ( 0 (t), 2 (t)) is a positively oriented

0
, ,

k (t)k 3 basis of T(t) M
g (t) =
k 0 (t) 2 (t)k

, otherwise.
k 0 (t)k3
Proposition 4.13.3. If : I M Rn (n 3) is a regular curve on a
hypersurface M , 1 is its first curvature in Rn , g is its geodesic curvature
in M , and k(t) is the normal curvature function of M at (t), then
1 (t)2 = g (t)2 + k(t) ( 0 (t))2 .
Proof. According to equation (4.14),
00 (t) = 2 (t) + II(t) ( 0 (t), 0 (t))N(t) ,
thus,
0 (t) 00 (t) = 0 (t) 2 (t) + II(t) ( 0 (t), 0 (t)) 0 (t) N(t) .
Since N(t) is orthogonal to both 0 (t) and 2 (t), the bivector 0 (t) N(t)
is orthogonal to 0 (t) 2 (t) by formula (1.9). Thus, by the Pythagorean
theorem,
k 0 (t) 00 (t)k2 = k 0 (t) 2 (t)k2 + (II(t) ( 0 (t), 0 (t)))2 k 0 (t) N(t) k2 .

(4.21)
Using formula (1.9) again to compute k 0 (t) N(t) k2 we obtain

0 2
k (t)k 0
k 0 (t) N(t) k2 = det = k 0 (t)k2 .
0 1
Substituting back into (4.21) and dividing by k 0 (t)k6 , we get

2 2 2
k 0 (t) 00 (t)k k 0 (t) 2 (t)k II(t) ( 0 (t), 0 (t))

= + ,
k 0 (t)k3 k 0 (t)k3 k 0 (t)k2
which is exactly what we wanted to show.

Proposition 4.13.4. The following properties of a regular curve : I M
lying on a hypersurface M of Rn1 are equivalent.
(1) is a pre-geodesic curve.
(2) The geodesic curvature of is constant 0.

(3) The curvature of equals the normal curvature of the hypersurface in the
direction of 0 at each point of the curve.
(4) The acceleration vector 00 (t) is in the plane spanned by 0 (t) and the
normal vector N(t) for all t I.
(5) At each point where the osculating plane exists, it contains the normal
line of the hypersurface at that point.
(6) For any constant speed reparameterization of , the acceleration vector
field 00 is orthogonal to the hypersurface.
Proof. (1) (6). A curve is pre-geodesic if and only if any of its
constant speed reparameterizations is geodesic. is geodesic if and only
if 0 0 = 0. Since 0 0 (t) is the orthogonal projection of 00 (t) onto the
tangent space of the hypersurface at (t), 0 0 (t) = 0 if and only if 00 (t) is
orthogonal to the hypersurface at (t).
If h = , then the osculating plane of at t is the same as the osculating
plane of at h(t).
(2) (3) follows from Proposition 4.13.3.
4.13.2 Clairauts Theorem

Alexis Claude Clairaut (1713-1765) was a French mathematician, astronomer
and geophysicist. He considered great circles on a sphere centered at the
origin. He observed that if we denote by r the distance of a point P on a great
circle C from the z-axis and by the angle at which the great circle intersects
the latitudinal circle at P , then r cos() is constant along C. Clairauts
relation was later extended to geodesic curves on surfaces of revolution.
We prove below a generalization of Clairauts relation to Riemannian man-
ifolds. The theorem presented here is still not the most general version of
the assertion as it follows from a more general principle of Hamiltonian and
Lagrangian mechanics known as Noethers theorem, and saying that symme-
tries of Lagrangian or Hamiltonian mechanical systems generate invariants
of motion.
Definition 4.13.5. A vector field X on a Riemannian manifold (M, g) is
a Killing vector field if and only if LX g = 0, or equivalently, if the flow t
generated by X preserves the Riemannian metric in the sense that t (g) = g
for all t R.
Theorem 4.13.6. Let X be a Killing field on a Riemannian manifold (M, g).
Then for any geodesic curve : I M , the function t 7 hX(t) , 0 (t)i is
constant.
316 4. Manifolds
Proof. The statement is obvious if is constant. If is not constant, then

the length v = k 0 k of its speed vector is a nonzero constant.
Take two points a, b I, and consider the variation
: [, ] [a, b] M, (s, t) = s (t) = s ((t))
of , where {t }tR is the flow generated by X. The variation will be defined

if > 0 is small enough. As X is a Killing field, t is an isometry for all t,
hence s |[a,b] has the same length l(s) as |[a,b] for any |s| . Application
of the first variation formula for the length (Theorem 4.12.9) gives
0 (b) 0 (a)

0
0 = l (0) = 1 (0, b), 0 1 (0, a), 0
k (b)k k (a)k
0 (b) 0 (a)

= X(b) , X(a) , ,
v v
which proves the theorem.
Consider now a surface of revolution in R3 . If its symmetry axis is the z-axis,
then it has a parameterization of the form
r(u, v) = (x(u) cos v, x(u) sin v, z(u)),
where u 7 (x(u), 0, z(u)) is the generatrix in the (x, z)-plane. The group
of rotations about the z axis acts on the surface of revolution by isometries.
This isometry group is the flow of the vector field
rv (u, v) = (x(u) sin v, x(u) cos v, 0),
therefore, rv is a Killing vector field on the surface.

Assume now that : I M = im r, (t) = r(u(t), v(t)) is a geodesic curve
on the surface. Then the quantity
hrv (u(t), v(t)), 0 (t)i = krv (u(t), v(t))k k 0 (t)k cos((t)),
where (t) is the angle between rv (u(t), v(t)) and 0 (t), is constant. As rv
is tangent to the circles of latitude, (t) is the angle at which the geodesic
curve intersects the circle of latitude through (t). It is also clear from
the formula for rv that krv (u, v)k = |x(u)| is the distance of the point r(u, v)
from the z-axis. Since k 0 k is constant for any geodesic curve, we obtain the
following corollary of the theorem for surfaces of revolution.
Corollary 4.13.7 (Clairaut Relation for Surfaces of Revolution). If is a
non-constant geodesic curve on a surface of revolution, r(t) is the distance of
(t) from the axis of the surface, (t) is the angle at which crosses the circle
of latitude through (t), then the function t 7 r(t) cos((t)) is constant.
4.13.3 Moving Orthonormal Frames Along a Hypersur-

face
Let M Rn be a smooth hypersurface, x : M Rn be its embedding map
e1 , . . . , en : M Rn be a moving orthonormal frame along M , such that en
is the unit normal vector field of M . We consider the maps x, e1 , . . . , en to
be vector-valued functions on M . Then their differentials are vector-valued
differential 1-forms, that can be written as a linear combination of the basis
fields ei with real valued 1-forms as coefficients
n
X
dx = i ei , (4.22)
i=1
n
X
dei = ij ej . (4.23)
i=1
Proposition 4.13.8. If v Tp M is a tangent vector of M , then dxp (v) = v

and pi (v) = hv, ei i. In particular, n 0, since en (p) is orthogonal to Tp M .
Proof. If : (, ) M is a smooth curve that represents v, which means

that (0) = p and 0 (0) = v, then by the definition of the differential of
functions we have
n
X n1
X
(dx)p (v) = (x )0 (0) = 0 (0) = v = hv, ei (p)iei (p) = hv, ei (p)iei (p).
i=1 i=1
We can arrange the 1-forms ij into a matrix using i as the row index and
j as the column index. can be thought of as a matrix valued 1-form as
j
well, which assigns to a tangent vector v Tp M the matrix i,p (v) .
Proposition 4.13.9. is skew-symmetric, i.e., ij + ji 0 for all 1

i, j n.
Proof. Since the frame e1 , . . . , en is orthonormal, hei , ej i ij . Differentiat-

ing we obtain 0 = hdei , ej i + hei , dej i = ij + ji .
The one forms i and ij are not independent of one another. The identity
d d = 0 for the exterior differentiation of differential forms yields some
318 4. Manifolds
compatibility relations between them. Considering ddx we obtain

n1
X n
X
0 = ddx = d i ei i ij ej
i=1 j=1
n1 n1
! n1
!
X X X
= j
d i
ij ej i
in en ,
j=1 i=1 i=1
which gives
n1
X n1
X
d j i ij = 0 for 1 j n 1, and i in = 0. (4.24)
i=1 i=1
Vanishing of ddei gives

