Variational Methods For Crystalline Microstructure - Analysis and Computation - G. Dolzmann (Springer, 2003) WW

Lecture Notes in Mathematics
Editors:
J.M. Morel, Cachan
F. Takens, Groningen
B. Teissier, Paris
1803
3
Berlin
Heidelberg
New York
Hong Kong
London
Milan
Paris
Tokyo
Georg Dolzmann
Variational Methods for

Crystalline Microstructure Analysis and Computation
13
Author
Georg Dolzmann
Department of Mathematics
University of Maryland
College Park
MD 20742
Maryland, USA
e-mail: dolzmann@math.umd.edu
http://www.math.umd.edu/dolzmann/
Cataloging-in-Publication Data applied for

A catalog record for this book is available from the Library of Congress.
Bibliographic information published by Die Deutsche Bibliothek
Die Deutsche Bibliothek lists this publication in the Deutsche Nationalbibliografie;
detailed bibliographic data is available in the Internet at http://dnb.ddb.de
Mathematics Subject Classication (2000): 74B20, 74G15, 74G65, 74N15, 65M60

ISSN 0075-8434
ISBN 3-540-00114-X Springer-Verlag Berlin Heidelberg New York
This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specif ically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting,
reproduction on microf ilm or in any other way, and storage in data banks. Duplication of this publication
or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965,
in its current version, and permission for use must always be obtained from Springer-Verlag. Violations are
liable for prosecution under the German Copyright Law.
Springer-Verlag Berlin Heidelberg New York a member of BertelsmannSpringer
Science + Business Media GmbH
http://www.springer.de
Springer-Verlag Berlin Heidelberg 2003
Printed in Germany
The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply,
even in the absence of a specif ic statement, that such names are exempt from the relevant protective laws
and regulations and therefore free for general use.
Typesetting: Camera-ready TEX output by the author
SPIN: 10899540
41/3142/ du - 543210 - Printed on acid-free paper
Preface
The mathematical modeling of microstructures in solids is a fascinating topic

that combines ideas from dierent elds such as analysis, numerical simulation, and materials science. Beginning in the 80s, variational methods have
been playing a prominent r
ole in modern theories for microstructures, and
surprising developments in the calculus of variations were stimulated by questions arising in this context.
This text grew out of my Habilitationsschrift at the University of Leizpig,
and would not have been possible without the constant support and encouragement of all my friends during the past years. In particular I would like
to thank S. M
uller for having given me the privilege of being a member of
his group during my years in Leipzig in which the bulk of the work was
completed.
Finally, the nancial support through the Max Planck Institute for Mathematics in the Sciences, Leipzig, my home institution, the University of Maryland at College Park, and the NSF through grant DMS0104118 is gratefully
acknowledged.
College Park, August 2002
Georg Dolzmann
Contents
1.
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.1 Martensitic Transformations and Quasiconvex Hulls . . . . . . . .
1.2 Outline of the Text . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1
3
8
2.
Semiconvex Hulls of Compact Sets . . . . . . . . . . . . . . . . . . . . . . .

2.1 The Eight Point Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Sets Invariant Under SO(2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3 The Thin Film Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.4 An Optimal Taylor Bound . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.5 Dimensional Reduction in Three Dimensions . . . . . . . . . . . . . . .
2.6 The Two-well Problem in Three Dimensions . . . . . . . . . . . . . . .
2.7 Wells Dened by Singular Values . . . . . . . . . . . . . . . . . . . . . . . . .
11
13
26
49
51
53
55
57
3.
Macroscopic Energy for Nematic Elastomers . . . . . . . . . . . . . .

3.1 Nematic Elastomers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 The General Relaxation Result . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3 An Upper Bound for the Relaxed Energy . . . . . . . . . . . . . . . . . .
3.4 The Polyconvex Envelope of the Energy . . . . . . . . . . . . . . . . . . .
3.5 The Quasiconvex Envelope of the Energy . . . . . . . . . . . . . . . . . .
69
70
72
75
77
80
4.
Uniqueness and Stability of Microstructure . . . . . . . . . . . . . . .

4.1 Uniqueness and Stability in Bulk Materials . . . . . . . . . . . . . . . .
4.2 Equivalence of Uniqueness and Stability in 2D . . . . . . . . . . . . .
4.3 The Case of O(2) Invariant Sets . . . . . . . . . . . . . . . . . . . . . . . . . .
4.4 Applications to Thin Films . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.5 Applications to Finite Element Minimizers . . . . . . . . . . . . . . . . .
4.6 Extensions to Higher Order Laminates . . . . . . . . . . . . . . . . . . . .
4.7 Numerical Analysis of Microstructure A Review . . . . . . . . . .
83
86
101
102
109
115
120
122
5.
Applications to Martensitic Transformations . . . . . . . . . . . . . .

5.1 The Cubic to Tetragonal Transformation . . . . . . . . . . . . . . . . . .
5.2 The Cubic to Trigonal Transformation . . . . . . . . . . . . . . . . . . . .
5.3 The Cubic to Orthorhombic Transformation . . . . . . . . . . . . . . .
5.4 The Tetragonal to Monoclinic Transformations . . . . . . . . . . . . .
5.5 Reduction by Symmetry Operations . . . . . . . . . . . . . . . . . . . . . .
127
128
134
135
143
151
VIII
Contents
6.
Algorithmic Aspects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153

6.1 Computation of Envelopes of Functions . . . . . . . . . . . . . . . . . . . 154
6.2 Computation of Laminates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
7.
Bibliographic Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.2 Semiconvex Hulls of Compact Sets . . . . . . . . . . . . . . . . . . . . . . . .
7.3 Macroscopic Energy for Nematic Elastomers . . . . . . . . . . . . . . .
7.4 Uniqueness and Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.5 Applications to Martensitic Transformations . . . . . . . . . . . . . . .
7.6 Algorithmic Aspects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
177
177
178
179
180
181
182
A. Convexity Conditions and Rank-one Connections . . . . . . . . . 183

A.1 Convexity Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
A.2 Existence of Rank-one Connections . . . . . . . . . . . . . . . . . . . . . . . 189
B. Elements of Crystallography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
C. Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
1. Introduction
Many material systems show fascinating microstructures on dierent length

scales in response to applied strains, stresses, or electromagnetic elds. They
are at the heart of often surprising mechanical properties of the materials
and a lot of research has been directed towards the understanding of the underlying mechanisms. In this text, we focus on two particular systems, shape
memory materials and nematic elastomers, which display similar microstructures, see Figure 1.1, despite being completely dierent in nature. The reason
for this remarkable fact is that the oscillations in the state variables are triggered by the same principle: breaking of symmetry associated with solid to
solid phase transitions. In the rst system we nd an austenite-martensite
transition, while the second system possesses an isotropic to nematic transition.
An extraordinarily successful model for the analysis of phase transitions
and microstructures in elastic materials was proposed by Ball&James and
Chipot&Kinderlehrer based on nonlinear elasticity. They shifted the focus
from the purely kinematic theory studied so far to a variational theory. The
fundamental assumption in their approach is that the observed microstructures correspond to elements of minimizing sequences rather than minimizers
for a suitable free energy functional with an energy density that reects the
breaking of the symmetry by the phase transition. This leads to a variational
problem of the type: minimize

J (u, T ) =
W (Du(x), T )dx,
where R denotes the reference conguration of the elastic body, x

the spatial variable, u : R3 the deformation, T the temperature, and
W : M33 R+ R+ the energy density. The precise form of W depends
on a large number of material parameters and is often not explicitly known.
However, the strength of the theory is that no analytical formula for the
energy density is needed. The behavior of deformations with small energy
should be driven by the structure of the set of minima of W , the so-called
energy wells, which are entirely determined by the broken symmetry.
These considerations lead naturally to the following two requirements for
the energy density W . First, the fundamental axiom in continuum mechanics
that the material response be invariant under changes of observers, i.e.,
3
G. Dolzmann: LNM 1803, pp. 110, 2003.

c Springer-Verlag Berlin Heidelberg 2003
1. Introduction
Fig. 1.1. Microstructures in a single crystal CuAlNi (courtesy of Chu&James,

University of Minnesota, Minneapolis) and in a nematic elastomer (courtesy of
Kundler&Finkelmann, University Freiburg).
W (RF, T ) = W (F, T ) for all R SO(3).
(1.1)
Secondly, the invariance reecting the symmetry of the high temperature

phase, i.e.,
W (RT F R, T ) = W (F, T ) for all R Pa ,
(1.2)
where Pa is the point group of the material in the high temperature phase.
Here we restrict ourselves to invariance under the point group since the assumption that the energy be invariant under all bijections of the underlying
crystalline lattice onto itself leads to a very degenerated situation with a
uid-like behavior of the material under dead-load boundary conditions. The
two hypotheses (1.1) and (1.2) have far reaching consequences which we are
now going to discuss briey (see the Appendix for notation and terminology).
We focus on isothermal situations, and we assume therefore that W 0 and
that the zero set K is not empty,
K(T ) = {X : W (X, T ) = 0} = for all T.
We deduce from (1.1) and (1.2) that
U K(T )
QU R K(T ) for all Q SO(3), R Pa .
(1.3)
This implies that K(T ) is typically a nite union so-called energy wells,
K(T ) = SO(3)U1 . . . SO(3)Uk .
(1.4)
We refer to sets with such a structure often as multi-well sets. Here the
matrices Ui describe the k dierent variants of the phases and k is determined
from the point groups of the austenite and the martensite alone. A set of the
form SO(3)Ui will in the sequel frequently be called energy well.
We now describe the framework for the mathematical analysis of martensitic transformations and its connection with quasiconvex hulls.
1.1 Martensitic Transformations and Quasiconvex Hulls
T > Tc
T < Tc
Fig. 1.2. The cubic to tetragonal phase transformation.

A fundamental example of an austenite-martensite transformation is the cubic to tetragonal transformation that is found in single crystals of certain
Indium-Thallium alloys. The cubic symmetry of the austenitic or high temperature phase is broken upon cooling of the material below the transformation temperature. The three tetragonal variants that correspond to elongation
of the cubic unit cell along one of the three cubic axes and contraction in
the two perpendicular directions, are in the low temperature phase states of
minimal energy, see Figure 1.2. If we use the undistorted austenitic phase as
the reference conguration of the body under consideration, then the three
tetragonal variants correspond to ane mappings described by the matrices
2 0 0
1 0 0
1 0 0
U1 = 0 1 0 , U2 = 0 2 0 , U3 = 0 1 0
0 0 1
0 0 1
0 0 2
with 2 > 1 > 1 > 0 (if the lattice parameter of the cubic unit cell is equal
to one, then 1 and 2 are the lattice parameters of the tetragonal cell, i.e.,
are the lengths of the shorter and the longer sides of the tetragonal cell,
respectively). In accordance with (1.3), the variants are related by
U2 = R2T U1 R2 ,
U3 = R3T U1 R3 ,
where R2 and R3 are elements in the cubic point group given by
01 0
0 01
R2 = 1 0 0 , R3 = 0 1 0 .
0 0 1
1 00
A short calculation shows that no further variants can be generated by the
action of the cubic point group.
The origin for the formation of microstructure lies in the fact that the different variants can coexist in one single crystal. They can be formed purely
1. Introduction
Fig. 1.3. Formation of an interface between two variants of martensite in a single

crystal.
displacively, without diusion of the atoms in the underlying lattice. This is

illustrated in Figure 1.3. Consider a cut along a plane with normal (1, 1, 0)
The upper part is stretched in direction (1, 0, 0) while the lower part is elongated in direction (0, 1, 0). This corresponds to transforming the material into
the phases described by U1 and U2 , respectively. After a rigid rotation of the
upper part, the pieces match exactly and the local neighborhood relations of
the atoms have not been changed.
The austenite-martensite transition has an important consequence: the
so-called shape memory eect, which leads to a number of interesting technological applications. A piece of material with a given shape for high temperatures can be easily deformed at low temperatures by rearranging the
martensitic variants. Upon heating above the transformation temperature,
the material returns to the uniquely determined high temperature shape, see
Figure 1.4.
Mathematically, the existence of planar interfaces between two variants of
martensite is equivalent to the existence of rank-one connections between the
corresponding energy wells. Here we say that two wells SO(3)Ui and SO(3)Uj ,
i = j, are rank-one connected if there exists a rotation Q SO(3) such that
QUi Uj = a n
(Hadamards jump condition),
(1.5)
where the matrix an is dened by (an)kl = (ak nl ) for a, n R3 . If (1.5)

holds, then n is the normal to the interface. More importantly, the existence
of rank-one connections together with the basic assumption that the energy
density W be positive outside of K(T ) implies that W cannot be a convex
function along rank-one lines. We conclude that W cannot be quasiconvex
since rank-one convexity is a necessary condition for quasiconvexity. Recall
that a function W : Mmn R is said to be quasiconvex if the inequality

W (F + D)dx
W (F )dx
[0,1]n
[0,1]n
holds for all F Mmn and C0 ([0, 1]n ; Rm ). Quasiconvexity of W

is (under suitable growth and coercivity assumptions) equivalent to weak
sequential lower semicontinuity of the corresponding energy functional and
0000000000
1111111111
00000000000
11111111111
0000000000
1111111111
00000000000 T<T
11111111111
111111
000000
0000000000
00000000000
11111111111
c
000000 1111111111
111111
0000000000
1111111111
00000000000
11111111111
000000
111111
0000000000
1111111111
00000000000
11111111111
000000
111111
0000000000
1111111111
00000000000
11111111111
000000
111111
0000000000
000000 1111111111
111111
000000
111111
1111111111
000000
111111
000000000000000000000
11111111111
0000000000
1111111111
000000
00000000000
11111111111
0000000000111111
1111111111
000000
111111
00000000000
11111111111
0000000000
000000
111111
000000000001111111111
11111111111
0000000000
1111111111
000000
00000000000
11111111111
0000000000111111
1111111111
000000
111111
000000000001111111111
11111111111
0000000000
000000
111111
00000000000
11111111111
0000000000
1111111111
000000
111111
00000000000
11111111111
0000000000
1111111111
000000
00000000000
11111111111
0000000000111111
1111111111
000000
111111
00000000000
11111111111
0000000000
1111111111
000000
00000000000
11111111111
0000000000111111
1111111111
00000000000
11111111111
00000000000
11111111111
deform
heat
T>Tc
00000000
11111111
0000
1111
11111111
00000000
0000
1111
00000000
11111111
0000
1111
00000000
11111111
0000
1111
00000000
11111111
0000
1111
00000000
11111111
0000
1111
00000000
11111111
0000
1111
00000000
11111111
0000
1111
00000000
11111111
0000
1111
0000000000
1111111111
000000000
111111111
00000000
11111111
0000
1111
0000000000
1111111111
000000000
111111111
00000000
11111111
0000
1111
0000000000
1111111111
000000000
111111111
00000000
11111111
0000
1111
0000000000
1111111111
000000000
111111111
00000000
11111111
0000
1111
0000000000
1111111111
000000000
111111111
00000000
11111111
0000
1111
0000000000
1111111111
000000000
111111111
0000000000
1111111111
000000000
111111111
00000000000
11111111111
00000
11111
0000000000
1111111111
000000000
111111111
00000000000
11111111111
00000111111111
11111
0000000000
1111111111
000000000
00000000000
11111111111
00000
11111
0000000000
1111111111
000000000
00000000000
11111111111
00000111111111
11111
0000000000
1111111111
000000000
111111111
00000000000
11111111111
00000111111111
11111
0000000000
1111111111
000000000
00000000000
11111111111
00000
11111
0000000000
1111111111
000000000
111111111
00000000000
11111111111
00000
11111
0000000000
1111111111
000000000
111111111
00000000000
11111111111
00000
11111
00000000000
11111111111
00000
11111
00000000000
11111111111
00000
11111
00000000000
11111111111
00000
11111
heat
T<Tc
Fig. 1.4. The shape memory eect. The gure shows two macroscopically dierent
deformations of the same lattice by using dierent arrangements of the martensitic
variants. The upper conguration uses just two deformation gradients, the lower
one six.
therefore the crucial notion of convexity in the vector valued calculus of

variations (see Section A.1 for details about the relevant notions of convexity).
The mathematical interest in these problems lies exactly in this lack of
quasiconvexity, that excludes a (nave) application of the direct method in
the calculus of variations. Many questions concerning existence, regularity,
or uniqueness of minimizers remain open despite considerable progress in
the past years. In this text we focus mainly on the question of for which
ane boundary data the inmum of the energy is zero. This is not only a
challenging mathematical problem, but also of considerable interest in applications since it characterizes for example all ane deformations of a shape
memory material that can be recovered upon heating. From now on we drop
the explicit dependence on the temperature T in the notation since we are
restricting ourselves to isothermal processes. We therefore dene the quasiconvex hull K qc of the compact zero set K of W (at xed temperature)
as

K qc = F :
inf
W (Du)dx = 0 .
uW 1, (;R3 )
u(x)=F x on
The following exemplary construction shows that K qc is typically nontrivial, i.e., K qc = K, and provides at the same time a link between the
computation of K qc and the ne-scale oscillations (or microstructure) between dierent variants observed in experiments, see Figure 1.1. Suppose that
QU1 U2 = a n and let
1. Introduction
F = QU1 + (1 )U2 = U2 + a n.
(1.6)
Then F K for all (0, 1), except for possibly a nite number of values
since the assumption Fi = U2 + i a n = Qi U with Qi SO(3), i = 1, 2,
and {1, . . . , k} leads to
(1 2 )a UT n = Q1 Q2 .
This contradicts the fact that there are no rank-one connections in SO(3).
We now choose (0, 1) such that
F = QU1 + (1 )U2 = U2 + a n K,
and assert that F K qc . To prove this, we construct a minimizing sequence
uj such that J (uj ) 0 as j . Let be the one-periodic function
R {0, 1} with (0) = 0 given in (0, 1 ] by

0 if t (0, ],
(t) =
1 if t (, 1 ],
and dene

(t) =
0
(s)ds,
,j (t) =
1
(jt) for j N.
j
Then

j (x) = U2 x + ,j x, n a
u
j converges to F x strongly in L and
satises D
uj (x) {QU1 , U2 } a.e., u
1,
j close to the boundary of
weakly in W
, and we only need to modify u
in order to correct the boundary data. This can be done for example by
choosing a cut-o function C ([ 0, )) such that 0 in [ 0, 12 ) and
1 in [ 1, ). Then

uj (x) = 1 (j dist(x, )) F x + (j dist(x, ))
uj (x)
(1.7)
has the desired properties, and a short calculation shows that the energy
in the boundary layer converges to zero as j . Hence J (uj ) 0,
while J (F x) > 0. Therefore it is energetically advantageous for the material
to form ne microstructure, i.e., minimizing sequences develop increasingly
rapid oscillations. This argument shows that F K qc .
It is clear that this process for the construction of oscillating sequences
and elements in K qc can be iterated. In fact, if F and G are matrices with
the foregoing properties that satisfy additionally rank(F G ) = 1, then
F + (1 )G K qc for [ 0, 1 ].
We denote by K lc the set of all matrices that can be generated in nitely
many iterations, the so-called lamination convex hull of K. It yields an important lower bound for the quasiconvex hull K qc :
K lc K qc .
This is an extremely useful way to construct elements in K qc , which we will
refer to as lamination method, but it has also its limitations, mainly due to
the fact that it requires to nd explicitly rank-one connected matrices in the
set K. Therefore the following equivalent denition of K qc , which is in nice
analogy with the dual denition of the convex hull of a compact set, is of
fundamental importance

K qc = F : f (F ) sup f (X) for all f : Mmn R quasiconvex . (1.8)
XK
Thus K qc is the set of all matrices that cannot be separated from K by quasiconvex functions. While this characterization seems to be at a rst glance
of little interest (the list of quasiconvex functions that is known in closed
form is rather short), it allows us to relate K qc to two more easily accessible
hulls of K, the rank-one convex hull K rc and the polyconvex hull K pc which
are dened analogously to (1.8) by replacing quasiconvexity with rank-one
convexity and polyconvexity, respectively (see Section A.1 for further information). All these hulls will be referred to as semiconvex hulls. The method
for calculating the dierent semiconvex hulls based on this denition - separating points from a set by semiconvex functions - will be called the separation method in the sequel. Since rank-one convexity is a necessary condition
for quasiconvexity and polyconvexity a sucient one, we have the chain of
inclusions
K lc K rc K qc K pc ,
and frequently the most practicable way to obtain formulae for K qc is to
identify K lc and K pc .
There exists an equivalent characterization of the semiconvex hulls of K
that turns out to be a suitable generalization of the representation of the
convex hull of K as the set of all centers of mass of (nonnegative) probability
measures supported on K. This formulation arises naturally by the search
for a good description of the behavior of minimizing sequences for the energy
functional. The sequence uj constructed in (1.7) converges weakly in W 1, to
the ane function u(x) = F x. This limit does not provide any information
about the oscillations present in the sequence uj . The right limiting object,
that encodes essential information about these oscillations, is the Young measure {x }x generated by the sequence of deformation gradients Duj . This
approach was developed by L. C. Young in the context of optimal control
problems, and introduced to the analysis of oscillations in partial dierential
equations by Tartar. By the fundamental theorem on Young measures (see
Section A.1 for a statement) we may choose a subsequence (not relabeled)
of the sequence uj such that the sequence Duj generates a gradient Young
measure {x }x that allows us to calculate the limiting energy along the
subsequence via the formula
1. Introduction

lim J (uj ) = lim

W (Duj )dx =
M33
W (A)dx (A)dx.
We conclude that supp x K a.e. since J (uj ) 0 as j . The averaging technique for gradient Young measures ensures the existence of a
homogeneous gradient Young measure with
, id = F and supp K.
Here we say that the gradient Young measure {x }x is homogeneous if
there exists a probability measure such that x = for a.e. x, see Section
A.1 for details. For example, the sequence uj in (1.7) generates the homogeneous gradient Young measure = QU1 + (1 )U2 which is usually
referred to as a simple laminate. It turns out that K qc is exactly the set
of centers of mass of homogeneous gradient Young measures supported on
K. Since this special class of probability measures is by the work of Kinderlehrer and Pedregal characterized by the validity of Jensens inequality for
quasiconvex functions we obtain

K qc = F = , id : P(K), f (, id) , f

for all f : Mmn R quasiconvex ,
where P(K) denotes the set of all probability measures supported on K.
This formula gives the convex hull of K if we replace in the denition quasiconvexity by convexity, since Jensens inequality holds for all probability
measures. The obvious generalizations to rank-one convexity and polyconvexity provide equivalent denitions for the other semiconvex hulls which
are extremely useful in the analysis of properties of generic elements in these
hulls. In particular, since the minors M of a matrix F are polyane functions
(i.e., both M and M are polyconvex) we conclude that the polyconvex hull
is determined from measures P that satisfy the so-called minor relations
, M = M (, id) for all minors M.
In the three-dimensional situation this reduces to

K pc = F = , id : P(K), cof F = , cof , det F = , det . (1.9)
1.2 Outline of the Text

With the denitions of the foregoing section at hand, we now briey describe
the topics covered in this text. A more detailed description can be found at
the beginning of the each chapter.
In Chapter 2 we focus on the question of how to nd closed formulae
for semiconvex hulls of compact sets. There does not yet exist a universal
method for the resolution of this problem, but three dierent approaches are
emerging as very powerful tools to which we refer to as the separation method,
the lamination method, and the splitting method, respectively. As a very
1.2 Outline of the Text
instructive example for the separation and the splitting method, we analyze
a discrete set of eight points. Then we characterize the semiconvex hulls for
compact sets in 22 matrices with xed determinant that are invariant under
multiplication form the left by SO(2). We thus nd a closed formula for all sets
arising in two-dimensional models for martensitic phase transformations. The
results are then extended to sets invariant under O(2, 3) which are relevant
for the description of thin lm models proposed by Bhattacharya and James.
As a preparation for the relaxation results in Chapter 3 we conclude with the
analysis of sets dened by singular values.
Chapter 3 is inspired by the experimental pictures of striped domain patterns in nematic elastomers, see Figure 1.1, which arise in connection with a
nematic to isotropic phase transformation. For this material, Bladon, Terentjev and Warner derived a closed formula WBTW for the free energy density
which depends on the deformation gradient F and the nematic director n,
but not on derivatives of n. From the point of view of energy minimization,
one can rst minimize in the director eld n and one obtains a new energy W
that depends only on the singular values of the deformation gradient. This
is a consequence of the isotropy of the high temperature phase which has
in contrast to crystalline materials no distinguished directions. We derive an
explicit formula for the macroscopic energy W qc of the system which takes
into account the energy reduction by (asymptotically) optimal ne structures
in the material.
We begin the discussion of aspects related to the numerical analysis of
microstructures in Chapter 4. The standard nite element method seeks a
minimizer of the nonconvex energy in a nite dimensional space. Assume for
example that h is a quasiuniform triangulation of and that Sh (h ) is a
nite element space on h (a typical choice being the space of all continuous
functions that are ane on the triangles in h ). Suitable growth conditions
on the energy density imply the existence of a minimizer uh Sh (h ) that
satises J (uh ) J (v h ) for all v h Sh (h ). To be more specic, let us
assume that we minimize the energy subject to ane boundary conditions
of the form (1.6). If one chooses for v h an interpolation of the functions uj
in (1.7) with j carefully chosen depending on h, then one obtains easily that
the energy converges to zero at a certain rate for h 0,
J (uh ) ch ,
, c > 0.
(1.10)
The fundamental question is now what this information about the energy
implies about the nite element minimizer. Recent existence results for nonconvex problems indicate that minimizers of J are not unique (if they exist),
and in this case the bound (1.10) is rather weak. It merely shows that the
inmum can be well approximated in the nite element space (absence of
the Lavrentiev or gap phenomenon). On the other hand, if J fails to have a
minimizer, then it is interesting to investigate whether for a suitable set of
boundary conditions the minimizing microstructure (the Young measure )
10
1. Introduction
is unique and what (1.10) implies for uh as h tends to zero. In particular, if

is unique, then the sequence Duh should display very specic oscillations,
namely those recorded in . This is the motivation behind Luskins stability
theory for microstructures, and we present a general framework that allows
one to give a precise meaning to this intuitive idea. Our approach is inspired
by the idea that stability should be a natural consequence of uniqueness, and
we verify this philosophy for ane boundary conditions F K qc based on
an algebraic condition, called condition (Cb ), on the set K. The new feature
in our analysis is to base all estimates on inequalities for polyconvex measures. This method turns out to be very exible and we include extensions
to thin lm theories and more general boundary conditions that correspond
to higher order laminates.
We apply the general theory developed in Chapter 4 to examples of
martensitic phase transformations in Chapter 5. Our focus is to analyze the
uniqueness of simple laminates based on our algebraic condition (Cb ). It
turns out that typically simple laminates are uniquely determined from their
center of mass unless the lattice parameters in the denition of the set K satisfy a certain algebraic condition. In theses exceptional case, we provide explicit characterizations for the possible microstructures underlying the ane
deformation F = , id.
Algorithmic aspects in the numerical analysis and computation of microstructures are addressed in Chapter 6. Nonconvex variational problems
typically fail to have a classical solution or they have solutions with intrinsically complicated geometries that cannot be approximate numerically. A
natural remedy here is either to replace the original functional by its relaxation or to broaden the class of solutions, i.e., to pass from minimizers in the
classical sense to minimizing Young measures. The latter approach requires a
discretization of the space of all probability measures, and it is an open problem to nd an ecient way to accomplish this. However, (nite) laminates
are accessible to computations and the explicitly known examples indicate
that this subclass is in fact sucient in many cases (the Tartar square being
a notable exception, see Section 6.2). In this chapter we rst discuss an algorithm for the computation of the rank-one convex envelope of a given energy
density which is an upper bound for the relaxed or quasiconvexied energy
density in the relaxed energy functional. We then modify this algorithm to
nd minimizing laminates of nite order and we prove rigorously convergence
of the proposed algorithms under reasonable assumptions.
In Chapter 7 we present detailed references to literature closely related
to the text. Additional references can be found in the appendices in which
we summarize some background material that is not included in the text.
Appendix A contains information about notions of convexity and classical
criteria for the existence of rank-one connections between matrices. Basic
nomenclature in crystallography is summarized in Appendix B, and the mathematical notation used throughout the text is collected in Appendix C.
2. Semiconvex Hulls of Compact Sets
Quasiconvex hulls of sets and envelopes of functions are at the heart of the
analysis of phase transformations by variational techniques. In this chapter we
address the question of how to nd for a given compact set K its quasiconvex
hull K qc . In general, this is an open problem since all characterizations of K qc
are intimately connected to the notion of quasiconvexity of functions in the
sense of Morrey, and the understanding of quasiconvexity is one of the great
challenges in the calculus of variations. However, three conceptual approaches
can be identied that allow one to resolve the problem for important classes
of sets K.
The separation method. By denition, K qc is the set of all matrices F that
cannot be separated by quasiconvex functions from K. If an inner bound
for K qc is known, i.e., a set A with A K qc , for example A = K lc or
A = K rc , then it suces to nd for all F Mmn \ A a quasiconvex function
f with f 0 in A and f (F ) > 0. An example of this approach is our
proof for the formulae of the semiconvex hulls of the set K with eight points
aks examples of quasiconvex functions on
in Theorem 2.1.1 based on Sver
symmetric matrices.
The lamination method. Since quasiconvexity implies rank-one convexity,
the segments F1 + (1 )F2 , [ 0, 1 ] are contained in K qc if the end
points F1 and F2 belong to K qc and satisfy rank(F1 F2 ) = 1. This fact
motivated the denition of the lamination convex hull K lc which is one of the
fundamental inner bounds for K qc . The lamination method tries to identify
K lc and then to use K lc as an inner bound for the separation method.
The splitting method. This method is well adapted to situations where a
good outer bound A for K qc is known, i.e., K qc A. It can be interpreted
as a reversion of the lamination method. Assume for example that A is given
by a nite number of inequalities, as for example in Theorems 2.1.1 or 2.2.3
below. We call a point F an unconstrained point, if we have strict inequality
in all inequalities dening A, and a constrained point else. Assume that F is
an unconstrained point. By continuity, Ft = F (I + ta b) belongs to A for
all a Rm , b Rn and t suciently small. Since K is compact, all the hulls
are compact and we may suppose that A is compact. We now dene t to
be the smallest positive (largest negative) parameter t for which the matrix
Ft satises equality in at least one of the inequalities in the denition of A,
G. Dolzmann: LNM 1803, pp. 1168, 2003.

12
i.e., is a constrained point. Therefore it suces to prove that the constrained

points belong to K lc in order to show that A K lc . Then
K qc A K lc K qc
and thus A = K qc . Having equality in one of the inequalities dening A provides additional information and often simplies the proof that certain matrices belong to K lc . Moreover, this procedure can be iterated and is also applicable to sets with a determinant constraint, since det Ft = det F if a, b = 0.
A convenient choice of A is often K pc . We use this general strategy, which
we call the splitting method, extensively in the following sections.
A natural question that arises in this context is whether the polyconvex
hull K pc coincides with the rank-one convex hull K rc (or even the lamination
convex hull K lc ), since in these cases a characterization of K qc is automatically obtained. More generally, one can ask whether Mpc (K) = Mrc (K), i.e.,
whether the set of all polyconvex measures satisfying the minors relations is
equal to the set of all laminates characterized by Jensens inequality for all
rank-one convex functions. It turns out that this is typically not true, and to
illustrate this point we consider the two-well problem where K is given by

0
1/ 0
K = SO(2)U1 SO(2)U2 , U1 =
, U2 =
0 1/
0
with > 1. Then Mpc is the set of all probability measures that satisfy the
minors relation
det , id = , det .
(2.1)
Consider now the special class of all probability measures Mpc (K) that
are supported on three points,
= 1 X1 + 2 X2 + 3 X3 ,
Xi K, i (0, 1), 1 + 2 + 3 = 1.
We assert that every Mqc Mpc of this form is in fact a laminate.

aks results that at least two of the three matrices
Indeed, it follows from Sver
Xi must be rank-one connected. We may thus assume that Xi SO(2)Ui for
i = 1, 2 with rank(X1 X2 ) = 1. The minors relation and the expansion
det F =
3

1 2
det(X2 X1 )
1 3

2
1
3 (1 3 ) det
X1 +
X2 X3
1 3
1 3
i det Xi
i=1
now imply that

rank X3

1
2
X1 +
X2 = 1,
1 + 2
1 + 2
2.1 The Eight Point Example
13
i.e., that Mrc (K) is a second order laminate. In order to prove that
Mpc (K) = Mqc (K) it suces to construct a solution of (2.1) with matrices
Xi that are not rank-one connected. For a specic example, we choose > 1
to be the solution of ( + 1/)2 = 8 and we dene

0 1/
0
X1 = U1 , X2 = JU1 =
, X3 = JU2 =
,
0
1/ 0
where J denotes the counter-clockwise rotation by /2, and we x = 1/3.
Then

3

1
1
1
( + 1/)
det
= 1 + ( + )2 = 1,
i X i =
1/
9 + 1/
9
i=1
and consequently Mpc (K) \ Mqc (K). Note, however, that , id K qc .

We use the idea to nd elements in K pc by solving the minors relations for
example in Theorem 2.2.6 to nd an SO(2) invariant set K with K pc = K rc .
This chapter is divided into several sections in which we discuss quasiconvex hulls for various classes of sets K. We begin with a very illustrative
example of a discrete set in symmetric 2 2 matrices rst analyzed by Dacorogna and Tanteri. For this set, the separation method provides one with
an outer bound for K qc and the splitting method allows one to show that
K qc = K lc for a large range of parameters. We then turn towards sets with
constant determinant that are invariant under SO(2). This is the class of sets
relevant in two-dimensional models for phase transformations. Theorem 2.2.5
gives an explicit characterization of the semiconvex hulls. These results are
then extended to O(2, 3) invariant sets related to the modeling of thin lms.
Finally, we study sets dened by singular values and the results obtained here
are an important ingredient in the derivation of the quasiconvex envelope of
a model energy for nematic elastomers in Chapter 3.

We begin the analysis of semiconvex hulls with the following example of a
discrete set with eight points in symmetric 22 matrices that was introduced
by Dacorogna and Tanteri. Theorem 2.1.1 extends their results and the proof
illustrates the power of the separation and the splitting method. Before we
state the theorem, we discuss briey quasiconvexity in symmetric matrices.
Let S n Mnn denote the subspace of all symmetric matrices. A function
f : S n R is said to be quasiconvex if for all matrices F S n and all
W02, () the inequality

f (F + D2 )dx
f (F )dx
14
holds. The proof of our characterization of the quasiconvex hull relies on

aks result that the function
Sver

det F if F is positive denite,
g0 (F ) =
0
else,
is quasiconvex on symmetric matrices. This function gives new restrictions
on gradient Young measures supported on symmetric matrices since they
have to satisfy the inequality

(2.2)
g0 , id , g0 .
It is an open problem whether these functions can be extended to all n n
matrices and whether they can be used directly for separation on Mnn .
The statement of the following theorem emphasizes the fact that the description of K pc (K qc , K rc ) involves additional conditions compared to the
formulae for conv(K) (K pc , K qc ).
Theorem 2.1.1. Let

xy
K= F =
: |x| = a, |y| = b, |z| = c
yz
with constants a, b, c > 0. Then

xy
conv(K) = F =
: |x| a, |y| b, |z| c
yz
and

K pc = F conv(K) : (x a)(z + c) y 2 b2 , (x + a)(z c) y 2 b2 .
Moreover, the following assertions hold:

i) If ac b2 < 0 then

K (2) = K lc = K rc = F conv(K) : |y| = b .
ii) If ac b2 0 then K (4) = K lc = K rc = K qc and

K qc = F K pc : (x a)(z c) (|y| b)2 ,

(x + a)(z + c) (|y| b)2 .
Remark 2.1.2. Note that K qc is quasiconvex as a set in all 2 2 matrices,
not only as a subset of the symmetric matrices.
Remark 2.1.3. It is an open problem to nd a formula for the quasiconvex
hull of K in the case ac b2 < 0.
15
Remark 2.1.4. A short calculation shows that the additional inequalities in

the denition of K lc are true for all F K pc if ac b2 = 0 and that consequently K lc = K pc . This was already shown by Dacorogna and Tanteri.
The authors also obtained the formula for K lc in the case ac b2 < 0 and
observed that K lc is always contained in the intersection of the convex hull
of K with the exterior of the two hyperboloids (x a)(z + c) = y 2 b2 and
(x + a)(z c) = y 2 b2 . However, they did not identify the latter set as K pc .
We now turn towards the proof of the theorem which we split into several steps. First we derive the formula for the polyconvex hull of K. Then
the formulae for the lamination convex hulls in statements i) and ii) in the
theorem are established. Finally we present the proof for the representation
of the quasiconvex hull for ac b2 0.
The Polyconvex Hull of K. Among the dierent notions of convexity,
polyconvexity has the most similarities with classical convexity. One instance
is the following representation for the polyconvex hull K pc ,

,
K pc = F M22 : (F, det F ) conv(K)
(2.3)
where

= (F, det F ) : F K R5 .
K
and
By denition, K consists of symmetric matrices, and therefore K
5

conv(K) are contained in a four-dimensional subspace of R . We restrict
our calculations to this subspace by the identications

a
a
a
a
a
a
a
a
K = c , c , c , c , c , c , c , c
b
b
b
b
b
b
b
b
and

= (x, z, y, xz y 2 ) : (x, z, y) K .
K
.
,...,f
by f
We denote the eight points in K
1
8

Since K is a nite set, conv(K) is a polyhedron in R4 , which is the inter
section of a nite number of half spaces. Moreover, on each face of conv(K)

we must have at least four points in K that span a three-dimensional hyperplane in R4 . A short calculation shows that the following list of six normals

completely describes the convex hull of K:
n1 = (c, a, 0, 1),
n3 = (c, a, 0, 1),
n5 = (0, 0, 1, 0),
n2 = (c, a, 0, 1, ),
n4 = (c, a, 0, 1),
n6 = (0, 0, 1, 0).
16
It turns out that the hyperplanes dened by n1 , . . . , n4 contain six points in

K,
, n1 = f
, n1 = 3ac + b2 < ac + b2 = f
, n1 ,
f
4
8
i
2
2

f 3 , n2 = f 7 , n2 = 3ac b < ac b = f i , n2 ,
, n3 = f
, n3 = 3ac b2 < ac b2 = f
, n3 ,
f
2
6
i
, n4 = f
, n4 = 3ac + b2 < ac + b2 = f
, n4 ,
f
1
i {4, 8},
i {3, 7},
i {2, 6},
i {1, 5},
and that the faces of the polyhedron dened by n5 and n6 contain four
points,
, n5 = b < b = f
, n5 ,
f
j
i
i = 1, 2, 3, 4, j = 5, 6, 7, 8,
, n6 ,
, n6 = b < b = f
f
j
i
i = 5, 6, 7, 8, j = 1, 2, 3, 4.
We include a few details of the argument that leads to this characterization.

The general idea is to choose (at least) four of the eight points f i in K and
span a three-dimensional plane
to check whether the corresponding points f
i

in K. If this plane constitutes a separating plane (a face), then its normal is

added to the description of conv(K).
As an example we choose f i , i = 1, 3, 5, 7. We need to check whether
f
,
these points dene a plane in R4 , i.e., whether the three vectors f
1
7
, and f
f
are linearly independent. It turns out that these vectors
f
f
3
7
5
7
are linearly dependent, and thus it is possible to add a further vector to the
f
, j = 1, 2, 3, 5, are linearly
list of vectors, say f 2 . Now the vectors f
j
7
independent since the rank of the matrix
0 2a 0 0
2c 2c 0 2c
A=
2b 2b 2b 0
2ac 0 0 2ac
f
,
is three (if one deletes the second column which corresponds to f
2
7
then the rank is only two). The corresponding normal vector has to satisfy
AT n = 0 and this leads to the linear system
0 c b ac
a c b 0
0 0 b 0 n = 0
0 c 0 ac
that has the solution n1 = (c, a, 0, 1). The corresponding equation in the
description of K pc is
, n1 = cx + az (xz y 2 ) ac + b2 .
f
17
Geometrically this is a hyperboloid which contains in fact six of the eight

points in K. It turns out that there are four hyperboloids that are important
in the description of K pc . In the following diagrams we sketch the set K in R3
(with the x-axis to the left, the y-axis out of the paper plane, and the z-axis
oriented upwards) and we circle the points in K that dene the corresponding
separating planes or hyperboloids.
Separating hyperboloid 1: We choose j = 1, 2, 3, 5.
sf
The corresponding system for the normal is given by
P

P
Psf

sf

PP
0010
P
sf
s
0 1 0 an = 0
P
P

Ps

100c
sf

PP
sf
P
and the solution gives n1 = (c, a, 0, 1).
sf
P

P
Psf

sf

PP
1 0 0c
P
sf
sf
0 0 1 0n = 0
P
P

Psf

0 1 0 a
sP

P
s
P
sP

P
Ps

sf

PP
001 0
P
sf
sf
0 1 0 a n = 0
P
P

Psf

1 0 0 c
sf

PP
sf
P
sf
P

P
Psf

sP

P
1 0 0 c
P
s
sf
0 0 1 0 n = 0
P
P

Psf

0 1 0 a
sf

PP
sf
P
18
Four points on a b-face of the cube. The corresponding

system for the normal is given by
s
PP
P
f
s

sP

100c
P
Psf

0 1 0 an = 0
s
P

P

P
f
s
ac00

s

PP
P
sf
and the solution gives n5 = (0, 0, 1, 0). Replacing b by
b leads to n6 = (0, 0, 1, 0).
In view of the representation (2.3) for the polyconvex hull and the formulae for the normals, this implies that all points in K pc must satisfy the
convex inequality
|y| b
(2.4)
as well as the additional inequalities

cx + az (xz y 2 ) ac + b2 ,
cx + az + (xz y 2 ) ac b2 ,
cx az + (xz y 2 ) ac b2 ,
cx az (xz y 2 ) ac + b2 ,
which we can rewrite as

(x a)(z c) y 2 + b2 , (x + a)(z c) y 2 b2 ,
(x a)(z + c) y 2 b2 , (x + a)(z + c) y 2 + b2 .
(2.5)
We now assert that this system of inequalities is equivalent to the conditions

|x| a,
|z| c,
|y| b
(2.6)
describing the convex hull of K and the two additional inequalities

(x + a)(z c) y 2 b2 ,
(x a)(z + c) y 2 b2 .
(2.7)
This proves the formula for the polyconvex hull of K. In fact, the sum of the
two upper and the two lower inequalities in (2.5) implies
az ac
and
az ac,
and the sum of the two left and the two right inequalities, respectively, gives
cx ac
and
cx ac.
Therefore |z| c and |x| a and this proves that (2.4) and (2.5) imply (2.6)
and (2.7). Conversely, if the convex inequalities |x| a, |z| c, and |y| b
in (2.6) hold, then x a 0, z c 0 and y 2 + b2 0. Consequently
(x a)(z c) y 2 + b2 . Similarly, we have x + a 0, z + c 0 and
thus (x + a)(z + c) y 2 + b2 , as asserted. This concludes the proof of the
formula for K pc for all parameters a, b, c > 0.
19
The Lamination Convex Hull of K for ac b2 < 0. We now turn

towards proving the formula for K lc and we assume rst that ac b2 < 0.
We let

A = F conv(K) : |y| = b .
In this case, none of the matrices in A with y = b is rank-one connected to
any of the matrices in A with y = b, and the assertion follows essentially
from the following well-known locality property of the rank-one convex hull.
Proposition 2.1.5. Assume that K is compact and that K rc consists of two
compact components C1 and C2 with C1 C2 = . Then
K rc = (K C1 )rc (K C2 )rc .
(2.8)
Clearly, all elements in A can be constructed using the rank-one connections between the four matrices in K with y = b and y = b, respectively. The observation is now that the polyconvex hull is not connected, since
K pc {F : |y| } = for > 0 so small that 2 < b2 ac. Indeed, summation of the two inequalities in the denition of K pc implies ac xz b2 y 2
or, equivalently, 0 > ac b2 + y 2 xz. Thus necessarily either x > 0 and
z < 0 or x < 0 and z > 0. In the former case the rst inequality cannot hold
since
(z a)(z + c) y 2 b2
0 x(z + c) az ac b2 + y 2 < 0.
In the latter case the second inequality is violated. We may now apply Proposition 2.1.5 and we conclude that K lc = K rc = A.
The Lamination Convex Hull of K for ac b2 0. Assume now that
ac b2 0, and let A be given by

A = F K pc : (x a)(z c) (|y| b)2 , (x + a)(z + c) (|y| b)2 .
By symmetry, we may suppose in the following arguments that y 0. Then
this set is described by three types of inequalities, namely the stripes
|x| a,
|z| c,
|y| b
(2.9)
dening the convex hull of K, the hyperboloids

(x a)(z + c) y 2 b2 ,
(x + a)(z c) y 2 b2 ,
(2.10)
in the denition of K pc , and the cones

(x a)(z c) (y b)2 ,
To simplify notation, we write
(x + a)(z + c) (y b)2 .
(2.11)
20

F =
Since A is compact, it suces to prove that all points F A that satisfy

equality in at least one of the inequalities in the denition of A can be constructed as laminates. To see this, assume that F lies in the interior of A. The
idea is to split F along a rank-one line into two rank-one connected matrices
F that satisfy equality in at least one of the inequalities in the denition of
A. We set
t = sup{t < 0 : F + tw w satises at least one equality in A},
t+ = inf {t > 0 : F + tw w satises at least one equality in A}.
By assumption, t < 0 < t+ and we dene F = F + t w w. Then
F = (t F + t F + )/(t+ t ) and it suces to show that F are contained
in K lc .
Assume thus that F A satises equality in at least one inequality in the
denition of A. We have to prove that this implies F K lc . This is immediate
for the convex inequalities |x| a, |y| b, and |z| c. For example, if = a,
then by (2.10) || = b and by symmetry we may assume that = b. Then
(2.9) implies that = c + (1 )(c) for some [ 0, 1 ] and thus

ab
a b
ab
a b
F =
+ (1 )
,
= 2ce2 e2 .
bc
b c
bc
b c
The argument is similar for || = c. Finally, if || = b and 0, then

(, ) conv (a, c), (a, c), (a, c), (a, c) ,
and therefore F K (2) .
Assume next that F lies on the surface of one of the cones
(x a)(z c) (y b)2 ,
(x + a)(z + c) (y b)2 .
These cones are the rank-one cones centered at points in K, and we may
suppose that F is contained in the rank-one cone C1 given by

ab
= (x a)(z c) (y b)2 = 0 ;
C1 = F : det F
bc
the argument is similar in the other case. The cone C1 intersects the part of
the boundary of the convex hull of K that is contained in the plane {y = b},
which by the foregoing arguments is contained in K (2) . We will show that F
belongs to a rank-one segment between a point G in this intersection and the
point F1 K, where F1 and G are given by

x b
ab
and G =
F1 =
,
|x| a, |z| c.
b z
bc
21
This implies F K (3) K lc . In order to prove this fact, let

a b
.
R = F1 F =
b c
By assumption, det R = 0, and we seek a t R such that

a + t(a ) b + t(b )
= G.
F1 + tR =
b + t(b ) c + t(c )
This implies
t=
2b
b
and thus
x=a
2b(a )
,
b
z =c
2b(c )
.
b
Clearly x a and we only have to check that x a, or equivalently

a
a
.
b
b
To establish this inequality, we subtract the equality (x a)(z c) = (y b)2
in the denition of C1 from the inequality (x + a)(z c) y 2 b2 in the
denition of K pc , and we obtain that F satises
2a( c) (2b)(b ).
Therefore, again in view of the denition of C1 ,
b
a
a
=
,
b
c
b
and this proves the bounds for x; the arguments for z are similar. Since
G K (2) we conclude
b+
1
b
1+t
F1 G =
F1 +
G K (3) .
F =
t
t
2b
2b
It remains to consider the case that F A satises equality in one of the
inequalities dening the one-sheeted hyperboloids. Assume thus that
( + a)( c) = 2 b2 .
The idea is to use the geometric property of one-sheeted hyperboloids H
ak, namely that for each point F on H there exist two
already observed by Sver
straight lines intersecting at F that are contained in H, and that correspond
22
to rank-one lines in the space of symmetric matrices. More precisely, we seek

solutions w = (u, v) S1 of
F + tw w H
or
( + tu2 + a)( + tv 2 c) = ( + tuv)2 b2 .
This is equivalent to the quadratic equation

u2 ( c) + v 2 ( + a) = 2uv.
Since u = 0 and v = 0 are only solutions for = a and = c, respectively,
we may assume that u, v = 0. In this case there are two solutions for the
ratio = u/v, given by
1,2 =
b
.
c
The strategy is now to split F into two points F along one of these rank-one
lines that satisfy equality in at least two of the inequalities in the denition
of A. Let
t = sup{t < 0 : F + tw w realizes two equalities in A},
t+ = inf {t > 0 : F + tw w realizes two equalities in A}.
By assumption, t < 0 < t+ and we dene F = F + t w w. In view
of the foregoing arguments, the matrices F belong either to K (3) or to the
of the two hyperboloids,
intersection H

= F : (x + a)(z c) = y 2 b2 , (x a)(z + c) = y 2 b2 .
H
The formula for the lamination convex hull is therefore established if we show
K lc . By symmetry it suces again to prove this for all F H
with
that H

y 0. Now, if F H, then
az = cx,
and
xz ac = y 2 b2 .
Thus the intersection of the two hyperboloids can be parameterized for y 0

by

a

c
t
t2 + ac b2 , t,
t2 + ac b2 , {1}, t 0.
c
a
We may assume that = 1. In this case the inequality (xa)(z c) (y b)2
in the denition of A is equivalent to (ac b2 )(b t)2 0 and this implies
t = b, and thus F K, if ac b2 > 0. If ac b2 = 0, then the intersection of
the hyperboloids coincides with the rank-one line between

ab
a b
a b
a b
and
, or
and
,
bc
b c
b c
b c
and consequently F K (1) . This proves the formula for the lamination convex hull.
23
The Quasiconvex Hull of K for ac b2 0. It remains to prove that

for ac b2 0 all points in K pc \ K lc can be separated from K (or equivalently from K lc ) with quasiconvex functions. Recall that by Remark 2.1.4
the quasiconvex and the polyconvex hull coincide for ac b2 = 0. We may
therefore assume in the following that ac b2 > 0. We divide the proof of
this assertion into three steps. First we show that the additional inequalities
in the denition of K lc are only active for x, z 0 or x, z 0. Then we
construct a suciently rich family of quasiconvex functions that separates
points from K, and nally we prove the theorem.
Reduction to the Case x, y, z 0. By symmetry we may always assume
that y 0. In this case the formula for K lc contains the additional inequalities
(x + a)(z + c) (y b)2 ,
(x a)(z c) (y b)2 .
(2.12)
Assume, for example, that F K pc with x 0 and z 0. The inequalities

in (2.12) can be rewritten as
(x a)(z c) b2 y 2 + 2y 2 2by.
It follows from F K pc that (x + a)(z c) b2 y 2 . The foregoing
inequalities are thus true if
(x a)(z c) (x + a)(z c) + 2y 2 2by
is satised. The equation with the minus and the plus sign are equivalent to
2x(z c) + 2y(b y) 0
and
2z(x + a) + 2y(b y) 0,
(2.13)
respectively. Since by assumption x 0, z c, and y [ 0, b ], the rst

inequality in (2.13) holds and this implies the rst inequality (2.12). Similarly,
the second inequality in (2.13) is true in view of z 0 and x a, and
consequently the second inequality in (2.12) follows.
Construction of Quasiconvex Functions. From now on we assume that
x, y, z 0 and that x = a, z = c and y = b. We need to show that all
points in K pc with (x a)(z c) < (y b)2 can be separated from K by
aks remarkable result
quasiconvex functions. This will be done using the Sver
that the functions

| det F | if the index of F is ,
g (F ) =
0
otherwise,
are quasiconvex on symmetric matrices. Here the index of the symmetric
matrix F is the number of its negative eigenvalues.
We begin by calculating the intersection of the boundary of the cone
(x a)(z c) (y b)2 with K pc for xed y [ 0, b). This intersection can
be parameterized by
24

t
t
y
y c + (y b)2 /(t a)
t Iy =
ay
b(y b)
,a +
,
b
c
and we write t F (y, t) or t Fy,t for simplicity. A short calculation shows

that |Iy | = (ac b2 )(b y)/(bc) > 0. We dene quasiconvex functions fy,t on
the space of all symmetric matrices by
fy,t (F ) = g0 (F Fy,t ),
y [ 0, b), t Iy ,
and show rst that fy,t = 0 on K. In order to do this, it suces to prove that
all the matrices of the form F F (y, t) with F K are not positive denite.
In fact,

a b
det
Fy,t = (a t)(c c) + (y b)2 (b y)2 0,
b c
and thus all matrices of the form F Fy,t , with F K and F11 = a are not
positive denite. Moreover,
"
!

a t
b y
a b
2
,
Fy,t =
b c
b y c c (yb)
ta
and consequently all the matrices X = F Fy,t with F K and F11 = a
satisfy X11 0 and are therefore not positive denite.
Separation of Points from K lc with Quasiconvex Functions. Recall
that we assume that

F =
with , , 0 and = a, = c, = b.
We have to show that all matrices F K pc with

( a)( c) < ( b)2
(2.14)
can be separated from K by a quasiconvex function. We will achieve this by

analyzing dierent regions for which are related to the intersection of K qc
with the plane y = . In this plane, the intersection of K qc with the quadrant
x 0 and z 0 is bounded by the three hyperbolic arcs
(x a)(z c) = ( b)2
and
(x a)(z c) = 2 b2 .
In the following we consider four dierent regions for 0 which are dened
by the points where two of these hyperbolic arcs intersect (see Figure 2.1).
More precisely, the hyperbola (xa)(zc) = (b)2 intersects the hyperbola
(x+a)(zc) = 2 b2 for x1 = a/b and the hyperbola (xa)(z+c) = 2 b2
for x2 = a + b( b)/c. The four cases now correspond to [ 0, x1 ],
(x1 , x2 ), = x2 , and (x2 , a), respectively. We begin with the last case
rst.
25
(xa)(z+c)=y2b2
z
(xa)(zc)=(yb) 2
x
(x+a)(z+c)=(yb)2
(x+a)(zc)=y 2b 2
Fig. 2.1. The polyconvex hull (bounded by the thick solid lines) and the quasiconvex
hull (the intersection of the four hyperbolic arcs) of K in the plane {y = > 0}.
Case a) Assume that > a + b( b)/c. If ( a)( + c) 2 b2 , then

c +
b2 2
c(b2 2 )
c
> c
= .
a
b( b)
b
We dene
b( b)
, Z = F G =
G = F , a +
c
a b( b)/c
0
0
c/b
Then Z is positive denite and in view of Section 2.1 the function g0 (F G )

separates F from K lc . On the other hand, if ( a)( + c) > 2 b2 , then F
does not belong to K pc .
Case b) Assume that = a + b( b)/c. We assert that in view of (2.14)
we may nd an x
I = (a/b, ) such that

x

0
Z = F F (, x
) =
x a)
0 c ( b)2 /(
is positive denite. This follows easily since F is positive denite if and only
if > x
and
c
( b)2
>0
x
a
or
(
x a)( c) ( b)2 < 0.
In view of (2.14) we can choose x

< close enough to x such that the latter
inequality holds. Therefore we can separate F from K lc with the function
g0 (F F (, x
)).

