DG 1
DG 1
DG 1
Gabriel Lugo
University of North Carolina Wilmington
i
This book is dedicated to my family, for without their love and support, this work would have not been
possible. Most importantly, let us not forget that our greatest gift are our children and our children’s
children, who stand as a reminder, that curiosity and imagination is what drives our intellectual pursuits,
detaching us from the mundane and confinement to Earthly values, and bringing us closer to the divine and
the infinite.
The 1992 edition of this document was reproduced and assigned an ISBN code, by the Printing services of the
University of North Carolina Wilmington, from a camera-ready copy supplied by the author. Thereafter, the
document has been made available for free download as the main text for the Differential Geometry course
offered in this department. The text was generated on a desktop computer using LATEX.
All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or trans-
mitted, in any form or by any means, electronic, mechanical, photocopying, recording or otherwise, without
the written permission of the author. Printed in the United States of America.
ii
Preface
These notes were developed as part a course on differential geometry at the advanced undergraduate,
first year graduate level, which the author has taught for several years. There are many excellent texts in
differential geometry but very few have an early introduction to differential forms and their applications to
Physics. It is purpose of these notes to:
1. Provide a bridge between the very practical formulation of classical differential geometry created by
early masters of the late 1800’s, and the more elegant but less intuitive modern formulation in terms
of manifolds, bundles and differential forms. In particular, the central topic of curvature is presented
in three different but equivalent formalisms.
2. Present the subject of differential geometry with an emphasis on making the material readable to
physicists who may have encountered some of the concepts in the context of classical or quantum
mechanics, but wish to strengthen the rigor of the mathematics. A source of inspiration for this goal is
rooted in the shock to this author as a graduate student in the 70’s at Berkeley, at observing the gasping
failure of communications between the particle physicists working on gauge theories and differential
geometers working on connection on fiber bundles. They seemed to be completely unaware at the time,
that they were working on the same subject.
3. Make the material as readable as possible for those who stand at the boundary between theoretical
physics and applied mathematics. For this reason, it will be occasionally necessary to sacrifice some
mathematical rigor or depth of physics, in favor of ease of comprehension.
4. Provide the formal geometrical background which constitute the foundation of the mathematical theory
of general relativity.
5. Introduce examples of other applications of differential geometry to physics that might not appear in
traditional texts used in courses for mathematics students.
This book should be accessible to students who have completed traditional training in advanced calculus,
linear algebra, and differential equations. Students who master the entirety of this material will have gained
insight in a very powerful tool in mathematical physics at the graduate level.
Gabriel Lugo, Ph. D.
Mathematical Sciences and Statistics
UNCW
Wilmington, NC 28403
lugo@uncw.edu
iii
iv
Contents
Preface iii
2 Differential Forms 27
2.1 One-Forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.2 Tensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.2.1 Tensor Products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.2.2 Inner Product . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.2.3 Minkowski Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.2.4 Wedge Products and 2-Forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.2.5 Determinants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.2.6 Vector Identities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.2.7 n-Forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.3 Exterior Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2.3.1 Pull-back . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.3.2 Stokes’ Theorem in Rn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.4 The Hodge ? Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
2.4.1 Dual Forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
2.4.2 Laplacian . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
2.4.3 Maxwell Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3 Connections 61
3.1 Frames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.2 Curvilinear Coordinates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.3 Covariant Derivative . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.4 Cartan Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4 Theory of Surfaces 75
4.1 Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.2 The First Fundamental Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4.3 The Second Fundamental Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.4 Curvature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.4.1 Classical Formulation of Curvature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
v
0 CONTENTS
References 147
Index 148
Chapter 1
1.1.1 Definition Euclidean n-space Rn is defined as the set of ordered n-tuples p(p1, . . . , pn), where pi ∈ R,
for each i = 1, . . . , n. We may associate a position vector p = (p1, . . . , pn) with any given point p in n-space.
Given any two n-tuples p = (p1, . . . , pn), q = (q1, . . . , qn) and any real number c, we define two operations:
p + q = (p1 + q 1 , . . . , pn + q n ), (1.1)
c p = (c p1 , . . . , c pn ).
These two operations of vector sum and multiplication by a scalar satisfy all the 8 properties needed to give
the set V = Rn a natural structure of a vector space. It is common to use the same notation Rn for the
space of n-tuples and for the vector space of position vectors. Technically we should write p ∈ Rn when
think of Rn as a metric space and p ∈ Rn when we think of it as vector space, but as most authors, we will
freely abuse the notation.
1
xi (p) = pi
for any point p = (p1 , . . . , pn ). The functions xi are then called the natural coordinate functions. When
convenient, we revert to the usual names for the coordinates, x1 = x, x2 = y and x3 = z in R3 . A small
awkwardness might occur in the transition to modern notation. In classical vector calculus, a point in Rn is
often denoted by x, in which case, we pick up the coordinates with the slot projection functions ui : Rn → R
defined by
ui (x) = xi .
1.1.3 Definition A real valued function in Rn is of class C r if all the partial derivatives of the function up
to order r exist and are continuous. The space of infinitely differentiable (smooth) functions will be denoted
by C ∞ (Rn ) or F (Rn ).
1 In these notes we will use the following index conventions:
In Rn , indices such as i, j, k, l, m, n, run from 1 to n.
In space-time, indices such as µ, ν, ρ, σ, run from 0 to 3.
On surfaces in R3 , indices such as α, β, γ, δ, run from 1 to 2.
Spinor indices such as A, B, Ȧ, Ḃ run from 1 to 2.
1
2 CHAPTER 1. VECTORS AND CURVES
1.1.4 Definition Let V and V 0 be finite dimensional vector spaces such as V = Rk and V 0 = Rn , and
let L(V, V 0 ) be the space of linear transformations from V to V 0 . The set of linear functionals L(V, R) is
called the dual vector space V ∗ . This space has the same dimension as V .
In calculus, vectors are usually regarded as arrows characterized by a direction and a length. Thus,
vectors are considered as independent of their location in space. Because of physical and mathematical
reasons, it is advantageous to introduce a notion of vectors that does depend on location. For example, if
the vector is to represent a force acting on a rigid body, then the resulting equations of motion will obviously
depend on the point at which the force is applied. In later chapters, we will consider vectors on curved
spaces; in these cases, the positions of the vectors are crucial. For instance, a unit vector pointing north at
the earth’s equator is not at all the same as a unit vector pointing north at the tropic of Capricorn. This
example should help motivate the following definition.
1.1.5 Definition A tangent vector Xp in Rn , is an ordered pair {x, p}. We may regard x as an ordinary
advanced calculus “arrow-vector” and p is the position vector of the foot of the arrow.
The collection of all tangent vectors at a point p ∈ Rn is called the tangent space at p and will be
denoted by Tp (Rn ). Given two tangent vectors Xp , Yp and a constant c, we can define new tangent vectors
at p by (X + Y )p =Xp + Yp and (cX)p = cXp . With this definition, it is clear that for each point p, the
corresponding tangent space Tp (Rn ) at that point has the structure of a vector space. On the other hand,
there is no natural way to add two tangent vectors at different points.
The set T (Rn ) (or simply T Rn ) consisting of the union of all tangent spaces at all points in Rn is called
the tangent bundle. This object is not a vector space, but as we will see later it has much more structure
than just a set.
1.1.6 Definition A vector field X in U ⊂ Rn is a section of the tangent bundle, that is, a smooth function
from U to T (U ). The space of sections Γ(T (U ) is also denoted by X (U ).
The difference between a tangent vector and a vector field is that in
the latter case, the coefficients v i of x are smooth functions of xi . Since in
general there are not enough dimensions to depict a tangent bundle and
vector fields as sections thereof, we use abstract diagrams such as shown
Figure 1.1. In such a picture, the base space M (in this case M = Rn )
is compressed into the continuum at the bottom of the picture in which
several points p1 , . . . , pk are shown. To each such point one attaches a
tangent space. Here, the tangent spaces are just copies of Rn shown as
vertical “fibers” in the diagram. The vector component xp of a tangent
vector at the point p is depicted as an arrow embedded in the fiber.
The union of all such fibers constitutes the tangent bundle T M = T Rn .
A section of the bundle amounts to assigning a tangent vector to every
point in the base. It is required that such assignment of vectors is done Fig. 1.2: Vector Field
in a smooth way so that there are no major “changes” of the vector field
between nearby points.
Given any two vector fields X and Y and any smooth function f , we can define new vector fields X + Y
1.1. TANGENT VECTORS 3
and f X by
(X + Y )p = Xp + Yp (1.2)
(f X)p = f Xp ,
so that X (U ) has the structure of a vector space over R. The subscript notation Xp indicating the location
of a tangent vector is sometimes cumbersome, but necessary to distinguish them from vector fields.
Vector fields are essential objects in physical applications. If we consider the flow of a fluid in a region,
the velocity vector field represents the speed and direction of the flow of the fluid at that point. Other
examples of vector fields in classical physics are the electric, magnetic, and gravitational fields. The vector
field in figure 1.2 represents a magnetic field around an electrical wire pointing out of the page.
1.1.7 Definition Let Xp = {x, p} be a tangent vector in an open neighborhood U of a point p ∈ Rn and
let f be a C ∞ function in U . The directional derivative of f at the point p, in the direction of x, is defined
by
Xp (f ) = ∇f (p) · x, (1.3)
where ∇f (p) is the gradient of the function f at the point p. The notation
Xp (f ) ≡ ∇Xp f,
is also commonly used. This notation emphasizes that, in differential geometry, we may think of a tangent
vector at a point as an operator on the space of smooth functions in a neighborhood of the point. The
operator assigns to a function f , the directional derivative of that function in the direction of the vector.
Here we need not assume as in calculus that the direction vectors have unit length.
It is easy to generalize the notion of directional derivatives to vector fields by defining
X(f ) ≡ ∇X f = ∇f · x, (1.4)
n
where the function f and the components of x depend smoothly on the points of R .
The tangent space at a point p in Rn can be envisioned as another copy of Rn superimposed at the
point p. Thus, at a point p in R2 , the tangent space consist of the point p and a copy of the vector space R2
attached as a “tangent plane” at the point p. Since the base space is a flat 2-dimensional continuum, the
tangent plane for each point appears indistinguishable from the base space as in figure 1.2.
Later we will define the tangent space for a curved continuum such as a surface in R3 as shown in figure
1.3. In this case, the tangent space at a point p consists of the vector space of all vectors actually tangent
to the surface at the given point.
1.1.9 Remark The space of smooth functions is a ring, ignoring a small technicality with domains. An
operator such as a vector field with the properties above, is called a linear derivation on F (Rn ). Proof
First, let us develop an mathematical expression for tangent vectors and vector fields that will facilitate
computation.
Let p ∈ U be a point and let xi be the coordinate functions in U . Suppose that Xp = {x, p}, where the
components of the Euclidean vector x are (v 1 , . . . , v n ). Then, for any function f , the tangent vector Xp
operates on f according to the formula
n
X ∂f
Xp (f ) = vi (p). (1.6)
i=1
∂xi
It is therefore natural to identify the tangent vector Xp with the differential operator
n
X ∂
Xp = vi (1.7)
i=1
∂xi p
∂ ∂
Xp = v1 + · · · + vn .
∂x1 p ∂xn p
Notation: We will be using Einstein’s convention to suppress the summation symbol whenever an expression
contains a repeated index. Thus, for example, the equation above could be simply written as
∂
Xp = v i . (1.8)
∂xi p
This equation implies that the action of the vector Xp on the coordinate functions xi yields the components
v i of the vector. In elementary treatments, vectors are often identified with the components of the vector,
and this may cause some confusion.
The operators ( )
∂ ∂
{e1 , . . . , ek }|p = ,...,
∂x1 p ∂xn p
form a basis for the tangent space Tp (Rn ) at the point p, and any tangent vector can be written as a
linear combination of these basis vectors. The quantities v i are called the contravariant components of the
tangent vector. Thus, for example, the Euclidean vector in R3
x = 3i + 4j − 3k
∂
Let X = v i be an arbitrary vector field and let f and g be real-valued functions. Then
∂xi
∂
X(af + bg) = vi (af + bg)
∂xi
∂ ∂
= v i i (af ) + v i i (bg)
∂x ∂x
i ∂f i ∂g
= av + bv
∂xi ∂xi
= aX(f ) + bX(g).
1.1. TANGENT VECTORS 5
Similarly,
∂
X(f g) = vi (f g)
∂xi
∂ ∂
= v i f i (g) + v i g i (f )
∂x ∂x
∂g ∂f
= f v i i + gv i i
∂x ∂x
= f X(g) + gX(f ).
To re-emphasize, any quantity in Euclidean space which satisfies relations 1.5 is a called a linear derivation
on the space of smooth functions. The word linear here is used in the usual sense of a linear operator in
linear algebra, and the word derivation means that the operator satisfies Leibnitz’ rule.
The proof of the following proposition is slightly beyond the scope of this course, but the proposition is
important because it characterizes vector fields in a coordinate-independent manner.
1.1.11 Example Given the point p(1, 1), the Euclidean vector x = (3, 4), and the function f (x, y) =
x2 + y 2 , we associate x with the tangent vector
∂ ∂
Xp = 3 +4 .
∂x ∂y
Then,
∂f ∂f
Xp (f ) = 3 +4 ,
∂x p ∂y p
= 3(2x)|p + 4(2y)|p ,
= 3(2) + 4(2) = 14.
∂f ∂f ∂f
X(f ) = 3x + 2y +z
∂x ∂y ∂z
= 3x(y 2 z 3 ) + 2y(2xyz 3 ) + z(3xy 2 z 2 ),
= 3xy 2 z 3 + 4xy 2 z 3 + 3xy 2 z 3 = 10xy 2 z 3 .
1.1.13 Definition Let X be a vector field in Rn and p be a point. A curve α(t) with α(0) = p is called
an integral curve of X if α0 (0) = Xp , and, whenever α(t) is the domain of the vector field, α0 (t) = Xα(t) .
In elementary calculus and differential equations, the families of integral curves of a vector field are called
the streamlines, suggesting the trajectories of a fluid with velocity vector X. In figure 1.2, the integral curves
would be circles that fit neatly along the flow of the vector field. In local coordinates, the expression defining
integral curves of X constitutes a system of first order differential equations, so the existence and uniqueness
of solutions apply locally. We will treat this in more detail in section ??
6 CHAPTER 1. VECTORS AND CURVES
Remarks
For each point p ∈ Rn , we say that the Jacobian induces a linear transformation F∗ from the tangent
space Tp Rn to the tangent space TF (p) Rm . In differential geometry we this Jacobian map is also called the
push-forward. If we let X be a tangent vector in Rn , then the tangent vector F∗ X in Rm is defined by
As shown in the diagram, F∗ X(f ) is evaluated at F (p) whereas X is evaluated at p. So, to be precise,
equation 1.11 should really be written as
As we have learned from linear algebra, to find a matrix representation of a linear map in a particular basis,
∂
one applies the map to the basis vectors. If we denote by { ∂x i } the basis for the tangent space at a point
1.2. DIFFERENTIABLE MAPS 7
p ∈ Rn and by { ∂y∂ j } the basis for the tangent space at the corresponding point F (p) ∈ Rm with coordinates
given by y j = f j (xi ), the push-forward definition reads,
∂ ∂
F∗ ( )(f ) = (f ◦ F ),
∂xi ∂xi
∂f ∂y j
= ,
∂y j ∂xi
∂ ∂y j ∂
F∗ ( i ) = .
∂x ∂xi ∂y j
In other words, the matrix representation of F∗ in standard basis is in fact the Jacobian matrix. In classical
notation, we simply write the Jacobian map in the familiar form,
∂ ∂y j ∂
i
= . (1.14)
∂x ∂xi ∂y j
(G ◦ F )∗ (X)(f ) = X(f ◦ (G ◦ F ),
= X((f ◦ G) ◦ F ),
= F∗ (X)(f ◦ G),
= G∗ (F∗ (X)(f )),
= (G∗ ◦ F∗ )(X)(f ).
1.2.3 Inverse Function Theorem. When m = n, mappings are called change of coordinates. In the
terminology of tangent spaces, the classical inverse function theorem states that if the Jacobian map F∗
is a vector space isomorphism at a point, then there exists a neighborhood of the point in which F is a
diffeomorphism.
1.2.4 Remarks
1. Equation 1.14 shows that under change of coordinates, basis tangent vectors and by linearity all tangent
vectors transform by multiplication by the matrix representation of the Jacobian. This is the source
of the almost tautological definition in physics, that a contravariant tensor of rank one, is one that
transforms like a contravariant tensor of rank one.
2. Many authors use the notation dF to denote the push-forward map F∗ .
1.3 Curves in R3
1.3.1 Parametric Curves
1.3.1 Definition A curve α(t) in R3 is a C ∞ map from an interval I ⊂ R into R3 . The curve assigns to
each value of a parameter t ∈ R, a point (α1 (t), α2 (t), α3 (t)) ∈ R3 .
α
I⊂R 7−→ R3
t 7−→ α(t) = (α1 (t), α2 (t), α3 (t))
One may think of the parameter t as representing time, and the curve α as representing the trajectory of a
moving point particle as a function of time. When convenient, we also use classical notation for the position
vector
x(t) = (x1 (t), x2 (t), x3 (t)), (1.15)
which is more prevalent in vector calculus and elementary physics textbooks. Of course, what this notation
really means is
xi (t) = (ui ◦ α)(t), (1.16)
where ui are the coordinate slot functions in an open set in R3
This curve is called the Temple of Viviani. Geometrically, this is the curve of intersection of a sphere
x2 + y 2 + z 2 = 4a2 of radius 2a, and the cylinder x2 + y 2 = 2ax of radius a, with a generator tangent to the
diameter of the sphere along the z-axis (See Figure 1.5b).
The Temple of Viviani is of historical interest in the development of calculus. The problem was posed
anonymously by Viviani to Leibnitz, to determine on the surface of a semi-sphere, four identical windows,
in such a way that the remaining surface be equivalent to a square. It appears as if Viviani was challenging
the effectiveness of the new methods of calculus against the power of traditional geometry.
It is said that Leibnitz understood the nature of the challenge
and solved the problem in one day. Not knowing the proposer of
the enigma, he sent the solution to his Serenity Ferdinando, as he
guessed that the challenge must have originated from prominent Ital-
ian mathematicians. Upon receipt of the solution by Leibnitz, Vi-
viani posted a mechanical solution without proof. He described it as
using a boring device to remove from a semisphere, the surface area
cut by two cylinders with half the radius, and which are tangential
to a diameter of the base. Upon realizing this could not physically
be rendered as a temple since the roof surface would rest on only
four points, Viviani no longer spoke of a temple but referred to the
shape as a “sail.”
At each point of the curve, the velocity vector is tangent to the curve and thus the velocity constitutes a
vector field representing the velocity flow along that curve. In a similar manner the second derivative α00 (t)
is a vector field called the acceleration along the curve. The length v = kα0 (t)k of the velocity vector is called
the speed of the curve. The classical components of the velocity vector are simply given by
1
dx dx2 dx3
dx
v(t) = ẋ ≡ = , , , (1.21)
dt dt dt dt
which appears in line integrals in advanced calculus is some sort of an infinitesimal tangent vector. The
norm kdxk of this infinitesimal tangent vector is called the differential of arc length ds. Clearly, we have
If one identifies the parameter t as time in some given units, what this says is that for a particle moving
along a curve, the speed is the rate of change of the arc length with respect to time. This is intuitively
exactly what one would expect.
The notion of infinitesimal objects needs to be treated in a more rigorous mathematical setting. At the
same time, we must not discard the great intuitive value of this notion as envisioned by the masters who
invented calculus, even at the risk of some possible confusion! Thus, whereas in the more strict sense of
modern differential geometry, the velocity is a tangent vector and hence it is a differential operator on the
space of functions, the quantity dx can be viewed as a traditional vector which, at the infinitesimal level,
represents a linear approximation to the curve and points tangentially in the direction of v.
1.3.2 Velocity
For any smooth function f : R3 → R , we formally define the action of the velocity vector field α0 (t) as
a linear derivation by the formula
d
α0 (t)(f ) |α(t) = (f ◦ α) |t . (1.25)
dt
The modern notation is more precise, since it takes into account that the velocity has a vector part as well
as point of application. Given a point on the curve, the velocity of the curve acting on a function, yields the
directional derivative of that function in the direction tangential to the curve at the point in question. The
diagram in figure 1.6 below provides a more visual interpretation of the velocity vector formula 1.25, as a
linear mapping between tangent spaces.
The map α(t) from R to R3 induces a push-forward map α∗ from the tangent space of R to the tangent
d
space of R3 . The image α∗ ( dt d
) in T R3 of the tangent vector dt is what we call α0 (t).
α∗ (d/dt) = α0 (t).
Since α0 (t) is a tangent vector in R3 , it acts on functions in R3 . The action of α0 (t) on a function f on R3 is
the same as the action of d/dt on the composition (f ◦ α). In particular, if we apply α0 (t) to the coordinate
functions xi , we get the components of the tangent vector
d i
α0 (t)(xi ) |α(t) = (x ◦ α)|t . (1.26)
dt
To unpack the above discussion in the simplest possible terms, we associate with the classical velocity vector
v = ẋ a linear derivation α0 (t) given by
d i
α0 (t) = (x ◦ α)t (∂/∂xi )α(t) ,
dt
dx1 ∂ dx2 ∂ dx3 ∂
= + + . (1.27)
dt ∂x1 dt ∂x2 dt ∂x3
So, given a real valued function f in R3 , the action of the velocity vector is given by the chain rule
If α(t) is a curve in Rn with tangent vector X = α0 (t), and F : Rn → Rm is differentiable map, then F∗ X
is a tangent vector to the curve F ◦ α in Rm . That is, F∗ maps tangent vectors of α to tangent vectors of
F ◦ α.
1.3.6 Definition If t = t(s) is a smooth, real valued function and α(t) is a curve in R3 , we say that the
curve β(s) = α(t(s)) is a reparametrization of α.
A common reparametrization of curve is obtained by using the arc length as the parameter. Using this
reparametrization is quite natural, since we know from basic physics that the rate of change of the arc length
is what we call speed
ds
v= = kα0 (t)k. (1.28)
dt
The arc length is obtained by integrating the above formula
s 2 2 2 3 2
dx1
Z Z
0 dx dx
s = kα (t)k dt = + + dt (1.29)
dt dt dt
In practice, it is typically difficult to find an explicit arc length parametrization of a curve since not only
does one have to calculate the integral, but also one needs to be able to find the inverse function t in terms
of s. On the other hand, from a theoretical point of view, arc length parameterizations are ideal, since any
curve so parametrized has unit speed. The proof of this fact is a simple application of the chain rule and
the inverse function theorem.
β 0 (s) = [α(t(s))]0
= α0 (t(s))t0 (s)
1
= α0 (t(s)) 0
s (t(s))
α0 (t(s))
= ,
kα0 (t(s))k
and any vector divided by its length is a unit vector. Leibnitz notation makes this even more self-evident
dx
dx dx dt dt
= = ds
ds dt ds dt
dx
dt
=
k dx
dt k
1.3.7 Example Let α(t) = (a cos ωt, a sin ωt, bt). Then
Z tp
s(t) = (−aω sin ωu)2 + (aω cos ωu)2 + b2 du
0
Z tp
= a2 ω 2 + b2 du
0
p
= ct, where, c = a2 ω 2 + b2 .
The vector T (s) is tangential to the curve and it has unit length. Hereafter, we will call T the unit tangent
vector. Differentiating the relation
T · T = 1, (1.31)
we get
2 T · T 0 = 0, (1.32)
so we conclude that the vector T 0 is orthogonal to T . Let N be a unit vector orthogonal to T , and let κ be
the scalar such that
T 0 (s) = κN (s). (1.33)
We call N the unit normal to the curve, and κ the curvature. Taking the length of both sides of last
equation, and recalling that N has unit length, we deduce that
κ = kT 0 (s)k. (1.34)
It makes sense to call κ the curvature because, if T is a unit vector, then T 0 (s) is not zero only if the direction
of T is changing. The rate of change of the direction of the tangent vector is precisely what one would expect
to measure how much a curve is curving. We now introduce a third vector
B = T × N, (1.35)
which we will call the binormal vector. The triplet of vectors (T, N, B) forms an orthonormal set; that is,
T · T = N · N = B · B = 1,
T · N = T · B = N · B = 0. (1.36)
B 0 · T + B · T 0 = 0.
we also conclude that B 0 must also be orthogonal to T . This can only happen if B 0 is orthogonal to the
T B-plane, so B 0 must be proportional to N . In other words, we must have
for some quantity τ , which we will call the torsion. The torsion is similar to the curvature in the sense that
it measures the rate of change of the binormal. Since the binormal also has unit length, the only way one
can have a non-zero derivative is if B is changing directions. This means that if in addition B did not change
directions, the vector would truly be a constant vector, so the curve would be a flat curve embedded into
the T N -plane.
The quantity B 0 then measures the rate of change in the up and down direction of an observer moving
with the curve always facing forward in the direction of the tangent vector. The binormal B is something
like the flag in the back of sand dune buggy.
1.3. CURVES IN R3 13
N 0 = aT + bB.
Taking the dot product of last equation with T and B respectively, we see that
a = N 0 · T, and b = N 0 · B.
N 0 · T = −N · T 0 = −N · (κN ) = −κ
N 0 · B = −N · B 0 = −N · (−τ N ) = τ.
N 0 = −κT + τ B.
The Frenet frame equations (1.38) can also be written in matrix form as shown below.
0
T 0 κ 0 T
N = −κ 0 τ N . (1.39)
B 0 −τ 0 B
The group-theoretic significance of this matrix formulation is quite important and we will come back to this
later when we talk about general orthonormal frames. Presently, perhaps it suffices to point out that the
appearance of an antisymmetric matrix in the Frenet equations is not at all coincidental.
The following theorem provides a computational method to calculate the curvature and torsion directly
from the equation of a given unit speed curve.
1.3.9 Proposition Let β(s) be a unit speed curve with curvature κ > 0 and torsion τ . Then
κ = kβ 00 (s)k
β 0 · [β 00 × β 000 ]
τ = (1.40)
β 00 · β 00
Proof If β(s) is a unit speed curve, we have β 0 (s) = T . Then
T 0 = β 00 (s) = κN,
β 00 · β 00 = (κN ) · (κN ),
β 00 · β 00 = κ2
κ2 = kβ 00 k2
14 CHAPTER 1. VECTORS AND CURVES
β 000 (s) = κ0 N + κN 0
= κ0 N + κ(−κT + τ B)
= κ0 N + −κ2 T + κτ B.
Then,
Therefore, ds/dt = r and s = rt, which we recognize as the formula for the length of an arc of circle of radius
r, subtended by a central angle whose measure is t radians. We conclude that
s s
β(s) = (r cos , r sin , 0)
r r
is a unit speed reparametrization. The curvature of the circle can now be easily computed
s s
T = β 0 (s) = (− sin , cos , 0),
r r
1 s 1 s
T 0 = (− cos , − sin , 0),
r r r r
κ = kβ 00 k = kT 0 k,
r
1 s 1 s
= 2
cos2 + 2 sin2 + 02 ,
r r r r
r
1 s s
= (cos2 + sin2 ),
r2 r r
1
= .
r
This is a very simple but important example. The fact that for a circle of radius r the curvature is κ = 1/r
could not be more intuitive. A small circle has large curvature and a large circle has small curvature. As
the radius of the circle approaches infinity, the circle locally looks more and more like a straight line, and
the curvature approaches 0. If one were walking along a great circle on a very large sphere (like the earth)
one would be perceive the space to be locally flat.
1.3.11 Proposition Let α(t) be a curve of velocity v, acceleration a, speed v and curvature κ, then
v = vT,
dv
a = T + v 2 κN. (1.41)
dt
1.3. CURVES IN R3 15
Proof Let s(t) be the arc length and let β(s) be a unit speed reparametrization. Then α(t) = β(s(t)) and
by the chain rule
v = α0 (t),
= β 0 (s(t))s0 (t),
= vT.
a = α00 (t),
dv
= T + vT 0 (s(t))s0 (t),
dt
dv
= T + v(κN )v,
dt
dv
= T + v 2 κN.
dt
Equation 1.41 is important in physics. The equation states that a particle moving along a curve in space
feels a component of acceleration along the direction of motion whenever there is a change of speed, and
a centripetal acceleration in the direction of the normal whenever it changes direction. The centripetal
Acceleration and any point is
v2
a = v2 κ =
r
ωs ωs bs p
β(s) = (a cos , a sin , ), where c = a2 ω 2 + b2 ,
c c c
aω ωs aω ωs b
β 0 (s) = (− sin , cos , ),
c c c c c
aω 2 ωs aω 2 ωs
β 00 (s) = (− 2 cos , − 2 sin , 0),
c c c c
aω 3 ωs aω 3 ωs
β 000 (s) = ( 3 sin , − 3 cos , 0),
c c c c
κ2 = β 00 · β 00 ,
a2 ω 4
= ,
c4
aω 2
κ = ± 2 .
c
(β 0 β 00 β 000 )
τ = ,
β 00 · β 00
" 2
#
b − aω ωs aω 2 ωs c4
2
c3 cos c − 2
c 3 sin c ,
= aω ωs aω ωs
,
c c2 sin c − c2 cos c a ω4
2
b a2 ω 5 c4
= .
c c5 a2 ω 4
Simplifying the last expression and substituting the value of c, we get
bω
τ = ,
a2 ω 2
+ b2
aω 2
κ = ± 2 2 .
a ω + b2
Notice that if b = 0, the helix collapses to a circle in the xy-plane. In this case, the formulas above reduce
to κ = 1/a and τ = 0. The ratio κ/τ = aω/b is particularly simple. Any curve for which κ/τ = constant, is
called a helix; the circular helix is a special case.
1.3.13 Example (Plane curves) Let α(t) = (x(t), y(t), 0). Then
α0 = (x0 , y 0 , 0),
α00 = (x00 , y 00 , 0),
000
α (x000 , y 000 , 0),
=
kα0 × α00 k
κ = ,
kα0 k3
| x0 y 00 − y 0 x00 |
= .
(x02 + y 02 )3/2
τ = 0.
s2 t2
β 0 (s) = (cos , sin , 0),
2c2 2c2
Since kβ 0 k = v = 1, the curve is of unit speed, and s is indeed the arc length. The curvature is given by
kα0 × α00 k2
κ2 = , (1.43)
kα0 k6
(α0 α00 α000 )
τ = , (1.44)
kα0 × α00 k2
where (α0 α00 α000 ) is the triple vector product [α0 × α00 ] · α000 .
Proof
α0 = vT,
00
α = v 0 T + v 2 κN,
α000 = (v 2 κ)N 0 + . . . ,
= v 3 κN 0 + . . . ,
= v 3 κτ B + . . . .
As the computation below shows, the other terms in α000 are unimportant here because α0 ×α00 is proportional
to B, so all we need is the B component to solve for τ .
curvature and torsion. Some geometrical insight into the significance of the curvature and torsion can be
gained by considering the Taylor series expansion of an arbitrary unit speed curve β(s) about s = 0.
