INTRODUCTION TO VECTORS AND TENSORS

Vector and Tensor Analysis

Volume 2

Ray M. Bowen
Mechanical Engineering
Texas A&M University
College Station, Texas

and

C.-C. Wang
Mathematical Sciences
Rice University
Houston, Texas

Copyright Ray M. Bowen and C.-C. Wang

(ISBN 0-306-37509-5 (v. 2))
____________________________________________________________________________
PREFACE

To Volume 2

This is the second volume of a two-volume work on vectors and tensors. Volume 1 is concerned
with the algebra of vectors and tensors, while this volume is concerned with the geometrical
aspects of vectors and tensors. This volume begins with a discussion of Euclidean manifolds. The
principal mathematical entity considered in this volume is a field, which is defined on a domain in a
Euclidean manifold. The values of the field may be vectors or tensors. We investigate results due
to the distribution of the vector or tensor values of the field on its domain. While we do not discuss
general differentiable manifolds, we do include a chapter on vector and tensor fields defined on
hypersurfaces in a Euclidean manifold.

This volume contains frequent references to Volume 1. However, references are limited to
basic algebraic concepts, and a student with a modest background in linear algebra should be able
to utilize this volume as an independent textbook. As indicated in the preface to Volume 1, this
volume is suitable for a one-semester course on vector and tensor analysis. On occasions when we
have taught a one –semester course, we covered material from Chapters 9, 10, and 11 of this
volume. This course also covered the material in Chapters 0,3,4,5, and 8 from Volume 1.

We wish to thank the U.S. National Science Foundation for its support during the
preparation of this work. We also wish to take this opportunity to thank Dr. Kurt Reinicke for
critically checking the entire manuscript and offering improvements on many points.

Houston, Texas R.M.B.

C.-C.W.

iii
__________________________________________________________________________
CONTENTS

Vol. 2 Vector and Tensor Analysis

Contents of Volume 1………………………………………………… vii

PART III. VECTOR AND TENSOR ANALYSIS

Selected Readings for Part III………………………………………… 296

CHAPTER 9. Euclidean Manifolds………………………………………….. 297

Section 43. Euclidean Point Spaces……………………………….. 297

Section 44. Coordinate Systems…………………………………… 306
Section 45. Transformation Rules for Vector and Tensor Fields…. 324
Section 46. Anholonomic and Physical Components of Tensors…. 332
Section 47. Christoffel Symbols and Covariant Differentiation…... 339
Section 48. Covariant Derivatives along Curves………………….. 353

CHAPTER 10. Vector Fields and Differential Forms………………………... 359

Section 49. Lie Derivatives……………………………………….. 359

Section 5O. Frobenius Theorem…………………………………… 368
Section 51. Differential Forms and Exterior Derivative………….. 373
Section 52. The Dual Form of Frobenius Theorem: the Poincaré
Lemma……………………………………………….. 381
Section 53. Vector Fields in a Three-Dimensiona1 Euclidean
Manifold, I. Invariants and Intrinsic Equations…….. 389
Section 54. Vector Fields in a Three-Dimensiona1 Euclidean
Manifold, II. Representations for Special
Class of Vector Fields………………………………. 399

CHAPTER 11. Hypersurfaces in a Euclidean Manifold

Section 55. Normal Vector, Tangent Plane, and Surface Metric… 407
Section 56. Surface Covariant Derivatives………………………. 416
Section 57. Surface Geodesics and the Exponential Map……….. 425
Section 58. Surface Curvature, I. The Formulas of Weingarten
and Gauss…………………………………………… 433
Section 59. Surface Curvature, II. The Riemann-Christoffel
Tensor and the Ricci Identities……………………... 443
Section 60. Surface Curvature, III. The Equations of Gauss and Codazzi 449

v
vi CONTENTS OF VOLUME 2

Section 61. Surface Area, Minimal Surface………........................ 454

Section 62. Surfaces in a Three-Dimensional Euclidean Manifold. 457

CHAPTER 12. Elements of Classical Continuous Groups

Section 63. The General Linear Group and Its Subgroups……….. 463
Section 64. The Parallelism of Cartan……………………………. 469
Section 65. One-Parameter Groups and the Exponential Map…… 476
Section 66. Subgroups and Subalgebras…………………………. 482
Section 67. Maximal Abelian Subgroups and Subalgebras……… 486

CHAPTER 13. Integration of Fields on Euclidean Manifolds, Hypersurfaces, and

Continuous Groups

Section 68. Arc Length, Surface Area, and Volume……………... 491

Section 69. Integration of Vector Fields and Tensor Fields……… 499
Section 70. Integration of Differential Forms……………………. 503
Section 71. Generalized Stokes’ Theorem……………………….. 507
Section 72. Invariant Integrals on Continuous Groups…………... 515

INDEX………………………………………………………………………. x
______________________________________________________________________________

CONTENTS

Vol. 1 Linear and Multilinear Algebra

PART 1 BASIC MATHEMATICS

Selected Readings for Part I………………………………………………………… 2

CHAPTER 0 Elementary Matrix Theory…………………………………………. 3

CHAPTER 1 Sets, Relations, and Functions……………………………………… 13

Section 1. Sets and Set Algebra………………………………………... 13

Section 2. Ordered Pairs" Cartesian Products" and Relations…………. 16
Section 3. Functions……………………………………………………. 18

CHAPTER 2 Groups, Rings and Fields…………………………………………… 23

Section 4. The Axioms for a Group……………………………………. 23

Section 5. Properties of a Group……………………………………….. 26
Section 6. Group Homomorphisms…………………………………….. 29
Section 7. Rings and Fields…………………………………………….. 33

PART I1 VECTOR AND TENSOR ALGEBRA

Selected Readings for Part II………………………………………………………… 40

CHAPTER 3 Vector Spaces……………………………………………………….. 41

Section 8. The Axioms for a Vector Space…………………………….. 41

Section 9. Linear Independence, Dimension and Basis…………….….. 46
Section 10. Intersection, Sum and Direct Sum of Subspaces……………. 55
Section 11. Factor Spaces………………………………………………... 59
Section 12. Inner Product Spaces………………………..………………. 62
Section 13. Orthogonal Bases and Orthogonal Compliments…………… 69
Section 14. Reciprocal Basis and Change of Basis……………………… 75

CHAPTER 4. Linear Transformations……………………………………………… 85

Section 15. Definition of a Linear Transformation………………………. 85

Section 16. Sums and Products of Linear Transformations……………… 93
viii CONTENTS OF VOLUME 2

Section 17. Special Types of Linear Transformations…………………… 97

Section 18. The Adjoint of a Linear Transformation…………………….. 105
Section 19. Component Formulas………………………………………... 118

CHAPTER 5. Determinants and Matrices…………………………………………… 125

Section 20. The Generalized Kronecker Deltas

and the Summation Convention……………………………… 125
Section 21. Determinants…………………………………………………. 130
Section 22. The Matrix of a Linear Transformation……………………… 136
Section 23 Solution of Systems of Linear Equations…………………….. 142

CHAPTER 6 Spectral Decompositions……………………………………………... 145

Section 24. Direct Sum of Endomorphisms……………………………… 145

Section 25. Eigenvectors and Eigenvalues……………………………….. 148
Section 26. The Characteristic Polynomial………………………………. 151
Section 27. Spectral Decomposition for Hermitian Endomorphisms…….. 158
Section 28. Illustrative Examples…………………………………………. 171
Section 29. The Minimal Polynomial……………………………..……… 176
Section 30. Spectral Decomposition for Arbitrary Endomorphisms….….. 182

CHAPTER 7. Tensor Algebra………………………………………………………. 203

Section 31. Linear Functions, the Dual Space…………………………… 203

Section 32. The Second Dual Space, Canonical Isomorphisms…………. 213
Section 33. Multilinear Functions, Tensors…………………………..….. 218
Section 34. Contractions…......................................................................... 229
Section 35. Tensors on Inner Product Spaces……………………………. 235

CHAPTER 8. Exterior Algebra……………………………………………………... 247

Section 36. Skew-Symmetric Tensors and Symmetric Tensors………….. 247

Section 37. The Skew-Symmetric Operator……………………………… 250
Section 38. The Wedge Product………………………………………….. 256
Section 39. Product Bases and Strict Components……………………….. 263
Section 40. Determinants and Orientations………………………………. 271
Section 41. Duality……………………………………………………….. 280
Section 42. Transformation to Contravariant Representation……………. 287

INDEX…………………………………………………………………………………….x
_____________________________________________________________________________
PART III

VECTOR AND TENSOR ANALYSIS

Selected Reading for Part III

BISHOP, R. L., and R. J. CRITTENDEN, Geometry of Manifolds, Academic Press, New York, 1964
BISHOP, R. L., and S. I. GOLDBERG, Tensor Analysis on Manifolds, Macmillan, New York, 1968.
CHEVALLEY, C., Theory of Lie Groups, Princeton University Press, Princeton, New Jersey, 1946
COHN, P. M., Lie Groups, Cambridge University Press, Cambridge, 1965.
EISENHART, L. P., Riemannian Geometry, Princeton University Press, Princeton, New Jersey, 1925.
ERICKSEN, J. L., Tensor Fields, an appendix in the Classical Field Theories, Vol. III/1.
Encyclopedia of Physics, Springer-Verlag, Berlin-Gottingen-Heidelberg, 1960.
FLANDERS, H., Differential Forms with Applications in the Physical Sciences, Academic Press,
New York, 1963.
KOBAYASHI, S., and K. NOMIZU, Foundations of Differential Geometry, Vols. I and II, Interscience,
New York, 1963, 1969.
LOOMIS, L. H., and S. STERNBERG, Advanced Calculus, Addison-Wesley, Reading, Massachusetts,
1968.
MCCONNEL, A. J., Applications of Tensor Analysis, Dover Publications, New York, 1957.
NELSON, E., Tensor Analysis, Princeton University Press, Princeton, New Jersey, 1967.
NICKERSON, H. K., D. C. SPENCER, and N. E. STEENROD, Advanced Calculus, D. Van Nostrand,
Princeton, New Jersey, 1958.
SCHOUTEN, J. A., Ricci Calculus, 2nd ed., Springer-Verlag, Berlin, 1954.
STERNBERG, S., Lectures on Differential Geometry, Prentice-Hall, Englewood Cliffs, New Jersey,
1964.
WEATHERBURN, C. E., An Introduction to Riemannian Geometry and the Tensor Calculus,
Cambridge University Press, Cambridge, 1957.
_______________________________________________________________________________
Chapter 9

EUCLIDEAN MANIFOLDS

This chapter is the first where the algebraic concepts developed thus far are combined with
ideas from analysis. The main concept to be introduced is that of a manifold. We will discuss here
only a special case cal1ed a Euclidean manifold. The reader is assumed to be familiar with certain
elementary concepts in analysis, but, for the sake of completeness, many of these shall be inserted
when needed.

Section 43 Euclidean Point Spaces

Consider an inner produce space V and a set E . The set E is a Euclidean point space if
there exists a function f : E × E → V such that:

(a) f ( x, y ) = f (x, z) + f ( z, y ), x, y , z ∈ E
and
(b) For every x ∈E and v ∈V there exists a unique element y ∈E such that
f ( x, y ) = v .

The elements of E are called points, and the inner product space V is called the translation space.
We say that f ( x, y ) is the vector determined by the end point x and the initial point y . Condition
b) above is equivalent to requiring the function f x : E → V defined by f x ( y ) = f ( x, y ) to be one to
one for each x . The dimension of E , written dimE , is defined to be the dimension of V . If V
does not have an inner product, the set E defined above is called an affine space.

A Euclidean point space is not a vector space but a vector space with inner product is made
a Euclidean point space by defining f ( v1 , v 2 ) ≡ v1 − v 2 for all v ∈V . For an arbitrary point space
the function f is called the point difference, and it is customary to use the suggestive notation

f ( x, y ) = x − y (43.1)

In this notation (a) and (b) above take the forms

297
298 Chap. 9 • EUCLIDEAN MANIFOLDS

x−y =x−z+z−y (43.2)

and

x−y =v (43.3)

Theorem 43.1. In a Euclidean point space E

(i ) x−x =0
(ii ) x − y = −( y − x ) (43.4)
(iii ) if x − y = x '− y ', then x − x ' = y − y '

Proof. For (i) take x = y = z in (43.2); then

x−x = x−x+x−x

which implies x − x = 0 . To obtain (ii) take y = x in (43.2) and use (i). For (iii) observe that

x − y ' = x − y + y − y ' = x − x '+ x '− y '

from (43.2). However, we are given x − y = x '− y ' which implies (iii).

The equation

x−y =v

has the property that given any v and y , x is uniquely determined. For this reason it is customary
to write

x=y+v (43.5)
Sec. 43 • Euclidean Point Spaces 299

for the point x uniquely determined by y ∈E and v ∈V .

The distance from x to y , written d ( x, y ) , is defined by

d (x, y ) = x − y = {( x − y ) ⋅ ( x − y )}
1/ 2
(43.6)

It easily fol1ows from the definition (43.6) and the properties of the inner product that

d ( x, y ) = d ( y , x ) (43.7)

and

d ( x , y ) ≤ d ( x , z ) + d ( z, y ) (43.8)

for all x, y , and z in E . Equation (43.8) is simply rewritten in terms of the points x, y , and z
rather than the vectors x − y , x − z , and z − y . It is also apparent from (43.6) that

d ( x, y ) ≥ 0 and d ( x, y ) = 0 ⇔ x = y (43.9)

The properties (43.7)-(43.9) establish that E is a metric space.

There are several concepts from the theory of metric spaces which we need to summarize.
For simplicity the definitions are sated here in terms of Euclidean point spaces only even though
they can be defined for metric spaces in general.

In a Euclidean point space E an open ball of radius ε > 0 centered at x 0 ∈E is the set

B( x 0 , ε ) = {x d ( x, x 0 ) < ε } (43.10)

and a closed ball is the set

300 Chap. 9 • EUCLIDEAN MANIFOLDS

B ( x 0 , ε ) = {x d ( x, x 0 ) ≤ ε } (43.11)

A neighborhood of x ∈E is a set which contains an open ball centered at x . A subset U of E is

open if it is a neighborhood of each of its points. The empty set ∅ is trivially open because it
contains no points. It also follows from the definitions that E is open.

Theorem 43.2. An open ball is an open set.

Proof. Consider the open ball B( x 0 , ε ) . Let x be an arbitrary point in B( x 0 , ε ) . Then

ε − d ( x, x 0 ) > 0 and the open ball B(x, ε − d (x, x 0 )) is in B(x 0 , ε ) , because if
y ∈ B( x, ε − d ( x, x 0 )) , then d ( y, x ) < ε − d ( x 0 , x ) and, by (43.8), d ( x 0 , y ) ≤ d ( x 0 , x ) + d ( x, y ) ,
which yields d ( x 0 , y ) < ε and, thus, y ∈ B( x 0 , ε ) .

A subset U of E is closed if its complement, E U , is open. It can be shown that closed

balls are indeed closed sets. The empty set, ∅ , is closed because E = E ∅ is open. By the same
logic E is closed since E E = ∅ is open. In fact ∅ and E are the only subsets of E which are
both open and closed. A subset U ⊂ E is bounded if it is contained in some open ball. A subset
U ⊂ E is compact if it is closed and bounded.

Theorem 43.3. The union of any collection of open sets is open.

Proof. Let {Uα α ∈ I } be a collection of open sets, where I is an index set. Assume that
x ∈ ∪ α∈IUα . Then x must belong to at least one of the sets in the collection, say Uα0 . Since Uα0 is
open, there exists an open ball B( x, ε ) ⊂ Uα0 ⊂ ∪ α∈IUα . Thus, B( x, ε ) ⊂ ∪ α∈IUα . Since x is
arbitrary, ∪ α∈IUα is open.

Theorem43.4. The intersection of a finite collection of open sets is open.

Proof. Let {U1 ,...,Uα } be a finite family of open sets. If ∩ni=1Ui is empty, the assertion is trivial.
Thus, assume ∩ni=1Ui is not empty and let x be an arbitrary element of ∩ni=1Ui . Then x ∈Ui for
i = 1,..., n and there is an open ball B( x, ε i ) ⊂ Ui for i = 1,..., n . Let ε be the smallest of the
positive numbers ε1 ,..., ε n . Then x ∈ B( x, ε ) ⊂ ∩ni=1Ui . Thus ∩ni=1Ui is open.
Sec. 43 • Euclidean Point Spaces 301

It should be noted that arbitrary intersections of open sets will not always lead to open sets.
The standard counter example is given by the family of open sets of R of the form ( −1 n ,1 n ) ,
n = 1, 2,3.... The intersection ∩ ∞n =1 ( −1 n ,1 n ) is the set {0} which is not open.

By a sequence in E we mean a function on the positive integers {1, 2,3,..., n,...} with values
in E . The notation {x1 , x 2 , x 3..., x n ,...} , or simply {x n } , is usually used to denote the values of the
sequence. A sequence {x n } is said to converge to a limit x ∈E if for every open ball B (x, ε )
centered at x , there exists a positive integer n0 (ε ) such that x n ∈ B( x, ε ) whenever n ≥ n0 (ε ) .
Equivalently, a sequence {x n } converges to x if for every real number ε > 0 there exists a positive
integer n0 (ε ) such that d ( x n , x ) < ε for all n > n0 (ε ) . If {x n } converges to x , it is conventional to
write

x = lim x n or x → x n as n → ∞
n →∞

Theorem 43.5. If {x n } converges to a limit, then the limit is unique.

Proof. Assume that

x = lim x n , y = lim x n
n →∞ n →∞

Then, from (43.8)

d ( x, y ) ≤ d ( x, x n ) + d ( x n , y )

for every n . Let ε be an arbitrary positive real number. Then from the definition of convergence
of {x n } , there exists an integer n0 (ε ) such that n ≥ n0 (ε ) implies d ( x, x n ) < ε and d ( x n , y ) < ε .
Therefore,

d ( x, y ) ≤ 2ε

for arbitrary ε . This result implies d ( x, y ) = 0 and, thus, x = y .

302 Chap. 9 • EUCLIDEAN MANIFOLDS

A point x ∈E is a limit point of a subset U ⊂ E if every neighborhood of x contains a

point of U distinct from x . Note that x need not be in U . For example, the sphere
{x d (x, x 0 ) = ε } are limit points of the open ball B(x 0 , ε ) . The closure of U ⊂ E , written U , is
the union of U and its limit points. For example, the closure of the open ball B( x 0 , ε ) is the
closed ball B ( x 0 , ε ) . It is a fact that the closure of U is the smallest closed set containing U .
Thus U is closed if and only if U = U .

The reader is cautioned not to confuse the concepts limit of a sequence and limit point of a
subset. A sequence is not a subset of E ; it is a function with values in E . A sequence may have a
limit when it has no limit point. Likewise the set of pints which represent the values of a sequence
may have a limit point when the sequence does not converge to a limit. However, these two
concepts are related by the following result from the theory of metric spaces: a point x is a limit
point of a set U if and only if there exists a convergent sequence of distinct points of U with x as
a limit.

A mapping f :U → E ' , where U is an open set in E and E ' is a Euclidean point space or
an inner product space, is continuous at x 0 ∈U if for every real number ε > 0 there exists a real
number δ ( x 0 , ε ) > 0 such that d ( x, x 0 ) < δ (ε , x 0 ) implies d '( f ( x 0 ), f ( x )) < ε . Here d ' is the
distance function for E ' . When f is continuous at x 0 , it is conventional to write

lim f ( x ) or f ( x ) → f ( x 0 ) as x → x 0
x→x0

The mapping f is continuous on U if it is continuous at every point of U . A continuous mapping

is called a homomorphism if it is one-to-one and if its inverse is also continuous. What we have
just defined is sometimes called a homeomorphism into. If a homomorphism is also onto, then it is
called specifically a homeomorphism onto. It is easily verified that a composition of two
continuous maps is a continuous map and the composition of two homomorphisms is a
homomorphism.

A mapping f :U → E ' , where U and E ' are defined as before, is differentiable at x ∈U if

there exists a linear transformation A x ∈ L (V ;V ') such that

f ( x + v ) = f ( x ) + A x v + o( x, v ) (43.12)
Sec. 43 • Euclidean Point Spaces 303

where

o( x, v )
lim =0 (43.13)
v →0 v

In the above definition V ' denotes the translation space of E ' .

Theorem 43.6. The linear transformation A x in (43.12) is unique.

Proof. If (43.12) holds for A x and A x , then by subtraction we find

(A x − A x ) v = o ( x, v ) − o( x, v )

By (43.13), ( A x − A x ) e must be zero for each unit vector e , so that A x = A x .

If f is differentiable at every point of U , then we can define a mapping

grad f :U → L (V ;V ') , called the gradient of f , by

grad f ( x ) = A x , x ∈U (43.14)

If grad f is continuous on U , then f is said to be of class C 1 . If grad f exists and is

itself of class C 1 , then f is of class C 2 . More generally, f is of class C r , r > 0 , if it is of class
r −1
C r −1 and its ( r − 1) st gradient, written grad f , is of class C 1 . Of course, f is of class C 0 if it is
continuous on U . If f is a C r one-to-one map with a C r inverse f −1 defined on f (U ) , then f
is called a C r diffeomorphism.

If f is differentiable at x , then it follows from (43.12) that

f (x + τ v) − f (x) d
A x v = lim = f (x + τ v ) (43.15)
τ →0 τ dτ τ =0
304 Chap. 9 • EUCLIDEAN MANIFOLDS

for all v ∈V . To obtain (43.15) replace v by τ v , τ > 0 in (43.12) and write the result as

f (x + τ v) − f (x) o( x, τ v )
Ax v = − (43.16)
τ τ

By (43.13) the limit of the last term is zero as τ → 0 , and (43.15) is obtained. Equation (43.15)
holds for all v ∈V because we can always choose τ in (43.16) small enough to ensure that x + τ v
is in U , the domain of f . If f is differentiable at every x ∈U , then (43.15) can be written

d
( grad f (x ) ) v = f (x + τ v) (43.17)
dτ τ =0

A function f :U → R , where U is an open subset of E , is called a scalar field. Similarly,

f :U → V is a vector field, and f :U → T q (V ) is a tensor field of order q . It should be noted
that the term field is defined here is not the same as that in Section 7.

Before closing this section there is an important theorem which needs to be recorded for
later use. We shall not prove this theorem here, but we assume that the reader is familiar with the
result known as the inverse mapping theorem in multivariable calculus.

Theorem 43.7. Let f :U → E ' be a C r mapping and assume that grad f ( x 0 ) is a linear
isomorphism. Then there exists a neighborhood U1 of x 0 such that the restriction of f to U1 is a
C r diffeomorphism. In addition

grad f −1 ( f ( x 0 )) = ( grad f ( x 0 ) )
−1
(43.18)

This theorem provides a condition under which one can asert the existence of a local inverse of a
smooth mapping.

Exercises

43.1 Let a sequence {x n } converge to x . Show that every subsequence of {x n } also

converges to x .
Sec. 43 • Euclidean Point Spaces 305

43.2 Show that arbitrary intersections and finite unions of closed sets yields closed sets.
43.3 Let f :U → E ' , where U is open in E , and E ' is either a Euclidean point space or an
inner produce space. Show that f is continuous on U if and only if f −1 (D ) is open in
E for all D open in f (U ) .
43.4 Let f :U → E ' be a homeomorphism. Show that f maps any open set in U onto an
open set in E ' .
43.5 If f is a differentiable scalar valued function on L (V ;V ) , show that the gradient of f
at A ∈ L (V ;V ) , written

∂f
(A)
∂A

is a linear transformation in L (V ;V ) defined by

⎛ ∂f ⎞ df
tr ⎜ ( A )BT ⎟ = (A + τ B)
⎝ ∂A ⎠ dτ τ =0

for all B ∈ L (V ;V )
43.6 Show that

∂μ1 ∂μ N
(A) = I ( A ) = ( adj A )
T
and
∂A ∂A
306 Chap. 9 • EUCLIDEAN MANIFOLDS

Section 44 Coordinate Systems

Given a Euclidean point space E of dimension N , we define a C r -chart at x ∈E to be a

pair (U , xˆ ) , where U is an open set in E containing x and xˆ :U → R N is a C r diffeomorphism.
Given any chart (U , xˆ ) , there are N scalar fields xˆ i :U → R such that

xˆ ( x ) = ( xˆ1 ( x ),..., xˆ N ( x ) ) (44.1)

for all x ∈U . We call these fields the coordinate functions of the chart, and the mapping x̂ is also
called a coordinate map or a coordinate system on U . The set U is called the coordinate
neighborhood.

Two charts xˆ :U1 → R N and yˆ :U2 → R N , where U1 ∩U2 ≠ ∅ , yield the coordinate
transformation yˆ xˆ −1 : xˆ (U1 ∩U2 ) → yˆ (U1 ∩U2 ) and its inverse xˆ yˆ −1 : yˆ (U1 ∩U2 ) → xˆ (U1 ∩U2 ) .
Since

yˆ (x ) = ( yˆ 1 ( x ),..., yˆ N ( x ) ) (44.2)

The coordinate transformation can be written as the equations

( yˆ (x),..., yˆ
1 N
( x ) ) = yˆ xˆ −1 ( xˆ1 ( x ),..., xˆ N ( x ) ) (44.3)

and the inverse can be written

( xˆ (x),..., xˆ
1 N
( x ) ) = xˆ yˆ −1 ( yˆ 1 ( x ),..., yˆ N ( x ) ) (44.4)

The component forms of (44.3) and (44.4) can be written in the simplified notation

y j = y j ( x1 ,..., x N ) ≡ y j ( x k ) (44.5)
Sec. 44 • Coordinate Systems 307

and

x j = x j ( y1 ,..., y N ) ≡ x j ( y k ) (44.6)

The two N-tuples ( y1 ,..., y N ) and ( x1 ,..., x N ) , where y j = yˆ j ( x ) and x j = xˆ j ( x ) , are the
coordinates of the point x ∈U1 ∩U2 . Figure 6 is useful in understanding the coordinate
transformations.
RN

U1 ∩U2

U1 U2

x̂
ŷ
xˆ(U1 ∩U2 ) yˆ(U1 ∩U2 )
yˆ xˆ −1

xˆ yˆ −1

Figure 6

It is important to note that the quantities x1 ,..., x N , y1 ,..., y N are scalar fields, i.e., real-
valued functions defined on certain subsets of E . Since U1 and U2 are open sets, U1 ∩U2 is open,
and because x̂ and ŷ are C r diffeomorphisms, x̂ (U1 ∩U2 ) and ŷ (U1 ∩U2 ) are open subsets of
R N . In addition, the mapping yˆ xˆ −1 and xˆ yˆ −1 are C r diffeomorphisms. Since x̂ is a
diffeomorphism, equation (44.1) written in the form

xˆ( x ) = ( x1 ,..., x N ) (44.7)

can be inverted to yie1d

308 Chap. 9 • EUCLIDEAN MANIFOLDS

x = x( x1 ,..., x N ) = x( x j ) (44.8)

where x is a diffeomorphism x : xˆ (U ) → U .

A C r -atlas on E is a family (not necessarily countable) of C r -charts {(U , xˆ

α α ) α ∈ I} ,
where I is an index set, such that

E = ∪Uα (44.9)
α∈I

Equation (44.9) states that E is covered by the family of open sets {Uα α ∈ I } . A C r -Euclidean
manifold is a Euclidean point space equipped with a C r -atlas . A C ∞ -atlas and a C ∞ -Euclidean
manifold are defined similarly. For simplicity, we shall assume that E is C ∞ .

A C ∞ curve in E is a C ∞ mapping λ : (a, b) → E , where (a, b) is an open interval of R .

A C ∞ curve λ passes through x 0 ∈E if there exists a c ∈ (a, b) such that λ ( c) = x 0 . Given a chart
(U , xˆ ) and a point x 0 ∈U , the jth coordinate curve passing through x 0 is the curve λ j defined by

λ j (t ) = x ( x01 ,..., x0j −1 , x0j + t , x0j +1 ,..., x0N ) (44.10)

for all t such that ( x01 ,..., x0j −1 , x0j + t , x0j +1 ,..., x0N ) ∈ xˆ (U ) , where ( x0k ) = xˆ (x 0 ) . The subset of U
obtained by requiring

x j = xˆ j (x ) = const (44.11)

is called the jth coordinate surface of the chart

Euclidean manifolds possess certain special coordinate systems of major interest. Let
{i ,..., i } be an arbitrary basis, not necessarily orthonormal, for V . We define N constant vector
1 N

fields i j : E → V , j = 1,..., N , by the formulas

Sec. 44 • Coordinate Systems 309

i j (x) = i j , x ∈U (44.12)

The use of the same symbol for the vector field and its value will cause no confusion and simplifies
the notation considerably. If 0E denotes a fixed element of E , then a Cartesian coordinate system
on E is defined by the N scalar fields zˆ1 , zˆ 2 ,..., zˆ N such that

z j = zˆ j ( x ) = ( x − 0E ) ⋅ i j , x ∈E (44.13)

If the basis {i1 ,..., i N } is orthonormal, the Cartesian system is called a rectangular Cartesian
system. The point 0E is called the origin of the Cartesian coordinate system. The vector field
defined by

r ( x ) = x − 0E (44.14)

for all x ∈E is the position vector field relative to 0E . The value r (x ) is the position vector of x .
If {i1 ,..., i N } is the basis reciprocal to {i1 ,..., i N } , then (44.13) implies

x − 0E = x ( z1 ,..., z N ) − 0E = zˆ j ( x )i j = z j i j (44.15)

Defining constant vector fields i1 ,..., i N as before, we can write (44.15) as

r = zˆ j i j (44.16)

The product of the scalar field zˆ j with the vector field i j in (44.16) is defined pointwise; i.e., if f
is a scalar field and v is a vector field, then fv is a vector field defined by

fv ( x ) = f ( x ) v ( x ) (44.17)

for all x in the intersection of the domains of f and v . An equivalent version of (44.13) is

zˆ j = r ⋅ i j (44.18)
310 Chap. 9 • EUCLIDEAN MANIFOLDS

where the operator r ⋅ i j between vector fields is defined pointwise in a similar fashion as in
(44.17). As an illustration, let { i 1 , i 2 ,..., i N } be another basis for V related to the original basis
by

i j = Qkj i k (44.19)

and let 0E be another fixed point of E . Then by (44.13)

z j = ( x − 0E ) ⋅ i j = ( x − 0E ) ⋅ i j + (0E − 0E ) ⋅ i j
(44.20)
= Qkj ( x − 0E ) ⋅ i k + (0E − 0E ) ⋅ i j = Qkj z k + c j

where (44.19) has been used. Also in (44.20) the constant scalars c k , k = 1,..., N , are defined by

c k = (0E − 0E ) ⋅ i k (44.21)

If the bases {i1 , i 2 ,..., i N } and { i 1 , i 2 ,..., i N } are both orthonormal, then the matrix ⎡⎣Qkj ⎤⎦ is
orthogonal. Note that the coordinate neighborhood is the entire space E .

The jth coordinate curve which passes through 0E of the Cartesian coordinate system is the
curve

λ (t ) = ti j + 0E (44.22)

Equation (44.22) follows from (44.10), (44.15), and the fact that for x = 0E , z1 = z 2 = ⋅⋅⋅ = z N = 0 .
As (44.22) indicates, the coordinate curves are straight lines passing through 0E . Similarly, the jth
coordinate surface is the plane defined by
Sec. 44 • Coordinate Systems 311

( x − 0E ) ⋅ i j = const

z3

i3

z2
0E i2

i1

z1

Figure 7

Geometrically, the Cartesian coordinate system can be represented by Figure 7 for N=3. The
coordinate transformation represented by (44.20) yields the result in Figure 8 (again for N=3).
Since every inner product space has an orthonormal basis (see Theorem 13.3), there is no loss of
generality in assuming that associated with every point of E as origin we can introduce a
rectangular Cartesian coordinate system.
312 Chap. 9 • EUCLIDEAN MANIFOLDS

z3
z3

r
r
i3 i3

z2 0E
0E i2 i2
i1
i1
z2
z1
z1

Figure 8

Given any rectangular coordinate system ( zˆ1 ,..., zˆ N ) , we can characterize a general or a
curvilinear coordinate system as follows: Let (U , xˆ ) be a chart. Then it can be specified by the
coordinate transformation from ẑ to x̂ as described earlier, since in this case the overlap of the
coordinate neighborhood is U = E ∩U . Thus we have

( z1 ,..., z N ) = zˆ xˆ −1 ( x1 ,..., x N )
(44.23)
( x1 ,..., x N ) = xˆ zˆ −1 ( z1 ,..., z N )

where zˆ xˆ −1 : xˆ (U ) → zˆ(U ) and xˆ zˆ −1 : zˆ(U ) → xˆ (U ) are diffeomorphisms. Equivalent versions

of (44.23) are

z j = z j ( x1 ,..., x N ) = z j ( x k )
(44.24)
x j = x j ( z1 ,..., z N ) = x j ( z k )
Sec. 44 • Coordinate Systems 313

As an example of the above ideas, consider the cylindrical coordinate system1. In this case
N=3 and equations (44.23) take the special form

( z1 , z 2 , z 3 ) = ( x1 cos x 2 , x1 sin x 2 , x 3 ) (44.25)

In order for (44.25) to qualify as a coordinate transformation, it is necessary for the transformation
functions to be C ∞ . It is apparent from (44.25) that ẑ xˆ −1 is C ∞ on every open subset of R 3 .
Also, by examination of (44.25), ẑ xˆ −1 is one-to-one if we restrict it to an appropriate domain, say
(0, ∞) × (0, 2π ) × ( −∞, ∞) . The image of this subset of R 3 under ẑ xˆ −1 is easily seen to be the set
{ }
R 3 ( z1 , z 2 , z 3 ) z1 ≥ 0, z 2 = 0 and the inverse transformation is

⎛ 1 2 2 2 1/ 2 −1 z
2
⎞
( x , x , x ) = ⎜ ⎡⎣( z ) + ( z ) ⎤⎦ , tan 1 , z 3 ⎟
1 2 3
(44.26)
⎝ z ⎠

which is also C ∞ . Consequently we can choose the coordinate neighborhood to be any open subset
U in E such that

{
zˆ(U ) ⊂ R 3 ( z1 , z 2 , z 3 ) z1 ≥ 0, z 2 = 0 } (44.27)

or, equivalently,

xˆ(U ) ⊂ (0, ∞) × (0, 2π ) × ( −∞, ∞)

Figure 9 describes the cylindrical system. The coordinate curves are a straight line (for x1 ), a
circle lying in a plane parallel to the ( z1 , z 2 ) plane (for x 2 ), and a straight line coincident with z 3
(for x 3 ). The coordinate surface x1 = const is a circular cylinder whose generators are the z 3
lines. The remaining coordinate surfaces are planes.

1
The computer program Maple has a plot command, coordplot3d, that is useful when trying to visualize coordinate
curves and coordinate surfaces. The program MATLAB will also produce useful and instructive plots.
314 Chap. 9 • EUCLIDEAN MANIFOLDS

z3 = x3

z2

x2
z1 x1

Figure 9

Returning to the general transformations (44.5) and (44.6), we can substitute the second into
the first and differentiate the result to find

∂y i 1 N ∂x
k
( x ,..., x ) ( y1 ,..., y N ) = δ ij (44.28)
∂x k ∂y j

By a similar argument with x and y interchanged,

∂y k 1 N ∂x
i
( x ,..., x ) k ( y1 ,..., y N ) = δ ij (44.29)
∂x j
∂y

Each of these equations ensures that

⎡ ∂y i ⎤ 1
det ⎢ k ( x1 ,..., x N ) ⎥ = ≠0 (44.30)
⎣ ∂x ⎦ det ⎡ ∂x ( y1 ,..., y N ) ⎤
j

⎢ ∂y l ⎥
⎣ ⎦

For example, det ⎡⎣( ∂z j ∂x k )( x1 , x 2 , x 3 ) ⎤⎦ = x1 ≠ 0 for the cylindrical coordinate system. The
determinant det ⎡⎣( ∂x j ∂y k )( y1 ,..., y N ) ⎤⎦ is the Jacobian of the coordinate transformation (44.6).
Sec. 44 • Coordinate Systems 315

Just as a vector space can be assigned an orientation, a Euclidean manifold can be oriented
by assigning the orientation to its translation space V . In this case E is called an oriented
Euclidean manifold. In such a manifold we use Cartesian coordinate systems associated with
positive basis only and these coordinate systems are called positive. A curvilinear coordinate
system is positive if its coordinate transformation relative to a positive Cartesian coordinate system
has a positive Jacobian.

Given a chart (U , xˆ ) for E , we can compute the gradient of each coordinate function xˆ i
and obtain a C ∞ vector field on U . We shall denote each of these fields by g i , namely

g i = grad xˆ i (44.31)

for i = 1,..., N . From this definition, it is clear that g i (x ) is a vector in V normal to the ith
coordinate surface. From (44.8) we can define N vector fields g1 ,..., g N on U by

g i = [ grad x ] (0,...,1,...,0) (44.32)

i

Or, equivalently,

x( x1 ,..., x i + t ,..., x N ) − x( x1 ,..., x N ) ∂x 1

g i = lim ≡ i ( x ,..., x N ) (44.33)
t →0 t ∂x

for all x ∈U . Equations (44.10) and (44.33) show that g i ( x ) is tangent to the ith coordinate curve.
Since

xˆ i ( x( x1 ,..., x N )) = x i (44.34)

The chain rule along with the definitions (44.31) and (44.32) yield

g i ( x ) ⋅ g j ( x ) = δ ij (44.35)

as they should.
316 Chap. 9 • EUCLIDEAN MANIFOLDS

The values of the vector fields {g1 ,..., g N } form a linearly independent set of vectors
{g1 (x ),..., g N (x )} at each x ∈U . To see this assertion, assume x ∈U and

λ 1g1 ( x ) + λ 2g 2 ( x ) + ⋅⋅⋅ + λ N g N ( x ) = 0

for λ 1 ,..., λ N ∈R . Taking the inner product of this equation with g j (x ) and using equation (44.35)
, we see that λ j = 0 , j = 1,..., N which proves the assertion.

Because V has dimension N, {g1 ( x ),..., g N ( x )} forms a basis for V at each x ∈U .

Equation (44.35) shows that {g1 ( x ),..., g N ( x )} is the basis reciprocal to {g1 ( x ),..., g N ( x )} . Because
of the special geometric interpretation of the vectors {g1 ( x ),..., g N ( x )} and {g1 (x ),..., g N ( x )}
mentioned above, these bases are called the natural bases of x̂ at x . Any other basis field which
cannot be determined by either (44.31) or (44.32) relative to any coordinate system is called an
anholonomic or nonintegrable basis. The constant vector fields {i1 ,..., i N } and {i1 ,..., i N } yield the
natural bases for the Cartesian coordinate systems.

If (U1 , xˆ ) and (U2 , yˆ ) are two charts such that U1 ∩U2 ≠ ∅ , we can determine the
transformation rules for the changes of natural bases at x ∈U1 ∩U2 in the following way: We shall
let the vector fields h j , j = 1,..., N , be defined by

h j = grad yˆ j (44.36)

Then, from (44.5) and (44.31),

∂y j 1
h j ( x ) = grad yˆ j ( x ) = ( x ,..., x N ) grad xˆ i (x )
∂x i
(44.37)
∂y j
= i ( x1 ,..., x N )g i ( x )
∂x

for all x ∈U1 ∩U2 . A similar calculation shows that

Sec. 44 • Coordinate Systems 317

∂x i 1
h j (x) = ( y ,..., y N )g i ( x ) (44.38)
∂y j

for all x ∈U1 ∩U2 . Equations (44.37) and (44.38) are the desired transformations.

Given a chart (U1 , xˆ ) , we can define 2N 2 scalar fields gij :U → R and g ij :U → R by

gij ( x ) = g i ( x ) ⋅ g j (x ) = g ji ( x ) (44.39)

and

g ij ( x ) = g i ( x ) ⋅ g j ( x ) = g ji ( x ) (44.40)

for all x ∈U . It immediately follows from (44.35) that

−1
⎣⎡ g ( x ) ⎦⎤ = ⎡⎣ gij ( x ) ⎤⎦
ij
(44.41)

since we have

g i = g ij g j (44.42)

and

g j = g ji g i (44.43)

where the produce of the scalar fields with vector fields is defined by (44.17). If θij is the angle
between the ith and the jth coordinate curves at x ∈U , then from (44.39)

gij (x )
cos θij = 1/ 2
(no sum) (44.44)
⎡⎣ gii ( x ) g jj ( x ) ⎤⎦
318 Chap. 9 • EUCLIDEAN MANIFOLDS

Based upon (44.44), the curvilinear coordinate system is orthogonal if gij = 0 when i ≠ j . The
symbol g denotes a scalar field on U defined by

g ( x ) = det ⎡⎣ gij ( x ) ⎤⎦ (44.45)

for all x ∈U .

At the point x ∈U , the differential element of arc ds is defined by

ds 2 = dx ⋅ dx (44.46)

and, by (44.8), (44.33), and (44.39),

∂x ∂x
ds 2 = ( xˆ ( x )) ⋅ j ( xˆ ( x ))dx i dx j
∂x i
∂x (44.47)
= gij ( x )dx dx j
i

If (U1 , xˆ ) and (U2 , yˆ ) are charts where U1 ∩U2 ≠ ∅ , then at x ∈U1 ∩U2

hij ( x ) = hi ( x ) ⋅ h j (x )
∂x k 1 N ∂x
l (44.48)
= ( y ,..., y ) ( y1 ,..., y N ) g kl ( x )
∂y i ∂y j

Equation (44.48) is helpful for actual calculations of the quantities gij (x ) . For example, for
the transformation (44.23), (44.48) can be arranged to yield

∂z k 1 N ∂z
k
gij ( x ) = ( x ,..., x ) ( x1 ,..., x N ) (44.49)
∂x i
∂x j

since i k ⋅ i l = δ kl . For the cylindrical coordinate system defined by (44.25) a simple calculation
based upon (44.49) yields
Sec. 44 • Coordinate Systems 319

⎡1 0 0⎤
⎢
⎡⎣ gij (x ) ⎤⎦ = 0 ( x1 ) 2 0⎥ (44.50)
⎢ ⎥
⎣⎢0 0 1 ⎦⎥

Among other things, (44.50) shows that this coordinate system is orthogonal. By (44.50) and
(44.41)

⎡1 0 0⎤
⎢
⎡⎣ g ( x ) ⎤⎦ = 0 1 ( x1 ) 2
ij
0⎥ (44.51)
⎢ ⎥
⎣⎢0 0 1 ⎦⎥

And, from (44.50) and (44.45),

g ( x ) = ( x1 ) 2 (44.52)

It follows from (44.50) and (44.47) that

ds 2 = ( dx1 ) 2 + ( x1 ) 2 ( dx 2 ) 2 + ( dx 3 )2 (44.53)

Exercises

44.1 Show that

∂x
ik = and i k = grad zˆ k
∂z k

for any Cartesian coordinate system ẑ associated with {i j } .

44.2 Show that

I = grad r (x )
320 Chap. 9 • EUCLIDEAN MANIFOLDS

44.3 Spherical coordinates ( x1 , x 2 , x 3 ) are defined by the coordinate transformation

z1 = x1 sin x 2 cos x 3
z 2 = x1 sin x 2 sin x 3
z 3 = x1 cos x 2

relative to a rectangular Cartesian coordinate system ẑ . How must the quantity ( x1 , x 2 , x 3 )

be restricted so as to make ẑ xˆ −1 one-to-one? Discuss the coordinate curves and the
coordinate surfaces. Show that

⎡1 0 0 ⎤
⎡⎣ gij ( x ) ⎤⎦ = ⎢0 ( x1 ) 2 0 ⎥
⎢ ⎥
⎢⎣0 0 ( x1 sin x 2 ) 2 ⎥⎦

44.4 Paraboloidal coordinates ( x1 , x 2 , x 3 ) are defined by

z1 = x1 x 2 cos x 3
z 2 = x1 x 2 sin x 3

z 3 = ( ( x1 ) 2 − ( x 2 ) 2 )
1
2

Relative to a rectangular Cartesian coordinate system ẑ . How must the quantity ( x1 , x 2 , x 3 )

be restricted so as to make ẑ xˆ −1 one-to-one. Discuss the coordinate curves and the
coordinate surfaces. Show that

⎡ ( x1 ) 2 + ( x 2 ) 2 0 0 ⎤
⎢ ⎥
⎡⎣ gij ( x ) ⎤⎦ = ⎢ 0 (x ) + (x )
1 2 2 2
0 ⎥
⎢ 0 0 ( x1 x 2 ) 2 ⎥⎦
⎣

44.5 A bispherical coordinate system ( x1 , x 2 , x 3 ) is defined relative to a rectangular Cartesian

coordinate system by
Sec. 44 • Coordinate Systems 321

a sin x 2 cos x 3
z =
1

cosh x1 − cos x 2
a sin x 2 sin x 3
z2 =
cosh x1 − cos x 2

and

a sinh x1
z =
3

cosh x1 − cos x 2

where a > 0 . How must ( x1 , x 2 , x 3 ) be restricted so as to make ẑ xˆ −1 one-to-one?

Discuss the coordinate curves and the coordinate surfaces. Also show that

⎡ a2 ⎤
⎢ 0 0 ⎥
⎢ ( cosh x1 − cos x 2 ) ⎥
2

⎢ ⎥
⎢ a2 ⎥
⎡⎣ gij ( x ) ⎤⎦ = ⎢ 0 0 ⎥
( cosh x − cos x 2 )
1 2
⎢ ⎥
⎢ ⎥
⎢ a 2 (sin x 2 ) 2 ⎥
0 0
⎢
⎣ ( cosh x − cos x ) ⎦
1 2 2⎥

44.6 Prolate spheroidal coordinates ( x1 , x 2 , x 3 ) are defined by

z1 = a sinh x1 sin x 2 cos x 3

z 2 = a sinh x1 sin x 2 sin x 3
z 3 = a cosh x1 cos x 2

relative to a rectangular Cartesian coordinate system ẑ , where a > 0 . How must ( x1 , x 2 , x 3 )

be restricted so as to make ẑ xˆ −1 one-to-one? Also discuss the coordinate curves and the
coordinate surfaces and show that
322 Chap. 9 • EUCLIDEAN MANIFOLDS

⎡ a 2 ( cosh 2 x1 − cos2 x 2 ) 0 0 ⎤
⎢ ⎥
⎡⎣ gij ( x ) ⎤⎦ = ⎢ 0 a 2 ( cosh 2 x1 − cos2 x 2 ) 0 ⎥
⎢ ⎥
⎢ 0 0 a 2 sinh 2 x1 sin 2 x 2 ⎥
⎣ ⎦

44.7 Elliptical cylindrical coordinates ( x1 , x 2 , x 3 ) are defined relative to a rectangular Cartesian

coordinate system by

z1 = a cosh x1 cos x 2
z 2 = a sinh x1 sin x 2
z3 = x3

where a > 0 . How must ( x1 , x 2 , x 3 ) be restricted so as to make ẑ xˆ −1 one-to-one?