Case c) Assume that a/b, a + b( b)/c . The
conclusion
follows

as in case b), since we can choose by continuity x
a/b, such that
F F (, x
) is positive denite.
26
Case d) Assume that [ 0, a/b ]. We assert that no point in K pc satises

the inequality (2.14). If (2.14) holds, then
>c+
( b)2
.
a
However, for
x=x
=
a
b
and
z = z = c +
( b)2
x
a
the inequality (x + a)(z c) 2 b2 is satised with equality. If
a
b
and
>c+
( b)2
,
x
a
then ( + a)( c) > 2 b2 , a contradiction. The same argument applied

to gradient Young measures supported on K shows in connection with the
inequality (2.2) that K is quasiconvex as a subset in the space of all 2 2
matrices. This concludes the proof of the theorem.
2.2 Sets Invariant Under SO(2)

We now turn to the case of SO(2) invariant sets K which is relevant for twodimensional models in elasticity. After a discussion of k-well problems with
nite k, we present the general result for arbitrary compact sets K with equal
determinant. The section is concluded by the surprising example of the set
K given by

0
0
K = SO(2)
SO(2)
SO(2), > > 1,
0
0
for which K rc and K pc do not coincide. This example shows that the general
result for sets with constant determinant cannot be extended to sets without
this constraint.
The One-well Problem. We begin with the case of one well, K = SO(2).
In view of the denition of the polyconvex hull we must have det F = 1 for
all F K pc . However, the only elements F in the convex hull of SO(2) with
det F = 1 are proper rotations and therefore
K rc = K qc = K pc = SO(2).
27
Incompatible Wells. The result for the one-well problem has the following
generalization. Assume that K = SO(2)U1 . . . SO(2)Uk , k 1, where the
wells SO(2)Ui and SO(2)Uj are incompatible for i = j, i.e., det(QUi Uj ) = 0
ak proved that the hulls are
for i = j and Q SO(2). In this situation, Sver
still trivial,
K rc = K qc = K pc = SO(2)U1 . . . SO(2)Uk .
Motivated by applications to phase transformations we analyze next the
so-called k-well problem in two dimensions where K is given by
F = SO(2)U1 . . . SO(2)Uk ,
k 2,
with positive denite matrices Ui that satisfy det Ui = > 0 for i = 1, . . . , k.

The Two-well Problem. This problem has attracted a lot of attention
since it is the basic model for martensitic transformations in two dimensions.
Here K is given by K = SO(2)U1 SO(2)U2 with det U1 , det U2 > 0. We may
assume that U1 = I and that U2 is diagonal, i.e., that K is of the form
K = SO(2) SO(2)H,
H = diag(, ), 0 < 1 .
(2.15)
In view of Proposition A.2.5, the two wells are compatible. A short calculation
shows that there exists only one rank-one connection between the wells if one
of the two eigenvalues of H is equal to one (in fact, rank(I H) = 1), and
that there exist two rank-one connections if both eigenvalues are dierent
from one.
Remark 2.2.1. Assume that > 1 = , (the case < 1 = is similar) and
let

0
K = SO(2) SO(2)U, U =
.
01
Then K pc = K qc = K rc = K lc = K (1) and

K qc = F : F = QUs , Q SO(2), Us =
s0
01

, s [ 1, ] .
We include a short proof of this fact which is a nice application of the minors
relation.
Proof. Assume that Mpc (K) is a polyconvex measure represented
by = t + (1 t)U , where and are probability measures supported on
SO(2) and t [ 0, 1 ]. Let

s1 s2
r1 r2
, S = , id =
conv(SO(2)).
R = , id =
r2 r1
s2 s1
28
Then the minors relations imply

F = tR + (1 t)SU,
det F = t + (1 t).
It follows from
R : cof(SU ) = (r1 s1 + r2 s2 )(1 + ),
the expansion (C.3) for the determinant, and Youngs inequality that
det F = t2 det R + t(1 t)R : cof(SU ) + (1 t)2 det S det U
t2 + t(1 t)( + 1) + (1 t)2
= t + (1 t)
with strict inequality unless
r1 = s1 , r2 = s2 , det R = 1, det S = 1.
It follows that R = S SO(2) and that is given by
= tR + (1 t)R U,
i.e., F K (1) . This proves the assertion.

ak. The crucial observaThe general two-well problem was solved by Sver
tion is that any element in the convex hull of K is of the form F = Y + ZH
where Y and Z are conformal matrices. Therefore it is natural to introduce
new coordinates in M22 by

z1 z2
0
y1 y2
+
F =
0
y2 y1
z2 z1
and to write F = (y, z) with y, z R2 .
Theorem 2.2.2. Assume that K is given by (2.15) with 0 < < 1 < .
Then K pc = K qc = K rc = K (3) . Moreover, for det H = 1,

K pc = X = (y, z) : |y| + |z| 1 and det F = 1 ,
while for det H > 1

det X 1
det X 1
.
K pc = X = (y, z) : |y| 1
and |z|
det H 1
det H 1
A dierent formula follows for the case of equal determinant ( = 1) from
the results below.
29
The k-well Problem with Equal Determinant. In the following we assume that
K = SO(2)U1 . . . SO(2)Uk ,
det Ui = > 0, i = 1, . . . , k.
The characterization of the semiconvex hulls in this case relies on the fact
that every pair of wells SO(2)Ui and SO(2)Uj , i = j, is rank-one connected,
see Proposition A.2.5. In order to describe the semiconvex hulls of K, we
= {C = F T F : F K} as a subset of all positive denite
consider the set K
pc , K
qc , and K
rc analogously. It is
and symmetric matrices. We dene K
an immediate consequence of the representations of the semiconvex hulls as
centers of mass of probability measures or as zero sets of semiconvex functions
that all matrices F K pc satisfy
|F e|2 max{|U1 e|2 , . . . , |Uk e|2 }
for all e S1 .
pc is contained in the half space

This implies that, for all e S1 , the set K

H = C symmetric with C : (e e) max{|U1 e|2 , . . . , |Uk e|2 }
in the three-dimensional space of all symmetric matrices. Moreover, if QUi
and Uj are rank-one connected, QUi Uj = a n, then |Ui n |2 = |Uj n |2 ,
and all the matrices F = QUi +(1)Uj on the rank-one segment between
QUi and Uj satisfy |F n |2 = |Ui n |2 = |Uj n |2 . Consequently the matrices
FT F = UjT Uj + UjT a n + n UjT a + 2 |a|2 n n
lie on the hyperplane {C : (n n ) = |Ui n |2 = |Uj n |2 } and satisfy
det(FT F ) = 2 since the determinant is ane on rank-one lines (and hence
constant on the entire rank-one segment since det QUj = det Ui = ). This
hyperplane intersects the surface S of constant determinant 2 in the space of
positive denite symmetric matrices in a one-dimensional curve. This suggests
that the rank-one convex hull of K corresponds to a polygon on S which is
bounded by curves generated from rank-one connections between the wells.
The next theorem asserts that these curves form indeed the boundary of
the semiconvex hulls if the matrices Ui and Uj have a certain maximality
property. This property states that all matrices U lie on one side of the
curve generated by (one of) the two rank-one connections between the wells
SO(2)Ui and SO(2)Uj . This implies that we may relabel the matrices Ui such
that
|U1 n |2 = . . . = |U n | > max |Uj n |2 .
j+1
(2.16)
The second part of the theorem asserts that there exists a nite description of
the hulls of K, in the sense that they are described by a nite number k of
inequalities. These inequalities are related to matrices Ui that correspond to
30
a corner of the hulls, i.e., to matrices for which there exists a vector v S1
such that
|Ui v|2 > max |Uj v|2 .
j=i
Theorem 2.2.3. Assume that U1 , . . . , Uk M22 , k 2, are symmetric

and positive denite matrices with det Ui = > 0 for i = 1, . . . , k, and that
SO(2)Ui = SO(2)Uj for i = j. Let U = {U1 , . . . , Uk },
K = SO(2)U1 . . . SO(2)Uk ,
and dene the set of corners C by

C = U U : v S1 with |U v|2 >
max
V U \{U }

|V v|2 .
(2.17)
Suppose that C contains elements, k. Then

K (2) = K rc = K qc = K pc
and there exists a set E = {e1 , . . . , e } S1 such that

K qc = F : det F = , |F ei |2 max |U ei |2 , i = 1, . . . , .
U C
Proof. Let

A = F : det F = , |F ei |2 max |U ei |2 , i = 1, . . . , .
U C
We rst show that the set A is dened by polyconvex conditions, that is,
K pc A, and then we apply the splitting method to prove that A K (2) .
Therefore K pc A K (2) K pc and all inclusions are equalities. This
establishes the theorem.
In order to demonstrate the inclusion K pc A we dene functions g and
hi : M22 R by
+
g(X) = (det X )2 , hi (X) = |Xei | max |U ei | , i = 1, . . . , ,
U C
where (t)+ = max{t, 0}. Then A is equivalently given by

A = F M22 : g(F ) 0, hi (F ) 0, i = 1, . . . , .
We conclude that K pc A since the functions g and hi are polyconvex with
g(F ) 0 and hi (F ) 0 for F K.
It remains to prove that A K (2) . This is done by showing that A is
compact and that the boundary of A consists of so-called maximal arcs
dened in the following way. Suppose that Ui , Uj C, i = j, and that there
exists an e S1 such that
|Ui e| = |Uj e|
max
U C\{Ui , Uj }
31
|U e|.
Proposition A.2.5 implies the existence of Qi SO(2) and si R such that

Qi Ui Uj = si Uj e e . Then we call the set

ij (e) = F = Q Uj + tsi Uj e e : Q SO(2), t [ 0, 1 ]

a maximal arc with endpoints Ui and Uj . By construction,
Uj + tsj Uj e e = (1 t)Uj + tQi Ui K (1) ,
and hence ij (e) K (1) . We now divide the proof into a series of steps in
which we rst construct the set E in the assertion of the theorem (Step 4)
and show that the boundary of A consists of maximal arcs (Step 6). Then
the assertion follows by an application of the splitting method.
Step 1: Suppose that i, j {1, . . . , k}. If there exists a v S1 and an
> 0 such that
|Ui w| = |Uj w|
for all w with |w v| < ,
then SO(2)Ui = SO(2)Uj .

Indeed, if this equality holds, then |Ui (v + tv )| = |Uj (v + tv )| for t
small enough and thus
|Ui v|2 = |Uj v|2 ,
Ui v, Ui v = Uj v, Uj v ,
|Ui v |2 = |Uj v |2 ,
and hence UiT Ui = UjT Uj . Since det Ui = det Uj , this implies by the polar
decomposition theorem that there exists a Q SO(2) with Ui = QUj .
Step 2: The set C is well-dened and contains at least two matrices, that
is, 2.
We dene for v S1 the set A(v)

A(v) = i {1, . . . , k} : |Ui v| = max |Uj v| .
j=1,...,k
We conclude by Step 1, that A(v) is a singleton for all but nitely many v.
Let

G = i : A(v) = {i} on a (relatively) open subset of S1
C = {Ui , i G}.
We have to show that G contains at least two points. Suppose that G contains
only one element, C = {Ui }. Then |Uj v| |Uj v| for all v S1 , j = i, and
therefore in particular max (Uj ) max (Ui ) and min (Uj ) min (Ui ). Since
det Ui = det Uj , we also have min (Uj ) min (Ui ) and this allows us to
conclude min (Uj ) = min (Ui ) as well as max (Uj ) = max (Ui ). Moreover,
the matrices UiT Ui and UjT Uj have the same system of eigenvectors, and
32
thus SO(2)Ui = SO(2)Uj , a contradiction. If M0 = {U1 , . . . , Uk } \ C = ,

then
|Ui v| max |U v| for all v S1 , Ui M0 ,
U C
and therefore M0 A, independently of the choice of the set C. It is easy to

see that (K M )qc = K qc for all M K qc , we may therefore assume in the
sequel that C = {U1 , . . . Uk }.
Step 3: Every matrix Ui C is the end point of exactly two distinct
maximal arcs. That is, for all Ui C there exist Up , Uq C, p = i, q = i, and
ep , eq S1 , ep = eq such that
|Ui ep | = |Up ep |
max
U C\{Ui , Up }
|U ep |, |Ui eq | = |Uq eq |
max
U C\{Ui , Uq }
|U eq |.
Note that Proposition 2.2.4 below implies that the inequality is necessarily
strict. We rst show that there are at least two maximal arcs with end point
Ui . For the arguments below, it is convenient to identify vectors e S1 with
angles in [ 0, ) by e = v() = (cos , sin ) or e = v(). We dene
g : [ 0, ] R by
g() = |Ui v()|
max
U C\{Ui }
|U v()|.
Then g is continuous and periodic, and since 2 there exist at least two
angles p , q with 0 p < q < and g(p ) = g(q ) = 0. Let ep and eq
be the corresponding vectors. Then
|Ui ep | =
max
U C\{Ui }
|U ep |,
|Ui eq | =
max
U C\{Ui }
|U eq |,
and there exist Up , Uq C \ {Ui } such that

|Ui es | = |Us es |
max
U C\{Ui ,Us }
|U es |,
s = p, q,
as asserted. Assume now that there exist three maximal arcs with end point
Ui , that is, there exist indices p, q, r dierent from i and corresponding vectors es = v(s ), s = p, q, r, with 0 p < q < r < such that
|Ui es | = |Us es |
max
U C\{Ui ,Us }
|U es |,
s = p, q, r.
By Proposition A.2.5, there exist Qs SO(2) and ts R such that

Qs Us Ui = ts Ui es e
s ,
s = p, q, r.
This implies for all v S1 that

2
2
|Us v|2 = |Ui v|2 + 2ts Ui v, Ui es e

s , v + ts |Ui es | es , v .
2
33
We dene P (s, v) = t2s |Ui es |2 e

s , v 0 and obtain
2
|Up er |2 = |Ui er |2 + 2tp Ui er , Ui ep e

p , er + P (p, er ),
|Ur ep |2 = |Ui ep |2 + 2tr Ui ep , Ui er e
r , ep + P (r, ep ),
|Uq er |2 = |Ui er |2 + 2tq Ui er , Ui eq e
q , er + P (q, er ),
|Ur eq |2 = |Ui eq |2 + 2tr Ui eq , Ui er e
r , eq + P (r, eq ),
|Uq ep |2 = |Ui ep |2 + 2tq Ui ep , Ui eq e
q , ep + P (q, ep ),
|Up eq |2 = |Ui eq |2 + 2tp Ui eq , Ui ep e
p , eq + P (p, eq ).
By assumption, |Ui es | always is maximal, and therefore the second term on
the right hand side of the foregoing identities must be less than or equal to
zero. Since
e
p , er 0,
e
r , ep 0,
e
q , er 0,
e
r , eq 0,
e
q , ep 0,
e
p , eq 0,
we obtain the following set of conditions:

tp Ui er , Ui ep 0,
tr Ui ep , Ui er 0,
tq Ui er , Ui eq 0,
tr Ui eq , Ui er 0,
tq Ui ep , Ui eq 0,
tp Ui eq , Ui ep 0.
(2.18a)
(2.18b)
(2.18c)
(2.18d)
(2.18e)
(2.18f)
Suppose now that tp 0 (the case tp 0 is analogous). Then Ui er , Ui ep 0

and tr 0 by (2.18b). Hence Ui eq , Ui er 0 by (2.18d), and we obtain from
(2.18c) that tq 0. This implies by (2.18e) that Ui ep , Ui eq 0, and we
conclude from (2.18f) that tp 0. This is only possible of tp = 0 and we
deduce that
|Up er | = |Ui er | = |Ur er |.
This contradicts the denition of a maximal arc unless p = r. The analogous
chain of implications shows also that q = r and hence p = q = r. In view of
Proposition A.2.5 we nd Qs SO(2) and ts R such that
Qs Up Ui = ts Ui es e
s ,
that is, there exist three dierent rank-one connections between the well
SO(2)Ui and SO(2)Up . However, the equation det(QUp Ui ) = 0 has at
most two solutions for i = p, and this contradiction establishes the assertion
of the step.
34
Step 4: Denition of E.
The results in Step 3 allow us to dene a graph G of degree two (that
is, all nodes are end points of exactly two edges in the graph) with nodes
corresponding to the matrices in C and edges corresponding to the maximal
arcs. Two nodes Ui , Uj C are connected by an edge if and only if Ui and
Uj are the endpoints of a maximal arc. It is easy to see that G is the union of
a nite number of disjoint cycles. We assert now that G consists of a single
cycle. Otherwise we choose one cycle in the graph and denote the set of
indices corresponding to the matrices in this cycle by M1 . By assumption,
M2 = {1, . . . , } \ M1 = and we dene
g(e) = max |Uj e| max |Uj e|.
jM1
jM2
By denition of C, there exist e such that g(e+ ) > 0 and g(e ) < 0. The
continuity of g implies the existence of a vector e with g(e) = 0. Hence
max |Uj e| = max |Uj e|.
jM1
jM2
Let p M1 and q M2 be indices such that

|Up e| = |Uq e|
max
U C\{Up , Uq }
|U e|.
Then Cpq (e) is a maximal arc and consequently there exists at least three
maximal arcs with end points Up and Uq . This contradicts the assertion of
Step 3. The graph G consists therefore of a single cycle with edges corresponding to maximal arcs given by distinct vectors e1 , . . . , e S1 . We
dene E = {e1 , . . . , e }.
Step 5: The set A is compact.
If follows from Step 3 that 2 and from Step 4 that A contains at
least two conditions of the form |F ei | Ci with linearly independent vectors
2 = e2 e1 , e2 e1 and dene an orthonormal basis with
e1 and e2 . Let v
#2 /|#
v 1 = e1 and v 2 = v
v 2 |. Then
|F v 1 | C1 ,
|F v 2 |
C1 + C2
for all F A.
|#
e2 |
Thus the coordinates of the vectors formed by the rows of F are uniformly
bounded in the orthonormal basis {v 1 , v 2 } and hence A is bounded. Since all
functions in the denition of A are continuous, A is closed and thus compact.
Step 6: The set of constrained points,

B = F A : e S1 such that |F e| = max |U e|
U C
is the union of the maximal arcs.

As a rst step, we show that B is in fact given by
35

B = F A : i {1, . . . , } such that |F ei | = max |U ei | .
(2.19)
U C
Assume thus that F B \ K and that there exists a Uq C with

|F e| = |Uq e|
max
U C\{Uq }
|U e|
with e E.
(2.20)
By Step 3, Uq is the end point of two maximal arcs pq (ep ) and qr (eq ). By
construction, ep and eq are not parallel and therefore e
p , eq = 0. Moreover,
e E and thus e , ep = 0 and e , eq = 0. We nd by Proposition A.2.5
rotations QF , Qp , Qr SO(2) and scalars tF , tp , tr R with tp = 0 and
tq = 0 such that
QF F Uq = tF Uq e e ,
Qp Up Uq = tp Uq ep e
p,
(2.21b)
e
q .
(2.21c)
Qr Ur Uq = tr Uq eq
(2.21a)
We now assume that tF = 0. If we deduce a contradiction to this assumption,

then F = QTF Uq K. Thus the constraints with e E are not active, and
we establish that the constrained points are given by (2.19).
We rst multiply (2.21a) by ep and eq and take the modulus. Since F A
and pq (ep ) and qr (eq ) are maximal arcs, the inequalities
|F ep |2 = |Uq ep |2 + 2tF Uq ep , Uq ee , ep + t2F |Uq e|2 e , ep ,
2
|F eq |2 = |Uq eq |2 + 2tF Uq eq , Uq ee , eq + t2F |Uq e|2 e , eq
imply that
2tF Uq ep , Uq ee , ep < 0,
2tF Uq eq , Uq ee , eq < 0.
(2.22)
Similarly,
2
2
|Up e|2 = |Uq e|2 + 2tp Uq e, Uq ep e
p , e + tp |Uq ep | ep , e ,
2
2
2
|Ur e|2 = |Uq e|2 + 2tr Uq e, Uq eq e
q , e + tr |Uq eq | eq , e ,
2
and by (2.20)
2tr Uq e, Uq eq e
q , e < 0,
2tp Uq e, Uq ep e
p , e < 0.
(2.23)
Finally,
2
2
|Up eq |2 = |Uq eq |2 + 2tp Uq eq , Uq ep e
p , eq + tp |Uq ep | ep , eq ,
2
2
2
|Ur ep |2 = |Uq ep |2 + 2tr Uq ep , Uq eq e
q , ep + tr |Uq eq | eq , ep ,
2
and since pq (ep ) and qr (eq ) are maximal arcs
36
2tr Uq ep , Uq eq e
q , ep < 0,
2tp Uq eq , Uq ep e
p , eq < 0.
(2.24)
It follows from e
p , e = e , ep and (2.22), (2.23) that
tF Uq ep , Uq ee , ep < 0,
tp Uq ep , Uq ee , ep < 0.
We get similarly from (2.22) and (2.23)

tF Uq eq , Uq ee , eq < 0,
tr Uq eq , Uq ee , eq < 0,
and nally from (2.24)

tr Uq ep , Uq eq e
q , ep < 0,
tp Uq ep , Uq eq e
q , ep < 0.
The foregoing estimates yield

tF tp < 0,
tF tr < 0,
tp tr < 0,
and this set of inequalities has no solution. Consequently the assumption

tF = 0 cannot hold, and this establishes the formula (2.19) for the constrained
points.
We may thus assume that F B \ K with
|F eq | = max |U eq |
U C
with eq E.
By denition of E there exist two matrices Uq , Ur C such that qr (eq ) is

a maximal arc with |F eq | = |Uq eq | = |Ur eq |. Moreover, there are indices
p, s {1, . . . , }, p = q, s = r and vectors ep , er C such that pq (ep ) and
rs (er ) are maximal arcs. By Proposition A.2.5 we nd Qr , QF SO(2) and
tr , tF R, tr = 0, tF = 0, such that
Qr Ur Uq = tr Uq eq e
q ,
QF F Uq = tF Uq eq e
q ,
(2.25)
F SO(2) and
q , Q
tq ,
tF R,
tq =
0,
tF =
0 with
as well as Q
q Uq Ur =
tq Ur eq e
Q
q ,
F F Ur =
Q
tF Ur eq e
q .
(2.26)
We show below that

tF [ 0, tr ] if tr > 0,
tF [ tr , 0 ] if tr < 0.
(2.27)
This implies tF = tr , = tF /tr [ 0, 1 ], and thus

(1)
.
QF F = Uq + tr Uq eq e
q = (1 )Uq + Qr Ur K
(2.28)
In order to prove (2.27), we rst observe that by (2.25)

2
2
|Ur ep |2 = |Uq ep |2 + 2 tr Uq ep , Uq eq e
q , ep + tr |Uq eq | eq , ep ,
2
2
2
|F ep |2 = |Uq ep |2 + 2 tF Uq ep , Uq eq e
q , ep + tF |Uq eq | eq , ep .
2
37
Since pq (ep ) is a maximal arc,

|Uq ep | |Ur ep |,
|Uq ep | |F ep |,
and we deduce
2
2
2 tr Uq ep , Uq eq e
q , ep + tr |Uq eq | eq , ep 0,
2
2
2
2 tF Uq ep , Uq eq e
q , ep + tF |Uq eq | eq , ep 0.
2
By assumption, the vectors ep and eq are not parallel, hence e

q , ep = 0.
This implies
2 tr Uq ep , Uq eq e
q , ep < 0,
2 tF Uq ep , Uq eq e
q , ep < 0,
and we infer that tF tr > 0, that is, tF and tr have the same sign. The same
arguments applied to the identity (2.26) with er instead of ep show that

tq > 0. It follows from the proof of Proposition 2.2.4 below that
tF

tF = tF t r ,

tq = tr ,
see formulae (2.30) and (2.31) with 2 = tr , j = tF ,

1 =
tq , and
j =
tF .
We obtain that
tF tr > 0,
(tF tr ) tr > 0.
If tr > 0, then tF > 0 and (tF tr ) > 0, hence tF tr and therefore

tr [ 0, tF ]. We conclude similarly if tr < 0 and this nishes the proof of
Step 6.
Step 7: The set A is contained in the lamination convex hull of K. More
precisely, A K (2) .
This is an immediate consequence of Step 6 which implies that the set
of all constrained points is a subset of K (1) . If F A is an unconstrained
point, then we apply the splitting method to prove that F K (2) . Choose
any rank-one curve t F (t) = F (I + tv v ) with v S1 and dene
t = sup{t < 0 : e E : |F (t)e|2 = max |U e|2 ,
U C
t = inf {t > 0 : e E : |F (t)e| = max |U e|2 .

+
U C
Since A is compact, the parameters t are nite. Let F = F (t ). Then

F B K (1) , F + F is a matrix of rank-one, and F is contained in the
rank-one segment between F + and F . Thus F K (2) and this concludes
the proof of the theorem.

The proof of Theorem 2.2.3 used the following fact.
38
Proposition 2.2.4. Assume that det Ui = > 0 for i = 1, . . . , with 3,

that SO(2)Ui = SO(2)Uj for i = j, and that there exists an e S1 such that
|U1 e|2 = = |U e|2 .
Then we may relabel the matrices in such a way that there exist j R and
Qj SO(2), j = 2, . . . , , such that 0 < 2 < . . . < and
Qj Uj U1 = j U1 e e ,
j = 2, . . . , .
In particular,
lc
{U2 , . . . , U1 } SO(2)U1 SO(2)U .

Proof. By Proposition A.2.5 there exist i = 0 and Qi SO(2) such that
Qi Ui U1 = i U1 e e .
We rst show that we may assume that 2 < < and j = 0. If
j = j+1 , j {2, . . . , 1}, then SO(2)Uj = SO(2)Uj+1 , and j = 0 implies
that SO(2)U1 = SO(2)Uj . Both identities contradict our assumption that the
wells SO(2)Uj are pairwise disjoint. If 2 > 0, then we dene j (0, 1) by
1 j =
j
(0, 1),
j = 2, . . . , 1.
We obtain

j U1 + (1 j )Q U = j U1 + (1 j ) U1 + U1 e e
= U1 + j U1 e e = Qj Uj .
Hence
lc
Uj SO(2)U1 SO(2)U ,
and the assertion of the proposition is immediate. If 2 < 0 then we use
i SO(2)
Proposition A.2.5 for U2 instead of for U1 and we nd
i = 0 and Q
with
i Ui U2 =
i U2 e e ,
Q
i = 1, 3, . . . , .
By denition,
Q2 U2 U1 = 2 U1 e e ,
1 U1 U2 =
Q
1 U2 e e ,
and thus

1 U1 e e .
1 Q2 U2 U2 +
1 U2 e e = 2 Q
Q
(2.29)
39
1 Q2 U2 U2 )e = 0
Multiplication of (2.29) with e from the right yields (Q
1 Q2 U2 U2 = 0 since there are no rank-one connecwhich is equivalent to Q
1 U1 e = U2 e that (
1 + 2 )U2 e = 0
tions in SO(2). We thus obtain from Q
and hence
1 = 2 > 0.
(2.30)
Similarly, for j 3,
and
j Uj U2 =
j U2 e e ,
Q
1 U1 U2 =
Q
1 U2 e e .
These equalities yield as before

1 U1
j Uj Q
j U2 e e
1 U2 e e =
Q
and
j Uj Q
1 Qj Uj = j Q
1 U1 e e + (
Q
j
1 )U2 e e .
j
1 = 0 and therefore
We conclude j +
1 + j = 2 + j 0.
j =
(2.31)
This proves the assertion.

We include two typical examples to illustrate Theorem 2.2.3. The threewell conguration is motivated by the analysis of microstructures in cubic to
orthorhombic transitions, see Section 5.3. The four-well problem arises as a
special case of a tetragonal to monoclinic transition in Section 5.4.
Examples. 1) A three-well problem. For > > 0 we dene
"
!
!
"

3
3
+ 0
2
2
2
2
U1 =
, U3 =
.
, U2 =
3
0
23 + 2
+ 2
2
A short calculation shows that the matrices
prop Ui satisfy the extremality
erty (2.17) with v 1 = (1, 0), v 2 = 12 (1, 3) and v 3 = 12 (1, 3) (see below).
The rank-one connections between the wells are given by Qi U1 U2 = ai ni ,
i = 1, 2, with

2
3
a1 =
=
,
n
,
1
3
3
2( 2 + 2 )
and
a2 =
2
2( + + 2 )
3
( + 2)

3
n2 =
;
3
40
moreover, Qi U1 U3 = ai ni , i = 3, 4, with

,
a3 =
2( 2 + 2 ) 3
and
a4 =
2
2( + + 2 )
3
+ 2

n3 =
n4 =

3
;
3
nally, Qi U2 U3 = ai ni , i = 5, 6, with

3(2 )
,
a5 = 2
3
+ 2
and
a6 = 2
+ + 2
3
3(2 + )
3
3

0
,
n5 =
1

1
,
n6 =
0
see Figure 2.2 for a sketch of the situation. Since

|U1 n1 |2 = |U2 n1 |2 = 2 + + 2 > ( )2 = |U3 n1 |2 ,
|U1 n2 |2 = |U2 n2 |2 = 2 + 2 < ( + )2 = |U3 n2 |2 ,
the extremal curve between U1 and U2 is dened by n1 . Also
|U1 n3 |2 = |U3 n3 |2 = 2 + + 2 > ( )2 = |U2 n3 |2 ,
|U1 n4 |2 = |U3 n4 |2 = 2 + 2 < ( + )2 = |U2 n4 |2 ,
and
|U2 n5 |2 = |U3 n5 |2 = 2 + + 2 > ( )2 = |U1 n5 |2 ,
|U2 n6 |2 = |U3 n6 |2 = 2 + 2 < ( + )2 = |U1 n6 |2 ,
and therefore the extremal curves between U1 and U3 and U2 and U3 are
given by n3 and n5 , respectively. Thus the semiconvex hulls of this threewell problem are described by Theorem 2.2.3 with E3 = {n1 , n3 , n5 }.
In view of the application to the cubic to orthorhombic transformation
in Section 5.3, we consider the simple laminates that can be formed with
the rank-one connections between the matrices in K that do not correspond
to maximal arcs. The question we want to address is whether the center of
mass F = , id has more than one representation as a simple laminate. By
symmetry, it suces to consider the rank-one connection between U1 and U2
with Q1 U1 U2 = a1 n1 . We dene
Ft = U2 + (1 t)a1 n1
C3
n3
n5
C2
41
C1
n1
Fig. 2.2. The three-well problem in two dimensions. The left gure shows the
matrices Ci = UiT Ui and the normals ni dening the extremal curves. The right
gure (generated by a ray tracing programme) displays the convex hulls for = 2
and = 1.5 on the surface det C = 2 . In this case, the quasiconvex hull of K is
given by the union of the quasiconvex hull of the two-well problems formed with
pairs of wells in K.
and barycenters of laminates in the other twinning systems that do not correspond to maximal arcs,
Gs = U3 + (1 s)a3 n3 ,
Hs = U3 + (1 s)a5 n5 .
We seek solutions t13 , s13 and t23 , s23 of

FtT Ft = GTs Gs
and
FtT Ft = HsT Hs ,
respectively. A short calculation shows that

t13 =
4 ,
3
s13 =
4 ,
3
t23 =
+ 1 ,
3
s23 =
4 .
3
and
This shows that a simple laminate supported on SO(2)U1 and SO(2)U2 is

uniquely determined from its center of mass F = (1 )Q1 U1 + U2 if and
only if is not equal to one of the two values
1 =
4 ,
3
2 =
+ 1 .
3
Since
1 2
5
+ ,
42
n3
C1
C3
e1
e2
C4
C2
n4
Fig. 2.3. The four-well problem. The left gure shows in a schematic sketch the
matrices Ci = UiT Ui and the normals dening the extremal curves. Here {e1 , e2 }
denotes the standard basis of R2 and n3 and n4 are parallel to e1 + e2 and e1 e2 ,
respectively. The right gure displays the semiconvex hulls for = 2, = 1.5 and
= 0.5 on the surface det C = 2 .
we obtain that
K qc =
SO(2)Ui SO(2)Uj
qc
i=j
+ .
2) A four-well problem. For > > 0, > 0 with 2 > 0 we dene

, U2 =
, U3 =
, U4 =
.
U1 =

As in the case of the three-well problem, it is easy to see that the matrices
Ui satisfy the maximality property (2.17). For example, taking
e parallel to
(1 + , 1) for > 0 small enough, one obtains

|U1
e|2 > max |U3
e|2 , |U4
e|2 , and |U1
e|2 > |U2
e|2 2 > 2 .
A short calculation shows that there exist rotations Qi and vectors ai R2
such that
Q1 U1 U2 = a1 e1 ,
Q7 U3 U4 = a7 e1 ,
Q2 U1 U2 = a2 e2 ,
Q8 U3 U4 = a8 e2 ,
where {e1 , e2 } denotes the standard basis of R2 ,

1
Q3 U1 U3 = a3 (e1 + e2 ),
2
1
Q4 U1 U3 = a4 (e1 e2 ),
2
1
Q5 U2 U4 = a5 (e1 e2 ),
2
1
Q6 U2 U4 = a6 (e1 + e2 ),
2
and
see Figure 2.3 for a sketch of the extremal curves.
43
The General Theorem. The results for k wells can easily be generalized
to any compact set of matrices with xed determinant.
Theorem 2.2.5. Assume that 0, that K {F M22 : det F = }
is compact, and that K is invariant under SO(2), i.e.,

K = SO(2)K = F = QU : Q SO(2), U K .
Then K (2) = K rc = K qc = K pc and

K qc = F M22 : det F = , |F e|2 max |U e|2 for all e S1 . (2.32)
U K
Proof. Let A denote the formula for the hulls in (2.32). Since A is dened
by polyconvex conditions, it is clear that K pc A and it remains to show
that A K (2) . This follows immediately form the splitting method if we
prove that F K (1) whenever F A is a constrained point, i.e., F satises
equality in one of the inequalities in the denition of A. We consider rst the
case > 0 and mention the necessary modications for = 0 at the end of
the proof. Assume thus that F K and that for at least one e S1
|F e|2 = max |U e|2 = |U1 e|2 , U1 K.
U K
By Proposition A.2.5 there exist 1 R and Q1 SO(2) such that

Q1 U1 F = 1 F e e .
We may assume that 1 > 0 (the proof for 1 < 0 is analogous). The idea is
to show by contradiction that there exist U2 K, Q2 SO(2) and 2 < 0
such that
Q2 U2 F = 2 F e e .
This implies
F =
2
1
Q1 U1 +
Q2 U2 ,
1 2
1 2
(1)
and thus F SO(2)U1 SO(2)U2
. In order to show this, assume that
> 0 is small enough such that B (F ) K = . By compactness, there exists
a c0 > 0 with
dist(F + sF e e , K) > c0 for all s < 0 such that F + sF e e B (F ),
and by continuity we may choose t0 > 0 such that
dist(F + sF et e
t , K) >
c0
2
(2.33)
44
for all s < 0 such that F + sF et e

t B (F ), |t| < t0 , where et = e + te .
We have et , e = Jet , e = et , Je = t, where J denotes the counterclockwise rotation by 2 , and hence

2
2
2 2
2
|(F + sF et e
t )e| = |F e| 2stF e, F et + s t |F et |

= |F e|2 2st |F e|2 + tF e, F e + s2 t2 |F et |2 .
Therefore we can nd a t1 (0, t0 ) such that

2
2
|(F + sF et e
t )e| |F e|
(2.34)
for all t [ t1 , 0), s > 0 with F + sF et e

t B (F ) with equality if and
only if s = 0. We now assert that
|F et |2 > max |U et |2 for t [ t1 , 0).
U K
(2.35)
Assume the contrary. Suppose rst that there exists a t [ t1 , 0) and a

U0 K such that |F et |2 = |U0 et |2 . By Proposition A.2.5 we may choose
Q0 SO(2) and 0 R with
Q0 U0 F = 0 F et e
t .
(2.36)
Clearly U0 B (F ) since B (F ) K = and inequality (2.33) implies that

0 > 0. We obtain from (2.34) and (2.36)
2
2
2
|Q0 U0 e|2 = |(F + 0 F et e
t )e| > |F e| = max |U e| ,
U K
a contradiction. It remains to consider the case

|F et |2 < max |U et |2 ,
U K
t [ t1 , 0).
We may choose an increasing sequence tk < 0, tk 0 and Uk = Utk K

with
|F ek |2 < |Uk ek |2 = max |U ek |2
U K
(2.37)
for all k, where ek = e + tk e . By compactness of K we may assume that

Uk U and therefore
lim |F ek |2 lim |Uk ek |2 = |U e|2 .
By Proposition A.2.5 there exists R and Q SO(2) with

Q U F = F e e .
(2.38)
By our assumptions > 0, and we assert next that we may nd k R,

Qk SO(2) and v k = e + k e R2 such that
Qk Uk F = k F v k v
k,
k ,
with
Qk Q,
45
v k e.
A short calculation shows that there exist always two rank-one connections
between SO(2)Uk and F ,
(i)
(i)
(i)
Qk Uk F = ak bk ,
i = 1, 2,
(2.39)
(i)
for which the rotations Qk do not have (for a suitable subsequence) the same
limit. We may pass to the limit in (2.39) and obtain
(i)
(i)
Q U F = a(i) b ,
(1)
(2)
= Q
and comparison with (2.38) establishes the assertion for a suitable subsequence since there are only two rank-one connections between SO(2)U . In
view of (2.34) we must have k > 0. We deduce
2
|Uk ek |2 = |F ek + k v
k , ek F v k |
2
2
= |F ek |2 + 2k v
k , ek F ek , F v k + k v k , ek |F v k |
2
and this contradicts (2.37) since k > 0 and tk < 0 imply
v
k , ek = e k e, e + tk e = tk k < 0,
F ek , F v k > 0
for k big enough. This establishes (2.35) and hence a contradiction to the
assumption F A. This concludes the proof of the theorem for > 0.
Assume now that = 0 and that K is thus contained in the rank-one
cone {X : det X = 0}. Suppose that F A is a constrained point, i.e., there
exists an e S1 such that
|F e|2 = max |U e|2 = |U1 e|2 ,
U K
U1 K.
If |F e|2 = 0, then |U1 e|2 = 0 and we may assume that F = se e and

U1 = te e with t s 0. Since U1 and U1 are contained in K, we
conclude F K (1) .
We now consider the case |F e|2 = 0. In this case the argument is identical
to the one used for > 0 since F e = 0 implies that F et = 0 with et = e+te
for t small enough. This concludes the proof of the theorem.

An SO(2) Invariant Set with K rc = K pc . The next example shows the
drastic dierences in the complexity of hulls of sets of matrices with equal
determinant and of sets without this constraint.
We assume in the following that > > 1 or that 1 > > > 0 and
that U1 and U2 M22 are given by

0
0
U1 =
, U2 =
.
0
0
46
Let
K = SO(2) SO(2)U1 SO(2)U2 .
(2.40)
Then any Mpc (K) can be represented as

= 0 + 1 1 U1 + 2 2 U2
(2.41)
with 0 , 1 , 2 [ 0, 1 ], 0 + 1 + 2 = 1, and , 1 , 2 P(SO(2)). Here

i Ui denotes the measure given by i Ui (E) = i (EUi1 ) where the set EUi1
is dened by EUi1 = {F : F Ui E}. The following theorems show that the
rank-one convex hull of K is given by the union of SO(2) and the rank-one
convex hull of the two compatible wells SO(2)U1 and SO(2)U2 . Every element
in K rc is therefore given by a second order laminate. Contrary to this, the
polyconvex hull is always considerably bigger than the rank-one convex hull.
However, if Mpc (K) satises F = , id K pc \ K rc , then the mass
of the support of on SO(2)U1 SO(2)U2 must be bigger than a constant
which depends only on and , see Theorem 2.2.6 below. It is exactly this
property which together with Proposition 2.1.5 allows us to prove that the
rank-one convex hull is given by K rc = SO(2) (SO(2)U1 SO(2)U2 )rc .
The following theorem summarizes our results.
Theorem 2.2.6. Suppose that either > > 1 or 1 > > > 0 and that
K is given by (2.40). Let
=
( )2
(0, 1).
( )2 + 4( 1)( 1)
Then there exists a polyconvex measure Mpc (K) with the representation
(2.41) and 0 {0, 1} if and only if 0 (0, ]. In particular,
K pc SO(2) (SO(2)U1 SO(2)U2 )pc .
Moreover, Mrc (K) does not contain any measure that is supported on SO(2)
and SO(2)U1 SO(2)U2 , i.e.,
Mrc (K) = Mrc (SO(2)) Mrc (SO(2)U1 SO(2)U2 ),
and in particular
rc
K rc = SO(2) SO(2)U1 SO(2)U2 .
Remark 2.2.7. For 0 = one obtains
F = I +
1
2
(U1 + U2 ) =
I K pc \ K rc .
2
+2
It is an open problem to nd a formula for K pc and K qc .
47
Proof. We begin with the assertion for polyconvex measures supported

on K. The idea is to nd R, S1 , S2 conv(SO(2)) and 0 , 1 , 2 (0, 1)
such that the minors relations hold, i.e.,
F = 0 R + 1 S1 U1 + 2 S2 U2 ,
det F = 0 + (1 + 2 ).
Then for arbitrary , 1 , 2 P(SO(2)) with , id = R, 1 , id = S1 , and

2 , id = S2 , the measure dened by = 0 + 1 1 U1 + 2 2 U2 belongs
to Mpc (K). Let

r1 r2
s1 s2
t1 t2
R=
, S1 =
, S2 =
.
r2 r1
s2 s1
t2 t1
It follows from (A.2) that R, S1 , S2 conv(SO(2)) if r12 + r22 1, s21 + s22 1,
and t21 + t22 1. We have to solve the minors relation
det F = 0 + (1 0 ),
where F is given by

0 r1 + 1 s1 + 2 t1 0 r2 1 s2 2 t2
F =
.
0 r2 + 1 s2 + 2 t2 0 r1 + 1 s1 + 2 t1
A short calculation shows that this is equivalent to
0 + (1 + 2 ) = 20 (r12 + r22 ) + 21 (s21 + s22 ) + 22 (t21 + t22 )

+ 0 1 (r1 s1 + r2 s2 ) + 2 (r1 t1 + r2 t2 ) ( + )
+ 1 2 (s1 t1 + s2 t2 )(2 + 2 ).
The right hand side is quadratic in the coecients of R, S1 and S2 and
can therefore be made arbitrarily small. In order to show the existence of a
solution of this equation we only need to derive conditions which imply that
the right hand side can in fact be made greater than or equal to the left hand
side. The right hand side is maximal if R = S1 = S2 SO(2) and we have to
nd 0 , 1 and 2 such that
0 + (1 + 2 )
20 + (21 + 22 ) + 0 (1 + 2 )( + ) + 1 2 (2 + 2 ).
This is equivalent to
0 (1 0 )( 1)( 1) + 1 2 ( )2 0.
(2.42)
Let 0 = 1 with > 0. The left hand side is maximal for 1 = 2 = 2 . It

follows that the inequality holds for 0 1 where = 0 is the solution of
(1 )( 1)( 1) +
2
( )2 = 0.
4
48
This implies
=
4( 1)( 1)
( )2 + 4( 1)( 1)
and it follows easily that 0 (0, ] is a necessary condition. On the other

hand, if 0 (0, ] then inequality (2.42) holds and we may dene volume
fractions 1 = 2 = (1 0 )/2 (0, 1) and r [ 0, 1 ] by
r2 =
0 + (1 + 2 )
.
20 + (21 + 22 ) + 0 (1 + 2 )( + ) + 1 2 (2 + 2 )
Consequently R = S1 = S2 = rI conv(SO(2)) and if , 1 , and 2 are

probability measures supported on SO(2) with , id = R, 1 , id = S1 ,
2 , id = S2 , then
= 0 I +

1 0
1 U1 + 2 U2 Mpc .
2
This proves the assertion about the polyconvex hull.

We now turn towards the proof that the rank-one convex hull is the union
of SO(2) with the rank-one convex hull of the two wells. As a rst step we show
that the fact that (nontrivial) polyconvex measures have minimal support on
SO(2)U1 SO(2)U2 implies that
dist(K pc \ SO(2), SO(2)) > 0.
(2.43)
Assume the contrary. Then there exists a sequence F (n) K pc with

(n)
(n)
(n)
(n)
(n)
F (n) = 0 R(n) + 1 S1 U1 + 2 S2 U2 SO(2)

and dist(F (n) , SO(2)) 0. It follows from the properties of the polyconvex
measures proven above that
(n)
(n)
1 + 2
and therefore we may choose a subsequence (again denoted by n) such that

+
F (n) F =
0 R
1 S1 + 2 S2 SO(2)
with
0 + 1 + 2 = 1, and 1 + 2 1 . On the other hand, the
minors relations for F (n) imply
1 = det F = lim det F (n)

n
(n)

(n)
(n)
= lim 0 + (1 + 2 )
0 + (1 + 2 ) = 1,
n
since = 1. This is a contradiction, and we conclude that (2.43) holds.
2.3 The Thin Film Case
49
The proof of the theorem is now a consequence of Proposition 2.1.5. In

view of (2.43) we may choose compact sets C1 and C2 such that
SO(2) C1 , K rc \ SO(2) C2 , C1 C2 = .
Then (A.5) implies
K rc = (K C1 )rc (K C2 )rc = SO(2) (SO(2)U1 SO(2)U2 )rc . (2.44)
We conclude from (2.44) that det F {1, } for all F K rc . Assume now
that Mrc (K) is given by = 0 +1 1 U1 +2 2 U2 with 0 {0, 1}. By
the minors relation for the determinant, det F = 0 + (1 0 ) {1, },
a contradiction. This proves the theorem.

2.3 The Thin Film Case

Bhattacharya and James used -convergence methods to derive a limiting
theory for martensitic thin lms. In their approach, the sets relevant for the
description of the dierent phases are given by
K = O(2, 3)U1 . . . O(2, 3)Uk ,
(2.45)
if the normal to the lm is suitably oriented with respect to the crystalline

axes. Here O(2, 3) denotes the set of all isometries of the plane into the threedimensional space, i.e., the set of all 3 2 matrices with F T F = I, and the
2 2 matrices Ui are positive denite and satisfy det Ui = for i = 1, . . . , k.
We dene
% : M22 M32 by
F11 F12
%(F ) = F21 F22 .