β 00 (0) 2 β 000 (0) 3
β(s) = β(0) + β 0 (0)s + s + s + ... (1.45)
2! 3!
Since we are assuming that s is an arc length parameter,
β 0 (0) = T (0) = T0
β 00 (0) = (κN )(0) = κ0 N0
000
β (0) = (−κ2 T + κ0 N + κτ B)(0) = −κ20 T0 + κ00 N0 + κ0 τ0 B0
Keeping only the lowest terms in the components of T , N , and B, we get the first order Frenet approximation
to the curve
. 1 1
β(s) = β(0) + T0 s + κ0 N0 s2 + κ0 τ0 B0 s3 . (1.46)
2 6
The first two terms represent the linear approximation to the curve. The first three terms approximate the
curve by a parabola which lies in the osculating plane (T N -plane). If κ0 = 0, then locally the curve looks
like a straight line. If τ0 = 0, then locally the curve is a plane curve contained on the osculating plane.
In this sense, the curvature measures the deviation of the curve from a straight line and the torsion (also
called the second curvature) measures the deviation of the curve from a plane curve. As shown in figure 1.9
a non-planar space curve locally looks like a wire that has first been bent in a parabolic shape in the T N
and twisted into a cubic along the B axis.
So suppose that p is an arbitrary point on a curve β(s) parametrized by arc length. We position the
curve so that p is at the origin so that β(0) = 0 coincides with the point p. We chose the orthonormal basis
vectors {e1 , e2 , e3 } in R3 to coincide with the Frenet Frame T0 , N0 , B0 at that point. then, the equation
(1.46) provides a canonical representation of the curve near that point. This then constitutes a proof of the
fundamental theorem of curves under the assumption the curve, curvature and torsion are analytic. (One
could also treat the Frenet formulas as a system of differential equations and apply the conditions of existence
and uniqueness of solutions for such systems.)
α = c1 α1 + c2 α2 + c3 , ci = constant vectors.
1.4.1 Isometries
1.4.3 Definition Let x and y be two column vectors in Rn and let xT represent the transposed row vector.
To keep track on whether a vector is a row vector or a column vector, hereafter we write the components
{xi } of a column vector with the indices up and the components {xi } of a row vector with the indices down.
Similarly, if A is an n × n matrix, we write its components as A = (ai j ). The standard inner product is given
by matrix multiplication of the row and column vectors
The inner product gives Rn the structure of a normed space by defining kxk =< x, x >1/2 and the structure
of a metric space in which d(x, y) = kx − yk. The real inner product is bilinear (linear in each slot), from
which it follows that
kx ± yk2 = kxk2 ± 2 < x, y > +kyk2 . (1.49)
Thus, we have the polarization identity
< x, y > = xT y,
= (xi αi )(y j ej ),
= (xi y j )αi (ej ) = (xi y j )δji .
= xi y i ,
y1
y2
= x1 x2 ... xn
...
yn
Since | cos θ| ≤ 1, it follows from equation 1.51, a special case of the Schwarz inequality
Let F be a linear transformation from Rn to Rn and B = {e1 , . . . en } be an orthonormal basis. Then, there
exists a matrix A = [F ]B given by
On the other hand, if A is a fixed n × n matrix, the map F defined by F (x) = Ax is a linear transformation
from Rn to Rn whose matrix representation in the standard basis is the matrix A itself. It follows that
given a linear transformation represented by a matrix A, we have
1.4.4 Definition A real n×n matrix A is called orthogonal if AT A = AAT = I. The linear transformation
represented by A is called an orthogonal transformation. Equivalently, the transformation represented by A
is orthogonal if
< x, Ay >=< A−1 x, y > . (1.57)
Thus, real orthogonal transformations are represented by symmetric matrices (Hermitian in the complex
case) and the condition AT A = I implies that det(A) = ±1.
1.4.5 Theorem If A is an orthogonal matrix, then the transformation determined by A preserves the
inner product and the norm.
Proof
Furthermore, setting y = x:
As a corollary, if {ei } is an orthonormal basis, then so is {fi = Aei }. That is, an orthogonal transformation
represents a rotation if det A = 1 and a rotation with a reflection if det A = −1.
1.4.6 Definition A mapping F : Rn → Rn called an isometry if it preserves distances. That is, if for all
x, y
1.4.7 Example (Translations) Let q be fixed vector. The map F (x) = x + q is called a translation. It
is clearly an isometry since kF (x) − F (y)k = kx + p − (y + p)k = kx − yk.
Proof Let F be an isometry represented by an orthogonal matrix A. Then, since the transformation is
linear and preserves norms, we have:
The composition of two isometries is also an isometry. The inverse of a translation by q is a translation
by −q. The inverse of an orthogonal transformation represented by A is an orthogonal transformation repre-
sented by A−1 . Consequently, the set of isometries consisting of translations and orthogonal transformations
constitutes a group. Given a general isometry, we can use a translation to insure that F (0) = 0. We now
prove the following theorem.
Now, using 1.49 and the norm preserving property above, we have:
To show F is linear, let ei be an orthonormal basis, which implies that fi = F (ei ) is also an orthonormal
basis. Then
n
X
F (ax + by) = < F (ax + by, fi > fi ,
i=1
Xn
= < F (ax + by), F (ei ) > fi ,
i=1
Xn
= < (ax + by), ei > fi ,
i=1
n
X n
X
=a < x, ei > fi + b < y, ei > fi ,
i=1 i=1
Xn n
X
=a < F (x), fi > fi + b < F (y), fi > fi ,
i=1 i=1
= aF (x) + bF (y).
F (x) = Ax + q, (1.59)
where A is orthogonal.
22 CHAPTER 1. VECTORS AND CURVES
Proof If F (0 = q, then Fe = F − q is an isometry with Fe(0) = 0 and hence by the previous theorem Fe is an
orthogonal transformation represented by an orthogonal matrix Fex = Ax. It follows that F (x) = Ax + q.
We have just shown that any isometry is the composition of translation and an orthogonal transformation.
The latter is the linear part of the isometry. The orthogonal transformation preserves the inner product,
lengths, and maps orthonormal bases to orthonormal bases.
1.4.11 Theorem If α is a curve in Rn and β is the image of α under a mapping F , then vectors tangent
to α get mapped to tangent vectors to β.
Proof Let β = F ◦ α. The proof follows trivially from the properties of the Jacobian map β∗ = (F ◦ α)∗ =
F∗ α∗ that takes tangent vectors to tangent vectors. If in addition F is an isometry, then F∗ maps the Frenet
frame of α to the Frenet frame of β.
1.4.12 Theorem (Fundamental theorem of curves) If C and C̃ are space curves such that κ(s) = κ̃(s),
and τ (s) = τ̃ (s) for all s, the curves are isometric.
Proof Given two such curves, we can perform a translation so that, for some s = s0 , the corresponding
points on C and C̃ are made to coincide. Without loss of generality, we can make this point be the origin.
Now we perform an orthogonal transformation to make the Frenet frame {T0 , N0 , B0 } of C coincide with
the Frenet frame {T̃0 , Ñ0 , B̃0 } of C̃. By Schwarz inequality, the inner product of two unit vectors is also a
unit vector, if and only if the vectors are equal. With this in mind, let
L = T · T̃ + N · Ñ + B · B̃.
A simple computation using the Frenet equations shows that L0 = 0, so L = constant. But at s = 0
the Frenet frames of the two curves coincide, so the constant is 3 and this can only happen if for all s,
T = T̃ , N = Ñ , B = B̃. Finally, since T = T̃ , we have β 0 (s) = β̃ 0 (s), so β(s) = β̃(s)+ constant. But since
β(0) = β̃(0), the constant is 0 and β(s) = β̃(s) for all s.
find an integral formula for the curve coordinates in terms of the curvature. Given a curve parametrized by
arc length, consider an arbitrary point with position vector x = (x, y) on the curve, and let ϕ be the angle
that the tangent vector T makes with the horizontal, as shown in figure 1.10. Then, the Euclidean vector
components of the unit tangent vector are given by
dx
= T = (cos ϕ, sin ϕ).
ds
1.4. FUNDAMENTAL THEOREM OF CURVES 23
dT dϕ dϕ
= (− sin ϕ , cos ϕ ) = κN,
ds ds ds
so that,
dT
dϕ
ds
= ds = κ.
We conclude that Z Z Z
x(s) = cos ϕ ds, y(s) = sin ϕ ds, where, ϕ= κ ds. (1.60)
Equations 1.60 are called the natural equations of a plane curve. Given the curvature κ, the equation of the
curve can be obtained by “quadratures,” the classical term for integrals.
The functions C(s) and S(s) are called Fresnel Integrals. In the standard classical function libraries of Maple
and Mathematica, they are listed as F resnelC and F resnelS respectively. The fast-increasing frequency of
oscillations of the integrands here make the computation prohibitive without the use of high-speed computers.
Graphing calculators are inadequate to render the rapid oscillations for s ranging from 0 to 15, for example,
and simple computer programs for the trapezoidal rule as taught in typical calculus courses, completely falls
apart in this range. The Cornu spiral is the curve x(s) = (x(s), y(s)) parametrized by Fresnel integrals (See
figure 1.11a). It is a tribute to the mathematicians of the 1800’s that not only were they able to compute
the values of the Fresnel integrals to 4 or 5 decimal places, but they did it for the range of s from 0 to
15 as mentioned above, producing remarkably accurate renditions of the spiral. Fresnel integrals appear in
the study of diffraction. If a coherent beam of light such as a laser beam, hits a sharp straight edge and
24 CHAPTER 1. VECTORS AND CURVES
a screen is placed behind, there will appear on the screen a pattern of diffraction fringes. The amplitude
and intensity of the diffraction pattern can be obtained by a geometrical construction involving the Fresnel
integrals. First consider the function Ψ(s) = kxk that measures the distance from the origin to the points
in the Cornu spiral in the first quadrant. The square of this function is then proportional to the intensity of
the diffraction pattern, The graph of |Ψ(s)|2 is shown in figure 1.11b. Translating this curve along an axis
coinciding with that of the straight edge, generates a three dimensional surface as shown from ”above” in
figure 1.11c. A color scheme was used here to depict a model of the Fresnel diffraction by the straight edge.
Derivation of formula 1.62 has fallen through the cracks in standard fat calculus textbooks, at best
relegated to an advanced exercise which most students will not do. Perhaps the reason is that the section
on polar coordinates is typically covered in Calculus II, so students have not yet been exposed to the tools
of vector calculus that facilitate the otherwise messy computation. To fill-in this gap, we present a short
derivation of this neat formula. For a plane curve in parametric polar coordinates, we have
x(t) = (r(t) cos θ(t), r(t) sin θ(t)),
ẋ = (ṙ cos θ − r sin θ θ̇, ṙ sin θ + r cos θ θ̇).
A direct computation of the dot product gives,
| < x, ẋ > |2 = (rṙ)2 .
On the other hand,
| < x, ẋ > |2 = kxk2 kẋk2 cos2 ψ,
= r2 (ṙ2 + r2 θ̇2 ) cos2 ψ.
Equating the two, we find,
ṙ2 = (ṙ2 + r2 θ̇2 ) cos2 ψ,
(sin2 ψ)ṙ2 = r2 θ̇2 cos2 ψ,
(sin ψ) dr = r cos ψ dθ,
r
tan ψ = .
dr/dθ
1.4. FUNDAMENTAL THEOREM OF CURVES 25
We leave it to the reader to do a direct computation of the curvature. Instead, we prove that if κ = 1/(as+b),
where a and b are constant, then the curve is a logarithmic spiral. From the natural equations, we have,
dθ 1
=κ= ,
ds as + b
1
θ = ln(as + b) + C, C = const,
a
eaθ = A(as + b), A = eC ,
1 ds
= Aeaθ = ,
κ dθ
ds = Aeas
Back to the natural equations, the x and y coordinates are obtained by integrating,
Z
x = Aeaθ cos θ dθ,
Z
y = Aeaθ sin θ dθ.
A whole family of meandering curves are obtained by letting κ = A sin ks. The meandering graph shown
in picture 1.13 was obtained by numerical integration for A = 2 and “wave number” k = 1. The larger the
value of A the larger the curvature of the “throats.” If A is large enough, the “throats” will overlap.
26 CHAPTER 1. VECTORS AND CURVES
Using superpositions of sine functions gives rise to a beautiful family of ”multi-frequency” meanders with
graphs that would challenge the most skillful calligraphists of the 1800’s. Figure 1.14 shows a rendition with
two sine functions with equal amplitude A = 1.8, and with k1 = 1, k2 = 1.2.
Chapter 2
Differential Forms
2.1 One-Forms
The concept of the differential of a function is one of the most puzzling ideas in elementary calculus. In
the usual definition, the differential of a dependent variable y = f (x) is given in terms of the differential of
the independent variable by dy = f 0 (x)dx. The problem is with the quantity dx. What does “dx” mean?
What is the difference between ∆x and dx? How much “smaller” than ∆x does dx have to be? There is
no trivial resolution to this question. Most introductory calculus texts evade the issue by treating dx as an
arbitrarily small quantity (lacking mathematical rigor) or by simply referring to dx as an infinitesimal (a
term introduced by Newton for an idea that could not otherwise be clearly defined at the time.)
In this section we introduce linear algebraic tools that will allow us to interpret the differential in terms
of a linear operator.
2.1.1 Definition Let p ∈ Rn , and let Tp (Rn ) be the tangent space at p. A 1-form at p is a linear map φ
from Tp (Rn ) into R, in other words, a linear functional. We recall that such a map must satisfy the following
properties:
A 1-form is a smooth assignment of a linear map φ as above for each point in the space.
∂ ∂xi
dxi ( j
)= = δji (2.4)
∂x ∂xj
The set of all linear functionals on a vector space is called the dual of the vector space. It is a standard
theorem in linear algebra that the dual of a finite dimensional vector space is also a vector space of the same
dimension. Thus, the space Tp? (Rn ) of all 1-forms at p is a vector space which is the dual of the tangent
space Tp (Rn ). The space Tp? (Rn ) is called the cotangent space of Rn at the point p. Equation (2.4)
indicates that the set of differential forms {(dx1 )p , . . . , (dxn )p } constitutes the basis of the cotangent space
27
28 CHAPTER 2. DIFFERENTIAL FORMS
∂ ∂
which is dual to the standard basis {( ∂x 1 )p , . . . ( ∂xn )p } of the tangent space. The union of all the cotangent
spaces as p ranges over all points in R is called the cotangent bundle T ∗ (Rn ).
n
2.1.3 Proposition Let f be any smooth function in Rn and let {x1 , . . . xn } be coordinate functions in a
neighborhood U of a point p. Then, the differential df is given locally by the expression
n
X ∂f i
df = i
dx (2.5)
i=1
∂x
∂f i
= dx
∂xi
Proof The differential df is by definition a 1-form, so, at each point, it must be expressible as a linear
combination of the basis elements {(dx1 )p , . . . , (dxn )p }. Therefore, to prove the proposition, it suffices to
show that the expression 2.5 applied to an arbitrary tangent vector coincides with definition 2.2. To see this,
∂
consider a tangent vector Xp = v j ( ∂x j )p and apply the expression above as follows:
∂f i ∂f i j ∂
( dx )p (Xp ) = ( dx )(v )(p) (2.6)
∂xi ∂xi ∂xj
∂f ∂
= v j ( i dxi )( j )(p)
∂x ∂x
i
∂f ∂x
= v j ( i j )(p)
∂x ∂x
∂f
= v j ( i δji )(p)
∂x
∂f
= ( i v i )(p)
∂x
= ∇f (p) · x(p)
= df (X)(p)
The definition of differentials as linear functionals on the space of vector fields is much more satisfactory than
the notion of infinitesimals, since the new definition is based on the rigorous machinery of linear algebra. If
α is an arbitrary 1-form, then locally
where the coefficients ai are C ∞ functions. Thus, a 1-form is a smooth section of the cotangent bundle and
we refer to it as a covariant tensor of rank 1, or simply a covector. The collection of all 1-forms is denoted
by Ω1 (Rn ) = T10 (Rn ). The coefficients (a1 , . . . , an ) are called the covariant components of the covector.
We will adopt the convention to always write the covariant components of a covector with the indices down.
Physicists often refer to the covariant components of a 1-form as a covariant vector and this causes some
confusion about the position of the indices. We emphasize that not all one forms are obtained by taking the
differential of a function. If there exists a function f , such that α = df , then the one form α is called exact. In
vector calculus and elementary physics, exact forms are important in understanding the path independence
of line integrals of conservative vector fields.
As we have already noted, the cotangent space Tp∗ (Rn ) of 1-forms at a point p has a natural vector
space structure. We can easily extend the operations of addition and scalar multiplication to the space of
all 1-forms by defining
2.2 Tensors
As we mentioned at the beginning of this chapter, the notion of the differential dx is not made precise
in elementary treatments of calculus, so consequently, the differential of area dxdy in R2 , as well as the
differential of surface area in R3 also need to be revisited in a more rigorous setting. For this purpose,
we introduce a new type of multiplication between forms that not only captures the essence of differentials
of area and volume, but also provides a rich algebraic and geometric structure generalizing cross products
(which make sense only in R3 ) to Euclidean space of any dimension.
2.2.1 Definition A map φ : X (Rn ) × X (Rn ) −→ R is called a bilinear map of vector fields, if it is linear
on each slot. That is, ∀Xi , Yi ∈ X (Rn ), f i ∈ F (Rn ), we have
2.2.2 Definition Let α and β be 1-forms. The tensor product of α and β is defined as the bilinear map
α ⊗ β such that
(α ⊗ β)(X, Y ) = α(X)β(Y ) (2.9)
for all vector fields X and Y .
Thus, for example, if α = ai dxi and β = bj dxj , then,
∂ ∂ ∂ ∂
(α ⊗ β)( k
, l) = α( k
)β( l )
∂x ∂x ∂x ∂x
∂ ∂
= (ai dx )( k )(bj dxj )( l )
i
∂x ∂x
= ai δki bj δlj
= ak bl .
A quantity of the form T = Tij dxi ⊗ dxj is called a covariant tensor of rank 2, and we may think of the set
{dxi ⊗ dxj } as a basis for all such tensors. The space of covariant tensor fields of rank 2 is denoted T20 (Rn ).
We must caution the reader again that there is possible confusion about the location of the indices, since
physicists often refer to the components Tij as a covariant tensor of rank two, as long is it satisfies some
transformation laws.
In a similar fashion, one can define the tensor product of vectors X and Y as the bilinear map X ⊗ Y
such that
three input slots. The map expects two functions in the first two slots and a vector in the third one. The
action of the map is bilinear on the two functions and linear on the vector. The output is a real number.
30 CHAPTER 2. DIFFERENTIAL FORMS
r
A tensor of type s is written in local coordinates as
∂ ∂
T = Tji11,...,j
,...,ir
i
⊗ · · · ⊗ ir ⊗ dxj1 ⊗ . . . dxjs (2.12)
s
∂x 1 ∂x
The tensor components are given by
∂ ∂
Tji11,...,j
,...,ir
= T (dxi1 , . . . , dxir , j
, . . . , js ). (2.13)
s
∂x 1 ∂x
The set Tsr |p (Rn ) of all tensors of type Tsr at a point p has a vector space structure. The union of all such
vector spaces is called the tensor bundle, and smooth sections of the bundle are called tensor fields Tsr (Rn );
that is, a tensor field is a smooth assignment of a tensor to each point in Rn .
The quantity g(X, Y ) is an example of a bilinear map that the reader will recognize as the usual dot product.
2.2.3 Definition A bilinear map g(X, Y ) ≡< X, Y > on vectors is called a real inner product if
2. g(X, X) ≥ 0, ∀X,
3. g(X, X) = 0 iff X = 0.
Since we assume g(X, Y ) to be bilinear, an inner product is completely specified by its action on ordered
pairs of basis vectors. The components gij of the inner product are thus given by
∂ ∂
g( i
, j ) = gij , (2.15)
∂x ∂x
where gij is a symmetric n × n matrix which we assume to be non-singular. By linearity, it is easy to see
∂ j ∂
that if X = ai ∂x i and Y = b ∂xj are two arbitrary vectors, then
In this sense, an inner product can be viewed as a generalization of the dot product. The standard Euclidean
inner product is obtained if we take gij = δij . In this case, the quantity g(X, X) =k X k2 gives the square
of the length of the vector. For this reason, gij is called a metric and g is called a metric tensor.
∂
Another interpretation of the dot product can be seen if instead one considers a vector X = ai ∂x i and a
= bj ai (dxj )( ∂x
∂
i)
= bj ai δij
= a i bi .
If we now define
bi = gij bj , (2.16)
we see that the equation above can be rewritten as
ai bj = gij ai bj ,
2.2. TENSORS 31
bi = g ij bj . (2.18)
We have mentioned that the tangent and cotangent spaces of Euclidean space at a particular point p are
isomorphic. In view of the above discussion, we see that the metric g can be interpreted on one hand as a
bilinear pairing of two vectors
g : Tp (Rn ) × Tp (Rn ) −→ R,
and on the other, as inducing a linear isomorphism
defined by
G[ X(Y ) = g(X, Y ), (2.19)
that maps vectors to covectors. To verify this definition is consistent with the action of lowering indices, let
∂ j ∂
X = ai ∂x i
i and Y = b ∂xj . We show that that G[ X = ai dx . In fact,
= ai bj dxi ( ∂x
∂
j ),
= ai bj δ i j ,
= ai bi = gij aj bi ,
= g(X, Y ).
for any 1-form α and tangent vector X. In Quantum Mechanics, it is common to use Dirac’s notation, in
which a linear functional α on a vector space V is called a bra-vector denoted by hα|, and a vector X ∈ V is
called a ket-vector, denoted by |Xi. The, action of a bra-vector on a ket-vector is defined by the bracket,
The mapping C : Tp∗ (Rn ) → R given by (α, X) 7→ hα|Xi = α(X) is called a contraction. In passing,
we introduce a related concept called the interior product, or contraction of a vector and a form. If α is a
(k + 1)-form and X a vector, we define
C(T ) = T i j hdxj | ∂x
∂
i i,
= T i j dxj ( ∂x
∂
i ),
= T i j δij ,
= T ii.
In other words, the contraction of the tensor is the trace of the n × n array that represents the tensor in the
given basis. The notion of raising and lowering indices as well as contractions can be extended to tensors of
all types. Thus, for example, we have
g ij Tiklm = T i klm .
A contraction between the indices i and l in the tensor above could be denoted by the notation
This is a very simple concept, but the notation for a general contraction is a bit awkward
because one needs
to keep track of the positions of the indices contracted. Let T be a tensor of type rs . A contraction Clk
yields a tensor of type r−1
s−1 . Let T be given in the form 2.12. Then,
where the “hat” means that these are excluded. Here is a very neat and most useful result. If S is a 2-tensor
with symmetric components Tij = Tji and A is a 2-tensor with antisymmetric components Aij = −Aji , then
the contraction
Sij Aij = 0 (2.24)
The short proof uses the fact that summation indices are dummy indices and they can be relabeled at will
by any other index that is not already used in an expression. We have
Sij Aij = Sji Aij = −Sji Aji = −Skl Akl = −Sji Aij = 0,
In terms of the vector space isomorphism between the tangent and cotangent space induced by the metric,
the gradient of a function f , viewed as a differential geometry vector field, is given by
or in components
(∇f )i ≡ ∇i f = g ij f,j , (2.26)
j
where f,j is the commonly used abbreviation for the partial derivative with respect to x .
In elementary treatments of calculus, authors often ignore the subtleties of differential 1-forms and tensor
products and define the differential of arc length as
x = r sin θ cos φ,
y = r sin θ sin φ,
z = r cos θ, (2.30)
Note: Technically speaking, Minkowski’s metric is not really a metric since η(X, X) = 0 does not imply
that X = 0. Non-zero vectors with zero length are called light-like vectors and they are associated with
particles that travel at the speed of light (which we have set equal to 1 in our system of units.)
The Minkowski metric ηµν and its matrix inverse η µν are also used to raise and lower indices in the space
in a manner completely analogous to Rn . Thus, for example, if A is a covariant vector with components
Aµ = (ρ, A1 , A2 , A3 ),
Aµ = η µν Aν
= (−ρ, A1 , A2 , A3 )
34 CHAPTER 2. DIFFERENTIAL FORMS
The alternating property is reminiscent of determinants of square matrices that change sign if any two
column vectors are switched. In fact, the determinant function is a model of an alternating bilinear map on
the space M2×2 of two by two matrices. Of course, for the definition above to apply, one has to view M2×2
as the space of column vectors.
2.2.7 Definition A 2-form φ is a map φ : T (Rn ) × T (Rn ) −→ R which is alternating and bilinear.
2.2.8 Definition Let α and β be 1-forms in Rn and let X and Y be any two vector fields. The wedge
product of the two 1-forms is the map α ∧ β : T (Rn ) × T (Rn ) −→ R, given by the equation
The proof of linearity on the second slot is quite similar and is left to the reader.
The wedge product of two 1-forms has characteristics similar to cross products of vectors in the sense
that both of these products anti-commute. This means that we need to be careful to introduce a minus sign
every time we interchange the order of the operation. Thus, for example, we have
if i 6= j, whereas
dxi ∧ dxi = −dxi ∧ dxi = 0
since any quantity that equals the negative of itself must vanish.
α = a dx + b dy,
β = c dx + d dy.
2.2. TENSORS 35
α ∧ β = ad dx ∧ dy + bc dy ∧ dx,
= ad dx ∧ dy − bc dx ∧ dy,
a b
= dx ∧ dy.
c d
The similarity between wedge products is even more striking in the next example, but we emphasize again
that wedge products are much more powerful than cross products, because wedge products can be computed
in any dimension.
2.2.11 Example For combinatoric reasons, it is convenient to label the coordinates as {x1 , x2 , x3 }. Let
dy ∧ dz = dx2 ∧ dx3 ,
dx ∧ dz = −dx1 ∧ dx3 ,
dx ∧ dy = dx1 ∧ dx2 .
If we consider vectors a = (a1 , a2 , a3 ) and b = (b1 , b2 , b3 ), we see that the result above can be written as
α ∧ β = (a × b)1 dx2 ∧ dx3 − (a × b)2 dx1 ∧ dx3 + (a × b)3 dx1 ∧ dx2 (2.37)
2.2.12 Example One could of course compute wedge products by just using the linearity properties. It
would not be as efficient as grouping into pairs, but it would yield the same result. For example, let
36 CHAPTER 2. DIFFERENTIAL FORMS
If we think of F as a matrix with components Fij , we know from linear algebra that we can write F uniquely
as a sum of a symmetric and an antisymmetric matrix, namely,
F = S + A,
1 1
= (F + F T ) + (F − F T ),
2 2
Fij = F(ij) + F[ij] ,
where,
1
F(ij) = (Fij + Fji ),
2
1
F[ij] = (Fij − Fji ),
2
are the completely symmetric and antisymmetric components. Since dxi ∧ dxj is antisymmetric, and the
contraction of a symmetric tensor with an antisymmetric tensor is zero, one may assume that the components
of the 2-form in equation 2.38 are antisymmetric as well. With this mind, we can easily find a formula using
wedges that generalizes the cross product to any dimension.
Let α = ai dxi and β = bi dxi be any two 1-forms in Rn , and Let X and Y be arbitrary vector fields.
Then
(α ∧ β)(X, Y ) = (ai dxi )(X)(bj dxj )(Y ) − (ai dxi )(Y )(bj dxj )(X)
= (ai bj )[dxi (X)dxj (Y ) − dxi (Y )dxj (X)]
= (ai bj )(dxi ∧ dxj )(X, Y ).
Because of the antisymmetry of the wedge product, the last of the above equations can be written as
n X
X n
α∧β = (ai bj − aj bi )(dxi ∧ dxj ),
i=1 j<i
1
= (ai bj − aj bi )(dxi ∧ dxj ).
2
In particular, if n = 3, the reader will recognize the coefficients of the wedge product as the components of
the cross product of a = a1 i + a2 j + a3 k and b = b1 i + b2 j + b3 k, as shown earlier.
Remark Quantities such as dx dy and dy dz which often appear in calculus II, are not really well defined.
What is meant by them are actually wedge products of 1-forms, but in reversing the order of integration,
the antisymmetry of the wedge product is ignored. In performing surface integrals, however, the surfaces
must be considered oriented surfaces and one has to insert a negative sign in the differential of surface area
component in the xz-plane as shown later in equation 2.83.
2.2. TENSORS 37
2.2.5 Determinants
The properties of n-forms are closely related to determinants, so it might be helpful to digress a bit and
review the fundamentals of determinants, as found in any standard linear algebra textbook such as [?]. Let
A ∈ Mn be an n × n matrix with column vectors
A = [v1 , v2 , . . . vn ]
2.2.13 Definition A function f : Mn → R is called multilinear if it is linear on each slot; that is,
2.2.14 Definition A function f : Mn → R is called alternating if it changes sign whenever any two
columns are switched; that is,
f [v1 , . . . , vi , . . . , vj , . . . , vn ] = −f [v1 , . . . , vj , . . . , vi , . . . vn ]
One can then prove that this defines the function uniquely. In particular, if A = (ai j), the determinant can
be expressed as
X
det(A) = sgn(π) a1π(1) a2π(2) . . . anπ(n) , (2.39)
π
where the sum is over all the permutations of {1, 2 . . . , n}. The determinant can also be calculated by the
cofactor expansion formula of Laplace. Thus, for example, the cofactor expansion along the entries on the
first row (a1 j ), is given by
X
det(A) = a1 k ∆k 1 , (2.40)
k
We set the Levi-Civita symbol with some or all the indices up, numerically equal to the permutation symbol
will all the indices down. The permutation symbols are useful in the theory of determinants. In fact, if
A = (ai j ) is an n × n matrix, then, equation (2.39) can be written as,
If one views the indices ik as labelling rows and jk as labelling columns of a matrix, we can represent the
completely antisymmetric symbol by the determinant,
i1
δj1 δji12 . . . δji1
i2 k
δ i2 i2
δji11 ij22...i j1 δj2 . . . δjk
...jk = (2.45)
k
. . . . . . . . . . . . . . . . . .
δ ik δ ik . . . δ ik
j1 j2 jk
Not surprisingly, the generalized Kronecker delta is related to a product of Levi-Civita symbols by the
equation
i1 i2 ...ik j1 j2 ...jk = δji11 ij22...i
...jk ,
k
(2.46)
which is evident since both sides are completely antisymmetric. In dimension 3, the only non-zero components
ij
of δkl are,
12 13 23 12 13 23
δ12 = δ13 = δ23 =1 δ21 = δ31 = δ32 = −1
21 31 32 21 31 32
δ21 = δ31 = δ32 =1 δ12 = δ13 = δ23 = −1.
b) ijk k
ijn = 2δn ,
c) ijk
ijk = 3!
