Professional Documents
Culture Documents
TO
DIFFERENTIAL GEOMETRY
Philippe G. Ciarlet
City University of Hong Kong
Contents
Preface
iii
5
5
6
9
11
16
18
29
36
101
PREFACE
The notes presented here are based on lectures delivered over the years by
the author at the Universite Pierre et Marie Curie, Paris, at the University of
Stuttgart, and at City University of Hong Kong. Their aim is to give a thorough
introduction to the basic theorems of Differential Geometry.
In the first chapter, we review the basic notions arising when a threedimensional open set is equipped with curvilinear coordinates, such as the metric
tensor, Christoffel symbols, and covariant derivatives. We then prove that the
vanishing of the Riemann curvature tensor is sufficient for the existence of isometric immersions from a simply-connected open subset of Rn equipped with
a Riemannian metric into a Euclidean space of the same dimension. We also
prove the corresponding uniqueness theorem, also called rigidity theorem.
In the second chapter, we study basic notions about surfaces, such as their
two fundamental forms, the Gaussian curvature, Christoffel symbols, and covariant derivatives. We then prove the fundamental theorem of surface theory,
which asserts that the Gau and Codazzi-Mainardi equations constitute sufficient conditions for two matrix fields defined in a simply-connected open subset
of R3 to be the two fundamental forms of a surface in a three-dimensional Euclidean space. We also prove the corresponding rigidity theorem.
In addition to such classical theorems, we also include in both chapters
very recent results, which have not yet appeared in book form, such as the
continuity of a surface as a function of its fundamental forms.
The treatment is essentially self-contained and proofs are complete. The
prerequisites essentially consist in a working knowledge of basic notions of analysis and functional analysis, such as differential calculus, integration theory
and Sobolev spaces, and some familiarity with ordinary and partial differential
equations.
These notes use some excerpts from Chapters 1 and 2 of my book Mathematical Elasticity, Volume III: Theory of Shells, published in 2000 by NorthHolland, Amsterdam; in this respect, I am indebted to Arjen Sevenster for his
kind permission to reproduce these excerpts. Otherwise, the major part of these
notes was written during the fall of 2004 at City University of Hong Kong; this
part of the work was substantially supported by a grant from the Research
Grants Council of Hong Kong Special Administrative Region, China [Project
No. 9040869, CityU 100803].
Hong Kong, January 2005
iii
Chapter 1
THREE-DIMENSIONAL DIFFERENTIAL
GEOMETRY
1.1
CURVILINEAR COORDINATES
To begin with, we list some notations and conventions that will be consistently
used throughout.
All spaces, matrices, etc., considered here are real.
Latin indices and exponents vary in the set {1, 2, 3}, except when they are
used for indexing sequences, and the summation convention with respect to
repeated indices or exponents is systematically used in conjunction with this
rule. For instance, the relation
g i (x) = gij (x)g j (x)
means that
g i (x) =
3
X
j=1
[Ch. 1
g2 (x)
x3
x
x
e
e2
x2
g1 (x)
E3
e1
x1
e2
e1
b E3 . The three
Figure 1.1-1: Curvilinear coordinates and covariant bases in an open set
b If the three
coordinates x1 , x2 , x3 of x are the curvilinear coordinates of x
b = (x) .
vectors g i (x) = i (x) are linearly independent, they form the covariant basis at x
b = (x)
and they are tangent to the coordinate lines passing through x
b.
1.2
METRIC TENSOR
Sect. 1.2]
Metric tensor
E3
E3
Figure 1.1-2: Two familiar examples of curvilinear coordinates. Let the mapping be
defined by
: (, , z) ( cos , sin , z) E3 .
Then (, , z) are the cylindrical coordinates of x
b = (, , z). Note that ( + 2k, , z) or
( + + 2k, , z), k Z, are also cylindrical coordinates of the same point x
b and that is
not defined if x
b is the origin of E3 .
Let the mapping be defined by
: (, , r) (r cos cos , r cos sin , r sin ) E3 .
Then (, , r) are the spherical coordinates of x
b = (, , r). Note that ( + 2k, + 2`, r)
or ( + 2k, + + 2`, r) are also spherical coordinates of the same point x
b and that
and are not defined if x
b is the origin of E3 .
1 1 2 1
(x) := 1 2 2 2
1 3 2 3
3 1
3 2 (x).
3 3
i 1
g i (x) := i (x) = i 2 (x),
i 3
i.e., g i (x) is the i-th column vector of the matrix (x) and let x = xi ei .
Then the expansion of about x may be also written as
(x + x) = (x) + xi g i (x) + o(x).
If in particular x is of the form x = tei , where t R and ei is one of
the basis vectors in R3 , this relation reduces to
(x + tei ) = (x) + tg i (x) + o(t).
[Ch. 1
In other words, the principal part with respect to x of the length between
the points (x + x) and (x) is {xi g i (x) g j (x)xj }1/2 . This observation
suggests to define a matrix (gij (x)) of order three, by letting
gij (x) := g i (x) g j (x) = ((x)T (x))ij .
The elements gij (x) of this symmetric matrix are called the covariant components of the metric tensor at x
b = (x).
Note that the matrix (x) is invertible and that the matrix (gij (x)) is
positive definite, since the vectors g i (x) are assumed to be linearly independent.
The three vectors g i (x) being linearly independent, the nine relations
g i (x) g j (x) = ji
unambiguously define three linearly independent vectors g i (x). To see this, let
a priori g i (x) = X ik (x)g k (x) in the relations g i (x) g j (x) = ji . This gives
X ik (x)gkj (x) = ji ; consequently, X ik (x) = g ik (x), where
(g ij (x)) := (gij (x))1 .
Hence g i (x) = g ik (x)g k (x). These relations in turn imply that
g i (x) g j (x) = g ik (x)g k (x) g j` (x)g ` (x)
and thus the vectors g i (x) are linearly independent since the matrix (g ij (x)) is
positive definite. We would likewise establish that g i (x) = gij (x)g j (x).
The three vectors g i (x) form the contravariant basis at the point x
b = (x)
and the elements g ij (x) of the symmetric positive definite matrix (g ij (x)) are
the contravariant components of the metric tensor at x
b = (x).
Sect. 1.3]
1.3
We now review fundamental formulas showing how volume, area, and length
b = () can be expressed either in
elements at a point x
b = (x) in the set
terms of the matrix (x) or in terms of the matrix (gij (x)) or of its inverse
matrix (g ij (x)).
These formulas thus highlight the crucial r
ole played by the matrix (gij (x))
for computing metric notions at the point x
b = (x). Indeed, the metric
tensor well deserves its name!
A domain in Rn is a bounded, open, and connected subset D of R3 with
a Lipschitz-continuous boundary, the set D being locally on one side of its
boundary. All relevant details needed here about domains are found in Necas
[1967] or Adams [1975].
Given a domain D R3 with boundary , we let dx denote the volume
element in D, d denote the area element along , and n = ni b
ei denote the
unit (|n| = 1) outer normal vector along (d is well defined and n is defined
d-almost everywhere since is assumed to be Lipschitz-continuous).
Note also that the assumptions made on the mapping in the next theorem
b ,
b
guarantee that, if D is a domain in R3 such that D , then {D}
b of D
b and D of D are related by
{(D)} = (D), and the boundaries D
b = (D) (see, e.g., Ciarlet [1988, Theorem 1.2-8 and Example 1.7]).
D
If A is a square matrix, Cof A denotes the cofactor matrix of A. Thus
Cof A = (det A)AT if A is invertible.
A mapping : E3 is an immersion if it is an immersion at each
x , i.e., if is differentiable in and the three vectors g i (x) = i (x) are
linearly independent at each x .
Theorem 1.3-1. Let be an open subset of R3 , let : E3 be an injective
b = ().
and smooth enough immersion, and let
b is given in terms of the volume
(a) The volume element db
x at x
b = (x)
element dx at x by
p
db
x = | det (x)| dx = g(x) dx, where g(x) := det(gij (x)).
b x) at
(b) Let D be a domain in R3 such that D . The area element d(b
b
x
b = (x) D is given in terms of the area element d(x) at x D by
q
p
b x) = | Cof (x)n(x)| d(x) = g(x) ni (x)g ij (x)nj (x) d(x),
d(b
10
[Ch. 1
where x = xi ei .
b x) is
Either expression of the length element given in (c) recalls that d`(b
by definition the principal part with respect to x = xi ei of the length
|(x + x) (x)|, whose expression precisely led to the introduction of the
matrix (gij (x)) in Section 1.2.
The relations found in Theorem 1.3-1 are used in particular for computing
b by means of integrals inside , i.e., in terms
volumes, areas, and lengths inside
b (Figure 1.3-1):
of the curvilinear coordinates used in the open set
3
b
b
Let D be a domain in R such that D , let D := (D), and let fb L1 (D)
be given. Then
Z
Z
p
b
f (b
x) db
x=
(fb )(x) g(x) dx.
b
D
b is given by
In particular, the volume of D
Z
Z p
b
g(x) dx.
vol D :=
db
x=
b
D
b := ()
Next, let := D, let be a d-measurable subset of , let
1
b and let b
b be given. Then
D,
h L ()
Z
Z
q
p
b
b x) = (b
h(b
x) d(b
h )(x) g(x) ni (x)g ij (x)nj (x) d(x).
b
b is given by
In particular, the area of
Z
Z p
q
b :=
b x) =
area
d(b
g(x) ni (x)g ij (x)nj (x) d(x).
b
Sect. 1.4]
11
x)
d(
d(x)
V
dx
x+x
x
dl(
x)
d
x
(x+x)
R3
(x) = x
E3
f
t
Figure 1.3-1: Volume, area, and length elements in curvilinear coordinates. The elements
b x) at x
b x), and d`(b
b are expressed in terms of dx, d(x), and x at x by
db
x, d(b
b = (x)
means of the covariant and contravariant components of the metric tensor; cf. Theorem 1.3-1.
Given a domain D such that D and a d-measurable subset of D, the corresponding
b = (D) ,
b the area of
b = () D,
b
relations are used for computing the volume of D
b = (C) ,
b where C = f (I) and I is a compact interval of R.
and the length of a curve C
1.4
b of E3 by means of its
Suppose that a vector field is defined in an open subset
b
Cartesian components vbi : R, i.e., this field is defined by its values vbi (b
x)b
ei
b where the vectors b
at each x
b ,
ei constitute the orthonormal basis of E3 ; see
Figure 1.4-1.
b is equipped with curvilinear coordinates
Suppose now that the open set
3
from an open subset of R , by means of an injective mapping : E3
b
satisfying () = .
12
[Ch. 1
v3 (
x)
vi (
x) ei
x3
v2 (
x)
x
v1 (
x)
e3
E3
e2
x2
e1
x1
b the vector
Figure 1.4-1: A vector field in Cartesian coordinates. At each point x
b ,
vbi (b
x)b
ei is defined by its Cartesian components vbi (b
x) over an orthonormal basis of E3 formed
by three vectors b
ei .
How to define appropriate components of the same vector field, but this time
in terms of these curvilinear coordinates? It turns out that the proper way to
do so consists in defining three functions vi : R by requiring that (Figure
1.4-2)
vi (x)g i (x) := vbi (b
x)b
ei for all x
b = (x), x ,
Sect. 1.4]
13
g3 (x)
ui (x)g i (x)
g2 (x)
u3 (x)
u2(x)
x3
x
u1(x)
x
e3
e
e
e3
E3
x2
g1 (x)
x1
Figure 1.4-2: A vector field in curvilinear coordinates. Let there be given a vector field
b by its Cartesian components vbi (b
in Cartesian coordinates defined at each x
b
x) over the
vectors b
ei (Figure 1.4-1). In curvilinear coordinates, the same vector field is defined at each
x by its covariant components vi (x) over the contravariant basis vectors g i (x) in such a
way that vi (x)g i (x) = vbi (b
x)ei , x
b = (x).