Xn n
X
0 = ddei = d ej ij = (dej ij + ej dij )
j=1 j=1
n
X n X
X n
= dek ik + ek jk ij .
k=1 j=1 k=1
This means that

n
X
dik + jk ij = 0, for all 1 i, j n, or in matrix form, d = 0.
j=1
4.13.4 Relation to Earlier Formulae for Parameterized

Hypersurfaces
Recall that for a regular parameterized hypersurface r : Rn , the shape
of the hypersurface and the way it is parameterized by r is encoded in the
matrix valued functions G : R(n1)(n1) and B : R(n1)(n1)
obtained form the first and second fundamental forms. From the viewpoint
of manifold theory, if r is injective, then M = r() is an injectively immersed
submanifold of Rn , the inverse of r is a chart on M . The first fundamental
form is the Riemannian metric inherited from the metric of Rn , and G is the
matrix of this Riemannian metric with respect to the chart r1 .
If ei is a moving orthonormal frame along M as above, then the first funda-
mental form is given by the fact that e1 , . . . , en1 is an orthonormal frame
in the tangent space. For two tangent vectors v, w Tp M we have
n1
X
Ip (v, w) = hv, wi = pi (v)pi (w).
i=1
Thus, the first fundamental form can be expressed by the differential forms
i .
To find the relation between the forms ij and the previously introduced
notions for parameterized hypersurfaces, let us start with the equation
X Y = X Y + II(X, Y )N,
where X and Y are tangential vector fields of M , and are the Levi-Civita
connections of Rn and M respectively, II is the second fundamental form. It
is clear that en = N, assume en is chosen to be equal to N. Applying this
formula for X = ei and Y = ej for 1 i, j n 1, we obtain
dei (ej ) = ej ei = ej ei + II(ei , ej )en .
Comparing this equation with (4.23), we get
n1
X
ej ei = ik (ej )ek ,
k=1
and
II(ei , ej ) = in (ej ). (4.25)
It is known that the Levi-Civita connection is determined by the first funda-
mental form. Thus, there should be a formula expressing ij for 1 i, j
n 1 in terms of the forms k (1 k n 1). Let us find this formula.
The forms pk form a basis in the dual space Tp M , so we can write the 1-
Pn1
forms ij as their linear combinations ij = k=1 bjik k , where the coefficients
bjik = ij (ek ) are smooth functions on M . Substituting into (4.24) we obtain
a system of linear equations
X
d j = (bjik bjki ) i k for 1 j n 1 (4.26)
1i<kn1
and X
0= (bnik bnki ) i k .
1i<kn1
The second equation is equivalent to the symmetry of the second fundamental

form
II(ei , ek ) = bnik = bnki = II(ek , ei ).
Evaluating (4.26) on the basis vectors ei and ek , taking the antisymmetry
bjki = bkji into account and permuting the role of i, j and k we obtain the
320 4. Manifolds
equations
d j (ei , ek ) = bjik + bkji ,

d k (ej , ei ) = bkji + bikj ,
d i (ek , ej ) = bikj + bjik .
Solving this system of equations for bjik , we get
1
bjik = (d i (ek , ej ) + d k (ej , ei ) d j (ei , ek )),
2
therefore,
n1
X 1
ij = (d i (ek , ej ) + d k (ej , ei ) d j (ei , ek )) k for 1 i, j n 1.
2
k=1
4.13.5 The GaussBonnet Formula

Consider a surface M R3 in R3 . Define a local orthonormal frame e1 , e2 , e3
over an open subset U of M . Define the one forms i and ij in the usual
way. Then the Levi-Civita connection of M is determined by the form
12 = 21 through the relations
X e1 = 12 (X)e2 ,
X e2 = 21 (X)e1 .
The structure equations for d12 yield the equation
d12 = 11 12 + 12 22 + 13 32 = 13 23 .
The form 13 23 is a multiple of the area form 1 2 of the surface M .

The multiplier f in the equation
13 23 = f 1 2
can be obtained by evaluating both sides on the orthonormal system e1 , e2 .

This gives
3 3
1 (e1 ) 1 (e2 )
f = 13 23 (e1 , e2 ) = det .
23 (e1 ) 23 (e2 )
According to equation (4.25), the matrix on the right-hand side is the matrix
of the second fundamental form of M with respect to the orthonormal basis
e1 , e2 , consequently, its determinant is the Gaussian curvature K of M . We

conclude that
d21 = 13 23 = K 1 2 .
Applying Stokes theorem this formula implies that fixing an orientation of
U , and a regular domain D U ,
Z Z
K 1 2 = 21 .
D D
Variants of this formula together with the geometrical interpretations of the

second integral are called local GaussBonnet formulae.
4.13.6 Steiners Formula

Steiners formula, named after the Swiss geometer Jacob Steiner (1796-1863),
asserts that for a compact convex subset K of Rn , the volume of the parallel
body
B(K, r) = {q Rn | d(q, K) r}
is a polynomial

n n n
n (B(K, r)) = W0 (K) + W1 (K)r + + Wn (K)rn
0 1 n
of degree n of r. The constant Wj (K) appearing in the coefficient of rj is a
geometrical invariant of the convex body K, called its jth quermassintegral.
n

The number Vj (K) given by the equation nj Vj (K) = nj Wnj (K),
where nj denotes the volume of the (n j)-dimensional Euclidean unit
ball, is the jth intrinsic volume of K.
For example, the constant term W0 (K) = Vn (K) is the volume of K. The
coefficient 2Vn1 (K) = nW1 (K) of r is the surface volume of the boundary
of K. The coefficient of rn1 is the average width of K multiplied with
a constant depending only on the dimension. The coefficient n V0 (K) =
Wn (K) of rn does not depend on K, it is the volume of the n-dimensional
unit ball for all K.
In this part we compute a smooth analogue of Steiners formula. A practical
formulation of the problem we want to solve is How much paint we need to
cover one side of a surface with a coat of paint of thickness r?. As the layer
of paint has positive thickness, it fills a part of space between the original
surface M and a parallel surface of M lying at distance r from M .
Definition 4.13.10. Let M be a hypersurface in Rn , with unit normal vector
field N and moving orthonormal frame e1 , . . . , en = N. For a real number
r R, we define the parallel surface Mr of M lying at distance r as the image
of the map xr : M Mr , xr = x + ren .
322 4. Manifolds
The following proposition summarizes some basic facts on the geometry of