Proof For part (a), we compute the determinant by cofactor expansion on the first row
i i
δni
δi δm
ijk j δj δj
imn = δi m n
δ k δ k δ k
i m n
j j
j j
δ δn i δi δnj i δi
j
δm
= δii m
k − δ + δ
δm δnk m k
δi δnk n k k
δi δm
j j
δ δnj δm δnj δnj δm j
= 3 m k − +
δm δnk δm k
δnk δnk δm k
j j
δ δnj δm δnj
= (3 − 1 − 1) m k =
δm δnk δm k
δnk
Here we used the fact that the contraction δii is just the trace of the identity matrix and the observation
that we had to transpose columns in the last determinant in the next to last line. for part (b) follows easily
2.2. TENSORS 39
ijk jk
inj = δjn ,
From this, part (c) is obvious. With considerably more effort, but inductively following the same scheme,
one can establish the general formula,
i ...i
i1 ...ik ,ik+1 ...in i1 ...ik ,jk+1 ...jn = k!δjk+1 n
k+1 ...jn
. (2.47)
a · b = δ ij ai bj = ai bi , (a × b)k = k ij ai bj (2.48)
2. Wedge product
α ∧ β = k ij (a × b)k dxi ∧ dxj . (2.49)
3. Triple product
a · (b × c) = δij ai (b × c)l ,
= δij ai j kl bk cl ,
= ikl ai bk cl ,
a · (b × c) = det([abc]), (2.50)
= (a × b) · c (2.51)
[a × (b × c)]l = l mn am (b × c)n
= l mn am (n jk bj ck )
= l mn n jk am bj ck )
= mnl jkn am bj ck )
k j
= (δm δl − δlj δm
k
)am bj ck
= bl am cm − cl am bm .
(a)
(∇ × ∇f )i = i jk ∇j ∇f = 0,
∇ × ∇f = 0 (2.56)
(b)
∇ · (∇ × A) = δ ij ∇i (∇ × a)j ,
= δ ij ∇i j kl ∇k al ,
= jkl ∇i ∇j ak ,
∇ · (∇ × A) = 0 (2.57)
where in the last step in the two items above we use the fact that a contraction of two symmetric
with two antisymmetric indices is always 0.
(c) The same steps as in the bac-cab identity give
[∇ × (∇ × A)]l = ∇l (∇m am ) − ∇m ∇m al ,
∇ × (∇ × A) = ∇(∇ · A) − ∇2 A,
where ∇2 A means the Laplacian of each component of A.
This last equation is crucial in the derivation of the wave equation for light from Maxwell’s equations
for the electromagnetic field.
2.2. TENSORS 41
2.2.7 n-Forms
2.2.17 Definition Let α1 , α2 , α2 , be one forms, and X1 , X2 , X3 ∈ X . Let π be the set of permutations
of {1, 2, 3}. Then
X
(α1 ∧ α2 ∧ α3 )(X1 , X2 , X3 ) = sign(π)α1 (Xπ(1) )α2 (Xπ(2) )α3 (Xπ(3) ),
π
= ijk α1 (Xi )α2 (Xj )α3 (Xk ).
where we assume that the wedge product of three 1-forms is associative but alternating in the sense that
if one switches any two differentials, then the entire expression changes by a minus sign. There is nothing
really wrong with using definition (2.58). This definition however, is coordinate-dependent and differential
geometers prefer coordinate-free definitions, theorems and proofs.
We can easily extend the concepts above to higher order forms.
where π is the set of permutations of {1, . . . , k}. A skew-symmetry covariant tensor of rank k at p, is called
a k-form at p. denote by Λk(p) (Rn ) the space of k-forms at p ∈ Rn . This vector space has dimension
n n!
dim Λkp (Rn ) = =
k k!(n − k)!
for k ≤ n and dimension 0 for k > n. We identify Λ0(p) (Rn ) with the space of C ∞ functions at p. The union
of all Λk(p) (Rn ) as p ranges through all points in Rn is called the bundle of k-forms and will be denoted by
[
Λk (Rn ) = Λkp (Rn ).
p
Sections of the bundle are called k-forms and the space of all sections is denoted by
A section α ∈ Ωk of the bundle technically should be called k-form field, but the consensus in the literature
is to call such a section simply a k-form. In local coordinates, a k-form can be written as
2.2.20 Definition The alternation map A : Tk0 (Rn ) → Tk0 (Rn ) is defined by
1 X
At(e1 , . . . , ek ) = (signπ)t(eπ(1) , . . . , eπ(k) ).
k! π
42 CHAPTER 2. DIFFERENTIAL FORMS
α ∧ β = (−1)kl β ∧ α. (2.62)
Now, for a little combinatorics. Factorials are unavoidable due to the permutation attributes of the wedge
product. The convention here follows Marsden [?] and Spivak [?], which reduces proliferation of factorials
later. Let us count the number of linearly independent differential forms in Euclidean space. More specifically,
we want to find a basis for the vector space of k-forms in R3 . As stated above, we will think of 0-forms
as being ordinary functions. Since functions are the “scalars”, the space of 0-forms as a vector space has
dimension 1.
R2 Forms Dim
0-forms f 1
1-forms f dx1 , gdx2 2
2-forms f dx1 ∧ dx2 1
R3 Forms Dim
0-forms f 1
1-forms f1 dx1 , f2 dx2 , f3 dx3 3
2-forms f1 dx2 ∧ dx3 , f2 dx3 ∧ dx1 , f3 dx1 ∧ dx2 3
3-forms f1 dx1 ∧ dx2 ∧ dx3 1
∂ ∂
T = T i1 ,...ir j1 ,...js ,k1 ,...,kp ⊗ ... ⊗ ⊗ dxj1 ⊗ · · · ⊗ dxjs ∧ dxk1 ∧ . . . ∧ dxkp . (2.63)
∂xi1 ∂xir
Thus, for example, the quantity
Ωi j = 12 Ri jkl dxk ∧ dxl
1
would be called the components of a 1 -valued 2-form
∂
Ω = Ωi j ⊗ dxj .
∂xi
The notion of the wedge product can be extended to tensor-valued forms using tensor products on the
tensorial indices and wedge products on the differential form indices.
2.3.1 Definition Let α be a one form in Rn . The differential dα is the two-form defined by
∂ i ∂ ∂ i ∂
dα(X, Y ) = fi dx − fi dx ,
∂xj ∂xk ∂xk ∂xj
∂ ∂
= j
(fi δki ) − k
(fi δji ),
∂x ∂x
∂ ∂ ∂fk ∂fj
dα , = −
∂xj ∂xj ∂xj ∂xk
The definition 2.64 of a differential of a 1-form can be refined to provide a coordinate-free definition in
general manifolds (see 6.28,) and it can be extended to differentials of m-forms. For now, the computation
immediately above suffices to motivate the following coordinate dependent definition (for a coordinate-free
definition for general manifolds, see ??):
2.3.2 Definition Let α be an m-form, given in coordinates as in equation (2.60). The exterior derivative
of α is the (m + 1)-form dα given by
2.3.3 Theorem
a) d : Ωm −→ Ωm+1
b) d2 = d ◦ d = 0
c) d(α ∧ β) = dα ∧ β + (−1)p α ∧ dβ ∀α ∈ Ωp , β ∈ Ωq (2.66)
Proof
a) Obvious from equation (2.65).
b) First, we prove the proposition for α = f ∈ Ω0 . We have
∂f
d(dα) = d( )
∂xi
∂2f
= dxj ∧ dxi
∂xj ∂xi
1 ∂2f ∂2f
= [ j i− ]dxj ∧ dxi
2 ∂x ∂x ∂xi ∂xj
= 0.
44 CHAPTER 2. DIFFERENTIAL FORMS
Now, suppose that α is represented locally as in equation (2.60). It follows from equation 2.65, that
By definition,
α ∧ β = Ai1 ...ip Bj1 ...jq (dxi1 ∧ . . . ∧ dxip ) ∧ (dxj1 ∧ . . . ∧ dxjq ).
Now, we take the exterior derivative of the last equation, taking into account that d(f g) = f dg + gdf for
any functions f and g. We get
The (−1)p factor comes into play since in order to pass the term dBji ...jp through p number of 1-forms of
type dxi , one must perform p transpositions.
2.3.1 Pull-back
2.3.5 Theorem
a) F ∗ (gα1 ) = (g ◦ F ) F ∗ α,
b) F (α1 + α2 ) = F ∗ α1 + F ∗ α2 ,
∗
(2.69)
c) F ∗ (α ∧ β) = F ∗ α ∧ F ∗ β,
d) F ∗ (dα) = d(F ∗ α.)
Part (d) is encapsulated in the commuting diagram in figure 2.2.
Fig. 2.2: d F ∗ = F ∗ d
Proof Part (a) is basically the definition for the case of 0-forms and part (b) is clear from the linearity of
the push-forward. We leave part (c) as an exercise and prove part (d). In the case of a 0-form, let g, be a
2.3. EXTERIOR DERIVATIVES 45
function and X a vector field in Rm . By a simple computation that amounts to recycling definitions, we
have:
d(F ∗ g) = d(g ◦ F ),
(F ∗ dg)(X) = dg(F∗ X) = (F∗ X)(g),
= X(g ◦ F ) = d(g ◦ F )(X),
F ∗ dg = d(g ◦ F ),
so, F ∗ (dg) = d(F ∗ g) is true by the composite mapping theorem. Let α be a k-form
α = Ai1 ,...ik dy i1 ∧ . . . dy ik ,
so that
dα = (dAi1 ,...ik ) ∧ dy i1 . . . dy ik .
Then, by part (c),
F ∗ α = (F ∗ Ai1 ,...ik ) F ∗ dy i1 ∧ . . . F ∗ dy ik ,
d(F ∗ α) = dF ∗ (Ai1 ,...ik ) ∧ F ∗ dy i1 ∧ . . . F ∗ dy ik ,
= F ∗ (dAi1 ,...ik ) ∧ F ∗ dy i1 ∧ . . . F ∗ dy ik ,
= F ∗ (dα).
To connect with advanced calculus, suppose that locally the mapping F is given by y k = f k (xi ). Then
the pullback of the form dg given the formula above F ∗ dg = d(g ◦ F ) is given in local coordinates by the
chain rule
∂g j
F ∗ dg = dx .
∂xj
In particular, the pull-back of local coordinate functions is given by
∂y i j
F ∗ (dy i ) = dx . (2.70)
∂xj
Thus, pullback for the basis 1-forms dy k is yet another manifestation of the differential as a linear map
represented by the Jacobian
∂y k
dy k = dxi . (2.71)
∂xi
In particular, if m = n,
dΩ = dy 1 ∧ dy 2 ∧ . . . ∧ dy n ,
∂y 1 ∂y 2 ∂y n
= . . . dxi1 ∧ dxi2 ∧ . . . dxin ,
∂xi1 ∂xi2 ∂xin
∂y 1 ∂y 2 ∂y n
= i1 i2 ...in i1 i2 . . . in dx1 ∧ dx2 ∧ . . . dxn ,
∂x ∂x ∂x
= |J| ∧ dx2 ∧ . . . dxn . (2.72)
gives rise to the integrand that appears in the change of variables theorem for integration. More explicitly,
let R ∈ Rn be a simply connected region, F be a mapping F : R ∈ Rn → Rm , with m ≥ n. If ω is a k-
form in Rm , then Z Z
ω= F ∗ω (2.73)
F (R) R
46 CHAPTER 2. DIFFERENTIAL FORMS
If F : Rn → Rn is a diffeomorphism, one can push-forward forms with the inverse of the pull-back F∗ =
(F −1 )∗ .
Z Z
F · dx = ω,
C
ZC
= ω,
φ(I)
Z
= φ∗ ω,
I
dxi
Z
= f i (x(t))
dt,
dt
ZI
dx
= F(x(t)) dt
I dt
If the polar coordinates map is denoted by F : R2 → R2 , then equation 2.74 is just the explicit expression
for the pullback of F ∗ ( f dA).
2.3.8 Example Polar coordinates are just a special example of the general transformation in R2 given
by,
∂x ∂x
x = x(u, v), dx = du + dv,
∂u ∂u
∂y ∂y
y = y(u, v), dy = du + dv,
∂u ∂u
2.3. EXTERIOR DERIVATIVES 47
for which
∂x ∂x
φ ∗ (dx ∧ dy) = ∂u ∂v
du ∧ dv (2.75)
∂y
∂u ∂y
∂v
2.3.10 Remark
1. The differential of area in polar coordinates is of course a special example of the change of coordinate
theorem for multiple integrals as indicated above.
2. As shown in equation 2.32 the metric in spherical coordinates is given by
ds2 = dr2 + r2 dθ2 + r2 sin2 θ dφ2 ,
so the differential of volume is
p
dV = det g dr ∧ dθ ∧ dφ,
2
= r sin θ dr ∧ dθ ∧ dφ.
48 CHAPTER 2. DIFFERENTIAL FORMS
dα = ( ∂P
∂x dx +
∂P
∂y ) ∧ dx + ( ∂Q
∂x dx +
∂Q
∂y ) ∧ dy
∂P ∂Q
= ∂y dy ∧ dx + ∂x dx ∧ dy
= ( ∂Q ∂P
∂x − ∂y ) dx ∧ dy. (2.76)
This example is related to Green’s theorem in R2 . For convenience, we include here a proof of Green’s
Theorem in a special case.
We say that a region D in the plane is of type I if it is enclosed between the graphs of two continuous
functions of x. The region inside the simple closed curve in figure 2.3 bounded by f1 (x) and f2 (x), between
a and b, is a region of type I. A region in the plane is of type II if it lies between two continuous functions
of y. The region in 2.3 bounded between c ≤ y ≤ d, would be a region of type II.
We first prove that for a type I region such as the one bounded between a and b shown in 2.3, we have
I Z Z
∂P
P dx = − dA (2.78)
C D ∂y
Where C comprises the curves C1 , C2 , C3 and C4. By the fundamental theorem of calculus, we have on the
right,
Z Z Z b Z f2 (x)
∂P ∂P
dA = dydx,
D ∂y a f1 (x) ∂y
Z b
= [P (x, f2 (x)) − P (x, f1 (x))] dx.
a
On the left, the integrals along C2 and C4 vanish, since there is no variation on x. The integral along C3 is
traversed in opposite direction of C1 , so we have,
2.3. EXTERIOR DERIVATIVES 49
I Z Z Z Z
P (x, y) dx = + +
+ P (x, y) dx,
C C1 C2 C3 C4
Z Z
= P (x, y) dx − P (x, y) dx,
C1 C3
Z b Z b
= P (x, f1 (x)) dx − P (x, f2 (x)) dx
a a
This establishes the veracity of equation 2.78 for type I regions. By a completely analogous process on type
II regions, we find that I Z Z
∂Q
Q dy = dA. (2.79)
C D ∂x
The theorem follows by subdividing R into a grid of regions of both types, all oriented in the same direction
as shown on the right in figure 2.3. Then one applies equations 2.78 or 2.79, as appropriate, for each of the
subdomains. All contributions from internal boundaries cancel since each is traversed twice, each in opposite
directions. All that remains of the line integrals is the contribution along the boundary δR.
Let α = P dx + Q dy. Comparing with equation 2.76, we can write Green’s theorem in the form
Z Z Z
α= dα. (2.80)
C D
It is possible to extend Green’s Theorem to more complicated regions that are not simple connected. Green’s
theorem is a special case in dimension of two of Stoke’s theorem.
Proof The proof can be done by pulling back to the uv-plane and using the chain rule, thus allowing us to
use Green’s theorem. Let ω = fi dxi and S be parametrized by xi = xi (uα ), where (u1 , u2 ) ∈ R ⊂ R2 . We
assume that the boundary of R is a simple closed curve. Then
Z Z
ω= fi dxi ,
C δS
∂xi
Z
= fi α
duα ,
δR ∂u
∂xi
Z Z
∂
= β
(fi α ) duβ ∧ duα ,
R ∂u ∂u
∂fi ∂xk ∂xi ∂ 2 xi
Z Z
= k ∂uβ ∂uα
+ f i β ∂uα
duβ ∧ duα ,
R ∂x ∂u
∂fi ∂xk ∂xi
Z Z
= k β α
duβ ∧ duα ,
R ∂x ∂u ∂u
∂fi ∂xk
Z Z i
β ∂x
= k β
du ∧ duα
R ∂x ∂u ∂uα
Z Z Z Z
∂fi k i
= k
dx ∧ dx = dfi ∧ dxi
S ∂x S
Z Z
= dω.
S
50 CHAPTER 2. DIFFERENTIAL FORMS
We present a less intuitive but far more elegant proof. The idea is formally the same, namely, we pull-back
to the plane by formula 2.73, apply Green’s theorem in the form given in equation 2.80, and then use the
fact that the pull-back commutes with the differential as in theorem 2.69.
Let φ : R ⊂ R2 → S denote the surface parametrization map. Assume that φ−1 (δS) = δ(φ−1 S), that is, the
inverse of the boundary of S is the boundary of the domain R. Then,
Z Z Z
∗
ω= φ ω= φ∗ ω,
δS φ−1 (δS) δ(φ−1 S)
Z Z
= d(φ∗ ω),
φ−1 S
Z Z
= φ∗ (dω),
φ−1 S
Z
= dω.
S
The proof of Stokes’ theorem presented here is one of those cases mentioned in the preface, where we have
simplified the mathematics for the sake of clarity. Among other things, a rigorous proof requires one to
quantify what is meant by the boundary (δS) of a region. The process involves either introducing simplices
(generalized segments, triangles, tetrahedra...) or singular cubes (generalized segments, rectangles, cubes...).
The former are preferred in the treatment of homology in algebraic topology, but the latter are more natural
to use in the context of integration on manifolds with boundary. A singular n-cube in Rn is the image under
a continuous map,
I n : [0, 1]n → Rn ,
of the Cartesian product of n copies of the unit interval [0, 1]. The idea is to divide the region S into formal
finite sums of singular cubes, called chains. One then introduces a boundary operator δ, that maps a singular
n-cube and hence n-chain, into P an (n − 1)-singular cube or (n − 1)-chain. Thus, in R3 for example, the
boundary of a cube, is the sum ci Fi of the six faces with a judicious choice of coefficients ci ∈ {−1, 1}.
With an appropriate scheme to label faces of singular cube and a corresponding definition of the boundary
map, one proves that δ ◦ δ = 0. For a thorough treatment, see the beautiful book Calculus on Manifolds by
M. Spivak [?].
2.3.13 Example Let α = M (x, y)dx + N (x, y)dy, and suppose that dα = 0. Then, by the previous
example,
dα = ( ∂N ∂M
∂x − ∂y ) dx ∧ dy.
Thus, dα = 0 iff Nx = My , which implies that N = fy and Mx for some function f (x, y). Hence,
α = fx dx + fy dy = df.
The reader should also be familiar with this example in the context of exact differential equations of first
order and conservative force fields.
2.3.15 Definition A differential form α is called exact if there exists a form β such that α = dβ.
Since d ◦ d = 0, it is clear that an exact form is also closed. The converse need not be true. The standard
counterexample is the form,
−y dx + x dy
ω= (2.82)
x2 + y 2
A short computation shows that dω = 0, so ω is closed. Let θ = tan−1 (y/x) be the angle in polar coordinates.
One can recognize that ω = dθ, but this is only true in R2 − L, where L is the non-negative x-axis,
2.4. THE HODGE ? OPERATOR 51
L = {(x, 0) ∈ R2 |x ≥ 0}. If one computes the line integral from (−1, 0) to (1, 0) along the top half of the
unit circle, the result is π. But the line integral along the bottom half of the unit circle gives −π. The
integral is therefore not path independent, so ω 6= dθ on any region that contains the origin. If one tries to
find another C 1 function f such that ω = df , one can easily show that f = θ + const, which is not possible
along L.
On the other hand, if one imposes the topological condition that the space is Contractible, then the
statement is true. A contractible space is one that can be deformed continuously to an interior point. We
have the following,
2.3.16 Poincaré Lemma. In a contractible space (such as Rn ), if a differential form is closed, then it is
exact.
To prove this lemma we need much more machinery than we have available at this point. We present the
proof in ??.
isomorphism then transforms a contravariant vector to a covariant vector. In terms of components, the
isomorphism is provided by the Euclidean metric that maps the components of a contravariant vector with
indices up to a covariant vector with indices down.
Another interesting example is provided by the spaces Λ1p (R3 ) and Λ2p (R3 ), both of which have dimension
3. It follows that these two spaces must be isomorphic. In this case the isomorphism is given as follows:
dx 7−→ dy ∧ dz
dy 7−→ −dx ∧ dz
dz 7−→ dx ∧ dy
(2.83)
n
More generally,
we have seen that the dimension of the space of k-forms in R is given by the binomial
n
coefficient k . Since
n n n!
= = ,
k n−k k!(n − k)!
it must be true that
Λkp (Rn ) ∼
= Λn−k
p (Rn ). (2.84)
To describe the isomorphism between these two spaces, we introduce the following generalization of deter-
minants,
2.4.1 Definition . Let φ : Rn → Rn be a linear map. The unique constant det φ such that,
φ∗ : Λn (Rn ) → Λn (Rn )
satisfies,
φ∗ ω = (det φ) ω, (2.85)
52 CHAPTER 2. DIFFERENTIAL FORMS
for all n-forms, is called the determinant of φ. This is congruent with the standard linear algebra formula
2.43, since in a particular basis, the Jacobian of a linear map is the same as the matrix the represents the
linear map in that basis. Let, g(X, Y ) be an inner product and {e1 , . . . , en } be an orthonormal basis with
dual forms {θ1 , . . . θn }. The element of arc length is, the bilinear symmetric tensor
ds2 = gij θi ⊗ θj .
2.4.2 Definition Let g be the matrix representing the components of the metric in Rn . The Hodge ?
operator is the linear isomorphism ? : Λnp (Rn ) −→ Λn−k p (Rn ) defined in standard local coordinates by the
equation, √
det g i1 ...ik
? (dxi1 ∧ . . . ∧ dxik ) = ik+1 ...in dx
ik+1
∧ . . . ∧ dxin , (2.87)
(n − k)!
√
For flat Euclidean space det g = 1, so the factor in the definition may appear superfluous. However, when
we consider more general Riemannian manifolds, we will have to be more careful with raising and lowering
indices with the metric, and take into account that the Levi-Civita √ symbol is not a tensor but something
slightly more complicated called a tensor density. Including the det g is done in anticipation of this more
general setting later. Since the forms dxi1 ∧ . . . ∧ dxik constitute a basis of the vector space Λkp (Rn ) and
the ? operator is assumed to be a linear map, equation (2.87) completely specifies the map for all k-forms.
In particular, if the components of a dual of a form are equal to the components of the form, the tensor is
called self-dual. Of course, this can only happen if the tensor and its dual are of the same rank.
A metric g on Rn induces an inner product on Λk (Rn ) as follows. Let {e1 , , . . . en } by an orthonormal
basis with dual basis θ1 , . . . , θn . If α, β ∈ Λk (Rn ), we can write
Clearly,
α ∧ ?β = ?α ∧ β
When it is evident that the inner product is the induced inner product on Λk (Rn ) the indicator (k) is often
suppressed. An equivalent definition of the induced inner product of two k-forms is given by
Z
< α, β >= (α ∧ ?β) d Ω. (2.90)
as
dx 0 1 dx
? = .
dy −1 0 dy
The reader might wish to peek at the symplectic matrix ?? in the discussion in chapter 5 on conformal
mappings. Given functions u = u(x, y) and v = v(x, y), let ω = u dx − v dy. Then,
dω = −(uy + vx ) dx ∧ dy, dω = 0 ⇒ uy = −vx ,
hence (2.96)
d?ω = (ux − uy ) dx ∧ dy, ?dω = 0 ⇒ ux = vy .
Thus, the equations dω = 0 and d ? ω = 0 are equivalent to the Cauchy-Riemann equations for a holomorphic
function f (z) = u(x, y) + iv(x, y). On the other hand,
du = ux dx + uy dy,
dv = vx dx + uy dy,
so the determinant of the Jacobian of the transformation T : (x, y) → (u, v)), with the condition above on
ω, is given by,
u uy
|J| = x = u2x + u2y = vx2 + vy2 .
vx vy
If |J| =
6 0, we can set ux = R cos φ, uy = R sin φ, for some R and some angle φ. Then,
R 0 cos φ sin φ
|J| = .
0 R − sin φ cos φ
Thus, the transformation is given by the composition of a dilation and a rotation. A more thorough discussion
of this topic is found in the section of conformal maps in chapter 5.
We leave it to the reader to complete the computation of the action of the ? operator on the other basis
forms. The results are
and
? (dx1 ∧ dx2 ∧ dx3 ) = 1. (2.99)
In particular, if f : R3 −→ R is any 0-form (a function), then,
2.4.5 Example Let α = a1 dx1 a2 dx2 + a3 dx3 , and β = b1 dx1 b2 dx2 + b3 dx3 . Then,
The previous examples provide some insight on the action of the ∧ and ? operators. If one thinks of the
quantities dx1 , dx2 and dx3 as analogous to ~i, ~j and ~k, then it should be apparent that equations (2.97) are
the differential geometry versions of the well-known relations
i = j × k,
j = −i × k,
k = i × j.
2.4.6 Example In Minkowski space the collection of all 2-forms has dimension 42 = 6. The Hodge ?
operator in this case splits Ω2 (M1,3 ) into two 3-dim subspaces Ω2± , such that ? : Ω2± −→ Ω2∓ .
More specifically, Ω2+ is spanned by the forms {dx0 ∧ dx1 , dx0 ∧ dx2 , dx0 ∧ dx3 }, and Ω2− is spanned by the
forms {dx2 ∧ dx3 , −dx1 ∧ dx3 , dx1 ∧ dx2 }. The action of ? on Ω2+ is
?(dx0 ∧ dx1 ) = 1 01
2 kl dx
k
∧ dxl = −dx2 ∧ dx3 ,
?(dx0 ∧ dx2 ) = 1 02
2 kl dx
k
∧ dxl = +dx1 ∧ dx3 ,
?(dx0 ∧ dx3 ) = 1 03
2 kl dx
k
∧ dxl = −dx1 ∧ dx2 ,
and on Ω2− ,
?(+dx2 ∧ dx3 ) = 1 23
2 kl dx
k
∧ dxl = dx0 ∧ dx1 ,
?(−dx1 ∧ dx3 ) = 1 13
2 kl dx
k
∧ dxl = dx0 ∧ dx2 ,
?(+dx1 ∧ dx2 ) = 1 12
2 kl dx
k
∧ dxl = dx0 ∧ dx3 .
In verifying the equations above, we recall that the Levi-Civita symbols that contain an index with value 0
in the up position have an extra minus sign as a result of raising the index with η 00 . If F ∈ Ω2 (M ), we will
2.4. THE HODGE ? OPERATOR 55
formally write F = F+ + F− , where F± ∈ Ω2± . We would like to note that the action of the dual operator
on Ω2 (M ) is such that ? : Ω2 (M ) −→ Ω2 (M ), and ?2 = −1. In a vector space a map like ?, with the
property ?2 = −1 is called a linear involution of the space. In the case in question, Ω2± are the eigenspaces
corresponding to the +1 and -1 eigenvalues of this involution. It is also worthwhile to calculate the duals of
1-forms in M1,3 . The results are,
2.4.2 Laplacian
Classical differential operators that enter in Green’s and Stokes’ Theorems are better understood as
special manifestations of the exterior differential and the Hodge ? operators in R3 . Here is precisely how
this works:
(? d ?) df = ∇ · ∇f = ∇2 f. (2.107)
The Laplacian definition here is consistent with 2.94 because in the case of a function f , that is, a 0-form,
δf = 0 so ∆f = δdf . The results above can be summarized in terms of short exact sequence called the de
Rham complex as shown in figure 2.4. The sequence is called exact because successive application of the
differential operator gives zero. That is, d ◦ d = 0. Since there are no 4-forms in R3 , the sequence terminates
as shown. If one starts with a function in Ω0 (R3 ), then (d ◦ d)f = 0 just says that ∇ × ∇f = 0, as in the case
of conservative vector fields. If instead, one starts with a one form α in Ω1 (R3 ), corresponding to a vector
56 CHAPTER 2. DIFFERENTIAL FORMS
field A, then (d ◦ d)f = 0 says that ∇ · (∇ × A) = 0, as in the case of incompressible vector fields. If one
starts with a function, but instead of applying the differential twice consecutively, one “hops” in between
with the Hodge operator, the result is the Laplacian of the function.
If we denote by R a simply connected closed region in Euclidean space whose boundary is δR , then
in terms of forms, the fundamental theorem of calculus, Stokes’ theorem (See ref 2.81), and the divergence
theorem in R3 can be expressed by a single generalized Stokes’ theorem.
Z Z Z
ω= dω. (2.108)
δR R
We find it irresistible to point out that if one defines a complex one-form,
This is Cauchy’s integral theorem. We should also point out the tantalizing resemblance of equations 2.96
to Maxwell’s equations in the section that follows.
dF = 0,
d?F = 4π ? J. (2.115)
2.4. THE HODGE ? OPERATOR 57
Proof The proof is by direct computation using the definitions of the exterior derivative and the Hodge ?
operator.
∂Ex ∂Ex
dF = − ∧ dx2 ∧ dt ∧ dx1 − ∧ dx3 ∧ dt ∧ dx1 +
∂x2 ∂x3
∂Ey ∂Ey
− 1 ∧ dx1 ∧ dt ∧ dx2 − ∧ dx3 ∧ dt ∧ dx2 +
∂x ∂x3
∂Ez ∂Ez
− 1 ∧ dx1 ∧ dt ∧ dx3 − ∧ dx2 ∧ dt ∧ dx3 +
∂x ∂x2
∂Bz ∂Bz
∧ dt ∧ dx1 ∧ dx2 − ∧ dx3 ∧ dx1 ∧ dx2 −
∂t ∂x3
∂By ∂By
∧ dt ∧ dx1 ∧ dx3 − ∧ dx2 ∧ dx1 ∧ dx3 +
∂t ∂x2
∂Bx ∂Bx
∧ dt ∧ dx2 ∧ dx3 + ∧ dx1 ∧ dx2 ∧ dx3 .
∂t ∂x1
Collecting terms and using the antisymmetry of the wedge operator, we get
∂Bx ∂By ∂Bz
dF = ( + + ) dx1 ∧ dx2 ∧ dx3 +
∂x1 ∂x2 ∂x3
∂Ey ∂Ez ∂Bx
( 3 − − ) dx2 ∧ dt ∧ dx3 +
∂x ∂x2 ∂t
∂Ez ∂Ex ∂By
( 1 − − ) dt ∧ dx1 ∧ x3 +
∂x ∂dx3 ∂t
∂Ex ∂Ey ∂Bz
( 2 − − ) dx1 ∧ dt ∧ dx2 .