(2.46)
0 0
It is easy to see that

K = SO(3)%
O(2)U1 . . . O(2)Uk ,
and therefore it is natural to consider rst sets invariant under O(2). Their
semiconvex hulls have open interior and are therefore much bigger than the
hulls for SO(2) invariant sets. This is due to the remarkable fact that any pair
of proper and improper rotations is rank-one connected in two dimensions.
Theorem 2.3.1. Assume that > 0, that K {F M22 : | det F | = }
is compact, and that K is invariant under O(2), i.e.,

K = O(2)K = F = QU : Q O(2), U K .
Then K (3) = K rc = K qc = K pc and

K qc = F M22 : | det F | , |F e|2 max |U e|2 e S1 .
U K
(2.47)
50
Proof. Let A denote the formula for the hulls in (2.47). Clearly A is a
polyconvex set, and it suces to prove that F K (2) for all F A with
det F 0 and |F e|2 = maxU K |U e|2 for at least one e S1 . The idea is
= det F for all U
K
such that the
K (1) with det U
to dene a set K
following assertions hold:
v|2 for all v S1 ,
|F v|2 max |U
K

U
e|2 .
|F e|2 = max |U
K

U
(2.48)
takes advantage of the remarkable fact that

The construction of the set K
rank(I R) = 1 for all R O (2). This fact motivates us to choose for all
w S1 a matrix Uw K with det Uw = and |Uw w|2 = maxU K |U w|2 .
w with det U
w = det F by
We then dene a matrix U
w = (1 w )Uw + w Qw Uw = Uw
U
det Uw det F
U T w w ,
T 2 w
(det Uw )|Uw
w |
where w [ 0, 1 ] and Qw O (2) are given by

w =
det Uw det F
,
2 det Uw
Qw = I 2
T
T
w
w
Uw
Uw
.
T
T
|Uw w | |Uw w |
Let

= QU
w : w S1 , Q SO(2) K (1) ,
K
and
M (v) = max |U v|2
U K
and
#(v) = max |U
v|2 .
M
K

U
= QU
w K

#(v) for all v S1 . In fact, if U
We rst assert that M (v) = M
1
with Q O(2) and w S , then by denition

w v|2 = | (1 w )Uw + w Qw Uw v|2 |Uw v|2 M (v)

v|2 = |U
|U
#(v) M (v). Similarly, M
#(v) |U
v v|2 = |Uv v|2 = M (v)
and therefore M
and this implies the assertion. This establishes (2.48) and we deduce from
(1) K (2) . This concludes the proof
the proof of Theorem 2.2.5 that F K
of the theorem.

We now present the result for thin lms.
Corollary 2.3.2. Assume that K is given by (2.45). Then
K (3) = K rc = K qc = K pc ,
and

K qc = F M32 : det(F T F ) 2 , |F e|2 max |U e|2 e S1 . (2.49)
U K
2.4 An Optimal Taylor Bound
51
Proof. Let A be the set in (2.49). For F M32 we dene adjij (F ) to be

the determinant of the (2 2)-matrix formed by the i-th and the j-th row of
F . Since
det(F T F ) = adj212 (F ) + adj223 (F ) + adj231 (F )
is a polyconvex function, we conclude that A is a polyconvex set and thus
K pc A. On the other hand, K is invariant under multiplication with elements in SO(3) from the left, i.e. SO(3)K = K, and for all F A there
exists a Q SO(3) such that QF =
%(F%) where F% = (F T F )1/2 and
% is the
2
T
%
embedding dened in (2.46). By denition, (det F ) = det(F F ) 2 and
|F%e|2 = |F e|2 for all e S1 , and therefore by Theorem 2.3.1
% lc ,
F% K
% = O(2)U1 . . . O(2)Uk .
K
The assertion of the corollary follows since

%(O(2)) O(2, 3).

2.4 An Optimal Taylor Bound

In this section, we consider a situation where additional invariances are generated by a dierent mechanism: intersections of sets rather than unions of sets.
This is typically the case in the analysis of bounds for the eective behavior
of polycrystals. As an example, we study a so-called nonlinear two-variant
elastic material as dened by Kohn and Niethammer [KN00] and we refer to
their paper for a general discussion of polycrystals and Taylor bounds. We
briey summarize the underlying ideas. For a single crystal (in some standard
orientation), the quasiconvex hull K qc describes the ane boundary conditions for which the eective energy is zero. A polycrystal on the other hand
consists of dierently oriented grains (which are themselves single crystals)
and the zero set of the eective energy for a grain rotated by Q SO(2) with
respect to the standard orientation is given by QK qc QT . An inner bound for
the set of ane boundary conditions for which the eective energy of the
polycrystal is zero, is therefore the intersection of all these sets,
&
1
qc
1
T =
SO(2) f f + f f SO(2) f f + f f
.
1
f S
This bound is called the Taylor bound and it ensures that the ane deformation can be accommodated in each grain individually; it does not take into
account compensation eects between dierent grains.
Theorem 2.4.1. Assume that K = SO(2)U1 SO(2)U2 describes a twovariant elastic material with U1 = diag(, 1 ) and U2 = diag( 1 , ). Then

1
T = F M22 : i (F ) [ , ], det F = 1 ,
52
where

1/2
1
,
= B + B 2 3
3
with B = 2 +
1
.
2
Proof. We dene for f S1
1

1
K(f ) = SO(2) f f + f f SO(2) f f + f f .
With this denition, it is easy to see that for all Q SO(2)

F K(f )pc
QF QT K(Qf )pc .
We will rst show that T is invariant under SO(2), i.e., RF Q T for all
R, Q SO(2) and F T . Indeed, if F T , then F K pc (f ) for all
f S1 , and since the sets K(f )pc are invariant under SO(2), we have that
QF K(f )pc for all f S1 and Q SO(2). Thus QF T . It thus suces
to show that QF QT T for all Q SO(2). Let us suppose that F T , and
that there exists a Q SO(2) such that QF QT T . Then there exists an
f S1 such that QF QT K(f )pc , i.e.,

1
pc
1
.
QF Q SO(2) f f + f f SO(2) f f + f f
This implies that F QT K(f )pc Q = K(Qf )pc , contradicting the assumption that F T .
We show next that is the maximal strain that can be recovered in any
basis in a single crystal in standard orientation. To do so, we rst consider
the quasiconvex hull for the grain in its standard orientation in some detail.
It follows from Theorem 2.2.3 that for K = K(e1 )
' '2

1
1 '
'
qc
22
K = F M
: det F = 1, 'F
' 2 + 2 .
1
As a rst step, we determine the direction of maximal rigidity of the grain

and the corresponding maximal principle strain, i.e., for a given basis {f , f }
of R2 we calculate the maximal strain = (f ) 1 such that
f f +
1
f f K qc .
We then minimize (f ) in f S1 , i.e., we consider the problem of nding

'

1 '2

1
1
'
'
min1 max : ' f f + f f

' 2 + 2 .
1
f S
If we write f = (cos , sin ), then this is equivalent to
2.5 Dimensional Reduction in Three Dimensions
2 +
53
1
1
1
2 2 sin cos 2 + 2 ,
2
and therefore, for xed , the maximal has to satisfy

2 +
1
1
1
+ 2 2 | sin cos | 2 + 2 .
2
Let A = | sin cos | [ 0, 12 ] and B = 2 + 12 . The solution for xed A and

B is given by
B B 2 + 4A2 4
2
(A) =
2(1 + A)
(the square root is always real since B > 2 for > 1). For xed B the
function A g(A) = 2 (A) satises
4(1 + A) B 2 B B 2 + 4A2 4
4(1 + A) 4 4A

g (A) =
<
=0
2
2
2
2B + 4A 4(1 + A)
2B 2 + 4A2 4(1 + A)2

and therefore 2+ (A) is minimal for A = 12 or 2 , 3
2 . We conclude
that the maximal strain that
is given by and that the
1can be recovered

corresponding basis is F =
(e1 e2 ) .
2
Assume now that there exists a matrix F = f f + 1 f f T
with > . Let Q = (f , f ) and observe that by the invariance of T under
SO(2) we have
QT F Q = e1 e1 +
1
e2 e2 T K pc ,
a contradiction. For we have f f + 1 f f K pc for all f S1

and hence this matrix is also contained in all the sets K(f )pc for all f S1 .
This concludes the proof of the theorem.

Remark 2.4.2. We recover the estimate Kohn and Niethammer for = 1 +
by expanding and keeping only lower order terms in . Indeed, = 1 + 2
(up to higher order terms) and in this case

1
1
1
1
1
1 22
T
C=F F =
.
, where F =
1
1
1
1
22 1
2
2
2.5 Dimensional Reduction in Three Dimensions

In the following sections we consider problems in 3 3 matrices. The rst
situation that one encounters is the case that the structure of the threedimensional wells is two-dimensional in the sense that the matrices have a
block diagonal structure.
54
To be more specic, assume that the matrices Ui , i = 1, . . . , k, are symmetric with a common eigenvector v and corresponding eigenvalue . In a
suitable basis, the matrices are therefore block diagonal,

i
U
i M22 ,
Ui =
, U
and the problem of characterizing the generalized convex hulls of

K = SO(3)U1 . . . SO(3)Uk
is equivalent to the corresponding two-dimensional problem for
k .
= SO(2)U
1 . . . SO(2)U
K
have been characterized in Theorem 2.2.3 in the case
The hulls for the set K
of equal determinant and therefore we assume that the matrices Ui satisfy
det Ui = for i = 1, . . . , k (see, however, Section 2.6 for the two well problem
with dierent determinant). This is not too restrictive an assumption, and
it is always satised in the analysis of martensitic phase transformations on
which we focus here. Examples include the orthorhombic to monoclinic transformation with two symmetry related wells and one variant of the tetragonal
to monoclinic transformations with four wells, see Section 5.5. This situation
is also encountered for special cases of the cubic to orthorhombic transformation if the boundary conditions reduce the number of the wells on which
the microstructure can be supported, see the analysis in Chapter 4.
Theorem 2.5.1. Assume that U1 , . . . , Uk M33 are positive denite and
symmetric with det Ui = > 0. Suppose that there exist > 0 and v S2
such that Ui v = v for i = 1, . . . , k. Let K = SO(3)U1 . . . SO(3)Uk .
Then K (2) = K lc = K rc = K qc = K pc . Moreover, there exists a discrete set
E = {e1 , . . . , e } S2 such that

K qc = F M33 : det F = , F T F v = 2 v,

|F ei |2 max |Uj ei |2 , i = 1, . . . , .
j=1,...,
Example. The rst nontrivial example of a quasiconvex hull in three

dimensions was given by Ball and James. Suppose that
2 0 0
1 0 0
1 0 0
(2.50)
U1 = 0 1 0 , U2 = 0 2 0 , U3 = 0 1 0 .
0 0 1
0 0 1
0 0 2
% = SO(3)U1 SO(3)U2 , is given
Then the quasiconvex hull of two wells, say K
by

% qc = F : det F = 12 2 , |F (e1 e2 )|2 12 + 22 , (F T F )e3 = 12 e3 .
K
2.6 The Two-well Problem in Three Dimensions
55
2.6 The Two-well Problem in Three Dimensions

We now consider the two well problem in three dimensions, i.e., the set
K = SO(3)U1 SO(3)U2
with U1 , U2 M33 positive denite. Changing the dependent and independent coordinates, we may assume that U1 = I and that U2 is a diagonal
matrix which we denote by H = diag(h1 , h2 , h3 ) with h1 , h2 , h3 > 0. In view
of Proposition A.2.1, the two wells are rank-one connected if and only if the
middle eigenvalue of H is equal to one. In this case, I and H have a common eigenvalue and one obtains again a reduction from the three-dimensional
problem to a two-dimensional one.
Theorem 2.6.1. Assume that
K = SO(3) SO(3)H,
with

H = diag h1 , h2 , h3
(2.51)
with h1 1 h2 and h3 = 1. Then

(

F%
lc
rc
qc
pc
lc
%
%
,
K =K =K =K = Q
: Q SO(3), F K
1
% = SO(2) SO(2) diag(h1 , h2 ).
where K
Remark 2.6.2. An explicit formula for
% lc = K
% (3) = K
% rc = K
% qc = K
% pc
K
aks result for the two-dimensional two-well problem in Thefollows from Sver
orem 2.2.2.
Theorem 2.6.3. Let the set K be given by (2.51) and assume that either
h1 h2 h3 > 1 or that h1 h2 h3 < 1. Then K pc is trivial, i.e.,
K pc = K.
We now turn to the most interesting case which shows the surprising
dierences between two-dimensional and three-dimensional models. Assume
that the two wells SO(3) and SO(3)H are incompatible and that the assumptions in Theorem 2.6.3 are not satised. In this case, it turns out that the
polyconvex hull is only trivial if the parameters hi satisfy additional conditions. If these are violated, then it is possible to solve explicitly the minors
relations which are now a system of nine quadratic and one cubic equation.
This has to be contrasted with the fact that the rank-one convex hull is
always trivial.
More precisely, we have the following results. Here we use the convention
that h4 = h1 and h3 = h0 .
56
Theorem 2.6.4. Suppose that K is given by

K = SO(3) SO(3)H, with H = diag h1 , h2 , h3 , h1 h2 h3 > 0,

and that K does not contain rank-1 connections, i.e., that h2 = 1. Assume,
in addition, that one the following two conditions holds:
i) there exists an i such that (hi 1)(hi1 hi+1 1) 0,
ii) h1 h2 > 1 > h3 > 13 or 3 > h1 > 1 > h2 h3 > 0.
Then
Mpc (K) is trivial and in particular
K pc = K.
The next theorem demonstrates that additional assumptions on the parameters hi are needed for the assertion to be true, even if both wells have
the same determinant.
Proposition 2.6.5. Let H = diag(h2 , h1 , h1 ) with h h where h is the
largest solution of the equation
2h
4 h
= 1.
(2.52)
1 + h 1 + h2
Suppose that K = SO(3) SO(3)H. Then K pc \ K rc = .
Proof. We have to nd (0, 1) and R, S conv(SO(3)) such that
F = R + (1 )SH satises the minors relations,
cof F =R + (1 )S cof H,
det F = + (1 ) det H = 1.
We x =
1
2
and seek R = S in the following form:

1 1
H = diag(h2 , , ),
h h
R = S = diag(r1 , r2 , r2 ).
Since all matrices in the construction are diagonal, the minors relations reduce
to three nonlinear equations,
r22
r1
1
1
(1 + )2 = (1 + 2 ),
4
h
2
h
r2
r1 r2
1
2
(1 + h )(1 + ) = (1 + h),
4
h
2
r1 r22
1 2
2
(1 + h )(1 + ) =1.
8
h
These equations can be solved explicitly for r1 and r2 ,
2.7 Wells Dened by Singular Values
57
2h
2 h
.
, r2 =
1 + h2
1+h
It remains to check whether R conv(SO(3)). Since r1 , r2 [ 0, 1 ] and
r1 r2 this is in view of (A.4) true if 2r2 r1 1, i.e. if h h where h is
the solution of (2.52).

r1 =
Remark 2.6.6. We nd h 8.35241 and h2 69.7628.

Remark 2.6.7. For h = 9, i.e. H = diag(81, 19 , 19 ), one obtains a particularly
nice matrix F K pc \ K rc = . In fact,
1 1
F = diag(9, , ) K pc \ K rc .
3 3
A short calculation shows that
9 3 3
R = diag( , , ) conv(SO(3)).
41 5 5
If we write D(x,y,z) = diag(x, y, z), then R has the representation
124
8
8
1
D(1,1,1) +
D(1,1,1) +
D(1,11) +
D(1,1,1) .
205
41
41
205
By construction, F = 12 R + 12 RH is the center of mass of a polyconvex
measure supported on eight matrices in K.
R=
In contrary to this, the rank-one convex hull of two incompatible wells is

always trivial.
Theorem 2.6.8. Suppose that K is given by (2.51) and that K contains no
rank-one connections. Then
Mrc (K) is trivial and in particular K rc = K.
The example of the two-well problem demonstrates the rich eects that
passage from two to three dimensional situations produces. The next section presents one of the few fully three-dimensional problems for which the
semiconvex hulls can be characterized.

We now study the case of sets K given by the singular values of the matrix
F , i.e.,

K = F Mnn : i (F ) = i , det F = ,
where 1 , . . . , n > 0 are given parameters and = 1 n . The results
in this section are motivated by the mathematical analysis of the eective
behavior of nematic elastomers in Chapter 3, and they allow us to nd an
explicit formula for the quasiconvex hull of the free energy of the system
proposed by Bladon, Terentjev, and Warner. Before we address the threedimensional situation, we consider the two-dimensional case.
58
The Two-dimensional Case. We now consider sets that are invariant under multiplication by SO(2) from the left and the right. We restrict ourselves
to the case of sets with constant determinant, since sets of this form are
important in the analysis of nematic elastomers in Chapter 3.
Assume that K is given by

K = F M22 : 1 (F ) = 1 , 2 (F ) = 2 , det F = 1 2 ,
(2.53)
where 0 < 1 2 and 0 1 (F ) 2 (F ) are the singular values of F .
Equivalently, K can be represented as
$

SO(2) 1 e e + 2 e e .
K=
eS1
The next theorem gives an explicit characterization of all the semiconvex

hulls.
Theorem 2.7.1. Assume that K is given by (2.53). Then

conv(K) = F : 1 2 min F : R max F : R 1 + 2 ,
RO+ (2)
1 2
min
RO (2)
RO+ (2)
F :R

max F : R 1 + 2 .
RO (2)
Moreover, K (2) = K rc = K qc = K pc and these sets are given by

K pc = F : det F = 1 2 , i (F ) [ 1 , 2 ] .
In particular, K pc = conv(K) {det F = 1 2 }.
Remark 2.7.2. The case that K is dened in terms of the singular values without the determinant constraint was treated in [DcT98], see also Figure 2.4.
Here conv(K) = {F : 2 2 , 1 + 2 1 + 2 }. The two conditions in
the denition of the convex hull follow immediately from general properties
of singular values since F 1 (F ) + 2 (F ) is a convex function.
Proof. Let M denote the right hand side of the formula in the assertion
of the theorem. In view of Proposition 2.7.8 it is clear that M is convex with
conv(K) M and that M is invariant under SO(2), i.e., SO(2)M SO(2) = M .
It therefore suces to show that all diagonal matrices = diag(1 , 2 ) in M
are contained in conv(K). However, if satises all the inequalities in the
denition of M , then

conv diag(1 , 2 ), diag(1 , 2 ), diag(2 , 1 ), diag(2 , 1 ) ,
see Figure 2.4. The remaining assertions are an immediate consequence of
the characterization of the semiconvex hulls of the two-well problem in Theorem 2.2.2. In fact, g(F ) = (det F 1 2 )2 is a polyconvex function and
therefore K pc {det X = 1 2 }. Moreover, if = diag(1 , 2 ) is a diagonal
matrix in conv(K), then
qc
SO(2) diag(1 , 2 ) SO(2) diag(2 , 1 ) ,
and this establishes the assertion.

59
2
( 1 ,2)
( 2, 1)
1
( 2, 1)
(1, 2)
Fig. 2.4. Semiconvex hulls of sets dened by singular values. The solid dots are the
four diagonal matrices in K in (2.53), the four circles are the additional diagonal
matrices in K if the condition of positive determinant is dropped. The convex hull
of the set K in (2.53) is the solid rectangle, the polyconvex hull consist of the two
hyperbolic arcs.
The Three-dimensional Case. We now extend the result in Theorem 2.7.1

to the three-dimensional situation. It turns out that the convex hull of K is
given by SO(3)SO(3) where = diag(s1 , s2 , s3 ) is a diagonal matrix such
that the vector (s1 , s2 , s3 ) is contained in the convex hull P of the diagonal
matrices in K. As a rst step we derive a formula for this set. It is important
to note that P contains matrices with negative determinant, e.g.,

1
diag (1 3 , 2 , (1 3 ) P.
2
Proposition 2.7.3. Assume that 0 < 1 2 3 and dene the set E by

E = (i i , j j , k k ) : i,j,k {1}, i j k = 1, {i, j, k} = {1, 2, 3} .
Then conv(E) = P where P is given by

P = R3 : |, | 1 + 2 + 3 , i {1}, 1 2 3 = 1,
|, | 1 + 2 + 3 , i {1}, 1 2 3 = 1,

|i | i , i = 1, 2, 3 .
Remark 2.7.4. If all i are distinct, then E contains 24 points and its convex
hull is the intersection of the 14 halfspaces dened by the normals ei and
1 (e1 e2 e3 ), see Figure 2.5.
3
Proof. It is clear that P contains conv(E) since P is a convex set that
contains all points in E. We therefore have to show that P conv(E). Since
P is a compact set it suces to show that P conv(E), where P consist
60

3
( 2 , 1 , 3 )
( 1 , 2 , 3 )
2
( 1 , 3 , 2 )
z 0
( 3 , 1 , 2 )
-2
( 2 , 3 , 1 )
-2
( 3 , 2 , 1 )
0
y
0
x
-2
Fig. 2.5. The convex hull of E. The left gure shows the generation of one of the
hexagons in the boundary of the convex hull, the right gure shows the convex hull
of E. The boundary of the polygon consists of six rectangles,with normals ei , six
small hexagons, with normals (1 , 2 , 3 ), i {1} with 1 2 3 = 1, and six large
hexagons, corresponding to 1 2 3 = 1, respectively.
of all points in P for which at least one of the inequalities in the denition
of P is an equality. We choose two representative cases, namely 3 = 3 and
1 + 2 + 3 = 1 + 2 + 3 ; all the other cases can be handled similarly.
Assume rst that p = (1 , 2 , 3 ) P with 3 = 3 . We assert rst that

p C = conv (1 , 2 , 3 ), (2 , 1 , 3 ), (1 , 2 , 3 ), (2 , 1 , 3 )
(see Figure 2.5). In fact, it follows from the inequalities in the denition of
P that
1 + 2 + 3 1 + 2 + 3
1 2 + 3 1 + 2 + 3
1 + 2 + 3 1 + 2 + 3
1 + 2
1 + 2 ,
1 2
1 + 2 ,
1 + 2 1 + 2 ,
1 2 + 3 1 + 2 + 3
1 2 1 + 2 .
These four inequalities characterize the convex hull of

(1 , 2 ), (2 , 1 ), (1 , 2 ), (2 , 1 )
(see Figure 2.4), and this proves that p conv(E).
Assume now that 1 + 2 + 3 = 1 + 2 + 3 . In this case we show that
that p is contained in the convex hull C of the six points

(1 , 2 , 3 ), (1 , 3 , 2 ), (2 , 3 , 1 ), (2 , 1 , 3 ), (3 , 1 , 2 ), (3 , 2 , 1 ) ,
(see Figure 2.5). The set C forms a hexagon in the ane subspace A given
by A = {1 + 2 + 3 = 1 + 2 + 3 }, and a point in A belongs to C if and
only if
1 1 ,
2 1 ,
61
3 1 ,
and
1 + 2 3 1 + 2 + 3 ,
1 2 + 3 1 + 2 + 3 ,
1 + 2 + 3 1 + 2 + 3 .
These inequalities are contained in the denition of P and thus p conv(E).
All the remaining cases can be handled similarly.

The next theorem describes the fundamental construction for the generation of the lamination convex hull of K. To simplify the notation, we assume
that 1 2 3 = 1. We use the convention that 0 = 3 , 4 = 1 and 0 = 3 ,
4 = 1 and we write (F ) = 1 (F ), 2 (F ), 3 (F )} for the set of the singular
values of F .
Theorem 2.7.5. Assume that 0 < 1 2 3 with 1 2 3 = 1 and that

K = F M33 : det F = 1, i (F ) = i , i = 1, 2, 3 .
Then the sets Mi dened by

Mi = F M33 : det F = 1, i (F ),

(F ) \ {i } [ min{i1 , i+1 }, max{i1 , i+1 } ]
are contained in K (1) for i = 1, 2, 3. Moreover, we have

K (1) = M1
if
1 = 2
and
K (1) = M3
if
2 = 3 .
Remark 2.7.6. In general the inclusion M1 M2 M3 K (1) is strict.

Proof. To prove the rst part of the theorem, we assume that i = 1, and
we write M = M1 . The argument is analogous for i = 2 and i = 3. The
assertion of the proposition is now equivalent to M K (1) where

M = F M33 : det F = 1, 1 (F ) = 1 , 2 (F ), 3 (F ) [ 2 , 3 ] .
Let F M . Since QF R M , for all Q, R SO(3) and F M , we may
suppose that F is diagonal, F = diag(1 , 2 , 3 ), with 2 , 3 [ 2 , 3 ]. Note
that 2 3 = 2 3 since det F = 1. There is nothing to prove if 2 = 2 or
3 = 3 , since the condition det F = 1 implies in this case that F K. We
now show that there exists for 2 , 3 (2 , 3 ) a > 0 (which depends on 2
and 3 ) such that

2
%
%
.
F = diag(1 , F ) K where F =
0 3
62
Then
F + F = 2e2 e3
and
We dene
% = (F% )T F% =
C
F =

1 + 1
F + F K (1) .
2
2

22 2
.
2 23 + 2
% are the solutions of

The eigenvalues t of C
% tI) = t2 (2 + 3 + 2 )t + 2 2 = 0,
det(C
2
3
2 3
and the requirement that t+ dened by

22 + 23 + 2 2
22 + 23 + 2
t =
22 23
2
2
be equal to 32 leads to
=
1
3
)
)
32 22 32 23 > 0.
(2.54)
Since t+ t = 22 23 = 22 32 , this choice of also yields t = 22 and we conclude

that for the value of given in (2.54) the matrices F are contained in K
and this proves the rst assertion of the theorem.
It remains to prove the characterization of K (1) if two of the parameters
in the description of K coincide. Without loss of generality we assume that
1 = 2 < 3 , and we have to prove that K (1) M1 . Suppose thus that
F K (1) \ K and choose F1 , F2 K such that there exists a (0, 1) and
a, n R3 , a, n = 0 with
F = F1 + (1 )F2 ,
F1 F2 = a n.
Since QF R K, for all Q, R SO(3) and F K, we may choose Q and

R SO(3) such that F 1 = QF1 R K is diagonal, F 1 = diag(3 , 1 , 1 ).
= Qa and
We dene analogously F 2 = QF2 R K, F = QF R K (1) , a
= RT n. Then
n
F = F 1 + (1 )F 2 ,
n
.
F 1 F 2 = a
The intersection of the plane spanned

by the unit vectors

e2 = (0, 1, 0) and
= 0 with normal n
at
e3 = (0, 0, 1) intersects the plane w R3 : w, n
least in a one-dimensional line through the origin parallel to some unit vector
v S2 . This implies
)v = (F 1 F 2 )v = 1 v F 2 v,
0 =
n, v
a = (
an
63
and therefore v is an eigenvector of F 1 and F 2 with corresponding eigenvalue

1 . Consequently, F v = 1 v and
1 (F ) = min (F ) = min2 |F e| |F v| = 1 .
eS
To prove that 1 is the smallest singular value of F , let ()+ denote the convex,
nondecreasing function t (t)+ = max{t, 0}. Then the functions
g1 (F ) =
g2 (F ) =
sup |F e| 3
eS2
sup | cof F e|
eS2
1 +
1
are polyconvex, and since F K (1) K pc we deduce i (F ) [ 1 , 3 ].

Therefore
1 (F ) = 1
and
1 = min{2 , 3 } 2 (F ) 3 (F ) max{2 , 3 }.
We obtain F M1 . The matrices F and F have the same singular values,

and hence F M1 . This concludes the proof of the theorem.

The foregoing theorem implies immediately a formula for the semiconvex
hulls in three dimensions.
Theorem 2.7.7. Assume that 0 < 1 2 3 with 1 2 3 = 1 and that

K = F M33 : det F = 1, i (F ) = i , i = 1, 2, 3 .
Then
K (2) = K lc = K rc = K qc = K pc ,
and these sets are given by

K qc = F M33 : det F = 1, i (F ) [ 1 , 3 ], i = 1, 2, 3 .
(2.55)
(2.56)
Moreover,

conv(K) = F M33 : 1 (F ), 2 (F ), 3 (F ) P ,
where P has been dened in Proposition 2.7.3. In particular,
K pc = conv(K) {F M33 : det F = 1}.
Proof. Let A be the set given in (2.56). Since det F = 1, we have
min (F ) = 1 (F ) =
1
1
=
,
max (cof F )
3 (cof F )
(2.57)
64
and therefore we may rewrite the denition of A as

A = F M33 : g1 (F ) 0, g2 (F ) 0, g3 (F ) 0 ,
where g1 and g2 were dened in the proof of Theorem 2.7.5 and
g3 (F ) = (det F 1)2 .
The functions gi are polyconvex, hence K pc A. It only remains to prove
that A K (2) . We then obtain that A K pc K (2) K pc and the
equation (2.55) is thus an immediate consequence. Since QF R A, for all
Q, R SO(3) and F A we may assume that F A is a diagonal matrix,
F = diag(1 , 2 , 3 ) with 1 1 2 3 3 . If 2 3 3 , then
3
2
1
3
3
and
2 3
3 .
3
By Theorem 2.7.5,

M1 = F M33 : det F = 1, 1 (F ) = 1 ,

2 3
2 3
, 3 }, 3 (F ) = max{
, 3 }
2 (F ) = min{
3
3
2 3
is contained in K (1) . Now 1 1 2
since
3
2
2 3
3
2 3 2 3
1 1 .
If 2 3 23 , then 3 (F ) = 3 for F M1 while for 2 3 > 23 one has

2 (F ) = 3 for F M1 . Another application of Theorem 2.7.5 with i = 3 or
i = 2, respectively, implies that

F M33 : i (F ) = i ,
(1)
M1
K (2) .
1 2
Suppose now that 1 1 2 3 2 . Then 1
2 , and we
1
conclude from Theorem 2.7.5 that

M3 = F M33 : det F = 1, 3 (F ) = 3 ,
1 2
1 2
}, 2 (F ) = max{1 ,
}
1 (F ) = min{1 ,
1
1
1 2
is contained in K (1) . In this situation,
2 3 3 , since
1
1 2
2
1
1 2 1 2
3 3 ,
and we can apply Theorem 2.7.5 once more (with i = 1 or i = 2) to deduce
(1)
F M33 : det F = 1, i (F ) = i , i = 1, 2, 3 M3
65
K (2) .
This establishes the result for the polyconvex hull.

In order to prove the formula for the convex hull, we conclude from Proposition 2.7.9 that for all R SO(3)
min U : R = 1 2 3 ,
U K
max U : R = 1 + 2 + 3 .
U K
(2.58)
Let A = {F M33 : (1 (F ), 2 (F ), 3 (F )) P}. By polar decomposition

and the denition of the (signed) singular values, F = Q1 diag(1 , 2 , 3 )Q2
belongs to A for all Q1 , Q2 SO(3) if and only if = diag(1 , 2 , 3 ) P.
In view of Proposition 2.7.3, P implies conv(K) and consequently
A conv(K). It remains to show that conv(K) A (it is not obvious
that A is a convex set). Assume that F conv(K). Since QKR = K for
all Q, R SO(3), we may assume that F = diag(1 , 2 , 3 ) is diagonal.
We have to show that (1 , 2 , 3 ) P. The formula for P shows that all
points (1 , 2 , 3 ) P with |1 | 2 3 are characterized by the three
inequalities
3 3 ,
1 + 2 + 3 1 + 2 + 3 .
By (2.58)
min F : R = 1 2 3 1 2 3 ,
RSO(3)
max F : R = 1 + 2 + 3 1 + 2 + 3 ,
RSO(3)
and by the convexity of the norm

3 = max2 |F e| max max2 |U e| = 3 ,
eS
U K eS
and these inequalities prove the desired estimates.

We nally show that K pc = conv(K) {F M33 : det F = 1 2 3 }.
It suces to prove this for diagonal matrices which we identify with vectors
in R3 . Intuitively this is clear: the surface 1 2 3 = 1 intersects the face F
of the convex hull of K given by 1 + 2 + 3 = 1 + 2 + 3 in R3+ in
a one-dimensional curve that contains the points (1 , 2 , 3 ) and (1 , 3 , 2 )
and consists of points with 1 < 1 between these two points. In order to
make the argument precise, consider the face F intersected with all points
that satisfy 2 = 3 . Obviously, the two points (1 , 12 (2 + 3 ), 12 (2 + 3 ))
and (0, 12 (1 + 2 + 3 ), 12 (1 + 2 + 3 )) are contained in this set. Since
1
1
2
2
4 (2 +3 ) 1 1 2 3 and 4 (1 +2 +3 ) 0 = 0 we conclude by continuity
the existence of a point (t, s, s) F with s2 t = 1 2 3 and 0 < t < 1 . This
proves the nal assertion in the theorem, since all points in the polyconvex
hull of K have to satisfy the inequality 1 1 . The proof of the corollary is
now complete.

66
In the characterization of the semiconvex hulls for sets depending on singular values we used the following fact. For simplicity, we rst state the
two-dimensional result.
Proposition 2.7.8. Assume that F M22 and that 1 2 are the signed
singular values of F which satisfy |1 | 2 . Then
max F : R = 1 + 2 ,
RO+ (2)
min
RO+ (2)
F : R = 1 2 ,
and
max F : R = 1 + 2 ,
RO (2)
min
RO (2)
F : R = +1 2 .
Proof. We may assume that F is diagonal, F = diag(1 , 2 ), since for all

Q1 and Q2 SO(2)
max F : R = max (Q1 F Q2 ) : R,
RO (2)
min
RO (2)
RO (2)
F :R=
min (Q1 F Q2 ) : R.
RO (2)
We rst prove the formula for the maximum in the compact set O+ . Assume
that R = ei (in complex notation) realizes one of the extremal values and
consider for > 0 any smooth curve : (, ) R with (0) = . Then

d ''
cos (t) sin (t)
F :
= ( sin )(
1 + 2 ) = 0.
'
sin (t) cos (t)
dt t=0
If 1 + 2 = 0, then F = 2 diag(1, 1) and F : R = 0 for all R O+ (2).
We may therefore assume that 1 + 2 = 0. In this case {0, }, and thus
R = I. This implies the assertion of the proposition. Similarly, if R O (2)
realizes one of the extrema, then

d ''
cos (t) sin (t)
F :
= (sin )(1 2 ) = 0.
'
sin (t) cos (t)
dt t=0
If 1 2 = 0, then F = 2 diag(1, 1) and F : R = 0 for all R O (2).
Otherwise, {0, } and therefore R = diag(1, 1) and this concludes
the proof of the proposition.

We now turn to the three-dimensional situation. To simplify the statement, we introduce some notation. Let D be any set. If f : D R and
f (x) f (y) for all y D then we write x argmax(f, D) or simply
x argmax(f ) if the domain D is clearly dened in the context. We dene argmin(f ) similarly.
Proposition 2.7.9. Assume that F M33 and that 1 , 2 , 3 are the
signed singular values which satisfy |1 | 2 3 are the signed singular
values of F . Then the following assertions hold:
67
1. We have for all F M33

max F : R = 1 + 2 + 3 ,
RO+ (3)
min
RO+ (3)
F : R = +1 2 3 ,
and
max F : R = 1 + 2 + 3 ,
RO (3)
min
RO (3)
F : R = 1 2 3 .
2. We now assume that F is diagonal, F = diag(1 , 2 , 3 ), with det F 0.

Assume rst that rank(F ) = 1. Then

% 1) : Q
% O+ (2) ,
argmax(F : R, O+ (3)) = Q = diag(Q,

% 1) : Q
% O (2) .
argmin(F : R, O+ (3)) = Q = diag(Q,
If rank(F ) = 2, then
argmax(F : R, O+ (3)) = {I},
argmin(F : R, O+ (3)) = {diag(1, 1, 1)}.
Finally, if det F > 0 and 1 = 2 = 3 then
argmax(F : R, O+ (3)) = {I},

argmin(F : R, O+ (3)) = Q = I + 2e e, e S2 },
while in all the other cases,
argmax(F : R, O+ (3)) = {I},
argmin(F : R, O+ (3)) = {diag(1, 1, 1)}.
3. If det F < 0, then
argmax(F : R, O (3)) = argmin(F : R, O+ (3)),
and
argmin(F : R, O (3)) = argmax(F : R, O+ (3)).
Proof. As in the two-dimensional situation we may assume that F
is a diagonal matrix, F = diag(1 , 2 , 3 ). The assertions are obvious if
rank(F ) 2, i.e., 1 0. We may thus suppose that rank(F ) = 3.
Since for every skew symmetric matrix X there exists a dierentiable curve
X : (, ) SO(3) with (0) = I and (0) = X (we may choose
(t) = exp(tX)), any Q SO(3) realizing one of the extrema must satisfy
d ''
(t)Q : F = XQ : F = 0,
'
dt t=0
d ''
Q(t) : F = QX : F = 0.
'
dt t=0
68
In view of the identity tr(AB) = tr(BA) for all A, B Mnn we have (recall
that F is diagonal) XQ : F = tr(F XQ) = tr(QF X) = QF : X = 0 and
QX : F = tr(F QX) = F Q : X = 0, and this implies that both the matrices
QF and F Q must be symmetric. The resulting equations QF = F QT and
F Q = QT F (or Qij j = Qji i and i Qij = j Qji ) can be rewritten in
dierent ways, in particular as
(i j )(Qij + Qji ) = 0,
(i + j )(Qij Qji ) = 0,
(i2 j2 )Qij = 0.
(2.59)
We assume rst that 1 > 0 and distinguish four dierent cases:

Case 1: 1 = 2 = 3 . Then Q is symmetric and thus either the identity or
a 180 rotation about an axis e given by Q = I + 2e e. The assertion in
this case follows from F : I = tr(F ) and F : (I + 2e e) = 1 .
Case 2: 1 < 2 = 3 . In this case, Q12 = Q21 = Q13 = Q31 = 0 and
% ) with Q O (2). If Q11 = 1 then
Q23 = Q32 . Thus Q = diag(1, Q
F : Q = 1 for all Q O (2), while Q : F = 1 + 2(cos )3 for Q11 = 1.

The maximum is therefore attained for = 0 while the minimum is realized
for Q = diag(1, 1, 1).
Case 3: 1 = 2 < 3 . It follows that Q13 = Q31 = Q23 = Q32 = 0 and that
% , 1) with Q
% O (2). If Q33 = 1 then
Q12 = Q21 . Thus Q = diag(Q
F : Q = 3 for all Q O (2), while Q : F = 2(cos )1 + 3 for Q33 = 1.

The maximum is therefore attained for = 0 and the minimum for = .
Case 4: 1 < 2 < 3 . By (2.59) the matrix Q is diagonal and the conclusion
is obvious.
Let us now assume that det F < 0. Then
max F : R =
RO (3)
min
RO (3)
F : (R) =
min
RO+ (3)
F : R.
This identity and the analogous one for the minimum prove the formulae in
parts 1 and 3 of the proposition.

3. Macroscopic Energy for Nematic Elastomers
One goal in the analysis of phase transformations via variational techniques

is to identify the so-called eective or relaxed energy of the system,

1
inf
W (Du)dx.
W qc (F ) =
uW 1, (;R3 ) ||
u(x)=F x on
It describes the energy of the system for ane boundary conditions, if the
system is allowed to form locally energetically optimal microstructures. In
particular, this formulation allows innitesimally ne structures by neglecting
all higher order eects such as surface energies or the atomistic structure of
the material that typically introduce nite length scales.
The energy W qc governs the macroscopic behavior of the system and is from a practical point of view - the right quantity for the computation of averaged quantities such as the macroscopic stress. Since W qc is quasiconvex, the
variational problem has a minimizer, and numerical schemes are expected
to provide reliable and mesh independent results without oscillations on a
scale comparable to the underlying triangulation. However, it is one of the
unresolved challenges in the eld to derive a characterization of W qc . Even
in the case of the cubic to tetragonal transition, which has been investigated
extensively and for which Ericksen and James proposed a quartic model energy, nothing in known about its relaxation. It is surprising or perhaps an
indication of the true complexity of the matter that the zero set of W qc , the
quasiconvex hull of the three martensitic wells in (5.3), has not been found
yet, despite numerous attacks.
This motivates us to go in this chapter beyond the analysis of crystalline
microstructures and to study a dierent physical system, nematic elastomers,
a class of polymers that undergo a nematic to isotropic phase transformation.
As a result of the isotropy of the high temperature phase, the energy in
the nematic phase depends only on the singular values of the deformation
gradient. We have seen in Section 2 that this invariance allows one for example
a characterization of the semiconvex hulls of the zero set, and we show in this
chapter how to nd the macroscopic energy for the system.
G. Dolzmann: LNM 1803, pp. 6981, 2003.

70

1111
0000
0000
1111
0000
1111
0000
0001111
111
0000
1111
000
111
000
111
000
111
111
000
000
111
000
111
000
111
000
111
00
11
11
00
00
11
00
11
00
11
00
11
000
111
111
000
000
111
1111
0000
0000
1111
000
111
111
000
000
111
000
111
000
111
00
11
11
00
00
11
00
11
00
11
00
11
0000
1111
1111
0000
111
000
000111
111
000
000
111
000
111
111
000
000
111
000
111
0000
1111
1111
0000
0000
1111
0000
1111
000
111
000
111
000
111
0000
1111
1111
0000
0000
1111
1111
0000
0000
1111
0000
1111
0000
1111
0000
1111
111
000
000
111
1111
0000
0000
1111
0000
1111
Fig. 3.1. The isotropic-nematic phase transformation in nematic elastomers, polymers consisting of weakly cross-linked side-chain polymer liquid crystals. The nematic elements (rigid, rod-like molecules) are attached to the backbone chain. They
have a random orientation in the high temperature (isotropic) phase due to thermal uctuations. In the low temperature (nematic) phase, a local alignment of the
mesogens causes a stretch of the network in direction of the director n (indicated
by the arrow in the right gure) and a contraction in the directions perpendicular
to it.
3.1 Nematic Elastomers

Nematic elastomers are a fascinating material that combine nematic properties with the rubber elasticity of the underlying polymeric network. They are
synthesized by cross-linking polymer liquid crystals close to their isotropic
to nematic transformation temperature. As a result, one obtains a solid with
an isotropic to nematic phase transformation at a comparable temperature.
The low temperature phase is characterized by the coupling of the elastic deformation to the orientation of the nematic director, see Figure 3.1. Bladon,
Terentjev and Warner proposed a formula for the free energy of nematic
elastomers modeling them as cross-linked networks of anisotropic Gaussian
chains within a continuum model. The state variables are the deformation
gradient F and a vector eld n, the director, describing the orientation of
the mesogens in the nematic phase. Using the undistorted high temperature
phase as the reference conguration, their expression can be written as

r1/3 |F |2 r 1 |F T n|2 3 if det F = 1,
r
WBTW (F, n) = 2
+
otherwise.
Here and r are two positive, temperature dependent material constants, the
rubber energy scale and the backbone anisotropy parameter (i.e., the mean
ratio of chain dimensions in the directions parallel and perpendicular to the
director: for r = 1 the chain is a spherical coil, while r > 1 corresponds to the
prolate case and r < 1 to the oblate one), respectively. The for elastomers
typical incompressibility is incorporated into the model by assuming that the
energy is innite if the determinant of the deformation gradient is not equal
to one. Above the transformation temperature, we have r = 1, and we recover
the standard form of a neo-Hookean, isotropic free energy density,
3.1 Nematic Elastomers
71

2 + 2 + 2 3 if det F = 1,
1
2
3
Wiso (F ) = 2
+
otherwise,
where 0 < 1 2 3 are the singular values of F . Below the transformation temperature we have r > 1 in the prolate case which we will consider
here. Since the formula for the energy does not contain derivatives of the director n, we can minimize rst in n and then consider the variational problem
for the deformation u. We dene
Wne (F ) = min2 WBTW (F, n)
nS
and obtain from

max |F T n| = 3 (F )
nS2
that

r1/3 2 + 2 + 1 2 3 , if det F = 1,
1
2
r 3
Wne (F ) = 2
+
otherwise.
For the mathematical analysis, it is convenient to consider Wne as a special

case of the family of energies
p
) p2 (F ) p3 (F )
1 (F
+
+
3 if det F = 1,
p
1
2p
3p
(3.1)
W (F ) =
+
else,
with 0 < 1 2 3 , 1 2 3 = 1, and p 2. The relaxation result that we
prove in the subsequent sections provides us with the following formula for
the relaxation of the free energy,
if F L
0

2
1/3 2
2 r1/6 1 (F ) + r 1 (F ) 3 if F I1 ,
qc
Wne
(F ) =
W (F )
if F S,
+
else,
where

L = F M33 : det F = 1, max (F ) r1/3 ,

I1 = F M33 : det F = 1, min (F )2max (F ) r1/2 ,

S = F M33 : det F = 1, min (F )2max (F ) r1/2 .
72

max (cof F)
I1
S
r 1/6
1
r 1/3
max (F)
Fig. 3.2. The macroscopic phase diagram for nematic elastomers. The phase
boundary between the intermediate phase and the solid phase is given by
min (F )2max (F ) = r1/2 .
The formula for the macroscopic energy reveals three dierent types of mechanical response of the system to applied strains, see Figure 3.2. In the liquid
phase L, the energy is identical to zero and the system shows no resistance
to applied stretches and has a completely soft behavior. In the intermediate
phase I1 , the expression for the energy depends only on the smallest singular
value 1 , and therefore the system behaves like a liquid along transformation paths that leave 1 xed; all other deformations lead to a change in the
stress. In the solid phase, nally, the material has the properties of a neoHookean rubber and all deformations change the stresses in the material. The
liquid like behavior has been reported in the experimental literature (within
certain limits) and rst numerical experiments with the macroscopic energy
qc
Wne
by Conti, DeSimone, and Dolzmann show a qualitative agreement with
the experimental results.
3.2 The General Relaxation Result

The goal of our analysis is to nd an explicit formula for the quasiconvex
envelope of the energy W given in (3.1). From a more general point of view,
the ideas for the characterization of relaxed energies are similar to those
developed for quasiconvex hulls of sets.
In the setting of hulls of compact sets, it was crucial to nd an inner
and an outer bound for the quasiconvex hull K qc , i.e., sets A and B with
A K qc B, and then to prove that the inclusion B A holds. In this
case, all the inclusions are equalities, hence A = K qc = B, and this establishes
a representation for K qc .
3.2 The General Relaxation Result
73
In the framework of envelopes of functions, the inner and the outer bound
are replaced by a lower and an upper bound for the relaxed energy, i.e.,
functions W and W with W W qc W . If it turns out that W W ,
then equality holds in this chain of inequalities and one obtains a formula for
W qc . As in the case of semiconvex hulls, there are two canonical choices for
W and W , namely the polyconvex envelope W pc and the rank-one convex
envelope W rc , and the goal is then to prove that W rc = W pc . If W is a
real valued function, then this identity implies that W rc = W qc = W pc and
one obtains a formula for the relaxed energy. In the situation at hand, the
incompressibility constraint causes a slight diculty since extended valued
quasiconvex functions are not necessarily rank-one convex. To resolve this
ak to show that W qc W rc .
issue, we use a construction by M
uller and Sver
Since polyconvexity implies quasiconvexity for extended valued functions, we
conclude W qc W rc = W pc W qc and this demonstrates that all the
envelopes are in fact equal.
Based on this approach, we prove in the subsequent sections the general
relaxation result. In order to simplify the notation, we dene the parameters
=
2
<1
3
and
2
> 1,
1
(3.2)
and sets of matrices L, I1 , I3 , S by

1
,
L = F M33 : det F = 1, max (F ) 3 , max (cof F )
1

1
I1 = F M33 : det F = 1, max (cof F ) ,
1

2max (F ) max (cof F ) 2max (F ) ,

I3 = F M33 : det F = 1, max (F ) 3 ,

max (F ) max (cof F ) max (F ) ,

S = F M33 : det F = 1,

max (F ) max (cof F ) 2max (F ) .
We also denote by L, I1 , I3 , S the corresponding subsets in the (s, t)-plane
which are formally given by replacing max (F ) by s and max (cof F ) by t in
the foregoing denition. It is easy to see that all matrices F M33 with
det F = 1 are contained in L I1 I3 S. We nally dene the sets E0 , E1 ,
and E3 by
1
] L,
1

1
E1 = (s, t) : t , t s2 ,
1

E3 = (s, t) : s 3 , t s ,
E0 = [ 0, 3 ] [ 0,
74
Fig. 3.3. The phase diagram for the relaxed energy.
see Figure 3.3 for a sketch of these domains in the phase plane.
We now state our relaxation result.
Theorem 3.2.1. Suppose that 0 < 1 2 3 with 1 2 3 = 1 and that
p 2. Let W : M33 R be given by
p
) p2 (F ) p3 (F )
1 (F
+
+
3 if det F = 1,
p
1
2p
3p
W (F ) =
(3.3)
+
else.
Then the relaxed energy W qc of the system is given by

if det F = 1,
max (F ), max (cof F )
qc
W (F ) =
+
else,
where the function : [ 0, ) [ 0, ) R is given by
0
if (s, t) E0 ,
1
+ 2(1 t)p/2 3
if (s, t) I1 E1 ,
(s, t) =
1 t

1 p + t p + s p 3 if (s, t) S,
2 s
3
1t
p
p/2
s
3
3
if (s, t) I3 E3 .
3 + 2 s
(3.4)
We split the proof into several steps in which we follow the general strategy
outlined above. The rst step is the construction of an upper bound W .
Then W is shown to be polyconvex and hence to be a lower bound and
this establishes the theorem, up to the additional argument that is needed in
order to show that W qc W rc . We conclude the proof in Section 3.5.
3.3 An Upper Bound for the Relaxed Energy
75
3.3 An Upper Bound for the Relaxed Energy

The lamination method described in Chapter 2 was based on the fact that
the segment F1 + (1 )F2 , [ 0, 1 ], is contained in K rc and hence in
K qc if rank(F1 F2 ) = 1 and if the end points F1 and F2 belong to K.
By the same argument, W rc (F1 + (1 )F2 ) max{W (F1 ), W (F2 )} along
rank-one lines. More generally
W rc (F ) max W (X)
XK
for all F K rc ,
and the idea behind the construction of W is to apply this inequality to

level sets of W , i.e., sets of the form {X : W (X) = w0 }, w0 R.
Proposition 3.3.1. The rank-one convex envelope W rc of W satises the
inequality W rc (F ) W (F ) where the function W (F ) is given by

p/2
1 p
+
2
3
1 p 1 p p
1
2
3
W (F ) =
+
+
3
1
2
3
p/2
p
3
3
+2
3
3
3
if F L,
if F I1 ,
if F S,
(3.5)
if F I3 ,
if det F = 1.
Moreover, for all F M33 with det F = 1 there exist pairs (i , Fi )i=1,...,k ,
k 4, with i [ 0, 1 ], Fi M33 , and det Fi = 1 satisfying condition Hk
such that
F =
k

i Fi ,
and
W (F ) = W (Fi ),
i = 1, . . . , k.
i=1
Remark 3.3.2. It turns out that k = 2 for matrices in the intermediate phases,
and that k = 4 for matrices in the liquid phase.
Proof. It follows from the inequality between the geometric and the arithmetic mean that W 0 and that

K = X : W (X) = 0 = F : det F = 1, i (F ) = i for i = 1, 2, 3 .
By the denition of L and Theorem 2.7.7, we have K rc = K (2) = L and thus
W rc (F ) max W (X) = 0
XK
for all F L.
76
It therefore suces to prove the assertion for the matrices contained in the
sets I1 and I3 . By the isotropy of W , we may assume that F is a diagonal
matrix, F = diag(1 , 2 , 3 ) with 1 2 3 = 1.
Assume rst that 1 (F ) = 1 1 . All diagonal matrices with this
property can be parameterized by

1
t
,
t 1 (t) = 1 , ,
t 1
1
3/2
3/2
t [ 1 , 1
where the domain of t is chosen in such a way that 1 is the smallest eigenvalue
of F . The energy along this curve is given by
p t p 1 p
1
3/2
+
+
3 if t [ 1 , 1],
1
2 1
t3 1
w1 (t) =
1 p 1 p t p
3/2
+
+
3 if t [ 1, 1
].
1
t2 1
3 1
3/2
The extrema of w1 on (1 , 1) satisfy

w1 (t) =
ptp1
p/2
2p 1
p
p/2
tp+1 3p 1
=0
t = t0 =
1/2
2
In view of 1 2 3 and 3 2 1 = 1,
1/2
2
3/2
= (22 1 )1/2 1
3/2
3/2
and hence t0 (1 , 1). Since

w1 (t) =
p(p 1)tp2
p/2
2p 1
p(p + 1)
p/2
tp+2 3p 1
> 0,
t (0, 1),
this extremum is a minimum at

3
2
1
= 1 , ,
,
1 (t0 ) = 1 ,
2 1
3 1
1 1
with
p/2 p
1
1
w1 (t0 ) = 2
+
3.
1
1
We have in view of the isotropy of the energy that W (F ) = w1 (t0 ) on

R,
Q,
R
SO(3)
,
,
K(t0 ) = Q diag 1 ,
1
1
3.4 The Polyconvex Envelope of the Energy
77
and the denition of quasiconvexity implies that W (F ) w1 (t0 ) for all points
in the quasiconvex hull of K(t0 ) which is given by Theorem 2.7.5 by
1
1
W (F ) w1 (t0 ) if 1 (F ) = 1 , 3 (F ) ,
.
1
1
Since 1 (F ) = max (cof F ), this establishes the formula for W in I1 . The
remaining case, F = diag(1 , 2 , 3 ) with 3 3 is analogous.


