This action might not be possible to undo. Are you sure you want to continue?
0iffe re nt ia G e o met ryD h
Theory and App ications
Philippt G C rlet TaTsien Li
A
A
1
0iffe re nt ia I G e o 1e t ry :
Theory and Applications
Series in Contemporary Applied Mathematics CAM
Honorary Editor: ChaoHao Gu (Fudun University) Editors: P. G. Ciarlet (City University ofHong Kong), TaTsien Li (Fudan University)
1. Mathematical Finance Theory and Practice (Eds. Yong Jiongmin, Rama Cont) 2. New Advances in Computational Fluid Dynamics Theory, Methods and Applications (Eds. F. Dubois, Wu Huamo) Actuarial Science Theory and Practice 3. (Eds. Hanji Shang, Alain Tosseti) 4. Mathematical Problems in Environmental Science and Engineering (Eds. Alexandre Ern, Liu Weiping) 5. GinzburgLandau Vortices (Eds. HaYmBrezis, TaTsien Li) 6. Frontiers and Prospects of Contemporary Applied Mathmetics (Eds. TaTsien Li, Pingwen Zhang) 7. Mathematical Methods for Surface and Subsurface Hydrosystems (Eds. Deguan Wang, Christian Duquennoi, Alexandre Ern) 8. Some Topics in Industrial and Applied Mathematics (Eds. Rolf Jeltsch, TaTsien Li, Ian Hugh Sloan) 9. Differential Geometry: Theory and Applications (Eds. Philippe G Ciarlet, TaTsien Li) .
Series in Contemporary Applied Mathematics CAM 9
0iffe Ie n t ia I G e o m e t ry :
Theory and Applications
editors
Philippe G Ciarlet
City University of Hong Kong, China
TaTsien Li
Fudan University, China
Higher Education Press
xe World Scientific
NEW J E R S E Y
*
LONDON
SINGAPORE
BElJlNG
*
SHANGHAI
*
HONG KONG
 TAIPEI
CHENNAI
Philippe G. Ciarlet
Department of Mathematics City University of Hong Kong 83 Tat Chee Avenue Kowloon, Hong Kong China
TaTsien Li
School of Mathematical Sciences Fudan University 220, Handan Road Shanghai, 200433 China
Editorial Assistants: Zhou ChunLian
Copyright @ 2008 by
Higher Education Press 4 Dewai Dajie, Beijing 100011, P. R. China, and World Scientific Publishing Co Pte Ltd
5 Toh Tuch Link, Singapore 596224
All rights reserved. No part of this book may be reproduced or transmittkd in any form
or by any means, electronic or mechanical, including photocopying, recording or by
any information storage and retrieval system, without permission in writing from the Publisher
ISBN 9787040222838
Printed in P. R. China
Preface
The ISFMACIMPA School on “Differential Geometry: Theory and Applications” was held on 07 August 18 August 2006, in the building of the ChineseFrench Institute for Applied Mathematics (ISFMA), Fudan University, Shanghai, China. This school was jointly organized by the ISFMA and the CIMPA (International Centre for Pure and Applied Mathematics), Nice, France. About sixty participants from China, Hong Kong, France, Cambodia, India, Iran, Pakistan, Philippines, Romania, Russia, SriLanka, Thailand, Turkey, Uzbekistan and Vietnam attended this highly successful event. The first objective of this school was to lay down in a selfcontained and accessible manner the basic notions of differential geometry, such as the metric tensor, the Riemann curvature tensor, the fundamental forms of a surface, covariant derivatives, and the fundamental theorem of surface theory etc. Although this field is with good reasons often considered as a “classical” one, it has been recently “rejuvenated”, thanks to the manifold applications where it plays an essential role. The second objective of this school was to present some of these applications, such as the theory of linearly and nonlinearly elastic shells, the implementation of numerical methods for shells, and mesh generation in finite element methods. To fulfill these objectives, four series of lectures, each series comprising ten 50minlectures, were delivered under the following titles: “Introduction to differential geometry” , “Introduction to shell theory” , “A differential geometry approach to mesh generation”, and “Numerical methods for shells”. This volume gathers the materials covered in these lectures. As such, this volume should be very useful to graduate students and researchers in pure and applied mathematics. The organizers take pleasure in thanking the various organizations for their generous support: The ISFMA, the CIMPA, the French Embassy in Beijing, the Consulate General of France in Shanghai, the National Natural Science Foundation of China, Fudan University, Higher Education Press and World Scientific. Finally, our special thanks are due to Mrs. Zhou ChunLian for her patient and effective work in editing this book.
~
Philippe G. Ciarlet and TaTsien Li February 2007
This page intentionally left blank .
.............. Cristinel Mardare: An Introduction to Shell Theory . .... ........ 222 ...... Ciarlet: An Introduction to Differential Geometry in R3 ........ 1 94 Dominique Chapelle: Some New Results and Current Challenges in the Finite Element Analysis of Shells ............ ........ .. 185 Pascal Frey: A Differential Geometry Approach to Mesh Generation ... Philzppe G... ................... ... ................................ Ciarlet. ..... ... ...................Contents Preface Philippe G....
This page intentionally left blank .
Kowloon. We also prove the corresponding uniqueness theorem. In Part 1. such as their two fundamental forms. the Gaussian curvature. Ciarlet Department of Mathematics. integration theory and Sobolev spaces. and some familiarity with ordinary and partial differential equations. these notes are extracted and adapted from my book “An Introduction to Differential Geometry with Applications to Elasticity”. Springer. and covariant derivatives. . arising when a threedimensional open set is equipped with curvilinear coordinates. In Part 2. ‘With the kind permission of SpringerVerlag. China [Project No.edu. 9040869. We then prove that the vanishing of the Riemann curvature tensor is sufficient for the existence of isometric immersions from a simplyconnected open subset of E3 equipped with a Riemannian metric into a threedimensional Euclidean space. CityU 100803 and Project No. with a special emphasis on those used in applications. Hong Kong. The prerequisites essentially consist in a working knowledge of basic notions of analysis and functional analysis.1 An Introduction to Differential Geometry in R3 Philippe G. which asserts that the Gauss and CodazziMainardi equations constitute sufficient conditions for two matrix fields defined in a simplyconnected open subset of R2 to be the two fundamental forms of a surface in a threedimensional Euclidean space. China Email: mapgc@cityu. We then prove the fundamental theorem of surface theory. CityU 100604]. the writing of which was substantially supported by two grants from the Research Grants Council of Hong Kong Special Administrative Region. also called rigidity theorem. The treatment is essentially selfcontained and proofs are complete. we review the basic notions. we study basic notions about surfaces.hk Introduction These notes’ are intended to give a thorough introduction to the basic theorems of differential geometry in R3. We also prove the corresponding rigidity theorem. such as differential calculus. 9040966. Dordrecht. such as the metric tensor and covariant derivatives. City University of Hong Kong 83 Tat Chee Avenue. 2005.
3) basic definitions and properties arising when the threedimensional open subset O(R) of E3 is equipped with the coordinates of the points of R as its curvilinear coordinates.gi . e.j . We begin by reviewing (Sections 1.. where (gpq) = (gij)'. Of fundamental importance is the metric tensor of the set O(R). g j and gt3 = g' . . . and let 0 : R 4 E3 be a smooth injective immersion. where g i = 8iO and gj . and lengths.2 Philippe G.4 the fundamental notion of covariant derivatives vilij of a vector field wigi : R 4 R3 defined by means of its covariant components wi over the contravariant bases gi. bases in the set O(R). whose covariant and contravariant components gij = gji : R + R and .31). by means of the functions gij and gij (Theorem 1. q E {1.3}): j 923 '.51). In particular. More specifically (Theorem 1. is expressed in terms of curvilinear coordinates. The vector fields g i : R + R3 and gj : R + R3 respectively form the covariant. necessarily. in the set O(R) are computed in terms of its curvilinear coordinates.2. and l7$ = g p q r i j q .. k . covariant derivatives naturally appear when a system of partial differential equations with a vector field as the unknown. its components and some of their partial derivatives must satisfy necessary conditions that take the form of the following relations (meant to hold for all i . It is shown in particular how volumes. the displacement field in elasticity.3}): Let the functions rijq and be defined by 1 rijq z(ajgiq+digjqdqgij) = Then. Ciarlet 1 Threedimensional differential geometry Outline Let R be an open subset of R3. g 3 . " . g i = hi. 2. We next introduce in Section 1. It is a basic fact that the symmetric and positivedefinite matrix field ( g i j ) defined on R in this fashion cannot be arbitrary. Covariant derivatives constitute a generalization of the usual partial derivatives of vector fields defined by means of their Cartesian components. and contravariant. E3 denote a threedimensional Eulet clidean space. 9'3 = gja : R + R are given by (Latin indices or exponents take their values in {I. .. areas.1 to 1.g.
aj0 there then exist a vector c E E3 and an orthogonal matrix Q of order three such that ~ ( x=) c + Q ~ ( X )for all 5 E R. i f R is connected.1 Curvilinear coordinates To begin with.  i . considered here are real. We conclude this chapter by indicating in Section 1. when they are used for indexing sequences.61 and 1. i.2. e. we list some notations and conventions that will be consistently used throughout.3}. to what extent is it unique? As shown in Theorems 1.71. defined in this fashion modulo isometries of E3. matrices. ~ in R. and second.a .g. All spaces.8 that the equivalence class of 0 . This means that. and the summation convention with respect to repeated indices or exponents . Latin indices and exponents range in the set {1. the necessary conditions R q i j k = 0 in R are also s u f i c i e n t for the existence of such a n immersion 0 . when does there exist a n immersion 0 : R 4 E3 such that g i j = &O ' a j 0 in R? If such an immersion exists. etc. this immersion i s unique up t o isometries of E3. 1. .. if 0 : R + E3 is any other smooth immersion satisfying g $3. kind and the functions are the covariant components of the R i e m a n n curvature tensor of the set 0 ( R ) .An Introduction to Differential Geometry in R3 3 The functions rij. the above existence and uniqueness theorems constitute an important special case of the fundamental theorem of Riemannian geometry and as such. Together. constitute the core of Part 1.. and are the Christoffel symbols of the first.. We then focus our attention on the reciprocal questions: Given an open subset R of R3 and a smooth enough symmetric and positivedefinite matrix field ( g i j ) defined on R.e. when is it the metric tensor field of an open set @(R) c E3. depends continuously o n the matrix field ( g i j ) with respect t o various topologies. especially the first one!): Under the assumption that R i s simplyconnected. the answers turn out t o be remarkably simple to state (but not so simple to prove. Besides. save when otherwise indicated.
The space a E3 is endowed with an orthonormal basis consisting of three vectors i?= Zi. Ciarlet is systematically used in conjunction with this rule.2.Let xi denote the coordinates of a point x E EX3 and let 13i := a / d x i . or 6aj according to the context. they form the covariant basis at 9= O(z) and they are tangent to the coordinate lines passing through 9 . and let l1 = f i denote the Euclidean norm of a E E3.3. This space will be identified with i R3. and aijk:= a 3 / d ~ i d ~ j a ~ k . the relation Si(X) = gij ( X ) d (XI means that 3 gi(x)= Cgij(x)gi(z) for i j=1 = 1. := d2/axidxj. For instance. Kronecker’s symbols are designated by dl. . let there be given a threedimensional vector space in which three vectors ea = e form a basis. dij. let a. i In addition. Let there be given an open subset 6 of E3 and assume that there exist an open subset R of R3 and an injective mapping 0 : R + E3 such h aij Figure 1. Let 2i denote the Cartesian coordinates of a point 2 E E3 and l let a := a/a2ii. Let E3 denote a threedimensional Euclidean space.x2.11: Curvilinear coordinates and covariant bases in a n open set 8 c E3. If the three vectors gi(z) = &O(z) are linearly independent. The thr? coordinates z1.4 Philippe G.z3 of x E R are the curvilinear coordinates of P = O ( x ) E R.b E E3.b and a A b denote the Euclidean inner product and exterior product of a.
Schwartz [1992] or Zeidler [1986. p.p . there are infinitely m a n y ways of defining curvilinear coordinates in a given open set R . If 0 E C1(R. Note that ( ( p + Z k ~$+ . but curvilinear coordinates may be defined only locally in this case: Given z E R. . p. Section 16. r ) or ('p 2 h .. e. $ . Section 16.rsin$) E E3.. E E3. Note that (p p.rcos$sinp.2 E R.$. z ) . Nirenberg [1974. h b z E3 Figure 1. an open subset R of B3 together with a mapping 0 : R 4 E3 are instead_ a priori given.g. the set fi is again open (for a proof. z ) are the cylindrical coordinates of P = 0(cp. Let the mapping 0 be defined by 0 : ( c p . E3) and the three vectors &O(z) are linearly independent at all 2 E R. + + + + In a different.Naturally. z ) or (p T + 2 k ~. If 0 E Co(R.11). Then each point 2 E fi can be unambiguously written as 2 = O ( 2 ) .T ) are also spherical coordinates of the same point P and that cp and $ are not defined if 2 is the origin of E3. p. e. Then (9. approach. ~ ) E C l i (rcos$cosp. z) Then (p. also cylindrical coordinates of the same . p sin 9. all that can be asserted (by the local inversion theorem) is the existence of an open neighborhood h .4]). h and the three coordinates z of 2 are called the curvilinear coordii nates of 2 (Figure 1. $ K + 2!~. depending on how the open set R and the mapping 0 are chosen! Examples of curvilinear coordinates include the wellknown cylindrical and spherical coordinates (Figure 1. T ) . but equally important.12). T ) are the spherical coordinates of P = 0 ( p . 171 or Zeidler [1986. 2!7r. Let the mapping 0 be defined by 0 : (cp. 2kn.12: Two familiar examples of curvilinear coordinates.E3) and 0 is injective. p.4]). z ) E R 4( p cos 'p.$. see. Corollary 2. see. z ) . are point 2 and that cp is not defined if Z is the origin of E3.g. k E Z.An Introduction to Differential Geometry in lit3 5 that O(R) = 6. and curvilinear coordinates inside R are unambiguously defined in this case. the set R := O(R) is open by the invariance of domain theorem (for a proof.
In this case. Then the three vectors gi(x) constitute the covariant basis at the point 2 = O ( x ) .. 6x3 8303 Let the three vectors gi(x) E R3 be defined by i. of V onto O ( V ) . hence an injection. Then the expansion of 0 about x may be also written as O(x + Sx) = O ( x )+ 6xigi(. Ciarlet V of x in R such that the restriction of 0 to V is a C1diffeomorphism.6 Philippe G. A mapping 0 : R + E3 is an immersion at x E R if it is differentiable at x and the matrix V O ( x ) is invertible or. if the three vectors gi(x) = &O(x) are linearly independent.) + o(6x) R and ei is If in particular Sx is of the form SX = Gtei. 1. equivalently. If (x Sx) E R.2 Let Metric tensor R be an open subset of R3 and let be a mapping that is diflerentiable at a point x E R.tl[ + fi(t) := O ( x tei) in a + . where bt E this relation reduces to one of the basis vectors in R3.then + SX is such that O(z + Sx) = O ( x ) VO(x)bx where the 3 x 3 matrix &I @ 8102 + + o(Sx). the last relation thus shows that each vector gi(x) i s tangent to the ath coordinate line passing through 2 = O ( x ) . Assume that the mapping 0 i s an immersion at x. gi(x) i s the ith column vector of the matrix VO(x).e.defined as the image by 0 of the points of R that lie on the line parallel to ei passing through x (there exist to and t l with to < 0 < t l such that the ith coordinate line is given by t E ]to. V O ( x ) and the column vector Sx are defined by &I @ 8202 8203 8 0 3 1 8302) (x) and Sx = 8103 f::) . O ( x + 6tei) = O ( x ) 6tgi(x) + + o(6t).
by letting gij(x) := gz(x) . The three vectors gi(x)being linearly independent. the nine relations g2(x). To see this.t h column of a matrix A is noted (A)ij.@(%)I2 = 6 x T V O ( x ) T V O ( x ) 6 x o(16z12) = 6xig&) + 'gj(x)6xj o(16s12). ~ the element at the ith row and j .12.see Figures 1. sw = (gik(z)g&)) . + Note that.where = (gij(x)):= (gij(2))y. The three vectors gi(x)form the contravariant basis a t the point f = O ( x ) and the elements gij(x) of the symmetric positive definite . since the vectors gi(x) are assumed to be linearly independent.gj(x)= 6 . For instance. unambiguously define three linearly independent vectors gi(x). These relations in turn imply that g i ( 4. In other words.. = gz"x)gje(x)gke(x) and thus the vectors gi(x) are linearly independent since the matrix (gij(z)) positive definite. 6xT stands for the transpose of the column vector 6x and V O ( Z )designates the transpose of the matrix VO(x). g&) =6 .An Introduction to Differential Geometry in R3 7 neighborhood of 2. This observation suggests to define a matrix (gij(x))of order three.11 and 1. in . here and subsequentIy. etc. hence fl(0) = & O ( x ) = gi(x)). (gje(4ge(4) = gZk(2)6% = g"(x). Note that the matrix VO(x) is invertible and that the matrix (gij(z)) is positive definite. Xik(x)= g i k ( x ) . Returning to a general increment 6x = 6xiei. gj(x) = (vo(x)Tvo(x))ij. consequently. let a priori gi(x)= Xik(z)gk(x) the relations gi(z).gj(x)Sxj}1/2. we also infer from the expansion of 0 about x that (recall that we use the summation convention) : p(x + 62) . The elements g i j ( x ) of this symmetric matrix are called the covariant components of the metric tensor at 2 = O ( x ) . This gives Xik(x)gkj(x) 6. We would likewise establish that gi(x) = is gij (4Sj (x). the principal part with respect to 6 x of the length between the points O ( x + 6 x ) and O ( x )is {Gxigi(x). Hence gi(x)= gik(x)gk(x). we use standard notations from matrix algebra.
g j ( x ) . we briefly explain in what sense the components of the “metric tensor” may be “covariant” or “contrawariant”.8 Philippe G. g j ( x ) and gZj(x) = g i ( x ) . g i ( x ) = g i j ( x ) g j ( x ) and g i ( x ) = g i j ( x ) g j ( x ) . If 0 : R + E3 is an immersion. 0). A mapping 0 : R + E3 is an immersion if it is an immersion at each point in R. . of the metric tensor at the same point O(z) = 6(Z) E E3. What is exactly the “secondorder tensor” hidden behind its covariant components g i j ( x ) or its contravariant exponents g i j ( z ) . Let g i j ( x ) and Tij ( 2 )be the covariant components. A simple computation then shows that where = 0 o 0 E C1(R. To conclude this section. Ciarlet matrix ( g i j ( x ) ) are the contravariant components of the metric tensor at 2 = O ( x ) . the vector fields gi : R + R3 and gi : R + R3 respectively form the covariant. i. while each exponent in g ij( x ) (‘varies like” that of the corresponding vector of the contravariant basis. Let g i ( x ) and $(Z) be the vectors of the corresponding contravariant bases at the same point 2. Let us record for convenience the fundamental relations that exist between the vectors of the covariant and contravariant bases and the covariant and contravariant components of the metric tensor at a point x E R where the mapping 0 is an immersion: g i j ( x ) = g i ( x ) . if 0 is differentiable in R and the three vectors g i ( x ) = & O ( x ) are linearly independent at each x E R.Let R and 6 be two domains in R3 and let 0 : R + E3 and 6 : R + E3 be two C’diffeomorphisms such that O(R) = 6(6)and such that the vectors g i ( x ) := & O ( x ) and Gi(Z) = of the covariant bases at the same point O(z) = 6(Z) E E3 are linearly independent.e.. and contravariant bases. Remark. Then a simple computation shows that = ( x j ) := x 1 These formulas explain why the components gij(x) and $ j ( z ) are respectively called “covariant” and ‘kontravariant”: Each index in g i j ( x ) “varies like” that of the corresponding vector of the covariant basis under a change of curmilinear coordinates. and let gzj ( x ) and ?j(Z) be the contravariant components.6) (hence Z = ~ ( z ) and ) (T):= x1 E C l ( 6 .
and the contravariant components d ( x ) and Gi(x) (both with selfexplanatory notations). areas. the components w{(x) “vary like” the vectors gi(z) of the covariant basis under a change of curvilinear coordinates. Indeed. and length elements at a point 2 = O ( x ) in the set R = O(R) can be expressed either in terms of the matrix VO(x). We leave it instead to the reader to verify in each instance that any index or exponent appearing in a component of a “tensor” indeed behaves according to its nature. that we shall no longer provide such commentaries in the sequel. it is also shown in ibid. in Boothby [1975]. or in terms of the matrix (Sij(X))..)iji(. the covariant components . Chapter 11. analogous justifications apply as well to the components of all the other “tensors” that will be introduced later on.3 Volumes.). for instance..= 3(. the set D being locally on .Marsden & Hughes [1983. particularly as regards its relevance to elasticity. that the same “tensor” also has “mixed” components gj(x). 0 In fact. is a bounded. however. and connected subset D d of Wdwith a Lipschitzcontinuous boundary. The reader interested by such questions will find exhaustive treatments of tensor analysis. Section 1. . 2 2. 1. Note. which turn out to be simply the Kronecker symbols 6. This is why they are respectively called %ovariant” and rrcontrauariant”. {) ( and Gi(x). These formulas thus highlight the crucial r61e played by the matrix ( g i j ( x ) ) for computing “metric” notions at the point 2 = O(z). and lengths in curvilinear coordinates We now review fundamental formulas showing hzw volume. open. Chapter 11. or Simmonds [1994]. ) In other words. g It is then easily verified that = ? J ~ ( x ) g i ( . Thus. area. Sections 1 to 31. while the components d ( x ) of a vector “vary like” the vectors g i ( x ) of the contravariant basis.4) ?Ji(x)gZ(z) G)) = z( (. of a vector at the same point O ( x )= 6 ( E ) satisfy (cf. A vector is an example of a “firstorder” tensor.An Introduction to Differential Geometry in W3 9 is beautifully explained in the gentle introduction to tensors given by Antman [1995. the “metric tensor” well deserves its name! A domain in Wd.
c R.10 Philippe G. The area element (b) Let D be a domain in R3 such that d f ( f ) at 3 = O(x) E 8 6 is given in terms of the area element dF(x) at x E d D by d f ( f ) = I C o f V O ( x ) n ( x ) l d r ( x )= m d n i ( x ) g i j ( x ) n j ( a : ) d I ’ ( x ) . n ( x ) = n. The second relation in (a) follows from the relation g(x) = I d e t V O ( x ) I 2 . d r denote the area element along r. All relevant details needed here about domains are found in NeEas [1967] or Adams [1975].g. where n ( x ) := ni(x)ei denotes the unit outer normal vector at x E d D . 112 . Given a domain D c R3 with boundary r.28 and Example 1. (c) The length element d?(f) at f = O(x) E fi is given by dT(f) = { S X ~ V O ( ~ ) ~ V O ( X ) =~ } ~ ” S {6xigij(x)6xJ} . { O ( D ) } . Ciarlet one side of its boundary. Let R be an open subset of R3. which itself follows from the relation (gz3(x))= v o ( x ) T v o ( z ) . The relation d2 = ) d e t V O ( x ) l dx between the volume elements is well known. where g(x) := det(gij(x)).71).(x) as its components). If A is a square matrix. Indications about the proof of the relation between the area elements d?(f) and d r ( x ) given in (b) are found in Ciarlet [1988. Theorem 1. C o f A denotes the cofactor matria: of A .= O(D). Using the relations C o f ( A T ) = .711 (in this formula. then {6}c R. Thus C o f A = (det A)AT if A is invertible.(x)e2 is identified with the column vector in R3 with n.and the boundaries 66 of and dD of ’ D are related by 8 6 = O ( a D ) (see.31. if D is a domain in R3 such that D C R. (a) The volume element d f at f = O(x) E R is given in terms of the volume element dx at x E R by  df = I det V O ( x ) l d z = m d x . and n = nii? denote the unit (In1 = 1) outer normal vector along I? ( d r is well defined and n is defined dralmost everywhere since r is assumed to be Lipschitzcontinuous) .let 0 : R 4 E3 be an injective and smooth enough immersion. Proof. Note also that the assumptions made on the mapping 0 in the next theorem guarantee that. where Sx = 6xiei. e. Ciarlet [1988. 5 Theorem 1. we let dx denote the volume element in D .. Theorem 1. and let R = @(a).
i. and lengths inside R by means of integrals insid: R .An Introduction to Differential Geometry in EX3 11 (CofA)T and Cof(AB) = (CofA)(CofB).31: Volume. and length elements an curvilinear coordinates.31 F e used in particular for computing volumes. the area of A = @ ( A ) c 8 2 . areas. . Then.dF(Z). Assume that R is a domain and that 0 is a C1diffeomorphism of R onto {a}. {fi}. cf. Theorem 1. given a domain V such that V C R and a drmeasurablesubset A of the corresponding relatio_ns are used for computing the volume of V = O(V) c R. let 5 := O ( D ) . and 6x at z E R by means of the covariant and contravariant components of the metric tensor. and d@Z) at P = O(z) E are expressed in terms of dz. area.2.31. 0 + The relations found in Theorem 1.we next have: I CofVo(x)n(x)/2 n ( x ) T Cof ( v o ( x ) T v o ( x ) ) n ( x ) = = g (x)i (x)gij( x ) n j(x) n . Then ~ 0 \t I R Figure 1.31): Let D be a domain in R3 such that 25 c 0. Either expression of the length element given in (c) recalls that dF(5) is by definition the principal part with respect to 6x = Gxaei of the length l@(x 62) . The elements dZ.O ( x ) / . in terms of the curvilinear coordinates used in the open set R (Figure 1. d f ( z ) . whose expression precisely led to the introduction of the matrix ( g i j ( x ) )in Section 1..and let f ^ L1(6) be given.e. fi 20. and the length of a curve (? = O ( C )c where C = f ( I ) and I is a compact interval of R.
12 Philippe G.. the volume of D is given by h volD := d2 = m d x . 1. dF(2) = m d w d l ? ( x ) . and let h E L1(C) be given.4 Covariant derivatives of a vector field Suppose that a vector field is defined in an open subset 6 of E3 by means of its Cartesian component_s Ci : 6 + R. Finally. Suppose now that the open set R is equipped with curvilinear coordinates from an open subset R of EX3. by means of an injective mapping 0 : R + E3 satisfying O(R) = 6.(2)= L(x o O)(x)mddl?(x). see Figure 1. the area of C is given by h areaC := s. this field is defined by its values Gi(2)2? at each 2 E R. but this time in terms of these curvilinear coordinates? It turns out that the proper way to do so consists in defining three functions u : R + R i by requiring that (Figure 1. Ciarlet h In particular. h In particular. Next. Then the length of the curve 6 := O ( C )c R is given by This relation shows in particular that the lengths of curves inside the open set O(R) are precisely those induced by the Euclidean metric of the space E3. consider a curve C = f ( I ) in R.e. the set O(R) is said to be isometrically immersed in E3.42) h . where the vectors Zi constitute the orthonormal basis of E3. For this reason.41. h(2)df. where I is a compact interval of R and f = f iei : I + R is a smoothenough injective mapping. lct r := do_. i. let CJe a drmeasurable subset of O ( C ) c d D . Then r. let 5 := s. How does one define appropriate components of the same vector field.
we immediately find how the old and new components are related. At each point 3 E 6.42:A vector field in curvilinear coordinates.^)over an orthonormal basis of E3 formed by three vectors Zi. the same vector field is defined at each z E R by its covariant components vi(z) over the contravariant basis vectors gi(z) in such a way that vi(z)gi(z) = Gi(3)ei. = O(z). Using the relations gi(x). Figure 1. In curvilinear coordinates.2j = $.41).. The three components vi(x)are called the covariant components .41: A vector field in Cartesian coordinates. the vector G(@ i?' is defined by its Cartesian components Gi(. g j ( z ) = 6. Let there be given a vector field in Cartesian coordinates defined at each 3 E 6 by its Cartesian components &(Z) over the vectors 8" (Figure 1. and 2 .2). viz.An Introduction to Differential Geometry in R3 13 Figure 1. 3 where the three vectors gi(x)form the contravariant basis at 2= O ( x ) (Section 1.
carrying out such a transformation naturally leads to a fundamental notion. 2113 := 8 . i. As we now show. Z ~ . v 3 2 23 P and rp. 2 where 21. . := Let O ( x ) = Ok(x)Zk and 6(2) = @(2)ei. . ajGi(2) = (vklle[gkli[gelj)(x). let vigi : R + R3 be the same field in curvilinear coordinates. (i) Another expression of [gi(x)]k g i ( x ) .and the three functions vi : R + R defined in this fashion are called the covariant components of the vector field vigi : R + E3. V 6 ( 2 ) V O ( x ) = I. v . Suppose next that we wish t o compute a partial derivative ajGi(2) at a point 2 = O ( x ) E R in terms of the partial derivatives & v k ( z ) and of the values vq(x) (which are also expected to appear by virtue of the chain rule). 23 and [Si(2)]k g y x ) . 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. Ciarlet of the vector v i ( x ) g i ( x )at 2. Given a vector field Gigi: fi + R3 in Cartesian coordinates with components Gi E C'(fi). := gp .e. . that of covariant derivatives of a vector field. the chain rule shows that the matrices VO(x) := ( a j O k ( x ) )(the row index is k ) and ? 6 ( 2 ) := ( a k W ( 2 ) ) (the row index is i) satisfy  h .t h component of g i ( x ) over the basis { Z I . Let R be a n open subset of R3 and let 0 : R + E3 be a n injective immersion that is also a C2diffeomorphism of R onto 6 := O(R).41. Z k := denotes the i . Z k . Z ~ } . h h Theorem 1..r ? .x E R. The following convention holds throughout this proof The simultaneous appearance of 2 and x in an equality means that they are related by 2 = O ( x )and that the equality in question holds for all x E R. Proof.= O ( x ) . Since 6 ( O ( x ) ) x for all = x E R.14 Philippe G. T h e n vi E C() and for all x E 'R h R. a i g j . that defined by h vi(x)e h 2 = vi(x)gz(x) for all 2= O ( x ) .where 6 : 6 + E3 denotes the inverse mapping of 0 : R + E3.
= (ii) The functions := g q h A . gk(x) = aegq(x). EL3) are given at each 2 = 0 ( x ) E 6 b y 5. These derivatives will be needed in (iii) for expressing the derivatives &Ci(2) as functions of LC (recall that Ci(2)= uk(z)[g'(x)]i). = ) h h Hence. the last of the Cartesian components of (iii) The partial derivatives the vector field Giza E C'(6.m(x)6r = r j m ( z ) g m ( x .23} can be also expressed in terms of the inverse mapping 0.22. noting that &(gq(x) . we observe that h r.k(x) gq(x). (x). We next compute the derivatives &gq(x) (the fields gq = gq'g' are of class C1 on R since 0 is assumed to be of class C2).kE Co(0). 3 . we may write a priori aesq(x) .k(x) r. gk(x))= 0 and [gq(x)lp aP@(2). E Co(R). 15 The components of the above column vector being precisely those of the vector g. gk(x).gj(x)= 6 j . = aegk(x) 5 p S q ( ~ ) a e k o p (rig(+ = = ~) Since 0 E C2(R. &g.Gi(2) = u k l l e ( 4 [&)li [ge(41j. Recalling that the vectors gk(x) form a basis.r & ( 4 g k ( 4 . . Hence the Icth component of gi(x)over the basis {21.Iw3) by assumption.E3) and 6% E relations show that r. = : R + Iw. as: h [gi(x)]k ako"2). components of the above row vector must be those the of the vector gi(x)since gi(x)is uniquely defined for each exponent i by the three relations gi(x). 2 . To find their thereby unambiguously defining functions expressions in terms of the mappings 0 and 0.An Introduction to Differential Geometry in Iw3 or equivalently. h C1(6. . j = 1 . = we obtain r.
2 can be also defined by the relations a j ( v i g i ) = vzl. 23 . by the chain rule and by (i). The functions = g p . and let there be given a vector field vigi : fl + R3 with covariant components v i E C'(fl).which are defined b y villj := ajvi .42. zgj a. = I?jqgi. since & g q ( z ) = .rP. where I'P.jgi * VZllj = {aj(wkg"} ' g i .v.)(ae[9q(z)li)[ge(z)lj = (aevk(z). h We compute the partial derivatives ajGi(2) as functions of z by means . 7 .% ( Z ) % ( 4 ) + [gk(~)li[ge(z)lj.1 ' & ( z ) g k ( z ) by (ii). . 3. Theorem 1.. = a. Ciarlet and [ g k ( z ) ] iand r&(z) are defined as in (i) and (ii).16 Philippe G.= gp . 5jGi (2)= 5j V k (6 2 ) [gk( z )i]+ vq (z)5j [gq(6 ( ) I)$( i = aevk(z)[ge(z)]j[Sk(")li v. (b) The Christoffel symbols I. Let the assumptions o n the mapping 0 : R + E3 be as in Theorem 1.gj and a j g . In particular then.(. The following result summarizes properties of covariant derivatives and Christoffel symbols that are constantly used. . 0 The functions v. we first note that a differentiable function w : R + R satisfies & w ( 6 ( 2 ) )= aew(z)@e(2) = aew(z)[ge(z)]j. := g p .41 are called the firstorder covariant derivatives of the vector field vigi : R + R3. d i g j = I?yi E Co(R) satisfy the ?' relations &gP = rP.v 4 3 2 23 P defined in Theorem 1.41. of the relation Zi(2) = v k ( z ) [ g k ( z ) ] iTo this end.I?$vp. & g j . are called the Christoffel symbols of the second kind (the Christoffel symbols of the first kind are introduced in the next section). (a) The firstorder cowariant derivatives villj E Co(R) of the vector field vigi : R + R3.
r : j g j (cf.41 by Villj = ajvi qjVp. It remains to verify that the covariant derivatives villj. ajg.  may be equivalently defined by the relations aj(Vig2) = vzlljgi These relations unambiguously define the functions vUillj { a j ( V k g ' ) } ' g i = since the vectors gz are linearly independent at all points of R by assumption. a j g . defined in Theorem 1. To this end. As shown in the next theorem. the Riemann curvature tensor It is remarkable that the components gij = g j i : R 4 R of the metric tensor of a n open set O(R) c E3 (Section 1.j + g p . = r:jgp.. we = 21' ZIlJSi. defined by a smooth enough immersion 0 : R + E3.5 Necessary conditions satisfied by the metric tensor. = r. g P 3% + g p .gi.An Introduction to Differential Geometry in R3 17 Proof.41). = m. r. cannot be arbitrary functions. s P ) g p Remark.dkrijq + I'plCqp = o in 0. d j g . 1.Virjkg' = r i q g i . the relation 8 j ( v i g i ) ( z )= (viiljgi)(z)reduces to aj(Gi(2)z)= (8jGi(2))Zi. . they must satisfy relations that take the form: d j r i k q .2). a covariant derivative of the first order constiIn tutes a generalization of a partial derivative of the first order in Cartesian h coordinates. by definition. ' To establish the other relations a j g . part (ii) of the proof of Theorem 1.this sense. hence aj(vigi)= (ajVi)gZ + Vidjgi = ( a j V i ) g i . z E If the affine space E3 is identified with R3 and O(z) := z for all R. we simply note that. The Christoffel symbols can be also defined solely in terms of the components of the metric tensor. see the proof of Theorem 1.51. note that o= Hence ' 9 .krjqp  . the Christoffel symbols satisfy & g p = . )= rp.
l j ' g q . let 0 E C3(R. 7 Then. necessarily.) := 2 3% . Theorem 1. and let g i j := aio. Let R be a n open set in EX3. Since a i k g j = & j g l c . let the three vectors gj(II:) be defined by the relations g j ( z ) . g i ( z ) = 6. k .18 Philippe G. Since we also have g j = g z j g . Let gi = It is then immediately verified that the functions rij. are also given by a. + 8. dikgj ' r:'rkqP. djrikq .51. + = Proof. g q = 8 j r i k q . := g P T i j q where ( g p q ) := ( g i j ) . j g p since aigj = (&gj .a4Qz.51 coincide with the Christoffel symbols introduced in the previous section. gq = akrijq . gP)g. this explains why they are denoted by the same symbol).8. . :I aigi = r .9. . the last relations imply that '? = &gj . it so happens that the functions Ij as defined in Theorem ' : 1. Consequently. d rp. according to the rule governing Latin indices and exponents. ajo denote the covariant components of the metric tensor of the set @(R). r t. E C1(R)be defined by Let the functions rijq C1(R) and E 1 rijq (a. Note that.j . .krjqpo in 0.3}.r:jrkqp. Ciarlet where the functions rijqand 17!j have simple expressions in terms of the functions g i j and of some of their partial derivatives (as shown in the next proof. dkgql = so that the above relations together give a i g j . g p .. we also have aikgj .q .2. . q E {1.j .akrijq rzjrkqpr.@. d k g . Therefore. q d For each II: E R.g q + aigj . 3 . dkgq = r : j g p . these relations are meant to hold for all i .E3) be a n immersion.r:krjqp.. Differentiating the same relations yields &rijq= dikgj .' .
2. We saw in Section 1.k :=a. the necessary conditions R q i j k = 0 thus simply constitute a rewriting of the relations a i k g j = a k i g j in the form of the equivalent relations d i k g j . E where l is the row index and j the column index (equivalently.S3. 1.. = g p q r i j q 23 = digj . The relations R q i j k = 0 found in Theorem 1.41 thus express that the Riemann curvature tensor of the set 0 ( R ) (equipped with the metric tensor with covariant components g i j ) vanishes.6 Existence of an immersion defined on an open set in R3 with a prescribed metric tensor Let M3. 3% are the Christoffel symbols of the first. 19 0 Remark. kinds. The functions and rP. . So far. akr. g q . f r : j r k q p 3 tkq z3q .r. Finally. the matrix representing the F’rkhet derivative at 2 E R of a differentiable mapping 0 = : R + E3 is denoted (el) VO(Z) := (ajo((2)) M3. thus allowing us to define a matrix field c = ( g q ) = VOTVO : R + S3>. the functions R w . and S denote the sets of all square matrices of order three.rfkrjqp are the covariant components of the Riemann curvature tensor of the set 0 ( R ) . The vectors g i and g j introduced above form the covariant and contravariant bases and the functions g i j are the contravariant components of the metric tensor (Section 1. As in Section 1.2).An Introduction to Differential Geometry in R3 and thus the required necessary conditions immediately follow. . 0 As shown in the above proof.4 that the Christoffel symbols of the second kind also naturally appear in a different context (that of covariant differentiation). and second. we have considered that we are given an open set R c R3 and a smooth enough immersion 0 : R + E3. of all symmetric matrices of order three. VO(z) is the matrix of order three whose jth column vector is a j 0 ) .g p = rP. and of all symmetric positive definite matrices of order three. . g q = a k i g j .
We now turn t o the recaprocal questaons: Given an open subset R of R3 and a smooth enough matrix field C = (gt3) : R 4 S. In order t o put these results in a wider perspective. which accordingly should be resolved by different means. let us make a brief incursion into Riemannian Geometry. For detailed treatments. a Riemannian metric on a manifold is a twice covariant. when does there exzst a n zmmerszon 0 .3>. More specifically. More generally. on both the set R and the smoothness of the field ( g i j ) . Note that these two results are established under different assumptions. the necessary condataons found in Theorem 1. viz. de WittMorette & DillardBleick [1977].in the sense that there exists an immersion 0 : R + E3 f . Marsden & Hughes [1983]. in or equavalently.61) and a uniqueness result (Theorem 1. Ciarlet where gz3: R t I are the covariant components of the metrac tensor of K the open set @(R) c E3. a. Hulin & Lafontaine [2004]. This particular Riemannian manifold (R. (gij )) possesses the remarkable property that its metric is the same as that of the surrounding space E3. positivedefinite tensor field acting on vectors in the tangent spaces t o the manifold (these tangent spaces coincide with R3 in the present instance). a manifold. This result comprises two essentially distinct parts. n Whether the immersion found in this fashion is globally injective is a different issue.( g i j ) ) is isometrically immersed in the Euclidean space E3. a global existence result (Theorem 1.. symmetric. or Gallot.e.@ in R? If such an immersion exists.@ . (a. R 4 c E3 such that c = V O ~ V O R. an open set R c R3 equipped with an immersion 0 : R + E3 becomes an example of a Riemannian manifold (0. such that gz3 = a. d j @ in R. this immersion is unique u p t o isometries a E3.. see classic texts such as ChoquetBruhat. Considered as a threedimensional manifold. when is C the metric tensor field of an open set @(a) E3? Equivalently.. If R is connected. viz.71). equipped with a Riemannian metric.i.( g i j ) ) . Berger [2003].20 Philippe G. t o what extent is it unique? The answers are remarkably simple: If R as szmplyconnected. the symmetric positivedefinite matrix field ( g i j ) : R St defined in this case by gij := &@ . the set R.71 are also suficient for the existence of such a n immersion.
71 in the next section): Let R be a connected open subset of R3. T h e n a Riemannian manifold (0. so named after Nash [1954]: A n y pdimensional Riemannian man + . in the sense that it can be isometrically immersed in a Euclidean space of the same dimension (three)? The answer to this question can then be rephrased as follows (compare with the statement of Theorem 1.An Introduction to Differential Geometry in R3 21 that satisfies the relations gij = &O . When the pdimensional manifold is an open subset of E%P+q.ajO in R? The answer is yes. this result becomes a special case of the fundamental theorem on flat Riemannian manifolds.( g i j ) ) with a Riemannian metric ( g i j ) of class C2 in R is flat i f and only if its Riemannian curvature tensor vanishes in R.7..3>. The first question above can thus be rephrased as follows: Given a n open subset R of R3 and a positivedefinite matrix field ( g i j ) : R + s.8 and 2. constitutes another important special case). T h e n the isometric immersions of a flat Riemannian manifold (R. The answer to the second question. ( g i j ) ) is the same as the length of its image by 0 in the Euclidean space E3 (see Theorem 1.. a j 0 . i. ( g i j ) ) into a Euclidean space E3 are unique up to isometries of E3.3>. Section 1. Another fascinating question (which will not be addressed here) is the following: Given again an open subset R of EX3 equipped with a symmetric. an outline of a selfcontained proof is given in Szopos [2005].e. Recast as such. but this time in a higherdimensional Euclidean space? Equivalently. Equivalently. assume this time that the Riemannian manifold (0. this result likewise becomes a special case of the socalled rigidity theorem.61 below): Let R be a simplyconnected open subset of R3. viz.9. can be rephrased as follows (compare with the statement of Theorem 1. Can such a Riemannian manifold still be isometrically immersed. This theorem addresses the same existence and uniqueness questions in the more general setting where R is replaced by a pdimensional manifold and E3 is replaced by a (p q)dimensional Euclidean space (the “fundamental theorem of surface theory”. established in Sections 2. 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. positivedefinite matrix field ( g i j ) : R + S. the issue of uniqueness. ( g i j ) ) flat. does there exist a Euclidean space Ed with d > 3 and does there exist an immersion 0 : R + Ed such that gij = &O . according to the following beautiful Nash theorem. the length of any curve in the Riemannian manifold (R.( g i j ) ) is n o longer flat. Recast as such.31). its Riemannian curvature tensor no longer vanishes in R. cf. when is the Riemannian manifold (R. which holds for a general finitedimensional Riemannian manifold.
22
Philippe G. Ciarlet
ifold equipped with a continuous metric can be isometrically immersed in a Euclidean space of dimension 2p with a n immersion of class C1; it can also be isometrically immersed in a Euclidean space of dimension (2p 1) with a globally injective immersion of class C1. 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 1.61. Let R be a connected and simplyconnected open set in R3 and let C = ( g i j ) E C2(R; S be a matrix field that satisfies );
R q i j k := djrikq  a k r i j q
+ r;’rkqpr;krjqp 0 in 0, =

where
1 r 239. . (a.. + a&,, . 2 3Q2q
’
an%),
:= ( g i j )  l .
rp. := Y
p r i j qwith
($9)
T h e n there exists a n immersion 0 E C3(R;E3) such that
c = V O ~ V O R. in
Pro05 The proof relies on a simple, yet crucial, observation. When a smooth enough immersion 0 = (Oe) : R + E3 is a priori given (as it was so far), its components Oe satisfy the relations &jOe = ryjapOe7 which are nothing but another way of writing the relations &gj = ryjgp (see the proof of Theorem 1.51). This observation thus suggests to begin by solving (see part (ii)) the system of partial differential equations
&Fej = rP.Fep in R, 23
whose solutions Fej : R + R then constitute natural candidates for the partial derivatives ajoe of the unknown immersion 0 = (Be) : R 4 E3 (see part (iii)). To begin with, we establish in (i) relations that will in turn allow us t o rewrite the sufficient conditions
djrikq  a k r i j q
+ rfjrkqPrykrjqpo in =
in a slightly different form, more appropriate for the existence result of part (ii). Note that the positive definiteness of the symmetric matrices ( g i j ) is not needed for this purpose. (i) Let R be a n open subset 0fR3 and let there be given a field ( g i j ) E C2(R; S 3 ) of symmetric invertible matrices. The functions rijq,;, and I’ ? gpq being defined by
rijq:= p j g i q + aigjq  aqgij), rTj := g p q r 23q7 .
1
( S P 4 ) := (gijll,
An Introduction to Differential Geometry in IR3
23
define the functions
R q i j k :=
R&
Then
ajrikq akrijq r:jrkqp rTkrjqP, + := a j q k akr$ rfkrye rfjqe. +

Rp.. = gPqRqijk and '23k
Using the relations
rjqet rejq = ajgqe
R p i j k = gpqR:jk.
and
r i k q = gqtrfk,
which themselves follow from the definitions of the functions and noting that
rijqand
(gPqajgqe+ gqeajgPq)= aj (gPqgqe) 0 , =
we obtain
gPq(ajrikq  rLkrjqT) = ajrYk rikqajgpq  r tkgPq(ajg,e e
=

rejq)
a j q k t F f k r y Q rfk(gPqajgqet gqeajgPq) = a j q k + &rye.
Likewise,
gpq(dkrijq  rLjrkqT) = akrYj rfjrie, and thus the relations R:jk = gPqRqijk are established. The relations Rq23k  gpqRpijk are clearly equivalent to these ones. .. We next establish the existence of solutions to the system
ai Fej
=':I ?
Fep in R.
(ii) Let R be a connected and simplyconnected open subset of IR3 and = E C'(R) satisfying the relations let there be given functions
ajqk akrfj rfkr;e rfjrge= o in R, +

which are equivalent t o the relations
ajrikq akrijq r$rkqp rTkrjqPo in R, + =
by part (i). Let a point xo E R and a matrix (F&) E M3 be given. T h e n there exists one, and only one, field (Fej) E C2(R; M3) that satisfies
& F Q ~ ( =) I'Fj(x)Fep(x), E 0, x x Fej(x0)= Fej. 0
24
Philippe G. Ciarlet
Let x1 be an arbitrary point in the set R , distinct f r o m xo. Since R is connected, there exists a path y = (ri) C1([O, l];R3) joining xo to E x1 in R; this means that
y(0) = xo,$1) = xl, and y ( t )E R for all 0
< t 6 1.
ryj(x)Fep(x),x
Assume that a matrix field (Fej) C1(R;M3) satisfies &Fej(x) = E E 0. Then, f o r each integer C E {1,2,3}, the three functions c j E C1([O, 11) defined by (for simplicity, the dependence on C is dropped) C j ( t ) := Fej(r(t)),6 t 6 1, 0
satisfy the following Cauchy problem for a Linear system of three ordinary differential equations with respect to three unknowns:
e c;, m
=
where the initial values
are given by
co := F&. 3
C = 1 , 2 , or 3 only differ by their initial values < .:
Note in passing that the three Cauchy problems obtained by letting
It is well known that a Cauchy problem of the form (with selfexplanatory notations)
has one and only one solution C E C1([0,1];R3) if A E Co([0,1];M3) (see, e.g., Schwartz [1992, Theorem 4.3.1, p. 3881). Hence each one of the three Cauchy problems has one and only one solution. Incidentally, this result already shows that, i f it exists, the unknown field (Fgj) is unique. In order that the three values cj (1)found by solving the above Cauchy problem for a given integer C E {1,2,3} be acceptable candidates for the three unknown values F j(XI), they must be of course independent of the e path chosen for joining xo t o xl. So, let yo E C1([O, l];R3) and y1 E C1([O, 1];R3)be two paths joining xo to x1 in R. The open set R being simplyconnected, there exists a homotopy G = (Gi) : [0,1]x [0,1]+ R3 joining yo to y1 in R, i.e., such that
G(.,O) = yo, G(., 1) = yl, G(t,A) G(0, A) = xo and G ( l , A)
E
R for all 0 6 t < 1, 0 < A 6 1, = x1 for all 0 6 X < 1,
An Introduction to Differential Geometry in R3 and smooth enough in the sense that
G E C1([O, 1 x [0, 1];R3) 1
25
and
Let <(.,A) = (cj(.,X)) E C'([0,1];R3) denote for each 0 6 X 1 the solution of the Cauchy problem corresponding to the path G(.,X) joining 5 to d. thus have ' We
(t,
<
853
aL
dGi A) = I':j(G(t,A))(&
at
X)&(t, A) for all 0
< t < 1,0 < X < 1,
<j(O, A) =
sj"for all o 6 X Q 1.
as, I (,
d X
A)
=0
Our objective is to show that for all 0
< X < 1,
as this relation will imply that [ j ( 1 , O ) = (j (1,l), as desired. For this purpose, a direct differentiation shows that, for all 0 ,< t 6 1, 0 ,< X 1,
where
dkr:j, etc., stand on the one hand (in the relations above and below, for r $ ( G ( . , .)),akr:j(G(., etc.). .)), On the other hand, a direct differentiation of the equation defining the functions aj shows that, for all 0 t 1,0 X 1,
< <
< <
But 2 =
at
ac.
dGi
at
so that we also have
Hence, subtracting the above relations and noting that 
d
(x) (G)and d~ (t) by assumption, we infer that
13 dGi
a[.
d
dGi
=
26
Philippe G. Ciarlet
= r;i combined with the assumed reThe assumed symmetries lations djI’yk  d,$$ + &rye  rfjrg,= 0 in R show that
a p kj
+ rg,r% r;jr;q = 0,
on the one hand. On the other hand,
since $(O, A) = Cj” and G(0, A) = zo for all 0 X 6 1. Therefore, for any fixed value of the parameter X E [0,1],each function oj (., A) satisfies a Cauchy problem for an ordinary differential equation, viz., doj dGi (t,X) =r:j(G(t,X))(t,X)nq(t,X), 0 at dt q ( 0 , A) = 0.
<
< t 6 1,
But the solution of such a Cauchy problem is unique; hence a (t, A) = j 0 for all 0 6 t 6 1. In particular then,
= 0 for all 0
6 X 6 1, 6 X 6 1, since G(1,X) = z1 for all
o<x<1.
(Fej) : R
and thus % ( l , X ) dX
= 0 for all 0
For each integer C, we may thus unambiguously define a vector field + R3 by letting
F e j ( d ) := Cj(1) for any z1E R,
where 7 E C1([O, 1 ;R3) is any path joining zo to z1in R and the vector 1 field ( C j ) E C1([O, 1 ) is the solution to the Cauchy problem 1
Cj(0) = Cj”,
corresponding to such a path. To establish that such a vector field is indeed the Cth rowvector field of the unknown matrix field we are seeking, we need to show that (Fej)&I E C1(R; R3) and that this field does satisfy the partial differential equations &Fej = TpjFep in R corresponding to the fixed integer C used in the above Cauchy problem.
An Introduction to Differential Geometry in R3
27
Let x be an arbitrary point in R and let the integer i E {1,2,3} be fixed in what follows. Then there exist x1 E R, a path y E C1([O, l];R3) joining xo to xl,T E ]0,1[, and an open interval I c [ O , l ] containing T such that y ( t ) = x (t  7 ) e i for t E I ,
+
where ei is the ith basis vector in R3. Since each function <j is contind<. dyz uously differentiable in [0,1] and satisfies " ( t ) = I?$(y(t))(t)&(t) at dt for all 0 t 1, we have
< <
<j(t)
= <(. j()
= <j.(T)
+ (t
+ (t
 T)(T)
dSj
dt
+
O ( t  'T)

T rj ( Y ( T ) ) < p ( T ) f ):
o(t  T )
for all t E I . Equivalently,
Fej(x + (t  .)ei) = F t j ( x ) (t  T)I':j(x)Fep(x) o(t  x). +
This relation shows that each function Fej possesses partial derivatives in the set 0, given at each x E R by
+
d i ~ e p ( ~r:j(x)Fep(x). = )
Consequently, the matrix field (Fej) is of class C1 in R (its partial derivatives are continuous in Q) and it satisfies the partial differential equations 8iFe.j = I'$Fep in R, as desired. Differentiating these equations shows that the matrix field ( F Q )is in fact of class C2 in 0. In order to conclude the proof of the theorem, it remains to adequately choose the initial values F& at xo in part (ii). (iii) Let R be a connected and simplyconnected open subset of R3 and ; ) let ( g i j ) E C2(Q; S be a matrix field satisfying
djrikq dkrijq+ r;jrkqpr:Jjqp o in R, =
the functions
1
rijq, ,I': ?
and gpq being defined by

rijq:= (ajg,, + aigj, 2
aqgij),
ryj := p r i j q , ( p ) ( g i j )  l . :=
; Given a n arbitrary point xo E R, let ( F g ) E S denote the square ; root of the matrix ( g $ ) := (gij (xo))E S. Let (Fej) E C2(R;M3) denote the solution t o the corresponding system
& F e j ( ~= r:j(x)Fep(x),2 E R, )
Fej(xo)= F&,
matrix field (gi g j ) E C2(R. together with initial values at xo. . let x1 E R be distinct from xo and let y E C1([O. . by construction. ) = 923 . = g” in R.. Sj) in a. M3) satisfies &(gz . a j 0 in R.R3) be any path joining xo to x1 in R. . we note that. . To see this. It remains to show that there exists a n immersion 0 E C3(R. g 23 To this end. To begin with. gj (xO = gj”.28 Philippe G. these fields satisfy &gj = q j g P in R.(g. satisfy a Cauchy problem for a linear system of nine ordinary differential equations and this system has at most one solution. as in part (ii). 0 Viewed as a system of partial differential equations. and dkgij = r i k j r’$ imply that + r j k i and rijq gPqF$ = C2(R. ’ (Si g j 1(xO)= go.0 t 1.S?) satisfies Hence the matrix field (gij) E d k g i j = r$grni Sij(Z 0 + r Z g r n j in R .with respect to the matrix field ( g i j ) : R + M3. i. that g i ’ g j = gij.Hence the . we show that the three vector fields defined by gj := (Fej).E3) < < such that &0 = g i in R. 11. 1 where g: is the jth column vector of the matrix (F&) E S53>. E3) such that 8j@e = Fej and g i j = &0 . the above system can have at most one solution in the space C2(R. M3). Ciarlet which exists and is unique by parts (i) and (ii).R3) satisfy 9a.=l E C2(R.. gi) + rK(g. Then the nine functions g i j ( T ( t ) ) . . Then there exists an E immersion 0 = (el) C3(R. 23 The definitions of the functions Fij. S j ) = ’ r.3. An inspection of the two above systems therefore shows that their solutions are identical.e.
equivalently. any solution (Fej) E C2(O. viz. &Fej(z) = r y j ( z ) F e p ( z ) z E R. or. such that the mapping 0 := (el) C3(R. The open set R being simplyconnected. for all i. for defining ad hoc initial vectors gp. p . or Spivak [1999]) shows that. 2. if these equations have a solution. It is no surprise that these necessary conditions are of the same nature as those of Theorem 1. ( 2 ) The assumed positive definiteness of the matrices ( g i j ) is used only in part (iii). for each integer l . k . E3) satisfies E &O = gi in R That 0 is an immersion follows from the assumed invertibility of the matrices ( g i j ) .An Introduction to Differential Geometry in JR3 Since the functions r$ satisfy MI3) of the system 29 = I':i. p. The proof is thus complete. The definitions of the functions r$ and rij. simply expressing that. then necessarily dikFej = &Fej in a. a simple computation shows that they are also necessary conditions. 2342351. see.. . Vol.51. 0 Remarks. . Conversely. j .g. (1) The assumptions made in part (ii) on the functions I'rj = I'yi are thus suficient conditions for the equations &Fej = r$Fep in R to have solutions. Flanders [1989]. e.Schwartz [1992. there exists a function Oe E C3((n) such that ai@e= Fei in R. Poincare"s lemma (for a proof. imply that the functions satisfy.. &gj = &jgk in R. Theorem 59 and Corollary 1. Fej(xo)= Fe"j satisfies &Fej = aj Fei in R.
2. viz.d ) . as for all p E D(R).. . The regularity assumption on the components g i j of the symmetric positive definite matrix field C = (gij) made in Theorem 1. k .S d . First. Mardare [2004] has shown that the existence theorem still holds if gij E WLCm(R). Mardare [2004a]have also shown that.61 are then assumed to hold only in the sense of distributions. This extension only demands that Latin indices and exponents now range in the set .e. C. {1. a quick look at its proof reveals that it holds verbatim in any dimension d 3 2. with a resulting mapping 0 in the space C2(R. Ciarlet These relations in turn imply that the eightyone suficient conditions Rqijk = O in R for all i .3}. S y ) a. 4 ) . but again satisfied if the boundary dR is Lipschitzcontinuous). As expected. The existence result has also been extended “up to the boundary of the set R” by Ciarlet & C. again owing to the relations satisfied by the functions R q i j k for all i.Ed). then a matrix field (gij) E C 2 ( a . .. a. j . that g i j E C2(R).61. with R3 replaced by Rd and E3 by a ddimensional Euclidean space Ed.2.d } and that the sets of matrices MI3.61. . there are other sets of six relations that will suffice as well. and Sd>. q E {1. the E sufficient conditions R q i j k = 0 in of Theorem 1.. if in addition the geodesic distance is equivalent to the Euclidean distance on R (a property stronger than the “geodesic property”. . i. can be significantly weakened. Ciarlet & C. Mardare [2003] has shown that the existence theorem still holds if gij E C’(fl). satisfied in particular if dR is Lipschitzcontinuous) and that the functions gij and their partial derivatives of the symmetric order 6 2 can be extended by continuity to the closure matrix field extended in this fashion remaining positivedefinite over the set Then the immersion 0 and its partial derivatives of order 6 3 can be also extended by continuity to 2. k . Then S. Mardare [2004b]. More specifically. To conclude. More specifically. with a resulting mapping 0 in the space W$‘T((R.. are satisfied if and only if the six relations R1212 = R1213 = R1223 = R1313 = 121323 = R 2 3 2 3 = 0 in fl are satisfied (as is immediately verified. S3. viz. assume that the set R satisfies the “geodesic property” (in effect. we briefly review various extensions of the fundamental existence result of Theorem 1. j .30 Philippe G. a mild smoothness assumption on the boundary dR. and S be replaced by their ddimensional counterparts Md.
7 Uniqueness up to isometries of immersions with the same metric tensor In Section 1. b E R3. Fhe threedimensional vector space R3 is identified throughout this proof with the Euclidean space E3.71.E3) and 0 E C1(R. In particular then. ( . be two immersions such that their associE3) ated metric tensors satisfy VOTVO = ~ 6 in R.6. b (= 1). R3 inherits the inner product and norm of E3. The terminology “rigidity theorem” reflects that such a geometric transformation indeed corresponds to the idea of a “rigid transformation” of the set 0(52)(provided a symmetry is included in this definition). This uniqueness result is the object of the next theorem. aptly called a rigidity theorem in view of its geometrical interpretation: It asserts that. ~ ~ 6 T h e n there exist a vector c E E3 and a n orthogonal matrix Q 6 O3 such that o(IL:) c + Q ~ ( x ) all IL: E R. Let 52 be a connected open subset of R3 and let 0 E C1(R.S. We now turn to the question of uniqueness of such immersions.An Introduction to Differential Geometry in W3 31 with a Riemann curvatu_re tensor vanishing in R can be extended to a matrix field (&) E C2(R. then the set O(52) is obtained by subjecting the set 6 ( R ) either to a rotation (represented by an orthogonal matrix Q with det Q = l). if two immersions 6 E C1(R. provided the given metric tensor field satisfies ad hoc sufficient conditions. This result relies on the existence of continuous extensions to of the immersion 0 and its partial derivatives of order 6 3 and on a deep extension theorem of Whitney [1934]. The spectral n o r m of a matrix A E M3 is denoted ( A (:= sup{lAb(. then by subjecting the rotated set to a translation (represented by a vector c ) .n>)defined on a connected open set 6 containing and whose Riemann curvature tensor still vanishes in 6. we have established the existence of an immersion 0 : 52 c W3 + E3 giving rise to a set @(a) with a prescribed metric tensor.E3) share the same metric tensor field. a 1. = for Proof. Let O3 denote the set of all orthogonal matrices of order three.or to a symmetry with respect t o a plane followed by a rotation (together represented by an orthogonal matrix Q with det Q = l). Theorem 1. E3) and 6 E C1(R.
e.. Q E V.32 Philippe G.O(x)I < sup IVO(z)l/y. Schwartz [1992]) can be applied. The matrix VO'(Z) being thus orthogonal for all 2 E p. Besides. The chain rule applied to the relation O'(O(z)) = x for all z E V then shows that VO'(Z) = V O ( ~ ) . apply the local inversion theorem first to the restriction of 0 to B . thereis no loss of generality in assuming that V is contained in B and that V is convex (to see this. The issue of uniqueness reduces in this case to finding 0 E C1(R. Since the matrix VO(zo) is invertible. Since the set B is convex. h .O(x)I = Iy . thus producing a "first" neighborhood V' of zo contained in B . Let 0l : p + V denote the inverse mapping of 0 : V + p. E3) that satisfies V O ( ~ ) ~ V O ( I~for all x E R = ) i s locally a n isometry: Given a n y point xo E 0.g. then to the restriction of the inverse mapping obtained in this fashion to an open ball V centered at O ( x o )and contained in O ( V ' ) ) .z / for all z.the meanvalue theorem applied in the convex set shows that le'(Q)W'(Z)I < 1 ~ 21for all Z .xi for all z. see. we consider the special case where 6 : R + E3 = R3 is the identity mapping. the meanvalue theorem (for a proof. see. E3) such that V O ( ~ ) ~ V O (I~for all x E R. Ciarlet To begin with. Schwartz [1992]) shows that there exist an open neighborhood V of xo contained in R and an open neighborhood p of O ( x o )in E3 such that the restriction of 0 to V is a C1diffeomorphism from V onto p. the local inversion theorem (for a proof..y E B. there exists a n open neighborhood V of xo contained in R such that IO(y) . (i) We first establish that a mapping 0 E C1(R.'for all z= x E V. we thus have lO(y) . It shows that I O ( y ) .O(x)1 < ly .x for all z. e. y E V i Let B be an open ball centered at xo and contained in R.g. = ) Parts (i) to (iii) are devoted to solving these equations. zElx>Yl Since the spectral norm of an orthogonal matrix is one.y E B.
a classical result from differential calculus (see. Hence for all z. e.ZI for all x . (ii) We next establish that. in the sense that. given any xo E R. The restriction of the mapping 0 to the open neighborhood V of zo is thus an isometry.E3) is locally a n isometry. that there exists a vector c E E3 and a matrix Q E M3 such that (the notation oz designates the column vector with components zi) O(z) = c + Qoz for all x E 0.g.101) shows that the mapping 0 is a f i n e in R. i. Y ) := Then F ( z . Theorem 3. The open set R being connected. y E V by (i). y E V.. in matrix form. that 33 ( y . For a fixed y E V. (z.e. if a mapping 0 E C1(R. y ) (WY)  @ e ( x ) ) ( @ e ( y.An Introduction to Differential Geometry in R3 or equivalently. Letting y = xo in this relation shows that VO(z) = V O ( X ' )for all z E V. y E V . Therefore the mapping 0 : R + E3 is twice n differentiable and its second Fre'chet derivative vanishes a a.z ( < ( O ( y ). then its derivative i s locally constant. . Consequently. y E V . (iii) By (ii). or equivalently..Q)(Y! )  ze) =0 for all z. Schwartz [1992. each function Gi(.. in the sense that VO(Z)= VO(Z') for all z E K The set V being that found in (i).O(z)(= Iy .7.@e(z)). the mapping V O : R + M3 is differentiable and its derivative vanishes in R.(Ye .y ) : V differentiable and its derivative vanishes. let the differentiable function F : V x V + R be defined for all x = ( z p ) V and all y = ( y p )E V by E F ( X . there exists an open neighborhood V of xo contained in R such that I O ( y ) . + J is R dGa dXi d@e dOe y ) = (y)(z) dya dzj + 6 i j = 0 for all z. y E V .O(z)(for all z.