The three components vi (x) are called the covariant components of the
vector vi (x)g i (x) at x
b, and the three functions vi : R defined in this fashion
are called the covariant components of the vector field vi g i : E3 .
Suppose next that we wish to compute a partial derivative bj vbi (b
x) at a point
b
x
b = (x) in terms of the partial derivatives ` vk (x) and of the values vq (x)
(which are also expected to appear by virtue of the chain rule). Such a task is
required for example if we wish to write a system of partial differential equations
whose unknown is a vector field (such as the equations of nonlinear or linearized
elasticity) in terms of ad hoc curvilinear coordinates.
As we now show, carrying out such a transformation naturally leads to a
fundamental notion, that of covariant derivatives of a vector field.
Theorem 1.4-1. Let be an open subset of R3 and let : E3 be an
b := (). Given a
immersion that is also a C 2 -diffeomorphism of onto
i b
3
b
b
vector field vbi e : R in Cartesian coordinates with components vbi C 1 (),
let vi g i : R3 be the same field in curvilinear coordinates, i.e., that defined
by
vbi (b
x)b
ei = vi (x)g i (x) for all x
b = (x), x .
Then vi C 1 () and for all x ,
where
and
bj vbi (b
x) = vkk` [g k ]i [g ` ]j (x), x
b = (x),
14
[Ch. 1
Proof. The following convention holds throughout this proof: The simultaneous appearance of x
b and x in an equality means that they are related by
x
b = (x) and that the equality in question holds for all x .
(i) Another expression of [g i (x)]k := g i (x) b
ek .
b x) =
b i (b
b :
b E3 denotes the
Let (x) = k (x)b
ek and (b
x)ei , where
3
b
inverse mapping of : E . Since ((x))
= x for all x , the chain
rule shows that the matrices (x) := (j k (x)) (the row index is k) and
b (b
b x) := (bk
b i (b
or equivalently,
j 1 (x)
b i (b
b i (b
b i (b
b i (b
bk
x)j k (x) = b1
x ) 2
x ) 3
x) j 2 (x) = ji .
j 3 (x)
The components of the above column vector being precisely those of the
vector g j (x), the components of the above row vector must be those of the
vector g i (x) since g i (x) is uniquely defined for each exponent i by the three
relations g i (x) g j (x) = ji , j = 1, 2, 3. Hence the k-th component of g i (x) over
b
the basis {b
e1 , b
e2 , b
e3 } can be also expressed in terms of the inverse mapping ,
as:
b i (b
[g i (x)]k = bk
x).
(ii) The functions q`k := g q ` g k C 0 ().
b q (b
Hence, noting that ` (g q (x) g k (x)) = 0 and [g q (x)]p = bp
x), we obtain
b q (b
q`k (x) = g q (x) ` g k (x) = bp
x)`k p (x) = qk` (x).
b C 1 (;
b R3 ) by assumption, the last relations
Since C 2 (; E3 ) and
q
0
show that `k C ().
Sect. 1.4]
15
where
and [g k (x)]i and q`k (x) are defined as in (i) and (ii).
= ` vk (x)[g ` (x)]j [g k (x)]i + vq (x) ` [g q (x)]i [g ` (x)]j
vikj = j vi pij vp
16
[Ch. 1
j g q = (j g q g p )g p = pqj g p .
Remark. The Christoffel symbols pij can be also defined solely in terms of
the components of the metric tensor; see the proof of Theorem 1.5-1.
If the affine space E3 is identified with R3 and = id , the relation
j (vi g i )(x) = (vikj g i )(x) (Theorem 1.4-2 (a)), reduces to bj (b
vi (b
x)b
ei ) = (bj vbi (b
x))b
ei .
In this sense, a covariant derivative of the first order constitutes a generalization
of a partial derivative of the first order in Cartesian coordinates.
1.5
Sect. 1.5]
17
they are denoted by the same symbol). Note that, according to the rule governing Latin indices and exponents, these relations are meant to hold for all
i, j, k, q {1, 2, 3}.
Theorem 1.5-1. Let be an open set in R3 , let C 3 (; E3 ) be an immersion, and let
gij := i j
denote the covariant components of the metric tensor of the set (). Let the
functions ijq C 1 () and pij C 1 () be defined by
1
(j giq + i gjq q gij ),
2
pij := g pq ijq where (g pq ) := (gij )1 .
ijq :=
Then, necessarily,
j ikq k ijq + pij kqp pik jqp = 0 in .
Proof. Let g i = i . It is then immediately verified that the functions ijq
are also given by
ijq = i g j g q .
For each x , let the three vectors g j (x) be defined by the relations g j (x)
g i (x) = jj . Since we also have g j = g ij g i , the last relations imply that pij =
i g j g p . Therefore,
i g j = pij g p
since i g j = (i g j g p )g p . Differentiating the same relations yields
k ijq = ik g j g q + i g j k g q ,
so that the above relations together give
i g j k g q = pij g p k g q = pij kqp .
Consequently,
Remark. The vectors g i and g j introduced above form the covariant and
contravariant bases and the functions g ij are the contravariant components of
the metric tensor (Section 1.2).
18
[Ch. 1
As shown in the above proof, the necessary conditions Rqijk = 0 thus simply constitute a re-writing of the relations ik g j = ki g j in the form of the
equivalent relations ik g j g q = ki g j g q .
The functions
ijq =
and
1
(j giq + i gjq q gij ) = i g j g q = jiq
2
pij = g pq ijq = i g j g p = pji
are the Christoffel symbols of the first, and second, kinds. We saw in
Section 1.4 that the same Christoffel symbols pij also naturally appear in a
different context (that of covariant differentiation).
Finally, the functions
Rqijk := j ikq k ijq + pij kqp pik jqp
are the covariant components of the Riemann curvature tensor of the
set (). The relations Rqijk = 0 found in Theorem 1.4-1 thus express that
the Riemann curvature tensor of the set () (equipped with the metric tensor
with covariant components gij ) vanishes.
1.6
Let M3 , S3 , and S3> denote the sets of all square matrices of order three, of
all symmetric matrices of order three, and of all symmetric positive definite
matrices of order three.
As in Section 1.2, the matrix representing the Frechet derivative at x of
a differentiable mapping = (` ) : E3 is denoted
(x) := (j ` (x)) M3 ,
where ` is the row index and j the column index (equivalently, (x) is the
matrix of order three whose j-th column vector is j ).
So far, we have considered that we are given an open set R3 and a
smooth enough immersion : E3 , thus allowing us to define a matrix field
C = (gij ) = T : S3> ,
where gij : R are the covariant components of the metric tensor of the
open set () E3 .
We now turn to the reciprocal questions:
Given an open subset of R3 and a smooth enough matrix field C = (gij ) :
S3> , when is C the metric tensor field of an open set () E3 ? Equivalently, when does there exist an immersion : E3 such that
C = T in ,
19
20
[Ch. 1
The answer to the second question, viz., the issue of uniqueness, can be
rephrased as follows (compare with the statement of Theorem 1.7-1 in the next
section): Let be a connected open subset of R3 . Then the isometric immersions
of a flat Riemannian manifold (; (gij )) into a Euclidean space E3 are unique
up to isometries of E3 . Recast as such, this result likewise becomes a special
case of the so-called rigidity theorem; cf. Section 1.7.
Recast as such, these two theorems together constitute a special case (that
where the dimensions of the manifold and of the Euclidean space are both equal
to three) of the fundamental theorem of Riemannian Geometry. This
theorem addresses the same existence and uniqueness questions in the more
general setting where is replaced by a p-dimensional manifold and E3 is replaced by a (p + q)-dimensional Euclidean space (the fundamental theorem of
surface theory, established in Sections 2.8 and 2.9, constitutes another important special case). When the p-dimensional manifold is an open subset of Rp ,
an outline of a self-contained proof is given in Szopos [2005].
Another fascinating question (which will not be addressed here) is the following: Given again an open subset of R3 equipped with a symmetric, positivedefinite matrix field (gij ) : S3 , assume this time that the Riemannian
manifold (; (gij )) is no longer flat, i.e., its Riemannian curvature tensor no
longer vanishes in . Can such a Riemannian manifold still be isometrically
immersed, but this time in a higher-dimensional Euclidean space? Equivalently,
do there exist a Euclidean space Ed with d > 3 and an immersion : Ed
such that gij = i j in ?
The answer is yes, according to the following beautiful Nash theorem, so
named after Nash [1954]: Any p-dimensional Riemannian manifold equipped
with a continuous metric can be isometrically immersed in a Euclidean space
of dimension 2p with an immersion of class C 1 ; it can also be isometrically
immersed in a Euclidean space of dimension (2p + 1) with a globally injective
immersion of class C 1 .
Let us now humbly return to the question of existence raised at the beginning
of this section, i.e., when the manifold is an open set in R3 .
Theorem 1.6-1. Let be a connected and simply-connected open set in R 3
and let C = (gij ) C 2 (; S3> ) be a matrix field that satisfies
Rqijk := j ikq k ijq + pij kqp pik jqp = 0 in ,
where
1
(j giq + i gjq q gij ),
2
pij := g pq ijq with (g pq ) := (gij )1 .
ijq :=
21
Proof. The proof relies on a simple, yet crucial, observation. When a smooth
enough immersion = (` ) : E3 is a priori given (as it was so far), its
components ` satisfy the relations ij ` = pij p ` , which are nothing but
another way of writing the relations i g j = pij g p (see the proof of Theorem
1.5-1). This observation thus suggests to begin by solving (see part (ii)) the
system of partial differential equations
i F`j = pij F`p in ,
whose solutions F`j : R then constitute natural candidates for the partial
derivatives j ` of the unknown immersion = (` ) : E3 (see part (iii)).
To begin with, we establish in (i) relations that will in turn allow us to
re-write the sufficient conditions
j ikq k ijq + pij kqp pik jqp = 0 in
in a slightly different form, more appropriate for the existence result of part (ii).
Note that the positive definiteness of the symmetric matrices (gij ) is not needed
for this purpose.
(i) Let be an open subset of R3 and let there be given a field (gij )
C (; S3 ) of symmetric invertible matrices. The functions ijq , pij , and g pq
being defined by
2
ijq :=
1
(j giq + i gjq q gij ),
2
pij := g pq ijq ,
(g pq ) := (gij )1 ,
Then
p
p
Rijk
= g pq Rqijk and Rqijk = gpq Rijk
.
22
[Ch. 1
p
and thus the relations Rijk
= g pq Rqijk are established. The relations Rqijk =
p
gpq Rijk are clearly equivalent to these ones.
We next establish the existence of solutions to the system
23
G
G
=
C 0 ([0, 1] [0, 1]; R3 ).
t
t
as this relation will imply that j (1, 0) = j (1, 1), as desired. For this purpose,
a direct differentiation shows that, for all 0 t 1, 0 1,
j
Gk Gi
Gi
Gi
= {qij pkq + k pij }p
+ pij p
+ q qij
,
t
t
t
t
24
[Ch. 1
where
j
Gk
pkj p
,
on the one hand (in the relations above and below, qij , k pij , etc., stand for
qij (G(, )), k pij (G(, )), etc.).