the parallel surfaces.
Proposition 4.13.11. The differential of the map xr : M Mr has maxi-
mal rank (n 1) at p M if and only if 1/r is not a principal curvature of
M at p. If p is such a point, then
the tangent spaces Tp M and Txr (p) Mr are parallel;
the principal directions of M at p are also principal directions of Mr
at xr (p). If the principal curvatures of M and Mr are computed with
respect to the unit normal vector field en , then if the principal curvature
of M corresponding to the principal direction v Tp M is i , then the
principal curvature of Mr corresponding to the same principal direction
v Txr (p) Mr is i /(1 ri ).
Pn1
Proof. The differential of the map xr is dxr = dx+rden = i=1 ei ( i +rni ).
In a coordinate free way, dxr takes the tangent vector v Tp M to the vector
v rLp (v), where Lp is the Weingarten map of M at p M . This means
that the map dxr has maximal rank at p M if and only if 1/r is not a
principal curvature of M at p. Assuming that xr has maximal rank, the
tangent space of Mr at xr (p) is parallel to im(ITp M rLp ) = Tp M . Thus,
we can choose e1 x1 1
r , . . . , en xr as an orthonormal frame on Mr . By
the chain rule, the Weingarten map of Mr at xr (p) is d(en x1 r )(xr (p)) =
Lp (ITp M rLp )1 . This proves that if v is a principal direction of M
with principal curvature i , then v is also a principal direction of Mr with
principal curvature i /(1 ri ).
Proposition 4.13.12. If D M is a compact connected regular domain in

M such that 1/r is not a principal curvature of M at any point of D, then
the volume measure of Dr = xr (D) Mr can be expressed by the formula
n1 Z
X
n1 (Dr ) = (1)i Ki (p)dp ri ,

i=0 D
as the absolute value of a degree n 1 polynomial of r, where Ki (p) denotes

the ith elementary symmetric polynomial
K0 (p) = 1,
K1 (p) = 1 (p) + + n1 (p),
..
.
X
Ki (p) = j1 (p) j2 (p) ji (p), (4.27)
1ji <j2 <<ji n1
..
.
Kn1 (p) = 1 (p) 2 (p) n1 (p)
of the principal curvatures 1 (p), . . . , n1 (p) at p.
Proof. As the derivative of xr at p is the linear map (ITp M rLp ), the pull-
back of the volume form of Mr to M is det(ITx(.) M rLx(.) )1 n1 .
For this reason,
Z
n1 (Dr ) = | det(ITp M rLp )|d(p),
D
where the integral is taken with respect to the volume measure of M . The
integrand can be expressed with the help of the characteristic polynomial of
Lp as
n1
Y 1
1
det(ITp M rLp )) = rn1 det ITp M Lp ) = rn1 i (p)
r i=1
r
n1 n1
X 1 X
= rn1 (1)i Ki (p) = (1)i Ki (p)ri .
i=0
rn1i i=0
Since 1/r is not a principal curvature at any point of D, det(ITp M rLp )

does not vanish on D, therefore, as D is connected, it has constant sign. As
a corollary we obtain that the integral of its absolute value is the absolute
value of its integral. Combining these observations,
Z n1 Z
X
(1)i Ki (p)dp ri ,

n1 (Dr ) = det(ITp M rLp )d(p) =

D D
i=0

as we wanted to prove.
The layer of thickness r over the regular domain D M , lying between D

and Dr is parameterized by the map
h : M [0, r] Rn , h(p, s) = p + sen (p).
We claim that if D is compact and r > 0 is sufficiently small, then h is

a diffeomorphism. Using the natural decomposition T(p,s) (M [0, r]) =
Tp M Ts [0, r], the tangent space T(p,s) (M [0, r]) is spanned by the vectors
e1 (p), . . . , en1 (p) and d(s), where the vector field d on R is the unit vector
field corresponding to the derivation of functions with respect to their single
variable.
324 4. Manifolds
The images of these tangent vectors under the derivative map of h are e1 (p)
sLp (e1 (p)), . . . , en1 (p)sLp (en1 (p)) and en (p). These vectors are linearly
independent if 1/s is not a principal curvature of M at p. If M has a positive
principal curvature at a point p of D, then let
+ = max max i (p)

pD 1in1
be the maximal value of principal curvatures at all points of D, otherwise

set + = 0. From the previous arguments, if 0 < r < 1/+ , then the
derivative of h has maximal rank n at each point of D [0, r], hence h is a
local diffeomorphism by the inverse function theorem. Suppose that h is not
a diffeomorphism for any r > 0. Then, since h is a local diffeomorphism for
small values of r, the reason why it is not a diffeomorphism for these rs, is
that h is not injective. If h is not injective for the r = 1/2, 1/3, . . . , 1/k, . . . ,
then we can find two sequences of pairs (pk , sk ) 6= (pk , sk ) such that pk , pk
M , sk , sk (0, 1/k), and pk + sk en (pk ) = pk + sk en (pk ). Since D is compact,
there is a convergent subsequence of the sequence pk , say pki which tends to
p as i tends to infinity. Then the pairs (pki , ski ) and (pki , ski ) tend to the pair
(p, 0) But then (p, 0) would not have a neighborhood on which h is injective.
This contradicts that h is a local diffeomorphism.
Proposition 4.13.13. Using the above notation, if h is a diffeomorphism

the volume of the layer of thickness r over D parameterized by h is equal to
the polynomial
n
(1)k1
X Z
Kk1 (p)d(p) rk .
k D
k=1
Proof. Introduce on M [0, r] the Riemannian metric, in which the basis

e1 (p), . . . , en1 (p), d(s) is orthonormal. The pull-back of the volume form of
Rn by h is the form
det(ITp M sLp ))d 1 d n1 ds,
where the sign depends on the orientations chosen. Thus,

[ Z r Z
n Ds = | det(ITp M sLp )|d(p)ds.
s[0,r] 0 D
As h is diffeomorphism, it does not vanish on D [0, r]. It is also clear that

det(ITp M sLp ) = 1 for s = 0, therefore det(ITp M sLp ) > 0 on D [0, r],
and the absolute value can be omitted. We conclude

[ Z r n1
X Z !
i
n Ds = (1) Ki (p)dp si ds
s[0,r] 0 i=0 D
n
(1)k1
X Z
= Kk1 (p)d(p) rk ,
k D
k=1
as it was to be proven.
Exercise 4.13.14. Prove that the above proposition implies Steiners for-
mula for the volume of parallel bodies of a convex body K, the boundary
M = K of which is a smooth hypersurface in Rn . Show that we get the
following explicit expression for the coefficients of the polynomial:
n Z
X 1
n (B(K, r)) = n (K) + |Kk1 (p)|d(p) rk .
k D
k=1
Exercise 4.13.15. Show that if K is a convex body in Rn bounded by

a smooth hypersurface M with GaussKronecker curvature function K =
Kn1 , then Z
|K(p)|d(p) = nn ,
M
where n is the volume of the n-dimensional unit ball (consequently, nn is

the surface volume of the unit sphere Sn1 in Rn ).
4.13.7 Minkowskis Formula

Being important geometrical invariants of a convex body, the coefficients of
the polynomial in Steiners formula are studied thoroughly. In this part, we
focus on the coefficient rn1 , which is, up to some constant multiplier, the
integral of the (n 1)st elementary symmetric polynomial of the principal
curvatures of the boundary hypersurface of K. Hermann Minkowski (1864-
1909) proved that if K is a compact convex set in Rn , then this coefficient is
the average width of K up to some constant, depending only on the dimension
n.
We prove below a formula which is true for any compact hypersurface M in
Rn which, in the special case when M is the boundary of a convex set reduces
to Minkowskis formula.
Proposition 4.13.16 (Minkowkis Formula). Let M be a smooth compact
hypersurface in Rn , en : M Rn be a unit normal vector field on M . Let
pn : M R, pn (p) = hen , pi denote the signed distance of the origin from
326 4. Manifolds
the affine tangent space of M at p. Then denoting by the surface volume

measure on M , we have
1
Z Z
Kn2 d = pn Kn1 d,
n1 M M
where the elementary symmetric polynomials Ki of the principal curvatures of

M taken with respect to the unit normal vector field en are defined by (4.27).
Proof. Let x : M Rn be the inclusion map, and choose a positively oriented

orthonormal frame e1 , . . . , en over an open subset U of M and orient M so
that the vector fields (e1 , . . . , en1 ) give positively oriented bases at points of
U . Consider the functions pi : U R, pi (p) = hei , pi. Using the fundamental
equations of hypersurfaces, we can express the differentials of the functions
pi by the equations
* n + * n1
+
X X
dpi = hdei , xi + hei , dxi = ej ij , x + ei , ej j
j=1 j=1

Pn p j + i , if i n 1,
j=1 j i
= P n j
j=1 pj n , if i = n.