~ The continuous mapping E : V +M3 defined in this fashion is thus locally constant in R.with c E E3 and Q E O3 (an analogous definition holds verbatim in any Euclidean space of dimension d 2 2). let . the matrix Q is orthogonal.'for all 2= ~ ( z )z.E V. E e65(2)T6&(2) all z = I for By parts (i) to (iii). noting that they also read  vo(z)Tvo(z) v G ( ~ ) ~ v G (for )all z E R.34 Philippe G. . = ~ Given any point zo E R. Ciarlet = Since Q = VO(zo) and VO(zO)TVO(zO) I by assumption. As in part (iii). Hence the assumed relations V O ( ~ ) ' V O ( ~= v ) imply that G ( ~ ) ~ v G (for )all z E R ~ E V. the mapping 0 is an immersion. we conclude from the assumed connectedness of C that the mapping E is constant in R. a n isometry preserves distances in the sense that IJ(y) . Consider the composite mapping c & := 6 o 0l : 9 + E3. Clearly. E3) and v q q = VG(z)V@yq = v G ( ~ ) v E I ( ~ ) .$1 for all z. = hence such that =(z) := v G ( ~ ) v o ( ~ ) = Q for all z E V. 6 E C1@. 13 An isometry of E3 is a mapping J : E3 + E3 of the form J(z) = c+Q OX for all z E E3.J(z)I = Iy .he neighborhoods V of zo and of O(zo) and the mapping 01V + V be defined as in part (i) (by : assumption. E V. there thus exist a vector c Q E O3 such that R3 and a matrix 6(2) G(z) = c + QO(Z) for all 2= ~ ( z )z. Clearly. hence the matrix V@(zo) '"is invertible). (iv) We now consider the general equations gij = g23 in R. Thus the proof 2 is complete. y E R.
while part (iv) provides the solution to the equations 9 23. according t o the following corollary to Theorems 1. T h e n 0 is a n isometry of Rd. E3) and 6 E C1(R.E3) and 6 E C1(R. Let the assumptions o n the set R and o n the matrix field C be as in Theorem 1.61. such that 0 = J o 6 . y E R. = c ( x o ) . then 0 is a n isometry. where R is a n open connected subset of R3. E3) is a given immersion. which asserts the following: Let R be a connected subset in Rd. T h e n there exists one and only one immersion 0 E C3(R. where 6 E C1(R. This theorem thus asserts that if a mapping 0 E C1(R. Parts (ii) and (iii) of the above proof thus provide a proof of this theorem under the additional assumption that the mapping 0 is of class C1 (the extension from R3 to Rd is trivial).E3) that satisfies = for v @ ( x ) ~ ~ @ ( x ) C ( X ) all x E R. In Theorem 1. O ( x o )= 0 and V O ( x 0 )= Fo.71. parts (i) to (iii) of the above proof provide the solution to the equations gij = &i in R. In terms of covariant components gij of metric tensors.. Theorem 1. and let 0 : R + Rd be a mapping that satisfies I@(Y)  @(%)I = ly  XI for all x .61 are thus only defined up to isometries in E3. E3) is such that V O ( x )E O3 for all x E R. Theorem 1.e.61 and 1.72. I Remark. according to the classical MazurUlam theorem.An Introduction to Differential Geometry in R3 35 Remarkably.. E3) share the same metric tensor field over a n open connected subset R of R3 i f and only i f they are isometrically equivalent. they become uniquely determined if they are required t o satisfy ad hoc additional conditions. Given any immersion a E C3(R.E3) are said to be isometrically equivalent if there exist c E E3 and Q E O3 such that 0 = c f QO in R.71 thus asserts that two immersions 0 E C1(R.where J is an isometry of E3. and let Fo E M3 be any matrix that satisfies FTF. 0 While the immersions 0 found in Theorem 1. i. let a point xo E R be given. .71. Two mappings 0 E C1(R. aj6 in R . the special case where 0 is the identity mapping of R3 identified with E3 is the classical Liouville theorem. E3) that satisfies Va(z)TVa(x) C ( x ) = . the converse is also true.&0 . . Proof.
E3) be two immersions that satisfy + v @ ( ~ ) ~ v= V + ()S ) ~ V + (for )all x E R. ) + Remark. o(x ~ Hence there exist (by Theorem 1. To see this. that covers the case where one of the mappings belongs t o the Sobolev space H1(R. it is uniquely determined. Ciarlet for all x E R (such immersions exist by Theorem 1.73. The Euclidean space E3 is identified with the space R3 throughout the proof.36 Philippe G.. due to Ciarlet & C. in R and VOTVO = 06’06 a. Besides.61). E3) be a mapping that satisfies det 06 > 0 a. The way the result in part (i) of the next proof is derived is due to F’riesecke. Theorem 1. of all orthogonal matrices Q E O3 with det Q = 1. Theorem 1. let 0 C1(R.71) c E R3 and Q E O3 such that a(x)= c QO(x) for all 5 E $2. E3) be a mapping that satisfies d e t V O > 0 in R. in that both mappings 0 and 6 are assumed to be in the space C1(R.e.71 constitutes the “classical” rigidity theorem. E3).+(zo)) for all x E R. The next theorem is an extension. 6(z) = c + Q@(Z) for almost all x E R. let 0 E C3(R. E and let 6 E H1(R. : i.e. Let R be a connected open subset of R3. the result of part (i) itself goes back to Reshetnyak [1967]. Let 0 denote the set of all proper orthogonal matrices of order three. E3).e. Proof.E3) and E C3(R. One possible choice for the matrix F is the square root of o 0 the symmetric positivedefinite matrix C (x 0 ) . so that V@(x) = Q V O ( x ) for all x E R. Then it is immediately verified that this mapping 0 satisfies the announced properties. in R. . let the mapping 0 : R 3 R3 be defined by O ( x ) := FoV+(xO)l(+(x) . James & Muller [2002]. Mardare [2003]. The relation VO(x0) = V+(xo) then implies that Q = I and 0 the relation O(z0) = ~ ( Z O in turn implies that c = 0. T h e n there exist a vector c E E3 and a matrix Q E 0 such that .
g. in for a proof of this identity.~ .1(2) = v o ( x ) v o ( x ) . div C o f V 6 = 0 in ( D ' ( B ) ) 3 in any open ball B such that B c R (to see this. . Hence : C o f V G ( z ) = (det V 6 ( x ) ) V 6 ( ~ ) =~ ~ ( X ) . combine the density of C2(B) H 1 ( B )with the classical Piola identity in the space C2(B). on the one hand.' . see.. E a(a&aiGj) For such mappings. there thus exist bounded open neighborhoods U of zo and U of Q(Q) satisfying U c R and { } c Q(R). Let xo E R be given.An Introduction to Differential Geometry in R3 37 (i) To begin with. together with the relations AC3j = 0 and &Oj&6j = 3 in R. on the other hand. (ii) Consider next the general case. 6 = (6j) (C"(R))3.V for almost all II: E R.we conclude that A6 Hence = divCofV6 = 0 in ( D ' ( B ) ) 3 .1]). it follows that 6 E H1(6. V ~ ( Z )0 for almost all II: E 0) such that E . Ciarlet [1988. Let 0' : 64 U denote the inverse mapping of O U . The assumed connectedness of R then implies that there exist a vector c E E3 and a matrix Q E 0 (by assumption. the identity = aai6jai(~6j) + 2t4k6jaik6j. In other words. e. 6 Since from E H1(U) and 0 ' can be extended t o a C1diffeornorphism . . shows that aikC3j = 0 in R. the local inversion theorem can b_e applied.    6 ( x )= c + Q ox for almost all x E R. we arc given a mapping 6 E H1(R) that satisfies vO(z) E 0 for almost all x E R.R3) and that 6VG(2) = V6(2)60.7. Theorem 1. we consider the special case where Q ( x ) = x for all x E 0. { } onto g . Since 0 is a n immersion. which therefore satisfies VQG'(2) = VO(x)l for all 2 = Q ( x ) E (the notation 6 indicates that differentiation is carried out with respect to the variable 2E Define the composite mapping u 6 G). such that the restriction Q U of 0 to U can 6be extended to a C1diffeomorphism from onto {6}. Since.
(1) The existence of 6 E H1(R. if a mapping 0 E H1(R. there thus exist c E E3 and Q E 0 such that By : &(2)= 6(z) = c + Q 02 for almost all P = ~ ( z E) U .e. E3). Schwartz [1966. let O(z) = z. the assumption d e t V O > 0 in R cannot be replaced by the weaker assumption det VO > 0 a.: for all z E R. in R. together imply that 6 & ( P ) E 0 for : almost all 2 E 6.z 1 .71).73). l Since the point zo E R is arbitrary. By a classical result from distribution theory (cf. To see this. in R. det V 6 > 0 a.e. and consequently that 6(z) = c + Q O ( ~for almost all z E Q. given a connected open subset R of R3. To see this. let for instance R be an open ball centered at the origin in R3. Hence the assumptions det VO > 0 in R. Section 2.38 Philippe G. let + . (3) The result established in part (i) of the above proof asserts that. or equivalently. an assumption about the sign of det 06 becomes necessary. such that ~ ( z:= ) h v O ( z ) v ~ ( z ) = Q for almost all z E U. (i). (5) Surprisingly.g.E3) and 6 E C1(R. 2 3 ) if z < 0. and 6 VOTVO = V G T V 6 a. (4)By contrast. we conclude from the assumed connectedness of R that E(z) = Q for almost all z E R.6]). let for instance R be an open ball centeredat the origin in R3. 2 2 . Then 6 E H1(R. Adams [1975. ) 0 Remarks. in R. Chapter 31).e.73 thus implies that 0 € H1(R. since 6 ( R ) c {z E R3. but then it can only be concluded that Q E 03: This is the classical rigidity theorem (Theorem 1. in R. Ciarlet for almost all P = O(z) E (see.E3) and 1 1 V 6 E O3 a. 1 3 0 z ) (this counterexample was kindly communicated to the author by Sorin Mardare). the assumptions d e t V O > 0 in R and d e t V 6 > 0 in R are no longer necessary. E3) is such that VO(z) E 0%for almost all z E R. if the mapping 0 is assumed to be instead in the space H1(R.E3). (2) If 6 E C1(R. This : result thus constitutes a generalizatio_n of Liouville 's theorem. then there exist c E E3 and Q E 0 such that O(z) = c &ox for almost all z E R. e. and let 6(z) = z if z 2 0 and O(z) = ( . yet there does not exist any orthogonal matrix such that 6(z) = Qo. E3) satisfying the assumptions of Theorem 1. of which Liouville's theorem is the special case corresponding to O(z) = z for all z E R. this relation shows that E E Ltoc(R).e..E3)(as in Theorem 1.
8 Continuity of an immersion as a function of its metric tensor Let R be a connected and simplyconnected open subset of R3.61. The assumption that det VO > 0 in R made in Theorem 1. a welldefined element 3 ( C ) in the quotient set C3(R. we define the seminorms where stands for the standard multiindex notation for partial derivatives. if R is a connected open set and O E C1(R.W). and means that K is a compact subset of R. The notation K R 0. and K @ R.(R. = where the functions rij. 0 Like the existence results of Section 1.61 and 1. A natural question thus arises as to whether the mapping 3 defined in this fashion is continuous.S satisfying ).71 establish the existence of a mapping 3 that associates with any matrix field C = ( g i j ) E C2(R. S3) and C3(R. R q i j k := ajrikq + r. and) let 23 2 0 and 6 ( x ) = ( .E3)is replaced by the weaker assumption O ( x ) = O(z) if 22 39 6 E Hk. a positive answer to this question has been provided by Ciarlet & Laurent [2003]. 1 2 0. If g E C'(R.73 is simply intended to f x ideas (a similar result clearly holds under the i other assumption).E3). .2 2 . ~ ) if 22 < 0 (this counterexample was kindly communicated to the author by HervB Le Dret). then either det VO > 0 in R or det VO < 0 in R. Conversely. then 0 is an immersion. (6) If a mapping 0 E C1(R. is a n immersion a continuous function of its metric tensor? When both spaces C2(R.q x i . 6 )E R means that there exist a vector a E E3 and a matrix Q E CD3 such that O(z) = a + Q6(z) for all 2 E R. E3) are equipped with their natural Frkchet topologies.t K cz R.$'kqp  I'ykrjqp0 in 0. ~ 2 . the uniqueness theorems of this section hold verbatim in any dimension d 2 2. and l7$ are defined in terms of the functions g i j as in Theorem 1. E3)/R. If 0 E Ce(R.An Introduction to Differential Geometry in R3 O ( x ) = ( x 1 ~ ~ . M3). More specifically. Together. Equivalently. with R3 replaced by Rd and Ed by a ddimensional Euclidean space. E3) is an immersion. 1. Theorems 1. let R be an open subset of R3.where ( O . we likewise aa .6.73 is still valid if 6 E H1(R. (7) A little further ado shows that the conclusion of Theorem 1. E3) or A E Ce(R.E3) satisfies d e t V O > 0 in 0.
K. Ciarlet set where 1. e. that there exist a vector c E E3 and a matrix Q E O3 such that 0 ( x ) = c + Q G ( x ) for all x E R.e. Then.G) &(6.7 z inf. Furthermore. Let C3(R. x)= { z:: inf &(GIc + Q+). this topology is metrizable: Let (Ki)+o be any sequence of subsets of R that satisfy K CE L? and Ki i Then C int Ki+l for all i 3 0. Then a sequence (On). ~ 0 for all = K G R.0) 0. see. defined above..O l l e . E3) := C3(R.40 Philippe G../le.g. E3) by the equivalence relation R.O l l e . E3)/R denote the quotient set of C3(R. E3) becomes a metric space when it is equipped with the distance d 3 defined by (0.Rd) becomes a locally convex topological space when it is equipped with the Fre'chet topology defined by the family of seminorms /l. and R = u 00 Ki. Chapter 11. Yosida [1966.where E R means that 0 and 6 are isometrically equivalent (Section 1. . = 0 for all K G R ?Ltco + lim n+m &(On. i.+) = { where .7). the space Ce(L?. 6 denotes the equivalence class of 0 modulo R. K R. i=O 1" ~ lim 10 . = where For details about F'rkchet topologies. for any integers C 3 0 and d 3 1. dg()i. Then it is easily verified that the set C3(R.1 denotes either the Euclidean vector norm or the matrix spectral norm.20 converges t o 0 with respect t o this topology if and only if ncc lim /I@" .
Conceivably. Another extension. S) d2) . 0 As shown by Ciarlet & C. R q i j k =0 in a}.”(R. E denote the equivalence class modulo R of any 0 E C3(R2. and. T h e n the mapping F : {C. as follows.E3) that satisfies VOTVO = C in R.  {C3(R. Let Cz(R. in R as a possible deformation of B when B is subjected to ad hoc applied forces and boundary conditions.E3)that is almosteverywhere injective and satisfies det V 0 > 0 a. Thanks mostly to the landmark existence theory of Ball [1977]. the associated metric space is denoted { X . is the following: Let R be a bounded and connected subset of W3. Theorem 1..S3) the for continuous extensions of the symmetric matrix fields C . according to the following result (if d is a metric defined on a set X . and C3@. The almosteverywhere injectivity of 0 (understood in the sense of Ciarlet & NeEas [1987])and the restriction on the sign of det V O mathematically express (in an arguably weak way) the noninterpenetrability and orientationpreserving conditions that any physically realistic deformation should satisfy.S : let F(C) C3(R. the mapping F can be extended to a mapping that is locally Lipschitzcontinuous with respect to the topologies of the Banach spaces C2(a. associated with a deformation 0 E H1(R. it is now customary in nonlinear threedimensional elasticity to view any mapping 0 E H1(R.e. essentially motivated by nonlinear threedimensional elasticity. ) := {(gij) E C2(R.. s”. Mardare [2004b].E3)pervades the mathematical modeling of threedimensional nonlinear elasticity (extensive treatments of this subject are found in Marsden & Hughes [1983] and Ciarlet [1988]). If R is bounded and has a Lipschitzcontinuous boundary.d 3 ) defined in this fashion is continuous. E3) ) . .2(R. E3) for the continuous extensions of the immersions 0.d } ) . and let B be an elastic body with R as its reference configuration. an alternative approach to the existence theory in threedimensional elasticity could thus regard the CauchyGreen tensor as the primary unknown.81. Let R be a connected and simplyconnected open subset ofR3. instead of the deformation itself as is usually the case. the above continuity result can be extended “up to the boundary of the set R”. E3). It then turns out that the metric tensor field VOTVO E L1(R. S 3 ) . S . given any matrix field C = ( g i j ) E C.An Introduction to Differential Geometry in R3 41 The continuity of an immersion as a function of its metric tensor has then been established by Ciarlet & Laurent [2003]. also known as the CauchyGreen tensor field in elasticity.
which shows that. E~) That a vector b and an orthogonal matrix R should appear in the lefthand side of such an inequality is of course reminiscent of the classical rigidity theorem (Theorem 1. ) One application of the above key inequality is the following sequential continuity property: Let O k E H1(R. since the CauchySchwarz inequality immediately shows that the mapping 0 E H1(O. the CauchyGreen tensors depend continuously o n the deformations. under appropriate smoothness and orientationpreserving assumptions. Then Ciarlet & C.0 )of order three (i. there exist a proper orthogonal matrix R = R ( 9 . the converse holds.E3) a mapping satisfying be det V 0 > 0 in Then there exists a constant C(0) with the following property: For each orientationpreserving mapping 9 E H1(QE3). if two mappings 6 E C1(R.V@TV@/I$n. C(@)IIV@V@. there exist a vectorb in E3 and a proper orthogonal matrix R of order three such that 0 ( z ) = b RO(z) for all z E 0. More generally. Then there exist a constant C ( 0 ) . i. This continuity result is itself a simple consequence of the following nonlinear Korn inequality.E3).. Ciarlet Clearly. Mardare [2004c] have shown that.. (by contrast with the orientationpreserving condition.E3) k be orientationpreserving mappings.: Let R be a bounded and connected open subset of R3 with a Lipschitzcontinuous boundary and let 0 E C1(n. an'orthogonal matrix of order three with a determinant equal to one) and a vector b = b ( 9 . E3) f VOTVO E L1(O. the isS3) sue of noninterpenetrability turns out to be irrelevant to this issue). E3) are isometrically equivalent if there exist a vector b in E3 and an orthogonal matrix R of order three (a proper one in this case) such that + 6(z) = b + R e ( % for almost all z E R..s3). the topologies being those of the same spaces H1(R. E3) and 0 E C1(R.0 )in E3 such that n. i. 1 . E3) and 0 € H1(R. S3) is continuous (irrespectively of whether the mappings 0 are almosteverywhere injective and orientationpreserving).e. e. and 0 E C1(2. then the two mappings are isometrically equivalent.E3)and L1(0.71). we shall say that two orientationpreserving mappings 6 € H1(R.e.42 Philippe G. E3) satisfying det 06 > 0 and det V 0 > 0 in an open connected subset R of R3 have the same metric tensor field. the deformations depend continuously o n their CauchyGreen tensors.(b 1 9 + R O ) I ( H ~ ( ~6. which constitutes the main result of ibid. 3 1.
is unambiguously defined (up to isometries of E3) by a single tensor field. whose covariant components g i j = gji : R + R are given by 923 ’ ai0 .where R is an open set in R3 and 0 : R + E3 is a smooth injective immersion. in particular for handling the part of the energy that takes into account the applied forces and the boundary conditions.e. such a sequential continuity could thus prove to be useful when considering injimizing sequences of the total energy. E3).. which implied continuity at rigid body deformations. a j 0 . The recent and noteworthy continuity result of Reshetnyak [2003]for quasiisometric mappings is in a sense complementary to the above one (it also deals with Sobolev type norms). James & Muller [2002]. John [1972] and Kohn [1982].k 2 1. 2 Differential geometry of surfaces Outline We saw in Part 1 that an open set 0 ( R ) in E3. ”Consider instead a surface G = O(w) in E3. by making use in particular of a fundamental “geometric rigidity lemma” recently proved by F’riesecke. . They key inequality is first established in the special case where 0 is the identity mapping of the set R. Should the CauchyGreen strain tensor be viewed as the primary unknown (as suggested above). at a mapping 0 that is isometrically equivalent to the identity mapping of R. 3 as converges to V O in ) k )El k+m. that are isometrically equivalent to O k such that Hence the sequence (0 converges to 0 in H1(R.thanks in particular to a methodology that bears some similarity with that used in Ciarlet & Laurent [2003]. It is then extended to an arbitrary mapping 0 E C1@ EXn) satisfying det V 0 > 0 in 2.An Introduction to Differential Geometry in R3 k 43 and orientationpreserving mappings 0 E H1(R. E3) as k + 00 if the sequence ( ( v o ~ ) ~ v o ~ ) ~ = . Such results are to be compared with the earlier. _ . such a “twodimensional manifold” equipped with the coordinates of the points of w as its curvilinear coordinates. Then by contrast. ~ V O L ~ ( R~. which are both naturally expressed in terms of the deformation itself. i. the metric tensor field. where w is a twodimensional open set in R2 and 8 : w + E3 is a smooth injective immersion. pioneering estimates of John [1961].
such as the principal curvatures and the Gaussian curvature. a g and b. (T.5.. are introduced and briefly discussed in Section 2. & u p .41).rPTp b.aUb. Mardare [2007]in this Volume. These two fundamental forms are introduced and studied in Sections 2. equations (Theorem 2.p = a. We establish in this process the formulas of Gauss and Weingarten (Theorem 2. Ciarlet [2005].) and I’Ep = a U T a p T . Ciarlet [2000].p.p)’. their components and some of their partial derivatives must satisfy necessary conditions taking the form of the following relations (meant to hold for all a .71): Let the functions rapT and be defined b Y 1 r a p T = a(dpa. o n the surface G are computed in terms of its curvilinear coordinates by means of the components a. the unknown displacement vector field of the middle surface of a shell) pervade the equations of shell theory.1 to 2. In particular. “metric notions”. necessarily.31). constitute the contravariant bases along 2.Their covariant components a..  8. E T { 1.. We next introduce in Section 2.44 Philippe G.&a. which respectively constitute the Gauss.61). apb.g) and the symmetric matrix field (hap) defined on w in this fashion cannot be arbitrary. It is also shown how the curvature of a curve on G can be similarly computed. &upT .. Other classical notions about “curvature”. The vector fields ai : w ’+ R3 defined in this fashion constitute the covariant bases along the surface G I while the vector fields ai : w R3 defined by the relations ai .2}): + a. : w + R and bag = bp..the first and second fundamental forms of G. : w R are respectively given by (Greek indices or exponents take their values in {1.. in = .bapbU.g.5.2}).. . . It is a basic fact that the symmetric and positive definite matrix field (a. Ciarlet requires two tensor fields for its definition (this time up to proper isometries of E3). and CodazziMainardi.bpT . see.e. it is shown how areas and lengths.d’. of a vector field qiai : w +R3 o n G I thus defined here by means of its covariant components qi over the contravariant bases az.p = ap. More specifically.g + rgubfip I ’ ~ o b u p O in w.  W.p of the first fundamental form (Theorem 2. Then. + r&ruTp r:. i. e. Covariant derivatives of vector fields on a surface (typically.6 the fundamental notion of covariant derivatives qil.p = a3 . this time by means of the components of both fundamental forms (Theorem 2.. aj = 6.. f + where (auT):= (a. = W. p.or Ciarlet & C.
the necessary conditions expressed by the Gauss and CodazziMainardi equations are also suficient for the existence of such an immersion 8.An Introduction to Differential Geometry in R3 45 The functions r a p 7 and rgp are the Christoffel symbols of the first. when does there exist an immersion 8 .p) together with a smooth enough symmetric matrix field (b. the above existence and uniqueness theorems constitute the fundamental theorem of surface theory. w + E3 such that < If such an immersion exists.81 and 2. they constitute the core of Part 2. to what extent is it unique? As shown in Theorems 2. Together.91 (like those of their “threedimensional counterparts” in Sections 1.7. the Gaussian curvature is a given function (the same for any surface) of the components of the first fundamental form and their partial derivatives of order 2 at the same point. the answers turn out to be remarkably simple: Under the assumption that w is simplyconnected. Besides. depends continuously o n the matrix fields (a. especially that of the existence). kind.10 how the equivalence class of 8. and second. We then turn to the reciprocal questions: Given an open subset w of R2 and a smooth enough symmetric and positive definite matrix field (a. if 8 : w 3 E3 is any other smooth immersion satisfying there then exist a vector c E E3 and a proper orthogonal matrix Q of order three such that qY)c + Q@(.72) the celebrated Theorema Egregium of Gauss: At each point of a surface.6 and 1.p) defined over w. We conclude Part 2 by indicating in Section 2.e. . i. This means that. when are they the first and second fundamental forms of a surface 8 ( w ) C E3.. if w is connected. their proofs are by no means easy. As such. this irnmersion is unique up t o proper isometries in E. defined in this fashion modulo proper isornetries of E3.) for all = E W. another important special case of the fundamental theorem of Riemannian geometry already alluded to in Part 1. We also establish in passing (Theorem 2.p) and (baij) with respect t o appropriate topologies.
1 Curvilinear coordinates on a surface In addition to the rules governing Latin indices that we set in Section 1. let there be given a twodimensional vector space. let a . and the vector product of vectors a. of 6 (Figure 2. . there are infinitely many other ways of defining curvilinear coordinates on 2. := d l d y . For instance. in which two vectors ea = e form a basis. If the mapping 0 : w + E3 is injective. b in the space E3.13.2) are given in Figures 2. Kronecker’s symbols are designated by b t . denote the coordinates of a point y E EXz and let 8.1 a threedimensional Euclidean space E3.1.2} and that the summation convention be systematically used in conjunction with these rules.p := d2/dyadyp. we henceforth require that Greek indices and exponents vary in the set {1. according t o each context). Wellknown examples of surfaces and of curvilinear coordinates and their corresponding coordinate lines (defined in Section 2.2. This space will be identified . E G can be YEW. Ciarlet 2.12 and 2. The set G := qW) is called a surface in E3. and the two coordinates y of y are called the curvilinear coordinates . For instance. b.baflr73)aP + (r/3/a b!7p)a3 + means that. and d. with EX2. once a surface 2 is defined as 2 = 8(w). Let y. for a = 1. and a A b denote the Euclidean inner product. lal. depending on how the domain w and the mapping 8 are chosen. a portion 2 of a sphere may be represented by means of Cartesian coordinates. the relation da(’%aZ)= (qflla .46 Philippe G. Let there be given as in Section 1. and the Euclidean norm. Finally. or Sap according to the context. In addition. Sap. Naturally.equipped with an orthonormal basis consisting of three vectors i e = Zi.11). each point unambiguously written as i j = 8(Y). let there be given an open subset w of R2 and a smooth enough mapping 0 : w 4 E3 (specific smoothness assumptions on 8 will be made later.
then + .13).11: Curvilinear coordinates on a surface and covariant and contravariant bases of the tangent plane.yz of y E w are the curvilinear coordinates of $ = B(y) E G. The two vectors a"(y) from this tangent plane defined by a*(y) . Let G = B(w) be a surface in E3. If the two vectors a.An Introduction t o Differential Geometry in R3 47 Figure 2. ao(y) = 6 . The two coordinates y1. or stereographic coordinates (Figure 2. spherical coordinates. If 6y is such that (y 6y) E w . this example illustrates the variety of restrictions that have to be imposed on i3 according to which kind of curvilinear coordinates it is equipped with! 2. they are tangent to the coordinate lines passing through y^ and they form the covariant basis of the tangent plane t o 2 at $ = B(y).2 First fundamental form Let w be an open subset of R2 and let be a mapping that is dzflerentiable at a point y E w. Incidentally.(y) = &B(y) are linearly independent. form its contravariant basis.
A portion of C that excludes a neighborhood of both “poles” and of a “meridian” (to fix ideas) can be represented by means of spherical coordinates. with a maDDina 8 of the form: 2R2v u2 v2 .R2 e : (u. { R 2 . with a mapping 8 of the form: 8 : (cp.12: Several systems of curvilinear coordinates on a sphere. A portion of C that excludes a neighborhood of the “North pole” can be represented bv means of StereomaDhic coordinates.$) E w + (Rcos$coscp. E w t (2.w) E + + _ _ + The corresponding explanatory graphical conventions.(z2 y2)}l/’) E E3. Let C C E3 be a sphere of radius R.Rsin$) E E3. y) y. with a mapping 8 of the form: 8 : (e.Rcos$sincp. Ciarlet X Dy Figure 2. .48 Philippe G. A portion of C contained “in the northern hemisphere” can be represented by means of Cartesian coordinates.
A portion ij of a circular cylinder of radius R can be represented by a mapping 8 of the form 8 : (p. Rsin p. . with selfexplanatory graphical conventions. ~ s i nE E 3 .13: Two familiar examples of surfaces and curvilinear coordinates. The corresponding coordinate lines are represented in each case.An Introduction to Differential Geometry in IR3 49 cp / I I I w/ Figure 2. ( R + r c o s x ) s i n p .z ) E w + (Rcos p. z ) E E3. ~E)w + ( ( R + r c o s ~ ) c o s p . A portion G of a torus can be represented by a mapping 8 of the form 8 :( p . x) with R > T .
8(y) are linearly independent.(y) + o ( 6 t ) .w = )(w2 : (Y). i.11. . and 2. or equivalently if the two vectors a.(y)). Ciarlet where the 3 x 2 matrix V8(y)and the column vector 6y are defined by Let the two vectors a. passing through y (there exist t o and tl with t o < 0 < tl such that the ath coordinate line is given by t E ]to.(y) is tangent t o the a . Then the 8(Y + SY) = O(Y) + 6Y"aa(Y) + O(6Y). The vectors a. This observation suggests to define a matrix (a. ( t ) := O(y t e a ) in a neighborhood of & hence f&(0) = da8(y) = a. a a ( y ) is the a . Returning t o a general increment Sy = 6 y a e a .13. the principal part with respect to Sy of the length between the points e(y++y) and 8(y) is {6yaa. A mapping 8 : w + E3 is an immersion at y E w if it is differentiable at y and the 3 x 2 matrix V8(y)is of rank two.e.q Y ) l Z = 6 Y T v q Y ) T v q l J ) 6 Y + O(l6Yl2) = b " a a ( y ) .(y) .t h coordinate line passing through @ = 8(y). we also infer from the expansion of 8 about y that (6yT and V8(y)T respectively designate the transpose of the column vector 6y and the transpose of the matrix + VO(Y)) 1% + SY) .. 2.ap(y)6yP}l12. this relation reduces to e(y + bte.(y). ap(Y)6Yp + O(l6YI"). Assume from now on in this section that the mapping 8 is an immersion at y. where 6 t E R and e.(y) := E R3 be defined by a. see Figures 2. defined as the image by 8 of the points of w that lie on a line parallel to e. In this case.p(y)) of order two by letting a d y ) := d Y ) .12. is one of the basis vectors in EX2. form the covariant basis of the tangent plane to ij at see Figure 2...11.(y) = 8. a p ( y ) = (ve(Y)Tve(Y))a.) + Sta.t h column vector of the matrix expansion of 8 about y may be also written as VB(y). = B(y) If in particular 6y is of the form 6y = Ste. the last relation shows that each vector a.50 Philippe G.tl[+ f . which thus span the tangent plane t o the surface 2 a t = 8(y). In other words.
of the surface 2 at $ = 8(y). where (aap(Y)) := ( a a a ( y ) ) . i. a : w + R3 respectively form the covariant. a"(y) = a""Y)adY). the mapping 8 : w c R2 + E3 "replacing" the mapping 0 : R c R3 + E3. A word of caution.(y).11) and the elements a " p ( y ) of the symmetric matrix ( a " p ( y ) ) are called the contravariant components of the first fundamental form.p(y)) is positive definite since the vectors a. If 8 : w + E3 is an immersion. ( Y ) . the f o u r relations a"(T/) d y ) = 6. There are indeed strong similarities between . . hence Y*"(y) = ~ " " ( y ) .(y) being thus defined.. A mapping 8 : w + E3 is an immersion if it is an immersion at each point in w . The two vectors a. or metric tensor. To see this.2. and contravariant. The presentation in this section closely follows that of Section 1. the vector fields a : w + R3 and . let a priori a"(y) = Y""(y)ao(y) in the relations a*(y) .(y) are assumed to be linearly independent.p(y) = 6. These relations in turn imply that ' a a ( Y )4 Y ) = a a u ( Y ) a w a .e. Note that the m a t r i z (a. ag(Y) and % Y = aap(Y)aP(Y) and () a"P(d = a"(Y) . a p ( ~ )= d.(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 * p ( y ) = a.An Introduction to Differential Geometry in R3 51 The elements a a p ( y ) of this symmetric matrix are called the covariant components of the first fundamental form.. The two vectors a*(y) form the contravariant basis of the tangent plane to the surface 2 at g = 8(y) (Figure 2. also called the metric tensor. This gives Y""(y)a. = aau(y)a"(y)am7(y) UAY) = aau(y)6t = u"p(y). if 8 is differentiable in w and the two vectors d a 8 ( y ) are linearly independent at each y E w . We would likewise establish that a a ( y ) = aaa(Y)aP(d. unambiguously dejine two linearly independent vectors a"(y) in the tangent plane. " bases of the tangent planes..aP(Y). of the surface 2 at g = 8(y). and thus the vectors a"(y) are linearly independent since the matrix ( a " p ( y ) ) is positive definite.l Hence a " ( y ) = a""(y)a.
Indeed. let 8 : w 4 E3 be a n injective and smooth enough immersion. but there are also sharp differences. The elements dZi(g) and d&) at y^ = 8(y) E is are related t o dy and 6y by means of the covariant components of the metric tensor of the surface 2 . .3 Areas and lengths on a surface We now review fundamental formulas expressing area and length elements at a point g = 8(y) of the surface w^ = O(w) in terms of the matrix (a. These formulas highlight in particular the crucial r6le played by the matrix (a. the first fundamental form well deserves “metric tensor” as its alias! Figure 2. cf. while the matrix VO(z) is square! 0 2. In particular. where C = f ( I ) and I is a compact interval of R. such as the way the metric tensor is defined in both cases. where a(y) := det(a.p(y)).52 Philippe G.31.31.p(y)).31.p(y)) for computing “metric” notions at y^ = 8(y).31:Area and length elements on a surface. Let w be a n open subset of R2. Ciarlet the two presentations. Theorem 2. (a) The area element dZ@) at g= 8(y) E 2 is given in terms of the area element dy at y E w by dZ@) = m d y . Theorem 2. The corresponding relations are used for computing the area of a surface A = @ ( A ) is and the length of a curve C c^ = 8 ( C ) C 2. see Figure 2. and let ij = 8 ( w ) . the matrix V8(y) is not a square matrix.
Section 1.8(y)/.An Introduction to Differential Geometry in R3 53 (b) The length element dZ(fj) at c = 8(y) E w^ is given by dZ(fj) = {by"~. We assume again that 8 : w t E3 is an injective and smooth enough immersion.3). in terms of the curvilinear coordinates used for defining the surface ij (see again Figure 2. Proof. areaA := Consider next a curve C = f(1) in w . Then the Zength of the curve := O(C) c w^ is given by 21 The last relation shows in particular that the lengths of curves inside the surface 8 ( w ) are precisely those induced by the Euclidean metric of the space E3. It c%n also be deduced directly from the relation between the area elements d r ( 2 ) and dr(z) given in Theorem 1. 0 The relations found in Theorem 2. let := @ ( A ) and let f ^ E L1(A^) be given.. Then A^ In particular.e. 2.31).i.p(y)by~}~'~. where I is a compact interval of R and f = f "e" : I 4 w is a smooth enough injective mapping.31 (b) by means of an ad hoc "threedimensional extension" of the mapping 8. The expression of the length element in (b) recalls that dZ@) is by definition the principal part with respect to 6y = by"ea of the length l8(y f 6y) . the area of h A^ is given by dZ(fj) = m d y .p(y)). cf. the surface 8 ( w ) is said to be isometrically immersed in E3. open.4 Second fundamental form. . Let A be a domain in R2 such that ?I c w (a domain in pz is a bounded. whose expression precisely led to the introduction of the matrix (a. curvature on a surface While the image O(R) c E3 of a threedimensional open set R c R3 by a smooth enough immersion 0 : R c R3 4 E3 is well defined by its . The relation (a) between the area elements is well known. For this reason. and connected subset of R2with a Lipschitzcontinuous boundary.31 are used as follows for computing surface integrals and lengths on the surface w^ by means of integrals inside w .
Theorems 1.provided ad hoc compatibility conditions arc satisfied by the covariant components g i j : R + R of its metric tensor (cf.54 Philippe G. 20 and G2. Ciarlet “metric” uniquely up to isometrics in E3. Yet it should be clear that in general 20 and GI. Section 2.1 a surface ij = 8(w)in E3.41: A metric alone does not define a surface an E3. For each y E w . As intuitively suggested by Figure 2. A flat surface GO may be deformed into a portion Gi of a cylinder or a portion 2 2 of a cone without altering the length of any curve drawn on it (cylinders and cones are instances of “developable surfaces”. cf. or 2 1 and G2. which naturally appears for this purpose through its covariant components (Theorem 2.the vector . solving this question relies on the knowledge of the second fundamental form of a surface.5). As shown in this section.41 the missing information is provided by the “curvature” of a surface. where w is an open subset of R2 and 8 : w C R2 + E3 is a smooth enough immersion. I Figure 2. are not identical surfaces modulo an isometry of E3! or Consider as in Section 2.71)1a surface given as the image O(w) c E3 of a twodimensional open set w c R2 by a smooth enough immersion 8 : w c R2 + E3 cannot be defined by its metric alone.41). A natural way to give substance to this otherwise vague notion consists in specifying how the curvature of a curve o n a surface can be computed.61 and 1.
it is customary to still write 1 the curvature as . hence in particular of class C2 o n w. the radius of curvature of the curve 6 at y^ is said to be infinite. has Euclidean norm one. since the symmetric matrix (a.An lntroduction to Differential Geometry in R3 55 is thus well defined.The intersection = P n 2 is thus a planar curve on G. the center of curvature is intrinsically defined. i. it is remarkable that the curvature of at y^ can be computed by means of the covariant components a.E3): The radius of curvature of a curve corresponding to a constant y2 vanishes at y1 = 0. R 1 Note that the real number . This implies in particular that the radius of curvature never vanishes along a curve o n a surface 8 ( w ) defined by a mapping 8 satisfying the assumptions of the next theorem. It is easily seen that the denominator in the definition of a3(y) may be also written as l a d y ) A az(?/)I = where a(y) := det(a. 1 If the algebraic curvature of at y^ is # 0. the mapping 8 “cannot be too smooth”. Figure 2. .p(y) of the first fundamental form of the surface G = 8(u)introduced in Section 2. Think of a surface made of two portions of planes intersecting along a segment. m. y2) = )ylJ1+“for some 0 < Q < 1. Remark.p(y)) is positive definite.p(y) of the “second” fundamental form of 2.41. a plane that contains the vector ay(y>. This means that the center of curvature of the curve C at y^ is the point (y^+Ras(y)). for this reason.e. which thus constitutes a fold on the surface..E3)but 8 !$ C2(w.43.p(y)) a ( y ) = 0. and 21 RR is then called the algebraic radius of curvature of the_curve C at y^.2. it can be written as .is always well defined by the formula R given in the next theorem. and is normal to the surface G at the point y^= 8(y). so that 8 E C1(u. Or think of a surface 8(u) with 0 E w and B(y1. 0 and that this relation holds in fact even if Fix y E w and consider a plane P normal to G at y^ = 8(y).in this case. see 21 as its sign changes in any system of curvilinear coordinates where the normal vector as(y)is replaced by its opposite.42. While R is not intrinsically defined. together with the covariant components b. The definition of the curvature of a planar curve is recalled in Figure 2. As shown in Theorem 2. If the curvature of C at Q is 0. It is intuitively clear that if R = 0.
.41. which is intrinsically defined. its inverse R is called the “algebraic radius of curvature” of y at p ( s ) (the sign of R depends on the orientation chosen on y). called the “curvature” of y at p ( s ) . + + + When As + 0.p ( y ) are the covariant components of the first fundamental f o r m . i. where f = f Q e a : I + R is a smooth enough injective mapping that satisfies . the ratio ~ this limit is nonzero. Consequently. Consider two points p ( s ) and p ( s As) with curvilinear abscissae s and s As and let Aq5(s) be the algebraic angle between the two normals u ( s ) and u ( s As) (oriented in the usual way) to y at those points.of the planar curve C at 1 h R c is given by the ratio where a. let 8 E C2(w. Let w be an open subset of R2. called “la d6velopp6e” in French. If + + + Theorem 2. the centers of curvature of y lie on a curve (dashed on the figure).( t ) e .42: Curvature of a planar curve. where t E I is such that df“ dt y = f ( t ) (Figure 2. parametrized by its curvilinear abscissa s. The center of curvature is also the limit as As + 0 of the intersection of the normals u ( s ) and u ( s As).e. e c e Then the curvature . is called the “center of curvature” of y at p ( s ) : It is the center of the “osculating circle” at p ( s ) . Ciarlet Figure 2. the restriction of C to this neighborhood is the image f ( I ) of an open interval I c R. The intersection P n 2 is a curve on 2 . that is tangent to the normals to y. Consider a plane P normal to 2 = 8 ( w ) at the point = 8 ( y ) .E3) be an injective immersion. in a suficiently small neighborhood of y. which is the image = 8 ( C ) of a curve C in the set W. the limit as As + 0 of the circle tangent to y at p ( s ) that passes through the point p ( s As). # 0 .56 Philippe G. Assume that. As has a limit. Let y be a smooth enough planar curve. The point p ( s ) Ru(s). and let y E w be fixed.43).
the center of R curvature of the curve c^ at Q is the point (y^+ Ra3(y)). we note R that sinAq!(s) = v ( s ) . o f G at y (Section 2. Let P be a plane containing the vector .1) and Proof. + As).An Introduction to Differential Geometry in R3 57 Figure 2. . The algebraic bl(Y) A an(y)l 1 curvature . (i) We first establish a wellknown formula giving the curva1 ture . which is intrinsically defined in the Euclidean space E3. T ( S so that + As) = .of the planar curve C = P n ij = e ( C ) at 9 = e(y) is given by the ratio R a3(Y) = h A where a.43: Curvature on a surface.p(y) are the covariant components of the first and secdfa ond fundamental forms of the surface G at y^ and . If .# 0.{ v ( s + As) .of a planar curve.v ( s ) } . Using the notations of Figure 2.p(y) and b.42.(t)are the components of the dt 1 vector tangent to the curve C = f ( Z ) at y = f ( t ) = f " ( t ) e . T(S . which is normal to the surface w^ = O(w).
it suffices to note that R 1 by definition of the functions b.31 (b)) The knowledge of the curvatures of curves contained in planes normal any curve on i3.p(y) of the symmetric matrix (b. is given by ' 1 By (i). the curvature . such that (0 o f ) ( t ) = p(s) and (a3 o f ) ( t ) = v ( s ) for all t E f. The elements b.41.58 Philippe G. Stoker [1969. Ciarlet (ii) The curve ( 0 o f ) ( I ) . see.= (s) 1 du R ds where T(S). Chapter 4. Section 121.is R . There thus exist an interval f c I and a mapping p : J + P . More specifically..of R . where J c R is an interval. can be also parametrized by its curvilinear abscissa s in a neighborhood of the point g. s E J. R R' 1 at Q and a3(y) and .p and that (Theorem 2. e.41 are called the covariant components of the second fundamental form of the surface i3 = 0 ( u ) at y^ = O(y). the radius of curvature R at of any smooth enough curve coscp 1 C (planar or not) on the surface 6 is given by = .p(y)) defined in Theorem 2.which is a przori parametrized by t E I .g. T(S) = (S) dP ds = ____ 4 0 O f)(t) ds dt g To obtain the announced expression for .where cp is to i3 suffices for computing the cuFature of  e the angle between the "principal normal" to given in Theorem 2.
I ( P ) c R is an open interval and f ( P ) = f a ( P ) e . . and let c ( P )denote the intersection of P with G. For a fixed y E w. where dt ~ . they "stay away from infinity".there is a unique pair of orthogonal planes Rl(Y) Ra(Y)' Pl E P and P E P such that the curvatures of the associated planar 2 1 1 curves PI n G and P n G are precisely and 2 (c) Rib) RdYY Proof.(Y) := aP"(Y)bao(Y).2) and the matrix (b. Theorem 2.P E P. (a) Let the assumptions and notations be as in Theorem 2. the radius of curvature of the curve P n G is infinite for at least one such plane P. As a first step in this direction. (i) Let A ( P )denote the intersection of P E P with the tangent plane T to the surface G at Q.# . Note that this compact interval contains 0 if.41. and only if.An Introduction to Differential Geometry in B3 59 2. Rl(Y) R2(Y) (b) Let the matrix (bt(y)). T h e n the set of curvatures of the associated planar curves P n G. Q being the row index. we show that these curvatures span a compact interval of R. : I ( P ) + R2 is a smooth enough injective mapping that satisfies df"(P) (t)ea# 0. T h e n 1 1 0 .p(y)) is defined as in Theorem 2.is a compact interval of 1 1 R.41. where (a"P(y)) = (a. be defined by [ 1 bp. Gaussian curvature The analysis of the previous section suggests that precise information about the shape of a surface G = O(w) in a neighborhood of one of its points $ = O(y) can be gathered by letting the plane P turn around the normal vector a 3 ( ~and by following in this process the variations of ) the curvatures at Q of the corresponding planar curves P n G. consider the set P of all planes P normal to the surface G = O(w) at Q = O(y).51.p(y))l (Section 2.denoted .Hence A ( P ) is tangent to E ( P ) at Q E 2 .41.as given in Theorem 2. In particular then.5 Principal curvatures. In a sufficiently small neighborhood2f Q the restriction of the curve 6 ( P )to this neighborhood is given by C ( P )= (8 o f ( P ) ) ( I ( P ) ) where .
the curvature of the corresponding curves C = C(P) at takes the same values as where E" := df"(P) (t) and + (ii) Let the symmetricmatrices A and B of order two be defined by A := (a.. This proves (a). since Al = (a@(y)). Ciarlet = f(P)(t). Note that BA' = (bt(y)) with bi(y) := aoU(y)b.~ B c .311).(y) are linearly independent. respectively denoted and . dt Since the line {y pLJae. Since A is positive definite. Hence Theorem 2. When Q varies in R2. the relation c ( c .(y). Hence the ratio is nothing but the Rayleigh quotient associated with the symmetric matrix ClBC'. i.p(y)) and B := (b. it has a (unique) square root C. and the set of all lines supporting the nonzero tangent vectors to the curves C(P).e.~ = B A . P E P . Ciarlet [1982. Hence the relations in (b) simply express that the sum and the product of the eigenvalues of the matrix BAI are respectively equal to its . a symmebric positive definite matrix C such that A = C2.p(y)).of the matrix . there exists a bijection between the set of all lines A(P) c T .~ ) .60 t E I(P)is such that y Philippe G. this Rayleigh quotient thus spans the compact interval of R whose endpoints are the smallest and 1 1 largest eigenvalue. Hence the line A ( P ) is given by A ( P ) = { Q + X d ( e o ~ ( P ) ) ( t ) .{ 0 } .. Theorem 1. QI being the row index. see. } taea # 0 by assumption.2 shows that the eigenvalues of the symmetric matrix C'BCl coincide with those of the (in general nonsymmetric) matrix BA'.g. Rl(Y) R() 2Y ClBC' (for a proof. X ~ R ={Q+X["U.XEE%}.. F'urt hermore.p E R} is tangent to the curve C ( P ) := K 1 ( e ( P )at y E w (the mapping 8 : w + R3is injective by assumption) ) for each such parametrizing function f ( P ) : I(P) + R2 and since the vectors a.(~). e.~= ~c c.41 shoTs tha: when P varies in P ..
fjis 1 Rib) called an umbilical point. which may be also written as det(bC@ 9 ) ) ( d e t ( a 4(Y)) since BAl = (bt(y)).51 are called the mixed components of the second fundamental form of the surface i3 = O(w) at 9 = 8(y). i. the Rib) RAY) 1 1 point = 8(y) is called a planar point. the corresponding lines A(P1) and A(P2) of the tangent plane are parallel to the vectors Eya.An Introduction to Differential Geometry in R3 61 trace and to its determinant.(y) and &‘ap(y).respectively. If = . if all the points of i3 are umbilical. then i3 is a portion of a sphere. Stoker [1969. see. Hence 0 = qTq2 = T C Ct2 = (:At2 = 0. corresponding to the eigenvalues R (Y) 1 symmetric matrix ClBCl R2(Y)’ 1 and .# 0. p.with = ) ( : . i f all the points of i3 are planar. then i3 is a portion of a plane. e. Rl(Y) RZ(Y) . If __ .e. T since CT = C. be two orthogonal (qTq2 = 0) eigenvectors of the 1 . for all P E ‘P.Paa(d} = %a(Y)Ix. For proofs.g.. It is remarkable that.0. This proves (c).51 are called the principal curvatures of If1  Rib) R2(Y) Lj at the curvatures of the planar curves P n 2 are the RZ(Yl) 1 1 same in all directions.P = E T M 2 1 1 the directions of the vectors ql and q2 are If Rl(Y) # uniquely determined and the lines A(P1) and A(P2) are likewise uniquely determined. ma7 We are now in a position to state several fundamental definitions: The elements b t ( y ) of the (in general nonsymmetric) matrix ( b t ( y ) ) defined in Theorem 2. This proves (b). 991. By (i). 87 and p. (iii) Let q1 = and E2 = (:a) ($) rll = Ctl and q2 = (“a) V2 = C e z .. which are orthogonal since { I F % m } . {t. 1 1 The real numbers and (one or both possibly equal to 0) found in Theorem 2. Likewise.
As intuitively suggested by Figure 2. parabolic.g. the Gaussian curvature at a point can also be expressed solely in terms of the functions a.61. see. or < 0.4.2 in the next section. Then the two orthogonal lines tangent to the planar curves PI n ^and P nw^(Theorem w 2 2. point according as its Gaussian curvature is > 0. e.121... But quite surprisingly. according to the convention made in Section 2. See. through its first fundamental form alone.. An asymptotic line is a curve on a surface that is everywhere tangent to a direction along which the radius of curvature is infinite.6. the real numbers R l ( y ) and & ( Y ) are Rl(Y) RdY) 1 called the principal radii of curvature of 2 at If. are respectively called the mean curvature and the Gaussian.51. R I ~ Y )O’ = the corresponding radius of curvature R l ( y ) is said to be infinite. A point on a surface is an elliptic. the associated centers of curvature are intrinsically defined. c If # 0 and # 0.351 or do Carmo [1994. a surface in R3 cannot be defined by its metric alone. or total. curvature of the surface w^ at y^.5l).51 (c)) are called the principal directions at y^. any point possesses a neighborhood where two intersecting families of asymptotic lines can be chosen as coordinate lines. It can be shown that.62 Philippe G.g. Lemma 3..6.= 0 but it is not a planar point. in other words. i f all the points of a surface are hyperbolic. Klingenberg [1973. since its curvature must be in addition specified through its second fundamental form.41.. any point along an asymptotic line is thus either parabolic or hyperbolic. i. Klingenberg [1973. ( + ) . It can be shown that a point that is neither planar n o r umbilical possesses a neighborhood where two orthogonal families of lines of curvature can be chosen as coordinate lines. See. or hyperbolic.g.g.and 2 Rl(Y) R 2 ( Y ) R1 ( Y M 2 ( Y ) ’ principal invariants of the matrix ( b E ( y ) ) (Theorem 2. so named after Gauss [1828] and Bonnet [1848] (for a “modern” proof. Ciarlet Let = 8 ( y ) E w^ be a point that is neither planar nor umbilical. see Figure 2. Theorem 6.6. Lemma 3. Klingenberg [1973. A line of curvature is a curve on 2 that is tangent to a principal direction at each one of its points. 1 1 1 which are the The numbers .p and their derivatives! This is the celebrated Theorema Egregium (“astonishing theorem”) of Gauss [1828]. e. Chapter 6. see Theorem 2. the principal curvatures at S are not equal.e. While the principal radii of curvature may simultaneously change their signs in another system of curvilinear coordinates. e. Theorem 11): Let S be a 1 1 c. e. Another striking result involving the Gaussian curvature is the equally celebrated Gauss Bonnet theorem.
such as a sphere or a torus. and compact surface in R3 (a “closed” surface is one “without boundary”. A point is hyperbolic if the Gaussian curvature is < 0 or equivalently.An Introduction to Differential Geometry in R3 63 h A I I Figure 2. “orientable”. A point is parabolic if exactly one of the two principal radii of curvature is infinite. “closed”. smooth enough. . if the two principal radii of curvature are of different signs. if the two principal radii of curvature are of the same sign. A point is elliptic if the Gaussian curvature is > 0 or equivalently. the surface then intersects its tangent plane along two curves. the surface is again locally on one side of its tangent plane. the surface is then locally on one side of its tangent plane.51: Different kinds of points o n a surface.
Then K ( i 7 ) d W ) = 2 ~ ( . a developable surface is one whose Gaussian curvature vanishes everywhere. a sphere has genus zero. or a portion of a surface spanned by the tangents to a skewed curve. Section 5.11) together with the vector a 3 ( y ) = (which is normal t o i3 and has Euclidean norm one) form the covariant basis at y^. the only one of a developable surface all points of which are planar..6 Covariant derivatives of a vector field defined on a surface. The interest of developable surfaces is that they can be. and let Then the vectors a. Spivak [1999]. More details about these various notions are found in classic texts such as Stoker [1969]. a t least locally. 2 where the genus g(S) is the number of “holes” of S (for instance.64 Philippe G.71).2g(S)) is the Euler characteristic of According to the definition of Stoker [1969. see Stoker [1969. or “developed” (hence their name). Klingenberg [1973. continuously “rolled out”. Klingenberg [1973. x(s) 2.2 and 2. Klingenberg [1973]. Any developable surface all points of which are parabolic can be likewise fully described: It is either a portion of a cylinder. Berger & Gostiaux [1987]. which exclude for instance Klein bottles.. Developable surfaces are otherwise often defined as “ruled” surfaces whose Gaussian curvature vanishes everywhere. . are defined in. A portion of a plane provides a first example. or Kiihnel [2002]. onto a plane. e. s. see Figure 2. without changing the metric of the intermediary surfaces in the process.g.(y) (which form the covariant basis of the tangent plane to 2 a t ?jB(y). Chapter 5. where 8 : w c R2 + E3 is a smooth enough injective immersion. at least locally. Section 21. consider a surface i3 = 8 ( w ) in E3.4. Section 3. e.2g(S)). Chapter 5. while a torus has genus one). do Carmo [1976]. The integer defined by x(S):= (2 . 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. as in. the Gauss and Weingarten formulas As in Sections 2.51) and let K : S + R denote its Gaussian curvature. or a portion of a cone. Ciarlet “orientable” surfaces. Sections 2 to 61).g.
11) and a3(y)= b l ( Y ) A az(y)l the contravariant basis at 2. This means that q2(y)aZ(y) is the vector at the point Q.61).61.e.. Our objective in this section is to compute the partial derivatives &(qiui) of such a vector field. One way to define such a field in terms of the curvilinear coordinates used for defining the surface i3 consists in writing it as viai : w 4 B3.11) together with the vector u3(y) form the contravariant basis at 5.61: Contravariant bases and vector fields along a surface. An arbitrary vector field defined on G may then be defined by its covariant components qi : w 4 R. in specifying its covariant components qi : w + IK over the vector fields u formed by the contravariant bases. see Figure 2. the three vectors a2(y). u p ( y ) = 6.. Then the vectors u a ( y ) (which form the contravariant basis of the tangent plane at see again Figure 2. Suppose that a vector field is defined on the surface G.. . Note that the vectors of the covariant and contravariant bases at y^ satisfy uZ(y). c= Figure 2. At each point y^ = 8(y) E G = O(w). U j ( Y ) = 8. This means that qi(y)ui(y) is the z vector at each point 8(y) 6 (Figure 2. form basis of the tangent plane to ij at y^ (Figure 2.An Introduction to Differential Geometry in IK3 65 Let the vectors u a ( y ) of the tangent plane to G at be defined by the relations u " ( y ) . i. These are found in the next theorem. where a " ( y ) form the contravariant al(Y)A az(y) .
. a‘)a.61.5l).4 and 1. are thus denoted by the same symbols as the “threedimensional” Christoffel symbols introduced in Sections 1.66 Philippe G. Let w be a n open subset of R2 and let 6 E C2(w. = b. a‘)a. (b) Let there be given a vector field qiaa : w + R3 with covariant components qi E C1(w).5. r/3la := 8. where vpla := aaqp . = b“a cr’ a . implies in turn that = (&a3 . No confusion should arise. together with the definition of the functions b! (Theorem 2.p(aP .41 and 2. The ChristofSel symbols “on a surface” and the covariant derivatives of a vector field defined on a surface are also naturally introduced in this process..= b. Then Viai E C’(w) and the partial derivatives &(viai) E C o ( w ) are given by 8a(rliai)= (&qp  rzpqu bap773)aP+ (&q3  + btqp)a3 = (vpla: . This formula.p and bc of the second fundamental form of 2 are defined in Theorems 2.ap = I’&a. and since &a3 .bapr/3)ap + ( v 3 / a+ b t q p ) a 3 . viz. those of Gauss and Weingarten.C p q U and Proof. a p ) a p = ( c . and r a p T . where the covariant and mixed components b. + bapa3 and bapaP = = &a3 = &a3 = I’guaa + b i a 3 . a3 = .. & u p . a3)= 0). The Christoffel symbols “on a surface” introduced in this section and the next one. a g ) a .4l).a. it follows that &a3 = (&a3 . since &a3 is in the tangent plane (&a3 .p (Theorem 2. 0 Theorem 2. a p ) a P = bapaP. b. A word of caution. however.E3) be an immersion. (a) The derivatives of the vectors of the covariant and contravariant bases are given by 8.paPua. Since any vector c in the tangent plane can be expanded as c = (c .51 and FEp := a..773. u p = b. Ciarlet as immediate consequences of two basic formulas.
+ (&a0 and bag. 67 In a.ap . 01' since & a P .a3 and &aP = J?iUau bta3 + 0.61) a.4. E3). (viai). and the functions I'Eg := a. a0 = That viai E C'(w) if vi E C'(w) is clear since ai E C ' ( w ) if 8 E C2(w.61 is highly reminiscent of the definition of the covariant derivatives w i i l j = djwi rpjwpof a vector field defined on an open set O(R) given in Section 1.gaP = b"a 01 respectively constitute the formulas of Gauss and Weingarten. 0 The definition of the covariant derivatives v g = 8077. . = (c . by definition of &aP = ( & a p . the former are more subtle to apprehend than the latter. a3)a3= r&a. 0 The relations (found in Theorem 2. a3)a3 = riuau+ bEa3.pa3. recall that the covariant derivatives willj = djvi .ag = &a". ag are the Christoffel symbols of the second kind (the Christoffel symbols of the first kind are introduced in the next section). a j ) a j . + ba0a3 a.l?EPvuof a . ai)ai = ( c . &a3 = bEa.ap = rgpa.. + b. . l vector field defined on a surface 8 ( w ) given in Theorem 2. The Christoffel symbols rEg can be also defined solely in terms of the covariant components of the first fundamental form. The formulas established supra immediately lead to the announced expression of d. . . Remark. However. and = &a3 = b..I?~'wp may be also defined by the relations (Theorem 1.jgJ = aj(wigz).71. a. The functions (also found in Theorem 2.a B . au)a.An Introduction to Differential Geometry in R3 Any vector c can be expanded as c particular. To see this.)a" + (&ap . Finally. see the proof of Theorem 2. d.% and q3la = 801773 are the firstorder covariant derivatives of the vector field viai : w + R3.61) vgIa =  r".ag = (d. a3 = .42) vi.
1paa ) + b. The reason is that a surface has in general a nonzero curvature.e. in w.71. E {I.aa). vector fields viai : w + EX3 for which 773 = 0). where P denotes the projection operator on the tangent plane in the direction of the normal vector (i.77aa3 for such tangential fields by Theorem 2. their couariant derivatives qaip = doq.b.p + I'gubpp  r$b. As shown in the next theorem. Gauss' Theorema Ggregium It is remarkable that the components a. let 0 E C3(w. = bap = 0. and I?& have simple expressions in terms of the functions a.. Theorem 2. : w + E of the first and second fundamental f o r m s of a surface % e(u).. it so happens that the functions as defined in Theorem 2. Ciarlet By contrast... which are meant to hold for all a .. .bo7 =  .. they must satisfy relations that take the form:  aura. this explains why they are denoted by the same symbol). These relations. where the functions rap. P(ciu2)':= c.91 (b) reduces to aa(77iai)= (77i.e. since in this case b. cannot be arbitrary functions. the formula giving the partial derivatives in Theorem 2. c. manifesting itself here by the "extra term" b. = 0 in w.+ r$rorp r&J'pTp b. Note that. again in this case. p.p = bp. sob.7 Necessary conditions satisfied by the first and second fundamental forms: the Gauss and CodazziMainardi equations. 2}.. This term vanishes in w if D is a portion of a plane.b.61. since a p ( ~ a a " = 77.)Ui.p and of some of their partial derivatives (as shown in the next proof.. 2. r" apqu satisfy only the relations Valpa" = p {a/3(rlaaa)) . defined by a smooth enough immersion 8 : w + E3. even if only tangential vector fields qaaa o n the surface e ( w ) are considered (i. and let . and CodazziMainardi.71 coincide with the Christoffel symbols introduced in the previous section. equations.pb. Let w be a n open subset of R2. reT spectively constitute the Gauss.68 Philippe G.qaa3.a.E3) be an immersion.p = upa : w + R and b.
p + bapa3.. in W.rEsb.Differentiating the same since 8. = W .Tp + bapbur. A a2 and. &ar = r$a. .ap .p + r. Since &a3 = ( & a 3 . = 8. a3. where ( a u T ):= (u.. a3)a3. a')a. (&a.aura.rpTp b d p . .pb. duar = 17~pI'. = dpr. .pb. + (&up = %. the last relations imply that I'gp = &ap . + r$rUTp r:. .. Proof..   b. Hence the Gauss equations immediately follow.a. . &a. .An Introduction to Differential Geometry in R3 69 denote the covariant components of the first and second fundamental f o r m s of the surface O ( W ) ... Since we also have ap = uQBa. = 0 in w .. hence that aa = r z p a . dpraUT. relations yields %rap. It is then immediately verified that the functions raprare also given by ragr Let a = 3 la1 = & u p . := Z(8paar := uurl?. r$&. . +&up ..bp. &b. a... .up. a. and a3 = a3. a. W. a.baubpr. we also have dauap.ap = (&a. = apb.up = aapa. &. + bapa3 . Since &. Let a. so that the above relations together give &up . .... a.  Consequently. &a. for each y E A a1 2 let the three vectors a j ( y ) be defined by the relations aj ( y ) . a i ( y ) = 6 i .... we have + = bau = &a3 = bauau. Let the functions rapT E C ' ( W ) and rgPE C'(W) be defined by r a p . Then.~:J'p.)a' &a..pT 1 + &upT  &aap).O. a. necessarily. p .p)'. a3)a3and &a.  b.
pas and &a3 This relation and the relations &ap = b. Hence the CodazziMainardi equations immediately follow. the functions sTapu aprauTaUrap.up.6). We recall that the Christoffel symbols of the second kind also naturally appeared in a different context (that of covariant differentiation. . a3.up . and second.ap = & p a a . aT = & p a . Finally.p obtain 8. we Differentiating the relations b.r$b.p = a. + b. Since a.ap . the Gauss and CodazziMainardi equations thus simply constitute a rewriting of the relations &. The functions and = a U T a p r= d.ap . aT and &. cf. a = r. and the face Q ( u ) .up = & p a . &.2 and 2. u p . Remark. As shown in the above proof. The vectors a and u p introduced above respectively form .. in the form of the equivalent relations &.pa. a3 + 8. The definitions of the functions and r a p r imply that the sixteen Gauss equations are satisfied if and only if they are satisfied for a: = .ap . a 3 = i3. Consequently..3).. a3 = a.r:pruTp rzurpTP := + are the covariant components of the Riemann curvature tensor of the surface Q ( u ) .aU together imply that = Lieup ' &a3 = r$bup.70 Philippe G.a ' are the Christoffel symbols of the first. we also have a. a3 = apb. kind..baP I.. &a. u p . the covariant and contravariant bases of the tangent plane to the surthe unit vector a3 = a3 is normal to the surface.. + I'z. functions affP are the contravariant components of the first fundamental form (Sections 2.bpp.. . Ciarlet = &up a3. rZpa. Section 2. . b..
r = 1 in the Gauss equations gives in particular S1212 = det(b.and Rz(y) satisfy Theorem 2.det(aap(y)) function .72 constitutes the famed Theorema Egregium of Gauss [1828]. . the Gaussian curvature. viz. the Gauss equations and the CodazziMainardi equations in fact respectively reduce to one and two equations.det(bap(y) (Theorem 2. u = 1 and a = 1.. u = 2 (other choices of indices with the same properties are clearly possible).91212.An Introduction to Differential Geometry in R3 71 1. r ( r ~ ~ . . Let w be a n open subset of R2. p = 1.51). p = 2. we thus reach the astonishing conclusion that. at each point of the surface. the principal curvatures 1 1 .ape denote the covariant components of the first fundamental f o r m of the surface O(w). In other words. let uap = d.E3) be a n immersion. and let the functions rapT and S l 2 1 2 be defined by rap7:= Z(apaaT + &upT 1  dTaap). 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: 1 < Theorem 2.p) Consequently. viz. p = 2. u = 2. By inspection of the R l ( Y ) R 2 ( Y ) .B . Then.r11ar22B). a notion involving the “curvature” of the surface. r = 2 and that the CodazziMainardi equations are satisfied if and only if they are satisfied for a = 1. s~~~~ 1 2 a 1 2 := p R (Y) 1 1  &la22 .72. is entirely determined by the knowledge of the “metric” of the surface at the same point. so named by Gauss who had been himself astounded by his discovery.D = 2. let 8 E C3(w. at each point O ( y ) of the surface O(w). the Gaussian curvature at each point O ( y ) of the surface O(w) can be written as since .. 0 = 1. Letting a = 2.dz2all) + ~ ~ ~ .
p : w + R S and b.we now give a selfcontained. Whether an immersion found in this fashion is injective is a different issue. So far. this immersion is unique up to isometries in E3. Following Ciarlet & Larsonneur [2001]. and of all symmetric.p : w + R are the covariant components of the first and second fundamental forms of the surface 0(w) c E3. we have considered that we are given an open set w c R2 and a smooth enough immersion 0 : w + E3.p) : w + S2.e. which turn out to be remarkably simple.p) : w + : and (b.81. where a. S of all symmetric matrices of order two. the necessary conditions of Theorem 2. This proof has also the merit t o shed light on the analogies (which cannot remain unnoticed!) between the assumptions and conclusions of both existence results (compare Theorems 1..p) : w + S2. A direct proof of the fundamental theorem of surface theory is given in Klingenberg [1973.8. Note that the immersion 0 need not be injective in order that these matrix fields be well defined. This proof amounts t o showing that it can be established as a simple corollary t o the fundamental theorem o n flat Riemannian manifolds established in Theorems 1.72 Philippe G.61 and 1. the Gauss and CodazziMainardi equations. analogous to that used in part (ii) of the proof of Theorem . positive definite matrices of order two. when does there exist an immersion 0 : w + E3 such that If such an immersion exists. i.71 when the manifold is an open set in R3.8 Existence of a surface with prescribed first and second fundamental forms Let M2.71 and 2.81) and both uniqueness results (compare Theorems 1. We now turn to the reciprocal questions: Given an open subset w of R2 and two smooth enough matrix fields (a. Ciarlet 2.p) : w + : and (b. which accordingly should be resolved by different means. and essentially elementary. If w is connected.91). when are they the first and second S fundamental forms of a surface O(w) c E3. constitute the fundamental theorem of surface theory: If w is simplyconnected. complete. where the global existence of the mapping 0 is based on an existence theorem for ordinary differential equations. and : denote the sets of all square matrices of order two. i. proof of this wellknown result.. Theorem 3. to what extent is it unique? The answers to these questions.71. thus allowing us to define the fields (a.61 and 2. are also suficient for the existence of such an immersion.e.S2.