On the other hand, a direct differentiation of the equation defining the functions j shows that, for all 0 t 1, 0 1,
j :=
q o Gk
j j n p Gi
Gi
=
+ i kj
p + qkj
+ pij p
.
t
t
t
t
j
Gi
= pij
p , so that we also have
t
t
j j
Gi Gk
Gi
=
+ {i pkj + qkj piq }p
+ pij p
.
t
t
t
t
j
j
Hence, subtracting the above relations and noting that
=
t
t
Gi
Gi
and
=
by assumption, we infer that
t
t
But
j
Gk Gi
Gi
+ {i pkj k pij + qkj piq qij pkq }p
qij
q = 0.
t
t
t
The assumed symmetries pij = pji combined with the assumed relations
j pik k pij + `ik pj` `ij pk` = 0 in show that
i pkj k pij + qkj piq qij pkq = 0,
on the one hand. On the other hand,
j (0, ) =
Gk
j
(0, ) pkj (G(0, ))p (0, )
(0, ) = 0,
= 0 for all 0 1,
j (1, ) =
and thus
25
j
(1, ) = 0 for all 0 1, since G(1, ) = x1 for all 0 1.
For each integer `, we may thus unambiguously define a vector field (F`j ) :
R3 by letting
F`j (x1 ) := j (1) for any x1 ,
where C 1 ([0, 1]; R3 ) is any path joining x0 to x1 in and the vector field
(j ) C 1 ([0, 1]) is the solution to the Cauchy problem
d i
dj
(t) = pij ((t))
(t)p (t), 0 t 1,
dt
dt
j (0) = j0 ,
j (t) = j ( ) + (t )
for all t I. Equivalently,
26
[Ch. 1
1
(j giq + i gjq q gij ),
2
pij := g pq ijq ,
(g pq ) := (gij )1 .
0
Given an arbitrary point x0 , let (F`j
) S3> denote the square root of
0
the matrix (gij
) := (gij (x0 )) S3> .
Let (F`j ) C 2 (; M3 ) denote the solution to the corresponding system
which exists and is unique by parts (i) and (ii). Then there exists an immersion
= (` ) C 3 (; E3 ) such that
j ` = F`j and gij = i j in .
To begin with, we show that the three vector fields defined by
g j := (F`j )3`=1 C 2 (; R3 )
satisfy
g i g j = gij in .
To this end, we note that, by construction, these fields satisfy
i g j = pij g p in ,
g j (x0 ) = g 0j ,
0
where g 0j is the j-th column vector of the matrix (F`j
) S3> . Hence the matrix
field (g i g j ) C 2 (; M3 ) satisfies
m
k (g i g j ) = m
kj (g m g i ) + ki (g m g j ) in ,
0
(g i g j )(x0 ) = gij
.
27
satisfies
i F`j = j F`i in .
The open set being simply-connected, Poincares theorem (for a proof, see,
e.g., Schwartz [1992, Vol. 2, Theorem 59 and Corollary 1, p. 234235]) shows
that, for each integer `, there exists a function ` C 3 () such that
i ` = F`i in ,
or, equivalently, such that the mapping := (` ) C 3 (; E3 ) satisfies
i = g i in .
That is an immersion follows from the assumed invertibility of the matrices
(gij ). The proof is thus complete.
Remarks. (1) The assumptions
j pik k pij + `ik pj` `ij pk` = 0 in ,
made in part (ii) on the functions pij = pji are thus sufficient conditions for
the equations i F`j = pij F`p in to have solutions. Conversely, a simple
computation shows that they are also necessary conditions, simply expressing
that, if these equations have a solution, then necessarily ik F`j = ki F`j in .
It is no surprise that these necessary conditions are of the same nature as
those of Theorem 1.5-1, viz., ik g j = ij g k in .
(2) The assumed positive definiteness of the matrices (gij ) is used only in
part (iii), for defining ad hoc initial vectors g 0i .
28
[Ch. 1
The definitions of the functions pij and ijq imply that the functions
Rqijk := j ikq k ijq + pij kqp pik jqp
satisfy, for all i, j, k, p,
Rqijk = Rjkqi = Rqikj ,
Rqijk = 0 if j = k or q = i.
These relations in turn imply that the eighty-one sufficient conditions
Rqijk = 0 in for all i, j, k, q {1, 2, 3},
are satisfied if and only if the six relations
R1212 = R1213 = R1223 = R1313 = R1323 = R2323 = 0 in
are satisfied (as is immediately verified, there are other sets of six relations that
will suffice as well, again owing to the relations satisfied by the functions Rqijk
for all i, j, k, q).
To conclude, we briefly review various extensions of the fundamental existence result of Theorem 1.6-1. First, a quick look at its proof reveals that it
holds verbatim in any dimension d 2, i.e., with R3 replaced by Rd and E3 by
a d-dimensional Euclidean space Ed . This extension only demands that Latin
indices and exponents now range in the set {1, 2, . . . , d} and that the sets of matrices M3 , S3 , and S3> be replaced by their d-dimensional counterparts Md , Sd ,
and Sd> .
The regularity assumptions on the components gij of the symmetric positive
definite matrix field C = (gij ) made in Theorem 1.6-1, viz., that gij C 2 (),
can be significantly weakened. More specifically, C. Mardare [2003a] has shown
that the existence theorem still holds if gij C 1 (), with a resulting mapping
in the space C 2 (; Ed ); likewise, S. Mardare [2004] has shown that the existence
2,
theorem still holds if gij Wloc
(), with a resulting mapping in the space
2,
Wloc
(; Ed ). As expected, the sufficient conditions Rqijk = 0 in of Theorem
1.6-1 are then assumed to hold only in the sense of distributions, viz., as
Z
{ikq j + ijq k + pij kqp pik jqp } dx = 0
Sect. 1.7]
29
Ciarlet & C. Mardare [2004a] have also shown that, if in addition the geodesic
distance is equivalent to the Euclidean distance on (a property stronger than
the geodesic property, but again satisfied if the boundary is Lipschitzcontinuous), then a matrix field (gij ) C 2 (; Sn> ) with a Riemann curvature
e Sn ) defined
tensor vanishing in can be extended to a matrix field (e
gij ) C 2 (;
>
e containing and whose Riemann curvature tensor
on a connected open set
e This result relies on the existence of continuous extensions
still vanishes in .
to of the immersion and its partial derivatives of order 3 and on a deep
extension theorem of Whitney [1934].
1.7
30
[Ch. 1
e : E3 = R3 is
To begin with, we consider the special case where
the identity mapping. The issue of uniqueness reduces in this case to finding
C 1 (; E3 ) such that
(x)T (x) = I for all x .
Parts (i) to (iii) are devoted to solving these equations.
(i) We first establish that a mapping C 1 (; E3 ) that satisfies
(x)T (x) = I for all x
is locally an isometry: Given any point x0 , there exists an open neighborhood
V of x0 contained in such that
|(y) (x)| = |y x| for all x, y V.
Let B be an open ball centered at x0 and contained in . Since the set B is
convex, the mean-value theorem (for a proof, see, e.g., Schwartz [1992]) can be
applied. It shows that
|(y) (x)| sup |(z)||y x| for all x, y B.
z]x,y[
b 1 (b
The matrix
x) being thus orthogonal for all x
b Vb , the mean-value
theorem applied in the convex set Vb shows that
|1 (b
y ) 1 (b
x)| |b
yx
b| for all x
b, yb Vb ,
or equivalently, that
Sect. 1.7]
31
1 F
`
(x, y) =
(y)(` (y) ` (x)) i` (y` x` ) = 0
2 yi
yi
for all x, y V . For a fixed y V , each function Gi (, y) : V R is differentiable and its derivative vanishes. Consequently,
`
`
Gi
(x, y) =
(y)
(x) + ij = 0 for all x, y V,
xi
yi
xj
or equivalently, in matrix form,
(y)T (x) = I for all x, y V.
Letting y = x0 in this relation shows that
(x) = (x0 ) for all x V.
(iii) By (ii), the mapping : M3 is differentiable and its derivative
vanishes in . Therefore the mapping : E3 is twice differentiable and
its second Frechet derivative vanishes in . The open set being connected,
a classical result from differential calculus (see, e.g., Schwartz [1992, Theorem
3.7.10]) shows that the mapping is affine in , i.e., that there exists a vector
c E3 and a matrix Q M3 such that
(x) = c + Qox for all x .
Since Q = (x0 ) and (x0 )T (x0 ) = I by assumption, the matrix
Q is orthogonal.
(iv) We now consider the general equations gij = geij in , noting that they
also read
T
e
e
(x)T (x) = (x)
(x)
for all x .
32
[Ch. 1
b C 1 (Vb ; E3 ) and
Clearly,
b (b
b x) = (x)
e
b 1 (b
x)
1
e
= (x)(x) for all x
b = (x), x V.
imply that
T
e
e
(x)T (x) = (x)
(x)
for all x
b (b
b x )T
b (b
b x) = I for all x V.
1
e
(x) := (x)(x)
= Q for all x V.
Sect. 1.7]
33
e j
e in ,
while part (iv) provides the solution to the equations gij = i
1
3
e
where C (; E ) is a given immersion.
(2) The classical Mazur-Ulam theorem asserts the following: Let be a
connected subset in Rd , and let : Rd be a mapping that satisfies
|(y) (x)| = |y x| for all x, y .
Then there exist a vector c Rd and an orthogonal matrix Q of order d such
that
(x) = c + Qox for all x .
Parts (ii) and (iii) of the above proof thus provide a proof of this theorem
under the additional assumption that the mapping is of class C 1 (the extension
from R3 to Rd is trivial).
While the immersions found in Theorem 1.6-1 are thus only defined up to
isometries in E3 , they become uniquely determined if they are required to satisfy
ad hoc additional conditions, according to the following corollary to Theorems
1.6-1 and 1.7-1.
Theorem 1.7-2. Let the assumptions on the set and on the matrix field C
be as in Theorem 1.6-1, let a point x0 be given, and let F0 M3 be any
matrix that satisfies
FT0 F0 = C(x0 ).
Then there exists one and only one immersion C 3 (; E3 ) that satisfies
(x)T (x) = C(x) for all x ,
(x0 ) = 0 and (x0 ) = F0 .
Proof. Given any immersion C 3 (; E3 ) that satisfies (x)T (x) =
C(x) for all x (such immersions exist by Theorem 1.6-1), let the mapping
: R3 be defined by
(x) := F0 (x0 )1 ((x) (x0 )) for all x .
Then it is immediately verified that this mapping satisfies the announced
properties.
Besides, it is uniquely determined. To see this, let C 3 (; E3 ) and
C 3 (; E3 ) be two immersions that satisfy
(x)T (x) = (x)T (x) for all x .
Hence there exist (by Theorem 1.7-1) c R3 and Q O3 such that (x) =
c + Q(x) for all x , so that (x) = Q(x) for all x . The relation
(x0 ) = (x0 ) then implies that Q = I and the relation (x0 ) = (x0 )
in turn implies that c = 0.
34
[Ch. 1
Remark. One possible choice for the matrix F0 is the square root of the
symmetric positive-definite matrix C(x0 ).
Theorem 1.7-1 constitutes the classical rigidity theorem, in that both ime are assumed to be in the space C 1 (; E3 ). The next theorem
mersions and
is an extension, due to Ciarlet & C. Mardare [2003], that covers the case where
one of the mappings belongs to the Sobolev space H 1 (; E3 ).
The way the result in part (i) of the next proof is derived is due to Friesecke,
James & M
uller [2002]; the result of part (i) itself goes back to Reshetnyak
[1967].
Let O3+ denote the set of all rotations, i.e., of all orthogonal matrices Q O3
with det Q = 1.
Theorem 1.7-3. Let be a connected open subset of R3 , let C 1 (; E3 ) be
a mapping that satisfies
det > 0 in ,
e H 1 (; E3 ) be a mapping that satisfies
and let
e = 0 in (D0 (B))3
div Cof
in any open ball B such that B (to see this, combine the density of C 2 (B)
in H 1 (B) with the classical Piola identity in the space C 2 (B); for a proof of this
identity, see, e.g., Ciarlet [1988, Theorem 1.7.1]), we conclude that
e = div Cof
e = 0 in (D0 (B))3 .
e = (
e j ) (C ())3 . For such mappings, the identity
Hence
e j i
e j ) = 2i
e j i (
e j ) + 2ik
e j ik
e j,
(i
Sect. 1.7]
35
e j = 0 and i
e j i
e j = 3 in , shows that ik
ej =
together with the relations
0 in . The assumed connectedness of then implies that there exist a vector
e
c E3 and a matrix Q O3+ (by assumption, (x)
O3+ for almost all
x ) such that
e
(x)
= c + Q ox for almost all x .
b
{U}
(), such that the restriction U of to U can be extended to a
1
b .