Define the differential (n 2)-form on U as
n1
X
= (1)i+1 pi 1n n
cin n1 ,
i=1
where the hat above in means that in is omitted. Compute the differential
of . Clearly,
n1
X
d = (1)i+1 dpi 1n n
cin n1 +
i=1
n1
X X
+ (1)i+j pi 1n djn n
cin n1 +
i=1 1j<i
n1
X X
+ (1)i+j1 pi 1n
cin djn n1
n
,
i=1 i<jn1
hence
n1
XX n
d = (1)i+1 pj ij 1n n
cin n1 +
i=1 j=1
n1
X
+ (1)i+1 i 1n n
cin n1 +
i=1
n1
X n
X X
+ (1)i+j pi 1n (jk kn ) n
cin n1 +
i=1 1j<i k=1
n1
X X n
X
+ (1)i+j1 pi 1n
cin (jk kn ) n1
n
.
i=1 i<jn1 k=1
The sum of the last two terms can be simplified as

n1
X X
(1)i+j pi 1n (ji in ) n
cin n1 +
i=1 1j<i
n1
X X
+ (1)i+j1 pi 1n
cin (ji in ) n1
n
i=1 i<jn1
n1
X X
= (1)(i+j)+(j1)+(ij1) pi ji 1n n
cjn n1 +
i=1 1j<i
n1
X X
+ (1)(i+j1)+(j2)+(ji1) pi ji 1n n
cjn n1
i=1 i<jn1
X n1
n1 X
= (1)j pi ji 1n n
cjn n1 .
i=1 j=1
Substituting back into the last formula for d we obtain

n1
X
d =pn (1)i+1 in 1n n
cin n1 +
i=1
n1
X
+ (1)i+1 i 1n n
cin n1
i=1
n1
X
=(n 1)pn 1n n1
n
+ (1)i+1 i 1n n
cin n1 .
i=1
(4.28)
328 4. Manifolds
Pn1
The forms , pn 1n n1 n
and i=1 (1)
i+1 i
1n
cin
n
n1 are natural forms on M in the sense that they do not depend
on the choice of the orthonormal frame. Indeed, if v1 , . . . , vn2 Tp M ,
then p (v1 , . . . , vn2 ) is the signed volume of the parallelepiped spanned
Pn1
by the orthogonal projection (p) = i=1 pi (p)ei (p) of p onto Tp M and
Lp (v1 ), . . . , Lp (vn2 ), where Lp is the Weingarten map at p. The differential
form pn 1n n1 n
is the volume form of M multiplied by the product of
the GaussKronecker curvature Kn1 and the signed distance pn of the origin
from the tangent plane. Naturality of the third form follows from equation
(4.28) and naturality of the first two forms. Computing the third form using
a frame, whose vectors point in principal directions at a point, we see easily
that
n1
X
(1)i+1 i 1n n
cin n1 = Kn2 1 n1 .
i=1
The importance of naturality is that although the forms , pn 1n n1

n
Pn1
and i=1 (1)i+1 i 1n n
cin n1 were defined locally, on an
open set, on which a local orthonormal frame e1 , . . . , en can be defined, these
forms are globally defined on M and relation (4.28) also holds globally on M .
If M is a compact hypersurface (with no boundary), then by Stokes theorem
Z Z
d = = 0,
M M
thus, integrating (4.28) we obtain the formula
1
Z Z
Kn2 d = pn Kn1 d
n1 M M
we wanted to show.
Exercise 4.13.17. Prove that for a compact hypersurface M Rn ,

Z
en Kn1 d = 0.
M
Hint: Apply Minkowskis formula with different choices of the origin of the
coordinate system.
In the special case, when M is strictly convex and the Gauss map en : M
Sn1 is a diffeomorphism between M and Sn1 , the second integral can be
rewritten as an integral on the sphere.
Since the derivative map Tp en : Tp M Ten (p) Sn1 is Lp , the pull-back of

the volume form of Sn1 by en is (1)n1 Kn1 1 n1 . Consequently,
Z Z
(1)n pn Kn1 d = pn (e1
n (u))du
M Sn1
Z
1
= (pn (e1 1
n (u)) + pn (en (u)))du.
2 Sn1
Definition 4.13.18. The width wK (u) of a compact set K Rn in the

direction u Sn1 is the width of the narrowest slab bounded by hyperplanes
orthogonal to u that contains K. The width is given by the formula
wK (u) = maxhu, xi min hu, xi.

xK xK
Definition 4.13.19. The mean width w(K) of a compact set K Rn is the

average of its widths in different directions, that is
Z
1
w(K) = wK (u)du.
nn Sn1
If K is a compact convex regular domain with boundary M = K, and

en : M Rn is the exterior unit normal vector field along M , then for
u Sn1 , hu, xi attains its maximum (or minimum respectively) on K at
the points x M of the boundary, at which en (x) = u (or en (x) = u re-
spectively). At such a point hu, xi = pn (x) (or hu, xi = pn (x) respectively).
In the special case, when the Gauss map is a diffeomorphism between M and
Sn1 , we get
Z Z
(pn (e1
n (u)) + pn (e1
n (u)))du. = wK (u)du = nn w(K).
Sn1 Sn1
Thus, we get the following corollary of Proposition 4.13.16, also known as

Minkowskis Formula.
Corollary 4.13.20 (Minkowskis Formula on the Mean Width). Let K Rn

be a compact convex regular domain with boundary M = K, such that the
exterior unit normal vector field en : M Rn along M is a diffeomorphism
between M and Sn1 . Then the coefficient of rn1 in the polynomial express-
ing the volume of B(K, r) is equal to the following quantities
(1)n2
Z
nn
nWn1 (K) = n1 V1 (K) = Kn2 d = w(K).
n1 M 2
330 4. Manifolds
4.13.8 Rigidity of Convex Surfaces