17zp := aurI'.p) E C2(w.Jacobowitz .. Theorem 2.6 .S.p).p = in w. Note that these two results are established under different assumptions on the set w and on the smoothness of the fields (a. is found in. e. These existence and uniqueness results together constitute the fundamental theorem of surface theory.araap). and on the theory of differential forms. + 8aagr .61 relies on the following elementary observation: Given a smooth enough . This result is another special case of the fundamental theorem of Riemannian geometry alluded to in Section 1. equations and that the corresponding isometric immersions are unique up to isometries in the Euclidean space.g.p = d. [1982]..61. which constitutes Bonnet's theorem. do Carmo [1976]. which usually rely on basic notions of Riemannian geometry. Proof. and CodazziMainardi.p.6. A substantial literature has been devoted to this theorem and its various proofs.p) E C 2 ( w . the latter being also called rigidity theorem.81.) and (b. 1 where (aur) := (a.81) and a uniqueness result (Theorem 2. viz.and Szopos [ZOOS]. such as connections or normal bundles. T h e n there exists a n immersion 6 E C3(w. Let w be a connected and simplyconnected open subset of JR2 and let (a.. a global existence result (Theorem 2. where rapr:= 5(dpa. Like the fundamental theorem of threedimensional differential geometry. E3) such that a.An Introduction to Differential Geometry in R3 73 1.p)'. this theorem comprises two essentially distinct parts.9I). The proof of this theorem as a corollary to Theorem 1.p) and (b. S 2 )be two matrix fields that satisfy the Gauss and CodazziMainardi equations. dp6 and b. We recall that this theorem asserts that a simplyconnected Riemannian manifold of dimension p can be isometrically immersed into a Euclidean space of dimension ( p + q ) if and only if there exist tensors satisfying together generalized Gauss. See in particular the earlier papers of Janet [1926]and Cartan [1927]and the more recent references of Szczarba [1970] Tenenblat [1971]. A proof of the "local" version of this theorem.
as consequences of the “twodimensional” Gauss and CodazziMainardi equations.p  2x3bap + x 3 c a ~ gi3 = hi3 in w x I&. . numbered (i) to (ix).p + x3cap and gi3 := hi3 in w x R 2 ..€[ + S. albeit sometimes lengthy.2x3b. let the matrix field ( g i j ) E C2(w x R. which accordingly will be omitted for the most part. computations. there then exists an immersion 0 : w x I&. provided of course that the iithreedimensional” suficient conditions of this theorem. viz.61.S) and (b. the proof is broken into nine parts. of course.p) E C 2 ( w . let the mapping 0 : w x 1E. see parts (i) and (viii) below). Ciarlet + E3 and E > 0. where aap and bap are the covariant components of the first and second fundamental forms of the surface 8 ( w ) and cap := aU7b. but the resulting difficulty is easily circumvented..E[ and let gij := 820 ’ 1818A 8281 ’ A ajo. Then the field ( g i j ) : w x I&. Assume that the matrices ( g i j ) constructed in this fashion are invertible. To avoid confusion with the “threedimensional” Christoffel symbols.x3) := 8 ( y ) where a3 := + x s a s ( y ) for all ( y . By Theorem 1.S 2 ) .”> becomes a natural candidate for applying the “threedimensional” existence result of Theorem 1..bp7. (i) Given two matrix fields (a. x 3 ) E w x 1E. + E3 O ( y .E[ . It thus remains to check that 0 := O(.S3) be defined by gap := aao . only the main intermediate results will be recorded. a j 0 in w x ]el &[. those “on a surface” will be denoted C& and Cap. and 2 E[ . E [ i E3 that satisfies gij = a i 0 .0) indeed satisfies (see part (ix)) The actual implementation of this program essentially involves elementary. & [ (they need not be. For clarity. over the set w x ]&.p) E C2(w. ajrikq akrijq rTjrkqpr Q j q P= o in R. Then an immediate computation shows that g a p = a. in this proof (and only in this proof). hence positive definite. +  can be shown to hold.74 immersion 8 : w be defined by Philippe G. That this is indeed the case is the essence of the present proof (see parts (i) to (vii)).61.
where (B)! := b! and (B")! := b . It is then easily verified that the functions h$ are given by h$ = ( n so that n)O + l)aag(Bn)!. the elements of the inverse matrix gap = are given in 20b y x(n+ n>O l)x. E o [ .a""(B"): and gi3 = di3.e. The above series are absolutely convergent in the space C2 Let a priori g"p = Cn20 x:hEp where h. bEnPl for n 2 2.. x3 designates the variable in R).E o . where cap := bLbpr and b: := auTbau in w . so that the relations gapgp7 = 6 read ."p are functions of y E GO only. where (gpq) Ro := wo x ] . (no). . E O ] ) if EO > 0 is small enough. Let wo be an open subset of R2 such that WO is a compact subset of w . : i. C2(W0x It is clear that such a series is absolutely convergent in the space [Eo. . Then there exists EO = E O ( W O ) > 0 such that the symmetric matrices (gij) are positive definite at all points in no.n 2 0. (B")! designates for any n 2 0 the element at the ath row and pth column of the matrix B".An Introduction t o Differential Geometry in EX3 75 (the variable y E w is omitted. Besides. .. (ii) The functions Cgp being defined by C& where := aaT Cap.
+ b g ( d p p + C&cup + CEucpp.la and baulp = baplu. (B"):.gij) and rfj:= gPqrij. = = = C p = c:p  c.x3cap. x z (B"+')E. the assumed CodazziMainardi equations imply that b7. (iii) The functions gij E C2(D0) and g i j E C2(Do)being defined as in part (i). We simply point out that the assumed CodazziMainardi equations are needed to conclude that the factor of x3 in the function r a p u is indeed that announced above.(.lp and bapiu. E C1(Qo) and E C1(no) by r i j q:= z(8jgiq + &gj.b0. = = ra33 r3p3 r33q 0.. where the functions cap. define the functions rij.x3cap. .1pbTu + C&cTu). raP3 ra3P baP .76 define the functions Philippe G. Capo  1  8. rt3= C n > O r..lpaTU 2C&bT. The above relations follow from straightforward computations based on the definitions of the functions b7. . All computations are straightforward.) + + Xi(b7. r i p= bap . We also note that the computation of the factor of xg in r a p u relies in particular on the easily established relations &cpo = b'.lp = b'. Then the functions expressions: rap.. and bLlp are defined as in parts (i) and (ii).+'bZlIp(B"):. x. They are recorded here because they play a pervading r61e in the subsequent computations. Ciarlet Furthermore. = Fijq = rjiq and have the following x3(b7. = rg3= 0.
in order that the relations hold. define the functions R q i j k E co((sz.j.a3rZpa + r.3rpap Ra2p3 = 0 satisfy in Ro.An Introduction to Differential Geometry in IR3 77 (iv) The functions rijq E C l ( n 0 ) and rfjE C'(n0) being defined as in part (iii). the relations R c u 2 p 3 = 0 imply that R q i j k = 0 if exactly one index is equal to 3. and finally. Then.. (v) The functions satisfy Ra3p3 = 0 in no.r. the relations R a 3 p 3 = 0 imply that R q i j k = 0 if exactly two indices are equal to 3. Note that neither the Gauss equations a? nor the CodazziMainardi equations are needed here. the relation R 1 2 1 2 = 0 implies that Rapg7 = 0. rijq (part The definitions of the functions gap (part (i)) and show that (iii)) . These relations immediately follow from the expressions found in part (iii) for the functions rijqand rp. (vi) The functions R a 2 P 3 := apr2301 .+rijq k r y ? k q p rYkrjqp. it is suficient that rijqand The above definition of the functions R q i j k and that of the functions I?$ (part (iii)) together imply that. Consequently. for all i.pr3cup .k . q .a.) by R q i j k := d j r i k q .
(vii) The function R~~~~ := satisfies  d2rzll+ rglrZlprg2rllp = 0 in Rlzl2 no. We simply record the main intermediary steps.~ 1 1 ~ 2 2 ) . Ra3p3 = 0 follow by making use of the relations dacpu = bilabur + b f 4 l a b p p + CLpcup + Cgucpp together with the relations d2bap  dab20 + C&b2u  CZpbau = 0.22 b 1 2 ~ 1 2+ b 2 2 ~ 1 i )+ xz(cf2 . The computations leading to this relation are fairly lengthy and they require some care. the expressions found in part (iii) for the functions r a p u and yield. after some manipulations: . which are special cases of the assumed CodazziMainardi equations. the expressions found in part (iii) for the functions r a p 3 easily yield r123r123rll3rZz3b f 2 =( + ~ ( b i i ~ .pr3aprg3rpapr.ar2pu r. which consist in evaluating separately the various terms occurring in the function R 1 2 1 2 rewritten as R~~~~ = (w221 + (r72r12u d2rzll) r71r22u) + (r123r123  ~113r223). First.3rapu =  = CZpbu  c. Ciarlet Then the expressions found in part (iii) show that r.78 Philippe G. z Second.pbau + and the relations ~ 3 ( b g l p b a u bzlpb2u  + Cipcau  C&c20).
By the assumed Gauss equations. (viii) Let w be a connected and simplyconnected open subset of R 2 .b\b\)}.An Introduction to Differential Geometry in R3 79 Third.hib22){blbl .b\b\). it is finally found that the function R1212 has the following remarkable expression: #1212 = {S1212 .C11C22 = (biibi2 . Then there exist open subsets we. which in particular make use of the relations established in part (ii) about the functions 6^/3 and bap\a.e 2 0. of R2 such that We is a compact subset of w for each e0 and w= . it is found that where the functions are precisely those appearing in the lefthand sides of the Gauss equations. after somewhat delicate computations. 51212 = 6ll&22 Hence R1212 = 0 as announced.(&11&22 " ^126l2)}{l .X3(b\ + 6) + x\{b\b22 . It is then easily seen that the above equations together yield Rl212 = {S1212 ~ — S3{ hxj{SaTl2blbT2 + (c12c12 Since C12C12 .
there exist a mapping p1E Co([O. 1 x [0.X) for all 0 6 t 6 1. a set Re := we x I .R2) is a loop in w since T ( 0 ) = rr(y(0)) = x(y(1)) = y(1) and y(t) E w for all 0 Q t 6 1.E ~ . a mapping y E Co([O.l]. by 1 definition of the set R.X):= (1. To show that R is Re simplyconnected. po(t.2) = yo for all 0 6 t 6 1 and p l ( t . i. R := U Re e>o is connected and simplyconnected.1) = 7 and p l ( t . 0 < X 6 1. where Re Finally. is a homotopy in R that reduces the loop y to the point ( y o l o ) E R. Let we. O Q X 6 1.E ~ . Let the projection operator 7r : R + w be defined by ~ ( y2 . 1 x [0..R3) defined by 1 p ( t .A) = pl(t. for Then pois a continuous mapping such that p o ( [ O .1] x [0.X)y(t)+ X7r(y(t)) all 0 Q t 6 I.3 ) = y for all (y. x3) E R.80 Philippe G.1].R2) and a point yo E w such that 1 p l ( t . Then the mapping p E Co([O.E [ [ .1] t R3 be defined by po(t. 1 x [0. 1 Q X 6 2. 1 6 X 6 2.O)= y ( t ) and p o ( t . Furthermore.A) for all 0 Q t 6 1. there exists Ee = &e(we) > 0 such that the symmetric matrices ( g i j ) are positive definite at all points in Ge.1]) c R. 1]. the open set := we x I . 2]. each C. X ) = po(t. p ( t . X )E w for all 0 6 t < 1.R3) that satisfies y(0) = y(1) and y ( t ) E R for all 0 6 t 6 1. Hence the set R is simplyconnected. let y be a loop in R . Ciarlet Furthermore. for The mapping y : = 7 r o y ECO([O. Since w is simply connected. for each C 2 0 . t 2 0. be open subsets of w with compact closures We C w such that w = Ue20~e. l )= 7r(y(t)) all t E [0.e. . It is clear that the set R := Ue20 is connected. I]. E ~ [ can then For be constructed in the same way that the set Ro was constructed in part (9. and let the mapping po: [0.
8 l = 1.E 3 ) such that @ ( Y . the functions 81 E rij. 8 p a 3 = . O ) for all y satisfies E w: Let gi := 8i0. gi3 = g i . a3. g. where the matrix field ( g i j ) E C2(R. = = a3 a3 := These relations imply that 8.B. 8 j 0 in R.An Introduction t o Differential Geometry in R3 (ix) B y parts (iv) to (viii). Then 8 3 3 0 = 83g3 = I'g3gp = 0. = 8. there thus exists a n immersion 0 E C3(R.8. B y Theorem 1. cf.2x3dap8. 8. part (i)) then show that + X ~ @ ( Y ) for all ( y . ' C1(R) constructed as in part (iii) satisfy in the connected and simplyconnected open set R .61.p  2~3b. ( 8 ~+ x38pa3) 8 = 8. Hence either 8l or (da8 + ~ 3 8 ~ 8 ' ) = 0 and 8l . The relations 0. where dl8 A 828 ld18 A 8281 ' But 8' = a3 is ruled out since ( 8 1 8 A 828). 8p8 . g 3 = &3 (cf. E C ' ( R ) and I' . gap = ( a d+ x d a a 3 ) . on the one hand. a3 = 0 implies 8.2 3 ) E 0.E 3 ) defined by 8 ( y ) = 0 ( y . T h e n the mapping 8 E C3(w. 8pa3 in R. 8l 8 1 = a3 in w . part (iii). 8l Noting that = det(gij)lz3.2 3 ) = O ( Y ) and consequently.p + x3cap and gi3 = 2 6i3 in R.8 + ~ 3 8 ~ 8 ' and g 3 = 8 l .8 4 3 .8 .0 > 0. E3) such that gij = 8iO.8. Hence there exists a mapping 8l E C3(w. a3 + xidaa3 . we obtain. .S3>) is defined by gap = a.
introduced in part (i) are the covariant components of the third fundamental form of the surface O(w).S2>) and LFoc(w. ape and b. Philippe G.p(w. [Ciarlet.82 Since..g.p = dap8. Remarks. that the existence of an immersion 8 E C3(w. (4) The Gauss equations are used only once in the above proof.S.p) can be significantly relaxed in several ways.) and (baa) E C o ( w . = d.p) and (b. 0 as desired. ) S2) matrix fields that satisfy the Gauss and CodazziMainardi equations in the sense of distributions.who established that if (asp) E Wll.plg introduced in part (ii) are the covariant derivatives of the second fundamental form of the surface 8(w). on the other hand.21. . In fact.a 3 in w. Ciarlet by part (i). to those of the spaces W:d.8."F(u. S and (hap) E Lgc(w.S2)for any p > 2..o(n + The regularity assumptions made in Theorem 2. for showing that R l 2 1 2 = 0 in part (vii). for details.81 on the matrix fields (a.'. all other assumptions of Theorem 2. Hartman & Wintner [1950] had already shown the stronger result that the existence theorem still holds if (a. Section 4..81 holding verbatim.p) are the fundamental forms of the surface 8 ( w ) . This completes the proof. 0 c.S2). Their result has been itself superseded by that of S. cf. E3) such that (a. Mardare [2005]. e.a3 .":(w. with a resulting mapping 8 in the space W.E3) still holds under the weaker (but certainly more natural.p = d. Naghdi [1972c (3) The functions b:lp and b.p) and (b. spa.. First. are two .g.p) E C'(w. in view of the regularity of the resulting immersion 8) assumption that (b. but not trivial. we conclude that a. Mardare [2003b].p) E C1(w. (2) The series expansion gap = l ) x ~ a a f f ( B "found in )~ part (i) is known. e.E3). who was able to further reduce these regularities. 2005.. see.with a resulting mapping 8 in the space C2(u. modification of the proof given here. (1) The functions cap = bzbp. S 2 ) . Cristinel Mardare has shown by means of an ad hoe. then there exists a mapping 8 E W."(w. E3). The last word in this direction seems to belong to S.
Let w be a connected open subset of R2 and let 8 E C2(w.E3) same fundamental forms. As shown by Ciarlet & Larsonneur [2001] (whose proof is adapted here). We now turn t o the question of uniqueness of such immersions. t o remind that it corresponds to the idea of a “rigid” one in E3.> gap := a. Arguments similar to those used in parts (i) and (viii) of the proof of Theorem 2. such that the symmetric matrices ( g i j ) defined by .An Introduction to Differential Geometry in JR3 83 2. where cap := a“‘b. if two immersions 2 E C 2 ( q E 3 ) and 8 E C2(u. T h e n there exist a vector c E E3 and a matrix Q E 0 such that : 8(y) = c + QZ(y) for all y E w.p 2 e and gi3 = bi3. Theorem 2. then the surface 8(w) is obtained by subjecting the surface g(w) to a rotation (represented by an orthogonal matrix Q with det Q = l).. the issue of uniqueness can be resolved as a corollary to its “threedimensional counterpart” .p = b. In addition. which constitutes another rigidity theorem. C 0. we have established the existence of an immersion 8 : w c IR2 + E3 giving rise t o a surface 8 ( w ) with prescribed first and second fundamental forms.81 show that there exist open subsets we of w and real numbers ~g > 0.p  in w. Proof. we let 0 := {Q E 0 3 .9 Uniqueness up to proper isometries of surfaces with the same fundamental forms In Section 2.p  and b. E3) and 2 E C2(w. Such a geometric transformation of the surface G(w) is sometimes called a “rigid transformation” .E~[. matrices of order three.p = a. like the issue of existence. d e t Q = 1) denote the set of all proper orthogonal . This observation motivates the terminology “rigidity theorem” . It asserts share the that. We recall that O3 denotes the set of all orthogonal matrices of order three.91. are positive definite in the set Q := U W e)o x IE~. . This is the object of the next theorem.bp.p . provided these forms satisfy ad hoc sufficient conditions.. called the rigidity theorem for surfaces. E3) be two immersions such that their associated first and second fundamental forms satisfy (with selfexplanatory notations) a.8.2x3bap + x3c.then by subjecting the rotated surface to a translation (represented by a vector c ) .
2 3 ) E 0.E3) defined by + 2303(y) and 6 ( y . a simple computation shows that ) wliere for all ( y . 2 3 ) = det Q det V 6 ( y . Therefore det Q = 1. + Remark. := aP“(Y)b*o(Y). 2 3 ) = c + Q&.23) E 0. 3 E w . 2 3 ) for all (y.91 constitutes the “classical” rigjdity theorem for surfaces. 0 Theorem 2.91 thus asserts that two immersions 8 E C1(w.z3) = det V 6 ( y . det VO(y. + Q G ( ~x3) for all (y.71) involves isometries of E3 that may not be proper.2 3 ) On the other hand. .71. which shows that the orthogonal matrix Q is in fact proper. Hence.84 Philippe G. 23) for all (y. there exist a vector c E E3 and an orthogonal matrix Q such that ~ ( yx 3 ) = c . q ) R therefore satisfy E = & in 0.E3) and (with selfexplanatory notations) O(y. so that det VO(y. By Theorem 1. The conclusion then follows by letting 2 3 = 0 in the relation @(y. in the sense that both immersions 8 and 8 are assumed to be in the space C2(w.E3) and is E C1 (w. E3). . on the one hand. E3) share the same fundamentaJ forms over a n open connected subset w of R3 i f and only if 8 = J+ o 8 . with c E E3 and Q E 0. 2 3 ) E R. 2 3 ) := 8(y) + 23&(y) gij for all (y. the “threedimensional” rigidity theorem (Theorem 1. where J+ is a proper isometry of E3. ~ E 0. 2 3 ) := 8 ( y ) 6 E C1(R. for all (y.. By contrast. Ciarlet The two immersions 0 E C1(R. A proper isometry of E3 is a mapping J+ : E3 + E3 of the form J+(z) = c Qoz for all 2 E E3. x 3 ) E R. Theorem 2.
92. .e. Mardare [2004a]. The proof essentially relies on the extension t o a Sobolev space setting of the “threedimensional” rigidity theorem established in Theorem 1. is nevertheless well defined a. in w and the assumption that 8 E C1(w.p = b. in w for an arbitrary mapping 5 E H1(w. we observe that the relations Zap = a.E 3 ) . Proof. in w. and  b. in w. in w for those mappings G that satisfy the assumptions of the next theorem.p = a. a w. E 3 )by .dpa3. E 3 )and G3 := 6 A62 1 1 1 A 621 6 E H1( w . we now show that a similar result holds under the assumptions that G E H 1 ( w . 3 ) is a n immersion together imply that E la1 A 6 2 1 = I d m Jdcto> o = 63.  . are well b. the vector field defined a.73.p = bp. in w .e. Let w be a connected open subset of R2 and let 8 E a 3 E C1(w. which is not necessarily well defined a.p a. E 3 ) . 8 p 6 3 are likewise well defined a. in w. 3 ) be an immersion that satisfies E Theorem 2.p a. Then there exist a vector c E E 3 and a matrix Q E 0 such that :  8(y) = c + Q8(y) for almost all y E w .e.E 3 ) .e.p := a. Naturally. in w. 3 ) . we establish that and thus the functions b. our first task will be to verify that the vector field 6 3 . 53 E H 1 ( w .An Introduction t o Differential Geometry in R3 85 As a preparation to our next result. means of the definition la1 A a21 b. .e.e. Assume that E there exists a vector field 5 E H 1 ( w . we record several technical preliminaries. E 3 ) and a3 = a1 A a 2 E C 1 ( w .  I C1(w. (i) To begin with.p.p := acu. Consequently.p = bp.e.p a. in w and b. Following Ciarlet & C. n Second. we note that the second fundamental form of the surface 8 ( w ) can still be defined under the weaker assumptions that 8 E C 1 ( w .. First. a.e. which evidently coincides with the usual one when 8 E C2(w. . 3 ) that satisfies E a.e. This fact will in turn imply that the functions b. a.E 3 ) (with selfexplanatory notations).
8pa3 = a p . To see this. imply that a.. Then we claim that . that a. Given any cp E D ( w ) . established.72.e. 8pa3 E D ' ( w ) .:= d.6) can thus be applied a. let U denote an open subset of R2 such that supp cp c U and TI is a compact subset of w . Ciarlet i.e. E 3 ) and a 3 E C 1 ( w . 23) Ew x R.92).Zap = cap a.e.e. in w . E 3 ) together. 0) 1 pa3 ) H A ( . .0 in . Third. . in w . To this end. (a.e.and the announced symmetries are .86 Philippe G.Zi3. or the assumptions 8 E H 1 ( w . cap aPZ3and cap := & a 3 .~3 ) 7 u we reach the conclusion that the expression (8. a 3 = 0 a. in w .e. Denoting by X ' (. &a3 in w and Zi.8 . we next construct a set R and a mapping 0 that satisfy the assumptions of Theorem 1. . we note that either the assumptions 8 E C 1 ( w . in w and that = H . we have V'(W) ' 8pa3i p ) v ( W ) = cpaCXO ' dy Observing that a. 8pZi3 = Zip ' daZ3 a. 8pa3. The formula of Weingarten (Section 2..  (ii) Starting from the set w and the mapping 0 (as given in the statement of Theorem 2. &a. showing that Cap = Zu7bu.p a. More precisely. we note that the matrix fields (Zap) := (Zap)' and ( a m p ) := (a. Then the mapping 0 := w x I + IE3 defined in this fashion is clearly R continuously differentiable on w x R and . let O(y.pO = 8p. dpa3 = ap0 .p a.e.D'(U). in w since 8 is an immersion and Zap = a.p)l are well defined and equal a. 2 3 ) := 8 ( g ) + 23a3(y) for all ( y .)xthe duality pairing between a topological vector space X and its dual X ' .p = b. in w . E 3 ) and a 3 E H 1 ( w .e.e.8. in w . .1 ( u . . let .. ~) 3 (8. in w . E 3 ) together. Hence 8 8 . in LtOc(w). 8pa3 = 8.8pa3 E LiOc(w). c p ) ~ ( ~is) symmetric with respect to Q and p since 8. hence that 8.e. @a.b7p a.0 . The assertion then follows from the assumptions b.
Hence 6 E H 1 ( R .e. in w by part (i).   (iv) By Theorem 1. E 3 ) since R c . l[.e. in R since Zap = a. < 1 (this Then it is clear that R is a connected open subset of R3 and that the mapping 0 E C1(R. I 8(y) + x363(y) = c + Q(B(y) + 23a3(y)) for almost all (y.e.e.E ~.72. 2 3 ) E zn x [. w x 11. are equal. 2 3 ) := 8(y) + 23?i3(y) for all (y. Likewise. established in (i) are used here) gap = a.p and the assumption that 8 is an immersion together imply that.e. Then the continuity of the functions a. .p + x$c. Finally. where 87 Let wnr n 2 0 ..p. we construct a mapping 0 that satisfies the assumptions of Theorem 1. in R since the functions bg := Zp"b.x3) E w x R. there exist a vector c E E3 and a matrix Q E 0 such that . for each n 3 0. we define a mapping 6 : R + IE3 by letting O(y. in w by assumption and Ep = cap . which are well defined a.92).x3) E R. det V 6 = det V O a. To this end.. Besides. n20 E. Let then R := (J(wn x ]En. be open subsets of R2 such that Zn is a compact subset of w and w = Un20wn. where the set R is defined as in (ii). in w .&.p  223b. to the functions bg. in w . note that the covariant components gij E Co(R) of the metric tensor field associated with the mapping 0 are given by (the symmetries b. asp.@ and b. > 0 such that det VO(y. again a. (iii) Starting with the mapping 8 . ga3 = 0. E3) satisfies det V O > 0 in R. there exists E. g33 = 1.[). 5 3 ) > 0 for all (y.p a.given in the statement of The(as orem 2. the components & E L1(R) of the metric tensor field associated with the mapping 0 satisfy & = gij a.p = bp.E ~ ] . there is no loss of generality in assuming that property will be used in part (iii)).72.p = b. Besides. 23) E R. b.p. a.An Introduction to Differential Geometry in W3 for all (y.e.
10 Continuity of a surface as a function of its fundamental forms Let w be a connected and simplyconnected open subset of R2.8.C 2 0. 2. If A E Ce(w.E3) and u3 E H 1 ( w . If f E Ce(w.K = SUP la"A(Y)I. To begin with. 1 denotes the matrix spectral norm .K := SUP { r$<e IW(Y)I I I ~ I I:= .E3) satisfying the assumptions of Theorem 2. we introduce the following twodimensional analogs to the notations used in Section 1.. E3). The notation K G w means that K is a compact subset of w .p).81 and 2. Theorems 2. Equivalently.R) or 8 E Ce(w. + Remarks. e :{: I : where 1. Let w be an open subset of W3.E3) and 6 3 E H 1 ( w .E3). (1) The existence of i? E H 1 ( w . is a surface a continuous function of its fundamental forms? When both spaces C2(w.. : A natural question thus arises as to whether the mapping F defined in this fashion is continuous. SUP ~ ~ &: : I{ . Hence 0 8(y) = c Q8(y) for almost all y E w as announced.p) E .S and (b. where (0. and K G w .88 Philippe G.(w. E3).(W. Ciarlet Differentiating with respect to 2 3 in this equality between functions in H1(!22. E3) and G3 E C1(w. iayy)i.€MI3)). and that 8 E H 1 ( w . we let C + \\f\\t. 2 0. (hap)) in the quotient set C3(w. (2) It is easily seen that the conclusion of Theorem 2. and K G w . Together.S2) and C3(w. a positive answer to this question has been provided by Ciarlet [2003] along the following lines.p) E C2(w.92 implies that i? E C1(w.E3)/R..E3) are replaced by the weaker assumptions i? E H:.S2) satisfying the Gauss and CodazziMainardi equations in w a welldefined element F((a. where d" stands for the standard multiindex notation for partial derivatives and 1.E3) are equipped with their natural F'rkchet topologies. ) C2(w.92 is still valid if the assumptions i? E H 1 ( w .E3).91 establish the existence of a mapping F that associates with any pair of matrix fields (a. we likewise let llAlle.s) R means that there exists a vector c E E3 and a matrix E Q E 0 such that 8(y) = c Q&) for all y E w . 1 denotes the Euclidean norm in the latter definition. E3) and Z3 E H.E3) shows that &(y) = Qu3(y) for almost all y E w . .
p C:. the space Ce(w.bp. ."(w. 8)= 0 . The continuity of a surface as a function of its fundamental forms has then been established by Ciarlet [2003]. O ) E R means that there exist a vector c E E3 and a matrix Q E 0 such that 8 ( y ) = c + Q G ( y ) for : all y E w.E3) by the equivalence relation R.p)) E C2(w.~)+o be any sequence cw and IF. the associated metric space is denoted { X . according to the following result (if d is a metric defined on a set X .. l l e .&b.0 converges to 8 with respect to this topology if and only if Furthermore.^ c int I F . n+m where Let C3(w. Let w be connected and simply connected open subset ofR2. G w .E3) := C3(w. Then a sequence (8". this topology is metrizable: Let of subsets of w that satisfy IF.C&rCprp = baubpT . Then the set C 3 ( w .e'chet topology defined by the family of seminorms l l .aucapr C&Curp .An Introduction to Differential Geometry in R3 89 We recall (see Section 1. IF.. .K 7100 =0 for all IF. Let C.E 3 ) / R denote the quotient set of C3(w.^ i=O Then lim I18n8jle. for any integers k 2 0 and d 2 1.S x S2):= {((a. where ( 8 .8) that. Gw e lim de(8". and w = I u 00 IF.. ~ +for all i 2 0.101.C$$. : )5 : S2).Rd) becomes a locally convex topological space when it is equipped with the &. E 3 )becomes a metric space when it is equipped with the distance d 3 defined by  where 6 denotes the equivalence class of 8 modulo R. (b.d } ) : Theorem 2.).^ (IF.bapbur in apb. K . = 0 in w } + W.p). apcaur . + .2 x C2(w.
WITTMORETTE DILLARDBLEICK DE & [1977]: Analysis. Academic Press. W. CHOQUETBRUHAT. PrenticeHall. + {C3(w.G. Universitext. Journal de 1’Ecole Polytechnique 19. BALL. DO CARMO. B. Springer. 1989).dz} : S defined in this fashion is continuous. E C3(w. GOSTIAUX.M. Anal. [1977]: Convexity conditions and existence theorems in nonlinear elasticity.337403. ANTMAN. 1971). x S2).1146. SpringerVerlag. SpringerVerlag. BONNET. [1848]: Memoire sur la theorie generale des surfaces. Rational Mech. Pura e Aplicada. R.p)) E C . Berlin (English translation of: Formas Diferenciais e Apli~$es. CIARLET.90 Philippe G.S. (b.P. . New York. [1927]: Sur la possibilite de plonger un espace riemannien donne E. Cambridge University Press.A. Berlin. Annales de la SocGtk Polonaise de Mathkmatiques 6. Paris (English translation: Introduction to Numerical Linear Algebra and Optimasation. J. [1982]: Introduction d 1’Analyse NumCrique Matricielle et d 1 ’Optimisation. [1975]: An Introduction to Differentiable Manifolds and Riemannian Geometry. [2003]: A Panoramic View of Riemannian Geometry.. x S ) let F(((a.d3 ) } 0 The above continuity results have been extended “up t o the boundary of the set w” by Ciarlet & C.E 3 . Berlin (Second Edition: 2004). [1994]: Differential Forms and Applications. Courbes et Surfaces. Arch. M. Instituto da Matematica. dans un espace euclidien. Mardare [2005]. [1975]: Soboleu Spaces. CARTAN. BERGER. M. [1987]: GComktrie Diffkrentielle: V a d G s . Masson. 0. Rio de Janeiro. S.M.E3) denote the equivalence class modulo R of any immersion 8 E C 3 ( u . References ADAMS. Ciarlet Given any element ((a. Academic Press. New York. ( u . 63. M. Paris. Presses Universitaires de France.p). NorthHolland. S . Manzfolds and Physics. DO CARMO.P. P. (hap))) ’. Amsterdam (Revised Edition: 1982). BERGERM.17. [1976]: Differential Geometry of Curves and Surfaces.p). Englewood Cliffs. Cambridge. BOOTHBY.E3) that satisfies T h e n the mapping F : { C ~ ( W . [1995]: Nonlinear Problems of Elasticity.
Arch.G. JAMES. [1989]: Differential Forms with Applications to the Physical H. NECAS. [2001]: On the recovery of a surface with prescribed first and second fundamental forms. LARSONNEUR. J. a function of its CauchyGreen tensor in L1...[1982]: The GaussCodazzi equations. 14611506. New York. Amer. NorthHolland. 171188. 83.. P. [1988]: Mathematical Elasticity.An Introduction to Differential Geometry in R3 91 CIARLET. Arch.R. [2004b]: Recovery of a manifold with boundary C. MARDARE. 1522. Elasticity (2005)). 81. J.G. NorthHolland.S. Commentationes societatis regiae scientiarum Gottingensis recentiores 6. J. Math. 811843. Nonlinear Sci. 415427. FLANDERS. . [2004c]: Continuity of a deformation in H 1 as C.G. tion of its CauchyGreen tensor. placements in shell theory. Volume I: ThreeDimensional P Elasticity. HARTMAN. MARDARE.. [2003]: Continuity of a deformation as a funcF. [2007]: An introduction to shell theory. 167185. H..F.G. [2004]: Riemannian Geometry. Math. [2005]: Recovery of a surface with boundary C. WINTNER. Pures Appl. CIARLET. CIARLET. Third Edition. [1828]: Disquisitiones generales circas superficies curvas. 253274. [2002]: A theorem on geometric G. P.G. Sciences. and its continuity as a function of its metric tensor. P. 13.G. P. CIARLET. P. MARDARE.G. Pures Appl. Models Methods Appl. 19. CIARLET. Springer. [2005]: A n Introduction to Differential Geometry with Applications to Elasticity. LAFONTAINE. Pure Appl. Analysis and Applications. Comm. J. CIARLET. 82.. Amsterdam. [2003]: The continuity of a surface as a function of its two fundamental forms. [2004a]: On rigid and infinitesimal rigid disC. Math. CIARLET. Amsterdam. JACOBOWITZ. MARDARE. MARDARE.D. Dover. 7879 of J . Math. 99117. LAURENT.S. CIARLET. Tensor. P. FRIESECKE. N. P. F. GALLOT.G. [2000]: Mathematical Elasticity. MULLER. . Berlin. J .G. CIARLET. and its continuity as a function of its two fundamental forms. Rational Mech. 14. placements in threedimensional elasticity. Pures Appl.. rigidity and the derivation of nonlinear plate theory from three dimensional elasticity. HULIN. 3.G.. P. Anal.. Math. Pures Appl.. 5 5 .553564. Volume 111: Theory of Shells. MARDARE.G. A.. GAUSS.D. 15891598. Springer. J. C. J. P. Rational Mech. CIARLET. 83. Math. P. [1987]: Injectivity and selfcontact in nonlinear elasticity.. 167. 115. Anal. [2003]: On rigid and infinitesimal rigid disC. P. 255269. S. 39. Dordrecht (also appeared as Vol. . Sci. CIARLET.. [1950]: On the embedding problem in differential P.. J.G. C CIARLET. geometry. Gottingen.G. CIARLET. 72. P. Math.
Pure Appl Math.J. dans un espace euclidien. Lecture L. [1972]: Bounds for deformations in terms of average strains. 44. JOHN. pp. T. Math.391413. MARDARE. [2003]: Mappings of domains in R" and their metric tensors. Berlin. Notes. Shisha.251290. Annals of Mathematics.G. SpringerVerlag. Vieweg & Sohn. [1983]: Mathematical Foundations of Elasticity.E. Berlin. [1994]: A Brief o n Tensor Analysis. Analysis and Applications 1. [1954]: C1 isometric imbeddings. [2004]: On isometric immersions of a Riemannian space with S. J. JOHN. NAGHDI. 6 0 . [1992]: Analyse 11: Calcul Diffe'rentiel et Equations Diffe'rentielles. 383396. Rational Mech. Arch. Physik. [1926]: Sur la possibilitb de plonger un espace riemannien donne M. Providence. to appear. KLINGENBERG. Academic Press. [1973]: Eine Vorlesung uber Differentialgeometrze. 8. Pures Appl. [1967]: Les Me'thodes Directes e n The'orie des Equations Elliptiques. [1961]: Rotation and strain.6985. (English translation: Differential Geometry: CurvesSurfacesManifolds. TRUESDELL. Annales de la Socie'te' Polonaise de Mathimatiques 5 .L. P.R. VIa/2 ( S . Siberian Math.F.G. New York University (Second Edition: American Mathematical Society. 193226. Vol. NASH. PrenticeHall. FLUGGE C. tions in differential geometry. J. RESHETNYAK. J. NIRENBERG..M. . RESHETNYAK. 1994). [2003b]: The fundamental theorem of surface theory for surfaces S. J . 14. J. Second Edition. Berlin... Berlin (English translation: A Course in Differential Geometry. Paris. [2005]: On Pfaff systems with L p coefficients and their applicaS. Wiesbaden W. [1972]: The theory of shells and plates. [2002]: Differentialgeometrie. Ciarlet JANET. F.R. Courant Institute. 78. MARDARE. MARDARE. KOHN. L.G. New 1 York. NECAS. in Inequalities 1 1 (0. American Mathematical Society. Providence. 332345. SpringerVerlag. Anal. 1978). SIMMONDS. 129144. SCHWARTZ. [1967]: Liouville's theory on conformal mappings under Y minimal regularity assumptions. 2002). in Handbuch der & Editors). Paris. SpringerVerlag.J. Englewood Cliffs (Second Edition: 1999).131172.J. W. 425640. J. tensor. Comm. with little regularity. MARDARE.92 Philippe G. SpringerVerlag.V. MARSDEN. Fried.433453. Sibirskii Math. Hermann. 3843. Hermann. Masson. Editor). [1966]: The'orie des Distributions. Elasticity 73. KUHNEL. [2003]: On the recovery of a manifold with prescribed metric C. pp. little regularity. Paris. SCHWARTZ. Y. [1974]: Topics in Nonlinear Functional Analysis. HUGHES. [1982]: New integral estimates for deformations in terms of their nonlinear strains. Analysis and Applications 2.
[1966]: Functional Analysis. John Wiley. I: FixedPoint Theorems. Boletim da Sociedade Brasileira de Matemcitica 2.[1971]: On isometric immersions of Riemannian manifolds. TENENBLAT. E. Vol.3145. Trans. SZOPOS. Berlin. [1986]: Nonlinear Functional Analysis and its Applications. SZCZARBA. YOSIDA. Amer. Math.6389. . K. Publish or Perish. Volumes I to V. Berkeley. K. STOKER. WHITNEY.An Introduction to Differential Geometry in R3 93 SPIVAK. 119143. J. M. R.J. Analysis and Applications 3. New York.H. in closed sets. [1999]: A Comprehensive Introduction to Differential Geometry. [1969]: Differential Geometry. 2336. SOC. [2005]: On the recovery and continuity of a submanifold with boundary. [1970]: On isometric immersions of Riemannian manifolds in Euclidean space. Third Edition.36. Boletim da Sociedade Brasileira de Matemcitica 1. ZEIDLER. [1934]: Analytic extensions of differentiable functions defined H. M. SpringerVerlag.
75005 France Email: mardare@ann. Dordrecht. Paris. from which the threedimensional theory of shells is obtained simply by replacing the reference configuration of a general body with that of a shell. Hong Kong.f r Introduction These notes’ are intended to provide a thorough introduction to the mathematical theory of elastic shells. The first part of this article is devoted to the threedimensional theory of elastic bodies. . which may be defined either over a threedimensional set or over a twodimensional set.jussieu. the writing of which was substantially supported by two grants from the Research Grants Council of Hong Kong Special Administrative Region. China Email: mapgcQcityu. Springer.94 An Introduction to Shell Theory Philippe G. The particular shape of the reference configuration of the shell does not play any rBle in this theory.hlc Cristinel Mardare Universite‘ Pierre et Marie CurieParis6 UMR 7598 Laboratoire JacquesLouis Lions. some portions of these notes are extracted and adapted from the book by the first author “An introduction t o Differential Geometry with Applications t o Elasticity”. City University of Hong Kong 8 3 Tat Chee Avenue. CityU 1006041. this theory is specific to shells. China [Project No. ‘With the kind permission of SpringerVerlag. 2005. edu. CityU 100803 and Project No. In contrast to the threedimensional theory. depending on whether the shell is viewed in its reference configuration as a threedimensional or as a twodimensional body (the latter being an abstract idealization of the physical shell when its thickness is “small”). The second part is devoted to the twodimensional theory of elastic shells. since it essentially depends on the geometry of the reference configuration of a shell. Ciarlet Department of Mathematics. Such a prediction is made either by solving a system of partial differential equations or by minimizing a functional. F. 9040869. Kowloon. The main objective of shell theory is to predict the stress and the displacement arising in an elastic shell in response to given forces. 9040966.
about distributions with derivatives in “negative” Sobolev spaces (Section 1.An Introduction to Shell Theory 95 For a more comprehensive exposition of the theory of elastic shells.6 study the existence and uniqueness of solutions to the equations of linearized threedimensional elasticity. Sections 1. the applied forces are “small”. we refer the reader to Ciarlet [18] and the references therein for the first part of the article. The equations of linearized threedimensional elasticity are obtained from the nonlinear ones by linearization with respect to the displacement field. or for that matter in any threedimensional elastic body. and to Ciarlet [20] and the references therein for the second part. This system is formed either by the equations of nonlinear threedimensional elasticity or by the equations of linearized threedimensional elasticity. the minimization problem associated with the equations of nonlinear threedimensional elasticity has a solution by a fundamental theorem of John Ball (Section 1.6 the fundamental Korn inequality. coerciveness. . If the constituting material is hyperelastic and the associated stored energy function satisfies certain conditions of polyconvexity. which fall into two distinct categories: If the data are regular. the equations of nonlinear threedimensional elasticity have a solution by the implicit function theorem (Section 1. In sections 1. 1 Threedimensional theory Outline In this first part of the article.21.L. and the boundary condition does not change its nature along connected portions of the boundary.8.51. in response to given loads are predicted by means of a system of partial differential equations in three variables (the coordinates of the physical space).8). which in turn implies that the equations of linearized threedimensional elasticity have a unique solution. the equations of equilibrium and the constitutive equations. due to J. we establish in Section 1. Using a fundamental lemma. Sections 1. we study the existence of solutions t o the equations of nonlinear threedimensional elasticity. The equations of nonlinear threedimensional elasticity are then obtained by adjoining appropriate boundary conditions to these equations.5).4 are devoted to the derivation of the equations of threedimensional elasticity in the form of two basic sets of equations. and growth.71. Lions. the displacement and the stress arising in an elastic shell.7).
w vectors in R3. a point x in space is defined by its Cartesian coordinates xi.). The physical space is identified with the threedimensional vector space R3 by fixing an origin and a Cartesian basis (e1. etc.The exterior product of two vectors u.e2. are real. w and with the Euclidean norm 11where u. The space M". The Kronecker symbol is denoted dl. The notation ( a i j ) designates the matrix in M" with aij as its element at the ith row and jth column. and some basic formulas All spaces.. The identity matrix in M" is denoted I := (6.1 Notation. matrices. The cofactor matrix associated with an invertible matrix A E M" is defined by CofA := (det A)AT. and its subspaces A" and S" are equipped with the inner product A : B and with the spectral norm (A1defined by i>j 1A) := sup{lAwl. where A = ( a i j ) and B = ( b i j ) denote matrices in M". we define the following spaces or sets of real square matrices of order n: M": the space of all square matrices. denote u. Let R be an open subset of R3. second. and third order are also denoted di := d / d x i . w E R". dijk := d 3 / 8 ~ l a ~ 2 a ~ 3 . E R3 is denoted w xi UAW. Sy : the set of all positivedefinite symmetric matrices. and + + .96 Philippe G.Partial derivative operators of order m 2 1 acting on functions or distributions defined over R are denoted where Ic = (ki)E N3 is a multiindex satisfying llcl := kl k2 k3 = m. The determinant and the trace of a matrix A = ( a i j ) are denoted det A and tr A.x2.e3). 0": set of all orthogonal matrices. Cristinel Mardare 1.In this way.23 or by the vector x := xiei. For any integer n 2 2. the 07:the set of all orthogonal matrices R E 0" with det R = 1. S": the space of all symmetric matrices. definitions. My: the set of all matrices A E Mn with det A > 0. dij := d2/dxidxj.The space R3 is equipped with the Euclidean inner product u . Partial derivative operators of the first.IwI < I}. Ciarlet. A": the space of all antisymmetric matrices.
denote the outer normal derivative operator along 8 R (since is Lipschitzcontinuous. and WmJ’(R) for any integer m 3 1 and any p 2 1. u The space WT1p(R)is the closure of D(R) in WmJ’(R) and the dual of the space Wrlp(R) is denoted WPm+”(R). Finally. The duality bracket between a distribution T and a test function cp E D(R) is denoted ( T . etc. where p’ = If the boundary of R is Lipschitzcontinuous and if r c 80 is a relatively o open subset of the boundary of R. where 8. The space of all indefinitely derivable functions cp : R 4 R with compact support contained in R is denoted D(R) and the space of all distributions over R is denoted D’(R). Y ) designates the spaces of all mappings from R into Y whose components in Y are respectively in Cm(R). a unit outer normal vector ( u i ) exists 80almost everywhere along 80. The usual Lebesgue and Sobolev spaces are respectively denoted Lp(R).. is defined as that consisting denoted Cm(R). we let 5. where the notation flu designates the restriction to the set U of a function f and is a compact set that satisfies . and the divergence of the same vector field is the function div v := Ci &vi.or simply C o ( X )if Y = R. = ui8i).Y is equipped with If em@. = o on w:. together with all their partial derivatives of order 6 m. LP(R) and Wm>P(n). and thus 8.An Introduction to Shell Theory The gradient of a function 97 4 IR is the vector field gradf := The gradient of a vector field v = (vi) : R 4 R” is the matrix field Vv := (djwi). we use the notation Hm(R) = Wm>’(R).). 8. the notation Cm(R. w:R :( ) := {f E w’J’(R). Mn. .oT’p(R) the space of is all mesurable functions such that flu E WmJ’(U) for all U G R.f = o on ro}. If R is bounded.LP(R. where i is the row index. em@).the space of all realvalued functions that are m times continuously differentiable over R is m 2 1. The space of all vectorvalued functions f E P ( R ) that. i is the row index. em@). the divergence of a matrix field T = ( t i j ) : R + M” is the vector field divT with components 8jtij)i.the notation U G R means that u c R.cp).}. If p = 2. the space C ” ( 2 ) equipped with the norm (aif). Y ) . is a Banach space.f := f= r. The space W. Y For any integer m 2 1 and any open set R c It3.P(~){f E w~J’(R). Y ) . Y ) and W”iP(R. possess continuous extentions to the closure of R. The space of all continuous functions from a topological space X into a normed space Y is denoted C o ( X . If Y is a finite dimensional vectorial space (such as R”. ) . where f :R (j= C”.
then the following formula of integration by parts is satisfied s.e. Any other configuration that the body might occupy when subjected to applied forces will be defined by means of a deformation. v da for all smooth enough matrix field F : R 4 M k and vector field v : R + Rk. Grisvard [54].. Y )are respectively equipped with the norms and Throughout this article. Cristinel Mardare the norm I ' 1 . the set R being locally on the same side of its boundary.98 Philippe G. I F : Vv d x + l a ( F n ). e. no interpenetration of matter occurs).2 Equations of equilibrium In this section. or NeEas [73]. The image is called the deformed configuration of the body defined +(a) . div F . Evans & Gariepy [47]). The notation da designates the area element induced on the surface dR by the volume element d x . det V + ( x ) > 0 for all z E 2)and injective on the open set R (i. that is. If R c R" is a domain. for such instances. Adams [2]. henceforth called the reference configuration of the body.g. then the spaces LP(R2. we begin our study of the deformation arising in an elastic body in response to given forces. a mapping that is orientation preserving (i.. See. a domain in R" is a bounded and connected open set with a Lipschitzcontinuous boundary.. see.Y )and W">P(R. Ciarlet.g. e. 1 (smooth enough means that the regularity of the fields F and k v is such that the above integrals are well defined. We also record the Stokes formula: 1..e. We consider that the body occupies the closure of a domain R c R3 in the absence of applied forces. v d x =  .
Then. a body fi c W3subjected to applied forces of densities f : fi t R3 and g : f’1 + R3 is in equilibrium if there exists a vector field z : {a}. or may not. It is sometimes more convenient to describe the deformed configuration of a body in terms of the displacement u instead of the deformation CP. the ones that are in “static equilibrium” in the presence of applied forces. R is the mass density in the deformed configuration.3). Our aim is thus to determine equations that a deformation CP corresponding to the static equilibrium of the loaded body should satisfy. called pressure. and g ( 5 ) = . s := {v 2 E R3. where id : + is the identity map. 5 denotes a generic point in n(5)is the unit outer normal to afi. The threedimensional equations of elasticity will then be obtained by combining these equations with a “constitutive equation” (Section 1. These examples illustrate that an applied force density may. according to the stress principle of Euler and Cauchy. where f’. 1v1 = l} denote the set of all unit vectors in R3. and T is a constant. let the applied forces acting on a specific deformed configuration fi := CP(Cl) be represented by the densities a a f : fi + R3 and g : f’1 + R3.x s 2 + R3 .where g is the gravitational constant.~ f i ( 5 ) . we first derive the “equations of equilibrium” from a fundamental axiom due to Euler and Cauchy. Let {a}. among all possible deformed configurations of the body. which is the vector field defined by u := CP . The “difference”between a deformed configuration and the reference configuration is given by the displacement.c i3fi is a relatively open subset of the boundary of fi. More specifically.id. 5 : fi + J .An Introduction to Shell Theory 99 by the deformation CP. notably when the body is expected to undergo small deformations (as typically occurs in linearized elasticity). Our objective in this section is to determine. To this end. If the body is subjected for instance to the gravity and to a uniform pressure on f ‘ ~then the densities and g are given by f(5) = g5(5)e3 . depend on the unknown deformation.
In other words. by means of Stokes’ formula (see Section 1.22. Proof. it can be conveniently reformulated in terms of functions defined over the reference configurence R of the body.21) is defined over the deformed configuration which is unknown. which is known. then i : (62)x S’J+ R3 is linear with respect to the second variable. /A f d2 + / LJA 2 A f d2 + LA i ( 2 . To this end. see. i i )= T(2)ii for all 2 E (62)and all ii E S’J. I f Z ( . =~) ?(z) E s3 for all 2E 6. The following theorem. and f : {6}.n ( 2 ) )dii = 0. (1.100 Philippe G. This axiom postulates in effect that the “equilibrium” of the body to the applied forces is reflected by the existence of a vector field i that depends only on the two variables 2 and i i ( 2 ) . . 0 {a}.the following equations of equilibrium in the deformed configuration: Theorem 1. The matrix field T : (62)+ M3 satisfies divp(2) = f(2) for all 2 E 6. Theorem 1. Combining Cauchy’s theorem with the stress principle of Euler and Cauchy yields.21. The system (1. Cristinel Mardare such that.1). for all domains A C 6. there exists a matrix field ?’ : {6}. where i i ( 2 ) denotes the exterior unit normal vector at Z to the surface aA (because is a domain.21) ? ‘ ( ~ ) i i ( g(2) for all 2 E Fl.g.+ R3 is of class C1 f o r all ii E S’J. shows that the dependence of on the second variable is necessarily linear: z i Z . which is due to Cauchy. i i ( 2 ) exists for d2almost all 2 E d A ) . + R3 is continuous f o r all 2 E {6}..) : Sz (. Ciarlet. Fortunately. e. we use the change of variables a. i i ) : {6}. 2 A i ( 2 .+ R3 is continuous. Ciarlet [18] or Gurtin & Martins [55]. The proof consists in applying the stress principle to particular subdomains in For details.i i ( 2 ) )dii = 0.+ M3 of class C1 such that i ( 2 . .
viz. assumed to be injective. ~ ( x:= T(+(x))CofV@(x)for all x E ) is called the first PiolaKirchhoff stress tensor field. V @(x) T (x) = V@ (x) [det V@x?( 9(x) V+ (x) which follows from the definition of T ( x ) and from the expression of the inverse of a matrix in terms of its cofactor matrix. we deduce that. a + M3 appearing in the second integral. or may not. g(5) dSi = g(x) da.. the above relation read which implies that the matrix field T satisfies the following partial differential equation: divT(x) Using the identity = f (x) for all x E R. f i ( 5 ) dSi = CofV@(x)n(x) da. ()' )] T. We also define the vector fields f : R + R3 and g : Fl f(5)d5 = f (x)dx. + +. depend on the unknown deformation First of all.2l). the fields f and g may. In terms of this tensor.and (with selfexplanatory notations) a a d5 = I det V@(x)/ dx. ' . + R3 by Note that. assuming that is smooth enough and using the change of variables 4j : 2 + (6)in the first equation of (1. for all domains A c R. like the fields f and g. f ( x ) dx The matrix field T : defined by + 1. that a.' T ( x ) ) is also symmetric. T(@(x))CofV+(x)n(x) da = 0.An Introduction to Shell Theory 101 5 = @(x) defined by the unknown deformation @ : + {fi}. we furthermore deduce from the symmetry of the matrix T(5) that the matrix ( V @ ( x ) . and the following formulas between the volume and area elements in {fi}.
for all z E rl. according to the following definition: A material is elastic if there exists a function Tn: x M t 4 M3 such that n ~ ( z= @(z.22) or (1. (1. a material is elastic if there exists a function Xfl : a x M $ S3 such that ~ ( z= @(z. . where the subset rl of a R is defined by = +(ri).102 Philippe G. (1. + n + 1.3 Constitutive equations of elastic materials It is clear that the stress tensor field should depend on the deformation induced by the applied forces. Either function Tnor E is called the response function of the mated rial. This dependence is reflected by the constitutive equation of the material.~ T ( ~ ) for all z E ) R. v+(z)) all z E ) for n.In order to determine these unknowns.22) E s3 for all z E R. by means of a response function. we will consider one class of such materials. In this article. specific to the material considered. Finally.23) has to be supplemented with an equation relating these fields. 4 Equivalently. either system (1. Then the equations of equilibrium defined in the reference configuration take the equivalent form div (V+(z)E(z))= f ( z ) for all z E R. let the second PiolaKirchhoff stress tensor field E : R + S3 be defined by ~ ( z:= V + ( Z ) . This is the object of the next section.and the stress field inside the body defined by the fields T : + MI3 or X : S3. v+(z)) for all z E ) a. Cristinel Mardare It is then clear that the equations of equilibrium in the deformed configuration (see eqns.21)) are equivalent with the following equations of equilibrium in the reference configuration: divT(z) = f(z) T ( z ) n ( z = g(z) ) V+(Z)~T(Z) for all z E R .23) in terms of the symmetric tensor field E. Ciarlet. The unknowns in either system of equations of equilibrium are the deformation of the body defined by the vector field : 2 + R3. (V+(z)E(z))n(z) g(z) for all z E = r1. (1.
An Introduction to Shell Theory
103
A response function cannot be arbitrary, because a general axiom in physics asserts that any “observable quantity” must be independent of the particular orthogonal basis in which it is computed. For an elastic material, the “observable quantity” computed through a constitutive equation is the stress vector field i. Therefore this vector field must be independent of the particular orthogonal basis in which it is computed. This property, which must be satisfied by all elastic materials, is called the axiom of material frameindifference. The following theorem translates this axiom in terms of the response function of the material.
Theorem 1.31. An elastic material satisfies the axiom of material frameindifference i f and only if
Tn(z,Q F ) = QTn(z, ) for all x E F
a and Q E 0 and F E Mi$, :
and Q E 0 and F E M : . :
or equivalently, if and only i f
Xn(x,Q F ) = Xn(x,F ) for all x E
The second equivalence implies that the response function E depend n on F only via the symmetric positive definite matrix U := ( F T F ) ’ l 2 , the square root of the symmetric positive definite matrix ( F T F )E S;. To see this, one uses the polar factorisation F = RU, where R := FU’ E 0 ,in the second equivalence of Theorem 1.31to deduce that :
En(,, F ) = Xfl(x,U ) for all z E
n and F = RU E M.:
This implies that the second PiolaKirchhoff stress tensor field E : + S3 depends on the deformation field @ : 32 + R3 only via the associated metric tensor field C := V@TV@, i.e.,
~ ( z = % z ~ ( z ) ) all ) (, for
z E a,
where the function
% : fi x S;
+ S3
is defined by
%(x, C ) := Xn(x,
for all z E
a and C E Sj3>.
We just saw how the axiom of material frameindifference restricts the form of the response function. We now examine how its form can be further restricted by other properties that a given material m a y possess. An elastic material is isotropic at a point z of the reference configuration if the response of the material “is the same in all directions”, i.e., if the Cauchy stress tensor $(Z) is the same if the reference configuration is rotated by an arbitrary matrix of 0 around the point z.An : elastic material occupying a reference configuration fi is isotropic if it is isotropic at all points of The following theorem gives a characterisation of the response function of an isotropic elastic material:
a.
104

Philippe G. Ciarlet, Cristinel Mardare
Theorem 1.32. An elastic material occupying a reference configuration R is isotropic i f and only if
@(z, F Q ) = Tfl(z, ) Q for all z E 2 and Q E 0 and F E M F : , :
or equivalently, i f and only if
Efl(z, Q ) = QTXfl(z, ) Q for all z E F F
and Q E 0 and F E M;. :
Another property that an elastic material may satisfy is the property of homogeneity: An elastic material occupying a reference configuration is homogeneous if its response function is independent of the particular point z E 2 considered. This means that the response function Tfl : 2 x M 4 M3, or equivalently the response function : Xfl : a x M : 4 S3,does not depend on the first variable. In other words, there exist mappings (still denoted) T n: M n+ M3 and Xfl : M 4 S3 L "4 : such that
a
T n ( z , F = T n ( F ) all x E ) for
and
n and F E M:,
E
X n ( x l F )= d ( F ) for all z E s2 and F
M .:
The response function of an elastic material can be further restricted if its reference configuration is a natural state, according to the following definition: A reference configuration is called a natural state, or equivalently is said to be stressfree, if
a
T # ( ~ , =) for all z E 2, I o
or equivalently, if
~ f l ( z ,) = I
o for all z E 2.
 We have seen that the second PiolaKirchhoff stress tensor field X : R 4 S3 is expressed in terms of the deformation field CP : 4 B3 as
X(z)
= %(z,C(z)), where
C(z) = VCP'(z)VCP(z) for all z E ;2.
If the elastic material is isotropic, then the dependence of X(z) in terms of C(z) can be further reduced in a remarkable way, according to the following RivlinEricksen theorem:
Theorem 1.33. If a n elastic material is isotropic and satisfies the principle of material frameindifference, t h e n there exists functions : R x B3 + EX, i E {1,2,3}, such that
,!
X(z) = +yo(z)I yi(z)C(z) yZ(z)c2(z) for all x
where ri(z)= $(z, tr C ,tr(CofC), det C ) .
+
+
E
2,
An Introduction to Shell Theory
105
Proof. See Rivlin & Ericksen [74] or Ciarlet [18].
0
Note that the numbers tr C(z), tr(CofC(z)), and det C(z) appearing in the above theorem constitute the three principal invariants of the matrix C(z). Although the RivlinEricksen theorem substantially reduces the range of possible response functions of elastic materials that are isotropic and satisfy the principle of frameindifference, the expression of X is still far too general in view of an effective resolution of the equilibrium equations. To further simplify this expression, we now restrict ourselves to deformations that are “close to the identity”. In terms of the displacement filed u : R + R3, which is related to the deformation @ : R + R3 by the formula
@(z)= z
+ ~ ( z for all z E 1, )
+ 2E(z),
the metric tensor field C has the expression
C ( X ) =I
1 E ( z ) :=  ( V u T ( z ) V u ( z ) V u T ( z ) V u ( z ) ) 2 denotes the GreenSt Venant strain tensor at z. Thanks to the above assumption on the deformation, the matrices E ( z ) are “small” for all z E and therefore one can use Taylor expansions to further simplify the expression of the response function given by the RivlinEricksen theorem. Specifically, using the Taylor expansions
where
+
+
a,
tr C(z) = 3 tr(CofC(z)) = 3 det C(z) = 1
+ 2 tr E ( z ) , + 4 t r E ( z ) + o((E(z)I), + 2 t r E ( z ) + o(lE(z)I), C2(z)= 1+ 4E(z) + o(IE(z)I),
tl and assuming that the functions yi are smooth enough, we deduce from the RivlinEricksen theorem that
where the realvalued functions X(z) and p ( z ) are independent of the displacement field u. If in addition the material is homogeneous, then X and p are constants. To sum up, the constitutive equation of a homogeneous and isotropic elastic material that satisfies the axiom of frameindifference must be such that X(X) = ~ t f ( z ,) I
+ X(tr E ( ~ ) )+I2 p ~ ( z+) o z ( l ~ ( z ) for) all z E 1 ~ .