Discuss the coordinate curves and coordinate surfaces. Also, show that

⎡ a 2 ( sinh 2 x1 + sin 2 x 2 ) 0 0⎤
⎢ ⎥
⎡⎣ gij ( x ) ⎤⎦ = ⎢ 0 a 2 ( sinh 2 x1 + sin 2 x 2 ) 0⎥
⎢ ⎥
⎢ 0 0 1⎥
⎣ ⎦

44.8 For the cylindrical coordinate system show that

g1 = (cos x 2 )i1 + (sin x 2 )i 2

g 2 = − x1 (sin x 2 )i1 + x1 (cos x 2 )i 2
g3 = i3

44.9 At a point x in E , the components of the position vector r (x ) = x − 0E with respect to the
basis {i1 ,..., i N } associated with a rectangular Cartesian coordinate system are z1 ,..., z N .
This observation follows, of course, from (44.16). Compute the components of r ( x ) with
respect to the basis {g1 ( x ), g 2 ( x ), g 3 ( x )} for (a) cylindrical coordinates, (b) spherical
coordinates, and (c) parabolic coordinates. You should find that
Sec. 44 • Coordinate Systems 323

r ( x ) = x1g1 ( x ) + x 3g 3 ( x ) for (a)

r ( x ) = x g1 ( x )
1
for (b)
1 1 1
r(x) = x g1 ( x ) + x 2 g 2 ( x ) for (c)
2 2

44.10 Toroidal coordinates ( x1 , x 2 , x 3 ) are defined relative to a rectangular Cartesian coordinate

system by

a sinh x1 cos x 3
z1 =
cosh x1 − cos x 2
a sinh x1 sin x 3
z =
2

cosh x1 − cos x 2

and

a sin x 2
z3 =
cosh x1 − cos x 2

where a > 0 . How must ( x1 , x 2 , x 3 ) be restricted so as to make ẑ xˆ −1 one to one? Discuss

the coordinate surfaces. Show that

⎡ a2 ⎤
⎢ 0 0 ⎥
⎢ ( cosh x1 − cos x 2 ) ⎥
2

⎢ ⎥
⎢ a2 ⎥
⎡⎣ gij ( x ) ⎤⎦ = ⎢ 0 0 ⎥
( cosh x − cos x 2 )
1 2
⎢ ⎥
⎢ ⎥
⎢ a 2 sinh 2 x 2 ⎥
0 0
⎢
( cos x1 − cos x 2 ) ⎥⎦
2

⎣
324 Chap. 9 • EUCLIDEAN MANIFOLDS

Section 45. Transformation Rules for Vectors and Tensor Fields

In this section, we shall formalize certain ideas regarding fields on E and then investigate
the transformation rules for vectors and tensor fields. Let U be an open subset of E ; we shall
denote by F ∞ (U ) the set of C ∞ functions f :U → R . First we shall study the algebraic structure
of F ∞ (U ) . If f1 and f 2 are in F ∞ (U ) , then their sum f1 + f 2 is an element of F ∞ (U ) defined by

( f1 + f 2 )( x ) = f1 ( x ) + f 2 ( x ) (45.1)

and their produce f1 f 2 is also an element of F ∞ (U ) defined by

( f1 f 2 )( x ) = f1 (x ) f 2 (x ) (45.2)

for all x ∈U . For any real number λ ∈R the constant function is defined by

λ (x) = λ (45.3)

for all x ∈U . For simplicity, the function and the value in (45.3) are indicated by the same
symbol. Thus, the zero function in F ∞ (U ) is denoted simply by 0 and for every f ∈ F ∞ (U )

f +0= f (45.4)

It is also apparent that

1f = f (45.5)

In addition, we define

− f = ( −1) f (45.6)
Sec. 45 • Transformation Rules 325

It is easily shown that the operations of addition and multiplication obey commutative, associative,
and distributive laws. These facts show that F ∞ (U ) is a commutative ring (see Section 7).

An important collection of scalar fields can be constructed as follows: Given two charts
(U1 , xˆ ) and (U2 , yˆ ) , where U1 ∩U2 ≠ ∅ , we define the N 2 partial derivatives ( ∂y i ∂x j )( x1 ,..., x N )
at every ( x1 ,..., x N ) ∈ xˆ (U1 ∩U2 ) . Using a suggestive notation, we can define N 2 C ∞ functions
∂y i ∂x j :U1 ∩U2 → R by

∂y i ∂y i
( x ) = xˆ (x ) (45.7)
∂x j ∂x j

for all x ∈U1 ∩U2 .

As mentioned earlier, a C ∞ vector field on an open set U of E is a C ∞ map v :U → V ,

where V is the translation space of E . The fields defined by (44.31) and (44.32) are special cases
of vector fields. We can express v in component forms on U1 ∩U2 ,

v = υ i gi = υ j g j (45.8)

υ j = g jiυ i (45.9)

As usual, we can computer υ i :U ∩U1 → R by

υ i (x ) = v( x ) ⋅ g i (x ), x ∈U1 ∩U2 (45.10)

and υi by (45.9). In particular, if (U2 , yˆ ) is another chart such that U1 ∩U2 ≠ ∅ , then the
component form of h j relative to x̂ is

∂x i
hj = gi (45.11)
∂y j
326 Chap. 9 • EUCLIDEAN MANIFOLDS

With respect to the chart (U2 , yˆ ) , we have also

v = υ k hk (45.12)

where υ k :U ∩U2 → R . From (45.12), (45.11), and (45.8), the transformation rule for the
components of v relative to the two charts (U1 , xˆ ) and (U2 , yˆ ) is

∂x i
υ = jυ
i j
(45.13)
∂y

for all x ∈U1 ∩U2 ∩U .

As in (44.18), we can define an inner product operation between vector fields. If

v1 :U1 → V and v 2 :U2 →V are vector fields, then v1 ⋅ v 2 is a scalar field defined on U1 ∩U2 by

v1 ⋅ v 2 (x) = v1 (x) ⋅ v 2 (x), x ∈U1 ∩U2 (45.14)

Then (45.10) can be written

υ i = v ⋅ gi (45.15)

Now let us consider tensor fields in general. Let Tq∞ (U ) denote the set of all tensor fields of
order q defined on an open set U in E . As with the set F ∞ (U ) , the set Tq∞ (U ) can be assigned
an algebraic structure. The sum of A :U → T q (V ) and B :U → T q (V ) is a C ∞ tensor field
A + B :U → T q (V ) defined by

( A + B ) (x) = A(x) + B(x) (45.16)

for all x ∈U . If f ∈ F ∞ (U ) and A ∈ Tq∞ (U ) , then we can define f A ∈ Tq∞ (U ) by

Sec. 45 • Transformation Rules 327

f A ( x) = f ( x) A ( x ) (45.17)

Clearly this multiplication operation satisfies the usual associative and distributive laws with
respect to the sum for all x ∈U . As with F ∞ (U ) , constant tensor fields in Tq∞ (U ) are given the
same symbol as their value. For example, the zero tensor field is 0 :U → T q (V ) and is defined by

0( x ) = 0 (45.18)

for all x ∈U . If 1 is the constant function in F ∞ (U ) , then

− A = (−1) A (45.19)

The algebraic structure on the set Tq∞ (V ) just defined is called a module over the ring F ∞ (U ) .

The components of a tensor field A :U → T q (V ) with respect to a chart (U1 , xˆ ) are the N q
scalar fields Ai1 ...iq :U ∩U1 → R defined by

Ai1 ...iq (x) = A (x)(g i1 (x),..., g iq (x)) (45.20)

for all x ∈U ∩U1 . Clearly we can regard tensor fields as multilinear mappings on vector fields
with values as scalar fields. For example, A (g i1 ,..., g iq ) is a scalar field defined by

A (g i1 ,..., g iq )(x) = A (x)(g i1 (x),..., g iq (x))

for all x ∈U ∩U1 . In fact we can, and shall, carry over to tensor fields the many algebraic
operations previously applied to tensors. In particular a tensor field A :U → T q (V ) has the
representation

A = Ai1...iq g i1 ⊗ ⋅⋅⋅ ⊗ g q
i
(45.21)
328 Chap. 9 • EUCLIDEAN MANIFOLDS

for all x ∈U ∩U1 , where U1 is the coordinate neighborhood for a chart (U1 , xˆ ) . The scalar fields
Ai1...iq are the covariant components of A and under a change of coordinates obey the
transformation rule

∂xi1 ∂x q
i

Ak1...kq = k1 ⋅⋅⋅ kq Ai1...iq (45.22)

∂y ∂y

Equation (45.22) is a relationship among the component fields and holds at all points
x ∈U1 ∩U2 ∩U where the charts involved are (U1 , xˆ ) and (U2 , yˆ ) . We encountered an example of
(45.22) earlier with (44.48). Equation (44.48) shows that the gij are the covariant components of a
tensor field I whose value is the identity or metric tensor, namely

I = gij g i ⊗ g j = g j ⊗ g j = g j ⊗ g j = g ij g i ⊗ g j (45.23)

for points x ∈U1 , where the chart in question is (U1 , xˆ ) . Equations (45.23) show that the
components of a constant tensor field are not necessarily constant scalar fields. It is only in
Cartesian coordinates that constant tensor fields have constant components.

Another important tensor field is the one constructed from the positive unit volume tensor
E . With respect to an orthonormal basis {i j } , which has positive orientation, E is given by (41.6),
i.e.,

E = i1 ∧ ⋅⋅⋅ ∧ i N = ε i1 ⋅⋅⋅iN i i1 ⊗ ⋅⋅⋅ ⊗ i iN (45.24)

Given this tensor, we define as usual a constant tensor field E : E → T ˆN (V ) by

E( x ) = E (45.25)

for all x ∈E . With respect to a chart (U1 , xˆ ) , it follows from the general formula (42.27) that

E = Ei1⋅⋅⋅iN g i1 ⊗ ⋅⋅⋅ ⊗ g iN = E i1⋅⋅⋅iN g i1 ⊗ ⋅⋅⋅ ⊗ g iN (45.26)

Sec. 45 • Transformation Rules 329

where Ei1⋅⋅⋅iN and E i1⋅⋅⋅iN are scalar fields on U1 defined by

Ei1⋅⋅⋅iN = e g ε i1⋅⋅⋅iN (45.27)

and

e i1⋅⋅⋅iN
E i1⋅⋅⋅iN = ε (45.28)
g

where, as in Section 42, e is +1 if {g i ( x )} is positively oriented and −1 if {g i ( x )} is negatively

oriented, and where g is the determinant of ⎡⎣ gij ⎤⎦ as defined by (44.45). By application of (42.28),
it follows that

E i1⋅⋅⋅iN = g i1 j1 ⋅⋅⋅ g iN jN E j1⋅⋅⋅ jN (45.29)

An interesting application of the formulas derived thus far is the derivation of an expression
for the differential element of volume in curvilinear coordinates. Given the position vector r
defined by (44.14) and a chart (U1 , xˆ ) , the differential of r can be written

dr = dx = g i ( x )dx i (45.30)

where (44.33) has been used. Given N differentials of r , dr1 , dr2 , ⋅⋅⋅, drN , the differential volume
element dυ generated by them is defined by

dυ = E ( dr1 , ⋅⋅⋅, drN ) (45.31)

or, equivalently,

dυ = dr1 ∧ ⋅ ⋅ ⋅ ∧ drN (45.32)

330 Chap. 9 • EUCLIDEAN MANIFOLDS

If we select dr1 = g1 ( x )dx1 , dr2 = g 2 ( x )dx 2 , ⋅⋅⋅, drN = g N ( x )dx N , we can write (45.31)) as

dυ = E ( g1 ( x ), ⋅⋅⋅, g N ( x ) ) dx1dx 2 ⋅⋅⋅ dx N (45.33)

By use of (45.26) and (45.27), we then get

E ( g1 ( x ), ⋅⋅⋅, g N ( x ) ) = E12⋅⋅⋅N = e g (45.34)

Therefore,

dυ = gdx1dx 2 ⋅⋅⋅ dx N (45.35)

For example, in the parabolic coordinates mentioned in Exercise 44.4,

dυ = (( x ) + ( x ) ) x x dx dx dx
1 2 2 2 1 2 1 2 3
(45.36)

Exercises

45.1 Let v be a C ∞ vector field and f be a C ∞ function both defined on U an open set in E .
We define v f :U → R by

v f (x) = v(x) ⋅ grad f (x), x ∈U (45.37)

Show that

v (λ f + μ g ) = λ ( v f ) + μ (v g)

and
Sec. 45 • Transformation Rules 331

v ( fg ) = ( v f ) g + f (v g)

For all constant functions λ , μ and all C ∞ functions f and g . In differential geometry,
an operator on F ∞ (U ) with the above properties is called a derivation. Show that,
conversely, every derivation on F ∞ (U ) corresponds to a unique vector field on U by
(45.37).
45.2 By use of the definition (45.37), the Lie bracket of two vector fields v :U →V and
u :U → V , written [u, v ] , is a vector field defined by

[ u, v ] f =u (v f )−v (u f) (45.38)

For all scalar fields f ∈ F ∞ (U ) . Show that [u, v ] is well defined by verifying that (45.38)
defines a derivation on [u, v ] . Also, show that

[u, v ] = ( grad v ) u − ( grad u ) v

and then establish the following results:

(a) [ u, v ] = − [ v , u ]
(b) ⎡⎣ v, [u, w ]⎤⎦ + ⎡⎣u, [ w, v ]⎤⎦ + ⎡⎣ w, [ v, u ]⎤⎦ = 0

for all vector fields u, v and w .

(c) Let (U1 , xˆ ) be a chart with natural basis field {g i } . Show that ⎡⎣ g i , g j ⎤⎦ = 0 .

The results (a) and (b) are known as Jacobi’s identities.

45.3 In a three-dimensional Euclidean space the differential element of area normal to the plan
formed from dr1 and dr2 is defined by

dσ = dr1 × dr2

Show that

dσ = Eijk dx1j dx2k g i (x)

332 Chap. 9 • EUCLIDEAN MANIFOLDS

Section 46. Anholonomic and Physical Components of Tensors

In many applications, the components of interest are not always the components with
respect to the natural basis fields {g i } and {g j } . For definiteness let us call the components of a
tensor field A ∈ Tq∞ (U ) is defined by (45.20) the holonomic components of A . In this section, we
shall consider briefly the concept of the anholonomic components of A ; i.e., the components of
A taken with respect to an anholonomic basis of vector fields. The concept of the physical
components of a tensor field is a special case and will also be discussed.

Let U1 be an open set in E and let {e a } denote a set of N vectors fields on U1 , which are
linearly independent, i.e., at each x ∈U1 , {e a } is a basis of V . If A is a tensor field in Tq∞ (U )
where U1 ∩U ≠ ∅ , then by the same type of argument as used in Section 45, we can write

A = Aa1a2 ...aq e a1 ⊗ ⋅⋅⋅ ⊗ e

aq
(46.1)

or, for example,

A= A1 e b1 ⊗ ⋅⋅⋅ ⊗ e bq
b ... bq
(46.2)

where {e a } is the reciprocal basis field to {e a } defined by

e a ( x ) ⋅ e b ( x ) = δ ba (46.3)

for all x ∈U1 . Equations (46.1) and (46.2) hold on U ∩U1 , and the component fields as defined by

Aa1a2 ...aq = A(e a1 ,..., e aq ) (46.4)

and

= A(e b1 ,..., e q )
b ... bq b
A1 (46.5)
Sec. 46 • Anholonomic and Physical Components 333

are scalar fields on U ∩U1 . These fields are the anholonomic components of A when the bases
{e } and {e } are not the natural bases of any coordinate system.
a
a

Given a set of N vector fields {e a } as above, one can show that a necessary and sufficient
condition for {e a } to be the natural basis field of some chart is

[e a , e b ] = 0

for all a, b = 1,..., N , where the bracket product is defined in Exercise 45.2. We shall prove this
important result in Section 49. Formulas which generalize (45.22) to anholonomic components can
easily be derived. If {eˆ a } is an anholonomic basis field defined on an open set U2 such that
U1 ∩U2 ≠ ∅ , then we can express each vector field eˆ b in anholonomic component form relative to
the basis {e a } , namely

eˆ b = Tba e a (46.6)

where the Tba are scalar fields on U1 ∩U2 defined by

Tba = eˆ b ⋅ e a

The inverse of (46.6) can be written

e a = Tˆabeˆ ba (46.7)

where

Tˆab ( x )Tca ( x ) = δ cb (46.8)

and
334 Chap. 9 • EUCLIDEAN MANIFOLDS

Tca ( x )Tˆbc (x ) = δ ba (46.9)

for all x ∈U1 ∩U2 . It follows from (46.4) and (46.7) that

Aa1 ...aq = Tˆab11 ⋅⋅⋅ Tˆaqq Aˆb1 ...bq

b
(46.10)

where

Aˆb1 ...bq = A(eˆ b1 ,..., eˆ bq ) (46.11)

Equation (46.10) is the transformation rule for the anholonomic components of A . Of course,
(46.10) is a field equation which holds at every point of U1 ∩U2 ∩U . Similar transformation rules
for the other components of A can easily be derived by the same type of argument used above.

We define the physical components of A , denoted by A a ,...,a , to be the anholonomic

1 q

components of A relative to the field of orthonomal basis g i { } whose basis vectors g i

are unit
vectors in the direction of the natural basis vectors g i of an orthogonal coordinate system. Let
(U1 , xˆ ) by such a coordinate system with gij = 0, i ≠ j . Then we define

g i (x) = gi (x) gi (x) (no sum) (46.12)

at every x ∈U1 . By (44.39), an equivalent version of (46.12) is

g i ( x ) = g i ( x ) ( gii ( x ))1 2 (no sum) (46.13)

Since {g i } is orthogonal, it follows from (46.13) and (44.39) that g i { } is orthonormal:

g a ⋅ g b = δ ab (46.14)
Sec. 46 • Anholonomic and Physical Components 335

as it should, and it follows from (44.41) that

⎡1 g11 ( x ) 0 ⋅ ⋅ ⋅ ⎤
0
⎢ 0 1 g 22 ( x ) 0 ⎥
⎢ ⎥
⎢ ⋅ ⋅ ⋅ ⎥
⎡⎣ g ij ( x ) ⎤⎦ = ⎢ ⎥ (46.15)
⎢ ⋅ ⋅ ⋅ ⎥
⎢ ⋅ ⋅ ⋅ ⎥
⎢ ⎥
⎣ 0 0 ⋅ ⋅ ⋅ 1 g NN (x ) ⎦

This result shows that g i can also be written

gi (x )
g i (x) = = ( gii ( x ))1 2 g i ( x ) (no sum) (46.16)
( g ii ( x ))1 2

Equation (46.13) can be viewed as a special case of (46.7), where

⎡1 g11 0 ⋅ ⋅ ⋅ ⎤ 0
⎢ ⎥
⎢ 0 1 g 22 0 ⎥
⎢ ⋅ ⋅ ⋅ ⎥
⎡Tˆab ⎤ = ⎢ ⎥ (46.17)
⎣ ⎦ ⎢ ⋅ ⋅ ⋅ ⎥
⎢ ⎥
⎢ ⋅ ⋅ ⋅ ⎥
⎢ 0 0 ⋅ ⋅ ⋅ 1 g NN ⎥⎦
⎣
By the transformation rule (46.10), the physical components of A are related to the covariant
components of A by

( )
−1 2
A a a ... a ≡ A(g a1 , g a2 ,..., g a ) = g a1a1 ⋅⋅⋅ g aq aq Aa1 ... aq (no sum) (46.18)
1 2 q q

Equation (46.18) is a field equation which holds for all x ∈U . Since the coordinate system is
orthogonal, we can replace (46.18)1 with several equivalent formulas as follows:
336 Chap. 9 • EUCLIDEAN MANIFOLDS

( )
12
A a a ...a = g a1a1 ⋅⋅⋅ g aq aq
a ... aq
A1
1 2 q

( ) (g )
−1 2 12
= g a1a1 ⋅⋅⋅ g aq aq
a2 ... aq
a2 a2 Aa1
⋅ (46.19)
⋅
⋅

( ) (g )
−1 2 12
= g a1a1 ⋅⋅⋅ g aq−1aq−1
aq
aq aq Aa1 ... aq−1

In mathematical physics, tensor fields often arise naturally in component forms relative to
product bases associated with several bases. For example, if {e a } and {eˆ b } are fields of bases,
possibly anholonomic, then it might be convenient to express a second-order tensor field A as a
field of linear transformations such that

Ae a = Ab a eˆ b ,
ˆ
a = 1,..., N (46.20)

In this case A has naturally the component form

A = Ab a eˆ b ⊗ e a
ˆ
(46.21)

Relative to the product basis {eˆ b ⊗ e a } formed by {eˆ b } and {e a } , the latter being the reciprocal
basis of {e a } as usual. For definiteness, we call {eˆ b ⊗ e a } a composite product basis associated
with the bases {eˆ b } and {e a } . Then the scalar fields Ab̂ a defined by (46.20) or (46.21), may be
called the composite components of A , and they are given by

Ab a = A(eˆ b ⊗ e a )
ˆ
(46.22)

Similarly we may define other types of composite components, e.g.,

Abaˆ = A(eˆ b , e a ), Aba = A(eˆ b , e a )

ˆ
(46.23)

etc., and these components are related to one another by

Sec. 46 • Anholonomic and Physical Components 337

Abaˆ = Acˆ a gˆ cbˆ ˆ = Abˆ c g ca = Acdˆ gˆ bcˆ ˆ g da (46.24)

etc. Further, the composite components are related to the regular tensor components associated
with a single basis field by

Abaˆ = AcaTbˆc = Abcˆ ˆTˆacˆ , etc. (46.25)

where Tbˆa and Tˆbaˆ are given by (46.6) and (46.7) as before, In the special case where {e a } and
{eˆ b } are orthogonal but not orthonormal, we define the normalized basis vectors e a and eˆ a as
before. Then the composite physical components A bˆ,a of A are given by

A bˆ ,a = A(eˆ bˆ , e a ) (46.26)

and these are related to the composite components by

A bˆ,a = ( gˆ bbˆ ˆ ) Ab a ( g aa )
12 ˆ 12
(no sum) (46.27)

Clearly the concepts of composite components and composite physical components can be
defined for higher order tensors also.

Exercises

46.1 In a three-dimensional Euclidean space the covariant components of a tensor field A

relative to the cylindrical coordinate system are Aij . Determine the physical components of
A.
46.2 Relative to the cylindrical coordinate system the helical basis {e a } has the component form
338 Chap. 9 • EUCLIDEAN MANIFOLDS

e1 = g 1
e 2 = ( cos α ) g 2 + ( sin α ) g 3 (46.28)
e 3 = − ( sin α ) g 2 + ( cos α ) g 3

where α is a constant called the pitch, and where g α { } is the orthonormal basis associated
with the natural basis of the cylindrical system. Show that {e a } is anholonomic. Determine
the anholonomic components of the tensor field A in the preceding exercise relative to the
helical basis.
46.3 Determine the composite physical components of A relative to the composite product basis
{e a ⊗g b }
Sec. 47 • Christoffel Symbols, Covariant Differentiation 339

Section 47. Christoffel Symbols and Covariant Differentiation

In this section we shall investigate the problem of representing the gradient of various
tensor fields in components relative to the natural basis of arbitrary coordinate systems. We
consider first the simple case of representing the tangent of a smooth curve in E . Let
λ : ( a, b ) → E be a smooth curve passing through a point x , say x = λ (c) . Then the tangent vector
of λ at x is defined by

dλ (t )
λ = (47.1)
x dt t =c

Given the chart (U , xˆ ) covering x , we can project the vector equation (47.1) into the natural basis
{gi } of x̂ . First, the coordinates of the curve λ are given by

xˆ ( λ (t ) ) = ( λ 1 (t ),..., λ N (t ) ) , λ ( t ) = x ( λ 1 (t ),..., λ N (t ) ) (47.2)

for all t such that λ ( t ) ∈U . Differentiating (47.2)2 with respect to t , we get

∂x d λ j
λ = j (47.3)
x ∂x dt t =c

By (47.3) this equation can be rewritten as

dλ j
λ = g j (x) (47.4)
x dt t =c

Thus, the components of λ relative to {g j } are simply the derivatives of the coordinate
representations of λ in x̂ . In fact (44.33) can be regarded as a special case of (47.3) when λ
coincides with the ith coordinate curve of x̂ .

An important consequence of (47.3) is that

340 Chap. 9 • EUCLIDEAN MANIFOLDS

xˆ ( λ (t + Δt ) ) = ( λ 1 (t ) + υ 1Δt + o( Δt ),..., λ N (t ) + υ N Δt + o( Δt ) ) (47.5)

where υ i denotes a component of λ , i.e.,

υ i = d λ i dt (47.6)

In particular, if λ is a straight line segment, say

λ (t ) = x + tv (47.7)

so that

λ (t ) = v (47.8)

for all t , then (47.5) becomes

xˆ ( λ (t ) ) = ( x1 + υ 1t ,..., x N + υ N t ) + o(t ) (47.9)

for sufficiently small t .

Next we consider the gradient of a smooth function f defined on an open set U ⊂ E .

From (43.17) the gradient of f is defined by

d
grad f (x) ⋅ v = f ( x + τ v ) τ =0 (47.10)
dτ

for all v ∈V . As before, we choose a chart near x ; then f can be represented by the function

f ( x1 ,..., x N ) ≡ f x( x1 ,..., x N ) (47.11)

Sec. 47 • Christoffel Symbols, Covariant Differentiation 341

For definiteness, we call the function f x the coordinate representation of f . From (47.9), we
see that

f ( x + τ v ) = f ( x1 + υ 1τ + o(τ ),..., x N + υ Nτ + o(τ )) (47.12)

As a result, the right-hand side of (47.10) is given by

d ∂f ( x ) j
f ( x + τ v ) τ =0 = υ (47.13)
dτ ∂x j

Now since v is arbitrary, (47.13) can be combined with (47.10) to obtain

∂f j
grad f = g (47.14)
∂x j

where g j is a natural basis vector associated with the coordinate chart as defined by (44.31). In
fact, that equation can now be regarded as a special case of (47.14) where f reduces to the
coordinate function xˆ i .

Having considered the tangent of a curve and the gradient of a function, we now turn to the
problem of representing the gradient of a tensor field in general. Let A ∈ Tq∞ (U ) be such a field
and suppose that x is an arbitrary point in its domain U . We choose an arbitrary chart x̂ covering
x . Then the formula generalizing (47.4) and (47.14) is

∂A(x)
grad A(x) = ⊗ g j ( x) (47.15)
∂x j

where the quantity ∂A(x) ∂x j on the right-hand side is the partial derivative of the coordinate
representation of A , i.e.,

∂A(x) ∂
= j A x( x1 ,..., x N ) (47.16)
∂x j
∂x

From (47.15) we see that grad A is a tensor field of order q + 1 on U , grad A ∈ Tq∞+1 (U ) .
342 Chap. 9 • EUCLIDEAN MANIFOLDS

To prove (47.15) we shall first regard grad A(x) as in L (V ;T q (V )) . Then by (43.15),

when A is smooth, we get

d
( grad A(x) ) v = A ( x + τ v ) τ =0 (47.17)
dτ

for all v ∈V . By using exactly the same argument from (47.10) to (47.13), we now have

∂A( x ) j
( grad A(x ) ) v = υ (47.18)
∂x j

Since this equation must hold for all v ∈V , we may take v = g k (x) and find

∂A(x)
( grad A(x) ) g k = (47.19)
∂x k

which is equivalent to (47.15) by virtue of the canonical isomorphism between L (V ;T q (V )) and

T q+1 (V ) .

Since grad A(x) ∈ T q+1 (V ) , it can be represented by its component form relative to the
natural basis, say

grad A(x) = ( grad A(x) ) 1 g i1 (x) ⋅⋅⋅ g iq (x) ⊗ g j (x)

i ...iq
j
(47.20)

Comparing this equation with (47.15), we se that

∂A(x)
( grad A(x) ) g i1 (x) ⋅⋅⋅ g iq (x) =
i1 ...iq
(47.21)
j
∂x j
Sec. 47 • Christoffel Symbols, Covariant Differentiation 343

for all j = 1,..., N . In applications it is convenient to express the components of grad A (x) in terms
of the components of A (x) relative to the same coordinate chart. If we write the component form
of A (x) as usual by

A(x) = A 1 q (x)g i1 (x) ⊗ ⋅⋅⋅ ⊗ g iq (x)

i ...i
(47.22)

then the right-hand side of (47.21) is given by

∂A(x) ∂A 1 q (x)
i ...i

= g i1 (x) ⊗ ⋅⋅⋅ ⊗ g iq (x)

∂x j ∂x j
(47.23)
⎡ ∂g i (x) ∂g i (x) ⎤
+ A 1 q (x) ⎢ 1 j ⊗ ⋅⋅⋅ ⊗ g iq (x) + g i1 (x) ⊗ ⋅⋅⋅ ⊗ q j ⎥
i ...i

⎣ ∂x ∂x ⎦

From this representation we see that it is important to express the gradient of the basis vector g i in
component form first, since from (47.21) for the case A = g i , we have

∂g i (x)
= ( grad g i (x) ) j g k (x)
k
(47.24)
∂x j

or, equivalently,

∂g i ( x ) ⎧ k ⎫
= ⎨ ⎬ g k (x) (47.25)
∂x j ⎩ij ⎭

⎧k ⎫ ⎧k ⎫
We call ⎨ ⎬ the Christoffel symbol associated with the chart x̂ . Notice that, in general, ⎨ ⎬ is a
⎩ij ⎭ ⎩ij ⎭
function of x , but we have suppressed the argument x in the notation. More accurately, (47.25)
should be replaced by the field equation

∂g i ⎧ k ⎫
= ⎨ ⎬ gk (47.26)
∂x j ⎩ij ⎭
344 Chap. 9 • EUCLIDEAN MANIFOLDS

which is valid at each point x in the domain of the chart x̂ , for all i, j = 1,..., N . We shall consider
some preliminary results about the Christoffel symbols first.

From (47.26) and (44.35) we have

⎧ k ⎫ ∂g i k
⎨ ⎬ = j ⋅g (47.27)
⎩ij ⎭ ∂x

By virtue of (44.33), this equation can be rewritten as

⎧k ⎫ ∂ 2x
⎨ ⎬ = i j ⋅g
k
(47.28)
⎩ ⎭
ij ∂x ∂x

It follows from (47.28) that the Christoffel symbols are symmetric in the pair (ij ) , namely

⎧k ⎫ ⎧ k ⎫
⎨ ⎬=⎨ ⎬ (47.29)
⎩ij ⎭ ⎩ ji ⎭

for all i, j , k = 1,..., N . Now by definition

gij = g i ⋅ g j (47.30)

Taking the partial derivative of (47.30) with respect to x k and using the component form (47.26),
we get

∂gij ∂g i ∂g
= ⋅ g j + g i ⋅ kj (47.31)
∂x k
∂x k
∂x

When the symmetry property (47.29) is used, equation (47.31) can be solved for the Christoffel
symbols:
Sec. 47 • Christoffel Symbols, Covariant Differentiation 345

⎧ k ⎫ 1 kl ⎛ ∂gil ∂g jl ∂gij ⎞
⎨ ⎬= g ⎜ j + i − l ⎟ (47.32)
⎩ij ⎭ 2 ⎝ ∂x ∂x ∂x ⎠

where g kl denotes the contravariant components of the metric tensor, namely

g kl = g k ⋅ g l (47.33)

The formula (47.32) is most convenient for the calculation of the Christoffel symbols in any given
chart.

As an example, we now compute the Christoffel symbols for the cylindrical coordinate
system in a three-dimensional space. In Section 44 we have shown that the components of the
metric tensor are given by (44.50) and (44.51) relative to this coordinate system. Substituting those
components into (47.32), we obtain

⎧ 1⎫ ⎧ 2⎫ ⎧ 2⎫ 1
⎨ ⎬ = −x , ⎨ ⎬=⎨ ⎬= 1
1
(47.34)
⎩22 ⎭ ⎩12 ⎭ ⎩12 ⎭ x

and all other Christoffel symbols are equal to zero.

Given two charts x̂ and ŷ with natural bases {g i } and {hi } , respectively, the
transformation rule for the Christoffel symbols can be derived in the following way:

⎧k ⎫ k ∂g i ∂x k l ∂ ⎛ ∂y s ⎞
⎨ ⎬ = g ⋅ = h ⋅ j ⎜ i hs ⎟
⎩ij ⎭ ∂x j ∂y l ∂x ⎝ ∂x ⎠
∂x k l ⎛ ∂ 2 y s ∂y s ∂h s ∂y l ⎞
= h ⋅ ⎜ j i s h + ⎟
∂y l ⎝ ∂x ∂x ∂x i ∂y l ∂x j ⎠
(47.35)
∂x k ∂ 2 y l ∂x k ∂y s ∂y l l ∂h s
= l + h ⋅ l
∂y ∂x j ∂x i ∂y l ∂x i ∂x j ∂y
∂x k ∂ 2 y l ∂x k ∂y s ∂y l ⎧ l⎫
= + ⎨ ⎬
∂y l ∂x j ∂x i ∂y l ∂x i ∂x j ⎩ st ⎭
346 Chap. 9 • EUCLIDEAN MANIFOLDS

⎧k ⎫ ⎧ l⎫
where ⎨ ⎬ and ⎨ ⎬ denote the Christoffel symbols associated with x̂ and ŷ , respectively. Since
⎩ij ⎭ ⎩ st ⎭
(47.35) is different from the tensor transformation rule (45.22), it follows that the Christoffel
symbols are not the components of a particular tensor field. In fact, if ŷ is a Cartesian chart, then
⎧l⎫
⎨ ⎬ vanishes since the natural basis vectors h s , are constant. In that case (47.35) reduces to
⎩ st ⎭

⎧ k ⎫ ∂x k ∂ 2 y l
⎨ ⎬= l (47.36)
⎩ij ⎭ ∂y ∂x ∂x
j i

⎧k ⎫
and ⎨ ⎬ need not vanish unless x̂ is also a Cartesian chart. The formula (47.36) can also be used
⎩ij ⎭
to calculate the Christoffel symbols when the coordination transformation from x̂ to a Cartesian
system ŷ is given.

Having presented some basis properties of the Christoffel symbols, we now return to the
general formula (47.23) for the components of the gradient of a tensor field. Substituting (47.26)
into (47.23) yields

∂A( x ) ⎡ ∂A 1 q ( x ) ki ...i ⎧ i ⎫ i ...i k ⎧ iq ⎫ ⎤

i ...i

=⎢ + A 2 q ⎨ 1 ⎬ + ⋅⋅⋅ + A 1 q −1 ⎨ ⎬⎥ g i1 ( x ) ⊗ ⋅⋅⋅ ⊗ g iq ( x ) (47.37)

∂x j
⎢⎣ ∂x
j
⎩kj ⎭ ⎩kj ⎭⎥⎦

Comparing this result with (47.21), we finally obtain

∂A 1 q ⎧ i1 ⎫ i1 ...iq −1k ⎧ iq ⎫
i ...i

A 1 q , j ≡ ( grad A ) 1 = +A 2 q ⎨ ⎬ + ⋅⋅⋅ + A
i ...i i ...iq ki ...i
⎨ ⎬ (47.38)
j
∂x j
⎩kj ⎭ ⎩ kj ⎭

i ...i
This particular formula gives the components A 1 q , j of the gradient of A in terms of the
i ...iq
contravariant components A 1 of A . If the mixed components of A are used, the formula
becomes
Sec. 47 • Christoffel Symbols, Covariant Differentiation 347

∂Ai1 ...ir j1 ... js ⎧i⎫ ⎧ ir ⎫

, ≡ ( grad A ) = + Ali2 ...ir j1 ... js ⎨ 1 ⎬ + ⋅⋅⋅ + Ai1 ...ir −1l j1 ... js
i1 ...ir i1 ...ir
A ⎨ ⎬
∂x
j1 ... js k j1 ... js ,k k
⎩lk ⎭ ⎩lk ⎭
(47.39)
⎧l ⎫ ⎧l ⎫
−Ai1 ...ir
lj2 ... js ⎨ ⎬ − ⋅⋅⋅ − A 1 r j1 ... js −1l ⎨ ⎬
i ...i

⎩ j1k ⎭ ⎩ js k ⎭

We leave the proof of this general formula as an exercise. From (47.39), if A ∈ Tq∞ (U ) , where
q = r + s , then grad A ∈ Tq∞+1 (U ) . Further, if the coordinate system is Cartesian, then Ai1 ...ir j1 ... js ,k
reduces to the ordinary partial derivative of Ai1 ...ir j1 ... js wih respect to x k .

Some special cases of (47.39) should be noted here. First, since the metric tensor is a
constant second-order tensor field, we have

gij ,k = g ij ,k = δ ij ,k = 0 (47.40)

for all i, l , k = 1,..., N . In fact, (47.40) is equivalent to (47.31), which we have used to obtain the
formula (47.32) for the Christoffel symbols. An important consequence of (47.40) is that the
operations of raising and lowering of indices commute with the operation of gradient or covariant
differentiation.

Another constant tensor field on E is the tensor field E defined by (45.26). While the sign
of E depends on the orientation, we always have

Ei1 ...iN ,k = E i1 ...iN ,k = 0 (47.41)

If we substitute (45.27) and (45.28) into (47.41), we can rewrite the result in the form

1 ∂ g ⎧ k⎫
=⎨ ⎬ (47.42)
g ∂x
i
⎩lk ⎭

where g is the determinant of ⎡⎣ gij ⎤⎦ as defined by (44.45).

348 Chap. 9 • EUCLIDEAN MANIFOLDS

Finally, the operation of skew-symmetrization K r introduced in Section 37 is also a

constant tensor. So we have

δ ij ......ij ,k = 0
1
1
r
r
(47.43)

in any coordinate system. Thus, the operations of skew-symmetrization and covariant

differentiation commute, provided that the indices of the covariant differentiations are not affected
by the skew-symmetrization.

Some classical differential operators can be derived from the gradient. First, if A is a
tensor field of order q ≥ 1 , then the divergence of A is defined by

div A = Cq ,q +1 ( grad A )

where C denotes the contraction operation. In component form we have

div A = A 1 ,k g i1 ⊗ ⋅⋅⋅ ⊗ g iq −1
i ...iq −1k
(47.44)

so that div A is a tensor field of order q − 1 . In particular, for a vector field v , (47.44) reduces to

div v = υ i ,i = g ijυi , j (47.45)

By use of (47.42) and (47.40), we an rewrite this formula as

div v =
1 ∂ gυ
i
( ) (47.46)
g ∂x i

This result is useful since it does not depend on the Christoffel symbols explicitly.

The Laplacian of a tensor field of order q + 1 is a tensor field of the same order q defined
by
Sec. 47 • Christoffel Symbols, Covariant Differentiation 349

ΔA = div ( grad A ) (47.47)

In component form we have

ΔA = g kl A 1 q ,kl g i1 ⊗ ⋅⋅⋅ ⊗ g iq
i ...i
(47.48)

For a scalar field f the Laplacian is given by

1 ∂ ⎛ ∂f ⎞
Δf = g kl f ,kl = l ⎜
g g kl k ⎟ (47.49)
g ∂x ⎝ ∂x ⎠

where (47.42), (47.14) and (47.39) have been used. Like (47.46), the formula (47.49) does not
i ...i
depend explicitly on the Christoffel symbols. In (47.48), A 1 q ,kl denotes the components of the
second gradient grad ( grad A ) of A . The reader will verify easily that A 1 q ,kl , like the ordinary
i ...i

second partial derivative, is symmetric in the pair (k , l ) . Indeed, if the coordinate system is
Cartesian, then A 1 q ,kl reduces to ∂ 2 A 1 ∂x k ∂x l .
i ...i i ...iq

Finally, the classical curl operator can be defined in the following way. If v is a vector
field, then curl v is a skew-symmetric second-order tensor field defined by

curl v ≡ K 2 ( grad v ) (47.50)

where K 2 is the skew-symmetrization operator. In component form (47.50) becomes

curl v =
1
2
(υi , j −υ j ,i ) g i ⊗ g j (47.51)

By virtue of (47.29) and (47.39) this formula can be rewritten as

1 ⎛ ∂υi ∂υ j ⎞ i
curl v = ⎜ − i ⎟g ⊗ g j (47.52)
2 ⎝ ∂x j
∂x ⎠
350 Chap. 9 • EUCLIDEAN MANIFOLDS

which no longer depends on the Christoffel symbols. We shall generalize the curl operator to
arbitrary skew-symmetric tensor fields in the next chapter.

Exercises

47.1 Show that in spherical coordinates on a three-dimensional Euclidean manifold the nonzero
Christoffel symbols are

⎧ 2 ⎫ ⎧ 2 ⎫ ⎧ 3 ⎫ ⎧ 3⎫ 1
⎨ ⎬=⎨ ⎬=⎨ ⎬=⎨ ⎬= 1
⎩ 21⎭ ⎩12 ⎭ ⎩31⎭ ⎩13⎭ x
⎧1⎫
⎨ ⎬ = −x
1

⎩ 22 ⎭
⎧1⎫
⎨ ⎬ = − x (sin x )
1 2 2

⎩33⎭
⎧ 3⎫ ⎧ 3⎫
⎨ ⎬ = ⎨ ⎬ = cot x
2

⎩32 ⎭ ⎩ 23⎭
⎧ 2⎫
⎨ ⎬ = − sin x cos x
2 2

⎩33⎭

47.2 On an oriented three-dimensional Euclidean manifold the curl of a vector field can be
regarded as a vector field by

∂υ j
curl v ≡ − E ijkυ j ,k g i = − E ijk gi (47.53)
∂x k

where E ijk denotes the components of the positive volume tensor E . Show that
curl ( curl v ) = grad ( div v ) − div ( grad v ) . Also show that

curl ( grad f ) = 0 (47.54)

For any scalar field f and that

Sec. 47 • Christoffel Symbols, Covariant Differentiation 351

div ( curl v ) = 0 (47.55)

for any vector field v .

47.3 Verify that

⎧k ⎫ ∂g k
⎨ ⎬ = − ⋅ gi (47.56)
⎩ij ⎭ ∂x j

Therefore,

∂g k ⎧k ⎫
= − ⎨ ⎬ gi (47.57)
∂x j
⎩ij ⎭

47.4 Prove the formula (47.37).

47.5 Prove the formula (47.42) and show that

1 ∂ g
div g j = (47.58)
g ∂x
j

47.6 Show that for an orthogonal coordinate system

⎧k ⎫
⎨ ⎬ = 0, i ≠ j, i ≠ k , j ≠ k
⎩ij ⎭
⎧ j⎫ 1 ∂gii
⎨ ⎬=− if i≠ j
⎩ii ⎭ 2 g jj ∂x j
⎧ i⎫ ⎧ i ⎫ 1 ∂gii
⎨ ⎬=⎨ ⎬=− if i≠ j
⎩ij ⎭ ⎩ ji ⎭ 2 gii ∂x j
⎧ i⎫ 1 ∂gii
⎨ ⎬=
⎩ii ⎭ 2 gii ∂x
i

Where the indices i and j are not summed.

352 Chap. 9 • EUCLIDEAN MANIFOLDS

47.7 Show that

∂ ⎧k ⎫ ∂ ⎧ k ⎫ ⎧t ⎫ ⎧ k ⎫ ⎧t ⎫ ⎧ k ⎫
⎨ ⎬− l ⎨ ⎬+ ⎨ ⎬⎨ ⎬− ⎨ ⎬⎨ ⎬ = 0
∂x j ⎩il ⎭ ∂x ⎩ij ⎭ ⎩il ⎭ ⎩tj ⎭ ⎩ij ⎭ ⎩tl ⎭

The quantity on the left-hand side of this equation is the component of a fourth-order tensor
R , called the curvature tensor, which is zero for any Euclidean manifold2.

2
The Maple computer program contains a package tensor that will produce Christoffel symbols and other important
tensor quantities associated with various coordinate systems.
Sec. 48 • Covariant Derivatives along Curves 353

Section 48. Covariant Derivatives along Curves

In the preceding section we have considered covariant differentiation of tensor fields which
are defined on open submanifolds in E . In applications, however, we often encounter vector or
tensor fields defined only on some smooth curve in E . For example, if λ : (a , b) → E is a smooth
curve, then the tangent vector λ is a vector field on E . In this section we shall consider the
problem of representing the gradients of arbitrary tensor fields defined on smooth curves in a
Euclidean space.

Given any smooth curve λ : (a, b) → E and a field A : (a , b) → T q (V ) we can regard the
value A(t ) as a tensor of order q at λ (t ) . Then the gradient or covariant derivative of A along λ
is defined by

d A( t ) A(t + Δt ) − A(t )
≡ lim (48.1)
dt Δ t → 0 Δt

for all t ∈ (a, b) . If the limit on the right-hand side of (48.1) exists, then d A(t ) dt is itself also a
tensor field of order q on λ . Hence, we can define the second gradient d 2 A(t ) dt 2 by replacing
A by d A(t ) dt in (48.1). Higher gradients of A are defined similarly. If all gradients of A
exist, then A is C ∞ -smooth on λ . We are interested in representing the gradients of A in
component form.

Let x̂ be a coordinate system covering some point of λ . Then as before we can

characterize λ by its coordinate representation ( λ i (t ), i = 1,..., N ) . Similarly, we can express A in
component form

A( t ) = A 1 (t )g i1 ( λ (t )) ⊗ ⋅⋅⋅ ⊗ g iq ( λ (t ))
i ...iq
(48.2)

where the product basis is that of x̂ at λ (t ) , the point where A(t ) is defined. Differentiating
(48.2) with respect to t , we obtain
354 Chap. 9 • EUCLIDEAN MANIFOLDS

i ...i
d A(t ) dA 1 q (t )
= g i1 ( λ (t )) ⊗ ⋅⋅⋅ ⊗ g iq ( λ (t ))
dt dt
(48.3)
⎡ dg i ( λ (t )) dg i ( λ (t )) ⎤
+ A 1 q (t ) ⎢ 1 ⊗ ⋅⋅⋅ ⊗ g iq ( λ (t )) + ⋅⋅⋅ + g i1 ( λ (t )) ⊗ ⋅⋅⋅ ⊗ q
i ...i
⎥
⎣ dt dt ⎦

By application of the chain rule, we obtain

dg i ( λ (t )) ∂g i ( λ (t )) d λ j (t )
= (48.4)
dt ∂x j dt

where (47.5) and (47.6) have been used. In the preceding section we have represented the partial
derivative of g i by the component form (47.26). Hence we can rewrite (48.4) as

dg i ( λ (t )) ⎧k ⎫ d λ j (t )
=⎨ ⎬ g k ( λ (t )) (48.5)
dt ⎩ij ⎭ dt

Substitution of (48.5) into (48.3) yields the desired component representation

d A(t ) ⎧⎪ dA 1 q (t ) ⎡ ki2 ...iq ⎧ i1 ⎫ ⎧ iq ⎫⎤ d λ j (t ) ⎫⎪

i ...i

=⎨ + ⎢A (t ) ⎨ ⎬ + ⋅⋅⋅ + A 1 q −1 (t ) ⎨ ⎬⎥
i ...i k
⎬
dt ⎪⎩ dt ⎢⎣ ⎩kj ⎭ ⎩kj ⎭⎦⎥ dt ⎪⎭ (48.6)
×g i1 ( λ (t )) ⊗ ⋅⋅⋅ ⊗ g iq ( λ (t ))

where the Christoffel symbols are evaluated at the position λ (t ) . The representation (48.6) gives
the contravariant components ( d A(t ) dt ) 1
i ...iq
of d A(t ) dt in terms of the contravariant components
i ...i
A 1 q (t ) of A(t ) relative to the same coordinate chart x̂ . If the mixed components are used, the
representation becomes

d A(t ) ⎪⎧ dA 1 r j1 ... js (t ) ⎡ ki2 ...ir ⎧ i1 ⎫ ⎧ ir ⎫

i ...i

=⎨ + ⎢A j1 ... js (t ) ⎨ ⎬ + ⋅⋅⋅ + A 1 r −1 j1 ... js (t ) ⎨ ⎬

i ...i k

dt ⎪⎩ dt ⎣ ⎩kl ⎭ ⎩kl ⎭
⎧k⎫ ⎧ k ⎫ ⎤ d λ l (t ) ⎫
− A kj2 ... js (t ) ⎨ ⎬ − ⋅⋅⋅ − A j1 ... js−1k (t ) ⎨ ⎬ ⎥
i1 ...ir i1 ...ir
⎬ (48.7)
⎩ j1l ⎭ ⎩ js l ⎭ ⎦ dt ⎭
×g i1 (λ (t )) ⊗ ⋅⋅ ⋅ ⊗ g ir (λ (t )) ⊗ g j1 (λ (t )) ⊗ ⋅⋅⋅ ⊗ g js (λ (t ))
Sec. 48 • Covariant Derivatives along Curves 355

We leave the proof of this general formula as an exercise. In view of the representations (48.6) and
(48.7), we see that it is important to distinguish the notation

i1 ...ir
⎛ dA ⎞
⎜ ⎟
⎝ dt ⎠ j1 ... js

which denotes a component of the covariant derivative d A dt , from the notation

dAi1 ...ir j1 ... js

dt

which denotes the derivative of a component of A . For this reason we shall denote the former by
the new notation

DAi1 ...ir j1 ... js

Dt

As an example we compute the components of the gradient of a vector field v along λ . By

(48.6)

dv (t ) ⎧ dυ i (t ) ⎧ i ⎫ d λ j (t ) ⎫
=⎨ + υ (t ) ⎨ ⎬
k
⎬ g i ( λ (t )) (48.8)
dt ⎩ dt ⎩kj ⎭ dt ⎭
or, equivalently,

Dυ i (t ) dυ i (t ) ⎧ i ⎫ d λ j (t )
= + υ k (t ) ⎨ ⎬ (48.9)
Dt dt ⎩kj ⎭ dt

In particular, when v is the ith basis vector g i and then λ is the jth coordinate curve, then (48.8)
reduces to
356 Chap. 9 • EUCLIDEAN MANIFOLDS

∂g i ⎧k ⎫
= ⎨ ⎬ gk
∂x j ⎩ij ⎭

which is the previous equation (47.26).