Cauchys rigidity theorem for convex polytopes says that the shape of a 3-
dimensional convex polytope is uniquely determined by the shapes of the
facets and the combinatorial structure describing which are the common edges
of the neighboring facets. More formally, if we have two convex 3-dimensional
polytopes P1 and P2 and a bijection : P1 P2 , which maps each facet
of P1 isometrically onto a facet of P2 , then extends to an isometry of the
whole space R3 .
We want to prove a smooth analogue of Cauchys rigidity theorem here. In-
stead of two convex polytopes, we shall consider two convex compact regular
domains, K1 and K2 in R3 , and require that the bijection : K1 K2
between their boundaries be a bending. This means that should preserve
the lengths of curves lying on the boundary of K1 .
Let M be a smooth hypersurface in Rn , e1 , . . . , en be a local orthonormal
frame on an open subset U M , h : M M be a bending of M . Then
the images of e1 , . . . , en1 under the derivative of h yield an orthonormal
tangential frame e1 , . . . , en1 along h(U ), which can be extended uniquely to
an orthonormal frame e1 , . . . , en having the same orientation as e1 , . . . , en .
Using this frame on M , we can introduce the differential 1-forms i , ij and
the functions pi , pi in the usual way. Instead of working with these differential
forms and functions, we prefer to work with their pull-back forms i = h i ,
ij = h ij and the functions pi = pi h. Since h is an isometry of Riemannian
manifolds, i = i . The Riemannian metric also determines the Levi-Civita
connection, which is encoded by the one forms ij for 1 i, j n 1, so we
also have ij = ij for 1 i, j n 1.
As the following proposition claims, the exterior shapes of U and h(U ) are
determined by the forms in and in .
Proposition 4.13.21. If U is connected, then h|U extends to an orientation-

preserving isometry of the whole space if and only if in = in for 1 i n1.
Proof. The easier part of the statement is that if the bending extends to an
n
orientation preserving isometry of the space, then in = 0 i . We leave the
details of this direction to the reader.
Assume now that in = in for all 1 i n 1. Choose a point p U
and an orientation preserving isometry of Rn such that (p) = h(p) and
Tp (ei (p)) = ei (h(p)) for 1 i n 1. Then h = 1 h is a bending of
M which fixes p and the frame e1 (p), . . . , en1 (p). Denote by e1 , . . . , en the
frame induced on h(U ) in the same way as the frame e1 , . . . , en was obtained
on h(U ).
Let : [0, 1] U be a smooth curve starting from (0) = p. Consider the
2
collection of the vector valued functions (e1 , . . . , en ) : [0, 1] Rn . It
satisfies the system of differential equations

n
d X j
(ei )( ) = i ( 0 ( ))(ej )( ), (1 i n).
dt j=1
The same system of differential equations is satisfied also by the vector valued
2
function (e1 h , . . . , en h ) : [0, 1] Rn . Both solutions start from
the same initial vectors at 0, therefore, by the uniqueness of the solutions of
ordinary differential equations with given initial value, ei = ei h for
i = 1, . . . , n. From this we obtain
n1
X n1
X
0 ( ) = i ( 0 ( ))ei (( )) = i ( 0 ( ))ei (h(( ))) = (h )0 ( ),
i=1 i=1
which implies by (0) = h((0)) = p that = h . Since U is connected,

(1) can be any point in U , consequently h1 h = idM , that is, extends
h|U to an isometry of the whole space.
The main tool of proving the rigidity of convex surfaces in R3 is a gener-

alization of Minkowskis formula due to Gustav Herglotz (1881-1953). This
formula has the same form as Minkowskis formula, but some of the ingre-
dients of the formula are taken not from the hypersurface M but from a
bending M of M . This way, Herglotzs formula relates the geometries of
M and M to each other, and in the special case, when M = M , it returns
Minkowskis formula.
Proposition 4.13.22 (Herglotzs Formula). Let M be a compact hypersur-

face in Rn , h : M Rn be a bending, h(M ) = M . Starting with a lo-
cal orthonormal frame e1 , . . . , en over an open subset U of M define the
j j
forms
Pi , i and the functions Kn2 , pi and pi as above. Then the forms
n1 n n n 1 n1
pn i=1 1 i n1 and Kn2 do not de-
pend on the choice of the frame, consequently they are properly defined over
M , and
Z n1
! Z
X
pn 1n in n
n1 = Kn2 1 n1 .
M i=1 M
Proof. The proof follows the same line of computation as the proof of Min-
kowskis formula, the only difference is that some terms are marked with a
bar to indicate that they come from the bent surface M . Due to the similarity
of the two proofs, we skip some details. Define the differential (n 2)-form
332 4. Manifolds
on M by the equation
n1
X
= (1)i+1 pi 1n n
cin n1 ,
i=1
and compute its differential as

n1
X
d = (1)i+1 dpi 1n n
cin n1 +
i=1
n1
X X
+ (1)i+j pi 1n djn n
cin n1 +
i=1 1j<i
n1
X X
+ (1)i+j1 pi 1n
cin djn n1
n
,
i=1 i<jn1
hence
n1
XX n
d = (1)i+1 pj ij 1n n
cin n1 +
i=1 j=1
n1
X
+ (1)i+1 i 1n n
cin n1 +
i=1
n1
X n
X X
+ (1)i+j pi 1n (jk kn ) n
cin n1 +
i=1 1j<i k=1
n1
X X n
X
+ (1)i+j1 pi 1n
cin (jk kn ) n1
n
i=1 i<jn1 k=1
The sum of the last two terms can be simplified to

n1
X n1
X
(1)j pi ji 1n n
cjn n1 .
i=1 j=1
Substituting back we obtain

n1
!
X
d = pn 1n in n
n1 + Kn2 1 n1 . (4.29)
i=1
Pn1
The forms , pn i=1 1n in n1
n
and Kn2 1 n1 do not
depend on the choice of the orthonormal frame e1 , . . . , en1 , hence they are
defined globally. For example, if v1 , . . . , vn2 Tp M , then p (v1 , . . . , vn2 )

is the signed volume of the parallelepiped spanned by (Tp h)1 ((h(p))) and
Lp (v1 ), . . . , Lp (vn2 ), where (h(p)) is the orthogonal projection of h(p) onto
Th(p) M , Lp is the Weingarten map at p. Naturality of the third form was
obtained during the proof of the Minkowski formula, that of the second form
follows form the relation (4.29) between the three forms.
If M is a compact hypersurface with no boundary, then integrating (4.29) we
obtain Herglotzs formula.
Theorem 4.13.23. Let M and M be the boundary surfaces of the compact

convex regular domains C and C in R3 . Assume that M has positive Gaussian
curvature K2 > 0 and that there is a bending h : M M . Then h extends
to an isometry of the whole space. In particular, C and C are congruent.
Proof. Orient M and M so that augmenting a positively oriented basis of

one of the tangent spaces by the exterior unit normal of C or C respectively
give a positively oriented basis of the space R3 . We may assume without loss
of generality the h : M M is orientation preserving, otherwise consider
instead of h a composition of h with a reflection in a plane.
Let us use the same notations as in Minkowskis and Herglotzs formulae.
Translating C and C we may assume that the origin is in the interior of C
and C. Then p3 > 0 and p3 > 0.
The integral of the Minkowski curvature H = K1 /2 of M over M can be
expressed both by Minkowskis and Herglotzs formulae as
Z Z Z
1
Hd = p3 13 23 = p3 (13 23 + 13 23 ).
M M 2 M
Changing the role of M and M yields

Z Z Z
1
Hd = p3 13 23 = p3 (13 23 + 13 23 ).
M M 2 M
Taking into account that M and M has the same Gaussian curvature at
corresponding points by Theorema Egregium, we get
13 23 = K2 1 2 = 13 23 ,
and Z Z Z
1
Hd Hd = p3 (13 13 ) (23 23 ).
M M 2 M
The differential form (13 13 ) (23 23 ) can be written as a multiple
f 1 2 = (13 13 ) (23 23 )
334 4. Manifolds
of the volume form 1 2 . We show that the function f is non-positive.