106
Philippe G. Ciarlet, Cristinel Mardare
If in addition R is a natural state, a natural candidate for a constitutive
equation is thus
~ ( z= X(tr E )
( ~ ) )+ 2 p ~ ( z for all z E I )
a,
and in this case X and p are then called the Lam6 constants of the material. A material whose constitutive equation has the above expression is called a St VenantKirchhoff material. Note that the constitutive equation of a St VenantKirchhoE material is invertible, in the sense that the field E can be also expressed as a function of the field E as
~ ( z= E(x) ) 2P
1

v (tr E(Z))I for all z E 2.
E
Remark. The Lami: constants are determined experimentally for each elastic material and experimental evidence shows that they are both strictly positive (for instance, X = 106kg/cm2 and p = 820000kg/cm2 for steel; X = 40000kg/cm2 and p = 1200kg/cm2 for rubber). Their explicit values do not play any r81e in our subsequent analysis; only their positivity will be used. The Lami: coefficients are sometimes expressed in terms of the Poisson coeficient v and Young modulus E through the expressions X P(3X + 2cL) v= 2 ( X + p ) and = X+P
'
1 4 The equations of nonlinear and linearized . threedimensional elasticity
It remains to combine the equations of equilibrium (equations (1.23) in Section 1.2) with the constitutive equation of the material considered (Section 1.3) and with boundary conditions on r := d R \ I ' l . Assuming o that the constituting material has a known response function given by T uor by Xu and that the body is held fked on ro, we conclude in this fashion that the deformation arising in the body in response to the applied forces of densities f and g satisfies the nonlinear boundary value problem: divT(s) = f(z), z E R, a.= 2, () 2 E r,,, (1.41) T ( z ) n ( z = g ( z ) , 5 r1, )
An Introduction to Shell Theory where
107
T ( z )= Tn(z,V@(z)) V+(z)Xn(z,v@(z)) all z = for
E
D.
(1.42)
The equations (1.41) constitute the equations of nonlinear threedimensional elasticity. We will give in Sections 1.7 and 1.8 various sets of assumptions guaranteeing that this problem has solutions. Consider a body made of an isotropic and homogeneous elastic material such that its reference configuration is a natural state, so that its constitutive equation is (see Section 1.3):
where X > 0 and p > 0 are the Lam6 constants of the material. The equations of linearized threedimensional elasticity are obtained from the above nonlinear equations under the assumption that the body will undergo a "small" displacement, in the sense that
+(z) = z
+ u ( z )with IVu(z)l << 1 for all z E 2.
Then, for all z E 2,
and
~ ( z= ~ + ( z ) ~ =z I ) ( ()

+ Vu(z))(X(tr E ( ~ ) )+I2 p ~ ( z ) ) X = tr(VuT(z) + ~ u ( z ) )p ( ~ u ' ( z )+ vu(z))+ oz(Ivu(z)l). + 2
Therefore the equations of linearized threedimensional elasticity, which are obtained from (1.41) by replacing T ( z )by its linear part with respect to Vu(z), are given by diva(z)
= f(z), z E = 0,
5
R,
(1.43)
a(++)
where
= g(z),
ro, 2 E rl,
E
~ ( z= X(tre(z))I+ 2pe(z) and )
e(z) = (Vu'(z)
1 2
+ V u ( z ) ) . (1.44)
We show in the next section that this linear system has a unique solution in appropriate function spaces.
108
Philippe G. Ciarlet, Cristinel Mardare
1.5
A fundamental lemma of J.L. Lions
We first review some essential definitions and notations, together with a fundamental lemma of J.L. Lions (Theorem 1.51). This lemma will play a key r61e in the proofs of Korn's inequality in the next Section. Let R be a domain in IW". We recall that, for each integer m 1,H"(R) and H,"(R) denote the usual Sobolew spaces. In particular,
H1(R) := {w E L2(R); aiv E P p ) , 1 6 i 6 n } , H y n ) := { w E H1(R); &jV E L2(R), 1 < 2, j ,< n } ,
where aiv and aijv denote partial derivatives in the sense of distributions, and H,~(R):= { w E ~ ' ( 0 ) ;= o on r}, w where the relation w = 0 on I' is t o be understood in the sense of trace. The norm in L2(R) is noted II.IIL2(n) and the norm in H"(R), m 2 1, is noted II.IIHm(n). In particular then,
We also consider the Sobolev space
H  l ( R ) := dual space of HA(R).
Another possible definition of the space HA (0)being
&(R) = closure of D(R) with respect to
II.IIHl(n),
where D(R) denotes the space of infinitely differentiable realvalued functions defined over R whose support is a compact subset of R , it is clear that
w
(., .)):
E
L2(R)
+ w E Hl(S2)
and &w E H'(R), 1 6 i 6 n,
since (the duality between the spaces D(R) and D'(R) is denoted by
I(v,cp)I =
I J' vcpdzl 6
n
R
//~llL2(n)llcpllH1(R),
I(aiv,cp)I = I 
haicp)I =
1 J'

ua,cpdz( 6 ll~IIL~(n)IIcpIIHl(n)
for all cp E D(R). It is remarkable (but also remarkably difficult to prove!) that the converse implication holds:
4) as a solution to the variational equations (1. Amrouche & Girault [6. Geymonat & Suquet [52].61) for all smooth vector fields v : R + R3 that satisfy w 1 2 =0 on ro. because the matrix field u is symmetric. 1 < i < n } ==+ u E L2(R). 8iv E H’(R). i. where u = X(tr e ( u ) ) I 2pe(u) and e(u) = (VuT + +Vu). p. the integrand in the lefthand side can be also written as u : V v = u : e(v). j v E Hmfl(R) holds for arbitrary integers m E Z. In order to verify the hypotheses of this lemma. Then {u E Hl(R) and.51. Let R be a domain in Rn and let u be a distribution o n R. 1 <i 6 n} .( v v T V W ) . we first need to establish the following classical. Proposition 2. Lions. 0 This implication was first proved by J. as stated in Magenes & Stampacchia [66. 320. and fundamental. where e(v) := .101 even proved that the more general implication {w E D’(R) and 8 i w E H”(R). p. inequality: 1 + . Note (27)].L. with Lipschitzcontinuous boundaries. 1 1 . are given in Bolley & Camus [14]. Various extensions 11 to “genuine” domains. 1. and Borchers & Sohr [15].L. for this reason. another proof was also given by Tartar [84].An Introduction to Shell Theory 109 Theorem 1.e. 2 The existence of a solution to the above variational problem follows from the LaxMilgram lemma. Note that. Its first published proof for domains with smooth boundaries appeared in Duvaut & Lions [46. Lions. it will be henceforth referred to as the lemma of J.6 Existence theory in linearized threedimensional elasticity We define a weak solution to the equations of linearized threedimensional elasticity (Section 1. .
$3)2 CIITJ II H1 (n. (ii) The space E(R.. R3) H v E E ( R .g. such that ll4l L2(0.R3) equipped with the norm ll4H 1 := 112) I I L z ( n .W3):= (Iv I/L 2 ( 0 .where that If v E L2(f12. This implies that the space E(R.~) all for E E(R. R3) is clearly linear.w3). R 3 ) + 11 11 L2(n. R3).R~).R3) 7 is clearly a Hilbert space. R ~ := {v ) E ~ ~ ( 0 2 (.110 Philippe G. Several proofs are available in the mathematical literature for this remarkable inequality. T h e n there exists a constant C o such that II 4. Proof. We proceed in several steps: (i) Korn's inequality is a consequence of the identity d i j w k = &ejk(v) + ajeik(v)  &eij(v) relating the matrix fields Vv = (ajwi) and e(v) = (eij(v)).R3) for d i E H & ( R .S3).61 (Korn's inequality).w3) 6 cllvIIE(n. bijective (thanks to the step (i)). there exists a constant C such that 114~l(n. o on ) = ro}.R .) IIP(n.L. )~E ~ ~ (. the open mapping theorem (see.51) shows that 8 j V k E L2(R). equipped with the norm 11 11 E(n. e. ) (Q. the lemma of J. Lions (Theorem 1. We adapt here that given in Duvaut & Lions [46].the above identity shows & j w k E Hl(R). Let R be a domain in R3 and let r c 8 R be such that arearo 0 . and continuous. Since the identity mapping id : v E H1(R.R3) . R ~w) .Iw3) + 1 4 I1L2(n. R ~ := { T JE H ~ ( R . Cristinel Mardare Theorem 1. ~ 3 ) t II v v II L2(n. Yosida [87]) shows that id is also an open mapping. Therefore. as is the space H1(R. R3). or equivalently. Ciarlet.83) 2 c1114H1 1. e w L ) S3)} coincides with the Sobolev space H'(R.R3) and e(v) E L2(R2. Since the functions a j v k also belong to the space H'(R).
R3) that satisfies e ( w ) = 0 must also satisfy : n &jv. Therefore.R3) that such + wu(n) 4 w in H ~ ( R .R3)is compact by the RellichKondrasov theorem.61 then follows by contradiction. there would exist a sequence ( w n ) n E ~ in H (R.S3).R3). Therefore the sequence ( ~ ~ ( is a Cauchy sequence with ~ 1 ) respect to the norm 11 . which shows that any field w E H (R.R3) and L2(R.An Introduction to Shell Theory 111 for all w E H1(R. R3). (iii) We establish that.R3) by the inequality established in Part (ii). But this contradicts the relation = 1. e. l l ~ ( n . = 0 in R. p )hence with respect to the norm 11 . the locus of all points z satisfying a + Ax = 0 is either a line in R3 or an empty set. = no3 it follows that w = 0 by Part (iii). which is precisely e ( w ) . and the proof is complete. Since its limit satisfies e(w) = lim e(vucn)) 0. i.e. where b E R3 and A E M3. the matrix A must be in addition antisymmetric. vanishes in 0. Hence A = 0 and b = 0. and thus w = 0 in R. If the inequality were false. there exists w E Hio(R. N is an increasing function. the field w must be affine. that converges in L2(R. R ~ ) . R3) being complete as a closed subspace of H1(R. This is a consequence of the identity of Part (i). The space H. IIHl(n. depending on whether the linear system a Aa:= 0 has solutions or not.[. Since the sequences ( w ~ ( ~and (e(w. But a Ax = 0 on r and arearo > 0. 11wIIHl(n. contains a subsequence ( w ~ ( ~ where a : N it )).then w = 0. Since the symmetric part of the gradient of w .R3) satisfies e(w) = 0.o (0. if w E H&(R. lle(2'n)llL2(n. of the form w(z)= b Aa: for all IC E R . by a classical result about distributions (see.R3) c L2(O. Since the rank of a nonzero antisymmetric matrix of order three is necessarily two. Schwartz [80]).p) = limn+oo 0 .. R3)such that : n + + + = 1 for all n. The sequence (on) being bounded in H 1 ( QR3).they are Cauchy sequences in the same spaces.))) converge respectively in )) the spaces L2(R.g. . the inclusion H1(R.s13) + 0 as n + Because the set R is a domain. o (iv) The Korn inequality of Theorem 1.R3) as n + m.
JR.. The uniqueness result established in Part (iii) of the proof is called the infinitesimal rigid displacement lemma.62.TR3).61) is a consequence of Korn's inequality established in the previous theorem combined with the positiveness of the Lam6 constants. a vector field w € H1(R. Korn's inequality is a l 1VvI2d z for all w E H. for all w E H.112 Philippe G. the variational problem (1. 0 We are now in a position to establish that the equations of linearized threedimensional elasticity have weak solutions. R3) satisfying e(w) = 0. We distinguish two cases depending on whether arearo > 0 or not.b E R3 Remark. i.1(R. the coerciveness of the bilinear form appearing in the lefthand side of the equation (1.61). which together imply that. In the special case where I'o trivial consequence of the identity Ie(w)12 z d = = dQ.E%3).R3) and g E L4/3(l?l. In particular. where a.61) has a unique solution f in the space Proof. p Theorem 1. Ciarlet.1). Assume that the Lame' constants satisfy X 3 0 and > 0 and that the densities of the applied forces satisfy f E L6//5(R. . I areal70 > 0. R3). Cristinel Mardare The inequality established in Part (ii) of the proof is called Korn's inequality without boundary conditions. itself obtained by applying twice the formula of integration by parts (see Section 1. It suffices to apply the LaxMilgram lemma to the variational equation (1. It shows that an infinitesimal rigid displacement field. since all its assumptions are clearly satisfied.e. is necessarily of the form w(z)= a + b A x for all z E 0.
+ + Hence there exists an increasing function a : N 4 N such that the subsequence ( w ~ ( ~is )a Cauchy sequence in H1(R.wda=O f o r a l l w e H 1 ( R .R3)~~ G I I W ~ ( ~) wIIH1(n2.61) is then established as a consequence of another Korn's inequality: lle(qllLyn. then the variational problem (1.61). with Part (iii) adapted as follows: The sequence ( w n ) n E ~ is now defined in H1(R.R3). where I& is the subspace of H1(R. 7100 Therefore w E This implies that by Part (iii). By the infinitesimal rigid displacement lemma (see Part (ii) of the proof of Theorem 1. Ro is the finitedimensional space {w : R + R3.61) has a solution in H1(R.61) is well defined over the quotient space H1(R. = lle(+n)llLyn. This space ) being complete. unique u p to a n infinitesimal rigid displacement field. R 3 ) satisfying e(w) = 0 .63.s3) 0 as n m.w) ) ~ + 0 in H1(R. The proof of this inequality follows that of Theorem 1. and its limit satisfies e ( w ) = lim e(wa(n)) = 0 .R3) and g E L4/3(dR. Il~a(n)IIH~(n.wdJ:+Jang.61).R3).R3).R3)/Ro 1 for all 12.An Introduction to Shell Theory 113 Theorem 1. R3)/& and satisfies /l+nIIH1(R.R3). . Assume that the Lame' constants satisfy X 3 0 and p > 0 and that the densities of the applied forces satisfy f E L6/5(R.R3).R3)/I&. hence ( w ~ ( . this time defined over the quotient space H1(R. w(x) = a + b A 2. It is again based on the LaxMilgram lemma applied to the variational equations (1. R3) such that wa(") + w in H1(R.b E R3}.61.s3) 3 Cll4lH1(n2.R3) consisting of all the infinitesimal rigid displacements fields. The compatibility relations satisfied by the applied forces imply that the variational equation (1. there exists w E H1(R.W3)/Rg all for + E H1(QR 3 ) / I & . which is a Hilbert space with respect to the norm The coerciveness of the bilinear form appearing in the lefthand side of the equation (1.R3) + 0 as n 00. a.R3)/I&. Sketch of proof. Ifarearo = O a n d J n f.
R3).61) is in the space W"f2~p(0. Ciarlet.W3)and there exist a constant C such that sn +. and dR arc regular and if there is no change of boundary condition along a connected portion of dR. w da = 0 for all vector fields v E H1(R. the following regularity results hold (indications about the proof arc given in Ciarlet [18. Theorem 1. Assume that rl = dR and f .61) is in the space Wmf2J'(C12.R3). x E 80. Assume that ro = 80.g E W"+11/PJ'(I'~. x E 0.g . ern+' Furthermore. Furthermore. and dR is of class for some integer m 2 0 and real number 1 < p < co satisfying p 2 &.7 Existence theory in nonlinear threedimensional elasticity by the implicit function theorem The question of whether the equations of nonlinear threedimensional elasticity have solutions has been answered in the affirmative when the . x E dR.114 Philippe G. fT(x)n(x) = g(x). Theorem 1. x E R.61) is called a pure displacement problem when r = do.s . More specifically. then the solution u to the variational equation (1. Theorem 6. 0 Since the system of partial differential equations associated with the linear threedimensional variational model is elliptic.P(n.)llHi(n. R3) satisfying e(w) = 0 .p)/~~1 for all n. u satisfies the boundary value problem: divu(x) U) ( .W3)and there exists a constant C such that ll~IIWm+2. 1. The variational problem (1. = 0. and o a displacementtraction problem when arearo > 0 and areal71 > 0. R3) and 80 is of class C m f 2 for some integer m 2 0 and real number 1 < p < co satisfying p 3 &. Cristinel Mardare = which contradicts the relation ((i~.64 (pure displacement problem).(. a pure traction problem when rl = dR.W3) G CllfllW~+2~P(n. u satisfies the boundary value problem: divu(z) = f(x). Iff E W">p(R.R3).361). = f. then any solution u to the variational equation (1.65 (pure traction problem). I f f E W"J'(R. we expect the solution of the latter to be regular if the data f. w dx g .
The existence result is then the following Theorem 1. there are two theories of existence.8).An Introduction to Shell Theory 115 data satisfy some specific assumptions.72) has a solution u E W2>p(R.71. and the applied forces are small enough. w = o on an}. T h e nonlinear boundary value problem (1. := (w E .Hence the equations of nonlinear threedimensional elasticity assert that the displacement field u : R 4 R3 inside the body is the solution to the boundary value problem :R f > 0 and p > 0 are the Lam6 constants of the material and R3 is the unknown displacement field.R3) is small enough.e.VuT Vu 2 where X I J Y + + VuTVu . f E Lp(R.71). (1. and one.R3) and IlfllLp(n.e. or nowhere along their boundary (i. but remains open in the other cases. The first requirement essentially means that the bodies are either held fixed along their entire boundary (i. based on the minimization of functionals. R3).72) where the field X is given in terms of the unknown field u by means of (1. i f R is a domain with a boundary dR of R3) class C2.. = dR). The existence theory based on the implicit function theorem asserts that the equations of nonlinear threedimensional elasticity have solutions if the solutions to the associated equations of linearized threedimensional elasticity are smooth enough. Proof. R~). ~ ) u = 0 on dR. rl = 80).71) (the case where rl = dR requires some extra care because the space of infinitesimal rigid displacements fields does not reduce to (0)). one based on the implicit function theorem. Y := U ( R . For the existence theorem based on the minimization of functionals we will only sketch of the proof of John Ball (Section 1. Define the spaces x w2>p(R. We restrict our presentation to the case of elastic bodies made of a St VenantKirchhoff material. and f o r some p > 3. To this day. we assume throughout this section that X = X(tr E ) I + 2pE and E = . We state here the existence theorems provided by both theories but we will provide the proof only for the existence theorem based on the implicit function theorem.71)(1. ) (1. In other words. due to John Ball. We assume that r = dR o div ( ( I+ V U ) = f in R..
g. It remains to prove that the assumptions of the inverse function theorem are indeed satisfied. In order to solve the nonlinear equation F(u) f . the function F is well defined (i. : X 4 Y is of class F . as we shall see. for all f E V . we expect the above equation to have solution if the linear equation F”(0)U f = has solutions in X .116 Philippe G. an operator that is linear.. e.e. Ciarlet. it is thus natural = to apply the inverse function theorem (see.. = Furthermore. Cristinel Mardare Define the nonlinear mapping F :X 4 Y by F(w) := div ( ( I + V W ) ~ ) any w E X . we expect the norm of u to be small too. Second. Hence. If the norm of f is small.64. this equation has solutions in X thanks to Theorem 1. the mapping f EVHUEU is of class C1. then there exist two open sets U c X and V c Y with 0 E U and 0 = F ( 0 ) E V such there exists a unique element u E U satisfying the that. so that the above equation can be written as Since F ( 0 ) = 0. : bijective. According to this theorem.e. equation F(U) f . First. for where x = X(tr E ) I + 2 p and E = ~  2 Y v w T + v w+ VW~VW It suffices t o prove that the equation 3‘(u) =f has solutions in X provided that the norm of f in the space Y is small enough. and continuous with a continuous inverse). The idea for solving the above equation is as follows. if F : X 4 Y is of class C1 and the F’rkchet derivative F’(0) X 4 Y is a n isomorphism (i.. Taylor [85]). F(u) Y for all u E X)since the space W’J’(R) is an algebra (thanks E t o the assumption p > 3). the function . But this equation is exactly the system of equations of linearized threedimensional elasticity.
71)(1.3) an elastic material has a constitutive equation of the form ~ ( z:= ~ f l ( z . R3) if f belongs t o the neighborhood V of the origin in L*(Cl. from which we infer that the equation F’(0)u= f is exactly the equations of linearized threedimensional elasticity (see (1.8 Existence theory in nonlinear threedimensional elasticity by the minimization of energy (John Ball’s approach) We begin with the definition of hyperelastic materials. F is even of class P ) .72) has solutions for all Ilfll~~cn. Ciarlet & Destuynder [25]. with selfexplanatory notation f. the Frkchet derivative of .71.71)(1. who simultaneously and independently established the existence of solutions to the equations of nonlinear threedimensional elasticity via the implicit function theorem. R ’ ) . Theorem 1. is given by F F’(0)u diva. < 0 The unique solution u in the neighborhood U of the origin in W2+’(Cl. ) v+(z))for all z E K. .44) with To = an).e.71 can be found in Valent [86]. Recall that (see Section 1. = where c := A ( t r e ) l + 2 p e and e := (VuT 2 1 + Vu). then the problem (1. This shows that. Marsden & Hughes [SS].R3). Therefore.43)(1. under the assumptions of Theorem 1. I w ~ ) u. + f in L * ( R .Third. the system of equations of nonlinear threedimensional elasticity is wellposed.71)(1.An Introduction to Shell Theory 117 C1 since it is multilinear (in fact. j f u in W ~ ~ ~ ( R . if S > 0 is the radius of a ball B(0.72) have a unique solution in the neighborhood U of the origin in W2iP(Q. R3) of the equations of nonlinear threedimensional elasticity (1. In particular. the equations of nonlinear threedimensional elasticity (1.72) depends continuously on f . 6. 1. where Til : x M$ + M3 is the response function of the material and T ( z )is the first PiolaKirchhoff stress tensor at 2. Existence results such as Theorem 1.S) contained in V ..i.64 shows that the function F’(0) : X 4 Y is an isomorphism Since all the hypotheses of the inverse function theorem are satisfied.
wherep>2. and 11 . Ciarlet [18. 11 is the norm defined by IlFll := {tr(FTF)}'12 for all F E M3. the Euler equation associated with a minimization problem (this property only holds formally because.d>O.118 Philippe G. . a t each x E D. is given by W ( x .. and growth.b>O. e. for such materials. F )is the unique matrix in M3 that satisfies for all F E M$ and H E M3 (a detailed study of hyperelastic materials can be found in.a>O.4): where.41)) when the material constituting the body is hyperelastic has solutions if the function W satisfies certain physically realistic conditions of polyconvexity. Chap. In other words.g. in general.c>O. g ( x . called the stored energy function. A typical example of such a function W .F ) = allFllP + bllCofFllq + cI det 3' ' 1  dlog(det F ) forallFEM:. John Ball [9] has shown that the minimization problem formally associated with the equations of nonlinear threedimensional elasticity (see (1. To see this. the solution to the minimization problem does not have the regularity needed to properly establish the Euler equation associated with the minimization problem). the equations of nonlinear threedimensional elasticity are. Cristinel Mardare Then an elastic material is hyperelastic if there exists a function + R. which is called the stored energy function of the material. consider first the equations of nonlinear threedimensional elasticity (see Section 1. The major interest of hyperelastic materials is that. coerciveness.q> ~. such that its response function Tn be fully reconstructed from W by means of the can relation W : s2 x M$ where denotes the F'rkchet derivative of W with respect to the variable F . for simplicity.. 41). we have assumed that the applied forces do not depend on the unknown deformation a.r>l. at least formally. Ciarlet.
81 (John Ball). then T#(x. where J’ is the Friichet derivative of the functional J defined by J ( 9 ) := 1 W(x.9 d x  1.V*(x))dx  f . X I ) .An Introduction to Shell Theory 119 A weak solution @ to the boundary value problem (1. at least formally. If the material is hyperelastic. H .F ) = W ( x . + + for almost all x E R and f o r all F E M . The functional J is called the total energy. det F ) for all F E M . V@(s)) s ( x . . CofF. Therefore the variational equations associated with the equations of nonlinear threedimensional elasticity are.f3 E R such that W ( x .81) is then the solution to the following variational problem. ) 3 a( llFllP IICofFIIq I det FIT) P F  3. Let R be a domain in R3 and let W be a polyconvex function that satisfies the following properties: F The function W ( . John Ball’s theory provides an existence theorem for this minimization problem when the function W satisfies the following fundamental definition (see [9]): A stored energy function W : x Mr”+ R is said 4 to be polyconvex if.. and . There exist numbers p 3 2. m) 4 R such that a. a W ( x . 6 ) : R 3 R is measurable f o r all ( F . . a > 0 . also known as the principle of virtual works: for all smooth enough vector fields v : 4 R3 such that = 0 on ro. the Euler equations associated with the minimization problem J(+) = *‘EM min J ( 9 ) . = and the above equation can be written as a J’(+)w = 0. : Theorem 1. 6 ) E M3 x M3 x (0. for each x f there exists a convex function W ( x .H . where M is an appropriate set of all admissible deformations @ : R + R3 (an example is given in the next theorem). for all smooth enough vector fields 9 : 4 R3 such that 9 = id on ro. g.V Q ( x ) ) . q 3 r > 1.) : M3 x M3 x (0.9du.F . : .
EX3) and g E L4/3(171. there exist subsequences (@. T h e n there exists @ E M such that J ( @ )= inf J ( 9 ) .) + infgiEM J ( 9 ) < m.(n))).R3). is convex.R ~ ) . define the functional J(9.20 Philippe G. in R .1.(.. i.)) 6m(n) := det(V9. Finally. assume that arearo > 0 and that infgiiM J ( @ ) < 00. and (det(V@. a sequence of vector .(n) := Cof(V9. m) + IR is the function given by the polyconvexity assumption on W . Let r1 be a relatively open subset of dR. W ( ZV @ ( x ) .) : M3 x M3 x (0. and L'(R).(. EX3).V@(z)) z d and the set f (z) . L4(R. and let there be given fields f E L6l5(R. for each x E R.)) 3 H 2 6 in Lq(R. For all a E W1J'(R. 6) := H J.))) such that (3 denotes weak convergence) +u(n) 3 in w'J'(R. (Cof(V@.EX3). fields G n E M such that J(@. in L'(R).)). H.(.for a detailed proof). . Let @ be a infimizing sequence of the functional J . Since W ( x .e..H ( x ) S(x)) d x . EX3).e.M3). H E LQ(R. and 6 E L'(R) with S > 0 almost everywhere in R. let I'o := dR \ rl.(. Ciarlet. Define the functional W ( x .).)). Since these spaces are reflexive. W ( x . Cristinel Mardare For almost all x E R. The coerciveness assumption on W implies that the sequences (a.)) are bounded respectively in the spaces W1>p(R. @ = id o n ro}.) the above weak convergences imply that . . F )+ +m i f F E M i s such that : det F + 0. *EM Sketch of proof (see Ball 191 or Ciarlet [18]. M3).M3). @(z)d x  M := {@ E W17p(R. where. W ( x . (Cof (V@. . det(V@) E L'((R).M3). C d e t ( V 9 ) > 0 a. and (det(V@. o f ( V @ ) E L4(R.
the weak limit 9 of the . but not polyconvex.R3)and since the trace operator is linear. 6 ) = inf. A compactness by compensation argument applied to the weak convergences above then shows that H = C o f ( V 9 ) and 6 = d e t ( V 9 ) . However. The property that W ( F )4 +m if F E M+ is such that det F + 0.M J(*). In particular.H u ( n ) . it satisfies the coerciveness inequality: W b ( F )3 a(llF112+ /ICofFII’ + ( d e t F ) 2 )+ p. b > 0.1 1 1 3 ) . it follows that 9 = id on ro. with a > 0 and p E R. which is given by can be “approximated” with polyconvex stored energy functions in the following sense: There exists polyconvex stored energy functions of the form W b ( F )= allF112 + bllCofFI12 + cI det FI2 . e E R. =) ~F + A stored energy function of this form possesses all the properties required for applying Theorem 1. + 2 0 A St VenantKirchhoff material with Lam6 constants X > 0 and p > 0 is hyperelastic... Ciarlet & Geymonat [26] have shown that the stored energy function of a S t VenantKirchhoff material.that satisfy w ~ ( F W ( F ) C ? ( I I F. 6) = inf*.) satisfies the conditions of the theorem. . 6 .81. Hence J ( + ) = J ( 9 H . since 9 .An Introduction t o Shell Theory + 121 But J ( @ u ( n ) . c > 0. Finally. sequence aU(. then implies that d e t ( V 9 ) > 0 a. in R.dlog(det F ) + e with a > 0.e. 2 Twodimensional theory Outline In the first part of the article. 9 in W1>P(!2.p. . d > 0.M J(*). we have seen how an elastic body subjected t o applied forces and appropriate boundary conditions can be modeled . Since J ( 9 ) = J ( 9 H . ) It remains to prove that E M . Hence 9 E M . 6 u ( n ) ) = J ( @ c ( n ) )and J ( @ n ) inf*EM J(*). Therefore J ( 9 H .
In the second part of the article. which were written in Cartesian coordinates in the first part of the article.2. the equations of nonlinear or linearized threedimensional elasticity. the existence and uniqueness of solutions to the linear Koiter shell equations are established.10.This surface is the middle surface of the shell and E is its halfthichness. we define the reference configuration of a shell as the set in R3 formed by all points within a distance 6 E from a given surface in R3.6 to the equations of linearized threedimensional elasticity in curvilinear coordinates to derive the linearized membrane and flexural shell models. The linear Koiter shell model is then defined by linearizing the above tensor fields. These new equations may be viewed as a simplification of the equations of threedimensional elasticity. In Sections 2. In Section 2. we give a brief account of the derivation of nonlinear membrane and flexural shell models by letting the thickness E approach zero in the equations of nonlinear threedimensional elasticity in curvilinear coordinates.72. Clearly. In the next section. which is nothing but an elastic body whose reference configuration has a particular shape. In Section 2. Finally. . We then define a system of threedimensional curvilinear coordinates inside the reference configuration of a shell. we study the nonlinear and linear Koiter shell models.3. Of special importance are their first and second fundamental forms. Cristinel Mardare by the equations of nonlinear or linearized threedimensional elasticity. obtained by eliminating some of the terms of lesser order of magnitude with respect to the thickness of the shell. In Section 2. This simplification i s done by exploiting the special geometry of the reference configuration of the shell.122 Philippe G. The energy of the nonlinear Koiter shell model is defined in terms of the covariant components of the change of metric and change of curvature tensor fields associated with a displacement field of the middle surface of the reference configuration of the shell. Ciarlet. we begin our study with a brief review of the geometry of surfaces in R3 defined by curvilinear coordinates. thanks to a fundamental K o r n inequality o n a surface and to an infinitesimal rigid displacement lemma o n a surface. The same program is applied in Section 2. and especially. as a preliminary step toward the derivation of twodimensional shell theories. the assumed “smallness” of the thickness of the shell. are recast in terms of these natural curvilinear coordinates.5. these equations can be used in particular to model an elastic shell. we will show how an elastic shell can be modeled by equations defined on a twodimensional domain.
Ciarlet ~ 2 231). . and the summation convention with respect to repeated indices and exponents is systematically used. As a consequence. 3 } (save when they are used for indexing sequences).) range in the set {1. rzp:= a'. e.R3). y ) denote a generic point in the set 9 and let d. The covariant and contravariant components a. Greek indices and exponents (except v in the notation 8. aap := aa ' ap. q Y ) are linearly independent at all points y E G. . These two vectors thus span the tangent plane to the surface s := e(w) a t the point O(y). constitute the contravariant basis at the point 0(y) E S. dpa. any vector field rl : w + R3 can be decomposed over either of these bases as r] = via2= $ai.p and affp of the first and the covariant fundamental f o r m of S . see.g.i.p := a3 .2}.. .. The three vectors %(y) constitute the covariant basis at the point 0(y). 2 . .p and bf of the second fundamental f o r m of S are then defined by letting: a. Note that a3(y) = a3(y) and that the vectors aa(y) are also in the tangent plane to S at O(y). a mapping such that the two vectors %(Y) := 8 .An Introduction to Shell Theory 123 2. where the coefficients vi and vi are respectively the covariant and the contravariant components of q. while the three vectors ai(y)defined by the relations a2(y) . we briefly recapitulate some important notions of differential geometry of surfaces (for detailed expositions.() jY = q.. Let w be a domain in EX2. dpaa. Let y = ( . Latin indices and exponents range in the set { 1 .. b t := ap'bua. and the unit vector is normal to S at the point O(y). b. := d/dy.e. the Christoffel symbols and mixed components b. Let there be given an immersion 0 E C3(9.p := a. where dj is the Kronecker symbol.1 A quick review of the differential geometry of surfaces in R3 To begin with. ap.
. then one can see that all these functions vanish.)). (i. + rrar.. imply that These relations are equivalent to the Gauss and CodazziMainardi equations. These equations. bauau. where a := det(a. bLa. Note that one also has & = la1 A a2l.pa3. p .a). where R:.(&a. namely..  arr.124 Philippe G.@). save for R12.(auaa)= 6'. . This function is the Gaussian curvature of the surface S . ? := U"~R?.. given by We will see that the sign of the Gaussian curvature plays an important r6le in the twodimensional theory of shells... If R .. Ciarlet. Likewise..e. The derivatives of the vector fields ai can be expressed in terms of the Christoffel symbols and of the second fundamental form by means of the equations of Gauss and Weingarten: a. a. the derivatives of the vector fields aj satisfy &ar &a 3 = = r.. combined with the symmetry of the second derivatives of the vector field a.. + b.q. Cristinel Mardare The area element along S is &dy.aY + bza3.ap &a3 = = r .@ rgarYp  are the mixed components of the Riemann curvature tensor associated with the metric (a. := a.~.
E ] onto @ a x [ . T h e n there exists E > 0 such that the mapping 0 is a C2diffeomorphism f r o m W x [ . e. The coordinates y1.) := 8(y) f xga3(y) for all ( y .x.. Naturally. according to the following result (due to Ciarlet [20.as) > 0 in W. E]. g3) > 0 on W x [E. locally.y2.a3 and g3 = a30 = a 3 imply that det(gl.1. Busse [16] and S. see.g. The more general case of shells with variable thickness or with a middle surface described by several charts (such as an ellipsoid or a torus) can also be dealt with. Therefore. and let 0 : W x R 4 R3 be defined by 0 ( y . The assumed regularity on 8 implies that 0 E C2(W x E > 0. the implicit function theorem can be applied if E is small enough: It shows that..:= @@‘).E .xj) E a‘. R3) + xsd. where (cf.e. E ] if E > 0 is small enough. where gi := &0. Mardare [67]. of the points y E W constitute a system of “twodimensional” curvilinear coordinates for describing the middle surface of the reference configuration of the shell. this definition makes sense physically only if the mapping 0 is globally injective on the set Fortunately. Theorem 2. the mapping 0 is a C2diffeomorphism: Given any y E W.21. Figure 2. consider an elastic shell with middle surface S = O(Z) and (constant) thickness 2~ > 0 .x3) E w x R.g3)lZ3=o det(al. = Hence det(g.g2.&I. E ] ) and det(g. the surface S = 8(G) will be identified with the middle surface of a shell before deformation occurs. g2. let 8 E C3(G..g2.. i. an elastic body whose reference configuration is the set {fi‘}. 23) := B ( y ) + 23a3(y) for all ( y . this is indeed the case if the immersion 8 is itself globally injective on the set Z and E is small enough.E .An Introduction to Shell Theory 125 2. there exist a neighborhood U ( y ) of y in G and ~ ( y> 0 ) . [E. More specifically. i. Theorem 3.21) RE := w x (E. S is the middle surface of the reference configuration of the shell.g3) > 0 in ( w x [&.az. In what follows. a‘. I11). E) and 0 ( y . The relations g.R3) be a n injective immersion.. Let w be a domain in R2.e.2 Geometry of a shell Let the set w c R2 and the mapping 8 : CJ 4 R3 be as in Section 2. = &0 = a. for any Proof.
such as the sets W x [&. To establish that the mapping 0 : W x [ . there exist E" > 0. which is almost invariably given for functions defined over open sets. zg) = B(y) +zga3(y) for all (y. @(a") such that 0 is a C2diffeomorphism from the set U(y) x [E(Y). there exist y E W and . Let w be a domain in R2.R3) be an immersion. y" E W.z. Cristinel Mardare Figure 2.. E ) . there and Since the set W is compact. The coordinates (yl.z. E ] + R3 is injective provided E > 0 is small enough.. (yyn. see.21)...&I. such n3 that E.g. RE = w x ( .E(Y)] onto O ( U ( y ) x [ . Schwartz [81. let 8 E C3((w. ~ ( y ) ] )See. Q E 9. (y".E . we proceed by contradiction: If this property is false. Ciarlet.. f7 E W. 0(p. IZFI < E". and let the let mapping 0 : @ + R3 be defined by 0(y. Stein [SZ]).).21: T h e reference Configuration of a n elastic shell.g. = 6 E. Chapter 3 (the . and (p". e.~ ( y ) .) # (y". the set may be viewed as the reference configuration of an elastic shell with thickness 2~ and middle surface S = 8 ( W ) . 1 proof of the implicit function theorem. +0 as n + 00. can be easily extended to functions defined over closures of domains. z ) .. Then the mapping 0 is globally injective on globally injective on W and E > 0 is small enough (Theorem 2.) and O ( y n . In this case.).zg) are then viewed as curvilinear coordinates of the of an arbitrary point z E E point 2" = O ( x E ) of the reference configuration of the shell.y2.126 Philippe G. E?).E . e.zg) E  if the immersion 8 is RE.2. 0.
An Introduction to Shell Theory
127
exists an increasing function : N
N such that
Hence
by the continuity of the mapping 0 and thus y = g since the mapping 8 is injective by assumption. But these properties contradict the local injectivity (noted above) of the mapping 0.Hence there exists E > 0 such that 0 is injective on the set i x [  E , E ] . Z 0 In what follows, we assume that E > 0 is small enough so that the conclusions of Theorem 2.21 hold. The reference configuration of the considered shell is then defined by
{fi"}
:=
0(2'),
where R' := w x (  E , E ) and 6' := @(RE). Let x" = (xg)denote a generic point in the set G" (hence xz = y a ) and let 2" = ( 2 ; ) denote a generic point in the reference configuration {&}. The reference configuration of the shell can thus be described either in terms of the "threedimensional" curvilinear coordinates y1, y2, xj, or in terms of the Cartesian coordinates 2:, 25,2j, of the same point 2 = 0 ( x ' ) E {&}. ' To distinguish functions and vector fields defined in Cartesian coordinates from the corresponding functions and vector fields defined in curvilinear coordinates, we henceforth adopt the following convention of notation: Any function or vector field defined on flEis denoted by letters surmounted by a hat (e.g., ijEis a function defined on &, is a vector field defined on !?F,etc.). The corresponding functions and vector fields defined in curvilinear coordinates are then denoted by the same letters, but without the hat (e.g., g" is the function defined on RE by gE(xE)= t j & ( P for all xE E RE, f " is the vector field defined on 0' by ) f"(zCE) f (2.) for all xE E RE,etc., the points 2" and zE being related = by PE = @ x ) . (') Let 8; := (hence = d / d y a ) and let 8; := ala2if.For each xE E D', the three linearly independent vectors gz(x') := afO(x") constitute the cowarian? basis at the point O ( x E )and the three (likewise , linearly independent) vectors gj?" (x")defined by the relations gj"(x') . gg(xE) = d i constitute the contravariant basis at the same point. As a consequence, any vector field uE: RE + R3 can be decomposed over either basis as
i"
A &
a/ax:
a/axz
u"= UZgi?" z Q l E g ; ,
128
Philippe G. Ciarlet, Cristinel Mardare
where the coefficients uf and uz+are respectively the covariant and the contravariant components of U . ' The functions gtj(zE) gf(zE). g(' := ;z) and g i j > " ( z E ) gi>"(x') . := gjiE(z') are respectively the covariant and contravariant components of the metric tensor induced by the immersion 0. The volume element in O ( n E is then f l d z ' , where )
gE := det(gzj).
For details about these notions of threedimensional differential geometry, see Ciarlet [23, Sections 1.11.31
2.3
The threedimensional shell equations
In this section, we consider an elastic shell whose reference configuration is (6') := O ( n E ) (see Section 2.2), and we make the following assumptions. The shell is subjected to applied body forces given by their densities : & + R3 (this means that d P is the body force applied to the volume d2' at each 2 E 6'). For ease of exposition, we assume that ' there are no applied surface forces. The shell is subjected to a homogeneous boundary condition of place along the portion @(yo [E, €1) of its lateral face 0 ( d w x [  E , E]), where x yo is a measurable subset of the boundary aw that satisfies lengthy0 > 0. This means that the displacement field of the shell vanishes on the set @(yo x [&,El). The shell is made of a homogeneous hyperelastic material, thus characterized by a stored energy function (see Section 1.8)
1"
1"
: M3 +
R.
Such a shell problem can thus be modeled by means of a minimization problem (Section 1.8), which is expressed in Cartesian coordinates, in the sense that all functions appearing in the integrands depend on three variables, the Cartesian coordinates 2 = ( 2 ; ) of a point in the ' reference configuration {&} of the shell. We now recast this problem in terms of the curvilinear coordinates zE = (zz) describing the reference configuration {&} = of the same shell. This will be the natural point of departure for the twodimensional approch to shell theory described in the next sections. More specifically, the minimization problem consists in finding a minimizer 8' : {h'} + R3 of the functional (see Section 1.8) defined bv
@(a')
s'
An Introduction to Shell Theory
129
over a set of smooth enough vector fields 9"= {A"}R3 satisfying &"(2') = 2" for all 2" E @y x [&,&I). Recall that the functional k (o is the total energy of the shell. This minimization problem can be transformed into a minimization problem posed over the set i.e., expressed in terms of the "natural" curvilinear coordinates of the shell, the unknown : f R3 of this new problem being defined by
a",
+' a"
xE E
+ & ( x E )= 2 & ( i E ) all for
2" =
a'. a"
If E > 0 is small enough, the mapping 0 is a C'diffeomorphism of onto its image {AE} = and det(VEO)> 0 in (Theorem 2.21). The formula for changing variables in multiple integrals then shows that (P" is a minimizer of the functional J" defined by
@(a")
a'
~ " ( 9 " ) %(V9"(x")(VEO(xE))') det V'O dz" :=
1,.

In.
f"(xE) . 9 " ( x E )det V"O dx",
where the matrix field V"9" : + M3 is defined by V"W = (cf. Section 1.1)and the vector field f" : + R3 is defined by
a'
n"
(aj$;)
f E ( x E ):=
]"(P) all 2' for
=
~ ( x ' )x~ E RE. ,
Note that the function det V"O is equal to the function f l , where gE = det(gtj); cf. Section 2.2. Consider next a linearly elastic shell with Lam6 constants X > 0 and p > 0. In this case, the minimization problem associated with the equations of linearized threedimensional elasticity (Section 1.4) consists in finding a minimizer & : {A"} + R3 over a set of smooth enough vector fields 9 = {&} + R3 satisfying G"(2")= 2" for all 2" E ' @y x [&,&I) of the functional k defined by (o
where
X i @ ( = (tr(FT ~) 8
+ F  21))' + ~ 4
I I F ~ 21112 for all F E
+
~3
(this stored energy function for a linearly elsatic material easily follows from the equations of linearized threedimensional elasticity given in Section 1.4). Its expression shows that the functional J' is well defined if 9 E Hl(A"; R3). '
130
Philippe G. Ciarlet, Cristinel Mardare
As in the nonlinear case, this minimization problem can be recast in curvilinear coordinates. As such, it consists in finding a minimizer W : D" + R3 over the set of all vector fields * I E E H 1 ( f i ER3)satisfying ; W = 0 on yo x [  E , E ] of the functional J" defined by
As usual in linearized elasticity, it is more convenient to express this energy in terms of the displacement field U" : 2" R3, defined by
f
w(x")
=
~ ( x " )u E ( x Efor all x E E 52". + )
Likewise, let v E : 4 R3 be such that forward calculation shows that
nE
W
=0
+ v. '
Then a straight
r/ir( V E Q V E O )  ' ) AijkelE (v')eie (v"), (E = efj
e:j(vE):=  ( d f v E . g ;
2
1
+ d;v'
.g:).
The functions AijkeiE and e:j (u") denote respectively the contravariant components of the threedimensional elasticity tensor in curvilinear coordinates, and the covariant components of the linearized strain tensor associated with the displacement field v".It is then easy to see that U" is a minimizer over the vector space
v ( V ):= { u " = ufgi+; uf E H'(R"), u = o on 7 0 x :
of the functional J" defined by
(E,E)},
This minimization problem will be used in Section 2.6 as a point of departure for deriving twodimensional linear shell models.
2.4
The twodimensional approach to shell theory
In a twodimensional approach, the above minimization problems of Section 2.3 are "replaced" by a, presumably much simpler, twodimensional problem, this time "posed over the middle surface S of the shell". This means that the new unknown should be now the deformation cp : + R3 of the points of the middle surface S = t9(U), or, equivalently, the displacement field : U + R3 of the points of the same surface S (the
F
c
An Introduction t o Shell Theory
131
deformation and the displacement fields are related by the equation ‘p = 8 6); cf. Figure 2.41. The twodimensional approach to shell theory yield a variety of twodimensional shell models, which can be classified in two categories (the same classification applies for both nonlinear and linearized shell models) : A first category of twodimensional models are those that are obtained from the threedimesional equations of shells “by letting E go to zero”. Depending on the data (geometry of the middle surface of the shell, boundary conditions imposed on the displacement fields, applied forces) one obtains either a membrane shell model, or a flexural shell model, also called a bending shell model. A brief description of these models and of their derivation is given in Sections 2.5 and 2.6.
+
h
Y
E > 0 “small enough” and data of ad hoc orders of magnitude, the threedimensional shell problem is “replaced” by a “twodimensional shell problem”. This means that the new unknowns are the three covariant components Ci : GJ + E of the % displacement field Ciai : GJ + R3 of the points of the middle surface S = O@). In this process, the “threedimensional” boundary conditions on !o need to ? be replaced by ad hoc L‘twOdimensional’’ boundary conditions on yo. For = 0 on 70 (used instance, the “boundary conditions of clamping” Ci = in Koiter’s linear equations; cf. Section 2.8) mean that the points of, and the tangent spaces to, the deformed and undeformed middle surfaces coincide along the set O(y0).
Figure 2.41: An elastic shell modeled as a twodimensional problem. For
132
Philippe G. Ciarlet, Cristinel Mardare
A second category of twodimensional models are those that are obtained from the threedimensional model by restricting the range of admissible deformations and stresses by means of specific a priori assumptions that are supposed to take into account the “smallness” of the thickness (e.g., the Cosserat assumptions, the KirchhoffLove assumptions, etc.). A variety of twodimensional models of shells are obtained in this fashion, as, e.g., those of Koiter, Naghdi, etc. A detailed description of Koiter’s model is given in Sections 2.7 and 2.8.
2.5
Nonlinear shell models obtained by rconvergence
Remarkable achievements in the asymptotic analysis of nonlinearly elastic shells are due to Le Dret & Raoult [64] and to F’riesecke, James, Mora & Miiller [49], who gave the first (and only ones to this date) figorous proofs of convergence as the thickness approaches zero. In so doing, they extended to shells the analysis that they had successfully applied to nonlinearly elastic plates in Le Dret & Raoult [62] and F’riesecke, James & Miiller [48]. We begin with the asymptotic analysis of nonlinearly elastic membrane shells. H. Le Dret and A. Raoult showed that a subsequence of deformations that minimize (or rather “almost minimize” in a sense explained below) the scaled threedimensional energies weakly converges in W1>p(R;R3)as E 4 0 (the number p > 1 is governed by the growth properties of the stored energy function). They showed in addition that the weak limit minimizes a “membrane” energy that is the rlimit of the (appropriately scaled) energies. We now give an abridged account of their analysis. Let w be a domain in R2 with boundary y and let 8 E C2(G; R3) be an injective mapping such that the two vectors a ( ) = d,8(y) are ,y linearly independent at all points y = (ye) E SLi. Consider a family of elastic shells with the same middle surface S = 8(G) and whose thickness 2~ > 0 approaches zero. The reference configuration of each shell is thus the image @@&) c R3 of the set & s1 c R3 through a mapping 0 : + R3 defined in Section 2.2. By Theorem 2.21, if the injective mapping 8 : W 4 R3 is smooth enough, the mapping 0 : 4 R3 is also injective for E > 0 small enough and y1, y2, x$ then constitute the “natural” curvilinear coordinates for describing each reference configuration O ( @ ) . Assume that all the shells in the family are made,of the same hyperelastic homogeneous material (see Section 1.8), satisfying the following properties : The stored energy function I’ : M3 4 R of the hyperelastic material %
") := O ( x " ) + U 8 ( Z E ) .An Introduction to Shell Theory satisfies the following assumptions: There exist constants C 0.Z E E a". which is given by P @ ( F ) = 1IF 4 + FT  where JJFJJ FTF}'12. where ff E Lq(fi") and $ = 1. where U" : fiE 4 R3 is its displacement . a > It can be verified that the stored energy function of a St VenantKirchhoff material. it is assumed that each shell is subjected to a boundary condition of place along its entire lateral face O(y x [&. of the reference configuration. an example of applied surface force that depends on the unknown deformation. the stored energy function of a linearly elastic material.. for each E > 0. Remark. By contrast. which is given by satisfies such inequalities with p = 4. &]). but 0 It is further assumed that.satisfies the first inequality with p := (tr not the second one. where y := dw. but are omitted for simplicity. i. = 2. The threedimensional problem is then posed as a minimization probl e m in terms of the unknown deformation field +i a'(. Applied surface forces on the "upper" and "lower" faces of the shells could be likewise considered. see in this respect Le Dret & Raoult [64] who consider a pressure load. Finally. and that these densities do not depend on the unknown deformation. that the displacement vanishes there. p E EX.and 1 < p < ca such that 133 > 0.e. the shells are subjected in their interior t o applied body forces of density f" = ffgZ+ : fi" 4 R3 per unit volume.
b2. If" 9" V"0dx". 2 2 .51). det This minimization problem may have no solution. The above minimization problem is then transformed into an analogous one. z E a.3): It consists in finding a" such that J E ( S E )where .8 2 s . n . in the sense that there exists a vector field f E L2(R. Ciarlet. 2 3 ) E R + 2" = (2") = ( 2 1 . 1 R y x [I.1.R3) independent of E such that fE(x") = f ( 2 ) for all zE = T"Z. 1 I@( (819. = + O ( E ) on 9 = inf J ( E ) ( ~where ). E Z g ) E i n". b3 (in this order). Should applied surface forces act on the upper and lower faces of the shells. b2. however.W3). With each point x E we associate the point xE E through the bijection a. 2 3 ) denote a generic point in and let & := 8/8xi. we assume that the applied body forces are of order O(1) with respect to E . 2 2 . R) J ( E ) ( * ) := *'EM(". but now posed over the fixed domain R := W X ] . These scalings and assumptions then imply that the scaled deformation % ( E ) satisfies the following minimization problem: a(&) M ( & . 2 2 . 0 In what follows. Cristinel Mardare field (Section 2. &*)(G(E))') det G ( Edz & ) r / n f . and J ( E ) ( @ ( E ) ) E 0) M(E.134 Philippe G. this is not a shortcoming as only the existence of a "diagonal infimizing family".IW3). I } ]. Let x = ( 2 1 . = 0 on y x E 9" . the notation (bl. [E. whose existence is always guaranteed. we would then assume that they are of order O(E) with respect to E . a + R3 We then define the unknown scaled deformation a(&) : by letting @ ( E ) ( Z ) := @ " ( x E ) for all 2" = T ' Z . Remark. a a" 7r" : 2 = (21. +det G(E) dz. a"E M(R") and J"(a") = M(R') := {a" W11p(R". b3) stands for the matrix in M3 whose three column vectors are bl. W€M(W) inf E]}. is required in the ensuing analysis (Theorem 2. n) := {9E W1>p(R.IL: E a. . Finally. 1[.
a function I : M m x n 4 R is quasiconvex if.+ R. 1 + (&w.R"). 721 (an account of its importance in the calculus of variations is provided in Dacorogna [38. 11) w Io~ . 411 (see also De Giorgi & . and all 6 = (&)gl W.X is quasiconvex and X <W). Given any function W : MImXn. E where V6 denotes the matrix (aj<i) E M"'". its quasiconvex envelope QW : MImxn + R is the function defined by QW := sup{$ wmxn 4 R.R~) J ( E ) ( w:= ) h {w E w ' > P ( s ~ .'l"(U. x E := ) n.who explicitly determine the quasiconvex envelope of the stored energy function of a St VenantKarchhoff material. Remark.a+)(&) therefore satisfies the following minimization problem: U(E) E W(R. a powerful theory initiated by De Giorgi [40. d2w. &w)(G(&))') det G ( Edx ) v€w(n. The scaled displacement U(E) := +(&) .IR3) and J(E)(u(E)) = := w(R. ('Po(&) + w)det G(E) dx. Also central to the ensuing analysis is the notion of rconvergence. x E '32. Chap 51): Let M m x n denote the space of all real matrices with m rows and n columns. := and the matrix field G(E): + M3 is defined by G ( E ) ( z ) V e O ( x Efor all xE = T ' X . Central to the ensuing result of convergence is the notion of quasiconvexity.An Introduction to Shell Theory where the vector field +o(E) 135 : fi + R3 is defined for each E > 0 by + o ( ~ ) ( z ) @(xE) for all x E = T ' Z . for all bounded open subsets @ U c R".R3) inf J(E)(w).= R on)y x [I. where %(I 5 f .all F E M m x n . An illuminating instance of actual computation of a quasiconvex envelope is found in Le Dret & Raoult [63]. due to Morrey [71.
136
Philippe G. Ciarlet, Cristinel Mardare
F’ranzoni [43]); an illuminating introduction is found in De Giorgi & Dal Maso [42] and thorough treatments are given in the books of Attouch [8] and Dal Maso [39]. As shown by Acerbi, Buttazzo & Percivale [l] for nonlinearly elastic strings, by Le Dret & Raoult [62, 641 for nonlinearly elastic planar membranes and membrane shells, by F’riesecke, James & Muller [48] for nonlinearly elastic flexural plates, and by F’riesecke, James, Mora & Muller [49] for nonlinearly elastic flexural shells, this approach has thus far provided the only known convergence theorems for justifying lowerdimensional nonlinear theories of elastic bodies. See also Ciarlet [19, Section 1.111 for an application to linearly elastic plates. We then recall the fundamental definition underlying this theory: Let V be a metric space and let J ( E ) : V + R be functionals defined for all E > 0. The family ( J ( E ) ) ~ > o is said to rconverge as E + 0 if there exists a functional J : V + R U {+co}, called the Flimit of the functionals J ( E ) , such that
V(E)
+
v as E
+
0 + J ( v ) < liminf J ( E ) ( v ( E ) ) ,
E+O
on the one hand and, given any v E V, there exist V ( E ) 6 V, E > 0, such that V ( E ) + v as E + 0 and J ( v ) = lim J ( E ) ( v ( E ) ) ,
E+O
on the other. As a preparation to the application of I?convergence theory, the scaled energies J ( E ): M(R) + R found above are first extended to the larger space LP(R; EX3) by letting
J(E)(W) =
Such an extension, customary in Fconvergence theory, has inter alia the advantage of “incorporating” the boundary condition into the extended functional. Le Dret & Raoult [64] then establish that the family ( J ( E ) ) ~ > ~ of extended energies Fconverges as E + 0 in LP(R; EX3) and that its rlimit can be computed by means of quasiconvex envelopes. More precisely, their analysis leads to the following remarkable convergence theorem, where the limit minimization problems are directly posed as twodimensional problems (part c)); this is licit since the solutions of these limit problems do not depend on the transverse variable (part (b)). Note thaq, while minimizers of J ( E )over M ( Q ) need not exist, the existence of a “diagonal infimizing family” in the sense understood below is always guaranteed because inf,,M(n) J ( E ) ( w > co. ) In what follows, the notation ( b l ; b2) stands for the matrix in M3x2 with bl, b2 (in this order) as its column vectors and &dy denotes as usual the area element along the surface S .
i
J ( E ) ( w if w E M ( Q ) , ) f c o if w E LP(R;R3) but w $ M(R).
An Introduction to Shell Theory
137
Theorem 2.51. A s s u m e that the applied body forces are of order O(1) with respect to E , and that there exist C > 0 , a > 0, p E R, and 1 < p < 00 such that the stored energy function W : M3 3 R satis$es the following growth conditions:
for all F E M3, IW(F)I < C ( I + IF\”) for all F E M ~ W(F) b a\FI” + p (W(F)  W/(G)I C(I + l ~ l p  l + I G \ ~  ~ IG I for all F , G E ~ 6 ) F
,
3 .
Let the space M(R) be defined by
M(R) := {v E W11p(R;R3);v = 0 on y x [I, I]},
and let (U(E)),>O be a “diagonal infimizing family” of the scaled energies, i.e., a family that satisfies
U(E)
E M(C2)
and J(E)(u(E)) 6
VEM(n)
inf
J ( E ) ( v ) h(&) all + for
E
> 0,
where h is any positive function that satisfies h(&) 0 as E + 0 . Then: + (a) The family ( U ( E ) ) , > ~ lies in a weakly compact subset of the space Wl~”(R;R3). (b) T h e limit u E M(R) as E + 0 of any weakly convergent subsequence of ( U ( E ) ) ~ > O satisfies 8321 = 0 in R and is thus independent of the transverse variable. 1 (c) The vector field 6 := S, u dx3 satisfies the following minimization problem:
C E W ~ 1 p ( w ; R 3 ) j ~ ( 6= and )
where
inf
17EW;’P(W;P3)
j~(rl),
G ( Y ) (ai(~); ( ~ )as(^)), := a2 ;
the vectors a i ( y ) forming for each y E 2 the covariant basis at the point O(y) E S , and Q W o ( y , .) denotes for each y E G the quasiconvex envelope of W o ( Y , .). 0
138
Philippe G. Ciarlet, Cristinel Mardare
It remains to descale the vector field C. In view of the scalings performed on the deformations, we are naturally led to defidng for each E > 0 the limit displacement field C" : Z +EX3 of the middle surface by
s
C"
:= 6.
It is then immediately verified that C" satisfies the following minimization problem (the notations are those of Theorem 2.51):
C"
E W~'*(w;EX3) j$(C") = and
inf
1)EW,1'P(W;R3)
jG(r]),where
The unknown r] in the above minimization problem appears only by means of its firstorder partial derivatives a,r] in the stored energy function
r] E
W1Jyw;R3)
f
€&*&
(a1 8 7 ;a 17 2
+
+
a2r]>>
found in the integrand of the energy jh. Assume that the original stored energy function is frameindifferent, in the sense that
I@(RF) I@(F) for all F E M3 and R 6 0;. =
This relation is stronger than the usual one, which holds only for F E M3 with det F > 0 (see Ciarlet [18, Theorem 4.211); it is, however, verified by the kinds of stored energy functions to which the present analysis applies, e.g., that of a St VenantKirchhoff material. Under this stronger assumption, Le Dret & Raoult [64, Theorem 101 establish the crucial properties that the stored energy function found in jh, once expressed as a function of the points of S, is frameindifferent and that it depends only on the metric of the deformed middle surface. For this reason, this theory is a frameindifferent, nonlinear "membrane" shell theory. It is remarkable that the stored energy function found in jh can be explicitly computed when the original threedimensional stored energy function is that of a St VenantKirchhoff material; see Le Dret & Raoult [64, Section 61. Again for a St VenantKirchhoff material, Genevey [50] has furthermore shown that, when the singular values of the 3 x 2 matrix fields (8,vi) associated with a field r] = viai belong to an appropriate compact subset of R2 (which can be explicitely identified), the expression
An Introduction to Shell Theory
139
j & ( q ) takes the simpler form
where
aapuT . %a@aUT
a,O(q)
x + 2p := a,(e + viai) . a,(e
+ 2p(aa‘7aP7 + aOrTaPU),
+ qjaj).
This is precisely the expression of j & ( q ) that was found by Miara [69] to hold “for all fields q” (i.e., without any restriction on the fields q such as that found by Genevey [50]),by means of a formal asymptotic analysis. This observation thus provides a striking example where the limit equations found by a formal asymptotic analysis “do not always coincide” with those found by means of a rigorous convergence theorem. Le Dret & Raoult [64, Section 61 have further shown that, if the stored energy function is frameindifferent and satisfies l@(F) $ ( I ) for all 3 F E M3 (as does the stored energy function of a St VenantKirchhoff material), then the corresponding shell energy i s constant under compression. This result has the striking consequence that “nonlinear membrane shells o#er n o resistance t o crumpling. This is an empirical fact7 witnessed by anyone who ever played with a deflated balloon” (to quote H. Le Dret and A. Raoult). We now turn our attention to the asymptotic analysis, by means of I?convergence theory, of nonlinearly elastic flexural shells. In its principle, the approach is essentially the same (although more delicate) as ‘that used for deriving the nonlinear membrane shell equations. There are, however, two major differences regarding the assumptions that are made at the onset of the asymptotic analysis. A first difference is that the applied body forces are now assumed to be of order C ~ ( E with respect to E (instead of 0(1)), the sense that ~) in there exists a vector field f E L2(a; independent of E such that Pi3)
fg(xE)= E ~ ~ ( x all x E = T‘X for )
E s2.
A second difference (without any counterpart for membrane shells) is that the set denoted M F ( win the next theorem contains other fields ) than 8 (the interpretation of this key assumption is briefly commented upon after the theorem). Under these assumptions, F’riesecke, James, Mora & Muller [49] have proved the following result. The notations as($), asp($), and hap($)
140
Philippe G. Ciarlet, Cristinel Mardare
used in the next statement are selfexplanatory: Given an arbitrary VFCtor field E H2(u;JR3),
+
Theorem 2.52. A s s u m e that the applied body forces are of order C ~ ( E ~ ) with respect t o E . A s s u m e in addition that the stored energy function W : M3 4 R satisfies the following properties: It is measurable and of class C2 in a neighborhood of Ot, satisfies it
* ( I ) = 0 and i @ ( R F )= $ ( F ) for all F E M3 and R E O , :
and, finally, it satisfies the following growth condition: There exists a constant C > 0 such that
IL+(F)I
c REO? inf
I F ~1~
for all F E M ~ .
Finally, assume that the set
M F ( w ) {+ :=
E
H2(w;JR3); a,p(+) = aap in w ; = 8 and as(+) = a3 on YO},
+
contains other vector fields than 8 . T h e n the scaled energies J ( E ) , E > 0 , constitute a family that I?converges as E + 0 in the following sense: A n y "diagonal infimizing family" (defined as in Theorem 2.51) contains a subsequence that strongly converges in H1(R; R3). Besides, the limit of any such subsequence is independent of the transverse variable, and the vector field cp :=
where
An introduction to Shell Theory
141
The assumption that the set M F ( u )contains other vector fields than 8 means that there exist nonzero displacement fields viai of the middle surface 6 ( 9 ) that are both inextensionaZ, in the sense that the surfaces 8(Z) and +(G), where := 8 viai, have the same metric (as reflected by the assumption a,p(+) = a,p in w ) , and admissible, in the sense that the points of, and the tangent spaces to, the surfaces 8(Z) and +(Z) coincide along the set 6(yo) (as reflected by the boundary conditions = 8 and a3(+) = a on 70). 3 it follows that the “descaled” unknown deformation pE Z + R3 of : the middle surface of the shell is a minimizer over the set M F ( w ) the of functional j g defined by
+
+
+
When the original threedimensional stored energy function is that of a St Venant Kirchhof material, the expression j$(+) takes the simpler form
where
+ E M F ( w by means of a formal asymptotic analysis by Lods & Miara )
interestingly, exactly the same expression jg(+) was found for all
[65], as the outcome of sometimes exceedingly delicate computations. This observation is thus in sharp contrast with that made for a membrane shell, whose limit equations cannot always be recovered by a formal approach, as noted earlier.