∂x ∂x1 ∂t
Therefore, dF = 0 iff
∂Bx ∂By ∂By
1
+ 2
+ = 0,
∂x ∂x ∂x3
which is the same as
∇ · B = 0,
and
∂Ey ∂Ez ∂Bx
− − = 0,
∂x3 ∂x2 ∂t
∂Ez ∂Ex ∂By
− − = 0,
∂x1 ∂x3 ∂t
∂Ex ∂Ey ∂Bz
− − = 0,
∂x2 ∂x1 ∂t
which means that
∂B
−∇×E− = 0. (2.116)
∂t
To verify the second set of Maxwell equations, we first compute the dual of the current density 1-form (2.114)
using the results from example 2.4.1. We get
? J = [−ρdx1 ∧ dx2 ∧ dx3 + J1 dx2 ∧ dt ∧ dx3 + J2 dt ∧ dx1 ∧ dx3 + J3 dx1 ∧ dt ∧ dx2 ]. (2.117)
We could now proceed to compute d ? F , but perhaps it is more elegant to notice that F ∈ Ω2 (M ), and so,
according to example (2.4.1), F splits into F = F+ + F− . In fact, we see from (2.112) that the components
of F+ are those of −E and the components of F− constitute the magnetic field vector B. Using the results
of example (2.4.1), we can immediately write the components of ?F :
1
?F = Bx dt ∧ dx1 + By dt ∧ dx2 + Bz dt ∧ dx3 +
2!
Ez dx1 ∧ dx2 − Ey dx1 ∧ dx3 + Ex dx2 ∧ dx3 ], (2.118)
58 CHAPTER 2. DIFFERENTIAL FORMS
or equivalently,
0 Bx By By
?
−Bx 0 Ez −Ey
Fµν =
−By
. (2.119)
−Ez 0 Ex
−Bz Ey −Ex 0
Effectively, the dual operator amounts to exchanging
E 7−→ −B
B 7−→ +E,
in the left hand side of the first set of Maxwell equations. We infer from equations (2.116) and (2.117) that
∇ · E = 4πρ
and
∂E
∇×B− = 4πJ.
∂t
Most standard electrodynamic textbooks carry out the computation entirely tensor components, To connect
with this approach, we should mention that it F µν represents the electromagnetic tensor, then the dual
tensor is √
? det g
Fµν = µνστ F στ . (2.120)
2
Since dF = 0, in a contractible region there exists a one form A such that F = dA. The form A is called the
4-vector potential. The components of A are,
A = Aµ dxµ ,
Aµ = (φ, A) (2.121)
where φ is the electric potential and A the magnetic vector potential. The components of the electromagnetic
tensor F are given by
∂Aν ∂Aµ
Fµν = µ
− , (2.122)
∂x ∂xν
The classical electromagnetic Lagrangian is
1
LEM = − Fµν F µν + J µ Aµ , (2.123)
4
with corresponding Euler-Lagrange equations
∂ ∂L ∂L
− = 0. (2.124)
∂xµ ∂Aµ ∂Aµ
∂Aµ
To carry out the computation we first use the Minkowski to write the Lagrangian with the indices down.
The key is to keep in mind that Aµ,ν are treated as independent variables, so the derivatives of Aα,β vanish
unless µ = α and ν = β. We get,
∂L 1 ∂L
=− (Fαβ F αβ ),
∂(Aµ,ν ) 4 ∂(Aµ,ν )
1 ∂L
=− (Fαβ Fλσ η αλ η βσ ),
4 ∂(Aµ,ν )
1
= − η αλ η βσ [Fαβ (δλµ δ µ σ − δσµ δ µ λ) + Fλσ (δαµ δ µ β − δβµ δ µ α)
4
1 αµ βν
= − [η η Fαβ + η µλ η νσ Fλσ − η αν η βµ Fαβ − η νλ η µσ (Fλσ ],
4
1 µν
= − [F + F µν − F νµ − F νµ ],
4
= −F µν .
2.4. THE HODGE ? OPERATOR 59
Connections
3.1 Frames
This chapter is dedicated to professor Arthur Fischer. In my second year as an undergraduate at Berkeley,
I took the undergraduate course in differential geometry which to this day is still called Math 140. The driving
force in my career was trying to understand the general theory of relativity, which was only available at the
graduate level. However, the graduate course (Math 280 at the time) read that the only prerequisite was
Math 140. So I got emboldened and enrolled in the graduate course taught that year by Dr. Fischer. The
required book for the course was the classic by Adler, Bazin, Schiffer. I loved the book; it was definitely
within my reach and I began to devour the pages with the great satisfaction that I was getting a grasp of the
mathematics and the physics. On the other hand, I was completely lost in the course. It seemed as if it had
nothing to do with the material I was learning on my own. Around the third week of classes, Dr. Fischer
went through a computation with these mysterious operators, and upon finishing the computation he said if
we were following, he had just derived the formula for the Christoffel symbols. Clearly, I was not following,
they looked nothing like the Christoffel symbols I had learned from the book. So, with great embarrassment
I went to his office and explained my predicament. He smiled, apologized when he did not need to, and
invited me to 1-1 sessions for the rest of the two-semester course. That is how I got through the book he
was really using, namely Abraham-Marsden. I am forever grateful.
As noted in Chapter 1, the theory of curves in R3 can be elegantly formulated by introducing orthonormal
triplets of vectors which we called Frenet frames. The Frenet vectors are adapted to the curves in such a
manner that the rate of change of the frame gives information about the curvature of the curve. In this
chapter we will study the properties of arbitrary frames and their corresponding rates of change in the
direction of the various vectors in the frame. These concepts will then be applied later to special frames
adapted to surfaces.
3.1.1 Definition A coordinate frame in Rn is an n-tuple of vector fields {e1 , . . . , en } which are linearly
independent at each point p in the space.
In local coordinates {x1 , . . . , xn }, we can always express the frame vectors as linear combinations of the
standard basis vectors
n
X ∂
ei = Aj i j = ∂j Aji , (3.1)
j=1
∂x
∂
where ∂j = ∂x j. Placing the basis vectors ∂j on the left is done to be consistent with the summation
convention, keeping in mind that the differential operators do not act on the matrix elements. We assume
the matrix A = (Aji ) to be nonsingular at each point. In linear algebra, this concept is called a change of
basis, the difference being that in our case, the transformation matrix A depends on the position. A frame
field is called orthonormal if at each point,
61
62 CHAPTER 3. CONNECTIONS
Throughout this chapter, we will assume that all frame fields are orthonormal. Whereas this restriction is
not necessary, it is convenient because it results in considerable simplification in computions.
3.1.2 Proposition If {e1 , . . . , en } is an orthonormal frame, then the transformation matrix is orthogonal
(ie, AAT = I)
Proof The proof is by direct computation. Let ei = ∂j Aji . Then
Hence
Given a frame {ei }, we can also introduce the corresponding dual coframe forms θi by requiring that
Since the dual coframe is a set of 1-forms, they can also be expressed in local coordinates as linear combina-
tions
θi = B ik dxk .
It follows from equation( 3.3), that
Therefore, we conclude that BA = I, so B = A−1 = AT . In other words, when the frames are orthonormal,
we have
ei = ∂k Aki
i
θ = Aik dxk . (3.4)
∂ ∂
The vectors ∂r , and ∂z are clearly unit vectors.
∂
To make the vector ∂θ a unit vector, it suffices to divide it by its length r. We can then compute the
dot products of each pair of vectors and easily verify that the quantities
∂ 1 ∂ ∂
e1 = , e2 = , e3 = , (3.6)
∂r r ∂θ ∂z
are a triplet of mutually orthogonal unit vectors and thus constitute an orthonormal frame. The surfaces with
constant value for the coordinates r, θ and z respectively, represent a set of mutually orthogonal surfaces
at each point. The frame vectors at a point are normal to these surfaces as shown in figure 3.1. Physicists
often refer to these frame vectors as {r̂, θ̂, ẑ}, or as {er , eθ , ez .}.
to an axis or to the origin. In such situations, transformations to the appropriate coordinate system often
result in considerable simplification of the field equations involved in the problem. It has been shown that
the Laplace operator that appears in the potential, heat, wave, and Schrödinger field equations, is separable
in exactly twelve orthogonal coordinate systems. A simple and efficient method to calculate the Laplacian
in orthogonal coordinates can be implemented using differential forms.
3.2.1 Example In spherical coordinates the differential of arc length is given by (see equation 2.31) the
metric:
ds2 = dr2 + r2 dθ2 + r2 sin2 θdφ2 .
Let
θ1 = dr,
2
θ = rdθ,
3
θ = r sin θdφ. (3.8)
Note that these three 1-forms constitute the dual coframe to the orthonormal frame derived in equation( 3.7).
Consider a scalar field f = f (r, θ, φ). We now calculate the Laplacian of f in spherical coordinates using the
methods of section 2.4.2. To do this, we first compute the differential df and express the result in terms of
the coframe.
∂f ∂f ∂f
df = dr + dθ + dφ,
∂r ∂θ ∂φ
∂f 1 1 ∂f 2 1 ∂f 3
= θ + θ + θ .
∂r r ∂θ r sin θ ∂φ
The components df in the coframe represent the gradient in spherical coordinates. Continuing with the
scheme of section 2.4.2, we next apply the Hodge ? operator. Then, we rewrite the resulting 2-form in terms
of wedge products of coordinate differentials so that we can apply the definition of the exterior derivative.
∂f 2 1 ∂f 1 1 ∂f 1
?df = θ ∧ θ3 − θ ∧ θ3 + θ ∧ θ2 ,
∂r r ∂θ r sin θ ∂φ
∂f 1 ∂f 1 ∂f
= r2 sin θ dθ ∧ dφ − r sin θ dr ∧ dφ + r dr ∧ dθ,
∂r r ∂θ r sin θ ∂φ
∂f ∂f 1 ∂f
= r2 sin θ dθ ∧ dφ − sin θ dr ∧ dφ + dr ∧ dθ,
∂r ∂θ sin θ ∂φ
∂ 2 ∂f ∂ ∂f
d ? df = (r sin θ )dr ∧ dθ ∧ dφ − (sin θ )dθ ∧ dr ∧ dφ +
∂r ∂r ∂θ ∂θ
1 ∂ ∂f
( )dφ ∧ dr ∧ dθ,
sin θ ∂φ ∂φ
1 ∂2f
∂ 2 ∂f ∂ ∂f
= sin θ (r )+ (sin θ ) + dr ∧ dθ ∧ dφ.
∂r ∂r ∂θ ∂θ sin θ ∂φ2
1 ∂2f 1
1 ∂ 2 ∂f ∂ ∂f
d ? df = 2 sin θ (r )+ (sin θ ) + θ ∧ θ2 ∧ θ3 .
r sin θ ∂r ∂r ∂θ ∂θ sin θ ∂φ2
1 ∂2f
2 1 ∂ 2 ∂f 1 1 ∂ ∂f
∇ f= 2 r + 2 (sin θ ) + . (3.9)
r ∂r ∂r r sin θ ∂θ ∂θ sin2 θ ∂φ2
The derivation of the expression for the spherical Laplacian by differential forms is elegant and leads naturally
to the operator in Sturm-Liouville form.
3.2. CURVILINEAR COORDINATES 65
The process above can be carried out for general orthogonal transformations. A change of coordinates
xi = xi (uk ) leads to an orthogonal transformation if in the new coordinate system uk , the line metric
Classically, the quantities {h1 , h2 , h3 } are called the weights. Please note that, in the interest of connecting
to classical terminology, we have exchanged two indices for one and this will cause small discrepancies with
the index summation convention. We will revert to using a summation symbol when these discrepancies
occur. To satisfy the duality condition θi (ej ) = δji , we must choose the corresponding frame vectors ei as
follows:
1 ∂ 1 ∂
e1 = √ = ,
g11 ∂u1 h1 ∂u1
1 ∂ 1 ∂
e2 = √ = ,
g22 ∂u2 h2 ∂u2
1 ∂ 1 ∂
e3 = √ = .
g33 ∂u3 h3 ∂u3
∂f
df = dxk ,
∂xk
∂f ∂ui k
= dx ,
∂ui ∂xk
∂f i
= du ,
∂ui
X 1 ∂f
= i ∂ui
θi .
i
h
= ei (f )θi .
As expected, the components of the gradient in the coframe θi are the just the frame vectors.
1 ∂ 1 ∂ 1 ∂
∇= , , . (3.11)
h1 ∂u1 h2 ∂u2 h3 ∂u3
Curl. Let F = (F1 , F2 , F3 ) be a classical vector field. Construct the corresponding 1-form F = Fi θi in the
coframe. We calculate the curl using the dual of the exterior derivative.
F = F1 θ1 + F2 θ2 + F3 θ3 ,
= (h1 F1 )du1 + (h2 F2 )du2 + (h3 F3 )du3 ,
= (hF )i dui , where (hF )i = hi Fi .
1 ∂(hF )i ∂(hF )j
dF = − dui ∧ duj ,
2 ∂uj ∂ui
1 ∂(hF )i ∂(hF )j
= − dθi ∧ dθj .
hi hj ∂uj ∂ui
1 ∂(hF )i ∂(hF )j
?dF = ijk [ − ] θk = (∇ × F )k θk .
hi hj ∂uj ∂ui
66 CHAPTER 3. CONNECTIONS
F = F1 θ1 + F2 θ2 + F3 θ3
?F = F1 θ 2 ∧ θ 3 + F2 θ 3 ∧ θ 1 + F3 θ 1 ∧ θ 2
= (h2 h3 F1 )du2 ∧ du3 + (h1 h3 F2 )du3 ∧ du1 + (h1 h2 F3 )du1 ∧ du2
∂(h2 h3 F1 ) ∂(h1 h3 F2 ) ∂(h1 h2 F3 )
d ? dF = + + du1 ∧ du2 ∧ du3 .
∂u1 ∂u2 ∂u3
Therefore,
1 ∂(h2 h3 F1 ) ∂(h1 h3 F2 ) ∂(h1 h2 F3 )
∇ · F = ?d ? F = + + . (3.12)
h1 h2 h3 ∂u1 ∂u2 ∂u3
3.3.1 Definition Given a pair (X, Y ) of arbitrary vector field in Rn , we associate a new vector field ∇X Y ,
so that ∇X : X (Rn ) −→ X (Rn ). The quantity ∇ called a Koszul connection if it satisfies the following
properties:
1. ∇f X (Y ) = f ∇X Y,
3. ∇X (Y1 + Y2 ) = ∇X Y1 + ∇X Y2 ,
4. ∇X f Y = X(f )Y + f ∇X Y,
for all vector fields X, X1 , X2 , Y, Y1 , Y2 ∈ X (Rn ) and all smooth functions f . Implicit in the properties, we
set ∇X f = X(f ). The definition states that the map ∇X is linear on X but behaves as a linear derivation
on Y. For this reason, the quantity ∇X Y is called the covariant derivative of Y in the direction of X.
∂
3.3.2 Proposition Let Y = f i ∂x i be a vector field in R
n
, and let X another C ∞ vector field. Then the
operator given by
∂
∇X Y = X(f i ) i (3.13)
∂x
defines a Koszul connection.
Proof The proof just requires verification that the four properties above are satisfied, and it is left as an
exercise.
The operator defined in this proposition is the standard connection compatible with the Euclidean metric.
The action of this connection on a vector field Y yields a new vector field whose components are the directional
derivatives of the components of Y .
Then,
∂ ∂
∇X Y = X(x2 ) + X(xy 2 ) ,
∂x ∂y
∂ ∂ ∂ ∂ ∂ ∂
= [x (x2 ) + xz (x2 )] + [x (xy 2 ) + xz (xy 2 )] ,
∂x ∂y ∂x ∂x ∂y ∂y
∂ ∂
= 2x2 + (xy 2 + 2x2 yz) .
∂x ∂y
The coefficients can be more succinctly expressed using the compact index notation,
∇X ei = ej ω j i (X). (3.17)
Hence,
< ∇X ei , ek >= ωki (X) (3.19)
The left-hand side of the last equation is the inner product of two vectors, so the expression represents an
array of functions. Consequently, the right-hand side also represents an array of functions. In addition, both
expressions are linear on X, since by definition, ∇X is linear on X. We conclude that the right-hand side can
be interpreted as a matrix in which each entry is a 1-forms acting on the vector X to yield a function. The
matrix valued quantity ω ij is called the connection form. Sacrificing some inconsistency with the formalism
of differential forms for the sake of connecting to classical notation, we sometimes write the above equation
as
< dei , ek >= ωki , (3.20)
where {ei } are vector calculus vectors forming an orthonormal basis.
3.3.5 Definition Let ∇X be a Koszul connection and let {ei } be a frame. The Christoffel symbols
associated with the connection in the given frame are the functions Γkij given by
The Christoffel symbols are the coefficients that give the representation of the rate of change of the frame
vectors in the direction of the frame vectors themselves. Many physicists therefore refer to the Christoffel
symbols as the connection, resulting in possible confusion. The precise relation between the Christoffel
symbols and the connection 1-forms is captured by the equations,
or equivalently
ω ki = Γkij θj . (3.23)
n 2 3
In a general frame in R there are n entries in the connection 1-form and n Christoffel symbols. The
number of independent components is reduced if one assumes that the frame is orthonormal.
If T = T i ei is a general vector field, then
∇ej T = ∇ej (T i ei )
= T,ji ei + T i Γkji ek
= (T,ji + T k Γijk )ei , (3.24)
i
Tkj = T,ji + Γijk T k . (3.25)
Here, the comma in the subscript means regular derivative. The equation above is also commonly written
as
∇ej T i = ∇j T i = T,ji + Γijk T k ,
We should point out the accepted but inconsistent use of terminology. What is meant by the notation ∇j T i
above is not the covariant derivative of the vector but the tensor components of the covariant derivative of
the vector; one more reminder that most physicists conflate a tensor with its components.
3.3.6 Proposition Let {ei } be an orthonormal frame and ∇X be a Koszul connection compatible with
the metric . Then
ωji = −ωij (3.26)
Proof Since it is given that < ei , ej >= δij , we have
0 = ∇X < ei , ej >,
= < ∇X ei , ej > + < ei , ∇X ej >,
= < ω ki ek , ej > + < ei , ω kj ek >,
= ω ki < ek , ej > +ω kj < ei , ek >,
= ω ki gkj + ω kj gik ,
= ωji + ωij .
Let us compute the covariant derivative of a one-form ω with respect to vector field X. The contraction of
ω ⊗ Y is the function iY ω = ω(Y ). Taking the covariant derivative, we have,
Hence, the coordinate-free formula for the covariant derivative of one-form is,
∇X (θi ⊗ ej ) = ∇X θi ⊗ ej + θi ⊗ ∇X ej .
The contraction of iej θi = θi (ej ) = δji , Hence, taking the contraction of the equation above, we see that the
left-hand side becomes 0, and we conclude that,
Let ω = Ti θi . Then,
Classically, we write
∇k Tj = Tjkk = Tj,k − Γi jk Ti . (3.30)
In general, let T be a tensor of type rs ,
T = Tji11,...,j
,...,ir
e ⊗ · · · ⊗ eir ⊗ θj1 ⊗ . . . θjs .
s i1
(3.31)
Since we know how to take the covariant derivative of a function, a vector, and a one form, we can use
Leibnitz rule for tensor products and property of the covariant derivative commuting with contractions, to
get by induction, a formula for the covariant derivative of an rs -tensor,
(∇X T )(θi1 , ..., θir , ej1 , ...ejs ) = X(T (θi1 , ..., θir , ej1 , ..., ejs ))
− T (∇X θi1 , ..., θir , ej1 , ..., ejs ) − · · · − T (θi1 , ..., ∇X θir , ej1 , ..., ejs )...
− T (θi1 , ..., θir , ∇X ej1 , ..., ejs ) − · · · − T (θi1 , ..., θir , ej1 , ..., ∇X ejs ). (3.32)
The covariant derivative picks up a term with a positive Christoffel symbol factor for each contravariant
index
and a term with a negative Christoffel symbol factor for each covariant index. Thus, for example, for
a 12 tensor, be components of the covariant derivative in classical notation are
In other words, a connection is compatible with the metric just means that the metric is covariantly constant
along any vector field.
70 CHAPTER 3. CONNECTIONS
In an orthonormal frame in Rn the number of independent coefficients of the connection 1-form is (1/2)n(n−
1) since by antisymmetry, the diagonal entries are zero, and one only needs to count the number of entries in
the upper triangular part of the n × n matrix ωij . Similarly, the number of independent Christoffel symbols
gets reduced to (1/2)n2 (n − 1). Raising one index with g ij , we find that ω ij is also antisymmetric, so in
R3 the connection equations become
ω 1 2 (X) ω 1 3 (X)
0
∇X [e1 , e2 , e3 ] = [e1 , e2 , e3 ] −ω 1 2 (X) 0 ω 2 3 (X) (3.35)
1 2
−ω 3 (X) −ω 3 (X) 0
Comparing the Frenet frame equation (1.39), we notice the obvious similarity to the general frame equations
above. Clearly, the Frenet frame is a special case in which the basis vectors have been adapted to a curve,
resulting in a simpler connection in which some of the coefficients vanish. A further simplification occurs
in the Frenet frame, since in this case the equations represent the rate of change of the frame only along
the direction of the curve rather than an arbitrary direction vector X. To elaborate on this transition
from classical to modern notation, consider a unit speed curve β(s). Then, as we discussed in section
0 dxi ∂
1.15, we associate with the classical tangent vector T = dx ds the vector field T = β (s) = ds ∂xi . Let
∂
W = W (β(s)) = wj (s) ∂x j be an arbitrary vector field constrained to the curve. The rate of change of W
3.4.1 Theorem Let {ei } be a frame with connection ω i j and dual coframe θi . Then
Θi ≡ dθi + ω i j ∧ θj = 0. (3.36)
Proof Let
ei = ∂j Aj i
be a frame, and let θi be the corresponding coframe. Since θi (ej ), we have
θi = (A−1 )ij dxj .
Let X be an arbitrary vector field. Then
∇X e i = ∇X (∂j Aji ).
ej ω ji (X) = ∂j X(Aji ),
= ∂j d(Aji )(X),
= ek (A−1 )kj d(Aji )(X).
ω ki (X) = (A−1 )kj d(Aji )(X).
3.4. CARTAN EQUATIONS 71
Hence,
ω ki = (A−1 )kj d(Aji ),
or, in matrix notation,
ω = A−1 dA. (3.37)
On the other hand, taking the exterior derivative of θi , we find that
dθ = −ω ∧ θ. (3.38)
In other words
dθi + ω ij ∧ θj = 0.
d(dθi ) + d(ω ij ∧ θj ) = 0,
dω ij ∧ θj − ω ij ∧ dθj = 0.
dω ij ∧ θj − ω ij ∧ (−ω jk ∧ θk ) = 0,
dω ij ∧ θj + ω ik ∧ ω kj ∧ θj = 0,
(dω ij + ω ik ∧ ω kj ) ∧ θj = 0,
dω ij + ω ik ∧ ω kj = 0.
72 CHAPTER 3. CONNECTIONS
Ωij ≡ dω ij + ω ik ∧ ω kj . (3.43)
3.4.4 Theorem Let θ be a coframe with connection ω in Rn . Then the curvature form vanishes:
Ω = dω + ω ∧ ω = 0. (3.44)
Proof Given that there is a non-singular matrix A such that θ = A−1 dx and ω = A−1 dA, we have
dω = d(A−1 ) ∧ dA.
Therefore, dω = −ω ∧ ω.
There is a slight abuse of the wedge notation here. The connection ω is matrix valued, so the symbol ω ∧ ω
is really a composite of matrix and wedge multiplication.
Hence,
sin θ cos φ sin θ sin φ cos θ
A−1 = cos θ cos φ cos θ sin φ − sin θ ,
− sin φ cos φ 0
and
cos θ cos φ dθ − sin θ sin φ dφ − sin θ cos φ dθ − cos θ sin φ dφ − cos φ dφ
dA = cos θ sin φ dθ + sin θ cos φ dφ − sin θ sin φ dθ + cos θ cos φ dφ − sin φ dφ .
− sin θ dθ − cos θ dθ 0
Since the ω = A−1 dA is antisymmetric, it suffices to compute:
A slicker computation of the connection form can be obtained by a method of educated guessing working
directly from the structure equations. We have that the dual one forms are:
θ1 = dr,
θ2 = r dθ,
θ3 = r sin θ dφ.
Then
Change of Basis
We briefly explore the behavior of the quantities Θi and Ωij under a change of basis. Let ei be frame
in M = Rn with dual forms θi , and let ei be another frame related to the first frame by an invertible
transformation.
ei = ej B ji , (3.45)
which we will write in matrix notation as e = eB. Referring back to the definition of connections (3.17), we
introduce the covariant differential ∇ which maps vectors into vector-valued forms,
∇ : Ω0 (M, T M ) → Ω1 (M, T M )
given by the formula
∇ei = ej ⊗ ω ji
= ej ω ji
∇e = e ω (3.46)
where, once again, we have simplified the equation by using matrix notation. This definition is elegant
because it does not explicitly show the dependence on X in the connection (3.17). The idea of switching
from derivatives to differentials is familiar from basic calculus. Consistent with equation 3.20, the vector
calculus notation for equation 3.46 would be
dei = ej ω j i . (3.47)
74 CHAPTER 3. CONNECTIONS
However, we point out that in the present context, the situation is much more subtle. The operator ∇ here
maps a vector field to a matrix-valued tensor of rank 11 . Another way to view the covariant differential
is to think of ∇ as an operator such that if e is a frame, and X a vector field, then ∇e(X) = ∇X e. If f
is a function, then ∇f (X) = ∇X f = df (X), so that ∇f = df . In other words, ∇ behaves like a covariant
derivative on vectors, but like a differential on functions. The action of the covariant differential also extends
to the entire tensor algebra, but we do not need that formalism for now, and we delay discussion to section
6.4 on connections on vector bundles. Taking the exterior differential of (3.45) and using (3.46) recursively,
we get
∇e = (∇e)B + e(dB)
= e ωB + e(dB)
= eB −1 ωB + eB −1 dB
= e[B −1 ωB + B −1 dB]
= eω
provided that the connection ω in the new frame e is related to the connection ω by the transformation law,
(See 6.62)
ω = B −1 ωB + B −1 dB. (3.48)
It should be noted than if e is the standard frame ei = ∂i in Rn , then ∇e = 0, so that ω = 0. In this
case, the formula above reduces to ω = B −1 dB, showing that the transformation rule is consistent with
equation (3.37). The transformation law for the curvature forms is,
Ω = B −1 ΩB. (3.49)
3.4.6 Example Suppose that B is a change of basis consisting of a rotation by an angle θ about e3 . The
transformation is a an isometry that can be represented by the orthogonal rotation matrix
cos θ − sin θ 0
B = sin θ cos θ 0 . (3.50)
0 0 1
Carrying out the computation for the change of basis 3.48, we find:
ω 1 2 = ω 1 2 − dθ,
ω 1 3 = cos θ ω 1 3 + sin θ ω 2 3 ,
ω 2 3 = − sin θ ω 1 3 + cos θ ω 2 3 . (3.51)
The B −1 dB part of the transformation only affects the ω 1 2 term, and the effect is just adding dθ much like
the case of the Maurer-Cartan form for SO(2) above.
Chapter 4
Theory of Surfaces
4.1 Manifolds
x : V ⊂ R2 −→ R3
x
(u, v) 7−→ (x(u, v), y(u, v), z(u, v)) (4.1)
Each set U = x(V ) is called a coordinate neighborhood of M . We require that the Jacobian of the map has
maximal rank. In local coordinates, a coordinate chart is represented by three equations in two variables:
It will be convenient to use the tensor index formalism when appropriate, so that we can continue to take
advantage of the Einstein summation convention. The assumption that the Jacobian J = (∂xi /∂uα ) be of
maximal rank allows one to invoke the implicit function theorem. Thus, in principle, one can locally solve
for one of the coordinates, say x3 , in terms of the other two, to get an explicit function
x3 = f (x1 , x2 ). (4.3)
The loci of points in R3 satisfying the equations xi = f i (uα ) can also be locally represented implicitly by an
expression of the form
F (x1 , x2 , x3 ) = 0. (4.4)
4.1.2 Definition Let Ui and Uj be two coordinate neighborhoods of a point p ∈ M with corresponding
charts x(u1 , u2 ) : Vi −→ Ui ⊂ R3 and y(v 1 , v 2 ) : Vj −→ Uj ⊂ R3 with a non-empty intersection Ui ∩ Uj 6= ∅.
75
76 CHAPTER 4. THEORY OF SURFACES
On the overlaps, the maps φij = x−1 y are called transition functions or coordinate transformations. (See
figure 4.2 )
4.1.3 Definition A differentiable manifold of dimension 2, is a space M together with an indexed collection
{Uα }α∈I of coordinate neighborhoods satisfying the following properties:
1. The neighborhoods {Uα } constitute an open cover M . That is, if p ∈ M , then p belongs to some chart.
2. For any pair of coordinate neighborhoods Ui and Uj with Ui ∩ Uj 6= ∅, the transition maps φij and
their inverses are differentiable.
3. An indexed collection satisfying the conditions above is called an atlas. We require the atlas to be
maximal in the sense that it contains all possible coordinate neighborhoods.
The overlapping coordinate patches represent different parametrizations for the same set of points in R3 .
Part (2) of the definition insures that on the overlap, the coordinate transformations are invertible. Part (3)
is included for technical reasons, although in practice the condition is superfluous. A family of coordinate
neighborhoods satisfying conditions (1) and (2) can always be extended to a maximal atlas. This can be
shown from the fact that M inherits a subspace topology consisting of open sets which are defined by the
intersection of open sets in R3 with M .
If the coordinate patches in the definition map from Rn to Rm n < m we say that M is a n-dimensional
submanifold embedded in Rm . In fact, one could define an abstract manifold without the reference to
the embedding space by starting with a topological space M that is locally Euclidean via homeomorphic
coordinate patches and has a differentiable structure as in the definition above. However, it turns out that
any differentiable manifold of dimension n can be embedded in R2n , as proved by Whitney in a theorem
that is beyond the scope of these notes.
A 2-dimensional manifold embedded in R3 in which the transition functions are C ∞ , is called a smooth
surface. The first condition in the definition states that each coordinate neighborhood looks locally like a
subset of R2 . The second differentiability condition indicates that the patches are joined together smoothly
as some sort of quilt. We summarize this notion by saying that a manifold is a space that is locally Euclidean
and has a differentiable structure, so that the notion of differentiation makes sense. Of course, Rn is itself
an n dimensional manifold.
The smoothness condition on the coordinate component functions xi (uα ) implies that at any point
x (u0 + hα ) near a point xi (uα
i α i
0 ) = x (u0 , v0 ), the functions admit a Taylor expansion
∂xi ∂ 2 xi
i 1
x (uα
0
α
+h )=x i
(uα
0) +h α
+ hα hβ + ... (4.5)
∂uα 0 2! ∂uα ∂uβ 0
must have maximal rank. At points where J has rank 0 or 1, there is a singularity in the coordinate patch.