Next if v is the tangent vector λ of λ , then (48.8) reduces to

d 2 λ (t ) ⎧ d 2 λ i (t ) d λ k (t ) ⎧ i ⎫ d λ j (t ) ⎫
=⎨ + ⎨ ⎬ ⎬ g i (λ (t )) (48.10)
dt 2 ⎩ dt
2
dt ⎩kj ⎭ dt ⎭

where (47.4) has been used. In particular, if λ is a straight line with homogeneous parameter, i.e.,
if λ = v = const , then

d 2 λ i (t ) d λ k (t ) d λ j (t ) ⎧ i ⎫
+ ⎨ ⎬ = 0, i = 1,..., N (48.11)
dt 2 dt dt ⎩kj ⎭

This equation shows that, for the straight line λ (t ) = x + tv given by (47.7), we can sharpen the
result (47.9) to

⎛ 1 ⎧ 1⎫ 1 ⎧ N⎫ ⎞
xˆ ( λ (t ) ) = ⎜ x1 + υ 1t − υ kυ j ⎨ ⎬ t 2 ,..., x N + υ N t − υ kυ j ⎨ ⎬ t 2 ⎟ + o(t 2 ) (48.12)
⎝ 2 ⎩ kj ⎭ 2 ⎩ kj ⎭ ⎠

for sufficiently small t , where the Christoffel symbols are evaluated at the point x = λ (0) .

Now suppose that A is a tensor field on U , say A ∈ Tq∞ , and let λ : (a, b) → U be a curve
in U . Then the restriction of A on λ is a tensor field

ˆ (t ) ≡ A (λ (t )),
A t ∈ ( a, b) (48.13)

In this case we can compute the gradient of A along λ either by (48.6) or directly by the chain rule
of (48.13). In both cases the result is
Sec. 48 • Covariant Derivatives along Curves 357

d A(λ (t )) ⎡ ∂A 1 q (λ (t )) ⎧ i1 ⎫ ⎧ iq ⎫⎤ d λ j (t )
i ...i

=⎢ + + ⋅⋅⋅ +
ki2 ...iq i1 ...iq −1k
A (t ) ⎨ ⎬ A (t ) ⎨ ⎬⎥
dt ⎣⎢ ∂x j ⎩kj ⎭ ⎩kj ⎭⎥⎦ dt (48.14)
×g i1 (λ (t )) ⊗ ⋅⋅⋅ ⊗ g iq (λ (t ))

or, equivalently,

d A(λ (t ))
= ( grad A(λ (t )) ) λ (t ) (48.15)
dt

where grad A(λ (t )) is regarded as an element of L (V ;T q (V )) as before. Since grad A is given

by (47.15), the result (48.15) also can be written as

d A(λ (t )) ∂A(λ (t )) d λ i (t )
= (48.16)
dt ∂x j dt

A special case of (48.16), when A reduces to g i , is (48.4).

By (48.16) it follows that the gradient of the metric tensor, the volume tensor, and the skew-
symmetrization operator all vanish along any curve.

Exercises

48.1 Prove the formula (48.7).

48.2 If the parameter t of a curve λ is regarded as time, then the tangent vector

dλ
v=
dt

is the velocity and the gradient of v

358 Chap. 9 • EUCLIDEAN MANIFOLDS

dv
a=
dt

is the acceleration. For a curve λ in a three-dimensional Euclidean manifold, express the

acceleration in component form relative to the spherical coordinates.
______________________________________________________________________________
Chapter 10

VECTOR FIELDS AND DIFFERENTIAL FORMS

Section 49. Lie Derivatives

Let E be a Euclidean manifold and let u and v be vector fields defined on some open set U in E.
In Exercise 45.2 we have defined the Lie bracket [ u, v ] by

[ u, v ] = ( grad v ) u − ( grad u ) v (49.1)

In this section first we shall explain the geometric meaning of the formula (49.1), and then we
generalize the operation to the Lie derivative of arbitrary tensor fields.

To interpret the operation on the right-hand side of (49.1), we start from the concept of
the flow generated by a vector field. We say that a curve λ : ( a , b ) → U is an integral curve of a
vector field v if

dλ ( t )
= v ( λ (t )) (49.2)
dt

for all t ∈ ( a , b ) . By (47.1), the condition (49.2) means that λ is an integral curve of v if and
only if its tangent vector coincides with the value of v at every point λ ( t ) . An integral curve
may be visualized as the orbit of a point flowing with velocity v. Then the flow generated by v
is defined to be the mapping that sends a point λ ( t0 ) to a point λ ( t ) along any integral curve of
v.

To make this concept more precise, let us introduce a local coordinate system x̂ . Then
the condition (49.2) can be represented by

d λ i ( t ) / dt =υ i ( λ ( t ) ) (49.3)

where (47.4) has been used. This formula shows that the coordinates ( λ i ( t ) , i = 1,… , N ) of an
integral curve are governed by a system of first-order differential equations. Now it is proved in
the theory of differential equations that if the fields υ i on the right-hand side of (49.3) are
smooth, then corresponding to any initial condition, say

359
360 Chap. 10 • VECTOR FIELDS

λ ( 0) = x 0 (49.4)

or equivalently

λ i ( 0 ) = x0i , i = 1,… , N (49.5)

a unique solution of (49.3) exists on a certain interval ( − δ , δ ) , where δ may depend on the
initial point x 0 but it may be chosen to be a fixed, positive number for all initial points in a
sufficiently small neighborhood of x 0 . For definiteness, we denote the solution of (49.2)
corresponding to the initial point x by λ ( t , x ) ; then it is known that the mapping from x to
λ ( t , x ) ; then it is known that the mapping from x to λ ( t , x ) is smooth for each t belonging to
the interval of existence of the solution. We denote this mapping by ρt , namely

ρt ( x ) = λ ( t , x ) (49.6)

and we call it the flow generated by v . In particular,

ρ 0 ( x ) = x, x ∈U (49.7)

reflecting the fact that x is the initial point of the integral curve λ ( t , x ) .

Since the fields υ i are independent of t, they system (49.3) is said to be autonomous. A
characteristic property of such a system is that the flow generated by v forms a local one-
parameter group. That is, locally,

ρt1 +t2 = ρt 1 ρt2 (49.8)

or equivalently

λ ( t1 + t2 , x ) = λ ( t2 , λ ( t1 , x ) ) (49.9)

for all t1 , t2 , x such that the mappings in (49.9) are defined. Combining (49.7) with (49.9), we see
that the flow ρt is a local diffeomorphism and, locally,

ρt−1 = ρ − t (49.10)
Sec. 49 • Lie Derivatives 361

Consequently the gradient, grad ρt , is a linear isomorphism which carries a vector at any point
x to a vector at the point ρt ( x ) . We call this linear isomorphism the parallelism generated by
the flow and, for brevity, denote it by Pt . The parallelism Pt is a typical two-point tensor whose
component representation has the form

Pt ( x ) = Pt ( x ) j g i ( ρt ( x ) ) ⊗ g j ( x )
i
(49.11)

or, equivalently,

⎡⎣ Pt ( x ) ⎤⎦ ( g j ( x ) ) = Pt ( x ) j g i ( ρt ( x ) )
i
(49.12)

{ }
where {g i ( x )} and g j ( ρt ( x ) ) are the natural bases of x̂ at the points x and ρt ( x ) ,
respectively.

Now the parallelism generated by the flow of v gives rise to a difference quotient of the
vector field u in the following way: At any point x ∈U , Pt ( x ) carries the vector u ( x ) at x to
the vector ( Pt ( x ) ) ( u ( x ) ) at ρt ( x ) , which can then be compared with the vector u ( ρt ( x ) ) also at
ρt ( x ) . Thus we define the difference quotient

1
⎡ u ( ρt ( x ) ) − ( Pt ( x ) ) ( u ( x ) )⎤⎦ (49.13)
t⎣

The limit of this difference quotient as t approaches zero is called the Lie derivative of u with
respect to v and is denoted by

1
L u ( x ) ≡ lim ⎡⎣ u ( ρt ( x ) ) − ( Pt ( x ) ) ( u ( x ) ) ⎤⎦ (49.14)
v t →0 t

We now derive a representation for the Lie derivative in terms of a local coordinate
system xˆ. In view of (49.14) we see that we need an approximate representation for Pt to within
first-order terms in t. Let x 0 be a particular reference point. From (49.3) we have

λ i ( t , x 0 ) = x0i + υ i ( x 0 ) t + o ( t ) (49.15)

Suppose that x is an arbitrary neighboring point of x 0 , say x i = x0i + Δx i . Then by the same
argument as (49.15) we have also
362 Chap. 10 • VECTOR FIELDS

λ i ( t, x ) = xi +υ i ( x ) t + o ( t )
∂υ i ( x 0 ) (49.16)
= x0i + Δx i + υ i ( x 0 ) t +
∂x j
( Δx j ) t + o ( Δ x k ) + o ( t )

Subtracting (49.15) from (49.16), we obtain

∂υ i ( x 0 )
λ i ( t , x ) − λ i ( t , x 0 ) = Δx i +
∂x j
( Δ x j ) t + o ( Δx k ) + o ( t ) (49.17)

It follows from (49.17) that

∂υ i ( x 0 )
Pt ( x 0 ) j ≡ ( grad ρt ( x o ) ) j = δ ij + t + 0 (t )
i i
(49.18)
∂x j

Now assuming that u is also smooth, we can represent u ( ρt ( x 0 ) ) approximately by

∂u i ( x 0 ) j
u i ( ρt ( x 0 ) ) = u i ( x 0 ) + υ ( x0 ) t + o (t ) (49.19)
∂x j

where (49.15) has been used. Substituting (49.18) and (49.19) into (49.14) and taking the limit,
we finally obtain

⎡ ∂u i ( x 0 ) j ∂υ i ( x 0 ) j ⎤
L u ( x0 ) = ⎢ υ − u ( x0 )⎥ gi ( x0 ) (49.20)
⎣ ∂x ∂x
v j j
⎦

where x 0 is arbitrary. Thus the field equation for (49.20) is

⎡ ∂u i j ∂υ i j ⎤
L u = ⎢ j υ − j u ⎥ gi (49.21)
v
⎣ ∂x ∂x ⎦

By (47.39) or by using a Cartesian system, we can rewrite (49.21) as

L u = ( ui , j υ j − υ i , j u j ) gi (49.22)
v

Consequently the Lie derivative has the following coordinate-free representation:

L u = ( grad u ) v − ( grad v ) u (49.23)

v
Sec. 49 • Lie Derivatives 363

Comparing (49.23) with (49.1), we obtain

L u = [ v, u ] (49.24)
v

Since the Lie derivative is defined by the limit of the difference quotient (49.13),
L u vanishes if and only if u commutes with the flow of v in the following sense:
v

u ρt = Pu
i (49.25)

When u satisfies this condition, it may be called invariant with respect to v . Clearly, this
condition is symmetric for the pair ( u, v ) , since the Lie bracket is skew-symmetric, namely

[ u, v ] = − [ v, u] (49.26)

In particular, (49.25) is equivalent to

v ϕ t = Qt v (49.27)

where ϕ t and Qt denote the flow and the parallelism generated by u .

Another condition equivalent to (49.25) and (49.27) is

ϕ s ρt = ρt ϕ s (49.28)

which means that

ρt ( an integral curve of u ) = an integral curve of u

ϕ s ( an integral curve of v ) = an integral curve of v (49.29)

To show that (49.29) is necessary and sufficient for

[ u, v ] = 0 (49.30)

we choose any particular integral curve λ : ( −δ , δ ) →U for v , At each point λ ( t ) we define an

integral curve μ ( s, t ) for u such that

μ ( 0, t ) = λ ( t ) (49.31)
364 Chap. 10 • VECTOR FIELDS

The integral curves μ (⋅, t ) with t ∈ ( −δ , δ ) then sweep out a surface parameterized by ( s, t ) .
We shall now show that (49.30) requires that the curves μ ( s, ⋅) be integral curves of v for all s.

As before, let x̂ be a local coordinate system. Then by definition

∂μ i ( s, t ) / ∂s = u i ( μ ( s, t ) ) (49.32)

and

∂μ i ( 0, t ) / ∂s = υ i ( μ ( 0, t ) ) (49.33)

and we need to prove that

∂μ i ( s, t ) / ∂s = υ i ( μ ( s, t ) ) (49.34)

for s ≠ 0. We put

∂μ i ( s, t )
ζ i ( s, t ) ≡ − υ i ( μ ( s, t ) ) (49.35)
∂t

By (49.33)

ζ i ( 0, t ) = 0 (49.36)

We now show that ζ i ( s, t ) vanishes identically. Indeed, if we differentiate (49.35) with respect
to s and use (49.32), we obtain

∂ζ i ∂ ⎛ ∂μ i ⎞ ∂υ i ∂μ j
= −
∂s ∂t ⎜⎝ ∂s ⎟⎠ ∂x j ∂s
∂u i ∂μ j ∂υ i j
= j − u (49.37)
∂x ∂t ∂x j
∂u i ⎛ ∂u i ∂υ i ⎞
= j ζ j + ⎜ j υi − j u j ⎟
∂x ⎝ ∂x ∂x ⎠

As a result, when (49.30) holds, ζ i are governed by the system of differential equations

∂ζ i ∂u i
= ζ j
(49.38)
∂s ∂x j
Sec. 49 • Lie Derivatives 365

and subject to the initial condition (49.36) for each fixed t. Consequently, ζ i = 0 is the only
solution. Conversely, when ζ i vanishes identically on the surface, (49.37) implies immediately
that the Lie bracket of u and v vanishes. Thus the condition (49.28) is shown to be equivalent to
the condition (49.30).

The result just established can be used to prove the theorem mentioned in Section 46 that
a field of basis is holonomic if and only if

⎡⎣ hi , h j ⎤⎦ = 0 (49.39)

for all i, j − 1,… , N . Necessity is obvious, since when {hi } is holonomic, the components of
each hi relative to the coordinate system corresponding to {hi } are the constants δ ij . Hence
from (49.21) we must have (49.39). Conversely, suppose (49.39) holds. Then by (49.28) there
exists a surface swept out by integral curves of the vector fields h1 and h 2 . We denote the
surface parameters by x1 and x2 . Now if we define an integral curve for h3 at each surface point
( x1 , x2 ) , then by the conditions

[ h1 , h3 ] = [h2 , h3 ] = 0

we see that the integral curves of h1 , h 2 , h3 form a three-dimensional net which can be regarded
as a “surface coordinate system” ( x1 , x2 , x3 ) on a three-dimensional hypersurface in the N-
dimensional Euclidean manifold E . By repeating the same process based on the condition
(49.39), we finally arrive at an N-dimensional net formed by integral curves of h1 ,…, h N . The
corresponding N-dimensional coordinate system x1 ,…, x N now forms a chart in E and its
natural basis is the given basis {hi } . Thus the theorem is proved. In the next section we shall
make use of this theorem to prove the classical Frobenius theorem.

So far we have considered the Lie derivative L u of a vector field u relative to v only.
v

Since the parallelism Pt generated by v is a linear isomorphism, it can be extended to tensor

fields. As usual for simple tensors we define

Pt ( a ⊗ b ⊗ ) = ( Pt a ) ⊗ ( Pt b ) ⊗ (49.40)

Then we extend Pt to arbitrary tensors by linearity. Using this extended parallelism, we define
the Lie derivative of a tensor field A with respect to v by

1
L A ( x ) ≡ lim ⎡⎣ A ( ρt ( x ) ) − ( Pt ( x ) ) ( A ( x ) ) ⎤⎦ (49.41)
v t →0 t
366 Chap. 10 • VECTOR FIELDS

which is clearly a generalization of (49.14). In terms of a coordinate system it can be shown that

( ) ∂Ai1…ir j1… js
i1…ir
L A = υk
v j1… js ∂x k

∂υ i1 ∂υ ik
− A ki2…ir j1… js − − Ai1…ir −1k j1… js (49.42)
∂x k ∂x k
∂υ k ∂υ k
+A i1…ir
k… js + A i1…ir
j1… js −1k
∂x j1 ∂x js

which generalizes the formula (49.21). We leave the proof of (49.42) as an exercise. By the
same argument as (49.22), the partial derivatives in (49.42) can be replaced by covariant
derivatives.

It should be noted that the operations of raising and lowering of indices by the Euclidean
metric do not commute with the Lie derivative, since the parallelism Pt generated by the flow
generally does not preserve the metric. Consequently, to compute the Lie derivative of a tensor
field A, we must assign a particular contravariant order and covariant order to A. The formula
(49.42) is valid when A is regarded as a tensor field of contravariant order r and covariant order
s.

By the same token, the Lie derivative of a constant tensor such as the volume tensor or
the skew-symmetric operator generally need not vanish.

Exercises

49.1 Prove the general representation formula (49.42) for the Lie derivative.
49.2 Show that the right-hand side of (49.42) obeys the transformation rule of the components
of a tensor field.
49.3 In the two-dimensional Euclidean plane E , consider the vector field v whose
components relative to a rectangular Cartesian coordinate system ( x1 , x 2 ) are

υ 1 = α x1 , υ 2 = α x2

where α is a positive constant. Determine the flow generated by v. In particular, find

the integral curve passing through the initial point

( x , x ) = (1,1)
1
0
2
0

49.4 In the same two-dimensional plane E consider the vector field such that
Sec. 49 • Lie Derivatives 367

υ1 = x2 , υ 2 = x1

Show that the flow generated by this vector field is the group of rotations of E . In
particular, show that the Euclidean metric is invariant with respect to this vector field.

49.5 Show that the flow of the autonomous system (49.3) possesses the local one-parameter
group property (49.8).
368 Chap. 10 • VECTOR FIELDS

Section 50. The Frobenius Theorem

The concept of the integral curve of a vector field has been considered in some detail in
the preceding section. In this section we introduce a somewhat more general concept. If v is a
non-vanishing vector field in E , then at each point x in the domain of v we can define a one-
dimensional Euclidean space D ( x ) spanned by the vector v ( x ) . In this sense an integral curve
of v corresponds to a one-dimensional hypersurface in E which is tangent to D ( x ) at each
point of the hypersurface. Clearly this situation can be generalized if we allow the field of
Euclidean spaces D to be multidimensional. For definiteness, we call such a field D a
distribution in E , say of dimension D. Then a D-dimensional hypersurface L in E is called an
integral surface of D if L is tangent to D at every point of the hypersurface.

Unlike a one-dimensional distribution, which corresponds to some non-vanishing vector

field and hence always possesses many integral curves, a D-dimensional distribution D in
general need not have any integral hypersurface at all. The problem of characterizing those
distributions that do possess integral hypersurfaces is answered precisely by the Frobenius
theorem. Before entering into the details of this important theorem, we introduce some
preliminary concepts first.

Let D be a D-dimensional distribution and let v be a vector field. Then v is said to be

contained in D if v ( x ) ∈D ( x ) at each x in the domain of v . Since D is D-dimensional, there
exist D vector fields {vα , α = 1,… , D} contained in D such that D ( x ) is spanned by
{vα ( x ) , α = 1,…, D} at each point x in the domain of the vector fields. We call
{vα , α = 1,…, D} a local basis for D . We say that D is smooth at some point x 0 if there exists a
local basis defined on a neighborhood of x 0 formed by smooth vector fields vα contained in D .
Naturally, D is said to be smooth if it is smooth at each point of its domain. We shall be
interested in smooth distributions only.

We say that D is integrable at a point x 0 if there exists a local coordinate system x̂

defined on a neighborhood of x 0 such that the vector fields {gα , α = 1,… , D} form a local basis
for D . Equivalently, this condition means that the hypersurfaces characterized by the conditions

x D +1 = const,… , x N = const (50.1)

are integral hypersurfaces of D . Since the natural basis vectors g i of any coordinate system are
smooth, by the very definition D must be smooth at x 0 if it is integrable there. Naturally, D is
said to be integrable if it is integrable at each point of its domain. Consequently, every
integrable distribution must be smooth.
Sec. 50 • The Frobenius Theorem 369

Now we are ready to state and prove the Frobenius theorem, which characterizes
integrable distributions.

Theorem 50.1. A smooth distribution D is integrable if and only if it is closed with respect to
the Lie bracket, i.e.,

u, v ∈D ⇒ [ u, v ] ∈D (50.2)

Proof. Necessity can be verified by direct calculation. If u, v ∈D and supposing that x̂ is a local
coordinate system such that {gα , α = 1,… , D} forms a local basis for D , then u and v have the
component forms

u = uα gα , v = υ α gα (50.3)

where the Greek index α is summed from 1 to D. Substituting (50.3) into (49.21), we see that

⎛ ∂υ α β ∂uα β ⎞
[ u, v ] = ⎜ β u − β υ ⎟ g α (50.4)
⎝ ∂x ∂x ⎠

Thus (50.2) holds.

Conversely, suppose that (50.2) holds. Then for any local basis {vα , α = 1,… , D} for D
we have

γ
⎡⎣ vα , v β ⎤⎦ = Cαβ vγ (50.5)

γ
where Cαβ are some smooth functions. We show first that there exists a local basis
{uα , α = 1,…, D} which satisfies the somewhat stronger condition

⎡⎣ uα , u β ⎤⎦ = 0, α , β = 1,…, D (50.6)

To construct such a basis, we choose a local coordinate system ŷ and represent the basis {vα }
by the component form

vα = υα1 k1 + + υαD k D + υαD +1k D +1 + + υαN k N

(50.7)
≡ υαβ k β + υαΔ k Δ
370 Chap. 10 • VECTOR FIELDS

where {k i } denotes the natural basis of ŷ , and whre the repeated Greek indices β and Δ are
summed from 1 to D and D + 1 to N , respectively. Since the local basis {vα } is linearly
independent, without loss of generality we can assume that the D × D matrix ⎡⎣υαβ ⎤⎦ is
nonsingular, namely

det ⎡⎣υαβ ⎤⎦ ≠ 0 (50.8)

Now we define the basis {uα } by

uα ≡ υ −1αβ v β , α = 1,… , D (50.9)

where ⎡⎣υ −1αβ ⎤⎦ denotes the inverse matrix of ⎡⎣υαβ ⎤⎦ , as usual. Substituting (50.7) into (50.9), we
see that the component representation of uα is

uα = δ αβ k β + uαΔ k Δ = k α + uαΔ k Δ (50.10)

We now show that the basis {uα } has the property (50.6). From (50.10), by direct
calculation based on (49.21), we can verify easily that the first D components of ⎡⎣ uα , u β ⎤⎦ are
zero, i.e., ⎡⎣ uα , u β ⎤⎦ has the representation

Δ
⎡⎣ uα , u β ⎤⎦ = Kαβ kΔ (50.11)

but, by assumption, D is closed with respect to the Lie bracket; it follows that

γ
⎡⎣ uα , u β ⎤⎦ = Kαβ uγ (50.12)

Substituting (50.10) into (50.12), we then obtain

γ Δ
⎡⎣ uα , u β ⎤⎦ = Kαβ k γ + Kαβ kΔ (50.13)

γ
Comparing this representation with (50.11), we see that Kαβ must vanish and hence, by (50.12),
(50.6) holds.

Now we claim that the local basis {uα } for D can be extended to a field of basis {ui } for
V and that
Sec. 50 • The Frobenius Theorem 371

⎡⎣ ui , u j ⎤⎦ = 0, i, j = 1,… , N (50.14)

This fact is more or less obvious. From (50.6), by the argument presented in the preceding
section, the integral curves of {uα } form a “coordinate net” on a D-dimensional hypersurface
defined in the neighborhood of any reference point x 0 . To define u D +1 , we simply choose an
arbitrary smooth curve λ ( t , x 0 ) passing through x 0 , having a nonzero tangent, and not belonging
to the hyper surface generated by {u1 ,… , u D }. We regard the points of the curve λ ( t , x 0 ) to
have constant coordinates in the coordinate net on the D-dimensional hypersurfaces, say
( xo1 ,…, xoD ) . Then we define the curves λ ( t, x ) for all neighboring points x on the hypersurface
of x o , by exactly the same condition with constant coordinates ( x1 ,… , x D ) . Thus, by definition,
the flow generated by the curves λ ( t , x ) preserves the integral curves of any uα . Hence if we
define u D +1 to be the tangent vector field of the curves λ ( t , x ) , then

[ u D+1 , uα ] = 0, α = 1,… , D (50.15)

where we have used the necessary and sufficient condition (49.29) for the condition (49.30).

Having defined the vector fields {u1 ,… , u D +1} which satisfy the conditions (50.6) and
(50.15), we repeat that same procedure and construct the fields u D + 2 , u D +3 ,… , until we arrive at a
field of basis {u1 ,… , u N }. Now from a theorem proved in the preceding section [cf. (49.39)] the
condition (50.14) is necessary and sufficient that {ui } be the natural basis of a coordinate system
xˆ. Consequently, D is integrable and the proof is complete.

From the proof of the preceding theorem it is clear that an integrable distribution D can
be characterized by the opening remark: In the neighborhood of any point x 0 in the domain of
D there exists a D-dimensional hypersurface S such that D ( x ) is the D-dimensional tangent
hyperplane of S at each point x in S .

We shall state and prove a dual version of the Frobenius theorem in Section 52.

Exercises

50.1 Let D be an integrable distribution defined on U . Then we define the following

equivalence relation on U : x 0 ∼ x1 ⇔ there exists a smooth curve λ in U joining x 0 to x1
and tangent to D at each point, i.e., λ ( t ) ∈D ( λ ( t ) ) , for all t. Suppose that S is an
equivalence set relative to the preceding equivalence relation. Show that S is an
372 Chap. 10 • VECTOR FIELDS

integral surface of D . (Such an integral surface is called a maximal integral surface or a

leaf of D .)
Sec. 51 • Differential Forms, Exterior Derivative 373

Section 51. Differential Forms and Exterior Derivative

In this section we define a differential operator on skew-symmetric covariant tensor fields.

This operator generalizes the classical curl operator defined in Section 47.

Let U be an open set in E and let A be a skew-symmetric covariant tensor field of order
r on U ,i.e.,

A: U → T ˆr (V ) (51.1)

Then for any point x ∈U , A ( x ) is a skew-symmetric tensor of order r (cf. Chapter 8).
Choosing a coordinate chart xˆ onU as before, we can express A in the component form

A = Ai1…ir g i1 ⊗ ⊗ g ir (51.2)

where {g i } denotes the natural basis of xˆ . Since A is skew-symmetric, its components obey the
identify

Ai1… j…k…ir = − Ai1…k… j…ir (51.3)

for any pair of indices ( j, k ) in ( i1 ,..., ir ) . As explained in Section 39, we can then represent A
by

1
A= ∑
i1 < <ir
Ai1 ...ir g i1 ∧ ∧ g ir = Ai ...i g i1 ∧
r! 1 r
∧ g ir (51.4)

where ∧ denotes the exterior product.

Now if A is smooth, then it is called a differential form of order r, or simply an r-form.

In the theory of differentiable manifolds, differential forms are important geometric entities,
since they correspond to linear combinations of volume tensors on various hypersurfaces (cf. the
book by Flanders1). For our purpose, however, we shall consider only some elementary results
about differential forms.

1
H. Flanders, Differential Forms with Applications to the Physical Sciences, Academic Press, New York-London,
1963.
374 Chap. 10 • VECTOR FIELDS

We define first the notion of the exterior derivative of a differential form. Let A be the r-
form represented by (51.2) or (51.4). Then the exterior derivative d A is an ( r + 1) –form given
by any one of the following three equivalent formulas:

N ∂Ai1…ir
dA = ∑∑
i1 < < ir k =1 ∂x k
g k ∧ g i1 ∧ ∧ g ir

1 ∂Ai1…ir k
= g ∧ g i1 ∧ ∧ g ir (51.5)
r ! ∂x k

1 ∂Ai1…ir j1
= δ kij1…
1 …ir
g ∧ ∧ g jr+1
r !( r + 1)! ∂x k
jr +1

Of course, we have to show that d A as defined by (51.5), is independent of the choice of the
chart xˆ. If this result is granted, then (51.5) can be written in the coordinate-free form

dA = ( −1) ( r + 1) !K r +1 ( grad A )
r
(51.6)

where K r+1 is the skew-symmetric operator introduced in Section 37.

We shall now show that d A is well-defined by (51.5), i.e., if ( x i ) is another coordinate

system, then
∂Ai1…ir ∂Ai1…ir
g k ∧ g i1 ∧ ∧ g ir = g k ∧ g i1 ∧ g ir (51.7)
∂x k
∂x k

or, equivalently,

∂Ai1 ∂Ai1 ∂x l1 ∂x lr +1
δ jki… ij
1 r ir
= δ l1ki…1 lrir+1
ir
(51.8)
1 r +1
∂x k ∂x k ∂x j1 ∂x jr +1

for all j1 ,… , jr +1 , where Ai1…ir and { g i } denote the components of A and the natural basis
corresponding to ( x i ) . To prove (51.8), we recall first the transformation rule

∂x j1 ∂x jr
Ai1 = Aj1 (51.9)
∂x i1 ∂x ir
ir jr

for any covariant tensor components. Now differentiating (51.9) with respect to x k , we obtain
Sec. 51 • Differential Forms, Exterior Derivative 375

∂Ai1 ∂Aj1… jr ∂x l ∂x j1 ∂x jr
=
ir

∂x k ∂x l ∂x k ∂x i1 ∂x ir
∂ 2 x j1 ∂x j2 ∂x jr
+ Aj1… jr k i1 (51.10)
∂x ∂x ∂x i2 ∂x ir
∂x j1 ∂x jr −1 ∂ 2 x jr
+ + Aj1… jr
∂x i1 ∂x ir −1 ∂x k ∂x ir

Since the second derivative ∂ 2 x j ∂x k ∂x i is symmetric with respect to the pair ( k , i ) , when we
form the contraction of (51.10) with the skew-symmetric operator K r +1 the result is (51.8). Here
we have used the fact that K r +1 is a tensor of order 2 ( r + 1) , so that we have the identities

∂x k ∂x i1 ∂x ir ∂x p1 ∂x pr +1
δ jki… ij = δ plm… pm
1 r 1 r
1 r +1 1 r +1
∂x 1 ∂x m1 ∂x mr ∂x j1 ∂x jr +1
(51.11)
∂x k ∂x i1 ∂x ir ∂x p1 ∂x pr +1
=δ lm1 mr
p1… pr +1
∂x l ∂x m1 ∂x mr ∂x j1 ∂x jr +1

From (38.2) and (51.5), if A is a 1-form, say A = u, then

⎛ ∂u ∂u ⎞
du = ⎜ ij − ij ⎟ g j ⊗ g i (51.12)
⎝ ∂x ∂x ⎠

Comparing this representation with (47.52), we see that the exterior derivative and the curl
operator are related by

2 curl u = − du (51.13)

for any 1-form u. Equation (51.13) also follows from (51.6) and (47.50). In the sense of (51.13)
, the exterior derivative is a generalization of the curl operator from 1-forms to r-forms in
general. We shall now consider some basic properties of the exterior derivative.

I. If f is a smooth function (i.e., a 0-form), then the exterior derivative of f coincides

with the gradient of f,

df = grad f (51.14)

This result follows readily from (51.5) and (47.14).

II. If w is a smooth covariant vector field (i.e., a 1-form), and if u and v are smooth
vector fields, then
376 Chap. 10 • VECTOR FIELDS

u ⋅ d ( v ⋅ w ) − v ⋅ d ( u ⋅ w ) = [ u, v ] ⋅ w + dw ( u, v ) (51.15)

We can prove this formula by direct calculation. Let the component representations of
u, v, w in a chart x̂ be

u = ui gi , v = υ i gi , w = wi g i (51.16)

Then

u ⋅ w = u i wi , v ⋅ w = υ i wi (51.17)
From I it follows that

∂ ( u i wi ) ⎛ ∂w ∂u i ⎞ k
d (u ⋅ w) = g k = ⎜ u i ki + wi k ⎟g (51.18)
∂x k
⎝ ∂x ∂x ⎠

and similarly

∂ (υ i wi ) ⎛ ∂w ∂υ i ⎞
d (v ⋅ w) = g k = ⎜υ i ki + wi k ⎟ g k (51.19)
∂x k ⎝ ∂x ∂x ⎠

Consequently,

⎛ ∂υ i ∂u i ⎞ i k ∂wi
u ⋅ d ( v ⋅ w ) − v ⋅ d (u ⋅ w ) = ⎜ uk k − υ k k ⎟ wi + ( u υ − u υ ) k (51.20)
k i

⎝ ∂x ∂x ⎠ ∂x

Now from (49.21) the first term on the right-hand side of (51.20) is simply [ u, v ] ⋅ w. Since the
coefficient of the second tensor on the right-hand side of (51.20) is skew-symmetric, that term
can be rewritten as

⎛ ∂w ∂w ⎞
u kυ i ⎜ ki − ik ⎟
⎝ ∂x ∂x ⎠

or, equivalently,

d w ( u, v )

when the representation (51.12) is used. Thus the identity (51.15) is proved.
Sec. 51 • Differential Forms, Exterior Derivative 377

III If A is an r-form and B is an s-form, then

d ( A ∧ B ) = d A ∧ B + ( −1) A ∧ d B
r
(51.21)

This formula can be proved by direct calculation also, so we leave it as an exercise.

IV. If A and B are any r-forms, then

d ( A + B ) = d A + dB (51.22)

The proof of this result is obvious from the representation (51.5). Combining (51.21) and (51.22)
, we have

d (α A + β B ) = α d A + β d B (51.23)

for any r-forms A and B and scalars α and β.

V. For any r-form A

d 2A ≡ d ( d A) = 0 (51.24)

This result is a consequence of the symmetry of the second derivative.

Indeed, from (51.5) for the ( r + 1) − form dA we have

1 1 1 ∂ 2 Ai1…ir
d A=
2
δ lj1 jr +1
δ ki1 ir
g p1 ∧ ∧ g pr + 2
r ! ( r + 2)! ( ( r + 1)!)
p1… pr + 2 j1… jr +1
2
∂x ∂x
l k

1 1 1 ∂2 A
= δ plki1…1 pirr + 2 l i1…ikr g p1 ∧ ∧ g pr + 2 (51.25)
r ! ( r + 1) ! ( r + 2 ) ! ∂x ∂x
=0

where we have used the identity (20.14).

VI. Let f be a smooth mapping from an open set U in E into an open set U in E , and
suppose that A is an r-form on U . Then

d (f ∗ ( A) ) = f ∗ (d A) (51.26)

where f ∗ denotes the induced linear map (defined below) corresponding to f , so that f ∗ ( A ) is an
r-form on U , and where d denotes the exterior derivative on U . To establish (51.26), let
378 Chap. 10 • VECTOR FIELDS

dimE = N and dimE = M . We choose coordinate systems ( x i , i = 1,…, N ) and

(x α
, α = 1,… , M ) on U and U , respectively. Then the mapping f can be characterized by

x i = f i ( x 1 ,… , x M ) , i = 1,…, N (51.27)

Suppose that A has the representation (51.4)2 in ( x i ) . Then by definition f ∗ ( A ) has the
representation

1 ∂x i1 ∂x ir α1
f ∗ (A) = Ai1…ir α1 g ∧ ∧ g αr (51.28)
r! ∂x ∂x αr

where { g α } denotes the natural basis of ( x α ) . Now by direct calculation of the exterior
derivatives of A and f ∗ ( A ) , and by using the symmetry of the second derivative, we obtain the
representation

d ( f ∗ ( A) ) = f ∗ ( d A)
1 ∂Ai1…ir ∂x k ∂x i1 ∂x ir β (51.29)
= g ∧ g α1 ∧ ∧ g αr
r ! ∂x k ∂x β ∂x α1 ∂x αr

Thus (51.26) is proved.

Applying the result (51.26) to the special case when f is the flow ρt generated by a vector
field v as defined in Section 49, we obtain the following property.
VII. The exterior derivative and the Lie derivative commute, i.e., for any differential form
A and any smooth vector field v

( )
d L A = L ( d A)
v v
(51.30)

We leave the proof of (51.30), by direct calculation based on the representations (51.5) and
(49.42), as an exercise.

The results I-VII summarized above are the basic properties of the exterior derivative. In
fact, results I and III-V characterize the exterior derivative completely. This converse assertion
can be stated formally in the following way.

Coordinate-free definition of the exterior derivative. The exterior derivative d is an operator

from an r-form to an ( r + 1) − form and satisfies the conditions, I, III-V above.
Sec. 51 • Differential Forms, Exterior Derivative 379

To see that this definition is equivalent to the representations (51.5), we consider any r-
form A given by the representation (51.4)2. Applying the operator d to both sides of that
representation and making use of conditions I and III-V, we get

(
r ! dA = d Ai1…ir g i1 ∧ ∧ g ir )
(
= d Ai1…ir ) ∧g i1
∧ ∧ g ir + Ai1…ir d g i1 ∧ g i2 ∧ ∧ g ir
− Ai1…ir g i1 ∧ dg i2 ∧ g i3 ∧ ∧ g ir + (51.31)
⎛ ∂Ai …i ⎞
= ⎜ 1 k r g k ⎟ ∧ g i1 ∧ ∧ g ir + 0 − 0 +
⎝ ∂x ⎠
∂Ai …i
= 1 k r g k ∧ g i1 ∧ ∧ g ir
∂x

where we have used the fact that the natural basis vector g i is the gradient of the coordinate
function x i , so that, by I,

g i = grad x i = dx i (51.32)

and thus, by V,

d gi = d 2 x i = 0 (51.33)

Consequently (51.5) is equivalent to the coordinate-free definition just stated.

Exercises

51.1 Prove the product rule (51.21) for the exterior derivative.
51.2 Given (47.53), the classical definition of the curl of a vector field, prove that

curl v = D 2 ( dv ) (51.34)

where D is the duality operator introduced in Section 41.

51.3 Let f be a 0-form. Prove that

Δf = D3d (D1 df ) (51.35)

in the case where dimE =3.

51.4 Let f be the smooth mapping from an open set U in E onto an open set U in E
discussed in VI. First, show that
380 Chap. 10 • VECTOR FIELDS

∂x i
grad f = g ⊗ gα
α i
∂x

Next show that f ∗ , which maps r-forms on U into r-forms on U , can be defined by

f ∗ ( A ) ( u1 ,… , u r ) = A ( ( grad f ) u1 ,…, ( grad f ) u r )

for all u1 ,… , u r in the translation space of E . Note that for r = 1 , f ∗ is the transpose of
grad f .

51.5 Check the commutation relation (51.30) by component representations.

Sec. 52 • Dural Form of Frobenius Theorem; Poincaré Lemma 381

Section 52. The Dual Form of the Frobenius Theorem; the Poincaré Lemma
The Frobenius theorem as stated and proved in Section 50 characterizes the integrability
of a distribution by a condition on the Lie bracket of the generators of the distribution. In this
section we shall characterize the same by a condition on the exterior derivative of the generators
of the orthogonal complement of the distribution. This condition constitutes the dual form of the
Frobenius theorem.

As in Section 50, let D be a distribution of dimension D defined on a domain U in E .

Then at each point x ∈U , D ( x ) is a D -dimensional subspace of V . We put D ⊥ ( x ) to be the
orthogonal complement of D ( x ) . Then dimD ⊥ ( x ) = N − D and the field D ⊥ on U defined in
this way is itself a distribution of dimension N − D. In particular, locally, there exist 1-forms
{z Γ , Γ = 1,…, N − D} which generate D ⊥ . The dual form of the Frobenius theorem is the
following theorem.

Theorem 52.1. The distribution D is integrable if and only if

d z Γ ∧ z1 ∧ ∧ z N − D = 0, Γ = 1,… , N − D (52.1)
Proof. As in Section 50, let {vα , α = 1,… , D} be a local basis for D . Then

vα ⋅ z Γ = 0, α = 1,… , D, Γ = 1,… , N − D (52.2)

By the Frobenius theorem D is integrable if and only if (50.5) holds. Substituting (50.5) and
(52.2) into (51.15) with u = vα , v = v β , and w = z Γ , we see that (50.5) is equivalent to

d z Γ ( vα , v β ) = 0, α , β = 1,… , D, Γ = 1,…, N − D (52.3)

We now show that this condition is equivalent to (52.1).

To prove the said equivalence, we extend {vα } into a basis {v i } in such a way that its
dual basis {v i } satisfies v D +Γ = z Γ for all Γ = 1,… , N − D. This extension is possibly by virtue of
(52.2). Relative to the basis {v i } , the condition (52.3) means that in the representation of d z Γ
by

2d z Γ = ζ ijΓ v i ∧ v j = 2ζ ijΓ v i ⊗ v j (52.4)

the components ζ ijΓ with 1 ≤ i, j ≤ D must vanish. Thus

382 Chap. 10 • VECTOR FIELDS

2d z Γ = ζ iΓΔ+ D v i ∧ z Δ (52.5)

which is clearly equivalent to (52.1).

As an example of the integrability condition (52.1), we can derive a necessary and

sufficient condition for a vector field z to be orthogonal to a family of surfaces. That is, there
exist smooth functions h and f such that

z = h grad f (52.6)

Such a vector field is called complex-lamellar in the classical theory. Using the terminology
here, we see that a complex-lamellar vector field z is a vector field such that z ⊥ is integrable.
Hence by (52.1), z is complex-lamellar if and only if dz ∧ z = 0. In a three-dimensional space
this condition can be written as

( curl z ) ⋅ z = 0 (52.7)

which was first noted by Kelvin (1851). In component form, if z is represented by

z = zi g i = zi dx i (52.8)

then dz is represented by

1 ∂zi k
dz = dx ∧ dx i (52.9)
2 ∂x k

and thus dz ∧ z = 0 means

∂zi k
zj dx ∧ dx i ∧ dx j = 0 (52.10)
∂x k

or, equivalently,

∂zi
δ pqr
kij
zj = 0, p, q, r = 1,… , N (52.11)
∂x k

Since it suffices to consider the special cases with p < q < r in (52.11), when N=3 we have

∂zi ∂z
0 = δ123
kij
zj = z jε kij ki (52.12)
∂x k
∂x
Sec. 52 • Dural Form of Frobenius Theorem; Poincaré Lemma 383

which is equivalent to the component version of (52.7). In view of this special case we see that
the dual form of the Frobenius theorem is a generalization of Kelvin’s condition from 1-forms to
r-forms in general.

In the classical theory a vector field (1-form) z is called lamellar if locally z is the gradient
of a smooth function (0-form) f, namely

z = grad f = df (52.13)

Then it can be shown that z is lamellar if and only if dz vanishes, i.e.,

curl z = 0 (52.14)

The generalization of lamellar fields from 1-forms to r-forms in general is obvious: We say that
an r-form A is closed if its exterior derivative vanishes

dA = 0 (52.15)

and we say that A is exact if it is the exterior derivatives of an (r-1)-form, say

A = dB (52.16)

From (51.24) any exact form is necessarily closed. The theorem that generalizes the classical
result is the Poincaré lemma, which implies that (52.16) and (52.15) are locally equivalent.

Before stating the Poincaré lemma, we defined first the notion of a star-shaped (or
retractable) open set U in E : U is star-shaped if there exists a smooth mapping

ρ : U ×R → U (52.17)

such that

⎧x when t ≤ 0
ρ ( x, t ) = ⎨ (52.18)
⎩x 0 when t ≥ 1

where x 0 is a particular point in U . In a star-shaped open set U all closed hypersurfaces of

dimensions 1 to N − 1 are retractable in the usual sense. For example, an open ball is star-
shaped but the open set between two concentric spheres is not, since a closed sphere in the latter
is not retractable.
384 Chap. 10 • VECTOR FIELDS

The equivalence of (52.13) and (52.14) requires only that the domain U of z be simply
connected, i.e., every closed curve in U be retractable. When we generalize the result to the
equivalence of (52.15) and (52.16), the domain U of A must be retractable relative to all r-
dimensional closed hypersurfaces. For simplicity, we shall assume that U is star-shaped.

Theorem 52.2. If A is a closed r-form defined on a star-shaped domain, then A is exact, i.e.,
there exists an (r-1)-form B on U such that (52.16) holds.

Proof. We choose a coordinate system ( x i ) onU . Then a coordinate system

(y α
, α = 1,… , N + 1) on U ×R is given by

( y ,… , y ) = ( x ,… , x , t )
1 N +1 1 N
(52.19)

From (52.18) the mapping ρ has the property

ρ (⋅ , t ) = idU when t ≤ 0 (52.20)

and
ρ (⋅ , t ) = xo when t ≥ 0 (52.21)

Hence if the coordinate representation of ρ is

x i = ρ i ( yα ) , i = 1,… , N (52.22)

Then

∂ρ i ∂ρ i
= δ ij , =0 when t≤0 (52.23)
∂y j ∂y N +1

and

∂ρ i ∂ρ i
=0 =0 when t ≥1 (52.24)
∂y j ∂y N +1

As usual we can express A in component form relative to ( x i )

1
A= Ai1…ir g i1 ∧ ∧ g ir (52.25)
r!
Sec. 52 • Dural Form of Frobenius Theorem; Poincaré Lemma 385

Then ρ∗ ( A ) is an r-form on U ×R defined by

1 ∂ρ i1 ∂ρ ir α1
ρ∗ ( A ) = Ai1…ir α1 h ∧ ∧ hαr (52.26)
r! ∂y ∂y αr

where {hα , α = 1,… , N + 1} denotes the natural basis of ( y α ) , i.e.,

hα = dy α (52.27)

where d denotes the exterior derivative on U ×R .

Now from (52.19) we can rewrite (52.26) as

r !ρ∗ ( A ) = X i1…ir hi1 ∧ ∧ hir + rYi1…ir −1 hi1 ∧ ∧ hir −1 ∧ h N +1 (52.28)

where

∂ρ j1 ∂ρ jr
X i1…ir = A j1… jr (52.29)
∂y i1 ∂y ir

and

∂ρ j1 ∂ρ jr −1 ∂ρ jr
Yi1…ir −1 = Aj1… jr (52.30)
∂y i1 ∂y ir −1 ∂y N +1

We put

1
X= X i1…ir hi1 ∧ ∧ hir (52.31)
r!

and

1
Y= Y j1… jr −1 h j1 ∧ ∧ h jr −1 (52.32)
( )
r − 1 !

Then X is an r-form and Y is an ( r − 1) − form on U ×R . Further, from (52.28) we have

ρ∗ ( A ) = X + Y ∧ h N +1 = X + Y ∧ dt (52.33)
386 Chap. 10 • VECTOR FIELDS

From (52.23), (52.24), (52.29), and (52.30), X and Y satisfy the end conditions

X i1…ir ( x, t ) = Ai1…ir ( x ) , Yi1…ir −1 ( x, t ) = 0 when t ≤ 0 (52.34)

and

X i1…ir ( x, t ) = 0, Yi1…ir −1 ( x, t ) = 0 when t ≥ 1 (52.35)

for all x ∈U . We define

B≡
( −1) 1Y
( r − 1) ! ∫0 i …i (
(⋅ , t ) dt 1 r −1 )g ∧ i1
∧ g ir −1 (52.36)

and we claim that this ( r − 1) -form B satisfies the condition (52.16).