C -diffeomorphism from U onto {U}
1
b
Let U : U U denote the inverse mapping of U , which therefore
1
b 1 (b
b (the notation
b indicates
satisfies
for all x
b = (x) U
U x) = (x)
b Define
that differentiation is carried out with respect to the variable x
b U).
the composite mapping
b :=
e 1 : U
b R3 .
U x) = (x)(x)
bU
+
3
3
(i), there thus exist c E and Q O+ such that
b x) = (x)
e
b,
(b
= c + Q ob
x for almost all x
b = (x) U
1
e
(x) := (x)(x)
= Q for almost all x U.
Since the point x0 is arbitrary, this relation shows that L1loc ().
By a classical result from distribution theory (cf. Schwartz [1966, Section 2.6]),
we conclude from the assumed connectedness of that (x) = Q for almost all
x , and consequently that
e
(x)
= c + Q(x) for almost all x .
36
[Ch. 1
1.8
Let be a connected and simply-connected open subset of R3 . Together, Theorems 1.6-1 and 1.7-1 establish the existence of a mapping F that associates
with any matrix field C = (gij ) C 2 (; S3> ) satisfying
Rqijk := j ikq k ijq + pij kqp pik jqp = 0 in ,
Sect. 1.8]
37
where the functions ijq and pij are defined in terms of the functions gij as in
Theorem 1.6-1, a well-defined element F(C) in the quotient set C 3 (; E3 )/R,
e R means that there exist a vector a E3 and a matrix Q O3
where (, )
e
such that (x) = a + Q(x)
for all x .
A natural question thus arises as to whether there exist natural topologies on
the space C 2 (; S3 ) and on the quotient set C 3 (; E3 )/R such that the mapping
F defined in this fashion is continuous.
Equivalently, is an immersion a continuous function of its metric tensor?
The object of this section, which is based on Ciarlet & Laurent [2003], is to
provide an affirmative answer to this question (see Theorem 1.8-5).
Note that such a question is not only clearly relevant to differential geometry per se, but it also naturally arises in nonlinear three-dimensional elasticity,
where a smooth enough immersion : E3 may be thought of as a deformation of the set viewed as a reference configuration of a nonlinearly elastic
body (although such an immersion should then be in addition injective and
orientation-preserving in order to qualify for this definition; for details, see, e.g.,
Ciarlet [1988, Section 1.4] or Antman [1995, Section 12.1]). In this context, the
associated matrix
C(x) = (gij (x)) = (x)T (x),
is called the (right) Cauchy-Green tensor at x and the matrix
(x) = (j i (x)) M3 ,
representing the Frechet derivative of the mapping at x, is called the deformation gradient at x.
The Cauchy-Green tensor field C = T : S3> associated with a
deformation : E3 plays a major role in the theory of nonlinear threedimensional elasticity, since the response function, or the stored energy function,
of a frame-indifferent elastic, or hyperelastic, material necessarily depends on
the deformation gradient through the Cauchy-Green tensor (see, e.g., Ciarlet
[1988, Chapters 3 and 4]. As already suggested by Antman [1976], the CauchyGreen tensor field of the unknown deformed configuration could thus also be
regarded as the primary unknown rather than the deformation itself as is
customary.
To begin with, we list some specific notations that will be used in this section
for addressing the question raised above. Given a matrix A M3 , we let (A)
denote its spectral radius and we let
|A| := sup
bR
b6=0
|Ab|
= {(AT A)}1/2
|b|
38
[Ch. 1
where stands for the standard multi-index notation for partial derivatives.
If C ` (; E3 ) or A C ` (; M3 ), ` 0, and K b , we likewise set
||`,K =
|A|`,K =
sup | (x)|
and
sup | A(x)|
and
xK
||=`
xK
||=`
kk`,K =
kAk`,K =
sup | (x)|,
xK
||`
sup | A(x)|,
xK
||`
where || denotes either the Euclidean vector norm or the matrix spectral norm.
The next sequential continuity results (Theorems 1.8-1, 1.8-2, and 1.8-3)
constitute key steps toward establishing the continuity of the mapping F (see
n
Theorem 1.8-5). Note that the functions Rqijk
occurring in their statements are
n
meant to be constructed from the functions gij
in the same way that the functions Rqijk are constructed from the functions gij . To begin with, we establish
the sequential continuity of the mapping F at C = I.
Theorem 1.8-1. Let be a connected and simply-connected open subset of
n
n
R3 . Let Cn = (gij
) C 2 (; S3> ), n 0, be matrix fields satisfying Rqijk
= 0 in
, n 0, such that
lim kCn Ik2,K = 0 for all K b .
Sect. 1.8]
39
Since
lim |Ap | = lim
((Ap )T Ap ) =
p
(I) = 1,
the sequence (Ap )p0 is bounded. Therefore there exists a further subsequence
(Aq )q0 that converges to a matrix S, which is orthogonal since
ST S = lim (Aq )T Aq = I.
q
But then
lim ST Aq = ST S = I,
which contradicts inf RO3 |RAq I| for all q 0. This proves (i).
In the remainder of this proof, the matrix fields Cn , n 0, are meant to be
those appearing in the statement of Theorem 1.8-1.
(ii) Let mappings n C 3 (; E3 ), n 0, satisfy (n )T n = Cn in
(such mappings exist by Theorem 1.6-1). Then
lim |n id|`,K = lim |n |`,K = 0 for all K b and for ` = 2, 3.
1
(j giq + i gjq q gij )g q .
2
1
n
n
n
q gij
)(g q )n , n 0,
(j giq
+ i gjq
2
40
[Ch. 1
n
n
since limn |gij
|1,K = limn |gij
ij |1,K = 0 by assumption. On the other
hand, the norms |(g q )n |0,K are bounded independently of n 0; to see this,
observe that (g q )n is the q-th column vector of the matrix (n )1 , then that
n 1 1/2
n 1
= |{((gij
) )} |0,K {|(gij
) |0,K }1/2 ,
Consequently,
lim |n id|2,K = lim |n |2,K = 0 for all K b .
n
n
Observing that limn |gij
|`,K = limn |gij
ij |`,K = 0 for ` = 1, 2 by
q n
assumption and recalling that the norms |(g ) |0,K are bounded independently
of n 0, we likewise conclude that
e n C 3 (; E3 ) that satisfy (
e n )T
e n = Cn
(iii) There exist mappings
in , n 0, and
n
e n C 3 (; E3 ), n 0, defined by
Then the mappings
satisfy
e n )T
e n = Cn in ,
(
Sect. 1.8]
41
e C 2 (; M3 ) satisfy
so that their gradients
e n |0,K = lim |
e n |2,K = 0 for all K b ,
lim |i
The mappings
clearly satisfy
n
e {
e n (x0 ) x0 } C 3 (; E3 ), n 0,
n :=
(n )T n = Cn in , n 0,
n (x0 ) = x0 , n 0.
Again applying the theorem about the differentiability of the limit of a sequence of mappings used in part (iii), we conclude from the last two relations
that the mappings n uniformly converge on every compact subset of toward
a limit C 1 (; E3 ) that satisfies
(x) = lim n (x) = I for all x .
n
42
[Ch. 1
This shows that ( id) is a constant mapping since is connected. Consequently, = id since in particular (x0 ) = limn n (x0 ) = x0 . We have
thus established that
lim |n id|0,K = 0 for all K b .
b of ,
b let K := (
b K).
b Since limn kg n
Given any compact subset K
ij
gij k2,K = 0 because K is a compact subset of , the definition of the functions
n
b R and the chain rule together imply that
gbij
:
n
lim kb
gij
ij k2,K
b = 0.
b let bi = /b
b n denote the functions conGiven x
b = (b
xi ) ,
xi . Let R
qijk
n
structed from the functions gbij in the same way that the functions Rqijk are
constructed from the functions gij . Since it is easily verified that these funcb n = 0 in ,
b Theorem 1.8-1 applied over the set
b shows that
tions satisfy R
qijk
n
3 b
3
b
there exist mappings C (; E ) satisfying
such that
b n bj
b n = gbn in ,
b n 0,
bi
ij
c b = 0 for all K
b n idk
b b ,
b
lim k
3,K
Sect. 1.8]
43
where c
id denotes the identity mapping of E3 , identified here with R3 . Define
the mappings n C 3 (; S3> ), n 0, by letting
b n (b
b x) .
n (x) =
x) for all x = (b
Proof. The proof is broken into four parts. In what follows, C and Cn
designate matrix fields possessing the properties listed in the statement of the
theorem.
(i) Let C 3 (; E3 ) be any mapping that satisfies T = C in .
ThenSthere exist a countable number of open balls BSr , r 1, such that
r
= r=1 Br and such that, for each r 1, the set s=1 Bs is connected and
the restriction of to Br is injective.
Given any x , there exists anSopen ball Vx such that the restriction
of to Vx is injective. Since = x Vx can also be written as a countable
union of compact subsets of , there already
exist countably many such open
S
balls, denoted Vr , r 1, such that =
r=1 Vr .
Let r1 := 1, B1 := Vr1 , and r2 := 2. If the set Br1 Vr2 is connected,
let B2 := Vr2 and r3 := 3. Otherwise, there exists a path 1 in joining
the centers of Vr1 and Vr2 since is connected. Then there exists a finite set
I1 = {r1 (1), r1 (2), , r1 (N1 )} of integers, with N1 1 and 2 < r1 (1) < r1 (2) <
< r1 (N1 ), such that
[
1 V r1 V r2
Vr .
rI1
44
[Ch. 1
S 1 +2
If the set ( N
r=1 Br ) Vr3 is connected, let BN1 +3 := Vr3 . Otherwise, apply
the same argument as above to a path 2 in joining the centers of Vr2 and
Vr3 , and so forth.
The iterative procedure thus produces a countable number of open
Sballs
Br , r 1, that possess the announced properties. In particular, = r=1 Br
since, by construction, the integer ri appearing at the i-th stage satisfies ri i.
en
(ii) By Theorem 1.8-2, there exist mappings n1 C 3 (B1 ; E3 ) and
2
C 3 (B2 ; E3 ), n 0, that satisfy
(n1 )T n1 = Cn in B1
e n )T
e n = Cn in B2
(
2
2
and
and
2
1
Then we assert that
Let (Qp )p0 be a subsequence of the sequence (Qn )n0 that converges to
a (necessarily orthogonal) matrix Q and let x1 denote a point in the set B1
e p (x1 ) Qp 1 (x1 ) and limn
e p (x1 ) = limn p (x1 ) =
B2 . Since cp =
2
2
1
p
(x1 ), the subsequence (c )p0 also converges. Let c := limp cp . Thus
e p2 (x)
(x) = lim
p
on the one hand. On the other hand, the differentiability of the mapping
implies that
(x) = (x1 ) + (x1 )(x x1 ) + o(|x x1 |) for all x B1 B2 .
Note that (x1 ) is an invertible matrix, since (x1 )T (x1 ) = (gij (x1 )).