4.1.4 Example Consider the local coordinate chart for the unit sphere obtained by setting r = 1 in the
equations for spherical coordinates 2.30
x = sin θ cos φ,
y = sin θ sin φ,
z = cos φ. (4.6)
Clearly, the surface represented by this chart is part of the sphere x2 +y 2 +z 2 = 1. The chart cannot possibly
represent the whole sphere because, although a sphere is locally Euclidean, (the earth is locally flat) there
is certainly a topological difference between a sphere and a plane. Indeed, if one analyzes the coordinate
chart carefully, one will note that at the North pole (θ = 0, z = 1, the coordinates become singular. This
happens because θ = 0 implies that x = y = 0 regardless of the value of φ, so that the North pole has an
infinite number of labels. In this coordinate patch, the Jacobian at the North Pole does not have maximal
rank. To cover the entire sphere, one would need at least two coordinate patches. In fact, introducing an
exactly analogous patch y(u.v) based on South pole would suffice, as long as in overlap around the equator
functions x−1 y, and y−1 x are smooth. One could conceive more elaborate coordinate patches such as those
used in baseball and soccer balls.
The fact that it is required to have two parameters to describe a patch on a surface in R3 is a manifestation
of the 2-dimensional nature of the surfaces. If one holds one of the parameters constant while varying the
other, then the resulting 1-parameter equation describes a curve on the surface. Thus, for example, letting
φ = constant in equation (4.6), we get the equation of a meridian great circle.
Here the angle u traces points around the z-axis, whereas the angle v traces points around the circle C. (At
the risk of some confusion in notation, (the parameters in the figure are bold-faced; this is done solely for
the purpose of visibility.) The projection of a point in the surface of the torus onto the xy-plane is located
at a distance (R + r cos u) from the origin. Thus, the x and y coordinates of the point in the torus are just
the polar coordinates of the projection of the point in the plane. The z-coordinate corresponds to the height
of a right triangle with radius r and opposite angle u.
78 CHAPTER 4. THEORY OF SURFACES
4.1.8 Notation Given a parametrization of a surface in a local chart x(u, v) = x(u1 , u2 ) = x(uα ), we will
denote the partial derivatives by any of the following notations:
∂x ∂2x
xu = x1 = , xuu = x11 =
∂u ∂u2
∂x ∂2x
xv = x2 = , xvv = x22 = ,
∂v ∂v 2
or more succinctly,
∂x ∂2x
xα = , xαβ = (4.10)
∂uα ∂uα ∂v β
as an object embedded in Euclidean space. An equivalent version of the metric (4.11) can be obtained by
using a more traditional calculus notation:
dx = xu du + xv dv
ds2 = dx · dx
= (xu du + xv dv) · (xu du + xv dv)
= (xu · xu )du2 + 2(xu · xv )dudv + (xv · xv )dv 2 .
E = g11 = xu · xu
F = g12 = xu · xv
= g21 = xv · xu
G = g22 = xv · xv .
That is
gαβ = xα · xβ =< xα , xβ > .
Just as in the case of curves, when we revert back to classical notation to describe a surface as xi (uα ), what
we really mean is (xi ◦ α)(uα ), where xi are the coordinate functions in R3 . Particular examples of tangent
80 CHAPTER 4. THEORY OF SURFACES
vectors on M are given by the push-forward of the standard basis of T R2 . These tangent vectors which
earlier we called xα are defined by
∂ ∂
α∗ (
α
)(f ) |α(uα ) = (f ◦ α) |uα
∂u ∂uα
In this formalism, the first fundamental form I is just the symmetric bilinear tensor defined by the induced
metric,
I(X, Y ) = g(X, Y ) =< X, Y >, (4.16)
where X and Y are any pair of vector fields in X (M ).
A simpler way to obtain this result is to recall that parametric directions are given by xu and xv , so
< xu , xv > F
cos θ = =√ . (4.20)
kxu k · kxv k EG
There are many interesting curves on a sphere, but amongst these the loxodromes have a special role in
history. A loxodrome is a curve that winds around a sphere making a constant angle with the meridians.
In this sense, it is the spherical analog of a cylindrical helix and as such it is often called a spherical helix.
The curves were significant in early navigation where they are referred as rhumb lines. As people in the late
1400’s began to rediscover that earth was not flat, cartographers figured out methods to render maps on flat
paper surfaces. One such technique is called the Mercator projection which is obtained by projecting the
sphere onto a plane that wraps around the sphere as a cylinder tangential to the sphere along the equator.
As we will discuss in more detail later, a navigator travelling a constant bearing would be moving on
a straight path on the Mercator projection map, but on the sphere it would be spiraling ever faster as
one approached the poles. Thus, it became important to understand the nature of such paths. It appears
as if the first quantitative treatise of loxodromes was carried in the mid 1500’s by the Portuguese applied
mathematician Pedro Nuñes, who was chair of the department at the University of Coimbra.
As an application, we will derive the equations of loxodromes and compute the arc length. A general
spherical curve can be parametrized in the form γ(t) = x(θ(t), φ(t)). Let σ be the angle the curve makes
82 CHAPTER 4. THEORY OF SURFACES
with the meridians φ = constant. Then, recalling that < xθ , xφ >= F = 0, we have:
dθ dφ
γ 0 = xθ + xφ .
dt dt
< xθ , γ 0 > E dθ
dt dθ
cos σ = 0
= √ =a .
kxθ k · kγ k ds
E dt ds
a2 dθ2 = cos2 σ ds2 ,
a2 sin2 σ dθ2 = a2 cos2 σ sin2 θ dφ2 ,
sin σ dθ = ± cos σ sin θ dφ,
csc θ dθ = ± cot σ dφ.
The convention used by cartographers, is to measure the angle θ from the equator. To better adhere to the
history, but at the same time avoiding confusion, we replace θ with ϑ = π2 − θ, so that ϑ = 0 corresponds to
the equator. Integrating the last equation with this change, we get
sec ϑ dϑ = ± cot σ dφ
ln tan( ϑ2 + π4 ) = ± cot σ(φ − φ0 ).
Thus, we conclude that the equations of loxodromes and their arc lengths are given by
φ = ±(tan σ) ln tan( ϑ2 + π4 ) + φ0 (4.22)
s = a(θ − θ0 ) sec σ, (4.23)
where θ0 and φ0 are the coordinates of the initial position. Figure 4.2 shows four loxodromes equally
distributed around the sphere.
Loxodromes were the bases for a number of beautiful drawings and woodcuts by M. C. Escher. figure 4.2
also shows one more beautiful manifestation of geometry in nature in a plant called Aloe Polyphylla. Not
surprisingly, the plant has 5 loxodromoes which is a Fibonacci number. We will show later under the discus-
sion of conformal (angle preserving) maps in section 5.2.2, that loxodromes map into straight lines making
a constant angle with meridians in the Mercator projection (See Figure ??).
b) Surface of Revolution
x = (r cos θ, r sin θ, f (r)),
xr = (cos θ, sin θ, f 0 (r)),
xθ = (−r sin θ, r cos θ, 0),
E = xr · xr = 1 + f 02 (r), Fig. 4.6: S. of Revolution
F = xr · xθ = 0,
G = xθ · xθ = r2 ,
ds2 = [1 + f 02 (r)]dr2 + r2 dθ2 .
c) Pseudosphere
u
x = (a sin u cos v, a sin u sin v, a(cos u + ln(tan )),
2
E = a2 cot2 u,
F = =0
G = a2 sin2 u,
ds2 = a2 cot2 u du2 + a2 sin2 u dv 2 .
4.2. THE FIRST FUNDAMENTAL FORM 83
The pseudosphere is a surface of revolution in which the profile curve is a tractrix. The tractrix curve was
originated by a problem posed by Leibnitz to the effect of finding the path traced by a point initially placed
on the horizontal axis at a distance a from the origin, as it was pulled along the vertical axis by a taught
string of constant length a, as shown in figure 4.7. The tractrix was later studied by Huygens in 1692.
Colloquially this is the path of a reluctant dog at (a, 0) dragged by a man walking√ up the z-axis. The
tangent segment is the hypothenuse of a right triangle with base x and height a2 − x2 , so the slope is
√
dz/dx = − a2 − x2 /x. Using the trigonometric substitution x = a sin u, we get z = a (cos2 u/ sin u) du,
R
which leads to the appropriate form for the profile of the surface of revolution. The pseudosphere was
studied by Beltrami in 1868. He discovered that in spite of the surface extending asymptotically to infinity,
the surface area is finite with S = 4πa2 as in a sphere of the same radius, and the volume enclosed is half
that sphere. We will have much more to say about this surface.
d) Torus
f) Catenoid
u
x = (u cos v, u sin v, c cosh−1 ), This is a catenary of revolution.
c
u2
E = ,
u2 − c2
F = 0,
G = u2 ,
u2
ds2 = du2 + u2 dv 2 , (4.26)
u2 − c2
84 CHAPTER 4. THEORY OF SURFACES
A conical helix a curve γ(t) = x(r(t), φ(t)), that makes a constant angle σ with the generators of the cone.
Similar to the case of loxodromes, we have
dr dφ
γ 0 = xr + xφ .
dt dt
< xr , γ 0 > E dr √ dr
dt
cos σ = 0
= √ = E .
kxr k · kγ k ds
E dt ds
E dr2 = cos2 σ ds2 ,
csc2 α dr2 = cos2 σ(csc2 α dr2 + r2 dφ2 ),
csc2 α sin2 σ dr2 = r2 cos2 σ dφ2 ,
1
dr = cot σ sin α dφ.
r
Therefore, the equations of a conical helix are given by
As shown in figure 4.9, a conical helix projects into the plane as a logarithmic spiral. Many sea shells and
other natural objects in nature exhibit neatly such conical spirals. The picture shown here is that of lobatus
gigas or caracol pala, previously known as strombus gigas. The particular one is included here with certain
degree of nostalgia, for it has been a decorative item for decades in our family. The shell was probably
found in Santa Cruz del Islote, Archipelago de San Bernardo, located in the Gulf of Morrosquillo in the
Caribbean coast of Colombia. In this densely populated island paradise, which then enjoyed the pulchritude
of enchanting coral reefs, the shells are now virtually extinct as the coral has succumbed to bleaching with
rising temperatures of the waters. The shell shows a cut in the spire which the island natives use to sever
the columellar muscle and thus release the edible snail.
4.3. THE SECOND FUNDAMENTAL FORM 85
Now, consider a curve on the surface given by uβ = uβ (s). Without loss of generality, we assume that
the curve is parametrized by arc length s so that the curve has unit speed. Let e = {T, N, B} be the Frenet
frame of the curve. Recall that the rate of change ∇T W of a vector field W along the curve correspond
to the classical vector w0 = dw
ds , so ∇W is associated with the vector dw. Thus the connection equation
∇e = eω is given by
0 −κ ds 0
d[T, N, B] = [T, N, B] κ ds 0 −τ ds (4.30)
0 τ ds 0.
Following ideas first introduced by Darboux and subsequently perfected by Cartan, we introduce a new
orthonormal frame f = {T, g, n, } adapted to the surface, where at each point, T is the common tangent to
the surface and to the curve on the surface, n is the unit normal to the surface and g = n × T. Since the two
orthonormal frames must be related by a rotation that leaves the T vector fixed, we have f = eB, where B
is a matrix of the form
1 0 0
B = 0 cos θ − sin θ . (4.31)
0 sin θ cos θ
0 −κ cos θ ds −κ sin θ ds
d[T, g, n] = [T, g, n] κ cos θ ds 0 −τ ds + dθ , (4.32)
κ sin θ ds τ ds − dθ 0
0 −κg ds κn ds
= [T, g, n] κg ds 0 −τg ds , (4.33)
κn ds τg ds 0
where:
κn = κ sin θ is called the normal curvature,
κg = κ cos θ is called the geodesic curvature; Kg = κg g the geodesic curvature vector, and
τg = τ − dθ/ds is called the geodesic torsion.
86 CHAPTER 4. THEORY OF SURFACES
We conclude that we can decompose T0 and the curvature κ into their normal and surface tangent space
components (see figure 4.11)
T0 = κn n + κg g, (4.34)
2
κ = κ2n + κ2g . (4.35)
The normal curvature κn measures the curvature of x(uα (s)) resulting from the constraint of the curve to lie
on a surface. The geodesic curvature κg measures the “sideward” component of the curvature in the tangent
plane to the surface. Thus, if one draws a straight line on a flat piece of paper and then smoothly bends
the paper into a surface, the line acquires some curvature. Since the line was originally straight, there is
no sideward component of curvature so κg = 0 in this case. This means that the entire contribution to the
curvature comes from the normal component, reflecting the fact that the only reason there is curvature here
is due to the bend in the surface itself. In this sense, a curve on a surface for which the geodesic curvature
vanishes at all points reflects locally the shortest path between two points. These curves are therefore called
geodesics of the surface. The property of minimizing the path between two points is a local property. For
example, on a sphere one would expect the geodesics to be great circles. However, travelling from Los
Angeles to San Francisco along one such great circle, there is a short path and a very long one that goes
around the earth.
If one specifies a point p ∈ M and a direction vector Xp ∈
Tp M , one can geometrically envision the normal curvature by
considering the equivalence class of all unit speed curves in M
that contain the point p and whose tangent vectors line up
with the direction of X. Of course, there are infinitely many
such curves, but at an infinitesimal level, all these curves can
be obtained by intersecting the surface with a “vertical” plane
containing the vector X and the normal to M . All curves in
this equivalence class have the same normal curvature and their
geodesic curvatures vanish. In this sense, the normal curvature
is more of a property pertaining to a direction on the surface
at a point, whereas the geodesic curvature really depends on Fig. 4.11: Curvature
the curve itself. It might be impossible for a hiker walking on
the undulating hills of the Ozarks to find a straight line trail, since the rolling hills of the terrain extend in
all directions. It might be possible, however, for the hiker to walk on a path with zero geodesic curvature
as long the same compass direction is maintained. We will come back to the Cartan structure equations
associated with the Darboux frame, but for computational purposes, the classical approach is very practical.
Using the chain rule, we se that the unit tangent vector T to the curve is given by
dx dx duα duα
T= = α = xα . (4.36)
ds du ds ds
To find an explicit formula for the normal curvature we first differentiate equation (4.36)
dT
T0 = ,
ds
d duα
= (xα ),
ds ds
α
d du d2 uα
= (xα ) + xα 2 ,
ds ds ds
dxα duβ duα d2 uα
= ( β ) + xα 2 ,
du ds ds ds
duα duβ d2 uα
= xαβ + xα 2 .
ds ds ds
4.3. THE SECOND FUNDAMENTAL FORM 87
Taking the inner product of the last equation with the normal and noticing that < xα , n >= 0, we get
duα duβ
κn = < T0 , n >=< xαβ , n > ,
ds ds
bαβ duα duβ
= , (4.37)
gαβ duα duβ
where
bαβ =< xαβ , n > (4.38)
It follows that the equation for the normal curvature (4.37) can be written explicitly as
We should pointed out that just as the first fundamental form can be represented as
To see this, it suffices to note that differentiation of the identity, < xα , n >= 0, implies that
Therefore,
4.3.3 Definition Directions on a surface along which the second fundamental form
e du2 + 2f du dv + g dv 2 = 0 (4.42)
vanishes, are called asymptotic directions, and curves having these directions are called asymptotic curves.
This happens for example when there are straight lines on the surface, as in the case of the intersection of
the saddle z = xy with the plane z = 0.
For now, we state without elaboration, that one can also define the third fundamental form by
III =< dn, dn >=< nα , nβ > duα duβ . (4.43)
From a computational point a view, a more useful formula for the coefficients of the second fundamental
formula can be derived by first applying the classical vector identity
A·C A·D
(A × B) · (C × D) = , (4.44)
B·C B·D
to compute
kxu × xv k2 = (xu × xv ) · (xu × xv ),
xu · xu xu · xv
= det ,
xv · xu xv · xv
= EG − F 2 . (4.45)
Consequently, the normal vector can be written as
xu × xv xu × xv
n= =√ .
kxu × xv k EG − F 2
It follows that we can write the coefficients bαβ directly as triple products involving derivatives of (x). The
expressions for these coefficients are
(x x x )
e = √ u v uu ,
EG − F 2
(xu xv xuv )
f = √ ,
EG − F 2
(x x x )
g = √ u v vv . (4.46)
EG − F 2
The first fundamental form on a surface measures the square of the distance between two infinitesimally
separated points. There is a similar interpretation of the second fundamental form as we show below.
The second fundamental form measures the distance from a point on the surface to the tangent plane at
a second infinitesimally separated point. To see this simple geometrical interpretation, consider a point
x0 = x(uα α α
0 ) ∈ M and a nearby point x(u0 + du ). Expanding on a Taylor series, we get
1
x(uα α α α β
0 + du ) = x0 + (x0 )α du + (x0 )αβ du du + . . . .
2
We recall that the distance formula from a point x to a plane which contains x0 is just the scalar projection
of (x − x0 ) onto the normal. Since the normal to the plane at x0 is the same as the unit normal to the
surface and < xα , n >= 0, we find that the distance D is
D = < x − x0 , n >,
1
= < (x0 )αβ , n > duα duβ ,
2
1
= II0 .
2
The first fundamental form (or, rather, its determinant) also appears in calculus in the context of calculating
the area of a parametrized surface. If one considers an infinitesimal parallelogram subtended by the vectors
xu du and xv dv, then the differential of surface area is given by the length of the cross product of these two
infinitesimal tangent vectors. That is,
dS = kxu × xv k dudv,
Z Z p
S = EG − F 2 dudv.
The second fundamental form contains information about the shape of the surface at a point. For example,
the discussion above indicates that if b = |bαβ | = eg − f 2 > 0 then all the neighboring points lie on the same
side of the tangent plane, and hence, the surface is concave in one direction. If at a point on a surface b > 0,
the point is called an elliptic point, if b < 0, the point is called hyperbolic or a saddle point, and if b = 0,
the point is called parabolic.
4.4 Curvature
The concept of curvature and its relation to the fundamental forms, constitute the central object of
study in differential geometry. One would like to be able to answer questions such as “what quantities
remain invariant as one surface is smoothly changed into another?” There is certainly something intrinsically
different between a cone, which we can construct from a flat piece of paper, and a sphere, which we cannot.
What is it that makes these two surfaces so different? How does one calculate the shortest path between
two objects when the path is constrained to lie on a surface?
These and questions of similar type can be quantitatively answered through the study of curvature. We
cannot overstate the importance of this subject; perhaps it suffices to say that, without a clear understanding
of curvature, there would be no general theory of relativity, no concept of black holes, and even more
disastrous, no Star Trek.
The notion of curvature of a hypersurface in Rn (a surface of dimension n − 1) begins by studying the
covariant derivative of the normal to the surface. If the normal to a surface is constant, then the surface is
a flat hyperplane. Variations in the normal are indicative of the presence of curvature. For simplicity, we
constrain our discussion to surfaces in R3 , but the formalism we use is applicable to any dimension. We will
also introduce in the modern version of the second fundamental form.
vector to a curve on the surface. We seek the directions in which the normal curvature attains the extrema.
dv
For this purpose, let the curve on the surface be given by v = v(u) and let λ = du. Then we can write the
normal curvature 4.41 in the form
II ∗ e + 2f λ + gλ2
κn = ∗ = , (4.47)
I E + 2F λ + Gλ2
where II ∗ and I ∗ are the numerator and denominator respectively. To find the extrema, we take the
derivative with respect to λ and set it equal to zero. The resulting fraction is zero only when the numerator
is zero, so from the quotient rule we get
It follows that,
II ∗ f + gλ
κn = = . (4.48)
I∗ F + Gλ
On the other hand, combining with equation 4.47 we have,
(e + f λ) + λ(f + gλ) f + gλ
κn = = .
(E + F λ) + λ(F + Gλ) F + Gλ
Equation 4.50 determines two directions du dv along which the normal curvature attains the extrema, except
for special cases when either bαβ = 0, or bαβ and gαβ are proportional, which would cause the determinant
to be identically zero. These two directions are called principal directions of curvature, each associated with
an extremum of the normal curvature. We will have much more to say about these shortly.
Solving each equation for λ we can eliminate λ instead, and we are lead to a quadratic equation for κn which
we can write as
e − Eκn f − F κn
f − F κn g − Gκn
= 0. (4.51)
In other words, the extrema for the values of the normal are the solutions of the equation
Had it been the case that gαβ = δαβ , the reader would recognize this as a eigenvalue equation for a symmetric
matrix giving rise to two invariants, that is, the trace and the determinant of the matrix. We will treat this
formally in the next section. The explicit quadratic expression for the extrema of κn is
eg − f 2
K = κ1 κ2 = , (4.53)
EG − F 2
and
1 Eg − 2F f + Ge
M = 12 (κ1 + κ2 ) = .. (4.54)
2 EF − G2
The quantity K is called the Gaussian curvature and M is called the mean curvature. To understand better
the deep significance of the last two equations, we introduce the modern formulation which will allow is to
draw conclusions from the inextricable connection of these results with the linear algebra spectral theorem
for symmetric operators.
4.4.1 Definition Let X be a vector field on a surface M in R3 and let N be the normal vector. The map
L, given by
LX = −∇X N, (4.55)
is called the Weingarten map. Some authors call this the shape operator. The same definition applies if M
is an n-dimensional hypersurface in Rn+1 .
Here, we have adopted the convention to overline the operator ∇ when it refers to the ambient space. The
Weingarten map is natural to consider, since it represents the rate of change of the normal in an arbitrary
direction tangential to the surface, which is what we wish to quantify.
4.4.2 Definition The Lie bracket [X, Y ] of two vector fields X and Y on a surface M is defined as the
commutator,
[X, Y ] = XY − Y X, (4.56)
meaning that if f is a function on M , then [X, Y ](f ) = X(Y (f )) − Y (X(f )).
and
4.4.6 Remark The two definitions of the second fundamental form are consistent. This is easy to see if one
chooses X to have components xα and Y to have components xβ . With these choices, LX has components
−na and II(X, Y ) becomes bαβ = − < xα , nβ >.
We also note that there is a third fundamental form defined by
In classical notation, the third fundamental form would be denoted by < dn, dn >. As one would expect,
the third fundamental form contains third order Taylor series information about the surface.
4.4.7 Definition The torsion of a connection ∇ is the operator T such that ∀X, Y,
∇X Y − ∇Y X = [X, Y ].
We will elaborate later on the importance of torsion-free connections. For the time being, it suffices to
assume that for the rest of this section, all connections are torsion-free. Using this assumption, it is possible
to prove the following important theorem.
Since < X, N >=< Y, N >= 0 and the connection is compatible with the metric, we know that
hence,
The central theorem of linear algebra is the spectral theorem. In the case of real, self-adjoint operators, the
spectral theorem states that given the eigenvalue equation for a symmetric operator
LX = κX, (4.60)
on a vector space with a real inner product, the eigenvalues are always real and eigenvectors corresponding
to different eigenvalues are orthogonal. Here, the vector spaces in question are the tangent spaces at each
point of a surface in R3 , so the dimension is 2. Hence, we expect two eigenvalues and two eigenvectors:
LX1 = κ1 X1 (4.61)
LX2 = κ2 X2 . (4.62)
4.4.9 Definition The eigenvalues κ1 and κ2 of the Weingarten map L are called the principal curvatures
and the eigenvectors X1 and X2 are called the principal directions.
Several possible situations may occur, depending on the classification of the eigenvalues at each point p
on a given surface:
It is also known from linear algebra, that in a vector space of dimension two, the determinant and the trace
of a self-adjoint operator are the only invariants under an adjoint (similarity) transformation. Clearly, these
invariants are important in the case of the operator L, and they deserve special names. In the case of a
hypersurface of n-dimensions, there would n eigenvalues, counting multiplicities, so the classification of the
points would be more elaborate
4.4.10 Definition The determinant K = det(L) is called the Gaussian curvature of M and H = 21 Tr(L)
is called the mean curvature .
Since any self-adjoint operator is diagonalizable and in a diagonal basis the matrix representing L is
diag(κ1 , κ2 ), if follows immediately that
K = κ1 κ2 ,
1
H = (κ1 + κ2 ). (4.63)
2
An alternative definition of curvature is obtained by considering the unit normal as a map N : M → S 2 ,
which maps each point p on the surface M , to the point on the sphere corresponding to the position vector
Np . The map is called the Gauss map.
4.4.11 Examples
3. The Gauss map of the top half of a circular cone sends all points on the cone into a circle. We may
envision this circle as the intersection of the cone and a unit sphere centered at the vertex.
4. The Gauss map of a circular hyperboloid of one sheet misses two antipodal spherical caps with bound-
aries corresponding to the circles of the asymptotic cone.
The Weingarten map is minus the derivative N∗ = dN of the Gauss map. That is, LX = −N∗ (X).
LX × LY = K(X × Y )
(LX × Y ) + (X × LY ) = 2H(X × Y ) (4.64)
Proof Since LX, LY ∈ X (M ), they can be expressed as linear combinations of the basis vectors X and Y .
LX = a1 X + b1 Y,
LY = a2 X + b2 Y.
Similarly
4.4.13 Proposition
eg − f 2
K = ,
EG − F 2
1 Eg − 2F f + eG
H = . (4.65)
2 EG − F 2
Proof Starting with equations (4.64), take the inner product of both sides with X × Y and use the vector
identity (4.44). We immediately get
< LX, X > < LX, Y >
< LY, X > < LX, X >
K = , (4.66)
< X, X > < X, Y >
< Y, X > < Y, Y >
< LX, X > < LX, Y > < X, X > < X, Y >
+
< Y, X > < Y, Y > < LY, X > < LY, Y >
2H =
< X, X > < X, Y >
. (4.67)
< Y, X > < Y, Y >
The result follows by taking X = xu and Y = xv . Not surprisingly, this is in complete agreement with the
classical formulas for the Gausssian curvature ( equation 4.53) and for the mean curvature (equation 4.54.
4.5. FUNDAMENTAL EQUATIONS 95
If we denote by g and b the matrices of the fundamental forms whose components are gαβ and bαβ
respectively, we can write the equations for the curvatures as:
b
K = det = det(g −1 b), (4.68)
g
b
2H = Tr = Tr(g −1 b) (4.69)
g
The next theorem due to Euler gives a characterization of the normal curvature in the direction of an
arbitrary unit vector X tangent to the surface M at a given point.
4.4.15 Theorem (Euler) Let X1 and X2 be unit eigenvectors of L and let X = (cos θ)X1 + (sin θ)X2 .
Then
II(X, X) = κ1 cos2 θ + κ2 sin2 θ. (4.70)
Proof Simply compute II(X, X) =< LX, X >, using the fact the LX1 = κ1 X1 , LX2 = κ2 X2 , and noting
that the eigenvectors are orthogonal. We get
< LX, X > = < (cos θ)κ1 X1 + (sin θ)κ2 X2 , (cos θ)X1 + (sin θ)X2 >
= κ1 cos2 θ < X1 , X1 > +κ2 sin2 θ < X2 , X2 >
= κ1 cos2 θ + κ2 sin2 θ.
4.4.16 Theorem The first, second and third fundamental forms satisfy the equation
III − 2H II + KI = 0 (4.71)
Proof The proof follows immediately from the fact that for a symmetric 2 by 2 matrix A, the characteristic
polynomial is κ2 − tr(A)κ + det(A) = 0, and from the Cayley-Hamilton theorem stating that the matrix is
annihilated by its characteristic polynomial.
the vectors xα live on the tangent space, but this is not necessarily true for the second derivative vectors
xαβ . Here, x(uα ) could refer to a coordinate patch in any number of dimensions, so all the tensor index
formulas that follow, apply to surfaces of codimension 1 in Rn . The set of vectors {xα , n} constitutes a
basis for Rn at each point on the surface, we can express the vectors xαβ as linear combinations of the basis
vectors. Therefore, there exist coefficients Γγαβ and cαβ such that,
Taking the inner product of equation 4.72 with n, noticing that the latter is a unit vector orthogonal to xγ ,
we find that cαβ =< xαβ , n >, and hence these are just the coefficients of the second fundamental form. In
other words, equation 4.72 can be written as
Equation 4.73 together with equation 4.76 below, are called the formulæ of Gauss. The covariant derivative
formulation of the equation of Gauss follows in a similar fashion. Let X and Y be vector fields tangent to
the surface. We decompose the covariant derivative of Y in the direction of X into its tangential and normal
components
∇X Y = ∇X Y + h(X, Y )N.
But then,
The quantity ∇X Y represents a covariant derivative on the surface, so in that sense, it is intrinsic to the
surface. If α(s) is a curve on the surface with tangent T = α0 (s), we say that a vector field Y is parallel-
transported along the curve if ∇T Y = 0. This notion of parallelism refers to parallelism on the surface, not
the ambient space. To illustrate by example, Figure 4.13 shows a vector field Y tangent to a sphere along the
circle with azimuthal angle θ = π/3. The circle has unit tangent T = α0 (s), and at each point on the circle,
the vector Y points North. To the inhabitants of the sphere, the vector Y appears parallel-transported on
the surface along the curve, that is ∇T Y = 0. However, Y is clearly not parallel-transported in the ambient
R3 space with respect to the connection ∇.
4.5. FUNDAMENTAL EQUATIONS 97
4.5.1 Proposition A Levi-Civita connection preserves length and angles under parallel transport.
Proof Let T = α0 (t) be tangent to curve α(T ), and X and Y be parallel-transported along α. By definition,
∇T X = ∇T Y = 0. Then
If one takes {eα } to be a basis of tangent space, the components of the connection in that basis are given
by the familiar equation
∇eα eβ = Γγ αβ eγ .
The Γ’s here are of course the same Christoffel symbols in the equation of Gauss 4.73. We have the following
important result:
4.5.2 Theorem In a manifold {M, g} with metric g, there exists a unique Levi-Civita connection.
The proof is implicit in the computations that follow leading to equation 4.76, which express the compo-
nents uniquely in terms of the metric. The entire equation (4.73) must be symmetric on the indices αβ, since
xαβ = xβα , so Γγαβ = Γγαβ is also symmetric on the lower indices. These quantities are called the Christoffel
symbols of the first kind. Now we take the inner product with xσ to deduce that
where we have lowered the third index with the metric on the right hand side of the last equation. The
quantities Γαβσ are called Christoffel symbols of the second kind. Here we must note that not all indices are
created equal. The Christoffel symbols of the second kind are only symmetric on the first two indices. The
notation Γαβσ = [αβ, σ] is also used in the literature.
The Christoffel symbols can be expressed in terms of the metric by first noticing that the derivative of
the first fundamental form is given by (see equation 3.34)
∂
gαγ,β = < xα , xγ >,
∂uβ
=< xαβ , xγ > + < xα , xγβ , >,
= Γαβγ + Γγβα .