To prove this, we take the exterior derivative of (52.33). By (52.26) and the fact that A is
closed, the result is

d X + d Y ∧ dt = 0 (52.37)

where we have used also (51.21) and (51.24) on the term Y ∧ dt . From (52.31) and (52.32) the
exterior derivatives of X and Y have the representations

∂X i1 ∂X i1
r !d X = ir
h j ∧ hi1 ∧ ∧ hir + ir
dt ∧ hi1 ∧ ∧ hir (52.38)
∂x j
∂t

and

∂Yi1 ∂Yi1
( r − 1) ! d Y = ir −1
h j ∧ hi1 ∧ ∧ hir −1 + ir −1
dt ∧ hi1 ∧ ∧ hir (52.39)
∂x j
∂t

Substituting these into (52.37), we then get

∂X i1 ir
h j ∧ hi1 ∧ ∧ hir = 0 (52.40)
∂x j

and
Sec. 52 • Dural Form of Frobenius Theorem; Poincaré Lemma 387

⎛ r ∂X i1 ∂Yi2 ⎞ i1
⎜ ( −1) +r ⎟h ∧ ∧ hir ∧ dt = 0
ir ir
(52.41)
⎝ ∂t ∂x i1
⎠

Consequently
1 ∂Y r ∂X j1
− δ ij11…irjr i2i1 ir = ( −1) jr
(52.42)
( r − 1) ! ∂x ∂t

Now from (52.36) if we take the exterior derivative of the ( r − 1) -form B on U , the
result is

( −1) ⎛ 1 ∂Yi i dt ⎞ gi ∧ ∧ gi
r

( r − 1) ! ⎜⎝ ∫0 ∂x i ⎟⎠
dB = 2
1
r 1 r

(52.43)
( −1) 1 δ i i ⎛ 1 ∂Yi i dt ⎞ g j ∧
r

( r − 1) ! r ! j … j ⎜⎝ ∫0 ∂x i ⎟⎠
= 1
1
r
r
2
1
r 1
∧ g jr

Substituting (52.42) into this equation yields

1 ⎛ 1 ∂X j1 ⎞
r ! ⎝ ∫0 ∂t
dB = ⎜− dt ⎟ g j1 ∧ ∧ g jr
jr

⎠
1
=
r!
(
X j1 jr ( ⋅ ,0 ) − X j1 jr ( ⋅ ,1) g j1 ∧ ) ∧ g jr (52.44)

1
= Aj j g j1 ∧ ∧ g jr = A
r! 1 r

where we have used the end conditions (52.34) and (52.35) on X.

The ( r − 1) -form B , whose existence has just been proved by (52.44), is not unique of
course. Since

ˆ ⇔ d B −B
dB = dB ˆ =0
( ) (52.45)

B is unique to within an arbitrary closed ( r − 1) -form on U .

Exercises
52.1 In calculus a “differential”
388 Chap. 10 • VECTOR FIELDS

P ( x, y , z ) dx + Q ( x, y , z ) dy + R ( x, y , z ) dz (52.46)

is called exact if there exists a function U ( x, y , z ) such that

∂U ∂U ∂U
dU = dx + dy + dz
∂x ∂y ∂z (52.47)
= P dx + Q dy + R dz

Use the Poincaré lemma and show that (52.46) is exact if and only if

∂P ∂Q ∂P ∂R ∂Q ∂R
= , = , = (52.48)
∂y ∂x ∂z ∂x ∂z ∂y

52.2 The “differential” (52.46) is called integrable if there exists a nonvanishing function
μ ( x, y , z ) , called an integration factor, such that the differential

μ ( P dx + Q dy + R dz ) (52.49)

is exact. Show that (52.46) is integrable if and only if the two-dimensional distribution
orthogonal to the 1-form (52.46) is integrable in the sense defined in this section. Then
use the dual form of the Frobenius theorem and show that (52.46) is integrable if and
only if

⎛ ∂R ∂Q ⎞ ⎛ ∂P ∂R ⎞ ⎛ ∂Q ∂P ⎞
P⎜ − ⎟ +Q⎜ − ⎟ +R ⎜ − ⎟=0 (52.50)
⎝ ∂y ∂z ⎠ ⎝ ∂z ∂x ⎠ ⎝ ∂x ∂y ⎠
Sec. 53 • Three-Dimensional Euclidean Manifold, I 389

Section 53. Vector Fields in a Three-Dimensional Euclidean Manifold, I.

Invariants and Intrinsic Equations
The preceding four sections of this chapter concern vector fields, distributions, and
differential forms, defined on domains in an N-dimensional Euclidean manifold E in general. In
applications, of course, the most important case is when E is three-dimensional. Indeed,
classical vector analysis was developed just for this special case. In this section we shall review
some of the most important results in the classical theory from the more modern point of view, as
we have developed so far in this text.

We recall first that when E is three-dimensional the exterior product may be replaced by
the cross product [cf.(41.19)]. Specially, relative to a positive orthonormal basis
{e1 , e 2 , e3} for V , the component representation of u × v for any u, v,∈V is

u × v = ε ijk u i v j e k
(53.1)
= ( u 2υ 3 − u 3υ 2 ) e1 + ( u 3υ 1 − u1υ 3 ) e 2 + ( u1υ 2 − u 2υ 1 ) e 3

Where

u = uie i , v = υ je j (53.2)

are the usual component representations of u and v and where the reciprocal basis {ei } coincides
with {e i } . It is important to note that the representation (53.1) is valid relative to a positive
orthonormal basis only; if the orthonormal basis {e i } is negative, the signs on the right-hand side
of (53.1) must be reversed. For this reason, u × v is called an axial vector in the classical theory.

We recall also that when E is three-dimensional, then the curl of a vector field can be
represented by a vector field [cf. (47.53) or (51.34)]. Again, if a positive rectangular Cartesian
coordinate system ( x1 , x 2 , x 3 ) induced by {e i } is used, then curl v has the components

∂υ j
curl v = ε ijk ek
∂x i
(53.3)
⎛ ∂υ ∂υ ⎞ ⎛ ∂υ ∂υ ⎞ ⎛ ∂υ ∂υ ⎞
= ⎜ 23 − 32 ⎟ e1 + ⎜ 31 − 13 ⎟ e 2 + ⎜ 12 − 21 ⎟ e 3
⎝ ∂x ∂x ⎠ ⎝ ∂x ∂x ⎠ ⎝ ∂x ∂x ⎠

where v is now a smooth vector field having the representation

v = υi e i = υ i e i (53.4)
390 Chap. 10 • VECTOR FIELDS

Since ( x i ) is rectangular Cartesian, the natural basis vectors and the component fields satisfy the
usual conditions

υ i = υi , ei = ei , i = 1, 2,3 (53.5)

By the same remark as before, curl v is an axial vector field, so the signs on the right-hand side
must be reversed when ( x i ) is negative.

Now consider a nonvanishing smooth vector field v. We put

s = v/ v (53.6)

Then s is a unit vector field in the same direction as v. In Section 49 we have introduced the
notions of internal curves and flows corresponding to any smooth vector field. We now apply
these to the vector field s. Since s is a unit vector, its integral curves are parameterized by the
arc length1 s. A typical integral curve of s is

λ = λ (s) (53.7)

where

dλ / ds = s ( λ ( s ) ) (53.8)

at each point λ ( s ) of the curve. From (53.6) and (53.8) we have

dλ
parallel to v ( λ ( s ) ) (53.9)
ds

but generally dλ / ds is not equal to v ( λ ( s ) ) , so λ is not an integral curve of v as defined in

Section 49. In the classical theory the locus of λ without any particular parameterization is
called a vector line of v.

Now assuming that λ is not a straight line, i.e., s is not invariant on λ , we can take the
covariant derivative of s on λ (cf. Section 48 and write the result in the form

1
Arc length shall be defined in a general in Section 68. Here s can be regarded as a parameter such that
dλ / ds = 1.
Sec. 53 • Three-Dimensional Euclidean Manifold, I 391

ds / ds = κ n (53.10)

where κ and n are called the curvature and the principal normal of λ , and they are characterized
by the condition

ds
κ= >0 (53.11)
ds

It should be noted that (53.10) defines both κ and n : κ is the norm of ds / ds and n is the unit
vector in the direction of the nonvanishing vector field ds / ds . The reciprocal of κ ,

r = 1/ κ (53.12)

is called the radius of curvature of λ .

Since s is a vector, we have

d ( s ⋅ s) ds
0= = 2s ⋅ = 2κ s ⋅ n (53.13)
ds ds

or, equivalently,

s⋅n = 0 (53.14)

Thus n is normal to s , as it should be. In view of (53.14) the cross product of s with n is a unit
vector

b ≡ s×n (53.15)

which is called the binormal of λ . The triad {s, n, b} now forms a field of positive orthonormal
basis in the domain of v. In general {s, n, b} is anholonomic, of course.

Now we compute the covariant derivative of n and b along the curve λ . Since b is a
unit vector, by the same argument as (53.13) we have

db
⋅b = 0 (53.16)
ds

Similarly, since b ⋅ s = 0, on differentiating with respect to s along λ , we obtain

392 Chap. 10 • VECTOR FIELDS

db ds
⋅ s = −b ⋅ = −κ b ⋅ n = 0 (53.17)
ds ds

where we have used (53.10). Combining (53.16) and (53.17), we see that db / ds is parallel to n,
say

db
= −τ n (53.18)
ds

where τ is called the torsion of the curve λ . From (53.10) and (53.18) the gradient of n along
λ can be computed easily by the representation

n = b×s (53.19)

so that
dn ds db
= b× + × s = b × κ n − τ n × s = −κ s + τ b (53.20)
ds ds ds

The results (53.10), (53.18), and (53.20) are the Serret-Frenet formulas for the curve λ .

So far we have introduced a field of basis {s, n, b} associated with the nonzero and
nonrectilinear vector field v. Moreover, the Serret-Frenet formulas give the covariant derivative
of the basis along the vector lines of v . In order to make full use of the basis, however, we need
a complete representation of the covariant derivative of that basis along all curves. Then we can
express the gradient of the vector fields s, n, b in component forms relative to the anholonomic
basis {s, n, b} . These components play the same role as the Christoffel symbols for a holonomic
basis. The component forms of grad s , grad n and grad b have been derived by Bjørgum.2 We
shall summarize his results without proofs here.

Bjørgum shows first that the components of the vector fields curl s , curl n , and curl b
relative to the basis {s, n, b} are given by

curls = Ωss + κ b
curln = − ( div b ) s + Ω n n + θ b (53.21)
curlb = (κ + div n ) s + η n + Ω b b

where Ωs , Ω n , Ω b are given by

2
O. Bjørgum, “On Beltrami Vector Fields and Flows, Part I. A Comparative Study of Some Basic Types of Vector
Fields,” Universitetet I Bergen, Arbok 1951, Naturvitenskapelig rekke Nr. 1.
Sec. 53 • Three-Dimensional Euclidean Manifold, I 393

s ⋅ curl s = Ωs , n ⋅ curl n = Ω n , b ⋅ curl b = Ω b (53.22)

and are called the abnormality of s, n, b , respectively. From Kelvin’s theorem (cf. Section 52)
we know that the abnormality measures, in some sense, the departure of a vector field from a
complex-lamellar field. Since b is given by (53.15), the abnormalities Ωs , Ω n , Ω b are not
independent. Bjørgum shows that

Ω n + Ω b = Ωs − 2τ (53.23)

where τ is the torsion of λ as defined by (53.18). The quantities θ and η in (53.21) are defined
by

b ⋅ curl n = θ , n ⋅ curl b = η (53.24)

and Bjørgum shows that

θ − η = div s (53.25)

Notice that (53.21) implies that

n ⋅ curl s = 0 (53.26)

but the remaining eight components in (53.21) are generally nonzero.

Next Bjørgum shows that the components of the second-order tensor fields grad s ,
grad n , and grad b relative to the basis {s, n, b} are given by

grad s = κ n ⊗ s + θ n − ( Ω n + τ ) n ⊗ b
+ ( Ωb + τ ) b ⊗ n − ηb ⊗ b
grad n = −κ s ⊗ s − θ s ⊗ n + ( Ω n + τ ) s ⊗ b
(53.27)
+τ b ⊗ s − ( div b ) b ⊗ n + (κ + div n ) b ⊗ b
grad b = − ( Ω b + τ ) s ⊗ n + ηs ⊗ b − τ n ⊗ s
+ ( div b ) n ⊗ n − (κ + div n ) n ⊗ b

These representations are clearly consistent with the representations (53.21) through the general
formula (47.53). Further, from (48.15) the covariant derivatives of s, n, and b along the integral
curve λ of s are
394 Chap. 10 • VECTOR FIELDS

ds
= ( grad s ) s = κ n
ds
dn
= ( grad n ) s = −κ s + τ b (53.28)
ds
db
= ( grad b ) s = −τ n
ds

which are consistent with the Serret-Frenet formulas (53.10), (53.20), and (53.18).

The representations (53.27) tell us also the gradients of {s, n, b} along any integral curves
μ = μ ( n ) of n and ν = ν ( b ) of b. Indeed, we have

ds / dn = ( grad s ) n = θ n + ( Ω b + τ ) b
dn / dn = ( grad n ) n = −θ s − ( div b ) b (53.29)
db / dn = ( grad b ) n = − ( Ω b + τ ) s + ( div b ) n

and

ds / db = ( grad s ) b = − ( Ω n + τ ) n − η b
dn / db = ( grad n ) b = ( Ω n + τ ) s + (κ + div n ) b (53.30)
db / db = ( grad b ) b = ηs − (κ + div n ) n

Since the basis {s, n, b} is anholonomic in general, the parameters ( s, n, b ) are not local
coordinates. In particular, the differential operators d / ds, d / dn, d / db do not commute. We
derive first the commutation formulas3 for a scalar function f.

From (47.14) and (46.10) we verify easily that the anholonomic representation of grad f is

df df df
grad f = s+ n+ b (53.31)
ds dn db

for any smooth function f defined on the domain of the basis {s, n, b} . Taking the gradient of
(53.31) and using (53.27), we obtain

3
See A.W. Marris and C.-C. Wang, “Solenoidal Screw Fields of Constant Magnitude,” Arch. Rational Mech. Anal.
39, 227-244 (1970).
Sec. 53 • Three-Dimensional Euclidean Manifold, I 395

⎛ df ⎞ df ⎛ df ⎞
grad ( grad f ) = s ⊗ ⎜ grad ⎟ + ( grad s ) + n ⊗ ⎜ grad ⎟
⎝ ds ⎠ ds ⎝ dn ⎠
df ⎛ df ⎞ df
+ ( grad n ) + b ⊗ ⎜ grad ⎟ + ( grad b )
dn ⎝ db ⎠ db
⎡ d df df ⎤
=⎢ −κ ⎥s⊗s
⎣ ds ds dn ⎦
⎡ d df df df ⎤
+⎢ −θ − ( Ωb + τ ) ⎥ s ⊗ n
⎣ dn ds dn db ⎦
⎡ d df df df ⎤
+⎢ + ( Ωn + τ ) + η ⎥ s ⊗ b
⎣ db ds dn db ⎦
⎡ df d df df ⎤
+ ⎢κ + −τ ⎥ n ⊗ s
⎣ ds ds dn db ⎦
⎡ df d df df ⎤
+ ⎢θ + + ( div b ) ⎥ n ⊗ n
⎣ dn dn dn db ⎦
⎡ df d df df ⎤
+ ⎢ − ( Ωn + τ ) + − (κ + div n ) ⎥ n ⊗ b
⎣ ds db dn db ⎦
⎡ df d df ⎤
+ ⎢τ + b⊗s
⎣ dn ds db ⎥⎦
(53.32)
⎡ df df d df ⎤
+ ⎢( Ω b + τ ) − ( div b ) + b⊗n
⎣ ds dn dn db ⎥⎦
⎡ df df d df ⎤
+ ⎢ −η + (κ + div n ) + b⊗b
⎣ ds dn db db ⎥⎦

Now since grad ( grad f ) is symmetric, (53.32) yields

d df d df df df df
− =κ +θ + Ωb
ds ds ds dn ds dn db
d df d df df df
− = Ωn +η (53.33)
ds db db ds dn db
d df d df df df df
− = Ωs − ( div b ) + (κ + div n )
db dn dn db ds dn db

where we have used (53.23). The Formulas (53.33)1-3 are the desired commutation rules.

Exactly the same technique can be applied to the identities

curl ( grad s ) = curl ( grad n ) = curl ( grad b ) = 0 (53.34)

396 Chap. 10 • VECTOR FIELDS

and the results are the following nine intrinsic equations4 for the basis {s, n, b} :

d dθ
− ( Ωn + τ ) − + (κ + div n )( Ω b − Ωn ) − (θ + η ) div b + Ωsκ = 0
dn db
dη d
− − ( Ω b + τ ) − ( Ω b − Ω n ) div b − (θ + η )(κ + div n ) = 0
dn db
dκ d
+ ( Ω n + τ ) − η ( Ωs − Ω b ) + Ω nθ = 0
db ds
dτ d
− (κ + div n ) − Ω n div b + η ( 2κ + div n ) = 0
db ds
dη
− + η 2 − κ (κ + div n ) − τ 2 − Ω n ( Ωs − Ω n ) = 0
ds
dκ dθ
− − κ 2 − θ 2 + ( 2Ω s − 3τ ) + Ω n ( Ω s − Ω n − 4τ ) = 0
dn ds
dτ d
+ div b − κ ( Ωs − Ω n ) + θ div b − Ω b (κ + div n ) = 0
dn ds
d d
(κ + div n ) + div b − θ n + ( div b ) + (κ + div n )
2 2

dn db
+ Ωsτ + ( Ω n + τ )( Ω b + τ ) = 0
d Ωs dκ (53.35)
+ + Ωs (θ − η ) + κ div b = 0
ds db

Having obtained the representations (53.21) and (53.27), the commutation rules (53.33),
and the intrinsic equations (53.35) for the basis {s, n, b} , we can now return to the original
relations (53.6) and derive various representations for the invariants of v. For brevity, we shall
now denote the norm of v by υ . Then (53.6) can be rewritten as

v = υs (53.36)

Taking the gradient of this equation yields

grad v = s ⊗ grad υ +υ grad s

dυ dυ dυ
= s⊗s+ s⊗n+ s⊗b
ds dn db (53.37)
+ υκ n ⊗ s + υθ n ⊗ n − υ ( Ω n + τ ) n ⊗ b
+υ ( Ω b + τ ) b ⊗ n − υη b ⊗ b

4
A. W. Marris and C.-C. Wang, see footnote 3.
Sec. 53 • Three-Dimensional Euclidean Manifold, I 397

where we have used (53.27)1. Notice that the component of grad v in b ⊗ s vanishes
identically, i.e.,

b ⋅ ( ( grad v ) s ) = [ grad v ] ( b, s ) = 0 (53.38)

or, equivalently,

dv
b⋅ =0 (53.39)
ds

so that the covariant derivative of v along its vector lines stays on the plane spanned by s and n.
In differential geometry this plane is called the osculating plane of the said vector line.

Next we can read off from (53.37) the representations5

dυ
div v = + υθ − υη
ds
dυ
= + υ div s
ds
(53.40)
dυ ⎛ dυ ⎞
curl v = υ ( Ω b + Ω n + 2τ ) s + n + ⎜υκ − ⎟b
db ⎝ dn ⎠
dυ ⎛ dυ ⎞
= υΩss + n + ⎜υκ − ⎟b
db ⎝ dn ⎠

where we have used (53.23) and (53.25). From (53.40)4 the scalar field Ω = v ⋅ curl v is given by

Ω = υ s ⋅ curl v = υ 2Ωs (53.41)

The representation (53.37) and its consequences (53.40) and (53.41) have many
important applications in hydrodynamics and continuum mechanics. It should be noted that
(53.21)1 and (53.27)1 now can be regarded as special cases of (53.40)4 and (53.37)2, respectively,
with υ = 1.

Exercises
53.1 Prove the intrinsic equations (53.35).
53.2 Show that the Serret-Frenet formulas can be written

5
O. Bjørgum, see footnote 2. See also J.L. Ericksen, “Tensor Fields,” Handbuch der Physik, Vol. III/1, Appendix,
Edited by Flügge, Springer-Verlag (1960).
398 Chap. 10 • VECTOR FIELDS

ds / ds = ω × s, dn / ds = ω × n, db / ds = ω × b (53.42)

where ω = τ s + κ b .
Sec. 54 • Three-Dimensional Euclidean Manifold, II 399

Section 54. Vector Fields in a Three-Dimensional Euclidean Manifold, II.

Representations for Special Classes of Vector Fields

In Section 52 we have proved the Poincaré lemma, which asserts that, locally, a
differential form is exact if and only if it is closed. This result means that we have the local
representation

f = dg (54.1)

for any f such that

df = 0 (54.2)

In a three-dimensional Euclidean manifold the local representation (54.1) has the following two
special cases.
(i) Lamellar Fields

v = grad f (54.3)

is a local representation for any vector field v such that

curl v = 0 (54.4)

Such a vector field v is called a lamellar field in the classical theory, and the scalar function f is
called the potential of v . Clearly, the potential is locally unique to within an arbitrary additive
constant.
(ii) Solenoidal Fields

v = curl u (54.5)

is a local representation for any vector field v such that

div v = 0 (54.6)

Such a vector field v is called a solenoidal field in the classical theory, and the vector field u is
called the vector potential of v . The vector potential is locally unique to within an arbitrary
additive lamellar field.

In the representation (54.3), f is regarded as a 0-form and v is regarded as a 1-form,

while in the representation (54.5), u is regarded as a 1-form and v is regarded as the dual of a 2-
form; duality being defined by the canonical positive unit volume tensor of the 3-dimensional
Euclidean manifold E .
400 Chap. 10 • VECTOR FIELDS

In Section 52 we remarked also that the dual form of the Frobenius theorem implies the
following representation.
(iii) Complex-Lamellar Fields

v = h grad f (54.7)

is a local representation for any vector field v such that

v ⋅ curl v = 0 (54.8)

Such a vector field v is called complex-lamellar in the classical theory. In the representation
(54.7) the surfaces defined by

f ( x ) = const (54.9)

are orthogonal to the vector field v .

We shall now derive some other well-known representations in the classical theory.

A. Euler’s Representation for Solenoidal Fields

Every solenoidal vector field v may be represented locally by

v = ( grad h ) × ( grad f ) (54.10)

Proof. We claim that v has a particular vector potential û which is complex-lamellar. From the
remark on (54.5), we may choose û by

uˆ = u + grad k (54.11)

where u is any vector potential of v . In order for û to be complex-lamellar, it must satisfy the
condition (54.8), i.e.,

0 = ( u + grad k ) ⋅ curl ( u + grad k )

(54.12)
= ( u + grad k ) ⋅ curl u = v ⋅ ( u + grad k )

Clearly, this equation possesses infinitely many solutions for the scalar function k, since it is a
first-order partial differential equation with smooth coefficients. Hence by the representation
(54.7) we may write
Sec. 54 • Three-Dimensional Euclidean Manifold, II 401

uˆ = h grad f (54.13)

Taking the curl of this equation, we obtain the Euler representation (54.10):

v = curl uˆ = curl ( h grad f ) = ( grad h ) × ( grad f ) (54.14)

It should be noted that in the Euler representation (54.10) the vector v is orthogonal to
grad h as well as to grad f , namely

v ⋅ grad h = v ⋅ grad f = 0 (54.15)

Consequently, v is tangent to the surfaces

h ( x ) = const (54.16)

or
f ( x ) = const (54.17)

For this reason, these surfaces are then called vector sheets of v . If v ≠ 0, then from (54.10),
grad h and grad f are not parallel, so that h and f are functionally independent. In this case the
intersections of the surfaces (54.16) and (54.17) are the vector lines of v.

Euler’s representation for solenoidal fields implies the following results.

B. Monge’s Representation for Arbitrary Smooth Vector Fields

Every smooth vector field v may be represented locally by

v = grad h + k grad f (54.18)

where the scalar functions, h, k, f are called the Monge potentials (not unique) of v .

Proof. Since (54.10) is a representation for any solenoidal vector field, from (54.5) we can write
curl v as

curl v = ( grad k ) × ( grad f ) (54.19)

It follows that (54.19) that

curl ( v − k grad f ) = 0 (54.20)

402 Chap. 10 • VECTOR FIELDS

Thus v − k grad f is a lamellar vector field. From (54.3) we then we have the local
representation

v − k grad f = grad h (54.21)

which is equivalent to (54.18).

Next we prove another well-known representation for arbitrary smooth vector fields in
the classical theory.

C. Stokes’ Representation for Arbitrary Smooth Vector Fields

Every smooth vector field v may be represented locally by

v = grad h + curl u (54.22)

where h and u are called the Stokes potential (not unique) of v .

Proof. We show that there exists a scalar function h such that v − grad h is solenoidal.
Equivalently, this condition means

div ( v − grad h ) = 0 (54.23)

Expanding (54.23), we get

Δh = div v (54.24)

where Δ denotes the Laplacian [cf. (47.49)]. Thus h satisfies the Poisson equation (54.24). It is
well known that, locally, there exist infinitely many solutions (54.24). Hence the representation
(54.22) is valid.

Notice that, from (54.19), Stokes’ representation (54.22) also can be put in the form

v = grad h + ( grad k ) × ( grad f ) (54.25)

Next we consider the intrinsic conditions for the various special classes of vector fields.
First, from (53.40)2 a vector field v is solenoidal if and only if

dυ / ds = −υ div s (54.26)
Sec. 54 • Three-Dimensional Euclidean Manifold, II 403

where s,υ , and s are defined in the preceding section. Integrating (54.26) along any vector line
λ = λ ( s ) defined before, we obtain

( s
υ ( s ) = υ0exp -∫ div s ds
s0 ) (54.27)

where υ0 is the value of υ at any reference point λ ( s0 ) . Thus1 in a solenoidal vector field v
the vector magnitude is determined to within a constant factor along any vector line λ = λ ( s ) by
the vector line pattern of v .

From (53.40)4, a vector field v is complex-lamellar if and only if

υΩ s = 0 (54.28)
or, equivalently,

Ωs = 0 (54.29)

since in the intrinsic representation v is assumed to be nonvanishing. The result (54.29) is

entirely obvious, because it defines the unit vector field s, parallel to v, to be complex-lamellar.

From (53.40)4 again, a vector field v is lamellar if and only if, in addition to (54.28) or
(54.29), we have also

dυ / db = 0, dυ / dn = υκ (54.30)

It should be noted that, when v is lamellar, it can be represented by (54.3), and thus the potential
surfaces defined by

f ( x ) = const (54.31)

are formed by the integral curves of n and b. From (54.30)1 along any b − line

υ ( b ) = υ ( b0 ) = const (54.32)

while from (54.30)2 along any n − line

υ ( n ) = υ ( n0 ) exp ( ∫ κ dn)
n

n0
(54.33)

1
O. Bj∅rgum, see footnote 2 in Section 53.
404 Chap. 10 • VECTOR FIELDS

Finally, in the classical theory a vector field v is called a screw field or a Beltrami field if
v is parallel to its curl, namely

v × curl v = 0 (54.34)

or, equivalently,

curl v = Ω s v (54.35)

where Ω s is the abnormality of s, defined by (53.22)1. In some sense a screw field is just the
opposite of a complex-lamellar field, which is defined by the condition that the vector field is
orthogonal to its curl [cf. (54.8)]. Unfortunately, there is no known simple direct representation
for screw fields. We must refer the reader to the three long articles by Bjørgum and Godal (see
footnote 1 above and footnotes 4 and 5 below), which are devoted entirely to the study of these
fields.

We can, of course, use some general representations for arbitrary smooth vector fields,
such as Monge’s representation or Stokes’ representation, to express a screw field first. Then we
impose some additional restrictions on the scalar fields involved in the said representations. For
example, if we use Monge’s representation (54.18) for v , then curl v is given by (54.19). In
this case v is a screw field if and only if the constant potential surfaces of k and f are vector
sheets of v , i.e.,

v ⋅ grad k = v ⋅ grad f = 0 (54.36)

From (53.40)4 the intrinsic conditions for a screw field are easily fround to be simply the
conditions (54.30). So the integrals (54.32) and (54.33) remain valid in this case, along the
b − lines and the n − lines. When the abnormality Ω s is nonvanishing, the integral of υ along
any s − line can be found in the following way: From (53.40) we have

dυ / ds = div v − υ div s (54.37)

Now from the basic conditions (54.35) for a screw field we obtain

d Ωs
0 = div ( Ωs v ) = Ωs div v + υ (54.38)
ds

So div v can be represented by

υ d Ωs
div v = − (54.39)
Ωs ds
Sec. 54 • Three-Dimensional Euclidean Manifold, II 405

Substituting (54.39) into (54.37) yields

dυ ⎛ 1 d Ωs ⎞
= −υ ⎜ div s + ⎟ (54.40)
ds ⎝ Ω s ds ⎠

or, equivalently,

d (υΩ s )
= −υΩ s div s (54.41)
ds

The last equation can be integrated at once, and the result is

υ ( s) =
υ ( s0 ) Ω s ( s0 )
Ω s ( s0 ) (
exp − ∫ ( div s ) ds
s

s0 ) (54.42)

which may be rewritten as

υ ( s ) Ω s ( s0 )
=
υ ( s0 ) Ω s ( s ) s (
exp − ∫ ( div s ) ds
s

0
) (54.43)

since υ is nonvanishing. From (54.43), (54.33), and (54.32) we see that the magnitude of a
screw field, except for a constant factor, is determined along any s − line, n − line, or b − line by
the vector line pattern of the field.2

A screw field whose curl is also a screw field is called a Trkalian field. According to a
theorem of Mémenyi and Prim (1949), a screw field is a Trkalian field if and only if its
abnormality is a constant. Further, all Trkalian fields are solenoidal and successive curls of the
field are screw fields, all having the same abnormality.3

The proof of this theorem may be found also in Bjørgum’s article. Trkalian fields are
considered in detail in the subsequent articles of Bjørgum and Godal4 and Godal5

2
O. Bjørgum, see footnote 2, Section 53.
3
P. Nemenyi and R. Prim, “Some Properties of Rotational Flow of a Perfect Gas,” Proc. Nat. Acad. Sci. 34, 119-
124; Erratum 35, 116 (1949).
4
O. Bjørgum and T. Godal, “On Beltrami Vector Fields and Flows, Part II. The Case when Ω is Constant in Space,”
Universitetet i Bergen, Arbok 1952, Naturvitenskapelig rekke Nr. 13.
5
T. Godal, “On Beltrami Vector Fields and Flows, Part III. Some Considerations on the General Case,” Universitete
i Bergen, Arbok 1957, Naturvitenskapelig rekke Nr.12.
________________________________________________________________

Chapter 11

HYPERSURFACES ON A EUCLIDEAN MANIFOLD

In this chapter we consider the theory of (N-1)-dimensional hypersurfaces embedded in an N-

dimensional Euclidean manifold E . We shall not treat hypersurfaces of dimension less than N-
1, although many results of this chapter can be generalized to results valid for those
hypersurfaces also.

Section 55. Normal Vector, Tangent Plane, and Surface metric

A hypersurface of dimension N-1 in E a set S of points in E which can be

characterized locally by an equation

x ∈ N ⊂ S ⇔ f (x ) = 0 (55.1)

where f is a smooth function having nonvanishing gradient. The unit vector field on N

grad f
n= (55.2)
grad f

is called a unit normal of S , since from (55.1) and (48.15) for any smooth curve λ = λ (t ) in S
we have

df λ 1
0= = (grad f ) ⋅ λ = n⋅λ (55.3)
dt grad f

The local representation (55.1) of S is not unique, or course. Indeed, if f satisfies (55.1),
so does –f, the induced unit normal of –f being –n. If the hypersurface S can be represented
globally by (55.1), i.e., there exists a smooth function whose domain contains the entire
hypersurface such that

x ∈ S ⇔ f (x) = 0 (55.4)

then S is called orientable. In this case S can be equipped with a smooth global unit normal
field n. (Of course, -n is also a smooth global unit normal field.) We say that S is oriented if a
particular smooth global unit normal field n has been selected and designated as the positive unit
normal of S . We shall consider oriented hypersurfaces only in this chapter.

407
408 Chap. 11 • HYPERSURFACES

Since grad f is nonvanishing, S can be characterized locally also by

x = x( y1 ,…, y N −1 ) (55.5)

in such a way that ( y1 ,… , y N −1 , f ) forms a local coordinate system in S . If n is the positive unit
normal and f satisfies (55.2), then the parameters ( y1 ,… , y N −1 ) are said to form a positive local
coordinate system in S when ( y1 ,… , y N −1 , f ) is a positive local coordinate system in E . Since
the coordinate curves of y Γ , Γ = 1,… , N − 1, are contained in S the natural basis vectors

h Γ = ∂x / ∂y Γ , Γ = 1,… , N − 1 (55.6)

are tangent to S . Moreover, the basis {h Γ , n} is positive in E . We call the ( N − 1) -

dimensional hyperplane S x spanned by {h Γ ( x )} the tangent plane of S at the point x ∈ S.

Since

h Γ ⋅ n = 0, Γ = 1,…, N − 1 (55.7)

The reciprocal basis of {h Γ , n} has the form {h Γ , n} where h Γ are also tangent to S , namely

h Γ ⋅ n = 0, Γ = 1,… , N − 1 (55.8)

and

h Γ ⋅ h Δ = δ ΔΓ , Γ, Δ = 1,…, N − 1 (55.9)

In view of (55.9) we call {h Γ } and {h Γ } reciprocal natural bases of ( y Γ ) on S .

Let v be a vector field on S , i. e., a function v; S → V . Then, for each x ∈ S , v can

be represented in terms of the bases {h Γ , n} and {h Γ , n} by

v = υ Γh Γ + υ N n = υ Γh Γ + υ N n (55.10)

where, from (55.7)-(55.9),

υ Γ = v ⋅ hΓ , υΓ = v ⋅ hΓ , υN = υ N = v ⋅ n (55.11)
Sec. 55 • Normal Vector, Tangent Plane, and Surface Metric 409

We call the vector field

vS ≡ υ Γ h Γ = υ Γ h Γ = v − ( v ⋅ n ) n (55.12)

the tangential projection of v , and we call the vector field

v n ≡ υ N n = υ N n = v − vS (55.13)

the normal projection of v . Notice that in (55.10)-(55.12) the repeated Greek index is summed
from 1 to N-1. We say that v is a tangential vector field on S if v n = 0 and a normal vector
field if vS = 0 .

If we introduce a local coordinate system ( x i ) in E , then the representation (55.5) may

be written

x i = x i ( y1 ,… , y N −1 ), i = 1,… , N (55.14)

From (55.14) the surface basis {h Γ } is related to the natural basis {g i } of ( x i ) by

∂x i
h Γ = hΓi g i = gi , Γ = 1,… , N − 1 (55.15)
∂y Γ

while from (55.2) the unit normal n has the component form

∂f / ∂x i
n= gi (55.16)
( g ab (∂f / ∂x a )(∂f / ∂x b ))1/ 2

where, as usual {g i } , is the reciprocal basis of {g i } and {g ab } is the component of the Euclidean
metric, namely

g ab = g a ⋅ g b (55.17)

The component representation for the surface reciprocal basis {h Γ } is somewhat harder
to find. We find first the components of the surface metric.

∂x i ∂x i
aΓΔ ≡ h Γ ⋅ h Δ = gij (55.18)
∂y Γ ∂y Δ
410 Chap. 11 • HYPERSURFACES

where gij is a component of the Euclidean metric

gij = g i ⋅ g j (55.19)

Now let ⎡⎣ a ΓΔ ⎤⎦ be the inverse of [ aΓΔ ] ,i.e.,

a ΓΔ aΔΣ = δ ΣΓ , Γ, Σ = 1,…, N − 1 (55.20)

The inverse matrix exists because from (55.18), [ aΓΔ ] is positive-definite and symmetric. In
fact, from (55.18) and (55.9)

a ΓΔ = h Γ ⋅ h Δ (55.21)

so that a ΓΔ is also a component of the surface metric. From (55.21) and (55.15) we then have

Γ ΓΔ ∂x i
h =a gi , Γ = 1,… , N − 1 (55.22)
∂y Δ

which is the desired representation for h Γ .

At each point x ∈ S the components aΓΔ ( x ) and a ΓΔ ( x ) defined by (55.18) and (55.21)
are those of an inner product on S x relative to the surface coordinate system ( y Γ ) , the inner
product being the one with induced by that of V since S x is a subspace of V . In other words
if u and v are tangent to S at x , say

u = u Γh Γ (x) v = υ Δ h Δ (x ) (55.23)

then
u ⋅ v = aΓΔ ( x )u Γυ Δ (55.24)

This inner product gives rise to the usual operations of rising and lowering of indices for tangent
vectors of S . Thus (55.23) 1 is equivalent to

u = aΓΔ ( x )u Γ h Δ ( x ) = uΔ h Δ ( x ) (55.25)
Sec. 55 • Normal Vector, Tangent Plane, and Surface Metric 411

Obviously we can also extend the operations to tensor fields on S having nonzero components
in the product basis of the surface bases {h Γ } and {h Γ } only. Such a tensor field A may be
called a tangential tensor field of S and has the representation

A = AΓ1 ... Γr h Γ1 ⊗ ⊗ h Γ2
(55.26)
= AΓ1 Γ2 ... Γr h Γ1 ⊗ h Γ2 ⊗ ⊗ h Γr , etc.

Then

AΓ1 Γ2 ... Γr = aΓ1Δ AΔΓ2 ... Γr , etc. (55.27)

There is a fundamental difference between the surface metric a on S and the Euclidean
metric g on E however. In the Euclidean space E there exist coordinate systems in which the
components of g are constant. Indeed, if the coordinate system is a rectangular Cartesian one,
then gij is δ ij at all points of the domain of the coordinate system. On the hypersurface S ,
generally, there need not be any coordinate system in which the components aΓΔ or a ΓΔ are
constant unless S happens to be a hyperplane. As we shall see in a later section, the departure
of a from g in this regard can be characterized by the curvature of S .

Another important difference between S and E is the fact that in S the tangent planes at
different points generally are different (N-1)-dimensional subspaces of V . Hence a vector in V
may be tangent to S at one point but not at another point. For this reason there is no canonical
parallelism which connects the tangent planes of S at distinct points. As a result, the notions of
gradient or covariant derivative of a tangential vector or tensor field on S must be carefully
defined, as we shall do in the next section.

The notions of Lie derivative and exterior derivative introduced in Sections 49 and 51,
however, can be readily defined for tangential fields of S . We consider first the Lie derivative.

Let v be a smooth tangent field defined on a domain in S . Then as before we say that a
smooth curve λ = λ (t ) in the domain of v is an integral curve if

dλ (t ) / dt = v( λ (t )) (55.28)

at all points of the curve. If we represent λ and v in component forms

yˆ( λ (t )) = (λ 1 (t ),… , λ N −1 (t )) and v = υ Γ h Γ (55.29)

relative to a surface coordinate system ( y Γ ) , then (55.28) can be expressed by

d λ Γ (t ) / dt = υ Γ ( λ (t )) (55.30)
412 Chap. 11 • HYPERSURFACES

Hence integral curves exist for any smooth tangential vector field v and they generate flow, and
hence a parallelism, along any integral curve. By the same argument as in Section 49, we define
the Lie derivative of a smooth tangential vector field u relative to v by the limit (49.14), except
that now ρt and Pt are the flow and the parallelism in S . Following exactly the same derivation
as before, we then obtain

⎛ ∂u Γ ∂υ Γ ⎞
L u = ⎜ Δ υ Δ − Δ u Δ ⎟ hΓ (55.31)
v
⎝ ∂y ∂y ⎠

which generalizes (49.21). Similarly if A is a smooth tangential tensor field on S , then L A is

v

defined by (49.41) with ρt and Pt as just explained, and the component representation for L A is
v

∂AΓ1 ... Γr Δ1 ... Δ s ∂υ Γ1

(S A) Γ1 ... Γr Δ1 ... Δ s = υ Σ − AΣΓ 2. ... Γ r
Δ1 ... Δ s
v ∂y Σ ∂y Σ
∂υ Γr ∂υ Σ
− − AΓ1 ... Γr −1Σ Δ1…Δ s + AΓ1 ... Γ r
ΣΔ 2…Δ s (55.32)
∂y Σ ∂y Δ1
∂υ Σ
+ + AΓ1 ... Γr Δ1…Δ s −1Σ
∂y Δ s

which generalizes (49.42).

Next we consider the exterior derivative. Let A be a tangential differential form on S ,

i.e., A is skew-symmetric and has the representations

A = AΓ1 ... Γr h Γ1 ⊗ ⊗ h Γr
= Σ AΓ1 ... Γr h Γ1 ∧ ∧ h Γr (55.33)
Γ1 <⋅⋅⋅<Γ r

1
= AΓ1 ... Γr h Γ1 ∧ ∧ h Γr
r!

Then we define the surface exterior derivative as d A

Sec. 55 • Normal Vector, Tangent Plane, and Surface Metric 413

N −1 AΓ1 ... Γr
dA = Σ Σ Δ
h Δ ∧ h Γ1 ∧ ∧ h Γr
Γ1 <⋅⋅⋅<Γ r Δ=1 ∂y
1 ∂AΓ1…Γr Δ
= h ∧ ∧ h Γr (55.34)
r ! ∂y Δ
1 ∂A
= δ ΣΔΓ1…1…ΣrΓ+1r Γ1…Δ Γr hΣ1 ∧ ∧ h Σ r +1
r !( r + 1)! ∂y

which generalizes (51.5). However, since the surface covariant derivative ∇A has not yet been
defined, we cannot write down the equation that generalizes (51.6) to the surface exterior
derivative. But we shall achieve this generalization in the next section. Other than this
exception, all results of Sections 49-52 can now be generalized and restated in an obvious way
for tangential fields on the hypersurface. In fact, those results are valid for differentiable
manifolds in general, so that they can be applied to Euclidean manifolds as well as to
hypersurfaces therein.

Excercises

55.1 For each x ∈ S define a linear transformation L x :V → V by

L x v = vS

for all v ∈V . Show that L x is an orthogonal projection whose image space is S x and
whose kernel is the one-dimensional subspace generated by n at x .

55.2 Show that

L x = h Γ ( x ) ⊗ h Γ (x ) = aΓΔ ( x )h Γ ( x ) ⊗ h Δ (x )
∂x i ∂x j
= a ΓΔ ( x ) gi (x ) ⊗ g j (x )
∂y Δ ∂y Γ

Thus L x is the linear transformation naturally isomorphic to the surface matric tensor at
x.

55.3 Show that

I = Lx + n ⊗ n

and
414 Chap. 11 • HYPERSURFACES

∂x i ∂x j
g ij ( x ) = a ΓΔ ( x ) Δ Γ
+ n i ( x )n j ( x )
∂y ∂y

55.4 Show that

∂x l Σ
L x g j ( x ) = g jl ( x ) h (x )
∂y Σ

and

∂x j Σ
Lxg j (x) = h (x)
∂y Σ

55.5 Compute the components aΓΔ for (a) the spherical surface defined by

x1 = c sin y1 cos y 2
x 2 = c sin y1 sin y 2
x 3 = c cos y1

and (b) the cylindrical surface

x1 =c cos y1 , x 2 =c sin y1 , x3 = y 2

where c is a constant.

55.6 For N=3 show that

a11 = a22 / a, a12 = − a12 / a , a 22 = a11 / a

where a = a11a22 − a122 = det [ aΔΓ ] .

55.7 Given an ellipsoid of revolution whose surface is determined by

x1 = b cos y1 sin y 2
x 2 = b sin y1 sin y 2
x 3 = c cos y1

where b and c are constants and b2 > c 2 show that

Sec. 55 • Normal Vector, Tangent Plane, and Surface Metric 415

a11 = b2 sin 2 y 2 , a12 = 0

and

a22 = b2 cos2 y 2 + c 2 sin 2 y 2

416 Chap. 11 • HYPERSURFACES

Section 56. Surface Covariant Derivatives

As mentioned in the preceding section, the tangent planes at different points of a
hypersurface generally do not coincide as subspaces of the translation space V of E. Consequently,
it is no longer possible to define the gradient or covariant derivative of a tangential tensor field
by a condition that formally generalizes (47.17). However, we shall see that it is possible to
define a type of differentiation which makes (47.17) formally unchanged. For a tangential tensor
field A represented by (55.26) we define the surface gradient ∇A by

∇A = AΓ1…Γr ,Δ h Γ1 ⊗ ⊗ h Γr ⊗ h Δ (56.1)

where

∂AΓ1…Γr ⎧ Γ1 ⎫ ⎧Γ ⎫
AΓ1…Γr ,Δ = + AΣΓ2…Γr ⎨ ⎬+ + AΓ1…Γr −1Σ ⎨ r ⎬ (56.2)
∂y Δ ⎩ΣΔ ⎭ ⎩ΣΔ ⎭

and

⎧ Ω ⎫ 1 ΩΣ ⎛ ∂aΓΣ ∂aΔΣ ∂aΓΔ ⎞

⎨ ⎬= a ⎜ Δ + Γ − Σ ⎟ (56.3)
⎩ ΓΔ ⎭ 2 ⎝ ∂y ∂y ∂y ⎠

⎧Ω⎫
The quantities ⎨ ⎬ are the surface Christoffel symbols. They obey the symmetric condition
⎩ ΓΔ ⎭

⎧Ω⎫ ⎧Ω⎫
⎨ ⎬=⎨ ⎬ (56.4)
⎩ ΓΔ ⎭ ⎩ ΔΓ ⎭

We would like to have available a formula like (47.26) which would characterize the surface
Christoffel symbols as components of ∂h Δ ∂y Γ . However, we have no assurance that
∂h Δ ∂y Γ is a tangential vector field. In fact, as we shall see in Section 58, ∂h Δ ∂y Γ is not
generally a tangential vector field. However, given the surface Christoffel symbols, we can
formally write

Dh Δ ⎧ Ω ⎫
= ⎨ ⎬ hΩ (56.5)
Dy Γ ⎩ ΓΔ ⎭

With this definition, we can combine (56.2) with (56.1) and obtain
Sec. 56 • Surface Covariant Derivatives 417

∂AΓ1…Γr
∇A = h Γ1 ⊗ ⊗ h Γr ⊗ h Δ
∂y Δ
(56.6)
⎧ Dh Γ1 Dh Γ r ⎫
+ AΓ1…Γr ⎨ Δ
⊗ ⊗ h Γr + + h Γ1 ⊗ ⊗ ⎬⊗h
Δ

⎩ Dy Dy Δ ⎭

If we adopt the notional convention

DAΓ1…Γr ∂AΓ1…Γr
= (56.7)
Dy Δ ∂y Δ

then (56.6) takes the suggestive form

DA
∇A = Δ
⊗ hΔ (56.8)
Dy

The components AΓ1…Γr ,Δ of ∇A represents the surface covariant derivative. If the mixed
components of A are used, say with

A = AΓ1…Γr Δ1…Δ s h Γ1 ⊗ ⊗ h Γr ⊗ h Δ1 ⊗ ⊗ hΔs (56.9)

then the components of ∇A are given by

Γ1 …Γ r
∂AΓ1…Γr Δ1…Δ s
A , =
Δ1 …Δ s Σ
∂y Σ
⎧Γ ⎫ ⎧ Γr ⎫
+ AΩΓ2 …Γr Δ1…Δ s ⎨ 1 ⎬ + + AΓ1…Γr −1Ω Δ1 Δs ⎨ ⎬ (56.10)
⎩ΩΣ ⎭ ⎩ΩΔ ⎭
⎧Ω ⎫ ⎧ Ω ⎫
− AΓ1…Γr ΩΔ2 …Δ s ⎨ ⎬− − AΓ1…Γr Δ1…Δ s−1Ω ⎨ ⎬
⎩ Δ1Σ ⎭ ⎩Δ s Σ ⎭

which generalizes (47.39). Equation (56.10) can be formally derived from (56.8) if we adopt the
definition

Dh Γ ⎧Γ⎫
Σ
= − ⎨ ⎬ hΔ (56.11)
Dy ⎩ΣΔ ⎭

When we apply (56.10) to the surface metric, we obtain

418 Chap. 11 • HYPERSURFACES

aΓΔ ,Σ = a ΓΔ ,Σ = δ ΔΓ,Σ = 0 (56.12)

which generalizes (47.40).

The formulas (56.2) and (56.10) give the covariant derivative of a tangential tensor field
in component form. To show that AΓ1…Γr ,Δ and AΓ1…Γr Δ1…Δ s ,Σ are the components of some
tangential tensor fields, we must show that they obey the tensor transformation rule, e.g., if
( y Γ ) is another surface coordinate system, then
∂y Γ1 ∂y Γr ∂y Ω Σ1…Σr
AΓ1…Γr ,Δ = A ,Ω (56.13)
∂y Σ1 ∂y Σr ∂y Δ

where AΓ1…Γr ,Δ are obtained from (56.2) when all the fields on the right-hand side are referred to
( y ).
Γ
To prove (56.13), we observe first that from (56.3) and the fact that aΓΔ and a ΓΔ are
components of the surface metric, so that

∂y Γ ∂y Δ ΣΘ ∂y Σ ∂y Ω
a ΓΔ = a , aΓΔ = aΣΩ (56.14)
∂y Σ ∂y Θ ∂y Γ ∂y Δ

we have the transformation rule

⎧ Σ ⎫ ∂y Σ ∂ 2 y Ω ∂y Σ ∂y Θ ∂y Φ ⎧ Ω ⎫
⎨ ⎬= Ω Γ Δ + Ω Γ Δ ⎨ ⎬ (56.15)
⎩ ΓΔ ⎭ ∂y ∂y ∂y ∂y ∂y ∂y ⎩ ΘΦ ⎭

which generalizes (47.35). Now using (56.15), (56.3), and the transformation rule

∂y Γ1 ∂y Γr Δ1…Δ r
AΓ1…Γr = A (56.16)
∂y Δ1 ∂y Δ r

we can verify that (56.13) is valid. Thus ∇A , as defined by (56.1), is indeed a tangential tensor
field.