Fix a point q M and choose a positively oriented orthonormal basis v1 , v2
in Tq M then
13 (v1 ) 13 (v2 )
B=
23 (v1 ) 23 (v2 )
is the matrix of the second fundamental form and the Weingarten map of M
at q with respect to the basis (v1 , v2 ). For this reason, B is negative definite,
and has determinant K2 (q). Similarly,

13 (v1 ) 13 (v2 )
B =
23 (v1 ) 23 (v2 )
is the matrix of the second fundamental form and the Weingarten map of M
at h(q) with respect to the basis (Tq h(v1 ), Tq h(v2 )), and it is also negative
definite, and has determinant K2 (q).
Consider the quadratic polynomial
D(c) = det(B cB) = K2 (q)(c2 tr(B B 1 )c + 1).
Since B is negative definite, BcB is positive definite for large c, consequently,

D must have a positive root c0 > 0. Since the product of the two roots is 1,
the second root is 1/c0 . Then D(c) = K2 (q)(c c0 )(c 1/c0 ) and
(1 c0 )2
f (q) = D(1) = K2 (q) 0.
c0
Equality f (q) = 0 holds if and only if tr(B B 1 ) = 2. But we also know that
det(B B 1 ) = 1, hence f (q) = 0 if and only if both eigenvalues of B B 1 are
equal to 1. Since both B and B are symmetric, and B is negative definite,
B B 1 is diagonalizable because of the principal axis theorem. This way, the
eigenvalues of B B 1 are equal to 1 if and only if B B 1 = I and B = B.
As a corollary, we obtain
Z Z Z Z
1 3 3 3 3 1
Hd Hd = p3 (1 1 ) (2 2 ) = p3 f 1 2 0,
M M 2 M 2 M
which implies Z Z
Hd Hd.
M M
However, the role of M and M is symmetric, so the reversed inequality

Z Z
Hd Hd
M M
must also be true, therefore we must have equality in both inequalities. Equal-
ity implies f 0, which can hold only if the matrices B and B are equal for
each q M and each choice of v1 v2 Tq M , that is, if 13 = 13 and 23 = 23 .
However this condition is equivalent to the extendability of h to an orientation
preserving isometry of the space R3 .
Bibliography
[1] W. Blaschke, H. Reichardt, Einfuhrung in die Differentalgeometrie.

Zweite aufl., Springer-Verlag Berlin-Gottingen-Heidelberg, 1960.
[2] W.M. Boothby, An Introduction to Differentiable Manifolds and Rieman-
nian Geometry. Academic Press, Inc. Orlando 1986.
[3] M.P. do Carmo, Differential Geometry of Curves and Surfaces. Prentice-

Hall, Inc., Englewood Cliffs New Jersey 1976.
[4] M.P. do Carmo, Riemannian Geometry. Birkhauser Boston 1992.
[5] D.L. Cohn, Measure Theory. Birkhauser, Boston-Basel-Stuttgart, 1980.
[6] A. Hatcher, Algebraic Topology. Cambridge University Press 2002.

[7] S.Mac Lane, Categories for the Working Mathematician. Graduate Texts
in Mathematics, Springer Verlag New York Heidelberg Berlin 1998.
[8] J.M. Lee, Riemannian Manifolds: An Introduction to Curvature. Gradu-
ate Texts in Mathematics, Springer Verlag New York 1997.
[9] J.R. Munkres, Topology. 2nd ed. Prentice Hall, Inc. Upper Saddle River,
2000.
[10] J.J. Stoker, Differential Geometry. Wiley Classics Library, John Wiley
& Sons, Inc. 1989.
[11] F.W. Warner, Foundations of Differentiable Manifolds and Lie Groups.
Graduate Texts in Mathematics, Springer-Verlag Berlin-Heidelberg-New
York 1983.
337
Index
adjoint map, 7 basis, 4

affine BeltramiEnneper theorem, 181
combination, 37 bending of a hypersurface, 157, 330
coordinate system, 39 Bernoullis lemniscate, 129
flag, 97 Bianchi
Grassmann manifold, 35 identity
independence, 38 first , 161, 277
space, 34 second , 279
subspace, 35 tensor, 281
spanned by a set, 38 transform, 191
transformation, 34 bilinear map, 11
affine connection, see connection bivector, 15
algebraic curvature tensor, 281 Borel
algebraic plane curve, 82 algebra, 68
algebraic Weyl tensors, 293 measure, 69
alternating inner regular , 69
k-linear map, 15 outer regular , 69
k-form, 20 regular , 69
tensor, 17 set, 68
angle boundary of a singular chain, 255
between affine subspaces, 43 boundary point, 50
between linear subspaces, 43 bump function, 66
enclosed by two vectors, 22
anti-discrete topology, 50 cardioid, 123
astroid, 107 Cartans formula, 249
asymptotic Cartans theorem, 208
curve, 180 Cartesian coordinate system, 43
direction, 179 category, 1
atlas, 201 of affine spaces, 34
automorphism, 2 of bilinear maps, 12
autoparallel curve, 305 of C r -manifolds, 209
of k-linear maps, 15
Backlund transform, 191 of linear spaces, 4
base space of a bundle, 233 of pairs of linear spaces, 14
338
Index 339
of pointed topological spaces, 60 conformally equivalent metrics, 298

of sets, 2 connection, 267
catenoid, 179, 198 at a point, 267
Cauchy sequence, 59 compatible with a Riemannian met-
Cauchys rigidity theorem, 329 ric, 272
CauchySchwarz inequality, 21 symmetric , 270
center of curvature, 118 torsion free , 270
chain curve, 123 continuity, 54
chain rule, 63, 65, 219 at a point, 54
characteristic polynomial, 10 contravariant functor, 2
Chebyshev net, 185 convergent sequence, 54
Christoffel symbols convex
of a connection, 268 plane curve, 127
of a hypersurface, 153 coordinate
circle of latitude, 163 line, 141, 169
Clairauts relation, 316 surfaces, 169
closed coordinate system
ball, 50 affine , 39
differential form, 251 Cartesian , 43
subset, 50 isothermal local , 304
closed curve, 125 local , 201
simple , 125 local on R3 , 169
closure of a set, 50 normal , 307
CodazziMainardi equations, 155 polar , 85
cofinite topology, 57 coordinates
commutative diagram, 2 of a point, 201
commutator of linear endomorphisms, of a tangent vector, 214
222 of a tensor field, 159
commuting vector fields, 225 of a vector, 5
complete of a vector field, 220
measure space, 68 of tensors, 13
metric space, 59 Cotton tensor, 304
surface, 188 countability axioms, 58
completion of a measure space, 68 covariant derivation
complexification of a linear space, 10 of tensor fields, 278
components, see coordinates of vector fields, 267
concatenation of loops, 60 along a curve, 269
cone, 151 along a parameterized surface,
over a curve, 174 271
configuration space, 205 covariant functor, 2
confocal second order surfaces, 171 covering
conformal Weyl tensor field, 295 k-fold , 62
340 Index
map, 61 difference quotients of higher order,

universal , 62 95
C r -compatible differential form, 236
atlas, 201 differential of a function, 245
charts, 201 dimension
C r -diffeomorphic manifolds, 203 of a linear space, 4
C r -diffeomorphism, 203 of an affine space, 34
C r -manifold, 203 dimension formula, 7
critical point of a function, 148 direction
Crofton formula space of an affine subspace, 35
for spherical curves, 90 vectors of an affine subspace, 35
in the plane, 87 directrix, 173
in the space, 89 discrete topology, 50
cross product, 31 distance
curvature in a Euclidean space, 40
functions of a curve, 100 in a metric space, 40
curvature tensor preserving map, 41
of a connection, 275 distribution
curve of general type, 97 integral manifold of a , 229
cycloid, 85 involutive , 229
cylinder over a curve, 174 divided differences of higher order, 95
dot product, 21
de Rham cohomology dual
algebra, 251 basis, 7
functor, 252 space, 7
spaces, 251 Dupins theorem, 170
degree of a covering map, 62
degree of a map on differential forms, eigenvalue, 10
243 eigenvector, 10
dense subset, 50 Einstein manifold, 295
derivation of functions, 222 Einsteins convention, 5
at a point, 214 ellipsoid, 150
derivative elliptic paraboloid, 151
directional , 63 elliptical helix, 107
of vector fields along a hyper- endomorphism, 2
surface, 145 equation
of a map between manifolds, 218 of a k-sphere, 48
of a multivariable function, 62 of a linear hyperplane, 44
partial , 64 of a linear subspace, 44
determinant of an affine subspace, 45
of a matrix, 8 Euclidean
of linear endomorphisms, 9 linear space, 21
Index 341
space, 40 Gauss frame, 152