We are now going to prove that the function W is in fact polyconvex and
therefore
W pc W rc W W pc
W = W pc = W rc .
The polyconvexity of W follows almost immediately from the following

proposition which asserts that the function is convex and nondecreasing.
Proposition 3.4.1. The function dened in (3.4) is convex and nondecreasing in its arguments.
Proof. In order to simplify the notation we dene the three functions
g1 : E1 I1 R,
g2 : S R,
g3 : E3 I3 R
by
1 p
+ 2(1 t)p/2 3,
1 t
1 p
t p
s p
g2 (s, t) =
+
+
3,
1 t
2 s
3
s p
3 p/2
+2
3.
g3 (s, t) =
3
s
g1 (s, t) =
We rst prove that is continuous. To see this, it suces to consider on

(E3 I3 ) and (E1 I1 ). If s = 3 , then g3 (3 , t) = 0, and along the curve
t = s we have
1 p/2
1 p/2
s p
+
+
3 = g3 (s, t).
1 2 s
1 2 s
3

Similarly, g1 (s, 11 ) = 0 and along the curve s = t/ we obtain
g2 (s, t) =
g2 (s, t) =
1 p
t p/2
t p/2
+
+
3 = g1 (s, t).
1 t
2 3
2 3
78
In order to prove that is nondecreasing, we calculate
Dg1 (s, t) = 0,
Dg2 (s, t) =
Dg3 (s, t) =
p
1p tp+1
p

p/2
+ p1 tp/21 ,
pt
psp1
p
ptp1
,
+
,
+
2p sp+1
3p
1p tp+1
2p sp
p/2
psp1

p3
+ p/2+1
,0 .
p
3
s
We obtain from these formulae that

t g1 (s, t) 0
s g2 (s, t) 0
t g2 (s, t) 0
1 3p/2
t3p/2
,
1
t p
s2p
,
t2p ( s)p ,
s g3 (s, t) 0
s3p/2 3
3p/2
All these inequalities are satised in the domains of the functions gi , and we
conclude that is nondecreasing in its arguments. We now show that is
continuously dierentiable. Since
1
) = 0,
1
we only need to check this along the curves t = s and t = s2 , respectively. A short calculation shows that
s g3 (3 , t) = 0,

2 1/2
s = Dg3 (s, t),
Dg2 s,
1
t g1 (s,
Dg2
3 1/2
t, t = Dg1 (s, t),
2
and this establishes the dierentiability of the function .

It remains to prove the convexity of . It is clear that g1 and g3 are convex
since the functions s sq and s sq are convex on R+ for q 1. We
obtain for g2 that
p(p 1)sp2
p2 tp1
p(p + 1)tp
p p+1
p sp+2 +
3p
2 s
2
D2 g2 (s, t) =
2
p1
p(p + 1) p(p 1)tp2

p t
p p+1
+
2 s
1p tp+2
2p sp
and thus the determinant of the matrix of the second derivatives is given by
p2 (p + 1)2 3p
p2 t2p2
p2 (p2 1)2p sp2
p2 (p 1)2 1p tp2
+
+
.
sp+2 t2
tp+2
s2
22p s2p+2
79
By assumption, 12 t and thus for p 2 and (p 1)2 1,

s
p2 (p 1)2 1p tp2
p2 t2p2
p2 1p tp2
p2 t2p2
2p
p p p 2 = 0.
2
2
s
s
3 2 t s
2 s2p+2
Since also (D2 g)11 > 0, we conclude that g2 is convex on its domain and this
nishes the proof of the proposition.

The next proposition is the analogue of the fact that convex and nondecreasing functions of convex functions are convex. We use this to show that
(max (F ), max (cof F )) is a polyconvex function since the maximal singular
value of a matrix is a convex function.
Proposition 3.4.2. Assume that : R2+ R is given by (3.4). Then the
function 1 : M33 R, given by

1 (F ) = max (F ), max (cof F )
is polyconvex.
Proof. By denition, 1 is polyconvex if there exists a convex function
g : M33 M33 R R
such that 1 (F ) = g(F, cof F, det F ). We dene

g(X, Y, ) = sup |Xe|, sup |Y e| .
eS2
eS2
It follows that for all matrices X1 , X2 , Y1 , Y2 M33 , scalars 1 , 2 R and

[ 0, 1 ]

g (X1 , Y1 , 1 ) + (1 )(X2 , Y2 , 2 )

= sup |(X1 + (1 )Y1 )e|, sup |(X2 + (1 )Y2 )e|
2
2
eS
eS

sup |X1 e| + (1 ) sup |Y1 e|, sup |X2 e| + (1 ) sup |Y2 e|
eS2
eS2
eS2
eS2

= sup |X1 e|, sup |X2 e| + (1 ) sup |Y1 e|, sup |Y2 e|
eS2
eS2
eS2
eS2

sup |X1 e|, sup |Y1 e| + (1 ) sup |X2 e|, sup |Y2 e|
2
2
2
2
eS
eS
eS
eS

= g X1 , Y1 , 1 + (1 )g X2 , Y2 , 2 .
Here we used the triangle inequality for the norm and the fact that is
nondecreasing in the rst inequality, and the convexity of for the second
inequality. This establishes the polyconvexity of 1 and concludes the proof.

We are now in a position to establish a characterization for the rank-one
convex and the polyconvex envelope of W .
80
Theorem 3.4.3. The rank-one convex and the polyconvex envelope of W

coincide and are given by
max (F ), max (cof F ) if det F = 1,

W rc (F ) = W pc (F ) =
+
else.
Proof. We dene
0
2 (F ) = I1 (det F )
where
I1 (t) =
if t = 1,
else.
+ (F ) = 1 (F ) + 2 (F ) is a polyconvex function which is nite only

Then W

on the set F M33 : det F = 1 . This implies that
1 (F ) = min (F ) =
1
max (cof F )
and
2 (F ) =
1
min (F )max (F )
whenever the energy is nite. In view of the denition of ,
0
if F L,

p
p/2
min (F )
1
+2
3
if F I1 ,
1
min F
+ (F ) = W (F )
W
if F S,
p/2

max (F ) p + 2
3
3 if F I3 ,
(F
)
max
+
else,
+ = W W rc . Therefore
and a comparison with (3.5) shows that W
+ W pc W rc ,
W rc W = W
and hence equality holds throughout this chain of inequalities. This proves
the assertion of the theorem.

3.5 The Quasiconvex Envelope of the Energy

The nal step in the construction is to prove that the quasiconvex envelope
is actually equal to the rank-one convex and the polyconvex envelope. This
ak.
is accomplished using the following construction by M
uller and Sver
3.5 The Quasiconvex Envelope of the Energy
81
Lemma 3.5.1. Let be given by

= {F Mmn : M (F ) = t},
where M is a minor of F and t = 0. Let V be an open set in , let F V rc ,
and let > 0. Then there exists a piecewise linear map u : Rn Rm
such that Du V rc a.e. in and
'
'
' {x : Du(x) V } ' < ||,
u(x) = F x on .
After these preparations, we prove the relaxation result in Theorem 3.2.1.
Proof of Theorem 3.2.1. We have to construct for all F M33 with
det F = 1 and for all > 0 a function F, W 1, (; R3 ) such that F, =
F x on and

W (DF, )dx ||W pc (F ) + O(),
where O() 0 as 0. This implies W qc (F ) W pc (F ), and since W pc

is quasiconvex, we conclude W qc = W pc .
We give the proof for the situation that W pc is obtained from W by
averaging with respect to laminates within laminates. If follows from Theorem 2.7.7 that there exist pairs (i , Fi )i=1,...,4 such that
F =
4

i Fi ,
and
W pc (F ) = W (Fi ), i = 1, . . . , 4.
i=1
(2) where K
= {F1 , F2 , F3 , F4 }. We choose
Moreover, F K

= F M33 : det F = 1 ,
and dene for > 0

< , = sup W (X) : X V W pc (F ).
V = F : dist(F, K)
Since W is continuous on we have 0 as 0. Lemma 3.5.1
guarantees the existence of a piecewise linear map F, : R3 with
DF, (x) Vrc a.e. and
'
'
F, (x) = F x on , and ' x : DF, (x) V ' ||.
Therefore, if M is an upper bound for W on V1 ,

'
'

W (DF, )dx ' DF, (x) V ' W pc (F ) + + M ||

||W pc (F ) + || + M .
The assertion of the theorem follows as 0.

4. Uniqueness and Stability of Microstructure
The numerical analysis of nite element minimizers of nonconvex variational

problems leads to the question of how one can quantitatively describe the
behavior of functions with small energy. Motivated by this problem, Luskin
analyzed in a sequence of papers uniqueness and stability of microstructures
corresponding to gradient Young measures for martensitic phase transformations in two and three dimensions. The goal of this chapter is two-fold.
First, we present a mathematically rigorous denition of Luskins intuitive
concept of stability. Secondly, we introduce a general framework for the analysis of stability based on an algebraic condition on the set K and the barycenter
F = , id. Starting from this condition, we develop a theory that claries
the relation between uniqueness and stability and identies simple laminates
as the only (known) class of microstructures that allows for a closed theory.
In particular, this theory includes all results in the literature and extends
them to an n-dimensional framework. It is based on a Young measure approach and underlines the general concepts underlying the estimates. We
nally prove that our results are optimal for two-dimensional problems and
we demonstrate the exibility of our method by extending it to the setting
of thin lms and to higher order laminates.
The main focus of our analysis is the identication of sucient conditions for uniqueness of microstructure. The key point of our approach is to
state estimates for nite element minimizers as a corollary of the uniqueness theory. In order to motivate the algebraic condition we dene below,
see Denition 4.1.4, we discuss a representative example in two dimensions
that actually appears in the analysis of the cubic to orthorhombic phase
transformation in three dimensions in Section 5.3. For , > 0 let
0
,
U1 =
0 +
3
2 2
,
U2 =
2 + 2

U3 =

3
2
3
2
and dene K = SO(2)U1 SO(2)U2 SO(2)U3 . Guided by the results in

Section 2.2 and by Figure 2.2, it is clear that the only microstructures that
are unique correspond to simple laminates for which the barycenter is a constrained point in K qc . By Step 6 in the proof of Theorem 2.2.3 these points
G. Dolzmann: LNM 1803, pp. 83126, 2003.

84
form the boundary of K qc which consists of maximal arcs between the matrices Ui . Consequently these points satisfy for some e S1 the inequality
|F e| = |Ui e| = |Uj e| > |Uk e|,
{i, j, k} = {1, 2, 3}.
This is exactly the condition (Cb ) in Denition 4.1.4 since in two dimensions
conditions on the cofactor of a matrix are equivalent to conditions on the
matrix itself.
The main results in this chapter can now be summarized as follows: We
prove in Section 4.1 for bulk materials (i.e., three-dimensional models) that

F K qc satises (Cb ) Mqc (K; F ) is unique is stable .
In the two-dimensional situation, our results are optimal in the sense that
F K qc satises (Cb ) Mqc (K; F ) is unique is stable.
We extend our theory in Section 4.4 to the analysis of thin lms and we
obtain again an optimal criterion for uniqueness and stability since
F K qc satises (Ctf ) Mqc (K; F ) is unique is stable.
The most detailed analysis is given in Section 4.1 for bulk materials in two and
three dimensions. We briey sketch the underlying ideas in the framework of
the three-well problem mentioned above. Suppose that
F = Q2 U2 + (1 )Q3 U3 ,
Q2 U2 Q3 U3 = a b,
is the center of mass of a simple laminate supported on

K = SO(2)U1 SO(2)U2 SO(2)U3
and that the twinning system corresponds to one of the twinning systems that
lead to a unique microstructure (i.e., the center of mass F is a constrained
point in K pc ). Then
|F e| = |U2 e| = |U3 e| > |U1 e|
for some e S1 , see the example following Proposition 2.2.4. Suppose that
Mpc (K; F ) and that is represented by
= 1 1 + 2 2 + 3 3 ,
Then
i P(SO(2)Ui ), i 0, 1 + 2 + 3 = 1.
|F e|
|Ae|d(A) = 1 |U1 e| + 2 |U2 e| + 3 |U3 e| |F e|

supp
and thus 1 = 0. Hence
(4.1)
85
= Q2 U2 + (1 )Q3 U3
is the unique gradient Young measure with center of mass equal to F . Now let
d(, ) be a distance on the set of all probability measures. The foregoing observation can be stated equivalently by saying that for every Mpc (K; F )
with

dist(A, K)d(A) = 0 d(, ) = 0.
M22
The main focus of this chapter is to generalize this result to probability

measures Mpc (M22 ; F ) (i.e., to probability measures supported on the
space of all 22 matrices, not only on K) and to prove the following stability
estimate: For all > 0 there exists a > 0 such that

dist(A, K)d(A) < d(, ) < .
(4.2)
M22
The proof of this result combines two fundamental estimates:

i) If the integral in (4.2) is small, then the support of is close to K.
Moreover, a suitable perturbation of the inequality (4.1) implies that the
mass of close to SO(2)U1 is small. For the mathematical statement of this
idea we dene the nearest point projection : M22 K by
|X (X)| = dist(X, K)
for all X M22 .
The goal is then to derive an estimate of the type

X supp : (X) SO(2)U1 c
dist(A, K)d(A).
M22
In particular, if the support of is contained in K, then there is no mass on

the well SO(2)U1 .
ii) The second estimate provides information about the sets

M = X supp : (X) SO(2)U , for = 2, 3.
The fact that the microstructure is unique implies that the points in the
set M = M2 M3 are close to the support of , i.e., to the matrices Q2 U2
and Q3 U3 . To state this precisely, we dene the so-called excess rotation
R(X) SO(2) for all X M by (X) = R(X)Q U if X M . The crucial
estimate can then be formulated by saying that R is close to the identity
matrix I in the sense that

2
|R(A) I| d(A)
dist(A, K)d(A).
M
M22
The precise formulations of these estimates are given in Propositions 4.1.9

and 4.1.12. In Sections 4.2-4.6 we extend these ideas to the analysis of thin
lm problems and second laminates in three dimensions. We conclude this
chapter by a review of the development of this approach to the numerical
analysis of microstructure and we relate our results to the existing literature.
86
4.1 Uniqueness and Stability in Bulk Materials

In this section, we present the general framework for the analysis of uniqueness and stability of microstructure supported on sets K of the from
K = SO(n)U1 . . . SO(n)Uk .
(4.3)
Throughout this section we assume that the matrices Ui Mnn are positive
denite with det Ui = > 0 for i = 1, . . . , k. Moreover, we assume that
SO(n)Ui = SO(n)Uj for i = j. We present applications of our approach to
several three-dimensional models of phase transformations in Chapter 5.
We begin with a precise denition of uniqueness. Since we try to identify the weakest possible conditions for uniqueness, we carry out most of
the analysis for polyconvex measures which form a much larger class than
gradient Young measures realizing microstructures. We dene uniqueness of
microstructures analogously.
Denition 4.1.1. A polyconvex measure Mpc (K) is said to be unique
if Mpc (K; , id) = {}. Similarly, a microstructure Mqc (K) is said to
be unique if Mqc (K; , id) = {}.
Assume now that is unique and let F = , id. The main motivation
behind our denition of stability is to give a quantitative statement of the
intuitive idea that any polyconvex measure with center of mass F and
support close to K should be close to in a suitable distance d on the space
of all probability measures. It turns out that a convenient way to measure
the distance of the support of to K is the integral

E() =

dist(A, K) + distmax{2,n1} (A, K) d(A).
Mnn
Here dist() denotes the Euclidean distance in Mnn . In order to dene the
distance d(, ) on the set of all probability measures supported on Mmn , let

M1 = M(Mmn ) : M 1
and let C0 (Mmn ) be the closure of C0 (Mmn ) in the supremum norm (in
our applications below, we have either m = n or m = 3 and n = 2). Choose
a dense set {i }iN of Lipschitz continuous functions in C0 (Mmn ) with
compact support and i 0 for all i N. We then dene a metric d(, ) that
induces the weak- topology on M1 by
d(, ) =

i=1
2i
|, i , i |
.
i 1,
We are now in a position to formulate our notion of stability.
87
Denition 4.1.2. A polyconvex measure Mpc (K) with underlying global

deformation F = , id is said to be stable if for all > 0 there exists a > 0
such that for all Mpc (Mmn ; F ) with

dist(A, K) + distmax{2,n1} (A, K) d(A)
Mmn
the estimate d(, ) holds.

The main ingredient in our analysis is the estimate in Proposition 4.1.9
below that estimates the mass of a measure Mpc (Mnn ; F ) close to the
wells SO(n)Uj , j = 1, . . . , k. Before we begin with the details, we introduce
the relevant notation for measures satisfying condition (Cb ); here and
are the indices in the denition of condition (Cb ), see Denition 4.1.4 below.
In particular, F = , id has a representation as the center of mass of a simple
laminate supported on SO(n)U SO(n)U , i.e., F = Q U + (1 )Q U
with Q , Q SO(n) and [ 0, 1 ]. We x Borel measurable functions (see
Proposition 4.1.3) : Mnn K with
|(X) X| = dist(X, K)
for all X Mnn ,
and : Mnn {Q U , Q U } with

|(X) X| = dist X, {Q U , Q U }
for all X Mnn .
(4.4)
(4.5)
Moreover, we may choose the projection without loss of generality in such

a way that (A) SO(n)U if (A) SO(n)U , for = , . The existence
of these functions is guaranteed by Proposition 4.1.3, and the rates in the
estimates below do not depend on the specic choice. For Mpc (Mnn )
we dene the set M of points in the support of that are close to SO(n)U
in the sense that

M = A supp : (A) SO(n)U , = 1, . . . , k,
+ of points close
see Figure 4.1. It is also convenient to introduce the sets M
to SO(n)U by

+ = A supp : (A) SO(n)U ,
M
= , ,
+ \ M has small measure.

and Proposition 4.1.9 shows that the dierence M
+ for a xed radius > 0
We nally introduce a local version of the sets M
according to

+, = A supp : (A) SO(n)U , |A Q U | < , = , .
M
For these sets we dene the corresponding volume fractions by
88
SO(n)U3
SO(n)U2
M3
M2
M1
SO(n)U1
Fig. 4.1. Sketch of the sets Mi dened by the projection for k = 3 (see the example
in the introduction to this chapter). The three wells are indicated by dots and the
projection denes three regions in the matrix space in which the nearest point in K
is contained in SO(n)Ui . The estimates on the volume fractions ensure that most
of the support of a polyconvex measure (indicated by the shaded areas) is close to
two of the wells.

= (M ) =
1 d(A),
(4.6)
M
and
%=

M
1 d(A),
% =

,
M
1 d(A).
(4.7)
M .
(4.8)
Finally, we set
M = M M ,
N=
$
{,}
To simplify the notation, we do not indicate explicitly the dependence of

in our formulae.
The construction of the projections and used the following fact.
Proposition 4.1.3. Assume that K Mmn is a compact set. Then there
exists a Borel measurable function : Mmn K such that
|(F ) F | = dist(F, K).
Proof. The following short proof is due to Kirchheim and Kristensen. Let
Dk be the family of all dyadic cubes of side length 2k in Mmn . We dene
(linear) orderings <k on each of the families D(k) in such a way that
1 , C
2 Dk , Ci C
i and C
1 <k C
2 implies C1 <k+1 C2 .
C1 , C2 Dk+1 , C
89
For a closed cube C let FC denote its corner minimizing |F |. We dene

k (F ) = FC(k,F ) ,
where

C(k, F ) = min C Dk : dist(F, K) = dist(F, K C) .
Since the sets {F : dist(F, K) = dist(F, K C)} are closed for all dyadic
cubes, it is easy to see that k is Borel measurable. Moreover, the denition
of the linear orderings implies that C(k + 1, F ) C(k, F ) and therefore
the functions k converge uniformly to a Borel measurable function that
satises |(F ) F | = dist(F, K) for all F Mmn .

Now we dene rst the condition (Cb ) on which our estimates are based,
and then we discuss the implications of the two hypotheses in (Cb ).
The Condition (Cb ). We begin with the denition of the condition (Cb )
that is at the heart of our analysis of uniqueness and stability of microstructure underlying a global deformation F in bulk materials.
Denition 4.1.4 (Condition (Cb ) for uniqueness in bulk materials).
Let U1 , . . . , Uk be positive denite n n matrices with det Ui = > 0 for
i = 1, . . . , k. Assume that K is given by
K = SO(n)U1 . . . SO(n)Uk ,
(4.9)
and that F K pc . Then F is said to satisfy condition (Cb ) if F has the

following properties:
i) If n 3, then there exist two indices and (not necessarily distinct)
and a simple laminate Mpc (K; F ) supported on SO(n)U SO(n)U .
ii) There exist two indices , {1, . . . , k} ( and not necessarily distinct and identical to the indices in i) for n 3) and a permutation
{1, . . . , k} \ {, } = {i1 , . . . , i } of the remaining indices with the following
property: for all j {1, . . . , } there exists a vector wj such that one of the
following two sets of inequalities holds:
|F wj | = |U wj | = |U wj |
max
{ij+1 ,...,i }
|U wj |,
and
|F wj | > |Uij wj |,
or
| cof F wj | = | cof U wj | = | cof U wj |
max
{ij+1 ,...,i }
and
| cof F wj | > | cof Uij wj |.
Here = k 2 if = and = k 1 if = .
| cof U wj |
90
Remark 4.1.5. Throughout the rest of the chapter we are mostly concerned
with the case = . The same arguments imply for = that the microstructure is supported on one well and thus a Dirac mass placed at a
single matrix in SO(n)U . In this case we refer also to a single Dirac mass as
a simple laminate in order to simplify the notation.
Remark 4.1.6. An apparently weaker formulation of condition (Cb ) would be
to allow dierent indices and in i) and ii) for n 3. However, the subsequent analysis shows that they are automatically identical and this justies
the formulation in Denition 4.1.4.
Remark 4.1.7. We show in Theorem 4.2.1 that ii) implies i) for n = 2.
In the next two propositions, we analyze the importance of the two assumptions i) and ii). The observation here is that the algebraic condition ii)
on the matrix F and the set K ensures that every polyconvex measure
Mpc (K; F ) is supported on the two wells SO(n)U SO(n)U , see
Proposition 4.1.9. The relevance of the rst condition, namely that F has
a representation as a barycenter of a simple laminate, lies in the fact that
simple laminates supported on two wells are unique, and this is the assertion
of the rst proposition.
Proposition 4.1.8. Assume that U1 , U2 Mnn with det U1 = det U2 > 0,
and that SO(n)U1 = SO(n)U2 . If F is the center of mass of a simple laminate
supported on the wells SO(n)U1 SO(n)U2 , then the Mpc (K; , id) = {}.
Proof. Assume rst that SO(n)U1 and SO(n)U2 are not rank-one connected and let F = Q1 U1 + (1 )Q2 U2 be the representation of F as the
barycenter of a simple laminate which exists by assumption. Since any polyconvex measure supported on two incompatible matrices has to be a single
Dirac mass, we conclude {0, 1}, and the assertion is immediate. We may
thus assume that Q1 U1 and Q2 U2 are rank-one connected, and we have to
show that Q1 , Q2 and are uniquely determined.
To simplify the notation, we let A = Q1 U1 , B = Q2 U2 , and we dene
a, b Rn by
Q1 U1 Q2 U2 = A B = a b.
(4.10)
Changing coordinates, we may associate with in a unique way a polyconvex

measure such that
supp SO(n) SO(n)AB 1 ,
F = F B 1 = AB 1 + (1 )I =
, id.
Step 1: The matrix AB 1 has one as an eigenvalue with algebraic multiplicity at least equal to n 2, i.e., there exist n 2 orthonormal eigenvectors
v i , i = 1, . . . , n 2, with
AB 1 v i = v i ,
i = 1, . . . , n 2.
91
Moreover, the vector a in (4.10) is orthogonal to v 1 , . . . , v n2 and satises

|F a|2 = |a|2 = |AB 1 a|2 .
It follows from det A = det(B + a b) = (det B)(1 + B 1 a, b) that
1
B a, b = 0 and thus
0 = B 1 a, ba = a, B T ba
= (a b)B 1 a = (A B)B 1 a = AB 1 a a,
and hence
F a = AB 1 a + (1 )a = a.
We conclude |F a| = |a| = |AB 1 a| and it only remains to construct the
eigenvectors v 1 , . . . , v n2 orthogonal to a. By the polar decomposition theorem, there exist R, Q SO(n) such that
RQAB 1 RT = = diag(1 , . . . , n )
with 0 < 1 2 n . Let
= RQRT ,
Q
= RQa,
a

b = RB T b.
=a
= a

Then A B = a b is equivalent to Q
b or Q
b and
T Q
= I = T T a

+ |
Q
b
b.
b
b T a
a|2
. Then T I is a matrix of rank one
Suppose rst that
b is parallel to T a
T

and has at least n1 eigenvalues equal to one. Since det = 1 it follows
that = I and this contradicts the assumption that the wells SO(n)A and
are not parallel.
SO(n)B are distinct. We may thus assume that
b and T a
In this situation, let
= span{
},
L
b, T a
=L
.
H
v = v
and since is diagonal with positive
if and only if v
H,
Then T
entries in the diagonal, we obtain
v = v

T
v = v

H.
v
Therefore has at least n 2 eigenvalues equal to one, and since there exists
a rank-one connection between and SO(n) we conclude that
= diag(1 , 1, . . . , 1, n ),
1 1 n .
Moreover,
= span {
= span {e2 , . . . , en1 },
} H
L
b, T a
92
is (a subspace of) the eigenspace corresponding to the eigenvalue

where H
This implies
one for .

or (QAB 1 )(RT v
) = RT v
for all v
H,
(RQAB 1 RT )
v=v
and thus
L = RT L
= span {B T Aa, B T b}.
v H = RT H
QAB 1 v = v
Hence B T b is perpendicular to H, and since QAB 1 Q = Qa B T b,

we deduce from the foregoing formula that
v = QAB 1 v = Qv
v H.
Therefore Q is the identity on H and hence maps L onto L. It remains to

v = v
for all
show that a is perpendicular to H. To see this, recall that
and thus
H,
v
v = T a

=
, v
= 0

a, v
a,
v H,
and hence in L.
Therefore RT a
is perpendicular to H
= Qa L and
i.e., a
the fact that Q maps L onto L implies that Qa L if and only if a L.
The choice of an orthonormal basis v 1 , . . . , v n2 of H concludes the proof of
Step 1.
= SO(n)U
1 SO(n)U
2 and F K
pc be the center of
Step 2: Let K
Assume that U
1 and U
2 are
mass of a simple laminate supported on K.
2 . Suppose that U
1 and U
2 have n 2
1 = det U
positive denite with det U
orthonormal common eigenvectors v i , i = 1, . . . , n 2, with corresponding
eigenvalues i . Assume, furthermore, that there exists an e Sn1 orthogonal
to v 1 , . . . , v n2 with
1 e|2 = |U
2 e|2 .
|F e|2 = |U
(4.11)
F ) = {
Then Mpc (K;
}.
We rst show that {F v 1 , . . . , F v n2 , F e} is an orthogonal system. It
i v i = i v i and the minors relations that
follows from U
|F v i |2 2i ,
1 )2
(det U
| cof F v i |2
,
2i
i = 1, . . . , n 2,
and consequently
'1
'2
i
1
2i
0 ' F v i
cof F v i ' = 2 |F v i |2 2 +
| cof F v i |2 0.
1
1 )2
i
i
det U
(det U
This implies
F v i =
93
2i
cof F v i ,
1
det U
thus F T F v i = 2i v i and hence

F v i , F v j = 2i ij ,
F v i , F e = 0,
i, j = 1, . . . , n 2.
Therefore
|F v i |2 = 2i = |Av i |2
for i = 1, . . . , n 2 and A supp , and thus

0
|F v i Av i |2 d(A)
supp

2

F v i , Av i d(A) +
|Av i |2 d(A)
= |F v i | 2
supp
supp

2

|Av i |2 d(A) = 0.
= |F v i | +
(4.12)
supp
In addition we obtain from (4.11) that

1 e|2 ,
1 e|2 = |F e|2
|Ae|2 d(A) |U
|U
supp
and hence equality holds throughout this inequality. We get

0
|F e Ae|2 d(A) = 0.
(4.13)
supp
Equations (4.12) and (4.13) show that for a.e. A supp the identities
Av i = F v i ,
Ae = F e
1 with Q SO(n). Then Q satises

hold. Assume rst that A = QU
1 v i = F v i ,
QU
1 e = F e,
QU
1 v i , U
1 e onto the
i.e., the rotation Q maps the n 1 orthogonal vectors U

n 1 orthogonal vectors F v i , F e. This uniquely determines Q and a similar
2 . Consequently the support of is uniquely
argument holds for A = QU
determined and is given by = Q1 U1 + (1 )Q2 U2 with [ 0, 1 ] and
Q1 , Q2 SO(n). To prove uniqueness of the volume fraction , assume that
there exist [ 0, 1 ], = , such that = Q1 U1 + (1 )Q2 U2 . Then
1 = SO(n)U
2 and this contradicts
1 Q2 U
2 ) = 0. Hence SO(n)U
( )(Q1 U
2 . The proof of the assertion of the
1 = SO(n)U
the assumption that SO(n)U
second step is thus complete.
The assertion of the proposition is now an immediate consequence of
Steps 1 and 2.

94
We now prove that hypothesis ii) in condition (Cb ) implies that the support of any Mpc (K; F ) is contained in the two wells SO(n)U SO(n)U .
This result is a special case of the proposition below since E() = 0 for measures with support in K. The more general statement which we provide here
is an important ingredient in the applications to stability and error estimates
for nite element minimizers in Section 4.5.
Proposition 4.1.9. Assume that K is given by (4.3), that F satises condition ii) in (Cb ) and that Mpc (Mnn ; F ). Then there exists a constant
c that depends only on F and n, but not on such that
cE()
for all {1, . . . , k} \ {, }.
(4.14)
In particular, if Mpc (K; F ) then is supported on SO(n)U SO(n)U .

Remark 4.1.10. The proof shows that the estimate for the volume fractions
can be improved to

c
dist(A, K) + dist2 (A, K) d(A), {1, . . . , k} \ {, }
supp
if the condition on the cofactor in condition (Cb ) is not used in the estimates.
The additional term involving distn1 (A, K) reects the fact that the cofactor
is a polynomial of degree n 1 in the entries of A. The same is true for the
estimates below that are based on these estimates for the volume fractions.
Proof. By denition of , we have for all w Sn1

k

2
2
|F w| |U w| =
=1

=

|F w|2 |(A)w|2 d(A)
supp

2F w, Aw (A)w |(A)w F w|2 d(A),
supp
and thus

k

|F w|2 |U w|2 +
=1
|(A)w F w|2 d(A)
supp
2|F w|
(4.15)
dist(A, K)d(A) cE().

supp
We obtain similarly from the minors relation for the cofactor that
k

| cof F w|2 | cof U w|2
=1

2 cof F w, cof Aw cof((A))w | cof((A))w cof F w|2 d(A).
supp
95
In view of the expansion

cof(A + B) =
n1

Li (A, B) = cof B +
i=0
n1

Li (A, B)
i=1
with functions Li (A, B) that are homogeneous of degree i in A and of degree

n 1 i in B we deduce that
'
'
'
'
' cof((A)) cof A' = ' cof((A)) cof A (A) + (A) '
c
n1

'
' '
'
'A (A)'i '(A)'n1i .
i=1
Since ai a + an1 for i = 2, . . . , n 2 and a 0, we conclude

k

| cof F w|2 | cof U w|2
=1

| cof (A)w cof F w|2 d(A) cE().
(4.16)
supp
The proof follows now by induction. We may assume that = k 1, = k

and ij = j for j = 1, . . . , k 2. Suppose that for j = 1 the assumption on F
in condition ii) in (Cb ) is satised. Thus there exists a vector w1 such that
|F w1 |2 |U1 w1 |2 > 0 and |F w1 |2 |Uj w1 |2 0,
j = 2, . . . , k.
It follows from (4.15) that 1 cE() with a constant c that depends only
on K and F . We conclude similarly with inequality (4.16) if the assumption
on the cofactor holds. Assume now that the assertion has been proved for
= 1, . . . , j 1 < k 2, and that the assumption on F in condition ii) in
(Cb ) is satised for the index j. Thus there exists a vector wj such that
|F wj | > |Uj wj | and |F wj | > |U wj |,
= j + 1, . . . , k.
It follows from (4.15) and the estimates for , = 1, . . . , j 1 that

k

|F w|2 |U w|2 cE(),
=j
and hence j cE(). The conclusion is analogous if the condition on the

cofactor holds, and this proves the assertion of the proposition.
The proof of the proposition is now complete.

96
Uniqueness of Microstructure. After these preparations, we can now

state and prove our uniqueness theorem for polyconvex measures in bulk
materials.
Theorem 4.1.11. Let K be given by (4.3). Suppose that Mpc (K) and
that F = , id satises condition (Cb ). Then is unique, and

Mpc (K; , id) = {} = Q U + (1 )Q U
with [ 0, 1] and Q U Q U = a b.
Proof. By Proposition 4.1.9,

dist(A, K) + distmax{2,n1} (A, K) d(A) = 0

Mnn
and hence is supported on the two wells SO(n)U and SO(n)U . The assertion of the theorem is an immediate consequence of Proposition 4.1.8.

Stability of Microstructure. We now turn towards the analysis of stability
of microstructure. Throughout this section we assume that n 3, i.e., we
concentrate on the situation with the additional hypothesis in (Cb ) that
F K pc has a representation as the barycenter of a simple laminate given
by
= Q U + (1 )Q U ,
Q U Q U = a b
(4.17)
with Q , Q SO(n) and [ 0, 1 ]. Furthermore we tacitly assume that

= and {0, 1}. If = , then the polyconvex measure is supported
on SO(n)U and hence a Dirac mass placed at a single matrix. The adaption
of the proof in this case is obvious.
Suppose now that F = , id with as in (4.17) satisfying condition
(Cb ) and that Mpc (Mnn ; F ). The stability analysis for requires two
ingredients. Proposition 4.1.9 provides estimates for the volume fractions
and shows that is bounded by E() for {, }. In addition to this,
it is important to control the support of on the two wells SO(n)U and
SO(n)U and to have information about the part of the support that is not
close to the two points Q U and Q U . This is done in the next proposition
which provides bounds on the excess rotation R(A) SO(n) dened by
(A) = R(A)(A)
on M = M M .
Proposition 4.1.12. Assume that K is given by (4.3) and that F satises

condition (Cb ). Let Mpc (Mnn ; F ). Then

|R(A) I|2 d(A) cE().
(4.18)
supp
97
Q U
Fig. 4.2. Sketch of the estimate for the excess rotation. The set SO(n)U is represented by a circle. The projection maps the points in M onto SO(n)U . Proposition 4.1.12 ensures that the projected points are close to Q U , the points in the
support of the unique microstructure realizing the center of mass.
Proof. Let be the simple laminate with center of mass equal to F

in (4.17). Choose an orthonormal basis {w1 , . . . , wn1 , b} of Rn where b is
the vector in the representation (4.17) for , and let
W = span{F w1 , . . . , F wn1 }.
Fix a rotation Q SO(n) such that
QW = {x Rn : xn = 0},

I| since
and let R(A)
= QR(A)QT . It suces to estimate |R(A)

|R(A)
I|2 d(A) =
|R(A) I|2 d(A).
supp
supp
In view of the estimate (4.14), it suces to prove this estimate on M . Since

wi is orthogonal to b, we have F wi = Q U wi = Q U wi and therefore
R(A)F wi = (A)wi on M . Then (4.15) implies

'
'2
'
'

'(R(A)
'
'(QR(A)QT QQT )QF wi '2 d(A)
I)QF wi d(A) =
M
M
'
'
'(R(A) I)F wi '2 d(A)
=
M
'
'
'((A) F )wi '2 d(A)
cE(),
and by the denition of W ,
98
'
'2

'(R(A)
I)ei ' d(A) cE()
for i = 1, . . . , n 1.
(4.19)

In order to prove (4.18) it thus suces to establish that (R(A)
I)en is
2
bounded in L (M, d). Since cof Q = Q for all Q SO(n), we have that
jn = cof jn (R)
where cof jn (R)
is (up to the sign (1)j+n ) the determinant
R
obtained by deleting the jth row and the nth column of
of the submatrix of R

R. A cofactor expansion for this subdeterminant down the jth column shows
that for j < n all terms in the expansion contain at least one o-diagonal
(the diagonal element in the jth row has been crossed out), and
element of R

thus Rjn is bounded by E(), i.e.,

2
jn
(A)d(A) cE().
R
M
is given as a sum of s terms pk , each of which is a product

For j = n, cof nn R
11 R
(n1)(n1) be the term corresponding to
of n 1 factors. Let p1 = R
We may estimate
the product of the rst n 1 diagonal entries in R.
s
s

'
'2
nn 1|2 = '
pk 1' c|p1 1|2 + c
|pk |2 .
|R
k=1
(4.20)
k=2
Then
'
'
11 1) (R
(n1)(n1) 1)'2 + . . . + c|R(n1)(n1) 1|2 ,
|p1 1|2 c'(R
where the dots stand for products with less than n 1 terms. In view of
(4.19) we deduce

|p1 1|2 d(A) cE()
M
and the assertion follows now from (4.20) since the products pk with k 2
which is again bounded by (4.19)
contain at least one o-diagonal entry in R

With the estimates for the volume fractions and the excess rotation at
hand, we can prove the following list of estimates for microstructures with
small energy.
Theorem 4.1.13. Suppose that K is given by (4.3) and that F satises
condition (Cb ). Let Mpc (K; F ) given by (4.17) be the unique polyconvex
measure with center of mass F . Then the following estimates hold for all
Mpc (Mnn ; F ):
i) Estimate orthogonal to the layering normal: for w Sn1 with w, b = 0,

|Aw F w|2 d(A) cE().
supp
99
ii) Estimate for the distance to the support of :

|A (A)|2 d(A) cE().
supp
% and
% be dened as in (4.7). Then
iii) Let
% | cE 1/2 (),
|
% |
|
c 1/2
E ().

iv) For all f W 1, (Mnn ; R),

|, f , f | c Lip(f )E 1/2 ().
Proof. To prove i), note that by assumption |U w| = |U w| = |F w|
for all w Sn1 with w, b = 0. Thus by (4.15) and the estimates for the
volume fractions

k

'
'
|((A) F )w|2 d(A) cE() +
'|F w|2 |Uk w|2 ' cE(),
supp
=1
and the assertion follows by the triangle inequality,

|(A F )w|2 d(A)
supp

|(A (A))w|2 + |((A) F )w|2 d(A) cE().
2
supp
Here we use the fact that by denition |A (A)|2 = dist2 (A, K).
We observe for the proof of ii) that by the triangle inequality

2
|A (A)|2 + |(A) (A)|2 d(A).
|A (A)| d(A) 2
supp
supp
By the denition of M and N in (4.8) and the estimates for the volume
fractions, (N ) cE(). Thus by (4.18)

|(A) (A)|2 d(A)
supp

2
|(R(A) I)(A)| d(A) +
|(A) (A)|2 d(A) cE(),
=
M
and ii) follows easily.

The proof of iii) is an immediate consequence of ii). In fact,
'
'
% U F '
% | |Q U Q U | = '%
Q U + (1 )Q
|
'
'
1/2
'
'
='
((A) A)d(A)'
|(A) A|2 d(A)
.
supp
supp
100
To prove the second assertion in iii), we observe that

% | |Q U Q U |
|
% | |Q U Q U | + |
% |
% |Q U Q U |.
|
By denition,
%
% =
A supp : (A) SO(n)U , |A Q U |
and therefore
%
% 1

M
|A Q U |d(A)
c 1/2
E ().

To prove iv), we estimate

'
'
'
' '
'
', f , f ' ', f , f ' + ', f , f '
'
'
1/2
Lip(f ), |id |2 + ', f , f '.
The rst term is bounded by ii), and the second can be estimated by
'
'
', f , f '
'
'
% (Q U ) f (Q U ) (1 )f (Q U )'
'%
f (Q U ) + (1 )f
% Lip(f ) |Q U Q U |.
| |
The assertion follows now from the two foregoing inequalities and iii).

The following stability theorem is an immediate consequence of the foregoing estimates.
Theorem 4.1.14. Suppose that Mpc (K) and that F = , id satises
condition (Cb ). Then is stable.
Proof. We have to prove that for all > 0 there exists a > 0 such that
for all Mpc (Mnn ; F ) with E() the estimate d(, ) holds. By
the denition of d(, ) and Theorem 4.1.13,
d(, ) =
2i
i=1
i=1
|, i , i |
i 1,
21
i 1, E 1/2 ()
cE 1/2 ().
i 1,

The assertion follows therefore for (/c)2 .
Equivalence of uniqueness and stability of microstructure is now an immediate consequence of the foregoing analysis.
4.2 Equivalence of Uniqueness and Stability in 2D
101
Corollary 4.1.15. Assume that K is given by (4.3), and that is a microstructure supported on K, i.e., Mqc (K; F ). Then
is stable
is unique.
Remark 4.1.16. If F satises condition (Cb ), then we obtain that uniqueness

and stability are equivalent.
Proof. Assume the contrary. Then there exists a gradient Young measures Mqc (K; F ) that is stable, but not unique. We may thus choose
Mqc (K; F ) with = . Since has compact support, we nd by
Zhangs Theorem on the truncation of sequences generating gradient Young
measures a sequence of Lipschitz functions ui with uniformly bounded Lipschitz constants such that the sequence Dui generates the Young measure .
Let i = Av Dui () . By construction, i as i . It is easy to see that

for all p 1

distp (Dui , K)dx 0 for i .
Hence E(i ) tends to zero as i . By assumption = , and thus there

exists an > 0 such that

d Av Dui () ,
for i suciently large. This contradicts the denition of stability and concludes the proof.

4.2 Equivalence of Uniqueness and Stability in 2D

The analysis of uniqueness and stability of microstructure in Section 4.1
relied on the result in Proposition 4.1.8 that simple laminates supported
on two wells are uniquely determined from their center of mass. In general,
it is an open problem to decide whether uniqueness of a microstructure
implies that is a simple laminate. However, this is true in two dimensions
and in this case we obtain an optimal characterization of stability. A further
simplication is that the condition (Cb ) reduces to conditions on F since
| cof F e| = |F e |. We assume in this section that
K = SO(2)U1 . . . SO(2)Uk
(4.21)
where the matrices Ui are positive denite with det Ui = for i = 1, . . . , k.

Theorem 4.2.1. Suppose that Mpc (K). Then is unique if and only
if F = , id satises condition (Cb ).
102
Proof. In view of Theorem 4.1.11, we only need to prove that uniqueness

of a polyconvex measure implies that F = , id satises condition (Cb ).
Assume thus that is unique. Then F cannot be an unconstrained point in K
since the splitting method generates a family of distinct polyconvex measures
with barycenter F by decomposing F into simple laminates supported on the
set of constrained points along directions in the rank-one cone.
Suppose now that is a constrained point. Let C be the set of corners
as dened in (2.17), and set U = {U1 . . . , Uk }. By Step 6 in the proof of
Theorem 2.2.3 we may assume that F is contained in a maximal arc pq (e)
with Up , Uq C and e S1 . It suces to show that uniqueness of implies
|Up e| = |Uq e| = |F e| >
max
U U \{Up ,Uq }
|U e|.
If strict inequality does not hold, then there exists an Ur U \ {Up , Uq } such
that
|Up e| = |Uq e| = |Ur e| = |F e|,
and Proposition 2.2.4 implies that we may assume without loss of generality
that Ur is contained in the polyconvex hull of SO(2)Up SO(2)Uq . We then
nd by the construction in (2.28) in the proof of Theorem 2.2.3 rotations
Qq , Qr , QF SO(2) and scalars tr , tF [ 0, 1 ] with tr = 0, tF = 0, such
that
QF F = (1 F )Up + F Qq Uq ,
Qr Ur = (1 r )Up + r Qq Uq .
If r = F , then F = QTF Qr Ur and we are done. If r = F , we may assume

that F < r and in this case
QF F =
r F
F
Up +
Qr Ur ,
r
r
and we conclude again that is not unique. This establishes the assertion of
the theorem.

We summarize the results in the following corollary.
Corollary 4.2.2. Let K be given by (4.21) and Mqc (K). Then
stable
unique
F = , id satises (Cb ).
Proof. This is an immediate consequence of Theorems 4.1.11 and 4.2.1.