Let b := (x1 ) and A := (x1 ). Together, the last two relations imply
that
b + A(x x1 ) = c + Qb + QA(x x1 ) + o(|x x1 |),
Sect. 1.8]
45
Then
(n2 )T n2 = Cn in B1 B2
The plane containing the intersection of the boundaries of the open balls B1
and B2 is the common boundary of two closed half-spaces in R3 , H1 containing
the center of B1 , and H2 containing that of B2 (by construction, the set B1 B2
is connected; see part (i)). Any compact subset K of B1 B2 may thus be written
as K = K1 K2 , where K1 := (K H1 ) B1 and K2 := (K H2 ) B2 (that
the open sets found in part (i) may be chosen as balls thus play an essential r
ole
here). Hence
lim kn2 k3,K1 = 0 and lim kn2 k3,K2 = 0,
r
[
Bs ,
s=1
r
[
Bs .
s=1
Since the restriction of to Br+1 is injective (part (i)), Theorem 1.8-2 shows
e nr+1 C 3 (Br+1 ; E3 ), n 0, that satisfy
that there exist mappings
e n )T
e n = Cn in Br+1 ,
(
r+1
r+1
n
e
lim k
r+1 k3,K = 0 for all K b Br+1 ,
46
[Ch. 1
S
and since the set r+1
s=1 Bs is connected (part (i)), Theorem 1.7-1 shows that
there exist vectors cn E3 and matrices Qn O3 , n 0, such that
e n (x) = cn + Qn n (x) for all x
r+1
r
r
[
s=1
Bs Br+1 .
Then an argument similar to that used in part (ii) shows that limn Qn = I
and limn cn = 0, and an argument similar to that used in part (iii) (note
that the ball Br+1 may intersect more
Sr than one of the balls Bs , 1 s r)
shows that the mappings nr+1 C 3 ( s=1 Bs ; E3 ), n 0, defined by
nr+1 (x) := nr (x) for all x
satisfy
r
[
Bs ,
s=1
r
[
Bs .
s=1
r
[
s=1
Bs , r 1,
possess all the required properties:SThey are unambiguously defined since for all
r
s > r, ns (x) = nr (x) forSall x s=1 Bs by construction; they are of class C 3
r
n
since the mappings r : s=1 Bs E3 are themselves of class C 3 ; they satisfy
n T
n
n
n
(
r satisfy the same relations
Sr ) = C in since the mappings
n
in s=1 Bs ; and finally, they satisfy limn k
S k3,K = 0 for all K b
since any compact subset of is contained in rs=1 Bs for r large enough. This
completes the proof.
It is easily seen that the assumptions Rqijk = 0 in are in fact superfluous
in Theorem 1.8-3 (as shown in the next proof, these relations are consequences
n
of the assumptions Rqijk
= 0 in , n 0, and limn kCn Ck2,K = 0 for
all K b ). This observation gives rise to the following corollary to Theorem
1.8-3, in the form of another sequential continuity result, of interest by itself. The
novelties are that the assumptions are now made on the immersions n , n 0,
and that this result also provides the existence of a limit immersion .
Theorem 1.8-4. Let be a connected and simply-connected open subset of
R3 . Let there be given immersions n C 3 (; E3 ), n 0, and a matrix field
C C 2 (; S3> ) such that
lim k(n )T n Ck2,K = 0 for all K b .
Sect. 1.8]
47
e n C 3 (; E3 ), n 0, of the form
Then there exist mappings
e n = c n + Q n n , c n E 3 , Qn O 3 ,
e n )T
e n = (n )T n in for all n 0, and there
which thus satisfy (
exists a mapping C 3 (; E3 ) such that
n
n
Proof. Let the functions Rqijk
, n 0, and Rqijk be constructed from the
n
components gij and gij of the matrix fields Cn := (n )T n and C in the
n
usual way (see, e.g., Theorem 1.6-1). Then Rqijk
= 0 in for all n 0, since
these relations are simply the necessary conditions of Theorem 1.5-1.
We now show that Rqijk = 0 in . To this end, let K be any compact subset
of . The relations
n
limn kRqijk
hence that
Rqijk k0,K = 0. This shows that Rqijk = 0 in K,
hence that Rqijk = 0 in since K is an arbitrary compact subset of .
By the fundamental existence theorem (Theorem 1.6-1), there thus exists a
mapping C 3 (; E3 ) such that T = C in . Theorem 1.8-3 can now
e n C 3 (; E3 ) such that
be applied, showing that there exist mappings
n
e )T
e = Cn in , n 0, and lim k
e k3,K for all K b .
(
n
Finally, the rigidity theorem (Theorem 1.7-1) shows that, for each n 0,
e n = cn + Qn n in because
there exist cn E3 and Qn O3 such that
n
e and n share the same metric tensor field and the set is
the mappings
connected.
48
[Ch. 1
Furthermore, this topology is metrizable: Let (Ki )i0 be any sequence of subsets
of that satisfy
Ki b and Ki int Ki+1 for all i 0, and =
Ki .
i=0
Then
lim kn k`,K = 0 for all K b lim d` (n , ) = 0,
where
d` (, ) :=
X
1 k k`,Ki
.
i 1 + k k
2
`,Ki
i=0
cE
QO3
Sect. 1.8]
49
Proof. Since {C02 (; S3> ); d2 } and {C3 (; E3 ); d3 } are both metric spaces, it
suffices to show that convergent sequences are mapped through F into convergent sequences.
Let then C C02 (; S3> ) and Cn C02 (; S3> ), n 0, be such that
lim d2 (Cn , C) = 0,
i.e., such that limn kCn Ck2,K = 0 for all K b . Given any
F(C), Theorem 1.8-3 shows that there exist n F(Cn ), n 0, such that
limn kn k3,K = 0 for all K b , i.e., such that
lim d3 (n , ) = 0.
Consequently,
As shown by Ciarlet & C. Mardare [2004b], the above continuity result can
be extended up to the boundary of the set , as follows. If is bounded and
has a Lipschitz-continuous boundary, the mapping F of Theorem 1.8-5 can be
extended to a mapping that is locally Lipschitz-continuous with respect to the
topologies of the Banach spaces C 2 (; S3 ) for the continuous extensions of the
symmetric matrix fields C, and C 3 (; E3 ) for the continuous extensions of the
immersions (the existence of such continuous extensions is briefly commented
upon at the end of Section 1.6).
Another extension, motivated by three-dimensional nonlinear elasticity, is
the following: Let be a bounded and connected subset of R3 , and let B
be an elastic body with as its reference configuration. Thanks mostly to
the landmark existence theory of Ball [1977], it is now customary in nonlinear
three-dimensional elasticity to view any mapping H 1 (; E3 ) that is almosteverywhere injective and satisfies det > 0 a.e. in as a possible deformation
of B when B is subjected to ad hoc applied forces and boundary conditions. The
almost-everywhere injectivity of (understood in the sense of Ciarlet & Necas
[1987]) and the restriction on the sign of det mathematically express (in
an arguably weak way) the non-interpenetrability and orientation-preserving
conditions that any physically realistic deformation should satisfy.
As mentioned earlier, the Cauchy-Green tensor field T L1 (; S3 )
associated with a deformation H 1 (; E3 ) pervades the mathematical modeling of three-dimensional nonlinear elasticity. Conceivably, an alternative approach to the existence theory in three-dimensional elasticity could thus regard
the Cauchy-Green tensor as the primary unknown, instead of the deformation
itself as is usually the case.
Clearly, the Cauchy-Green tensors depend continuously on the deformations,
since the Cauchy-Schwarz inequality immediately shows that the mapping
H 1 (; E3 ) T L1 (; S3 )
50
[Ch. 1
One application of the above key inequality is the following sequential continuity property: Let k H 1 (; Ed ), k 1, and C 1 (; Ed ) be orientationpreserving mappings. Then there exist a constant C() and orientation-prese k H 1 (; Ed ), k 1, that are isometrically equivalent to
erving mappings
k
such that
k
1/2
e kH 1 (;Ed ) C()k(k )T k T k 1
k
.
L (;Sd )
e k ) converges to in H 1 (; Ed ) as k if the
Hence the sequence (
k=1
T
1
d
sequence ((k )T k )
k=1 converges to in L (; S ) as k .
Sect. 1.8]
51
Chapter 2
DIFFERENTIAL GEOMETRY OF SURFACES
2.1
In addition to the rules governing Latin indices that we set in Section 1.1, we
henceforth require that Greek indices and exponents vary in the set {1, 2} and
that the summation convention be systematically used in conjunction with these
rules. For instance, the relation
(i ai ) = (| b 3 )a + (3| + b )a3
means that, for = 1, 2,
3
X
i=1
2
2
X
X
i ai =
(| b 3 )a + 3| +
b a3 .
=1
=1
is called a surface in E3 .
b := ()
53
54
[Ch. 2
a2 (y)
(y)
a2 (y)
a1 (y)
a1 (y)
=()
E3
y2
R2
y1
Figure 2.1-1: Curvilinear coordinates on a surface and covariant and contravariant bases of
the tangent plane. Let
b = () be a surface in E3 . The two coordinates y1 , y2 of y are
the curvilinear coordinates of yb = (y)
b . If the two vectors a (y) = (y) are linearly
independent, they are tangent to the coordinate lines passing through yb and they form the
covariant basis of the tangent plane to
b at yb = (y). The two vectors a (y) from this tangent
plane defined by a (y) a (y) = form its contravariant basis.
Sect. 2.1]
55
y
x
x
u
Figure 2.1-2: Several systems of curvilinear coordinates on a sphere. Let be a sphere
of radius R. A portion of contained in the northern hemisphere can be represented by
means of Cartesian coordinates, with a mapping of the form:
: (x, y) (x, y, {R2 (x2 + y 2 )}1/2 ) E3 .
A portion of that excludes a neighborhood of both poles and of a meridian (to fix
ideas) can be represented by means of spherical coordinates, with a mapping of the form:
: (, ) (R cos cos , R cos sin , R sin ) E3 .
A portion of that excludes a neighborhood of the North pole can be represented by
means of stereographic coordinates, with a mapping of the form:
2R2 u
2R2 v
u2 + v 2 R 2
: (u, v)
,
,
R
E3 .
u2 + v 2 + R 2 u2 + v 2 + R 2
u2 + v 2 + R 2
The corresponding coordinate lines are represented in each case, with self-explanatory
graphical conventions.
1
56
[Ch. 2
Figure 2.1-3: Two familiar examples of surfaces and curvilinear coordinates. A portion
b
of a circular cylinder of radius R can be represented by a mapping of the form
: (, z) (R cos , R sin , z) E3 .
A portion
b of a torus can be represented by a mapping of the form
: (, ) ((R + r cos ) cos , (R + r cos ) sin , r sin ) E3 ,
with R > r.
The corresponding coordinate lines are represented in each case, with self-explanatory
graphical conventions.
Sect. 2.2]
2.2
57
1 1 2 1
(y) := 1 2 2 2 (y).
1 3 2 3
1
a (y) := (y) = 2 (y),
3
i.e., a (y) is the -th column vector of the matrix (y) and let y = y e .
Then the expansion of about y may be also written as
(y + y) = (y) + y a (y) + o(y).
If in particular y is of the form y = te , where t R and e is one of
the basis vectors in R2 , this relation reduces to
(y + te ) = (y) + ta (y) + o(t).
A mapping : E3 is an immersion at y if it is differentiable at
y and the 3 2 matrix (y) is of rank two, or equivalently if the two vectors
a (y) = (y) are linearly independent.
Assume from now on in this section that the mapping is an immersion
at y. In this case, the last relation shows that each vector a (y) is tangent
to the -th coordinate line passing through yb = (y), defined as the image
by of the points of that lie on a line parallel to e passing through y
(there exist t0 and t1 with t0 < 0 < t1 such that the -th coordinate line is
given by t ]t0 , t1 [ f (t) := (y + te ) in a neighborhood of yb; hence
f 0 (0) = (y) = a (y)); see Figures 2.1-1, 2.1-2, and 2.1-3.