The surface covariant derivative ∇A just defined is not the same as the covariant
derivative defined in Section 47. First, the domain of A here is contained in the hypersurface
⎧Σ⎫
S , which is not an open set in E . Second, the surface Christoffel symbols ⎨ ⎬ are generally
⎩ ΓΔ ⎭
nonvanishing relative to any surface coordinate system ( y Γ ) unless S happens to be a
hyperplane on which the metric components a ΓΔ and aΓΔ are constant relative to certain
Sec. 56 • Surface Covariant Derivatives 419

“Cartesian” coordinate systems. Other than these two points the formulas for the surface
covariant derivative are formally the same as those for the covariant derivative on E .

In view of (56.3) and (56.10), we see that the surface covariant derivative and the surface
exterior derivative are still related by a formula formally the same as (51.6), namely

d A = ( −1) r ( r + 1)!K r +1 (∇A) (56.17)

for any surface r-form. Here K r +1 denotes the surface skew-symmetric operator. That is, in
terms of any surface coordinate system ( y Γ )

1 Γ1 Γp Δp
Kp = δΔ Δp h Δ1 ⊗ ⊗h ⊗ h Γ1 ⊗ ⊗ hΓ p (56.18)
p! 1

for any integer p from 1 to N-1. Notice that from (56.18) and (56.10) the operator K p , like the
surface metric a , is a constant tangential tensor field relative to the surface covariant derivative,
i.e.,
Γ1…Γ p
δ Δ …Δ ,Σ = 0
1 p
(56.19)

As a result, the skew-symmetric operator as well as the operators of raising and lowering of
indices for tangential fields both commute with the surface covariant derivative.

In Section 48 we have defined the covariant derivative along a smooth curve λ in E .

That concept can be generalized to the surface covariant derivative along a smooth curve λ in
S . Specifically, let λ be represented by ( λ Γ ( t ) ) relative to a surface coordinate system ( y Γ ) in
S , and suppose that A is a tangential tensor field on λ represented by

A ( t ) = AΓ1…Γr (t )h Γ1 ( λ ( t ) ) ⊗ ⊗ h Γr ( λ ( t ) ) (56.20)

Then we define the surface covariant derivative by the formula

DA ⎡ dAΓ1…Γr ⎛ ΔΓ2…Γr ⎧ Γ1 ⎫ ⎧ Γ ⎫⎞ dλΣ ⎤

=⎢ +⎜A ⎨ ⎬+ + AΓ1…Γr ⎨ 1 ⎬ ⎟ ⎥ h Γ1 ⊗ ⊗ h Γr (56.21)
Dt ⎢⎣ dt ⎝ ⎩ ΔΣ ⎭ ⎩ ΔΣ ⎭ ⎠ dt ⎥⎦

which formally generalizes (48.6). By use of (56.15) and (56.16) we can show that the surface
covariant derivative D A / Dt along λ is a tangential tensor field on λ independent of the choice
of the surface coordinate system ( y Γ ) employed in the representation (56.21). Like the surface
covariant derivative of a field in S , the surface covariant derivative along a curve commutes
420 Chap. 11 • HYPERSURFACES

with the operations of raising and lowering of indices. Consequently, when the mixed
component representation is used, say with A given by

A ( t ) = AΓ1…Γr Δ1…Δ s ( t ) h Γ1 ( λ ( t ) ) ⊗ ⊗ h Γ1 ( λ ( t ) ) ⊗ h Δ1 ( λ ( t ) ) ⊗ ⊗ hΔs ( λ ( t ) ) (56.22)

the formula for DA / Dt becomes

DA ⎛ dA 1 r Δ1…Δ s
Γ …Γ
⎧Γ ⎫ ⎧Γ ⎫
=⎜ + ( AΣΓ2…Γr Δ1…Δ s ⎨ 1 ⎬ + + AΓ1…Γr −1Σ Δ1…Δ s ⎨ r ⎬
Dt ⎜⎝ dt ⎩ΣΩ ⎭ ⎩ΣΩ ⎭
(56.23)
⎧ Σ ⎫ ⎧ Σ ⎫ dλΩ ⎞
− AΓ1…Γr ΣΔ2…Δ s ⎨ ⎬ − − A Γ1…Γ r
Δ1…Δ s −1Σ ⎨ ⎬)
Δ1
⎟ h Γ1 ⊗ ⊗ h Γ1 ⊗ h ⊗ ⊗ hΔs
⎩ Δ1Ω ⎭ ⎩ Δ s Ω ⎭ dt ⎠

which formally generalizes (48.7).

As before, if A is a tangential tensor field and λ is a curve in the domain of A , then the
restriction of A on λ is a tensor of the form (56.20) or (56.22). In this case the covariant
derivative of A along λ is given by

DA ( λ ( t ) ) Dt = ⎡⎣∇A ( λ ( t ) )⎤⎦ λ ( t ) (56.24)

which generalizes (48.15). The component form of (56.24) is formally the same as (48.14). A
special case of (56.24) is equation (56.5), which is equivalent to

Dh Γ Dy Δ = [ ∇h Γ ] h Δ (56.25)

since from (56.10)

⎧ Σ ⎫
∇h Γ = ⎨ ⎬ h Σ ⊗ h Ω (56.26)
⎩ ΓΩ ⎭

If v is a tangential vector field defined on λ such that

Dv Dt = 0 (56.27)

then v may be called a constant field or a parallel field on λ . From (56.21), v is a parallel field
if and only if its components satisfy the equations of parallel transport:

dυ Γ ⎧Γ⎫
+ υ Δ ⎨ ⎬ λ Σ = 0, Γ = 1,…, N − 1 (56.28)
dt ⎩ ΔΣ ⎭
Sec. 56 • Surface Covariant Derivatives 421

⎧Γ⎫
Since λ Σ ( t ) and ⎨ ⎬ ( λ ( t ) ) are smooth functions of t, it follows from a theorem in ordinary
⎩ ΔΣ ⎭
differential equations that (56.28) possesses a unique solution

υ Δ = υ Δ (t ) , Δ = 1,… , N − 1 (56.29)

provided that a suitable initial condition

υ Δ ( 0 ) = υ0Δ , Δ = 1,… , N − 1 (56.30)

is specified. Since (56.28) is linear in v , the solution v ( t ) of (56.27) depends linearly on v ( 0 ) .

Thus there exists a linear isomorphism

ρ0,t : S λ (0) → S λ ( t ) (56.31)

defined by

ρ0,t ( v ( 0 ) ) = v ( t ) (56.32)

for all parallel fields v on λ . Naturally, we call ρ0,t the parallel transport along λ induced by
the surface covariant derivative.

The parallel transport ρ0,t preserves the surface metric in the sense that

ρ0,t ( v ( 0 ) ) ⋅ ρ0,t ( u ( 0 ) ) = v ( 0 ) ⋅ u ( 0 ) (56.33)

for all u ( 0 ) , v ( 0 ) in S λ (0) but generally ρ0,t does not coincide with the Euclidean parallel
transport on E through the translation space V . In fact since S λ (0) and S λ (t ) need not be the
same subspace in V , it is not always possible to compare ρ0,t with the Euclidean parallelism. As
we shall prove later, the parallel transport ρ0,t depends not only on the end points λ ( 0 ) and
λ ( t ) but also on the particular curve joining the two points. When the same two end points
λ ( 0 ) and λ ( t ) are joined by another curve μ in S , generally the parallel transport along μ from
λ ( 0 ) to λ ( t ) need not coincide with that along λ .

If the parallel transport ρt , t along λ from λ ( t ) to λ ( t ) is used, the covariant derivative

Dv Dt of a vector field on λ can be defined also by the limit of a difference quotient, namely
422 Chap. 11 • HYPERSURFACES

Dv v ( t + Δt ) − ρt ,t +Δt ( v ( t ) )
= lim (56.34)
Dt Δt →0 Δt

To prove this we observe first that

⎛ Γ dυ Γ ( t ) ⎞
v ( t + Δt ) = ⎜ υ ( t ) + Δt ⎟ h Γ ( λ ( t + Δt ) ) + o ( Δt ) (56.35)
⎝ dt ⎠

From (56.28) we have also

ρt ,t +Δt ( v ( t ) )
⎛ ⎧Γ⎫ ⎞ (56.36)
= ⎜ υ Γ ( t ) − υ Δ ( t ) ⎨ ⎬ ( λ ( t ) ) λ Σ ( t ) Δt ⎟ h Γ ( λ ( t + Δ t ) ) + o ( Δ t )
⎝ ⎩ ΔΣ ⎭ ⎠

Substituting these approximations into (56.34), we see that

v(t + Δt ) − ρt ,t +Δt ( v ( t ) ) ⎛ dυ Γ ⎧Γ ⎫ ⎞
lim =⎜ + υ Δ ⎨ ⎬ λ Σ ⎟ hΓ (56.37)
Δt → 0 Δt ⎝ dt ⎩ ΔΣ ⎭ ⎠

which is consistant with the previous formula (56.21) for the surface covariant derivative
Dv Dt of v along λ .

Since the parallel transport ρt , t is a linear isomorphism from S λ (t ) to S λ ( t ) , it gives rise to

various induced parallel transport for tangential tensors. We define a tangential tensor field A
on λ a constant field or a parallel field if

D A Dt = 0 (56.38)

Then the equations of parallel transport along λ for tensor fields of the forms (56.20) are

dAΓ1…Γr ⎛ ΔΓ2…Γr ⎧ Γ1 ⎫ ⎧ Γ ⎫⎞ dλΣ

+⎜A ⎨ ⎬+ + AΓ1…Γr −1Δ ⎨ r ⎬ ⎟ =0 (56.39)
dt ⎝ ⎩ ΔΣ ⎭ ⎩ ΔΣ ⎭ ⎠ dt

If we donate the induced parallel transport by P0,t , or more generally by Pt , t , then as before we
have

DA A ( t + Δt ) − Pt ,t +Δt ( A ( t ) )
= lim (56.40)
Dt Δt →0 Δt
Sec. 56 • Surface Covariant Derivatives 423

The coordinate-free definitions (56.34) and (56.40) demonstrate clearly the main
difference between the surface covariant derivative on S and the ordinary covariant derivative
on the Euclidean manifold E . Specifically, in the former case the parallelism used to compute
the difference quotient is the path-dependent surface parallelism, while in the latter case the
parallelism is simply the Euclidean parallelism, which is path-independent between any pair of
points in E .

In classical differential geometry the surface parallelism ρ , or more generally P, defined

by the equation (56.28) or (56.39) is known as the Levi-Civita parallelism or the Riemannian
parallelism. This parallelism is generally path-dependent and is determined completely by the
surface Christoffel symbols.

Exercises

56.1 Use (55.18), (56.5), and (56.4) and formally derive the formula (56.3).

56.2 Use (55.9) and (56.5) and the assumption that Dh Γ / Dy Σ is a surface vector field and
formally derive the formula (56.11).

56.3 Show that

∂h Δ Dh Δ ⎧ Ω ⎫
hΩ ⋅ Γ
= hΩ ⋅ =⎨ ⎬
∂y Dy Γ ⎩ ΓΔ ⎭

and

∂h Γ Dh Γ ⎧Γ⎫
hΔ ⋅ Σ
= h Δ ⋅ Σ
= −⎨ ⎬
∂y Dy ⎩ΣΔ ⎭

These formulas show that the tangential parts of ∂h Δ ∂y Γ and ∂h Γ ∂y Σ coincide with
Dh Δ Dy Γ and Dh Γ Dy Σ , respectively.

56.4 Show that the results of Exercise 56.3 can be written

⎛ ∂h ⎞ Dh Δ ⎛ ∂h Γ ⎞ Dh Γ
L ⎜ ΓΔ ⎟ = Γ
and L⎜ Σ ⎟ = Σ
⎝ ∂y ⎠ Dy ⎝ ∂y ⎠ Dy

where L is the field whose value at each x ∈ S is the projection L x defined in Exercise
55.1.
424 Chap. 11 • HYPERSURFACES

56.5 Use the results of Exercise 56.3 and show that

⎛ ∂A ⎞
∇A = L∗ ⎜ Δ ⎟ ⊗ h Δ
⎝ ∂y ⎠

where L∗ is the linear mapping induced by L .

56.6 Adopt the result of Exercise 56.5 and derive (56.10). This result shows that the above
formula for ∇A can be adopted as the definition of the surface gradient. One advantage
of this approach is that one does not need to introduce the formal operation DA Dy Δ . If
needed, it can simply be defined to be L∗ ( ∂A ∂y Δ ) .

56.7 Another advantage of adopting the result of Exercise 56.5 as a definition is that it can be
used to compute the surface gradient of field on S which are not tangential. For
example, each g k can be restricted to a field on S but it does not have a component
representation of the form (56.9). As an illustration of this concept, show that

⎧ j⎫ ∂x q ∂x l
∇g k = ⎨ ⎬ g jq Σ Δ hΣ ⊗ h Δ
⎩kl ⎭ ∂y ∂y

and

⎛ ∂2 xk ⎧ Σ ⎫ ∂x k ⎧ k ⎫ ∂x l ∂x j ⎞ ∂x s
∇L = ⎜ Γ Δ − ⎨ ⎬ Σ + ⎨ ⎬ Δ Γ ⎟ g ks Φ hΦ ⊗ h Δ ⊗ h Γ = 0
⎝ ∂y ∂y ⎩ ΔΓ ⎭ ∂y ⎩ jl ⎭ ∂y ∂y ⎠ ∂y

Note that ∇L = 0 is no more than the result (56.12).

56.8 Compute the surface Christoffel symbols for the surfaces defined in Excercises 55.5 and
55.7.

56.9 If A is a tensor field on E, then we can, of course, calculate its spatial gradient, grad A.
Also, we can restrict its domain S and compute the surface gradient ∇A . Show that

∇A = L∗ ( gradA )

56.10 Show that

∂ ( det [ aΔΓ ])
12
12 ⎧ Φ ⎫
Σ
= ( det [ aΔΓ ] ) ⎨ ⎬
∂y ⎩ΦΣ ⎭
Sec. 57 • Surface Geodesics, Exponential Map 425

Section 57. Surface Geodesics and the Exponential Map

In Section 48 we pointed out that a straight line λ in E with homogeneous parameter can
be characterized by equation (48.11), which means that the tangent vector λ of λ is constant
along λ , namely

dλ dt = 0 (57.1)

The same condition for a curve in S defines a surface geodesic. In terms of any surface
coordinate system ( y Γ ) the equations of geodesics are

d 2λ Γ d λ Σ d λ Δ ⎧Γ⎫
+ ⎨ ⎬ = 0, Γ = 1,… , N − 1 (57.2)
dt 2 dt dt ⎩ΣΔ ⎭

Since (57.2) is a system of second-order differential equations with smooth coefficients, at each
point

x0 = λ ( 0) ∈ S (57.3)

and in each tangential direction

v0 = λ ( 0) ∈ S x 0 (57.4)

there exists a unique geodesic λ = λ ( t ) satisfying the initial conditions (57.3) and (57.4).

In classical calculus of variations it is known that the geodesic equations (57.2) represent
the Euler-Lagrange equations for the arc length integral

( λ (t ) ⋅ λ (t ))
12
s (λ) = ∫
t1
dt
t0
12 (57.5)
⎛ dλ Γ dλ Δ ⎞
= ∫ ⎜ aΓΔ ( λ ( t ) )
t1

dt dt ⎟⎠
dt
t0
⎝

between any two fixed points

x = λ ( t0 ) , y = λ ( t1 ) (57.6)
426 Chap. 11 • HYPERSURFACES

on S . If we consider all smooth curves λ in S joining the fixed end points x and y , then the
ones satisfying the geodesic equations are curves whose arc length integral is an extremum in the
class of variations of curves.

To prove this, we observe first that the arc length integral is invariant under any change
of parameter along the curve. This condition is only natural, since the arc length is a geometric
property of the point set that constitutes the curve, independent of how that point set is
parameterized. From (57.1) it is obvious that the tangent vector of a geodesic must have constant
norm, since
D (λ ⋅ λ ) Dλ
=2 ⋅λ = 0⋅λ = 0 (57.7)
Dt Dt

Consequently, we seek only those extremal curves for (57.5) on which the integrand on the right-
hand of (57.5) is constant.

Now if we denote that integrand by

( ) (
L λ Γ , λ Γ ≡ aΓΔ ( λ Ω ) λ Γλ Δ )
12
(57.8)

then it is known that the Euler-Lagrange equations for (57.5) are

d ⎛ ∂L ⎞ ∂L
⎜ ⎟− = 0, Δ = 1,… , N − 1 (57.9)
dt ⎝ ∂λ Δ ⎠ ∂λ Δ

From (57.8) we have

∂L
∂λ Δ
=
1
2L
( 1
L
)
aΓΔ λ Γ + aΔΓλ Γ = aΓΔ λ Γ (57.10)

Hence on the extermal curves we have

d ⎛ ∂L ⎞ 1 ⎧ ∂aΓΔ Ω Γ Γ⎫
⎜ Δ ⎟ = ⎨ Ω λ λ + aΓΔ λ ⎬ (57.11)
dt ⎝ ∂λ ⎠ L ⎩ ∂y ⎭

where we have used the condition that L is constant on those curves. From (57.8) we have also

∂L 1 ∂aΓΩ Γ Ω
= λ λ (57.12)
∂λ Δ
2 L ∂yΔ

Combining (57.11) and (57.12), we see that the Euler-Lagrange equations have the explicit form
Sec. 57 • Surface Geodesics, Exponential Map 427

1⎡ 1 ⎛ ∂aΓΔ ∂aΩΔ ∂aΓΩ ⎞ Γ Ω ⎤

⎢ aΓΔ λ + ⎜ Ω + Γ − Δ ⎟ λ λ ⎥ = 0 (57.13)
L⎣ 2 ⎝ ∂y ∂y ∂y ⎠ ⎦

where we have used the symmetry of the product λ Γλ Ω with respect to Γ and Ω . Since
L ≠ 0 (otherwise the extremal curve is just one point), (57.13) is equivalent to

1 ⎛ ∂a ∂a ∂a ⎞
λ Θ + a ΘΔ ⎜ ΓΔ Ω
+ ΩΔΓ
− ΓΩΔ ⎟
λ Γλ Ω = 0 (57.14)
2 ⎝ ∂y ∂y ∂y ⎠

which is identical to (57.2) upon using the formula (56.3) for the surface Christoffel symbols.

The preceding result in the calculus of variations shows only that a geodesic is a curve
whose arc length is an extremum in a class of variations of curves. In terms of a fixed surface
coordinate system ( y Γ ) we can characterize a typical variation of the curve λ by N − 1 smooth
functions η Γ ( t ) such that

η Γ ( t0 ) = η Γ ( t1 ) = 0, Γ = 1,… , N − 1 (57.15)

A one-parameter family of variations of λ is then given by the curses λ α with representations.

λαΓ ( t ) = λ Γ ( t ) + αη Γ ( t ) , Γ = 1,… , N − 1 (57.16)

From (57.15) the curves λ α satisfy the same end conditions (57.6) as the curve λ , and
λ α reduces to λ when α = 0. The Euler-Lagrange equations express simply the condition that

ds ( λ a )
=0 (57.17)
d α α =0

for all choice of η Γ satisfying (57.15). We note that (57.17) allows α = 0 to be a local minimum,
a local maximum, or a local minimax point in the class of variations. In order for the arc length
integral to be a local minimum, additional conditions must be imposed on the geodesic.

It can be shown, however, that in a sufficient small surface neighborhood N x0 of any

point x 0 in S every geodesic is, in fact, a curve of minimum arc length. We shall omit the
proof of this result since it is not simple. A consequence of this result is that between any pair of
points x and y in the surface neighborhood N x0 there exists one and only one geodesic λ (aside
428 Chap. 11 • HYPERSURFACES

from a change of parameter) which lies entirely in N x0 . That geodesic has the minimum arc
kength among all curves in S not just curves in N x0 , joining x to y .

Now, as we have remarked earlier in this section, at any point x 0 ∈ S , and in each
tangential direction v 0 ∈ S x0 there exists a unique geodesic λ satisfying the conditions (57.3) and
(57.4). For definiteness, we donate this particular geodesic by λ v0 . Since λ v0 is smooth and,
when the norm of v 0 is sufficiently small, the geodesic λ v0 ( t ) , t ∈ [0,1] , is contained entirely in
the surface neighborhood N x0 of x 0 . Hence there exists a one-to-one mapping

expxo :Bxo → N xo (57.18)

defined by
exp x0 ( v ) ≡ λ v (1) (57.19)

for all v belonging to a certain small neighborhood Bx0 of 0 in S x0 . We call this injection the
exponential map at x 0 .

Since λ v is the solution of (57.2), the surface coordinates ( y Γ ) of the point

y ≡ exp x0 ( v ) = λ v (1) (57.20)

depends smoothly on the components υ Γ of v relative to ( y Γ ) ,i.e., there exists smooth functions

y Γ = expxΓ0 (υ 1 ,… ,υ N −1 ) , Γ = 1,… , N − 1 (57.21)

where (υ Γ ) can be regarded as a Cartesian coordinate system on Bx0 induced by the basis
{h ( x )} , namely,
Γ 0

v = υ Γh Γ ( x 0 ) (57.22)

In the sense of (57.21) we say that the exponential map expx 0 is smooth.

Smoothness of the exponential map can be visualized also from a slightly different point
of view. Since N x0 is contained in S , which is contained in E , expx0 can be regarded also as a
mapping from a domain Bx0 in an Euclidean space S x0 to the Euclidean space E , namely

expx0 : Bx0 → E (57.23)

Sec. 57 • Surface Geodesics, Exponential Map 429

Now the smoothness of expx0 has the usual meaning as defined in Section 43. Since the surface
coordinates ( y Γ ) can be extended to a local coordinate system ( y Γ , f ) as explained in Section
55, smoothness in the sense of (57.21) is consistant with that of (57.23).

As explained in Section 43, the smooth mapping expx0 has a gradient at any point v in the
( )
domain Bx0 . In particular, at v = 0, grad expx0 ( 0 ) exists and corresponds to a linear map

grad expx0 ( 0 ) : S x0 → V (57.24)

We claim that the image of this linear map is precisely the tangent plane S x0 , considered as a
( )
subspace of V ; moreover, grad expx0 ( 0 ) is simply the identity map on S x0 . This fact is more
( )
or less obvious, since by definition the linear map grad expx0 ( 0 ) is characterized by the
condition that

⎣ ( 0 ⎦ )
⎡ grad exp x ( 0 ) ⎤ ( v ( 0 ) ) = d expx ( v ( t ) )
dt 0
(57.25)
t =0

for any curve v = v ( t ) such that v ( 0 ) = 0 . In particular, for the straight lines

v ( t ) = vt (57.26)

with v ∈Bx0 , we have v ( 0 ) = v and

exp x0 ( vt ) = λ vt (1) = λ v ( t ) (57.27)

so that
d
expx0 ( vt ) = v (57.28)
dt
Hence

⎣ ( 0 ⎦ )
⎡ grad expx ( 0 ) ⎤ ( v ) = v, v ∈Bx0 (57.29)

( )
But since grad expx0 ( 0 ) is a linear map, (57.29) implies that the same holds for all v ∈ S x0 ,
and thus
430 Chap. 11 • HYPERSURFACES

( )
grad expx0 ( 0 ) = idS x
0
(57.30)

It should be noted, however, that the condition (57.30) is valid only at the origin 0 of S x0 ; the
same condition generally is not true at any other point v ∈Bx0 .

Now suppose that {e Γ , Γ = 1,… , N − 1} is any basis of S x0 . Then as usual it gives rise to a
Cartesian coordinate system ( w Γ ) on S x0 by the component representation

w = w Γe Γ (57.31)

for any w ∈ S x0 . Since expx0 is a smooth one-to-one map, it carries the Cartesian system
( w ) on B
Γ
x0 to a surface coordinate system ( z Γ ) on the surface neighborhood N x0 of x 0 in S .
For definiteness, this coordinate system ( z Γ ) is called a canonical surface coordinate system at
x 0 . Thus a point z ∈ N x0 has the coordinates ( z Γ ) if and only if

z = expx0 ( z Γe Γ ) (57.32)

Relative to a canonical surface coordinate system a curve λ passing through x 0 at t = 0 is a

surface geodesic if and only if its representation λ Γ ( t ) in terms of ( z Γ ) has the form

λ Γ ( t ) = υ Γt , Γ = 1,… , N − 1 (57.33)

for some constant υ Γ . From (57.32) the geodesic λ is simply the one denoted earlier by λ v ,
where
v = υ Γe Γ (57.34)

An important property of a canonical surface coordinate system at x 0 is that the

corresponding surface Christoffel symbols are all equal to zero at x 0 . Indeed, since any curve λ
given by (57.33) is a surface geodesic, the geodesic equations imply

⎧Σ⎫ Γ Δ
⎨ ⎬υ υ = 0 (57.35)
⎩ ΓΔ ⎭

where the Christoffel symbols are evaluated at any point of the curve λ . In particular, at t = 0
we get
Sec. 57 • Surface Geodesics, Exponential Map 431

⎧Σ⎫
⎨ ⎬ ( x 0 )υ υ = 0
Γ Δ
(57.36)
ΓΔ
⎩ ⎭

Now since υ Γ is arbitrary, by use of the symmetry condition (56.4) we conclude that

⎧Σ⎫
⎨ ⎬ ( x0 ) = 0 (57.37)
⎩ ΓΔ ⎭

In general, a surface coordinate system ( y Γ ) is called a geodesic coordinate system at a

point x 0 if the Christoffel symbols corresponding to ( y Γ ) vanish at x 0 . From (57.37), we see
that a canonical surface coordinate system at x 0 is a geodesic coordinate system at the same
point. The converse, of course, is not true. In fact, from the transformation rule (56.15) a
surface coordinate system ( y Δ ) is geodesic at x 0 if and only if its coordinate transformation
relative to a canonical surface coordinate system ( z Γ ) satisfies the condition

∂2 y Δ
=0 (57.38)
∂z Γ∂z Ω x0

which is somewhat weaker that the transformation rule

y Δ = eΓΔ z Γ (57.39)

for some nonsingular matrix ⎡⎣ eΓΔ ⎤⎦ when ( y Δ ) is also a canonical surface coordinate system at x 0 ,
the ⎡⎣ eΓΔ ⎤⎦ being simply the transformation matrix connecting the basis {e Γ } for ( z Γ ) and thebasis
{eΓ } for ( y Γ ) .

A geodesic coordinate system at x 0 plays a role similar to that of a Cartesian coordinate

system in E . In view of the condition (57.37), we see that the representation of the surface
covariant derivative at x 0 reduces simply to the partial derivative at x 0 , namely

∂AΓ1…Γr ( x 0 )
∇A ( x 0 ) = e Γ1 ⊗ ⊗ e Γ1 ⊗ e Δ (57.40)
∂z Δ
432 Chap. 11 • HYPERSURFACES

It should be noted, however, that (57.40) is valid at the point x 0 only, since a geodesic
coordinate system at x 0 generally does not remain a geodesic coordinate system at any
neighboring point of x 0 . Notice also that the basis {e Γ } in S x0 is the natural basis of ( z Δ ) at x 0 ,
this fact being a direct consequence of the condition (57.30).

In closing, we remark that we can choose the basis {e Γ } in S x0 to be an orthonormal basis.

In this case the corresponding canonical surface coordinate system ( z Γ ) satisfies the additional
condition

aΓΔ ( x 0 ) = δ ΓΔ (57.41)

Then we do not even have to distinguish the contravariant and the covariant component of a
tangential tensor at x 0 . In classical differential geometry such a canonical surface coordinate
system is called a normal coordinate system or a Riemannian coordinate system at the surface
point under consideration.
Sec. 58 • Surface Curvature, I 433

Section 58. Surface Curvature, I. The Formulas of Weingarten and Gauss

In the preceding sections we have considered the surface covariant derivative at

tangential vector and tensor fields on S . We have pointed out that this covariant derivative is
defined relative to a particular path-dependent parallelism on S , namely the parallelism of Levi-
Civita as defined by (56.28) and (56.39). We have remarked repeatedly that this parallelism is
not the same as the Euclidean parallelism on the underlying space E in which S is embedded.
Now a tangential tensor on S , or course, is also a tensor over E being merely a tensor having
nonzero components only in the product basis of {h Γ } , which can be regarded as a subset of the
basis {h Γ , n} for V . Consequently, the spatial covariant derivative of a tangential tensor field
along a curve in S is defined. Similarly, the special covariant derivative of the unit normal n of
S along a curve in S is also defined. We shall study these spatial covariant derivatives in this
section.

In classical differential geometry the spatial covariant derivative of a tangential field is a

special case of the total covariant derivative, which we shall consider in detail later. Since the
spatial covariant derivative and the surface covariant derivative along a surface curve often
appear in the same equation, we use the notion d dt for the former and D Dt for the latter.
However, when the curve is the coordinate curve of y Γ , we shall write the covariant derivative as
∂ ∂y Γ and D Dy Γ , respectively. It should be noted also that d dt is defined for all tensor fields
on λ , whether or not the field is tangential, while D Dt is defined only for tangential fields.

We consider first the covariant derivative dn dt of the unit normal field of S on any
curve λ ∈ S . Since n is a unit vector field and since d dt preserves the spatial metric on S , we
have

dn dt ⋅ n = 0 (58.1)

Thus dn dt is a tangential vector field. We claim that there exists a symmetric second-order
tangential tensor field B on S such that

dn dt = − Bλ (58.2)

for all surface curves λ . The proof of (58.2) is more or less obvious. From (55.2), the unit
normal n on S is parallel to the gradient of a certain smooth function f , and locally S can be
characterized by (55.1). By normalizing the function f to a function

f (x)
w (x) = (58.3)
grad f ( x )
434 Chap. 11 • HYPERSURFACES

we find

n ( x ) = grad w ( x ) , x ∈S (58.4)

Now for the smooth vector field grad w we can apply the usual formula (48.15) to compute the
spatial covariant derivative ( d dt )( grad w ) along any smooth curve in E . In particular, along
the surface curve λ under consideration we have the formula (58.2), where

B = − grad ( grad w ) S (58.5)

As a result, B is symmetric. The fact that B is a tangential tensor has been remarked after
(58.1).

From (58.2) the surface tensor B characterizes the spatial change of the unit normal n of S .
Hence in some sense B is a measurement of the curvature of S in E . In classical differential
geometry, B is called the second fundamental form of S , the surface metric a defined by
(55.18) being the first fundamental form. In component form relative to a surface coordinate
system ( y Γ ) , B can be represented as usual

B = bΓΔ h Γ ⊗ h Δ = b ΓΔ h Γ ⊗ h Δ = bΔΓ h Γ ⊗ h Δ (58.6)

From (58.2) the components of B are those of ∂n ∂y Γ taken along the y Γ -curve in S , namely

∂n ∂y Γ = −bΓΔ h Δ = −bΓΔ h Δ (58.7)

This equation is called Weingarten’s formula in classical differential geometry.

Next we consider the covariant derivatives dh Γ dt and dh Γ dt of the surface natural basis
vectors h Γ and h Γ along any curve λ in S . From (56.21) and (56.23) we have

Dh Γ ⎧ Δ ⎫ Σ
= ⎨ ⎬ λ hΔ (58.8)
Dt ⎩ ΓΣ ⎭
and
Dh Γ ⎧Γ⎫
= − ⎨ ⎬ λ ΣhΔ (58.9)
Dt ⎩ ΔΣ ⎭

By a similar argument as (58.2), we have first

Sec. 58 • Surface Curvature, I 435

dh Γ
= −CΓ λ (58.10)
dt

where CΓ is a symmetric spatial tensor field on S . In particular, when λ is the coordinate curve
of y Δ , (58.10) reduces to

∂h Γ
= −CΓ h Δ (58.11)
∂y Δ

Since CΓ is symmetric, this equation implies

∂h Γ ∂h Γ
−hΣ ⋅ = − h Δ ⋅ = CΓ ( h Δ , hΣ ) (58.12)
∂y Δ ∂y Σ

We claim that the quantity given by this equation is simply the surface Christoffel symbol,
⎧Γ⎫
⎨ ⎬
⎩ ΔΣ ⎭
∂h Γ ⎧ Γ ⎫
−hΣ ⋅ Δ = ⎨ ⎬ (58.13)
∂y ⎩ ΔΣ ⎭

Indeed, since both d dt and D Dt preserve the surface metric, we have

∂δ ΣΓ ∂ ( h Σ ⋅ h )
Γ
∂h Γ ∂hΣ Γ
0= Δ = = hΣ ⋅ Δ + Δ ⋅ h (58.14)
∂y ∂y Δ ∂y ∂y

Thus (58.13) is equivalent to

∂hΣ ⎧ Γ ⎫
hΓ ⋅ =⎨ ⎬ (58.15)
∂y Δ ⎩ΣΔ ⎭

But as in (58.14) we have also

∂aΓΔ ∂ ( h Γ ⋅ h Δ ) ∂h ∂h
Σ
= Σ
= h Γ ⋅ ΣΔ + h Δ ⋅ ΣΓ
∂y ∂y ∂y ∂y
(58.16)
⎛ ∂h ⎞ ⎛ ∂h ⎞
= aΓΩ ⎜ hΩ ⋅ ΣΔ ⎟ + aΔΩ ⎜ hΩ ⋅ ΣΓ ⎟
⎝ ∂y ⎠ ⎝ ∂y ⎠
436 Chap. 11 • HYPERSURFACES

Further, from (58.14) and (58.12) we have

∂h Δ ∂h
hΩ ⋅ Σ
= hΩ ⋅ ΔΣ (58.17)
∂y ∂y

Comparing (58.17) and (58.16) with (56.4) and (56.12), respectively, we see that (58.15) holds.

On differentiating (55.8), we get

∂ ( n ⋅ hΓ ) ∂h Γ ∂n
0= = n⋅ + hΓ ⋅ Δ (58.18)
∂y Δ ∂y Δ
∂y

Substituting Weingarten’s formula (58.7) into (58.18), we then obtain

∂h Γ
n⋅ = bΔΓ (58.19)
∂y Δ

The formulas (58.19) and (58.13) determine completely the spatial covariant derivative of
h Γ along any y Δ -curve:

∂h Γ Γ ⎧Γ⎫ Σ Γ Dh Γ
= bΔ n − ⎨ ⎬ h = bΔ n + (58.20)
∂y Δ ⎩ ΔΣ ⎭ Dy Δ

where we have used (58.9) for (58.20) 2 . By exactly the same argument we have also

∂h Γ ⎧Σ ⎫ Dh Γ
Δ
= bΓΔ n + ⎨ ⎬ hΣ = bΓΔ n + (58.21)
∂y ⎩ ΓΔ ⎭ Dy Δ

As we shall see, (58.20) and (58.21) are equivalent to the formula of Gauss in classical
differential geometry.

The formulas (58.20), (58.21), and (58.7) determine completely the spatial covariant
derivatives of the bases {h Γ , n} and {h Γ , n} along any curve λ in S . Indeed, if the coordinates
of λ in ( y Γ ) are ( λ Γ ) , then we have
Sec. 58 • Surface Curvature, I 437

dn
= −bΓΔ λ Δ h Γ = −bΔΓλ Δ h Γ
dt
dh Γ ⎧Γ⎫ Dh Γ
= bΔΓλ Δ n − ⎨ ⎬ λ Δ hΣ = bΔΓλ Δ n + (58.22)
dt ⎩ ΔΣ ⎭ Dt
dh Γ ⎧Σ ⎫ Dh Γ
= bΓΔ λ Δ n + ⎨ ⎬ λ Δ hΣ = bΓΔ λ Δ n +
dt ⎩ ΓΔ ⎭ Dt

From these representations we can compute the spatial covariant derivative of any vector or
tensor fields along λ when their components relative to {h Γ , n} or {h Γ , n} are given. For
example, if v is a vector field having the component form

v ( t ) = υn ( t ) n ( λ ( t ) ) + υ Γ ( t ) hΓ ( λ ( t ) ) (58.23)

along λ , then

dv ⎛ ⎧Γ⎫ ⎞
( )
= υ n + bΓΔυ Γλ Δ n + ⎜υ Γ − υ n bΔΓλ Δ + ⎨ ⎬υ Σλ Δ ⎟ h Γ (58.24)
dt ⎝ ⎩ΣΔ ⎭ ⎠

In particular, if v is a tangential field, then (58.24) reduces to

dv ⎛ ⎧Γ⎫ ⎞
= bΓΔυ Γλ Δ n + ⎜υ Γ + ⎨ ⎬υ Σλ Δ ⎟ h Γ
dt ⎝ ⎩ΣΔ ⎭ ⎠ (58.25)
Dv
= bΓΔυ Γλ Δ n +
Dt

where we have used (56.21). Equation (58.25) shows clearly the difference between d dt and
D Dt for any tangential field.

Applying (58.25) to the tangent vector λ of λ , we get

dλ Dλ
= B ( λ, λ ) n + (58.26)
dt Dt

In particular, when λ is a surface geodesic, then (58.26) reduces to

dλ dt = B ( λ , λ ) n (58.27)
438 Chap. 11 • HYPERSURFACES

Comparing this result with the classical notions of curvature and principal normal of a curve (cf.
Section 53), we see that the surface normal is the principal normal of a surface curve λ if and
only if λ is a surface geodesic; further, in this case the curvature of λ is the quadratic form
B ( s, s ) of B in the direction of the unit tangent s of λ .

In general, if λ is not a surface geodesic but it is parameterized by the arc length, then
(58.26) reads

ds Ds
= B ( s, s ) n + (58.28)
ds Ds

where s denotes the unit tangent of λ , as usual. We call the norms of the three vectors in
(58.28) the spatial curvature, the normal curvature, and the geodesic curvature of λ and denote
them by κ , κ n , and κ g , respectively, namely

ds Ds
κ= , κ n = B ( s, s ) , κg = (58.29)
ds Ds

Then (58.28) implies

κ 2 = κ n2 + κ g2 (58.30)

Further, if n0 and n g are unit vectors in the directions of ds ds and Ds Ds , then

κ n0 = κ n n + κ g n g (58.31)

Or equivalently,

n 0 = ( κ n κ ) n + (κ g κ ) n g (58.32)

which is called Meunier’s equation.

At this point we can define another kind of covariant derivative for tensor fields on S .
As we have remarked before, the tangent plane S x of S is a subspace of V . Hence a tangential
vector v can be regarded either as a surface vector in S x or as a special vector in V . In the
former sense it is natural to use the surface covariant derivative Dv Dt , while in the latter sense
we may consider the spatial covariant derivative dv dt along any smooth curve λ in S . For a
tangential tensor field A , such as the one represented (56.20) or (56.22), we may choose to
recognize certain indices as spatial indices and the remaining ones as surface indices. In this
Sec. 58 • Surface Curvature, I 439

case it becomes possible to define a covariant derivative tha is a mixture of d dt and D Dt . In

classical differential geometry this new kind of covariant derivative is called the total covariant
derivative.

More specifically, we first consider a concrete example to explain this concept. By

definition, the surface metric tensor a is a constant tangential tensor relative to the surface
covariant derivative. In terms of any surface coordinate system ( y Γ ) , a can be represented by

a = aΓΔ h Γ ⊗ h Δ = h Δ ⊗ h Δ = a ΓΔ h Γ ⊗ h Δ (58.33)
and we have

Da D ( h Δ ⊗ h )
Δ
Dh Δ Dh Δ
0= = = hΔ ⊗ + ⊗ hΔ (58.34)
Dt Dt Dt Dt

The tensor a , however, can be regarded also as the inclusion map A of S x in V , namely

A : S x →V (58.35)

in the sense that for any surface vector v ∈ S x , Av ∈V is given by

Av = ( h Δ ⊗ h Δ ) ( v ) = h Δ ( h Δ ⋅ v ) = υ Δ h Δ (58.36)

In (58.36) it is more natural to regard the first basis vector h Δ in the product basis h Δ ⊗ h Δ as a
spatial vector and the second basis vector h Δ as a surface vector. In fact, in classical differential
geometry the tensor A given by (58.35) is often denoted by

∂x i
A = hΔi g i ⊗ h Δ = g ⊗ hΔ
Δ i
(58.37)
∂y

where ( x i ) is a spatial coordinate system whose natural basis is {g i } . When the indices are
recognized in this way, it is natural to define a total covariant derivative of A , denoted by
δ A δ t , by

δ A dh Δ Dh Δ
= ⊗ hΔ + hΔ ⊗
δt dt Dt
(58.38)
d ⎛ ∂x i ⎞ Δ Dh Δ
= ⎜ Δ gi ⎟ ⊗ h + hΔ ⊗
dt ⎝ ∂y ⎠ Dt
440 Chap. 11 • HYPERSURFACES

Since a mixture of d dt and D Dt is used in defining δ δ t , it is important to recognize the

“spatial” or “surface” designation of the indices of the components of a tensor before the total
covariant derivative is computed.

In classical differential geometry the indices are distinguished directly in the notation of
the components. Thus a tensor A represented by

A ( t ) = Aikj……ΔΓ…… ( t ) g i ⊗ g j ⊗ g k ⊗ ⊗ hΓ ⊗ hΔ (58.39)

has an obvious interpretation: the Latin indices i, j, k … are designated as “spatial” and the Greek
indices Γ, Δ,… are “surface.” So δ A δ t is defined by

δ A d ik…Γ…
= ( Aj…Δ… g i ⊗ g j ⊗ g k ⊗ )⊗h Γ ⊗ hΔ
δ t dt
+ Aikj……ΔΓ……g i ⊗ g j ⊗ g k ⊗ ⊗
D
Dt
( hΓ ⊗ hΔ )
⎛d ⎞
= ⎜ Aikj……ΔΓ…… ⎟ g i ⊗ g j ⊗ g k ⊗ ⊗ hΓ ⊗ hΔ (58.40)
⎝ dt ⎠
⎡d
+ Aikj……ΔΓ…… ⎢ ( g i ⊗ g j ⊗ g k ⊗ )⊗h Γ ⊗ hΔ
⎣ dt

+gi ⊗ g j ⊗ g k ⊗ ⊗
D
Dt
( hΓ ⊗ hΔ )⎤⎥⎦

The derivative dAikj……ΔΓ…… dt is the same as DAikj……ΔΓ…… Dt , of course, since for scalars there is but
one kind of parallelism along any curve. In particular, we can compute explicitly the
representation for δ A δ t as defined by (58.38):

δ A d ⎛ ∂x i ⎞ ∂x i ⎛ dg Dh Δ ⎞
= ⎜ Δ ⎟ gi ⊗ hΔ + Δ ⎜ i ⊗ hΔ + gi ⊗ ⎟
δ t dt ⎝ ∂y ⎠ ∂y ⎝ dt Dt ⎠
(58.41)
⎛ ∂2 xi ∂x j ∂x k ⎧ i ⎫ ∂x i ⎧ Σ ⎫ ⎞
= ⎜ Δ Γ + Δ Γ ⎨ ⎬ − Σ ⎨ ⎬ λ Γg i ⊗ h Δ ⎟
⎝ ∂y ∂y ∂y ∂y ⎩ jk ⎭ ∂y ⎩ ΓΔ ⎭ ⎠

As a result, when λ is the coordinate curve of ( y Γ ) , (58.41) reduces to

δ A ⎛ ∂ 2 xi ∂x j ∂x k ⎧ i ⎫ ∂x i ⎧ Σ ⎫⎞ Δ
= ⎜ + ⎨ ⎬− Σ ⎨ ⎬ ⎟ gi ⊗ h (58.42)
δ y Γ ⎝ ∂y Δ ∂y Γ ∂y Δ ∂y Γ ⎩ jk ⎭ ∂y ⎩ ΓΔ ⎭ ⎠
Sec. 58 • Surface Curvature, I 441

Actually, the quantity on the right-hand side of (58.42) has a very simple representation.
Since δ A δ y Γ can be expressed also by (58.38) 1 , namely

δ A ∂h Δ Δ Dh Δ
= ⊗ h + h Δ ⊗ (58.43)
δ y Γ ∂y Γ Dy Γ

from (58.21) we have

δA Δ Dh Δ Δ Dh Δ
= bΔΓ n ⊗ h + ⊗ h + h Δ ⊗
δ yΓ Dy Γ Dy Γ

Γ ( Δ
h ⊗ hΔ )
D
= bΔΓ n ⊗ h Δ + (58.44)
Dy
= bΔΓ n ⊗ h Δ

where we have used (58.34). The formula (58.44) 3 is known as Gauss’ formula in classical
differential geometry. It is often written in component form

x i ;ΓΔ = bΓΔ n i (58.45)

where the semicolon in the subscript on the left-hand side denotes the total covariant derivative.

Exercises

58.1 Show that

1 ⎛ ∂n ∂x ∂x ∂n ⎞
(a) bΓΔ = − ⎜ Δ ⋅ Γ + Γ ⋅ Γ ⎟
2 ⎝ ∂y ∂y ∂y ∂y ⎠

(b) bΔΓ = gij x i ;ΔΓ n j

∂ 2x
(c) bΔΓ = n ⋅
∂y Δ ∂y Γ

58.2 Compute the quantities bΔΓ for the surfaces defined in Exercises 55.5 and 55.7.

58.3 Let A be a tensor field of the form

A = A j1 jr
Γ1 Γ s g j1 ⊗ ⊗ g jr ⊗ h Γ1 ⊗ ⊗ h Γs
442 Chap. 11 • HYPERSURFACES

show that

δA
= Aj 1 jr
Γ1 Γ s ;Δ g j1 ⊗ ⊗ g jr ⊗ h Γ1 ⊗ ⊗ h Γs
δy Δ

where

∂A j1… jr Γ1…Γs
A j1 jr
Γ1 Γ s ;Δ =
∂y Δ
r
⎧ jβ ⎫ j j lj j ∂x k
+ ∑ ⎨ ⎬ A 1 β −1 β +1 r Γ1…Γs Δ
β =1 ⎩ lk ⎭ ∂y
s
⎧ Λ ⎫ j1 jr
−∑ ⎨ ⎬A Γ1 Γβ −1ΛΓβ +1 Γs
β =1 ⎩ Γ β Δ ⎭

Similar formulas can be derived for other types of mixed tensor fields defined on S .

58.4 Show that (58.7) can be written

∂x i
n i ;Γ = −bΓΔ
∂y Δ
Sec. 59 • Surface Curvature, II 443

Section 59. Surface Curvature, II. The Riemann-Christoffel Tensor and the Ricci
Identities

In the preceding section we have considered the curvature of S by examining the change
of the unit normal n of S in E . This approach is natural for a hypersurface, since the metric on
S is induced by that of E . The results of this approach, however, are not entirely intrinsic to
S , since they depend not only on the surface metric but also on the particular imbedding of S
into E . In this section, we shall consider curvature from a more intrinsic point of view. We
seek results which depend only on the surface metric. Our basic idea is that curvature on S
corresponds to the departure of the Levi-Civita parallelism on S from a Euclidean parallelism.

We recall that relative to a Cartesian coordinate system x̂ on E the covariant derivative

of a vector field v has the simplest representation

∂υ i
grad v = υ i , j g i ⊗ g j = gi ⊗ g j (59.1)
∂x j

Hence if we take the second covariant derivatives, then in the same coordinate system we have

∂ 2υ i
grad ( grad v ) = υ , jk g i ⊗ g ⊗ g = j k g i ⊗ g j ⊗ g k
i j k
(59.2)
∂x ∂x

In particular, the second covariant derivatives satisfy the same symmetry condition as that of the
ordinary partial derivatives:

υ i , jk = υ i ,kj (59.3)

Note that the proof of (59.3) depends crucially on the existence of a Cartesian coordinate
system relative to which the Christoffel symbols of the Euclidean parallelism vanish identically.
For the hypersurface S in general, the geodesic coordinate system at a reference point is the
closest counterpart of a Cartesian system. However, in a geodesic coordinate system the surface
Christoffel symbols vanish at the reference point only. As a result the surface covariant
derivative at the reference point x 0 still has a simple representation like (59.1)

grad v ( x 0 ) = υ Γ ,Δ ( x 0 ) e Γ ( x 0 ) ⊗ e Δ ( x 0 )
∂υ Γ (59.4)
= Δ eΓ ( x0 ) ⊗ eΔ ( x0 )
∂z x0

[cf. (57.40)]. But generally, in the same coordinate system, the representation (59.4)(59.4) does
not hold any neighboring point of x 0 . This situation has been explained in detail in Section 57.
444 Chap. 11 • HYPERSURFACES

In particular, there is no counterpart for the representation (59.2) 2 on S . Indeed the surface
second covariant derivatives generally fail to satisfy the symmetry condition (59.3) valid for the
spatial covariant derivatives.