4.3 The Case of O(2) Invariant Sets

We have seen in Section 2.3 that the analysis of thin lm problems is closely
related to the analysis of sets invariant under O(2). Therefore we begin with
103
the extension of the analysis of uniqueness and stability to sets with a multiwell structure invariant under O(2). We assume that
K = O(2)U1 . . . O(2)Uk
(4.22)
where the matrices Ui are positive denite and satisfy det Ui = > 0 for
i = 1, . . . , k. Furthermore we suppose that O(2)Ui = O(2)Uj for i = j. As a
rst step we formulate the analogue of condition (Cb ) for bulk materials for
the case of O(2) invariant sets.
Denition 4.3.1. Assume that K is given by (4.22) and that F M22 .
b ) if one of the following criteria is
Then F is said to satisfy condition (C
satised:
i) | det F | = and (Cb ) holds for F with K = SO(2)U1 . . .SO(2)Uk where
F = F if det F = and F = diag(1, 1)F if det F = ;
ii) there exists an {1, . . . , k} and an e S1 such that
|F e| = |U e| > max |Uj e|;
j=
iii) may relabel the matrices such that there exists an {1, . . . , k}, 2,
and an e S1 such that
|F e| = |U1 e| = . . . = |U e| > max |Uj e|.
j+1
Additionally, for some {1, . . . , } there exists a vector v S1 with

|U v| >
max
j{1,...,k}, j=
|Uj v|,
and
U e, U e = F e, F e .
Remark 4.3.2. The three cases are not exclusive, for example F = Ui satises
the assumptions in i), ii), and iii).
As in the case of bulk materials, our analysis reveals that uniqueness and
stability of polyconvex measures is closely related to simple laminates. We
b ).
show that Mpc (K; F ) is a simple laminate if satises condition (C
We represent this simple laminate by
= Q U + (1 )Q U ,
Q U Q U = a b
with [ 0, 1 ], Q , Q O(2), and and not necessarily distinct since

the support of can be contained in O(2)U . Throughout this section, we
assume that {0, 1}, i.e., that is not a Dirac mass placed at a single
matrix. The modications in the case of a Dirac mass placed at one matrix
in K are obvious since all the three cases are included in i).
By Proposition 4.1.3 we may choose as before Borel measurable projections : M22 K with
104
|(X) X| = dist(X, K)
for all X M22 ,
and : M22 {Q U , Q U } with

|(X) X| = dist X, {Q U , Q U }
for all X M22 .
Moreover, we may assume that (A) SO(2)U if (A) SO(2)Q U , for

= , . As before, the rates in the estimates below do not depend on our
specic choice. For Mpc (M22 ) we dene the sets (see also Section 4.1)

= 1, . . . , k,
M = A supp : (A) O(2)U , det (A) = ,

+ = A supp : (A) SO(2)Q U ,
= , ,
M

+, = A supp : (A) SO(2)Q U , |A Q U | < , = , ,
M
and the volume fractions
= (M ),
and
%=

M
1 d(A),
= + + ,
% =

,
M
1 d(A).
To simplify the notation, we do not indicate explicitly the dependence on

in our formulae. Finally, we set

M = A supp : (A) SO(2)Q U , = , ,
= (M ), = , ,
and
M = M M ,
N = supp \ M.
The energy E is now given by

dist(A, K) + dist2 (A, K) d(A),
E() =
supp
and we let
b ),
0 if condition i) or ii) holds in (C
b ).
1 if condition iii) holds in (C
Proposition 4.3.3. Suppose that Mpc (M22 ; F ) and that F satises

b ). Then
condition (C

c E() + E 1/2 ()
for
105

1 , . . . , k \ , .

b ) with
Moreover, { , } = { , } if assumption i) holds in condition (C
det F = and { , } = { } if assumptions ii) and iii) hold, respectively. Finally,

|R(A) I|2 d(A) c E() + E 1/2 () .
supp
Proof. The estimates are based on three fundamental inequalities. The

rst inequality was already derived in (4.15) and yields for all w S1 that
k

j |F w|2 |Uj w|2 +

|(A)w F w|2 d(A) cE().
(4.23)
supp
j=1
The second estimate relies on the fact that the determinant is a nullLagrangian and that therefore
k

j ( det Uj ) det F =
j=1

det (A) det(A) d(A).
supp
We deduce from the expansion

'
' '
'
' '
' det(A) det (A)' ' det A (A) ' + ' cof((A)) : (A (A))'

c |A (A)|2 + |A (A)| ,
that
k
'
'
'
'
j ( det Uj ) det F ' cE().
'
(4.24)
j=1
The third estimate is in the same spirit as the two foregoing ones, but not
b ). Then
based on a null-Lagrangian. Assume that iii) holds in condition (C
k
'
'
'
'
j Uj e, Uj e F e, F e '
'
j=1
'
'

'
'
|(A)e, (A)e F e, Ae d(A)'
='
supp

'
' '
'
'((A) F )e, (A)e ' + 'F e, (A (A))e 'd(A)
supp

c
supp

1/2
|((A) F )e|2 d(A)
+c
|A (A)|d(A).
supp
We obtain from the denition of E() and (4.23) that
106

k
'
''

'
j Uj e, Uj e F e, F e ' c E() + E 1/2 () .
'
j=1
b ).
We now distinguish three cases corresponding to the three conditions in (C

Case 1: Condition i) holds in (Cb ). We obtain from (4.23) that j E()
for j {, } and from (4.24) and the arguments in Section 4.1, see in
particular the proof of Theorem 4.1.9, that j cE() if det F = . This
proves the assertion.
b ). In this case we conclude from (4.23)
Case 2: Condition ii) holds in (C
that j E() for j = .
b ). We rst derive some algebraic inforCase 3: Condition iii) holds in (C
mation about the matrices U1 , . . . , U (here has been dened in the state b ); we relabel the matrices again, if necessary). Proposiment of condition (C
tion 2.2.4 ensures the existence of rotations Qj SO(2) and scalars j R,
j = 2, . . . , such that
0 < 2 < . . . < .
In particular, U2 , . . . , U1 (SO(2)U1 SO(2)U )qc and

Uj e, Uj e = Qj Uj e, Qj Uj e = U1 e, U1 e + j |U1 e|2 .
We now turn to the estimates for the volume fractions. By (4.23)
j E()
for j = + 1, . . . , k.
The assumption |U v| > maxj= |Uj v| implies that = 1 or = . We may

assume that = 1. Then Uj e, Uj e > F e, F e for j = 2, . . . , and thus

'
'
j 'Uj e, Uj e F e, F e '
j=2

'
''
'
j Uj e, Uj e F e, F e '
='
j=2
k
'
''
'
'
j Uj e, Uj e F e, F e ' + cE()
j=1
c E() + E 1/2 () .
It remains to prove the estimate for the excess rotations R. This is analogous
to the proof of Proposition 4.1.12. By assumption,
F b = Q U b = Q U b = (A)b
and thus by (4.23)
for A M
|(R(A) I)F b |2 d(A) =
107
|((A) F )b |2 d(A) cE().
Since |(R(A) I)(F b ) | = |(R(A) I)F b | and (N ) c(E() + E 1/2 ()),

the assertion follows easily.

b ) is necessary
We now state the main result that ensures that condition (C
and sucient for uniqueness.
Theorem 4.3.4. Let K be given by (4.22). Then Mpc (K) is unique if
b ).
and only if F = , id satises condition (C
b ).
Proof. We assume rst that F satises condition (C

Case 1: Condition i) holds in (Cb ). We may assume that det F = , and
thus we are in the setting of the bulk materials in Section 4.1; the uniqueness
follows as in Theorem 4.1.11.
b ). It follows from E() = 0 and PropoCase 2: Condition ii) holds in (C
sition 4.3.3 that supp O(2)U . Moreover,
'
'
'
'
Aed(A)'
|Ae|d(A) |U e|,
|F e| = '
O(2)U
O(2)U
and therefore in view of the assumption |F e| = |U e|

|F e Ae|2 d(A) 0.
0
supp
We deduce that Ae = F e for almost all A supp . There are exactly two
matrices Q O(2) with Q U e = F e and consequently
= Q+ U + (1 )Q U .
The minors relation det F = 2 1 uniquely determines and hence is
unique.
b ). This is analogous to Case 2 since by
Case 3: Condition iii) holds in (C
Proposition 4.3.3 the measure is supported on O(2)U .
We now turn to the converse implication that uniqueness of implies the
b ) for F = , id. By the splitting method, cannot
validity of condition (C
be an unconstrained point. We may thus assume that F is a constrained
point. For | det F | = the argument is identical to one in the proof of
Theorem 4.2.1 for bulk materials. Suppose now that | det F | < , and that
there exists an {1, . . . , k}, 2, and an e S1 such that
|F e| = |U1 e| = . . . = |U e| > max |Uj e|.
j+1
Note that we only need to consider the case 2 since = 1 corresponds

b ). We argue by contradiction. Thus assume that
to case ii) in condition (C
there does not exist an i {1, . . . , } and a v S1 such that
108
|Ui v| > max |Uj v|,
Ui e, Ui e = F e, F e .
and
j=i
(4.25)
As in the proof of Proposition 2.2.4 we nd Qj SO(2)Uj and j R for

j = 2, . . . , with 0 < 2 < . . . < and
Qj Uj U1 = j U1 e e , Uj e, Uj e = U1 e, U1 e + j |U1 e|2 . (4.26)
Finally there exists QF SO(2) and F R such that
QF F U1 = F U1 e e ,
F e, F e = U1 e, U1 e + F |U1 e|2 .
This implies U2 , . . . , U1 (SO(2)U1 SO(2)U )qc , and F (0, ). Indeed,

if F = 0 or F = , then
F e, F e = U1 e, U1 e or F e, F e = U e, U e ,
respectively. Since U1 and U are corners of the set K in the sense of Theorem 2.2.3, we conclude that U1 or U satisfy (4.25), contradicting our assumption that none of the matrices satises this condition. The goal is
now to construct a one-parameter family of polyconvex measures supported
on SO(2)U1 SO(2)U with center of mass F . There are unique rotations
Q
1 , Q O(2) such that Q1 U1 e = F e and Q U1 e = F e. Moreover, for all
2
1 , [ 0, 1 ] we nd vectors a, b, c R with
Q+
1 U1 Q1 U1 = a e ,
Q+
U Q U = b e ,
and
V1 V = c e ,
where
!
Vj =
+
j
+
j +
Q+
j Uj
j
+
j +
"
Q
j Uj
j = 1, .
Consequently,
= +
1 Q+ U1 + 1 Q U1 + Q+ U + Q U
1
is a polyconvex measure. We now assert that

Mpc (F ) = = +
1 Q+ U1 + 1 Q U1 + Q+ U + Q U :
1
1 , [ 0, 1 ], 1 + 1 + + = 1,
+
(+
1 + ) (1 + ) = det F,
(+
1 + 1 )U1 e, U1 e + ( + )U e, U e = F e, F e
4.4 Applications to Thin Films
109
is a one-parameter family of measures. By construction, all measures are

second laminates and thus polyconvex measures. We only need to show that
, id = F . Let X = , id. Then
Xe = F e,
Xe , F e = F e , F e,
det X = det F,
and consequently
Xe , (F e) = Xe , (Xe) = det X = det F = F e , (F e) .
We obtain that Xe = F e and hence X = F . The assertion follows now if
we can solve the linear system
1+
det F

1 1
F
1 =
1 1 1 1
1

where we used the notation x = Ux e, Ux e for x {1, , F }. The augmented matrix of this system is row equivalent to
1 0 0 1 1 1 2
0 1 0 1
1
0 0 1 1
2
with
1 =
F 1
(0, 1),
1
2 =
det F
1
1
(0, 1),
2
and the general solution is given by
1 1 2
1
+s
s
2
0
1
1
1 .
1
The solutions
1 , satisfy the constraints 1 , [ 0, 1 ] for
s [ max{0, 1 + 2 1}, min{1 , 2 } ]

and this concludes the proof of the theorem.


In this section, we extend the results presented so far for bulk materials to
thin lms. If the normal to the thin lm is suitably oriented with respect to
the crystallographic directions in the material, then the set K is given by
110
K = O(2, 3)U1 . . . O(2, 3)Uk
(4.27)
where U1 , . . . , U M22 satisfy det Ui = > 0 for i = 1, . . . , k. We recall

from Section 2.3 that all F K qc can be written as
qc
F = Q%
(F%) with Q SO(3), F% SO(2)U1 . . . SO(2)Uk
and
% : M22 M32 dened by
F11 F12
F11 F12
%) = F F .

%
(
F
F% =
21 22
F21 F22
0 0
The results in this section show that as is the case of the two-dimensional
theories in Sections 4.2 and 4.3 uniqueness and stability are equivalent for
thin lms. The main dierence is the at rst sight surprising result that
the microstructures underlying globally ane deformations F K qc are not
unique unless F is area preserving. This is a consequence of the extraordinarily rich folding patterns for thin lms in three dimensions and can be nicely
illustrated by the following example.
Let U1 = diag(, ) and F =
%(diag(, t)) with t (1, 1). We dene
(0, 1) by 2 1 = t and (0, ) by cos = t. Then
F =
0
0
0
0
0
1
1
+ (1 ) 0 = 2 0 cos + 2 0 cos
,
0
0 0
0 sin
0 sin
and we nd two laminates supported on K with center of mass F since
0
0
0 cos = 0 cos U1 O(2, 3)U1 .
0 sin
0 sin
The rst construction is a two-dimensional one, but the second construction
is the limit of genuinely three-dimensional folding patterns. This demonstrates that the behavior of thin lms is qualitatively dierent from the twodimensional setting in which
111
O(2, 3)U1 qc
0 t
determines a unique microstructure.
We begin our analysis with the denition of condition (Ctf ) for thin lms
which replaces condition (Cb ) for bulk materials.
Denition 4.4.1. Assume that K is given by (4.27). Then F K pc (K) is
said to satisfy condition (Ctf ) if det(F T F ) = 2 and if there exists an e S1
and , {1, . . . , k} such that
|F e| = |U e| = |U e| >
max
j{1,...,k}\{,}
|Uj e|.
We dene Borel measurable projections and as well as the volume

fractions with the obvious modications analogously to (4.4)-(4.8). More precisely, if (A) O(2, 3)U , then we dene the excess rotation in SO(3) in the
following way: let S = (A)Q1
, then R(A) is the matrix with the columns
R(A) = (S%
e1 , S%
e2 , S%
e1 S%
e2 ).
In this section, we dene the energy E by

E() =
dist(A, K) + dist2 (A, K) d(A).
supp
Proposition 4.4.2. Suppose that Mpc (M32 ) and that F = , id satises condition (Ctf ). Then
cE()
and
for {1, . . . , k} \ {, },

|R(A) I|2 d(A) c E() + E 1/2 () .
supp
Moreover, the microstructure underlying F is unique, i.e.,

Mpc (K; F ) = {Q(Q U ) + (1 )Q(Q U )
with Q SO(3), Q , Q SO(2) such that rank(Q U Q U ) = 1, and
[ 0, 1 ].
Proof. We may assume that F =
%(F%) with F% M22 and det F% = .
As in the proof of Proposition 4.1.9,

k

|F w|2 |U w|2 +
=1
supp
|((A) F )w|2 d(A) cE(),
112
and the bounds on the volume fractions follow from (Ctf ). If Mpc (K; F ),
then E() = 0 and is supported on O(2, 3)U O(2, 3)U . We rst show
that

(4.28)
% SO(2)U SO(2)U .
supp
Let adjij (A) denote the determinant of the 2 2 matrix formed by the ith
and the jth row in A M32 . Then

= adj12 (F )
| adj12 (A)|d(A) ,
supp
and thus adj12 (A) = for almost all A supp . A short calculation shows
that
det(F T F ) = adj212 (F ) + adj223 (F ) + adj213 (F )
for all F M32 ,
and hence F det(F T F ) is a polyconvex function. We obtain

adj212 (A) + adj223 (A) + adj213 (A) d(A) 2 ,
2 = det(F T F )
supp
and consequently adj13 (A) = adj23 (A) = 0 for almost all A supp . We
deduce that for almost all A supp the third row of A has to be parallel
to the rst and the second row and therefore the third row must be equal
to the zero vector. This establishes (4.28). The analysis has therefore been
reduced to the case of two-dimensional bulk materials and we conclude by
Theorem 4.1.1 that the polyconvex measures are unique,
= (Q U ) + (1 )(Q U ) ,
Q U Q U = a b
(4.29)
with a, b R2 , [ 0, 1 ] and Q SO(2) for = , .

We now turn to the proof of the estimate for the excess rotation R(A).
For simplicity, we frequently write R instead of R(A). We write ei for the
% SO(2)
ei for the standard basis in R2 . Let Q
standard basis in R3 and %
% F%b = |F%b |%
% 1) SO(3).
be the rotation with Q
e1 , and dene Q = diag(Q,
I| where
As in the proof of Proposition 4.1.12, it suces to estimate |R
T

%
R = QRQ . By denition, F b = Q U b = Q U b and thus we deduce
that R(A)F b = (A)b on M and that

|(R(A)
I)e1 |2 d(A)
|F%b |2
M

|(QR(A)QT I)QF b |2 d(A)
=
M

(4.30)
2
=
|(R(A) I)F b | d(A)
M
=
|((A) F )b |2 d(A) E().
M
113
1 S2 , this estimate can be improved to an L1 estimate for R

11 1.
Since Re
In fact,

11 1)2 + 1 R
11 1)2 + R
11 = 1 (R
2 = 1 (R
2 + R
2 ,
1R
11
21
31
2
2
and thus by (4.30)

11 |d(A) 1
I)e1 |2 d(A) cE 1/2 ().
|1 R
|(R
(4.31)
2
M
I)e2 .
The crucial estimate that requires some care is the estimate for (R
We suppress the dependence on A in the following calculations. Since Q is a
block diagonal matrix,
21 R

22 R
12 = cof 33 (R)
11 R
R
R11 R12

%
%T
Q
= det Q
R21 R22

% 2 R11 R22 R21 R12 ,
= (det Q)
and

R11 R12

%
Q
adj12 (A) = det
U
R21 R22

= R11 R22 R12 R21 on M.
% = 1 that
Hence we infer from det Q

21 R
11 R
22 R
12 = 1 adj12 (A)
R
on M,
and hence
adj12 F = R33 = R11 R22 R12 R21 = adj12 (A).
We obtain from the fact that adj12 () is a null-Lagrangian that

| adj12 F adj12 (A)|d(A)
M

=
adj12 F adj12 (A) d(A)
adj12 F adj12 (A) d(A)
supp
N

adj12 A adj12 (A) d(A)
=
supp
N

| adj12 A adj12 (A)|d(A) + c
.
supp
{,}
114
We have by (C.3) for all A, B M22 that

det A det B = det (A B) + B det B
= det(A B) + cof(A B) : B,
and we therefore conclude that

M
|A (A)|2 + |(A)| |A (A)| d(A) + cE() (4.32)

supp
cE().
In view of
11 R
12 R
22 ) + R
22 (1 R
11 ) + R
12 R
22 R
21 ) = (1 R
21 ,
1 (R
we infer
22 = R
22 (1 R
11 ) R
12 R
11 R
12 R
21 + 1 (R
22 R
21 )
1R

11 ) R
12 R
21 + 1 adj12 (F ) adj12 (A) .
22 (1 R
= R
We conclude in view of (4.32) and (4.31) that

22 1|d(A) 1
|R
M
M

2
11 1| + 1 R
2 d(A)
12
+
|R
(4.33)
+R
21
4
M
1 2

2 d(A).
12 + R
c E() + E 1/2 () +
R
21
4
M
2 S2 , and therefore
On the other hand, Re

2
2
2

22 |d(A)
(R12 + R32 )d(A) =
(1 R22 )d(A) 2
|1 R
M
M
M
1 2

2 d(A).
12 + 2R
c E() + E 1/2 () +
R
21
M 2
2 on the right hand side can now be absorbed on the
The term involving R
12
left hand side and we infer in view of (4.30)
1 2

2 d(A) c E() + E 1/2 () .
12 + R
R
32
2
M
This in turn implies, again in view of (4.33) that
4.5 Applications to Finite Element Minimizers
22 1|2 d(A) 2
|R
and
115

22 1|d(A) c E() + E 1/2 () ,
|R

I)e2 |2 d(A) c E(u) + E 1/2 (u) .
|(R
Finally,

2

1 Re
2 e1 e2 |2 d(A)
|(R I)e3 | d(A) =
|Re
M
M

1 e1 ) Re
2 + e1 (Re
2 e2 )|2 d(A)
=
|(Re
M

1 e1 |2 + |Re
2 e2 |2 d(A)
c
|Re
M

c E() + E 1/2 () .
This concludes the proof of the estimate for the excess rotation.

Theorem 4.4.3. Let K be given by (4.27) and Mpc (K). Then

F = , id satises (Ctf )
is unique.
Proof. We already proved in Proposition 4.4.2 that is unique if F

satises condition (Ctf ). Assume thus that is unique and let F = , id. We
may suppose that F =
%(F%) with F% M22 . It follows from the discussion
at the beginning of this section that det(F T F ) = 2 and we may assume
that det F% = . The proof of Proposition 4.4.1 shows that
%
supp
%(K)
% = SO(2)U SO(2)U ,
K
with
and we are therefore in the two-dimensional situation. In this case, condition

% and the assertion follows from The(Ctf ) is equivalent to (Cb ) for F% and K,
orem 4.2.1.

Based on the estimates for the volume fractions and the excess rotation
in Proposition 4.4.2, it is easy to deduce the following stability result.
Corollary 4.4.4. Let K be given by (4.27) and Mqc (K). Then
F = , id satises (Ctf )
is unique
is stable.

In this section, we discuss applications of the uniqueness and stability results
to nite element methods for nonconvex variational problems. Here we focus
116
on bulk materials since the adaption of the techniques to the case of thin
lms is straight forward.
Suppose that is a polygonal domain and that we want to minimize

1
J (u) =
W (Du)dx
(4.34)
||
in the class of admissible functions

AF = u W 1,max{2,n1} (; Rn ) : u(x) = F x on .
We assume furthermore that the zero set K of W has a multi-well structure (4.3) and that W satises the coercivity condition

W (X) dist(X, K) + distmax{2,n1} (X, K) , > 0.
(4.35)
The simplest nite element method for the numerical solution of the minimization problem is obtained by choosing a triangulation Th of and by minimizing J in AF Sh where Sh is the space of all continuous functions that
are piecewise ane on the elements in Th . The coercivity assumption (4.35)
implies the existence of a nite element minimizer uh Sh and the goal of
the numerical analysis is to describe the qualitative behavior of uh . In order to obtain rigorous estimates, one needs uniqueness of the minimizer and
this leads naturally to the situation in Section 4.1. We thus assume that F
satises condition (Cb ). Consequently, the inmum of J is not attained in
AF , but the underlying microstructure is unique by Theorem 4.1.11. The
question is therefore whether the oscillations in the nite element minimizer
Duh have the statistics recorded in . In order to compare Duh and we
follow Collins, Kinderlehrer, and Luskin and associate with Duh the gradient
Young measure

h = Duh (x) x
and pass to its average Av h which is dened via duality by

1
Duh (x) , dx
Av h , = Av Duh () , =
||

1
(Duh (x))dx for all C0 (Mnn ).
=
||
It turns out that the stability results in Section 4.1 together with some standard interpolation results lead to explicit error estimates for uh , see Corollary 4.5.2 below. We begin with a more general statement about functions
u with small energy that does not require that u be contained in a nite
element space. We dene for = , the sets

= x : (Du(x)) SO(3)U ,

, = x : (Du(x)) SO(3)U , |Du(x) Q U | < .
117
Theorem 4.5.1. Let > 0 and suppose that F satises condition (Cb ).
Assume that u AF with J (u) . Then there exists a constant c that
depends only on F , n, and such that the following assertions hold:
i) (estimates for directional derivatives tangential to the layering direction):
for all w Sn1 with w, b,

'
'
1
'(Du(x) F )w'2 dx c,
||
ii) (estimates for the deformation)

1
|u(x) F x|dx c,
||
iii) (total distance form the wells)

1
|Du (A)|2 dx c,
||
iv) (estimates for volume fractions)
' | |
' c
' ,
'
' 1/2 ,
'
||

' | |
'
'
'
' c 1/2 ,
'
||
v) (weak convergence, averages): the following estimate holds for all

with Lipschitz boundary,
'
'
'
'
' (Du F )dx' c(, ) 1/2 .
vi) (weak convergence, functions): for all f W 1, (Mnn )

' 1
''
'
f (Du) [ f (Q U ) + (1 )f (Q U ) ] dx' cf 1, 1/2 .
'
||
Proof. Let = Av Du(x) be the homogeneous gradient Young measure
associated with . Then

dist(A, K) + distmax{2,n1} (A, K) d(A)
E() =
supp

1
(4.36)
dist(Du, K) + distmax{2,n1} (Du, K) dx
=
||

1 1
W (Du)dx .
||
Proof of i) By statement i) in Theorem 4.1.13,

1
2
|(Du F )w| dx =
|Aw F w|2 d(A) cE() c.
||
supp
118
Proof of ii) Since is bounded in every direction, we may use Poincares

inequality

|u|2 dx c
|Du, w|2 dx for all w Sn1 , u W01,2 ()
for the component functions of u.

Proof of iii) This follows from statement ii) in Theorem 4.1.13.
%
Proof of iv) Let (A) = {(A)SO(n)U } . By denition of ,

1
| |
%
,
(A)d(A) =
(Du(x))dx =
=
||
||
supp
%, = |, |/||. The assertion is therefore a consequence of

and similarly
statement iii) in Theorem 4.1.13.
Proof of v) This is a consequence of the inequality

'
'

'
'
' Du dx' c() uL2 () + (uL2 () DuL2 () )1/2 , u W 1,2 (),
which follows from embedding theorems, see also (4.42) below.

Proof of vi) This follows from statement iv) in Theorem 4.1.13.

The foregoing general theorem implies the following explicit error estimate
for nite element minimizers.
Corollary 4.5.2. Let be a polygonal domain and Th a regular triangulation
in the sense of Ciarlet. Assume that F satises the condition (Cb ) and that
uh is a minimizer of J in A Sh . Then there exists a constant c which
depends only on F , n, and the shape of the triangles in Th , but not on h,
such that

1
|uh (x) F x|2 dx ch1/2 .
||
Moreover, the L2 norm of the error of the gradient is bounded in directions
w Sn1 orthogonal to the layering direction b by

'
'
1
'(Duh F )w'2 dx ch1/2 .
||
Finally,

d Av Duh (x) ch1/4 .
Proof. It is easy to see that there exists a nite element function uh with

1
W (Duh )dx ch1/2 .
||
119
The informal idea of this construction is the following. Consider for > 0
the function
u (x) = Q U x (x, b)a.
Here (x) = (x/) and : R R is the continuous function with (0) = 0
and
0 if s (0, )

(z + s) =
for all z Z.
1 if s (, 1)
Then Du {Q U , Q U } and u is ane on layers separated by ane
hyperplanes with normal b. The goal is now to choose for h > 0 xed in
such a way that the nodal interpolation of u onto Th has minimal energy.
This requires to balance the following two contributions to the energy.
(a) For each interface across which Du changes from Q U to Q U one
needs a neighborhood of diameter O(h) which contributes a term of order
O(h/) to the energy.
(b) The function u satises |u (x) F x| = O() and thus a boundary
layer along of width O() is required to interpolate u and the correct
boundary values F x.
This shows that one can construct a function u,h with energy

1
W (Du,h ) = O(h/) + O().
||
The choice = O( h) proves the assertion.

The estimates in the corollary follow now immediately from the assertions
in Theorem 4.5.1.

Remark 4.5.3. In the three-dimensional situation, it is natural to assume
quadratic growth of the energy W , i.e., to assume that
W (X) dist2 (X, K),
> 0.
In this case, we have to replace the estimate (4.36) by

1
dist(Du, K) + dist2 (Du, K) dx
E() =
||

1
1/2
1
dist2 (Du, K)dx

+
dist2 (Du, K)dx
||
||

1/2 1 1
1 1
W (Du)dx
+
W (Du)dx
||
||

c h1/2 + h1/4 .
120
This leads to the estimates

1
|uh (x) F x|2 dx c h1/2 + h1/4 ,
||

'
'

1
'(Duh F )w'2 dx c h1/2 + h1/4 ,
||
d Av Duh () , c h1/4 + h1/8 ,

which follow from Corollary 4.5.2.
4.6 Extensions to Higher Order Laminates

The theory developed in Section 4.1 provides a sucient condition for uniqueness and stability of microstructure which is essentially an algebraic condition
on the center of mass F and the set K. However, it only applies to simple
laminates and Corollary 4.1.15, which asserts that stability implies uniqueness, already indicates that extensions of the theory to higher order laminates
are subtle. It is an open question whether uniqueness is sucient for stability.
Theorem 4.2.1 shows that this is true for two-dimensional problems.
The advantage of the approach described here is that it clearly distinguishes between those arguments that rely on uniqueness and those that do
not. Therefore some of the estimates can be extended to higher order laminates and we demonstrate this for the case of the cubic to tetragonal phase
transformation in three dimensions. Suppose that 2 > 1 and dene
U1 = diag(2 , 1 , 1 ),
U2 = diag(1 , 2 , 1 ),
U3 = diag(1 , 1 , 2 ).
For the rest of this section we consider the set K given by

K = SO(3)U1 SO(3)U2 SO(3)U3 .
(4.37)
The polyconvex hull of two of the wells, say SO(3)U1 SO(3)U2 , is equal to
the second lamination convex hull and a polyconvex measure underlying a
global ane deformation F is unique if and only if F is a rst order laminate
(or a matrix in K). Iqbals results show that even the mass of on the two
wells is not uniquely determined from its center of mass.
Suppose now that Mpc (SO(3)U1 SO(3)U2 ) (in particular, could be
a second order laminate). It follows from the example following Theorem 2.5.1
that F = , id satises
| cof F e3 | = | cof U1 e3 | = | cof U2 e3 | > | cof U3 e3 |.
(4.38)
Moreover, if Mpc (K; F ) and if F satises (4.38), then F is the center of

mass of a second order laminate supported on SO(3)U1 SO(3)U2 . The strict
4.6 Extensions to Higher Order Laminates
121
inequality in (4.38) implies estimates analogously to those in Section 4.1. As

before, we dene

E() =
dist(A, K) + dist2 (A, K) d(A).
supp
Theorem 4.6.1. Suppose that F satises (4.38) and that Mpc (M33 ).
Then

|Ae3 F e3 |2 d(A) cE().
supp
Proof. We conclude as in the proof of Proposition 4.1.9 that

3

|F e3 |2 |U e3 |2 +
|(A)e3 F e3 |2 d(A) cE()
supp
=1
and
3

| cof F e3 |2 | cof U e3 |2 cE().
=1
The second inequality implies with (4.38) that 3 cE(). The rst estimate
yields in view of this bound and |F e3 | = |U1 e3 | = |U2 e2 | that

|((A) F )e3 |2 d(A) cE().
(4.39)
supp
The assertion follows now by the triangle inequality from the two foregoing
estimates since |(A) A| = dist(A, K).

As a corollary, we obtain estimates similar to those in Section 4.1 for rst
order laminates.
Corollary 4.6.2. Let K be given by (4.37) and suppose that F Mpc (K)
satises (4.38). Assume that u AF with J (u) with > 0 where J has
been dened in (4.34). Then there exists a constant c that depends only on F
and such that the following assertions hold:
i) (estimates for directional derivatives tangential to the layering direction)

'
'
1
'(Du(x) F )e3 '2 dx c,
||
ii) (estimates for the deformation)

1
|u(x) F x|dx c,
||
iii) (weak convergence, averages): the following estimate holds for all
with Lipschitz boundary,
'
'
'
'
' (Du F )dx' c(, ) 1/2 .
122
Proof. The rst two assertions follow from Theorem 4.6.1 and Poincares
inequality. The proof of last result is identical to the proof of v) in Theorem 4.5.1.
4.7 Numerical Analysis of Microstructure A Review

In order to put our results in a more general context, and to trace back the
origins of the ideas presented here, we include in this section a brief discussion
of related work. At the same time, this allows us to sketch some aspects in
the development of the numerical analysis of microstructures that began only
about ten years ago.
The rst numerical simulations of microstructures in elastic materials undergoing solid to solid phase transformations were reported in [Si89] for a hypothetic hyperelastic material and in [CL89, CLR93] for a three-dimensional
model of the cubic to tetragonal transformation with a stored energy function
proposed in [Er86] and a choice of parameters suggested by James.
The numerical analysis of nite element schemes for the minimization of
nonconvex problems was initiated in [CKL91] in the scalar, one-dimensional
setting. Let
1

E(v) =
(v (y)) + (v(y) f (y))2 dy
0
where 0 satises (s) = 0 if and only if s {sL , sU } and f is ane

with f s (sL , sU ). It is easy to see that the inmum of the energy in the
space of all Lipschitz continuous functions is zero, but that there does not
exist a minimizer for the functional in this class. In fact, v has to approximate
the ane function f with slope s, but the potential vanishes only in the
two points sL and sU . Minimizing sequences {uk } develop increasingly ner
oscillations, converge weakly to the ane function f , and the sequence {uk }
generates the unique gradient Young measure {x }x(0,1) with
x = = sL + (1 )sU ,
s = sL + (1 )sU .
The same behavior is expected for nite element minimizers, and the authors
note: ... we shall show that uh (x) and nonlinear functions of uh (x) converge
weakly. We show below that the topology of this convergence is metrizable
since it is convergence in the weak- topology of a suitable Banach space,
and we give an error estimate for this convergence in an appropriate metric.
([CKL91], page 322). Indeed, the fundamental estimate in their paper is
d(Av uh () , ) ch1/4 ,
(4.40)
where d denotes a metric that metrizes the weak- convergence. The crucial
estimate in the proof of this convergence result states that the distribution
123
of the values of uh is determined by the unique Young measure. The values
of uh lie in a neighborhood of sL and sU on a set of measure close to and
1 , respectively. Surprisingly, estimates in the metric d were not further
pursued in the literature. Our notion of stability in Denition 4.1.4 is based
on exactly the same ideas.
The analysis of higher dimensional scalar problems started in [C91, CC92],
where properties of minimizers in nite element spaces of the functional

E(v) =
(Dv(x)) + v(x) a, x dx
were investigated. Here is a bounded domain in Rn and Dirichlet conditions

u(x) = a, x are imposed on a subset 0 of (the lower order term can be
omitted if 0 = ). The function : Rn R and : R R are assumed to
be nonnegative with (w) = 0 if and only if w K = {w1 , . . . , wk } Rn ,
k 2, and (s) = 0 only for s = 0 unless 0. The existence of nite
element minimizers follows if and satisfy suitable coercivity assumptions.
Moreover, based on an explicit construction, one can show that the energy
in the nite element space is bounded by c0 h if is bounded on bounded
sets, where > 0 is related to the growth of . For example, = 12 if 0.
Also in this case the functional E does not have a Lipschitz minimizer,
but minimizing sequences generate a unique gradient Young measure under
the condition that a has a unique representation,
a=
k

i wi .
i=1
In fact, {x }x is homogeneous with

x =
k

i wi for a.e. x .
i=1
Thus all minimizing sequences generate the same Young measure and consequently the distribution of the gradients of functions with small energy should
be determined from the Young measure. This intuitive statement is made precise by what the authors in [C91, CC92] call the probabilistic analysis of the
oscillations.
Dene the projection : Rn K such that () = wi where i is the
smallest index such that | wi | = minj=1,...,k | wj |. Suppose that and
satisfy the growth conditions
() 1 | ()|p
for all Rn , p > 1, 1 > 0,
and
(t) 2 |t|q
for all t R, q > 1,
124
and that is any subdomain of with Lipschitz boundary. To simplify the

statements we assume in the sequel that p = q = 2. It is an immediate
consequence of the foregoing hypotheses that

|Dv (Dv)|2 dx cE(v)
(4.41)
(see Lemma 1 in [CC92]), and the estimate

'
'

'
'
' Du dx' c() uL2 () + (uL2 () DuL2 () )1/2 ,
u W 1,2 ()
(4.42)
implies that
'
'

'
'
' (Dv a)dx' c E 1/4 (v) + E 1/8 (v)
(4.43)
(see Lemma 3 in [CC92]). These estimates allow one to prove estimates for
the distribution of the values of Dv. For > 0 with < 12 mini=j |wi wj |
we dene

i, (v) = x : Dv(x) B(wi , ) ,
where B(wi , ) is the ball with center wi and radius . Then we have for all
v W 1,2 () with u(x) = a, x on 0
'
'

'
'
for i = 1, . . . , k.
(4.44)
'|i, (v)| i ||' c E 1/2 (v) + E 1/16 (v)
One of the many consequences of this precise control of the volume fractions
is the error estimate in Theorem 6 in [CC92]. Let f : Rn R be Lipschitz
continuous in its second argument, i.e.
|f (x, ) f (x, )| L| |
for all , Rn .
Then
k
'
'

'
'
f (x, Dv(x))
i f (x, wi ) dx' c(L) E 1/2 (v) + E 1/16 (v) ,
'
i=1
where the constant c depends on f only via the Lipschitz constant L.

These results were then extended in [CCK95] to the case of point wells
in Mmn under the assumption that the wells are pairwise compatible, i.e.
rank(Wi Wj ) = 1 for i = j. Then the matrices Xi = W1 Wi = ai bi
are pairwise compatible, and this is only possible if either all the vectors
ai or the vectors bi are parallel (see Lemma 2.1 in [CCK95]). A suitable
change of coordinates allows one to reduce the problem essentially to the
125
scalar situation and the estimates for the volume fractions can be obtained
as in the scalar case.
The paper [CCK95] also presents an adaption of the methods used for
point wells to the physically relevant case of energies with potential wells
in the exemplary case of the two-well problem in two dimensions (see also
[Gd94] for results with nonconforming elements on grids with specic, problem adapted orientations). Assume that 0 satises (F ) = (RF ) for all
R SO(2) and that (X) = 0 if and only if X K = SO(2)U1 SO(2)U2
where Q1 U1 and Q2 U2 with Q1 , Q2 SO(2) are rank-one connected. Let
A = Q1 U1 + (1 )Q2 U2 with (0, 1)
(4.45)
and
Q1 U1 Q2 U2 = a n,
a, n R2 ,
and consider the minimization problem

(Dv) + (v(x) Ax) dx.
inf
uW 1, (;R2 )
u(x)=Ax on
Due to the choice of the boundary data, there does not exist a minimizer
for the energy. However, minimizing sequences show exactly the expected
behavior: the estimates (4.41) and (4.43) follow as in the case of point wells
(with slightly modied exponents) and only the proof for the estimates (4.44)
of the volume fractions requires some modications. Dene
(F ) = QUi where QUi satises dist(F, K) = |QUi F |,
and

i, (v) = x : (Dv(x)) SO(2)Ui , |Dv(x) (Dv(x))|

(if F has the same distance to more than one well, then we choose the one
with the smallest index). In this situation the following estimates hold:
'
'
'
'
'|
2, | (1 )||' cE 2/5 (v).
1, | ||' cE 2/5 (v), '|
The analysis of three-dimensional models begins in [L96a] for a two-well problem with SO(3) invariance described by the energy

E(v) =
(Dv)dx.
Here 0 and (F ) = 0 if and only if F K = SO(3)U1 SO(3)U2 and

satises the growth condition (F ) |F (F )|2 . The wells are again
assumed to be compatible and boundary conditions that correspond to the
126
center of mass of a simple laminate (4.45) are imposed. This paper improves
the foregoing estimates in several aspects. Dene : M33 {Q1 U1 , Q2 U2 }
by
|(X) X| = dist(X, SO(3)U1 SO(3)U2 ),
and let

i, (v) = x : (Dv(x)) = Qi Ui , |Dv(x) (Dv(x))| .
The estimate

|(Dv A)w|2 c E(v) + E 1/2 (v) ,
for all w R3 , w, n = 0,
and (4.42) imply immediately

|v(x) F x|2 dx c E(v) + E 1/2 (v) ,
'
'

'
'
' (Dv A)dx' c E 1/2 (v) + E 1/8 (v) .
Bounds for the volume fractions i,R ,

'
'

'|1, | ||' c E 1/2 (v) + E 1/8 (v) ,
'
'

'|2, | (1 )||' c E 1/2 (v) + E 1/8 (v)
follow from the sharper inequality

|Dv (Dv)|2 dx c E(v) + E 1/2 (v) ,
which is based on an L2 estimate for the excess rotation R(Du) which we

dene by (Du(x)) = R(Du(x))(Du(x)), see e.g. inequality (4.18). These
estimates were subsequently applied to various phase transformations using
conforming and nonconforming nite element methods, see e.g. [L96b, LL98a,
LL98b, BLL99, BL00, LL00] or [GP99, GP00, Pr00] for models including
penalizations.
5. Applications to Martensitic Transformations
Many technological applications of shape memory materials are based on

alloys that undergo cubic to tetragonal, cubic to orthorhombic, or tetragonal
to monoclinic transformations. In this chapter we apply the general theory
developed in Chapter 4 to these transformation and investigate the question
of uniqueness and stability via the validity of condition (Cb ). We restrict
our attention to three-dimensional bulk materials and we focus therefore on
simple laminates. The application of the stability theory and the resulting
error estimates for nite element minimizers developed in Section 4.5 are
obvious and therefore not stated explicitly.
Suppose that the set K describing one of the transformations is given by
K = SO(3)U1 . . . SO(3)Uk with symmetric and positive denite matrices
Ui that satisfy det Ui = > 0 for i = 1, . . . , k. Simple laminates are obtained
by solving the twinning relation
QUi Uj = a n,
Q SO(3), a, n R3 ,
(5.1)
for which the general solutions have been given in Proposition A.2.1. Since
all the transformations considered here lead to symmetry related wells, that
is,
Ui = RUj R,
R a 180 degree rotation,
(5.2)
we may apply the result in Proposition A.2.4 which provides explicit formulae
for the two solutions of (5.1). We refer to the solutions as type-I and type-II
twinning systems, respectively. In the special case that there are two distinct
rotations that satisfy (5.2) we call the twinning systems a compound twinning
system. We summarize the information for the various transformations in
Tables 5.1-5.5, in which we omit occasionally lengthy expressions that are
not needed in the text. In these tables we write Rij

and Ri for the 180
rotations with axes ei ej and ei , respectively.
Suppose now that
= Qi Ui + (1 )Uj ,
Qi Ui Uj = a n, [ 0, 1 ],
is a simple laminate supported on K, and let F = , id be its center of

mass. In order to apply the general theory based on condition (Cb ) in Denition 4.1.4, we need to nd so-called test vectors w R3 with
G. Dolzmann: LNM 1803, pp. 127152, 2003.

128
|Ui w|2 = |Uj w|2 > max |Uk w|2 ,

k=i,j
and
w, n = 0
or
| cof Ui w|2 = | cof Uj w|2 > max | cof Uk w|2 ,
k=i,j
and
w, Uj1 a = 0.
Then automatically |F w|2 = |Ui w|2 and | cof F w|2 = | cof Ui w|2 , respectively, and the assumptions ii) in condition (Cb ) are satised. The rst equality is immediate and the second follows from formula (C.2) for the cofactor
since det Ui = det Uj implies Uj1 a, n = 0. We frequently take advantage
of the fact that by Proposition A.2.4 the vectors n and Uj1 a are parallel if
the wells are symmetry related. The examples below illustrate that a proof
of uniqueness for type-I twins using a vector w as a test vector on F typically also proves uniqueness for the corresponding type-II twin by testing the
cofactor matrices by the same vector w.
Carrying out this program in the subsequent sections, we rederive some
of Luskins results as a simple application of our general approach. In particular we nd that for cubic to orthorhombic and tetragonal to orthorhombic
transformations simple laminates are uniquely determined by their center of
mass F unless the lattice parameters describing the energy wells satisfy a
certain algebraic condition. The new results are explicit characterizations of
the sets Mpc (K; F ) in these special cases, and this answers a question raised
by James.
5.1 The Cubic to Tetragonal Transformation

The prototype of a solid to solid phase transformation is the cubic to tetragonal transformation with three martensitic wells. An example is the phase
transformation in Indium rich InTl alloys which undergo a transformation
from face-centered cubic phase into face-centered tetragonal phase. For this
transformation we have the following result.
Theorem 5.1.1. Assume that 1 , 2 > 0, 1 = 2 , and that
2 0 0
1 0 0
1 0 0
,
U
,
U
=
=
U1 =
0 1 0 2 0 2 0 3 0 1 0
0 0 1
0 0 1
0 0 2
(5.3)
Let
K = SO(3)U1 SO(3)U2 SO(3)U3
and suppose that Mpc (K) is a simple laminate with F = , id. Then
is unique and Mpc (K; F ) = {}.
129
Table 5.1. Rank-one connections (twins) in the tetragonal variants (we abbreviate
in this table type by tp and compound by cp).
(ij)
tp
Uj1 a
(12)
R1+2
cp
1 (1, 1, 0)
2
2
2
2(2
1
)
(1 , 2 , 0)
2 + 2
2
1
2
2
2(2
1
)
(1, 1, 0)
2 + 2
2
1
(12)
R12
cp
1 (1, 1, 0)
2
2
2
2(2
1
)
(1 , 2 , 0)
2 + 2
2
1
2
2
2(2
1
)
(1, 1, 0)
2 + 2
2
1
Proof. In view of the symmetry relations in Section 5.5, is suces to

consider simple laminates supported on SO(3)U1 SO(3)U2 . There exist two
rank-one connections between the wells SO(3)U1 and SO(3)U2 which generate compound twins, see Table 5.1. In order to simplify the notation in
the following statements, we call a matrix F a type-I twin if it is generated
from the twinning system with normal n = 12 (1, 1, 0), and a type-II twin
otherwise.
In view of Theorem 4.1.11 we only need to prove the existence of vectors
w with
|F w|2 = |U1 w|2 = |U2 w|2 > |U3 w|2 .
In order to accomplish this, we assume rst that 2 > 1 . If F is generated
from the type-I twinning system, then
|F w|2 = |U1 w|2 = |U2 w|2 > |U3 w|2
1
for w = (1, 1, 0),
2
and if F is generated from the type-II twinning system, then

|F w|2 = |U1 w|2 = |U2 w|2 > |U3 w|2
1
for w = (1, 1, 0).
2
Assume now that 2 < 1 . In this case we have for both twinning systems
that
|F w|2 = |U1 w|2 = |U2 w|2 > |U3 w|2
for w = (0, 0, 1),
and uniqueness of is a consequence of the foregoing inequalities.

The following theorem complements the results in Theorem 5.1.1 by identifying the implications of assumption ii) in condition (Cb ) without the hypothesis that be a simple laminate.
Theorem 5.1.2. Let K = SO(3)U1 SO(3)U2 SO(3)U3 and F K pc .
Suppose that there exists a w S2 with
130
|F w|2 = |U1 w|2 = |U2 w|2 > |U3 w|2 .
(5.4)
Then
F (SO(3)U1 SO(3)U2 )(1)
if
2 > 1 ,
and
F (SO(3)U1 SO(3)U2 )(2) = (SO(3)U1 SO(3)U2 )pc
if
2 < 1 .
Remark 5.1.3. Assumption (5.4) for 2 < 1 can be interpreted as an assumption for cof F in the case 2 > 1 . Thus there is a surprising dierence
in the implications of assumption on F and cof F for matrices F K pc .
Proof. In view of Proposition 4.1.9, assumption (5.4) implies that any
polyconvex measure with barycenter F has to be supported on the two wells
SO(3)U1 SO(3)U2 . Suppose rst that 2 > 1 . In order to conclude from
Step 2 in the proof of Proposition 4.1.8 that F is the center of mass of a
simple laminate, we have to prove the existence of vectors v, e R3 such
that v, e = 0, v is a common eigenvector of U1 and U2 , and
|F e|2 = |U1 e|2 = |U2 e|2 .
In view of (5.3) we choose v = (0, 0, 1) and deduce from Theorem 2.5.1 that
F T F v = 12 v. Moreover, by assumption the vector w = (w1 , w2 , w3 ) satises
|U1 w|2 = |U2 w|2
(22 12 )w12 = (22 12 )w22
|U1 w|2 > |U3 w|2
(22 12 )w12 > (22 12 )w32 ,
and
and therefore |w1 | = |w2 | > |w3 |. We dene e = (w1 , w2 , 0) = (w1 , w1 , 0)

and observe by the foregoing estimate that e = 0 and that e, v = 0. We
now have
|F w|2 = |F (e + w3 v)|2 = |F e|2 + 2w3 F e, F v + w32 |F v|2 = |F e|2 + w32 12 .
Similarly,
|Ui w|2 = |Ui e|2 + w32 12 ,
and hence
|F e|2 = |U1 e|2 = |U2 e|2 .
The assertion of the theorem follows now easily from Proposition 4.1.8. Note
that this argument does not work if 1 > 2 , since the assumptions are
satised with w = (0, 0, 1) for all F in the polyconvex hull of the two wells
SO(3)U1 and SO(3)U2 .

131
Remark 5.1.4. Suppose that 2 > 1 . The proof of Theorem 5.1.2 shows in
fact the following implication. Assume that F is any matrix in the polyconvex
hull of the three martensitic wells (5.3). Suppose that there exists a w S2
with |w1 | |w2 | > |w3 | and
|F w|2 = max |U w|2 .
U K
Then F (SO(3)U1 SO(3)U2 )(1) . Moreover, if |w1 | > |w2 | then F K.

Thus equality in just one of the necessary conditions
|F w|2 max |U w|2
U K
w S2
already implies that F has to be the center of mass of a simple laminate

unless w = (w1 , w2 , w3 ) satises |w1 | = |w2 | = |w3 |, i.e., w is one of the eight
vectors 13 (1, 1, 1). Therefore it is not surprising that the polyconvex
hull of three wells cannot be characterized by

F : det F = 12 2 , |F e|2 max |U e|2 , | cof F e|2 max | cof U e|2 e S2 .
U K
U K
As a matter of fact, this formula misses a crucial compensation eect between

the dierent wells. The correct representation has therefore to include at least
the additional inequalities

F : det F = 12 2 ,

|F v|2 + c(v, w)| cof F w|2 max |U v|2 + c(v, w)| cof U w|2
(5.5)
U K

for all v, w S2 .
The foregoing arguments imply in particular that one needs to understand
matrices that realize equality in conditions involving
pairs of vectors (v,
w)
with |Ui v|2 + c(v, w)| cof Ui w|2 = maxU K |U v|2 + c(v, w)| cof U w|2 for
i = 1, 2, 3.
The following proposition shows that it is also dicult to construct explicitly laminates that are supported on all three wells. In fact, any laminate
supported on three wells must contain at least four Dirac masses.
Proposition 5.1.5. Suppose that K = SO(3)U1 SO(3)U2 SO(3)U3 , where
the matrices Ui are given by (5.3) with 2 > 1 . Suppose that Mqc (K)
is supported on all three wells, i.e., that has positive mass on SO(3)Ui for
i = 1, 2, 3. Then the support of must contain at least four points.
Proof. Assume that is supported on three points, i.e., that is given
by
= 1 X 1 + 2 X 2 + 3 X 3 ,
Xi SO(3)Ui , i > 0, 1 + 2 + 3 = 1,
132
and let F = , id. We divide the proof into three steps.

Step 1: Without loss of generality we may assume that
rank(X1 X2 ) = 1,
rank
1
2
X1 +
X2 X3 = 1.
1 3
1 3
aks results that any gradient Young measure

Indeed, it follows from Sver
supported on three incompatible matrices is a single Dirac mass that at least
two of the three matrices Xi must be rank-one connected. We may therefore
assume that X1 X2 = a n with a, n R3 , a, n = 0. In this situation,
we follow arguments by James and Kinderlehrer which imply that
rank

2
1
X1 +
X2 X3 = 1.
1 3
1 3
(5.6)
To show this, we use the identity

cof F =
3

1 2
cof(X2 X1 )
1 3

2
1
3 (1 3 ) cof
X1 +
X2 X3
1 3
1 3
i cof Xi
i=1
and insert the minors relations

F = 1 X 1 + 2 X 2 + 3 X 3 ,
cof F = 1 cof X1 + 2 cof X2 + 3 cof X3 ,
to obtain that
cof

2
1
X1 +
X2 X3 = 0.
1 3
1 3
Step 2: We have
= sup |Xe|
< 2
max (X)
eS2
where
=
X
1
2
X1 +
X2 .
1 3
1 3
= 2 . Then there exists an

e| = 2
e S2 such that |X
Assume that max (X)
and thus
1
=
2
e| |U
2 = |X
e| + (1 )|U
e| 2 with
1
.
1 3
In particular, |U1
e has to be pare| = 2 and |U2
e| = 2 which implies that
allel to e1 and e2 , respectively. This is only possible if
e = 0, a contradiction.
and SO(3)U3 are incompatible.
Step 3: The wells SO(3)X
Without loss of generality we may assume that X2 = U2 . Since there are
exactly two rank-one connections between the wells SO(3)U1 and SO(3)U2 ,
133
the matrix X1 has to be equal to either Q1 U1 or Q2 U1 . A short calculation

shows that
22 12
22 12
21 2
21 2
0
0
22 2+12 2 22 +12
222 +122 22 +12
2 1 21 2
21 2
1
Q2 = 22
Q1 = 2 +2 2 +2 0 ,
0 .
2 1 2 1
2 +12 22 +12
0
0 1
0
0
1
We may suppose that X1 = Q1 U1 , and thus
2 2 )
2 2 )
(
1 (
2
1
2
1
0
1 + 1 2 +
2
2
2
2 +1
2
1
2 2 )
2 2 )
2 (
= diag(X,
% 1 ) =
2
1
2
1
X
.
1 (
0
2
22 +12
22 +12
0
0
1
(5.7)
% SO(2) such that

By the polar decomposition theorem, there exists a Q
% 1)X
= (X
T X)
1/2 = diag(Y% , 1 ),
Y = diag(Q,

and Y% M22 is symmetric. There exists a rank-one connection between X
and the well SO(3)U3 if and only if there exists a rank-one connection between
Y and SO(3)U3 . In view of Proposition A.2.1, this rank-one connection exists
if and only if the middle eigenvalue mid of
T XU
1
Z = U31 Y 2 U31 = U31 X
3
is equal to one. By (5.7),
2
% 1 ),
Z = diag(Z,
22
% M22
Z
with
% = 1 Y% T Y% = 1 X
% T X,
%
Z
12
12
%=
det Z
22
.
12
By Step 2,
2

% < 2 for
max (Z)
{0, 1}.
12
This implies
mid
2
22
% = 2
>
=
det
Z
max
mid
12
12
{0, 1}.
and hence the middle eigenvalue of Z is strictly bigger than one for
The wells are therefore incompatible.
This contradicts (5.6) and the assertion of the proposition follows.

134
Table 5.2. Rank-one connections (twins) in the cubic to trigonal transformation.

(ij)
type
Uj1 a
(12)
R2
compound
(0, 1, 0)
2(2+)
(1, 0, 1)
a2 +2+3 2
(12)
R13
compound
1 (1, 0, 1)
2
2(2+)
(0, 1, 0)
a2 +2 2
5.2 The Cubic to Trigonal Transformation

If the lengths of the sides of the unit cell in the cubic and the trigonal phase
are the same, then this transformation is in a suitable basis characterized by
the four strains

U2 = ,
U1 = ,

U3 = ,
U4 = .