The vectors a (y), which thus span the tangent plane to the surface
b at
yb = (y), form the covariant basis of the tangent plane to
b at yb; see
Figure 2.1-1.
Returning to a general increment y = y e , we also infer from the expansion of about y that
|(y + y) (y)|2 = y T (y)T (y)y + o(|y|2 )
= y a (y) a (y)y + o(|y|2 ).
58
[Ch. 2
In other words, the principal part with respect to y of the length between
the points (y + y) and (y) is {y a (y) a (y)y }1/2 . This observation
suggests to define a matrix (a (y)) of order two by letting
a (y) := a (y) a (y) = (y)T (y) .
The elements a (y) of this symmetric matrix are called the covariant
components of the first fundamental form, also called the metric tensor,
of the surface
b at yb = (y).
Note that the matrix (a (y)) is positive definite since the vectors a (y) are
assumed to be linearly independent.
The two vectors a (y) being thus defined, the four relations
a (y) a (y) =
unambiguously define two linearly independent vectors a (y) in the tangent
plane. To see this, let a priori a (y) = Y (y)a (y) in the relations a (y)
a (y) = . This gives Y (y)a (y) = ; hence Y (y) = a (y), where
(a (y)) := (a (y))1 .
Hence a (y) = a (y)a (y). These relations in turn imply that
a (y) a (y) = a (y)a (y)a (y) a (y)
and thus the vectors a (y) are linearly independent since the matrix (a (y))
is positive definite. We would likewise establish that a (y) = a (y)a (y).
The two vectors a (y) form the contravariant basis of the tangent plane
to the surface
b at yb = (y) (Figure 2.1-1) and the elements a (y) of the
symmetric matrix (a (y)) are called the contravariant components of the
first fundamental form, or metric tensor, of the surface
b at yb = (y).
Let us record for convenience the fundamental relations that exist between
the vectors of the covariant and contravariant bases of the tangent plane and
the covariant and contravariant components of the first fundamental tensor:
a (y) = a (y) a (y)
a (y) = a (y)a (y)
Sect. 2.3]
2.3
59
dS(
y)
(y+y) A
y)
dl(
(y) = y
E3
y y+y
dy
A
R2
C
t
b y ) at
Figure 2.3-1: Area and length elements on a surface. The elements db
a(b
y ) and d`(b
yb = (y)
b are related to dy and y by means of the covariant components of the metric
tensor of the surface
b ; cf. Theorem 2.3-1. The corresponding relations are used for computing
b = (A)
b = (C)
the area of a surface A
b and the length of a curve C
b , where C = f (I)
and I is a compact interval of R.
60
[Ch. 2
b y ) at yb = (y)
(b) The length element d`(b
b is given by
b y ) = y a (y)y 1/2 .
d`(b
Proof. The relation (a) between the area elements is well known. It can
b x) and
also be deduced directly from the relation between the area elements d(b
d(x) given in Theorem 1.3-1 (b) by means of an ad hoc three-dimensional
extension of the mapping .
b y ) is by definition
The expression of the length element in (b) recalls that d`(b
the principal part with respect to y = y e of the length |(y + y) (y)|,
whose expression precisely led to the introduction of the matrix (a (y)).
The relations found in Theorem 2.3-1 are used for computing surface integrals and lengths on the surface
b by means of integrals inside , i.e., in terms
of the curvilinear coordinates used for defining the surface
b (see again Figure
2.3-1).
Let A be a domain in R2 such that A (a domain in R2 is a bounded,
open, and connected subset of R2 with a Lipschitz-continuous boundary; cf.
b := (A), and let fb L1 (A)
b be given. Then
Section 1.3), let A
Z
Z
p
fb(b
y ) db
a(b
y ) = (fb )(y) a(y) dy.
b
A
b is given by
In particular, the area of A
Z
Z p
b :=
area A
db
a(b
y) =
a(y) dy.
b
A
The last relation shows in particular that the lengths of curves inside the
surface () are precisely those induced by the Euclidean metric of the space E3 .
For this reason, the surface () is said to be isometrically imbedded in E3 .
2.4
Sect. 2.4]
61
2
Figure 2.4-1: A metric alone does not define a surface in E3 . A flat surface
b0 may be
deformed into a portion
b1 of a cylinder or a portion
b 2 of a cone without altering the length
of any curve drawn on it (cylinders and cones are instances of developable surfaces; cf.
Section 2.5). Yet it should be clear that in general
b 0 and
b1 , or
b0 and
b2 , or
b1 and
b2 ,
are not identical surfaces modulo an isometry of E3 !
62
[Ch. 2
a(y),
Sect. 2.4]
63
(s)
p(s) + R(s)
(s + s)
(s + s)
(s)
p(s + s)
(s)
p(s)
Figure 2.4-2: Curvature of a planar curve. Let be a smooth enough planar curve,
parametrized by its curvilinear abscissa s. Consider two points p(s) and p(s + s) with
curvilinear abscissae s and s + s and let (s) be the algebraic angle between the two
normals (s) and (s + s) (oriented in the usual way) to at those points. When s 0,
(s)
has a limit, called the curvature of at p(s). If this limit is non-zero, its
the ratio
s
inverse R is called the algebraic radius of curvature of at p(s) (the sign of R depends on
the orientation chosen on ).
The point p(s) + R(s), which is intrinsically defined, is called the center of curvature
of at p(s): It is the center of the osculating circle at p(s), i.e., the limit as s 0 of the
circle tangent to at p(s) that passes through the point p(s + s). The center of curvature is
also the limit as s 0 of the intersection of the normals (s) and (s + s). Consequently,
the centers of curvature of lie on a curve (dashed on the figure), called la developpee in
French, that is tangent to the normals to .
1
b at yb is given by the ratio
of the planar curve C
R
df df
(t)
(t)
1
dt
dt
=
,
R
df
df
a (f (t))
(t)
(t)
dt
dt
b (f (t))
where a (y) are the covariant components of the first fundamental form of
b
at y (Section 2.1) and
b (y) := a3 (y) a (y) = a3 (y) a (y) = b (y).
64
[Ch. 2
a3(y)
y =(y)
y + Ra3(y)
E3
e2
e1
R2
y = f (t)e
df
(t)e
dt
f = f e
t
b
the curve C = f (I) at y = f (t) = f (t)e . If
6= 0, the center of curvature of the curve C
R
at yb is the point (b
y + Ra3 (y)), which is intrinsically defined in the Euclidean space E3 .
1
of
Proof. (i) We first establish a well-known formula giving the curvature
R
a planar curve. Using the notations of Figure 2.4-2, we note that
sin (s) = (s) (s + s) = {(s + s) (s)} (s + s),
so that
1
(s)
sin (s)
d
:= lim
= lim
= (s) (s).
s0 s
s0
R
s
ds
Sect. 2.5]
65
1
b is given by
of C
R
1
d
= (s) (s),
R
ds
where
d(a3 f ) dt
df dt
d
(s) =
(t)
= a3 (f (t))
(t) ,
ds
dt
ds
dt
ds
dp
d( f ) dt
(s) =
(s) =
(t)
ds
dt
ds
df dt
df dt
= (f (t))
(t)
= a (f (t))
(t) .
dt
ds
dt
ds
Hence
df df
1
= a3 (f (t)) a (f (t))
(t)
(t)
R
dt
dt
dt
ds
2
1
, it suffices to note that
R
66
2.5
[Ch. 2
The analysis of the previous section suggests that precise information about the
shape of a surface
b = () in a neighborhood of one of its points yb = (y)
can be gathered by letting the plane P turn around the normal vector a3 (y)
and by following in this process the variations of the curvatures at yb of the
corresponding planar curves P
b , as given in Theorem 2.4-1.
As a first step in this direction, we show that these curvatures span a compact
interval of R. In particular then, they stay away from infinity.
Note that this compact interval contains 0 if, and only if, the radius of
curvature of the curve P
b is infinite for at least one such plane P .
Theorem 2.5-1. (a) Let the assumptions and notations be as in Theorem 2.41. For a fixed y , consider the set P of all planes P normal to the surface
where (a (y)) = (a (y))1 (Section 2.2) and the matrix (b (y)) is defined as
in Theorem 2.4-1. Then
1
1
+
= b11 (y) + b22 (y),
R1 (y) R2 (y)
1
det(b (y))
= b11 (y)b22 (y) b21 (y)b12 (y) =
.
R1 (y)R2 (y)
det(a (y))
1
1
6=
, there is a unique pair of orthogonal planes P1 P
R1 (y)
R2 (y)
and P2 P such that the curvatures of the associated planar curves P1
b and
1
1
P2
b are precisely
and
.
R1 (y)
R2 (y)
(c) If
Proof. (i) Let (P ) denote the intersection of P P with the tangent plane
b ) denote the intersection of P with
T to the surface
b at yb, and let C(P
b . Hence
b ) at yb
(P ) is tangent to C(P
b.
b )
In a sufficiently small neighborhood of yb the restriction of the curve C(P
b
to this neighborhood is given by C(P ) = ( f (P ))(I(P )), where I(P ) R
is an open interval and f (P ) = f (P )e : I(P ) R2 is a smooth enough
df (P )
injective mapping that satisfies
(t)e 6= 0, where t I(P ) is such that
dt
y = f (P )(t). Hence the line (P ) is given by
d( f (P ))
(P ) = yb +
(t); R = {b
y + a (y); R} ,
dt
Sect. 2.5]
67
df (P )
(t) and e 6= 0 by assumption.
dt
b ))
Since the line {y+ e ; R} is tangent to the curve C(P ) := 1 (C(P
3
at y (the mapping : R is injective by assumption) for each
such parametrizing function f (P ) : I(P ) R2 and since the vectors a (y)
are linearly independent, there exists a bijection between the set of all lines
(P ) T , P P, and the set of all lines supporting the nonzero tangent
vectors to the curve C(P ).
Hence Theorem 2.4-1 shows that when P varies in P, the curvature of the
b = C(P
b ) at yb takes the same values as does the ratio
corresponding curves C
b (y)
when := ( ) varies in R2 {0}.
a (y)
where :=
a (y)
T
A
is nothing but the Rayleigh quotient associated with the symmetric matrix
C1 BC1 . When varies in R2 {0}, this Rayleigh quotient thus spans the
compact interval of R whose end-points are the smallest and largest eigenvalue,
1
1
respectively denoted
and
, of the matrix C1 BC1 (for a proof,
R1 (y)
R2 (y)
see, e.g., Ciarlet [1982, Theorem 1.3-1]). This proves (a).
Furthermore, the relation
C(C1 BC1 )C1 = BC2 = BA1
shows that the eigenvalues of the symmetric matrix C1 BC1 coincide with
those of the (in general non-symmetric) matrix BA1 . Note that BA1 =
(b (y)) with b (y) := a (y)b (y), being the row index, since A1 =
(a (y)).
Hence the relations in (b) simply express that the sum and the product of
the eigenvalues of the matrix BA1 are respectively equal to its trace and to its
det(b (y))
determinant, which may be also written as
since BA1 = (b (y)).
det(a (y))
This proves (b).
(iii) Let 1 = (1 ) = C 1 and 2 = (2 ) = C 2 , with 1 = (1 ) and
2 = (2 ), be two orthogonal ( T1 2 = 0) eigenvectors of the symmetric matrix
1
1
and
, respectively.
C1 BC1 , corresponding to the eigenvalues
R1 (y)
R2 (y)
Hence
0 = T1 2 = T1 CT C 2 = T1 A2 = 0,
68
[Ch. 2
since CT = C. By (i), the corresponding lines (P1 ) and (P2 ) of the tangent
plane are parallel to the vectors 1 a (y) and 2 a (y), which are orthogonal
since
1 a (y) 2 a (y) = a (y)1 2 = T1 A 2 .