Eulers formula, 149 Gauss lemma, 309
evolute, 119 GaussBonnet formula, 320
exact differential form, 251 GaussKronecker curvature, 149
exponential map, 306 general linear group, 4
exterior generatrix
algebra, 17 of a ruled surface, 173
powers of a linear space, 15 of a surface of revolution, 163
product, 15 geodesic curvature, 313
exterior differentiation, 245 geodesic curve, 305
exterior point, 50 geodesically complete Riemannian man-
extreme value theorem, 59 ifold, 307
gradient vector field, 64, 298
f -related vector fields, 226 Gram matrix, 24
factor GramSchmidt orthogonalization, 23
map, 7 Grassmann
space algebra, 17
for linear spaces, 6 manifold, 6, 203
topology, 53 topology of , 51
FaryMilnor theorem, 139
Fenchels theorem, 136 Hausdorff space, 57
fiber of a bundle, 233 Hazzidakis formula, 188
first fundamental form, 146 HeineBorel theorem, 59
first variation helicoid, 179
of the surface volume, 195 helix, 86
flag, 97 Herglotz formula, 331
flow Hesse form, 299
of a vector field, 76, 221 Hilberts non-immersion theorem, 189
of an o.d.e., 76 Hodge star operator, 30
four vertex theorem, 130 homeomorphism, 54
Frenet homotopic
formulae, 100 loops, 60
frame, 98 maps
distinguished , 98 piecewise smothly , 252
Frobenius theorem, 78, 230 homotopy equivalence, 254
Fubinis theorem, 71 piecewise smooth , 252
fundamental group, 60 homoty class of loops, 60
fundamental theorem hyperbolic paraboloid, 151
of curve theory, 105 hyperplane, 35
of hypersurface theory, 155 hypersphere, 46
of Riemannian geometry, 273 equations, 46
Gauss equations, 155, 289 image of a linear map, 6

342 Index
indicator function, 70 Klein bottle, 212

induced Euclidean structure knot, 139
on dual space, 29 trivial , 139
on exterior powers, 28 Koszuls formula, 274
on tensor spaces, 2729 Kronecker delta symbol, 7
infinitesimal variation KulkarniNomizu product, 291
of a hypersurface, 195
initial object of a category, 12 Lagrange identity, 32
inner product, 21 Lebesgue measure, 69
integral length
of a cohomology class over a ho- of a curve, 83
mology class, 259 of a vector, 21
of a form over a chain, 256 of the evolute, 121
of a form over a regular domain, Levi-Civita connection, 273
263 Lie
of a measurable function, 70 algebra, 222
integral curve of a Lie group, 227
of a vector field, 220 bracket of vector fields, 223
maximal , 220 derivative, 238
of an o.d.e., 75 group, 204
maximal , 76 line of curvature, 166
integral manifold, 229 linear
integration combination, 4
by change of variables, 71, 72 0-weight , 36
in spherical coordinates, 73 affine , 37
interior point, 50 hull, 6
interior product, 243 independence, 4
intrinsic volume, 321 isomorphism, 4
inverse function theorem, 65 map, 4
involute, 122 space, 3
involutive distribution, 229 subspace, 6
isometry, 41 generated by a set, 6
group, 41 spanned by a set, 6
isomorphic loop, 60
linear spaces, 4
objects, 2 mapping of class C r , 203
isomorphism, 2 matrix
of a linear map, 5
Jacobi identity, 32 of a bilinear function, 25
Jacobian matrix, 63 mean curvature, 149
mean width of a compact set, 329
kernel of a linear map, 6 measurable
Killing vector field, 315 function, 70
Index 343
measurable set, 68 open

measure, 68 ball, 50
product , 69 subsets, 49
space, 68 open subset
complete , 68 of a metric space, 50
completion of a , 68 opposite category, 3
meridian, 163 ordinary differential equation, 75
metric linear , 76
topology, 50 orientation
metric space, 40 of a linear space, 11
complete , 59 of a manifold, 259
totally bounded , 59 preserving isometry, 41
Meusniers theorem, 145 preserving reparameterization, 83
minimal hypersurface, 198 orthogonal
Minkowski curvature, 149 linear transformation, 41
Minkowskis formula, 325, 329 matrix, 41, 106
moment curve, 105 orthogonal vectors, 22
morphisms of a category, 1 orthonormal
motion of a Euclidean space, 41 system, 22
moving frame orthonormal frame
along a curve, 98 along a curve, 98
along a hypersurface, 317 osculating
multiplicity of an eigenvalue, 10 circle, 118
flag, 97
natural
k-plane, 94
isomorphism, 3
k-sphere, 113
between linear spaces, 14
reparameterization, 85
transformation, 3 parallel
neighborhood of a point, 50 curves of a curve, 120
norm of a vector, 21 surfaces of a hypersurface, 321
normal surfaces of a surface, 179
coordinate system, 307 transport
curvature, 144 along a curve, 270
plane of a curve, 131 for affine spaces, 40
section of a hypersurface, 144 vector field along a curve, 270
vector parameter line, 141
of a hyperplane, 44 parameterization
of a hypersurface, 142 local of a set, 201
nowhere dense subset, 50 of a linear subspace, 44
of a simple arc, 81
objects of a category, 1 of a subset, 83
one-sheeted hyperboloid, 150 of an affine subspace, 44
344 Index
parameterized reparameterization, 83
curve, 82 by arc length, 85
hypersurface, 141 Ricci
permutation group, 2 curvature, 287
permutation rule for mixed product, tensor, 286, 292
32 RiemannChristoffel curvature tensor,
Plucker embedding, 16 280
Poincare lemma, 253 Riemannian
Poincares half-plane model of the hy- curvature tensor
perbolic plane, 160 of a hypersurface, 159
pointed topological space, 60 of a Riemannian manifold, 280
polar coordinates, 85 manifold, 264
position vector, 35 metric, 265
positive definite symmetric bilinear rotation number, 125
function, 21 theorem, 125
positively oriented basis, 11
pre-geodesic curve, 306 Sards lemma, 72
principal scalar curvature, 288, 294
curvature, 148 Schurs theorem, 285
principal axis theorem, 27 second fundamental form, 146
principal direction, 148 section of a bundle, 235
product sectional curvature, 283
measure, 69 self-adjoint linear endomorphism, 26
of manifolds, 203 separation axioms, 5658, 58
topology, 54 shape operator, 146
projection of a bundle, 233 shpere
projective plane, 210 in Euclidean space, 46
pseudosphere, 165 -algebra, 68
quadratic form of a bilinear function, sign
26 of a permutation, 8
quermassintegral, 321 simple arc, 81
sine-Gordon equation, 187
radius of curvature, 118 singular point
rank of a linear map, 6 of a curve, 121
rectifiable curve, 83 smooth singular
rectifying plane, 131 chain, 254
regular homology space, 256
domain, 260 simplex, 254
parameterized curve, 85 Sorgenfrey plane, 58
parameterized hypersurface, 142 space form, 284
reparameterization, 83 spaces of constant curvature, 284
topological space, 57 special orthogonal
Index 345
group, 106 of a manifold, 217