To see this fact we choose an arbitrary surface coordinate system ( y Γ ) with natural basis
{h Γ } and {h Γ } on S as before. From (56.10) the surface covariant derivative of a tangential
vector field

v = υ Γh Γ (59.5)

is given by

∂υ Γ ⎧Γ⎫
υ Γ ,Δ = Δ
+υΣ ⎨ ⎬ (59.6)
∂y ⎩ΣΔ ⎭

Applying the same formula again, we obtain

∂ ⎛ ∂υ Γ Σ ⎧ Γ ⎫⎞ ⎛ ∂υ Φ Σ ⎧ Φ ⎫⎞ ⎧ Γ ⎫
υ Γ ,ΔΩ = Ω ⎜ Δ
+ v ⎨ ⎬⎟ ⎜ Δ +υ ⎨ ⎬⎟ ⎨
+ ⎬
∂y ⎝ ∂y ⎩ΣΔ ⎭ ⎠ ⎝ ∂y ⎩ΣΔ ⎭ ⎠ ⎩ΦΩ ⎭
⎛ ∂υ Γ ⎧ Γ ⎫⎞ ⎧ Ψ ⎫
− ⎜ Ψ +υΣ ⎨ ⎬⎟ ⎨ ⎬
⎝ ∂y ⎩ΣΨ ⎭ ⎠ ⎩ ΔΩ ⎭
(59.7)
∂ 2υ Γ ∂υ Σ ⎧ Γ ⎫ ∂υ Φ ⎧ Γ ⎫ ∂υ Γ ⎧Ψ⎫
= Δ Ω + Ω ⎨ ⎬+ Δ ⎨ ⎬− Ψ ⎨ ⎬
∂y ∂y ∂y ⎩ΣΔ ⎭ ∂y ⎩ΦΩ ⎭ ∂y ⎩ ΔΩ ⎭
⎛⎧ Φ ⎫⎧ Γ ⎫ ⎧ Γ ⎫⎧ Ψ ⎫ ∂ ⎧ Γ ⎫⎞
+υ Σ ⎜ ⎨ ⎬ ⎨ ⎬− ⎨ ⎬⎨ ⎬+ Ω ⎨ ⎬⎟
⎝ ⎩ΣΔ ⎭ ⎩ΦΩ ⎭ ⎩ΣΨ ⎭ ⎩ ΔΩ ⎭ ∂y ⎩ΣΔ ⎭ ⎠

In particular, even if the surface coordinate system reduces to a geodesic coordinate system
( y Γ ) at a reference point x0 , (59.7) can only be simplified to
∂ 2υ Γ ∂ ⎧Γ⎫
υ Γ ,ΔΩ ( x 0 ) = + υ Σ ( x0 ) ⎨ ⎬ (59.8)
∂z Δ ∂z Ω X0 ∂z Ω ⎩ΣΔ ⎭ X0

which contains a term in addition to the spatial case (59.2))2. Since that additional term
generally need not be symmetric in the pair ( Δ, Ω ) , we have shown that the surface second
covariant derivatives do not necessarily obey the symmetry condition (59.3).
Sec. 59 • Surface Curvature, II 445

From (59.7) if we subtract υ,ΓΩΔ from υ,ΓΔΩ then the result is the following commutation
rule:

υ Γ ,ΔΩ −υ Γ ,ΩΔ = −υ Σ R Γ ΣΔΩ (59.9)

where

∂ ⎧ Γ⎫ ∂ ⎧ Γ ⎫ ⎧ Φ ⎫⎧ Γ ⎫ ⎧ Φ ⎫ ⎧ Γ ⎫
R Γ ΣΔΩ ≡ Δ ⎨ ⎬− Ω ⎨ ⎬+ ⎨ ⎬⎨ ⎬− ⎨ ⎬ ⎨ ⎬ (59.10)
∂y ⎩ΣΩ ⎭ ∂y ⎩ΣΔ ⎭ ⎩ΣΩ ⎭ ⎩ΦΔ ⎭ ⎩ΣΔ ⎭ ⎩ΦΩ ⎭

Since the commutation rule is valid for all tangential vector fields v, (59.9) implies that under a
change of surface coordinate system the fields R Γ ΣΔΩ satisfy the transformation rule for the
components of a fourth-order tangential tensor. Thus we define

R ≡ R Γ ΣΔΩ h Γ ⊗ hΣ ⊗ h Δ ⊗ hΩ (59.11)

and we call the tensor R the Riemann-Christoffel tensor of L . Notice that R depends only on
the surface metrica a , since its components are determined completely by the surface Christoffel
symbols. In particular, in a geodesic coordinates system ( z Γ ) at x 0 (59.10) simplifies to

∂ ⎧ Γ⎫ ∂ ⎧ Γ⎫
R Γ ΣΔΩ ( x 0 ) = ⎨ ⎬ − ⎨ ⎬ (59.12)
∂z Δ ⎩ΣΩ ⎭ X0 ∂z Ω ⎩ΣΔ ⎭ X0

In a certain sense the Riemann-Christoffel tensor R characterizes locally the departure of

the Levi-Civita parallelism from a Euclidean one. We have the following result.

Theorem 59.1. The Riemann-Christoffel tensor R vanishes identically on a neighborhood of a

point x 0 ∈ S if and only if there exists a surface coordinate system ( z Γ ) covering x 0 relativwe
⎧ Γ⎫
to which the surface Christoffel symbols ⎨ ⎬ vanishing identically on a neighborhood of x 0 .
⎩ ΔΣ ⎭
In view of the representation (59.10), the sufficiency part of the preceding theorem is
obvious. To prove the necessity part, we observe first the following lemma.

Lemma. The Riemann-Christoffel tensor R vanishes identically near x 0 if and only if the
system of first-order partial differential equations

∂υ Γ ⎧ Γ ⎫
0 = υ Γ ,Δ = Δ
+υ Ω ⎨ ⎬, Γ = 1,… N − 1 (59.13)
∂y ⎩ΩΔ ⎭
446 Chap. 11 • HYPERSURFACES

can be integrated near x 0 for each prescribed initial value

υ Γ ( x 0 ) = υ0Γ , Γ = 1,… N − 1 (59.14)

Proof. Necessity. Let {v Σ , Σ = 1,… , N − 1} be linearly independent and satisfy (59.13). Then we
can obtain the surface Christoffel symbols from

∂υΣΓ ⎧ Γ⎫
0 = Δ + υΣΩ ⎨ ⎬ (59.15)
∂y ⎩ΩΔ ⎭
and
⎧ Γ⎫ Σ ∂υ Σ
Γ

⎨ ⎬ = −uΩ (59.16)
⎩ΩΔ ⎭ ∂y Δ

where ⎡⎣ uΩΣ ⎤⎦ denotes the inverse matrix of ⎡⎣υΣΩ ⎤⎦ . Substituting (59.16) into (59.10), we obtain
directly

R Γ ΣΔΩ = 0 (59.17)

Sufficiency. It follows from the Frobenius theorem that the conditions of integrability for
the system of first-order partial differential equations

∂υ Γ ⎧ Γ⎫
Δ
= −υ Ω ⎨ ⎬ (59.18)
∂y ⎩ΩΔ ⎭

are

∂ ⎛ Ω ⎧ Γ ⎫⎞ ∂ ⎛ Ω ⎧ Γ ⎫⎞
⎜υ ⎨ ⎬ ⎟ − ⎜υ ⎨ ⎬ ⎟ = 0, Γ = 1,… N − 1 (59.19)
∂y Σ ⎝ ⎩ΩΔ ⎭ ⎠ ∂y Δ ⎝ ⎩ΩΣ ⎭ ⎠

If we expand the partial derivatives in (59.19) and use (59.18), then (59.17) follows as a result of
(59.10). Thus the lemma is proved.

Now we return to the proof of the necessity part of the theorem. From (59.16) and the
condition (56.4) we see that the basis {v Σ } obeys the rule

∂υΣΓ Δ ∂υΩΓ Δ
υΩ − Δ υΣ = 0 (59.20)
∂y Δ ∂y
Sec. 59 • Surface Curvature, II 447

These equations are the coordinate forms of

[ v Σ , v Ω ] = 0, Σ, Ω = 1,… , N − 1 (59.21)

As a result there exists a coordinate system ( z Γ ) whose natural basis coincides with {v Σ } . In
particular, relative to ( z Γ ) (59.16) reduces to

⎧ Γ⎫ Ω ∂δ Σ
Γ

⎨ ⎬ = −δ Σ =0 (59.22)
⎩ΩΔ ⎭ ∂z Δ

Thus the theorem is proved.

It should be noted that tangential vector fields satisfying (59.13) are parallel fields
relative to the Levi-Civita parallelism, since from (56.24) we have the representation

Dv ( λ ( t ) ) / Dt = υ Γ ,Δ λ Δ h Γ (59.23)

along any curve λ. Consequently, the Levi-Civita parallelism becomes locally path-
independent. For definiteness, we call this a locally Euclidean parallelism or a flat parallelism.
Then the preceding theorem asserts that the Levi-Civita parallelism on S is locally Euclidean if
and only if the Riemann-Christoffel tensor R based on the surface metric vanishes.

The commutation rule (59.9) is valid for tangential vector fields only. However, we can
easily generalize that rule to the following Ricci identities:

AΓ1 ... Γr Δ1 ... Δ s ,ΣΩ − AΓ1 ... Γr Δ1 ... Δ s ,ΩΣ

= − AΦΓ2 ... Γr Δ1 ... Δ s R Γ1 ΦΣΩ
(59.24)
− − AΓ1 ... Γr −1Φ Δ1 ... Δ s R Γr ΦΣΩ + AΓ1 ... Γr ΨΔ2 ... Δ s R Ψ Δ1ΣΩ
+ + AΓ1 ... Γr Δ1 ... Δ s −1Ψ R Ψ Δ sΣΩ

As a result, (59.17) is also the integrability condition of the system

AΓ1 ... Γr Δ1 ... Δ s ,Σ = 0 (59.25)

for each prescribed initial value at any reference point x 0 ; further, a solution of (59.25)
corresponds to a parallel tangential tensor field on the developable hyper surface S .
448 Chap. 11 • HYPERSURFACES

In classical differential geometry a surface S is called developable if its Riemann-

Christoffel curvature tensore vanishes identically.
Sec. 60 • Surface Curvature, III 449

Section 60. Surface Curvature, III. The Equations of Gauss and Codazzi

In the preceding two sections we have considered the curvature of S from both the
extrinsic point of view and the intrinsic point of view. In this section we shall unite the results of
these two approaches.

Our starting point is the general representation (58.25) applied to the y Δ -coordinate
curve:

∂v Dv
Δ
= bΓΔυ Γ n + (60.1)
∂y Dy Δ

for an arbitrary tangential vector field v. This representation gives the natural decomposition of
the spatial vector field ∂v / ∂y Δ on S into a normal projection bΓΔυ Γ n and a tangential
projection Dv / Dy Δ . Applying the spatial covariant derivative ∂ / ∂y Σ along the y Σ− coordinate
curve to (60.1)(60.1), we obtain

∂ ⎛ ∂v ⎞ ∂ ∂ ⎛ Dv ⎞
Σ ⎜ Δ ⎟
= Σ ( bΓΔυ Γ n ) + Σ ⎜ Δ ⎟
∂y ⎝ ∂y ⎠ ∂y ∂y ⎝ Dy ⎠
(60.2)
∂ D ⎛ Dv ⎞
= Σ ( bΓΔυ Γ n ) + bΓΣυ Γ ,Δ n +
∂y Dy Σ ⎜⎝ Dy Δ ⎟⎠

where we have applied (60.1) in (60.2)2 to the tangential vector field Dv / Dy Δ . Now, since the
spatial parallelism is Euclidean, the lefthand side of (60.2) is symmetric in the pair ( Σ, Δ ) ,
namely
∂ ⎛ ∂v ⎞ ∂ ⎛ ∂v ⎞
Σ ⎜ Δ ⎟
= Δ⎜ Σ⎟ (60.3)
∂y ⎝ ∂y ⎠ ∂y ⎝ ∂y ⎠

Hence from (60.2) we have

D ⎛ Dv ⎞ D ⎛ Dv ⎞ ∂
Δ ⎜ Σ ⎟
− Σ ⎜ Δ ⎟
= Σ ( bΓΔυ Γ n ) − ( bΓΣυ Γ n )
Dy ⎝ Dy ⎠ Dy ⎝ Dy ⎠ ∂y (60.4)
+ ( bΓΣυ Γ
,Δ − bΓΔυ ,Σ ) n
Γ

This is the basic formula from which we can extract the relation between the Riemann-
Christoffel tensor R and the second fundamental form B.
450 Chap. 11 • HYPERSURFACES

Specifically, the left-hand side of (60.4) is a tangaential vector having the component
form

D ⎛ Dv ⎞ D ⎛ D v ⎞
− = (υ Γ ,ΔΣ − υ Γ ,ΣΔ ) h Γ
Dy Δ ⎜⎝ Dy Σ ⎟⎠ Dy Σ ⎜⎝ Dy Δ ⎟⎠ (60.5)
= −υ Ω R Γ ΩΣΔ h Γ

as required by (59.9), while the right-hand side is a vector having the normal projection

⎛ ∂ ( bΓΔυ Γ ) ∂ ( bΓΣυ Γ ) ⎞
⎜ − + b υ Γ
− b υ Γ
,Σ ⎟ n (60.6)
⎜ ∂y Σ ∂y Δ
ΓΣ ,Δ ΓΔ
⎟
⎝ ⎠

and the tangential projection

( −b ΓΔ υ ΓbΣΦ + bΓΣυ ΓbΔΦ ) hΦ (60.7)

where we have used Weingarten’s formula (58.7) to compute the covariant derivatives
∂n / ∂y Σ and ∂n / ∂y Δ . Consequently, (60.5) implies that

∂ ( bΓΔυ Γ ) ∂ ( bΓΣυ Γ )
Σ
− Δ
+ bΓΣυ Γ ,Δ − bΓΔυ Γ ,Σ = 0 (60.8)
∂y ∂y

and that

υ Γ R Φ ΓΣΔ = bΓΔυ Γ bΣΦ − bΓΣυ ΓbΔΦ (60.9)

for all tangential vector fields v.

If we choose v = h Γ , then (60.9) becomes

R Φ ΓΣΔ = bΓΔ bΣΦ − bΓΣbΔΦ (60.10)

or, equivalently,

RΦΓΣΔ = bΓΔ bΦΣ − bΓΣbΦ Δ (60.11)

Sec. 60 • Surface Curvature, III 451

Thus the Riemann-Christoffel tensor R is completely determined by the second

fundamental form. The important result (60.11) is called the equations of Gauss in classical
differential geometry. A similar choice for v. in (60.8) yields

∂bΓΔ ⎧ Φ ⎫ ∂bΓΣ ⎧ Φ⎫
− bΦΔ ⎨ ⎬ − + bΦΣ ⎨ ⎬=0 (60.12)
∂y Σ ⎩ ΓΣ ⎭ ∂y
Δ
⎩ ΓΔ ⎭

By use of the symmetry condition (56.4) we can rewrite (60.12) in the more elegant form

bΓΔ ,Σ − bΓΣ ,Δ = 0 (60.13)

which are the equations of Codazzi.

The importance of the equations of Gauss and Codazzi lies not only in their uniting the
second fundamental form with the Riemann Christoffel tensor for any hypersurfaces S in E ,
but also in their being the conditions of integrability as asserted by the following theorem.

Theorem 60.1. Suppose that aΓΔ and bΓΔ are any presceibed smooth functions of ( y Ω ) such
that [ aΓΔ ] is positive-definite and symmetric, [ bΓΔ ] is symmetric, and together [ aΓΔ ] and [ bΓΔ ]
satisfy the equations of Gauss and Codazzi. Then locally there exists a hyper surface S with
representation

xi = xi ( y Ω ) (60.14)

on which the prescribed aΓΔ and bΓΔ are the first and second fundamental forms.

Proof. For simplicity we choose a rectangular Cartesian coordinate system x i in E . Then in

component form (58.21) and (58.7) are represented by

⎧ Σ⎫
∂hΓi / ∂y Δ = bΓΔ n i + ⎨ ⎬ hΣi , ∂n i / ∂y Γ = −bΓΔ hΔi (60.15)
⎩ ΓΔ ⎭

We claim that this system can be integrated and the solution preserves the algebraic conditions

δ ij hΓi hΔj = aΓΔ , δ ij hΓi n j = 0, δ ij n i n j = 1 (60.16)

In particular, if (60.16) are imposed at any one reference point, then they hold identically on a
neighborhood of the reference point.
452 Chap. 11 • HYPERSURFACES

The fact that the solution of (60.15) preserves the conditions (60.16) is more or less
obvious. We put

f ΓΔ ≡ δ ij hΓi hΔj − aΓΔ , f Γ ≡ δ ij hΓi n j , f ≡ δ ij n i n j − 1 (60.17)

Then initially at some point ( y01 , y02 ) we have

f ΓΔ ( y01 , y02 ) = f Γ ( y01 , y02 ) = f ( y01 , y02 ) = 0 (60.18)

Now from (60.15) and (60.17) we can verify easily that

∂f ΓΔ ∂f Γ ⎧ Δ⎫ ∂f
= bΓΣ f Δ + bΔΣ f Γ , = − bΣΔ f ΓΔ + ⎨ ⎬ f Δ + bΓΣ f , = −2bΣΔ f Δ (60.19)
∂y Σ ∂y Σ
⎩ ΓΣ ⎭ ∂y Σ

along the coordinate curve of any y Σ . From (60.18) and (60.19) we see that f ΓΔ , f Γ , and f
must vanish identically.

Now the system (60.15) is integrable, since by use of the equations of Gauss and Codazzi we
have

∂ ⎛ ∂hΓi ⎞ ∂ ⎛ ∂hΓi ⎞
Σ ⎜ Δ ⎟
− Δ ⎜ Σ ⎟ = ( R Ω ΓΣΔ − bΓΣbΔΩ + bΓΔ bΣΩ ) hΩi
∂y ⎝ ∂y ⎠ ∂y ⎝ ∂y ⎠ (60.20)
+ ( bΓΣ ,Δ − bΓΔ ,Σ ) n i = 0

and

∂ ⎛ ∂n i ⎞ ∂ ⎛ ∂n i ⎞
⎟ = ( bΣ ,Γ − bΓ ,Σ ) hΩ = 0
Ω Ω
⎟−
i
⎜ ⎜ (60.21)
∂y Σ ⎝ ∂y Γ ⎠ ∂y Γ ⎝ ∂y Σ ⎠

Hence locally there exist functions hΓi ( y Ω ) and n i ( y Ω ) which verify (60.15) and (60.16).

Next we set up the system of first-order partial differential equations

∂x i / ∂y Δ = hΔi ( y Ω ) (60.22)

where the right-hand side is the solution of (60.15) and (60.16). Then (60.22) is also integrable,
since we have
Sec. 60 • Surface Curvature, III 453

∂ ⎛ ∂x i ⎞ ∂hΔi ⎧ Ω⎫ i ∂ ⎛ ∂x i ⎞
⎜ ⎟ = = bΔΣ n i
+ ⎨ ⎬ Ω h = ⎜ ⎟ (60.23)
∂y Σ ⎝ ∂y Δ ⎠ ∂y Σ ⎩ ΔΣ ⎭ ∂y Δ ⎝ ∂y Σ ⎠

Consequently the solution (60.14) exists; further, from (60.22), (60.16), and (60.15) the first and
the second fundamental forms on the hypersurface defined by (60.14) are precisely the
prescribed fields aΓΔ and bΓΔ , respectively. Thus the theorem is proved.

It should be noted that, since the algebraic conditions (60.16) are preserved by the
solution of (60.15), any two solutions ( hΓi , n j ) and ( hΓi , n j ) of (60.15) can differ by at most a
transformation of the form

hΓi = Q ij hΓj , n i = Q ij n j (60.24)

where ⎡⎣Q ij ⎤⎦ is a constant orthogonal matrix. This transformation corresponds to a change of

rectangular Cartesian coordinate system on E used in the representation (60.14). In this sense
the fundamental forms aΓΔ and bΓΔ , together with the equations of Gauss andCodazzi, determine
the hyper surface S locally to within a rigid displacement in E .

While the equations of Gauss show that the Riemann Christoffel tensor R of S is
completely determined by the second fundamental form B, the converse is generally not true.
Thus mathematically the extrinsic characterization of curvature is much stronger than the
intrinsic one, as it should be. In particular, if B vanishes, then S reduces to a hyperplane which
is trivially developable so that R vanishes. On the other hand, if R vanishes, then S can be
developed into a hyperplane, but generally B does not vanish, so S need not itself be a
hyperplane. For example, in a three-dimensional Euclidean space a cylinder is developable, but
it is not a plane.

Exercise

60.1 In the case where N = 3 show that the only independent equations of Codazzi are

bΔΔ ,Γ − bΔΓ ,Δ = 0

for Δ ≠ Γ and where the summation convention has been abandoned. Also show that the
only independent equation of Gauss is the single equation

R1212 = b11b22 − b122 = det B

454 Chap. 11 • HYPERSURFACES

Section 61. Surface Area, Minimal Surface

Since we assume that S is oriented, the tangent plane S x of each point x ∈ S is

equipped with a volume tensor S . Relative to any positive surface coordinate system ( y Γ ) with
natural basis {h Γ } , S can be represented by

S = a h1 ∧ ∧ h N −1 (61.1)

where

a = det [ aΓΔ ] (61.2)

If U is a domain in S covered by ( y Γ ) , then the surface area σ (U ) is defined by

σ (U ) = ∫ ⋅⋅ ⋅∫ ady1 dy N −1 (61.3)
U

where the ( N − 1) -fold integral is taken over the coordinates ( y Γ ) on the domain U . Since a
obeys the transformation rule

⎡ ∂y Γ ⎤
a = a det ⎢ Δ ⎥ (61.4)
⎣ ∂y ⎦

relative to any positive coordinate systems ( y Γ ) and ( y Γ ) , the value of σ (U ) is independent of

the choice of coordinates. Thus σ can be extended into a positive measure on S in the sense of
measure theory.

We shall consider the theory of integration relative to σ in detail in Chapter 13. Here we
shall note only a particular result wich connects a property of the surface area to the mean
curvature of S . Namely, we seek a geometric condition for S to be a minimal surface, which
is defined by the condition that the surface area σ (U ) be an extremum in the class of variations
of hypersurfaces having the same boundary as U . The concept of minimal surface is similar to
that of a geodesic whichwe have explored in Section 57, except that here we are interested in the
variation of the integral σ given by (61.3) instead of the integral s given by (57.5).

As before, the geometric condition for a minimal surface follows from the Euler-
Lagrange equation for (61.3), namely,
Sec. 61 • Surface Area, Minimal Surface 455

∂ ⎛∂ a ⎞ ∂ a
⎜ ⎟− =0 (61.5)
∂y Γ ⎝ ∂hΓi ⎠ ∂x i

where a is regarded as a function of x i and hΓi through the representation

∂x i ∂x j
aΓΔ = δ ij hΓi hΔi = δ ij (61.6)
∂y Γ ∂y Δ

For simplicity we have chosen the spatial coordinates to be rectangular Cartesian, so that
aΓΔ does not depend explicitly on x i . From (61.6) the partial derivative of a with respect to
hΓi is given by

∂ a
= a a ΓΔ δ ij hΔi = a δ ij h Γj (61.7)
∂hΓi

Substituting this formula into (61.5), we see that the condition for S to be a minimal surface is

⎛⎧ Ω ⎫ ∂h Γj ⎞
a δ ij ⎜ ⎨ ⎬ h Γj + Γ ⎟ = 0, i = 1,… , N (61.8)
⎝ ⎩ΩΓ ⎭ ∂y ⎠

or, equivalently, in vector notation

⎛⎧ Ω ⎫ ∂h Γ ⎞
a ⎜ ⎨ ⎬ hΓ + Γ ⎟ = 0 (61.9)
⎝ ⎩ΩΓ ⎭ ∂y ⎠

In deriving (61.8) and (61.9), we have used the identity

∂ a ⎧Ω ⎫
=⎨ ⎬ a (61.10)
∂y Γ ⎩ΩΓ ⎭

which we have noted in Exercise 56.10. Now from (58.20) we can rewrite (61.9) in the simple
form

a bΓΓ n = 0 (61.11)
456 Chap. 11 • HYPERSURFACES

As a result, the geometric condition for a minimal surface is

I B = tr B = bΓΓ = 0 (61.12)

In general, the first invariant I B of the second fundamental form B is called the mean
curvature of the hyper surface. Equation (61.12) then asserts that S is a minimal surface if and
only if its mean curvature vanishes.

Since in general B is symmetric, it has the usual spectral decomposition

N −1
B = ∑ β Γc Γ ⊗ c Γ (61.13)
Γ=1

where the principal basis {c Γ } is orthonormal on the surface. In differential geometry the
direction of c Γ ( x ) at each point x ∈ S is called a principal direction and the corresponding
proper number β Γ ( x ) is called a principal (normal) curvature at x. As usual, the principal
(normal) curvatures are the extrema of the normal curvature

κ n = B ( s, s ) (61.14)

in all unit tangents s at any point x [cf. (58.29)2]. The mean curvature I B , of course, is equal to
the sum of the principal curvatures, i.e.,

N −1
IB = ∑ β Γ (61.15)
Γ=1
Sec. 62 • Three-Dimensional Euclidean Manifold 457

Section 62. Surfaces in a Three-Dimensional Euclidean Manifold

In this section we shall apply the general theory of hypersurfaces to the special case of a
two-dimensional surface imbedded in a three-dimensional Euclidean manifold. This special case
is the commonest case in application, and it also provides a good example to illustrate the
general results.

As usual we can represent S by

xi = xi ( y Γ ) (62.1)

where i ranges from 1 to 3 and Γ ranges from 1 to 2. We choose ( x i ) and ( y Γ ) to be positive

spatial and surface coordinate systems, respectively. Then the tangent plane of S is spanned by
the basis

∂x i
hΓ ≡ gi , Γ = 1, 2 (62.2)
∂y Γ

and the unit normal n is given by

h1 × h 2
n= (62.3)
h1 × h 2

The first and the second fundamental forms aΓΔ and bΓΔ relative to ( y Γ ) are defined
by

∂h Γ ∂n
aΓΔ = h Γ ⋅ h Δ . bΓΔ = n ⋅ Δ
= −h Γ ⋅ Δ (62.4)
∂y ∂y

These forms satisfy the equations of Gauss and Codazzi:

RΦΓΣΔ = bΓΔ bΦΣ − bΓΣbΦ Δ , bΓΔ ,Σ − bΓΣ ,Δ = 0 (62.5)

Since bΓΔ is symmetric, at each point x ∈ S there exists a positive orthonormal basis
{c ( x )} relative to which B ( x ) can be represented by the spectral form
Γ

B ( x ) = β1 ( x ) c1 ( x ) ⊗ c1 ( x ) + β 2 ( x ) c 2 ( x ) ⊗ c 2 ( x ) (62.6)
458 Chap. 11 • HYPERSURFACES

The proper numbers β1 ( x ) and β 2 ( x ) are the principal (normal) curvatures of S at x. In

general {c Γ ( x )} is not the natural basis of a surface coordinate system unless

[c1 , c 2 ] = 0 (62.7)

However, locally we can always choose a coordinate system ( z Γ ) in such a way that the natural
basis {h Γ } is parallel to {c Γ ( x )} . Naturally, such a coordinate system is called a principal
coordinate system and its coordinate curves are called the lines of curvature. Relative to a
principal coordinate system the components aΓΔ and bΓΔ satisfy

b12 = a12 = 0 and h1 = a11 c1 , h 2 = a22 c 2 (62.8)

The principal invariants of B are

I B = tr B = β1 + β 2 = a ΓΔ bΓΔ
(62.9)
II B = det B = β1β 2 = det ⎡⎣bΔΓ ⎤⎦

In the preceding section we have defined I B to be the mean curvature. Now II B is called the
Gaussian curvature. Since

bΔΓ = a ΓΩbΔΩ (62.10)

we have also

det [ bΔΩ ] det [bΔΩ ]

II B = = (62.11)
det [ aΔΩ ] a

where the numerator on the right-hand side is given by the equation of Gauss:

det [bΔΩ ] = b11b22 − b122 = R1212 (62.12)

Notice that for the two-dimensional case the tensor R is completely determined by R1212 , since
from (62.5)1, or indirectly from (59.10), RΦΓΣΔ vanishes when Φ = Γ or when Σ = Δ.

We call a point x ∈ S elliptic, hyperbolic, or parabolic if the Gaussian curvature threre

is positive, negative, or zero, respectively. These terms are suggested by the following geometric
considerations: We choose a fixed reference point x 0 ∈ S and define a surface coordinate
Sec. 62 • Three-Dimensional Euclidean Manifold 459

system ( y Γ ) such that the coordinates of x 0 are ( 0, 0 ) ) and the basis {h Γ ( x )} is orthonormal,
i.e.,

aΓΔ ( 0,0 ) = δ ΓΔ (62.13)

In this coordinate system a point x ∈ S having coordinate ( y1 , y 2 ) near ( 0, 0 ) is located at a

normal distance d ( y1 , y 2 ) from the tangenet plane S xo with d ( y1 , y 2 ) given approximately by

d ( y1 , y 2 ) = n ( x 0 ) ⋅ ( x − x 0 )
⎛ 1 Dh Γ ⎞
≅ n ( x0 ) ⋅ ⎜ hΓ ( x0 ) y Γ + yΓ yΔ ⎟ (62.14)
⎜ 2 Dy Δ ⎟
⎝ x0 ⎠
1
≅ bΓΔ ( x 0 ) y Γ y Δ
2

where (62.14)2,3 are valid to within an error of third or higher order in y Γ . From this estimate
we see that the intersection of S with a plane parallel to S xo is represented approximately by
the curve

bΓΔ ( x 0 ) y Γ y Δ = const (62.15)

which is an ellipse, a hyperbola, or parallel lines when x 0 is an elliptic, hyperbolic, or parabolic

point, respectively. Further, in each case the principal basis {c Γ } of B ( x 0 ) coincides with the
principal axes of the curves in the sense of conic sections in analytical geometry. The estimate
(62.14) means also that in the rectangular Cartesian coordinate system ( x i ) induced by the
orthonormal basis {h1 ( x 0 ) , h 2 ( x 0 ) , n}, the surface S can be represented locally by the
equations

1
x1 = y 1 , x2 = y2 , x3 ≅ bΓΔ ( x 0 ) y Γ y Δ (62.16)
2

As usual we can define the notion of conjugate directions relative to the symmetric
bilinear form of B ( x 0 ) . We say that the tangential vectors u, v ∈ S x0 are conjugate at x 0 if

B ( u, v ) = u ⋅ ( Bv ) = v ⋅ ( Bu ) = 0 (62.17)
460 Chap. 11 • HYPERSURFACES

For example, the principal basis vectors c1 ( x 0 ) and c 2 ( x 0 ) are conjugate since the components
of B form a diagonal matrix relative to {c Γ } . Geometrically, conjugate directions may be
explained in the following way: We choose any curve λ ( t ) ∈ S such that

λ ( 0 ) = u ∈ S xo (62.18)

Then the conjugate direction v of u corresponds to the limit of the intersection of S λ ( t ) with S x0
as t tends to zero. We leave the proof of this geometric interpretation for conjugate directions as
an exercise.

A direction represented by a self-conjugate vector v is called an asymptotic direction. In

this case v satisfies the equation

B ( v, v ) = v ⋅ ( Bv ) = 0 (62.19)

Clearly, asymptotic directions exit at a point x o if and only if x o is hyperbolic or parabolic. In

the former case the asympototic directions are the same as those for the hyperbola given by
(62.15), while in the latter case the asymptotic direction is unique and coincides with the
direction of the parallel lines given by (62.15).

If every point of S is hyperbolic, then the asymptotic lines form a coordinate net, and
we can define an asymptotic coordinate system. Relative to such a coordinate system the
components bΓΔ satisfy the condition

b11 = b22 = 0 (62.20)

and the Gaussian curvature is given by

II B = −b122 / a (62.21)

A minimal surface which does not reduce to a plane is necessarily hyperbolic at every point,
since when β1 + β 2 = 0 we must have β1β 2 < 0 unless β1 = β 2 = 0.

From (62.12), S is parabolic at every point if and only if it is developable. In this case
the asymptotic lines are straight lines. In fact, it can be shown that there are only three kinds of
developable surfaces, namely cylinders, cones, and tangent developables. Their asymptotic lines
are simply their generators. Of course, these generators are also lines of curvature, since they are
in the principal direction corresponding to the zero principal curvature.
Sec. 62 • Three-Dimensional Euclidean Manifold 461

Exercises
62.1 Show that

a = n ⋅ ( h1 × h 2 )

62.2 Show that

1 ∂ 2 x ⎛ ∂x ∂x ⎞
bΓΔ = Δ
⋅
Γ ⎜
× 2⎟
a ∂y ∂y ⎝ ∂y ∂y ⎠
1

62.3 Compute the principal curvatures for the surfaces defined in Exercises 55.5 and 55.7.
________________________________________________________________
Chapter 12

ELEMENTS OF CLASSICAL CONTINUOUS GROUPS

In this chapter we consider the structure of various classical continuous groups which are formed
by linear transformations of an inner product space V . Since the space of all linear transformation
of V is itself an inner product space, the structure of the continuous groups contained in it can be
exploited by using ideas similar to those developed in the preceding chapter. In addition, the group
structure also gives rise to a special parallelism on the groups.

Section 63. The General Linear Group and Its Subgroups

In section 17 we pointed out that the vector space L (V ;V ) has the structure of an algebra,
the product operation being the composition of linear transformations. Since V is an inner product
space, the transpose operation

T :L (V ;V ) → L (V ;V ) (63.1)

is defined by (18.1); i.e.,

u ⋅ A T v = Au ⋅ v, u, v ∈V (63.2)

for any A ∈L (V ;V ) . The inner product of any A, B∈L (V ;V ) is then defined by

(cf. Exercise 19.4))

A ⋅ B = tr ( ABT ) (63.3)

From Theorem 22.2, a linear transformation A ∈L (V ;V ) is an isomorphism of V if and

only if

det A ≠ 0 (63.4)

Clearly, if A and B are isomorphisms, then AB and BA are also isomorphisms, and from Exercise
22.1 we have

463
464 Chap. 12 • CLASSICAL CONTINUOUS GROUP

det ( AB ) = det ( BA ) = ( det A )( det B ) (63.5)

As a result, the set of all linear isomorphisms of V forms a group GL (V ) , called the general
linear group ofV . This group was mentioned in Section 17. We claim that GL (V ) is the disjoint
union of two connected open sets in L (V ;V ) . This fact is more of less obvious since GL (V ) is
the reimage of the disjoint union ( −∞,0 ) ∪ ( 0, ∞ ) under the continuous map

det : L (V ;V ) →R

If we denote the primeages of ( −∞, 0 ) and ( 0, ∞ ) under the mapping det by GL (V )

−
and
GL (V ) , respectively, then they are disjoint connected open sets in L (V ;V ) , and
+

GL (V ) =GL (V ) ∪GL (V )
− +
(63.6)

Notice that from (63.5) GL (V ) is, but GL (V )

+ −
is not, closed with respect to the group
operation. So GL (V
) is, but GL (V ) is not, a subgroup ofGL (V ) . We call the elements of
+ −

GL (V ) and GL (V ) proper and improper transformations ofV , respectively. They are

+ −

separated by the hypersurface S defined by

A ∈S ⇔ A ∈L (V ;V ) and det A = 0

Now since GL (V ) is an open set in L (V ;V ) , any coordinate system in L (V ;V )

corresponds to a coordinate system in GL (V ) when its coordinate neighborhood is restricted to a
subset ofGL (V ) . For example, if {e i } and {ei } are reciprocal bases forV , then {ei ⊗ ei } is a basis
for L (V ;V ) which gives rise to the Cartesian coordinate system { X i j } . The restriction of { X i j }
to GL (V ) is then a coordinate system onGL (V ) . A Euclidean geometry can then be defined on
GL (V ) as we have done in general on an arbitrary open set in a Euclidean space. The Euclidean
geometry on GL (V ) is not of much interest, however, since it is independent of the group
structure onGL (V ) .

The group structure on GL (V ) is characterized by the following operations.

1. Left-multiplication by any A ∈GL (V ),
Sec. 64 • The General Linear Group 465

LA : GL (V ) →GL (V )

is defined by

LA ( X ) ≡ AX , X ∈GL (V ) (63.7)

2. Right-multiplication by any A ∈GL (V )

RA :GL (V ) →GL (V )

is defined by

RA ( X ) ≡ XA, X ∈GL (V ) (63.8)

3. Inversion

J :GL (V ) →GL (V )

is defined by

J ( X ) = X −1 , X ∈GL (V ) (63.9)

Clearly these operations are smooth mappings, so they give rise to various gradients which
are fields of linear transformations of the underlying inner product space L (V ;V ) . For example,
the gradient of LA is a constant field given by

∇ LA ( Y ) = AY, Y∈L (V ;V ) (63.10)

since for any Y ∈L (V ;V ) we have

d d
LA ( X + Yt ) = ( AX + A Yt ) = AY
dt t =0 dt t =0

By the same argument ∇ RA is also a constant field and is given by

466 Chap. 12 • CLASSICAL CONTINUOUS GROUP

∇ RA ( Y ) = YA, Y ∈L (V ;V ) (63.11)

On the other hand, ∇J is not a constant field; its value at any point X ∈GL (V ) is given by

⎡⎣ ∇J ( X ) ⎤⎦ ( Y ) = − X −1YX, Y ∈ L (V ;V ) (63.12)

In particular, at X = I the value of ∇ J ( I ) is simply the negation operation,

⎡⎣ ∇J ( I ) ⎤⎦ ( Y ) = − Y, Y ∈L (V ;V ) (63.13)

We leave the proof of (63.13) as an exercise.

The gradients ∇ LA , A ∈GL (V ) , can be regarded as a parallelism on GL (V ) in the

following way: For any two points X and Y in GL (V ) there exists a unique A ∈GL (V ) such
that Y=LA X . The corresponding gradient ∇ LA at the point X is a linear isomorphism

⎡⎣ ∇ LA ( X ) ⎤⎦ :GL (V ) X →GL (V ) Y (63.14)

where GL (V )X denotes the tangent space of GL (V ) at X , a notation consistent with that

introduced in the preceding chapter. Here, of course, GL (V ) X coincides with L (V ;V ) X , which
is a copy of L (V ;V ) . We inserted the argument X in ∇ LA ( X ) to emphasize the fact that
∇ LA ( X ) is a linear map from the tangent space of GL (V ) at X to the tangent space of GL (V ) at
the image point Y. This mapping is given by (63.10) via the isomorphism of GL (V ) X and
GL (V ) Y with L (V ;V ) .

From (63.10) we see that the parallelism defined by (63.14) is not the same as the Euclidean
parallelism. Also, ∇LA is not the same kind of parallelism as the Levi-Civita parallelism on a
hypersurface because it is independent of any path joining X and Y. In fact, if X and Y do not
belong to the same connected set of GL (V ) , then there exists no smooth curve joining them at all,
but the parallelism ∇LA ( X ) is still defined. We call the parallelism ∇LA ( X ) with A ∈GL (V ) the
Cartan parallelism on GL (V ) , and we shall study it in detail in the next section.

The choice of left multiplication rather than the right multiplication is merely a convention.
The gradient RA also defined a parallelism on the group. Further, the two parallelisms ∇LA and
∇RA are related by
Sec. 64 • The General Linear Group 467

∇LA =∇J D ∇RA −1 D ∇J (63.15)

Since LA and RA are related by

LA = J D RA −1 D J (63.16)

for all A ∈GL (V ).

Before closing the section, we mention here that besides the proper general linear group
GL (V ) , several other subgroups of GL (V ) are important in the applications. First, the special
+

linear group SL (V ) is defined by the condition

A ∈SL (V ) ⇔ det A = 1 (63.17)

Clearly SL (V ) is a hypersurface of dimension N 2 − 1 in the inner product space L (V ;V ) .

Unlike GL (V ) , SL (V ) is a connected continuous group. The unimodular group UM (V ) is
defined similiarly by

A ∈UM (V ) ⇔ det A = 1 (63.18)

Thus SL (V ) is the proper subgroup of UM (V ) , namely

UM (V ) = SL (V )
+
(63.19)

Next the orthogonal group O (V ) is defined by the condition

A ∈O (V ) ⇔ A −1 = A T (63.20)

From (18.18) or (63.2) we see that A belongs to O (V ) if and only if it preserves the inner product
of V , i.e.,

Au ⋅ Av = u ⋅ v, u, v ∈V (63.21)

Further, from (63.20) if A ∈O (V ) , then det A has absolute value 1. Consequently, O (V ) is a

subgroup of UM (V ) . As usual, O (V ) has a proper component and an improper component, the
468 Chap. 12 • CLASSICAL CONTINUOUS GROUP

former being a subgroup of O (V ) , denoted by SO (V ) , called the special orthogonal group or the
rotational group of V . As we shall see in Section 65, O (V ) is a hypersurface of dimension
1
N ( N − 1) . From (63.20), O (V ) is contained in the sphere of radius N in L (V ;V ) since if
2
A ∈O (V ) , then

A ⋅ A = tr AA T = tr AA −1 = tr I = N (63.22)

As a result, O (V ) is bounded in L (V ;V ).

SinceUM (V ) , SL (V ) , O (V ) , and SO (V ) are subgroups of GL (V ) , the restrictions of

the group operations of GL (V ) can be identified as the group operations of the subgroups. In
particular, if X, Y∈SL (V ) and Y = LA ( X ) , then the restriction of the mapping ∇ LA ( X ) defined
by (63.14) is a mapping

⎡⎣ ∇LA ( X ) ⎤⎦ : SL (V ) X → SL (V ) Y (63.23)

Thus there exists a Cartan parallelism on the subgroups as well as onGL (V ) . The tangent spaces
of the subgroups are subspaces of L (V ;V ) of course; further , as explained in the preceding
chapter, they vary from point to point. We shall characterize the tangent spaces of SL (V ) and
SO (V ) at the identity element I in Section 66. The tangent space at any other point can then be
obtained by the Cartan parallelism, e.g.,

SL (V ) X = ⎡⎣∇ LX ( I )⎤⎦ (SL (V )I )

for any X∈SL (V ) .

Exercise
63.1 Verify (63.12) and (63.13).
Sec. 64 • The Parallelism of Cartan 469

Section 64. The Parallelism of Cartan

The concept of Cartan parallelism on GL (V ) and on its various subgroups was introduced
in the preceding section. In this section, we develop the concept in more detail. As explained
before, the Cartan parallelism is path-independent. To emphasize this fact, we now replace the
notation ∇LA ( X ) by the notation C ( X ,Y ) , where Y=AX .
Then we put

C ( X ) ≡ C ( I, X ) ≡ ∇LX ( I ) (64.1)

for any X ∈GL (V ). It can be verified easily that

C ( X, Y ) = C ( Y ) D C ( X )
−1
(64.2)

for all X and Y. We use the same notation C ( X, Y ) for the Cartan parallelism from X to Y for
pairs X, Y in GL (V ) or in any continuous subgroup of GL (V ) , such as SL (V ). Thus
C ( X, Y ) denotes either the linear isomorphism

C ( X , Y ) :GL (V ) X →GL (V ) Y (64.3)

or its various restrictions such as

C ( X, Y ) : SL (V ) X → SL (V ) Y (64.4)

when X, Y belong to SL (V ).

Let V be a vector field on GL (V ) , that is

V :GL (V ) → L (V ;V ) (64.5)

Then we say that V is a left-invarient field if its values are parallel vectors relative to the Cartan
parallelism, i.e.,

⎡⎣C ( X , Y ) ⎤⎦ ( V ( X ) ) = V ( Y ) (64.6)
470 Chap. 12 • CLASSICAL CONTINUOUS GROUP

for any X, Y in GL (V ) . Since the Cartan parallelism is not the same as the Euclidean parallelism
induced by the inner product space L (V ;V ) , a left-invarient field is not a constant field. From
(64.2) we have the following representations for a left-invarient field.

Theorem 64.1. A vector field V is left-invariant if and only if it has the representation

V ( X ) = ⎡⎣C ( X ) ⎤⎦ ( V ( I ) ) (64.7)

for all X ∈GL (V ).

As a result, each tangent vector V ( I ) at the identity element I has a unique extension into
a left-variant field. Consequently, the set of all left-invariant fields, denoted by g l (V ) , is a copy
of the tangent space GL (V )I which is canonically isomorphic to L (V ;V ). We call the
restriction map

I : gl (V ) →GL (V ) ≅ L (V ;V )
I
(64.8)

the standard representation of gl (V ) . Since the elements of g l (V ) satisfy the representation

(64.7), they characterize the Caratan parallelism completely by the condition (64.6) for
all V ∈ g l (V ) . In the next section we shall show that g l (V ) has the structure of a Lie algebra,
so we call it the Lie Algebra ofGL (V ) .

Now using the Cartan parallelism C ( X, Y ) , we can define an opereation of covariant

derivative by the limit of difference similar to (56.34), which defines the covariant derivative
relative to the Levi-Civita parallelism. Specifically, let X ( t ) be a smooth curve in GL (V ) and let
U ( t ) be a vector field on X ( t ) . Then we defince the covariant derivative of U ( t ) along
X ( t ) relative to the Cartan parallelism by

DU ( t ) U ( t + Δ t ) −⎡⎣C ( X ( t ) , X ( t +Δ t ) )⎤⎦ ( U ( t ) )
= (64.9)
Dt Δt

Since the Cartan parallelism is defined not just on GL (V ) but also on the various continuous
subgroups of GL (V ) , we may use the same formula (64.9) to define the covariant derivative of
tangent vector fields along a smooth curve in the subgroups also. For this reason we shall now
derive a general representation for the covariant derivative relative to the Cartan parallelism
without restricting the underlying continuous group to be the general linear groupGL (V ) . Then
Sec. 64 • The Parallelism of Cartan 471

we can apply the representation to vector fields on GL (V ) as well as to vector fields on the
various continuous subgroups ofGL (V ) , such as the special linear group SL (V ).

The simplest way to represent the covariant derivative defined by (64.9) is to first express
the vector field U ( t ) in component form relative to a basis of the Lie algebra of the underlying
continuous group. From (64.7) we know that the values {EΓ ( X ) , Γ = 1,..., M } of a left-invariant
field of basis {EΓ ( X ) , Γ = 1,..., M } form a basis of the tangent space at X for all X belonging to
the underlying group. Here M denotes the dimension of the group; it is purposely left arbitrary so
as to achieve generality in the representation. Now since U ( t ) is a tangent vector at X ( t ) , it can be
{
represented as usual by the component form relative to the basis EΓ ( X ( t ) ) , say }
U ( t ) = Uˆ Γ ( t ) EΓ ( X ( t ) ) (64.10)

where Γ is summed from 1 to M . Substituting (64.10) into (64.9) and using the fact that the basis
{EΓ } is formed by parallel fields relative to the Cartan parallelism, we obtain directly

D U ( t ) d Uˆ Γ ( t )
= EΓ ( X ( t ) ) (64.11)
Dt dt

This formula is comparable to the representation of the covariant derivative relative to the
Euclidean parallelism when the vector field is expressed in component form in the terms of a
Cartesian basis.