Remark. Although rconvergence automatically provides the existence of a minimizer of the rlimit functional, the existence of a minimizer of the functional j~ over the set M F ( u )can be also established by means of a direct method of calculus of variations; cf. Ciarlet & Coutand [24].c3
R 3 ) . Their elastic material is thus characterized by two Lam6 constants X > 0 and p > 0.Their reference configurations are thus of the form O(fi“). Then Ciarlet & Lods [27. This means that. and on the order of magnitude of the applied forces. More specifically. a 3 ( y ) for all ( y . All the shells in the family are made with the same homogeneous isotropic elastic material and that their reference Configurations are natural states. or twodimensional flexural. Another major step was achieved by SanchezPalencia [77]. ~ : ) . as the outcome of an asymptotic analysis of the equations of threedimensional linearized elasticity as E + 0 . where w is a domain in R2 with boundary y. The shells are subjected to body forces and that the corresponding applied body force density is O(EP) with respect t o E . they established convergence theorems in H I . The asymptotic analysis of elastic shells has been a subject of considerable attention during the past decades. where a convergence theorem for “membrane shells” was “almost proved”. equations when the method of formal asymptotic expansions is applied to the variational equations of threedimensional linearized elasticity (see also Caillerie & SanchezPalencia [17] and Miara & SanchezPalencia [70]). on the boundary conditions. x ~:= 6 ( y ) ) + x . 281 and Ciarlet. and rigorously justified. in L 2 . Ciarlet. After the landmark attempt of Goldenveizer [53]. or in ad hoc completion spaces.who clearly delineated the kinds of geometries of the middle surface and boundary conditions that yield either twodimensional membrane. consider a family of linearly elastic shells of thiclcness 2~ that satisfy the following assumptions: All the shells have the same middle surface S = 8(C)c R3. for some ad hoc power p (which will be specified later). that justify either the linear twodimensional equations of a “membrane shell”. for each E > 0.6 Linear shell models obtained by asymptotic analysis In this section.142 Philippe G. Cristinel Mardare 2. or those of a “generalized membrane shell”. we briefly review the genesis of those twodimensional linear shell theories that can be found. where RE := w x (&.&) and the mapping 0 is defined by @ ( y . or those of a ‘‘j?exural shell”. Lods & Miara [31] carried out an asymptotic analysis of linearly elastic shells that covers all possible cases: Under three distinct sets of assumptions on the geometry of the middle surface.E > 0.a major step for linearly elastic shells was achieved by Destuynder [44] in his Doctoral Dissertation. .and 8 E C 3 ( Z .
€1) of its lateral face. This assumption is in fact one in disguise about the geometry of the surface S and on the set 7 0 .3): uEE V ( V ) and J E ( u E ) = min V"EV(W) JE(v'). it is satisfied if S is a portion of a cylinder and O(70) is contained in one or two generatrices of S .ap) and pOlg(rl)= ( % g q .u : = o on TO x (E. each shell is subjected to a boundary condition of place on the portion @y x [E. . In Ciarlet. and linearized change of curvature. qa E ~ ' ( w ) . Then the displacement field of the shell satisfies the following minimization problem associated with the equations of linearized threedimensional elasticity in curvilinear coordinates (see Section 2.E)}.(w) := { q = viai.An Introduction to Shell Theory the contravariant components f = fi)'gS are of the form ' fz?" 143 E L2(RE)of the body force density f i l E ( y E Z ~= & p f i ( y . in which case the shells are plates. := E H ~ ( R ) .773 E H ~ ( w ) . Lods & Miara [31] it is first assumed that the space of linearized inextensional displacements (introduced by SanchezPalencia PI) v. and the functions f i E L2(R) are independent of E (surface forces acting on the "upper" and "lower" faces of the shell could be as well taken into account but will not be considered here. ) I[. where y is a fixed portion (o o of y. z3) E R := w x 11. a3 acu denote as usual the covariant components of the linearized change of metric.z3) for all (y. tensors.gi+. or if S is contained in a plane. let f 1 y d v ) = ~ ( d p v . . v = dyv3 = i 0 on yo. w. for simplicity of exposition). Let then the functions pi>' E L2(w)be defined for each E > 0 by J € Finally. ycyp(q)= 0 in w} contains nonzero functions.r&a. where V ( R E ) { v E= v. ' For any displacement field q = V i a i : w R3.q) .with lengthy0 > 0. + aav. this problem has one and only one solution u E V ( R ) . For instance. For each E > 0.
as E + 0. forces. the twodimensional equations of a linearly elastic f # 'Vexural shell" are therefore justified.144 Philippe G. Ciarlet. where it was assumed that y = y and that the surface S is elliptic. and where for all v = viai E V F ( Wwhere the functions ). are precisely the familiar contravariant components of the shell elasticity tensor. Hence a more refined asymptotic analysis is needed in this case. = the only information it provides is that u d & J+ 0 in H1(w.. i f the applied body force density is O ( e 2 ) with respect to E . R3) := &ai belongs t o the space V F ( W ) and where the limit vector field satisfies the equations of a linearly elastic "flexural shell" . A first instance of such a refinement was given by Ciarlet & Lods [27]. I V F ( W ){ 0 } . However. Equivalently. R3) ' 2 & E as E 0. since both sides of these variational equations are of the same order (viz. the vector field 6 satisfies the following constrained minimization problem: 5 E V F ( W ) G(5) = inf&(v). If V F ( W )( 0 } . Lods & Miara [31]showed that. As shown in Ciarlet & Lods [27] and Ciarlet & SanchezPalencia [35]. indeed imply that V p ( w ) = (0). E" S. + SE . together with ad hoc regularity assumptions. viz. independent of E .. then ! 2E 1' & uE dz: + C in H 1(w. Cristinel Mardare Under this assumption Ciarlet. these two conditions. o in the sense that its Gaussian curvature is > 0 everywhere.the above convergence result still applies. E ~ ) because of the assumptions made on the applied . a a p c T p u T ( < ) p f f p ( v ) h d = b p i l ~ v i h d y 3 y for all 77 = viai E V F ( WObserve in passing that the limit 6 is indeed ).
E ) . and i f their density is again 0(1)with respect t o E . Ciarlet & Lods [30] studied all the "remaining" cases where VF(w) { 0 } . where the limit vector field C := &ai belongs t o the space H.yap(q). or when = S is for instance a portion of a hyperboloid of revolution. v a E and solves the equations of a linearly elastic "membrane shell". &aapUTyur C ) ~ a (q)& d~ = ( p pi"vi& dy for all q = viai E V M ( W ) . o the twodimensional equations of a linearly elastic "membrane shell" are therefore justified.g. where the functions a"Pu7. Observe that the limit C is again independent of E . the field C satisfies the following unconstrained minimization problem: where &(v) := 1 J I _ E U U " " ' m T ( q ) y a p ( l ) ~ d Y l p z 3 E v i & d y .. Ciarlet & Lods [28] showed that. when S is elliptic but lengthyo < lengthy. e.a. if the applied body force density is 0(1)with respect to E . In this case. To give a flavor of their results. etc. then 2E /' pE u: dxg + <a in H1(w) and 2E J' u dx. ~3 E ~ ~ ( w ) ) . If y = y and S is elliptic. since both sides of these variational equations are of the same order (viz. viz . then . Ciarlet & Lods [30] showed that. i f the applied body forces are "admissible" in a specific sense (but a bit too technical t o be described here).l(w). o. and pi)€ have the same meanings as above.  2 Finally.An Introduction to Shell Theory 145 In this case. consider the important special case where the seminorm becomes a n o r m over the space W ( W ) := { q E H ' ( ~ . V M ( W )(77 := = viai.. g f c3 E in L 2 ( w )as E f 0. Equivalently. I w ~ ) o on y) q=. because of the assumptions made on the applied forces..
c = for where B L is the unique extension to V # . va E ~ 1 W ' ( w ) q. the vector field that satisfies C" E V ( w ) = { q = viai. vi = ~ ~ = o on 7 3 'yo}. and L g : V & w ) + R is an ad hoc linear form.9).Iw of the quotient space W ( w ) / W ..93) and on the Korn inequality ((onan elliptic surface" (end of Section 2. where V F ( W )( 0 ) but :1. ( w ) of the bilinear form BM defined by i. B ( q ) } ~ d ~ + ~ . Ciarlet.il M lished.e.10. whose twodimensional equations are therefore justified.. ( w ) solves "limit" variational equa E B ~ ( q )~ . where W .8 to 2. 781 (see also SanchezHubert & SanchezPalencia [79]). again in all possible cases. C E V # . Suffice it to say here that they crucially hinge on the Korn inequality "with boundary conditions" (Theorem 2.'yap(q) = O in w } . { & ~ " " " ' ~ P T ( C ' ) ~ ~ ~~( U ") " " ' p u " ' ( C " ) p .1 = is not a norm over the space W ( w ) . ( w ) = {q E W ( w ) . a similar convergence result can be estab. viz.146 Philippe G.~ g ( q ) all q E v ~ ( w ) .Iw M . ( w ) . Either one of the above variational problems corresponding to the %emaining" cases where V F= {0} constitute the equations of a linearly elastic "generalized"membrane shell. In the "last" remaining case.(W) with respect of (. Ciarlet & Lo& [28. but only in the completion V. Cristinel Mardare where V L ( w ):= completion of W ( w )with respect to Furthermore. 76.3 E H'(w).102) ' to the linear Koiter shell equations. E B Mis the bilinear form found above for a linearly elastic "membrane shell". the limit field tions of the form ]. The proofs of the above convergence results are long and technically difficult. Combining these convergences with earlier results of Destuynder [45] and SanchezPalencia [75. Let C denote for each E > 0 the unique solution (Theorem 2. determined ( by the behavior as E + 0 of the admissible body forces. 291 have also justified as follows the linear Koiter shell equations studied in Sections 2.
r(7))raa(rl) ~aa@'PuT(17)Pa/3(7)} I E3 is thus exactly the s u m of the stored energy function of a linearly elastic "membrane shell" and of that of a linearly elastic 'yexural shell". viz. have ' exactly the same asymptotic behavior as E + 0 .. in precisely the same function spaces that were found in the asymptotic analysis of the threedimensional solution. etc. we begin our study of the equations proposed by W. generalized membrane.T. the vector fields C' and  1 PE where u denotes the solution of the threedimensional problem. membrane. or flexural. It is all the more remarkable that Koiter's equations can be fully justified f o r all types of shells. gener1 alized membrane. or flexural). such as those of Naghdi. for a detailed description and analysis of other linear shell models.An Introduction to Shell Theory or equivalently. for a linearly elastic shell. it is in a sense the "best" twodimensional one for linearly elastic shells! One can thus only marvel at the insight that led W.. 77 { iaapcT rv.7 The nonlinear Koiter shell model In this section. Koiter to conceive the "right" equations. These equations are "twodimen . since it is clear that Koiter's equations cannot be recovered as the outcome of an asymptotic analysis of the threedimensional equations. the unique solution to the minimization problem 147 C' where E V ( w ) andj(6') = inf vEV(w) j(7) Observe in passing that. shells exhausting all such possible outcomes! So. out of purely mechanical and geometrical intuitions! We refer to Ciarlet [20] for a detailed analysis of the asymptotic analysis of linearly elastic shells. the stored energy function found in Koiter's energy.T. the twodimensional equations of linearly elastic. Then. whose versatility is indeed remarkable. Novozilov. Budiansky and Sanders. and for an extensive list of references. 2. Koiter for modeling thin elastic shells. even though Koiter's linear model is not a limit model. f o r each category of linearly elastic shells (membrane.
asserts that the state of stress inside the shell m + .3). It asserts that any point situated on a normal to the middle surface remains on the normal to the deformed middle surface after the deformation has taken place and that. y2. Given an arbitrary displacement field 77 := viai : 3 4 R3 of the surface S with smooth enough components vi : W + R. Then the functions Rap(v) := b a p ( v ) . where aa(v) := aa(6 + v). in addition. Then the functions + denote the covariant components of the change of metric tensor associated with the displacement field 77 = vzazof S. ap(v). let 1 b a p ( v ) := aa. in the sense that they are expressed in terms of two curvilinear coordinates used for defining the middle surface of the shell. is the KirchhoffLove assumption. Ciarlet. To begin with. let aap(v) := aa(v) . If the two vectors aa(v) are linearly independent at all points of w .(@ v). of a geometrical nature. Note that both surfaces 6(W)and (6 q)(W) equipped with the are same curvilinear coordinates y1. consider an elastic shell made of a St VenantKichhoff material modeled as a threedimensional problem (Section 2. 7) denote the covariant components of the second fundamental form of the deformed surface (6 + v)(W). Note that = la1 (v)A a2 (v) I. As a point of departure. of a mechanical nature. Cristinel Mardare sional”. so named after Koiter [59] and since then a twodimensional nonlinear . The nonlinear twodimensional equations proposed by Koiter [59] for modeling such an elastic shell are then derived from those of nonlinear threedimensional elasticity on the basis of two a priori assumptions: One assumption.{a1(v) A a2(v)l. denote the covariant components of the first fundamental form of the deformed surface (6 ~)(a). The other assumption.bap denote the covariant components of the change of curvature tensor field associated with the displacement field v = vzazof S.148 Philippe G. ~ m + where 4 7 := det(aap(v)). model of choice in computational mechanics. we describe the nonlinear Koiter shell equations. the distance between such a point and the middle surface remains constant.
571).6.p(q) = bCYp(q) b. should be a stationary point. W.1). Koiter then reached the conclusion that the displacement field 6" = <faiof the middle surface S := O(Z) of the shell. where the functions are unknowns. they are defined by The above functional j is called Koiter's energy for a nonlinear elastic shell. over a set of smooth enough vector fields q = viai : 55 R3 satisfying ad hoc boundary conditions on 7 0 . The stored energy function W K found in Koiter's energy j is thus defined by w K ( q )= . i.T.p(rl). Koiter [59. of the functional j defined by (cf. 6 E3 Another closely related set of nonlinear shell equations "of Koiter's type" has been proposed by Ciarlet [all.2). thereby avoiding the possibility of a vanishing denominator in the expression WK(T. This expression is the sum of the & "membrane" part WM(V) = 5a'"""'Gu7(~)Gap(~) and of the 'yexural" part WF(rl)= aa~u"'Ru'(rl)R. eqs. In these equations.e.3)]): < : j  ~Pz~"vi&dy.p is simply replaced by 6. and (8. the denominator that appears in the functions R. in particular a minimizer. Taking these a priori assumptions into account.An Introduction to Shell Thcory 149 is planar and parallel to the middle surface (this second assumption is itself based on delicate a priori estimates due to John [56. where the functions a'lpur and pi>"E L 2 ( w ) are the same as in Section 2.~ ~ f l ~ ~ ~ ~ apuTRu. (8.I). (q)Rap(77)~ ~ ( q ) ~ ( r ~ ) ~ 2 for ad hoc vector fields q = & + &3 viai.. Then Ciarlet & Roquefort Ja(rl) . (4.
for smooth enough vector fields r] = viai : W + R3. defined by (cf. This result thus raises hopes that a rigorous justification.p(r]) of the change of metric. by means of rconvergence theory. coincides with that found by a formal asymptotic analysis of the threedimensional equations. and change of curvature. respectively denoted r a p ( r ] ) and pap(r]) below. and denotes the . tensors associated with a displacement field r] = viai of the middle surface S = O(W) of the shell. " b y definition". To begin with.let the function yap(r]) : W 4 R be defined by where a.One of its virtues is that the integrands of the first two integrals are quadratic expressions in terms of the covariant components Gap(r])and R. Given a displacement field r] := viai of the surface S = O(w) with smooth enough covariant components v i : W + R.7)  pi'Erli& d y .asp) and Radr])= b a d r ] ) . . Section 2. which we are about to describe.150 Philippe G.bap.R3) be an immersion. [. of these tensors by their linear parts with respect t o r].8 The linear Koiter shell model Consider the Koiter energy j for a nonlinearly elastic shell. Let w be a domain in R2 and let 8 E C2(W. of a nonlinear shell model of Koiter's type might be possible. Ciarlet.p(r]) = 2 ( a a p ( r ] ) . Theorem 2.p and a a p ( r ] ) are the covariant components of the first fundamental f o r m of the surfaces O(D)and (@+r])(g). 2. Accordingly. our first task consists in finding explicit expressions of such linearized tensors.81. In order to obtain the energy corresponding to the linear equations of Koiter [60]. to replace the covariant components 1 G. we compute the components r a p ( r ] ) . it suffices. with the thickness 2s as the "small" parameter. Cristinel Mardare [34] have shown that the leading term of a formal asymptotic expansion of a solution to this twodimensional model.
.uau + b a3 .v3)a' + (&v3 + bzv7)a3.1) &ar = r.p(q) of the metric tensor of the surface (0 v)(W) by definition given by are + The relations then show that hence that The other expressions of 'yap(v) immediately follow from the relation aav = &(viai)= (&vu   b. . We next compute the components pap(v).a". + a..lp are defined by valp = dpv.bapv3   1 ( ~ o T . v.vu. ] . 17 The functions yap(q) are called the covariant components of the linearized change of metric tensor associated with a displacement 77 = viai of the surface S . . itself a consequence of the Gauss and Weingarten equations (see Section 2. r".+ a a v ~. 7 7 Proof.r&vu ) bap773. The covariant components a. L &a3 = b. In particular then.An Introduction to Shell Theory 151 linear part with respect t o 11 in the expression [. .  where the covariant derivatives v. E H1(w) and v3 E L2(w) * ~ ~ ~ ( E L 2)( u ) . T h e n ~ ( v= )2 (apv 1 = #l'"lp  1 * a.v. ap) = ypa(v) + vpla) .
0. Given a displacement field r. = da8 satisfy (al A a2(= &.p and b.a l = b:/p. Ciarlet. and bzla are defined b y I n particular then. . In parts (i) and (ii). Let A denote the matrix of order three with al. satisfy the symmetry relations b.(a.bapl? where b.r.152 Philippe G.  where the covariant derivatives qalp. b. Proof.cpr/3 .82.b~Or/3+ b.let the functions p a p ( q ) : 2 + R be defined by P a d r l ) := [bap(rl) . I . Cristinel Mardare Theorem 2.p(v) are the covariant components of the second fundamental form of the surfaces 8(w) and ( 8 v)(W). E and 73 E H 2 ( 4 =+ Pap(77) E L 2 ( 4 . a3 = ppa(rl) r/3l0$  b. 7.pbz)r/T.r/T +(aab$ frLUb. l1 a A a21 Besides. Then + P a p ( r l ) = (aaprl   c p a u r l ) . where a = det (asp ) . r.b". i det A = (al A az) . the proof is divided into five parts.R3) be an := immersion. Let w be a domain in R2 and let 8 E C3(W.a 3 as its column vectors.r/glp + b z r / ~ j af b. we establish elementary relations satisfied by the vectors ai and a of the covariant and contravariant bases along S. The functions b. Consequently. .l. and denotes the linear part with respect to rl in the expression [.l.r/~ rz~r/U) aapr/3  r g p a u r / 3 . (det A ) 2= det(ATA) = det(aap) = a . 2 (i) The two vectors a. a2. . a3 = (a1 A a2) . For convenience. [. .Ur/7) +bg(dpr/U + r/31.~ Viai of the surface S = 8(W) with smooth enough and "small enough" covariant components r/i : w + R. a A a 2 = /a1A a21.
(ii) The vector fields ai and aa are related by a1 A a3 = *a2 and &a’.  where ((h.o.9 arc linearly independent in W and the fields rl = via. Hence lal A a3 A a 2 = a21 = 153 &. established.e. = &a2. a3 &a2 .t.) . p ( r l ) := [bap(rl) . Consequently.bap]lin = (daprl r ~ ~ & .p + (daprl . as) . e. ~ . = a. ai = d . a p = a. . Hence a1 A a 3 . since ai . a3 = Sa3. (a1 A (a1A as) .R3).= (&ap m + aaprl) . .2.~are “small enough”. on the one hand. on the other hand.rZPa.3}.S.0. = 0 and &a2 .(rl). a3 = fi. To prove that two vector fields c and d coincide. a 2 = (a1 A a 2 ) .g. Let + where Then 1 .p(rl) = b. (a1A a 2 + a1 A 8217 + dlrl A a2 + h.t. al = 0 and (a1 A a3) . it suffices to prove that c .. a3 = 0.0.a 2 = &. The other relation is similarly (iii) The covariant components b.r 3 ) pp. terms of order higher than linear with respect to 7. In the present case.aj = 6 .rl) a3 + h.. al = Since the vectors a. with respect to the norm of the space C1(i. i. a1 = &a2.p and a.An Introduction to Shell Theory since a.t. ai for i E {1.p(q) satisfy b.. The following computations arc therefore licit as they apply to a linearization around rl = 0.  since &a3 = al A a2 by (i). are smooth enough by assumption. the vectors & ( 9 q ) are also linearly independent in w provided the fields r.” stands for “higherorder terms”.
likewise.t. we find the announced expansion. Next. since [a.~ .~ .p(q)). ^ Consequently.0.. Noting that there are no linear terms with respect to q in the product (1 .p)(l + 28. . we note that det(A + H) = (detA with A := (asp) and A + H := (a. consequently. (r^a C T + ba0as) • (ai A d2v) = r ^ a 2 • (ai A d2ri) . To this end.o. • a0 (Theorem 2.t.). A a 2 ) = Va(r^g#i77 • a3 + b^dxT} • a 1 ). Therefore.t. b. aa + h..bapd2r] • (ai A a 3 ) • a2). There remains to find the linear term with respect to q in the expansion —1 . . Hence H = aprl .o.} .~. Cristinel Mardare since baj3 = daa0 • a3 and daa0 = r^a C T + bap&3 by the formula of Gauss.rl) . since A' = (amp).p(V) — aap]lin = aprl . a. + 8. since. i ^ 1 + a ^ • aCT) + ( ^ ^ .81). a a ) ( l + 8. ) . a2 • (ax A d2r)) = d2ij • (ai A a 2 ) = y/ad2r\ • a3 and ai A a3 = —y/aa2. a3 + h. • a" + h. Ciarlet. + a .0.154 Philippe G.a.p(V)) = det(a.~ .1 (11 a. viz. . ap + h.t.t.). .r £/A»7) • a 3 + h.p + {dapv ~ r&a.—(11 + .V . by (ii). a ' ) . a(v) = det(a.~.o. (iv) The components pap(ij) can be also written as where the functions % a / 3 and b^\a are defined as in the statement of the theorem.. a.
. da*T = T^a" + Vaa3. because of the formulas of Gauss and Weingarten.d0aaT = da ( . viz.r ^ T a T + b^a 3 ) a 3 = da& • a3 = ~batTa • a3 = 0.An Introduction to Shell Theory The Gauss and Weingarten equations.bT0)Vr ib0da b a17 • a3 = ( . = b^(d0r)a — I since .da0aT . daa3 = batTaT..r ^ 6 > 3 + (dabT0)a3 ~ b0baaa° . then that da0V • a3 = da 3 {(<H. it may be easily the functions r\r\a and bp~. Again.I ^ X + b}e?) . The relation TTacrbaprjr . we can write 0 .d0 {TTaaar + braa3) = (9arT0a)^ + T ^ r ^ a " .TaaTbT0r]a = 0! involves only the covariant derivatives rewritten so as to involve in addition stratagem simply consists in using the This gives (v) The functions bila are symmetric with respect to the indices Q and /3. While this relation seemingly rj3\af3 and rj^p. imply that \a '3 + & ^rja 3 155 . We thus obtain • a3 3 a ^. + V07]r )daa 3 •a 3 VT.r^Wfc)«i" + (5/3% f~b0Vr )a }a 3 ~\~{pa0ri3 + [dQ: ^ K + ^ Q ? ? T ) a 3 . a 3 +( .
. Remarks. we immediately infer from the definition of the functions b i l . Cristinel Mardare 0 = (&paT. but (2) The symmetry p a p ( q ) = pp. By contrast.8. The functions r/31ap = 8. Together with the expression of the components ~ ~ p ( in terms of q (Theorem 2.  dpb7. The functions p a ~ ( q are called the covariant components of the ) linearized change of curvature tensor associated with a displacement q = viai of the surface S .(q) must be linearly independent in a). (1) The functions b a p ( v ) are not always well defined (in order that they be.81)..pqr'". Koiter [60]).D (cf. that we also have and thus the proof is complete.. defined here by means of its mixed components b. On the other hand. we were led t o assume that 8 E C 3 ( W . 0 While the expression of the components pap(q) in terms of the covariant components qi of the displacement field is fairly complicated but well known (see.la appearing in the functions pap(q).la = dab.0773 .p8..8paaT) a3 = dab. The culprits responsible for this regularity are the functions b'.q). + rLUbz respectively represent a secondorder covariant derivative of the vector field viai and a firstorder covariant derivative of the second fundamental form of S . E H1(w) and 773 E H 2 ( w ) .R3) in Theorem 2. part (v) of the proof of Theorem 2.8l).156 Consequently.p(q) requires proving first that the covariant derivatives bzla are themselves symmetric with respect to the indices a and . the vectors a. who showed that their principal merit is to afford the definition of the components ~ ~ p ( q ) and pa0 (q) under substantially weaker regularity assumptions on the mapping 8. this simpler expression q) was efficiently put t o use by Blouza & Le Dret [13]. the functions p a p ( q ) are always well defined. deriving the same symmetry from the other expression of p.773 and b. Philippe G.  on the one hand. Ciarlet. e. that in terms of q = viai is remarkably simple but seems t o have been mostly ignored.82 in order to insure that pap(q) E L 2 ( w ) if q = viai with 7. Otherwise . More specifically. + r'a" b"P I7iUb.g.r.(q) follows immediately by inspection of the expression pap(q) = (d. although it already appeared in Bamberger [lo].a3 found there.
Let y be a measurable subset of y = dw that satisfies lengthyo > o 0. q E ~ ~ ( w ) .An Introduction to Shell Theory 157 Blouza & Le Dret [13] have shown how this regularity assumption on O can be weakened if only the expressions of y a p ( q ) and pap(q) in terms of the field q are considered. Equivalently. denote the outer normal derivative operator along dw. the boundary conditions qi = 8. 3 qi Then the displacement field C" = <Faiof the middle surface S = O(is) of the shell (the covariant components are unknown) should be a stationary point over the space V ( w ) of the functional j defined by < . yap(q) and pap(q) denote the covariant components of the linearized change of metric. and change of curvature.= dvr/3 = o on yo}. ) va E ~ ' ( w ) .773 = 0 on y express that the shell is clamped along the portion O ( y 0 ) o of its middle surface (see Figure 2.41). We arc now in a position to describe the linear Koiter shell equations. for all q = viai E V(w).&E L 2 ( w ) account for the applied forces. This functional j is called Koiter's energy for a linearly elastic shell. The choice of the function spaces H1(w) and H 2 ( w ) for the tangential components qa and normal components 73 of the displacement fields q = viai is guided by the natural requirement that the functions y a p ( q ) . the vector field = <fai E V(w) should satisfy the variational equations <" We recall that the functions denote the contravariant components of the shell elasticity tensor (A and p are the Lam6 constants of the elastic material constituting the shell). Finally. let 8. and let the space V(w) be defined by ~ ( w := { q = viai. and the given functions pi. tensors associated with a displacement field q = viai of S .
E pao(rl) := { (aaprl r g p a D r l ) . %P Such a key inequality is an instance of a Korn inequality on a surface. see Section 2.so that the energy is in turn well defined for 71 E V ( w ) . 2. Ciarlet. The objective of this section is to establish such an inequality.101) that. as a consequence of the lemma of J. under the assumptions 3X + 2p > 0 and p > 0. Otherwise these choices can be weakened to accommodate shells whose middle surfaces have little regularity (see Blouza & Le Dret [131). For more details. Given 77 = viai with qa E H 1 ( w ) and 7 E H 2 ( w ) .+. 0 Chapter 71 and the references therein. there exists a constant ce > 0 such that [ t a p ( 2 ceaa~uT(y)tuTtap c a.6.51). To begin with. a3) E ~ ’ ( w ) ya)’ap(rl) :=  . Ciarlet & Miara [la]). am+ aarl. Let w be a domain in R2 and let 8 E C3(W. we shall see (Theorem 2.P < for all y E and all symmetric matrices ( t a p ) . Remark. “without boundary conditions”. we establish a Korn’s inequality on a surface. Theorem 2. Lions (cf. see Ciarlet [20.R3) be an injective immersion. Theorem 1.158 Philippe G. To this end.L. Koiter’s linear equations can be fully justified by means of an asymptotic analysis of the “threedimensional” equations of linearized elasticity as E + 0. We follow here Ciarlet & Miara [33] (see also Bernadou. When lengthyo > 0. Cristinel Mardare and pag(q) be both in L 2 ( w ) . a@)} ~2(w). the existence and uniqueness of a solution to this variational problem by means of the LaxMilgram lemma will then be a consequence of the existence of a constant c such that c Ilvall%l(w)+ +c { a 1/2 llv3IlLyd)} < c{ c %P IlryaP(rl)llLz(w) 2 II~~~(~])JI. 3 let { z1( a p r l .) for all rl = qiai E ~ ( w ) .91.9 Korn’s inequalities on a surface Our objective in the next sections is to study the existence and uniqueness of the solution to the variational equations associated with the linear Koiter model.
: bz . with qa E L 2 ( w ) and 7 3 E H1(w).P llraP(r]>112L2(w) c %P llPaP(r])l12L2(w)}1’21 the space W ( w ) is a Hilbert space.r’zp b:) qT p} d y .(b:cp) + b:r. tensors associated with the displacement field r ] = viai of the surface S = @(a). are used in this proof. / YCYP(r])(PdY =  W s.82. . (i) Define the space W ( w ) := { q = viai.(8..vTcp + vTaa(b. P ~ ~ ( v ) ( P= Y ~  {8av3%3cp + JZp%v3cp + b3. equipped with the n o r m 11r]11W(w) := ~ ~ ~ ~ ~ w ( w by defined ) { c a Ilva11~22(w) + llv311:l(w) + + c aY.An Introduction to Shell Theory 159 denote the covariant components of the linearized change of metric. simply because they are more convenient for its purposes. . belongs to W ( w ) if there exist functions in L 2 ( w ) . @ )such that f o r all q = viai with qa E H1(w) and 73 E H2(w).cp) + b. Proof. as found in Theorems 2. They mean that a vector field r ] = viai. va E L ~ ( w ) E ~ l~( w ) . denoted r a p ( r ] ) and pap(?). b i + r.81 and 2.vocp . The relations “yap(r]) E L2(w)” and “ p a p ( r ] ) E L2(w)” appearing in the definition of the space W ( w ) are to be understood in the sense of distributions. { / W 1 Z(VPd*cp + va%(P) + r : p w + bapV3cp} dY. such that for all cp E D ( w ) . E Then.r:. The “fully explicit” expressions of the functions rap(q) and p a p ( r ] ) . and linearized change of curvature. h e n there exists a constant T c0 = c o ( w .p773cp + vua. ~ raa(r]) L2(w)1 P a d r ] ) E L 2 ( 4 } .
To prove the 7 other inclusion.aasp&d) E f. {q = %ai. and ad. let q = v a E W ( w ) . Lions (Theorem 1. Since both spaces are complete (cf. The identity mapping L from the space { r ] = viai. continuous. . ~) and surjective by (ii). yap E L 2 ( w ) .letting lc 00 in the relations y a p ( qk )cpdw = .51) and thus qa E H 1 ( w ) .v.~is also continuous or equivalently.o(rl) + @7sap(v) . the continuity over W of the functions bop. ( ( w ( w ) shows that there exist va E L 2 ( w ) .p are then imply that sap(?) E L 2 ( w ) since the functions continuous on w. Given a function cp E D ( w ) .and pap 7.r/3 E E l 2 ( # ) } equipped with the norm rl = viai c IlvallLl(w)+ c a 11v311L2(w)}1'2 into the space W ( w ) equipped with (1 . since x E L 2 ( w ) implies a.L.160 Philippe G. 0 . that the inequality of Korn's type without boundary conditions holds.173 E H 1 ( w ) . .x E H'(w). E H 1 ( w ) . . The definition of the functions pap(q). ( ( w (is injective. + Pap in L 2 ( 4 + as k + 00.. hence that q 3 E H 2 ( w ) . the open mapping theorem then shows that the inverse mapping L . 73 E H 2 ( w ) } c W ( w ) . Clearly. 77. (ii) The spaces W(w) and {q = viai. The relations i2 sap(rl) := 1 pave + 43va> = yap(r1) + r". ( then shows that yap = y a p ( q ) and pap = pap(q). + bapv3 and b.r/3 E H 2 ( w ) } coincide. . (iii) Korn's inequality without boundary conditions. Cristinel Mardare with elements q k = Let there be given a Cauchy sequence (qk)& vtai E W ( w ) . and L p ap qk)cpdw = . Ciarlet. Hence aova E L2(w)by the lemma of J. T$ + 7ap(vk)+ Yap in L2(4. aova E Hl(w).va E H 1 ( w ) . and the relations p a p ( q ) E L2(w) then imply that dap773 E L2(w).. va E H1(w). The definition of the norm ( ( .Therefore. (i)). 4 E L 2 ( w ) such that va in ~ ' ( w ) . ap(aova) = {aps. pap(vk) 73 in ~ ' ( w ) .l(w). bg.
i. immersion 8 E C3(w. Ciarlet & Miara [12. Theorem 61. such as the one originally found by Bernadou & Ciarlet [ll] is surprisingly “technical”). see also Bernadou. . we need an infinitesimal rigid displacement lemma “on a surface” (the adjective “infinitesimal” reminds that only the linearized parts r.l(w) c Q + l177311. I I L Z ( ~ ) and in the righthand side of the above inequality.11 and 5. E H 1 ( w ) and 773 E H 2 ( w ) .p) and (b.An Introduction to Shell Theory 161 In order to establish a Korn’s inequality “with boundary conditions” . E H 1 ( w ) and 773 E H 2 ( w ) be such that r a p ( r l ) = pap(77) = 0 in w .2(w)}1’2. In the next proof. To this end. “with boundary conditions”. “without boundary conditions”. “threedimensional infinitesimal rigid displacement lemma in Cartesian coordinates” used in part (iii) of the next proof (a “direct” proof. The one given here is not the shortest. Let there be given a domain w in EX2 and an injective EX3).61 and Blouza & Le Dret [13. the functions E i j ( G ) and eij(v) are the Cartesian and covariant components of the “threedimensional” linearizecstrain tensor respectively associated with a displacement field G i 8 : (0). (a) Let q = viai with 7. we have to identify classes of boundary conditions to be imposed on the fields 71 = viai. delicate. but it is a natural one: It relies on the classical.a. Any proof of this theorem is. situations where the seminorm becomes a norm. while part (b) is a n infinitesimal rigid displacement lemma on a surface. Part (a) in the next theorem is an infinitesimal rigid displacement lemma on a surface..5 and 2.in order that we can “get rid” of the norms ) ) ~ ~ .p) are required to vanish in w ) . Lemmas 2. which is due to Bernadou & Ciarlet [ll. the functions are the Christoffel symbols (of the second kind) associated with the mapping 0 : 2 f {0}. h Theorem 2.4 EX3 and a displacement field vigi : EX3. the functions vuiiljare the covariant derivatives .92. which should be in addition equivalent to the norm rl = via2 c llvall.p(q) .p(r]) and p a p ( q ) of the “full” change of metric and curvature tensors i ( a a p ( r l ) .finally. 0 of the vector field vig’ : 0 + EX3.211.b. at least to some extent. and much easier t o prove. with r].e.Theorems 5. .
i We now show that. for all x E 0. gjGi(2)= (uUk. &gj.. in R3 and 0 : + R3 deConsequently. .162 Philippe G. (i) I n parts (i) to (iii).G3}.are linearly indewhere pendent at all points 3: = (xi) E 3 and the three vectors gi(x)defined by gi(z)gj(z) 6. i t h any vector W . notes a C2diffeomorphism from fi onto its image the three vectors gi(x):= &0(3:).r ?‘3 vP and r?. and let R := @(a). n Proof.the chain rule shows that the matrices VO(x) ( a j @ ( z ) ) (the := row index is k) and V6@) (dkO’(2)) (the row index is i ) satisfy := h h . Gi := denotes the ith component of gk(z) over the basis {Gl.R3) and gz E C1(a. 2113 . are likewise linearly independent at all points 3: E = Note also that gi E C1(a.+ R3 by letting i a @(a). Cristinel Mardare Then there exist two vectors a. . for It is then clear that G E H 1 ( 6 ) . let 2i deno_te the Cartesian coordinates of a point 2 E EX3. field v = uig’ : R + R3 with u E H’(R). (b) Let y be a dymeasurable subset of y = dw that satisfies length o y > 0 and let a vector field q = viai with va E H 1 ( w ) and 773 E H 2 ( w ) o be such that yap(q) = . G i ( $ ) 2 = vi(x)gi(x) all 2 = ~ ( x )x. = Then q = 0 a w .Iw3). where 6 : 6 + R3 denotes the inverse mapping of 0 : R + R3.oaa(q) 0 in w and qi = dVv3 = 0 on yo. R denotes a domain. b E R3 such that q ( y ) = a + b A O(y) for all y E a. The proof is divided into five parts. x ) .G2. E G. di = d/axi. let di := a/d2i. the simultaneous appearance of 2 and x in an equality means that they are related by 2 = O ( x ) and that the equality in question holds for all 3: E R.83. . Ciarlet. Let O(3:) = O‘(cc)Gk and 6(2) = @(2)ei. Let Gz denote the b@s of R3. numbered (i) to (v). 2=q where u. = . a. . Since 6(0(3:))cc for all = cc E R. In what follows. V6(2)V@(x) I.~e[~”li[~elj)(”). we then associate a vector field i 3 = Gg : {6}. v .:= gp .‘3 2 and [gk(x)]z g“3:) .
An Introduction to Shell Theory or equivalently. we first note that a differentiable function w : R t IFS satisfies al.R3) and 6 E C'(6.. & } can be also expressed in terms of the inverse mapping 0. this TO end. we observe that r.r:k(+%(x)) + + 4 (~e[S"4li) [ge(41j [Sk(41i[ge(x)lj. noting that &(gq(x) . 2 . the last h relations show that E C0(R). Hence the kth component of gZ(x) over the basis { & .w(6(2)) aew(X)5j@((a) = by the chain rule. 5&2) = 5. = We next compute the derivatives &gq(x) (it is easily seen that the fields gq = gqTgTare of class C1 on R since 0 is assumed to be of class C2). we may write a priori amq. We are now in a position to compute the partial derivatives @$i(?) as functions of x. = aew(x)[ge(41j. gk(x)) = 0 and [gq(x)lP= aPOq(2).k(x)= r . [gi(x)]k dkO"2).R3) by assumption. as: h h A . Recalling that the vectors gk(x)form a basis. m ( x ) q = r. 163 The components of the above column vector being precisely those of the vector gj(x). To find their expressions in terms of the mappings 0 and 0. These derivatives will be needed below for expressing the derivatives GGi(2) as functions of x (recall that Gi(2) = uk(z)[gk(x)]i). gj(x) = d. = A h Hence.j = 1 . aegk(x) ap@q(2)dlk@p(2) r & ( x ) . 3 .vk(6(2))[g~(z)]ivq(z)8j[gq(6(2))li = aevk(~)[ge(~)lj[s"~)li 4 = ( a e v k ( 4 . In particular then.gk(x) aegq(x) . G 2 .( ) = r&(x)gk (x) 1 A thereby unambiguously defining functions : R + R.gk(x). = = Since 0 E C2(R. components of the above row vector must be those the of the vector gi(x)since gi(x)is uniquely defined for each exponent i by the three relations gi(x). we obtain h h r & ( x ) = gq(x). .by means of the relation Gi((a) = vk(z)[gk(~)]i.m(x)gm(x) .
It follows from part (ii) that Then the identity . T h e n there exist two vectors a. G E H 1 ( 6 ) we i and.i 311 = . e ( 4[ g k ( 4 1 i [ g e ( x ) l j .r:jvup. h If the Jields w and 3 are related as in (i). and [ g k ( x ) ]and r&(z)are defined as above.E R. Ciarlet. Cristinel Mardare since &gg(x) = . (x) x (iii) Let a vector field w = vigi with vi E H1(R) be such that eij(w) = 0 in R. 2 .b E R3 such that the associated vector field viga is of the f o r m w(x) = a + b A O(z) for all x E a.r & ( x ) g k ( x ) .q k ( x ) V q ( Z ) . let l?o be a drmeasurable subset of the boundary 6'0 that satisfies arearo > 0.(ajvi 1 + aivj). and let a vector field w = vigi with vi E H 1 ( 0 ) be such that eij(w) = 0 in R and w = 0 on Fo. i (ii) With any vector field w = vigz : R + R3 with vi E H1(R). with any vector field G = Gigi : 6 + EX3 with associate the functions Z i j ( G ) E L 2 ( 6 )defined by h e i j ( 3 ) := (ajGi 2 1  + aiGj). Next. ' k l l e ( 4 := aevk(x) . it t h e n immediately follows f r o m (i) that Zij(G)(Z) = (eke(w)[gk]i[ge]j) for all 2 = ~ ( x ) .164 Philippe G. We have therefore shown that @i(Z) where = v k l . T h e n w = 0 in R. we next associate the functions eij(w) E L2(R) defined by ' e i j ( v ) := (villj 2 1 + v .
.g. z 3 ) g i ( Y . p. where the vectors g i are defined by g . T h e n the covariant components I J ~ of the vector field vigz are in H 1 ( R ) and the corresponding functions e i j ( v ) E L2(R) defined as in part (ii) are given by eayP(v) = . Theorem 6. Hence there exist two vectors a. each function Gi istherefore a polynomial of degree 1 in the variables z j . let there be associated the vector field v = vigi defined o n 2 by v z ( Y .. + where EO > 0 has been chosen in such a way that the mapping 0 is a (Theorem 2. With C2diffeomorphism from into its image any vector field v = Viai with covariant components r]. e.21).An Introduction to Shell Theory further shows that 165 By a classical result from distribution theory (Schwartz [80.Yaap(77) . see. b E R3 such that (2denotes the column vector with components 2i) G~(z)c? + b A 2 for all 5 E 6.x g ) := qY) z g a g ( y ) for all x = (xi):= ( y . z 3 )= r ] i ( Y ) a z ( Y ). E < h h + h But Eij ( G ) = 0 also implies that bij = bji. Ciarlet [18.z3PaL3(77) .&g). = Since the set where such a vector field G i 2 vanishes is always of zero area unless a = b = 0 (as is easily proved. and we let the mapping 0 : 2 4 R3 be defined by O ( y . There thus exist constants ai and bij such that q ( z ) = ai b i j 2 j for all 2 = (2i) R. (iv) W e now let R := wx (Ei3. such that vi(z)gi(z) a + b A O ( x )for all x E R. the assumption arearo > 0 implies that G = 0.z g ) E 2. in H 1 ( w ) and r]3 in H 2 ( w ) . SO]). =a or equivalently. since the set R is connected.34]). g j = 6.53(&xr]3 b:r]u)(Y)aa(Y) @(a) + z f o r all ( y .xg) E R.
but the dependence with respect to y E W is omitted. a + Then the functions vi are in H1(R). = a. by the formula of Weingarten (see Section 2. To prove the above assertion. bza.y~(q) in terms of the functions qi (Theorems 2. + xa = (&PI3 + bEVu). g i E H 1 ( R ) since g i E The announced expressions for the functions eij(v) are then obtained by simple computations.81 and 2. let the vector field v = vigi be defined on by vigi = viai x3 x. First.1). Hence. as in the above expressions of the functions e.a3 = = 0. based on the relations q j = { a j ( v d ) )' g i (part (i)) and e i j ( v ) = i(uillj vjlli).aa.166 Philippe G. The explicit expressions of the functions y a p ( q ) and p.yp(v).x3b:au and g 3 = a s . we note that a. . throughout the proof.g i i = (Vjaj + x3Xaa") . E H 1 ( w ) and 773 E H 2 ( w ) . the functions eiilj(v) defined as in part (ii) are given by covariant derivatives of the fields qiai and Xiai with X3 To see this. we show that. the vectors of the covariant basis associated with the mapping 0 = 8 x3a3 are given by g . the functions e i j ( v ) above take the expressions announced in the statement of part (iv). given . Besides. Ciarlet. the dependence on x3 is explicit. + The assumed regularities of the functions qi and X . we proceed in two stages. Cristinel Mardare Note that. Second. . functions v X . when C'(a).82) are used in this part of the proof. imply that u = ( v j g j ) .
part (iii) implies that w = 0 in on J and part (b) of the theorem is established. .p(w) is equal to . gi = 0 on := 70 x [ED. the assumption that yap(r])= pap(r]) = 0 in w implies that eij(v)= 0 in R.O + XOIa .+."bC7?73) + b. z3 i.bpu773) .b.yUT(rl). I Hence and part (a) of the theorem is established..'%lap . 2 3 )E T.b F u l p .% + b:qula) )  = bE (PP&?) b&U + bpTLT773) + bg ( P A ? ) . .77~ f bgbop773.) vanish on yo. / 3 1 . b E vi(y. there exist two vectors a. expressions and using the symmetry relations bglp = b.T(r]) i."77+  + b.since 773 = dv773 = 0 on y implies 8.Ja (Theorem 2.. gi = (7j. Finally. E ~ ) and let the vector field w = vigi with wi E N1(R) be defined as in part (iv). If in addition qi = o &773 = 0 on yo.An Introduction to Shell Theory 167 We first note that X E H1(w) (since bg E C1(a)) and that e a 3 ( w ) = .'%Ia .b .773 bgq.773 = 0 on yo.. (v) Let the set fi = w x ( . b g q g ) .. Since arearo > 0."j. Therefore. E ~ ] .e. "with boundary conditions". the factor of $ in e.P. By part (iv). G.b:(77rlp . is indeed as announced. Replacing the functions X by their .bil. hence that q =0 0 We are now in a position to prove the announced Korn's inequality on a surface.b. 0 when X = (&vs .e. ~+ b.Ppdrl) 2b:b.p(w> in this case.bryT773)} . = (ujgJ) .p a p ( q ) .b.82). It thus remains to find the explicit forms of the functions e. Let next y c y be such that lengthy0 > 0. the functions xa = (8.  in e. the factor of . by part (iii). x3)gi(y7 = a z3) R3 such that + b A {B(y) + z 3 a d y ) ) for all ( ~ . bE (r]31pg + b&3r/~ + h z v ~ l p + bg ( 1 .p(w) = b.J + z3Xaaa) .bz'%lP .E o . we find that + 1 S{XaI. Part (iv) then shows that o + v. b T / .
and let the space V ( w ) be defined as: V(W) := { q = viai. It was later given other proofs by Ciarlet & Miara [33] and Bernadou. 3 Given q = viai with qa E H 1 ( w ) and q 3 E H'(w). Let the space v ( w ) := { q = qiai. let y be a dymeasurable subset of y = dw that satisfies lengthyo > 0. E~ ~ ( w ) . v 3 E equipped with the norm  ff2(~)} be If the announced inequality is false. { (aapv  ~ 0 8. there exists a sequence of vector fields q'"E V ( w ) such that ($)El .93. Ciarlet & Miara [12]. Let w be a domain in R'. let 8 E C3(L7. Theorem 2. Then there exists a constant c = c(w. 0) such that f o r all q = viai E ~ ( w ) .then by Blouza & Le Dret 1131. as} E~ ~ ( w ) ~ ~ ) denote the covariant components of the linearized change of metric and linearized change of curvature tensors associated with the displacement field q = viai of the surface S = 8(W). ~ ap} . qa E H 1 ( w ) .168 Philippe G. qa E ~ ' ( w ) . proof. let 1 ~ ~ p ( r := {. Ciarlet & Miara [12]. v ~= dVr/3= 0 on yoyo). then by Akian [4] and Ciarlet & S. Mardare [32]. We follow here the proof of Bernadou. Ciarlet. r ]E H ' ( w ) . who showed that it still holds under a less stringent smoothness assumption on the mapping 8. who showed that it can be directly derived from the threedimensional Korn inequality in curvilinear coordinates (this idea goes back to Destuynder [45]).yo. l ) ~ a D ( r l:= ) +8 .(am.R3) be an o injective immersion. a. Cristinel Mardare This inequality was first proved by Bernadou & Ciarlet [Ill.
( ~ ~ p ( r f ( ~ ) andF( ~ a p ( v u ( k ) ) ) & also converges in L2(w) (to 0. and the proof is complete. +w Hence 7 = 0 by Theorem 2. each sequence the RellichKondraSov theorem. the inequality of Theorem 2. and . then 773 = 0): for all q3 E H 2 ( w ) satisfying 7 3 = dvv3 = 7 0 on yo. But this contradicts the relations ~ ~) Ilv4 k ) I I H ~ ( ~ ) ~ H ~ ( ~ ) ~ H =( 1 for all /C 2 1.91). y y ~ = (yl. but )) p this information is not used at this stage) since The subsequence the norm (v"('))g1 is thus a Cauchy sequence with respect to hence with respect to the norm II.An Introduction to Shell Theory 169 Since the sequence is bounded in P ( w ) .93 reduces to two distinct inequalities (obtained by letting first qa = 0. a subsequence (qg(k))r=l (u : N 4 N is an increasing function) converges in e ( w ) by Furthermore.O) for all ( y l . IIPaP(v)IIL2(W) = klim I I P a p ( v u ( k ) ) l l L 2 ( W )= 0. If the mapping 0 is of the form O(yl.92.y2) E ) i3. The space V ( w ) being complete as a closed subspace of the space V ( w ) .y2.IIHl(w)xHl(w)xH2(W) by Korn's inequa2ity without boundary conditions (Theorem 2. there exists v E V ( w ) such that vU(lc)v in P ( w ) + and the limit v satisfies Il"laP(rl)llL2(w) :=: 2 IITap(v'(k))IIL~(w)= 0.
The second inequality is the twodimensional K o r n inequality in Cartesian coordinates. Theorems 1.51 and 1. and Anicic. It is remarkable that. As a result.1 and 6. Cristinel Mardare for all 17. By contrast.93) is now replaced by the norm 11173IIL~(~). a compact surface S “without boundary’’ (such as an ellipsoid or a torus) is defined by means of a finite number I 2 2 of injective immersions 8 E C3(Wi). in the sense that all its points are elliptic (this means that the Gaussian curvature is > 0 everywhere on S ) . the regularity assumptions made on the mapping 8 and on the field 11 in both the infinitesimal rigid displacement lemma and the Korn inequality on a surface of Theorems 2. Ciarlet [19. 1 Remarks. for specific geometries and boundary conditions.. (1) The norm )lr/311H2(W) appearing in the lefthand side of the Korn inequality on a “general” surface (Theorem 2. More specifically. 8)> 0 such that for all 1 = qiai with q. Both play a central rijle in the existence theory for linear twodimensional plate equations (see. 0 Only compact surfaces defined by a single injective immersion 8 E C3((w) have been considered so far.91) and the infinitesimal rigid displacement . Mardare [67]. As shown by S.IW3) be an injective immersion with the property that the surface S = 8(a)is elliptic. Le Dret & Raoult [7]. E H 1 ( w ) satisfying q.521). Ciarlet. This replacement reflects that it is enough that qa E H 1 ( w ) and 173 E L2(w) in order that yap(^) E L 2 ( w ) .51.g.92 and 2. The first inequality is a wellknown property of Sobolev spaces. e. 1 6 i 6 I . E H:(w) and q3 E L2(w). Le Dret [61]. no boundary condition can be imposed on 7l3. Theorems 6.170 Philippe G. where Q = Viaz. Then there exists a constant C M = C M ( W . Ciarlet & Lods [27] and Ciarlet & SanchezPalencia [35] have established the following Korn inequality “on an elliptic surface”: Let w be a domain in R2 and let 8 E C2’’(W. i where the sets wi are domains in EX2. in such a way that S = UiE1Bi(wi). a Korn inequality can be established that only involves the linearized change of metric tensors. As shown by Blouza & Le Dret [13]. under the additional assumption that the Co(5)norms of the Christoffel symbols are small enough. (2) The Korn inequality on an elliptic surface was first established by Destuynder [45. the Korn inequality “without boundary conditions” (Theorem 2. = 0 on yo.93 can be substantially weakened.
(w. let 6 E C3(W. we establish the uniform positivedefiniteness of the elasticity tensor of the shell. = c. uniqueness. A. let amp denote the contravariant components of the metric tensor of the surface 8(Z). In what follows. E H ~ ( w ) qi = d. let y be a measurable subset of y = d w that o satisfies lengthyo > 0.10 Existence.92) can be both extended to such surfaces without boundary. Let w be a domain in EX2. Our primary objective consists in showing that the bilinear form B : V(w) x V(w) 4 IR defined by for all E V(w) x V(w) is V(w)elliptic. denote the outer normal derivative operator along dw. given here by means of its contravariant components aapuT (note that the assumptions on the Lam6 constants. and regularity of the solution to the linear Koiter shell model Let w be a domain in R2. 3 X 2p > 0 and p elastic materials).101. 2.R3)be an immersion] and let the space V ( w ) be defined by q ~ . let d. where yo is a dymeasurable subset of y := dw that satisfies lengthyo > 0. T h e n there exists a constant for all y EW and all symmetric matrices ( t a p ) . > 0. As a preliminary. Proof. p ) > 0 such that > 0 . let 8 E C3(W. M2 and S2 respectively designate the set of all real matrices of order two and the set of all real symmetric matrices of order two. 8 . qa E ~ l ( w ) . let the contravariant components of the twodimensional elasticity tensor of the shell be given by and assume that 3 X + 2 p > 0 and 11 c. (c.q3 = o on yo). V(w) := { q = viai.~) + viz..An Introduction to Shell Theory 171 lenima on a surface "without boundary conditions" (Theorem 2. . are weaker than those usually made for Theorem 2.IR3) a n be injective immersion.
e. this inequality holds with y = 21. Then there exists a constant y = y(x. let A(!/) (a%)) = and T = (tap). Cristinel Mardare (i) To begin with. Since there clearly for any B = A . Given any matrix I E M2.p < x < 0 and p > 0.p > 0. we thus obtain X(tr B ) 2 + 2p tr(BTB) = ( A B ): B = C : A'C > 0.p) > 0 such that B :B < . Let x and p be two constants satisfying x + p > 0 and p > 0.p) > 0 such that ytr(BTB) < X(trB)2 + 2ptr(BTB) for all B E M2. If x 2 0 and p > 0. let K ( y ) E S2 be the unique square root of A ( y ) (i. The linear mapping A : M2 f M2 defined in this fashion can be easily inverted if x t. ' = p(x.172 Philippe G.l C E MI2.p <x <0 and p > 0. p ) > 0 is the constant found in (i).l C E M2 if . we establish a crucial inequality.p < x < 0 and p b exists a constant . and let )~ B ( y ) := K ( y ) T K ( y ) S2. (ii) We next show that. a"fluT(Y)tuTtap 3 yaau(Y)aP7(y)to. where y = y(x. the announced inequality also holds if . define the matrix C E M2 by 3 C =AB := X(tr B)I + 2pB.p # 0 and p # 0.t. Ciarlet. as Noting that the bilinear mapping (BIG')E M2 x & + B : C := t r ( B T C ) defines an inner product over the space M2. It thus remains to consider the case where . for any y E W and any nonzero symmetric matrix ( t a p ) . Given any y E W and any symmetric matrix ( t a p ) .BC : C for all B = A .. with y = (2pp)l in this case. E . the unique positivedefinite symmetric matrix that satisfies ( K ( Y )= A ( y ) ) .
Finally. Theorem 2. o R3) be an injective immersion. let there be given constants X and p that satisfy 3X 2p > 0 and p > 0.e. with ce := (yb)l.102. there thus exists a constant a ( X . Let w be a domain in R2..93).Ttap a. J 0 Combined with Korn's inequality "with boundary conditions" (Theorem 2. a solution to the variational equations of the linear Koiter shell model. or equivalently.tap for all y E i and all symmetric matrices ( t a p ) . let y be a subset of y = dw o with length y > 0.An Introduction to Shell Theory Then 173 By the inequality established in part '(i).q3 E H 2 ( w ) . Then there is one and only one solution = <fai to the variational problem: + <" <' E V(W) = { Q = Viai.qi = dyq3 = o on yo}. we infer that Hence ltapI2 6 aau(Y)aflT(Y)t. the positive definiteness of the elasticity tensor leads to the existence of a weak solution. p ) > 0 such that aaOn~ 2 1 (y)tU. i. qa E H1(w). and functions pa>€E L T ( w ) f o r some r > 1 and p3+ E L 1 ( w ) . .tap 2 a t r (B(Y)TB(Y)) if x + p > 0 and p > 0. if 3X + 2p > 0 and p > 0 (iii) Conclusion: Since the mapping is continuous and its domain of definition is compact. and let 8 E C3(w..P and thus c %B ItapI2 6 ceaapUT(y)t.
Ciarlet.101) together imply that 2 min {e.93) and the uniform positive definiteness of the elasticity tensor of the shell (Theorem 2. Since the symmetric matrix (u. The continuous embeddings of the space H1(W) into the space Ls(w) for any s 3 1 and of the space H2(w) into the space C°(W) show that the linear form defined by the righthand side is continuous over the same space. The assumptions made on the mapping 8 ensure in particular that the vector fields ai and ai belong to C2(w.T^daW • a = (dpr] • aa + dar) • a/3) 3 The field C e V(u) is also the unique solution to the minimization problem: where / J UJ Proof. Hence the bilinear form B is V(w)elliptic. r]a G H1(w).s. there exists a0 such that a(y) 3 a0 > 0 for all y E i.p(y)) is positivedefinite for all y E a. Finally.773 e H2(LJ)} equipped with the hilbertian norm 1 /9 := { Z^ ll7?"! the space V(W) is a Hilbert space. As a closed subspace of the space V(UJ) := {n = ^ a l .174 where aap*T = 1 Philippe G. and pap{r}) = (da0Ti . the Korn inequality "with boundary conditions" (Theorem 2. Fgp. Hence the bilinear form defined by the lefthand side of the variational equations is continuous over the space V(W). and a arc continuous on the compact set w. y j c j V 2 ^ ^ \\va\\m {uj) for all 17 = v i a i E V(w). . Cristinel Mardare J±t_aa0a°r + 2ll(a a °a^ + aaTa^).R3) and that the functions a apuT .
. In what follows. equivalent to the variational equations of Theorem 2. 7 1 := .2(w)}.R3) a n be injective immersion.Let w be a domain in R2 and let 0 E C3(W.103.e. i. := a a ( t a p l p ) + r. for a n arbitrary tensor field with smooth enough covariant components t a p : w 4 R.. we give the proof only in the case where 7 = y. (ua)is the unit outer normal vector along 7.102 reduces to V(W) = (7= %ai. Theorem 2.l(w). those that correspond to lengthyo > 0. and arx := rJax denotes the tangential derivative of x in the direction of the vector (ra). 0 i.2 4 . 72 := u1. and.102 is smooth enough. when the space V ( w ) of Theorem 2.e. . taqp taplap + rpaqtPq+ r$qtaq.q(taPlp).An Introduction to Shell Theory 175 The LaxMilgram lemma then shows that the variational equations have one and only one solution. then C" is also a solution to the following boundary value problem: where n a p := EaapqrYu7(<") and mag := aapqr 3 E3 Pur (C"). If the solution C" = <:ai t o the variational equations found in Theorem 2. We next derive the boundary value problem that is. this solution is also the unique solution of the minimization problem stated in the theorem.r/3 E H.V a E H. For simplicity.102. The extension to the case where lengthy1 > 0 is straightforward. Assume that the boundary y of w and the functions pi>€are smooth enough. 7 1 := 7\70. at least formally. 0 The above existence and uniqueness result applies to linearized pure displacement and displacementtraction problems. := a p t ap Proof. Since the bilinear form is symmetric.
ap = r&aU + bapa3 (see Section 2.176 Philippe G. (ii) Using the Green formula in Sobolev spaces (see.a2. This gives.g. hence a fortiori for all q = viai with va E H t ( w ) and r/3 E H i ( w ) . we assume that the solution C' is "smooth enough" in the sense that n a p E H1(w) and map E H2(w). Consequently. for all q = viai with 71.1).a3 as its column vectors. since 8. e..we first transform the first integral appearing in the lefthand side of the variational equations. E H. (iii) We then likewise transform the second integral appearing in the . so that f i = det A (see part (i) of the proof of Theorem 2. (i) We first establish the relations Let A denote the matrix of order three with a1. NeEas [73]) and assuming that the functions nap = npa are in H 1 ( w ) .(w) and q 3 E L 2 ( w ) . Cristinel Mardare In what follows. Ciarlet.102).
b~b.. part (i)). we obtain &(apmaP + r&.)mup f b. viz. & { m a p l a p . 177 forhallq = qzatwith qa E H i ( w ) and q 3 E H i ( w ) .Q(mUplp) . Using the symmetry map= m p a .pmap}mdy. and the same Green formula as in part (ii).An Introduction to Shell Theory lefthand side of the variational equations."mUP)lp = (bFI. Using in this relation the easily verified formula (b. h = f i r & (cf. the relation a p .map+ l?:pmuP)daq3 dy Consequently. + S.
a solution of the associated boundary value problem.103.m a p l a p + p 3 ' € } q 3 d y = 0 for all 7 = viai with va E H.. The functions a .(w) and 7 3 E H i ( w ) .s + R. is also a classical solution. The proof of this result.178 Philippe G. Ciarlet. viz. t ap lap = aa(t"PIp) + r:. tapla = which have naturally appeared in the course of the proof of Theorem 2. here by means of its contravariant components t a p : i. the solution of the variational equations. The announced partial differential equations are thus satisfied in w . which is due to Alexandrescu [5]. Cristinel Mardare = and the symmetry relations bFlu obtain bglp (Theorem 2.. 0 The functions nap = &aapuTynT(SE) are the contravariant components of the linearized stress resultant tensor field inside the shell. and second order. we state a regularity result that provides an instance where the weak solution.r a@ + rpantPu+ r P u tau. . constitute examples of firstorder. covariant derivatives of a tensor field defined o n a surface. Finally. the variational equations 1{ w 1 aaPuTru~(<">rap(~) p u T p u T ( S E ) p a p ( ~ ) piiEqi}&dy = 0 + pa  imply that +l &{bapnap + bgbupmaP . tensor field inside the shell. is long and delicate and for this reason is only briefly sketched here.82).(taPlp).P &3 = affPu. or linearized bending moment. viz. we finally (iv) By parts (ii) and (iii). and the functions POT (C") 3 are the contravariant components of the linearized stress couple.
102. p. To begin with. J. E H4(w) if pa)" E H 1 ( w ) and p3>"E L2(w). [a]Adams. E. D. y is of class 9 E Cmf4(W.. Douglis & Nirenberg [3] implies that C" E W"+3i4(w) and < a : E W"f4 )q (~).An Introduction to Shell Theory  179 Theorem 2. Academic Press. Douglis & Nirenberg [3].103 (which is of the third order with respect to the unknowns and of the fourth order with respect to the unknown < ) : is uniformly elliptic and satisfies the supplementing condition o n L and the complementing boundary conditions. assume that the boundary y is of class C4 and the mapping 9 belongs to the space C4(sS. y is of class Cm+4 and 9 E C"+4(sj. satisfies < : cz < : E H 3 ( w ) and < . Theorem 3. Percivale.: Sobolev Spaces.. Lemma 3.104. One first verifies that the linear system of partial differential equations found in Theorem 2. p. assume that y = y. EX3).102 satisfies < : E W"+3. One then verifies that the same system is also strongly elliptic in the sense of NeEas [73. p3>€ W " .R3).q(w) and < : E W"+4+7(w) Sketch of proof. with components E Hi(w)and 53" E Hi(w)since y = y by o assumption. 2601 then shows that the weak solution C" = <fa2 found in Theorem 2. q ( w ) . References [l] Acerbi. in the sense of Agmon. 1975.A. R3). Assume that. Buttazzo. Finally.51 about the index of the associated linear operator then implies that < : E W3>4(w) and c: E W414(w) if pa+ E W1i'J(w) and p3lE E LQ(w)for some q > 1. R. . G. New York. A regularity result of NeEas [73. Then a regularity result of Agmon. for some integer m 3 0 and some real number q > 1. 137148. palE E W m + l i q ( ~ ) . Assume finally that. T h e n and E o the weak solution C' = <faifound in Theorem 2. A result of Geymonat [51. Elasticity 25 (1991).2.: A variational definition of the strain energy for an elastic string. Let w be a domain in R2 with boundary y and let 9 : w 4 R3 be a n injective immersion. for some integer m 3 1 and some real number q > 1. 1851.