In this situation, we have the following result.
Corollary 5.2.1. Assume that > , > 0 and that the set K is given by
K = SO(3)U1 . . .SO(3)U4 . Suppose that Mpc (K) is a simple laminate
and let F = , id. Then Mpc (K; F ) = {}.
Proof. By symmetry and invariance under the point group it suces to
prove the statement of the corollary for i = 1 and j = 2. If F is generated
from the compound twinning system with n = e2 (cf. Table 5.2), then
|U1 w|2 = |U2 w|2 = ( + )2 + 2 2 > ( )2 = |U3 w|2 = |U4 w|2
with w = (e1 + e3 )/ 2. We conclude from Proposition 4.1.9 that must be

supported on SO(3)U1 SO(3)U2 and since e1 +e3 is a common eigenvector
of U1 and U2 (with eigenvalue ) we obtain from Proposition 4.1.8 the
uniqueness of . Similarly, if F is generated using the second normal parallel
to e1 + e3 , then
| cof U1 w|2 = | cof U2 w|2 = (2 + 2 2 )2
> ( )2 (2 + 2 2 ) = | cof U3 w|2 = | cof U4 w|2 ,
where w =
1 (e1
2
e3 ). We conclude as before.

5.3 The Cubic to Orthorhombic Transformation
135
Fig. 5.1. The second type of the cubic to orthorhombic phase transformation. The
tetragonal unit cell within the cubic lattice is stretched.

The orthorhombic phase is characterized by three mutually perpendicular
axes of twofold symmetry. If these three axes are parallel to the edges of the
cubic cell of the parent phase, then the transformation strains are described
by the six matrices
00
0 0
00
U1 = 0 0 , U2 = 0 0 , U3 = 0 0 ,
0 0
0 0
0 0
0 0
0 0
0 0
U4 = 0 0 , U5 = 0 0 , U6 = 0 0 .
00
0 0
00
There seems to be no material known with this symmetry in the martensitic
phase. A second type of cubic to orthorhombic transformations occurs in
materials with a fcc parent phase, see Figure 5.1. It is characterized by two
axes of symmetry along face diagonals of the cubic cell and one along the
edge orthogonal to this face. The six variants are usually described by the
matrices
0
0
0
U1 = 0 , U2 = 0 , U3 = 0 0 , (5.8)
0 0
0 0
0
0
0 0
0 0
U4 = 0 0 , U5 = 0 , U6 = 0
0
0
0
with
=
>=
> 0,
2
2
= > 0.
(5.9)
This transformation is for example found in CuAlNi alloys, the high temperature phase is body centered cubic, the low temperature phase has orthorhombic symmetry.
136
Table 5.3. Rank-one connections (twins) in the second variant of a cubic to

orthorhombic transformation.
(ij)
type
Uj1 a
(12)
R1
compound
(1, 0, 0)
4
(, , 0)
2 +2
4
(0, 1, 0)
2 +2
(12)
R2
compound
(0, 1, 0)
4
(, , 0)
2 +2
4
(1, 0, 0)
2 +2
(13)
R23
type-I
(0, 1, 1)
2
(13)
R23
type-II
1 (, , )
2
(0, 1, 1)
2
Theorem 5.3.1. Assume that K = SO(3)U1 . . . SO(3)U6 where the matrices Ui are dened in (5.8). Suppose that Mpc (K) is a simple laminate
supported on K and let F = , id.
1. Suppose the rank-one connection used in the denition of F determines
a compound twin. Then Mpc (K; F ) = {} for all the parameters , ,
and satisfying (5.9), except those such that
2 = 2 + 2 ,
or
2 =
( 2 2 )2
.
2 + 2
If these identities hold, then Mpc (K; F ) consists of a one-parameter family of simple laminates supported on at most four matrices. Moreover, F
has two distinct representations as a simple laminate if and only if = 12 .
2. If F is generated by a type-I twinning system, then Mpc (K; F ) = {}
for all the parameters , , and satisfying (5.9), except those such that
= + .
If this identity holds, then Mpc (K; F ) can be obtained from a three-well
problem in two dimensions. Moreover, F has two distinct representations
as a simple laminate with mass on one well and mass 1 on the other
well in the type-I twinning system if and only if is equal to one of the
two values 1 and 2 dened by
1 =

1
4
3
and
2 =
1
+ 1
3
and 1 and 2 lie in the open interval (0, 1).

3. If F is generated by a type-II twinning system, then Mpc (K; F ) = {}
for all the parameters , , and satisfying (5.9), except those such that
137
= .
If this identity holds, then Mpc (K; F ) can be obtained from a three-well
problem in two dimensions. Moreover, F has two distinct representations
as a simple laminate if and only if the mass on one of the two wells is
equal to one of the two values 1 and 2 dened in 2.
Remark 5.3.2. It follows from the results in Chapter 2 that the semiconvex
hulls are all equal for the three-well problems. Therefore the characterizations
of the polyconvex hulls in the exceptional cases are also characterizations for
the quasiconvex hulls.
Proof. We divide the proof into two parts. We rst prove the uniqueness
assertions and then the characterizations for the sets Mpc (K; F ) for the
exceptional cases.
In order to prove uniqueness, it suces to consider laminates for which
one of the matrices is in SO(3)U1 . Assume rst that F is a compound twin
with normal n = e1 .
We have
|U1 e3 |2 = |U2 e3 |2 > max |Ui e3 |2 ,
i=3,...,6
|U1 e2 |2 = |U2 e2 |2 = |U5 e3 |2 = |U6 e3 |2 > |U3 e2 |2 = |U4 e2 |2 ,
2 > 2 + 2 ,
2 < 2 + 2 .
Moreover,
| cof U1 e1 |2 = . . . = | cof U4 e1 |2 > | cof U5 e1 |2 = | cof U6 e1 |2
for ( 2 + 2 ) 2 > ( 2 2 )2 and
| cof U1 e3 |2 = | cof U2 e3 |2 > max | cof Ui e3 |2
i=3,...,6
for ( 2 + 2 ) 2 < ( 2 2 )2 , and we can therefore choose appropriate vectors wi if the assumptions in the theorem hold. The nonuniqueness part is
discussed in detail below.
Suppose now that F is obtained from the type-I twinning system between
SO(3)U1 and SO(3)U3 . Let w = (, , ). Then
|U1 w|2 = |U3 w|2 = |U5 w|2 > |U2 w|2 = |U4 w|2 = |U6 w|2 .
(5.10)
Therefore it only remains to show that U5 cannot be part of the microstructure. We have to prove the existence of a vector v = (t, s, s) such that
|U1 v|2 = |U3 v|2 > |U5 v|2
(since n is parallel to (0, 1, 1) we have automatically that |F v|2 = |U1 v|2 ).
This is equivalent to
138

g(s, t) = s2 2 + (t + s)2 + (s + t)2 t2 2 + 2s2 ( + )2 > 0.
Since g(t, t) = 0 and
s g(t, t) = 2t( 2 2( + )2 ),
this is possible if 2 = ( + )2 , i.e., the hypothesis of the theorem holds.
Similarly, if F is a type-II twin, then we have for w = (, , ) that
| cof U1 w|2 = | cof U3 w|2 = | cof U6 w|2
> | cof U2 w|2 = | cof U4 w|2 = | cof U5 w|2
and we only have to exclude U6 from the microstructure. Let v = (s, t, t).
Then
| cof U1 v|2 = | cof U3 v|2
and we have to determine s, t R such that
g(s, t) = | cof U1 v|2 | cof U6 v|2 > 0.
Since g(t, t) = 0 and
s g(t, t) = 2t( + )2 ( 2 ( )2 )
this is possible if the assumptions in ii) hold. Finally, the remaining cases
follow by symmetry.
We now turn towards the characterizations in the case that the polyconvex
measure is not uniquely determined. We discuss the three dierent cases
separately.
1) We assume rst that 2 = 2 + 2 and that
F = tQU1 + (1 t)U2 = U2 + ta n,
where Q and a are given by Proposition A.2.4 with n = e1 , i.e.
2
(1 4t
2 + 2 ) 0
F = (1 + 4t
.
)
0
2 + 2
0
0
Assume that Mpc (K) satises , id = F . Since
|F e2 |2 = |Ui e2 |2 ,
|F e3 |2 = |Ui e3 |2 ,
i = 1, . . . , 6,
and
| cof F e1 |2 = = | cof U4 e1 |2 > | cof U5 e1 |2 > | cof U6 e1 |2 ,
(5.11)
139
must be supported on SO(3)U1 . . .SO(3)U4 and thus can be represented

as
=
4

i i Ui ,
i P(SO(3)),
i 0,
1 + 2 + 3 + 4 = 1.
i=1
Let Xi = i , id conv SO(3). Thus , id =
,4
i=1
i Xi Ui and
4
4
'
'
'
'
|F e2 | = '
i Xi Ui e2 '
i |Xi Ui e2 | |F e2 |.
i=1
i=1
Therefore we have equality everywhere in this chain of inequalities and the

strict convexity of the norm implies that all the vectors Xi Ui e2 , i = 1, . . . , 4,
must have the same direction and maximal length 2 + 2 . The same assertion
follows for Xi Ui e3 and Xi cof Ui e1 , i = 1, . . . , 4. We conclude in view of (5.8)
that

0
0
X 1 = X 2 = X3 = X 4 ,
0
0
0
0
and
0
0
X1 0 = X2 0 = X3 0 = X4 0 ,
and since cof(AB) = cof(A) cof(B) we have additionally that
X1 = X2 = X3 0 = X4 0 .
0
0
Since
0 '2
'
'2 '
'
'
'
'
'X1 ' = 'X1 0 ' = 2 + 2 = 2 ,
0
the matrices Xi are length preserving on an orthogonal frame and have singular values bounded by one, since they are a convex combination of matrices
in SO(3). This implies by Lemma 5.3.3 below that Xi O(3). On the other
hand, Xi = i , id with i P(SO(3)) and thus Xi = Qi SO(3). The
proper rotation mapping one orthogonal frame {v 1 , v 2 , v 3 } onto another orthogonal frame {w1 , w2 , w3 } is given by Q = w1 v 1 + w2 v 2 + w3 v 3 ,
and we obtain
140
0
1 24
2
+
X2T X1 U1 = 1 + 24
,
0
2
+
0
0
2
0
1 22
+ 2

0
X2T X3 U3 = 1 + 22
,
2
+
2
2 2
0
+

0
1 22
2
+

0.
X2T X4 U4 = 1 + 22
+ 2
2
0
2
2
+
We conclude
F 11 0
X2T F = F 21 0
F 31 0
(5.12)
with
4 2
2 2
+ 2 + (3 + 4 ) 1 2
,
2
+
+ 2
4 2
2 2
,
= 1 1 + 2
+
(
+
1
+
2
3
4
+ 2
2 + 2
2
= (3 4 )
.
2 + 2
F 11 = 1 1
F 21
F 31
We obtain from (5.11) and (5.12) that X2 = I and

3 = 4 ,
1 + 3 = t,
1 + 2 + 3 + 4 = 1,
i 0.
If we consider 3 = s as a free parameter, we can represent all solutions to

these equations by
1 (s) = t s,
2 (s) = 1 t s,
3 (s) = s,
4 (s) = s
for s [ 0, min{t, 1 t} ]. In the special case t = 12 we may choose s = 0

and s = 12 , and obtain two representations of F as a simple laminate with
1 = 2 = 12 and 3 = 4 = 12 , respectively. On the other hand, if t = 21 ,
then it is easy to see that the only representation of F as a simple laminate
corresponds to 1 = t, 2 = 1t. If we dene = 12s, s [ 0, min{t, 1t} ],
and
141
1
2
ts
1ts
X1 U1 +
U2 =
X1 U1 +
U2 ,
1 23
1 23
1 2s
1 2s
1
G2 = (X3 U3 + X4 U4 ),
2
G1 =
then G1 and G2 are rank-one connected and

F = G1 + (1 )G2 ,
[ max{1 2t, 2t 1}, 1 ].
This proves the assertion for the case that n = e1 in the denition of the
compound twins. The other case with n = e2 follows by symmetry. Assume
now that 2 = ( 2 2 )2 /( 2 + 2 ), i.e. 2 < 2 + 2 . If F is generated from
the compound twin system with n = e1 , then |Ui e2 |2 = |F e2 |2 > |U3 e2 |2 =
|U4 e2 |2 for i = 1, 2, 5, 6, and thus the polyconvex measure must be supported
on SO(3)U1 SO(3)U2 SO(3)U5 SO(3)U6 . Since | cof Ui e1 |2 =| cof F e1 |2
and | cof Ui e3 |2 =| cof F e3 |2 for i = 1, 2, 5, 6, we conclude as before. The last
case, the compound twin with n = e2 , follows again by symmetry.
2) Assume now that = + . We consider the case that F is obtained
from the type-I twinning system between SO(3)U1 and SO(3)U3 , i.e.
0
1
F = U3 + ta n, n = 1
2 1
and a is given by Proposition A.2.4. It follows from (5.10) that every measure
Mpc (K) with , id is supported on SO(3)U1 SO(3)U3 SO(3)U5 . Since
1 (1, 1, 1) is a common eigenvector of U1 , U3 and U5 under the assumption
3
that = + we are in a situation in which Theorem 2.5.1 applies. In order
to simplify the calculations, we choose a new basis B of R3 by
1
1
1
1
1
1
B = 1 , 1 , 1 ,
3
6 2
2
1
0
i of Ui in this basis as
and obtain the representations U
%1 ),
1 = diag( + , U
U
with
%1 =
U
3 = diag( + , U
%3 ),
U
%3 =
U
2
23
3
2
+ 2
5 = diag( + , U
%5 )
U
%5 =
U
3
2
3
2
+ 2
%3 and in order to prove

%1 + (1 t)U
The matrix F% satises F% = tQ1 U
nonuniqueness of the polyconvex measure we rst solve the system F% =
%3 ) for s and t with R = R(s) SO(2). The solutions are
%5 (1 s)U
R(sQT5 U
%3 ) and (s, t) = (
given by (s, t) = (0, 0) (corresponding to F = U
s, t) with
142
1

4
.
s = t =
3
This implies that we have two representations of F as a simple laminate for

parameters t (0, 1). If t (0, 1), then the two curves on the surface of
constant determinant det F = 2 2 corresponding to the (symmetrized)
%1 )T U
%5 )T U
%1 and U T U3 and (U
%5 and U T U3 , respecrank-one lines between (U
3
3
tively, do not intersect. However, there is a one parameter family of second
order laminates that generate F which can for example be generated by xing
%2 and dening G2 to be the
%1 + (1 )U
any matrix of the form G1 = Q2 U
intersection point of the rank-one line through G1 and F with the boundary
%1 SO(2)U
%3 SO(2)U
%5
of the quasiconvex hull of SO(2)U
3) We assume nally that = and that F is the barycenter of a
simple laminate given by the type-II twins between SO(3)U1 and SO(3)U3 .
For completeness, we summarize the arguments in this case. As in the proof of
Theorem 5.3.1 we deduce with w = (1, 1, 1) and v = (s, t, t) as test vectors
that any Mpc (K; F ) is supported on SO(3)U1 SO(3)U3 SO(3)U6 . Since
1 (1, 1, 1) is a common eigenvector of U1 , U3 and U6 under the assumption
3
that = we are in a situation in which Theorem 2.5.1 applies. In order
to simplify the calculations, we choose a new basis B of R3 by

1
1
1
1
1
1
B = 1, 1, 1 .
3
6 2
2 0
1
The transformed matrices are given by
%1 ),
1 = diag( , U
U
with
%1 =
U
3 = diag( , U
%3 ),
U
%3 =
U
2
23
3
2
2
6 = diag( , U
%6 )
U
%6 =
U
3
2
3
2
2
The analysis has thus been reduced to the three-well problem in Section 2.2,
%j J T are identical to the matrices considered there. The
since the matrices J U
proof of the theorem is now complete.

The proof of the foregoing theorem used the following fact.
Lemma 5.3.3. Assume that A Mnn satises |Av i | = 1 on an orthonormal basis {v 1 , . . . , v n } and that the singular values of A are bounded by one.
Then A O(n).
Proof. We may assume that A is symmetric. Let {i , ei } be an eigensystem
,
, (j)
(j)
for A, A = i ei ei , and dene i by vj = i ei . Then
5.4 The Tetragonal to Monoclinic Transformations

n
n

'
(j)
(j)
1 = |Av j |2 = '
i i ei |2 =
2i (i )2 1
i=1
143
for all j,
i=1
(j)
since |i | 1. If |i | < 1, then i = 0 for all j and hence v j = 0, contradicting the assumption that |Av j | = 1. Hence i {1}, and we conclude

that AT A = I with det A {1}.

It is an open problem to characterize uniqueness of microstructure in the
cubic to monoclinic phase transformation. However, the ideas described in
the analysis of the cubic to tetragonal transformation can be applied to three
possible tetragonal to monoclinic transformations, which we will refer to as
type-I through type-III. The matrices for the type-I transformation are given
by
1 0 4
1 0 4
U1 = 0 2 0 , U2 = 0 2 0 ,
4 0 3
4 0 3
2 0 0
2 0 0
U3 = 0 1 4 ,
U4 = 0 1 4 .
0 4 3
0 4 3
The matrices Ui are positive denite for 1 , 2 , 3 > 0 and 1 3 42 > 0, and
we may assume that 4 > 0. We summarize the relevant rank-one connections
in Table 5.4.
The matrices for the type-II transformation are given by
1 3 4
1 3 4
U2 = 3 1 4 ,
U1 = 3 1 4 ,
4 4 2
4 4 2
1 3 4
1 3 4
U3 = 3 1 4 ,
U4 = 3 1 4 ,
4 4 2
4 4 2
and we assume that
1 , 2 > 0,
12 > 32 ,
det Ui = (1 + 3 )(2 (1 3 ) 242 ) > 0. (5.13)
These conditions imply that the matrices Ui are positive denite. The relevant
rank-one connections are summarized in Table 5.5. Finally, the transformation of type-III is described by the following matrices:
144
Table 5.4. Rank-one connections for the type-I tetragonal to monoclinic transformation.
(ij)
type
Uj1 a parallel to
(12)
R1
compound
(1, 0, 0)
(0, 0, 1)
(12)
R3
compound
(0, 0, 1)
(1, 0, 0)
(13)
R12
type-I
1 (1, 1, 0)
2
(13)
R12
type-II
(1, 1, 0)
Table 5.5. Rank-one connections for the type-II tetragonal to monoclinic transformation.
(ij)
type
Uj1 a parallel to
(12)
R2
type-I
(0, 1, 0)
(12)
R2
type-II
(0, 1, 0)
(13)
R3
compound
(0, 0, 1)
(1, 1, 0)
(13)
R12
compound
1 (1, 1, 0)
2
(0, 0, 1)
1 4 0
1 4 0
U1 = 4 2 0 , U2 = 4 2 0 ,
0 0 3
0 0 3
2 4 0
2 4 0
U3 = 4 1 0 , U4 = 4 1 0
0 0 3
0 0 3
with 1 , 2 , 3 > 0, 4 = 0 and 1 2 42 > 0. This case is an example for a

three dimensional situation which can be reduced to a two-dimensional one,
see Theorem 2.5.1, and we do not discuss it in the following theorem.
Theorem 5.4.1. Assume that the matrices Ui , i = 1 . . . , 4, describing the
tetragonal to monoclinic transformations of type I and II and the corresponding vectors a and n in the representation of the rank-one connections between
145
the wells are given by the expressions in Tables 5.45.5. Let

K = SO(3)U1 . . . SO(3)U4
and suppose that F describes a global deformation corresponding to the center
of mass of a simple laminate, i.e.
F = QUi + (1 )Uj ,
QUi Uj = a n,
i = j, [ 0, 1 ], and Q SO(3). Then the microstructure underlying F is

unique, and in fact Mpc (K; F ) = {} with = QUi + (1 )Uj , unless
the situation is symmetry related to one of the following exceptional choices
of lattice parameters and twinning systems:
1. Tetragonal to monoclinic transformation of type I: If i = 1 and j = 2,
i.e., F is generated from the compound twinning system between SO(3)U1
and SO(3)U2 , then F is not unique if
n = (1, 0, 0)
and
22 (32 + 42 ) = (1 3 4 )2
or
n = (0, 0, 1)
and
22 = 12 + 42 .
2. Tetragonal to monoclinic transformation of type II: If i = 1 and j = 3,

i.e., F is generated from the compound twinning system between SO(3)U1
and SO(3)U3 , then F is not unique if
n = (0, 0, 1) and 21 3 = 42 .
In all the exceptional cases, Mpc (K; F ) consists of a one-parameter family
of laminates and the matrix F has two dierent representations as a simple
laminate if and only if = 12 .
Remarks 5.4.2. 1) The proof of the theorem shows that uniqueness of the
microstructure can be obtained based on condition (Cb ) with n = e3 in
Denition 4.1.4. Therefore stability of microstructure is an immediate consequence of the results in Chapter 4.
2) The analysis of the transformation of type-III can be reduced in view
of Theorem 2.5.1 and the examples following Proposition 2.2.4 to the twodimensional situation, and the uniqueness results follow from the corresponding analysis in Chapter 4.
Proof. We sketch the proof of the theorem in the setting of our general
framework for uniqueness based on Denition 4.1.4 and Theorem 4.1.14
Proof for the tetragonal to monoclinic transformation of type-I. We consider separately the dierent twinning systems and establish the existence
146
of the test vectors w which imply uniqueness of microstructure based on

Denition 4.1.4.
Case i = 1, j = 2, n = e1 : Assume rst that 22 (32 + 42 ) = (1 3 4 )2 .
Since U21 a is parallel to (0, 0, 1), we may use w1 = (1, 0, 0) and w2 = (0, 1, 0)
as test vectors for the cofactor matrices. We obtain
|F w1 |2 =| cof U1 w1 |2 = | cof U2 w1 |2 = 22 (32 + 42 ),
| cof U3 w1 |2 = | cof U4 w1 |2 = (1 3 4 )2 ,
and
| cof F w2 |2 =| cof U1 w2 |2 = | cof U2 w2 |2 = (1 3 4 )2 ,
| cof U3 w1 |2 = | cof U4 w1 |2 = 22 (32 + 42 ),
and thus we may choose w correspondingly to ensure uniqueness. It remains
to establish the characterization of Mpc (K; F ) for 22 (32 + 42 ) = (1 3 4 )2 .
Suppose that
F = QU1 + (1 )U2 = U2 + a n
1 0 4
4 0 0
2(
+
)
1
3
4
0 0 0 .
= 0 2 0 +
32 + 42
3 0 0
4 0 3
We assume as in the proof of Theorem 5.3.1 that Mpc (K; F ) is given by
= 1 X1 U1 + 2 X2 U2 + 3 X3 U3 + 4 X4 U4
with Xi conv SO(3). It follows that the matrices Xi map the orthogonal
vectors {cof Ui e1 , cof Ui e2 , Ui e3 }, i = 1, . . . , 4, onto the orthogonal vectors
{cof F e1 , cof F e2 , F e3 } of same length. We may apply Lemma 5.3.3 and
conclude that Xi SO(3) for i = 1, . . . , 4. Therefore the matrices Xi are
uniquely determined and a short calculation shows that
2 2 (1 +3 )
0 4
1 42 +
2
3
4
X1 U1 =
0
2 0 ,
1 +3 )
4 + 23 42(
0
2
3
+
3
that X2 = I and therefore X2 U2 = U2 , that
2 (1 +3 )
2 (1 +3 )
1 42 +
42 +
4
2
2
3
3
4
4
X3 U3 =
0
2
0 ,
1 +3 ) 3 4 (1 +3 )
4 + 3 42(+
2
2
2
3
+
3
147
and nally that
X4 U4 =
42 (1 +3 )
32 +42
42 (1 +3 )
32 +42
0
4 +
3 4 (1 +3 )
32 +42
1 +3 )
3 42(+
2
3
0 .
We may now solve for the unknown volume fractions i , and if we consider
4 = s [ 0, 1 ] as a free parameter, then we nd the solutions
1 = s,
2 = 1 s,
3 = 4 = s.
Let
Z1 =
s
1s
X1 U1 +
U2 ,
1 2s
1 2s
Z2 =
1
1
X3 U3 + X4 U4 .
2
2
With these denitions, it is easy to see that the polyconvex measures is in

fact a laminates since
rank(X1 U1 U2 ) = 1,
rank(X3 U3 X4 U4 ) = 1,
rank(Z1 Z2 ) = 1,
and hence

Mpc (K; F ) = =( s)X1 U1 + (1 s)U2 + sX3 U3 + sX4 U4 ,

s [ 0, min{, 1 } ] .
A short calculation shows that there exist two dierent representations of F
as a simple laminate if and only if = 12 .
Case i = 1, j = 2, n = e3 = (0, 0, 1): We may use w1 = e1 and w2 = e2
as test vectors for the matrices and we deduce
|F w1 |2 = |U1 w1 |2 = |U2 w1 |2 = 12 + 42 ,
|U3 w1 |2 = |U4 w1 |2 = 22 ,
and
|F w2 |2 = |U1 w2 |2 = |U2 w2 |2 = 22 ,
|U3 w2 |2 = |U4 w2 |2 = 12 + 42 ,
respectively. This proves the uniqueness of the microstructure unless we have

12 + 42 = 22 . Assume now that this identity holds and that F is a simple
laminate generated from this twinning system,
F = QU1 + (1 )U2 = U2 + a n
1 0 4
0 0 1
2(
+
)
1
3
4
0 0 0 .
= 0 2 0 +
22
4 0 3
0 0 4
148
We suppose again that the polyconvex measure Mpc (K; F ) is given by

= 1 X1 U1 + 2 X2 U2 + 3 X3 U3 + 4 X4 U4
with Xi conv SO(3). Now the matrices Xi , i = 1, . . . , 4, map the orthogonal vectors {Ui e1 , Ui e2 , cof Ui e3 }, i = 1, . . . , 4, onto the orthogonal vectors
{F e1 , F e2 , cof F e3 } of same length. We may apply Lemma 5.3.3 and conclude
that Xi SO(3) for i = 1, . . . , 4. Therefore the matrices Xi are uniquely determined and a short calculation shows that
1 +3 )
1 0 4 + 21 4 (
2
2
X1 U1 = 0 2
,
0
2
2 ( + )
4 0 3 4 12 3
2
that X2 = I and thus X2 U2 = U2 , that
1 0 4 + 1 4 (21 +3 )
2
4 (1 +3
X3 U3 = 0 2
,
2
2
42 (1 +3 )
4 0 3
2
2
and nally that
X4 U4 = 0 2
4 0
1 4 (1 +3 )
22
4 (1 +3
.
2
2
2
( + )
3 4 12 3
2
1 0 4 +
As before, we can solve for the unknowns i and we obtain with 4 = s as a

free parameter the solutions
1 = s,
2 = 1 s,
3 = 4 = s.
It is easy to check that corresponds again to a second order laminate.

Case i = 1, j = 3: If F is a type-I twin, then the interface normal is
parallel to e1 e2 and we may choose w = e1 + e2 + e3 as a test vector. We
nd
|F w|2 = |U1 w|2 = |U3 w|2 = 22 + (1 + 4 )2 + (3 + 4 )2 ,
|U2 w|2 = |U4 w|2 = 22 + (1 4 )2 + (3 4 )2 .
In the type-II twinning system, U31 a is parallel to e1 e2 , and the choice
of w = e1 + e2 e3 as a test vector for the cofactor matrices shows that
149
| cof F w|2 = | cof U1 w|2 = | cof U3 w|2

= 22 (1 + 4 )2 + 22 (3 + 4 )2 + (1 3 42 )2 ,
and that
| cof U2 w|2 = | cof U4 w|2 = 22 (1 4 )2 + 22 (3 4 )2 + (1 3 42 )2 ;
these inequalities prove the uniqueness of for all choices of the lattice parameters.
Proof for the tetragonal to monoclinic transformation of type-II.
We begin our analysis with the uniqueness results for the type-I/II twinning system, then we turn towards the characterization for the compound
twins between SO(3)U1 and SO(3)U3 .
Case i = 1, j = 2: If F is generated form a type-I twinning system, then
we choose w = (1, 0, 1) as a test vector for F and obtain
|F w|2 = |U1 w|2 = |U2 w|2 = (1 + 4 )2 + (2 4 )2 + (3 4 )2 ,
|U3 w|2 = |U4 w|2 = (1 4 )2 + (2 4 )2 + (3 + 4 )2 .
We conclude
|U1 w|2 |U3 w|2 = 4(1 + 2 3 )4
and we obtain uniqueness with the appropriate choice of the sign since (5.13)
implies that 1 + 2 3 > 0. For the type-II twinning system we use the
same vector w for the cofactor matrices, and we get
| cof F w|2 = | cof U1 w|2 = | cof U2 w|2
2
= (1 + 3 )2 (1 3 4 )2 + (1 + 3 )4 (1 2 42 )
2
+ (1 + 3 )4 (2 3 + 42 )
and
| cof U3 w|2 = | cof U4 w|2
2
= (1 + 3 )2 (1 3 4 )2 + (1 + 3 )4 (1 2 42 )
2
+ (1 + 3 )4 (2 3 + 42 ) ,
and hence
| cof U1 w|2 | cof U3 w|2 = 4(1 + 2 3 )(1 + 3 )2 4 .
We conclude uniqueness as before.
Case i = 1, j = 3, n = e3 = (0, 0, 1): Assume rst that 21 3 = 42 . We
choose w = (1, 1, 0) and obtain
150
|F w|2 = |U1 w|2 = |U3 w|2 = 2(1 3 )2 + (4 4 )2 ,

|U2 w|2 = |U4 w|2 = 2(1 3 )2 + (4 4 )2 ,
and thus
|U1 w|2 |U2 w|2 = 4(21 3 42 ),
and uniqueness of microstructure is an immediate consequence. Assume now
that 21 3 = 42 , and that F is a simple laminate generated from this twinning
system,
0 0 1 3
1 3 4
2
(
+
)
4 1
2
3
0 0 (1 3 ) .
F = U3 + a n = 3 1 4 +
(1 + 3 )2
4 4 2
00
24
To simplify notation, we dene
=
24 (1 + 2 3 )
(1 + 3 )2
We assume again that the polyconvex measure Mpc (K; F ) is given by

= 1 X1 U1 + 2 X2 U2 + 3 X3 U3 + 4 X4 U4
with Xi conv SO(3). We conclude as before that the matrices Xi map the
orthogonal vectors {Ui e1 , Ui e2 , cof Ui e3 }, i = 1, . . . , 4, onto the orthogonal
vectors {F e1 , F e2 , cof F e3 } of same length. We may apply Lemma 5.3.3 and
obtain that Xi SO(3) for i = 1, . . . , 4. Therefore the matrices Xi are
uniquely determined and a short calculation shows that
1 3 4 + (1 3 )
X1 U1 = 3 1 4 (1 3 ) ,
4 4
2 24
that
1 3 4 + 1
X2 U2 = 3 1 4 + 3
4 4 2 4
that X3 = I and therefore X3 U3 = U3 , and nally that
1 3 4 3
X4 U4 = 3 1 4 1 .
4 4 2 4
5.5 Reduction by Symmetry Operations
151
We solve for the volume fractions i and nd the solutions

1 = s,
2 = s,
3 = 1 s,
4 = s.
As before, we obtain a one-parameter family of second order laminates with

center of mass equal to F .
Case i = 1, j = 3, n = e3 = (0, 0, 1): We use w = (1, 1, 0) as a test
vector for F and obtain
| cof F w|2 = | cof U1 w|2 = | cof U3 w|2 = 2(1 + 3 )2 (22 + 242 )
2
| cof U2 w|2 = | cof U4 w|2 = 2 (2 (3 1 ) + 242 ,

and since 1 2 42 > 0 we deduce
| cof U1 w|2 | cof U2 w|2 = 4(1 2 42 )(2 3 + 42 ) + 222 (1 + 3 )2 > 0.
This implies uniqueness of the microstructure and concludes the proof of the
theorem.

5.5 Reduction by Symmetry Operations

In the foregoing sections, we discussed the uniqueness of simple laminates
for representative twinning systems. The analysis for the remaining ones can
be reduced to the presented cases by symmetry operations. To simplify the
notation, we dene
R{Ui , Uj }RT = {RUi RT , RUj RT }.
Suppose now that we have analyzed the twins generated by the system
Q1,2 Ui Uj = a1,2 n1,2 and that {Uk , U } = R{Ui , Uj }RT . Then
Q1,2 Ui Uj = a1,2 n1,2
RQ1,2 R RUi RT RUj RT = Ra1,2 Rn1,2

1,2 Uk U = a
1,2 n
1,2
Q
T
1,2 = RQ1,2 RT , a
= Ra, n
= Rn. The results for {Uk , U } follow
with Q
therefore from those for {Ui , Uj }.
The Cubic to Tetragonal Transformation. In this case it suces to
analyze the simple laminates supported on the two wells SO(3)U1 SO(3)U2
since
R23
{U1 , U3 }R23
= R13
{U2 , U3 }R13
= {U1 , U2 }.
152
The Cubic to Trigonal Transformation. In this case it suces to analyze

the simple laminates supported on the two wells SO(3)U1 SO(3)U2 and
SO(3)U1 SO(3)U3 since
R12
{U1 , U4 }R12
= {U1 , U2 },
R3 {U3 , U4 }R3 = {U1 , U2 },
R23
{U2 , U4 }R23
= {U2 , U1 },
R12
{U2 , U3 }R12
= {U2 , U1 }.
The Cubic to Orthorhombic Transformation. In this case it suces

to analyze the simple laminates supported on SO(3)U1 SO(3)U2 and on
SO(3)U1 SO(3)U3 since
R3 {U1 , U4 }R3 = {U1 , U3 },
R12
{U1 , U6 }R12
= {U1 , U3 },
R1 {U2 , U4 }R1 = {U1 , U3 },
R12
{U2 , U6 }R12
= {U2 , U3 },
R12
{U1 , U5 }R12
= {U1 , U3 },
R13
{U3 , U5 }R13
= {U2 , U3 },
R13 {U4 , U5 }R13 = {U1 , U4 },
R13
{U5 , U6 }R13
= {U1 , U2 }.
R13
{U3 , U6 }R13
= {U2 , U3 },
R13 {U4 , U6 }R13 = {U1 , U4 },
R2 {U2 , U3 }R2 = {U1 , U3 },
R12
{U2 , U5 }R12
= {U2 , U3 },
R23
{U3 , U4 }R23
= {U1 , U2 },
The Tetragonal to Monoclinic Transformation. It suces for both the

tetragonal to monoclinic transformations analyzed in Section 5.4 to analyze
the simple laminates supported on the pairs of wells SO(3)U1 SO(3)U2 and
on SO(3)U1 SO(3)U3 . In fact, we have for the type-I transformation
R2 {U1 , U4 }R2 = {U1 , U3 },
R3 {U2 , U4 }R3 = {U1 , U3 },
R1 {U2 , U3 }R1 = {U1 , U3 },
R12
{U2 , U4 }R12
= {U1 , U2 },
and for the type-II transformation
R12
{U1 , U4 }R12
= {U1 , U2 },
R1 {U2 , U4 }R1 = {U3 , U1 },
R12
{U2 , U3 }R12
= {U2 , U1 },
R3 {U2 , U4 }R3 = {U1 , U2 }.
6. Algorithmic Aspects
The numerical analysis of nonconvex variational problems, microstructures,

and Young measures poses a wealth of challenging problems, both from the
analytical and the computational point of view. One of the fundamental issues in this context is the question of what the appropriate objects are that
one wants to obtain as the output of a computation: a (highly oscillating)
minimizer of the nonconvex energy in a nite element space, a minimizer of
the relaxed problem, or a discretization of the Young measure.
It turns out that each of these choices has its own advantages. We discussed in Chapter 4 properties of minimizers of nonconvex variational problems in nite element spaces, and we described the information they contain
about the minimizing microstructure for ane boundary conditions that are
the barycenter of a simple laminate. In this chapter we focus on the other
two aspects - computation of relaxed functionals and of Young measures.
The idea to replace the original variational principle by the relaxed one
is motivated by the observation that the implementation of a direct energy
minimization in a nite element space faces two serious obstacles. First, there
is no class of steepest descent type algorithms known that produces reliable
results for nonconvex energies without ingenious guesses for the initialization
of the routine. Secondly, the rigidity results for Lipschitz functions whose
gradients are functions of bounded variation show that minimizers of variational problems modeling solid to solid phase transformations must have an
intrinsically complicated structure that renders them inaccessible to nave
approximation schemes.
The relaxed energy has under suitable growth conditions on the density
W an integral representation given by

qc
J (u) =
W qc (Du)dx.
Here W
qc
denotes the quasiconvex envelope of W dened by

1
qc
W (F ) =
inf
W (Du)dx.
uW 1, (;R3 ) ||
(6.1)
u(x)=F x on
Equivalently, W qc can be dened as the largest quasiconvex function less

than or equal to W . If we dene the rank-one convex envelope W rc and the
G. Dolzmann: LNM 1803, pp. 153175, 2003.

154
polyconvex envelope W pc analogously, then

W rc W qc W pc .
We therefore obtain an upper bound for the relaxed functional by replacing
W qc by (an approximation of) W rc . There are no general, structural conditions known that ensure the equality of W rc and W qc , but these envelopes
coincide in all examples for which explicit characterizations have been obtained. The inmum of the nonconvex energy is equal to the minimum of the
relaxed energy, and it is therefore expected that a numerical minimization
of the relaxed functional does not generate mesh sensitive oscillations. The
drawback of this approach is that exactly these oscillations contain a lot of
information about the underlying microstructure which is dicult to recover
from the minimizers of the relaxed energy. One approach to recover this information is the algorithm for the computation of laminates that we describe
in Section 6.2.
Another motivation for the computation of rank-one convex envelopes of
functions is the computation of rank-one convex hulls of sets. The rank-one
convex hull of a compact set K can be characterized by

K rc = X : distrc (X, K) = 0 ,
and is thus equal to the zero set of the rank-one convexication of the distance
function to K.
6.1 Computation of Envelopes of Functions

Our approach to the approximation of rank-one convex functions is rather
in the spirit of nite dierence schemes than nite element methods. The
algorithm computes a function f h dened on a uniform grid
Gh = {h F : F Zmn }
in the space of all matrices. The reason for this choice is a surprising rigidity
of nite element spaces. To illustrate this, consider for n = 2, m = 1, and
N N a uniform triangulation Th of the square Q = [ 0, 1 ]2 with nodes
nij = (i/N, j/N ) and triangles, the edges of which are parallel to the vectors
{(1, 0), (0, 1) (1, 1)}, see Figure 6.1. Assume that we want to approximate the
convex function
f (x1 , x2 ) = |x1 + x2 1|
in the space Sh (Th ) of all continuous functions which are ane on the elements in the triangulation. At a rst glance, it seems natural to seek

fh = sup gh f : gh S1 (h ), gh is D-convex .
155
jh
(j1)h
(i1)h
ih
Fig. 6.1. The function f (x1 , x2 ) = |x1 + x2 1| cannot be approximated

by D
convex functions gh f in S1 (h ) with D = (1, 0), (0, 1), (1, 1) . The dashed
line corresponds to the zero set of f , and the arrows on the right hand side of the
gure indicate the directions of convexity.
However, there are two fundamental diculties with this approach. First, the
maximum of two functions in S1 (h ) is not necessarily in S1 (h ). Secondly,
nite element spaces are too rigid to approximate convex functions from
below. Indeed, we assert that
|gh f |; 1
for all gh Sh , gh f and gh is D-convex.
Consequently the scheme does not converge for h 0.

In order to prove this estimate, we x a D-convex function gh S1 (h )
such that gh f and consider gh on the cube
Ci,j = [ (i 1)h, ih ] [ (j 1)h, jh ],

see Figure 6.1. Let Mi,j = (i 12 )h, (j 12 )h be its center. Since gh is ane
on the two triangles in Ci,j , we have
gh (Mi,j ) =
gh (ni1,j1 ) + gh (ni,j ) ,
2
while the convexity of gh in direction (1, 1) implies that

gh (Mi,j )
gh (ni1,j ) + gh (ni,j1 ) .
2
We dene
di,j = gh (ni1,j ) gh (ni1,j1 ) + gh (ni,j1 ) gh (ni,j )
and we deduce from the two foregoing estimates that
di,j 0.
156
If we take the sum of all these inequalities for i, j {1, . . . , 2k }, we obtain

k
2

di,j = gh (0, 1) gh (0, 0) + gh (1, 0) gh (1, 1) 0.
i,j=1
By assumption, gh f and therefore gh (0, 1) 0, gh (1, 0) 0, and thus

gh (0, 0) + gh (1, 1) gh (0, 1) + gh (1, 0) 0.
This implies that at least one of the two values gh (0, 0) and gh (1, 1) has to
be less than or equal to zero and therefore the L norm of f gh is at least
one. The same diculty arises if one denes alternatively (for smooth enough
functions)

f%h = sup gh 1 f : gh S1 (h ), gh is D-convex ,
where 1 given by 1 f (p ) = f (p ) for all nodes p in the triangulation
denotes the interpolation operator onto S1 (h ).
The Algorithm for the Computation of Rank-one Convex Envelopes. Our approach to the computation of an approximation f h of the
D-convex envelope of a given function f uses a discretization for both the
function and the set of directions D. For h > 0 xed we let

Gh = hF : F Zmn ,
and we choose a set of directions Dh Gh as an approximation of D. A
typical choice which allows us to obtain explicit estimates is

Dh = h(a b) : a Zm , b Zn , |a| , |b| h1/3 ,
see the statements of the theorems below. The advantage of this choice is
that it leads naturally to a robust denition of a discretely Dh -convex approximation of f .
The idea behind the algorithm is to perform convexications along rankone lines until the function is stable under this operations. In view of our
denitions, the line
(F, R) = {F + tR, t R}
intersects the grid Gh in innitely many, evenly spaced points for all matrices
F Gh and all directions R Dh . We consider f along this line as a function
of one variable dened
in a set of nodes and denote its piecewise ane inter'
polation by f '(F,R) , see Figure 6.3. We now dene the notion of discrete
D-convexity for functions dened on Gh .
ST ART (f, Gh,Q , Dh )
157
i := 0; fih := f |Gh,Q
?
g := fih
h
R Dh , F Gh,Q
g h := convexify(g h , F, R, Gh,Q )
h
fi+1
:= g h ; i := i + 1
?
HH
H

H
H
NO
h
fih fi1

<
EP
S H
HH

HH

HH

Y ES
ST OP (fih )
Fig. 6.2. The algorithm for the computation of the rank-one convex envelope.
Denition 6.1.1. Let Dh Gh . We say that

' a function f : Gh R is
discretely Dh -convex on Gh if the functions f '(F,R) are convex (as functions
of one variable) for all F Gh and all R Dh . The discretely Dh -convex
envelope f h,D of f is the largest discretely Dh -convex function g h f .
Remark 6.1.2. The pointwise maximum of two discretely Dh -convex functions is a discretely Dh -convex function. Thus f h,D is well-dened.
Remark 6.1.3. For computations it is necessary to use a nite domain, and
we denote by Gh,Q the intersection of the grid with a cube Q Mmn . We
then dene discretely Dh -convex functions on Gh,Q analogously.
The algorithm we are now going to describe is based on an idea by Kohn
and Strang for the case of rank-one convexity, i.e., for D being the space of
all rank-one matrices. Let f0 = f and dene iteratively
158
0.4
0.2
2
-0.2
-0.4
-0.6

Fig. 6.3. Sketch of the functions g h (F,R) (solid line) and gh (dotted line) in
the algorithm for the computation of discretely D-convex functions.

fi+1 (F ) = inf fi (A) + (1 )fi (B) : F = A + (1 )B, [ 0, 1 ],

and A B parallel to a non-zero direction in D .
Then
f D = lim fi .
i
A dierent way to dene the functions fi is the following. For F , R Mmn

dene fF,R (t) = f (F + tR). Then

fi+1 (F ) = inf (fi )F,R (0) : R D ,
where g denotes the convex envelope of the function g. This interpretation
of the formula can easily be translated into an algorithm which we describe
schematically in Figure 6.2. In the ow chart we use a parameter EP S to
control the performance of the algorithm. Since it is a priori not clear whether
the discretely rank-one convex function can be determined with nitely many
iterations, one needs to introduce a stop criterion. The proposed algorithm
h
terminates if the dierence between the functions fih and fi+1
are small.
The input consists of the restriction of the given function f to a uniform
grid of width h and a set Dh Gh of directions; the output is (for EP S > 0
an approximation to) a discretely Dh convex function dened on Gh,Q . The
central part of the algorithm is a subroutine convexify(g h , F, R, Gh,Q ) which
computes the convexication of the restriction of the function g h to the intersection of the grid Gh with a line (F, R). The convexication
is dened as
'
the largest convex and piecewise ane function gh g h '(F,R) restricted
to the nodes in Gh on (F, R), see Figure 6.3.
The next theorem describes the properties of the functions fih .
Theorem 6.1.4. Assume that h > 0 and that f : Mmn R is continuous.
Then there exists a discretely Dh -convex function f h : Gh,Q R such that
the functions fih dened in the ow-chart of the algorithm in Figure 6.2 with
EP S = 0 converge to f h .
159
The case of the rank-one convex envelope, which corresponds to the choice
of all rank-one matrices for D, has attracted a lot of attention because of
its close connection to the quasiconvex envelope. In particular, a reliable
algorithm for the computation of the rank-one convex envelope can be used
to get upper bounds on the quasiconvex envelope. There are two important
assumptions one has to make in order to state an explicit convergence result:
On the one hand, the D-convexication of any function f depends in general
on the values of f on the entire space Mmn , not only on a nite neighborhood
of a given point. Since computations can be performed only on compact sets,
one has to assume a condition of the form (f D )|Q = (f|Q )D (see also the
remark following the theorem below). This will ensure convergence of the
functions f h to f D as h 0. A sucient condition is, for example, that
f g D on Mmn with f = g D on Mmn \ Q. On the other hand, the proof
relies on the representation
f rc (F ) = inf
N

i (Fi ) : (i , Fi ) HN , F =
N

i=1

i Fi ,
(6.2)
i=1
where the inmum is taken over all pairs (i , Fi )i=1,...,N satisfying condition
HN in Denition A.1.4. There is no criterion known that ensures that the
envelope for a given function can be obtained with a nite N . Therefore
convergence can only be obtained if this is assumed. In this situation the
following convergence result holds.
Theorem 6.1.5. Assume that f is Lipschitz continuous and that there exists
a rank-one convex function g : Mmn R such that f g on Q and f = g
on Mmn \ Q. Suppose in addition that f rc can be computed by formula (6.2)
with N N0 , and that

Dh = h(a b) : a Zm , b Zn , |a| , |b| h1/3 .
Then there exists a constant C which depends only on m, and n such that
f rc f h L (Gh,Q ) C|f |1,;Q h1/3 ,
where f h is the rank-one convex function the existence of which is guaranteed
in Theorem 6.1.4.
Remark 6.1.6. The condition that f coincides with a rank-one convex function outside a compact set can of course be weakened. For example, to get
convergence of the algorithm at the established rate on a compact set Q1 , it
is sucient that the rank-one convex envelope can be obtained by using only
points in some compact set Q2 . Then one runs the algorithm on Q2 .
The proof of the theorem relies on the following approximation result.
160
0
-0.5
-1
-1.5
-2
-1.4 -1.2
-0.8 -0.6 -0.4
Fig. 6.4. The logarithm (base 10) of the L -norm versus the logarithm of the
width of the grid for the modied Kohn-Strang example. The dashed line has slope
one.
Lemma 6.1.7. Assume that h (0, 1), that f and Dh are as in Theorem 6.1.5. and that the pairs (i , Fi ), i = 1, . . . , N satisfy condition HN
,N
with F = i=1 i Fi . Suppose that F h Gh satises |F F h | c0 h1/3 . Then
there exist pairs (hi , Fih ), i = 1, . . . , N , which satisfy condition HN and a
constant c1 which depends only on m, n, and maxi=1,...,N |Fi | such that
,N
h
i) Fih Gh and
F h = i=1 hi F
;
i 1/3
h
ii) |Fi Fi | c0 + (N 1)c1 h
for i = 1, . . . , N ;
iii)we have the estimate
N
'

''
'
i f (Fi ) hi f (Fih ) ' c0 + (N 1)c1 |f |1, h1/3 .
'
i=1
This lemma is also the key ingredient in the convergence proof for the
computation of laminates in Section 6.2 below.
We now present the results of some numerical experiments for the computation of envelopes of functions and sets. In our examples we use the sets
Dh of rank-one directions dened by
Dh,k = {hR : R = a b, a Zm , b Zn , |a| , |b| k}.
The Kohn-Strang Example. One classical example in this context is the

function

1 + |F |2 if F = 0,
f (F ) =
0
else,
which was originally studied in the context of optimal design problems by
Kohn and Strang. This example is very appealing because the semiconvex
envelopes can be calculated explicitly. In fact, f rc = f pc where
161
Table 6.1. Numerical results for the modied Kohn-Strang function in (6.4), see
also Figure 6.4. The formula for the minimizing laminate (6.3) shows that the full
rank-one cone is used in the constructions. This is reected in the fact that the error
decreases as the parameter k in the sets Dk increases. In addition, a third iteration
of the algorithm further reduces the error.
h
Dk
1st iteration
2nd iteration
3rd iteration
0.125
0.070 718
0.067 708
0.067 188
0.035 938
0.031 250
0.031 250
0.034 636
0.031 250
0.031 250
0.034 636
0.031 250
0.031 250
0.077 139
0.076 384
0.076 384
0.022 042
0.021 856
0.021 354
0.019 142
0.013 951
0.013 238
0.019 142
0.013 951
0.013 238
0.076 660
0.076 489
0.076 488
0.027 269
0.026 795
0.026 795
0.012 500
0.009 863
0.009 732
0.010 326
0.004 561
0.004 390
0.0625
0.03125
1 + |F |2 if (F ) 1,
2( D) if (F ) 1,

where D = | det F | and (F ) = |F |2 + 2D. We sketch the proof of this
formula following Kohn and Strang for the convenience of the reader and in
order to emphasize two important consequences of the calculation.
Since f is convex on all rank-one lines that do not pass through zero, it
is clear that
0
if F = 0,
2|F | if rank(F ) = 1, |F | 1,
f1 (F ) =
1 + |F |2 otherwise.
f rc (F ) =
The calculation of f2 requires an inspired guess. We use polar decomposition

to write F as
F = QU = 1 u1 v 1 + 2 u2 v 2 ,
where 1 , 2 are the singular values of F and {u1 , u2 } and {v 1 , v 2 } are
orthonormal vectors in R2 . Then F can be rewritten with (0, 1) as

2
u2 v 2 ,
F = 1 u1 v 1 + (1 ) 1 u1 v 1 +
1
(6.3)
162
and thus

2
u2 v 2 .
f2 (F ) f1 1 u1 v 1 + (1 )f2 1 u1 v 1 +
1
For 1 1 this implies the estimate

f2 (F ) 21 + (1 ) 1 + 21 +
22
.
(1 )2
In case that 1 + 2 1 this expression is minimized for 1 = 2 /(1 1 )

and this leads to

if 1 + 2 1,
1 + 21 + 22
f2 (F ) =
2(1 + 2 ) 21 2 if 1 + 2 1.
Hence f rc f2 and since f2 turns out to be polyconvex, f rc = f2 . The
derivation of the formula has two important implications.
First, f rc (F ) can be obtained for all F as a laminate which is supported
on at most four points, and since f and f rc agree with the convex function 1 + |F |2 outside of the compact set |F | 1, all the assumptions in
Theorem 6.1.5, except the continuity of f , are satised. Replacing f by a
continuous function f with f rc f f all the hypotheses are fullled and
the algorithm converges at least at the established rate. One possible choice
for f is

1
+ |F |2 if |F | 2 1,
f (F ) =
(6.4)
2 2|F | if |F | 2 1.
We use this function in our numerical experiments and summarize our results
in Figure 6.4.
Secondly, the entire rank-one cone is used in the construction of the second order laminates. The impact of this on the computation can be seen in
Table 6.1: the approximation error decreases as we pass from D1 to D4 , and
the fact that the error is further reduced in a third iteration of the convexication routine is another manifestation of the fact that not all the directions
used in the construction of the relaxation are contained in the discrete sets
Dh .
The Eight Point Example. We now consider the eight point set in Section 2.1. Let f be the 1 -distance to the set K given by
x y

K=
: |x| = a, |y| = b, |z| = c
yz
with a, b, c > 0 and ac b2 0. For our experiments we chose a = 34 , b = 12 ,
and c = 14 . In this case,
K (4) = K lc = K rc = K qc K pc ,
see Theorem 2.1.1. The results of our computation with Q = [ 1, 1]4 and
h = 1/16 are summarized in Figure 6.5.
6.2 Computation of Laminates
163
Fig. 6.5. Computation of the rank-one convex hull for the eight point set. The left
gure shows the intersection of K pc with the diagonal matrices (y = 0) which is
formed by the intersection of two hyperbolae. The right gure displays the smaller
set K rc , which is K pc intersected with two additional cones. The dots correspond to
the grid points in Gh with h = 1/16 in which the discrete rank-one convex envelope
of the 1 -distance to K is smaller than 0.0001.