1
1
6=
, the directions of the vectors 1 and 2 are uniquely
R1 (y)
R2 (y)
determined and the lines (P1 ) and (P2 ) are likewise uniquely determined.
This proves (c).
If
Sect. 2.5]
69
Figure 2.5-1: Different kinds of points on a surface. A point is elliptic if the Gaussian
curvature is > 0 or equivalently, if the two principal radii of curvature are of the same sign;
the surface is then locally on one side of its tangent plane. A point is parabolic if exactly one
of the two principal radii of curvature is infinite; the surface is again locally on one side of its
tangent plane. A point is hyperbolic if the Gaussian curvature is < 0 or equivalently, if the
two principal radii of curvature are of different signs; the surface then intersects its tangent
plane along two curves.
70
[Ch. 2
where the genus g(S) is the number of holes of S (for instance, a sphere
has genus zero, while a torus has genus one). The integer (S) defined by
(S) := (2 2g(S)) is the Euler characteristic of
b.
According to the definition of Stoker [1969, Chapter 5, Section 2], a developable surface is one whose Gaussian curvature vanishes everywhere. Developable surfaces are otherwise often defined as ruled surfaces whose Gaussian
curvature vanishes everywhere, as in, e.g., Klingenberg [1973, Section 3.7]). A
portion of a plane provides a first example, the only one of a developable surface
all points of which are planar. Any developable surface all points of which are
parabolic can be likewise fully described: It is either a portion of a cylinder,
or a portion of a cone, or a portion of a surface spanned by the tangents to a
skewed curve. The description of a developable surface comprising both planar
and parabolic points is more subtle (although the above examples are in a sense
the only ones possible, at least locally; see Stoker [1969, Chapter 5, Sections 2
to 6]).
The interest of developable surfaces is that they can be, at least locally,
continuously rolled out, or developed (hence their name), onto a plane,
without changing the metric of the intermediary surfaces in the process.
Sect. 2.6]
2.6
71
a1 (y) a2 (y)
,
|a1 (y) a2 (y)|
y .
Then the vectors a (y) (which form the covariant basis of the tangent plane to
b at yb = (y); see Figure 2.1-1) together with the vector a3 (y) (which is normal
to
b and has Euclidean norm one) form the covariant basis at yb.
i (y)ai (y)
3 (y)
a3 (y)
a2 (y)
2 (y) a1 (y)
1 (y)
= ( )
E3
y2
R2
y1
Figure 2.6-1: Contravariant bases and vector fields along a surface. At each point yb =
(y)
b = (), the three vectors ai (y), where a (y) form the contravariant basis of the
a1 (y) a2 (y)
, form the contravariant
tangent plane to
b at yb (Figure 2.1-1) and a3 (y) =
|a1 (y) a2 (y)|
basis at yb. An arbitrary vector field defined on
b may then be defined by its covariant
components i : R. This means that i (y)a i (y) is the vector at the point yb.
72
[Ch. 2
a = C
a + b a3 and a = C
a + b a3 ,
3
a3 = a = b a = b a ,
where the covariant and mixed components b and b of the second fundamental
form of
b are defined in Theorems 2.4-1 and 2.5-1 and
C
:= a a .
(i ai ) = ( C
b 3 )a + ( 3 + b )a3
= (| b 3 )a + (3| + b )a3 ,
where
| := C
and 3| := 3 .
Sect. 2.6]
73
a = ( a a )a + ( a a3 )a3 = C
a + b a3 ,
by definition of C
and b . Finally,
a = ( a a )a + ( a a3 )a3 = C
a + b a3 ,
since
a a3 = a a3 = b a a = b .
a = C
a + b a3 and a = C
a + b a3
and
a3 = a3 = b a = b a ,
respectively constitute the formulas of Gau and Weingarten. The functions
(also found in Theorem 2.6-1)
| = C
and 3| = 3
C
:= a a = a a
74
[Ch. 2
for such tangential fields by Theorem 2.6-1. The reason is that a surface has in
3
general a nonzero curvature, manifesting itself here by the extra term b
a .
2.7
C C + C
C C
C = b b b b in ,
b b + C
b C
b = 0 in ,
n o
1
2
|1 2 |
75
denote the covariant components of the first and second fundamental forms of
C :=
Then, necessarily,
C C + C
C C
C = b b b b in ,
b b + C
b C
b = 0 in .
a1 a2
and, for each y , let the three vectors aj (y) be defined
|a1 a2 |
by the relations aj (y) ai (y) = ij . Since we also have a = a a and a3 = a3 ,
a = C
a + b a3 ,
a a = C
a a + b a3 a = C
C + b b .
Consequently,
a a = C C
C b b .
a a = C C
C b b .
76
[Ch. 2
a a3 = C
b .
Consequently,
a a3 = b + C
b .
a a3 = b + C
b .
1
( a + a a ) = a a = C
2
and
C
= a C = a a = C
are the Christoffel symbols of the first, and second, kind. We recall that
S := C C + C
C C
C
77
1
det(b (y)
=
(Theorem 2.5-1). By inspection of the function
R1 (y)R2 (y)
det(a (y))
S1212 , we thus reach the astonishing conclusion that, at each point of the surface,
a notion involving the curvature of the surface, viz., the Gaussian curvature,
is entirely determined by the knowledge of the metric of the surface at the
same point, viz., the components of the first fundamental forms and their partial
derivatives of order 2 at the same point! This startling conclusion naturally
deserves a theorem:
since
Theorem 2.7-2. Let be an open subset of R2 , let C 3 (; E3 ) be an immersion, let a = denote the covariant components of the first fundamental form of the surface (), and let the functions C and S1212 be defined
by
1
( a + a a ),
2
1
:= (212 a12 11 a22 22 a11 ) + a (C12 C12 C11 C22 ).
2
C :=
S1212
Then, at each point (y) of the surface (), the principal curvatures
and R21(y) satisfy
1
S1212 (y)
=
, y .
R1 (y)R2 (y)
det(a (y))
1
R1 (y)
2.8
Let M2 , S2 , and S2> denote the sets of all square matrices of order two, of all
symmetric matrices of order two, and of all symmetric, positive definite matrices
of order two.
So far, we have considered that we are given an open set R2 and a
smooth enough immersion : E3 , thus allowing us to define the fields
(a ) : S2> and (b ) : S2 , where a : R and b : R
are the covariant components of the first and second fundamental forms of the
surface () E3 .
Note that the immersion need not be injective in order that these matrix
fields be well defined.
We now turn to the reciprocal questions:
Given an open subset of R2 and two smooth enough matrix fields (a ) :
S2> and (b ) : S2 , when are they the first and second fundamental
forms of a surface () E3 , i.e., when does there exist an immersion :
E3 such that
n o
1
2
a := and b :=
in ?
|1 2 |
78
[Ch. 2
79
C C + C
C C
C = b b b b in ,
b b + C
b C
b = 0 in ,
where
1
( a + a a ),
2
:= a C where (a ) := (a )1 .
C :=
1 2
, and let
|1 2 |
gij := i j .
80
[Ch. 2
where
(B) := b and (Bn ) := b1 bn1 for n 2,
i.e., (Bn ) designates for any n 0 the element at the -th row and -th
column of the matrix Bn . The above series are absolutely convergent in the
space C 2 (0 ).
P
h
0 a + x3 (h1 a 2h0 b )
X
+
xn3 (h
n a 2hn1 b + hn2 c ) = .
n2
h
(Bn ) , n 0,
n = (n + 1)a
so that
g =
n0
(n + 1)xn3 a b1 bn1 .
81
C
:= a C ,
where
(a ) := (a )1 and C :=
1
( a + a a ),
2
b| := b C
b C
b = b| .
Then
b | = a b| and b| = a b | .
Furthermore, the assumed Codazzi-Mainardi equations imply that
b | = b | and b| = b| .
The above relations follow from straightforward computations based on the
definitions of the functions b | and b| . They are recorded here because they
play a pervading r
ole in the subsequent computations.
(iii) The functions gij C 2 (0 ) and g ij C 2 (0 ) being defined as in part (i),
define the functions ijq C 1 (0 ) and pij C 1 (0 ) by
ijq :=
1
(j giq + i gjq q gij ) and pij := g pq ijq .
2
Then the functions ijq = jiq and pij = pji have the following expressions:
= C x3 (b | a + 2C
b ) + x23 (b | b + C
c ),
3 = 3 = b x3 c ,
33 = 33 = 33q = 0,
X
= C
xn+1
b | (Bn ) ,
3
n0
3 = b x3 c ,
X
3 =
xn3 (Bn+1 ) ,
n0
33
p33
= 0,
where the functions c , (Bn ) , and b | are defined as in parts (i) and (ii).
All computations are straightforward. We simply point out that the assumed
Codazzi-Mainardi equations are needed to conclude that the factor of x3 in the
function is indeed that announced above. We also note that the computation of the factor of x23 in relies in particular on the easily established
relations
c + C
c .
c = b | b + b | b + C
82
[Ch. 2
R23 = 0,
R33 = 0 in 0 .
The above definition of the functions Rqijk and that of the functions ijq
and pij (part (iii)) together imply that, for all i, j, k, q,
Rqijk = Rjkqi = Rqikj ,
Rqijk = 0 if j = k or q = i.
Consequently, the relation R1212 = 0 implies that R = 0, the relations
R23 = 0 imply that Rqijk = 0 if exactly one index is equal to 3, and finally,
the relations R33 = 0 imply that Rqijk = 0 if exactly two indices are equal
to 3.
(v) The functions
R33 := 33 3 3 + p3 3p p33 p
satisfy
R33 = 0 in 0 .
These relations immediately follow from the expressions found in part (iii)
for the functions ijq and pij . Note that neither the Gau equations nor the
Codazzi-Mainardi equations are needed here.
(vi) The functions
R23 := 23 3 2 + p2 3p p23 p
satisfy
R23 = 0 in 0 .
The definitions of the functions g (part (i)) and ijq (part (iii)) show that
23 3 2 = (2 b b2 ) + x3 ( c2 2 c ).
Then the expressions found in part (iii) show that
p2 3p p23 p = 3 2 23
b
= C
b2 C2
+ x3 (b2 | b b | b2 + C2
c C
c2 ),
83
c = b | b + b | b + C
c + C
c
2 b b2 + C
b2 C2
b = 0,
12 12 11 22 = (C12
C12
C11
C22
)a
+ x3 {(C11 b2 |2 2C12
b1 |2 + C22
b1 |1 )a
+ 2(C11
C22
C12
C12
)b }
+ x23 {b1 |1 b2 |2 b1 |2 b1 |2 )a
+ (C11
b2 |2 2C12
b1 |2 + C22
b1 |1 )b
+ (C11
C22
C12
C12
)c }.
+ 2(C11
C22
C12
C12
)b }
+ x23 {S 12 b1 b2 + (b1 |1 b2 |2 b1 |2 b1 |2 )a
b1 |1 )b
+ (C11
b2 |2 2C12
b1 |2 + C22
+ (C11
C22
C12
C12
)c },
84
[Ch. 2
S := C C + C
C C
C
are precisely those appearing in the left-hand sides of the Gau equations.
It is then easily seen that the above equations together yield
R1212 = {S1212 (b11 b22 b12 b12 )}
x3 {S1212 (b11 b22 b12 b12 )b
}
Since
S 12 b1 b2 = S1212 (b11 b22 b21 b12 ),
c12 c12 c11 c22 = (b11 b12 b11 b22 )(b11 b22 b21 b12 ),
it is finally found that the function R1212 has the following remarkable expression:
R1212 = {S1212 (b11 b22 b12 b12 )}{1 x3 (b11 + b22 ) + x23 (b11 b22 b21 b12 )}.
By the assumed Gau equations,
S1212 = b11 b22 b12 b12 .
Hence R1212 = 0 as announced.
