高等固体力学

INTRODUCTION TO MICROMECHANICS
AND NANOMECHANICS
Lecture Notes (CE236/C214)
SHAOFAN LI
Department of Civil and Environmental Engineering,
University of California, Berkeley, CA94720, USA
University of California
Berkeley, California
Contents
1. INTRODUCTION 1
2. PRELIMINARY 4
2.1 Vectors and Tensors 4
2.2 Review of Linear Elasticity Theory 13
2.3 Exercises 19
3. HOMOGENIZATION I — CLASSICAL AVERAGING METHOD 21
3.1 Representative volume element 21
3.2 Average stress in an RVE 22
3.3 Average strain and strain rate 24
3.4 Definition of eigenstrain, eigenstress, and inclusion 26
3.5 Eshelby’s equivalent eigenstrain method I: Traction boundary
condition 27
3.6 Eshelby’s equivalent eigenstress method II: Displacement boundary
condition 28
3.7 Effective material properties via eigenstrain method 29
3.8 Jock Eshelby (I) 33
3.9 Exercises 35
4. GREEN’S FUNCTION AND FOURIER TRANSFORM 37
4.1 Green’s Function 37
4.2 Fourier transform 40
4.3 Examples of Green’s Function 46
4.4 Static Green’s function for 3D linear elasticity 49
4.5 Variation in a Theme: Radon Transform 54
4.6 Joseph Fourier(I) 58
vi INTRODUCTION TO MICROMECHANICS AND NANOMECHANICS
4.7 Exercises 61
5. EIGENSTRAIN THEORY 63
5.1 Fundamental equations of micro-elasticity 63
5.2 Method of Green’s Functions 66
5.3 Application I: Dislocation problems 68
5.4 Application II: Stress intensity factor for a flat ellipsoidal crack 71
5.5 Isotropic inclusion-Eshelby’s solution 78
5.6 Exterior Solution of Ellipsoidal Inclusion 83
5.7 Jock Eshelby (II): Lessons from J.D.Eshelby 88
5.8 Exercises 90
6. EFFECTIVE ELASTIC MODULUS 95
6.1 Effective elastic moduli for composites of dilute suspension 95
6.2 Self-consistent method 104
6.3 Mori-Tanaka methods 109
6.4 Rodney Hill 120
6.5 Exercises 124
7. INTRODUCTION OF DISLOCATION THEORY 127
7.1 Screw dislocation 127
7.2 Edge dislocation 134
7.3 The Peach-Koehle force 139
7.4 Configuration force: Eshelby’s energy-momentum tensor 144
7.5 Continuum theory of dislocation 150
7.6 Discrete Dislocation Dynamics (DD) 160
7.7 The Peierls-Nabarro Model 164
7.8 Dislocations in the epitaxial thin film 176
8. COMPARISON VARIATIONAL PRINCIPLES 188
8.1 Review of Variational Calculus 188
8.2 Extreme variational principles in linear elasticity 192
8.3 Hashin-Shtrikman variational principles 199
8.4 Review of Functional Analysis and Convex Analysis 205
8.5 Legendre Transformation and Duality 215
8.6 Legendre-Fenchel transformation in linear elasticity 222
8.7 Talbot-Willis variational principles 224
Contents vii
8.8 Exercises 227

9. BOUNDS ON EFFECTIVE PROPERTIES 230
9.1 Hashin-Shtrikman bounds 230
9.2 Microstructure Characterization 241
9.3 Exercises 249
10. PERIODIC MICROSTRUCTURE 251
10.1 Unit cell and Fourier series 252
10.2 Eigenstrain homogenization 256
10.3 Introduction to Asymptotic Homogenization 262
10.4 Variational Characterization 278
10.5 Multiscale Finite Element Method 282
10.6 G-, H-, and Γ- convergence 289
10.7 Exercises 300
10.8 Toshia Mura 301
11. MICROMECHANICS THEORY OF VOID GROWTH 306
11.1 Void Growth in Linear Viscous Solids 306
11.2 The McClintock solution 312
11.3 The Gurson model 318
11.4 Exercise 328
References 329
Introduction 1
Chapter 1
INTRODUCTION
What is micromechanics ? Generally speaking, micromechanics is a scien-