: Me‘canique de I’Ingknieur. Glowinski & J. [4] Akian. [8] Attouch. Ciarlet. A. Douglis. Quart. C.. 337403. 319 (1994). Editors). 4547. Girault. Lions. 473496. [lo] Bamberger. V. : Decomposition of vector spaces and application t o the Stokes problem in arbitrary dimension. B. Methods Appl. SanchezPalencia. 11. 12831299. [ll] Bernadou. : Sur l’ellipticitk du moditle linkaire de coques de W. [16] Busse S. R . J. P.: Asymptotic analysis of linearly elastic shells with variable thickness. Paris. M. Le Dret. S. Math. Math. Appl. 63 (1977). J. D. [13] Blouza. : Existence and uniqueness for the linear Koiter model for shells with little regularity. Hermann. 109140. NorthHolland. P. Sci. Cristinel Mardare [3] Agmon. [18] Ciarlet. [5] Alexandrescu. Lecture Notes in Economics and Mathematical Systems.T. Rational Mech. Revue Roumaine Math. Y. Ciarlet. 282 (1976). [12] Bernadou. Acad. Volume 11.. . : Existence theorems for twodimensional linear shell theories. A . Skr. : Regularit6 pour une classe de problitmes aux limites elliptiques dkg6nkres variationnels. Paris. Sci. 17 (1964). H. : Theorbme d’existence pour le modhle bidimensionnel de coque non lineaire de W.. H.R. 27 (2005). Comm. [15] Borchers. Se‘r. : Estimates near the boundary for solutions of elliptic partial differential equations satisfying general boundary conditions.. S o h . 553590. H. Camus. P. Amsterdam. 1988. [17] Caillerie.. 1997. : Elastic thin shells: asymptotic theory in the anisotropic and heterogeneous cases.G. 899902. : Mathematical Elasticity. Arch. 1981. 3592. J.. Sci. 57 (1999). C. Math.: Variational Convergence f o r Functions and Operators.L. : On the equations rot w = g and div u = f with zero boundary conditions. Nirenberg. Pure Appl. in Computing Methods in Applied Sciences and Engineering (R. 6787. A. Pitman. Analysis and Applications 1 (2003). A. : Mathematical Elasticity. I.: Convexity conditions and existence theorems in nonlinear elasticity. Elasticity 34 (1994). Czech. M.. Models Methods Appl. 1984. 134 (1976).G. Volume I: ThreeDimensional Elasticity. P. 111138..180 Philippe G. J. [6] Amrouche. W. Math. Le Dret. Amsterdam. Math.. Miara.h pp. 43 (1998). : A simple proof of the ellipticity of Koiter’s model. 0. [9] Ball. Sci. E. Hokkaido Math. Paris.G. J.G. : The infinitesimal rigid displacement lemma in Lipschitz coordinates and application t o shells with minimal regularity. C. SpringerVerlag. 44 (1994). H. NorthHolland.. J. [14] Bolley.L. Volume 11: TheonJ of Plates. 19 (1990). Raoult. Ciarlet. S. Acad. Koiter. 116. Koiter. 89136. 5 (1995). Pures Appl... [7] Anicic. L. P. Heidelberg. 317337. [19] Ciarlet. Boston. Anal.T.
Math. J. : On the ellipticity of linear shell models. B20 (1999). Springer. P. : An introduction t o differential geometry in R3. P. Amsterdam. [21] Ciarlet. Rational Mech. Volume III: Theory of Shells. 136 (1996).T. Ser. Elasticity 43 (1996). Arch. [23] Ciarlet. 75 (1996). 2000. Lods. P. 423426.. Dordrecht. D.G. V. Justification of flexural shell equations. : A justification of the twodimensional plate model. P. Miara. P.G.G. Destuynder. angew. Paris. A. [29] Ciarlet. B. [27] Ciarlet. Anal. Mardare. E. Models Methods Appl. Roquefort. Lods. : Asymptotic analysis of linearly elastic shells: “Generalized membrane shells”. : On the ellipticity of linear membrane shell equations. Math. V. 147188. Sci. S.R. [26] Ciarlet. 279296. P. B. Elasticity 50 (1998). Pures Appl. P. .G. G : Sur les lois de comportement en BlasticitB nonlinkaire compressible. [30] Ciarlet. J. 261277. Justification of Koiter’s shell equations. 43 (1992). J.G. Mcanique 18 (1979). : Un modkle bidimensionnel non lineaire de coque analogue B celui de W. : Mathematical Elasticity. 11.. Math. C. P. NorthHolland. [34] Ciarlet. [25] Ciarlet. 315344. 136 (1996).G. P.& Acad. : Asymptotic analysis of linearly elastic shells. : Direct Methods in the Calculus of Variations. [24] Ciarlet.G. 11 (2001).G. [32] Ciarlet. I. : Asymptotic analysis of linearly elastic shells. : Asymptotic analysis of linearly elastic shells.A n Introduction to Shell Theory 181 [20] Ciarlet. 7879 of J. SanchezPalencia.. 191200. Chinese Ann. Sir. 13791391. [33] Ciarlet. Rational Mech. D.. J.. Lods. Sci. 405410..G. Geymonat. Anal. B. Berlin. P.G.: Existence of a solution for a nonlinearly elastic plane membrane subject to plane forces. : An existence and uniqueness theorem for the twodimensional linear membrane shell equations. V. Arch. I.: Existence of minimizing solutions around ”extended states” for a nonlinearly elastic clamped plane membrane. P. D. 295 (1982). 136 (1996). [38] Dacorogna. Rational Mech.. 5167. [22] Ciarlet. Phys. 1989.Arch. 2005).G. Math. V.G. 11. 129244. Pures Appl. J. 243253. Elasticity. Coutand. Elasticity 53 (1998/99). Math. 22B (2001). : Justification of a twodimensional nonlinear shell model of Koiter’s type. 75 (1996). Math. P.G.: An existence theorem for nonlinearly elastic “flexural” shells. Acad. P..G. 147159. 331 (2000). Springer. P. Paris. [37] Coutand.. Lods.. 111. Koiter...G. Miara. P. P. Sci. 2. Justification of membrane shell equations. 163190. J. [31] Ciarlet. 119161. 2005 (also appeared as Vol. 107124. [36] Coutand. : On Korn’s inequalities in curvilinear coordinates. [28] Ciarlet. Chinese Ann. P. [35] Ciarlet. : A n Introduction to Differential Geometry with Applications to Geometry. Anal. Lods.G. V. C..
E.D.: rConvergence and Calculus of Variations.: Remarks on nonlinear membrane shell problems. 1993. Muller. Birkhauser. Se'r. G. R. Paris. 58 (1975). [56] John. 55 (2002). Acta Applicandae Mathematicae 4 (1985).. 1992. : Estimates for the derivatives of the stresses in a thin shell and interior shell equations. Math. 1976). A. Mat. Berlin. [50] Genevey.: Elliptic Problems in Nonsmooth Domains. F. Cristinel Mardare [39] Dal Maso. CRC Press. [45] Destuynder. T: Su un tip0 di convergenza variazionale.. [52] Geymonat. [55] Gurtin M. Mech. 5 (1977). 235267.F. C1.: Cauchy's theorem in classical physics.: Sulla convergenza di alcune successioni di integrali del tip0 dell' area. Sci.. : Functional spaces for NortonHoff materials. Lecture Notes in Mathematics. P. Boston.E.C.. 697702.C. Mech. [47] Evans L. [51] Geymonat. Suquet.. 1972 (English translation: Inequalities in Mechanics and Physics.G. M. Acad. Pitman. (1997). 305328. Boll. Atti Accad. James. Dunod. Comm. : A classification of thin shell theories. Lions. Math. Mora. Martins L..: rconvergenza e Gconvergenza. Pure Appl. Prikl. 593608. Paris. C. [48] Friesecke. 1980. Mat.. G. [42] De Giorgi. Ann. Muller. [44] Destuynder. Paris.: Measure Theory and Fine Properties of Functions. 1563.L. SpringerVerlag.182 Philippe G. Roma 8 (1975). : Derivation of an approximate theory of shells by means of asymptotic integration of the equations of the theory of elasticity. [54] Grisvard P. . E. Mat. 979. SpringerVerlag. K.: An Introduction to rConvergence. Universitk Pierre et Marie Curie. : Derivation of nonlinear bending theory for shells from three dimensional nonlinear elasticity by Gammaconvergence. P. 206222. 227294. : Les Ine'quations e n Me'canique et e n Physique.D. 60 (1976). 27 (1963). 18 (1965). Pura Appl. : A theorem on geometric rigidity and the derivation of nonlinear plate theory from threedimensional elasticity. 213220. Mat. Berlin. Methods Appl. Gariepy R. Rend. [49] Friesecke. G. Anal. 8 (1986). E. : Sui problemi ai limiti per i sistemi lineari ellittici. R. Rational Mech. [40] De Giorgi. 69 (1965). Naz. and Franzoni. Ciarlet. G. [53] Goldenveizer. Ital.: Sur une Justification des Moddes de Plaques et de Coques par les Mkthodes Asymptotiques. 215237. 1983. 14611506. Boston. P.L. [41] De Giorgi. Math. G. Lincei Rend. 207284. Sci. Math.. Doctoral Dissertation. 1985. S. Arch. I.. Comm. Mat. Un. 336 (2003). J. 842850. and Dal Maso. Sci. Vol. E. James. [43] De Giorgi.R. G. S. [46] Duvaut. Pure Appl. Solids 2. G.
Anal.B. 74 (1995). 365388. 1966.: Multiple Integrals in the Calculus of Variations. : On the nonlinear theory of thin elastic shells. J. SpringerVerlag. [65] Lods. Anal. I . [58] Klingenberg. A.An Introduction to Shell Theory 183 [57] John. : On the foundations of the linear theory of thin elastic shells. Stampacchia.: Stressdeformation relations for isotropic materials. : Inequality of Korn's type on compact surfaces without boundary. S i r . Pacific J. Proc. 2553. Analysis and Applications 2 (2004). 24B (2003). Arch. 549578.: Asymptotic analysis of linearly elastic shells. Akad. Math. Wetensch. 18 (1971). Pures Appl. Hughes. 154. .: The nonlinear membrane model as variational limit of nonlinear threedimensional elasticity. Paris.. E. Englewood Cliffs. The membrane model.T.R.. Proc. 411417. [71] Morrey. Arch.. [66] Magenes. Masson. Math. Akad. Kon. 11791192. B. W. PrenticeHall. 331353. [63] Le Dret. Chinese Annals Math.B. : Nonlinearly elastic shell models. H. Anal. [73] NeEas. [62] Le Dret. rational Mech.. Jr. Ned. Rational Mech. A. Wetensch. SpringerVerlag. Scuola Norm. J. H. SpringerVerlag. 1591 Koiter. 142 (1998). Cas de flexion pure non inhibke. [67] Mardare. E.: The quasiconvex envelope of the Saint VenantKirchhoff stored energy function. Paris. 355374.R. A. 5984. E. [60] Koiter. Asymptotic Anal.: Mathematical Foundations of Elasticity.. 323425. 1983. The flexural model. J. 11. B. Rational Mech. 6 (1996). [61] Le Dret. [72] Morrey.E. 2 (1952). Proc.. 191204 [68] Marsden. S. 12 (1996). Raoult.. 169195. Ericksen J. : Les Me'thodes Directes en The'orie des Equations Elliptiques. 1967.. Ann. 247358. : Wellposedness for Koiter and Naghdi shells with a GImidsurface. Raoult. C. 1978).. [64] Le Dret. B. Sci. V. T. I. 1973 (English translation: A Course in Differential Geometry. W. 4154. Arch. Berlin.. : I problemi a1 contorno per le equazioni differenziali di tip0 ellittico. H.S. : Refined interior equations for thin elastic shells.L. Raoult. B69 (1966). F. Jr. 142 (1998). B73 (1970). G. SanchezPalencia. [69] Miara. Royal SOC. [75] SanchezPalencia. : The membrane shell model in nonlinear elasticity: A variational asymptotic derivation. C. Berlin. H..J. : Statique et dynamique des coques minces. : Eine Vorlesung uber Differentialgeometrie.J. Berlin. Math. 4 (1955). Kon. Nonlinear Sci. I. Acad. Pure Appl. 235267. Miara. Ned. 309 (1989). [70] Miara. C. W. Comm. : Nonlinearly elastic shell models. Pisa 12 (1958). [74] Rivlin R. Sup. : Quasiconvexity and the lower semicontinuity of multiple integrals.T. Edinburgh 125A (1995).
[85] Taylor M. Math. Paris. I. 1992.: Partial Differential Equations I: Basic Theory. J. Pures Appl. E. John Wiley. Acad. Springer. Universite de ParisSud. L. : Analyse 11: Calcul Diffkrentiel et Equations Diff6rentielles. 78. [78] SanchezPalencia. E.: Functional Analysis. : Coques Elastiques Minces: Proprits Asymptotiques. Cristinel Mardare [76] SanchezPalencia.R. [81] Schwartz. T. [80] Schwartz. E. SanchezPalencia. Paris. 1969. 71 (1992). : The‘orie des Distributions. . : Statique et dynamique des coques minces. [83] Stoker.J. Springer Tracts in Natural Philosophy. Cas de flexion pure inhibee . Acad. Princeton University Press. C. Paris. Berlin. : Boundary Value Problems of Finite Elasticity. S h . 309 (1989).R. Hermann. Ciarlet. E. 379406. J .. 1978. 1978. 909916. : Singular Integrals and Differentiability Properties of Functions. Springer. [79] SanchezHubert. 11. S h . Berlin. L.13. SpringerVerlag.184 Philippe G. Sci. Sci. : Differential Geometry. Paris. [77] SanchezPalencia.E. Publications Mathhmatiques d’Orsay No. Masson. : Passages ? limite de 1’6lasticit6 tridimensionla i nelle B la thhorie asymptotique des coques minces. 11.: Topics in Nonlinear Analysis. [87] Yosida K.311 (1990). [86] Valent. 531537. Vol. C. 1988. L. : Asymptotic and spectral properties of a class of singularstiff problems. [84] Tartar.Approximation membranaire. [82] Stein. 1997. 1970. Berlin. Orsay. 31. 1966. J. Paris. E. Hermann. New York. 1996.
the concept of “shell finite element” is still somewhat fuzzy.g. and such analyses have been conducted ever since the early development of finite element methods.185 Some New Results and Current Challenges in the Finite Element Analysis of Shells Dominique Chapelle INRIARocquencourt B. as it may correspond to very different ideas and techniques in various actual implementations. and across the various disciplinary cultures involved. without the use of any shell theory. However. 1 Introduction Finite element procedures are widely used t o analyse the behaviour of shell structures in various areas of engineering (e. 78153 Le Chesnay Cedex. This is expected t o be very useful as it paves the way for further thorough mathematical analyses of shell elements. Chapelle@inria. and this situation has resulted into an incredibly abundant literature and also to some significant degree of confusion. aerospace and civil engineering industries). in the framework of the asymptotic behawiours of shell models. In particular.fr Abstract Finite element procedures are widely used in engineering practice t o analyse the behaviour of shell structures. France Email: Dominique. A particularly important motivation for this is the understanding and the treatment of the deficiencies associated with the analysis of thin shells (among which the locking phenomenon). In . P. and shell elementssuch as the general shell elementsderived from 3D formulations using some kinematical assumptions. We then survey these deficiencies. We conclude the article by giving some detailed guidelines t o numerically assess the performance of shell finite elements as regards these pathological phenomena. in the automotive. Over the time. Our first objective in this paper is to give a unified perspective of these two families of shell elements. which is essential t o design improved procedures. the concept of “shell finite element” has referred to very different ideas and techniques. however. a significant distinction can be made between shell elements that are obtained via the discretization of shell models. 105.
Moreover. see for example (Bathe 1996) and the references therein. In (Bernadou. shell finite element procedures obtained by directly discretizing a shell model can be thoroughly analysed and explicit error estimates are then obtained under some wellidentified conditions (pertaining to the regularity of the exact solutions. Various further refinements of the facetshell techniques have been introduced. even if a successfully converging scheme were to be found.g. the finite element methods derived from a shell model draw their justification from the assumed relevance of such models. however. what is typically lacking for such finite element procedures is the proof that the finite element solutions converge to some known limit solution when the mesh is being refined. Examples of this category of elements can be found in (Batoz and Dhatt 1992). (Ciarlet 2000)) and finite element techniques obtained from various otherusually engineeringrootedconsiderations. the accuracy of the approximation would be rather poor. etc. the characterization of such a limit solution would be extremely valuable to evaluate the whole approach. Additional deficiencies associated with the facet type approximation of the geometry are identified in (Akian and SanchezPalencia 1992). One of the earliest techniques used in shell analysis consists in simply combining plate bending and membrane behaviours in flat elements. Ducatel and Trouv6 1988). we can identify in the “zoology” of shell finite elements a particularly distinct “line of division” between finite element methods that result from the discretization of shell mathematical models (namely essentially twodimensional models with unknowns given on the midsurface of the shell. it is also shown that the corresponding consistency error is nonconverging. the collection of which approximates the geometry of the actual shell structure. see e. by construction the coarse approximation of the geometry involved in these procedures implies that. By contrast. the type of numerical integration used. see e. a basic facetshell procedure is analysed and proved to be a nonconforming discretization scheme of a shell model. and this result is consistent with previous observations of the lack of convergence exhibited by these methods. Indeed.g. The major drawback of this approach is. see (Bernadou 1996). but whether or not these developments actually counteract the abovementioned deficiencies is not clear (and indeed not proven). However. This is known as the facetshell element approach. On the other hand. see (Bernadou 1996) and the references therein. Although they are usually numerically tested using various benchmark problems. Roughly speaking.186 Dominique Chapelle this respect. the solution of which they can be shown to approximate within a certain accuracy using numerical analysis techniques.). other shell finite element methods are grounded on mechanical considerations used “at the element level”. (Irons and Ahmad 1980). that building a shell finite element procedure on a set shell model is clearly restrictive as to .
we show that we can identify a welldefined shell model that “underlies” general shell elements.e. due to its specific construction that does not rely on any shell theory. 187 the variety of situations that can be analysed.which is a situation commonly encountered in practice). the literature abounds with allegedly reliable or lockingfree . see (Chapelle and Bathe 1998). This methodology is based on the idea of “degenerating” a 3D solid finite element into a shell element by using some kinematical a s s u m p t i o n s for describing the variation of displacements across the thickness of the shell structure. see for example (Bathe 1996). This is indeed why these finite elements are called general shell e l e m e n t s (and sometimes also “degenerated solid elements”). a finite element procedure with proven reliability with respect to the variation of the thickness parameter is still to be found.Some New Results and Current Challenges . such that the finite element solutions converge to the exact solution of this model when the mesh is refined. and the references therein. general 3D constitutive laws can be employed. This unification of concepts paves the way for further mathematical analyses of these highly attractive numerical procedures. are now wellidentified. A particularly important motivation for a thorough mathematical analysis of shell finite elements is the understanding and the treatment of some serious numerical pathologies that are known to occur when the thickness of the shell structure is “small” (to fix the ideas. and they are closely related to the complexity and diversity of the asymptotic behaviours exhibited by shells. By contrast. however. The first objective of this article is to present some recent results that give a unified perspective of general shell elements in the framework of shell models and their discretizations. and adapting the models to account for an arbitrary (3D) constitutive equationr for a general nonlinear behaviouris not straightforward.. and also (Bathe 1996) and (Bischoff and Ramm 1997). these phenomena typically become very significant as soon as the ratio of the thickness over the other characteristic dimensions of the structure is of the order of 1%. see the seminal work of (Ahmad. This approach is widely used in practice and inumerable evidences of its effectiveness can be found in the literature. Note that this drawback in fact also applies to facet shell elements. However. for later generalizations of these concepts. among which the bynow classical numerical locking.. More specifically. a mathematical analysis of such a procedure is very difficult to achieve. On the other hand. In addition. i. In particular. most shell theories are based on a linear elastic behaviour. Since the only model used is the 3D model in consideration. we compare this model to other classical shell models by means of an asymptotic analysis. a third methodology has been developed to specifically allow for unrestricted versatility as regards shell modeling capabilities. The sources of these difficulties. From a numerical analysis point of view. Irons and Zienkiewicz 1970).
unless otherwise stated. This assumption is made only in order to simplify the formulas and the discussions. with the classical convention that C may take different values at different occurences (including when appearing several times in the same equation).3 by presenting some guidelines for assessing and improving the reliability of shell finite elements. The second main objective of this article is thus to survey the issue of the reliability of shell finite elements with respect to variations of the thickness parameter. Indeed.1) before focusing on computational issues (Section 3. this issue must be put into perspective with the variations of behaviours induced in the exact solutions by variations of the thickness. and does not represent a restriction in our analysis. namely the finite elements obtained by discretizing shell models (Section 2. In general. In all our discussions we focus on linear formulations. see (Chapelle and Bathe 2003) for a general analysis with arbitrary thickness. equivalently. namely with the analysis of the asymptotic behaviours of shell models. when not altogether irrelevant. is independent of all the parameters appearing in the same equation (and in particular of the mesh parameter h and of the thickness parameter E ) . Section 3 is then dedicated to the issue of the reliability of shell finite elements. hence we do not repeat the same definitions here and we only define the new notation (mainly related to finite element concepts).2). these models are based on kinematical assumptions that are used . on the 2D reference domain w from which the midsurface is obtained via the mapping 8 (that we call the chart). Our notation is based on the notation used in (Ciarlet 2001). Another simplification that we use is to assume that the thickness is constant (denoted by E ) over the whole shell structure. classical shell models are mathematical models in which the unknowns are defined on the midsurface of the shell body or. and the general shell elements (Section 2. In Section 2 we present the two families of shell finite elements that we want to analyse and compare. and we conclude in Section 3. hence it is natural to treat them in this framework first. This article is organized in the following manner. Of course. see also (Ciarlet 2000). In addition.1 Discretizations o classical shell models f As discussed in (Ciarlet 2000). we use the symbol C to denote a generic positive constant that. the difficulties that we want to analyse are already present in linear problems.188 Dominique Chapelle shell finite elements.1). but the assessment performed to reach this conclusion is often inadequate or incomplete.2). 2 Two families of shell finite elements 2. We start by surveying the asymptotic behaviours of shell models (Section 3.
2). for the Naghdi model the natural diplacement space is VN= {(v. 2 = saaa. When the assumption is that any normal fiber remains straight and unstretched during the deformation. the displacements can be described by a displacement field only..4) . the displacements in the whole shell body are completely described by the datumn wof a displacement field and a rotation field. Note that. s) n (2. (2. we will consider the model summarized in (Naghdi 1963).e. z ~is defined on w .e. a displacement at any point in RE is given by where s. As an example of this category of shell models. i. the additional assumption on the preserved orthogonality of fibers is equivalent to the rotation vector being given by sa = (773p +b 3 U ) . and the shell models that are based on this assumption the ReissnerMindlin type models. see (Reissner 1945) and (Mindlin 1951). 189 to describe the displacements of points located on material fibers that are orthogonal to the midsurface in the original configuration. 1. that we call the Naghdi model.3) We will call these shell models the KirchhoffLove type models.1) and ) (2. unstretched and normal t o the deformed midsurface (the socalled KirchhoffLove kinematical assumption) . E [H1(u)15} BC. Namely. is a vector tangent to the midsurface at the point of coordinates ( ~ 1 ~ i. Namely. comparing (2. 2. We call this kinemati~ s ) ~ cal assumption the ReissnerMindlin kinematical assumption. where r ] ( z l .. Shell models can then be (roughly speaking) divided into two categories according to the type of kinematical assumption made. When the assumption is that any normal fiber remains straight. as the vector that represents the effect of the rotation of the normal fiber. An important difference between these two categories of models is the respective regularities of their solution spaces (that we also call the displacement spaces).Some New Results and Current Challenges . The shell model proposed in (Koiter 1965) belongs to this category and will be henceforth used as an example (we refer to this model as the Koiter model).
in each case. Clearly.190 and VK = Dominique Chapelle {q E [~'(w)]~ x H'(w)} n BC (2. and for the Koiter model. be compatible with the nature of the functional space. When the boundary conditions are compatible with the functional space considered and such that rigid body motions are prevented. Note that the discrepancy in the regularity of the transverse component of the displacement field ( u s ) can be interpreted as the consequence of the additional constraint (2. where BC symbolically denotes the essential boundary conditions prescribed. These issues are thoroughly addressed in (Bernadou 1996). we will say that we have admissible boundary conditions. The main difficulty lies in the required use of C1conforming finite elements for the transverse displacements in models of KirchhoffLove type. a priori error estimates follow from interpolation bounds and are of the type for the Naghdi model.and on computations that require an extensive use of the chart 6 .3) acting on the solution space VN where s € [Hl(w)]2. where the constants C and the orders of convergence k depend on the regularity of the exact solution and on the finite element shape functions considered. in order to obtain wellposed variational formulations. see (Bernadou 1996) and the references therein. Since conforming methods are used. at least in principle. For the Naghdi and Koiter models (and indeed also for similar models from the two above categories). We emphasize that these estimates are not independent of the thickness parameter E . Note that this holds in particular when the displacement field is set to zero on some part of the boundary dw that does not correspond to a straight segment in the physical space. the discretization of classical shell models by conforming finite element methods is rather straightforward.6)(2. Let us emphasize that these finite element procedures are based on meshes that are constructed in the reference domain w .5) for the Koiter model. As a consequence. as the constants C in (2.7) above depend on E for two reasons: . some appropriate boundary conditions need be prescribed in the displacement space. it can be shown that the corresponding bilinear forms are coercive and bounded on their respective displacement spaces provided that the boundary conditions are sufficient to prevent any rigid body motion of the shell body. Of course these boundary conditions must.
Consider a 2D mesh of w given in the form of a set of points (the nodes) in w and of a connectivity that defines the elements. in the form 1 . 0 the regularity bounds are not independent of E in general. 2. In practice. 2. Unlike the previously discussed procedures. 5 2 ) coordinates and the local coordinates ( &). 1. ) ) de5. the shape functions depend .10) where B30 and L 3 0 respectively represent the external and internal virtual works.8) (21.52). and the quantities (zf). Consider a general 3D variational formulation posed on the 3D geometrical domain defined by the mapping @(21.. 191 the coercivity and continuity constants of the corresponding bilinear forms depend on E . Isoparametric elements are used. the type of finite element procedure that we have been describing in this section is not very often encountered.~) (2. but instead from a 3D variational formulation. This issue will be further addressed in Section 3.22. hence this defines a onetoone mapping inside each element between the ( 5 1 . .Z2) +23(113(21. A typical general shell element procedure is constructed as follows.Some New Results and Current Challenges . and infer a modified variational problem from the stress assumption D33 = .9) We symbolically denote this modified problem by B 3 D ( ~= L 3 D ( ~ ) . general shell elements are not derived from a shell model.22. (2. taking into account the essential boundary conditions. Of course. (2.0. in spite of the extensive use of shell finite elements in engineering.11) where the functions X i are the shape functions associated to the k nodes of the element considered.23) = 0(5l.&[. (2.2 General shell elements Most shell finite element procedures used in engineering practice fall into this category of general shell elements. note the nodal coordinates.. and V 3 D denotes the appropriate functional space.2.23) E W x ] &. VV E V 3 D .
192 Dominique Chapelle on the polynomial order chosen and on the geometric type of the element (namely quadrilateral or triangular). for any continuous scalar or vector function 4. Using this mapping. vw E v y . E 2 ) ( ~ ( 2 ) +QS(~)).8). the following approximation @ h ( 2 i 7 ~ z . = (2. and are interpolated inside the elements using the above shape functions.16) to be compared with (2. at) 0.15) where BiD and LiD are obtained from B30 and L30 by using. instead of the exact chart 0.(i) . hence the unknowns (degrees of freedom) are the 3 displacement components and the 2 rotation components at the nodes.13) where s(2) corresponds to a rotation at node i . ~ z ) 23Z(a3)(2i72z).15).23= ) C i=l k Xi(E1. 2 3 )= Z ( ~ ) ( ~ i . we define the interpolation operator Z such that i=l in every element of the mesh. it appears from the above description that general shell elements have two key advantages: . Note that the degrees of freedom are similar to those of a classical shell finite element procedure obtained by discretizing a ReissnerMindlin type model. hence it satisfies . 3.14) We call the space of functions of this type that satisfy the boundary conditions prescribed on the structure (note that boundary conditions can be prescribed here as they would be for a ReissnerMindlin type model). (2. Consider the problem Find u E V2D such that B i D ( uw ) = L ” h ( W ) . 4.22. + (2. namely ~(21. . (2. However. The general shell element procedure consists in solving Problem (2. Consider displacement functions defined over RE that satisfy ReissnerMindlin kinematical assumptions at all the nodes of the mesh.
r ) in VN such that Clearly. hence they can be easily used for arbitrary 3D material laws and also for nonlinear formulations. These connections are now much better identified.e. This is indeed why these procedures are called general shell elements. .e.2. C3. a function 4 is in vj if and only if 4 = z(4). since the unknowns are a displacement field and a rotation field.. a 3 = 0 at all the nodes of the mesh. i.3. for a given 3D material law we can consider the problem Find (C. see in particular (Chapelle and Bathe 2000). Moreover. q . Then E V2Dis equivalent to with the three conditions C1. 0 The connections between general shell elements and classical shell elements (i. Bi E Vj for i = 1.Some New Results and Current Challenges 0 . this defines a shell model of ReissnerMindlin type. This shell model is the natural candidate for being the model that underlies general shell elements.17). the derivation of the associated matrix problem does not require the datum of the exact midsurface chart 8 . i.e..15) as an internal approximation of Problem (2. we define Vj as the scalar finite element space associated with the isoparametric discretization above. 7 and S satisfy the proper boundary conditions. With a view to analysing Problem (2. Note that Vj c H1(w). this issue has probably been a source of major misunderstanding between the engineering community (primarily using and developing general shell elements) and the numerical analysis community (considering only classical shell elements). Bi E Vj and i . the model that provides the solution to which finite element solutions converge when the mesh parameter h (namely the diameter of the largest element) tends to zero. both being given on the midsurface (or in the reference domain). but from a general 3D formulation. This is extremely convenient in practice as most structures analysed are the result of design procedures (CAD systems) that do not provide these charts. but only the position of the points and the value of the normal vectors at the nodes. 193 They are not obtained from a specific shell model. those obtained by discretizing classical shell models) have remained unclear for many years. Indeed. As implied by the use of the approximate chart Oh. 5 . C2.
Note that the use of the projection operator is required to that purpose because we do not directly have 5 in the tangential plane at every point of w .r(B)) such that (11. Using the equivalence (2. can then reformulate the general shell ) we element procedure as follows.21) ) Comparing now this equation with the continuous problem (2. Clearly VNh is a (finitedimensional) subspace of VN. Note that we assume that the Lam6 constants do not'depend on E .C2. we can say that the general shell element procedure corresponds to a discretization scheme of the continuous problem with approximate bilinear and linear forms.e. Taking into account the ReissnerMindlin kinematical assumption we have .(5 ' u3)u3. (2.C3) above.19) we define the finite element space VNh as the set of functions (q. and we observe that consistency errors come from two sources: the approximation of the geometry corresponding t o Eq. In particular. as shown in (Chapelle and Bathe 2000).g.16) and the presence of the interpolation operator in (2.21). when considering Hooke's constitutive law we obtain. T ( 5 )= 5.17).s) satisfies the three conditions (Cl. (2. (Ciarlet 1978). When choosing specific 3D formulations. r ) VNh such that in B ~ D ( ~ + 5 3 Z ( r ) .20) T ( 5 )= s 5 = Z(s) * valid for any ( q . Find ( < . i.17) and the corresponding general shell element procedure can be further analyzed. v(q. q f Z 3 Z ( 3=)L3hD(q+Z3Z(s)). the connection between the shell model represented by (2. see e.194 Dominique Chapelle Denoting by T the operator that projects a vector field defined over w onto the tangential plane of the midsurface a t each point.~ ( 5 )in VNh. (2.s ) E VNh.
. .1 by showing that the two models are asymptotically equivalent as regards their asymptotic solutions when the thickness parameter tends to zero. &=&7 (2. taking 0 the firstorder approximation of the strains. Moreover. Provided that the boundary conditions are admissible..25) In addition..26) i.27) the zeroorder approximation of all geometric coefficients. This suggests that some close connections exist between the basic model and the Naghdi model. namely 9 affflu.28) we obtain exactly the bilinear form of the Naghdi model (with c = 8 in the equations recalled in(Ciar1et 2000)). and we assume f E [ L 2 ( f l E ) ]We then have the following result. see (Chapelle and Bathe 2000). and tcuT corresponds to a new tensor defined by 1 tcUT(s) = (b:s.lT 2 f b:s. We will further substantiate these connections in Section 3.T. under standard assump3. Proposition 1. 195 where we recall that yuT.lU)' (2. f represents the applied distributed force.17) has a unique solution. It is then important to investigate the relations between this shell model (that we call the basic shell model. namely { eUl/T(rl 0 f x3s) 23s) ep. and also (Delfour 1998) where various combinations of truncations are considered and analysed. this solution is the limit of the solutions of Problem (2. This result shows that (at least in the case of Hooke's law) the shell model given by (2. pFT and 703 respectively represent the components of the tensors of membrane strains. gap M asp. see (Chapelle and Bathe 2003)) and classical shell models. bending (or flexural) strains and shear strains in the Naghdi model.22) by truncating the expansions in powers of x 3 of the expressions appearing under the integral sign.x3 p ~ ( ~ I s ) = TP3(rlI S > I (2. see (Chapelle and Bathe 2000). Problem (2. tions regarding the mesh and the finite element shape functions used.17) really is the mathematical model that underlies general shell element procedures. we consider the external virtual work given by (2. If we formally transform the expression (2.e.drl+ yUT(rl) .Some New Results and Current Challenges .15) when the mesh parameter h tends to zero.
196 Dominique Chapelle As a conclusion. typically on the fineness of the 2D mesh regardless of the third dimension. general shell elements arein factnonconforming discretization schemes of a shell model of ReissnerMindlin type which is itself very closely connected to the Naghdi model. although they do not seem (at first sight) to bear much resemblance to classical shell elements. in particular as regards the reliability of shell finite elements which is the subject of our next section. This means that we implicitly expect uniform convergence of the finite element solution with respect to the thickness parameter. allow us to specify the norms in which we seek uniform estimates for approximate solutions. (Malinen 2000)). This is extremely important. The reason for this will be analysed in Section 3. Unfortunately. as it allows to analyse finite element procedures of both “families” using similar tools.g. This motivation is based on the assumption (or the “hope”) that the accuracy of the solution obtained only depends on criteria that prevail in 2 0 analysis. We denote the norm symbol with a V’ to signify that we do not necessarily require that uniform convergence hold in the Sobolev space in which the problem is originally set (for example [H1(w)]’ x H’(w) for KirchhoffLove type models). compared to the cost of a full 3D analysis.e. it was soon recognized that standard finite element methods (such as the ones that we described in the previous section) do not provide such uniform bounds in general. we can say that. e. error bounds of the type where we now use E as a superscript to indicate that the solutions depend on this parameter. namely the asymptotic behaviours of shell models. namely the thickness of the structure. 3 Computational reliability issues for thin shells The essential motivation underlying the formulation of shell models and shell finite elements is the reduction of the cost of the analysis (in particular the computational cost).2. i. . As a matter of fact. This analysis will. by instead considering a 2D problem. but that we may tolerate uniform convergence in weaker norms to be defined. which is the objective of the forthcoming section. and where C (and also k) should not depend on E .e. To that purpose it is necessary to start by surveying the behaviour of the exact solution of the shell model when the thickness parameter tends to zero. in particular. further advances in the analysis of some specific general shell procedures have already been obtained with this approach (cf. i.
V: the solution space (we recall that the definition of this space takes into account the essential boundary conditions) . or this displacement and the rotation of the normal fiber for a ReissnerMindlin type model. 2": unknown solution. namely the displacement of the midthe surface for a KirchhoffLove type model. We also introduced E as a superscript in the righthand side of the formulation because it is unlikely %oobtain a weUposed asymptotic behaviour while keeping the loading constant over the whole sequence of problems. 4lJ. 197 3. (3. E : a corresponding test function that we call a "displacement" (even though it also contains a rotation for a ReissnerMindlin type model)..2) The meaning of the symbols appearing in this formulation is as follows. T We further emphasize that. T ) .. BF: a scaled representation of the bending (flexural) energy. h P d r l . More . as seen in (Koiter 1965).s ) ) = Sw[""P~77aP(6)7u7(. in (3. : For example. + EBM(Z&.EdY. L E ( E )the external virtual work associated with E .Some New Results and Current Challenges . or of the membrane energy and the shear energy for a ReissnerMindlin type model.2).1 Asymptotic behaviours of shell models Most classical shell models can be written in the generic following form: Find Z EE V such that E 3 B F ( Z EE ) . B M : a scaled representation of the membrane energy for a KirchhoffLove type model. E YE E V .) = L E ( E ) . B i m . the expressions of the bilinear forms BF and BM do not depend o n the thickness parameter E .I) (3. for the Koiter model we have. (77.6) + CPaaP7a3(C.
(2"). Proposition 2. this is equivalent to requiring that the scaled internal work E~PBF(Z". As discussed in (SanchezHubert and SanchezPalencia 1997) (see also (Ciarlet 2000) for the Koiter model). have a finite nonzero limit. F (3. it may very well happen that V = (0). we will say that the given scaling provides an admissible asymptotic behaviour.8) We can then show the following result. 2") E'PBM(Z". This situation induces an asymptotic behaviour very different from when pure bending is not inhibited. Brezzi and Lovadina 1999).198 Dominique Chapelle specifically. I n some cases the convergence m a y also be such that the sequence (2") tends to some limit in V .10) This subspace contains the displacements that have zero membrane energy (and also zero shear energy for ReissnerMindlin type models). classical shell models feature asymptotic behaviours that dramatically differ according to the contents of the subspace: VF = {EE v I B M ( E . I n addition. i f such a number exists we have 1GpG3.9) Remark 1. .~ ~ M2 "~ = ~ ( ) ~ . As a matter of fact. Of course. ) (3. for which the scaled external work G ( Z E converges to a finite and non) zero limit when E tends to zero. but we want to allow f o r more general situations in which a weaker convergence property m a y hold. E )= 0.11) In this case we say that pure bending is inhibited. these displacements are called pure. due to the severe kinematical constraints corresponding to zero membrane strains (3 equations versus 3 components of displacements). (3. since 2") + E 3 .~ ~ F ( 2 "Z . The existence of nonzero pure bending displacements is governed by the geometry of the midsurface and the boundary conditions prescribed on the structure (we further analyse this issue below). There is at most one exponent p that provides a n admissible asymptotic behaviour.bending displacements (since only the bending energy is nonzero). ) E +E 1 .7) where G is a function of V' independent of E and p is a real number. (Baiocchi and Lovadina 2002) and (Chapelle and Bathe 2003). see (Blouza. (3. what we will be looking for in the asymptotic analysis is a scaling of the righthand side in the form L'(E) =E~G(E). In this case. (3. For this reason.
This is why this asymptotic behaviour is also called a bendingdominated behaviour. Then.12) is Find Z o in V such that . the part corresponding to the membrane energy (and also the part corresponding to the shear energy when applicable) tends to zero. i. while the part that corresponds to bending deformations tends to a finite value.Some New Results and Current Challenges . Note that. the variational formulation can indeed be rewritten as 1 B F ( Z & . and also (Pitkaranta 1992) and (SanchezPalencia 1992). Z E converges strongly in V to Zo.14) shows that.e. if and only if G ( E )= 0. Z = (3. in the scaled deformation energy. VE E VF. Therefore we do not further consider this situation. By contrast. Proposition 3 shows that we have an admissible asymptotic behaviour for the scaling p = 3 provided that Zo is nonzero. This. when pure bending is not inhibited. E (3. however. < .e. E Since BF(ZO... this equation can be interpreted as the penalized f o r m of a problem in which the solution Zo is constrained to lie in the subspace VF.Namely. E ) 0. It can then be shown that a welldefined asymptotic behaviour exists for p = 3. A s s u m e that pure bending is not inhibited. although both may be encountered in practice. see (Chenais and Paumier 1994). i. E V E E VF.12) &2 is a small parameter.r(ZE. setting p = 3. since it corresponds to the restriction to VF. We then have the following result. VE E V.the solution of (3.13) Note that this problem is wellposed. namely of “larger magnitude” (corresponding to a bendingdominated behaviour. hence to effects of smaller asymptotic order.15) an other admissible scaling (for 1 p < 3) may exist. closed subspace of v. 199 We start by considering the case of noninhibited pure bending. when pure bending is inhibited (which is a situation more common than the previous one. Proposition 3. see (Ciarlet 2000) and (Chapelle and Bathe 2003)).15). In this case. of the minimization problem a equivalent to the original variational formulation. (3. (3. ) t BM(Z“. ) = G ( E ) . p = 3). does not correspond to a “physical situation” as small perturbations in the geometry (producing variations in VF) in the loading lead to a or violation of (3. the candidate limit problem for (3.14) E70 &2 Equation (3.13). If Zo = 0. E ) = G ( E ) . I n addition we have 1 lim Bn.
E'O zE) 0. Defining VMas the space obtained by completion of V for the membrane energy norm. Assume that pure bending is inhibited and that G E (VM)'. this problem is wellposed provided that G E (VM)'. due to (3. Note that this corresponds to a restriction on the loading since (VM)' V'. We call this new norm the membrane energy norm. satisfies llEllM 6 cllEllV. and also the shear part for ReissnerMindlin type models) tends to a finite value.= G ( E ) .r(Z'. Proposition 4.18) but is not equivalent to the original norm (cf.21) Since. E YE E V (3. this is equivalent to the same condition holding for any E E V M . + E 2 B ~ ( ZE ). Note that (under the assumptions of the proposition) we always have an admissible asymptotic behaviour since Z M cannot be zero (unless. we introduce the variational problem Find Z M in VMsuch that Bll.21). (3. = (3. see (Lions 1973). namely the bendingdominated behaviour and the membranedominated behaviour. the ' solution of (3.19) Clearly. It is . examples of the Koiter and Naghdi models above). Under this assumption we have an admissible asymptotic behaviour as shown in the following proposition.17) and we note that defines a norm over V which.200 Dominique Chapelle the asymptotic behaviour is very different and more complex. We consider the tentative scaling p = 1. Then. of course. the bending part of the scaled energy tends to zero while the remaining part (the membrane part. vE E vi (3. (3. setting p = 1. An C equivalent manner to write this condition is (G(E)I< C[Bn/r(E. this asymptotic behaviour is called a membranedominated behaviour. E ) = G ( E ) . In addition we have lim E ' B F ( z ~ . 2 converges strongly in VMto Z M . ) ] ' . The variational formulation then gives B M ( Z €E ) .20) since. We then recognize in (3.19). E b'E E v. E VM. We have therefore identified two very distinct categories of admissible asymptotic behaviours for classical shell models. by density considerations. of course.16) (3. G is zero).16) a singular perturbation problem.
201 the contents of the subspace VF that determine into which category a given problem falls.25) where go and g1 are in L2(w). (3. next section). see in particular (Pitkaranta and SanchezPalencia 1997) (cf.2). we assume that g is regular enough to provide the following expansion g ( x i . 5 3 ) . while 9 2 is a bounded function.. hence it is natural to seek uniform convergence of finite element solutions in the same norm (cf. These situations are more complex than the two “clearcut” asymptotic behaviours that we have discussed. 5 z . x 3 ) = gO ( x 1 .24). we consider the asymptotic assumption f = &Plg.r]+23S) = (3. Another significant difference is that the loading acting on the structure is threedimensional. In particular the membrane and bending deformation energies are coupled in the term (3. ‘ (3. Other asymptotic behaviours may exist and the above discussion shows that they can arise only when pure bending is inhibited and G $ (VM)’) ! (disregarding cases when (3.e. namely the shell model that underlies general shell element. for the abovementioned reason).23) In order to perform an asymptotic analysis consistent with the general one above. Find (C“.. Z 2 .26) . also Section 3. We recall that the general asymptotic analysis performed above is not directly applicable to this model since it cannot be written in the form (3. i.24) recalling that the external work involves an integration over the thickness. T “ ) the solution of the basic shell model. and they have not been completely elucidated yet.rE) V l such that E B 3 0 (6E +Z3TE. We denote by (C“. although significant advances have been made in their analysis. convergence of the solutions is obtained in a given norm.Some New Results and Current Challenges . We now consider the case of the basic shell model.22) as shown in (2. in the form (3.3). For each of these asymptotic behaviours.15) holds. Furthermore. x 2 ) k ( x 3 ) 2 g 2 ( x ~ . x 2 ) + x 3 9 1 ( 5 i .
ro) which is the element of that satisfies VF VF @%C0.31) .28) and the parameter c set to 8 in Bg. we then have the following asymptotic properties (assuming that the chart is smooth). for both KirchhoffLove type and ReissnerMindlin type models. We close this section by providing some guidelines for the analysis of the subspace of pure bending displacements. s ) ) = G(rl). .27). Proposition 5 .a + b&).t h e n for the scaling p = 1 both sequences (C". (11. (3. and V$ the completed space by the membrane energy norm when pure bending is inhibited. see (Chapelle and Bathe 2000). First of all we note that. namely =(~3. for a ReissnerMindlin type model pure bending displacements also have zero shear strains. r E )and T E ) converge in V to the { same limit ('r) C. we will say that the basic shell model and the Naghdi model are "asymptotically equivalent".r0).M which is the element of VG that satisfies (e. as the solution of the Naghdi problem (3. To summarize these properties.29) Alternatively. since the contents of this subspace have been shown to crucially determine the asymptotic behaviour of the shell problems.If pure bending is n o t inhibited in the Naghdi problem (3. pure bending displacements have identically zero membrane strains.27). ) E VF". In addition. Therefore.27). the solution of the basic shell model converges to the same limit solutions. However. q r l . the condition of zero shear strains can be written in the form of an explicit expression giving the rotation fhield q in terms of the displacement field s .202 Dominique Chapelle We also introduce the following Naghdi problem such that Find with (3. (3. Calling the purebending displacement subspace relative to this Naghdi problem (3.see (Ciarlet 2000).")'. T E ) and converge in VN t o the same limit (Co. and under the exact same assumptions.rE) ( g " . if pure bending is inhibited and if G E (V. . then for the scaling p = 3 both sequences (CE.
32) is the same as the geometric nature of the midsurface at the point in consideration.. In particular it shows how the geometry (i. System (3. 2. Furthermore. even in the case of ReissnerMindlin type models. By contrast. The differential nature (elliptic.e. This property is extremely valuable to analyse the contents of the subspace of pure bending displacements. see (Koiter 1965)). the condition of vanishing membrane strains YaL?(rl) = 0. . In addition. for which the finite clement solution Z. 2 . which is why elements of this subspace are also called inextensional displacements (we recall that the membrane strain tensor is the tensor of linearized change of metric of the midsurface. (3.32) enjoys the following remarkable property (cf. it is natural to require that 1.2 Asymptotic reliability of shell finite elements We recall that our objective is to obtain finite element procedures that behave uniformly well with respect to the thickness parameter. (Chapelle and Bathe 1998) for examples). We refer to (SanchezHubert and SanchezPalencia 1997) fo a more detailed analysis of this issue. (SanchezPalencia 1989a). i. Proposition 6.l satisfies an estimate of the type (3. In a bendingdominated case such a uniform estimate hold in the norm of the displacement space V .Some New Results and Current Challenges .e.32) make up an exactly determined differential system (three equations versus three unknowns) that must be satisfied by the displacement field only. 3. 203 va. In a membranedominated case such a uniform estimate hold in the membrane energy norm. when such a concept is relevant (namely in the hyperbolic andby extensionparabolic case) the characteristics of the system are also the asymptotic lines of the surface.. parabolic or hyperbolic) of System (3.g. (SanchezPalencia 1989b) and the references therein). Hence this is clearly the crucial condition that determines the contents of the pure bending displacement subspace. the type of the midsurface) and the boundary conditions can affect the contents of the subspace (cf. e.P = 1 . In the light of the above discussion regarding the asymptotic behaviours of shell models.33) where the norm used remains to be specified.
2).33) in the membrane energy norm. As is wellknown . VE E V .37) are rather unrestrictive.I\\M &IJZEZ. in a bendingdominated situation. VE E V . the problem to be considered is instead BF(Z.Z. We further point out that the interpolation assumptions (3. By contrast.I. ) = G ( E ) . lirn IIE . For example the Clement operator satisfies these two sets of assumptions when VMis a Sobolev space on which this operator is properly defined. defined respectively a VMand V .I.Emasl If we consider only the membrane energy norm in this uniform convergence result.E ) 1 + ~ B M ( Z . For a membranedominated problem wein factneed to consider the scaled problem BM(Z.36) and (3.38) ho EE10. VE E vh.I E v h such that E3~F(z:.EI ). E VE E v . h (3. (3.I. this bound is weaker than an estimate of the type (3.204 Dominique Chapelle We consider a standard finite element procedure obtained by discretizing the variational formulation (3.36) VE E V . h0 (3.&(E)IlV = 0 .  + (3. V E E VM. and such regularity results are not available. and both with values a v . Proposition 7. E) + & B ~ ( z .Illv} = 0. ) E + E'BF(Z.39) The major difficulty associated with this type of finite element formulation is the classical numerical locking phenomenon. n n h VE E V M . However. namely the problem Find Z.34) where Vj denotes the finite element displacement space used.35) We can then show that we have a uniform convergence property. this is clearly the best estimate that we can obtain without uniform regularity results on the exact solutions. I).= G ( E ) .37) Then for any fixed lirn sup E~~~ > 0 we have {llZE. as proved in (Chapelle and Bathe 1998). = L E ( E ) . E VE E Vj. (3. Assume that there exist two interpolation operators Zh and J h . (3.
Palma. A s s u m e that n o element edge in the mesh is part of a n asymptotic line.33) holds in this case. This being granted. locking occurs at its worst in situations where vh n VF = (0).40) is the common rule. Consider a regular hyperbolic shell fixed o n some part of its boundary.. Proposition 3 can now be applied with v h instead of V. keeping the finite element subspace (in particular the mesh) fixed. when E tends to zero the finite element solution 2 converges to . Consider an arbitrary element of the mesh. the result is immediately obtained by “propagating” the zero displacements from the boundary conditions to the whole domain. T h e n (3. this is a purely numerical artefact that makes the structure appear stiffer as it gets thinner. the pathological situation expressed by (3.40) holds. Proposition 8. and more specifically that. with 7 set to zero on any of its edges. ) = G ( E ) . h (3. SanchezPalencia and Vilarino 1998) and (SanchezHubert and SanchezPalencia 1997) for other results concerning (3. for any inextensional displacement 7. but in this case we still have ~1 = ~2 = 173 = 0 on a part of nonzero area of . for a given mesh. ). Even though the thickness is always finite in practice.. and a finite element scheme.41) and of course the solution of this problem is the zero displacement if (3. E YE E V n VF. in which the displacement components ~ 1 172 and 173 are approximated using . the finite element solution gets “smaller and smaller” when the thickness decreases (whereas the exact solution does not vanish). also (Choi’. beams.I. the solution of the limit problem Find 2 in v n VF such that : h BF(Z. . b) Or the element is only partially covered (Figure 1b). as illustrated in the following result. We will show that. 205 from other penalized formulations (such as nearly incompressible elasticity.Some New Results and Current Challenges . then 7 is identically zero over the whole element. then we have two possibilities: a) Either the lattice of asymptotic lines originating from this edge covers the element completely (like in Figure 1a) and the conclusion directly follows from Proposition 6. Hence. continuous piecewisepolynomial functions. unlike for beams or plates. plates. We thus have that. which explains the terminology “numerical locking”. for which we also give the proof because it is simple and illuminating as to how the geometry influences the asymptotic properties (cf. this implies that no uniform error bound of the type (3.39).40)). Proof. . 7 is if zero on some edge of any element of the mesh.40) Indeed.40) holds. (3. in the framework of (3. The specific difficulty with shells is that.
Recalling that the displacement components are given by polynomial functions. see in particular (Arnold 1981). which is much more general. hence locking always occurs using displacement. But of course. however. in order to avoid inhibiting the locking mechanism identified in the above proof. 0 In order to make the statement of this property simpler. A practical consequence of this discussion is that.206 Dominique Chapelle Figure 1 Asymptotic lines and inhibited region for an element the element. for a given thickness. that this result carries over to any case in which the same propagation technique can be applied. in industrial computations. For beams and plates.40) by raising the polynomial degrees of the discretization spaces. the errors in . we assumed that no single element edge corresponds to an asymptotic curve.based finite elem e n t formulations. this implies that they are zero over the whole element. We further emphasize that Proposition 8 reveals a significant difference between finite element methods for shells and for other structures. aligning meshes along lines of specific geometric properties is impossible in general). It is obvious. Here our argument is independent of the polynomial degree. it is indeed always possible to avoid (3. This holds in particular for cylindrical test problems for which the (very strong) temptation t o align the mesh along the axis should definitely be resisted if one wants to draw conclusions of somewhat general significance (note that. when using benchmark problems to numerically detect locking phenomena in shells. we should not align the mesh with the asymptotic lines of the midsurface when such lines exist.
cf. ~) + y B h ( Z g . Note that. for ReissnerMindlin plates this additional unknown corresponds to the shear stress. For instance. see e. When the analysis of the mixed formulation can be achieved. ) = 0.42) If we symbolically denote by ECE group of additional stress unknowns. see (Brezzi and Fortin 1991. Z) . = VE E V .44) where Bh is a bilinear form that “resembles B M ” .but for which the constraint B&(E. For example.E ) = 0 is “more easily satisfied” for finite element displacements than Bn/r(E. and that the mixed bilinear form M is continuous over V x L 2 . For shell formulations.. E this strategy is fairly close to the engineeringbased idea of reduced integration. (Lyly 2000). M(Z‘.12) then reads { B F ( Z €E. & 1 VE E Vh.g. The resulting finiteelement formulation can then be written as B F ( Z E. it indeed provides a proof that uniform convergence with respect to E holds (in the norm of V). we have that D(E. in which an additional unknowncorresponding to the “delinquent” term of the energyis introduced..E)+defines a norm which is equivalent to the L2norm. it is by now classical to resort to mixed formulations. the and by E the corresponding test function. In fact. In order to circumvent the locking phenomenon. ) + M ( E .43) where the second equation expresses the stressstrain relationship for the additional unknowns. for the Koiter shell model the additional unknown is the membrane stress tensor. S) VE E (3. . namely the tensor with contravariant components (3.&2D(EE. In general.Some New Results and Current Challenges . By construction. in order to allow the elimination of the stress degrees of freedom at the element level. = 0. in some cases it has been possible to substantiate some earlierproposed reducedintegration schemes using the theory of mixed methods. e. from this point of view.E ) = G ( E ) . L2.F) G ( E ) . both fields (displacements and stresses) are discretized in the finite element procedure. the mixed formulation equivalent to (3. (Arnold and Brezzi 1997). 207 the solution may be acceptably small if the order of the polynomials is sufficiently high and the mesh is sufficiently fine. and L2 denotes the product space obtained by taking L 2 ( w ) for each component of the stress term. Then.g. discontinuous shape functions are used for the stresses. (3. it is thus natural to introduce the membrane stresses (and the shear stresses for ReissnerMindlin type models) as auxiliary unknowns. Bathe 1996). This analysis.
see also (Iosilevich. This procedure has been shown to give very reliable results (i. (Bathe and Dvorkin 1986) and (Bucalem and Bathe 1993)).208 Dominique Chapelle however. have passed this numerical testing procedure. the dramatic dependance of the shell behaviour on the geometry (and the boundary conditions).e. two other reasons make it difficult to advocate the shell infsup test as strongly as for nearlyincompressible elasticity: 1. the problem of finding a couple of displacement/stress finite element spaces in order to satisfy the corresponding infsup condition is still open (cf. a numerical procedurecalled the infsup testhas been introduced in (Chapelle and Bathe 1993) t o test whether the infsup condition is satisfied without resorting to an actual mathematical proof. relies on a crucial infsup condition which can be written in the general form where s h denotes the discrete stress space and is a strictly positive constant independent of h and E. which implies that many confignrations . by clearly indicating whether the condition is satisfied or not) on numerous finite element procedures for which theoretical results where available. MITCS and MITC16 elements. However. Defining the seminorm (3. Bathe and Brezzi 1997) for plate formulations. cf. Iosilevich and Chapelle 2000) for shells. (Arnold and Brezzi 1997) for an attempt in this direction). Moreover. A similar test has been devised and used in (Bathe. the absence of theoretical results pertaining to shell infsup conditions does not allow a thorough validation of this test. due to the kinematical constraint introduced by the incompressibility condition.46) we have the equivalentmore tion classicalxpression of the infsup condi(3. which are based on a mixed formulation.47) This infsup condition requires a deep adequation between the displacement and stress finite element spaces in order to be fulfilled. whereas the corresponding standard finite elements (that are also called displacementbased methods to distinguish them from mixed methods) have all failed as expected. The MITC shell elements (MITC4. For shell formulations. In the case of nearlyincompressible elasticity where locking also arises.
the infsup condition that the test is based upon is in fact stronger than the infsup condition corresponding to the shell mixed formulation. But for shell problems to fall into that category.still more preoccupyingdifficulty. 2. instead of (3.. as regards locking as well as other difficulties featured by mixed finite elements for shells (see below).Some New Results and Current Challenges . Finally.48) This means that locking should not appear unless a very small thickness (or a coarse mesh) is used. 209 should probably be considered in order to reach an acceptable level of confidence (only one configuration was used for nearlyincompressible elasticity). convergence (and in particular uniform convergence) cannot be expected in this case unless the consistency of B& . However. the mixed formulation leads to a perturbation of the leading t e r m of the formulation. and a uniform error estimate is then obtained under the condition h 6 C&. (3. This method corresponds to a mixedstabilized formulation in which additional stabilizing terms are added to the mixed formulation in order to bypass the infsup condition. which means that a given finite element procedure could satisfy the actual infsup condition and yet fail in the numerical test.. another interesting idea. compared to Equation (3.45) is shown to hold. as regards the mathematical analysis of the locking of shell problems.44) we in fact need to consider B&(Zi. Therefore. a thorough numerical testing of this finite element procedure is lacking. to our best knowledge the only shell finite element method for which an analysis valid for general geometries has provided a uniform error estimate is the method given in (Chapelle and Stenberg 1998). Given the considerable difficulty to prove an infsup condition for shell finite elements. (3. which is that they are clearly tailored to the approximation of a problem in which a constraint is enforced by penalization.49) Hence.35) satisfied by the exact solution. . according to our above discussion the asymptotic behaviour must be of the bendingdominated type. Namely a mixed method is used for which a stability condition weaker than the infsup condition (3. is to relax the infsup condition while tolerating some degree of nonuniformity in the convergence of the approximate solutions. = G ( E ) . On the other hand. In the framework of shell models.E ) + c 2 B ~ ( Z EI). mixed formulations feature another . if the problem solved using a mixed method happens to be of the membranedominated type. proposed and analysed in (Bramble and Sun 1997). VE E Vh.
we believe that it is crucial. As a matter of fact. in general. Displacementbased finite elements work well in membranedominated situations but are subject to numerical locking in bendingdominated situations.3 Guidelines for assessing and improving the reliability of shell finite elements Whereas the complete mathematical analysis required to design and justify a general uniformlyconvergent finite element method for shell problems still seems out of reach. 3. Of course. benchmarks are both an essential interface and an invaluable means to assess the performance and the reliability of shell finite element methods. with extensive numerical testing on the other hand. in order to provide a rigorous and meaningful assessment tool. 0 Numerical computations must be performed and reported upon for several values of h (since we are concerned with the whole convergence behaviour and not only with the accuracy in one specific instance) and for several values of E (in order to assess the robustness of the convergence with respect to this parameter). in this reliability quest. . some detailed guidelines are given in (Chapelle and Bathe 1998). which rely on the following principles: A basic set of benchmarks must contain instances of the two fundamental asymptotic behaviours of shell structures. for the abovementioned mixedstabilized scheme that works well in bendingdominated cases. On the other hand. To that purpose. Several geometries must be considered. uniform convergence is not obtained in membranedominated situations. Therefore. to combine the insight provided by the analysis of the mathematical models and the elements of analysis available on their discretizations on the one hand. This condition is very difficult to enforce. since the asymptotic behaviour is highly sensitive on the geometry of the midsurface (for example a set of benchmarks only based on cylinders is not satisfactory) . we do not know of any shell finite element procedure that would have been proven to work well in both regimes. there is a major risk of seriously deteriorating the numerical solution in membranedominated cases. it appears that we are “trapped” in a dilemma caused by the dramatic discrepancy between the two major asymptotic states of shell problems. if we try to “fix” thebendingdominated behaviour by using mixed methods (or some kind of reduced integration strategy). and properly used. in particular because mixed methods are not designed to that purpose. In this interaction.210 Dominique Chapelle with respect to BM is strictly controlled. benchmarks must be very carefully chosen. As a matter of fact. hence also to improve these methods.
The shell shown in Figure 3 is clamped on one side and loaded. by self weight. respectively. noninhibited 0 . cf. cf. It can be shown that pure bending is inhibited for this problem (as a matter of fact. but we can use reference solutions obtained with a very fine mesh to compute error measures. Ovaskainen and Piila 1995) which also presented a procedure to obtain numerical solutions of arbitrary accuracy'.g. 1995) should be multiplied by (1. Figure 2.v 2 ) and (1 . This problem was originally proposed in (Pitkaranta. e. It consists of the following test problems: 0 Circular cylinder loaded by periodic pressure. 0 'The values of displacements and deformation energies given in (Pitkaranta et al. We do not have an analytical solution for this problem (see (Choi' 1999) where the exact solution of the limit problem is analysed).Some New Results and Current Challenges . Leino. Clamped hemispherical cap. wellposed membranedominated behaviour if both ends are clamped). ... It can be shown that pure bending is not inhibited. 211 P = PO cos(2sz) computational domain Figure 2 Cylinder loaded by periodic pressure In (Chapelle and Bathe 1998).v2)'.. Depending on the boundary conditions we can obtain either one of the two main asymptotic behaviours (bendingdominated behaviour if both ends are left free. Partly clamped hyperbolic paraboloid. a basic set of benchmarks is also proposed based on these principles. (Chapelle and Bathe 1998).
available for plates (see e. It is. . If these boundary layers are not properly taken into account in the choice of the discretization scheme (in particular by refining the mesh when and where appropriate). Pointwise displacements. Once a good set of benchmarks has been selected. it is crucial to measure the error made in the numerical computations by using proper error measures. is now becoming available for shells also through some more recent works (Pitkaranta. including the currently most widely used. Arnold and Falk 1996. they can strongly affect the convergence behaviour and induce wrong conclusions in the reliability assessment. Dauge and Yosibash 2000)). We now review and evaluate some possible choices.g. We also emphasize that a deep insight on the behaviour of the exact solutions is a n essential prerequisite in the assessment procedure. i.212 Dominique Chapelle Figure 3 Hyperbolic paraboloid pure bending is obtained for a hemispherical cap only when the boundary is completely free. If we choose an axisymmetric loading such as the one shown in Figure 4 we can compare the results of the shell analysis with the results of some (reliable) axisymmetric analysis scheme. This error measure is probably the most extensively used. (Haggblad and Bathe 1990. especially in the engineering literature. 1. Karamian. see (Ciarlet 2000). in the proper use of benchmarks. Matache and Schwab 1999. and also (Chapelle and Bathe 1998)).e. This insight. SanchezHubert and SanchezPalencia 2000). A particularly important feature of these solutions is the development of boundary layers which vary in amplitude and width when the thickness of the shell varies.
in general. Its major drawback. ) ]$ .Some New Results and Current Challenges . Z'). 2 . 213 t 2 p = po cos(2a) Figure 4 Hemispherical shell under axisymmetric loading however. in particular. Energy variation. This cannot be a good choice. namely Eh = E 2 B F ( 2 i .2. + E E 3. namely [E'BF(E. ) + BM(Zi. since the pointwise convergence of finite element methods is not guaranteed. (3. can scale the energy by any given power of E since we are concerned with relative errors).50) the energy of the finite element solution. Modified energy norm.51) .).d. 2. ) B!&.. namely [E'BF(E. 4. Energy norm.2) ' and &h +BM(Z'. beside the fact that its use with the exact solution (or any reference solution) is delicate (since the BLform is usually tailored to one particular discretization). is that the use of a nonphysical modification of the energy makes it difficult to interpret convergence in this norm. likely to be misleading. ) B M ( EE ) ]3 (note that we E .) (3..(E. namely I& . since it is the norm in which the displacementbased finite element solution is the best approximation.Ehl where & is the energy that corresponds to the exact (or reference) solution. namely + E = & 2 B F ( Z E . whereas this solution locks in bendingdominated situations. This choice is advocated in (Malinen and Pitkaranta n.
if the mesh is refined so that the boundary layers of a given test problem arc properly resolved for this norm. when boundary layers (namely. but this docs not hold for mixed methods. which shows that a displacementbased finite element procedure is unif o r m l y optimal with respect to E for the norm 1 1 . namelyas discussed in the introduction of this sectionthe norms in which exact solutions arc shown to converge to limit solutions. the norm of the displacement space for bendingdominated problems and the membrane energy norm for membranedominated problems. it can be shown that this measure equals the square of the error in the energy norm. Note that. Using the norms in which asymptotic convergence is obtained for the exact solutions.53)  vh Ellv}.d. since the Naghdi model is asymptotically equivalent we can use the corresponding energy expressions (and stiffness matrices) to compute adequate norms. However. see (Chapelle and Bathe 1998). Therefore. uniform optimal convergence is obtained. Indeed. This is illustrated in Figures 5 to 7. where the case of . when using general shell elements. as a consequence of our above discussion (cf. IIv. / 1 f e ~l l .214 Dominique Chapelle for displacementbased methods and for mixed methods.)). which makes the interpretation of this indicator very difficult (see also (Malinen and Pitkaranta n.1).e. for instance). ) + B M ( EE ) ]4 where L is a characteristic dimension E . for example [L2BF(E. M + 41ZE ZFlllV < C Einf ( 1 ‘ ell^ +&lIZE 12 E  (3. Asymptotic convergence norms. i. it is then natural to seek finite element procedures for which uniformlydecreasing errors can be measured on benchmarks such as those proposed above.. these computations are not straightforward since the membrane and bending energies are combined in the basic shell model. For practical purposes the membrane energy norm can be directly computed using its definition (and the corresponding stiffness matrices). Section 3. For displacementbased methods. the “nonsmooth part” of the exact solutions) are properly resolved. I/ZE zl . For the norm of the displacement space (typically a combination of Sobolev norms) it is more convenient to use an equivalent norm constructed with the deformation energy. 5. of the structure (the diameter. for a displacementbased finite element procedure used in a membranedominated case it can be shown that.