As we shall see in the next section, the left-variant basis {EΓ } used in the representation
(64.11) is not the natural basis of any coordinate system. Indeed, this point is the major difference
between the Cartan parallelism and the Euclidean parallelism, since relative to the latter a parallel
basis is a constant basis which is the natural basis of a Cartesian coordinate system. If we
introduce a local coordinate system with natural basis {H Γ ,Γ = 1,..., M } , then as usual we can
represent X ( t ) by its coordinate functions ( X Γ ( t ) ) , and EΓ and U ( t ) by their components

EΓ = EΓΔ H Δ and U ( t ) =U Γ ( t ) HΓ ( X ( t ) ) (64.12)

By using the usual transformation rule, we then have

Û Γ ( t ) = U Δ ( t ) FΔΓ ( X ( t ) ) (64.13)

where ⎡⎣ FΔΓ ⎤⎦ is the inverse of ⎡⎣ EΔΓ ⎤⎦ . Substituting (64.13) into (64.11), we get
472 Chap. 12 • CLASSICAL CONTINUOUS GROUP

DU ⎛ dUΔ ∂ FΩΓ d X Σ Δ ⎞
=⎜ +U Ω EΓ ⎟ H Δ (64.14)
Dt ⎝ d t ∂XΣ dt ⎠

We can rewrite this formula as

DU ⎛ dUΔ d XΣ ⎞
=⎜ + U Ω LΔΩΣ ⎟ HΔ (64.15)
Dt ⎝ d t dt ⎠

where

∂ FΩΓ Δ
Γ ∂ EΓ
LΔΩΣ = EΓΔ = − FΩ (64.16)
∂ XΣ ∂XΣ

Now the formula (64.15) is comparable to (56.37) with LΔΩΣ playing the role of the Christoffel
symbols, except that LΔΩΣ is not symmetric with respect to the indices Ω and Σ . For definiteness,
we call LΔΩΣ the Cartan symbols. From (64.16) we can verify easily that they do not depend on the
choice of the basis {EΓ } .

It follows from (64.12)1 and (64.16) that the Cartan symbols obey the same transformation
rule as the Christoffel symbols. Specifically, if ( X Γ ) is another coordinate system in which the
Δ
Cartan symbols are LΩΣ , then

Δ ∂X Δ ∂ X Ψ ∂ X Θ ∂ X Δ ∂2 X Φ
LΩΣ =LΦΨΘ + (64.17)
∂XΦ ∂XΩ ∂XΣ ∂XΦ ∂XΩ∂XΣ

This formula is comparable to (56.15). In view of (64.15) and (64.17) we can define the covariant
derivative of a vector field relative to the Cartan parallelism by

⎛ ∂U Δ ⎞
∇U = ⎜ Σ
+ U Ω LΔΩΣ ⎟ H Δ ⊗ HΣ (64.18)
⎝∂X ⎠

where {HΣ } denotes the dual basis of {H Δ } . The covariant derivative defined in this way clearly
possesses the following property.

Theorem 64.2. A vector field U is left-invariant if and only if its covariant derivative relative to
the Cartan parallelism vanishes.
Sec. 64 • The Parallelism of Cartan 473

The proof of this proposition if more or less obvious, since the condition

∂U Δ
Σ
+ U Ω LΔΩΣ = 0 (64.19)
∂X

is equivalent to the condition

∂ Uˆ Δ
=0 (64.20)
∂XΣ

where Û Δ denotes the components of U relative to the parallel basis {EΓ } . The condition (64.20)
means simply that Û Δ are constant, or, equivalent, U is left-invariant.

Comparing (64.16) with (59.16), we see that the Cartan parallelism also possesses the
following property.

Theorem 64.3. The curvature tensor whose components are defined by

∂ Γ ∂
R Γ ΣΔΩ ≡ L −
Δ ΣΩ
LΓΣΔ +LΦΣΩ LΓΦΔ − LΦΣΔ LΓΦΩ (64.21)
∂X ∂ XΩ

vanishes identically.

Notice that (64.21) is comparable with (59.10) where the Christoffel symbols are replaced
by the Cartan symbols. The vanishing of the curvature tensor

R Γ ΣΔΩ = 0 (64.22)

is simply the condition of integrability of equation (64.19) whose solutions are left-invariant fields.

From the transformation rule (64.17) and the fact that the second derivative therein is
symmetric with respect to the indices Ω and Σ , we obtain

∂XΔ ∂XΨ ∂XΘ

Δ
LΩΣ Δ
− LΣΩ = ( LΦΨΘ − LΦΘΨ ) (64.23)
∂XΦ ∂XΩ ∂XΣ

which shows that the quanties defined by

Δ
TΩΣ ≡ LΔΩΣ − LΔΣΩ (64.24)
474 Chap. 12 • CLASSICAL CONTINUOUS GROUP

are the components of a third-order tensor field. We call the T the torsion tensor of the Cartan
parallelism. To see the geometric meaning of this tensor, we substitute the formula (64.16) into
(64.24) and rewrite the result in the following equivalent form:

Δ Ω ∂ EΨΔ
Σ Ω
Δ
Ω ∂ EΦ
T E E =E
ΩΣ Φ Ψ − EΨ
Φ (64.25)
∂XΩ ∂XΩ

which is the component representation of

T ( EΦ , EΨ ) = ⎡⎣ EΦ , EΨ ⎤⎦ (64.26)

where the right-hand side is the Lie bracket of EΦ and EΨ , i.e.,

⎛ ∂ EΨΔ Ω ∂ EΦ ⎞
Δ
⎡⎣ EΦ , EΨ ⎤⎦ ≡ L EΨ = ⎜ EΦΩ − E Ψ ⎟ HΔ (64.27)
EΦ
⎝ ∂XΩ ∂ XΩ ⎠

Equation (64.27) is comparable to (49.21) and (55.31).

As we shall see in the next section, the Lie bracket of any pair of left-invariant fields is
itself also a left-invariant field. Indeed, this is the very reason that the set of all left-invariant fields
is so endowed with the structure of a Lie algebra. As a result, the components of ⎡⎣ EΦ , EΨ ⎤⎦
relative to the left-invariant basis {EΓ } are constant scalar fields, namely

⎡⎣ EΦ , EΨ ⎤⎦ = CΦΨ
Γ
EΓ (64.28)

Γ
We call CΦΨ the structure constants of the left-invariant basis {EΓ } . From (64.26), CΦΨ
Γ
are nothing
but the components of the torsion tensor T relative to the basis {EΓ } , namely

Γ
T = CΦΨ E Γ ⊗ EΦ ⊗ E Ψ (64.29)

Now the covariant derivative defined by (64.9) for vector fields can be generalized as in the
preceding chapter to covariant derivatives of arbitrary tangential tensor fields. Specifically, the
general formula is

D A ⎧⎪ d AΔ1...1...Δ sr ⎛ AΔ1... Δ s LΣΩ +⋅⋅⋅+ AΔ1... Δ s LΣΩ ⎞ d X Ω ⎫⎪

Γ Γ ΣΓ2... Γ r Γ1 Γ1... Γ r −1Σ Γ r
Δ
=⎨ +⎜ ⎟ ⎬ H Γ1 ⊗⋅⋅⋅⊗ H s (64.30)
Dt ⎪ d t ⎜ − AΣΔΓ1 ... Γ r Σ Γ1 ... Γ r Σ ⎟
LΔ1Ω −⋅⋅⋅− AΔ1... Δ s−1Σ LΔ s Ω ⎠ d t
⎩ ⎝ 2... Δ s ⎭⎪
Sec. 64 • The Parallelism of Cartan 475

which is comparable to (48.7) and (56.23). The formula (64.30) represents the covariant derivative
in terms of a coordinate system ( X Γ ) . If we express A in component form relative to a left-variant
basis {EΓ } , say

A ( t ) = Aˆ Δ1...1...ΔΓrs ( t ) EΓ1 ( X ( t ) ) ⊗⋅⋅⋅⊗ EΔ s ( X ( t ) )

Γ
(64.31)

then the representation of the covariant derivative is simply

D A ( t ) d Aˆ Δ1...1...Δ sr ( t )
Γ Γ

= EΓ1 ( X ( t ) ) ⊗⋅⋅⋅⊗ EΔ s ( X ( t ) ) (64.32)

Dt dt

The formulas (64.30) and (64.32) represent the covariant derivative of a tangential tensor field A
along a smooth curve X ( t ) . Now if A is a tangential tensor field defined on the continuous
group, then we define the covariant derivative of A relative to the Cartan parallelism by a formula
similar to (56.10) except that we replace the Christoffel symbols there by the Cartan symbols.

Naturally we say that a tensor field A is left-invariant if ∇A vanishes. The torsion tensor
field T given by (64.29) is an example of a left-invariant third-order tensor field. Since a tensor
field is left invariant if and only if its components relative to the product basis of a left-invariant
basis are constants, a representation formally generalizing (64.7) can be stated for left-invariant
tensor fields in general. In particular, if {EΓ } is a left-invariant field of bases, then

E = E1 ∧⋅⋅⋅∧ EM (64.33)

is a left-invariant field of density tensors on the group.

476 Chap. 12 • CLASSICAL CONTINUOUS GROUP

Section 65. One-Parameter Groups and the Exponential Map

In section 57 of the preceding chapter we have introduced the concepts of geodesics and
exponential map relative to the Levi-Civita parallelism on a hypersurface. In this section we
consider similar concepts relative to the Cartan parallelism. As before, we define a geodesic to be a
smooth curve X ( t ) such that


DX
=0 (65.1)
Dt

where X denotes the tangent vector X and where the covariant derivative is taken relative to the
Cartan parallelism.

Since (65.1) is formally the same as (57.1), relative to a coordinate system ( X Γ ) we have
the following equations of geodesics

d 2 X Γ dX ∑ dX Δ Γ
+ L∑ Δ = 0, Γ = 1,..., M (65.2)
dt 2 dt dt

which are comparable to (57.2). However, the equations of geodesics here are no longer the Euler-
Lagrange equations of the arc length integral, since the Cartan parallelism is not induced by a
metric and the arc length integral is not defined. To interpret the geometric meaning of a geodesic
relative to the Cartan parallelism, we must refer to the definition (64.9) of the covariant derivative.

 ( t ) of any smooth curve X ( t ) in

We notice first that if we express the tangent vector X
component form relative to a left-invariant basis {EΓ } ,

 (t ) = G Γ (t ) E ( X (t ))
X (65.3)
Γ

then from (64.11) a necessary and sufficient condition for X ( t ) to be a geodesic is that the
components G Γ ( t ) be constant independent of t . Equivalently, this condition means that

 (t ) = G ( X (t ))
X (65.4)

where G is a left-invariant field having the component form

G = G Γ EΓ (65.5)
Sec. 65 • One-Parameter Groups, Exponential Map 477

where G Γ are constant. In other words, a curve X ( t ) is a geodesic relative to the Cartan
parallelism if and only if it is an integral curve of a left-invariant vector field.

This characteristic property of a geodesic implies immediately the following result.

Theorem 65.1. If X ( t ) is a geodesic tangent to the left-invariant field G, then LA ( X ( t ) ) is also a

geodesic tangent to the same left-invariant field G for all A in the underlying continuous group.

A corollary of this theorem is that every geodesic can be extended indefinitely from t = − ∞
to t = + ∞ . Indeed, if X ( t ) is a geodesic defined for an interval, say t ∈[ 0,1] , then we can extend
X ( t ) to the interval t ∈[1, 2] by

X ( t + 1) ≡ LX ( I )X −1 (0 ) ( X ( t ) ) , t ∈ [ 0,1] (65.6)

and so forth. An important consequence of this extension is the following result.

Theorem 65.2. A smooth curve X ( t ) passing through the identity element I at t = 0 is a geodesic
if and only if it forms a one-parameter group, i.e.,

X ( t1 + t2 ) = X ( t1 ) X ( t2 ) , t1 , t2 ∈R (65.7)

The necessity of (65.7) is a direct consequence of (65.6), which may be generalized to

X ( t1 + t2 ) = LX ( t1 ) ( X ( t2 ) ) (65.8)

for all t1 and t2 . Here we have used the condition that

X ( 0) = I (65.9)

Conversely, if (65.7) holds, then by differentiating with respect to t2 and evaluating the result at
t2 = 0 , we obtain

 ( t ) = ⎡C ( X ( t ) ) ⎤ ( X
X  ( 0) ) (65.10)
1 ⎣ 1 ⎦

 ( t ) is the value of a particular left-invariant field at all X ( t ) , and thus X ( t ) is a

which shows that X
geodesic relative to the Cartan parallelism.
478 Chap. 12 • CLASSICAL CONTINUOUS GROUP

Now combining the preceding propositions, we see that the class of all geodesics can be
characterized in the following way. First, for each left-invariant field G there exists a unique one-
parameter group X ( t ) such that

 ( 0) = G ( I )
X (65.11)

where G ( I ) is the standard representation of G . Next, the set of all geodesics tangent to G can be
represented by LA ( X ( t ) ) for all A belonging to the underlying group.

As in Section 57, the one-to-one correspondence between G ( I ) and X ( t ) gives rise to the
notion of the exponential map at the indentity element I . For brevity, let A be the standard
representation of G , i.e.,

A ≡ G (I) (65.12)

Then we define

X (1) = exp I ( A ) ≡ exp A (65.13)

which is comparable to (57.19). As explained in Section 57, (65.13) implies that

X ( t ) = exp ( At ) (65.14)

for all t ∈R . Here we have used the extension property of the geodesic. Equation (65.14) formally
represents the one-parameter group whose initial tangent vector at the identity element I is A .

We claim that the exponential map defined by (65.13) can be represented explicitly by the
exponential series

1 2 1 3 1
exp ( A ) = I + A + A + A + ⋅⋅⋅+ A n + ⋅⋅⋅ (65.15)
2! 3! n!

Clearly, this series converges for each A ∈ L (V ;V ) . Indeed, since we have

An ≤ A
n
(65.16)

for all positive integers n , the partial sums of (65.15) form a Cauchy sequence in the inner product
space L (V ;V ) . That is,
Sec. 65 • One-Parameter Groups, Exponential Map 479

1 n 1 m 1 1
A + ⋅⋅⋅+ A ≤ A + ⋅⋅⋅+
n m
A (65.17)
n! m! n! m!

and the right-hand side of (65.17) converges to zero as n and m approach infinity.

Now to prove that (65.15) is the correct representation for the exponential map, we have to
show that the series

1 22 1
exp ( At ) = I + At + A t + ⋅⋅⋅+ A n t n + ⋅⋅⋅ (65.18)
2! n!

defines a one-parameter group. This fact is more or less obvious since the exponential series
satisfies the usual power law

exp ( At ) exp ( At2 ) = exp ( A ( t1 + t2 ) ) (65.19)

which can be verified by direct multiplication of the power series for exp ( At1 ) and exp ( At2 ) .
Finally, it is easily seen that the initial tangent of the curve exp ( At ) is A since

d ⎛ 1 22 ⎞
⎜ I + At + A t + ⋅⋅⋅ ⎟ = A (65.20)
dt ⎝ 2 ⎠ t =0

This completes the proof of the representation (65.18).

We summarize our results as follows.

Theorem 65.3. Let G be a left-invariant field with standard representation A . Then the geodesics
X ( t ) tangent to G can be expressed by

⎛ 1 ⎞
X ( t ) = X ( 0 ) exp ( At ) = X ( 0 ) ⎜ I + At + A 2t 2 ⋅⋅⋅ ⎟ (65.21)
⎝ 2! ⎠

where the initial point X ( 0 ) is arbitrary.

In the view of this representation, we see that the flow generated by the left-invariant field
G is simply the right multiplication by exp( At ) , namely

ρt = Rexp( At ) (65.22)
480 Chap. 12 • CLASSICAL CONTINUOUS GROUP

for all t . As a result, if K is another left-invariant field, then the Lie bracket of G with K is given
by

K ( X exp ( At ) ) − K ( X ) exp ( At )
[G, K ] ( X ) = lim
t →0
(65.23)
t

This formula implies immediately that [ G, K ] is also a left-invariant field. Indeed, if the standard
representation of K is B , then from the representation (64.7) we have

K ( X ) = XB (65.24)

and therefore

K ( X exp ( At ) ) = X exp ( At ) B (65.25)

Substituting (65.24) and (65.25) into (65.23) and using the power series representation (65.18), we
obtain

[G, K ] ( X ) = X ( AB − BA ) (65.26)

which shows that [G, K ] is left-invariant with the standard representation AB − BA . Hence in
terms of the standard representation the Lie bracket on the Lie algebra is given by

[ A, B] = AB − BA (65.27)

So far we have shown that the set of left-invariant fields is closed with respect to the operation of
the Lie bracket. In general a vector space equipped with a bilinear bracket product which obeys the
Jacobi identities

[ A , B ] = − [ B, A ] (65.28)
and

⎣⎡ A, [ B, C]⎦⎤ + ⎣⎡ B, [C, A ]⎦⎤ + ⎣⎡ C, [ A, B ]⎦⎤ = 0 (65.29)

is called a Lie Algebra. From (65.27) the Lie bracket of left-invariant fields clearly satisfies the
identities (65.28) and (65.29). As a result, the set of all left-invariant fields has the structure of a
Lie algebra with respect to the Lie bracket. This is why that set is called the Lie algebra of the
underlying group, as we have remarked in the preceding section.
Sec. 65 • One-Parameter Groups, Exponential Map 481

Before closing the section, we remark that the Lie algebra of a continuous group depends
only on the identity component of the group. If two groups share the same identity component,
then their Lie algebras are essentially the same. For example, the Lie algebra of GL (V ) and
GL (V ) are both representable by L (V ;V ) with the Lie bracket given by (65.27). The fact that
+

GL (V ) has two components, namely GL (V ) and GL (V ) cannot be reflected in any way by the
+ −

Lie algebra g l (V ) . We shall consider the relation between the Lie algebra and the identity
component of the underlying group in more detail in the next section.

Exercises

65.1. Establish the following properties of the function exp on L (V ;V ):

(a) exp ( A + B ) = ( exp A )( exp B ) if AB = BA .
(b) exp ( − A ) = ( exp A ) .
−1

(c) exp 0 = I
(d) B ( exp A ) B −1 = exp BAB −1 for regular B .
(e) A = A Τ if and only if exp A = ( exp A ) .
Τ

(f) A = − A Τ if and only if exp A is orthogonal.

(g) det ( exp A ) = etr A .
65.2 If P is a projection, show that

exp λ P = I + ( eλ − 1) P

for λ ∈R .
482 Chap. 12 • CLASSICAL CONTINUOUS GROUP

Section 66. Subgroups and Subalgebras

In the preceding section we have shown that the set of left invariant fields is closed with
respect to the Lie bracket. This is an important property of a continuous group. In this section we
shall elaborate further on this property by proving that there exists a one-to-one correspondence
between a connected continuous subgroup of GL (V ) and a subalgebra of g l (V ) . Naturally we
call a subspace h of gl (V ) h is closed with respect to the Lie bracket, i.e.,
a subalgebra if
[G, H ]∈ h whenever both G and H belong to h . We claim that for each subalgebra h of
gl (V ) there corresponds uniquely a connected continuous subgroup H of GL (V ) whose Lie
algebra coincides with the restriction of h on H .

First, suppose that H is a continuous subgroup ofGL (V ) , i.e., H is algebraically a

subgroup of GL (V ) and geometrically a smooth hyper surface inGL (V ) . For example, H may
be the orthogonal group or the special linear group. Then the set of all left-invariant tangent vector
fields on H forms a Lie algebra h . We claim that every element V in h can be extended uniquely
into a left-invariant vector field V onGL (V ) . Indeed, this extension is provided by the
representation (64.7). That is, we simply define

V ( X ) = ⎡⎣C ( X ) ⎤⎦ ( V ( I ) ) (66.1)

for all X ∈GL (V ) . Here we have used the fact that the Cartan parallelism on H is the restriction
of that on GL (V ) to H . Therefore, when X ∈H , the representation (66.1) reduces to

V ( X ) = ⎡⎣C ( X ) ⎤⎦ ( V ( I ) )

since V is left-invariant on H .

From (66.1), V and V share the same standard representation:

A ≡ V (I)

As a result, from (65.27) the Lie bracket on H is related to that on GL (V ) by

[ V, U] = ⎡⎣ V, U ⎤⎦ (66.2)
Sec. 66 • Maximal Abelian Subgroups, Subalgebras 483

for all U and V in h . This condition shows that the extension h of h consisting of all left-
invariant fields V with V in h is a Lie subalgebra of gl (V ) . In view of (66.1) and (66.2), we
simplify the notation by suppressing the overbar. If this convention is adopted, then h becomes a
Lie subalgebra of gl (V ) .

Since the inclusion h ⊂ gl (V ) is based on the extension (66.1), if H 1 and H 2 are

continuous subgroups of GL (V ) having the same identity components, then their Lie algebras
h 1 and h 2 coincide as Lie subalgebras of gl (V ) . In this sense we say that the Lie algebra
characterizes only the identity component of the underlying group, as we have remarked at the end
of the preceding section. In particular, the Lie algebra of UM (V ) coincides with that of SL (V ) .

It turns out that every Lie subalgebra of gl (V ) can be identified as the Lie algebra of a
unique connected continuous subgroup of gl (V ) . To prove this, let h be an arbitrary Lie
subalgebra of gl (V ) . Then the values of the left-invariant field belonging to h form a linear
subspace of L (V ;V ) at each point of GL (V ) . This field of subspaces is a distribution on
GL (V ) as defined in Section 50. According to the Frobenius theorem, the distribution is
integrable if and only if it is closed with respect to the Lie bracket. This condition is clearly
satisfied since h is a Lie subalgebra. As a result, there exists an integral hypersurface of the
distribution at each point inGL (V ) .

We denote the maximal connected integral hypersurface of the distribution at the identity
by H . Here maximality means that H is not a proper subset of any other connected integral
hypersurface of the distribution. This condition implies immediately that H is also the maximal
connected integral hypersurface at any point X which belongs to H . By virtue of this fact we
claim that

LX ( H ) = H (66.3)

for all X ∈H . Indeed, since the distribution is generated by left-invariant fields, its collection of
maximal connected integral hypersurfaces is invariant under any left multiplication. In particular,
LX ( H ) is the maximal connected integral hypersurface at the point X, since H contains the
identity I . As a result, (66.3) holds.

Now from (66.3) we see that X ∈H implies X −1 ∈H since X −1 is the only possible element
such that LX ( X −1 ) = I . Similiarly, if X and Y are contained in H , then XY must also be contained
in H since XY is the only possible element such that LY−1 LX −1 ( XY ) = I . For the last condition we
have used the fact that
484 Chap. 12 • CLASSICAL CONTINUOUS GROUP

LY−1 LX −1 ( H ) = LY−1 ( H ) = H

which follows from (66.3) and the fact that X −1 and Y −1 are both in H . Thus we have shown that
H is a connected continuous subgroup of GL (V ) having h as its Lie algebra.

Summarizing the results obtained so far, we can state the following theorem.

Theorem 66.1. There exists a one-to-one correspondence between the set of Lie subalgebras of
gl (V ) and the set of connected continuous subgroups of GL (V ) in such a way that each Lie
subalgebra h of gl (V ) is the Lie algebra of a unique connected continuous subgroup H of
GL (V ) .

To illustrate this theorem, we now determine explicitly the Lie algebras s l (V ) and
s o (V ) of the subgroups SL (V ) and SO (V ) . We claim first

A∈s l (V ) ⇔ tr A = 0 (66.4)

where tr A denotes the trace of A . To prove this, we consider the one-parameter group exp ( At )
for any A ∈L (V ;V ) . In order that A ∈ sl (V ) , we must have

det ( exp ( At ) ) = 1, t ∈R (66.5)

Differentiating this condition with respect to t and evaluating the result at t = 0 , we obtain [cf.
Exercise 65.1(g)]

⎡det ( exp ( At ) )⎤⎦ = tr A

d
0= (66.6)
dt ⎣ t =0

Conversely, if tr A vanishes, then (66.5) holds because the one-parameter group property of
exp ( At ) implies

⎡det ( exp ( At ) )⎤⎦ = det ( exp ( At ) ) tr A = 0, t ∈R

d
dt ⎣

while the initial condition at t = 0 ,

Sec. 66 • Maximal Abelian Subgroups, Subalgebras 485

det ( exp 0 ) = det I = 1

is obvious. Thus we have completed the proof of (66.4).

From the representation (65.27) the reader will verify easily that the subspace of
L (V ;V ) characterized by the right-hand side of (66.4) is indeed a Lie subalgebra, as it should be.

Next we claim that

A ∈ so (V ) ⇔ AT = − A (66.7)

where A T denotes the transpose of A . Again we consider the one-parameter group exp ( At ) for
any A ∈L (V ;V ) . In order that A ∈ so (V ) , we must have

exp ( − At ) = ⎡⎣exp ( At ) ⎤⎦ = exp ( A T t )

T
(66.8)

Here we have used the identities

⎡⎣exp ( A ) ⎤⎦ = exp ( A T )
−1
⎡⎣ exp ( A ) ⎤⎦ = exp ( − A ) ,
T
(66.9)

which can be verified directly from (65.15). The condition (66.7) clearly follows from the
condition (66.8). From (65.27) the reader also will verify the fact that the subpace of
L (V ;V ) characterized by the right-hand side is a Lie subalgebra.

The conditions (66.4) and (66.7) characterize completely the tangent spaces of SL (V ) and
SO (V ) at the identity element I . These conditions verify the claims on the dimensions of
SL (V ) and SO (V ) made in Section 63.
486 Chap. 12 • CLASSICAL CONTINUOUS GROUP

Section 67. Maximal Abelian Subgroups and Subalgebras

In this section we consider the problem of determining the Abelian subgroups ofGL (V ) .
Since we shall use the Lie algebras to characterize the subgroups, our results are necessarily
restricted to connected continuous Abelian subgroups only. We define first the tentative notion of
a maximal Abelian subset H of GL (V ) . The subset H is required to satisfy the following two
conditions:

(i) Any pair of elements X and Y belonging to H commute.

(ii) H is not a proper subset of any subset of GL (V ) satisfying condition (i).

Theorem 67.1. A maximal Abelian subset is necessarily a subgroup.

The proof is more or less obvious. Clearly, the identity element I is a member of every
maximal Abelian subset. Next, if X belongs to a certain maximal Abelian subset H , then X −1 also
belongs to H . Indeed, X ∈H means that

XY = YX, Y ∈H (67.1)

Multiplying this equation on the left and on the right by X −1 , we get

YX −1 = X −1Y, Y ∈H (67.2)

As a result, X −1 ∈H since H is maximal. By the same argument we can prove also that
XY ∈H whenever X ∈H and Y ∈H . Thus, H is a subgroup ofGL (V ) .

In view of this theorem and the opening remarks we shall now consider the maximal,
connected, continuous, Abelian subgroups ofGL (V ) . Our first result is the following.

Theorem 67.2. The one-parameter groups exp ( At ) and exp ( Bt ) commute if and only if their
initial tangents A and B commute.

Sufficiency is obvious, since when AB = BA the series representations for exp ( At ) and
exp ( Bt ) imply directly that exp ( At ) exp ( Bt ) = exp ( Bt ) exp ( At ) . In fact, we have

exp ( At ) exp ( Bt ) = exp ( ( A + B ) t ) = exp ( Bt ) exp ( At )

Sec. 66 • Maximal Abelian Subgroups, Subalgebras 487

in this case. Conversely, if exp ( At ) and exp ( Bt ) commute, then their initial tangents A and B
must also commute, since we can compare the power series expansions for exp ( At ) exp ( Bt ) and
exp ( Bt ) exp ( At ) for sufficiently small t . Thus the proposition is proved.

We note here a word of caution: While the assertion

AB = BA ⇒ exp ( A ) exp ( B ) = exp ( B ) exp ( A ) (67.3)

is true, its converse is not true in general. This is due to the fact that the exponential map is local
diffeomorphism, but globally it may or may not be one-to-one. Thus there exists a nonzero
solution A for the equation

exp ( A ) = I (67.4)

For example, in the simplest case when V is a two-dimensional spce, we can check directly from
(65.15) that

⎡ 0 − θ ⎤ ⎡cos θ − sin θ ⎤
exp ⎢ ⎥=⎢ cos θ ⎥⎦
(67.5)
⎣ 0 0 ⎦ ⎣sin θ

In particular, a possible solution for (67.4) is the matrix

⎡0 − 2π ⎤
A=⎢ ⎥ (67.6)
⎣ 2π 0 ⎦

For this solution exp ( A ) clearly commutes with exp ( B ) for all B, even though A may or may not
commute with B. Thus the converse of (67.3) does not hold in general.

The main result of this section is the following theorem.

Theorem 67.3. H is a maximal connected, continuous Abelian subgroup of GL (V ) if and only if

it is the subgroup corresponding to a maximal Abelian subalgebra h of gl (V ) .
Naturally, a maximal Abelian subalgebra h of gl (V ) is defined by the following two
conditions:

(i) Any pair of elements A and B belonging to h commute, i.e., [ A, B] = 0 .

(ii) h is not a proper subset of any subalgebra of gl (V ) satisfying condition (i).
488 Chap. 12 • CLASSICAL CONTINUOUS GROUP

To prove the preceding theorem, we need the following lemma.

Lemma. Let H be an arbitrary connected continuous subgroup of GL (V ) and let N be a

neighborhood of I in H . Then H is generated by N , i.e., every element X of H can be expressed
as a product (not unique)

X = Y1Y2 ⋅⋅⋅ Yk (67.7)

where Yi or Yi−1 belongs to N . [The number of factors k in the representation (67.7) is arbitrary.]

Note. Since H is a hypersurface in the inner product space L (V ;V ) , we can define a

neighborhood system on H simply by the intersection of the Euclidean neighborhood system on
L (V ;V ) with H . The topology defined in this way on H is called the induced topology.

To prove the lemma, let H 0 be the subgroup generated by N . Then H 0 is an open set in
H since from (67.7) every point X ∈H 0 has a neighborhood LX ( N ) in H . On the other hand,
H 0 is also a closed set in H because the complement of H 0 in H is the union of LY ( H 0 ) ,
Y∈H which are all open sets in H . As a result H 0 must coincide with H since by hypothesis
H0
H has only one component.

By virtue of the lemma H is Abelian if and only if N is an Abelian set. Combining this
remark with Theorem 67.2, and using the fact that the exponential map is a local diffeomorphism at
the identity element, we can conclude immediately that H is a maximal connected continuous
Abelian subgroup of GL (V ) if and only if h is a maximal Abelian Lie subalgebra of gl (V ) .
This completes the proof.

It should be noticed that on a connected Abelian continuous subgroup of GL (V ) the

Cartan parallelism reduced to a Euclidean parallelism. Indeed, since the Lie bracket vanishes
identically on h , any invariant basis {EΓ } is also the natural basis of a coordinate system
{X } on H .
Γ
Further, the coordinate map defined by

X ≡ exp ( X Γ EΓ ) (67.8)

is a homomorphism of the additive group R M with the Abelian group H . This coordinate system
plays the role of a local Cartesian coordinate system on a neighborhood of the identity element
of H . The mapping defined by (67.8) may or may not be one-to-one. In the former case H is
isomorphic to R M , in the latter case H is isomorphic to a cylinder or a torus of dimension M .
Sec. 66 • Maximal Abelian Subgroups, Subalgebras 489

We say that the Cartan parallelism on the Abelian group H is a Euclidean parallelism
because there exists a local Cartesian coordinate system relative to which the Cartan symbols
vanish identically. This Euclidean parallelsin on H should not be confused with the Eucliedean
parallelism on the underlying inner product space L (V ;V ) in which H is a hypersurface. In
genereal, even if H is Abelian, the tangent spaces at different points of H are still different
subspaces of L (V ;V ) . Thus the Euclidean parallelism on H is not the restriction of the
Euclidean parallelism of L (V ;V ) to H .

An example of a maximal connected Abelian continuous subgroup of GL (V ) is a

dilatation group defined as follows: Let {e i , i = 1,..., N } be the basis ofV . Then a linear
transformation X of V is a dilatation with axes {e i } if each e i is an eigenvector of X and the
corresponding eigenvalue is positive. In other words, the component matrix of X relative to {e i } is
a diagonal matrix with positive diagonal components, say

⎡λ1 ⎤
⎢ ⋅ ⎥
⎢ ⎥
⎡⎣ X ij ⎤⎦ = ⎢ ⋅ ⎥ , λi > 0 i = 1,..., N (67.9)
⎢ ⎥
⎢ ⋅ ⎥
⎢⎣ λN ⎥⎦

where λi may or may not be distinct. The dilatation group with axes {e i } is the group of all
dilatations X . We leave the proof of the fact that a dilatation group is a maximal connected
Abelian continuous subgroup of GL (V ) as an exercise.

Dilatation groups are not the only class of maximal Abelian subgroup ofGL (V ) , of course.
For example, when V is three-dimensional we choose a basis {e1 , e 2 , e 3} forV ; then the subgroup
consisting of all linear transformations having component matrix relative to {e i } of the form

⎡ a b 0⎤
⎢0 a 0⎥
⎢ ⎥
⎢⎣b c a ⎥⎦

with positive a and arbitrary b and c is also a maximal connected Abelian continuous subgroup of
GL (V ) . Again, we leave the proof of this fact as an exercise.
490 Chap. 12 • CLASSICAL CONTINUOUS GROUP

For inner product spaces of lower dimensions a complete classification of the Lie
subalgebars of the Lie algebra gl (V ) is known. In such cases a corresponding classification of
connected continuous subgroups of GL (V ) can be obtained by using the main result of the
preceding section. Then the set of maximal connected Abelian continuous subgroups can be
determined completely. These results are beyond the scope of this chapter, however.
________________________________________________________________

Chapter 13

INTEGRATION OF FIELDS ON EUCLIDEAN MANIFORDS,

HYPERSURFACES, AND CONTINUOUS GROUPS

In this chapter we consider the theory of integration of vector and tensor fields defined on
various geometric entities introduced in the preceding chapters. We assume that the reader is
familiar with the basic notion of the Riemann integral for functions of several real variables.
Since we shall restrict our attention to the integration of continuous fields only, we do not need
the more general notion of the Lebesgue integral.

Section 68. Arc Length, Surface Area, and Volume

Let E be a Euclidean maniforld and let λ be a smooth curve in E . Then the tangent of λ
is a vector in the translation space V of E defined by

λ ( t + Δt ) − λ ( t )
λ ( t ) = lim (68.1)
Δt → 0 Δt

As usual we denote the norm of λ by λ ,

λ ( t ) = ⎡⎣ λ ( t ) ⋅ λ ( t ) ⎤⎦
1/ 2
(68.2)

which is a continuous function of t , the parameter of λ . Now suppose that λ is defined for t
from a to b. Then we define the arc length of λ between λ ( a ) and λ ( b ) by

∫ λ ( t ) dt
b
l= (68.3)
a

We claim that the arc length possesses the following properties which justify the definition
(68.3).

(i) The arc length depends only on the path of λ joining λ ( a ) and λ ( b ) ,
independent of the choice of parameterization on the path.

491
492 Chap. 13 • INTEGRATION OF FIELDS

Indeed, if the path is parameterized by t so that

λ (t ) = λ ( t ) (68.4)

with
t = t (t ) (68.5)

then the tangent vectors λ and λ are related by

dt
λ (t ) = λ ( t ) (68.6)
dt

As a result, we have

∫ λ ( t ) dt = ∫ λ ( t ) dt
b b
(68.7)
a a

which proves property (i).

(ii) When the path joining λ ( a ) and λ ( b ) is a straight line segment, the arc length is
given by

l = λ (b) − λ ( a ) (68.8)

This property can be verified by using the parameterization

λ (t ) = ( λ (b) − λ ( a )) t + λ ( a ) (68.9)

where t ranges from 0 to 1 since λ ( 0 ) = λ ( a ) and λ (1) = λ ( b ) . In view of (68.7), l is given by

∫ λ ( b ) − λ ( a ) dt = λ (b) − λ ( a )
1
l= (68.10)
0

(iii) The arc length integral is additive, i.e., the sum of the arc lengths from
λ ( a ) to λ ( b ) and from λ ( b ) to λ ( c ) is equal to the arc length from λ ( a ) to λ ( c ) .

Now using property (i), we can parameterize the path of λ by the arc length relative to a
certain reference point on the path. As usual, we assume that the path is oriented. Then we
assign a positive parameter s to a point on the positive side and a negative parameter s to a
point on the negative side of the reference point, the absolute value s being the arc length
Sec. 68 • Arc Length, Surface Area, Volume 493

between the point and the reference point. Hence when the parameter t is positively oriented,
the arc length parameter s is related to t by

s = s ( t ) = ∫ λ ( t ) dt
1
(68.11)
0

where λ ( 0 ) is chosen as the reference point. From (68.11) we get

ds / dt = λ ( t ) (68.12)

Substituting this formula into the general transformation rule (68.6), we see that the tangent
vector relative to s is a unit vector pointing in the positive direction of the path, as it should be.

Having defined the concept of arc length, we consider next the concept of surface area.
For simplicity we begin with the area of a two-dimensional smooth surface S in E . As usual,
we can characterize S in terms of a pair of parameters ( u Γ , Γ = 1, 2 ) which form a local
coordinate system on S

x ∈ S ⇔ x = ζ ( u1 , u 2 ) (68.13)

where ζ is a smooth mapping. We denote the tangent vector of the coordinate curves by

h Γ ≡ ∂ζ / ∂u Γ , Γ = 1, 2 (68.14)

Then {h Γ } is a basis of the tangent plane S x of S at any x given by (68.13). We assume that
S is oriented and that ( u Γ ) is a positive coordinate system. Thus {h Γ } is also positive for S x .

Now let U be a domain in S with piecewise smooth boundary. We consider first the
simple case when U can be covered entirely by the coordinate system ( u Γ ) . Then we define the
surface area of U by

σ= ∫∫ e ( u , u ) du
1 2 1
du 2 (68.15)
ζ −1
(U )

where e ( u1 , u 2 ) is defined by

e ≡ a = ( det [ aΓΔ ] ) = ( det [ h Γ ⋅ h Δ ] )

1/ 2 1/ 2
(68.16)
494 Chap. 13 • INTEGRATION OF FIELDS

The double integral in (68.15) is taken over ζ −1 (U ) , which denotes the set of coordinates ( u Γ )
for points belonging to U .

By essentially the same argument as before, we can prove that the surface area has the
following properties which justify the definition (68.15).

(iv) The surface area depends only on the domain U , independent of the choice of
parameterization on U .

To prove this, we note that under a change of surface coordinates the integrand e of
(68.15) obeys the transformation rule [cf. (61.4)]

⎡ ∂u Γ ⎤
e = e det ⎢ Δ ⎥ (68.17)
⎣ ∂u ⎦

As a result, we have

∫∫ e ( u , u ) du du 2 = ∫∫ e ( u , u ) du
1 2 1 1 2 1
du 2 (68.18)
ζ −1 (U ) ζ −1 (U )

which proves property (iv).

(v) When S is a plane and U is a square spanned by the vectors h1 and h 2 at the point
x 0 ∈ S , the surface area of U is

σ = h1 h 2 (68.19)

The proof is essentially the same as before. We use the parameterization

ζ ( u1 , u 2 ) = x 0 + u Γ h Γ (68.20)

From (68.16), e is a constant

e = h1 h 2 (68.21)

and from (68.20), ζ −1 (U ) is the square [ 0,1] × [ 0,1] . Hence by (68.15) we have

1 1
σ =∫ ∫ h1 h 2 du1 du 2 = h1 h 2 (68.22)
0 0
Sec. 68 • Arc Length, Surface Area, Volume 495

(vi) The surface area integral is additive in the same sence as (iii).

Like the arc length parameter s on a path, a local coordinate system ( u Γ ) on S is called
an isochoric coordinate system if the surface area density e ( u 1 , u 2 ) is identical to 1 for all
(u ) .
Γ
We can define an isochoric coordinate system ( u Γ ) in terms of an arbitrary surface
coordinate system ( u Γ ) in the following way. We put

u 1 = u 1 ( u1 , u 2 ) ≡ u1 (68.23)

and
u 2 = u 2 ( u1 , u 2 ) ≡ ∫ e ( u1 , t ) dt
u2
(68.24)
0

where we have assumed that the origin ( 0, 0 ) is a point in the domain of the coordinate system
(u ) .
Γ
From (68.23) and (68.24) we see that

∂u 1 ∂u 1 ∂u 2
= 1, = 0, = e ( u1 , u 2 ) (68.25)
∂u1 ∂u 2 ∂u 2

As a result, the Jacobian of the coordinate transformation is

⎡ ∂u Γ ⎤
det ⎢ Δ ⎥ = e ( u1 , u 2 ) (68.26)
⎣ ∂u ⎦

which implies immediately the desired result:

e ( u 1, u 2 ) = 1 (68.27)

by virtue of (68.17). From (68.26) the coordinate system ( u Γ ) and ( u Γ ) are of the same
orientation. Hence if ( u Γ ) is positively oriented, then ( u Γ ) is a positive isochoric coordinate
system on S .

An isochoric coordinate system x = ζ ( u Γ ) in corresponds to an isochoric mapping ζ

from a domain in R 2 onto the coordinate neighborhood of ζ in S . In general, ζ is not
isometric, so that the surface metric aΓΔ relative to ( u Γ ) need not be a Euclidean metric. In fact,
the surface metric is Euclidean if and only if S is developable. Hence an isometric coordinate
496 Chap. 13 • INTEGRATION OF FIELDS

system (i.e., a rectangular Cartesian coordinate system) generally does not exist on an arbitrary
surface S . But the preceding proof shows that isochoric coordinate systems exist on all S .

So far, we have defined the surface area for any domain U which can be covered by a
single surface coordinate system. Now suppose that U is not homeomorphic to a domain in R 2 .
Then we decompose U into a collection of subdomains, say

U = U1 ∪U2 ∪ ∪UK (68.28)

whee the interiors of U1 ,…,UK are mutually disjoint. We assume that each Ua can be covered by
a surface coordinate system so that the surface area σ (Ua ) is defined. Then we define σ (U )
naturally by

σ (U ) = σ (U1 ) + + σ (UK ) (68.29)

While the decomposition (68.28) is not unique, of course, by the additive property (vi) of the
integral we can verify easily that σ (U ) is independent of the decomposition. Thus the surface
area is well defined.

Having considered the concepts of arc length and surface area in detail, we can now
extend theidea to hypersurfaces in general. Specifically, let S be a hupersurface of dimension
M. Then locally S can be represented by

x ∈ S ⇔ x = ζ ( u1 ,… u M ) (68.30)

where ζ is a smooth mapping. We define

h Γ ≡ ∂ζ / ∂u Γ , Γ = 1,… , M (68.31)
and
aΓΔ ≡ h Γ ⋅ h Δ (68.32)

Then the {h Γ } span the tangent space S x and the aΓΔ define the induced metric on S x . We
define the surface area density e by the same formula (68.16) except that e is now a smooth
function of the M variables ( u1 ,… u M ) , and the matrix [ aΓΔ ] is also M × M .

Now let U be a domain with piecewise smooth boundary in S , and assume that U can
be covered by a single surface coordinate system. Then we define the surface area of U by
Sec. 68 • Arc Length, Surface Area, Volume 497

σ= ∫ ⋅ ⋅⋅∫ e ( u1 ,… , u M ) du1 du M (68.33)

ζ −1 (U )

By the same argument as before, σ has the following two properties.

(vii) The surface area is additive and independent of the choice of surface coordinate
system.
(viii) The surface area of an M-dimensional cube with sides h1 ,…, h M is

σ = h1 hM (68.34)

More generally, when U cannot be covered by a single coordinate system, we decompose U by

(68.28) and define σ (U ) by (68.29).

We can extend the notion of an isochoric coordinate system to a hypersurface in general.

To construct an isochoric coordinate system ( u Γ ) , we begin with an arbitrary coordinate system
(u ).
Γ
Then we put
u Γ = u Γ ( u1 ,…, u M ) = u Γ , Γ = 1,… , M − 1 (68.35)

u M = u M ( u1 ,…, u M ) ≡ ∫ e ( u1 ,…, u M −1 , t ) dt
uM
(68.36)
0

From (68.35) and (68.36) we get

⎡ ∂u Γ ⎤
det ⎢ Δ ⎥ = e ( u1 ,… u M ) (68.37)
⎣ ∂u ⎦

As a result, the coordinate system ( u Γ ) is isochoric since

e ( u 1 ,… , u M ) = e ( u1 ,…, u M )
1
=1 (68.38)
det ⎡⎣∂u Γ / ∂u Δ ⎤⎦

Finally, when M = N , S is nothing but a domain in E . In this case ( u1 ,… , u N )

becomes an arbitrary local coordinate system in E , and e ( u1 ,… , u N ) is just the Euclidean
volume relative to ( u i ) . The integral
498 Chap. 13 • INTEGRATION OF FIELDS

υ≡ ∫ ⋅⋅⋅∫ e ( u1 ,… , u M ) du1 du M (68.39)

ζ −1
(U )

now defines the Euclidean volume of the domain U .

Sec. 69 • Vector and Tensor Fields 499

Section 69. Integration of Vector Fields and Tensor Fields

In the preceding section we have defined the concept of surface area for an arbitrary M-
dimensional hyper surface imbedded in an N-dimensional Euclidean manifold E . When M = 1 ,
the surface reduces to a path and the surface area becomes the arc length, while in the case M=N
the surface corresponds to a domain in E , and the surface area becomes the volume. In this
section we shall define the integrals of various fields relative to the surface area of an arbitrary
hyper surface S . We begin with the integral of a continuous function f defined on S .

As before, we assume that S is oriented and U is a domain in S with piecewise

smoothboundary. We consider first the simple case when U can be covered by a single surface
coordinate system x = ζ ( u1 ,… u M ) . We choose ( u Γ ) to be positively oriented, of course. Under
these assumptions, we define the integral of f on U by

∫U
f dσ ≡ ∫
ζ −1 (U ) ∫ fe du1 du M (69.1)

where the function f on the right-hand side denotes the representation of f in terms of the
surface coordinates ( u Γ ) :

f ( x ) = f ( u1 ,…, u M ) (69.2)

where

x = ζ ( u1 ,… , u M ) (69.3)

It is understood that the multiple integral in (69.1) is taken over the positive orientation on
ζ −1 (U ) in R M .

By the same argument as in the preceding section, we see that the integral possesses the
following properties.

(i) When f is identical to 1 the integral of f is just the surface area of U , namely

σ (U ) = ∫ dσ (69.4)
U

(ii) The integral of f is independent on the choice of the coordinate system ( u Γ ) and is
additive with respect to its domain.
(iii) The integral is a linear function of the integrand in the sense that
500 Chap. 13 • INTEGRATION OF FIELDS

∫ (α
U
f + α 2 f 2 ) dσ = α1 ∫ f1 dσ + α 2 ∫ f 2 dσ
1 1
U U
(69.5)

for all constants α1 and α 2 . Also, the integral is bounded by

σ (U ) min f ≤ ∫ f dσ ≤ σ (U ) max f (69.6)

U U U

where the extrema of f are taken over the domain U .

Property (iii) is a standard result of multiple integrals in calculus, so by virtue of the

definition (69.1) the same is valid for the integral of f .

As before, if U cannot be covered by a single coordinate system, then we decompose U

by (68.28) and define the integral of f over U by

∫U
f dσ = ∫ f dσ +
U1
+∫
UK
f dσ (69.7)

By property (ii) we can verify easily that the integral is independent of the decomposition.

Having defined the integral of a scalar field, we define next the integral of a vector field.
Let v be a continuous vector field on S , i.e.,

v : S →V (69.8)

where V is the translation space of the underlying Euclidean manifold E . Generally the values
of v may or may not be tangent to S . We choose an arbitrary Cartesian coordinate system with
natural basis {e i }. Then v can be represented by

v ( x ) = υ i ( x ) ei (69.9)

where υ 1 ( x ) , i = 1,… , N , are continuous scalar fields on S . We define the integral of v by

∫ U
v dσ ≡ (∫
U
υ i dσ ei) (69.10)

Clearly the integral is independent of the choice of the basis {e i }.

More generally if A is a tensor field on S having the representation

Sec. 69 • Vector and Tensor Fields 501

A ( x ) = Ai1…ir j1… js ( x ) ei1 ⊗ ⊗ eir ⊗ e j1 ⊗ ⊗ e js (69.11)

where {ei } denotes the reciprocal basis of {e i } , then we define

∫U
A dσ ≡ (∫ U
)
Ai1…ir j1… js dσ ei1 ⊗ ⊗ e js (69.12)

Again the integral is independent of the choice of the basis {e i }.

The integrals defined by (69.10) and (69.12) possess the same tensorial order as the
integrand. The fact that a Cartesian coordinate system is used in (69.10) and (69.12) reflects
clearly the crucial dependence of the integral on the Euclidean parallelism of E . Without the
Euclidean parallelism it is generally impossible to add vectors or tensors at different points of the
domain. Then an integral is also meaningless. For example, if we suppress the Euclidean
parallelism on the underlying Euclidean manifold E , then the tangential vectors or tensors at
different points of a hyper surface S generally do not belong to the same tangent space or tensor
space. As a result, it is generally impossible to “sum” the values of a tangential field to obtain an
integral without the use of some kind of path-independent parallelism. The Euclidean
parallelism is just one example of such parallelisms. Another example is the Cartan parallelism
on a continuous group defined in the preceding chapter. We shall consider integrals relative to
the Cartan parallelism in Section 72.