We now present a variant of the algorithm for the computation of rank-one
convex envelopes that is designed for the computation of laminates. At the
same time, it provides information about oscillations in minimizing sequences.
Assume for example that we choose boundary conditions u(x) = F x which
enforce the formation of a unique microstructure in the nonconvex problem.
Then we can obtain an approximation of the microstructure from minimizers of the nonconvex energy based on the methods described in Chapter 4.
The relaxed problem, however, is also minimized by the ane deformation
u(x) = F x which does not provide any information about the underlying
microstructure. Suppose now that the rank-one convex and the quasiconvex
envelope of W coincide at F . Then there exists a laminate with center of
mass F such that W qc (F ) = W rc (F ) = , W . This laminate can be approximated by nite laminates N supported on N matrices, and Theorem 6.2.4
ensures that the proposed algorithm computes a laminate h supported on
on a discretization Gh of the space of all matrices with
|h , W | |N , W | + 5(N 1)c1 mn |f |1, h1/3 .

We therefore obtain for h small enough approximating laminates h for which
h , W is arbitrarily close to W rc (F ) = , W . These laminates represent
microstructures with small energy which constitute a good approximation to
the minimizing microstructure.
An algorithm similar in spirit was proposed by Aranda and Pedregal.
They focus on laminates realizing a given matrix F as the center of mass and
try to construct a generalized solution of the variational problem by splitting
F in an optimal way along rank-one lines. One therefore minimizes
164

; ; ; ; ; ; ; ; ;

; ; ; ; ; ;

;
;
;
;
;

; ; ; ; ; ;

; ; ; ; ; ; ; ; ;
2
Fig. 6.6. A typical problem with splitting algorithms if one tries to split a xed
matrix. The computation for the point in the center with directions (1, 0) and (0, 1)
(corresponding to separate convexity) gets stuck after the rst step, in which the
value is found to be 1. However, the correct value in the center is 2, see also
Table 6.2. The right gure shows the values of the function f1 in (6.5), which is generated at each node by minimizing among all splittings with direction {(1, 0), (0, 1)}.
(R, t+ , t ) =
t+
t
W
(F
t
R)
+
W (F + t+ R)
t + t+
t + t+
among all rank-one matrices R and parameters t and t+ 0. This step

involves a discretization of the set of all rank-one matrices, which can for
example be realized by taking R = sab, where (s, a, b) belongs to a suitable
discretization of a compact subset of RSm Sn . The same procedure is then
applied to F = F t F . However, at least any deterministic implementation
faces the typical diculty in the numerical analysis of nonconvex problems:
the algorithm nds a local minimum, but not the global one. See Figure 6.6
for a cartoon of this situation.
Here we propose a dierent approach, namely to compute simultaneously
Young measures for all matrices in a given subset of a discretization of Mmn .
In fact, the algorithm for the computation of D-convex envelopes can easily
be modied to nd certain laminates which consist of nitely many atoms
supported on an equidistant grid Gh . In the following we describe rst the
algorithm and prove its convergence, and then we present results of numerical
experiments.
We dene by induction classes of laminates Lk (Q) which are supported
on at most 2k matrices in the cube Q and can be generated by successively
splitting a single Dirac mass along rank-one lines. Let
L0 (Q) = {F : F Q}.
If

M

k =
i Fi Lk (Q), M 2k , i > 0,
i=1
M

165
i = 1,
i=1
then all laminates that can be generated by splitting (up to relabeling) the
rst M , M M , of the matrices Fi along rank-one lines belong to Lk+1 (Q):
k+1 =
M

+
i F + + i F +
i=1
M

i Fi Lk+1 (Q),
i=M +1
if
i > 0, i = i + i , and
Fi Q, rank(Fi+ Fi ) = 1, Fi =
i
Fi+ + i Fi
i
i
for i = 1, . . . , M .
For computations one needs to dene a subclass of laminates in Lk (Q) which

is supported on a nite set of points in Gh,Q and uses only a nite set Dh of
directions.
Denition 6.2.1. Let L0 (Gh,Q , Dh ) = {F : F Gh,Q }. If
k =
M

i Fi Lk (Gh,Q , Dh ), M 2k , i > 0,
i=1
M

i = 1,
i=1
then all laminates that can be generated by splitting (up to relabeling) the rst
M , M M , of the matrices Fi into matrices Fi+ , Fi Gh,Q along lines
parallel to directions in Dh belong to Lk+1 (Gh,Q , Dh ):
k+1 =
M

i=1
+
i Fi+
i Fi
M

i Fi Lk+1 (Gh,Q , Dh ),
i=M +1
if
i > 0, i = i + i , Fi Gh,Q , Fi Fi D Dh , and
rank(Fi+ Fi ) = 1, Fi =
i
Fi+ + i Fi
i
i
for i = 1, . . . , M .
Remark 6.2.2. If Lk (Gh,Q , Dh ), then can be written as a convex combination = + + (1 ) with + , Lk1 (Gh,Q , Dh ) such that
F + = + , id and F = , id belong to Gh,Q , F + and F are rank-one
connected with F + F D Dh , and F = F + + (1 )F . This follows from the fact that in the rst step F is split into two Dirac masses
centered in points in Gh,Q along a rank-one line parallel to a direction in
Dh , F = F + + (1 )F . The subsequent splitting steps generate Young
measures + and with center of mass equal to F + and F , if the weights
are rescaled by and 1 , respectively.
166
Remark 6.2.3. If
=
N

i Fi Lk (Gh,Q , Dh ),
i=1
then the pairs (i , Fi )i=1,...,N satisfy condition HN in Denition A.1.4.

The observation here is that the algorithm for the computation of the
rank-one convex envelope of a given function f implicitly constructs minimizing Young measures in Lk if the basic convexication routine is executed k times. Assume for simplicity that f is rank-one convex outside
Q = [ 1, 1 ]mn and that Gh,Q = Gh Q = {F1 , . . . , F }. For a given function
f : Gh,Q R, a given set of directions Dh Gh , and F Gh,Q we dene an
optimal splitting of F ,

F = (f, F )F + (f, F ) + 1 (f, F ) F (f, F )
with [ 0, 1 ] and F + F parallel to a direction D Dh by requiring
that

(f, F )f F + (f, F ) + 1 (f, F ) f F (f, F )

= min f (G+ ) + (1 )f (G ) : G+ , G Gh,Q , [ 0, 1 ],

G+ G D Dh , F = G+ + (1 )G .
The Algorithm for the Computation of Laminates. The algorithm is
now dened in the ow-chart in Figure 6.7. The output for the computation
of laminates in Lk for all points in Gh,Q consists of an k eld fractions
(which contains the volume fractions of F + in the optimal splittings) and
an (2k) eld atoms (which contains the matrices F in the optimal
splitting). In order to describe the performance of the algorithm and to prove
its convergence, we dene fih to be the function that is generated in the i-th
h
iteration of the algorithm, i.e., f0 = f |Gh,Q and fi+1
is generated from fih by
nding an optimal splitting of each matrix Fj with respect to fih ,

h
(Fj ) = (fih , Fj )fih F + (fih , Fj ) + 1 (fih , Fj ) fih F (fih , Fj ) .
fi+1
Similarly we construct from the output variables fractions and atoms laminates j,i Li (Gh,Q , Dh ) with center of mass Fj and i splitting levels by
j,0 = Fj
and

j,i+1 = fraction[j, i] ,i + 1 fraction[j, i] ,i
where
F = F + (fih , Fj ) = atoms[j, 2i],
and
F = F (fih , Fj ) = atoms[j, 2i+1],


167

Fig. 6.7. The algorithm for the computation of laminates.
are the matrices arising in the optimal splitting of Fj in the (i + 1)st step. It
follows from the foregoing denitions that
fih (Fj ) = j,i , f .
(6.5)
Indeed, f0h (Fj ) = f (Fj ) = j,0 , f , and if the assertion holds for fi and j,i ,
then

h
fi+1
(Fj ) = (fih , Fj )fih F + (fih , Fj ) + 1 (fih , Fj ) fih F (fih , Fj )

= (fih , Fj ),i , f + 1 (fih , Fj ) ,i , f
= j,i , f .
168
Table 6.2. Output of the algorithm with depth = 2 (many of the splitting steps have
more than one minimizing splitting, and therefore the output generated depends in
general on the implementation of the algorithm). A indicates that the measure
is not split in this step, and that only one atom is dened.
matrix
atoms
fractions
20
10
30
20
1
2
21
21
20
24
22
21
23
20
24
1
2
23
23
20
24
3
4
1
2
1
4
24
14
34
24
1
2
Before we state the main theorem of this section, we include an example

for the generation of the measures j,i . We assume that the points in Figure 6.6 have integer coordinates (0, 0) (lower left corner) and (4, 4) (upper
right corner). For simplicity we refer to the point with coordinates (i, j) as
ij, and we generate the measures for the central point 22. The output of the
algorithm with depth= 2 is summarized in Table 6.2. By denition,
22,1 =
1
1
21 + 23 ,
2
2
20,1 =
1
1
10 + 30 ,
2
2
24,1 =
1
1
14 + 34 ,
2
2
and hence
1
1
1
1
1
1
20,1 + 24,1 = 10 + 30 + 14 + 34 .
2
2
4
4
4
4
It is important to note that 22,2 is obtained by starting from the splitting
generated in the second iteration of the algorithm and not by further splitting
the matrices found in the rst splitting step. We now state our convergence
result for laminates in Lk (Q) which are supported on at most 2k points.
Recall that we assume that the points in Gh,Q have been labeled from 1 to .
22,2 =
Theorem 6.2.4. Assume that Q = [ 1, 1 ]mn and f W 1, (Q). Suppose

that Lk (Q) with , id = Fj [ 12 , 12 ]mn , j {1, . . . , } is supported
on N matrices and that h > 0 is small enough so that dened by
= 1 2(N 1)c1 h1/3
(6.6)
is greater than zero. Finally, let h = j,k be the Young measure computed by
the algorithm in Figure 6.7 with depth= k and

Dh = h(a b) : a Zm , b Zn , |a| , |b| h1/3 .
Then
|h , f | |, f | + 5(N 1)c1 mn|f |1, h1/3 ,
where c1 is the constant appearing in Lemma 6.1.7.
(6.7)
169
Proof. We divide the proof in two steps. The rst step shows that the measures j,i are minimizing in Li , and the second one proves the estimate (6.7)
based on the construction of discrete laminates in Lemma 6.1.7.
Step 1: Optimality of j,i . We have for i = 0, . . . , k, and j = 1, . . . , , that

j,i , f = min , f : Li (Gh,Q , Dh ), , id = Fj .
This statement is obvious for i = 0 and i = 1, since the algorithm chooses
an optimal splitting. Assume now that the assertion has been established for
i1, and suppose that it does not hold for i. Then there exists a j {1, . . . , }
such that

j,i , f > min , f : Li (Gh,Q , Dh ), , id = Fj ,
and let be an element in Li (Gh,Q , Dh ) realizing the minimum on the right
hand side (this minimum exists since Li (Gh,Q , Dh ) is a nite set). We may
split as
+ + (1 )
,
=
Li1 (Gh,Q , Dh ),
[ 0, 1 ]
(see the remark following Denition 6.2.1), such that F + =

+ , id and
F =
, id belong to Gh,Q with
F + + (1 )
F ,
Fj =
rank(F + F ) 1.
Since F Gh,Q , there exist indices , {1, . . . , } such that F + = F

and F = F . In view of the fact that the measures ,i1 and ,i1 are
minimizing, we conclude
fih (Fj ) = j,i , f >
, f

=
+ , f + (1 )
, f
,i1 , f + (1 )
,i1 , f
f h (F ) + (1 )
f h (F ).
=
i1
i1
This implies that there exists a better splitting for Fj in the i-th step, and
this contradicts the denition of the loop in the algorithm.
Step 2: Proof of the error estimate (6.7). We now construct explicitly
a Young measure h Lk (Gh,Q , Dh ) which is close to in the sense of
Lemma 6.1.7. The estimate (6.7) is then a consequence of the Lipschitz continuity of f . The only diculty here is that we cannot apply Lemma 6.1.7
directly, since the generated Young measure might not belong to Lk (Gh,Q , Dh )
(the constructed laminate is supported on matrices which lie in an O(h1/3 )neighborhood of the support of ). The remedy here is to construct from
by rescaling a Young measure Lk (Q) for which the distance of the atoms
170
in its support to the boundary of Q is suciently large. With > 0 given

by (6.6) we dene : Mmn Mmn by
(F ) = (F Fj ) + Fj .
Suppose that Lk (Q) is given by
=
N

i Xi ,
i=1
and let
=
N

i Yi
with
Yi = (Xi )
and
Fj =
i=1
N

i Yi .
i=1
Since is ane, it is clear that Lk (Q), and by construction

'
'
' '
|Xi Yi | = 'Xi (Xi Fj ) + Fj ' = 'Xi Fj (Xi Fj )'
= (1 )|Xi Fj | 4(N 1) mn c1 h1/3 ,

We now assert that the -distance of the matrices Yi to Q is at least
(N 1)c1 h1/3 . By denition
'
'
|Yi;k | = '(Xi;k Fi;k ) + Fi;k '
(1 )|Fi;k | + |Xi;k |,
and since Fj [ 12 , 12 ]mn , and Xi Q we obtain
1
|Yi;k | (1 ) +
2

1 1
= + 1 2(N 1)c1 h1/3

2 2
= 1 (N 1)c1 h1/3 .
Lemma 6.6 now ensures the existence of a laminate h Lk (Gh,Q , Dh ) represented by
h =
N

hi Yih
with Yih Gh,Q
and Fj =
i=1
N

i=1
Moreover, the points Yih are close to Yi in the sense that

|Yih Yi | (N 1)c1 h1/3
and we have the estimate
for i = 1, . . . , N,
hi Yih .
F4
171
F1
J3
J4
J2
J1
F2
F3
Fig. 6.8. The four point conguration in the diagonal matrices that supports an
innite laminate. The lamination convex hull is given by the four line segments and
the square with corners J1 , . . . , J4 .
N
'
''
'
i f (Yi ) hi f (Yih ) ' (N 1)c1 |f |1, h1/3 .
'
i=1
Since h = j,k is by Step 1 minimizing in Lk (Gh,Q , Dh ) we may estimate

|h , f | |
h , f |
N
'
'
'
'
='
hi f (Yih )'
i=1
N
N
'
'' ''
''
'
'
i f (Yih ) i f (Yi ) ' + '
i f (Yi ) i f (Xi ) '
i=1
N
'
'
'
'
+'
i f (Xi )'
i=1
i=1
5(N 1)c1 mn|f |1, h1/3 + |, f |.

The proof is now an immediate consequence of Steps 1 and 2.

Numerical Experiments for the Computation of Laminates. We conclude this section with two numerical experiments for the generation of laminates by the algorithm in Figure 6.7. We rst use our scheme to nd approximations of the innite laminate supported on four points. Then we report
on computations for eight point set in Theorem 2.1.1.
An Innite Laminate. A canonical example for the performance of an
algorithm is the following four point set. Let
172
1 1
1 1
,
), F2 = diag(
, ),
4 2
2 4
1 1
1 1
),
F3 = diag( , ), F4 = diag( ,
4 2
2 4
F1 = diag(
and dene K = {F1 , F2 , F3 , F4 }. Then K does not contain any rank-one

connection, but the lamination convex hull of K is nontrivial, see Figure 6.8
for a sketch. We used the algorithm for the computation of laminates with
center of mass equal to the zero matrix on a grid with 17 grid points on
the one dimensional axes in the cube [ 1, 1 ]4 . With ve splitting levels, we
obtained the Young measure
5 =
39
40
40
39
2
2
F +
F +
F +
F +
J +
J
162 1 162 2 162 3 162 4 162 2 162 4
and the rank-one tree shown in Figure 6.9. Here

1
F + = diag( , 0),
4
F = 0,
1
F = diag( , 0).
4
One obtains the corresponding subtrees if one uses less than ve splittings
(this is not surprising, since the algorithm is completely deterministic).
The Eight Point Example. Recall that the eight point set K is given by

xy
K=
: |x| = a, |y| = b, |z| = c
yz
with a, b, c > 0 and ac b2 > 0. In this case,
K (4) = K lc = K rc = K qc = K pc ,
see Theorem 2.1.1 for the precise statement. For our experiments we choose
f to be the 1 distance to K with a = 34 , b = 12 , and c = 14 . In our rst
experiment we chose F = 0 and computed on Gh,Q with h = 1/8 and Dh,2 , a
set of 64 directions. The algorithm correctly nds that F K (2) and produces
the laminate
=
with
34 14
,
F1 =
14 12
7
5
5
7
F + F + F + F
24 1 24 2 24 3 24 4
3
4
14
,
F2 =
14 12
34
F3 =
1
4
1
4
1
2
It is easy to see that is indeed a laminate since

3
F1 F2 = e1 e1 ,
2
3
F3 F4 = e1 e1 ,
2
F4 =
3 1
4 4
1 1
4 2
1/3
2/3
F2
F1
F4
2/3 1/3
F3
1/2
F1
1/2
F2
1/3 2/3
F+
2/3
F3
1/3
J1
2/3
F4
1/3
J2
2/3
1/3
J3
J4
Fig. 6.9. Approximation of an innite laminate with nite ones. The laminate computed with the algorithm
using ve splitting steps. The shaded squares correspond to the points in which the Young measure is supported.
The volume fractions in the splittings are indicated along the lines denoting rank-one connections.
J2
J1
1/3 2/3
J4
1/3 2/3
J3

173
F1
3/4
F2
1/3
F3
3/4 1/4
G2
F4
F5
G3
1/4 3/4
1/2
F6
1/3
H2
2/3
F7
1/4 3/4
G4
Fig. 6.10. The laminate generated for the eight point set with center of mass equal to zero.
1/4
G1
2/3
H1
1/2
F8
174
and
175
7
5

14 12
5
7
.
F1 +
F2
F3 +
F4 =
12
12
12
12
12 1
It is not surprising that the result is exact since all the necessary rank-one
directions are contained in the set Dh . Minimizing laminates are not unique,
and if we restrict Dh to Dh,1 , a set of 16 matrices, then we we obtain the
following laminate supported on all eight matrices in K and with center of
mass equal to zero:
=
6
2
3
1
1
3
2
6
F +
F +
F +
F +
F +
F +
F +
F ,
24 1 24 2 24 3 24 4 24 5 24 6 24 7 24 8
where (it is convenient to relabel the matrices)
3
1
34 14
34 14
4
, F2 =
, F3 = 4
F1 =
,
14 12
14 21
14 12
34 14
34 41
34 41
, F6 =
, F7 =
,
F5 =
1
1
1
1
1
1
4
2
4
2
4
2
3
4
F4 =
14
3
F8 = 4
1
4
1
4
1
2
1
4
1
2
In this notation, F1 F2 , F3 F4 , F5 F6 , F7 F8 are all parallel to e2 e2 .

If one denes
3
F1 +
4
1
G 3 = F5 +
4
G1 =
1
F2 ,
4
3
F6 ,
4
3
F3 +
4
1
G4 = F 7 +
4
G2 =
1
F4 ,
4
3
F8 ,
4
then G1 G2 and G3 G4 are parallel to e1 e1 . Finally, with H1 and H2

given by
H1 =
2
1
G1 + G2 ,
3
3
H2 =
1
2
G3 + G4 ,
3
3
the matrix H1 H2 is parallel to (e1 + e2 ) (e1 + e2 ) and

F =
1
1
H1 + H2 .
2
2
It remains a challenging problem to nd a class of reliable algorithms for

the minimization of nonconvex problems. Such a scheme would allow one to
design an integrated approach to the numerical analysis of microstructures by
combining computation of relaxed energy densities with the minimization the
relaxed functionals and an approximation of the underlying microstructures
in a postprocessing step via the computation of laminates.
7. Bibliographic Remarks
We begin by describing some important contributions that form the background for the mathematical theory in a broader context. Then we provide
detailed references for the material presented in the chapters of this text.
Ball&James [BJ87, BJ92] and Chipot&Kinderlehrer [CK88] derived the
mathematical description of fundamental mechanisms for the appearance of
microstructure in single crystals based on energy minimization, and Bhattacharya&Kohn [BhK96, BhK97] extended the theory to polycrystals. Magnetic eects, and the coupling of magnetic and elastic properties were analyzed by DeSimone [DS93], James&Kinderlehrer [JK93], Tartar [Ta95] and
DeSimone&James [DSJ97]. Kohn&M
uller [KM92, KM94] presented an analysis of domain branching which was then applied to magnetic domain patterns
in Choksi, Kohn&Otto [CKO99]. James&Hane [JH00] and Pitteri&Zanzotto
[PZ00] provided a detailed description of the crystallographic aspects. Further
applications include studies of formation of blisters by Ortiz&Gioia [OG94]
and Ben Belgacem, Conti, DeSimone&M
uller [BCDM00] and of dislocation
patterns by Ortiz&Repetto [OR99].
This research has led to new, simply stated, but deep questions in the calculus of variations which are closely related to Tartars earlier work and his far
reaching programme on oscillations in nonlinear partial dierential equations
and compensated compactness [Ta79, Ta83, Ta90]. At the heart of the analysis of microstructures in variational problems lie the notions of quasiconvexity
and of quasiconvex hulls which remain fty years after Morreys seminal work
[Mo52, Mo66] one of the fundamental challenges in the calculus of variations
with ramications to a number of other problems [Sv95, A98, B98].
7.1 Introduction
The fundamental contribution in the recent work by Ball&James [BJ87,
BJ92] and Chipot&Kinderlehrer [CK88] is to start from a variational approach and not from kinematic theories as in [BMK54, Er80, Er86, WLR53].
The degeneracies of the theory if one assumes invariance under all bijections
of the lattice were analyzed in [Er77, Er89, Fo87, Pa77, Pa81, Pi84, Za92].
Applications of the shape memory eect can be found in the conference proceedings [SMST97]. A beautiful account of the relations between the quasi-
G. Dolzmann: LNM 1803, pp. 177182, 2003.

178
convex hull of a set and properties of sequences converging to the set is given
in [Sv95]. The denition of the Young measure goes back to L. C. Young
and was introduced to the analysis of oscillations in partial dierential equations by Tartar, see, e.g., [Ta79, Ta83]. The version of the fundamental theorem on Young measures given in Section A.1 follows [B89]. Young measures
generated by sequences of gradients were characterized in [KP91], see Theorem A.1.6 for a statement of the result, and the averaging technique is one of
their technical tools. A detailed discussion of the dierent denitions for the
semiconvex hulls and the proofs of their equivalence can be found in [M99b].
Surprising existence results for nonconvex variational principles have been
ak by an adaption of Gromovs convex integration
obtained by M
uller&Sver
method [MS99a] and by Dacorogna&Marcellini [DcM99] using the Baire category argument. However, the rigidity results in [DM95] show that geometry
of these solutions is very dierent from laminates of nite order. The Lavrientiev phenomenon has been studied in [BM85, BK87, NM90] and in [Li95] in
connection with the approximation of singular minimizers. Approximations
of Young measures can be found in [NW93, NW95, CR00] for scalar problems
and in [AP00] for two-dimensional vector valued problems.
7.2 Semiconvex Hulls of Compact Sets

The fact that gradient Young measures supported on three matrices without
rank-one connections are trivial, i.e., a Dirac mass placed at one of the three
matrices, was rst proven in [Sv91]. The paper [Sv95] contains an overview
of the dierent semiconvex hulls of sets and their relation to compactness
questions. The example of the discrete set with eight points in Section 2.1 is
taken from [DcT00], see Remark 2.1.4 for an account of the partial results
in this paper. Generalizations of the notion of quasiconvexity in the sense
of Morrey to integrands that depend on higher derivatives were studied by
Meyers [M65]. New examples of quasiconvex functions on symmetric matrices
ak [Sv92b] and this paper contains a concise summary
were obtained by Sver
of the restrictions they impose on gradient Young measures. It is an open
problem whether these functions can be extended to the space of all matrices, see [MS99b] for partial results. The observation that one can nd on a
one-sheeted hyperboloid at every point two straight lines contained in the
ak [Sv93]
hyperboloid that correspond to rank-one directions is due to Sver
where he used this fact to derive the formula of the quasiconvex hull of two
martensitic wells, see the statement of his result in Theorem 2.2.2. This paper
also contains the proof that the semiconvex hulls of multi-well sets in twodimensions without rank-one connections are trivial. The locality property of
the rank-one convex hull in Proposition 2.1.5 was rst mentioned in [Pe93]
and later in the context of D-convex sets in [MP98]. Elegant proofs can be
found in [Kir00, Mt00]. The paper [BhD01] presents the semiconvex hulls
for k-well problems invariant under SO(2), O(2) and O(2, 3) in Sections 2.2
7.3 Macroscopic Energy for Nematic Elastomers
179
and 2.3 as well as the dimension reduction in Section 2.5. The proofs given
here simplify the original arguments. The theory of martensitic thin lms has
been described in [BJ99]. The results about the two-well problem in three dimensions can be found in [DKMS00]. Note that Theorem 2.6.4 establishes
the existence of a neighborhood of order one about the identity such that
the polyconvex hull of SO(3) SO(3)H is trivial for all H in this neighborhood. A rst result in this direction was obtained in [KL00] where the
authors prove that the quasiconvex hull is trivial if H is suciently close to
the identity. The proof relies on a linearization argument, Johns estimates
for deformations with bounded strains [Jo72a, Jo72b], and the quasiconvexity of the quadratic function q : M33 R, q(X) = (cof(X + X T ))33 ,
see Theorem A.1.3. The rst example of two incompatible wells with nontrivial polyconvex hull was constructed in [DKMS00], and the parameters
hi in this example violate the second set of inequalities in condition ii) in
Theorem 2.6.4 with h1 = h2 193.995 and h2 = h3 = h. Here h is the
larger of the two solutions of the equation h + 1/h = 14. The example given
in Proposition 2.6.5 slightly improves the original one. It is not known what
the optimal value for h is for which the polyconvex hull is trivial and how
the topology of the hulls changes as soon as they become larger. The formula
for the polyconvex hull of K in (2.53) has rst been proven in [DSD00]. This
paper contains also the result for the three-dimensional situation in the special case of 1 = 2 , and the same proof yields the result in the general case.
It follows also from the characterization of the semiconvex hulls without the
constraint that the determinant be positive in [DcT98], see Remark 11 in
[DcT00]. Our proof for the representation of K pc follows [DSD00]. The case
of arbitrary sets in 2 2 matrices that are given by conditions on the singular
values (with and without the restriction that the determinant be positive)
has been solved in [CT00]. The paper [B77] contains a characterization of
convex and polyconvex functions that depend on singular values. The function F 1 (F ) + 2 (F ) which we use is just an example of a rich family
of convex functions depending on the singular values of F . A more general
version of Theorem 2.7.5 can be found in [CDDMO01]. The proof presented
here is taken from [DSD01]. The problem to maximize the expression F : R
for R O(n) was also considered in [Ja86], Appendix 3.
7.3 Macroscopic Energy for Nematic Elastomers

The motivation behind the successful eorts to synthesize nematic elastomers
[FKR81, KpF91] was the attempt to reproduce the mesophases typical of a
liquid crystal within an amorphous, polymeric solid. The resulting physical
system has the translational order of a solid phase coupled to the orientational order of a nematic phase. Soft deformation modes whose occurrence
had been predicted by theory [GL89], were observed experimentally [KnF], in
association with the appearance of domain patterns with a characteristically
180
layered texture. The formation of these microstructures was explained by energy minimization in the framework of continuum models in [WB96, WT96].
The model considered here is due to Bladon, Terentjev&Warner [BTW93].
The mathematical analysis of this model started in [DSD00], where an explicit formula for the relaxation of the energy in a two-dimensional model
was obtained. The results presented here are based on [D01, DSD01]. The
crucial construction in the proof of the quasiconvexity of the envelope is due
ak [MS99a]. The relaxation result for nematic elastomers conto M
uller&Sver
stitutes one of the few explicitly know relaxation results and seems to be the
rst for an SO(3) invariant energy related to phase transformations, see e.g.
[Kh67, Pi91, K91, LDR95] for related relaxation questions. In particular it
allows one to explore the features of the relaxed energy in numerical simulations, and rst results in this direction have been reported in [CDD01]. A
dierent approach has been pursued in [ACF99], and an extension of the results to compressible models was proposed in [Sy01]. The maximal eigenvalue
of the cofactor matrix was also used in [BC97, BC99] to separate points from
polyconvex hulls of sets.
7.4 Uniqueness and Stability

The uniqueness and stability results are strongly inuenced by Luskins work
on uniqueness and stability of microstructure [L96a, L96b, LL98a, LL98b,
L98, BLL99, LL99a, LL99b, BL00, L00, EL01, BBL01]. Our approach to a
formulation of stability via the the distance d is inspired by the construction
of a distance in [CKL91] for the one-dimensional, scalar case. In our presentation we stress the point of view that the stability estimates should be
formulated for Young measures and not for functions. The idea to use the
cofactor in the proof of uniqueness of simple laminates is inspired by [BJ92]
and generalizes results in [BhD01]. The fact that polyconvex measures supported on two incompatible matrices have to be trivial (i.e. a single Dirac
mass placed at one of the two matrices) was shown in [BJ87]. The proofs in
Section 4.1 are an adaption of the ideas in [BLL99] to the framework with
polyconvex measures in the n-dimensional setting. In particular, the proof of
the estimates for the excess rotation becomes conceptually much clearer and
shorter than in [BLL99]. Possible extensions of this theory described in this
chapter are strongly limited by Iqbals results [Iq99] that the volume fractions
of laminates within laminates are typically not uniquely determined from the
center of mass of the Young measure. The rates for the nite element minimizer in Section 4.5 are based on a construction that has been known for a
long time. A more formal argument can be found in [C91, CC92, CCK95].
Based on ideas from [AcF84, AcF88, Liu77], Zhang constructed a truncation
method for functions generating gradient Young measures that was sharpened by M
uller (see [Z92, M99a]). The idea to formulate error estimates for
nite element minimizers in a distance d(, ) on the space of all probability
7.5 Applications to Martensitic Transformations
181
measures originates in [CKL91]. Our results cover all triangulations that are
regular in the sense of Ciarlet [Ci78].
7.5 Applications to Martensitic Transformations

The transformation mechanisms are well-understood in many materials and
we only mention a few typical examples that relate to the analytical results. Indium rich InTl alloys typically undergo a transformation from a
face-centered cubic phase into a face-centered tetragonal phase [Gt50]. The
transformation strains given in the text for the cubic to trigonal phase transformation correspond to the case that the lengths of the sides of the unit
cell in the cubic and the trigonal phase are the same, see [JH00]. This is
very nearly the case in AuCd [HS00] and in TbDyFe2 , a material with giant
magnetostriction [JK93]. The rst type of the cubic to orthorhombic phase
has not been observed in any material. The second type of the cubic to orthorhombic transformation occurs in materials with an fcc parent phase and
is for example found in CuAlNi alloys [DR64].
The uniqueness results follow also from Luskins work already quoted in
Section 7.4. However, the proofs become more transparent and shorter in
our general framework developed in Chapter 4. The quasiconvex hull of the
three wells describing the cubic to tetragonal transformation has attracted a
lot of attentions and is still unknown. A formula for the polyconvex hull of
the three wells was recently announced in [Fr00]. The paper [DK02] contains
a construction that shows that the interior of the lamination convex hull
of three wells with determinant equal to one contains the identity matrix
as an interior point. This answers a question raised in [M98] and predicts
an ideally soft or liquid-like behavior of shape memory materials. Proposition 5.1.5 shows that it is also dicult to construct explicitly laminates
that are supported on all three wells. In fact, any laminate supported on
three wells must contain at least four Dirac masses. This was shown in [Fr00]
based on a dierent proof for the polyconvex hull. However, it is worthwhile
to note that there exists a short proof for the quasiconvex hull which relies on
Theorem 9 in [Sv91] (see also [Sv92b]). This theorem states that any gradient
Young measure supported on three incompatible matrices must be a single
Dirac mass. The proof of the proposition given in the text then uses crucially
the identity (5.6) which is identity (5.7) in [JK89]. The uniqueness part in
the cubic to orthorhombic transformation in Theorem 5.3.1 follows also from
Theorem 3.1 in [BLL99]. For completeness of our presentation we nally discuss the three tetragonal to monoclinic transformations described in [BL00].
Theorem 5.4.1 summarizes the results in [BL00] and presents a short proof
in the general framework described in this text. The new ingredient is the
precise characterization of the sets Mpc (K; F ) in the exceptional cases for
the lattice parameters.
182
7.6 Algorithmic Aspects

The algorithm for the computation of rank-one convex (or more generally,
D-convex) envelopes of functions was dened in [D99] and its convergence
(in dependence on the discretization parameter h in the space of all matrices) was established in [DW00]. Theorems 6.1.4 and 6.1.5 and Lemma 6.1.7
can be found there. The algorithm is based on ideas in [KS86] how to nd
the rank-one convex envelope of a function by performing convexications
along rank-one lines. The book [Dc89] contains a lot of information about
nonconvex variational problems, their relaxation, and integral representations for the relaxed energy. Explicit relaxation formulae can be found in
[Kh67, Pi91, K91, LDR95, DSD01]. The rigidity results in [DM95] show that
every function u : R3 with Du SO(3)U1 SO(3)U2 is (locally) a
function of one variable if the set E = {x : Du(x) SO(3)U1 } is
a set of nite perimeter. Here U1 and U2 are two of the thee wells in the
cubic to tetragonal phase transformation. This result was extended to the
three-well problem in [Kir98]. Various characterizations of semiconvex hulls
of sets in terms of convexications of distance functions can be found in [Z98].
The Kohn-Strang example was rst discussed in [KS86] and generalizations
and questions concerning the existence of minimizers have been analyzed in
[DcM95, AF98]. The construction of the rank-one convex envelope of the
Kohn-Strang function in Section 6.1 follows Section 5D in [KS86]. An algorithm for the computation of separately convex envelopes of functions and
hulls was presented in [MP98]. See also [AP00] for related ideas. Approximation properties for laminates by laminates supported on nitely many points
have been obtained in [Pe93, MS99a]. The four point conguration with nontrivial rank-one convex hull was discovered independently by several authors
in dierent contexts (see e.g. [AH86, BFJK, CT93, Sch74, Ta93]).
A. Convexity Conditions and Rank-one

Connections
A.1 Convexity Conditions

Throughout the text we have seen the fundamental importance of various
convexity conditions, in particular of rank-one convexity, quasiconvexity and
polyconvexity. In this appendix, we give a precise denition of these notions
of convexity and we summarize some important results.
Let M (F ) denote the vector of all minors (i.e. all subdeterminants) of an
m n matrix F Mmn and d(m, n) its length. For m = n = 2 and n =
m = 3 we have M (F ) = det F and M (F ) = (F, cof F, det F ), respectively.
Denition A.1.1. A function f : Mmn R = (, ] is said to be
i) convex if
f (A + (1 )B) f (A) + (1 )f (B) A, B Mmn , [ 0, 1 ];
ii) polyconvex if there exists a convex function g : Rd(m,n) R such that
f (F ) = g(M (F ));
iii) quasiconvex if for every open and bounded set Rn with || = 0 one
has

f (F + D)dx
f (F )dx F Mmn , W01, (; Rm ),
whenever the integral on the left hand side exists;

iv) rank-one convex if for all A, B Mmn with rank(A B) = 1 and for
all [ 0, 1 ]
f (A + (1 )B) f (A) + (1 )f (B).
While all these notions coincide for m = 1 or n = 1, they are fundamentally
dierent for m 2 and n 2.
Proposition A.1.2. Assume that m, n 2 and that f : Mmn R is real
valued. Then we have the following implications:
G. Dolzmann: LNM 1803, pp. 183192, 2003.

184
A. Convexity Conditions and Rank-one Connections
f convex
f convex
f polyconvex
f polyconvex
and
f quasiconvex
f quasiconvex
f rank one convex
f rank one convex
(m 3)
If f : Mmn R, then
f convex
f polyconvex
f rank-one convex
(A.1)
and
f polyconvex
f quasiconvex.
A detailed discussion of these implications and various counterexamples can

be found in [Dc89], Theorem 4.1 and [M99b], Lemma 4.3. See also [BM84] for
a generalization of the concept of quasiconvexity, called W 1,p -quasiconvexity.
The question of whether rank-one convexity implies quasiconvexity was open
ak [Sv92a] for m 3. The case
for a long time and nally answered by Sver
m = n = 2 is still completely open. For smooth functions, rank-one convexity
is equivalent to the Legendre-Hadamard condition that the second gradient
be nonnegative on rank-one matrices, i.e. (D2 f ; R, R) 0 for all R Mmn
aks example, Kristensen recently showed
with rank(R) = 1. Based on Sver
that there does not exist an analogous local criterion for quasiconvexity and
polyconvexity, see [Kr99, Kr00]. Milton [Mi02] modied the construction of
the counterexample and found a set of seven matrices in M32 for which the
quasiconvex hull is bigger than the rank-one convex hull.
A special case are quadratic functions for which rank-one convexity and
quasiconvexity are known to be equivalent.
Theorem A.1.3 ([Dc89], Theorem 4.1.7). Let A be a symmetric matrix
in Mmnmn . Let
f (X) = AX, X,
where X Mmn . Then
i) f is rank-one convex if and only if f is quasiconvex;
ii) if m = 2 or n = 2 then
f polyconvex f quasiconvex f rank-one convex;
185
iii) if m, n 3 then in general

f rank-one convex f polyconvex.
An example for part iii) is due to Terpstra [Tr38] and Serre [Sr83]. A short
self-contained proof can be found in [B85].
Rank-one convexity can be viewed as a special case of the more general
concept of D-convexity. Let D Mmn be any set. Then a function f :
Mmn R is said to be D-convex if the functions
g : R R,
g(t) = f (F + tD)
are convex for all F Mmn and all D D (see [MP98] for a detailed
discussion).
Based on these notions of convexity, we dene semiconvex hulls of compact
sets K Mmn in the following way:
K pc = {F Mmn : f (F ) sup f (X) f : Mmn R polyconvex }.
XK
The quasiconvex hull K qc , the rank-one convex hull K rc , the convex hull
conv(K), and the D-convex hull K D are dened analogously (in [MP98] the
set K D is called the functionally D-convex hull). In particular for the description of constructions it is convenient to introduce a further hull which has
no immediate characterization in terms of convex functions. The lamination
convex hull K lc is dened in the following way (see [MS96]): Let K (0) = K
and dene

K (i+1) = A + (1 )B : A, B K (i) , rank(A B) = 1, (0, 1) K (i) .
Then
K lc =
K (i) .
i=0
The relations in the rst part of Proposition A.1.2 between the dierent
notions of convexity imply immediately the following chain of inclusions:
K lc K rc K qc K pc conv(K).
Similarly we dene semiconvex envelopes of functions f : Mmn R. The
largest polyconvex function less than or equal to f is called the polyconvex
envelope of f and denoted by f pc . The quasiconvex, the rank-one convex, the
convex, and the D-convex envelopes are dened correspondingly and denoted
by f qc , f rc , f , f D , respectively. In view of Proposition A.1.2 we have for
functions f : Mmn R the following inequalities,
f rc f qc f pc .
186
For extended valued functions f : Mmn R we dene the envelopes analogously, but in this case the inequalities between the envelopes might not
hold.
The condition HN has been introduced in [Dc85].
F1 s
s F6
G1 s
F2
s F3
s
F4 s
s
s F5
Fig. A.1. Graphic interpretation of condition HN : the solid lines in the gure
correspond to the rank-one connections, the dots at the end of the lines represent
the matrices Fi , and the dots in the middle of the lines indicate the matrices Gi
which are recursively constructed.
Denition A.1.4. We say that N pairs (i , Fi ), i = 1, . . . , N , with i

(0, 1) and Fi Mmn satisfy condition HN if the following holds:
if N = 2 then rank(F1 F2 ) = 1 and 1 + 2 = 1;
if N > 2, then (relabeling the matrices if necessary) rank(F1 F2 ) = 1 and
the pairs (i , Gi )i=1,...,N 1 dened by
1 = 1 + 2 ,
i = i+1 ,
1
2
F1 +
F2 ,
1 + 2
1 + 2
Gi = Fi+1 , i = 2, . . . , N 1,
G1 =
satisfy condition HN 1 .
It is an immediate consequence of the foregoing denitions that F K (i)
implies the existence of at most 2i pairs (i , Fi )i=1,...,k with Fi K satisfying
condition Hk .
We nally describe classes of (homogeneous) probability measures based
on these notions of convexity. We denote by M(K) the set of all Radon
measures supported on K equipped with the total variation norm M and
by P(K) the subset of all nonnegative Radon measures with mass one. We
write

, f =
f (A)d(A) P(K), f : K R,
K
187
whenever the right hand side exists.

It is convenient to introduce the following notation: if U is invertible and
M(K), then we denote by U the measure given by
(U )(E) = (EU 1 ) where EU 1 = {F : F U E}.
We refer to the Dirac measure placed at a point F Mmn frequently as the
Dirac mass at F .
As pointed out in the introduction, the link between the nonconvex minimization problem and probability measures is the Young measure generated
by (subsequences of) minimizing sequences.
Theorem A.1.5 ([B89], Section 2). Let be Lebesgue measurable, let
K Rm be closed, and let z j : Rm , j = 1, 2, . . ., be a sequence of
Lebesgue measurable functions satisfying z j K in measure as j , i.e.,
given any open neighborhood U of K in Rm , then
'
'
lim '{x : z j (x) U }' = 0.
j
Then there exists a subsequence z of z j and a family {x }x of positive

measures on R-m , depending measurably on x, such that
i) x M = Rm dx 1 for a.e. x ,
ii) supp x K for a.e. x , and
iii) f (z ) x , f in L for each continuous function f : Rm R satisfying f () 0 as || .

Suppose further that {z } satises the boundedness condition
'
'
lim sup '{x BR : |z (x)| k}' = 0
k
for every R > 0, where BR = B(0, R). Then x M = 1 for a.e. x (i.e.,
x is a probability measure), and given any measurable subset A of ,
f (z ) x , f
in L1 (A)
for any continuous function f : Rm R such that {f (z )} is sequentially

weakly relatively compact in L1 (A).
In analogy to Jensens inequality for convex functions we dene the class
of polyconvex measures Mpc (K) P(K) by

Mpc (K) = P(K) : f (, id) , f f : K R polyconvex ,
and we write Mpc (K; F ) for all measures Mpc (K) with center of mass
equal to F , i.e., , id = F . We dene the class of quasiconvex measures
Mqc (K) (usually called homogeneous gradient Young measures) and rankone convex measures Mrc (K) (also called laminates) analogously ([KP91,
Pe93, Sv95]). If the pairs (i , Fi ), i = 1, . . . , N satisfy condition HN , then
188
N

i Fi
i=1
denes a laminate of nite order. If N = 2, then is called simple laminate.

In case that N = 3 or N = 4 and (up to relabeling the matrices)
rank(F1 F2 ) = 1,
rank(F3 F4 ) = 1,
rank(G1 G2 ) = 1
where
G1 =
1
2
F1 +
F2 ,
1 + 2
1 + 2
G2 =
3
4
F3 +
F4 ,
3 + 4
3 + 4
the measure is called a second order laminate.

A fundamental result by Kinderlehrer&Pedregal states that quasiconvex
measures are exactly the Young measures that can be generated by sequences
of gradients.
Theorem A.1.6 ([KP91]). A (weakly* measurable) map : P(Mmn )
is a gradient Young measure if and only if x 0 a.e. and there exist a compact set K Mmn and u W 1, (; Rm ) such that the following three
conditions hold:
i) supp x K for a.e. x,
ii) x , id = Du(x)
for a.e. x, and
iii) x , f f x , id for a.e. x and for all quasiconvex functions f :

Mmn R.
We say that the Young measure {x }x is homogeneous if there exists a
P(Mmn ) such that x = for a.e. x. We identify in this case the family
of measures {x }x with the single measure .
The next theorem describes the close connection between these convex
hulls and the corresponding sets of Radon measures.
Theorem A.1.7. Assume that K Mmn is a compact set. Then

K qc = , id : Mqc (K)
and the analogous statement holds for K rc and K pc .
Since SO(2) is contained in an ane subspace of M22 it is not dicult
to see that
a b

, a2 + b2 1 .
conv(SO(2)) =
(A.2)
b a
The convex hull of SO(3) is given by
A.2 Existence of Rank-one Connections

conv(SO(3)) = F = QU : Q SO(3), U = U T ,
3

i i 1 for |i | = 1 and 1 2 3 = 1 ,
189
(A.3)
(A.4)
i=1
where 1 , 2 , 3 are the eigenvalues of the symmetric matrix U (see [Ja86]).

Moreover,
tr(F ) 2eT F e 1
e S2 , F conv SO(3).
We will frequently use the important fact that SO(n) does not support any
(nontrivial) gradient Young measure.
Theorem A.1.8 ([Ki88]). Suppose that Du SO(n) a.e. in . Then Du
is constant and u(x) = Qx + b with Q SO(n) and b Rn . If uj
W 1, (; Rn ) satises
dist(Duj , SO(n)) 0 in measure,
then Duj Q in measure where Q SO(n) is a constant.
The rank-one convex hull has a certain locality property which we will
use in order to show that certain rank-one convex hulls are trivial.
Proposition A.1.9 ([Pe93, MP98, Mt00, Kir00]). Assume that K is
compact and that K rc consists of two compact components C1 and C2 with
C1 C2 = . Then
K rc = (K C1 )rc (K C2 )rc .
(A.5)

The discussion in the introduction shows that formation of microstructure
is closely connected to the existence of rank-one connections between energy
wells. Here we say that two wells SO(m)U1 and SO(m)U2 with U1 U2 Mmn
symmetric and positive denite are rank-one connected or compatible if there
exists a rotation R SO(m) such that
RU2 U1 = a b
with a Rm , b Rn . Otherwise the wells are called incompatible.
In this section, we discuss the solvability of the so-called twinning equation
rank(QA B) = 1 in detail for (2 2) and (3 3) matrices. We begin with
the case m = n = 3. In this case the equation RU2 U1 = a m is after
postmultiplication by U11 equivalent to
190
RU2 U11 = I + a n,
n = U11 m.
By the polar decomposition theorem this equation holds if and only if the
matrix C = U11 U22 U11 has the representation
C = (I + n a)(I + a n),
1 + a, n > 0.
(A.6)
Therefore the existence of rank-one connections is equivalent to nding solutions a, n R3 of (A.6) with C given.
Proposition A.2.1 ([BJ87, Kh83]). Necessary and sucient conditions
for a symmetric (3 3)-matrix C = I with eigenvalues 1 2 3 to be
expressible in the form
C = (I + n a)(I + a n)
with 1 + a, n > 0 and a = 0, n = 0 are that 1 > 0 (i.e., C is positive
denite) and 2 = 1. The solutions are given by
.
.
3 (1 1 )
1 (3 1)
e1 +
e3 ,
b =
3 1
3 1

1 3 1

m=
1 1 e1 + 3 1e3 ,
3 1
where = 0 is a constant and e1 , e3 are normalized eigenvectors of C corresponding to 1 , 3 , respectively, and where can take the values 1.
In the n-dimensional situation we obtain after suitable transformations
the condition that at least n 2 eigenvalues have to be equal to one, while
the smaller eigenvalue is less than or equal to one, and the larger is larger
than or equal to one.
Proposition A.2.2 ([DM95], Proposition 5.2). Assume that A = I and
B = diag(1 , . . . , n ) with 0 < 1 n . Then the two wells SO(n)
and SO(n)B are rank-one connected if and only if 2 = = n1 = 1.
Moreover, the vectors a and b in the representations Q B = a b lie in
the plane spanned by e1 and en .
In twinning calculations one often encounters the situation that U1 and U2
are related by some rotation R SO(3), through U2 = RT U1 R and therefore
C2 = U2T U2 = RT U1T U1 R = RT C1 R = C1 .
(A.7)
In this situation, the solvability of the twinning equation

QU1 = U2 (I + a n),
is described by the following proposition.
Q SO(3),
(A.8)
191
Proposition A.2.3 ([Er91]). Assume that the symmetric and positive definite (3 3)-matrices C1 and C2 are related by a rotation in the sense of
identity (A.7). Then the following assertions hold:
1. If there exists a solution of the twinning equation (A.8), then at least one
of the following conditions must hold:
a) R represents a 180 rotation.
b) e is an eigenvector of C1 .
c) e is perpendicular to an eigenvector of C1 , which can be taken as (C1 e)
e, if (b) does not hold.
2. If any of the above conditions holds, and C2 = C1 , then the twinning
equation (A.8) can be solved.
3. If the twinning equation (A.8) can be solved, and (a) does not hold, then
with R
a 180 rotation.
(A.7) must hold with R replaced by R
Condition (1a) is particularly important, since in this case the two solutions can easily be obtained, as discussed in [Er81, Er85, Gr83]. The following
version can be found in [Bh92].
Proposition A.2.4. Let R = I+2ee be a 180 rotation about e S2 and
assume that U1 and U2 M33 are symmetric matrices with U1T U1 = U2T U2
and U2 = RT U1 R. Then there are two solutions of the equation
QU1 U2 = a n,
Q SO(3), a, n R3 ,
(A.9)
and they are given by

n1 = e,
U2T e

a1 = 2
U2 e
T 2
|U2 e|
and
n2 =
U T U2 e
2
e 2 2 ,

|U2 e|
a2 = U2 e,
where is chosen in such a way that |n2 | = 1. The corresponding rotations

can be found by solving the equation (A.9) for Q.
The two-dimensional situation allows some simplications.
Proposition A.2.5. Assume that C1 , C2 M22 are positive semidenite,
C1 = F1T F1 , C2 = F2T F2 . Let e S1 . Then the following four statements are
equivalent:
i) there exist Q SO(2) and a R2 such that QF1 F2 = a e ;
ii) we have |F1 e|2 = |F2 e|2 ;
iii) there exists a v R2 such that C1 = C2 + v e + e v;
iv) det(C1 C2 ) 0.
Moreover, the vector a in statement i) and the vector v in statement iii) are
related by v = F2T a + 12 |a|2 e . Finally, if det F1 = det F2 , then a = F2 e
with R and the two wells are always rank-one connected.
192
Remark A.2.6. The observation that two wells with equal determinant in two
dimensions are always rank-one connected is at the heart of the characterization of the semiconvex hulls in two dimensions. While this can be proven by a
direct calculation, it is worthwhile noting that this fact is also a consequence
the following decomposition of matrices (see, e.g., [CK88], Proposition 3.4).
Proposition A.2.7. Let A M33 with det A > 0. Then there is a rotation
Q SO(3) and vectors a1 , n1 , a2 , n2 R3 with ai , ni = 0, i = 1, 2, such
that
A = (det A)1/3 Q(I + a2 n2 )(I + a1 n1 ).
If A M22 , then the same result holds with a2 = n2 = 0 and the exponent
1
1
3 replaced by 2 .
B. Elements of Crystallography
A set of points L is called a Bravais lattice if and only if there exist three
linearly independent vectors g 1 , g 2 , g 3 R3 such that

L = L(g i ) = x =
3

n i g i , ni Z .
i=1
The vectors g i are not uniquely determined and a theorem in crystallography

(see [Er77]) states that two sets of linearly independent vectors g 1 , g 2 , g 3
1 , g
2 , g
3 generate the same lattice if and only if there exists a matrix
and g
M Z33 with det M = 1 such that
i =
g
3

Mij g j .
i=1
It turns out that there are fourteen dierent three-dimensional or Bravais

lattices1 which are conventionally grouped into seven crystal systems sharing some symmetry, see Table B.1: triclinic2,3 , monoclinic4 , orthorhombic5,6 ,
tetragonal7,8 , hexagonal9 , cubic10 .
The symmetry is described in terms of crystallographic point groups, i.e.,
groups of symmetry operations which map the lattice back to itself and leave
at least one point in the lattice xed. The symbols for the dierent groups
in Hermann-Mauguin notation can be found in Table B.1. Here n denotes a
rotation by 2/n, m a reection about a mirror plane, 1 an inversion through
1
2
3
4
5
6
7
8
9
10
Auguste Bravais (1811-63), professor at the Ecole

Polytechnique, classied the
fourteen space lattices in 1848. The German physicist M. L. Frankenheim found
erroneously 15 types in 1835.
gr. - three, thrice (in compound words)
gr. come to lean, make to slope or slant
gr.
oo alone, solitary; only
gr. o
o straight
gr. o
oo rhombus, i.e., a four sided gure with all the sides, but only the
opposite angles, equal
gr. four (in compound words)
gr. corner, angle
gr.
(indecl.) six
gr.
o, lat. cubus marked on all six sides; anything of cubic shape
G. Dolzmann: LNM 1803, pp. 193196, 2003.