53).. Oliveira and Bucalem 2003). when the infsup condition is satisfied. the boundary layers) as in the rest of the domain. Figure 5 shows the convergence curves obtained for the membrane norm and for decreasing values of E (t denotes the actual thickness of the shell. see (Pitkaranta et al. if no particular treatment of the boundary layers is applied.Some New Results and Current Challenges . see (Chapelle. we obtain the convergence curves displayed in Figure 6 where some significant sensitivity with respect to E is observed. All these numerical results are clearly consistent with Equation (3. We can see that the convergence rate is optimal for all values of the thickness. 1995).. namely 2e) when as many layers of elements are used within the layer of width along the boundaries (i. The Naghdi model is used. l L 215 t/R=l/lO + t/R=lll00 . and Q2 displacementbased elements are employed to discretize the problem. By contrast.x t/R=111000 x Quadratic rate ! 01 7 E B h : e 001 7 0001 7 0 0001 I 1 10 100 Figure 5 Clamped cylinder: adapted mesh. if the original norm of V is used instead of the membrane norm (Figure 7) we again detect some marked sensitivity. membrane norm a clamped circular cylinder is considered (as described in Fig. and that there is very little sensitivity of the convergence behaviour on the value of E . a). we have the following optimal error bound (see (Brezzi and Fortin 1991)): .e. Also. for bendingdominated problems classical results for mixed formulations show that. This can be shown to be the correct way to resolve the boundary layers. Likewise.
membrane norm 1 0.0001 1 10 N=llh 100 Figure 7 Clamped cylinder: adapted mesh. original norm .0001 Figure 6 Clamped cylinder: nonadapted mesh.1 E b : UI e 0.1 E b L 0.001 0.216 1 Dominique Chapelle 0.01 W ? ! 0.01 0.001 0.
cf. where the singular part consists of strong boundary layers or of “generalized layers” that lie along asymptotic lines of the surface. even though this infsup condition cannot be proven. 2000).. we emphasize that all the aboveproposed benchmarks are based on problems for which a welldefined asymptotic limit exists. Moreover the displacements that develop in these layers appear to be very similar to pure bending displacements. Finally. the regularity of the stresses also needs to be considered to control the error in the seminorm 1 . We point out that the “ScordelisLo roof” (for which the geometry and the boundary conditions are defined in Fig. converge . and more specifically for which the asymptotic behaviour is either membranedominated or bendingdominated. as it can be shown that it corresponds to a situation of inhibited pure bending with “G $ (VM)”’. it does not seem to be possible to identify a norm for which an asymptotic limit exists. and in particular in cases when pure bending is inhibited and “G $ (VM)’”. Even though more analysis is still needed.54) for this same test problem and o n the sequence of meshes used for the infsup test (as it is the infsup values that enter in the bounding constants). In these cases. as previously mentioned we do not know of any shell finite element procedure that is proven t o satisfy the infsup condition. Note however that a procedure which passes the infsup test on a given test problem automatically satisfies the estimate (3. see in particular (Pitkaranta and SanchezPalencia 1997) and (Pitkaranta et al. Namely.e. however. see (Pitkaranta et al. 1999) and also (Karamian et al. these recent results give hope that finite element procedures which would perform well (i. and the loading corresponds t o selfweight). 8. hence it is difficult to see what kind of uniform convergence could be sought in finite element solutions. 217 Of course. Nevertheless. which appears to be one of the most widely used test problems in the engineering literature. it is the see loading in the vicinity of the free boundaries which causes this difficulty).1) do exist. This is why the use of such problems cannot be recommended to assess the reliability of finite element procedures. As previously discussed. belongs to this category.Some New Results and Current Challenges . other types of asymptotic behaviours (as defined in Section 3. IS (note that this seminorm is always bounded by the L2norm. namely a smooth part (in which the membrane energy is asymptotically dominant like in membranedominated situations) and a singular part.. Moreover. 1995) for examples. some recent works have shown that in some cases (more specifically when the shell midsurface is parabolic or hyperbolic) it is possible to split the solution into several components. (Chapelle and Stenberg 1998)). (Chapelle and Bathe 1998) (in fact. this estimate is very useful to determine how the meshes should be refined in order t o ascertain uniformly converging errors in the numerical results obtained with finite element schemes that “would satisfy” the infsup condition (hence lockingfree finite elements) .
363369.315. Baiocchi and C. J. . 13591380.N. C. SanchezPalencia (1992). D. Irons and O. Ph6nombnes de blocage membranaire’.C. Akian and E. 37. B. References S. Math.e. Ahmad. Math. Math. ‘Approximation de coques 6lastiques minces par facettes planes. J. Paris t. Math.218 Dominique Chapelle clamped boundaries \ \ \ I I I Figure 8 ScordelisLo Roof uniformly) in bendingdominated and membranedominated situations would also perform well in these more complex situations. ‘Asymptotic analysis of the boundary layer for the ReissnerMindlin plate model’. Sci. Comp. Acad.S. 66(217). Anal. Internat. Falk (1996).M. Lovadina (2002).N.L. Nurner. which is attested by the joint publications listed in the references. C. ‘Discretization by finite elements of a model parameter dependent problem’. Zienkiewicz (1970). Brezzi (1997). SBrie I. ‘Lockingfree finite element methods for shells’. Methods Engrg. D. Arnold and F. R. D. ‘A shell classification by interpolation’. 114. Arnold (1981). Sci. 486514. 27(2).N. converge uniformly for each of the abovementioned components separately. 419451. Numer. 2 . Arnold and R.405421. 12(10). i. Acknowledgement The author is particularly indebted to Professor KlausJiirgen Bathe as many ideas and results presented in this article have originated from a fruitful collaboration with him. S I A M J. ‘Analysis of thick and thin shell structures by curved finite elements’. Models Methods Appl.
52.L. ‘Computations on thin noninhibited hyperbolic elastic shells. Numer. Acad. Bathe (1993). Finite Element Procedures. 67. J.J. ‘A formulation of general shell elementsthe use of mixed interpolation of tensorial components’.831836. 219 K.N. ‘On the locking phenomenon for a class of elliptic problems’. 22.131152. Journal of Computational and Applied Mathematics 89. E. M. Appl. Methods Engrg. Chapelle and K. Math. ‘The infsup test’. Prentice Hall. Numer. Chapelle and K. and M. K. Bathe (1993). 12931321. Palma. Chapelle and R. A. 537545. Bathe and E. 6 Structures 81. Numer. F. Mode‘lisation des Structures par Ele‘ments Finis. Blouza. D. 289313. Springer Verlag. Paumier (1994). Sci. Oliveira. J. Bathe (2003). 63 Structures 75(5). Bathe (2000). Meth. Mode‘lisation Mathe‘matique et Analyse Numkrique 32.. Bathe (1998). J. Trouvh (1988). F. F.J.L.L.427440. J. Brezzi and M. J. Dhatt (1992). A.697722. Paris. D. Lovadina (1999). ‘Fundamental considerations for the finite element analysis of shell structures’. Brezzi and C.Some New Results and Current Challenges . Methods Engrg. 36. ‘Stabilized finite element formulations for shells in a bending dominated state’. ChoY. 40. ‘An infsup test for shell finite elements’. D.L.37293754.J. Chapelle. ‘Sur la classification des coques linhairement hlastiques’.J. 119133. Paris.44274449. Math. Sci. 22.J. Methods Engrg. ‘Shear deformable shell elements for large strains and rotations’. Internat. Bernadou (1996). Sun (1997). Bischoff and E.328. Chapelle (ZOOO).J. Benchmarks for membrane locking’. SanchezPalencia and M. Chapelle and K. Mixed and Hybrid Finite Element Methods. Bucalem and K. Internat. Chenais and J.3273. Ducatel and P. D. D.187217. Finite Element Methods for Thin Shell Problems. M. 63 Structures 47(4/5). Bathe.3: Coques. Se‘rie I t.J. J. D.523533. Bramble and T. ‘MITC elements for a classical shell model’. Hermes. SpringerVerlag. ChoY(1999). SIAM Journal of Numerical Analysis 36(1). ~ . Bucalem (2003). D. ‘Approximation of a circular cylindrical shell by CloughJohnson flat plate finite elements’. K. D. C.. 439456. The Finite Element Analysis of Shells Fundamentals. R. Stenberg (1998). Dvorkin (1986). Numer. Comput. John Wiley & Sons. International Journal for Numerical Methods in Engineering 48(2). Chapelle and K. M. M. Iosilevich and D. Fortin (1991). Numer. Bathe (1996).1936. vol.C. ‘A lockingfree finite element method for Naghdi shells’. Vilarino (1998). Computers 63 Structures 66. ‘Higherorder MITC general shell elements’. Bernadou. Math. Comput.Ramm (1997). ‘Membrane locking in the finite element computation of very thin elastic shells’. D. Batoz and G. 3 D. Y . Comput. ‘The mathematical shell model underlying general shell elements’. Internat.
‘A model problem for boundary layers of thin elastic shells’.103214. R.).24432471.3138. ‘Boundary layer realization in thin elastic 3D domains and 2D hierarchic plate models’. 4. K. J. J. Wetensch. Methods Engrg. 61. Wiley and Sons. Malinen and J. Math. Brezzi (1997). 154. Ciarlet (ZOOO). Yosibash (ZOOO). Interfaces and Transitions (M. International Journal for Numerical Methods in Engineering 30. Ahmad (1980). American Mathematical Society.77107. The Finite Element Method for Elliptic Problems. Solids Structures 37. Mathematical Modelling and Numerical Analysis 34. Proc. in Progress in Solid Mechanics. J. P. Dauge and Z. ‘Influence of rotary inertia and shear on flexural motion of isotropic elastic plates’. Kon. NorthHolland. Haggblad and K. ‘On evaluating the infsup condition for plate bending elements’. 36393663.127134. J. Mathematical Elasticity . Ned. Numer.130. Bathe (1990). A.L. P. J. Numer. Math. 1990. Journal of Applied Mechanics 18. Ciarlet (1978). . Intrinsic differential geometric methods in the asymptotic analysis of linear thin shells. Iosilevich. Delfour (1998). Malinen (2000). Koiter (1965).G. R. M. P. SpringerVerlag. Lyly (2000). Methods Engrg. ed. 40. NorthHolland. P. J.Laboratory for Mechanics of Materials. B. SBrie IIb.220 Dominique Chapelle P. Karamian.T. Pitkaranta and E. J. Numer.J.J. Akad. London.G. M.d. Providence.9811011. NorthHolland. Bathe and F. M. Acad. 190. ‘The problem of membrane locking in finite element analysis of cylindrical shells’.M. in Boundaries. 85. ‘On the asymptotic behaviour of sensitive shells with small thickness’. Foundations of elastic shell theory. Internat. CRM Proceedings & Lecture Notes. ‘Mathematical modelling of linearly elastic shells’. Vol. SanchezPalencia (1997). Numer. On geometrically incompatible bilinear shell elements and classical shell models. W. ‘On the connection between some linear triangular ReissnerMindlin plate bending elements’. Irons and S. J. SanchezHubert and E. pp. M. Pitkaranta (n. B69. SanchezPalencia (2000). ‘On the nonlinear theory of thin elastic shells’. Techniques of Finite Elements.G. Helsinki University of Technology . C. B. Lions (1973). Sci. Internat. Pitkaranta (1992). Delfour. M.523542. Paris 325.Volume 111: Theory of Shells. Mindlin (1951). To appear in Internat. Research Report TKKLo30. A benchmark study of reducedstrain shell finite elements: quadratic schemes. Perturbations SinguliRres dans les ProblRmes aux Limites et en ContrGle Optimal. ‘Specifications of boundary conditions for Reissner/Mindlin plate bending finite elements’. Ciarlet (2001). Naghdi (1963). Acta Nurnerica 10.M.). pp.
E. I. Sci. ‘The effect of transverse shear deformation on the bending of elastic plates’. Coques Elastiques Minces . Paris t. Pures Appl.Proprie‘tb Asymptotiques. ‘Shell deformation states and the finite element method: a benchmark study of cylindrical shells’. 71. R. 81121.309. Cas de flexion pure non inhibge’.Some New Results a n d Current Challenges . R. SBrie I. SanchezPalencia (1992). 411417.379406. Cas de. 0. Math.. Acad. Engrg. SanchezPalencia (1989b). Paris t. Y . 128. Methods Appl. E. Ovaskainen and J. SanchezHubert and E. J. E. Research Report SAM 9918. 221 J. A. C. Pitkaranta. Schwab (1999). SanchezPalencia (1997). ‘Statique et dynamique des coques minces. Mech.flexion pure inhibke . Piila (1995). E. C. ‘Statique et dynamique des coques minces.Approximation membranaire’. ETH Zrich. . SanchezPalencia (1989a). J. Leino. Journal of Applied Mechanics 67. Comput. 11. Masson. Pitkaranta. 531537. ‘Asymptotic and spectral properties of a class of singularstiff problems’.309. Paris. Reissner (1945). Acad. Sci. Fourier mode analysis of layers in shallow shell deformations.M.A69A77. Matache and C.. J . SBrie I.
The term geometric refers here to the manner we are considering the numerical solution of a partial derivative equation. P. Their purposes is t o reproduce by calculation the behavior of a system usually described by a model based o n partial derivative equations.222 A Differential Geometry Approach to Mesh Generation Pascal Frey Laboratoire J. Lions. we present a mesh adaptation scheme based on a geometric error estimate. This allows to study the variations of the solution in terms of the calculus of variations and more precisely to take advantage of the powerful tool of differential geometry to properly characterize the surface. from biological sciences are based on partial differential . France Abstract In this chapter. Theoretical issues about error estimate in H 1 seminorm or in L p norms are discussed. engineering sciences. L.” Excerpt from the inaugural lesson of Professor Pierre Louis Lions at the Collkge de France. if not any. rue du Chevaleret. for instance a twodimensional Cartesian surface embedded in EX3. chemistry. finance and. mechanics. 75013 Paris. the numerical simulations appeared a privileged tool of investigation in science and technology. Lions shows clearly that many. Universite‘ Pierre et Marie Curie 175. the solution is envisaged as a manifold embedded in a higher dimensional space. even more recently. Paris. Numerical examples from various fields of applications are presented to emphasize the potential of this approach. meteorology. This short excerpt of the inaugural lesson of Pr.L. problems resulting from physics. the main concepts of mesh generation and adaptation are briefly presented and practical issues related to the implementation of the numerical algorithms are discussed. june 2003. Indeed. Introduction “In a f e w decades. The rise of the numerical simulations thus reinforce the need for the mathematical study (analysis) of these equations and for their numerical resolution.
c2u. linear (1st order) transport equation. This information is usually called an initial or final condition (thus referring to time) or a boundary condition (referring to space). Burgers equation (non linear. u u + UU. bounded). Here. Usually. can be . + The goal of numerical simulations. what is usually called an algorithm or a numerical method. Such a problem is characterized by the existence of a unique solution and it enjoys a property of stability that allows to control the solution by the data of the problem in certain norms. if the equation contains partial derivatives of functions of independent variables with respect to some of these variables.&Uzz = 0 . The general solution of an ODE involve constants whereas the general solution of a PDE contains arbitrary functions. the equation is then a general ordinary differential equation (ODE).uYY . = 0 . t utt .. The same applies for industrial fields (aeronautics. It remains to be defined the way the computer will carry out the computation. more precisely. at least. 2nd order). such that the difference J J u u h J Jin a given norm. the aim is to compute a numerical solution function u h as accurately as possible (ie. ut u . A% = f. Burgers with dissipation term. car manufacturers.close to the analytical solution or. Let us recall that a differential equation is an equation containing in addition to the independent variables (and functions of these variables). = 0 . the solution u of such problem is not known analytically. are written some traditional equations of the physics: ut + cu. In general.). = 0.. In this context. a fundamental notion emerges: the notion of a wellposed problem in the sense of Hadamard. etc. 1st order).A Differential Geometry Approach to Mesh Generation 223 equations (PDEs in short).. additional information on the boundary of the domain of interest (in space and time) are associated with the PDE. On the other hand. wave propagation (linear. then the equation is a partial differential equation (PDE) or.uZY= f(z.. biharmonic equation. nu = f. The classical objective of a numerical simulation is to predict by calculation the behavior of the system under investigation that is described by a mathematical model. y). for each PDE or each model. derivatives or differentials of these functions. MongeAmphre equation. There is a continuous dependence of the solution with re . Notice here the fundamental difference between ODE’S and PDE’s. If the functions that apppear in the differential equation depend on an independent variable. space agencies. u. Actually. a differential equation with partial derivatives. Poisson equation (equilibrium problems). Hence. all these phenomena involve values that are solutions of PDE systems.
It is thus a matter of studying the behavior of the approximations when the parameter of the discretization ( h or t ) tends toward 0 and of comparing these approximate solutions to the exact solution. physics. The socalled model is a simplified description of the phenomenon expressed as a set of differential equations supplied with boundary conditions. a new tool has rapidly proved essential in the context of numerical simulations: the adaptation of the domain discretization (the triangulation or the mesh Th) to the solution. Hence. Stage (4) involves the notions of algorithm. Stage (1) establishes the mathematical translation of a given problem. Consequently. it is important to prove mathematically that the numerical results are indeed accurate approximations of the solution. Formally speaking. we will mainly set the emphasis on the steps (3) and (4). Stage (3) is made necessary because we can only consider a finite (limited) number of numerical values. On the other hand. complexity as well as programming and is usually referred to as numerical simulation. The finite element method used to solve such a problem relies on the approximation of the solution using polynomial functions on each element K of a triangulation Th of R. For linear elliptic problems. Why this . In this paper. it is easy to prove the convergence of this method.. notably because of the nonlinear models arising in most of the physical problems. Similarly. Problem statement. it is of the utmost importance to raise the question of knowing whether the problem is indeed wellposed before even considering computing an approximation of its solution. Let us consider a PDE problem on a closed bounded domain R c Rd and let us denote by u the exact solution of this problem.224 Pascal Frey spect to the initial data.that involves h as well as the global regularity of u. Stage (2) has as an ambition to study the validity of the models by exhibiting properties satisfied by the solution of the equations. it is possible to compute an a priori error estimation of the difference between the exact solution u and the numerical (discrete) solution u h . when the maximal diameter h of the elements tends towards zero.) (1) (2) Figure 1 Flowchart of the numerical solving of PDE problems. the resolution of a general PDE problem can be sketched as Figure 1: Mathematical Analysis Algorithm Analysis (3) (maths. This is indeed a tedious task.
) is a continuous elliptic bilinear form on X x X and (. for the rather simple following reasons. Approximation of the solution. ii) it places the problem of searching solutions within the very convenient framework of the best approximation in the sense of the norm associated with a scalar product defined in Hilbert spaces. with respect to a quadratic form representing the energy. i. where u ( . one needs a tool to measure the discrepancy between the exact and the numerical solution. given a desired accuracy value. Or conversely. More precisely. Hence. a mesh adapted to the solution of the problem a t hand. almost everywhere). . in a few iterations only. Given a fixed number of degrees of freedom.. Let us go back for a short while to the concept of finite element approximation solutions for elliptic problems. However. Therefore.u)”. V u E X . such solutions appear as minimizing functions. .e. The variational formulation has numerous advantages among which we rank the following: i) from the theoretical point of view.. in numerous test cases and realistic problems. the variational method places the search for the solutions of elliptic equations within the general framework of Hilbert spaces. these regions are usually not known a priori. on which the approximation error is now below the target value. mesh adaptation allows to improve the accuracy of the numerical solution u h by adjusting the mesh density locally. on a subspace or a given convex domain. The aim of an error estimqte or an error indicator is to accurately identify the regions where the solution is discontinuous and to provide useful information about the local mesh density to reduce the error. the desired solution u presents some singularities in various regions of the domain and numerical experiments tend to show that the approximation error decreases as the triangulation is refined (elements are subdivided into smaller ones) at these places. it allows to dramatically reduce the number of degrees of freedom of the discrete problem as compared to a constant size mesh. Indeed. let us consider the variational formulation of an elliptic problem: ‘3nd u E Banach space X such that. it allows for the existence of a weak solution of the problem (this solution satisfies the equation and the boundary conditions only in a weak sense. .A Differential Geometry Approach to Mesh Generation 225 method is it become popular ? Well.And let us consider the finite element approximation of the problem as well: .) denotes the duality product between X f and X (f is supposed t o be in X f ) . This information will be then used in combination with a mesh refinement technique to construct. For instance.u)= (f. u(u.
u h ) ” . However. Another concern is that this computation is specific t o each . an a posteriori error estimation is of the following form: Ilu . the functions of of X h are polynomials on each element K E Th. On the other hand. the solution u h obtained using finite element methods is usually not interpolant: the numerical solution U h at the nodes 2 h of T h does not coincide with the exact value u(zh)at these nodes.uh) = ( f . Moreover. It allows t o control as well as to ensure the overall stability of the method. ) f its existence ensures the convergence of the methods when the solution u is sufficiently regular. All that contributes to the difficulty of explicitly computing the error llu .K E Th allow t o adaptively refine the mesh.U h ll < G(h. f ) in which the righthand side term can be explicitly computed this time as it involves the numerical solution obtained on the given triangulation of the domain. An a priori error estimation can be written as follows: IIu .W l l X 6 F ( h . Such an estimation corresponds for instance to the estimation of the energy: it ensures that the energy of the solution is controlled by the energy of the righthand side term f.U h . the socalled error estimate. essentially based on the calculus of residuals and the resolution of a global or local problem. Let q be the second term of the a posteriori estimation. They can be considered as efficient. a(uh. reliable and robust [18].uh((.226 Pascal Frey ‘snd u h E X h c X such that. Ideally. one would like to have inequalities like: crl(K) 6 IIUuhllX(K) 6 rl(K)> where X ( K ) denotes the space of the restrictions t o K of the functions of space X . it can be written as: the quantities q ( K ) .u. it is not even possible to guarantee that the numerical solution u h coincide with the exact solution u in at least one point in each element K E Th. Several error estimates have been proposed in the context of the finite element methods. By assumption. An interpolation error estimate. triangulations of 0. Vuh E X h . We introduce a regular set (Th)r.
controlling the interpolation error leads to controlling the approximation error as well. .). V h ) = L(vh). . here u(u. a ) is the constant related to the continuity (resp. and to deviate from the physical problem (related to the mathematical operators). Now. this problem involves: .A Differential Geometry Approach to Mesh Generation 227 type of equation and must be remade with each change of equation. The numerical solution u h is known t o converge toward the exact solution u as the diameter h tends toward zero. to achieve a higher level of abstraction. .v ) is a bilinear continuous and coercitive (aelliptic) form on V and L ( v ) is a linear continuous form on V . ” From the practical point of view. U ( U h .. According to Cea’s lemma for elliptic problems [8] the following inequality holds: .the adaptation of mesh Th to the metric specifications.the construction of an error estimate for the interpolation error. a ( u . hence. .” We have the following result regarding the error between the analytical and the numerical solution (known as the CBa’s lemma): where M > 0 (resp. the ellipticity) of the bilinear form u(. we consider that the weak formulation on V : “find u E V such that for all w E V .the definition of an anisotropic metric tensor field. construct a mesh TL # Th o n which the interpolation error IluPhull is bounded in all directions by a given tolerance value. 1u .Phull. one can deduce that the approximation error tends toward the interpolation error. Hence. h Let V be a real Hilbert space and let v c V be a subspace of V of finite dimension.U h l l X 1 < c llu  PhUllX Thus. Now. Here. the former adaptation problem can be rewritten as follows: “Given a linear interpolation Phu of u with respect to a mesh Th. And we would like to solve the projected problem: “find U h E v h such that f o r all Vh E vh. let PhU denotes the linear interpolate of the solution u on each element K E Th and let us consider the interpolation error IIu . given Th a triangulation of the domain R E EXd. v ) = L(v)” holds. we advocate here the use of an indirect approach to compute the error.
228 Pascal Frey Although it does not appears clearly in the previous statement. Paper outline.(.1 Triangulations and meshes Definition 1. 1 Preliminary definitions In this section. are called the barycentric coordinates of point x with respect . not all in the same hyperplane. such that the 1 matrix A of order d + 1 + + is invertible. to a j . we will focuss on mesh adaptation issues and on controlling the geometric accuracy in surface mesh generation.. we will introduce a geometric error estimate for the interpolation error and suggest various manners of computing approximations of second order derivatives of a discrete function uh.e. In Section 1. we will provide some definitiogs about basic notions of triangulations and metric tensors. this problem is closely related to another approximation problem: the generation of a piecewise linear approximation for a surface manifold embedded in EX3. Section 2. Finally. i. Section 3. metric tensors and differential geometry. < j 6 d 1. its associated matrax being A: j d j=1 The )A . x = (xi) is fully specified b y the data of d f 1 scalar values X j = Xj(x) that are solutions of the following linear system. All the items listed in the previous paragraph will be reviewed in the following sections. we will give and comment numerical examples to emphasize the potential of differential geometry applied to mesh generation. A simplex K whose vertices are the aj is the convex hull of the a j . This paper will attempt to reveal how these two problems are intimately related.1 (simplex). Let us consider d 1 points a j = ( ~ i j ) $E~ Rd. Every point x E EXd. 1. we will provide the reader with some basic notions and definitions about mesh generation. Section 4.
2. the condition ii) providing the conformity property (Figure 2). # 0.1simplex. K~ E R h . Figure 2 Conforming (left) and nonconforming triangulations (right).1. A simplicia1 set T is a triangulation of the domain R if: i) T is a simplicia1 covering up of R. iv) K Condition i) indicates that R and Rh must have the same topology. a 0face is a single point. 8. ii) VP vertex of dRh. We like to introduce Oh = U K as a simplicial set: the union of simplices of dimension d. Given these two definitions. Definition 1.n KJ= 0.3. A triangulation is also called a conformal covering up of R. . The Euler formula and the DehnSommerville relations relate the number of kfaces in a triangulation of R. ii) the intersection of two simplices is either empty or reduced to a d . Definition 1. Remark 1. Conditions iii) and iv) have been introduced to proscribe overlapping elements as well as degenerated (null volume) elements. it is now possible to define a triangulation of R. V K E Rh. a h is a simplicia1 covering up of i) iii) ah provided that: is homotopic to R. We like to call kface a simplex of dimension Ic < d . P E Oh. The previous definition does not give any hint about the existence nor about the uniqueness of the triangulation of the domain R.1. we define the notion of covering up. i # j .A Differential Geometry Approach to Mesh Generation 229 Let us consider a bounded domain R c Rd and denote 6 9 its boundary. v K i . Then.
Definition 1.2. The Delaunay triangulation of a given set of points in general positions' exists and is unique.1.k<d denotes the number of its kfaces. . More precisely. the above formule becomes: + n.230 Pascal Frey Definition 1. the socalled Delaunay criterion or the empty sphere criterion. Hence. T is the Delaunay triangulation of R if the open ball circumscribed t o any of its element contains no vertex of T . +nf  n. the following relation holds: + n. = 2 .  n.5. then n.na n f = 2. n f = 1. Remark 1. the number of holes). the Delaunay triangulation can still be defined using the previous definition. means that the open ball associated with each element K does not contain any vertex of T but the vertices of K . I n three dimensions.4. I f the points are not all in general positions.n. its Euler characteristic is equal t o 1 (l)d. This criterion.ne+nf=2c where c corresponds t o the number of connected components of 8 R . W h e n the triangulation is homeomorphic t o the topological sphere.I n two dimensions. . one triangulation is of particular interest in computational geometry: the Delaunay triangulation. if the domain R includes n o hole. Remark 1.3. + Among the various possible types of triangulations. A fundamental theoretical issue follows then: Theorem 1. The existence is given by construction and the uniqueness property is achieved since the points are in general position by assumption.. .29 where g stands for the genus of the surface (Le. Proof: this result is obtained trivially by involving the duality with the Voronoi' diagram associated with the set of points. 'Points in general positions: no more than d + 1 cocyclic or cospherical points. Let T be a triangulation of R. although it will contain nonsimplicia1 polytopes. thus considering a simplicia1 triangulation. The Euler characteristic x of a triangulation T is given by the alternate sum: d k=O where nk. the triangulation of a close surface is such that n.
providing physical information and attributes to the solver. and unstructured meshes.A Differential Geometry Approach to Mesh Generation 231 In the context of finite element methods. 1. thus relating the letter h to the local diameter of the elements. Then. where the connectivity is like the Cartesian grid using in finite differences schemes. let us recall that creating a mesh of a computational domain requires at first discretizing the domain boundaries as accurately as possible and then creating vertices within the coveringup of the domain. the accuracy and the quality of the discretization. the inradius or the length of the largest edge.1. We will call such a triangulation a mesh.1 Numerical simulations In numerical simulations based on the finite element method. To summarize. Defining the connectivity of a mesh as the type of connection between its vertices. exhibiting any non Cartesian type of connectivity. nodes and vertices are exactly coincident. computations usually involve the following steps: i) mesh generation: managing the geometric approximation (spatial discretization) of the domain boundaries. and not only the vertices. containing polyt. additional concerns must be stressed. iii) matrix assembly: discretization of the equations. ii) interpolation: for constructing finite elements. Obviously. The latter is usually a conforming coveringup or a conforming triangulation of R. regarding in particular the existence of the triangulation. boundary and/or initial conditions must be assigned to mesh elements. . that will be associated with the degrees of freedom. We denote by Th a mesh of a domain R. For Lagrange PI elements. It is easy to see that structured meshes are usually convenient for very simple geometries while unstructured meshes are more flexible to handle complex geometries. By extension. it is required to generate a discretization of the domain R. allow us to introduce the notion of structured meshes. It is also possible to consider nonsimplicia1 meshes or mixed meshes. These two stages can be carried out separately or simultaneously. one can also define the connectivity of a mesh element in terms of the connections between its vertices. In particular. for instance.opes of different dimensions. it is necessary to define all nodes. iv) numerical resolution of the resulting (linear) system.
Remark 1. Usually the domain geometry as well as the field of application are the key points t o take into account f o r chosing a method. . we suggest the following classification [20]: MM manual or semiautomatic methods: usually applicable to geometrically simple domains. Two types of such methods can be identified: algebraic interpolation methods (transfinite interpolation from boundary curves) and solutionbased methods (numerical resolution of a PDE on the computational domain). anisotropic). IM point insertion incremental methods: consist in creating and inserting vertices in the domain. regarding the number as well as the locations of mesh vertices ( i e . 1. octree) belong to this class. Advancingfront and Delaunaybased methods are representatives of this class. PM parameterization (mapping) methods: the mesh is the result of the inverse transformation mapping of a regular lattice of points in a parametric space to the physical (real) space.3 Mesh adaptation Solutionadaptive mesh generation is a very promising and challenging technique that is targeted at improving the accuracy of the numerical solution at a lower computational cost. close to a domain of reference. as mesh generation methods have been developed in different contexts and are usually targeted at different fields of applications. each submesh may have been created by any of the previous methods. The identification of the method suitable for solving the mesh generation problem at hand is a dificult and tedious task. and also on the shape of the mesh elements (isotropic vs. the density). Both block decomposition methods and spatial decomposition methods (e. Both methods rely on a given discretization of the domain boundaries. .2 A brief survey of mesh generation methods Despite many conceptual differences. CM constructive methods: the final mesh results from the merging of several submeshes using topological or geometrical transformations. DD domain decomposition methods: the mesh is the result of a topdown analysis consisting in splitting the domain into smaller geometrically simpler domains.1.232 Pascal Frey 1.4. Enumeration methods (mesh entities are userspecified) and explicit methods (taking advantage of the geometry of the domain) are representative of this class. It relies on a more efficient (optimal) point distribution.g.1.
1. Mesh adaptation can be carried out 1 u by using either a global approach (advancingfront or Delaunaybased method. is the canonical inner product of Euclidean space. for instance) or a local reconstruction (optimization) method or a combination of globallocal techniques.A Differential Geometry Approach to Mesh Generation 233 After a numerical computation has been performed on a given initial mesh Ti. in some regions. 5.2 Notion of metric tensor Lengths and distances play important roles in a mesh generation process.5. . resume the adaption loop: i = i + 1. but requires consequent modifications for unsteady (time dependent) problems [l]. More precisely.. i=l n . . the length of a vector or the distance between two points in space are based on the definition of the inner product2. The mesh adaptation scheme described here is well suited for steadystate problems. Remark 1. i 3. 2The scalar (or dot) product of two vectors u = [ui. In addition to adapting the mesh element size.v. i 6. estimation of the local error lleill and definition of a metric field Mi.u] v = [vi. Algebraic results indicate that the inner product is related to a positivedefinite nondegenerate symmetric bilinear form.. (u. . .] . E EXn. the socalled hmethod. go to 2. . interpolation (projection) of the solution u on Ti+l. the local error being then bounded by above on each element. the numerical solution is analyzed using an a posteriori error estimate and the result of this analysis is converted into a metric field providing shape and size specifications. A new mesh Ti+l is then created using an adaptive technique from the mesh Ti. construction of an initial mesh To (i = 0 ) . an alternative consists in adapting the order of the interpolation for the finite elements (the socalled pmethod). computation of the numerical solution u on Ti. using Lagrange Pk elements.uhll or the interpolation error 1 .. The whole procedure is repeated until convergence of this scheme. 4. where IleilI represents either the approximation error IIu . 2. reconstruction of a mesh Ti+l according to M i . in order to equidistribute the approximation/interpolation error over the mesh elements. .v) = C uivi. The overall mesh adaptation scheme can be summarized as follows: 1. lc > 1. From the mathematical point of view.IIhuII.
y . E V .x ) E F .y ) = ( y . a metric or a distance function is a function which defines a distance between the elements of a set. This notion has been first introduced by M.2) o n the manifold’s tangent spaces. The norm of a vector x is simply the expression of the length of vector x.1 Metric. it is a (covariant) tensor field of type (0. ii) linearity (in the first variable): iii) n o n negativity (positive definiteness): v x E v . satisfying the axioms: f i) conjugate symmetry:  V x . hence. So.6. the inner product can be defined as a map: (.234 Pascal F’rey Definition 1.7. also known as a Riemannian metric. at p E M . 1. x . we will present metricrelated operations. V x E V . scalar product and distance In mathematics.x ) . Then.2. Obviously.y) = ( y . Frkhet in his work in 1906 [19] . In this section.x ) 2 0 ( iv) n o n degeneracy: (x. At any point in the manifold.. m. A metric tensor i s a symmetric tensor field g of rank 2 that is used t o measure distance in a space. the . V y E V i i f x = O . if F = R then the first property is the symmetry of the inner product: ( x . Every inner product space has a natural norm defined as: )1x11 = the non negativity axiom of the definition making this definition always possible.y)=O. ( x . such a set is then called a metric space.x ) . this tensor takes a pair of vectors in the tangent space to that point and associates a real number. . we will adopt the differential geometry notion of metric tensor which refers to a structure defined only on a vector space. Definition 1. we will review the definition of a metric tensor and of the associated scalar product. Here. Considering a vector space V over the scalar field F .) : V x V F. given a smooth manifold. a metric induces a topology on the set. As a consequence. ( x .
The curves which locally join their points along shortest paths are then called geodesics. it becomes easy to define the length of a segment [a. . where gij = g (ddXj . on a connected manifold every Riemannian metric tensor becomes a metric space and induces a distance metric on M .M + R is symmetric and positive definite. known as Cartesian tensors. we refer the reader to Ciarlet [ll]and to the chapter devoted to differential geometry in this book for more details about this notion of metric tensor and. If a system of local coordinate xi is fixed.b] of a curve parameterized by t as: 1= ] a ddt. the covariant components of a metric tensor. these two notions are strictly equivalent for tensors in threedimensional Euclidean space. in a Riemannian manifold. given by: where the infimum is taken over all the rectifiable curves y : [0. A Riemannian metric tensor on a manifold M is not a distance metric on M . 1 + M 1 with y(0) = z and $1) = y . we can write: 9= [.A Differential Geometry Approach to Mesh Generation 235 bilinear form gp : T p M x T. If we consider the twodimensional Euclidean metric tensor. . the metric tensor appears as a matrix. It can be seen as an inner product.6. Remark 1. In other words. L dXi ’ While the distinction between covariant and contravariant indices is of importance for general tensors. denoted g = ( g i j ) and has a unique expression: n g= i+l gijdxi 8 d x j . However.el . in particular. and the length of a curve simply becomes the wellknown formula: At this point.
known as the CauchySchwarz inequality and the triangle inequality: Definition 1..2. according to the previous inequalities. 1 &xGj =d ziz. then a linear m a p T : V 4 V is called diagonalizable. . 1. 1 + Y Then. k Y E R".. vx # 0 E Rn is only satisfied if the matrix M is positive definite. For a Hermitian matrix (symmetric in the real case). it can be proven the following inequalities. i. if its n eigenvalues are positive.236 Pascal Frey As an example. ) M defined by: (X. . with the equality only if x and y are linearly dependent (colinear). we define the Euclidean n o r m of any vector in E" as: % 1211 = 111.10. From the axioms of Definition 1. . with positive scale factors and orthogonal directions of scaling. we use the characterization of a diagonalizable matrix. Let (. it actually measures the length of vector x according t o metric M . Given the inner product (. is an inner product.Y)M = x T M y .2 Metric decomposition We know that a positive definite (non degenerated) matrix M is the matrix of a positive quadratic form f on R.. or as the dot product of the results of directionally differential scaling of the two vectors x and y. E v .9. Definition 1. .) be a n inner product defined o n vector space V .. the norm defined by the inner product is indeed a norm on V . and we have: 12 + YII 6 0 ~ 1 1 IIYII . This can be interpreted as an orthogonal transformation. the condition xTMx > 0 . Here. it is easy to introduce: Definition 1.e. vx7Y E v . If V is afinitedimensional vector space.6. ) M in the vector space Rn. where M represents any positive definite matrix.8.Y)l G ll~ll. With these two definitions.11Yll. in the vector space R". the general form (. if there exists a basis of V with respect t o which T is represented by a diagonal matrix. yx. then we have: l(X.
e. is: 1. A square matrix A is diagonalizable if it is similar to a diagonal matrix or i f it exists an invertible matrix P such that PIAP is a diagonal matrix. we can write that Definition 1. $ R. a metric M will be represented by its unit ball. Thus.7. In R3. we will associate a metric tensor M with a symmetric positive definite matrix also denoted M . then M = XI2. where R is a matrix composed of the eigenvectors and A is the matrix constructed from the corresponding eigenvalues.R can be considered as a combination of rotation 2 matrices. it is possible to make this decomposition to be unique.2. However..A Differential Geometry Approach to Mesh Generation 237 The following sufficient condition is often used: the map T is diagonalizable if it has exactly n = d i m ( V ) distinct eigenvalues. . This indicates that M is the matrix representation of the linear map T with respect to some basis. For instance in R 2 . It follows that a symmetric positive definite matrix3 M can be decomposed in the special form: M=RAR~. the columns of R and A can be permuted. In R3. as any eigenvector of the eigen subspace is associated to an eigenvalue. In the case of two distinct d eigenvalues. 1. if we have a double eigenvalue A. i.3 Geometric representation Definition 1. The geometric shape of the unit ball is totally dependent o n the choice of the norm. where 0 denotes the center of the ellipsoid. The main axes of the ellipsoid are given by the eigenvectors of the matrix M and the corresponding radii are given by the square root of the inverse of the eigenvalues of M . Remark 1. E v 7 1 x1 < 1) 1l ' B y extension. its characteristic polynomial has n distinct roots. 1. R can be seen as a rotation (orthogonal) matrix of angle IT 0. 1 1. supplied with the norm 11. { d m 3For the sake of simplicity. It is sufficient to take R = I and A = M .12. an ellipsoid defined as: E = p E R 3 / l l O p l l ~ = = 1}.11. The closed unit ball in a normed vector space V . f ( x ) = 1) is called the unit sphere of V . i f V is a linear space with a real quadratic f o r m f : V then { x E V . Notice that this decomposition is not unique. 0 E [O.
hence. This will enable us to consider the notions of metric and its geometric representation as equivalent.a diagonal matrix of M. Similarly. we will define an intersection procedure. we would like to define the largest ellipsoid E included in the intersection of the two ellipsoids El and E2 associated with the two metric tensors: We consider two quadratic forms q and q' on a real vector space. we would like to deal with a single metric at the vertex. We have seen that a matrix M representing a metric can be associated with a quadratic form. For various reasons. the terms of the diagonal matrices associated to the metric tensors M and N are obtained using the Rayleigh formula: Xi = eTMei T and pi = ei N e i . If q is positive definite. Denoting by P = (el e2 e3) the invertible matrix of GL3(R) formed by the eigenvectors ei. The matrix N' is diagonalizable in R and admits a basis of eigenvectors (orthogonal) with respect to q.3. This is the aim of the simultaneous reduction of the quadratic forms. Given a n ellipsoid E of center 0. . 3 of the matrix N'. it is possible to find a basis of the vector space in which the matrices associated with q and q' are represented respectively by I and D . m a i n axes are given by the eigenvectors and radii are associated with the eigenvalues of M . 1. we have: . Considering the problem from the geometric point of view. e 2 . the matrix N' = M' N is introduced.2. it is possible to analyze its unit ball with respect to the metric. To achieve this result. including mesh generation purposes. i = 1 . Hence. it exists a unique metric tensor M such that the geometric locus of the points satisfying the relation ll0pll = 1 describes the ellipsoid. i = 1.8. e 3 )in which M = M ( q ) and N = M ( q ' ) are congruent to a diagonal matrix and then to deduce the metric tensor of the intersection. to analyze a metric tensor (or a matrix). In this basis.(R). the idea is to find a basis ( e l . In other words.238 Pascal Frey Remark 1. To this end.4 Metric intersection Let suppose that two metric tensors are specified at a point p E Rd.
CLz) 0 pl. We are looking for the metric tensor at t .t ) p + t q (a mesh edge./13) ) is a symmetric positive definite matrix as: det Mn = (P')z det A where A = (max(Xi. hence for a matrix Mt defined along the segment c ( t ) for any value of the parameter t E [0.2.9. ii) in the case of isotropic metric tensors. Remark 1.5 Metric interpolation Let us consider a parametrization of the segment p q as c : [0. Hence Mn is a metric tensor according to Definition 1.p i ) ) .1] + R2. i) this definition does not necessarily imply that the eigenvectors of any of the two metrics M or N are preserved. 1. are two symmetric positive definite matrices representing two metric tensors associated with the endpoints.7. then N' = N . c(t) = (1 . iii) i f N' admits a triple eigenvalue a.Pl) 0 0 0 m=(Xz.Xz). M = diag(X1) and N = diag(Xz). Figure 3 Example of metric intersection in two and three dimensions using the simultaneous reduction of two quadratic forms. then i f a 6 1.1]. 0 0 max(X3.A Differential Geometry Approach to Mesh Generation Finally. then Mn = diag(X) where X = max(X1. we have N' = M and i f a > 1. the intersection matrix Mn defined as: 239 Mn = (PT)' ( rnax(X1. The definition of this matrice Mt involves the interpolation of the eigenvalues of the matrices M p and . for instance) and suppose that Mp and M.
we will introduce the main concepts about the differential geometry of surfaces required to understand the point of view adopted in the following section. Denoting by P = (el e2 e3) the invertible matrix of G L 3 ( R ) formed by the eigenvectors ei of the matrix N = MplMql we compute (Xi)i=1. $m . in the basis {ei}i=1. In particular. Finally.3 and (pi)i=1. we define hi = l / f i and hi = 1 1 6 .291. 1. In other words. O where Hi(t)i=1. + Figure 4 Example of metric interpolation in two dimensions. for i = 1 .3 A differential geometry primer In this short section. 7 (t) HI 1 0 0 )F1<t<l..240 Pascal F’rey M. To this end.3 are affine functions such that: H i ( t ) = (1 . for which the metric at any point c ( t ) is given by: Mt = ( ( 1 O<t<l. Differential geometry is often seen as the calculus on manifolds or as the study of the different aspects of curvature. This procedure allows to define a continuous metric field along the segment. this scheme is similar to reducing simultaneously the two quadratic forms associated with the metrics.16. Then. Finding the interpolated metric tensor Mt requires to express the two matrices in a basis {ei} in which both are congruent to a diagonal matrix and then to deduce the metric tensor at point t. detailled proofs of the main results given here can be found in these textbooks. The reader interested by differential geometry is invited to read the chapter of this book devoted to this topic or to consult reference books like [ l l . This topic emerged with . 3 .t)hb th$.3.3 the eigenvalues of M p and M . we suggest a linear interpolation scheme. Figure 4 shows the result of linear interpolation of metric tensors in two dimensions. the metric tensor Mt is given by: Mt=(PT)l ( .
. The surface f is regular at p = (u.F. and p is then a regular point. f u ) is a basis of the vector space TpR3. in which this great mathematician explains that the whole geometry of a curved surface is fully determined from the formula for the line element d s . Definition 1. the tangent plan inherit from the Euclidean structure of the ambient space. Let f : U + R3 be a surface. i) A differentiable mapping f : U + R3 such that dfu : TpR2 + Tf(p)R3 injective f o r all p E U is a parameterized surface. T h e vector field: n := fu Ifu x fu x ful i s called the Gauss unit normal field along f.X ( p ) E TpR3. A vector field along f = (f: U + R3} is a differentiable mapping X : U + R3. the notion of curvature is related to the deviation from flatness at any point in a differential manifold.A Differential Geometry Approach to Mesh Generation 241 a paper of K.1 The first fundamental form. The moving frame (f u . Let U denotes a n open set in R2 and p = (u. area One of the most important structures carried by a surface is the first fundamental form. the differential study of surfaces aims at finding invariants that do not depend on the choice of the parameterization. f u . The tangent space at p is a vector space of dimension two. Note that ( f u . . In other words. which we will now introduce. Definition 1.A vector field X along f takes values in the tangent space of R3 restricted to the surface f.3. The natural inner product of U c R3 induces on each tangent plane TpS of a regular surface U an inner product (.) E R2 v be a point of the set U . 1. ii) The twodimensional linear subspace d f p ( R g ) c T f ( p ) R 3 called is the tangent space of f at p and is denoted as Tp ... One of the most important concept in differential geometry is that of tangent space. this in turn is determined by metric.14. isomorphic to R2. A mapping is f satisfying this condition is called regular. Gauss entitled Disquisition generales circa superficies curvas (General investigations of curved surfaces) published in 1827. Intuitively. The mapping n : U + S 2 c R3 is also referred to as the Gauss map. n) is called the Gauss frame of the surface f : U + R3. Tangent vectors are the partial derivatives of the position function.v ) E U if and only if fu x f u # 0.13. Elements of Tp are f f called tangent vectors off at p . ) p . More generally.
This f o r m is called the first fundamental form and is denoted by g p or Ip: I. F = According to the previous definition. The inner product of U c R3 which is a symmetric M bilinear form induces a quadratic form o n Tpf c Tf(p$R3 R3 by restriction. : TpR2 M R2 + R2 M Tp . where u : I + U c R2 is a curve in U . we define a tangent vector C ( t ) to f a t u ( t ) : C(t) = c i tii 0 u ( t ). the arclength parameter s ( t ) is uniquely defined by: o . The first fundamental f o r m I of a surface f : U + R3 is positive definite. Then. ) where = Commonly. k ( p ) . (gik(p))is also equivalent to the matrix notation I( e i . this form allows t o make measurements on the surface such as lengths of curves and areas of regions. (2) The inner product o n Tf(p!IK3 composed with the linear m a p df.i(p) . These two definitions can be identified when identifying TpR2with Tpf by means of the isometry df. Let f : U + R3 be a surface. The matrix representation of the first fundamental form. e2) of TpR2. for instance.16. Geometf rically. Definition 1.(W) = ( w .fw).1) ' d f U . arc@) f. we use Gauss' tensor notation: E = g(fu. fu). e k ) of I with respect to the canonical basis ( e l . with respect t o the basis f u l . = s(fv. : R2 + Tf(p)R3 induces a quadratzc form o n TpR2 which is also called the first fundamental f o r m and also denoted by g or I . w ) p= 1w12 2 0 . For a curwe c ( t ) = f u(t) o n a surface f .1.242 Pascal Frey Definition 1. and the length of c(t)is given by: I C( t )1 = c i>j g i j (u( ) ) 2 ) t i j (t) t (t Definition 1. f u 2 is denoted by: ( 9 i d = ( d f u i 7 fU. f .17. A curve on the surface f is a curve c : I + R3 which can be written as f o u. and we have: Proposition 1. f w G.15.
F 2 d u dv . is given by the formulas: = Lo tl dEU2 + 2FUV + GV2dt . f ( p ) = a } . If a regular surface is given by S = { p E R3. ( d udv = d E G . However. As mentioned. p H n ( p ) . A regular surface S c R3 is orientable i f and only if there exists a differentiable field of unit normal vectors n : S + R3 o n S. the surface S is orientable. Considering a differentiable field of unit normal vectors on an open set U c S . The element of area d A of a regular surface contained in the coordinate neighborhood of the parametrization X : U c R2 + S is given by the positive number: d A = IXu A X . where f : U c R3 + R is diflerentiable and a is a regular value of f .A Differential Geometry Approach to Mesh Generation 243 From the definition.18. Hence.2. 1. This is the main concept behind the definition of a Riemannian manifold.2 The second fundamental form. v(t0))and q = ( u ( t l )v ( t 1 ) ) . Proposition 1. a differentiable mapping n : U + R3 associating to each p E U a unit normal vector n ( p ) E R3 to S at q..19.3. The vector field of fu x f. we come to the following definition: Definition 1. i. The area does not depend on the choice of the parametrization. f for all u E U . We introduce the following definitions: . on a curved surface. the first fundamental form is not sufEcient to characterize the surface. We can also refer to ds2 = Edu2 2Fdudv Gdv2 as a Riemannian metric.e. the first fundamental form allows areas to be computed on a surface. + + Definition 1. curvatures Let f : U + R3 be a surface and let consider the Gauss map n : U + S2 c R3. it determines is a normal orientation of the tangent plane. for which n maps p into the unit normal vector n ( p ) E S2 to f at p . we can see that the length 1 of an arc of curve on the surface joining two points p = (u(to). the inner product induced by the Riemannian metric on the tangent space at every point changes as the point moves on the surface. X ( u ) ) E Tf(u)R3 orthogonal to T. Hence. is a normal vector field along f as (f(u).
f is: (hik) = (n.f u k ) = ( n. The normal curvature in the direction f X is the number: . If 1 is proportional t o I (11 = . then the associated value . Let X E T. f . More precisely.i .1. := drip o df.3 (Rodriguez).f is called the Weingarten map.(X. A vector X O E 5. By extension. Let f : U + R3 be a surface and let Sif = { X E T.(X)=rI(X. 1 = 1 be a unit tangent vector on a x 1 surface f. else there exists exactly two (up t o sign) principal directions orthogonal t o each other.(X) = . every direction is 1 a principal direction. f f . : TpR2x TpR2+ R is called the second fundamental form off at p and is denoted by 1 or 1 II.(X) = I I ( X . .I).21. or. X ). and the associated eigenvalues are the principal curvatures.f.22. The matrix representation of IIp with respect to the canonical basis { e i } of TpR2and the associated basis { f. = dn. Definition 1. Definition 1. f u i u k ) . o df’ : Tp + T.f + T. ~1 and 62 the largest and smallest eigenvalues. Let X E S i f .. re Corollary 1. The quadratic form: dn. if X Ois a principal direction.} of T.(Xo) is called a principal curvature of f at p .X)ER. We will denote spectively. T h e linear mapping L. is ’f said t o be a principal direction i f X O is a critical point of the function: X E S j f H. Principal curvatures are characterized by the result: Proposition 1. f lI.. ‘ df.20. T h e n X is a principal direction i f and only if X is a n eigenvector of the Weingarten map: L.’ : T.244 Pascal F’rey Definition 1. we will show the relation between the second fundamental form and the local curvatures of a surface. if using Gauss’ notation: (b3 = G::2 ) ‘ Now. X ) = 1) denotes the unit circle in T.
’ + + where the lefthand side is the characteristic polynomial of the Weingarten map L. 2 H ( p ) = TrL. Consequently. + R3. Regarding the intrinsic properties of a surface. a positive constant.5. PO is o ) o ) called a planar point. Definition 1. Let c ( t ) = f ou(t) be a curve o n a surface f : U c ( t ) is a geodesic if V C ( t ) / d t = 0 . t . T h e n f for suficiently small E > 0 there exists a unique geodesic c ( t ) = f o u(t). I1 < E satisfying the initial conditions u(0)= po and C(0) = X . we can state: Definition 1. I t i s spherical i f there exists xo E R3 such that J f ( p ) xo) = p. Let f : U + EX3 be a surface. is < 0. Definition 1. Let X E Tpo be a tangent vector to a surface f . A surface is said to be planar if n ( p ) is constant. A surface f : U elliptic parabolic hyperbolic + R3 is: >O =0 at po E U if detII.4. Definition 1.A Differential Geometry Approach t o Mesh Generation 245 We will now use the principal curvatures to introduce two fundamental notions.23.. in K .2Htc K . Theorem 1.26.24. and K ( p ) = detL.25. A point po is called an umbilic if ~ ~ ( p = ~ ~ ( p oI)n . More precisely. if K I C ~ O )= ~ ~ ( p= 0 . The curvature functions are determined by the equation: det(K id d n o df I) = K . Let f : U + R3 be a surface. A surface is composed of only umbilic points i f and only i f it is planar or spherical. Proposition 1.2. and we have: We will use these definitions to characterize points on the surface. addition. T h e Gauss curvature and the mean curvature off are the following functions o n U : Proposition 1.
we introduce the six Christoffel symbols: where gij.j+ C(rfjr2 rfkr.l and l?Fj. are quadratic forms o n TpW2. we briefly review the concept of error estimation . Let U be a n open.246 Pascal Frey As a consequence. all the nonconstant geodesics on a sphere are great circles. B an isometry of W3. g i k . This result concludes our short introduction to the differential geometry of curves and surfaces. we can give the fundamental result. l . This will allow us t o write the following result: Proposition 1 6 (Integrability conditions). p E U .) = C(hijhk1 . whose coeficients ( g i k ( p ) ) and ( h i k ( p ) ) are differentiable functions of p . known as Bonnet’s theorem: Theorem 1. then: i) there exists a surface f : U + W3 whose first and second fundamental forms are I. In this section. 2 A geometric error estimate Anisotropic error estimation has been studied extensively in the last two decades. hik. CodazziMainardi equations provide relations between g i k . The theorem asserts that the knowledge of the first and second fundamental forms determines a surface locally. If I.3 (Fundamental theorem of surface theory). h i k .k = d g i j / d u k . simply connected subset of W2.z: rF. and IIp. The Gauss formula and .k  r2. We will use these concepts in Section 3.  (3) I I Finally. and II.h i k h j z ) g z . when dealing with the tedious problem of meshing surfaces. To give compatibility equations for the theory of surfaces. ii) any two surfaces f and f” defined o n U which have the same first and second fundamental forms differ b y a n isometry: f =B of . Suppose I. is positive definite and the Gauss and CodazziMainardi equations are satisfied.
w)( A ( ( w (( (( ~ 1 1 . such that (u(w. for instance..wE V . We assume the reader to be familiar with the concept of Hilbert spaces and with variational (weak) formulation for elliptic problems. a ( u . we will recall the classical error bound given by Cea’s lemma [9] and already mentioned in the introduction of this chapter. A bilinear form a(.1 A priori error analysis We will primarily focuss on error estimation in the context of finite element methods (or FEM in short).. .) is continous means that: 3A4 > 0.wh) = 0.) is aelliptic or coercive if 3a > 0. For an elliptic problem of order 2. <4 and a(. Vwh E Vh indicates that u h is the projection of u on Vh according to the scalar product a4 and represents the 4Recall that for real vector spaces.1 Interpolation error in L2 norm and H 1 seminorm Let a be a bounded domain of Rd and let u be the solution of an elliptic differential problem on under Dirichlet boundary conditions on dol the domain boundary. 2. a is a bilinear continuous elliptic form.(( represents the norm defined in the space V . vw E V ” where V is a Hilbert space. In addition] we will explain how this bound will provide a characterization of the optimal mesh elements] both in terms of shape and of size.1.v) = l(v). 2.vw. the positivedefinite nondegenerate sesquilinear form called an inner product is actually a positivedefinite nondegenerate symmetric bilinear form. . Moreover] if the form a is symmetric.uh.A Differential Geometry Approach to Mesh Generation 247 and we propose an approach to obtain an anisotropic bound for the interpolation error suitable for linear simplicia1elements in any dimension. such that la(wlw)I < Q 11w((2l v E~V Let Vh be the internal approximation space constructed with a given finite element method (Vh c V ) . for which we can write a variational formulation: ‘fJind u E v such that a(u. The concepts presented in this section can however be transposed to other classical discretization methods such as finite differences or finite volumes. 1 is a linear continuous form on V . .the following inequality holds: where uh denotes the discrete solution associated with Vh and I].
. The previous inequalities are only valid i f the same forms a and 1 are used in the exact and approached problems.n H 1 norm on the domain 0. 6) represent the longest edge and the ) diameter of the inscribed circle of the element (resp. in other words.Weuse the following inequalities. then it exists a constant a! such that: . .2. The main idea of a priori error estimates consists in bounding the approximation error of the exact solution by the interpolation error by chosing the PhU interpolate of u as specific 2/h in the finite element space v . reference element) (Figure 5). h Let us denote by lI. Let assume the domain 0 is covered by a mesh Th. for all w E C1: IlVfjll < IlBll l l V 4 and IlVvll 6 IIBlll IlVfjll . the following relations hold: where h (resp i and p (resp. this requires to compute the integrals exactly. The first point is to notice that the norm on 0 can be obtained from the norms on each element of the mesh Th. it remains to evaluate the interpolation error in the L2 and H1 norms depending on the choice of the finite elements used for the approximation [lo]. let us consider the affine transformation from a reference triangle to an arbitrary simplicia1 element K . respectively. If the mesh Th is uniform ( i e . .Phu111. At this stage.1. We will the provide an upper bound on the approximation error 1u .2$ inI volving the interpolation error IIu . the interpolation error can be written as: Now. a better upper bound can be provided: Remark 2. In such a case.$.Uhl11.l1~.248 Pascal F'rey best approximation in a subspace of finite dimension. of constant size). Let B denotes the matrix of this transformation.2. In particular.
~defined as: we obtain. suficiently smooth (u E H'++'(R)). For finite elements of degree k (Pk or &k) and for an exact solution u. 2 .A Differential Geometry Approach to Mesh Generation 249 Figure 5 Reference element and arbitray triangle. . Theorem 2.1 (General). given the integral seminorms l w l m . for all w E H " ( K ) We use then the inequality: and we consider the interpolation error in the reference element for polynomials elements of degree k : as well as the Relations (6) and the hypothesis (7) to deduce the following fundamental result.the following bounds of the norms L2 and H 1 hold: where the constant C is independent of h. to deduce the inequalities] for all w E Cm: and thus.
Uh111.uh111.zi].I . bl (8) (9) and finally 11 .max (u”(x)[.xi11.Phu(x)I 6 .b] h .b ] .2 11 ’ < C h xE[a. it is easy to deduce that: ( P h U Y ( Z ) = U(Xil)WL(X) U(zi)w:(X). respectively. Vx E [a.uhJJ a constant mulby tiplied by the interpolation error IIu .xi]. .PhuII. + thus. for all z ) = in [zil.rnax Iu”(x)I.1. M a VVh E vh Chosing for vh the interpolate Phu of u in the finite element space Vh. This leads to the following results: Theorem 2. Their can restriction in the interval [ x ~ . E[a. . the following inequalities: ). We will use the upper bound on the approximation error introduced previously (Relation (5)): IIU .zi]. Vh is the space of the piecewise affine continous functions on each segment Ti. classically denote wi1 and wi the two basis we functions of v h associated with the points xi1 and xi.b] hence.~ w ~ . max + (Xi12u (Ei )i “ Proof: the sketch of the proof is classically based on Taylor expansions: .. U(5i1)= U(X) + (Xi1 U(Zi) = U(X) + (Xi  X)U’(X) X)U’(X) + (Xi  X ) z U”(Vi).Phu111.I ( z ziwZ(z) X .250 2. Here. 2 xt[a.2 < . x ~ ] be written as follows: Then.2 Pascal Frey P elements in one dimension I Let consider a given bounded interval [a.(. it i s possible to es tablish.bl Iu”(z)I.2.% 1 ) 1 .llu .. 2 . Denoting h = it{ 1.2 6 C h max Iu”(x)I. we have wii(z) wi(z)= 1 and x ~ . Iu’(x) PAu(x)I < h2 8 xE[a. IIu .I . allows us to bound the approximation error Jlu.well as the relation: as PhU(z) = U(Xil)Wil(X) + + + U(Zi)Wi(X). N } max (xi. In each interval [zi1.b] subdivided into N subintervals Ti = [xil.
it is possible to specify the size hi for each element Ki t o globally equidistribute the interpolation error as follows: 0 in L2 norm: hi in H 1 norm: hi c ( (rnaxIu”(z)I)i xEKi 0 0: (max Iu”(z)I)’.z)2w:1(z)u”(Ei) + (3%  z)2w:(z)21”(77i)] 6 Remark 2.z) wil(z) (Xi . xEKi It can be shown that each of these choices minimizes the L2 and H 1 global errors.Phu111.). . on the other hand.. we write: 1 . respectively.~ ( z ) u ”+E i ).A Differential Geometry Approach to Mesh Generation Using the previous Taylor expansions yields to: 251 + #Xi’ z) wzl(z)u”(&) + (Xi . I n each element Ki of length hi local error bounds are obtained according to: IIu .”(&) 2 + (Xi .2. = 1 ( (xi 2 PhU(Z) = u ( z )  2 2 1 On the one hand.z)”Wi(z)] 6 2 + .)2Wil(. using the relation: PhU(2) = u ( z ) 1 + [(5i1 .2 ) 2 w i ( 2 ) u ~ ~ ( r ] i ) ] .2.z)2wi(z)u”(?7i)] 1 and (Phu)‘(z) u’(z) + [(zi~ z) ~ ~ . we write the inequality: and we deduce the upper bound: [(Zil 1 2 .~ ) ~ w : ( z ) u ” ( 7 7 i ) ] . 8 h2 and. For instance for the L2 error. < Chi xEKi Iu”(z)I max Hence.K.
xi1). an interpolation error equal to 0 on u . . the maximum value of the interpolation error of the solution u is obtained for the midpoint of the interval.252 Pascal F’rey the interpolation error IIu . w e take for the approximation space v . the derivative of the interpolate of u. the approxih mation error IIu . By introducing hi = xi . is a constant function which is equal to the value of u at ’ the midpoint of the interval. 2. I n such a case.3. in the same way as we did in dimension one. xi]. we chose for vh the interpolate PhU of u in the finite element space v .x)(x .Uhll can be bounded by a constant multiplied by the interpolation error IIu . let us consider a polygonal domain R c R2 covered by a simplicia1 h mesh Th. This leads to the following relation: hi / /. Again. g = hz+i vi = 1. denoted PAu. we have a maximum value of the interpolation error on u and. then u’ is an a f i n e function and u” becomes a constant.. we obtain finally the following bound: The minimum is obtained by computing the gradient of this error with respect to the points xi.1. This explains why it is ‘ interesting to estimate the derivatives of the solution at the midpoints of the intervals.n is bounded by the term: The explicit computation of [(~i~x)~wi_~(x)+(xix)’w~(x)] to leads the simple expression: (xi.. we will make use of the fundamental result given by Relation (5): Here. I n other words.a. .xi1. simultaneously. at the midpoint of [xil. Indeed. Then.PhUllo. .PhuII.3 Lagrange PI elements in two dimensions Now.N  1 Remark 2.the space of piecewise affine continuous functions on each simplex Ki. Assuming u is a polynomial of degree 2.