98 CHAPTER 4. THEORY OF SURFACES
Adding the first two and subtracting the third of the equations above, and recalling that the Γ’s are symmetric
on the first two indices, we obtain the formula
1
Γαβγ = (gαγ,β + gβγ,α − gαβ,γ ). (4.75)
2
Raising the third index with the inverse of the metric, we also have the following formula for the Christoffel
symbols of the first kind (hereafter, Christoffel symbols refer to the symbols of the first kind, unless otherwise
specified.)
1
Γσαβ = g σγ (gαγ,β + gβγ,α − gαβ,γ ). (4.76)
2
The Christoffel symbols are clearly symmetric in the lower indices
Div X = ∇ · X = v α kα (4.79)
4.5. FUNDAMENTAL EQUATIONS 99
We get
Div X = v α ,α +Γα αγ v γ ,
∂ α 1 ∂
= v + (det(g))v γ ,
∂uα 2 det(g) ∂uγ
1 ∂ p
=p ( det(g)v α ). (4.80)
det(g) ∂uα
If f is a function on the manifold, df = f,β duβ so the contravariant components of the gradient are
∆f = Div(Gradf ),
1 ∂ p
=p ( det(g)g αβ f,β ) (4.82)
det(g) ∂uα
The quantity ∆ is called the Laplace-Beltrami operator on a function and it generalizes the Laplacian of
functions in Rn to functions on manifolds.
4.5.4 Example
As an example we unpack the formula for Γ111 . First, note that det(g) = kgαβ k = EG − F 2 . From
equation 4.76 we have
1 1γ
Γ111 = g (g1γ,1 + g1γ,1 − g11,γ ),
2
1 1γ
= g (2g1γ,1 − g11,γ ),
2
1 11
= [g (2g11,1 − g11,1 ) + g 12 (2g12,1 − g11,2 )],
2
1
= [GEu − F (2Fu − F Ev )],
2 det(g)
GEu − 2F Fu + F Ev
= .
2(EG − F 2 )
Due to symmetry, there are five other similar equations for the other Γ’s. Proceeding as above, we can derive
the entire set.
100 CHAPTER 4. THEORY OF SURFACES
∂ ∂
or in terms of the Euclidean R2 del operator ∇ =< ∂x , ∂y >,
∂ fx + ∂ q f y = 0,
q
∂x 2 2
1 + fx + fy ∂y 2 2
1 + fx + fy
" #
∇f
∇· p = 0. (4.90)
1 + k∇f k2
4.5. FUNDAMENTAL EQUATIONS 101
nα = −bγα xγ . (4.91)
xβδ = Γα
βδ xα + bβδ n,
xβδγ = Γα α
βδ,γ xα + Γβδ xαγ + bβδ,γ n + bβδ nγ
= Γα α µ α
βδ,γ xα + Γβδ [Γαγ xµ + bαγ n] + bβδ,γ n − bβδ bγ xα
µ
xβδγ = [Γα α α α
βδ,γ + Γβδ Γµγ − bβδ bγ ]xα + [Γβδ bαγ + bβδ,γ ]n, (4.92)
xβγδ = [Γα
βγ,δ + Γµβγ Γα
µδ − bβγ bα
δ ]xα + [Γα
βγ bαδ + bβγ,δ ]n. (4.93)
The last equation above was obtained from the preceding one just by permuting δ and γ. Subtracting that
last two equations and setting the tangential component to zero we get
Rαβγδ = bβδ bα α
γ − bβγ bδ , (4.94)
Technically we are not justified at this point in calling R a tensor since we have not established yet the
appropriate multi-linear features that a tensor must exhibit. We address this point in a later chapter.
Lowering the index above we get
Rαβγδ = bβδ bαγ − bβγ bαδ . (4.96)
The remarkable result is that the Riemann tensor and hence the Gaussian curvature does not depend on
the second fundamental form but only on the coefficients of the metric. Thus, the Gaussian curvature is an
intrinsic quantity independent of the embedding, so that two surfaces that have the same first fundamental
form have the same curvature. In this sense, the Gaussian curvature is a bending invariant!
Setting the normal components equal to zero gives
Γα α
βδ bαγ − Γβγ bαδ + bβδ,γ − bβγ,δ = 0 (4.98)
Computing the Riemann tensor is labor intensive since one must first obtain all the non-zero Christoffel
symbols as shown in the example above. Considerable gain in efficiency results from a form computation.
For this purpose, let {e1 , e2 , e3 } be a Darboux frame adapted to the surface M , with e3 = n. Let {θ1 , θ2 , θ3 }
be the corresponding orthonormal dual basis. Since at every point, a tangent vector X ∈ T M is a linear
combination of {e1 , e2 }, we see that θ3 (X) = 0 for all such vectors. That is, θ3 = 0 on the surface. As a
consequence, the entire set of the structure equations is
dθ1 = −ω 12 ∧ θ2 , (4.99)
2
dθ = −ω 21 ∧θ , 1
(4.100)
dθ3 = −ω 31 ∧ θ1 − ω 32 ∧ θ2 = 0, (4.101)
dω 12 = −ω 13 ∧ ω 32 , Gauss Equation (4.102)
dω 13 = −ω 12 ∧ ω 23 , Codazzi Equations (4.103)
dω 23 = −ω 21 ∧ ω 13 . (4.104)
dω 12 = K θ1 ∧ θ2 , (4.105)
ω 13 2
∧θ + ω 23 1
∧ θ = −2H θ ∧ θ . 1 2
(4.106)
Proof By applying the Weingarten map to the basis vector {e1 , e2 } of T M , we find a matrix representation
of the linear transformation:
Hence
dω 12 = K θ1 ∧ θ2 .
4.5.8 Definition A point of a surface at which K = 0 is called a planar point. A surface with K = 0 at
all points is called a flat or Gaussian flat surface. A surface on which H = 0 at all points is called a minimal
surface.
4.5. FUNDAMENTAL EQUATIONS 103
4.5.9 Example Sphere Since the first fundamental form is I = a2 dθ2 + a2 sin2 θ dφ2 , we have
θ1 = a dθ,
θ2 = a sin θ dφ,
dθ2 = a cos θ dθ ∧ dφ,
= − cos θ dφ ∧ θ1 = −ω 21 ∧ θ1 ,
ω 21 = cos θ dφ = −ω 12 ,
1
dω 12 = sin θ dθ ∧ dφ = (a dθ) ∧ (a sin θ dφ),
a2
1 1
= θ ∧ θ2 ,
a2
1
K = 2.
a
Thus, we have:
θ1 = a dθ,
θ2 = (b + a cos θ) dφ,
dθ2 = −a sin θ dθ ∧ dφ,
= sin θ dφ ∧ θ1 = −ω 21 ∧ θ1 ,
ω 21 = − sin θ dφ = −ω 12 ,
cos θ
dω 12 = cos θ dθ ∧ dφ = (a dθ) ∧ [(a + b cos θ) dφ],
a(b + a cos θ)
cos θ
= θ1 ∧ θ2 ,
a(b + a cos θ)
cos θ
K= .
a(b + a cos θ)
We have:
√
θ1 = E du,
2
√
θ = G dv,
√ √
dθ1 = ( E)v dv ∧ du = −( E)v du ∧ dv,
√
( E)v
=− √ du ∧ θ2 = −ω 12 ∧ θ2 ,
G
√ √
dθ2 = ( G)u du ∧ dv = −( G)u dv ∧ du
√
( G)u
=− √ dv ∧ θ2 = −ω 21 ∧ θ1 ,
E
√ √
( E)v ( G)u
ω 12 = √ du − √ dv
G E
√ ! √ !
∂ 1 ∂ G ∂ 1 ∂ E
dω 12 =− √ du ∧ dv + √ dv ∧ du,
∂u E ∂u ∂v G ∂v
" √ ! √ !#
1 ∂ 1 ∂ G ∂ 1 ∂ E
= −√ √ + √ θ1 ∧ θ2 .
EG ∂u E ∂u ∂v G ∂v
Therefore, the Gaussian Curvature of a surface mapped by a coordinate patch in which the parametric lines
are orthogonal is given by:
" √ ! √ !#
1 ∂ 1 ∂ G ∂ 1 ∂ E
K = −√ √ + √ . (4.107)
EG ∂u E ∂u ∂v G ∂v
Again, to connect with more classical notation, if a surface described by a coordinate patch x(u, v) has first
fundamental for given by I = E du2 + G dv 2 , then
dx = xu du + xv dv,
xu √ xv √
=√ E du + √ G dv,
E G
xu xv
= √ θ1 + √ θ2 ,
E G
dx = e1 θ + e2 θ2 ,
1
(4.108)
where
xu xv
e1 = √ , e2 = √ .
E G
Thus, when the parametric curves are orthogonal, the triplet {e1 , e2 , e3 = n} constitutes a moving orthonor-
mal frame adapted to the surface. The awkwardness of combining calculus vectors and differential forms in
the same equation is mitigated by the ease of jumping back and forth between the classical and the modern
formalism. Thus, for example, covariant differential of the normal in 4.104 can be rewritten without the
arbitrary vector in the operator LX as shown:
The equation just expresses the fact that the components of the Weingarten map, that is, the second
fundamental form in this basis, can be written as some symmetric matrix given by:
ω1 3 = l θ1 + m θ2 ,
ω2 3 = m θ1 + n θ2 . (4.111)
4.5. FUNDAMENTAL EQUATIONS 105
A slick proof of the theorema egregium can be obtained by differential forms. Let F : M → M f be an
isometry between two surfaces with metrics g and g̃ respectively. Let {eα } be an orthonormal basis for dual
basis {θα }. Define ẽα = F∗ eα . Recalling that isometries preserve inner products, we have
< ẽα , ẽβ >=< F∗ eα , F∗ eβ >=< eα , eβ >= δαβ .
f Let θ̃α be the dual forms and denote
Thus, {ẽα } is also an orthonormal basis of the tangent space of M.
with tilde’s the connection forms and Gaussian curvature of M.f
dθ̃α + ω̃ αβ ∧ θ̃β = 0,
F ∗ dθ̃α + F ∗ ω̃ αβ ∧ F ∗ θ̃β = 0,
dθα + F ∗ ω̃ αβ ∧ θβ = 0,
The connection forms are defined uniquely by the first structure equation, so
F ∗ ω̃ αβ = ω α β
.
c) In a similar manner, we compute the pull-back of the curvature equation:
So again by uniqueness, F ∗ K = K.
Gaussian curvatures K and K̃ follow from an easy computation using formula 4.107.
a2
p
−1 ∂ ∂
K=√ 2 2
u +a = 2 , (4.114)
u2 + a2 ∂u ∂u (u + a2 )2
√ √ !
ũ2 − a2 ∂ ũ2 − a2 a2
K̃ = − = − . (4.115)
ũ2 ∂ ũ ũ ũ4
It is immediately evident by substitution that as expected F ∗ K̃ = K. Figure 4.14 shows several stages of
a one-parameter family Mt of isometries deforming a catenoid into a helicoid. The one-parameter family of
coordinate patches chosen is
zt = (cos t) x + (sin t) y (4.116)
Writing the equation of the coordinate patch zt in complete detail, one can compute the coefficients of
the fundamental forms and thus establish the family of surfaces has mean curvature H independent of the
parameter t, and in fact H = 0 for each member of the family. We will discuss at a later chapter the geometry
of surfaces of zero mean curvature.
The tantalizing appearance of the Laplacian in this coordinate system gives an inkling that there is some
complex analysis lurking in the neighborhood. Readers acquainted with complex variables will recall that
the real and imaginary parts of holomorphic functions satisfy Laplace’s equations and that any holomorphic
function in the complex plane describes a conformal map. In anticipation of further discussion on this matter,
we prove the following:
4.5.15 Theorem Define the mean curvature vector H = Hn. If x(u, v) is an isothermal parametrization
of a surface, then
If follows that xuu + xvv is orthogonal to the surface and points in the direction of the normal n. On the
other hand,
Eg + Ge
= H,
2EG
g+e
= H,
2λ2
e + g = 2λ2 H,
< xuu + xvv , n > = 2λ2 H,
xuu + xvv = 2λ2 H.
108 CHAPTER 4. THEORY OF SURFACES
Chapter 5
Geometry of Surfaces
109
110 CHAPTER 5. GEOMETRY OF SURFACES
Chapter 6
Riemannian Geometry
stated, we are not assuming a priori that M is embedded (or immersed) in Euclidean space. If in addition the
space is equipped with a metric, the space is called a Riemannian manifold. If the signature of the metric is
of type g = diag(1, 1, . . . , −1, −1), with p ‘+’ entries and q ‘-’ entries, we say that M is a pseudo-Riemannian
manifold of type (p, q). As we have done with Minkowski’s space, we switch to Greek indices xµ for local
coordinates of curved space-times. We write the Riemannian metric as
ds2 = gµν dxµ dxν . (6.1)
We will continue to be consistent with earlier notation and denote the tangent space at a point p ∈ M as
Tp M , the tangent bundle as T M , and the space of vector fields as X (M ). Similarly, we denote the space of
differential k-forms by Ωk (M ), and the set of type rs tensor fields by Tsr (M ).
111
112 CHAPTER 6. RIEMANNIAN GEOMETRY
Product Manifolds
Suppose that M1 and M2 are differentiable manifolds of dimensions m1 and m2 respectively. Then,
M1 × M2 can be given a natural manifold structure of dimension n = m1 + m2 induced by the product of
coordinate charts. That is, if (φi1 , Ui1 ) is a chart in M1 in a neighborhood of p1 ∈ M1 , and (φi2 , Ui2 ) is a
chart in a neighborhood of p2 ∈ M2 in M2 , then the map
φi1× φi2 : Ui1× Ui2 → Rn
defined by
(φi1× φi2 )(p1 , p2 ) = (φi1(p1 ), φi2(p2 )),
is a coordinate chart in the product manifold. An atlas constructed from such charts, gives the differentiable
structure. Clearly, M1 ×M2 is locally diffeomorphic to Rm1×Rm2 . To discuss the tangent space of a product
manifold, we recall from linear algebra, that given two vector spaces V and W , the direct sum V ⊕ W is the
vector space consisting of the set of ordered pairs
V ⊕ W = {(v, w) : v ∈ V, w ∈ W },
together with the vector operations
(v1 , w1 ) + (v2 , w2 ) = (v1 + v2 , w1 + w2 ), for all, v1 , v2 ∈ V ; w1 , w2 ∈ W,
k(v, w) = (kv, kw), for all k ∈ R
People often say that one cannot add apples and peaches, but this is not a problem for mathematicians. For
example, 3 apples and 2 peaches plus 4 apples and 6 peaches is 7 apples and 8 peaches. This is the basic
idea behind the direct sum. We now have the following theorem:
More generally, if
ϕ : M1 × M2 → N
is a smooth manifold mapping, then we have a type of product rule formula for the Jacobian map,
This formula will be useful in the treatment of principal fiber bundles, in which case we have a bundle space
E, and a Lie group G acting on the right by a product manifold map µ : E × G → E.
6.2 Submanifolds
A Riemannian submanifold is a subset of a Riemannian manifold that is also Riemannian. The most
natural example is a hypersurface in Rn . If (x1 , x2 . . . xn ) are local coordinates in Rn with the standard
metric, and the surface M is defined locally by functions xi = xi (uα ), then M together with the induced
first fundamental form 4.12, has a canonical Riemannian structure. We will continue to use the notation ∇
for a connection in the ambient space and ∇ for the connection on the surface induced by the tangential
component of the covariant derivative
∇X Y = ∇X Y + H(X, Y ), (6.3)
where H(X, Y ) is the component in the normal space. In the case of a hypersurface, we have the classical
Gauss equation 4.74
∇X Y = ∇X Y + II(X, Y )N (6.4)
= ∇X Y + < LX, Y > N, (6.5)
where LX = −∇X N is the Weingarten map. If M is a submanifold of codimension n − k,P then there are k
normal vectors Nk and k classical second fundamental forms IIk (X, Y ), so that H(X, Y ) = k IIk (X, Y )Nk .
As shown by the theorema egregium, the curvature of a surface in R3 depends only on the first funda-
mental form, so the definition of Gaussian curvature as the determinant of the second fundamental form
does not even make sense intrinsically. One could redefine K by Cartan’s second structure equation as it was
used to compute curvatures in Chapter 4, but what we need is a more general definition of curvature that
is applicable to any Riemannian manifold. The concept leading to the equations of the theorema egregium
involved calculation of the difference of second derivatives of tangent vectors. At the risk of being somewhat
misleading, figure 4.95 illustrates the concept. In this figure, the vector field X consists of unit vectors
tangent to parallels on the sphere, and the vector field Y are unit tangents to meridians. If an arbitrary
tangent vector Z is parallel-transported from one point on an spherical triangle to the diagonally opposed
point, the result depends on the path taken. Parallel transport of Z along X followed by Y , would yield
114 CHAPTER 6. RIEMANNIAN GEOMETRY
a different outcome that parallel transport along Y followed by parallel transport along X. The failure
of the covariant derivatives to commute is reflection of the existence of curvature. Clearly, the analogous
parallel transport by two different paths on a rectangle in Rn yield the same result. This fact is the reason
why in elementary calculus, vectors are defined as quantities that depend only on direction and length. As
indicated, the picture is misleading, because, covariant derivatives, as is the case with any other type of
derivative, involves comparing the change of a vector under infinitesimal parallel transport. The failure of
a vector to return to itself when parallel-transported along a closed path is measured by an entity related
to the curvature called the holonomy of the connection. Still, the figure should help motivate the definition
that follows.
6.2.1 Definition On a Riemannian manifold with connection ∇, the curvature R and the torsion T are
defined by:
R(X, Y ) = ∇X ∇Y − ∇Y ∇X − ∇[X,Y ] , (6.6)
T (X, Y ) = ∇X Y − ∇Y X − [X, Y ]. (6.7)
6.2.2 Theorem The Curvature R is a tensor. At each point p ∈ M , R(X, Y ) assigns to each pair of
tangent vectors, a linear transformation from Tp M into itself.
Proof Let X, Y, Z ∈ X (M ) be vector fields on M . We need to establish that R is muiltilinear. Since clearly
R(X, Y ) = −R(Y, X), we only need to establish linearity on two slots. Let f be a C ∞ function. Then,
R(f X, Y ) = ∇f X ∇Y Z − ∇Y ∇f X Z − ∇[f X,Y ] Z,
= f ∇X ∇Y Z − ∇Y (f ∇X )Z − ∇[f XY −Y (f X)] Z,
= f ∇X ∇Y Z − Y (f )∇X )Z − f ∇Y ∇X Z − ∇f XY Z + ∇(Y (f )X+f Y X) Z,
= f ∇X ∇Y Z − Y (f )∇X )Z − f ∇Y ∇X Z − f ∇XY Z + ∇Y (f )X Z + ∇f Y X Z,
= f ∇X ∇Y Z − Y (f )∇X )Z − f ∇Y ∇X Z − f ∇XY Z + Y (f )∇X Z + f ∇Y X Z,
= f ∇X ∇Y Z − f ∇Y ∇X Z − f (∇XY Z − ∇Y X )Z,
= f R(X, Y )Z.
Similarly, recalling that [X, Y ] ∈ X , we get:
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−
[X, Y ](f )Z) − f ∇[X,Y ] (Z),
= f R(X, Y )Z.
We leave it as an almost trivial exercise to check linearity over addition in all slots.
Proof Since T (X, Y ) = −T (Y, X), it suffices to prove linearity on one slot. Thus,
T (f X, Y ) = ∇f XY − ∇Y (f X) − [f X, Y ],
= f ∇X Y − Y (f )X − f ∇Y X − f XY + Y (f X),
= f ∇X Y − Y (f )X − f ∇Y X − f XY + Y (f )X + f Y X,
= f ∇X Y − f ∇Y X − f [X, Y ],
= f T (X, Y ).
Again, linearity over sums is clear.
6.2.4 Theorem In a Riemannian manifold there exist a unique torsion free connection called the Levi-
Civita connection, that is compatible with the metric. That is:
[X, Y ] = ∇X Y − ∇Y X, (6.8)
∇X < Y, Z >= < ∇X Y, Z > + < Y, ∇X Z > . (6.9)
Proof The proof parallels the computation leading to equation 4.76.Let ∇ be a connection compatible with
the metric. By taking the three cyclic derivatives of the inner product, and subtracting the third from the
sum of the first two
(a) ∇X < Y, Z >= < ∇X Y, Z > + < Y, ∇X Z >,
(b) ∇Y < X, Z >= < ∇Y X, Z > + < X, ∇Y Z >,
(c) ∇Z < X, Y >= < ∇Z X, Y > + < X, ∇Z Y >,
(a) + (b) − (c) = < ∇X Y, Z > + < ∇Y X, Z > + < [X, Z], Y > + < [Y, Z], X >
=2 < ∇X Y, Z > + < [Y, X], Z > + < [X, Z], Y > + < [Y, Z], X >
Therefore:
< ∇X Y, Z >= 21 {∇X < Y, Z > +∇Y < X, Z > −∇Z < X, Y >
+ < [X, Y ], Z > + < [Z, X], Y > + < [Z, Y ], X >}. (6.10)
The bracket of any two vector fields is a vector field, so the connection is unique since it is completely
determined by the metric. In disguise, this is the formula in local coordinates for the Christoffel symbols
4.76. This follows immediately by choosing X = ∂/∂xα , Y = ∂/∂xβ and Z = ∂/∂xγ . Conversely, if one
defines ∇X Y by equation 6.10, a long but straightforward computation with lots of cancellations, shows that
this defines a connection compatible with the metric.
As before, if {eα } is a frame with dual frame {θa }, we define the connection forms ω, Christoffel symbols
Γ and torsion components in the frame by
∇X eβ = ω γβ (X) eγ , (6.11)
∇eα e β = Γγαβ eγ , (6.12)
T (eα , eβ ) = T γαβ eγ . (6.13)
As was pointed out in the previous chapter, if the frame is an orthonormal frame such as the coordinate
frame {∂/∂xµ } for which the bracket is zero, then T = 0 implies that the Christoffel symbols are symmetric
in the lower indices.
T γαβ = Γγαβ − Γγβα = 0.
For such a coordinate frame, we can compute the components of the Riemann tensors as follows:
R(eγ , eβ ) eδ = ∇eγ ∇eβ eδ − ∇eβ ∇eγ eδ ,
= ∇eγ (Γα α
βδ eα ) − ∇eβ (Γγδ eα ),
α µ
= Γα α µ α
βδ,γ eα + Γβδ Γγα eµ − Γγδ,β eα − Γγδ Γβα eµ ,
µ µ
= [Γα α α α
βδ,γ − Γβγ,δ + Γβδ Γγµ − Γβγ Γδµ ]eα ,
= Rαβγδ eα ,
116 CHAPTER 6. RIEMANNIAN GEOMETRY
µ µ
Rαβγδ = Γα α α α
βδ,γ − Γβγ,δ + Γβδ Γγµ − Γβγ Γδµ . (6.14)
Let X = X µ eµ be and α = Xµ θµ be a covariant and a contravariant vector field respectively. Using the
notation ∇α = ∇eα it is almost trivial to compute the covariant derivatives. The results are,
µ
∇β X = (X,β + X ν Γµ βν )eµ ,
∇β α = (Xµ,β − X ν Γν βµ )θµ , (6.15)
We show the details of the first computation, and leave the second one as an easy exercise
∇β X = ∇β (X µ eµ ), (6.16)
µ
= X,β eµ + X µ Γδβµ eδ , (6.17)
µ
= (X,β + X ν Γµ βν )eµ . (6.18)
In classical notation, the covariant derivatives X µ kβ and Xµkβ are given in terms of the tensor components,
µ
X µ kβ = X,β + X ν Γµ βν ,
Xµkβ = Xµ,β − Xν Γν βµ . (6.19)
Again, we show the computation for the first identity and leave the second as a exercise. We take the second
derivative. and then reverse the order,
µ
∇α ∇β X = ∇α (X,β eµ + X ν Γµ βν eµ ),
µ µ δ
= X,βα eµ + X,β Γαµ eδ + X ν,α Γµ ν µ ν µ δ
βµ eν + X Γβν,α eµ + X Γβν Γαν eδ ,
µ ν µ ν ν
∇α ∇β X = (X,βα + X,β Γαν + X,α Γβν + X ν Γµ ν δ µ
βν,α + X Γβν Γαδ )eµ ,
µ ν µ ν ν
∇β ∇α X = (X,αβ + X,α Γβν + X,β Γαν + X ν Γµ ν δ µ
αν,β + X Γαν Γβδ )eµ .
Subtracting the last two equations, only the last two terms of each survive, and we get the desired result,
2∇[α ∇β] (X) = X ν (Γµβν,α − Γµαν,β + Γδβν Γµαδ − Γδαν Γµβδ )eµ ,
2∇[α ∇β] (X µ eµ ) = (X ν Rµ ναβ )eµ .
At the risk of being repetitious, many authors use the notation ∇β X µ to denote the covariant derivative
X µ kβ , but it is really an (excusable) abuse of notation that arises from thinking of tensors as the components
of the tensors. The Ricci identities are the basis for the notion of holonomy, namely, the simple interpretation
that the failure of parallel transport to commute along the edges of an rectangle, indicates the presence of
curvature. With more effort with repeated use of Leibnitz rule, one can establish more elaborate Ricci
identities for higher order tensors. If one assumes zero torsion, the Ricci identities of higher order tensors
just involve more terms with the curvature. It the torsion is not zero, there are additional terms involving
the torsion tensor; in this case it is perhaps a bit more elegant to use the covariant differential introduced in
the next section, so we will postpone the computation until then.
The generalization of the theorema egregium to manifolds comes from the same principle of splitting
the curvature tensor of the ambient space into the tangential on normal components. In the case of a
6.2. SUBMANIFOLDS 117
If the ambient space is Rn , the curvature tensor R is zero, so we can set the horizontal and normal components
in the right to zero. Noting that the normal component is zero for all Z, we get:
R(X, Y )Z+ < LY, Z > LX− < LX, Z > LY = 0, (6.21)
∇X LY − ∇Y LX − L([X, Y ]) = 0. (6.22)
In particular, if n = 3, and at each point in the surface, the vectors X and Y constitute an a basis of the
tangent space, we get the coordinate-free theorema egregium
K =< R(X, Y )X, Y >=< LX, X >< LY, Y > − < LY, X >< LX, Y >= det(L). (6.23)
where we have introduced a coordinate-free definition of the differential of the one form θ by
dθ(X, Y ) = X(θ(Y )) − y(θ(X)) − θ([X, Y ]). (6.28)
It is easy to verify that this definition of the differential of a one form satisfies all the required properties of
the exterior derivative, and that it is consistent with the coordinate version of the differential introduced in
Chapter 2. We conclude that
Θα = dθα + ω αβ ∧ θβ , (6.29)
118 CHAPTER 6. RIEMANNIAN GEOMETRY
which is indeed the first Cartan equation of structure. Proceeding along the same lines, we compute:
Ωαβ = dω αβ + ω αγ ∧ ω γβ . (6.30)
The quantities connection and curvature forms are matrix-valued. Using matrix multiplication notation, we
can abbreviate the equations of structure as
Θ = dθ + ω ∧ θ,
Ω = dω + ω ∧ ω. (6.31)
Taking the exterior derivative of the structure equations gives some interesting results. Here is the first
computation,
dΘ = dω ∧ θ − ω ∧ dθ,
= dω ∧ θ − ω ∧ (Θ − ω ∧ θ),
= dω ∧ θ − ω ∧ Θ + ω ∧ ωθ,
= (dω + ω ∧ ω) ∧ θ − ω ∧ Θ,
= Ω ∧ θ − ω ∧ Θ,
so,
dΘ + ω ∧ Θ = Ω ∧ θ. (6.32)
Similarly, taking d of the second structure equation we get,
dΩ = dω ∧ ω + ω ∧ dω,
= (Ω − ω ∧ ω) ∧ ω + ω ∧ (Ω − ω ∧ ω).
Hence,
dΩ = Ω ∧ ω − ω ∧ Ω. (6.33)
Equations 6.32 and 6.33 are called the first and second Bianchi identities. The relationship between the
torsion and Riemann tensor components with the corresponding differential forms are given by
Θα = 21 T αγδ θγ ∧ θδ ,
Ωαβ = 21 Rαβγδ θγ ∧ θδ . (6.34)
In the case of a non-coordinate frame in which the Lie bracket of frame vectors does not vanish, we first
write them as linear combinations of the frame
α
[eβ , eγ ] = Cβγ eα . (6.35)
The components of the torsion and Riemann tensors are then given by
T αβγ = Γα α α
βγ − Γγβ − Cβγ ,
µ µ µ
Rαβγδ = Γα α α α α α σ
βδ,γ − Γβγ,δ + Γβδ Γγµ − Γβγ Γδµ − Γβµ Cγδ − Γ σβ C γδ . (6.36)
6.3. SECTIONAL CURVATURE 119
The Riemann tensor for a torsion-free connection has the following symmetries;
The last cyclic equation is the tensor version of the first Bianchi Identity with 0 torsion. It follows immediately
from setting Ω ∧ θ = 0 and taking a cyclic permutation of the antisymmetric indices {β, γ, δ} of the Riemann
tensor. The symmetries reduce the number of independent components in an n-dimensional manifold from
n4 to n2 (n2 − 1)/12. Thus, for a 4-dimensional space, there are at most 20 independent components. The
derivation of the tensor version of the second Bianchi identity from the elegant differential forms version,
takes a bit more effort. In components the formula
dΩ = Ω ∧ ω − ω ∧ Ω
reads,
Rα βκλ;µ θµ ∧ θκ ∧ θλ = (Γρ µβ Rα ρκλ − Γα µρ Rρ βκλ )θµ ∧ θκ ∧ θλ ,
where we used the notation,
∇µ Rα βκλ = Rα βκλ;µ .
Taking a cyclic permutation on the antisymmetric indices κ, λ, µ, and using some index gymnastics to show
that the right hand becomes zero, the tensor version of the second Bianchi identity for zero torsion becomes
Rα β[κλ;µ] = 0 (6.39)
R : Tp (M ) ⊗ Tp (M ) ⊗ Tp (M ) ⊗ Tp (M ) → R,
by
Now, consider any 2-dimensional plane Vp ⊂ Tp (M ) and let X, Y ∈ V be linearly independent. Then,
is a bilinear form that represents the area of the parallelogram spanned by X and Y . If we perform a linear,
non-singular change of of coordinates,
X 0 = aX + bY, y 0 = cX + dY, ad − bc 6= 0,
then both, G(X, Y.X, Y ) and R(X, Y, X, Y ) transform by the square of the determinant D = ad − bc, so that
ratio is independent of the choice of vectors. We define the sectional curvature of the subspace Vp by
R(X, Y, X, Y )
K(Vp ) = ,
G(X, Y, X, Y )
R(X, Y, X, Y )
= (6.42)
< X, X >< Y, Y > − < X, Y >2
The set of values of the sectional curvatures for all planes at Tp (M ) completely determines the Riemannian
curvature at p. For a surface in R3 the sectional curvature is the Gaussian curvature, and the formula is
equivalent to the theorema egregium. If K(Vp ) is constant for all planes Vp ∈ Tp (M ) and for all points
p ∈ M , we say that M is a space of constant curvature. For a space of constant curvature k, we have
6.3.1 Example
The model space of manifolds of constant curvature is a quadric hypersurface M of Rn+1 with metric
k 2 t2 + Σi (y i )2 = k 2 .