(viii) Let be a connected and simply-connected open subset of R2 . Then
there exist open subsets ` , ` 0, of R2 such that ` is a compact subset of
for each ` 0 and
[
=
` .
`0
Furthermore, for each ` 0, there exists ` = ` (` ) > 0 such that the symmetric matrices (gij ) are positive definite at all points in ` , where
` := ` ]` , ` [ .
Finally, the open set
:=
`0
85
and
for all 0 t 1, 0 1,
for all 0 t 1, 1 2,
n o
1
2
in .
|1 2 |
86
[Ch. 2
= 2x3 a3 + x23 a3 a3 in .
87
2
and (b ) Lloc (; S ) are two matrix fields that satisfy the Gau and CodazziMainardi equations in the sense of distributions, then there exists a mapping
2,
Wloc
() such that (a ) and (b ) are the fundamental forms of the surface ().
2.9
88
[Ch. 2
Proof. Arguments similar to those used in parts (i) and (viii) of the proof
of Theorem 2.8-1 show that there exist open subsets ` of and real numbers
` > 0, ` 0, such that the symmetric matrices (gij ) defined by
g := a 2x3 b + x23 c and gi3 = i3 ,
where c := a b b , are positive definite in the set
[
:=
` ]` , ` [ .
`0
e C 1 (; E3 ) defined by (with
The two immersions C 1 (; E3 ) and
self-explanatory notations)
e
e
e 3 (y)
(y, x3 ) := (y) + x3 a3 (y) and (y,
x3 ) := (y)
+ x3 a
gij = geij in .
e
det (y, x3 ) = det Q det (y,
x3 ) for all (y, x3 ) .
e
det (y, x3 ) = det (y,
x3 ) for all (y, x3 ) .
89
e2 |
|e
a1 a
H 1 (; E3 ) (with self-explanatory notations). Naturally, our first task will be to
e 3 , which is not necessarily well defined a.e. in for
verify that the vector field a
e H 1 (; E3 ), is nevertheless well defined a.e. in for
an arbitrary mapping
e that satisfy the assumptions of the next theorem. This fact
those mappings
e 3 are likewise well defined
will in turn imply that the functions eb := e
a a
a.e. in .
e 3 H 1 (; E3 ),
a
and
eb = b a.e. in .
Proof. The proof essentially relies on the extension to a Sobolev space setting
of the three-dimensional rigidity theorem established in Theorem 1.7-3.
(i) To begin with, we record several technical preliminaries.
First, we observe that the relations e
a = a a.e. in and the assumption
that C 1 (; E3 ) is an immersion together imply that
e2| =
|e
a1 a
q
q
det(e
a ) = det(a ) > 0 a.e. in .
90
[Ch. 2
e a
e3 = a
e a
e 3 a.e. in . To
i.e., that a a3 = a a3 in and a
this end, we note that either the assumptions C 1 (; E3 ) and a3 C 1 (; E3 )
together, or the assumptions H 1 (; E3 ) and a3 H 1 (; E3 ) together, imply
that a a3 = a3 L1loc (), hence that a3 D0 ().
Given any D(), let U denote an open subset of R2 such that supp U
and U is a compact subset of . Denoting by X 0 h, iX the duality pairing
between a topological vector space X and its dual X 0 , we have
Z
a3 dy
D 0 () h a3 , iD() =
Z
Z
=
(a3 ) dy ( ) a3 dy.
H 1 (U ;E3 )
h ( ), a3 iH01 (U ;E3 ) ,
91
Then it is clear that is a connected open subset of R3 and that the mapping
C 1 (; E3 ) satisfies det > 0 in .
Finally, note that the covariant components gij C 0 () of the metric tensor
field associated with the mapping are given by (the symmetries b = b
established in (i) are used here)
g = a 2x3 b + x23 c ,
g3 = 0,
g33 = 1.
well defined a.e. in , are equal, again a.e. in , to the functions b . Likewise, the
components geij L1 () of the metric tensor field associated with the mapping
e satisfy geij = gij a.e. in since e
a = a and eb = b a.e. in by
assumption and ce = c a.e. in by part (i).
(iv) By Theorem 1.7-2, there exist a vector c E3 and a matrix Q O3+
such that
e
e 3 (y) = c + Q((y) + x3 a3 (y)) for almost all (y, x3 ) .
(y)
+ x3 a
92
2.10
[Ch. 2
Let be a connected and simply-connected open subset of R2 . Together, Theorems 2.8-1 and 2.9-1 establish the existence of a mapping F that associates
to any pair of matrix fields (a ) C 2 (; S2> ) and (b ) C 2 (; S2 ) satisfying the Gau and Codazzi-Mainardi equations in a well-defined element
e R means that
F ((a ), (b )) in the quotient set C 3 (; E3 )/R, where (, )
3
3
e
there exists a vector c E and a rotation Q O+ such that (y) = c + Q(y)
for all y .
A natural question thus arises as to whether there exist ad hoc topologies on
the space C 2 (; S2 ) C 2 (; S2 ) and on the quotient set C 3 (; E3 )/R such that
the mapping F defined in this fashion is continuous.
Equivalently, is a surface a continuous function of its fundamental forms?
The purpose of this section, which is based on Ciarlet [2003], is to provide
an affirmative answer to the above question, through a proof that relies in an
essential way on the solution to the analogous problem in dimension three given
in Section 1.8.
Such a question is not only relevant to surface theory, but it also finds
its source in two-dimensional nonlinear shell theories, where the stored energy
functions are often functions of the first and second fundamental forms of the
unknown deformed middle surface (for an overview of nonlinear shell theories,
see, e.g., Ciarlet [2000]). For instance, the well-known stored energy function
wK proposed by Koiter [1966, Equations (4.2), (8.1), and (8.3)] for modeling
nonlinearly elastic shells made with a homogeneous and isotropic elastic material
takes the form:
wK =
3
a
(e
a a )(e
a a ) + a (eb b )(eb b ),
2
6
4
a a + 2(a a +a a ),
+ 2
> 0 and > 0 are the two Lame constants of the constituting material, a
and b are the covariant components of the first and second fundamental forms
of the given undeformed middle surface, (a ) = (a )1 , and finally e
a and
eb are the covariant components of the first and second fundamental forms of
the unknown deformed middle surface.
An inspection of the above stored energy functions thus suggests a tempting
approach to shell theory, where the functions e
a and eb would be regarded
as the primary unknowns in lieu of the customary (Cartesian or curvilinear)
components of the displacement. In such an approach, the unknown components
e
a and eb must naturally satisfy the classical Gau and Codazzi-Mainardi
equations in order that they actually define a surface.
To begin with, we introduce the following two-dimensional analogs to the
notations used in Section 1.8. Let be an open subset of R3 . The notation b
Sect. 2.10]
93
where stands for the standard multi-index notation for partial derivatives and
|| denotes the Euclidean norm in the latter definition. If A C ` (; M3 ), ` 0,
and b , we likewise let
kAk`, = nsup | A(y)|,
y
||`
n o
1
2
in
|1 2 |
n n n o
1
2
in , n 0,
|1 n 2 n |
such that
lim kn k3, = 0 for all b .
g := a 2x3 b + x23 c
n
g
:= an 2x3 bn + x23 cn
and gi3 := i3 ,
n
and gi3
:= i3 , n 0
94
[Ch. 2
c := b b , b := a b , (a ) := (a )1 ,
n
,n n
:= b,n
b,n
b , (a,n ) := (an )1 , n 0.
b ,
:= a
Sect. 2.10]
95
where
ijq :=
1
(j giq +i gjq q gij ) and pij := g pq ijq , with (g pq ) := (gij )1 ,
2
n
and let the functions Rqijk
C 0 (), n 0 be similarly defined from the matrix
n
2
3
fields (gij ) C (; S> ), n 0. Then
n
Rqijk = 0 in and Rqijk
= 0 in for all n 0.
it follows that
lim kCnp Cp k2,` = 0, ` k , p = 0, 1, 2,
where the matrices Cp and Cnp , n 0, p = 0, 1, 2, are those defined in the proof
of part (i). The definition of the norm kk2,` then implies that
lim kCn Ck2,` = 0, ` K .
n o
1
2
in ,
|1 2 |
1 2
, and let
|1 2 |
gij := i j .
96
[Ch. 2
n n n o
1
2
in .
n
|1 2 n |
Dropping the exponent n for notational convenience in this part of the proof,
let g i := i . Then 33 = 3 g 3 = p33 g p = 0, since it is easily verified that
the functions p33 , constructed from the functions gij as indicated in part (ii),
vanish in . Hence there exists a mapping 1 C 3 (; E3 ) such that
(y, x3 ) = (y) + x3 1 (y) for all (y, x3 ) .
Consequently, g = + x3 1 and g 3 = 1 . The relations gi3 = g i g 3 = i3
then show that
( + x3 1 ) 1 = 0 and 1 1 = 1.
These relations imply that 1 = 0. Hence either 1 = a3 or 1 = a3
in . But 1 = a3 is ruled out since we must have
{1 2 } 1 = det(gij )|x3 =0 > 0.
Noting that
a3 = 0 implies a3 = a3 ,
we obtain, on the one hand,
g = ( + x3 a3 ) ( + x3 a3 )
= 2x3 a3 + x23 a3 a3 in .
Sect. 2.10]
97
en C 3 (; E3 ) of the form
Then there exist immersions
en = cn + Qn n , cn E3 , Qn O3
98
[Ch. 2
en () and n ()
(hence the first and second fundamental forms of the surfaces
are the same for all n 0) and an immersion C 3 (, E3 ) such that a and
b are the covariant components of the first and second fundamental forms of
the surface (). Besides,
n
Proof. An argument similar to that used in the proof of Theorem 1.8-4 shows
that passing to the limit as n is allowed in the Gau and Codazzi-Mainardi
equations, which are satisfied in the spaces C 0 () and C 1 () respectively by the
functions an and bn for each n 0 (as necessary conditions; cf. Theorem
2.7-1). Hence the limit functions a and b also satisfy the Gau and CodazziMainardi equations.
By the fundamental existence theorem (Theorem 2.8-1), there thus exists an
immersion C 3 (; E3 ) such that
1 2
a = and b =
.
|1 2 |
Theorem 2.10-1 can now be applied, showing that there exist mappings (now
en C 3 (; E3 ) such that
denoted)
an
and
e n
en and bn =
en
=
e n 2
en
1
in , n 0,
e n 2
en |
|1
Finally, the rigidity theorem for surfaces (Theorem 2.9-1) shows that, for
each n 0, there exist cn E3 and Qn O3+ such that
en = cn + Qn n in ,
Sect. 2.10]
99
i .
i=0
Then
lim kn k`, = 0 for all b lim d` ( n , ) = 0,
where
d` (, ) :=
X
1 k k`,i
.
i 1 + k k
2
`,i
i=0
cE
QO3
C C + C
C C
C = b b b b in ,
b b + C
b C
b = 0 in }.
n o
1
2
in .
|1 2 |
100
[Ch. 2
Proof. Since {C02 (; S2> S); d2 } and {C 3 (; E3 ); d3 } are both metric spaces,
it suffices to show that convergent sequences are mapped through F into convergent sequences.
Let then ((a ), (b )) C02 (; S2> S2 ) and ((an ), (bn )) C02 (; S2>
2
S ), n 0, be such that
lim d2 (((an ), (bn )), ((a ), (b ))) = 0,
Let there be given any F (((a ), (b ))). Then Theorem 2.10-1 shows
that there exist n F (((an ), (bn ))), n 0, such that
lim kn k3, = 0 for all b ,
lim d3 (n , ) = 0.
Consequently,
lim d3 (F (((an ), (bn ))), F (((a ), (b )))) = 0,
The above continuity results have been extended up to the boundary of the
set by Ciarlet & C. Mardare [2005b].
References
101
102
References
References
103