194
a point and n
an improper rotation, i.e., a rotation by 2/n followed by an
inversion through a point on the axis. The position of the operation within
the symbol refers to a dierent direction in the lattice related to the operation
(axis of rotation, normal to the plane of reection), see [S69], Chapter 3.5 for
further details. For a given lattice L(g i ), we dene the point group P(g i ) to
be the maximal group under which the lattice is invariant. For example, the
point group of the cube is the group of 48 orthogonal transformations that
map the cube to the cube11 .
There are several ways to choose unit cells for the fourteen dierent lattices, and the standard cells (highest degree of symmetry) are shown in Figure B.1. The conventional symbols are P for primitive (the unit cell contains
just one point), C for C-centered (a lattice point in the center of the C side
of the unit cell), F or fc for face centered (lattice points in the center of each
face) and I or bc for body centered (a lattice point in the center of the unit
cell; German: innenzentriert) (see e.g. [S69] for a discussion of the notation).
Note that the face centered and the body centered tetragonal systems are
equivalent.
The dierent crystal systems are in particular of importance in diusionless solid-solid phase transformations which are characterized by a break of
symmetry of the underlying Bravais lattice. We call martensite12 the phase
that forms as the result of such a transformation. The high temperature phase
is frequently called austenite13 . A given system can undergo several transformations at several critical temperatures. For example, immediately after
solidication iron forms a body centered cubic (bcc) structure, called -ferrite.
Upon further cooling, iron transforms to a face centered cubic structure (fcc),
called or austenite. Finally, iron transforms back to the bcc structure at
lower temperatures; this structure is called , or ferrite. In higher carbon
steels, the fcc austenite transforms to a body centered tetragonal martensite,
thus showing the characteristic break of symmetry during the transformation.
The following examples describe typical phase transformations and the
corresponding transformation matrices where we always assume that the reference conguration is given in the undistorted austenitic phase. The number
N of variants in the low temperature phase is given by
N=
order of high temperature point group

order of low temperature point group
see e.g. [VTA74, BJ92] for a discussion of the mathematical concepts. A detailed analysis of the point groups associated with the dierent variants is
contained in [PZ00]. A nice summary of the matrices describing the energy
11
12
13
In the context of continuum models in elasticity the point group is frequently

dened as a subgroup of SO(3). This is also our convention in this text.
Adolf Martens (1850-1914), German metallurgist, professor at the TH Berlin.
Sir William Chandler Roberts-Austen (1843-1902), FRS, 19th-century pioneer in
metallurgy and alloy phase diagrams. He presented in 1897 his rst temperaturecomposition diagram for the Fe-C system [KPt98].
triclinic P
195
monoclinic P monoclinic C
orthorhombic P orthorhombic C orthorhombic I orthorhombic F
tetragonal P
tetragonal I
hexagonal P
cubic P
trigonal P
cubic I
cubic F
Fig. B.1. The fourteen Bravais lattices.
wells (under the assumption that the austenitic phase is the reference conguration) and the relation of their elements to the transformation strains can
be found in [H97, JH00]. We include some examples and collect the relevant
information about the twinning systems. We dene twins to be continuous
and piecewise homogeneous deformations with constant deformation gradients F1 and F2 in layers separated by a hyperplane. Necessarily the two
deformation gradients must be rank-one connected and the normal n on the
hyperplane is determined from F1 F2 = a n. Since we are interested in
zero-energy deformations, we focus on rank-one connections between energy
wells SO(3)Ui and SO(3)Uj , i = j. By Propositions A.2.3 and A.2.4 the situation Ui = RT Uj R with R a 180 degree rotation is particularly important.
196
Table B.1. Crystallographic point groups.
crystal system
char, symmetry
restrictions
symbol
order
Triclinic
onefold symmetry
none
Monoclinic
one diad
= = 90
Orthorhombic
three mutually
perpendicular diads
= = = 90
Tetragonal
one tetrad
a=b
= = = 90
Trigonal
one triad
Hexagonal
one hexad
a = b = c and
= = or
a=b
= = 90
= 120
a=b
= = 90
= 120
Cubic
four triads
1
2
m
2/m
222
mm2
mmm
4
4
4m
422
4mm
42m
4/mmm
3
3
32
3m
3m
6
6
6/m
622
6mm
6m2
6/mmm
23
m3
432
43m
m3m
1
2
2
2
4
4
4
8
4
4
8
8
8
8
16
3
6
6
6
12
6
6
12
12
12
12
24
12
24
24
24
48
a=b=c
= = = 90
We call the two solutions given in Proposition A.2.4 a type-I and a type-II
twin, respectively. A twin is called a compound twin if it is at the same time a
type-I and a type-II twin, i.e., if there exist two distinct 180 degree rotations
R1 , R2 with Ui = R1T Uj R1 and Ui = R2T Uj R2 .
C. Notation
We write,Rn for the n-dimensional real vector space with scalar product
u, v = i ui vi , Rn+ for the positive octant {x : xi 0 for i = 1, . . . , n} in
Rn , and Sn for the unit sphere in Rn+1 . We use Mmn for the space of all
T
nm
(real) (m n)-matrices with the norm |F |2 = tr(F T F ), where
,F M
is the transposed matrix, and the scalar product F : G =
i,j Fij Gij . If
e R2 , then e = Je, where J is the counter-clockwise rotation by 2 .
If m = n, then the cofactor matrix cof F Mnn of F is dened to be
the matrix of all (n 1) (n 1) minors of F which satises
cof F = (det F )F T i.e. (cof F )F T = (det F )I F Mnn ,
(C.1)
where for a Rm and b Rn the (mn)-matrix ab is given by (ab)i,j =

ai bj , and I = diag(1, . . . , 1) is the identity matrix in Mnn . We will frequently
use the following formula for the inversion of a matrix,
(I + a b)1 = I a b a, b Rn , a, b = 0.
In particular, suppose that F G = a b and let
H = F + (1 )G = G + a b.
If G and H are invertible and if G1 a, b = 0, then
H1 = G1 G1 a GT b,
cof H = cof G cof Gb G1 a.
(C.2)
The assumption G1 a, b = 0 is for example satised if det F = det G since

the formula
det(I + c d) = 1 + c, d,
c, d Rn
implies that
det F = det(G + a b) = (det G)(1 + G1 a, b).
We have for F M22 the expansion
G. Dolzmann: LNM 1803, pp. 197200, 2003.

198
C. Notation
det(A + B) = det A + A : (cof B) + det B,
(C.3)
while for F M33

cof(F I) = cof F (tr F )I + F T + I,
det(F I) = det F tr cof F + tr F det I.
(C.4)
(C.5)
We denote the group of all orthogonal matrices by O(n), the group of all
proper rotations by SO(n) or O+ (n) and we dene O (n) = O(n) \ O+ (n)
to be the set of all orthogonal matrices with determinant minus one. Finally
O(n, m) = {Q Mmn : QT Q = I}.
We will frequently use the following polar decomposition result (see e.g.
Theorem 9 in Chapter IX, 14 in [Gnt58]).
Proposition C.1. Any matrix F Mnn can be represented as F =
Q1 S1 = S2 Q2 with Q1 , Q2 O(n) and uniquely determined symmetric and
positive semidenite matrices S1 and S2 . Moreover, if det F = 0, then also
Q1 and Q2 are uniquely determined.
Denition C.2. Assume that F Mnn and that the polar decomposition
of F is given by F = QS with Q O(n) and S symmetric and positive
denite. We call the (nonnegative) eigenvalues 0 1 . . . n of S the
singular values of F . Moreover, if det F < 0, then we call the values i = i ,
i = 2, . . . , n, and 1 = 1 the signed singular values of F . By denition,
|1 | 2 . . . n , 1 < 0.
Remark C.3. Assume det F < 0 and that F = QS is the polar decomposition of F . If + = diag(1 , . . . , n ) denotes the diagonal matrix of the
singular values of F then S = QT1 + Q1 with Q1 SO(n). Similarly, if
= diag(1 , . . . , n ) is the diagonal matrix of the signed singular values
of F , then there exists a proper orthogonal matrix Q2 SO(n) such that
F = Q2 Q1 . The convention to put the minus sign in front of the smallest
singular value of F is convenient in the statements of Proposition 2.7.8 and
Proposition 2.7.9.
C. Notation
Vectors and matrices

Rn
u, v
Sn
Mmn
tr F
F :G
|F |
i (F )
i (F )
I
FT
cof F
M (F )
SO(n)
O(n)
O (n)
O(m, n)
Rij
, Ri
n-dimensional Euclidean space

inner product in Rn
unit sphere in Rn+1
space of all real m n matrices
trace of an n n matrix F
inner product in Mnn dened by F : G = tr(F T G)
Euclidean norm in Rn and Mmn , |F |2 = F : F
singular values of F , 1 n
signed singular values of F , |1 | 2 . . . n
identity matrix in Mnn
transposed matrix
cofactor matrix, (cof F )F T = (det F )I
vector of all minors of a matrix F
orthogonal matrices in Mnn with det F = 1
orthogonal matrices in Mnn with det F = 1
orthogonal matrices with determinant 1
isometries from Rm into Rn
180 rotations with axes ei ej and ei
ab
J
v
%
e1 , %
e2
e1 , e2 , e3
rank-one matrix (a b)ij = ai bj

counterclockwise rotation by /2 in the plane
Jv, v rotated counterclockwise by /2 in the plane
standard basis in R2
standard basis in R3
Hulls of matrices
K (i)
K lc
K rc
K qc
K pc
conv K
ith lamination hull

lamination convex hull of K
rank-one convex hull of K
quasiconvex hull of K
polyconvex hull of K
convex hull of K
199
200
C. Notation
Sets of measures
supp
, f
U
P(K)
Mrc (K)
Mrc (K; F )
Mqc (K)
Mqc (K; F )
Mpc (K)
Mpc (K; F )
support of the measure

integral of f with respect to
measure dened by (U )(E) = (EU 1 ) for U invertible and EU 1 = {F : F U E}
set of all probability measures supported on K
laminates supported on K
laminates supported on K with center of mass F
gradient Young measures supported on K
gradient Young measures supported on K with center
of mass F
polyconvex measures supported on K
polyconvex measures supported on K with center of
mass F
Functions and function spaces

Lp
W k,p
C k,
p
|x|1
|x|
(t)+
f rc
f qc
f pc
Lebesgue space
Sobolev space
H
older Spaces
space of p-summable sequences
|x1 | + . . . + |xn | for x Rn
maxi=1,...,n |xi | for x Rn
max{t, 0}
rank-one convex envelope of f
quasiconvex envelope of f
polyconvex envelope of f
embedding M22 M32
projection from Mmn onto a multi-well set K
projection from Mmn onto two wells
References
[AcF84] E. Acerbi, N. Fusco, Semicontinuity problems in the calculus of variations, Arch. Rational Mech. Anal. 86 (1984), 125-145
[AcF88] E. Acerbi, N. Fusco, An approximation lemma for W 1,p functions, in:
Material instabilities in continuum mechanics and related mathematical problems (J. M. Ball, ed.), Oxford University Press, 1988, 217-242
[AF98] G. Allaire, G. Francfort, Existence of minimizers for non-quasiconvex
functionals arising in optimal design, Ann. Inst. H. Poincare Anal. Non Lineaire
15 (1998), 301-339
[ACF99] D. R. Anderson, D. E. Carlson, E. Fried, A continuum-mechanical
theory for nematic elastomers, J. Elasticity 56 (1999), 33-58
[AP00] E. Aranda, P. Pedregal, On the computation of the rank-one convex
hull of a function, SIAM J. Sci. Comput. 22 (2000), 1772-1790
[A98] K. Astala, Analytic aspects of quasiconformality, Proceedings of the International Congress of Mathematicians, Vol. II (Berlin, 1998), Doc. Math. 1998,
Extra Vol. II, 617-626
[AH86] R. J. Aumann, S. Hart, Bi-convexity and bi-martingales, Israel J. Math.
54 (1986), 159-180
[B77] J. M. Ball, Convexity conditions and existence theorems in nonlinear elasticity, Arch. Rational Mech. Anal. 63 (1976/77), 337403
[B84] J. M. Ball, Dierentiability properties of symmetric and isotropic functions,
Duke Math. J. 51 (1984), 699-728
[B85] J. M. Ball, On the paper Basic calculus of variations, Pacic J. Math.
116 (1985), 7-10
[B89] J. M. Ball, A version of the fundamental theorem for Young measures, in:
PDEs and continuum models of phase transitions, Lecture Notes in Phys. 344,
Springer, 1989, 207-215
[B98] J. M. Ball, The calculus of variations and materials science. Current and
future challenges in the applications of mathematics (Providence, RI, 1997),
Quart. Appl. Math. 56 (1998), 719-740
[BC97] J. M. Ball, C. Carstensen, Nonclassical austenite-martensite interfaces,
J. Phys. IV France 7 (1997), 35-40
[BC99] J. M. Ball, C. Carstensen, Compatibility conditions for microstructures
and the austenite martensite transition, Mater. Sc. Engin. A 273-275 (1999),
231-236
[BJ87] J. M. Ball, R. D. James, Fine phase mixtures as minimizers of energy,
Arch. Rational Mech. Anal. 100 (1987), 13-52
[BJ92] J. M. Ball, R. D. James, Proposed experimental tests of a theory of ne
microstructure and the two-well problem, Phil. Trans. Roy. Soc. London A 338
(1992), 389-450
[BK87] J. M. Ball, G. Knowles, A numerical method for detecting singular minimizers, Numer. Math. 51 (1987), 181-197
202
References
[BM85] J. M. Ball, V. J. Mizel, One-dimensional variational problems whose

minimizers do not satisfy the Euler-Lagrange equation, Arch. Rational Mech.
Anal. 90 (1985), 325-388
[BM84] J. M. Ball, F. Murat, W 1,p -quasiconvexity and variational problems for
multiple integrals, J. Funct. Anal. 58 (1984), 225-253
[BBL01] P. Belik, T. Brule, M. Luskin, On the numerical modeling of deformations of pressurized martensitic thin lms, M2 AN Math. Model. Numer. Anal.
35 (2001), 525-548
[BL00] P. Belik, M. Luskin, Stability of microstructure for tetragonal to monoclinic martensitic transformations, M2 AN Math. Model. Numer. Anal. 34
(2000), 663-685
[BCDM00] H. Ben Belgacem, S. Conti, A. DeSimone, S. Mller, Rigorous
bounds for the F
oppl-von K
arm
an theory of isotropically compressed plates, J.
Nonlinear Sci. 10 (2000), 661-683
[Bh92] K. Bhattacharya, Self-Accommodation in Martensite, Arch. Rational
Mech. Anal. 120 (1992), 201-244
[BhD01] K. Bhattacharya, G. Dolzmann, Relaxation of some multi-well problems, Proc. Roy. Soc. Edinburgh Sect. A 131 (2001), 279-320
[BFJK] K. Bhattacharya, N. B. Firoozye, R. D. James, R. V. Kohn, Restrictions on microstructure, Proc. Roy. Soc. Edinburgh Sect. A 124 (1994),
843-878
[BJ99] K. Bhattacharya, R. D. James, A theory of thin lms of martensitic
materials with applications to microactuators, J. Mech. Phys. Solids 47 (1999),
531-576
[BhK96] K. Bhattacharya, R. V. Kohn, Symmetry, texture and the recoverable
strain of shape memory materials, Acta Materialia 25 (1996), 529-542
[BhK97] K. Bhattacharya, R. V. Kohn, Elastic energy minimization and the
recoverable strains of polycrystalline shape-memory materials, Arch. Rational
Mech. Anal. 139 (1997), 99-180
[BLL99] K. Bhattacharya, B. Li, M. Luskin, The simply laminated microstructure in martensitic crystals that undergo a cubic-to-orthorhombic phase transformation, Arch. Ration. Mech. Anal. 149 (1999), 123-154
[BTW93] P. Bladon, E. M. Terentjev, M. Warner., Transitions and instabilities in liquid-crystal elastomers, Phys. Rev. E 47 (1993), R3838-R3840
[BMK54] J. S. Bowles, J. K. MacKenzie, The crystallography of martensitic
transformations I and II, Acta Met. 2 (1954), 129-137, 138-147
[CT00] P. Cardaliaguet, R. Tahraoui, Equivalence between rank-one convexity
and polyconvexity for isotropic sets of R22 (part I), Cahiers de Mathematiques
du CEREMADE 0028 (2000)
c
, Numerical solution of the scalar double-well
[CP97] C. Carstensen, P. Plecha
problem allowing microstructure, Math. Comp. 66 (1997), 997-1026
c
, Numerical analysis of compatible phase
[CP00] C. Carstensen, P. Plecha
transitions in elastic solids, SIAM J. Numer. Anal. 37 (2000), 2061-2081
ek, Numerical approximation of Young mea[CR00] C. Carstensen, T. Roubic
sures in non-convex variational problems, Numer. Math. 84 (2000), 395-415
[CT93] E. Casadio Tarabusi, An algebraic characterization of quasiconvex functions, Ricerche Mat. 42 (1993), 11-24
[C91] M. Chipot, Numerical analysis of oscillations in nonconvex problems, Numer. Math. 59 (1991), 747-767
[CC92] M. Chipot, C. Collins, Numerical approximations in variational problems with potential wells, SIAM J. Numer. Anal. 29 (1992), 1002-1019
[CCK95] M. Chipot, C. Collins, D. Kinderlehrer, Numerical analysis of oscillations in multiple well problems, Numer. Math. 70 (1995), 259-282
References
203
[CK88] M. Chipot, D. Kinderlehrer, Equilibrium congurations of crystals,

[CKO99] R. Choksi, R. V. Kohn, F. Otto, Domain branching in uniaxial ferromagnets: a scaling law for the minimum energy, Comm. Math. Phys. 201
(1999), 61-79
[Ch95] J. W. Christian, S. Mahajan, Deformation Twinning, Progress in Materials Science 39 (1995), 1-157
[Ci78] P. G. Ciarlet, The nite element method for elliptic problems, NorthHolland, Amsterdam, 1978
[CKL91] C. Collins, D. Kinderlehrer, M. Luskin, Numerical approximation
of the solution of a variational problem with a double well potential, SIAM J.
Numer. Anal. 28 (1991), 321-332
[CL89] C. Collins, M. Luskin, The computation of the austenitic-martensitic
phase transition, in: PDEs and continuum models of phase transitions (Nice,
1988), Lecture Notes in Phys. 344, Springer, Berlin-New York, 1989, 34-50
[CLR93] C. Collins, M. Luskin, J. Riordan, Computational results for a twodimensional model of crystalline microstructure, in: Microstructure and phase
transition, (D. Kinderlehrer, R. D. James, M. Luskin and J. Ericksen, eds.),
IMA Vol. Math. Appl. 54, Springer, New York, 1993, 51-56
[CDD01] S. Conti, A. DeSimone, G. Dolzmann, Soft elastic response of
stretched sheets of nematic elastomers: a numerical study, J. Mech. Phys. Solids,
50 (2002), 1431-1451
ller, F. Otto,
[CDDMO01] S. Conti, A. DeSimone, G. Dolzmann, S. Mu
Multiscale problems in materials, in: Trends in Nonlinear Analysis, (M. Kirkilionis S. Kr
omker, R. Rannacher, F. Tomi, eds.), Springer (to appear)
[Dc85] B. Dacorogna, Remarques sur les notions de polyconvexite, quasi-convexite et convexite de rang 1, J. Math. Pures Appl. 64 (1985), 403-438
[Dc89] B. Dacorogna, Direct methods in the calculus of variations, Springer, 1989
[DcM95] B. Dacorogna, P. Marcellini, Existence of minimizers for nonquasiconvex integrals, Arch. Rational Mech. Anal. 131 (1995), 359-399
[DcM99] B. Dacorogna, P. Marcellini, Implicit partial dierential equations,
Birkh
auser, 1999
[DcT98] B. Dacorogna, C. Tanteri, On the dierent convex hulls of sets involving singular values, Proc. Roy. Soc. Edinburgh Sect. A 128 (1998), 1261-1280.
[DcT00] B. Dacorogna, C. Tanteri, Implicit partial dierential equations and
the constraints of nonlinear elasticity, Journal des Mathematiques Pures et Appliquees 81 (2002), 311-341
[DS93] A. DeSimone, Energy minimizers for large ferromagnetic bodies, Arch. Rational Mech. Anal. 125 (1993), 99-143
[DS99] A. DeSimone, Energetics of ne domain structures, Ferroelectrics 222
(1999), 275-284
[DSD00] A. DeSimone, G. Dolzmann, Material instabilities in nematic elastomers, Physica D 136 (2000), 175-191
[DSD01] A. DeSimone, G. Dolzmann, Macroscopic response of nematic elastomers via relaxation of a class of SO(3)-invariant energies, Arch. Rational
Mech. Anal. 161 (2002), 181 - 204
[DSJ97] A. DeSimone, R. D. James, A theory of magnetostriction oriented towards applications, J. Appl. Phys 81 (1997), 5706-5708
[D99] G. Dolzmann, Numerical computation of rank-one convex envelopes, SIAM
J. Numer. Anal. 36 (1999), 1621-1635
[D01] G. Dolzmann, Variational methods for crystalline microstructure theory
and computation, Habilitationsschrift, Universit
at Leipzig, 2001
204
References
[DK02] G. Dolzmann, B. Kirchheim, Liquid-like response of shape-memory materials, MPI MIS Preprint 53, 2002
ller, V. Sver
k, The two-well
[DKMS00] G. Dolzmann, B. Kirchheim, S. Mu
a
problem in three dimensions, Calc. Var. Partial Dierential Equations 10 (2000),
21-40
ller, Microstructures with nite surface energy: the
[DM95] G. Dolzmann, S. Mu
two-well problem, Arch. Rational Mech. Anal. 132 (1995), 101-141
[DW00] G. Dolzmann, N. J. Walkington, Estimates for numerical approximations of rank-one convex envelopes, Numer. Math. 85 (2000), 647-663
[DR64] M. J. Duggin, W. A. Rachinger, The nature of the martensite transformation in a copper-nickel-aluminum alloy, Acta Metall. 12 (1964), 529-535
[EL01] Y. Efendiev, M Luskin, Stability of microstructures for some martensitic
transformations, Math. Comput. Modelling 34 (2001), 1289-1305
[Er77] J. L. Ericksen, Special topics in nonlinear elastostatics, in: Advances in
applied mechanics 17 (C. S. Yih, ed.), Academic Press, 1977, 189-244
[Er80] J. L. Ericksen, Some phase transitions in crystals, Arch. Rational Mech.
Anal. 73 (1980), 99-124
[Er81] J. L. Ericksen, Continuous martensitic transitions on thermoelastic solids,
J. Thermal Stresses 4 (1981), 107-119
[Er85] J. L. Ericksen, Some surface defects in unstressed thermoelastic solids,
[Er86] J. L. Ericksen, Constitutive theory for some constrained elastic crystals,
Int. J. Solids and Structures 22, (1986), 951-964
[Er89] J. L. Ericksen, Weak martensitic transformations in Bravais lattices, Arch.
Rational Mech. Anal. 107 (1989), 23-36
[Er91] J. L. Ericksen, On kinematic conditions of compatibility, J. Elasticity 26
(1991), 65-74
[FKR81] H. Finkelmann, H. J. Koch, G. Rehage, Liquid crystalline elastomers
A new type of liquid crystalline materials, Macromol. Chem. Rapid Commun.
2 (1981), 317-325
[Fo87] I. Fonseca, Variational methods for elastic crystals, Arch. Rational Mech.
Anal. 97 (1987), 189-220
[FH90] R. L. Fosdick, B. Hertog, Material symmetry and crystals, Arch. Rational Mech. Anal. 110 (1990), 43-72
[Fr90] D. A. French, On the convergence of nite-element approximations of a
relaxed variational problem, SIAM J. Numer. Anal. 27 (1990), 419-436
[Fr00] G. Friesecke, Quasiconvex hulls and recoverable strains in shape-memory
alloys, TMR meeting Phase transitions in crystalline solids, Max Planck Institute for Mathematics in the Sciences, Leipzig, 2000
[Gnt58] F. R. Gantmacher, Matrizenrechnung, Teil I, VEB Deutscher Verlag der
Wissenschaften, Berlin, 1958
[GP99] M. K. Gobbert, A. Prohl, A discontinuous nite element method for
solving a multiwell problem, SIAM J. Numer. Anal. 37 (1999), 246-268
[GP00] M. K. Gobbert, A. Prohl, A comparison of classical and new nite
element methods for the computation of laminate microstructure, Appl. Num.
Math. 36 (2001), 155-178
[GL89] L. Golubovic, T. C. Lubensky, Nonlinear elasticity of amorphous solids,
Phys. Rev. Lett. 63 (1989), 1082-1085
[Gd94] P. A. Gremaud, Numerical analysis of a nonconvex variational problem
related to solid-solid phase transitions, SIAM J. Numer. Anal. 31 (1994), 111127
[Gr83] M. E. Gurtin, Two-phase deformations of elastic solids, Arch. Rational
Mech. Anal. 84 (1984), 1-29
References
205
[Gt50] L. Guttman, Crystal structures and transformations in Indium-Thallium

solid solutions, Transactions AIME, J. Metals 188 (1950), 1472-1477
[H97] K. F. Hane, Microstructures in Thermoelastic Materials, Dissertation, Department of Aerospace Engineering and Mechanics, University of Minnesota,
1997
[HS00] K. F. Hane, T. W. Shield, Microstructure in the Cubic to Trigonal Transition, Materials Science and Engineering A 291 (2000), 147-159
[Iq99] Z. Iqbal, Variational methods in solid mechanics, Ph.D. Thesis, University
of Oxford, 1999
[Ja86] R. D. James, Displacive phase transformations in solids, J. Mech. Phys.
Solids 34 (1986), 359-394
[JH00] R. D. James, K. F. Hane, Martensitic transformations and shape-memory
materials, Acta Mater. 48 (2000), 197-222
[JK89] R. D. James, D. Kinderlehrer, Theory of diusionless phase transitions,
in: PDEs and continuum models of phase transitions (Nice, 1988), Lecture Notes
in Phys. 344, Springer, Berlin, 1989, 51-84
[JK93] R. D. James, D. Kinderlehrer, Theory of magnetostriction with applications to Tbx Dy1 xFe2 , Phil. Mag. B 68 (1993), 237-274
[Jo72a] F. John, Bounds for deformations in terms of average strains, Inequalities,
III, Proc. Third Sympos., Univ. California, Los Angeles, Calif., 1969, Academic
Press, New York, 1972, 129-145
[Jo72b] F. John, Uniqueness of non-linear elastic equilibrium for prescribed boundary displacements and suciently small strains, Comm. Pure Appl. Math. 25
(1972), 617-635
[KPt98] F. X. Kayser, J. W. Patterson, Sir William Chandler Roberts-Austen
His r
ole in the development of binary diagrams and modern physical metallurgy,
J. Phase Equilibria 19 (1998), 11-18
[Kh67] A. Khachaturyan, Some questions concerning the theory of phase transformations in solids, Soviet Physics Solid State 8 (1967), 2163-2168
[Kh83] A. Khachaturyan, Theory of structural transformations in solids, Wiley,
1983
[Ki88] D. Kinderlehrer, Remarks about equilibrium congurations of crystals, in:
Material instabilities in continuum mechanics and related mathematical problems (J. M. Ball, ed.), Oxford University Press, 1988, 217-242
[KP91] D. Kinderlehrer, P. Pedregal, Characterization of Young measures
generated by gradients, Arch. Rational Mech. Anal. 115 (1991), 329-365
[KP95] D. Kinderlehrer, P. Pedregal, Remarks on the two-well problem (unpublished manuscript)
[Kir98] B. Kirchheim, Lipschitz minimizers of the 3-well problem having gradients
of bounded variation, MPI MIS Preprint #12, 1998
[Kir00] B. Kirchheim, Geometry and rigidity of microstructures, Habilitationsschrift, Universit
at Leipzig, 2001
[K91] R. V. Kohn, The relaxation of a double-well energy, Contin. Mech. Thermodyn. 3 (1991), 193-236
[KL00] R. V. Kohn, V. Lods, Some results about two incompatible elastic strains,
manuscript
ller, Branching of twins near an austenite-twinned[KM92] R. V. Kohn, S. Mu
martensite interface, Phil. Mag. A 66 (1992), 697-715
ller, Surface energy and microstructure in coherent
[KM94] R. V. Kohn, S. Mu
phase transitions, Comm. Pure Appl. Math. 47 (1994), 405-435
[KN00] R. V. Kohn, B. Niethammer, Geometrically nonlinear shape-memory
polycrystals made from a two-variant material, M2 AN Math. Model. Numer.
Anal. 34 (2000), 377-398
206
References
[KS86] R. V. Kohn, G. Strang, Optimal design and relaxation of variational

problems II, Comm. Pure Appl. Math. 39 (1986), 139-182
[Kr99] J. Kristensen, On the non-locality of quasiconvexity, Ann. Inst. H.
Poincare Anal. Non Lineaire 16 (1999), 1-13
[Kr00] J. Kristensen, On conditions for polyconvexity, Proc. Amer. Math. Soc.
128 (2000), 1793-1797
[KnF] I. Kundler, H. Finkelmann, Strain-induced director reorientation in nematic liquid single crystal elastomers, Macromol. Chem. Rapid Communications
16 (1995), 679-686
pfer, H. Finkelmann, Nematic liquid single crystal elastomers,
[KpF91] J. Ku
Macromol. Chem. Rapid Communications 12 (1991), 717-726
[LDR95] H. LeDret, A. Raoult, The quasiconvex envelope of the Saint VenantKirchho stored energy function, Proc. Roy. Soc. Edinburgh 125 (1995), 11791192
[LL98a] B. Li, M. Luskin, Nonconforming nite element approximation of crystalline microstructure, Math. Comp. 67 (1998), 917-946
[LL98b] B. Li, M. Luskin, Finite element analysis of microstructure for the cubic
to tetragonal transformation, SIAM J. Numer. Anal. 35 (1998), 376-392
[LL99a] B. Li, M. Luskin, Approximation of a martensitic laminate with varying
volume fractions, M2 AN Math. Model. Numer. Anal. 33 (1999), 67-87.
[LL99b] B. Li, M. Luskin, Theory and computation for the microstructure near
the interface between twinned layers and a pure variant of martensite, Materials
Science & Engineering A 273 (1999), 237-240
[LL00] B. Li, M. Luskin, Approximation of a martensitic laminate with varying
volume fractions, M2 AN Math. Model. Numer. Anal. 33 (1999), 67-87.
[Li95] Z. Li, A numerical method for computing singular minimizers, Numer. Math.
71 (1995), 317-330
[Li97] Z. Li, Simultaneous numerical approximation of microstructures and relaxed
minimizers, Numer. Math. 78 (1997), 21-38
[Li98a] Z. Li, Laminated microstructure in a variational problem with a non-rankone connected double well potential, J. Math. Anal. Appl. 217 (1998), 490-500
[Li98b] Z. Li, Rotational transformation method and some numerical techniques
for computing microstructures, Math. Models Methods Appl. Sci. 8 (1998), 9851002
[Liu77] F. C. Liu, A Luzin type property of Sobolev functions, Indiana Univ. Math.
J. 26 (1977), 645-651
[L96a] M. Luskin, Approximation of a laminated microstructure for a rotationally
invariant, double well energy density, Numer. Math. 75 (1996), 205-221
[L96b] M. Luskin, On the computation of crystalline microstructure, Acta Numer.
5 (1996), 191-257
[L98] M. Luskin, The stability and numerical analysis of martensitic microstructure, in ENUMATH 97, H. Bock, F. Brezzi, R. Glowinski, G. Kanschat, Y.
Kuznetsov, J. Periaux and R. Rannacher, editors, 1998, 54-69
[L00] M. Luskin, On the stability of microstructure for general martensitic transformations, in: Lectures on Applied Mathematics, H.-J. Bungartz and R. W.
Hoppe and C. Zenger, editors, Springer-Verlag, 2000, 31-44.
[Mt00] J. Matou
sek, On directional convexity, Discrete Comput. Geom. 25
(2001), 389-403
c
, On functional separately convex hulls, Dis[MP98] J. Matou
sek, P. Plecha
crete Comput. Geom. 19 (1998), 105-130
[M65] N. G. Meyers, Quasi-convexity and lower semi-continuity of multiple variational integrals of any order, Trans. Amer. Math. Soc. 119 (1965) 125-149
References
207
[Mi02] G. W. Milton, The theory of composites, Cambridge monographs on applied and computational mathematics, Cambridge University Press, 2002
[Mo52] C. B. Morrey, Quasi-convexity and the lower semicontinuity of multiple
integrals, Pacic J. Math. 2 (1952), 25-53
[Mo66] C. B. Morrey, Multiple integrals in the calculus of variations, Springer,
1966
ller, Microstructures, phase transitions and geometry, in: Proceed[M98] S. Mu
ings European Congress of Mathematics, Budapest, July 22-26, 1996 (A. Balog
et al., eds), Progress in Mathematics, Birkh
auser 1998
ller, A sharp version of Zhangs theorem on truncating sequences
[M99a] S. Mu
of gradients, Trans. Amer. Math. Soc. 351 (1999), 4585-4597
ller, Variational methods for microstructure and phase transitions,
[M99b] S. Mu
in: Proc. C.I.M.E. summer school Calculus of variations and geometric evolution problems, Cetraro, 1996, (F. Bethuel, G. Huisken, S. M
uller, K. Steen,
S. Hildebrandt, M. Struwe, eds.), Springer LNM 1713, 1999
ller, V. Sver
k, Attainment results for the two-well problem by
[MS96] S. Mu
a
convex integration, in: Geometric analysis and the calculus of variations (J.
Jost, ed.), International Press, 1996, 239-251
ller, V. Sver
k, Convex integration with constraints and applica[MS99a] S. Mu
a
tions to phase transitions and partial dierential equations, J. Eur. Math. Soc.
(JEMS) 1 (1999), 393-442
ller, V. Sver
k, Convex integration for Lipschitz mappings and
[MS99b] S. Mu
a
counterexamples to regularity, Ann. of Math. (2), to appear
n-Marrero, A numerical method for detecting singular min[NM90] P. V. Negro
imizers of multidimensional problems in nonlinear elasticity, Numer. Math. 58
(1990), 135-144
[NW93] R. A. Nicolaides, N. J. Walkington, Computation of microstructure
utilizing Young measure representations, Transactions of the Tenth Army Conference on Applied Mathematics and Computing (West Point, NY, 1992), ARO
Rep., 93-1, 1993, 57-68
[NW95] R. A. Nicolaides, N. J. Walkington, Strong convergence of numerical
solutions to degenerate variational problems, Math. Comp. 64 (1995), 117-127
[OG94] M. Ortiz, G. Gioia, The morphology and folding patterns of bucklingdriven thin-lm blisters, J. Mech. Phys. Solids 42 (1994), 531-559
[OR99] M. Ortiz, E. Repetto, Nonconvex energy minimization and dislocation
structures in ductile single crystals, J. Mech. Phys. Solids 47 (1999), 397-462
[Pa77] G. P. Parry, On the crystallographic point groups and on Cauchy symmetry, Math. Proc. Cambridge Philos. Soc. 82 (1977), 165-175
[Pa81] G. P. Parry, On the relative strengths of stability criteria in anisotropic
materials, Mech. Res. Comm. 7 (1980), 93-98
[Pe93] P. Pedregal, Laminates and microstructure, Europ. J. Appl. Math. 4
(1993), 121-149
[SMST97] SMST-97-Proceedings, Proceedings of the Second International Conference on Shape Memory and Superelastic Technology Conference held in March
1997, Asilomar, California, USA, (A. Pelton, D. Hodgson, S. Russell, and T.
Duerig, eds.), Shape Memory Applications, Inc., 2380 Owen Street, Santa Clara,
CA
[Pi91] A. C. Pipkin, Elastic materials with two preferred sates, Q. J. Mech. Appl.
Math. 44 (1991), 1-15
[Pi84] M. Pitteri, Reconciliation of local and global symmetries of crystals, J.
Elasticity 14 (1984), 175-190
[PZ98] M. Pitteri, G. Zanzotto, Generic and non-generic cubic-to-monoclinic
transitions and their twins, Acta. Mater. 46 (1998), 225-237
208
References
[PZ00] M. Pitteri, G. Zanzotto, Continuum models for phase transitions and

twinning in crystals, Chapman & Hall, London, 2002
[Pr00] A. Prohl, An adaptive nite element method for solving a double well problem describing crystalline microstructure, M2 AN Math. Model. Numer. Anal.
33 (1999), 781-796
[S69] D. Sands, Introduction to crystallography, W. A. Benjamin, Inc., 1969,
unabridged republication by Dover, 1993
[Sch74] V. Scheffer, Regularity and irregularity of solutions to nonlinear second
order elliptic systems of partial dierential equations and inequalities, Dissertation, Princeton University, 1974
[Sr83] D. Serre, Formes quadratiques et calcul des variations, J. Math. Pures
Appl. 62 (1983), 177-196
, Relaxation of a class of SO(n)-invariant energies related to

[Sy01] M. Silhav
y
nematic elastomers, manuscript
[Si89] S. A. Silling, Phase changes induced by deformation in isothermal elastic
crystals, J. Mech. Phys. Solids 37 (1989), 293-316
k, On regularity of the Monge-Amp`ere equation without convexity

[Sv91] V. Sver
a
assumptions, preprint, Heriot-Watt University, 1991
k, Rank-one convexity does not imply quasiconvexity, Proc. Roy.

[Sv92a] V. Sver
a
Soc. Edinburgh Sect. A 120 (1992), 293-300
k, New examples of quasiconvex functions, Arch. Rational Mech.

[Sv92b] V. Sver
a
Anal. 119 (1992), 293-300
k, On the problem of two wells, in: Microstructure and phase tran[Sv93] V. Sver
a
sitions, IMA Vol. Appl Math. 54, (D. Kinderlehrer, R. D. James, M. Luskin
and J. Ericksen, eds.), Springer, 1993, 183-189
k, Lower semicontinuity of variational integrals and compensated

[Sv95] V. Sver
a
compactness, in: Proc. ICM 1994 (S. D. Chatterji, ed.), vol. 2, Birkh
auser, 1995,
1153-1158
[Ta79] L. Tartar, Estimations de coecients homog
eneises, in: Computing methods in applied sciences and engineering, Lecture Notes in Math. 704, Springer,
1979, 364-373
[Ta83] L. Tartar, The compensated compactness method applied to systems of
conservation laws, in: Systems of nonlinear partial dierential equations, NATO
Adv. Sci. Inst. Ser. C: Math. Phys. Sci. 111, Reidel, 1983, 263-285
[Ta90] L. Tartar, H-measures, a new approach for studying homogenisation, oscillations and concentration eects in partial dierential equations, Proc. Roy.
Soc. Edinburgh Sect. A 115 (1990), 193-230
[Ta93] L. Tartar, Some remarks on separately convex functions, in: Microstructure
and phase transitions, IMA Vol. Appl Math. 54, (D. Kinderlehrer, R. D. James,
M. Luskin and J. Ericksen, eds.), Springer, 1993, 191-204
[Ta95] L. Tartar, Beyond Young measures, Meccanica 30 (1995), 505-526
[Tr38] F. J. Terpstra, Die Darstellung biquadratischer Formen als Summen von
Quadraten mit Anwendungen auf die Variationsrechnung, Math. Ann. 116
(1938), 166-180
[VTA74] G. Van Tendeloo, S. Amelinckx, Group-theoretical considerations concerning domain formation in ordered alloys, Acta Cryst. A30 (1974), 431-440
[WT96] M. Warner, E. M. Terentjev, Nematic elastomers a new state of
matter?, Prog. Polym. Sci. 21 (1996), 835-891
[WLR53] M. S. Wechsler, D. S. Lieberman, T. A. Read, On the theory of the
formation of martensite, Trans. AIME J. Metals 197 (1953), 1503-1515
[WB96] J. Weilepp, H. R. Brand, Director reorientation in nematic-liquidsingle-crystal elastomers by external mechanical stress, Europhys. Lett. 34
(1996), 495-500
References
209
[Yo37] L. C. Young, Generalized curves and the existence of an attained absolute

minimum in the calculus of variations, Comptes Rendues de la Societe des
Sciences et des Lettres de Varsovie, classe III 30 (1937), 212-234
[Za92] G. Zanzotto, On the material symmetry group of elastic crystals and the
Born rule, Arch. Rational Mech. Anal. 121 (1992), 1-36
[Z92] K. Zhang, A construction of quasiconvex functions with linear growth at
innity, Ann. Scuola Norm. Sup. Pisa Cl. Sci. (4) 19 (1992), 313-326
[Z98] K. Zhang, On various semiconvex hulls in the calculus of variations, Calc.
Var. Partial Dierential Equations 6 (1998), 143-16
Index
algorithm
for the computation of envelopes, 156
for the computation of laminates, 166
austenite, 194
austenite-martensite transformation, 3,
193
body centered unit cell, 194
Bravais lattice, 193
cofactor, 197
compatible wells, 189
compound twin, 196
computation
of laminates, 163
of rank-one convex envelopes, 154
condition HN , 186
conditions for uniqueness of microstructure
condition (Cb ), 89
b ), 103
condition (C
condition (Ctf ), 111
constrained point, 11
convexity conditions, 183
crystallographic point groups, 193
cubic, 193
D-convexity, 185
Dirac mass, 187
eight point set, 13
energy well, 2
excess rotation, 85, 96
face centered unit cell, 194
four-well problem, 42
incompatible wells, 27, 189

innite laminate, 171
laminate, 187
of nite order, 188
lamination convex hull, 185
lamination method, 7, 11
macroscopic phase diagram, 72
martensite, 194
minor relation, 8
monoclinic, 193
multi-well set, 2
nematic director, 70
nematic elastomers, 69
one-well problem, 26
orthogonal group, 198
orthorhombic, 193
point group, 194
polar decomposition, 198
polyane functions, 8
polyconvex hull, 7, 185
polyconvex measure, 187
polyconvexity, 183
quadratic forms, 184
quasiconvex hull, 5, 185
quasiconvexity, 183
Radon measures, 186
rank-one connection, 4
existence, 190
rank-one convex hull, 7, 185
rank-one convexity, 183
relaxed energy, 69
gradient Young measure, 187

hexagonal, 193
homogeneous Young measure, 187
second order laminate, 188

semiconvex hull, 7
separation method, 7, 11
212
Index
sets dened by singular values

in three dimensions, 59
in two dimensions, 58
signed singular values, 198
simple laminate, 8, 188
singular value, 198
splitting method, 11
stability, 87
of microstructure, 83
Taylor bound, 51
tetragonal, 193
three-well problem, 39
triclinic, 193
twin
compound, 196
type-I, 196
type-II, 196
twinning system, 195
two-well problem
in three dimensions, 55
in two dimensions, 27
unconstrained point, 11
uniqueness
in cubic to orthorhombic transformations, 135
in cubic to tetragonal transformations, 128
in cubic to trigonal transformations,
134
in tetragonal to monoclinic transformations, 143
of microstructure, 83
unit cell, 194
Young measure, 7

Variational Methods For Crystalline Microstructure - Analysis and Computation - G. Dolzmann (Springer, 2003) WW

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Variational Methods For Crystalline Microstructure - Analysis and Computation - G. Dolzmann (Springer, 2003) WW

Uploaded by

Copyright:

Available Formats

Lecture Notes in Mathematics

Variational Methods for

Cataloging-in-Publication Data applied for

Mathematics Subject Classication (2000): 74B20, 74G15, 74G65, 74N15, 65M60

41/3142/ du - 543210 - Printed on acid-free paper

The mathematical modeling of microstructures in solids is a fascinating topic

Semiconvex Hulls of Compact Sets . . . . . . . . . . . . . . . . . . . . . . .

Macroscopic Energy for Nematic Elastomers . . . . . . . . . . . . . .

Uniqueness and Stability of Microstructure . . . . . . . . . . . . . . .

Applications to Martensitic Transformations . . . . . . . . . . . . . .

Algorithmic Aspects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153

A. Convexity Conditions and Rank-one Connections . . . . . . . . . 183

Many material systems show fascinating microstructures on dierent length

where R denotes the reference conguration of the elastic body, x

G. Dolzmann: LNM 1803, pp. 110, 2003.

Fig. 1.1. Microstructures in a single crystal CuAlNi (courtesy of Chu&James,

W (RF, T ) = W (F, T ) for all R SO(3).

Secondly, the invariance reecting the symmetry of the high temperature

QU R K(T ) for all Q SO(3), R Pa .

1.1 Martensitic Transformations and Quasiconvex Hulls

1.1 Martensitic Transformations and Quasiconvex Hulls

where R2 and R3 are elements in the cubic point group given by

Fig. 1.3. Formation of an interface between two variants of martensite in a single

displacively, without diusion of the atoms in the underlying lattice. This is

(Hadamards jump condition),

where the matrix an is dened by (an)kl = (ak nl ) for a, n R3 . If (1.5)

holds for all F Mmn and C0 ([0, 1]n ; Rm ). Quasiconvexity of W

1.1 Martensitic Transformations and Quasiconvex Hulls

therefore the crucial notion of convexity in the vector valued calculus of

1.1 Martensitic Transformations and Quasiconvex Hulls

1.2 Outline of the Text

1.2 Outline of the Text

is unique and what (1.10) implies for uh as h tends to zero. In particular, if

2. Semiconvex Hulls of Compact Sets

G. Dolzmann: LNM 1803, pp. 1168, 2003.

2. Semiconvex Hulls of Compact Sets

i.e., is a constrained point. Therefore it suces to prove that the constrained

We assert that every Mqc Mpc of this form is in fact a laminate.

now imply that

2.1 The Eight Point Example

and consequently Mpc (K) \ Mqc (K). Note, however, that , id K qc .

2.1 The Eight Point Example

2. Semiconvex Hulls of Compact Sets

holds. The proof of our characterization of the quasiconvex hull relies on

Moreover, the following assertions hold:

2.1 The Eight Point Example

Remark 2.1.4. A short calculation shows that the additional inequalities in

2. Semiconvex Hulls of Compact Sets

It turns out that the hyperplanes dened by n1 , . . . , n4 contain six points in

We include a few details of the argument that leads to this characterization.

2.1 The Eight Point Example

Geometrically this is a hyperboloid which contains in fact six of the eight

2. Semiconvex Hulls of Compact Sets

Four points on a b-face of the cube. The corresponding

as well as the additional inequalities

which we can rewrite as

We now assert that this system of inequalities is equivalent to the conditions

describing the convex hull of K and the two additional inequalities

2.1 The Eight Point Example

The Lamination Convex Hull of K for ac b2 < 0. We now turn

dening the convex hull of K, the hyperboloids

in the denition of K pc , and the cones

2. Semiconvex Hulls of Compact Sets

and we write t F (y, t) or t Fy,t for simplicity. A short calculation shows

thus SO(2)Ui = SO(2)Uj , a contradiction. If M0 = {U1 , . . . , Uk } \ C = ,

We now assume that tF = 0. If we deduce a contradiction to this assumption,

|F eq |2 = |Uq eq |2 + 2tF Uq eq , Uq ee , eq + t2F |Uq e|2 e , eq