In view of (69.10) and (69.12) we see that the integral of a vector field or a tensor field
possesses the following properties.

(iv) The integral is linear with respect to the integrand.

(v) The integral is bounded by

∫ U
v dσ ≤ ∫ v dσ
U
(69.13)

and similarly

∫
U
A dσ ≤ ∫ A dσ
U
(69.14)

where the norm of a vector or a tensor is defined as usual by the inner product of V . Then it
follows from (69.6) that

∫ v d σ ≤ σ (U ) max v (69.15)
U U

and
502 Chap. 13 • INTEGRATION OF FIELDS

∫ A d σ ≤ σ (U ) max A (69.16)
U U

However, it does not follow from (69.6), and in fact it is not true, that σ (U ) minU v is a lower
bound for thenorm of the integral of v.
Sec. 70 • Integration of Differential Forms 503

Section 70. Integration of Differential Forms

Integration with respect to the surface area density of a hyper surface is a special case of
a more general integration of differential forms. As remarked in Section 68, the transformation
rule (68.17) of the surface area density is the basic condition which implies the important
property that the surface area integral is independent of the choice of the surface coordinate
system. Since the transformation rule (68.17) is essentially the same as that of the strict
components of certain differential forms, we can extend the operation of integration to those
forms also. This extension is the main result of this section.

We begin with the simple notion of a differential N-form Z on E . By definition, Z is a

completely skew-symmetric covariant tensor field of order N. Thus relative to any coordinate
system ( u i ) , Z has the representation

Z = Z i1…iN hi1 ⊗ ⊗ hiN = Z1…N h1 ∧ ∧ hN

(70.1)
= z h1 ∧ ∧ hN

where z is called the relative scalar or the density of Z . As we have shown in Section 39, the
transformation rule for z is

⎡ ∂u i ⎤
z = z det ⎢ j ⎥ (70.2)
⎣ ∂u ⎦

This formula is comparable to (68.17). In fact if we require that ( u i ) and ( u i ) both be positively
oriented, then (70.2) can be regarded as a special case of (68.17) with M = N . As a result, we
can define the integral of Z over a domain U in E by

∫ Z ≡∫ ∫ z ( u ,…, u ) du
1 N 1
du N (70.3)
U ζ −1 (U )

and the integral is independent of the choice of the (positive) coordinate system x = ζ ( u i ) .

Notice that in this definition the Euclidean metric and the Euclidean volume density e
are not used at all. In fact, (68.39) can be regarded as a special case of (70.3) when Z reduces to
the Euclidean volume tensor

E = e h1 ∧ ∧ hN (70.4)
504 Chap. 13 • INTEGRATION OF FIELDS

Here we have assumed that the coordinate system is positively oriented; otherwise, a negative
sign should be inserted on the right hand side since the volume density e as defined by (68.16) is
always positive. Hence, unlike the volume integral, the integral of a differential N-form Z is
defined only if the underlying space E is oriented. Other than this aspect, the integral of Z and
the volume integral have essentially the same properties since they both are defined by an
invariant N-tuple integral over the coordinates.

Now more generally let S be an oriented hypersurface in E of dimension M , and

suppose Z is a tangential differential M-form on S . As before, we choose a positive surface
coordinate system ( u Γ ) on S and represent Z by

Z = z h1 ∧ ∧ hM (70.5)

where z is a function of ( u1 ,… , u M ) where {h Γ } is the natural basis reciprocal to {h Γ } . Then

the transformation rule for z is

⎡ ∂u Γ ⎤
z = z det ⎢ Δ ⎥ (70.6)
⎣ ∂u ⎦

As a result, we can define the integral of Z over a domain U in S by

∫ Z ≡∫ z ( u ,… , u ) du
)∫
1 M 1
du M (70.7)
U ζ −1
(U

and the integral is independentof the choice of the positive surface coordinate system ( u Γ ) . By
the same remark as before, we can regard (70.7) as a generalization of (68.33).

The definition (70.7) is valid for any tangential M-form Z on S . In this definition the
surface metric and thesurface area density are not used. The fact that Z is a tangential field on
S is not essential in the definition. Indeed, if Z is an arbitrary skew-symmetric spatial
covariant tensorof order M on S , then we define the density of Z on S relative to ( u Γ ) simply
by
z = Z ( h1 ,… , h M ) (70.8)

Using this density, we define the integral of Z again by (70.7). Of course, the formula (70.8) is
valid for a tangential M-form Z also, since it merely represents the strict component of the
tangential projection of Z .

This remark can be further generalized in the following situation: Suppose that S is a
hyper surface contained in another hypersurface S 0 in E , and let Z be a tangential M-form on
Sec. 70 • Integration of Differential Forms 505

S 0 . Then Z gives rise to a density on S by the same formula (70.8), and the integral of Z over
any domain U in S is defined by (70.7). Algebraically, this remark is a consequence of the
simple fact that a skew-symmetric tensor over the tangent space of S 0 gives rise to a unique
skew-symmetric tensor over the tangent space of S , since the latter tangent space is a subspace
of the former one.

It should be noted, however, that the integral of Z is defined over S only if the order of
Z coincides with the dimension of S . Further, the value of theintegral is always a scalar, not a
vector or a tensor as in the preceding section. We can regard the intetgral of a vector field or a
tensor field as a special case of the integral of a differential form only when the fields are
represented interms of their Cartesian components as shown in (69.9) and (69.11).

An important special case of the integral of a differential form is the line integral in
classical vector analysis. In this case S reduces to an oriented path λ , and Z is a 1-form w.
When the Euclidean metric on E is used, w corresponds simply to a (spatial or tangential) vector
field on λ. Now using any positive parameter t on λ , we obtain from (70.7)

w = ∫ w ( t ) ⋅ λ ( t ) dt
b
∫λ a
(70.9)

Here we have used the fact that for an inner product space the isomorphism of a vector and a
covector is given by

w, λ = w ⋅ λ (70.10)

The tangent vector λ playes the role of the natural basis vector h1 associated with the parameter
t, and (70.10) is just the special case of (70.8) when M = 1.

The reader should verify directly that the right-hand side of (70.9) is independent of the
choice of the (positive) parameterization t on λ. By virtue of this remark, (70.9) is also written
as

∫ λ
w = ∫ w ⋅d λ
λ
(70.11)

in the classical theory.

Similarly when N = 3 and M = 2 , a 2-form Z is also representable by a vector field w,

namely

Z ( h1 , h 2 ) = w ⋅ ( h1 × h 2 ) (70.12)
506 Chap. 13 • INTEGRATION OF FIELDS

Then the integral of Z over a domain U in a two-dimensional oriented surface S is given by

∫ Z= ∫∫ w ⋅ ( h1 × h 2 ) du1du 2 ≡ ∫ w ⋅ d σ (70.13)
U U
ζ −1
(U )

where dσ is the positive area element of S defined by

dσ ≡ ( h1 × h 2 ) du1du 2 (70.14)

The reader will verify easily that the right-hand side of (70.14) can be rewritten as

dσ = en du1 du 2 (70.15)

where e is the surface area density on S defined by (68.16), and where n is the positive unit
normal of S defined by

h1 × h 2 1
n= = h1 × h 2 (70.16)
h1 × h 2 e

Substituting (70.15) into (70.13), we see that the integral of Z can be represented by

∫ Z= ∫∫ ( w ⋅ n ) e du du
1 2
(70.17)
U
ζ −1
(U )

which shows clearly that the integral is independent of the choice of the (positive) surface
coordinate system ( u Γ ) .

Since the multipleof an M-form Z by a scalar field f remains an M-form, we can define
the integral of f with respect to Z simply as the integral of f Z . Using a Cartesian coordinate
representation, we can extend this operation to integrals of a vector field or a tensor field relative
to a differential form. The integrals defined in the preceding section are special cases of this
general operation when the differential forms are the Euclidean surface area densities induced
bythe Euclidean metric on the underlying space E .
Sec. 71 • Generalized Stokes’ Theorem 507

Section 71. Generalized Stokes’ Theorem

In Section 51 we have defined the operation of exterior derivative on differential forms
on E . Since this operation does not depend on the Euclidean metric and the Euclidean
parallelism, it can be defined also for tangential differential forms on a hyper surface, as we have
remarked in Section 55. Specifically, if Z is a K-form on an M-dimensional hypersurface S , we
choose a surface coordinate system ( u Γ ) and represent Z by

Z= ∑
Γ1 < <Γ K
Z Γ1 ΓK h Γ1 ∧ ∧ h ΓK (71.1)

then the exterior derivative d Z of Z is a ( K + 1) − form given by

dZ = ∑
Γ1 < <Γ K
dZ Γ1 ΓK ∧ h Γ1 ∧ ∧ h ΓK (71.2)

where dZ Γ1 ΓK is defined by

∂Z Γ1 ΓK
dZ Γ1 ΓK ≡ Δ
hΔ (71.3)
∂u

In this section we shall establish a general result which connects the integral of d Z over a
( K + 1) − dimensional domain U in S with the integral of Z over the K-dimensional boundary
surface ∂U of U . We begin with a preliminary lemma about a basic property of the exterior
derivative.

Lemma. Let S 0 be a P-dimensional hypersurface in E with local coordinate system x = η ( y α )

and suppose that S is an M-dimensional hypersurface contained in S 0 and characterized by the
representation

η −1 ( y α ) = x ∈ S ⇔ y α = y α ( u Γ ) , α = 1,…, P (71.4)

Let W be a K-form on S 0 with the component form

W=
α
∑α Wα1 αK gα1 ∧ ∧ gα K (71.5)
1< < K
508 Chap. 13 • INTEGRATION OF FIELDS

where {gα , α = 1,… , P} denotes the natural basis of ( y α ) on S 0 , and suppose that Z is the
tangential projection of W on S , i.e.,

∂yα1 ∂yα K Γ1
Z=
α
∑α Wα1 αK
∂u Γ1 ∂u ΓK
h ∧ ∧ hΓK (71.6)
1< < K

where {h Γ
, Γ = 1,… , M } denotes thenatural basis of ( u Γ ) on S . Then the exterior derivatives
of Z coincides with the tangential projection of the exterior derivative of W . In other words,
theoperation of exterior derivative commutes with the operation of tangentialprojection.

We can prove this lemma by direct calculation of d W and d Z . From (71.5) and (71.2)
d W is given by

∂Wα1
dW =
α
∑α ∂y β
αK
g β ∧ gα1 ∧ ∧ gα K (71.7)
1< < K

Hence its tangential projection on S is

∂Wα1 ∂y β ∂y α1 ∂y α K Δ
α
∑α ∂y β
αK

∂u Δ ∂u Γ1 ∂u Γ K
h ∧ h Γ1 ∧ ∧ h ΓK (71.8)
1< < K

Similarly, from (71.6) and (71.7), d Z is given by

∂ ⎛ ∂y α1 ∂y α K ⎞ Δ
dZ = ∑
α1 < <α K
⎜ Wα <
∂u Δ ⎝ 1
<α K
∂u Γ1 ∂u ⎠ ΓK ⎟
h ∧ h Γ1 ∧ ∧ h Γ K
(71.9)
∂Wα1…α K ∂y β ∂y α1 ∂y α K Δ
= ∑
α1 < <α K ∂y β ∂u Δ ∂u Γ1 ∂u ΓK
h ∧ h Γ1 ∧ ∧ h Γ K

where we have used the skew symmetry of the exterior product and the symmetry of the second
derivative ∂ 2 y α / ∂u Γ∂u Δ with respect to Γ and Δ. Comparing (71.9) with (71.8), we have
completed the proof of the lemma.

Now we are ready to present the main result of this section.

Generalized Stokes’ Theorem. Let S and S 0 be hypersurfaces as defined in the preceding

lemma and suppose that U is an oriented domain in S with piecewise smooth boundary ∂U .
(We orient the boundary ∂U as usual by requiring the outward normal of ∂U be the positive
normal.) Then for any tangential ( M − 1) -form Z on S 0 we have
Sec. 71 • Generalized Stokes’ Theorem 509

∫∂U
Z = ∫ dZ
U
(71.10)

Before proving this theorem, we remark first that other than the condition

S ⊂ S0 (71.11)

The hypersurface S 0 is entirely arbitrary. In application we often take S 0 = E but this choice is
not necessary. Second, by virtue of the preceding lemma it suffices to prove (71.10) for
tangential ( M − 1) − forms Z on S only. In other words, we can choose S 0 to be the same as
S without loss of generality. Indeed, as explained in the preceding section, the integrals in
(71.10) are equal to those of tangential projection of the forms Z and d Z on S . Then by virtue
of the lemma the formula (71.10) amounts to nothing but a formula for tangential forms on S .

Before proving the formula (71.10) in general, we consider first the simplest special case
when Z is a 1-form and S is a two-dimensional surface. This case corresponds to the Stokes
formula in classical vector analysis. As usual, we denote a 1-form by w since it is merely a
covariant vector field on S . Let the component form of w in ( u Γ ) be

w = wΓh Γ (71.12)

where Γ is summed from 1 to 2. From (71.2) the exterior derivative of w is a 2-form

∂wΓ Δ ⎛ ∂w ∂w ⎞
dw = Δ
h ∧ h Γ = ⎜ 12 − 21 ⎟ h1 ∧ h 2 (71.13)
∂u ⎝ ∂u ∂u ⎠

Thus (71.10) reduces to

⎛ ∂w2 ∂w1 ⎞ 1 2
∫ζ −1
( ∂U )
wΓλ Γ dt = ∫∫ ⎜ 1 − 2 ⎟du du
⎝ ∂u ∂u ⎠
(71.14)
ζ −1
(U )

where ( λ 1 ( t ) , λ 2 ( t ) ) denotes the coordinates of the oriented boundary curve ∂U in the

coordinate system ( u1 , u 2 ) , the parameterization t on ∂U being positively oriented but otherwise
entirely arbitrary.

Now since the integrals in (71.14) are independent of the choice of positive coordinate
system, for simplicity we consider first thecase when ζ −1 (U ) is the square [ 0,1] × [0,1] in R 2 .
Naturally, we use the parameters u1 , u 2 , 1 − u1 , and 1 − u 2 on the boundary segments
510 Chap. 13 • INTEGRATION OF FIELDS

( 0, 0) → (1,0) , (1,0) → (1,1) , (1,1) → ( 0,1) , and ( 0,1) → ( 0,0 ) , respectively. Relative to this
parameterization on ∂U the left-hand side of (71.14) reduces to

w1 ( u1 ,0 ) du1 + ∫ w2 (1, u 2 ) du 2
1 1
∫0 0
(71.15)
− ∫ w1 ( u ,1) du − ∫ w2 ( 0, u ) du
1 1
1 1 2 2
0 0

Similarly, relative to the surface coordinate system ( u1 , u 2 ) the right-hand side of (71.14)
reduces to

1 1 ⎛ ∂w2 ∂w1 ⎞ 1 2
∫∫
0 0
⎜ 1 − 2 ⎟ du du
⎝ ∂u ∂u ⎠
(71.16)

which may be integrated by parts once with respect to one of the two variables ( u1 , u 2 ) and the
result is precisely the sameas (71.15). Thus (71.14) is proved in this simple case.

In general U may not be homeomorphic to the square [ 0,1] × [0,1] , of course. Then we
decompose U as before by (68.28) and we assume that each Ua , a = 1,… , K , can be represented
by the range

Ua = ζ a ([ 0,1] × [0,1]) (71.17)

By the result for the simple case we then have

∫ ∂Ua
w= ∫ ∂Ua
dw , a = 1,… , K (71.18)

Now adding (71.18) with respect to a and observing the fact that all common boundaries of
pairs of U1 ,…,UK are oriented oppositely as shown in Figure 10, we obtain

∫ ∂U
w= ∫ U
dw (71.19)

which is the special case of (71.10) when M = 2.

Sec. 71 • Generalized Stokes’ Theorem 511

U3 U5 U7

U2
U8
U6
U1 U4

Figure 10

The proof of (71.10) for the general case with an arbitrary M is essentially the same as the
proof of the preceding special case. To illustrate the similarity of the proofs we consider next the
case M = 3. In this case Z is a 2-form on a 3-dimensional hypersurface S . Let
( u Γ , Γ = 1, 2, 3) be a positive surface coordinate system on S as usual. Then Z can be
represented by

Z = ∑ Z ΓΔ h Γ ∧ h Δ
Γ<Δ (71.20)
= Z12 h ∧ h + Z13h ∧ h + Z 23h ∧ h
1 2 1 3 2 3

From (71.2) the exterior derivative of Z is given by

⎛ ∂Z ∂Z ∂Z ⎞
d Z = ⎜ 123 − 132 + 231 ⎟ h1 ∧ h 2 ∧ h3 (71.21)
⎝ ∂u ∂u ∂u ⎠

As before, we now assume that U can be represented by the range of a cube relative to a certain
( u Γ ) , namely
U = ζ ([0,1] × [ 0,1] × [0,1]) (71.22)

Then the right-hand side of (71.10) reduces to the triple integral

512 Chap. 13 • INTEGRATION OF FIELDS

1 1 1 ⎛ ∂Z12 ∂Z13 ∂Z 23 ⎞ 1 2 3
∫ ∫ ∫ ⎜⎝ ∂u
0 0 0 3
−
∂u 2
+ 1 ⎟ du du du
∂u ⎠
(71.23)

which may be integrated by parts once with respect to one of the three variables ( u1 , u 2 , u 3 ) . The
result consists of six terms:

Z12 ( u1 , u 2 ,1) du1 du 2 − ∫ Z12 ( u1 , u 2 , 0 ) du1 du 2

1 1 1 1
∫ ∫
0 0 0 ∫
0

Z13 ( u1 ,1, u 3 ) du1 du 3 + ∫ Z13 ( u1 , 0, u 3 ) du1 du 3

1 1 1 1
−∫ ∫ ∫ (71.24)
0 0 0 0

Z 23 (1, u 2 , u 3 ) du 2 du 3 − ∫ Z 23 ( 0, u 2 , u 3 ) du 2 du 3
1 1 1 1
+∫ ∫ ∫
0 0 0 0

which are precisely the representations of the left-hand side of (71.10) on the six faces of the
cube with an appropriate orientation on each face. Thus (71.10) is proved when (71.22) holds.

In general, if U cannot be represented by (71.22), then we decompose U as before by

(68.28), and we assume that each Ua may be represented by

Ua = ζ a ([ 0,1] × [0,1] × [0,1] ) (71.25)

for an appropriate ζ a . By the preceding result we then have

∫
∂Ua
Z= ∫
Ua
dZ, a = 1,… , K (71.26)

Thus (71.10) follows by summing (71.26) with respect to a.

Following exactly the same pattern, the formula (71.10) can be proved by an arbitrary
M = 4,5,6,…. Thus the theorem is proved.

The formula (71.10) reduces to two important special cases in classical vector analysis
when the underlying Euclidean manifold E is three-dimensional. First, when M = 2 and U is
two-dimensional, the formula takes the form

∫
∂U
w= ∫ U
dw (71.27)

which can be rewritten as

∫∂U
w ⋅ dλ = ∫
U
curl w ⋅ dσ (71.28)
Sec. 71 • Generalized Stokes’ Theorem 513

where dσ is given by (70.14) and where curl w is the axial vector corresponding to dw , i.e.,

dw ( u, v ) = curl w ⋅ ( u × v ) (71.29)

for any u, v in V . Second, when M = 3 and U is three-dimensional the formula takes the
form

∫∫ w ⋅ dσ = ∫∫∫ div wdυ

∂U U
(71.30)

where dυ is the Euclidean volume element defined by [see (68.39)]

dυ ≡ e du1 du 2 du 3 (71.31)

relative to any positive coordinate system ( u i ) , or simply

dυ = dx1 dx 2 dx 3 (71.32)

relative to a right-handed rectangular Cartesian coordinate system ( x i ) . In (71.30), w is the

axial vector field corresponding to the 2-form Z , i.e., relative to ( x i )

Z = w3e1 ∧ e 2 − w2e1 ∧ e 3 + w1e 2 ∧ e 3 (71.33)

which is equivalent to

Z ( u, v ) = w ⋅ ( u × v ) , u, v ∈V (71.34)

From (71.33), d Z is given by

⎛ ∂w ∂w ∂w ⎞
d Z = ⎜ 11 + 22 + 33 ⎟ e1 ∧ e 2 ∧ e 3 = ( div w ) E (71.35)
⎝ ∂x ∂x ∂x ⎠

As a result, we have

∫∂U
Z= ∫∂U
w ⋅ dσ (71.36)

and
514 Chap. 13 • INTEGRATION OF FIELDS

∫U
d Z = ∫∫∫ div wdυ
U
(71.37)

The formulas (71.28) and (71.30) are called Stokes’ theorem and Gauss’ divergence
theorem, respectively, in classical vector analysis.
Sec. 72 • Invariant Integrals on Continuous Groups 515

Section 72. Invariant Integrals on Continuous Groups

In the preceding chapter we have considered various groups which are also hypersurfaces
contained in the Euclidean space L (V ;V ) . Since the integrations defined so far in this chapter
can be applied to any hypersurface in a Euclidean manifold, in particular they can be applied to
the continuous groups. However, these integrations are generally unrelated to the group
structure and thus their applications are limited. This situation is similar to that about parallelism.
While the induced metric and its Levi-Civita parallelism certainly exist on the underlying
hypersurface of the group, they are not significant mathematically because they do not reflect the
group structure. This remark has led us to consider the Cartan parallelism which is defined by the
left-invariant fields on the group.
Now as far as integrationis concerned, the natural choice for a continuous group is the
integration based on a left-invarianct volume density. Specifically, if {e Γ } is a basis for the Lie
algebra of the group, then a volume tensor field Z is a left-invariant if and only if it can be
represented by

Z = ce1 ∧⋅⋅⋅∧ e M (72.1)

where c is a constant. The integral of Z over any domain U obeys the condition

∫U
Z =∫
LX (U )
Z (72.2)

for all elements X belonging to the group.

A left-invariant volume tensor field Z is also right-invariant if and only if it is invariant

under the inversion operation J when the dimension M of the group is even or it is mapped into
− Z by J J when M is odd. This fact is more or less obvious since in general J maps any left-
invariant field into a right-invariant field. Also, the gradient of J at the identity element
coincides with the negation operation. Consequently, when M is even, Z ( I ) is invaritant
under J , while if M is odd, Z ( I ) is mapped into − Z ( I ) by J . Naturally we call Z an invariant
volume tensor field if it is both left-invariant and right-invariant. Relative to an invariant Z the
integral obeys the condition (72.2) as well as the conditions

∫U
Z= ∫
RX (U )
Z, ∫ U
Z= ∫(
J U)
Z (72.3)

Here we have used the fact that J preserves the orientatin of the group when M is even, while
J reverses the orientation of the group when M is odd.
516 Chap. 13 • INTEGRATION OF FIELDS

By virtue of the representation (72.1) all left-invariant volume tensor fields differ from
one another by a constant multiple only and thus they are either all right-invariant or all not
right-invariant. As we shall see, the left-invariant volume tensor fields on GL (V ) , SL (V ) ,
O (V ) , and all continuous subgroups of O (V ) are right-invariant. Hence invariant integrals
exist on these groups. We consider first the general linear group GL (V ) .

To prove that the left-invariant volume tensor fields on the GL (V ) are also right-
invariant, we recall first from exterior algebra the transformation rule for a volume tensor under a
linear map of the underlying vector space. Let W be an arbitrary vector space of dimension M ,
and suppose that A is a linear transformation of W

A : W →W (72.4)

Then A maps any volume tensor E on W to ( det A ) E ,

A ∗ (E ) = ( det A ) E (72.5)

since by the skew symmetry of E we have

E ( Ae1 ,..., Ae M ) = ( det A ) E ( e1 ,..., e M ) (72.6)

for any {e1 ,..., e M } in W . From the representation of a left-invariant field on GL (V ) the value
of the field at any point X is obtained from the value at the identity I by the linear map

∇LX : L (V ;V ) → L (V ;V ) (72.7)

which is defined by

∇LX ( K ) = XK , K ∈L (V ;V ) (72.8)

Similiarly, a right-invariant field is obtained by the linear map ∇RX defined by

∇RX ( K ) = KX , K ∈L (V ;V ) (72.9)

Then by virtue of (72.5) a left-invariant field volume tensor field is also right-invariant if and
only if
Sec. 72 • Invariant Integrals on Continuous Groups 517

det ∇RX = det ∇LX (72.10)

Since the dimension of L (V ;V ) is N 2 , the matrices of ∇RX and ∇LX are N 2× N 2 . For
simplicity we use the product basis {ei ⊗ e j } for the space L (V ;V ) ; then (72.8) and (72.9) can
be represented by

( XK ) j = C ikjl K kl = X li Kkl = X liδ jk K kl

i
(72.11)

and

( KX ) j = D ikjl K kl = Kkl X kj = δ li X kj Kkl

i
(72.12)

To prove (72.10), we have to show that the N 2 × N 2 matrices ⎡⎣C ikjl ⎤⎦ and ⎡⎣ D ikjl ⎤⎦ have the same
determinant. But this fact is obvious since from (72.11) and (72.12), ⎡⎣C ikjl ⎤⎦ is simply the
transpose of ⎡⎣ D ikjl ⎤⎦ , i.e.,

D ikjl = Cljki (72.13)

Thus invariant integrals exist on GL (V ) .

The situation with the subgroup SL (V ) or UM (V ) is somewhat more complicated,

however, because the tangent planes at distinct points of the underlying group generally are not
the same subspace of L (V ;V ) . We recall first that the tangent plane of SL (V ) at the
identity I is the hyperplane consisting of all tensors K ∈L (V ;V ) such that

tr ( K ) = 0 (72.14)

This result means that the orthogonal complement of SL (V )I relative to the inner product on
L (V ;V ) is the one-dimensional subspace

l = {α I, α ∈R } (72.15)

Now from (72.8) and (72.9) the linear maps ∇RX and ∇LX coincide on l , namely

∇RX (α I ) = α X = ∇LX (α I ) (72.16)

518 Chap. 13 • INTEGRATION OF FIELDS

By virtue of (72.16) and (72.10) we see that the restrictions of ∇RX and ∇LX on SL (V )I give
rise to the same volume tensor at X from any volume tensor at I . As a result every left-
invariant volume tensor field on SL (V ) or UM (V ) is also a right-invariant, and thus invariant
integrals exist on SL (V ) andUM (V ) .

Finally, we show that invariant integrals exists on O (V ) and on all continuous subgroups
of O (V ) . This result is entirely obvious because both ∇LX and ∇RX preserve the inner product
on L (V ;V ) for any X ∈O (V ). Indeed, if K and H are any elements of L (V ;V ) , then

XK ⋅ XH = tr ( XKHT X T ) = tr ( XKHT X −1 )
(72.17)
= tr ( KHT ) = K ⋅ H

and similiarly

KX ⋅ HX = K ⋅ H (72.18)

for any X ∈O (V ). As a result, the Euclidean volume tensor field E is invariant on O (V ) and on
all continuous subgroups of O (V ) .

It should be noted that O (V ) is a bounded closed hypersurface in L (V ;V ). Hence the

integral of E over the whole group is finite

0< ∫ E<∞ (72.19)

O (V )

Such is not the case for GL (V ) or SL (V ) , since they are both unbounded in L (V ;V ) . By
virtue of (72.19) any continuous function f on O (V ) can be integrated with respect to E over
the entire group O (V ) . From (72.2) and (72.3) the integral possesses the following properties:

∫ f ( Q) E ( Q) = ∫ f ( Q 0Q ) E ( Q )
O (V ) O (V )
(72.20)
∫ f ( Q) E ( Q) = ∫ f ( QQ0 ) E ( Q )
O (V ) O (V )

for any Q0 ∈O (V ) and

Sec. 72 • Invariant Integrals on Continuous Groups 519

∫ f ( Q) E (Q) = ∫ f ( Q −1 ) E ( Q ) (72.21)
O (V ) O (V )

in addition to the standard properties of the integral relative to a differential form obtained in the
preceding section. For the unbounded groups GL (V ) , SL (V ) , and UM (V ) , some
continuous functions, such as functions which vanish identically outside some bounded domain,
can be integrated over the whole group. If the integral of f with respect to an invariant volume
tensor field exists, it also possesses properties similar to (72.20) and (72.21).

Now, by using the invariant integral on O (V ) , we can find a representation for

continuous isotropic functions

f : L (V ;V ) × ⋅ ⋅⋅ × L (V ;V ) →R (72.22)

which verify the condition of isotropy :

f ( QK 1QT ,..., QK P QT ) = f ( K 1 ,..., K P ) (72.23)

for all Q∈O (V ) , the number of variables P being arbitrary. The representation is

∫( g ( QK 1QT ,..., QK P QT ) E ( Q )
)
f ( K 1 ,..., K P )
O V
= (72.24)
∫( E ( Q)
O V )

where g is an arbitrary continuous function

g : L (V ;V ) × ⋅⋅⋅ × L (V ;V ) →R (72.25)

and where E is an arbitrary invariant volume tensor field on O (V ) . By a similar argument we

can also find representations for continuous functions f satisfying the condition

f ( QK 1 ,..., QK P ) = f ( K 1 ,..., K P ) (72.26)

for all Q∈O (V ). The representation is

∫( g ( QK 1 ,..., QK P ) E ( Q )
)
f ( K 1 ,..., K P )
O V
= (72.27)
∫( E (Q)
O V )
520 Chap. 13 • INTEGRATION OF FIELDS

Similiarly, a representation for functions f satisfying the condition

f ( K 1Q,..., K P Q ) = f ( K 1 ,..., K P ) (72.28)

for all Q∈O (V ) is

∫ O (V )g ( K1Q,..., K PQ ) E ( Q )
f ( K 1 ,..., K P ) = (72.29)
∫ O (V ) E ( Q )
We leave the proof of these representations as exercises.

If the condition (72.23), (72.26), or (72.28) is required to hold for all Q belonging to a
continuous subgroup G of O (V ) , the representation (72.24), (72.27), or (72.29), respectively,
remains valid except that the integrals in the representations are taken over the group G .
_______________________________________________________________________________
INDEX

The page numbers in this index refer to the versions of Volumes I and II of the online versions. In addition to the
information below, the search routine in Adobe Acrobat will be useful to the reader. Pages 1-294 will be found in the
online editions of Volume 1, pages 295-520 in Volume 2

Abelian group, 24, 33, 36, 37, 41 Associative algebra, 100

Absolute density, 276 Associative binary operation, 23, 24
Absolute volume, 291 Associative law for scalar multiplication, 41
Addition Asymptotic coordinate system, 460
Asymptotic direction, 460
for linear transformations, 93
Atlas, 308
for rings and fields, 33
Augmented matrix, 143
for subspaces, 55
Automorphism
for tensors, 228
of groups, 29
for vectors, 41
of linear transformations, 100, 109,
Additive group, 41 136, 168,
Adjoint linear transformation, 105-117 Axial tensor, 227
determinant of, 138 Axial vector densities, 268
matrix of, 138
Adjoint matrix, 8, 154, 156,279 Basis, 47
Affine space, 297 Anholonomic, 332
Algebra change of, 52, 76, 77, 118, 210, 225,
associative,100 266, 269, 275, 277, 287

tensor, 203-245 composite product, 336, 338

Algebraic equations, 9, 142-145 cyclic, 164, 193, 196

Algebraic multiplicity, 152, 159, 160, 161, dual, 204, 205, 215
164, 189, 191, 192 helical, 337
Algebraically closed field, 152, 188 natural, 316
Angle between vectors, 65 orthogonal, 70
Anholonomic basis, 332 orthonormal,70-75
Anholonomic components, 332-338 product, 221, 263
Associated homogeneous system of reciprocal, 76
equations, 147

x
INDEX xi

standard, 50, 116 Complete orthonormal set, 70

Beltrami field, 404 Completely skew-symmetric tensor, 248
Bessel’s inequality, 73 Completely symmetric tensor, 248
Bijective function, 19 Complex inner product space, 63
Bilinear mapping, 204 Complex numbers, 3, 13, 34, 43, 50, 51
Binary operation, 23 Complex vector space, 42, 63
Binormal, 391 Complex-lamellar fields, 382, 393, 400
Components
Cancellation axiom, 35 anholonomic, 332-338
Canonical isomorphism, 213-217 composite, 336
Canonical projection, 92, 96, 195, 197 composite physical, 337
Cartan parallelism, 466, 469 holonomic, 332
Cartan symbols, 472 of a linear transformation 118-121
Cartesian product, 16, 43, 229 of a matrix, 3
Cayley-Hamilton theorem, 152, 176, 191 physical, 334
Characteristic polynomial, 149, 151-157 of a tensor, 221-222
Characteristic roots, 148 of a vector, 51, 80
Characteristic subspace, 148, 161, 172, 189, Composition of functions, 19
191 Conformable matrices, 4,5
Characteristic vector, 148
Conjugate subsets of L (V ;V ) , 200
Christoffel symbols, 339, 343-345, 416
Continuous groups, 463
Closed binary operation, 23
Contractions, 229-234, 243-244
Closed set, 299
Contravariant components of a vector, 80,
Closure, 302 205
Codazzi, equations of, 451 Contravariant tensor, 218, 227, 235, 268
Cofactor, 8, 132-142, 267 Coordinate chart, 254
Column matrix, 3, 139 Coordinate curves, 254
Column rank, 138 Coordinate functions, 306
Commutation rules, 395 Coordinate map, 306
Commutative binary operation, 23 Coordinate neighborhood, 306
Commutative group, 24 Coordinate representations, 341
Commutative ring, 33 Coordinate surfaces, 308
Complement of sets, 14 Coordinate system, 308
xii INDEX

asymptotic, 460 Curl, 349-350

bispherical, 321 Curvature, 309
canonical, 430 Gaussian, 458
Cartesian, 309 geodesic, 438
curvilinear, 312 mean, 456
cylindrical, 312 normal, 438, 456, 458
elliptical cylindrical, 322 principal, 456
geodesic, 431
isochoric, 495 Decomposition, polar, 168, 173
normal, 432 Definite linear transformations, 167
orthogonal, 334 Density
paraboloidal, 320 scalar, 276
positive, 315 tensor, 267,271
prolate spheroidal, 321 Dependence, linear, 46
rectangular Cartesian, 309 Derivation, 331
Riemannian, 432 Derivative
spherical, 320 Covariant, 339, 346-347, 353, 411,
439
surface, 408
exterior, 374, 379
toroidal, 323
Lie, 359, 361, 365, 412
Coordinate transformation, 306
Determinant
Correspondence, one to one, 19, 97
of linear transformations, 137, 271-
Cosine, 66, 69
279
Covariant components of a vector, 80, 205
of matrices, 7, 130-173
Covariant derivative, 339, 346-347
Developable surface, 448
along a curve, 353, 419
Diagonal elements of a matrix, 4
spatial, 339
Diagonalization of a Hermitian matrix, 158-
surface, 416-417 175
total, 439 Differomorphism, 303
Covariant tensor, 218, 227, 235 Differential, 387
Covector, 203, 269 exact, 388
Cramer’s rule, 143 integrable, 388
Cross product, 268, 280-284 Differential forms, 373
Cyclic basis, 164, 193, 196
INDEX xiii

closed, 383 Dummy indices, 129

exact, 383
Dilitation, 145, 489 Eigenspace, 148
Dilatation group, 489 Eigenva1ues, 148
Dimension definition, 49 of Hermitian endomorphism, 158
of direct sum, 58 multiplicity, 148, 152, 161
of dual space, 203 Eigenvector, 148
of Euclidean space, 297 Element
of factor space, 62 identity,24
of hypersurface, 407 inverse, 24
of orthogonal group, 467 of a matrix, 3
of space of skew-symmetric tensors, of a set, 13
264 unit,24
of special linear group, 467 Empty set, 13
of tensor spaces, 221 Endomorphism
of vector spaces, 46-54 of groups, 29
Direct complement, 57 of vector spaces, 99
Direct sum Equivalence relation, 16, 60
of endomorphisms, 145 Equivalence sets, 16, 50
of vector spaces, 57 Euclidean manifold, 297
Disjoint sets, 14 Euclidean point space, 297
Distance function, 67 Euler-Lagange equations, 425-427, 454
Distribution, 368 Euler’s representation for a solenoidal field,
Distributive axioms, 33 400
Distributive laws for scalar and vector Even permutation, 126
addition, 41 Exchange theorem, 59
Divergence, 348 Exponential of an endomorphism, 169
Domain of a function, 18 Exponential maps
Dual basis, 204, 234 on a group, 478
Dual isomorphism, 213 on a hypersurface, 428
Dual linear transformation, 208 Exterior algebra, 247-293
Dual space, 203-212 Exterior derivative, 374, 379, 413
Duality operator, 280 Exterior product, 256
xiv INDEX

ε -symbol definition, 127 Gauss

transformation rule, 226 equations of, 433
formulas of, 436, 441
Factor class, 61 Gaussian curvature, 458
Factor space, 60-62, 195-197, 254 General linear group, 101, 113
Factorization of characteristic polynomial, Generalized Kronecker delta, 125
152 Generalized Stokes’ theorem, 507-513
Field, 35 Generalized transpose operation, 228, 234,
Fields 244, 247
Beltrami, 404 Generating set of vectors, 52
complex-lamellar, 382, 393, 400 Geodesic curvature, 438
scalar, 304 Geodesics, equations of, 425, 476
screw, 404 Geometric multiplicity, 148
solenoidal, 399 Gradient, 303, 340
tensor, 304 Gramian, 278
Trkalian, 405 Gram-Schmidt orthogonalization process, 71,
74
vector, 304
Greatest common divisor, 179, 184
Finite dimensional, 47
Group
Finite sequence, 20
axioms of, 23-27
Flow, 359
continuous, 463
Free indices, 129
general linear, 101, 464
Function
homomorphisms,29-32
constant, 324
orthogonal, 467
continuous, 302
properties of, 26-28
coordinate, 306
special linear, 457
definitions, 18-21
subgroup, 29
differentiable, 302
unitary, 114
Fundamental forms
first, 434
Hermitian endomorphism, 110, 139, 138-170
second, 434
Homogeneous operation, 85
Fundamental invariants, 151, 156, 169, 274,
278 Homogeneous system of equations, 142
Homomorphism
INDEX xv

of group,29 Invertible linear transformation, 99

of vector space, 85 Involution, 164
Irreduciable polynomials, 188
Ideal, 176 Isometry, 288
improper, 176 Isomorphic vector spaces, 99
principal, 176 Isomorphism
trivial, 176 of groups, 29
Identity element, 25 of vector spaces, 97
Identity function, 20
Jacobi’s identity, 331
Identity linear transformation, 98
Jordon normal form, 199
Identity matrix, 6
Image, 18
Kelvin’s condition, 382-383
Improper ideal, 178
Kernel
Independence, linear, 46
of homomorphisms, 30
Inequalities
of linear transformations, 86
Schwarz, 64
Kronecker delta, 70, 76
triangle, 65
generalized, 126
Infinite dimension, 47
Infinite sequence, 20
Lamellar field, 383, 399
Injective function, 19
Laplace expansion formula, 133
Inner product space, 63-69, 122, 235-245
Laplacian, 348
Integer, 13
Latent root, 148
Integral domain, 34
Latent vector, 148
Intersection
Law of Cosines, 69
of sets, 14
Least common multiple, 180
of subspaces, 56
Left inverse, 104
Invariant subspace, 145
Left-handed vector space, 275
Invariants of a linear transformation, 151
Left-invariant field, 469, 475
Inverse
Length, 64
of a function, 19
Levi-Civita parallelism, 423, 443, 448
of a linear transformation, 98-100
Lie algebra, 471, 474, 480
of a matrix, 6
Lie bracket, 331, 359
right and left, 104
xvi INDEX

Lie derivative, 331, 361, 365, 412 Maximal linear independent set, 47
Limit point, 332 Member of a set, 13
Line integral, 505 Metric space, 65
Linear dependence of vectors, 46 Metric tensor, 244
Linear equations, 9, 142-143 Meunier’s equation, 438
Linear functions, 203-212 Minimal generating set, 56
Linear independence of vectors, 46, 47 Minimal polynomial, 176-181
Linear transformations, 85-123 Minimal surface, 454-460
Logarithm of an endomorphism, 170 Minor of a matrix, 8, 133, 266
Lower triangular matrix, 6 Mixed tensor, 218, 276
Module over a ring, 45
Maps Monge’s representation for a vector field, 401
Continuous, 302 Multilinear functions, 218-228
Differentiable, 302 Multiplicity
Matrix algebraic, 152, 159, 160, 161, 164,
189, 191, 192
adjoint, 8, 154, 156, 279
geometric, 148
block form, 146
Negative definite, 167
column rank, 138
Negative element, 34, 42
diagonal elements, 4, 158
Negative semidefinite, 167
identity, 6
Negatively oriented vector space, 275, 291
inverse of, 6
Neighborhood, 300
nonsingular, 6
Nilcyclic linear transformation, 156, 193
of a linear transformation, 136
Nilpotent linear transformation, 192
product, 5
Nonsingular linear transformation, 99, 108
row rank, 138
Nonsingular matrix, 6, 9, 29, 82
skew-symmetric, 7
Norm function, 66
square, 4
Normal
trace of, 4
of a curve, 390-391
transpose of, 7
of a hypersurface, 407
triangular, 6
Normal linear transformation, 116
zero, 4
Normalized vector, 70
Maximal Abelian subalgebra, 487
Normed space, 66
Maximal Abelian subgroup, 486
INDEX xvii

N-tuple, 16, 42, 55 Parity

Null set, 13 of a permutation, 126, 248
Null space, 87, 145, 182 of a relative tensor, 226, 275
Nullity, 87 Partial ordering, 17
Permutation, 126
Odd permutation, 126 Perpendicular projection, 115, 165, 173
One-parameter group, 476 Poincaré lemma, 384
One to one correspondence, 19 Point difference, 297
Onto function, 19 Polar decomposition theorem, 168, 175
Onto linear transformation, 90, 97 Polar identity, 69, 112, 116, 280
Open set, 300 Polar tensor, 226
Order Polynomials
of a matrix, 3 characteristic, 149, 151-157
preserving function, 20 of an endomorphism, 153
of a tensor, 218 greatest common divisor, 179, 184
Ordered N-tuple, 16,42, 55 irreducible, 188
Oriented vector space, 275 least common multiplier, 180, 185
Orthogonal complement, 70, 72, 111, 115, minimal, 180
209, 216 Position vector, 309
Orthogonal linear transformation, 112, 201 Positive definite, 167
Orthogonal set, 70 Positive semidefinite, 167
Orthogonal subspaces, 72, 115 Positively oriented vector space, 275
Orthogonal vectors, 66, 71, 74 Preimage, 18
Orthogonalization process, 71, 74 Principal curvature, 456
Orthonormal basis, 71 Principal direction, 456
Osculating plane, 397
Principal ideal, 176
Parallel transport, 420 Principal normal, 391
Parallelism Product
of Cartan, 466, 469 basis, 221, 266
generated by a flow, 361 of linear transformations, 95
of Levi-Civita, 423 of matrices, 5
of Riemann, 423 scalar, 41
Parallelogram law, 69 tensor, 220, 224
xviii INDEX

wedge, 256-262 Relatively prime, 180, 185, 189

Projection, 93, 101-104, 114-116 Restriction, 91
Proper divisor, 185 Riemann-Christoffel tensor, 443, 445, 448-
449
Proper subgroup, 27
Riemannian coordinate system, 432
Proper subset, 13
Riemannian parallelism, 423
Proper subspace, 55
Right inverse, 104
Proper transformation, 275
Right-handed vector space, 275
Proper value, 148
Right-invariant field, 515
Proper vector, 148
Ring, 33
Pure contravariant representation, 237
commutative, 333
Pure covariant representation, 237
with unity, 33
Pythagorean theorem, 73
Roots of characteristic polynomial, 148
Row matrix, 3
Quotient class, 61
Row rank, 138
Quotient theorem, 227
r-form, 248

Radius of curvature, 391 r-vector, 248

Range of a function, 18
Rank Scalar addition, 41

of a linear transformation, 88 Scalar multiplication

of a matrix, 138 for a linear transformation, 93

Rational number, 28, 34 for a matrix, 4

Real inner product space, 64 for a tensor, 220

Real number, 13 for a vector space, 41

Real valued function, 18 Scalar product

Real vector space, 42 for tensors, 232

Reciprocal basis, 76 for vectors, 204

Reduced linear transformation, 147 Schwarz inequality, 64, 279

Screw field, 404
Regular linear transformation, 87, 113
Second dual space, 213-217
Relation, 16
Sequence
equivalence, 16
finite, 20
reflective, 16
infinite, 20
Relative tensor, 226
INDEX xix

Serret-Frenet formulas, 392 Standard basis, 50, 51

Sets Standard representation of Lie algebra, 470
bounded, 300 Stokes’ representation for a vector field, 402
closed, 300 Stokes theorem, 508
compact, 300) Strict components, 263
complement of, 14 Structural constants, 474
disjoint, 14 Subgroup, 27
empty or null, 13 proper,27
intersection of, 14 Subsequence, 20
open, 300 Subset, 13
retractable, 383 Subspace, 55
simply connected, 383 characteristic, 161
singleton, 13 direct sum of, 54
star-shaped, 383 invariant, 145
subset, 13 sum of, 56
union, 14 Summation convection, 129
Similar endomorphisms, 200 Surface
Simple skew-symmetric tensor, 258 area, 454, 493
Simple tensor, 223 Christoffel symbols, 416
Singleton, 13 coordinate systems, 408
Skew-Hermitian endomorphism, 110 covariant derivative, 416-417
Skew-symmetric endomorphism, 110 exterior derivative, 413
Skew-symmetric matrix, 7 geodesics, 425
Skew-symmetric operator, 250 Lie derivative, 412
Skew-symmetric tensor, 247 metric, 409
Solenoidal field, 399-401 Surjective function, 19
Spanning set of vectors, 52 Sylvester’s theorem, 173
Spectral decompositions, 145-201 Symmetric endomorphism, 110
Spectral theorem Symmetric matrix, 7
for arbitrary endomorphisms, 192 Symmetric operator, 255
for Hermitian endomorphisms, 165 Symmetric relation, 16
Spectrum, 148 Symmetric tensor, 247
Square matrix, 4 σ -transpose, 247
xx INDEX

for components of tensors, 225, 226,

Tangent plane, 406 239, 266, 287, 327
Tangent vector, 339 for components of vectors, 83-84,
210-211, 325
Tangential projection, 409, 449-450
for product basis, 225, 269
Tensor algebra, 203-245
for strict components, 266
Tensor product
Translation space, 297
of tensors, 224
Transpose
universal factorization property of,
229 of a linear transformation, 105
of vectors, 270-271 of a matrix, 7
Tensors, 218 Triangle inequality, 65
axial, 227 Triangular matrices, 6, 10
contraction of, 229-234, 243-244 Trivial ideal, 176
contravariant, 218 Trivial subspace, 55
covariant, 218 Trkalian field, 405
on inner product spaces, 235-245 Two-point tensor, 361
mixed, 218
polar, 226 Unimodular basis, 276
relative, 226 Union of sets, 14
simple, 223 Unit vector, 70
skew-symmetric, 248 Unitary group, 114
symmetric, 248 Unitary linear transformation, 111, 200
Terms of a sequence, 20 Unimodular basis, 276
Torsion, 392 Universal factorization property, 229
Total covariant derivative, 433, 439 Upper triangular matrix, 6
Trace
of a linear transformation, 119, 274, Value of a function, 18
278 Vector line, 390
of a matrix, 4 Vector product, 268, 280
Transformation rules Vector space, 41
for basis vectors, 82-83, 210 basis for, 50
for Christoffel symbols, 343-345 dimension of, 50
for components of linear direct sum of, 57
transformations, 118-122, 136-138
INDEX xxi

dual, 203-217
factor space of, 60-62
with inner product, 63-84
isomorphic, 99
normed, 66
Vectors, 41
angle between, 66, 74
component of, 51
difference,42
length of, 64
normalized, 70
sum of, 41
unit, 76
Volume, 329, 497

Wedge product, 256-262

Weight, 226
Weingarten’s formula, 434

Zero element, 24
Zero linear transformation, 93
Zero matrix,4
Zero N-tuple, 42
Zero vector, 42