If k = 0, the space is flat. If = 1, let (y 0 )2 = k 2 t2 and the quadric is isometric to a sphere of constant
curvature 1/k 2 . If = −1, Σi (xi )2 = −k 2 (1 − t2 ) > 0, then t2 < 1 and the surface is a hyperboloid of two
sheets. Consider the mapping from (R)n+1 to Rn given by
xi = y i /t.
We would like to compute the induced metric on the the surface. We have
−k 2 t2 + Σi (y i )2 = −k 2 t2 + t2 Σi (xi )2 = −k 2
so
−k 2
t2 = .
−k 2 + Σi (xi )2
Taking the differential, we get
−k 2 Σi (xi dxi )
t dt = .
−k 2 + Σi (xi )2
6.3. SECTIONAL CURVATURE 121
−k 2 (Σi xi dxi )2
dt2 = .
(−k 2 + Σi (xi )2 )3
It is not obvious, but in fact, the space is also of constant curvature (−1/k 2 ). For an elegant proof, see [?].
When n = 4 and = −1, the group leaving the metric
invariant, is the Lorentz group O(1, 4). With a minor modification of the above, consider the quadric
M : −k 2 t2 + (y 1 )2 + . . . (y 4 )2 = k 2 .
In this case, the quadric is a hyperboloid of one sheet, and the submanifold with the induced metric is
called the de Sitter space. The isotropy subgroup that leaves (1, 0, 0, 0, 0) fixed is O(1, 3) and the manifold is
diffeomorphic to O(1, 4)/O(1, 3). Many alternative forms of the de Sitter metric exist in the literature. One
that is particularly appealing is obtained as follows. Write the metric in ambient space as
y 0 = k sinh(τ /k),
y i = k cosh(τ /k).
Then, we have
where dΩ is the volume form for S 3 . The most natural coordinates for the volume form are the Euler
angles and Cayley-Klein parameters. The interpretation of this space-time is that we have a spatial 3-sphere
which propagates in time by shrinking to a minimum radius at the throat of the hyperboloid, followed by
an expansion. Being a space of constant curvature, the Ricci tensor is proportional to the metric, so this is
an Einstein manifold.
122 CHAPTER 6. RIEMANNIAN GEOMETRY
6.4 Big D
In this section we discuss the notion of a connection on a vector bundle E. Let M be a smooth manifold
and as usual we denote by Tsr (p) the vector space of type rs tensors at a point p ∈ M . The formalism
applies to any vector bundle, but in this section we are primarily concerned with the case where E is the
tensor bundle E = Tsr (M ). Sections Γ(E) = Tsr (M ) of this bundle are called tensor fields on M . For general
vector bundles, we use the notation s ∈ Γ(E) for the sections of the bundle. The section that maps every
point of M to the zero vector, is called the zero section. Let {eα } be an orthonormal frame with dual forms
{θα }. We define the space Ωp (M, E) tensor-valued p-form as sections of the bundle,
T = Tβα11,...β
,...αr
s ,γ1 ,...,γp
eα1 ⊗ . . . eαr ⊗ θβ1 ⊗ · · · ⊗ θβs ∧ θγ1 ∧ . . . ∧ θγp . (6.46)
A tensor-valued 0-form is just a regular tensor field T ∈ Tsr (M ) The main examples of tensor-valued forms
are the torsion and the curvature forms
Θ = Θ α ⊗ eα ,
Ω = Ωα β ⊗ eα ⊗ θβ . (6.47)
T = T α βγ eα ⊗ θβ ⊗ θγ ,
1
= T α βγ eα ⊗ θβ ∧ θγ ,
2
= eα ⊗ ( 21 T α βγ θβ ∧ θγ ).
since the tensor is antisymmetric in the lower indices. Similarly, the tensorial components of the curvature
are
1 α
Ω= R βγδ eα ⊗ θβ ⊗ θγ ∧ θδ ,
2
= eα ⊗ θβ ⊗ ( 21 Rα βγδ θγ ∧ θδ ).
∇X : Γ(Tsr ) → Γ(Tsr )
If instead of the tensor bundle we have a general vector bundle E, we replace the tensor fields in the definition
above by sections s ∈ Γ(E) of the vector bundle. The definition induces a derivation on the entire tensor
algebra satisfying the additional conditions,
5) ∇X (T1 ⊗ T2 ) = ∇X T1 ⊗ T2 + T1 ⊗ ∇X T2 ,
6) ∇X ◦ C = C ◦ ∇X , for any contraction C.
The properties are the same as a Koszul connection, or covariant derivative for tensor-valued 0 forms T .
Given an orthonormal frame, consider the identity tensor,
I = δ α β eα ⊗ θ β , (6.49)
and take the covariant derivative ∇X . We get
∇X eα ⊗ θα + eα ⊗ ∇X θα = 0,
eα ⊗ ∇X θα = −∇X eα ⊗ θα ,
= −eβ ω β α (X) ⊗ θα ,
eβ ⊗ ∇X θβ = −eβ ω β α (X) ⊗ θα ,
which implies that,
∇X θβ = −ω β α (X)θα . (6.50)
Thus as before, since we have formulas for the covariant derivative of basis vectors and forms, we are led
by induction to a general formula for the covariant derivative of an rs -tensor given mutatis mutandis by
the formula 3.32. In other words, the covariant derivative of a tensor acquires a term with a multiplicative
connection factor for each contravariant index and a negative term with a multiplicative connection factor
for each covariant index.
As a reminder of the definition of the inner product iX , condition (c) is equivalent to the equation,
∇T (θ1 , . . . , θr , X, X1 , . . . , Xs ) = (∇X T )(θ1 , . . . , θr , X1 , . . . , Xs ).
In particular, if X is vector field, then, as expected
∇X(Y ) = ∇X Y,
The operator ∇ is called the covariant differential. Again, for a general vector bundles, we denote the
sections by s ∈ Γ(E) and the covariant differential by ∇s.
θ = eθ ⊗ θ α
which has the same components as the identity 11 tensor. This is a kind of a odd creature. About the
dx = i dx + j dy + k dz,
which is sort of a mixture of a vector and a form. The vector of differential forms would then be written as
a column vector.
In a frame {eα }, the covariant differential of tensor-valued 0-form T is given by
∇T = ∇eα T ⊗ θα ≡ ∇α T ⊗ θα
In particular, if X = v α eα , we get,
∇X = ∇β X ⊗ θβ = ∇β (v α eα ) ⊗ θβ ,
= (∇β (v α ) eα + v α Γγβα eγ ) ⊗ θβ ,
α
= (v,β + v γ Γα
βγ )eα ⊗ θ
β
α
= vkβ eα ⊗ θ β ,
v α kβ = v α ,β + Γα γ
βγ v , (6.53)
α
for the covariant derivative components vkβ and the comma to abbreviate the directional derivative ∇β (v α ).
Of course, the formula is in agreement with equation 3.25. ∇X is a 11 -tensor.
∇α = ∇(vα ⊗ θα )
= ∇vα ⊗ θα − vβ ω β α ⊗ θα ,
= (∇γ vα θγ − vβ Γβαγ θγ ) ⊗ θα ,
= (vα,γ − Γα γ α
βγ vα ) θ ⊗ θ ,
hence,
vαkβ = vα,γ − Γα
βγ vα . (6.54)
As promised earlier, we now prove the Ricci identities for contravariant and covariant vectors when the
torsion is not zero. Ricci Identities with torsion. The results are,
∇X = ∇β X ⊗ θβ ,
∇2 X = ∇(∇β X ⊗ θβ ),
= ∇(∇β X) ⊗ θβ + ∇β X ⊗ ∇θβ ,
= ∇α ∇β X ⊗ θ β ⊗ θ α − ∇ β ⊗ ω β α ⊗ θ α ,
= ∇α ∇β X ⊗ θβ ⊗ θα − ∇µ X ⊗ Γµαβ θβ ⊗ θα ,
∇2 X = (∇α ∇β X − ∇µ X Γµαβ ) θβ ⊗ θα .
6.4. BIG D 125
On the other hand, we also have ∇X = ∇α X ⊗ θα , so we can compute ∇2 by differentiating in the reverse
order to get the equivalent expression,
∇2 X = (∇β ∇α X − ∇µ X Γµβα ) θα ⊗ θβ .
Subtracting the last two equations we get an alternating tensor, or a two-form that we can set equal to zero.
For lack of a better notation we call this form [∇, ∇]. The notations Alt(∇2 ) and ∇ ∧ ∇ also appear in the
literature. We get
[∇, ∇] = [∇α ∇β − ∇β ∇α )X − ∇µ X(Γµαβ − Γµβα )]θβ ∧ θα ,
= [∇α ∇β − ∇β ∇α − ∇[eα ,β] )X + ∇[eα ,β] X − ∇µ X(Γµαβ − Γµβα )]θβ ∧ θα ,
µ
= [R(eα , eβ )X + Cαβ ∇µ X − ∇µ X(Γµαβ − Γµβα ]θβ ∧ θα ,
µ
= [R(eα , eβ )X + Cαβ − ∇µ X(Γµαβ − Γµβα − Cαβ
µ
]θβ ∧ θα ,
= 21 (X ν Rµ ναβ − ∇µ X T µ αβ )θβ ∧ θα .
Θ = Dθ,
Ω = dω + ω ∧ ω, (6.59)
DΘ = Ω ∧ θ,
DΩ = 0 (6.60)
With apologies for the redundancy, we reproduce the change of basis formula 3.49. Let e0 = eB be an
orthogonal change of basis. Then
De0 = e ⊗ dB + DeB,
= e ⊗ dB + (e ⊗ ω)B,
= e0 ⊗ (B −1 dB + B −1 ωB),
= e0 ⊗ ω 0 ,
where,
ω 0 = B −1 dB + B −1 ωB. (6.61)
Multiply the last equation by B and take the exterior derivative d. We get.
Bω 0 = dB + ωB,
Bdω 0 + dB ∧ ω 0 = dωB − ω ∧ dB,
Bdω + (Bω 0 − ωB) ∧ ω 0 = dωB − ω ∧ (ω 0 B − ωB),
0
Ω0 = B −1 ΩB. (6.62)
As pointed out after equation 3.49, the curvature is a tensorial form of adjoint type. The transformation
law above for the connection has an extra term, so it is not tensorial. It is easy to obtain the classical
transformation law for the Christoffel symbols from equation 6.61. Let {xα } be coordinates in a patch
(φα , Uα ), and {y β } be coordinates on a overlapping patch (φβ , Uβ ). The transition functions φαβ are given
by the Jacobian of the change of coordinates,
∂ ∂xα ∂
= ,
∂y β ∂y β ∂xα
∂xα
φαβ = .
∂y β
Inserting the connection components ω 0α β = Γ0α γ
βγ dy , into the change of basis formula 6.61, with B = φαβ ,
1
we get ,
ω 0α β = (B −1 )α κ dB κ β + (B −1 )α κ ω κ λ B λ β ,
∂y α
κ
∂x ∂y α κ ∂xλ
= d + ω λ ,
∂xκ ∂y β ∂xκ ∂y β
∂y α ∂ 2 xκ ∂y α κ ∂xλ
Γ0α γ
βγ dy = κ σ β
dy σ + κ
Γλσ dxσ ,
∂x ∂y ∂y ∂x ∂y β
∂y α ∂ 2 xκ ∂y α κ ∂xσ ∂xλ
Γ0α
βγ = + Γ .
κ γ
∂x ∂y ∂y β ∂xκ λσ ∂y γ ∂y β
1 We use this notation reluctantly, to be consistent with most literature. The notation results in violation of the index
notation. We really should be writing φα β , since in this case, the transition functions are matrix-valued.
6.4. BIG D 127
Thus, we retrieve the classical transformation law for Christoffel symbols that one finds in texts on general
relativity.
α
κ ∂y ∂xσ ∂xλ ∂y α ∂ 2 xκ
Γ0α
βγ = Γλσ + (6.63)
∂xκ ∂y γ ∂y β ∂xκ ∂y γ ∂y β
6.4.4 Parallelism
When first introduced to vectors in elementary calculus and physics courses, vectors are often described
as entities characterized by a direction and a length. This primitive notion, that two such entities in Rn with
the same direction and length represent the same vector, regardless of location, is not erroneous in the sense
that parallel translation of a vector in Rn does not change the attributes of a vector as described. In
elementary linear algebra, vectors are described as n-tuples in Rn equipped with the operations of addition
and multiplication by scalar, and subject to eight vector space properties. Again, those vectors can be
represented by arrows which can be located anywhere in Rn as long as they have the same components.
This is another indication that parallel transport of a vector in Rn is trivial, a manifestation of the fact the
Rn is a flat space. However, in a space that is not flat, such as s sphere, parallel transport of vectors is
intimately connected with the curvature of the space. To elucidate this connection, we first describe parallel
transport for a surface in R3 .
6.4.2 Definition Let uα (t) be a curve on a surface x = x(uα ), and let V = α0 (t) = α∗ ( dt
d
) be the velocity
vector as defined in 1.25. A vector field Y is called parallel along α if
∇V Y = 0,
6.4.3 Definition A curve α(t) with velocity vector V = α0 (t) is called a geodesic or self-parallel if
∇V V = 0.
∇V V = ∇vT (vT ),
= v∇T (vT ),
dv
= v T + v 2 ∇T T,
dt
1 d
= 2 (v 2 )T + v 2 ∇T T
dt
We have < T, T >= 1, so < ∇T T, T >= 0 which shows that ∇T T is orthogonal to T . We also have v > 0.
Since both the tangential and the normal components need to vanish, the theorem follows.
If M is a hypersurface in Rn with unit normal n, we gain more insight on the geometry of geodesics as a
direct consequence of the discussion above. Without real loss of generality consider the geometry in the case
of n = 3. Since α is geodesic, we have kα0 k2 =< α0 , α0 >=constant. Differentiation gives < α0 , α00 >= 0, so
that the acceleration α00 is orthogonal to α0 . Comparing with equation 4.34 we see that T 0 = κn n, which
reinforces the fact that the entire curvature of the curve is due to the normal curvature of the surface as
128 CHAPTER 6. RIEMANNIAN GEOMETRY
a submanifold of the ambient space. In this sense, inhabitants constrained to live on the surface would be
unaware of this curvature, and to them, geodesics would appear locally as the straightest path to travel.
Thus, for a sphere in R3 of radius a, the acceleration α00 of a geodesic only has a normal component, and
the normal curvature is 1/a. That is, the geodesic must lie along a great circle.
6.4.5 Theorem Let α(t) by curve with velocity V . For each vector Y in the tangent space restricted to
the curve, there is a unique vector field Y (t) locally obtained by parallel transport.
Proof We choose local coordinates with frame field {eα = ∂u∂α }. We write the components of the vector
fields in terms of the frame
∂
Y = yβ ,
∂uβ
duα ∂
V = . then,
dt ∂uα
∇T V = ∇u˙α eα (y β eβ ),
= u˙α ∇eα (y β eβ ),
duα ∂y β
= + u˙α y β Γγ αβ eγ ,
dt ∂uα
dy γ duα γ
=[ + yβ Γ αβ ]eγ.
dt dt
So, Y is parallel along the curve iff,
dy γ duα γ
+ yβ Γ αβ = 0. (6.64)
dt dt
The existence and uniqueness of the coefficients y β that define Y are guaranteed by the theorem on existence
and uniqueness of differential equations with appropriate initial conditions.
∇V V = ∇u̇α eα [u˙β eβ ],
= u̇α ∇eα [u̇β eβ ],
∂ u̇β
= u̇α [ eβ + u̇β ∇eα eβ ],
∂uα
∂ u̇β
= u̇α α eβ + u̇α u̇β ∇eα eβ ,
∂u
du ∂ u̇β
α
= eβ + u̇α u̇β Γσαβ eσ ,
dt ∂uα
= üβ eβ + u̇α u̇β Γσαβ eσ ,
= [üσ + u̇α u̇β Γσαβ ]eσ .
The existence and uniqueness theorem for solutions of differential equations leads to the following theorem
6.4.6 Theorem Let p be a point in M and V a vector Tp M . Then, for any real number t0 , there exists
a number δ and a curve α(t) defined on [t0 − δ, t0 + δ], such that α(t0 ) = p, α0 (t0 ) = V , and α is a geodesic.
For a general vector bundles E over a manifold M , a section s ∈ Γ(E) of a vector bundle is called a parallel
section if
∇s = 0. (6.66)
6.5. LORENTZIAN MANIFOLDS 129
We discuss the length minimizing properties geodesics in section 6.6 and provide a number of examples for
surfaces in R3 and for Lorentzian manifolds. Since geodesic curves have zero acceleration, in Euclidean
space they are straight lines. In Einstein’s theory of Relativity, gravitation is fictitious force caused by the
curvature of space time, so geodesics represent the trajectory of free particles.
years earlier from a seminar at the Dublin Institute for Advanced Studies. The most general, spherically
symmetric, static solution of the Einstein vacuum equations is the Schwarzschild metric 2
1
ds2 = 1 − 2GM
dt2 − dr2 − r2 dθ2 − r2 sin2 θ dφ2
r 2GM
(6.74)
1− r
so that,
1
dt = du + dr
1 − 2m
r
ds2 = 2drdu + [1 − 2m
r ] du
2
− r2 dθ2 − r2 sin2 θ dφ2 . (6.75)
In these coordinates it is evident that the event horizon r = 2m is not a real singularity. The Vaidya metric
is the generalization
ds2 = 2drdu + [1 − 2m(u) 2 2 2 2 2
r ] du − r dθ − r sin θ dφ ,
2
(6.76)
where m(u) is now an arbitrary function. The geometry described by the Vaidya solution to Einstein
equations, represents the gravitational field in the exterior of a radiating, spherically symmetric star. In all
our previous curvature computations by differential forms, the metric has been diagonal; this is an instructive
example of one with a non-diagonal metric. The first step in the curvature computation involves picking out
a basis of one-forms. The idea is to pick out the forms so that in the new basis, the metric has constant
coefficients. One possible choice of 1-forms is
θ0 = du,
2m(u)
θ1 = dr + 12 [1 − r ] du,
θ2 = r dθ,
θ3 = r sin θ dφ. (6.77)
where
g01 = g10 = −g22 = −g33 = 1,
while all the other gαβ = 0. In the coframe, the metric has components:
0 1 0 0
1 0 0 0
gαβ = . (6.78)
0 0 −1 0
0 0 0 −1
Since the coefficients of the metric are constant, the components ωαβ of the connection will be antisymmetric.
This means that
ω00 = ω11 = ω22 = ω33 = 0.
We thus conclude that
ω 10 = g 10 ω00 = 0,
ω 01 = g 01 ω11 = 0,
2 The 2GM
Schwarzschild radius is r = c2
, but here we follow the common convention of setting c = 1.
6.5. LORENTZIAN MANIFOLDS 131
ω 22 = g 22 ω22 = 0,
ω 33 = g 33 ω33 = 0.
To compute the connection, we take the exterior derivative of the basis 1-forms. The result of this compu-
tation is
dθ0 = 0,
m m
dθ1 = −d[ m
r du] = 2
dr ∧ du = 2 θ1 ∧ θ0 ,
r r
2 1 1 2 1 0 2
dθ = dr ∧ dθ = θ ∧ θ − [1 − 2m r ]θ ∧θ ,
r 2r
dθ3 = sin θ dr ∧ dφ + r cos θ dθ ∧ dφ,
1 1 1
= θ1 ∧ θ3 − [1 − 2m r ] θ0 ∧ θ3 + cot θ θ2 ∧ θ3 . (6.79)
r 2 r
For convenience, we write below the first equation of structure [6.24] in complete detail.
dθ0 = ω 00 ∧ θ0 + ω 01 ∧ θ1 + ω 02 ∧ θ2 + ω 03 ∧ θ3 ,
dθ1 = ω 10 ∧ θ0 + ω 11 ∧ θ1 + ω 12 ∧ θ2 + ω 13 ∧ θ3 ,
dθ2 = ω 20 ∧ θ0 + ω 21 ∧ θ1 + ω 22 ∧ θ2 + ω 23 ∧ θ3 ,
dθ3 = ω 30 ∧ θ0 + ω 31 ∧ θ1 + ω 32 ∧ θ2 + ω 33 ∧ θ3 . (6.80)
Since the ω’s are one-forms, they must be linear combinations of the θ’s. Comparing Cartan’s first structural
equation with the exterior derivatives of the coframe, we can start with the initial guess for the connection
coefficients below:
m 0
ω 10 = 0, ω 11 = θ , ω 12 = A θ2 , ω 13 = B θ3 ,
r2
1 2m 2 1
ω 20 = − [1 − ]θ , ω 21 = θ2 , ω 22 = 0, ω 23 = C θ3 ,
2 r r
3 1 2m 3 1 3 1
ω 0 = − [1 − ]θ , ω 31 = θ , ω 32 = cot θ θ3 , ω 33 = 0.
2 r r r
Here, the quantities A, B, and C are unknowns to be determined. Observe that these are not the most
general choices for the ω’s. For example, we could have added a term proportional to θ1 in the expression
for ω 11 , without affecting the validity of the first structure equation for dθ1 . The strategy is to interactively
tweak the expressions until we set of forms completely consistent with Cartan’s structure equations.
We now take advantage of the skewsymmetry of ωαβ , to determine the other components. The find A,
B and C, we note that
ω 12 = g 10 ω02 = −ω20 = ω 20 ,
ω 13 = g 10 ω03 = −ω30 = ω 30 ,
ω 23 = g 22 ω23 = ω32 = −ω 32 .
From equations [??], [??], and [??], we find that
1 2m 1 2m 1
A = − [1 − ], B = − [1 − ], C = − cot θ. (6.81)
2 r 2 r r
Similarly, we have
ω 00 = −ω 11 ,
ω 02 = ω 21 ,
ω 03 = ω 31 ,
132 CHAPTER 6. RIEMANNIAN GEOMETRY
hence,
m 0
ω 00 = − θ ,
r2
1
ω 02 = − θ2 ,
r
0 1 3
ω 3= θ .
r
It is easy to verify that our choices for the ω’s are consistent with first structure equations, so by uniqueness,
these must be the right values.
There is no guesswork in obtaining the curvature forms. All we do is take the exterior derivative of the
connection forms and pick out the components of the curvature from the second Cartan equations [6.25].
Thus, for example, to obtain Ω11 , we proceed as follows.
Ω11 = dω 11 + ω 11 ∧ ω 11 + ω 12 ∧ ω 21 + ω 13 ∧ ω 31 ,
m 1 2m 1
= d[ 2 θ0 ] + 0 − 2 [1 − ] ω 3 ∧ ω 31 + (θ2 ∧ θ2 + θ3 ∧ θ3 ),
r 2r r
2m
= − 3 dr ∧ θ0 ,
r
2m 1
= − 3 θ ∧ θ0 .
r
The computation of the other components is straightforward and we just present the results.
1 dm 2 m
Ω1 2 = − θ ∧ θ0 − 3 θ1 ∧ θ2 ,
r2 du r
1 dm 3 m
Ω1 3 = −2
θ ∧ θ0 − 3 θ1 ∧ θ3 ,
r du r
2 m 2 0
Ω 1 = 3θ ∧ θ ,
r
m
3
Ω 1 = 3 θ3 ∧ θ0 ,
r
2m
Ω2 3 = 3 θ 2 ∧ θ 3 .
r
By antisymmetry, these are the only independent components. We can also read the components of the full
Riemann curvature tensor from the definition
1
Ωαβ = Rαβγδ θγ ∧ θδ . (6.82)
2
Thus, for example, we have
1
Ω11 = R11γδ θγ ∧ θδ ,
2
hence
2m
R1101 = −R1110 = 3 ; other R11γδ = 0.
r
Using the antisymmetry of the curvature forms, we see, that for the Vaidya metric Ω10 = Ω00 = 0,
Ω20 = −Ω12 , etc., so that
R00 = R2020 + R3030
= R1220 + R1330
Substituting the relevant components of the curvature tensor, we find that
1 dm
R00 = 2 (6.83)
r2 du
while all the other components of the Ricci tensor vanish. As stated earlier, if m is constant, we get the
Ricci flat Schwarzschild metric.
6.6. GEODESICS 133
6.6 Geodesics
Geodesics were introduced in the section on parallelism. The equation of geodesics on a manifold given by
equation 6.65 involves the Christoffel symbols. Whereas it is possible to compute all the Christoffel symbols
starting with the metric as in equation 4.76, this is most inefficient, as it is often the case that many of the
Christoffel symbols vanish. Instead, we show next how to obtain the geodesic equations by using variational
principles Z
δ L(uα , u̇α , s) ds = 0, (6.84)
to minimize the arc length. Then we can pick out the non-vanishing Christoffel symbols from the geodesic
equation. Following the standard methods of Lagrangian mechanics, we let uα and u̇α be treated as inde-
pendent (canonical) coordinates and choose the Lagrangian in this case to be
The choice will actually result in minimizing the square of the arc length, but clearly this is an equivalent
problem. It should be observed that the Lagrangian is basically a multiple of the kinetic energy 12 mv 2 . The
motion dynamics are given by the Euler-Lagrange equations.
d ∂L ∂L
− = 0. (6.86)
ds ∂ u̇γ ∂uγ
Applying this equations keeping in mind that gαβ is the only quantity that depends on uα , we get:
d α β α β α β
0= ds [gαβ δγ u̇ + gαβ u̇ δγ ] − gαβ,γ u̇ u̇
d β α α β
= ds [gγβ u̇ + gαγ u̇ ] − gαβ,γ u̇ u̇
= gγβ üβ + gαγ üα + gγβ,α u̇α u̇β + gαγ,β u̇β u̇α − gαβ,γ u̇α u̇β
= 2gγβ üβ + [gγβ,α + gαγ,β − gαβ,γ ]u̇α u̇β
= δβσ üβ + 12 g γσ [gγβ,α + gαγ,β − gαβ,γ ]u̇α u̇β
where the last equation was obtained contracting with 12 g γσ to raise indices. Comparing with the expression
for the Christoffel symbols found in equation 4.76, we get
sin2 θ φ̇ = k.
134 CHAPTER 6. RIEMANNIAN GEOMETRY
Rather than trying to solve the second Euler-Lagrange equation for θ, we evoke a standard trick that involves
reusing the metric. It goes as follows:
dφ
sin2 θ = k,
ds
sin2 θ dφ = k ds,
sin4 θ dφ2 = k 2 ds2 ,
sin4 θ dφ2 = k 2 (a2 dθ2 + a2 sin2 θ dφ2 ),
(sin4 θ − k 2 a2 sin2 θ) dφ2 = a2 k 2 dθ2 .
The last equation above is separable and it can be integrated using the substitution u = cot θ.
ak
dφ = p dθ,
sin θ sin2 θ − a2 k 2
ak
= 2
√ dθ,
sin θ 1 − a2 k 2 csc2 θ
ak
= p dθ,
sin2 θ 1 − a2 k 2 (1 + cot2 θ)
ak csc2 θ
= p dθ,
1 − a2 k 2 (1 + cot2 θ)
ak csc2 θ
= p dθ,
(1 − a2 k 2 ) − a2 k 2 cot2 θ
csc2 θ
= q dθ,
1−a2 k2 2
2
a k 2 − cot θ
−1 1−a2 k2
=√ du, where( c2 = a2 k2 ).
c2 − u2
φ = − sin−1 ( 1c cot θ) + φ0 .
Here, φ0 is the constant of integration. To get a geometrical sense of the geodesics equations we have just
derived, we rewrite the equations as follows:
cot θ = c sin(φ0 − φ),
cos θ = c sin θ(sin φ0 cos φ − cos φ0 sin φ),
a cos θ = (c sin φ0 )(a sin θ cos φ) − (c cos φ0 )(a sin θ sin φ.)
z = Ax − By, where A = c sin φ0 , B = c cos φ0 .
We conclude that the geodesics of the sphere are great circles determined by the intersections with planes
through the origin.
Of course, we could use the expressions for the equations of geodesics we just derived above, but since the
coefficients are functions of r only, it is just a easy to start from the Lagrangian,
L = (1 + f 02 ) ṙ2 + r2 φ̇2 .
Since there is no dependance on φ, the Euler-Lagrange equation on φ gives rise to a conserved quantity.
d
(2r2 φ̇) = 0,
ds
r2 φ̇ = c (6.89)
where c is a constant of integration. If the geodesic α(s) = α(r(s), φ(s)) represents the path of a free particle
constrained to move on the surface, this conserved quantity is essentially the angular momentum. A neat
result can be obtained by considering the angle σ that the tangent vector V = α0 makes with a meridian.
Recall that the length of V along the geodesic is constant, so let’s set kV k = k. From the chain rule we have
dr dφ
α0 (t) = xr + xφ .
ds ds
Then
< α0 , xφ > G dφ
cos σ = = √ds ,
kα0 k · kxφ k k G
1 √ dφ 1
= G = rφ̇.
k ds k
We conclude from 6.89, that for a surface of revolution, the geodesics make an angle σ with meridians that
satisfies the equation
r cos σ = constant. (6.90)
This result is called Clairaut’s relation. Writing equation 6.89 in terms of differentials, and reusing the
metric as we did in the computation of the geodesics for a sphere, we get
r2 dφ = c ds,
r4 dφ2 = c2 ds2 ,
= c2 [(1 + f 02 ) dr2 + r2 dφ2 ],
(r4 − c2 r2 ) dφ2 = c2 [(1 + f 02 ) dr2 ,
p p
r r2 − c2 dφ = c 1 + f 02 dr,
136 CHAPTER 6. RIEMANNIAN GEOMETRY
so
Z p
1 + f 02
φ = ±c √ dr. (6.91)
r r2 − c2
If c = 0, then the first equation above gives φ =constant, so the meridians are geodesics. The parallels
r =constant are geodesics when f 0 (r) = ∞ in which case the tangent bundle restricted to the parallel is a
cylinder with a vertical generator.
In the particular case of a cone of revolution with a generator that makes an angle α with the z-axis,
f (r) = cot(α)r, equation 6.91 becomes:
Z √
1 + cot2 α
φ = ±c √ dr
r r2 − c2
As shown in figure 6.4, a ribbon laid flatly around a cone follows the path of a geodesic. None of the
parallels, which in this case are the generators of the cone, are geodesics.