matrix, 106 of an affine space, 40
speed vector of a curve in a manifold, indicatrix, 137
214 of a curve, 94
sphere space
equations, 46 of a hypersurface, 142
exotic , 210 of a manifold, 214, 215
with g handles, 210 of a submanifold, 212
with standard smooth structure, of an affine space, 40
207 surface of a curve, 175
standard vector
basis of Rn , 6 of a manifold, 214, 215
dot product on Rn , 21 of a submanifold, 212
model of the Euclidean space, 43 of an affine space, 40
orientation of Rn , 11 tantrix, 137
simplex, 254 tensor bundle, 235
topology of an affine space, 51 tensor field of type (k, l)
Steiners formula, 321 over a hypersurface, 159
step function, 70 over a manifold, 235
Stokes theorem, 256 tensor product
submanifold of Rn , 207 of linear maps, 12
subspace distribution, see distribution of linear spaces, 11
subspace topology, 52 tensors
superderivation, 243 of type (k, l), 13
support of a differential form, 261 Theorema Egregium, 156
surface topological manifold, 202
of constant negative curvature, topological space, 49
183 compact , 58
area, 161 connected , 61
developable , 174 first-countable , 58
minimal , 198 Hausdorff , 57
non-orientable compact , 210 locally path-connected , 61
of revolution, 162 normal , 57
orientable compact , 210 path-connected , 61
ruled , 173 regular , 57
volume, 161 second-countable , 58
symmetric semilocally simply-connected ,
k-linear function, 21 61
tensor, 17 separable , 58
sequentially compact , 59
tangent simply connected , 61
bundle T0 -space, 57
346 Index
T1 -space, 57 along a hypersurface, 145

T2 -space, 57 tangential , 145
T3 -space, 57 unit normal , 142
T4 -space, 57 along a parameterized surface, 271
topology, 50 left invariant , 227
torus, 150 over Rn , 75
total over a manifold, 219
curvature vector space, 3
of planar curves, 124 vectorization of an affine space, 35
of space curves, 135 vertex of a curve, 129
differential equation, 77 volume form of an oriented Rieman-
total space of a bundle, 233 nian manifold, 265
totally bounded metric space, 59 volume measure of a Riemannian man-
trace ifold, 265
of a linear endomorphism, 11
of a matrix, 10 wedge product, 15
of a parameterized curve, 83 Weierstrass theorem, 59
tractrix, 124, 165 Weingarten map, 146
transit map, 201 Weyl tensor, 293
translation, 6 Weyl tensor field, 295
triangle inequality, 40 Whitneys theorem, 207
triple product expansion formula, 32 width of a compact set, 329
triply orthogonal system, 170
trivector, 15 Youngs theorem, 65
umbilical point, 151

planar , 151
spherical , 151
Umlaufsatz, 125
uniform continuity, 59
unit speed reparameterization, 85
variation
of a curve, 308
of a hypersurface, 195
compactly supported , 195
vector bundle, 233
vector field
along a curve, 98, 269
binormal , 131
principal normal , 131
unit normal in plane, 117
unit tangent , 117, 131

01 Csikos Differential Geometry

Uploaded by

Document Informationclick to expand document informationGeometry

Document Informationclick to expand document information

Copyright:

Available Formats

01 Csikos Differential Geometry

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

01 Csikos Differential Geometry

Uploaded by

Copyright:

Available Formats

DIFFERENTIAL GEOMETRY

Series of Lecture Notes and Workbooks for Teaching

Eotvos Lorand University

Referee: Arpad Kurusa

Creative Commons NonCommercial-NoDerivs 3.0 (CC BY-NC-ND 3.0)

ISBN 978 963 279 221 7

Made within the framework of the project Nr. TAMOP-4.1.2-08/2/A/KMR-

KEY WORDS: Curve, Frenet frame, curvature, torsion, hypersurface, funda-

SUMMARY: The aim of this textbook is to give an introduction to differ-

1.1 Categories and Functors

(ii) There is a (unique) morphism 1X MorC (X, X) for any object X,

MorC (X, Y ) and g MorC (Y, X), the identities

F (1X ) = 1F (X) for any object X ObC ;

F (g f ) = F (f ) F (g) for any f MorC (X, Y ) and g MorC (Y, Z).

To each category C, there is an opposite category C op . This category has the

is commutative, i.e., Y F (f ) = G(f ) X . A natural isomorphism between

1.2 Linear Algebra

of elements of V with is defined in such a way that the following identities

(ii) 0 V such that x + 0 = x for all x V ;

(iii) x V x V such that x + (x) = 0;

(vii) ()x = (x);

Definition 1.2.2. A linear combination of some vectors x1 , . . . , xk V is a

Definition 1.2.4. A map L : V W between the linear spaces V and W is

Linear spaces as objects and linear transformations as morphisms form a

x1 e1 + + xn en of the basis vectors. The numbers (x1 , . . . , xn ) are called

(1) Recall that Rn denotes the set of n-tuples of real numbers

It is clear that Rn is an n-dimensional linear space over the field of real

(A + B)(x) = A(x) + B(x), (A)(x) = (A(x)).

(3) A linear subspace of a linear space V is a nonempty subspace W V ,

(4) If L Hom(V, W ) is a linear transformation, then the image im L =

V /W , v 7 Tv (W ) is the factor map. Every linear map L : V W is a

(6) The linear space V = Hom(V, R) consisting of the linear functions on V

1.2.2 Determinant of Matrices and Linear

which can be defined by the formula

Exercise 1.2.6. Show that the sign of a permutation is always equal to 1.

Definition 1.2.7. The determinant of an n n matrix A = (aij )1i,jn is

Some important properties of the determinant are summarized in the follow-

(2) When we permute the columns of a square matrix the determinant is

det(a(1) , . . . , a(n) ) = sgn det(a1 , . . . , an ) for all Sn .

det A = det A> .

As a consequence, properties (1)-(4) hold also if the columns are replaced by

Recall that the product AB of two n n matrices A and B is also an n n

det(AB) = det(A) det(B).

Proposition 1.2.10. Let L : V V be a linear endomorphism of the linear

Proof. Let S : V V be the invertible linear endomorphism which takes the

Comparing the coefficients we obtain

Definition 1.2.12. The complexification C V of a linear space V over R

(x + iy)(v + iw) = (xv yw) + i(xw + yv).

A linear endomorphism of L : V V can be extended to the complexification

Comparing these expressions we can relate elementary symmetric polynomi-

The coefficient of ()n1 in pA () is equal to the sum of the diagonal

1.2.3 Orientation of a Linear Space

has positive determinant.

1.2.4 Tensor Product

One can consider the category of bilinear maps defined on V W , in which

(v1 + v2 , w) (v1 , w) (v2 , w), (w, v1 + v2 ) (w, v1 ) (w, v2 ),

for all v, v1 , v2 V , w, w1 , w2 W , R. Set V W = FV W /Z and let

L1 L2 (v1 v2 ) = L1 (v1 ) L2 (v2 ).