Y) = C U ( Z i .A Differential Geometry Approach to Mesh Generation 253 Let us denote by h the longest edge of a triangle. that are the restrictions to the triangle Ki of the basis functions of v h associated with the three vertices. consequently we deduce easily that: 3 V ( P h U ) ( Z . a= CYzxi(Z.Y)=Y. 26 C h and finally.YEn l l ~ 2 ~ ( x . for all ~ . Yi)Xi(Z. Vx. i=l cZiX2(Z.3. Y) . 3 3 3 cxi(Z.Y)=Z.~)II X. i=l Using the following Taylor expansions: _____ 5We recall that the spectral norm of a matrix A is the largest singular value of A or the square root of the largest eigenvalue of the positivesemidefinite matrix AA".YER SUP (11) X. by 00 the smallest internal angle among all mesh simplices.y) 11 its spectral It is now possible t o enounce the following result. and by D2u the Hessian matrix of a C 2 function u defined on R2: and by llD2u(x. i=l and. ~ ) l l (12) Proof.2: Theorem 2. llu . Then. Yi)VXi(Z. EyKi. we denote classically by X i the barycentric coordinates. . (see [24] for more details) in each triangle Ki. 2 6 C h SUP IID2u(x. Y) = c 3 U(Zi.Y)=l. E R.Y). the following inequalities hold: hence. echoing Theorem 2. y. i=l 1 and PhU(Z.P h ~ 1 1 1 .~ h l l 1 . IIu .
y ) Pascal Rey = u(x. On the other hand.D2u(&q)MAa (14) and.Y)I = 2lTI AjAk where h~ denotes the length of the longest edge in the mesh. the triangle for which this term is minimal is that f o r which sin80 is maximal. ) y 1 + i=3 i=l d __f & ( x . It will be minimal for a minimal h Among all triangles having a longest edge of length value of . Hence. Indeed.254 we are able to obtain: Ph'LL(X.4 (Optimality). we use the following inequality: 4 I 1 2 h$ sin(&) T to obtain: n and the desired result: Remark 2. using Relation (13). I n the upper bound o n the error. szn(80) h. it can easily be shown that: hence. the error on the gradient value can be bounded by above: Noticing that IVXi(X. where I 1 denotes T the area of T and OT represents the smallest interior angle in the mesh. we will consider the equilateral triangle as the optimal triangle.y)MAi. . the dominant term is related to the gradients. we have finally the equation: Obtaining the error bounds is now a simple exercise. I t is also the triangle of maximal surface inscribed within a given circle.
at the circumcenter (which coincide with the center of gravity). the gradient of u is an affine function and D2ubecomes a constant matrix. the Hessian matrix D2u is a scalar matrix. we rely on Relations (14) and (15). A more clever analysis consists in looking for the eigenvalues and the eigenvectors of the Hessian matrix D2u in each triangle and to consider as optimal triangle the triangle of maximal surface inscribed within an ellipse. the following expressions still hold: as well as: Again. this defines a new scalar product and consequently a new metric for which the concentric circles of the error isovalues become concentric ellipses of fixed excentricity. Then. when considering an equilateral triangle. The . The gradient VPhU of PU is a constant function equal to the value of V u at the center of h gravity of the triangle. which is null at the center and maximal at the triangle vertices. It is exactly the opposite for the error on V u . Hence.A Differential Geometry Approach to Mesh Generation 255 Let us suppose now that u is a polynomial function of degree 2. Moreover. Thus. it is the error on the gradients that will dictate the optimal shape of the triangles in the mesh. for non isotropic problems in which the solution changes rapidly in one privileged direction. but only diagonal. To this end. the previous analysis proves to be at least insufficient. In anisotropic problems. On the other hand. In each simplex. to ensure positive values. considering a homogeneous and isotropic problem. This analysis explains why it is desirable to estimate the derivatives of the solution u at the center of gravity of the elements. we are taking the absolute values of the eigenvalues. the Hessian matrices diagonalized in their eigenbasis are not scalar anymore. Hence. Hence. the maximum value of the interpolation error of the solution u is achieved at the center of the circumscribed circle (the circumcenter). we have simultaneously a maximum value of the interpolation error of u and a minimal (null) value of the interpolation error on Vu. we observe that the error on u is null on the ellipse of fixed excentricity circumscribed to the triangle and is maximal at its center. If we denote by: the new symmetric positivedefinite matrix associated with the scalar product.
as it is known by the H1 norm.1 Anisotropic formulation of the interpolation error In the previous section. let K denote a mesh element in any dimension.2. any vector v c K can be expre as a linear combination of the edges of the simplex K. More generally.y)\v). y£K v<ZK (16) where cd is a constant only related to the dimension of the domain (specifically. it will then nds on both the size and the shape of the triangles (Figurd 6).2 Mesh adaptation using the error estimate We will now approach the numerical and algorithmic aspects related to the implementation of the error bounds previously established.K max max{v. c2 = 219 and c3 = 9/32). Indeed. To overcome this problem. However.]D2u(x. we will first replace the second maximum (on the vector v c K) by introducing esh edges.256 Pascal Frey interpolation error on u depends only on the size of the elements. the global bound on the interpolation error can be expressed as: \\uPhu\\oo. Hence. max e€EK . in the context of mesh adaptation for numerical simulations. it follows the result: VvCK. this bound is not fully usable as it involves the estim of two maxima that cannot be computed numerically. 2. from a pra point of view. Regarding the approximation error. Figure 6 and i lues of the error and optimal triangle in the isotropic case n in the non isotropic case (righthand side). we have established a bound of the interpolation error of the solution u at each mesh point. 2.
only two points are sufficient to describe exactly the solution. the number of points and edges required to exactly describe it is a n infinite set.y)}. the node distribution must be proportional t o the local curvature.5. The answer to this problem requires the computation or the knowledge of the firstorder derivatives of the underlying parametrized arc to identify the extrema. To represent the curve with a given accuracy would require the knowledge of the intrinsic properties (the local curvature) of this curve. 11 max (17) N Remark 2. namely the secondorder derivatives. In other words. ii) W e could also notice that the interpolation error is related t o the lengths of the edges of a mesh. controlling the lengths of the mesh edges will enable us to control the interpolation error o n each element. as well as the variation of the derivatives. if the solution is represented by a curve (a polynomial function). Overcoming this problem requires to be able to find a metric tensor FK. If we consider a linear solution in one dimension (a line).y. the solution t o this problem can be seen as a parametric Cartesian surface embedded in IR3 : C = {(z. Then.PhulI. Hence. the previous bound can be rewritten as follows: But again. We will show hereafter how to compute the metric tensor & when approximating the metric ID2ul. i) This expression indicates that the interpolation error o n K is proportional t o the s q u a r e d t h e length of the longest edge of K measured according to the matrix M K . This will reveal extremely useful for mesh generation purposes. the interpolation error EK on the element K is evaluated as follows: E K = 1 2 . M K e ) .A Differential Geometry Approach t o Mesh Generation 257 where EK denotes the set of edges in the element K . Then. In two dimensions. On the other hand. The maximum over the metric field 1D2ul is not known. F w ) / 'v'w c K } has a minimal K surface.a(z. .~ = cd e E E K ( e . we obtain the explicit bound: IIu  Hence. the coefficient of proportionality being related to the desired accuracy (a tolerance value). hence a symmetric positive definite matrix on the element K such that: G G and such that the region defined by: { ( w .. the righthand side term is numerically tedious to compute.
MKe)= 1 ( ( e ( ( M K 1. the interpolation error on an element K is of order E if the length of all its edges is equal to 1 when computed according to the matrix M K . because it is usually more convenient . to introduce secondorder derivatives (the Hessian) of u in the evaluation.3 Metric construction A metric tensor.258 Pascal Frey Similarly.~. over a domain R c Rd is a data of infinite dimension that cannot be used as such for numerical purposes.M K . Hence. More precisely.. In other words. In such a case. where E &  = IluPphu((.2. we have yet to explain how to define it analytically and eventually how to compute it numerically. However. Such a mesh will be then called a unit mesh.MK e ) . If we introduce the matrix MK the edges must be such that: = cd . Let E denote the maximal error value on each mesh element K . the aim is to equidistribute the interpolation error in each direction over the mesh elements in order to control the approximation error.. the interpolation error on the edges of an element K is such that: EK = llu . at least. then (e.PhuII. This can explain why the error estimate for the interpolation error is naturally called a geometric error estimate. the solution can be seen as a hypersurface embedded in El4. And we can use the same reasoning to introduce secondorder derivatives in the expression of the interpolation error. we will consider a metric defined at each mesh vertex and represented by a symmetric positive definite matrix. According to Relation (17). Obtaining a mesh on which the interpolation error is equidistributed is a problem equivalent of constructing a mesh in which all edges have a length according to MK equal to 1. 2.K = cd max ( e . 2. as a symmetric tensor field. In three dimensions. the secondorder derivatives allow us to introduce anisotropy. the control of the piecewise linear approximation of the surface C requires. eEEK  b'e E E K .2. = b'e E E K . we will use a discrete approximation of the metric tensor. considering the mesh as the underlying support.2 Problem statement In most meshing problems. or at least numerically very close to 1. as the matrix M K indicated in the above relation is not known.
6. in this case. the interpolation error is still bounded by E . Remark 2.4 Evaluation of the Hessian matrix The expression of the bound on the interpolation error.2.. Notice that this metric tensor is naturally anisotropic. we define at each mesh vertex the anisotropic mesh metric as: and R is the matrix composed of the eigenvector. Relation (16). can be obtained using for instance a finite element method. y ) . the main directions are given by the eigenvectors and the sizes by the positive eigenvalues di. although it will not necessarily be equidistributed.cd is a constant value related only to the dimension. according to Relation (17). Thus. iii) to define an isotropic metric. .piecewise linear in this case. maximal) edge size. Let E denote the maximal interpolation error accepted over the mesh and let hmin (resp.A Differential Geometry Approach to Mesh Generation 259 for finite element simulations where the unknowns and the vertices coincide.d and M = c:) . Hence. it is highly desirable to have a good estimation of the Hessian in order to define an accurate metric tensor. Then. Relation (17). We have then to reconstruct an approximation of the Hessian of the solution u based on the discrete solution uh. It is also a way of controlling the timestepping in a flow solver. ii) the new scalar product provided b y the metric tensor is related to the Hessian of the variables of the problem at hand. only a discrete solution uh. h. involves the Hessian of the exact solution of the problem and this matrix appears also in the metric construction. 2.) represent the minimal (resp. However. the di represent the eigenvalues of the Hessian matrix D 2 u ( x . one can simply consider: d = max(di) a=l. the Hessian of u h is null in the interior of an element and is a Dirac mass on its boundary. This metric field is made continuous using a classical interpolation scheme. as the exact solution is not known. i) introducing a minimal and a maximal sizes is a practical way of avoiding unrealistic metn'cs.
we associate a basis function (pi E v h and we denote as si the support of (pi: s = supppi. In order t o compute the interpolation error. Let consider the local L2 projection operator. 0 a) A dual L2 projection scheme. all the terms present in the righthand side of the inequality are well known or can be computed. At each vertex p i . the By interpolation error can be bounded by above in the following manner: Ib*where. we recall the Clkment interpolate. we uses a double L2 projection and a Clkment interpolation operator to reconstruct the Hessian matrix.7. Let consider a mesh Th of a domain fl C @. This approach usually provides a good accuracy for the Hessian. Remark 2.and to evaluate the approximation error as the norm of the difference between the reconstructed solution u*and the discrete solution u h : ((uu h ( (   ((u*uh(( . a i s not t necessary to explicitly know u* o n fl. We denote by v = {w € Hi(fl2)such thatwllc E p k } h the approximation space associated with a Lagrange P finite element. We will now describe two different approaches to numerically compute the Hessian value of the solution: 0 on the first hand. k where Pk is the space of polynomials with scalar coefficients of degree less or equal t o k . i At first. as on the other hand. we use a Taylor expansion and a resolution of the linear system obtained by a leastsquare approximation. Moreover. The idea consists in projecting w E L2(fl) on P k independantly on each Si c Th.nalogy with Relation (16). u* may not be defined. the idea is to reconstruct a C2 continuous solution u*. it allows to reconstruct a Hessian matrix associated with a solution u*.uhll = IIu* . although it does not guarantee that the reconstructed Hessian is associated with u*. for w E L2(fl). it is suficient here to reconstruct its Hessian matrix. we can write IIu* . As such. as the solution u* coincide with uh at the mesh vertices. More precisely. we define now E L2(R) as follows: < < . 1 i n. based on the discrete solution u h . a.Phu*II.260 Pascal F’rey In order to evaluate the righthand side of Relation (16). in this case.
we can apply the same procedure to each component of this gradient. the weight being related to the surface of the elements. it becomes possible to define an operator to reconstruct the gradient as follows: We still have to explain how to reconstruct a PI gradient from uh. is Notice that this scheme is simply a reconstruction using a weighted average.A Differential Geometry Approach to Mesh Generation The Clkment operator lIc: L2 V is defined as: h  26 1 Now. we have by definition. the operator to reconstruct the gradient leads to: or similarly: VRUh(Pi)= KES. For each Si. in order to reconstruct the Hessian matrix. The operator to reconstruct the Hessian matrix is given by the following relation: H*(Uh) HR(Uh) = + H*(WJT 2 . for each vertex p i . b'p E Po: where /KI (resp. We consider the case where k = 0 and we apply the L2 projection operator t o vuh E L'(S2). As u h is the PI discrete solution on each element. ISil) the measure of the surface of K (resp. its gradient is a constant value on each element and it is not necessarily continuous. having reconstructed the gradient of the solution. c IKIWhIK) lSil (19) At this stage. Si). Hence.
1 <j 6 d: Remark 2. b) A leastsquare approximation of the Hessian matrix. This leads t o a usually overdetermined system of the form: A X = B . Let p be a vertex of the mesh Th and let B(p) be the ball of p .8. This approach is proved to be equivalent to a method in which we apply a dual L2 projection globally o n the space of Lagrange PI finite element functions with a treatment of the mass matrix by mass lumping. O”*(p) = ( y ). we suppose that C a r d ( B ( p ) )= n and n 2 3. and B is a vector of diyi) mension n given by the righthand side of the Relation (ZO). and function . this is usually true in two dimensions for an arbitrary unstructured mesh.. Without loss of generality. VU*(P) = (. Considering a Taylor expansion of u* at a vertex pi E B ( p ) and truncating it a t the second order leads t o the following relation: This relation can be also developed as follows: PPi = (.262 Pascal Frey where the matrix H * ( u ~ corresponds to: ) with for each component xj.) . with X T = ( a b c ) where A is a (n x 3) matrix function of (xi. the set of mesh vertices connected to p .) .
Physical phenomena can involve large scale variations (e. * ( p i ) ) . To solve this problem. capturing weak phenomena is crucial for abtaining an accurate solution and a good understanding of the behavior of the physical phenomenon.Relation (16) is normalized using the local absolute value of the variable u: . because they often ”hide” weak phenomena. as follows: ”Find X E EX3. A local error estimate can overcome this problem. i) I n some peculiar cases. However. for instance a Gauss pivoting algorithm. such that llAX . the system can be underdetermined (if C a r d ( B ( p ) ) < 3). This final system is solved using a classical method.5 An error estimate for CFD problems We now turn to a practical application of this concept of error estimate based on the interpolation error for numerical simulations in computational fluid mechanics (CFD).a local L2 projection as described before. Such accuracy requires taking into account all phenomena interactions in the main flow area. This system can be solved using a leastu square approximation.g. weak as well as strong shocks). multiscale phenomena. ii) this computation requires the knowledge of u* and of Vu* at the mesh vertices. several methods can be envisaged. for instance: . It is even harder to capture such phenomena when strong shocks are located in the flow. The latter consists in minimizing the distance between the vectors AX and B of W” by minimizing the square of the Euclidean norm of their difference. It is thus difficult to capture the weakest phenomena.the same approach but with truncating the Taylor expansion at the first order and solving a overdetermined linear system of two unknowns using a leastsquare approximation. even using mesh adaptation. yi. To evaluate Vu*.” It can be shown that the solution of this problem is the solution of the linear 3 x 3 system of normal equations [9]: Y EP3 A ~ A X = A ~ B . u*(p). Additional vertices connected to the vertices of B ( p ) can be taken into account to increase the number of equations. Following the ideas suggested by [6. 2.A Differential Geometry Approach to Mesh Generation 263 of ( a . recirculation.26]./3.B)I2. . we consider that u* and U h coincide at the mesh vertices.B1I2 = inf llAY . x i .9.2. Remark 2.
To this end. The term ~llull.. we will attempt to explain how the theoretical concepts of metric tensor and geometric error estimate introduced in the previous sections can be used to create meshes adapted to the simulation at hand.264 Pascal Frey where IuI. = max(lu/.y. This advocates adapting the size of the mesh elements as suggested in the previous section in order to equidistribute the interpolation error. thus dimensionless variables. .with 6 < ' 1 a constant. to increase the degree of the polynomials in the space 9. y is set to 1. in the context of anisotropic mesh adaptation for compressible flows. . Various techniques 271 have been proposed to overcome this problem (see [l. However. in order to define a single metric that accounts for all. other approaches suggest to adapt the approximation space and. capturing weak phenomena by means of Relation (21) leads to mesh the strong shock regions with isotropic elements. Here. To obtain a metric that takes into account all variables. These oscillations have a magnitude of the same order as weak phenomena. the error estimates involve the parameter h. Indeed. As mentioned previously. a large number of papers have been published presenting mesh adaptation techniques for numerical simulations based on finite element or finite volume methods. more precisely. we advocate to filter the oscillations with the local gradient of the solution in order to preserve the anisotropy. 7 = 1 . mesh adaptation allows to improve the numerical accuracy of the solution as well as the convergence rate. approach is in turn known as padaptation. the element size. This is made possible because the previous error estimation provides a relative error measure. On the other hand. we propose the following error estimate: (22) where h is the diameter ( i e . This is mainly due to the discretization of the solution that introduces virtual oscillations in the direction parallel to the shock. This In this section. Remark 2. the length of the longest edge) of element K and 0 < y < 1 is a parameter that is considered close to or equal to zero if strong shocks are involved in the flow. 3 Mesh adaptation using a geometric error estimate Over the last few years. we use the intersection procedure described in the previous Section. in convergence and accuracy results. This approach is usually known as hadaptation.n introduces a cutoff to avoid numerical problems.10. To achieve isotropic adaptation.cIIullm. for instance).n).
and let us consider B ( p ) as the open ball circumscribed to the simplex containing p .1 (Delaunay lemma). where H . . This gives a practical way of checking whether the triangulation T is a Delaunay triangulation or not. Now.Delaunay triangulations have been extensively studied.i) as K. Then. We know that z $! B(K.. especially in computational geometry [4].1 is adjacent to K.. the Delaunay criterion (the empty sphere criterion) holds. according to the assumption: q $l! B ( p ) H p $! B ( q ) . n. q ) the vertex opposite to this face in each simplex. x E H. Denoting by z the vertex of Ki not common with i Ki1. Lemma 3. The result follows: as z E B(zo). We consider the point g barycenter of KOand the line gz that intersects an ordered set of simplices Ki. This allows us to conclude. then this criterion holds globally and T is a Delaunay triangulation. represents the halfspace containing limited by the hyperplane support the common kface.1 The Delaunay triangulation In Section 1. (following Lawson 1311) At first.By we definition. we can notice that the criterion is symmetric. i = 1. adaptive Delaunay mesh generation Since the pioneering work of Delaunay [14]. we have defined the notion of Delaunay triangulation of the convex hull of a set of points S c Rd and Theorem 1. Proof. Let us consider a pair of adjacent simplices in EXd sharing a kface of dimension d .1. As z E Hi+1 and as B ( z i ) n H + c B(zi+l)n H + then i1 ii x E B(xi+l). 3. then z E B(z. If f o r each and every pair of adjacent simplices in T .1 gives a fundamental result about the existence and uniqueness of such a triangulation. What must be emphasized here is that a local property for any pair of two adjacent simplices leads to a global property for the whole triangulation. The following lemma introduces a local property of Delaunay triangulations. Indeed Delaunay triangulations are of great interest as they can serve to support efficient and versatile mesh generation methods [22]. according to the intersection point from g to z.1 and let us denote p (resp. Let T be a n arbitrary triangulation of the convex hull of a set of points S c Rd in general position.A Differential Geometry Approach to Mesh Generation 265 3. define Hi the halfspace containing Ki and not containing Ki1. it is easy to note that if q $! B ( p ) then B ( p ) n Hq C B ( q ) n Hq and B ( q ) n H p c B ( p ) n H p .1 Classical vs.i) which is in 0 contradiction with the assumption that z $l! B(zn1). Let suppose that the Delaunay criterion is true for each and any configuration of two adjacent simplices in the triangulation T and that a point x E T exists that is enclosed in the circumsphere of a simplex KOnot adjacent to any of the simplices having z as vertex.
Practically. Figure 7 Construction of the Delaunay triangulation using the Delaunay kernel (bold lines indicate the external edges of C(P)in two dimensions). we provide an incremental yet practical method for constructing Theorem 3. In this fundamental result. 3. C ( p ) stands for the cavity of point p : the set of simplices in Ti such that their circumspheres contain point p and B ( p ) denotes the ball of point p : the set of simplices formed by joining p to the external faces of C ( p ) (Figure 7). Let Ti be the Delaunay triangulation of the convex hull of a set of point S c Rd and let p $ S be a vertex enclosed in Ti. The Delaunay kernel procedure can be written as: Ti+l = Ti \ C ( P ) + B(P) 9 (23) and provides Ti+l. However.1 (Delaunay kernel). Proof: (can be found in [20]) based on the duality with the VoronoY diagram. the Delaunay triangulation of a set of points S E Rd is purely a triangulation of the convex hull of the set S.2 Constrained triangulation As mentioned previously. a Delaunay triangulation of the convex hull of S U { p } .1.266 a Delaunay triangulation. Pascal Frey Now. this result ensures that the cavity C ( p ) is a starshaped polytope with respect to point p . .
c) and K2 = ( a . c ) and K i = ( c .2. Proposition 3. Theorem 3. If the quadrilateral formed by the four vertices is convex. b ) sharing a common edge ab. the following result will provide the existence of a triangulation: Lemma 3.1. This is the case when the initial points all belong to the domain boundary. at least two specific problems arise. for instance) and ii) creating additional vertices in the triangulation that are not part of the initial set S. internal points must be inserted within the domain to enable numerical computations and provide information about the evolution of the solution within the domain. d d b Figure 8 Edge flip operation in two dimensions. a valid conforming triangulation Ti+1 having p as vertex can be created using the Delaunay kernel. an alternate configuration of adjacent triangles K . Notice that Ti+l is not a Delaunay triangulation either. Regarding the last requirement.b. Given a set of nonintersecting edge segments in E X ' . . = ( a .1. Proof: see [22] for a formal proof. d . it is always possible to generate a triangulation containing these segments as element edges.1 when the initial triangulation is not a Delaunay triangulation. Going from (K1.Kz) to ( K . This triangulation can be obtained from any arbitrary triangulation by using a edgeflip operator in a iterative manner. Proof: this is simply the extension of Theorem 3. 0 Let consider a pair of adjacent triangles K1 = ( a . d . 0 .b) exists and is a valid conforming triangulation of the set of points. Then. p not being a vertex of Ti.A Differential Geometry Approach to Mesh Generation 267 in numerical simulations. d . K i ) is the purpose of the edge flip operator (Figure 8). Let Ti be an arbitrary triangulation and let p E EXd be a point enclosed in Ti. related to the necessity of i) inserting specific entities in the triangulations (a given set of edges defining the domain boundary.
3. The circumcenter 0 can be computed as the solution of a linear system bf the form: Hence. 2. Although many numerical problem arise when implementing of the algorithm (especially in three dimensions).1.4 Creation and insertion of internal points Defining the cavity C ( p ) of a point p requires identifying all simplices having a circumsphere that contains point p .2. 3. The next section will present a method for dealing with the discretization of dR for R c R3. The overall scheme of this method can be sketched as follows. Let consider a closed bounded domain R c Rd and its boundary dR. Beware that the resulting triangulation Th may not be a Delaunay triangulation as entites have been enforced using the edge flip operation.268 3.1. an element K belongs to the cavity C ( p ) iif the following relation holds: . a twodimensional manifold embedded in R3. using Lemma 3. Here. the set S corresponds to the set of vertices of 1. creation and insertion of additional vertices pi in the triangulation using Theorem 3. O ) = IIOpiII. it involves the successive steps: 1. given a point p . generation of a Delaunay triangulation of the points of S using the Delaunay kernel and enforcement of the possibly missing entities among the required entities of the set &.1. we will attempt t o focuss here on a few issues that are of interest in the context of mesh adaptation for numerical simulations and especially on the creation of internal points. The radius of the sphere circumscribed to a simplex K corresponds to the distance between any vertex pi of K and the circumcenter 0 of K : r K = d ( p i . We suppose here that there exists a piecewise afine discretization & of dR composed of nonintersecting kfaces of dimension d.3 Pascal F'rey Delaunaybased meshing algorithm The previous results can be combined and used to develop a Delaunaybased mesh generation method. resulting in the triangulation Th of the domain.1. construction of a triangulation of the bounding box of & (for instance composed of only 2 triangles in R2).
Then. Proof: it consists in checking that the sodefined cavity is a starshaped poly0 gon with respect t o point P (see [ 2 2 ] ) . the set { K E Th .2. In mesh adaptation. Let us simply mention the following result in two dimensions: Proposition 3. p E K } of all simplices containing p .. hence a symmetric positive definite matrix M p at each point p E R". The extension t o the anisotropic case consists in introducing a metric tensor. as these two notions are strictly equivalent.A Differential Geometry Approach to Mesh Generation 269 a ( p . all distance checks involved in the Delaunay measure will be replaced by length checks according to the given matrix M p . K ) being called the Delaunay measure of point p with respect to the simplex K . namely: Obviously. provides a K) valid anisotropic Delaunay kernel. on which the approximate solution u h as been computed. Inserting a point p in a given (Delaunay) triangulation Th requires at first to identify all simplices that belong to the cavity C ( p ) . it is possible to enumerate the set C ( p ) (see [20] for more details on this procedure). . Although Theorem 3. Th' can is be achieved by finding the simplex or. In this sense. It is therefore highly recommended to consider all matrices related to the vertices of element K [23]. the computation of distances between vertices. in order to equidistribute the interpolation error.1 ensures that the cavity C ( p ) is starshaped polytope with respect to point p . T h e Delaunay measure. a ~ ~ ( p .< 1. This will allow us to consider the Euclidean norm of any vector in R" given the inner product (. given a . overcoming this problem is tedious and beyond the scope of this paper. Hence. a discrete metric field can be defined on Th using an a priori geometric error estimate as discussed in Section 2. According to Section 2. in some peculiar cases. this must be check and eventually enforced numerically. To answer this question in a practical manner. ) M . the metric or the matrix M is usually defined at the vertices of the current mesh Th. leads to solving a nonlinear system of equations. One question left is related to the decision of creating and eventually inserting a new point p or removing an existing vertex q in the triangulation Th. the geometric error estimate provides a symmetric positive definite matrix M at each and any mesh vertex so as to equidistribute the interpolation error over the mesh elements. By doing so however. we advocate a naive algorithm based on the edge length analysis or. taking into account only the metric tensor at the given point p is not sufficient and accurate in many applications. using adjacency relationships.
The result of this operation is to remove the edge and the set of its connected simplices.1. Among the many algorithms designed for surface mesh generation. 3.2. numerically speaking. To this end. provided the resulting triangulation remains valid. 3. as it is too conservative. then the edge is subdivided into subsegments of unit length. having a length much smaller than one.270 Pascal Frey userspecified tolerance value. we introduce the notion of a metric conforming edge: Definition 3. a linear approximation of f can be obtained by using Taylor formula for a point x close to a point a and by dropping secondorder terms: . a mesh edge ab is said to be matrix conforming iif: Once a new point p E K has been created and inserted in a triangulation Th.2 Surface mesh generation This is by far the most tedious problem among all mesh generation problems. Remember that the classical mesh adaptation scheme described previously assumed given a triangulation & of the boundary d o of a domain R c W d . the associated metric tensor Mp is obtained using an interpolation scheme based on the metric tensors of point p i E K (Section 1).1 Problem statement: the concept of geometric mesh Given a differentiable function f : R + R. are collapsed by merging the two endpoints into a single vertex. all mesh edges are analyzed and their lengths with respect to the metric tensor M are compared to one. If for a given edge. On the other hand.The boundary dR is then a twodimensional manifold embedded in R3. point insertion and point deletion operations cannot be based on the sole check Ilabll~ 1 on the = length of any edge ab. Given a matrix M . We have already stated that this problem is equivalent to constructing a unit mesh. short edges. The aim of this section is precisely to explain how to generate the triangulation & for a closed bounded domain R in R3. the length is much larger than one. we will focuss here on two issues: (i) the control of the geometric approximation for Cartesian surfaces and (ii) the adaptation of discrete piecewise linear surfaces. Hence. From a practical point of view.
D2fl(x)aa')I 9 2 9 x E K uCK 2 7 (24) Pha(Ico. . complete normed vector spaces). Given a surface C. f where of(. This smoothness property is related to the regularity of the surface approximation.K IILT  6 . we shall consider here an immersion 0 : R + R3 and perform the relevant analysis using the covariant basis. for which we can write: f ( x )= f ( a )+ o (a)(.max max (% IP"(4Iv) . this property is locally related to the measure of the gap between the normal to the element and the normal at any point of the surface discretized by the element.a ) . Proposition 3. a represents any vertex of element K and a' is the intersection point between the line a x and the edge of K opposite to vertex a. This corresponds indeed to an approximation property. it is usually sufficient to consider the notion of geometric continuity at the orders 0 and 1.PhC(lco. . .f ( a ) ) . We refer the reader to the chapter of this book dealing with differential geometry by P. Ciarlet and to his book [Ill for more details on this approach.v) H a(u. The interpolation error lla . Let R be a closed bounded domain of R2 and let consider the mapping6 a : R + R. For the sake of clarity. We adopt here a less general point of view.3 (Oorder axiom).A Differential Geometry Approach to Mesh Generation 271 It represents a tangent line approximation as the righthand side term matches the equation of the tangent line to the graph of f at point ( a . let us simply remark that these axioms allow to answer the following questions: (i) how close to the surface C is the triangulation Th? and (ii) how regular (smooth) is the approximation Th? Additional concerns may arise when dealing with affine elements. regarding local Taylor expansions. in view of numerical simulations for instance. In the context of surface meshing.Phall can be bounded by above as follows: IIa . a surface triangulation Th is said to be Uorder geometric if each element is close to the surface.)represents the Frkchet derivative o f f at point a . Formally speaking. This result can be extended to Banach spaces ( i e . mainly related to creating a minimal number of elements as well as creating wellshaped elements.x E K 1 (aa'. On the other hand.G. We have the following result: v). 6Actually. (25) where Pha denotes the linear interpolate of LT o n a n element K E Th. x = ( u . the 1order geometric property is satisfied when each linear element is close to the tangent plane at each point of the surface. It will ensure that the gap between the piecewise affine approximation and the underlying surface is bounded by above.K6 .
max I ( a a ’ . x is not located along the element edges.D2u(x tza)ax)dtl. D 2 u ( z + t x a ) a a ‘ ) d t ~ 0 As a is the site of x. norm of the interpolation error llu . ie.V(0  PhU)(Z)) + I’ (1. hence for an extremum of this function or a point x where the extremum is attained. then (u.Phu at site a of K .~hu)(x)I 6 . ~ ‘ u ( y aa’)~ ) or similarly: I(.P ~ ( T ) ( Z ) ~ . .Phu)(a)is equal to: (0.t ) X 2 ( a a ’ . thus: I(uPhu)(x))= IS’(l.D2a(z+tsa)ax)dt. ~ ~ a ( y ) a a ’ ) ~ . Now.Phu)(a)= 0 and at point x.Phu)(z)= 0. we have: (uPha)(x) =  I’ 0 (lt)(ax. X 2 < .272 Pascal Frey Proof: the sketch of the proof is based on classical Taylor expansions. let us consider the L .(I . V(O . We denote by a’ the intersection point between the line ax and the edge opposite to a in K . Then. it exists X E R such that ax = Xaa‘.and we find then the following bounds: 3 and finally: l(u . D2+ + t x a ) a+. This leads to solving the following equation: (v.~. We are looking for an upper bound on u .Phu. such that: V(u .PhfT)(Z)l= I /‘(I t)(ax.P h u ) ( ~ = 0 .PhO)(Z) + ( x u . (u. + (27) Let assume that x E K . with v) respect to any point x = (u. We develop the expression of u ...Phull.max yEaa’ < 2 9 I(aa’. E K . yEK 2 9 This result provides a global upper bound on the interpolation error on u at point z E K .  (26) . we have then: .t ) ( a x . We define the site a of point x E K as the element vertex closer to x than any other vertex of K . )) VV C K As the surface triangulation T h is supposed interpolant.
Obviously. 6 . we introduce the following bound on the interpolation error: 2 IIa .PhU at both vertices a and b with respect to any arbitrary point x along the edge ab.JD2a(w1.P ~ ~ ( I ~ . Notice that in Relation (as).wz)l 6 (w. along an edge of K . Let us denote 1D2cr1 the symmetric positive definite matrix: this allows to obtain the following inequality: Indeed. Considering Taylor expansions with an integral rest of a . Indeed. To overcome this problem. Being already symmetric. this last inequality is more coarse than Relation (29) as it does not involve points a and a' and hence is now independent of any point x E K .D2a(w1. looking at the eigenbasis of the matrix ID2al and denoting by (211.v2)l w ) w). the same conclusion holds when considering vertex b or vertex c of K in the analysis. thus meaning that the Hessian matrix must be converted into a symmetric positive definite matrix. the following relation holds: ldld + d2w221 6 Id1v11 + ldzv221. it comes by summation: . To be complete. where R corresponds to the matrix of the eigenvectors and the di are the eigenvalue. thus making the evaluation of the interpolation error rather tedious. from which we can conclude that: I(w. hence the righthand side term can be written as follows: Relation (29) involves an extremum value that is obviously not known a p r i ori. ID2a(y)I is a symmetric positive definite matrix. the Hessian matrix can be decomposed as follows: D 2a = R (2 :2) Rl. wz) the components of any twodimensional vector v.A Differential Geometry Approach t o Mesh Generation 273 We are willing to introduce a metric tensor in the previous inequality.max max(w. it remains to consider the case where the extremum x is not attained within the element K but on its boundary. ID'~(Y)I K 9 YEK vCK w) .
The second. Finally. it is easy to o show that l . following the same analysis. Hence. based on the intrinsic properties of the surface (i.. ~ ~ o ~ ( y ) a b ) l . A Taylor expansion at the second order at a point (u.max(ab. we suggest here two approaches to create an optimal mesh..1] such that ax = Xab and bx = (1.0) 1 + uuu + a v v + 2 (auuu2 2uuvuw+ uvvw2) o(u2 + 2 ).2 74 Pascal Frey where x1 = z + t xa and 2 2 = x t xb. 8 YEAB where ID20(y)I is defined as previously. based on the local deformation.(A’ 4 + (1  A ) ~ ) max ~ ( a b .v) = a(0. I~’o(y)IabI). l ~ ~ o ( y ) ( a b ) . K max(ab. Let us consider a vertex p as the origin of the local frame p = (O. 0 As stated before.~ h o ) ( x ) ( ~ . allows us to account for isotropic adaptation. This deviation can be measured by considering the Hessian along the normal to the surface at this vertex. yEAB As the polynomial P(X) = X2 + (1I(U . To this end.PN)(~)I has a maximum at A = .e. a(0.b] then 3X E [0. we will then keep and retain Relation (29) as the upper bound of the interpolation error.Pho)(x)Iis bounded by above as: + = 1 .2 Local deformation of a surface The main idea consists in characterizing the deviation of the surface mesh Th in the vicinity of a vertex p with respect to the tangent plane Tp to the surface at this vertex.O)). the problem of generating a geometric mesh Th of a domain R consists in defining an optimal mesh which would be as smooth as possible. and a.2.O. v) in the vicinity of vertex p can be written as follows: a ( u . The first. hence the second fundamental form of the surface. Relation (28).X)ba. As z E [a.  < 1 8 YEAB As f < $. this analysis leads to work out the following bound for the interpolation error: l(o . we have: 2 1 < 1 . represent the first and second partial derivatives of u with respect to the variable u. Given n o the unit normal vector to the . 3. the principal curvatures and directions) allows us to deal with anisotropic mesh generation. + +  where u.
Let consider the local frame F = ( p . e l . w) and Tpthe tangent plane at point p can be expressed as follows: &(PI = (720. we suggest the following definition. 0. w I o ( % w) . + Definition 3. Hence.we come easily to the relation: which is proportional to the second fundamental form of the surface for u2 w2 sufficiently small. o(u. In order to have an anisotropic control of the mesh elements. we will now introduce another approach based on the intrinsic properties of the surface.P P ~ ) VPi E B(P). e2) in Tprthen h ( p ) . a E R. Hence.3 Local curvature of a surface Let consider a vertex p = (u.) and if we introduce the vector e g = (0. then the distance E ( P ) between the point u. a ( O . 4 % .u. the optimal element in the mesh Th of dR must have a local size proportional to the inverse of the square root of ~ ( p ) : where E represents the maximal desired tolerance value and k is the average size of the elements in the vicinity of p .w)) on an arbitrary surface triangulaw. tion T h and let p 1 . As mentioned.) and (0.0. the element size h ( p ) . p 2 denote the minimal and maximal radii of curvature at p and e l .2.0 ( 0 . (30) This deformation measures the maximal distance between the vertices pi adjacent to p and the tangent plane Tp at vertex p . O . 3. this measure is only suitable for isotropic mesh adaptation as Relation (30) corresponds to a scalar value. w. l). a(u.2. a. e2 the corresponding unit vectors along the principal directions of curvature. a. The discrete local deformation of the surface at a vertex n p a a mesh Th is defined as: E ( P ) = max(np. Erom the given upper bound Relation (29).(0. we can assume that the local deformation at each vertex p is proportional to the square of the element size in the vicinity of p . 0 ) ) ) . (u.must be locally proportional to the minimal radius of curvature to guarantee that the local deviation remains bounded by above: h ( p ) = ap1. defined in the tangent plane T p . As no is orthogonal to both the vectors (1.A Differential Geometry Approach to Mesh Generation 275 surface at point p . As pointed out. O ) ) ) = (no1 (u.1.
j ) of the projection of h in the plane of 0. then the size specification in the usual frame (z. when the surface is not analytically defined but is only known given a triangulation Th. Projecting this circle onto the plane of 0. Relation (31) must be rewritten as: with a and r] two coefficients related to the minimal and maximal radii of curvature. j ) is given by the equation: where P = (w1 w2) and hi denotes the coordinates in ( z .4 Evaluation of the intrinsic properties of a discrete surface The expressions obtained so far involve computing the principal curvatures and the principal directions of curvature. . To this end. This equation corresponds to a circle of unit radius centered at point p in Tp.276 Pascal Frey can be written as a linear combination of el and e2 in F hlel Then. allow us to obtain the size specification at an arbitrary point x E Q. We are then looking for a numerical scheme to evaluate the intrinsic properties of this discrete surface. the size specification can be formulated as follows: + h2e2. the problem of estimating the second derivatives becomes accute. However. This equation defines an anisotropic metric tensor at point x E R. In such a case. we will now briefly review some of the methods that have been proposed to deal with this numerical evaluation problem. One could also consider to define the metric tensor so as to obtain anisotropic triangles on the surface. This equation corresponds to an ellipse in the tangent plane Tp that fully encloses the circle defined by Equation (31). z . If 211 and 212 stand for the two orthogonal projections of el and e2. i. 3.2. to prescribe a unit mesh. Relation (29) contains the Hessian of the function IJ that is supposed to be known analytically.
This equation can be reformulated by taking into account the equations of three circles passing through vertex p and two of its adjacent neighbors as: r. More precisely. a is the S 1 i angle between two successive edges ppi. K d S = 27r i=O ai .2) .. a unit normal vector np is computed as the average value of the adjacent face normals nKi. let us first notice that we have the following equation for the normal curvature: 1 1 ~. given a surface triangulation Th: n1 /Js. Indeed. Let consider a vertex p and its ball B ( p ) . (I 1c. + 2 2 where 8 represent the angle between any arbitrary direction u in the tangent plane Tpa t vertex p and the principal direction that corresponds t o 1c.( K I . b and c can be determined by solving the linear system corresponding t o the circle equations using a least square fit... c) A local quadric surface fitting.~)(c0s28cos2a sin28sin2a). Ki E B ( p ) : . = . The GaussBonnet theorem leads to the following equation for the gaussian curvature K. where I represents the sum of the areas of triangles Ki E B ( p ) .K c . a t each vertex p E Th. Assuming K is constant in a small neighborhood leads t o write: n1 27r  K=3 s i=o c ai . Vpi E B ( p ) and n = Card(B(p)).A Differential Geometry Approach to Mesh Generation 277 a) A local circle fitting.1 and a is the angle between the tangent direction corresponding to the normal curvature r. Finally. this approach is based on a least square approximation of a quadric surface. the principal curvatures and directions are given by the following set of equations: + IQ =a f Jb2fc2 and 8= 2 b) The GaussBonnet relation. This approach is based on the Euler and Meusnier theorems for finding an estimation of the principal curvatures [7]. and the direction u. where coefficients a .1c.4). Similar to the evaluation of the Hessian matrix (Section 2. with n = Card(B(p)) .2.=abcos2a+csin2a.
the principal curvatures of the surface are supposed to coincide with the principal curvatures of the paraboloid. N = 2c. in turn. G = 1 (2(bu CV))’ = 1. leads to solving the linear system: 2x3y . the main curvatures requires finding the roots of ti2 .K. F = ~ ( U U bv)(bu CW) = 0 . The coefficients of the first and second fundamental form of the quadric surface are such that: I(v)= Edu2 + 2 F d u d v + G d v 2 and II(w) = L d u 2 +2Mdudv +N d v 2 .2 78 Pascal Frey following an idea suggested by Hamann 1251. Solving this system is equivalent to solving the minimization problem: n i= 1 which.zY2 2 x 3 y 4 x 2 y 2 2xy3 ( x r i 2 2xy3 y4 ) (I) b = )I:( 2xyz . the numerical scheme approximates a small neighborhood of the mesh around vertex p by an osculating paraboloid (the quadric fitting). + . Considering that all pi E B ( p ) belong to the quadric surface leads to writing a usually overdetermined system of the form A X = B.O. r and 1 72 being two orthogonal vectors of the tangent plane Tp. Given the parameterization of the paraboloid.O): E = 1 ( 2 ( a u bv))2 = 1. Vertex p is transformed v). r 1 . At a regular surface point ( x . Locally. where a. where v = A71 +pr2 represents any vector of the tangent plane Tp. Z) = ax2 + 2bzy + cy2 . Finally. r 2 . n p ) such that np coincide with the z axes.% H = 0 . + + + + + + L = 2a.z = 0 . the Gaussian and the mean curvatures are computed as: K=4acb2 Computing ti1 < the polynomials: ti2 and H=a+c. M = 2b. c are coefficients to be numerically computed as the solution of a linear system of equations. z ) we can write the canonical equation of the surface: f ( z . is then easy to compute these coefficients and we found at p = (0.y .y . This C2 continuous quadric surface centered at vertex p is represented in the explicit form z = z(u. 0) of the local frame 3 = ( p . b. along with its adjacent vertices to the origin p = (O.
We assume that at any point p of the twodimensional parametric space Q = [4. This form allows us to prescribe the size and the stretching of the mesh elements tangentially to the algebraic curves of equations: r : x3  y2 +2  3xZ = 0.1)3 0.if X 2 1. ~  and R is the rotation matrix of angle 6 given by the formula: 6 = arctan 2Y 3(x2 . the anisotropic metric is represented by the matrix M p defined as follows: thus prescribing the sizes hl and h2 given by the set of equations: hl(X. min(0.2. we consider a twodimensional analytical example. X being considered here as the parameter. 4 Applications Finally. to help decide whether the point p is elliptic.2(1 x ) + 0. if x < 1. 0. we would like to illustrate the potential of a mesh generation approach based on the differential geometry concepts we have introduced in the previous sections on various examples covering analytic (academic) as well as realistic applications.005. Several heuristic methods have been proposed to provide information about the local neighborhood of a surface point p and. l). l).2(X .1 An academic example As a first example. if X < 1. in particular. parabolic or hyperbolic.z(X . min(0.A) ' .2.4] is given a symmetric positive definite quadratic form represented by a 2 x 2 matrix M p .4] x [4.Y) = h2(X.Y) = { { min(0.A Differential Geometry Approach to Mesh Generation 279 Once the two principal curvatures and radii of curvature have been computed. I).2(1  + + min(O. 4. At any point p E R. I). Euler's theorem will give access to the principal directions of curvature [all.q3+ 0. if X 2 1.05. We consider the previous approaches sufficiently accurate in the context of numerical simulations.
Although the quadratic form is defined analytically at each and any point of the domain R. Figure 9 Anisotropic mesh adaptation for an analytic metric specified on a twodimensional domain R = [4. we have used an interpolation scheme (cf. we deliberately used it in a discrete way: the only data supplied to the mesh generation algorithm was the matrix M p defined at the mesh vertices. the metric field can be considered as a discrete field. in this case. . In other words. Section 1. The only difference with a numerical simulation is that.4].280 Pascal Frey Figures 9 and 10 present the initial and adapted meshes generated using an adaptation scheme based on the given metric. From top to bottom and from left to right: adapted meshes at iterations 0 (initial). the exact and numerical solutions u and U h coincide exactly at all mesh vertices.5) to compute the matrix at any vertex location when inserting a new vertex in the existing mesh.4] x [4. In this sense.2. 1 and 6 and corresponding isovalue curves of the maximal eigenvalue of the matrix M p .
4. Introducing the Hamiltonian + H ( x . let suppose 70 = $3'0) is a starting curve in R2 and let C(t). t .0) = yo given.t). z E R" be defined by: t dimensional function u. < u(z. = O ) . More precisely.O < t T denotes a family of closed C2 curves embedded in W2 parameterized by y(z. Let assume that this family of curves is generated by moving 70 along its normal vector field with a speed F related to the local curvature: y(x.n. Enlargements of the adapted mesh at the final iteration. We can embed the interface curve as the zero level set of a higher thus let u(z. To fix the ideas. Derivating this equation permits to find a partial differential equation from the evolution of u: u~(z.t = 0 ) = k d where d represents the signed distance to the interface. let us briefly introduce the basic concepts of level sets functions. t ) = F(lc). u(z. Vu * y z t ) = 0 .2 Curvaturedriven evolution flows Now we will focuss on a physical dynamic problem in two dimensions.n being the unit normal vector. Following Sethian's work 1361. VU) = F ( K ) / V U / . let consider for instance a multi physics problem involving two fluids in parallel motion with two different velocities and densities.A Differential Geometry Approach to Mesh Generation 281 Figure 10 Anisotropic mesh adaptation for an analytic metric specified on a twodimensional domain R = [4. To this end. yielding an unstable interface for all speeds. the curve 70. known as the level set equation.4]. t) (. we can consider the general motion of an interface between two media in the normal direction under an external velocity field F ( K )related to the curvature of the interface.4] x [4.
. (33) A correct definition of the solution of such equation for which we have existence] unicity and stability results. the weak solutions are in W1@ and the viscosity solution represents the physical (good) solution among the weak solutions. Lemma 4 1 A classical solution of Equation (33) is a viscosity solution. vx $ ) ! u ( x . t o ) E Rx]O.1. Therefore] the level set equation for motion by mean curvature .T[ where u2. The interface y(xlt ) is evolved in time by an external velocity field V(xl t ) according to the simple advection equation: Ut(Z. . n1 Describing the interface in this way makes possible the natural change of topology of the interface.t ) E RxIO. V U ) 0 = ( 2 . t ) + v Vu(z. u(2. t ) < 0 . and a continuously diflerentiable viscosity solution is a classical solution.32]for first order HamiltonJacobi equations.T[+R is a viscosity solution of Equation (33) i f u is a sub. m i n i m u m ) then: The following lemma shows that the definition of viscosity solution is consistent with that of classical solutions. In order to deal with topological changes to the dynamically evolving interface.(resp. For such equation. T[ . ' Let consider here the case where the normal velocity is the mean curvature 6. the problem is to embed the interface of an open region R as the level set of a smooth (at least Lipschitz continuous) higher dimensional function u(xl t ) . vx E 8 0 . at any point ( 2 0 . Definition 4. Let us briefly introduce this powerful concept.t ) = 0 .282 Pascal Frey leads to consider the socalled parabolic HamiltonJacobi equation: Ut(Z. is given by the notion of viscosity solution. Qz E R. has a local m a x i m u m (resp. The standard theory of viscosity solution has been introducing by Crandall and Lions [12. t ) = 0 . A function u : Rx]O. super) solution: for all test function v E C1. t ) + H(x1t . The level set function u is closely related to the definition of an implicit function and enjoys the following properties: { u ( z 1 t > 0.
mean curvature flow is a geometric initial value problem. mesh adaptation seems a natural way of dealing with the accuracy problem as well as with capturing the interface shape. the solution y ( t ) evolves in time so that at each point its normal velocity vector is equal t o its mean curvature vector. where t represents the fictitious time. Sethian [36] and Osher [33]. the level function can be chosen as a Lipschitz continuous function. This operation can be achieved by solving to steady state (for infinite fictious time) the equation [38]: ut + sgn(uO)(lVul. represents the evolving surfaces as the level set of an auxiliary function solving an appropriate nonlinear differential equation. In other words. Unfortunately as Equation (34) is nonlinear it would not be straightforward to implement a n implicit method. small timesteps are required to maintain stability. to improve mass conservation and to reduce errors in the computations of geometrical quantities such as the interface curvatures. among others. the mesh must be refined near the interface while coarsening in regions farther away. But. with explicit methods. the level set function u must be reinitialized to be the signed distance to the interface. . As the interface may change or deform in time.A Differential Geometry Approach to Mesh Generation is given by the following equation?: 283 when noticing that the normal and the mean curvature can be written in terms of u as follows: n=. this reinitialisation process must be carried out as often as each time step. However. On the other hand. For surfaces of codimension one. 'In the classical setup. Depending on the numerical scheme used to solve Equation (33). the socalled signed distance function is known to produce more robust numerical results.vu lVul and K = div (5) . In principle. this equation is stiff and thus tedious to solve numerically. The conventional manner of solving this type of equation would be using an implicit scheme. Starting from a smooth initial surface 70 E Rd.1) = 0 . a completely different approach. partly because of numerical diffusion. However. This levelset approach has been extensively analytically and numerically developed since the seminal works of Dervieux [15]. a t least for linear equations. a semiimplicit method for computing mean curvature flows which is stable for large time steps has been suggested by Smereka [37]. To overcome these problems. Moreover.
. The error estimate described Section 2 has been used to prescribe the anisotropic metric tensor at each mesh vertex.y) = (y3. the solution at tn+' is computed on a refined grid with a maximal accuracy. This will ensure that the numerical solution is not lost or diffused when solving the advection problem. This allows to define a metric tensor at each point of the domain R. Here. respectively. 9 adaptations have been performed. To illustrate the level set method. rotating under the divergence free velocity field: V ( x . . on a domain R = [0.x3) E R2 (Figure 12).1]x [0. We solve the pseudoadvection problem using PI Lagrange finite elements and the method of characteristics that reduces the partial differential equation to an ordinary differential equation.1] and an analytical curve 7 0 as represented Figure 11. we consider a domain R = [0. The density variable has been chosen to adapt the meshes with the following parameters: E = 0. we impose a local size to be proportional to the radius of curvature at the interface and a smooth gradation while moving away from the interface.1] x [0. x) E W2. hmin = 0. the eigenvectors being colinear to the normal and tangent vectors to the curve. 4. the size in the tangent direction is set as a coefficient times the minimal size in the normal direction. involving highly anisotropic phenomena (shocks). the beginning and at the end of the at time stepping.284 Pascal Frey the local mesh size is directly related to the local curvature of the zero levelset curve 7 0 .1] and a curve yo evolving under the divergence free velocity field: V ( x .02. Thus.y) = (y. Numerically.01 m and h. that takes both metrics into account for all points in the domain. Mesh refinement is achieved using the metric intersection procedure described Section 2. The geometry of the computational domain is shown in Figure 13. the aim is to capture the behavior of the physical phenomenon and to emphasize the reduction of the number of degrees of freedom obtained thanks to the anisotropy. each 400 time steps of the Euler solver. it can be seen that two regions are refined corresponding to the initial tn and final position tnfl of the curve 70.3 A CFD example in two dimensions The next example is related to a Euler computation at Mach 3 in a scramjet configuration and can be considered as typical of numerical simulations in compressible fluids. On this Figure. We provide another example of an evolution flow. In order to achieve anisotropy. = 2 m.
Bottom: evolution of the levelset curve u = 0 for two consecutive time steps. 285 Figure 11 Curvaturebased mesh generation for a levelset approach. Top: anisotropic mesh adapted to the local curvature for two consecutive time steps(1eft). .A Differential Geometry Approach to Mesh Generation I . Middle: zoom on the curvature adaptated meshes.levelset curves of the distance function at time t (right).
Left: anisotropic meshes adapted to the local curvature for two consecutive time steps tn and tn+1.286 Pascal Prey Figure 12 Curvaturedriven mesh generation for a levelset approach. right: corresponding levelset curves of the distance function at times tn and tn+1. .
The aim was to combine the curvaturebased metric defined using a leastsquares approximation of a quadric surface (Section 3.5 Mesh adaptation in three dimensions Finally.0 . The expected result is to obtain mesh refinement in highly curved regions as well as in regions where the gradient of the SOlution is important. allowing us to define a discrete metric field over the mesh. The first case consists in adapting the mesh to the metric field prescribed by the geometric error estimate. Figure 14. In addition. 4. However. lI3 as: J f~(x. the exact solution and the numerical solution coincide at the mesh vertices. 4. the related metric is obtained using the interpolation scheme described Section 2. we would like to show an example of anisotropic mesh adaptation in three dimensions.2.A Differential Geometry Approach to Mesh Generation 287 Figure 13 Computational Fluid Dynamics: Euler computation at Mach 3 in a scramjet configuration. z ) = tanh ((x + 1.4 Curvaturebased surface mesh adaptation Let us consider an analytical solution corresponding to a Cartesian hypersurface in JR4 defined on the domain C = [1. Figure 15 clearly shows that the objective has been reached using the metric intersection procedure. 3 ) ’ ~ ) . The results. More details as well as numerical evaluation of this approach can be found in Reference [2]. when introducing new mesh vertices. A Euler .3)20(y . Initial and adapted meshes at the iterations 2 and 9 and corresponding isodensity distributions. Figure 15. This example concerns a classic numerical simulation of transonic air flow around the ONERA M6 wing. a second test has been performed with the same analytically defined solution on a supertoroid geometry. allow us to check the efficiency of Hessian based mesh adaptation. y. Here again.4) with the metric provided by the error estimate.
bottom: corresponding isovalues of the analytical solution. Top: initial isotropic mesh and final anisotropic mesh.288 Pascal Frey Figure 14 Example of anisotropic surface mesh adaptation. . Figure 15 Another example of anisotropic surface mesh adaptation: initial geometrybased mesh and final anisotropic mesh.
This transonic simulation case gives raise to a wellknown lambdashock. 5. The final mesh contains 23. Reasons explaining this loss of stretching can be found in [27]. Figure 16 shows the anisotropic adaptation. every 250 time steps. Figure 16 Onera M6 Wing test case for CFD computations. The mesh has been adapted 9 times.06 degrees.516 and 132.848 boundary triangles and 37.A Differential Geometry Approach to Mesh Generation 289 solution is computed for Mach number equal to 0.8395 with an angle of attack of 3. The initial mesh is a relatively coarse mesh containing 7. In this example. The variable used to adapt the mesh is the Mach number.676 tetrahedra.815 vertices. the maximal aspect ratio achieved for the anisotropic elements is about 10. .922 tetrahedra. Initial and final anisotropic meshes (iterations 0 and 9) and cut through the volume mesh. thus revealing the difficulty of achieving strong anisotropy numerically.
Yvinec. Paris) for providing me with the example of curvaturebased meshes for levelsets methods. to appear in Journal of Comput. Journal of Differential Geometry. Rocquencourt) for their valuable discussions and suggestions about the topics covered in this chapter. in particular J. Estimateur derreur g6om6trique et m6triques anisotropes pour l’adaptation de maillage. P.L.D. Alauzet. Sonner. Several application examples have been described to emphasize the efficiency of this approach with regard to the numerical solution as well as to the computational cost. Fudan University. Mohammadi. they must all be thanked for their kindness and assiduity. of Mathematics. Other people shall also be mentioned here to which i am indebted. [3] L. Paris) and F. Ciarlet (dept. Level set approach t o mean curvature flow in arbitrary codimension. as well as to Professor Li TaTsien (dept. Shanghai) for providing a more than friendly and convenient environment to the lecturers. INRIA (2003). Saiac (CNAM. Unstructured 3D mesh adaptation for time dependent problems. Jambu for their invitation in participating in the CIMPA schoolISFMA symposium on differential geometry: theory and applications. of Mathematics. Alauzet et P. we have presented a differential geometry approach to mesh generation and mesh adaptation. 43. In the context of numerical simulations. Application t o CFD simulations. Cambridge University Press (1997). Research report. Acknowledgments I am greatly indebted to Professors P. This approach is based on the definition of a tensor metric at each vertex of a given triangulation Th in Rd. Boissonnat and M. Algorithmic geometry. P. Frey. Ducrot (LJLL. RR4789. [a] F.290 Pascal Frey 5 Conclusions In this chapter. More details about mesh generation and mesh adaptation can be found in reference [20]. Frey. City University of Hong Kong) and M. George and B. 693737 (1996).H. [4] J. The symmetric positive definite matrix associated with a quadratic form is used to prescribe the element size as well as the anisotropic stretching of the mesh elements. . this matrix is defined based on a geometric error estimate related to the Hessian of the numerical solution.G. Partie 11: exemples d’applications. as well as V. Physics (2006). Over 50 participants of more than ten different countries have attended the summer school. References [l] F. Ambrosio and M. Alauzet (INRIA.
Palermo. [20] P. Nat. USA. 142 (1983). [22] P. Sci. secondorder partial differential equations. Thomasset. George and H. Int. Rendic. [15] A. Bull. Kimmel and G. Lions. Trans.J. [16] M. [11] P. Basic Error Estimates for Elliptic Problems. Premikres exphiences de maillage automatique par une me'thode de Delaunay anisotrope e n trois dimensions. Amer. B. Prentice Hall (1976). Delaunay triangulation and meshing. Amsterdam (1978). M. 22: 174 (1906). 17352 (1991). SOC. Ciarlet.G.L. Malcolm and C.J. P. Delaunay. Revue europe'nne d'e'lhents finis. SOC. George. PrenticeHall71977). Viscosity solutions of HamiltonJacobi Equations. SpringerVerlag (1984). Chen and F. . North Holland. Geodesic active contours. Bul. [7] X.A Differential Geometry Approach to Mesh Generation 291 [5] V. Applications to Finite Elements. A n introduction to Differential Geometry with Applications to Elasticity. Ishii. J . 1: 167 (1992). 9: 467486 (2000)..L. [23] P. Paris (1997).P.G. [12] M. Etude numerique d'estimations d'erreur a posteriori. 25: 475491 (1997). Anisotropic Unstructured Mesh Adaptation for Flow Simulations. Borouchaki... Numerical Methods for Nonlinear Problems . Intrinsic surface properties from surface triangulation. Technical Report INRIA RT0272 (2002). Math. Ciarlet. Caselles. in Handbook of Numerical Analysis. DoCarmo. [21] P. FrBchet. Forsythe.L.G. [14] B. Computer methods for mathematical computations. [6] M. Math. Cambridge University Press (1988). New Orleans. Amer. About surface remeshing. of the 9th Int. [9] P. Frey. Int. Lecture Notes in Physics. [lo] P. 277. Fortin. Meshing Roundtable. Hermks. Ciarlet. 0. 158163 (1981). Paris. F. Mat. [17] G. Pironneau. Circ. Hermks Science. Lions. Crandall and P. Finite Element methods (Part l). Schmitt. URSS. Acad. Mohammadi. [19] M. Springer (2005). Meth. George and H. The Finite Element Method for Elliptic Problems. Proc. Multifluid incompressible flows by a finite element method. Crandall. R.G.L. Erey and P.N. of Compuer Vision 22(1): 6179 (1997).G. 739743 (1992). Hecht. H. Lions Eds. 11. Glowinski. Sur quelques points du calcul fonctionnel. [8] P. Class. Proc. on Computer Vision. Sapiro. [18] M. Users guide to viscosity solutions of 27.L. vol 11. NorthHolland. Borouchaki.G. Moler. CastroDiaz. Dervieux and F. Fluids. LO.G.J.L. [24] R. Sci. Ciarlet. Oxford (2001). Numer. Ciarlet and J. Introduction to Numerical Linear Algebra and Optimization. 123136 (2002). Mesh generation: application to finite elements. J. Sur la sphere vide. 2nd European Conf. 793800 (1934). Differential geometry of curves and surfaces. [13] M.
Pironneau. [26] F. Wiley (1989). J. 3: 231246 (1986). Quelques mdthodes de rbolution des problbmes aux limites non line'aires. Hecht. Lagiie and F. A IAA paper. J . Hagen.. Paris (1983). Hamann.292 Pascal Fkey [25] B. 15th Int. and S. Computer Aided Design. Comp. San Diego. Hecht F and B.F. CA (2005). [38] M. Noltmeier and Knodel eds. Meshing Roundtable. Computing Suppl. Curvature and the Evolution of Fronts. 19: 13. Sethian. L. Properties of ndimensional triangulations. [31] C. Comput. Fronts propagating with curvature depending speed. . Smereka. A few snags in mesh adaptation loops. Hecht.L. in Geometric Modelling. Springer Verlag (1978).M. Cambridge University Press (1987). Optimal mesh for PI interpolation in H 1 seminorm. Introduction h l'analyse nume'rique des e'quations a m ddriue'es partielles. Mohammadi.. [33] S. Osher. Johnson. 8. Paris (1969). 970859 (1997). Osher and J. The finite element method for fluids. Meshing Roundtable. [28] C.A. 14th Int. NY (1993). Lawson. [30] J. [29] W. 439456 (2003). Communications of Mathematical Physics. Smereka.. Sussman. 114: 146159 (1994). Dunod. AL (2006). Raviart and J.A. 101:4 (1985). Farin. Journal of Scientific Computing. [34] 0. A Course in Differential Geometry. 791249. Masson. Numerical Solutions of Partial Differential Equations b y the Finite Element Method. Lions. Sethian. Phys. [35] P. [36] J. [37] P. (1988). Klingenberg. Semiimplicit level set method for curvature and surface diffusion motion. [32] J. P. A level set approach for computing solutions to incompressible twophase How. [27] F. Proc. Birmingham. Mesh adaptation by metric control for multiscale phenomena and turbulence. Curvature approximation for triangulated surfaces. Springer. Thomas. Proc. Phys.
The book presents some important applications to shells.corn.com . covariant derivatives.cn 6598 hc ISBNI3 9789812771469 World Scientific www.hep. and the fundamental theorem of surface theory in a selfcontained and accessible manner. thanks to the manifold applications where it plays an essential role. the fundamental forms of a surface. This volume will be very useful to graduate stydents and researchers in pure and applied mathematics. such as the theory of linearly and nonlinearly elastic shells. the Riemann curvature tensor. and mesh generation in finite element methods.worldscientific. such as the metric tensor. the implementation of numerical methods for shells. Although the field is often considered a “classical” one.Theory and Applications Series in Contemporary Applied Mathematics 0if f e re nt ia I G e 0 1e t ry : CAM 9 This book gives the basic notions of differential geometry. Higher Education Press www. it has recently been rejuvenated.