6.7 Geodesics in GR
We have dθ0 = dθ1 = 0. To find the connection forms we compute dθ2 and dθ3 , and rewrite in terms of the
6.7. GEODESICS IN GR 137
coframe. We get
l l
dθ2 = p dl ∧ dθ = − p dθ ∧ dl,
b2o + l2 b2o + l2
l
=− 2 θ2 ∧ θ1 ,
bo + l2
l p
dθ3 = p sin θ dl ∧ dφ + cos θ b2o + l2 dθ ∧ dφ,
b2o + l2
l cot θ
=− 2 θ3 ∧ θ1 − p θ3 ∧ θ2 .
bo + l2 b2o + l2
Comparing with the first equation of structure, we start with simplest guess for the connection forms ω’s.
That is, we set
l
ω2 1 = θ2 ,
b2o
+ l2
l
ω3 1 = 2 θ3 ,
bo + l 2
cot θ
ω3 2 =p θ3 .
b2o + l2
Using the antisymmetry of the ω’s and the diagonal metric, we have ω 2 1 = −ω 1 2 , ω 1 3 = −ω 3 1 , and
ω 2 3 = −ω 3 2 . This choice of connection coefficients turns out to be completely compatible with the entire
set of Cartan’s first equation of structure, so, these are the connection forms, all other ω’s are zero. We can
then proceed to evaluate the curvature forms. A straightforward calculus computation which results in some
pleasing cancellations, yields
b2o
Ω1 2 = dω 1 2 + ω 2 1 ∧ ω 2 1 = − θ1 ∧ θ2 ,
(b2o
+ l2 )2
b2
Ω1 3 = dω 1 3 + ω 1 2 ∧ ω 2 3 = − 2 o 2 2 θ1 ∧ θ3 ,
(bo + l )
b2
Ω2 3 = dω 2 3 + ω 2 1 ∧ ω 1 3 = 2 o 2 2 θ2 ∧ θ3 .
(bo + l )
Thus, from equation 6.36, other than permutations of the indices, the only independent components of the
Riemann tensor are
b2
R2323 = −R1212 = R1313 = 2 o 2 2 ,
(bo + l )
and the only non-zero component of the Ricci tensor is
b2o
R11 = −2 .
(b2o + l2 )2
Of course, this space is a 4-dimensional continuum, but since the space is spherically symmetric, we may get
a good sense of the geometry by taking a slice with θ = π/2 at a fixed value of time. The resulting metric
ds2 for the surface is
ds2 2 = dl2 + (b2o + l2 ) dφ2 . (6.94)
Let r2 = b2o + l2 . Then dl2 = (r2 /l2 ) dr2 and the metric becomes
r2
ds2 2 = dr2 + r2 dφ2 , (6.95)
r2 − b2o
1
= b2o
dr2 + r2 dφ2 . (6.96)
1 − r2
138 CHAPTER 6. RIEMANNIAN GEOMETRY
Comparing to 4.26 we recognize this to be a catenoid of revolution, so the equations of geodesics are given
by 6.91 with f (r) = b0 cosh−1 (r/b0 ). Substituting this value of f into the geodesic equation, we get
Z
1
φ = ±c p √ dr. (6.97)
r − bo r2 − c2
2 2
There are three cases. If c = bo , the integral gives immediately φ = ±(c/b0 ) tanh−1 (r/b0 ). We consider the
where F (s, k) is the well-known incomplete elliptic integral of the first kind.
Elliptic integrals are standard functions implemented in computer algebra systems, so it is easy to render
some geodesics as shown in figure 6.5. The plot of the elliptic integral shown here is for k = 0.9. The
plot shows clearly that this is a 1-1, so if one wishes to express r in terms of φ one just finds the inverse
of the elliptic integral which yields a Jacobi elliptic function. Thomas Muller has created a neat Wolfram-
Demonstration that allows the user to play with MT wormhole geodesics with parameters controlled by
sliders.
1
ds2 = −h(r) dt2 + dr2 + r2 (dθ2 + sin θ dφ2 ), (6.100)
h(r)
where
2GM
h(r) = 1 − . (6.101)
r
Thus, the Lagrangian is
1 2
L = −h ṫ2 + ṙ + r2 θ̇2 + r2 sin θφ̇2 . (6.102)
h
6.7. GEODESICS IN GR 139
If in the equation for g22 , one chooses initial conditions θ(0) = π/2, θ̇(0) = 0, we get θ(s) = π/2 along the
geodesic. We infer from rotation invariance that the motion takes place on a plane. Hereafter, we assume
we have taken these initial conditions. From the other two equations we obtain
dt
h = E,
ds
dφ
r2 = L.
ds
for some constants E and L. We recognize the conserved quantities as the “energy” and the angular
momentum. Along the geodesic of a massive particle, with unit time-like tangent vector, we have
dxµ dxν
− 1 = gµν (6.103)
ds ds
The equations of motion then reduce to
2
2 2
dt 1 dr 2 dφ
−1 = −h + +r ,
ds h ds ds
2
E2 1 dr L2
−1 = − + + 2,
h h ds r
2 2
dr L
E2 = +h 1+ 2 .
ds r
Hence, we obtain the neat equation,
2
dr
2
E = + V (r), (6.104)
ds
where V (r) represents the effective potential.
L2
2GM
V (r) = 1 − 1+ 2 ,
r r
2
2GM L 2M GL2
=1− + 2 − . (6.105)
r r r3
If we let V̂ = V /2 in this expression we recognize the classical 1/r potential, and the 1/r2 term corresponding
to the Coriolis contribution associated with the angular momentum. The 1/r3 term is a new term arising
from general relativity. Clearly we must have E 2 < V (r). There are multiple cases depending on the values
of E and L and the nature of the equilibrium points. Here we are primarily concerned with bounded orbits,
so we seek conditions for the particle to be in a potential well. This presents us with a nice calculus problem.
We compute V 0 (r) and set equal to zero to find the critical points
2
V 0 (r) = (GM r2 − L2 r + 3GM L2 ) = 0.
r4
The discriminant of the quadratic is
D = L2 − 12G2 M 2 .
140 CHAPTER 6. RIEMANNIAN GEOMETRY
If D < 0 there are no critical points. In this case, V (r) is a monotonically increasing function on the interval
(2M G, ∞), as shown in the bottom left graph in figure 6.6. The maple plots in this figure are in units with
GM = 1. In the case D < 0, all trajectories either fall toward the event horizon or escape to infinity.
If D > 0, there are two critical points
√
L2 − L L2 − 12G2 M 2
r1 = ,
√ 2GM
L2 + L L2 − 12G2 M 2
r2 = .
2GM
The critical point r1 is a local maximum associated with an unstable circular orbit. The critical point r2 > r1
gives a stable circular orbit. Using the standard calculus trick of multiplying by the conjugate of the radical
in the first term, we see that
r1 → 3GM,
L2
r2 → ,
GM
as L → ∞. For any L, the properties of the roots of the quadratic imply that r1 r2 = 3L2 . As shown in the
graph 6.6, as L gets larger, the inner radius approaches 3GM and the height of the bump increases, whereas
the outer radius recedes to infinity. As the value of D approaches 0, the two orbits coalesce at L2 = 12G2 M 2 ,
which corresponds to r = 6GM , so this is the smallest value of r at which a stable circular orbit can exist.
Since V (r) → 1 as r → ∞, to get bounded orbits we want a potential well with V (r1 ) < 1. We can easily
verify that when L = 4GM the local maximum occurs at r1 = 4GM , which results in a value of V (r1 ) = 1.
This case is the one depicted in the middle graph in figure 6.6, with the graph of V 0 (r) on the right showing
the two critical points at r1 = 4GM, r2 = 12GM . Hence the condition to get a bounded orbit is
√
2 3 GM < L < 4GM,
E 2 < V (r1 ), r > r1 ,
so that the energy results in the particle trapped in the potential well to the right of r1 . This is the case
that applies to the modification of the Kepler orbits of planets. If we rewrite
dr dr dφ L dr
= = 2
ds dφ ds r dφ
and substitute into equation 6.104, we get
2
L2 dr L2
2 2GM
=E − 1+ 2 1− .
r4 dφ r r
If now we change variables to u = 1/r, we obtain
du 1 dr dr
=− 2 = −u2 ,
dφ r dφ dφ
6.7. GEODESICS IN GR 141
The solution of the orbit equation is therefore reduced to an elliptic integral. If we expand the denominator
Z
Ldu
φ= p + φ0 ,
(E − 1) + 2GM u − L2 u2 + 2GM L2 u3
2
and neglect the cubic term, we can complete the squares of the remaining quadratic. The integral becomes
one of standard inverse cosine type; hence, the solution gives the equation of an ellipse in polar coordinates
1
u= = C(1 + e cos(φ − φ0 )),
r
for appropriate constants C, shift φ0 and eccentricity e. The solution is automatically expressed in terms of
the energy and the angular momentum of the system. More careful analysis of the integral shows that the
inclusion of the cubic term perturbs the orbit by a precession of the ellipse. While this approach is slicker,
we prefer to use the more elementary procedure of differential equations. Differentiating with respect to φ
the equation
2
du
L2 = (E 2 − 1) + 2GM u − L2 u2 + 2GM L2 u3 ,
dφ
and cancelling out the common chain rule factor du/dφ, we get
d2 u GM
2
= 2 − u + 3GM u2
dφ L
d2 u GM L2 2
+ u = + u . (6.106)
dφ2 L2 GM
The linear part of the equation corresponds precisely to Newtonian motion, and is small, so we can treat
the quadratic term as a perturbation
u = u0 + u1 + u2 2 + . . . .
Substituting u into equation 6.106, the first approximation is the linear approximation given by
GM
u000 + u = .
L2
The homogenous solution is of the form u = A cos(φ − φ0 ), where A and φ0 are the arbitrary constants, and
the particular solution is a constant. So the general solution is
GM
u0 = + A cos(φ − φ0 ),
L2
GM AL2
= 2 [1 + e cos(φ − φ0 )], e= .
L GM
142 CHAPTER 6. RIEMANNIAN GEOMETRY
Without loss of generality, we can align the axes and set φ0 = 0 . In the Newtonian orbit, we would write
u0 = 1/r, thus getting the equation of a polar conic.
GM
u0 = (1 + e cos φ) (6.107)
L2
In the case of the planets, the eccentricity e < 1, so the conics are ellipses. Having found u0 we reinsert u
into the differential equation 6.106 and keeping only the terms of order . We get
GM L2
(u0 + u1 )00 + (u0 + u1 ) = + (u0 + u1 )2 ,
L2 GM
GM L2 2
(u000 + u0 − 2 ) + (u001 + u1 ) = u .
L GM 0
Thus, the result is a new differential equation for u1 ,
L2 2
u001 + u01 = u ,
GM 0
L2
= [(1 + 12 e2 ) + 2e cos φ + 12 e2 cos 2φ].
GM
The equation is again a linear inhomogeneous equation with constant coefficients, so it is easily solved by
elementary methods. We do have to be a bit careful since we have a resonant term on the right hand side.
The solution is
L2
u1 = [(1 + 21 e2 ) + 2eφ cos φ − 16 e2 cos 2φ].
GM
The resonant term φ cos φ makes the solution non-periodic, so this is the term responsible for the precession
of the elliptical orbits. The precession is obtained by looking at the perihelion, that is, the point in the
elliptical orbit at which the planet is closest to the sun. This happens when
du d
≈ (u0 + u1 ) = 0,
dφ dφ
− sin φ + (sin φ + eφ cos φ + 31 e sin φ) = 0.
Starting with the solution φ = 0, after on revolution, the perihelion drifts to φ = 2π + δ. By the perturbation
assumptions, we assume δ is small, so to lowest order, the perihelion advance in one revolution is
6πG2 M 2
δ = 2π = . (6.108)
L2
From equation 6.107 for the Newtonian elliptical orbit, the mean distance a to the sun is given by the average
of the aphelion and perihelion distances, that is
1 L2 /GM L2 /GM L2
1
a= + = .
2 1+e 1−e GM 1 − e2
Thus, if we divide by the period T , the rate of perihelion advance can be written in more geometric terms as
6πGM
δ= .
a(1 − e2 )T
The famous computation by Einstein of a precession of 43.1” of an arc per century for the perihelion advance
of the orbit of Mercury, still stands as one of the major achievements in modern physics.
dxµ dxν
0 = gµν ,
ds ds
6.7. GEODESICS IN GR 143
d2 u
+ u = 3GM u2 .
dφ2
Consider the problem of light rays from a distant star grazing the sun as they approach the earth. Since the
space is asymptotically flat, we expect the geodesics to be asymptotically straight. The quantity 3GM is of
the order of 2km, so it is very small compared to the radius of the sun, so again we can use perturbation
methods. We let = 3GM and consider solutions of equation
u00 + u = u2 ,
of the form
u = u0 + u1 .
u0 = A cos φ + B sin φ,
1 = Ar cos φ + Br sin φ,
1 = Ax + By
Without loss of generality, we can align the vertical axis parallel to the incoming light with impact parameter
b (distance of closest approach)
1
u0 = cos φ.
b
As above, we reinsert the u into the differential equation and compare the coefficients of terms of order .
We get an equation for u1 ,
1 1
u001 + u1 = 2 cos2 φ = 2 (1 + cos 2φ).
b 2b
We solve the differential equation by the method of undetermined coefficients and thus we arrive at the
perturbation solution to order
1 2
u = cos φ + 2 − 2 cos2 φ.
b 3b 3b
To find the the asymptotic angle of the outgoing photons, we let r → ∞ or u → 0. Thus we get a quadratic
equation for cos φ.
2 2GM
cos φ = − = −
3b b
Set φ = π2 + δ. Since δ is small, we have sin δ ≈ δ, and we see that δ = 2GM/b is the approximation of the
deflection angle of one of the asymptotes. The total deflection is twice that angle
4GM
2δ = .
b
The computation results in a deflection by the sun of light rays from a distant star of about 1.75”. This
was corroborated in an experiment lead by Eddington during the total solar eclipse of 1919. The part of the
expedition in Brazil was featured in the 2005 movie, The House of Sand. For more details and more careful
analysis of the geodesics, see for example, Misner Thorne and Wheeler [?].
144 CHAPTER 6. RIEMANNIAN GEOMETRY
This section is dedicated to the memory of Professor S.-S. Chern. I prelude the section with a short
anecdote that I often narrate to my students. In June 1979, an international symposium on differential
geometry was held at the Berkeley campus in honor of the retirement of Professor Chern. The invited
speakers included an impressive list of the most famous differential geometers at the time, At the end of
the symposium, Chern walked on the stage of the packed auditorium to give thanks and to answer some
questions. After a few short remarks, a member of the audience asked Chern what he thought was the
most important theorem in differential geometry. Without any hesitation he answered, “there is only one
theorem in differential geometry, and that is Stokes’ theorem.” This was followed immediately by a question
about the most important theorem in analysis. Chern gave the same answer: “there is only one theorem
in analysis, Stokes’ theorem. A third person then asked Chern what was the most important theorem in
Complex Variables. To the amusement of the crowd, Chern responded, “There is only one theorem in
complex variables, and that that is Cauchy’s theorem. But if one assumes the derivative of the function is
continuous, then this is just Stokes’ theorem.” Now, of course it is well known that Goursat proved that the
hypothesis of continuity of the derivative is automatically satisfied when the function is holomorphic. But
the genius of Chern was always his uncanny ability to extract the essential of what makes things work, in
the simplest terms.
The Gauss-Bonnet theorem is rooted on the theorem of Gauss (4.72), which combined with Stokes’
theorem, provides a beautiful geometrical interpretation of the equation. This is undoubtedly part of what
Chern had in mind at the symposium, and also when wrote in his Euclidean Differential Geometry Notes
(Berkeley 1975 ) [?] that the theorem has “profound consequences and is perhaps one of the most important
theorems in mathematics.”
Let β(s) by a unit speed curve on an orientable surface M , and let T be the unit tangent vector. There
is Frenet frame formalism for M , but if we think of the surface intrinsically as 2-dimensional manifold,
then there is no binormal. However, we can define a “geodesic normal” taking G = J(T ), where J is the
symplectic form ??, Then the geodesic curvature is given by the Frenet formula
T 0 = κg G. (6.109)
6.8.1 Proposition Let {e1 , e2 } be an orthonormal on M , and let β(s) be a unit speed curve as above,
with unit tangent T . If φ is the angle that T makes with e1 , then
∂φ
κg = − ω 12 (T ). (6.110)
∂s
Proof Since {T, G} and {e1 , e2 } are both orthonormal basis of the tangent space, they must be related by
a rotation by an angle φ, that is
T cos φ sin φ e1
= , (6.111)
G − sin φ cos φ e2
that is,
Since T = β 0 , we have
∂φ ∂φ
β 00 = ∇T T = −(sin φ) e1 + cos φ∇T e1 + (cos φ) e2 + sin φ∇T e2 ,
∂s ∂s
∂φ ∂φ
= −(sin φ) e1 + (cos φ)ω 21 (T )e2 + (cos φ) e2 + (sin φ)ω 12 (T )e1 ,
∂s ∂s
∂φ ∂φ
=[ − ω 12 (T )][−(sin φ)e1 ] + [ − ω 12 (T )][(cos φ)e2 ]
∂s ∂s
∂φ
=[ − ω 12 (T )][−(sin φ)e1 + (cos φ)e2 ]
∂s
∂φ
=[ − ω 12 (T )]G,
∂s
= κg G.
6.8.2 Definition Let H C be a smooth closed curve on M parametrized by arc length with geodesic curvature
κg . The line integral C κg ds is called the total geodesic curvature. If the curve is piecewise smooth, the
total geodesic curvature is the sum of the integrals of each piece.
A circle of radius R gives an elementary example. The geodesic curvature is the constant 1/R, so the
total geodesic curvature is (1/R)2πR = 2π.
If we integrate formula 6.110 around a smooth simple closed curve C which is the boundary of a region
R and use Stokes’ Theorem, we get
I I I
κg ds = dφ − ω 12 ds,
C C C
I Z Z
= dφ − dω 12 .
C R
R
For a smooth simple closed curve, C
dφ = 2π. Using the Cartan-form version of the theorema egregium
4.106 we get immediately Z Z Z
K dS + κg ds = 2π. (6.114)
R C
If the boundary of the region consists of k piecewise continuous functions as illustrated in figure 6.7, the
change of the angle φ along C is still 2π, but the total change needs to be modified by adding the exterior
angles αk . Thus, we obtain a fundamental result called the Gauss-Bonnet formula,
6.8.3 Theorem Z Z Z X
K dS + κg ds + αk = 2π. (6.115)
R C k
Every interior ιk angle is the supplement of the corresponding exterior αk angle, so the Gauss Bonnet formula
can also be written as Z Z Z X
K dS + κg ds + (π − ιk ) = 2π. (6.116)
R C k
The simplest manifestation of the Gauss-Bonnet formula is for a triangle in the plane. Planes are flat
surfaces, so K = 0 and the straight edges are geodesics, so κg = 0 on each of the three edges. The interior
146 CHAPTER 6. RIEMANNIAN GEOMETRY
angle version of the formula then just reads 3π − ι1 − ι2 − ι3 = 2π, which just says that the interior angles
of a flat triangle add up to π. Since a sphere has constant positive curvature, the sum of the interior angles
of a spherical triangle is larger than π. That amount of this sum over 2π is called the spherical excess. For
example, the sum of the interior angles of a spherical triangle that is the boundary of one octant of a sphere
is 3π/2, so the spherical excess is π/2.
RR
6.8.4 Definition The quantity K dS is called the total curvature
6.8.5 Example A sphere of radius R has constant Gaussian Curvature 1/R2 . The surface area of the
sphere is 4πR2 , so the total Gaussian curvature for the sphere is 4π.
6.8.6 Example For a torus generated by a circle of radius a rotating about an axis with radius b as
in example (4.40), the differential of surface is dS = a(b + a cos θ) dθdφ, and the Gaussian curvature is
K = cos θ/[a(b + a cos θ)], so the total Gaussian curvature is
Z 2π Z 2π
cos θ dθdφ = 0.
0 0
6.8.9 Theorem Given a triangulation of a compact surface M , let V be the number of vertices, E the
number of edges and F the number of faces. Then the quantity
χ(M ) = V − E + F, (6.117)
is independent of the triangulation. In fact the quantity is independent of any “polyhedral” subdivision.
This quantity is a topological invariant called the Euler characteristic.
6.8.10 Example
1. A balloon-inflated tetrahedron has V = 4, E = 6, F = 4, so the Euler characteristic of a sphere is 2.
2. A balloon-inflated octahedron has V = 6, E = 12, F = 8, so we get the same number 2.
3. The diagram on the right of gigure 6.8 represents a topological torus. In the given rectangle, opposites
sides are identified in the same direction. The number of edges without double counting are shown in
red, and the number of vertices not double counted are shown in black dots. We have V = 6, E =
18 F = 12. So the Euler characteristic of a torus is 0.
4. In one has a compact surface, one can add a “handle”, that is, a torus, by the following procedure. We
excise a triangle in each of the two surfaces and glue the edges. We lose two faces and the number of
edges and vertices cancel out, so the Euler characteristic of the new surface decreases by 2. The Euler
characteristic of a pretzel is -4
5. The Euler characteristic of an orientable surface of genus g, that is, a surface with g holes is given by
χ(M ) = 2 − 2g.
where F is the number of triangles and the ιk ’s are the interior angles of triangle 4k . The line integrals of
the geodesic curvatures all cancel out since each edge in every triangle is traversed twice, each in opposite
directions. Rewriting the equation, we get
Z Z
K dS = −πF + S
M
where S is the sum of all interior angles. Since the manifold is locally Euclidean, the sum of all interior
angles at a vertex is 2π, so we have Z Z
K dS = −πF + 2πV
M
There are F faces. Each face has three edges, but each edge is counted twice, so 3F = 2E, and we have
F = 2E − 2F Substituting in the equation above, we get,
Z Z
K dS = −π(2E − 2F ) + 2πV = 2π(V − E + F ) = χ(M ).
M
This is a remarkable theorem because it relates the bending invariant Gaussian curvature to a topological
invariant. Theorems such as this one which cut across disciplines, are the most significant in mathematics.
Not surprisingly, it was Chern who proved a generalization of the Gauss-Bonnet theorem to general orientable
Riemannian manifolds of even dimensions [?].
148 CHAPTER 6. RIEMANNIAN GEOMETRY
Bibliography
[1] H. Abrams 1971 The World of M. C. Escher, Meulenhoff International, Netherlands (1971)
[2] Atiyah, M., Hitchin, N., and Singer, I. 1978 Self-duality in four-dimensional Riemannian geometry,
Proc. R. Soc. Lond. A. 362, (1978) pp 425-461.
[3] Belavin, A., Polyakov, A., Schwartz, A., and Tyupkin, Y. 1975 Pseudoparticle Solutions of the Yang-
Mills Equations, Physics Letters B, Vol. 59, No, 1, (1975) pp 85-87.
[4] Chern, S. S. 1975: Introduction to Differential Geometry. Notes, University of California, Berkeley, CA.
[5] Chern, S. S. 1944 A Simple Intrinsic Proof of the Gauss-Bonnet Formula for Closed Riemannian Man-
ifolds, Ann. Math. 45 (4), 1944, 747-752.
[6] Chern, S. S. 1955 An Elementary Proof of the Existence of Isoterhmal Parameters on a Surface, Pro-
ceedings of AMS, Vol. 6, No, 4, (1955), 771-782.
[7] Dray, T. 1984 1966 The Relationship between Monopole Harmonics and Spin-weighted Spherical Har-
monics, Journal of Mathematical Physics, 26, 1030 (1985). https://doi.org/10.1063/1.526533
[8] Eguchi, T., Gilkey, P., and Hanson, A. 1980 Gravitation, Gauge Theories and Differential Geometry,
Physics Reports, 66, No, 6 (1980). pp 213-293
[9] Georgi, H. 1999: Lie Algebras in Particle Physics. Frontiers in Physics V54, Westview (1999).
[10] Goldberg, J., Mcfarland, A., Newman, E., Rohrlich, F. and Sudarshan, C. 1966 Spin-s Spherical Har-
monics and ð, Proceedings of AMS, Vol. 6, No, 4, (1955), 771-782.
[12] Grey, Alfred, 2006: Modern Differential Geometry of Curves and Surfaces with Mathematica. 3rd ed.
Chapman & Hall, CRC, 977pp.
[13] Hall, B. 2015: Lie Groups, Lie Algebras, and Representations. 2nd ed. Springer, Graduate Texts in
Mathematics 222, Switzerland, (2015).
[14] Hicks, Noel, 1965: Notes on Differential Geometry. Van Nostrand Reinhold, Princeton, NJ.
[15] Hirsh, A. 2002 Extension of ’Villarceau-Section’ to Surfaces of Revolution with a Generating Conic , J.
of Geometry and Graphics. Vol 6, (2002) No, 2, pp121-132.
[16] Hoffman, K. and Kunze, R., 1971: Linear Algebra. 2nd ed. Prentice-Hall, 407pp.
[18] Kobayashi, S and Nomizu, K 1963: Foundations of Differential Geometry. Wiley & Sons, New York,
London, 1963.
[19] Eisenhart, L. P. 1960: A Treatise on the Differential Geometry of Curves and Surfaces. Ginn and
Company, Boston, 1909. Reprint by Dover Publications, New York, 1960.
149
150 BIBLIOGRAPHY
[20] Abraham, R. and Marsden, J. 1978: Foundations of Mechanics. 2nd ed. Addison-Wesley, 838pp.
[21] Misner, C. W., Thorne, K. S., and Wheeler, J. A. 1973: Gravitation. W.H. Freeman and Company,
1279pp.
[22] Morris, M.S. and Thorne, K.S., 1987: Wormholes in Spacetime and their use for Interstellar Travel: A
Tool for Teaching General Relativity. Am. J. Phys., 56(5), pp395-412.
[23] Müller, Thomas, 2008: Exact geometric optics in a Morris-Thorne wormhole spacetime. Phys. Rev. D,
77, 044043-1-044043-11.
[24] Newman, E. and Penrose, R. 1962. An Approach to Gravitational Radiation by a Method of Spin Coef-
ficients, Journal of Mathematical Physics 3, 566 (1962), pp 556-578. https://doi.org/10.1063/1.1724257
[25] Newman, E. and Janis, R. 1962. Note on the Kerr Spinning-Particle Metric, Journal of Mathematical
Physics 6, 915 (1965), pp 915-917. https://doi.org/10.1063/1.1704350
[26] O’Neill, Barrett, 2006: Elementary Differential Geometry. 2nd ed. Academic Press, 503pp.
[27] Oprea, J., 1997: Differential Geometry and Its Applications. Prentice-Hall, Englewood Cliffs, NJ.
[28] Oprea, J., 2000: The Mathematics of Soap Films. Student Mathematical, Library, Volume 10, AMS,
Providence, RI.
[29] Penrose, R. 1987. On the Origins of Twistor Theory. Gravitation and Geometry, a volume on honour
of I. Robinson, Biblipolis, Naples, 25pp.
[30] Penrose, R. 1976. Nonlinear Gravitons and Curved Twistor Theory, General Relativity and Gravitation
7 (1976), No. 1 pp 31-52. and Geometry, a volume on honour of I. Robinson, Biblipolis, Naples, 25pp.
[31] Rogers, C and Schief, W. K., 2002. On Regular Solutions of Euclidean Yang-Mills Equations. Physic
Cambridge University Press.
[32] Schwarz, A, 1977. Bäcklund and Darboux Transformations : Geometry and Modern Applications in
Soliton Theory. Physics Letters B, Vol. 67, No, 2, (1977) pp 172-174.
[33] Spivak, M. 1965: Calculus on Manifolds. Addison Wesley (1965), 159pp
[34] Spivak, M. 1979: A comprehensive introduction to Differential Geometry 5 volume set. 2nd ed. Publish
or Perish, Texas.
[35] Struik, Dirk J., 1961: Lectures on Classical Differential Geometry. 2nd ed. Dover Publications, Inc.,
232pp.
[36] Taub, A. H. 1939: Tensor Equations Equivalent to the Dirac Equations. Annals of Mathematics, Vol
40, Number 4, pp.937-947.
[37] Trautman, A. 1977: Solutions of the Maxwell and Yang-Mills Equations Associated with Hopf Fibrings.
International Journal of Theoretical Physics, Vol 16, Number 8 (1977), pp.561-565.
[38] Terng,C and Uhlenbeck, K, 2000: Geometry of Solitons. Notices of the AMS, Vol 47, Number 1.
[39] Ward, R. S., 1977: On Self-dual Gauge Fields. Physics Letters A, Vol. 61, No, 2, (1977) pp 81-82.
[40] Weisstein, E. W. From Mathworld- A Wolfram Web Resource: [Available online at http://mathworld.
wolfram.com/CostaMinimalSurface.html.]
Index
Acceleration Contraction, 31
along a curve, 9 Coordinate
Centripetal, 15 Cylindrical, 32
Functions, 1
Bianchi identities, 118, 119 Geodesic, 105
Bundle Isothermal, 106
Cotangent, 27 Local, 6
Dual basis, 27 Minkowski, 33
Section, 2 Polar, 6, 24, 46, 47
Tangent, 2, 79 Slot functions, 8
Tensor, 30 Spherical, 33, 47
Transformation, 6
Cartan equations Coordinate patch, see Patch
Connection form, 71, 115 Cornu spiral, 23
Curvature form, 72 Covariant derivative
First structure equations, 70 Divergence, 98
for surface in R3 , 102 Tensor fields, 68
Manifolds, 117 Vector fields, 66, 116
Second structure equation, 71 Covariant differential, 123
Catenoid of surface normal, 104
First fundamental form, 83 of tensor-valued 0-form, 124
Helicoid curvature, 106 of vector field, 124
Christoffel symbols, see Connection Curvature
Circle Form, see Cartan equations
Curvature, 14 Gaussian, see Gaussian curvature
Frenet frame, 14 Geodesic, 85
Clairut relation, 135 Normal, 85–87, 90
Cone of a curve, 12
First fundamental form, 84 Curves, 8–26
Geodesics, 136 Fundamental theorem, 17–22
Conformal map in R3 , 8–17
Mercator, 81 Natural equations, 22
Conical helix, 25, 84 Plane, 16
Connection Curvilinear Coordinates, 63
Affine, 123
Change of basis, 73, 126, 127 de Sitter space, 121
Christoffel symbols, 67, 97, 99, 115, 127 Determinants
Compatible with metric, 67, 68 By pull-back, 51
Curvature form, 72 Definition, 37
Frenet Equations, 70 Levi-Civita symbol, 37
Koszul, 66 Diffeomorphism, 6
Levi-Civita, 97, 115 Differentiable map, 6–7
Linear, 122 Jacobian, 6
Parallel transport, 96 Push-forward, 6
151
152 INDEX
Tangent bundle, 2
Tangent vector, 1–5
Contravariant components, 4
in Rn , 2
Tensorial form
of adjoint type, 74
Tensors
Antisymmetric, 32, 36
Bilinear map, 29
Bundle, 30
Components, 29
Contravariant, 29
Metric, 30
Riemann, see Riemannian, Riemann tensor
Self-dual, 52
Tensor product, 29
Torsion, see Riemannian
Theorema egregium, see Gaussian curvature