tific discipline that studies: (1) mechanical, electrical, and, in general, thermo-
dynamical behaviors of a material with microstruture, or (2) materials’ behav-
iors at micro (nano) or mesoscale.
In recent years, micromechanics has become an indispensible part of the-
oretical foundation for many engineering fields and emergying technologies
such as nanotechnology and biomedical technology.
The term “micromechanics” has become a truly interdiscipline jargon. It
has been used with different meanings in different contexts. Traditionally, in
the area of applied mechanics, micromechanics is referred to as a hierarchical
mechanics paradigm that deals the effective material properties that are statis-
tical averagies of a nested two level structure: microscopic and macroscopic
structures. A material point at a macrolevel can be viewed as an ensemble
microscope material space. The physical laws at macrolevel or the material
behaviors at macro-level are derived from the ensemble average of massive
micro-objects governed by the physical laws at microlevel. For instance, the
effective material properties at macrolevel are the average of material proper-
ties of microstructures at fine scale. In general, the two-level paradigm is a
special mathematical abstraction that is not associated with any fixed length
scale. When studying material properties of a metal, 1mm may be viewed
◦
as macroscale, and the length scale at microlevel may range from A to nm;
whereas studying the deformation of a dam, the macroscale could be up to 103
m, and the length scale at microlevel may be around 10−2 m. In this sense,
traditional micromechanics is essentially a particular (in some sense classical)
averaging theory that takes into account the overall effects of microstructures.
In practice, it deals with subjects of a broad spectrum: material properties of
2 INTRODUCTION TO MICROMECHANICS AND NANOMECHANICS
composite/synthetic materials, e.g. composite structures, cementitious materi-

als, geotechnical materials, and phase transformations; material properties of
bio-materials, e.g. constitutive modeling of bone, muscle, blood flow; environ-
mental problems e.g. air pollutions, ground water transport and diffusion, oil
spill in the ocean, etc.
In condensed matter physics and today in applied mechanics as well, the
term micromechanics is used to describe a three-level physics realm: microme-
◦
chanics at molecular or atomic level (A), meso-mechanics at nm length scale,
and macroscopic phenomenological theory at mm level or up.
The main task of contemporary micromechanics, or nano-mechanics, is to
seek unknown physical laws or mechanics regulations at the nano-scale. Dif-
ferent from traditional micromechanics, a salient feature of nanomechanics is
its multiscale and multi-physics character. It includes some features that are
present in quantum mechanics, or quantum statistical mechanics, a manifesta-
tion of the effects at atomic or sub-atomic level; on the other hand, it also shares
with many features from the description of continuum mechanics, because of
the size statistical ensemble.
The impetus for contemporary micromechanics or nano-mechanics is pri-
marily due to the emergence of nanoscience and bio-medical technology. It
appears that physics along is not sufficient to deal with the many problems that
are appearing from today’s nano-technologies and nano-engineering. There
is a call for a nano-mechanics and nano-computational mechanics to serve as
the infra-structure of these emerging engineering fields. For instatnces, much
attention has been focused on material properties of thin film, manufactur-
ing devices and components of a microelectromechanical system (MEMS),
e.g. sub-micro size sensors, motors, the mechanics of nanotube and nanowire,
computer-aided material design, and micro-biophysics/biochemistry systems,
e.g. protein/DNA interaction in biomolecular simulation (e.g. Schlik et al
[1999ab]), etc.
From the perspective of higher learning and intellectual advancement, mi-
cromechanics has developed into a rigorous mathematical theory, philosoph-
ical methodology, and beautiful computational realization. Forty years ago,
micro-elasticity started with simple definitions of eigenstrain and inclusion,
came along with Eshelby’s elegant equivalent homogenization theory (Eshelby
[1957],[1959],[1961]) and Hashin & Shtrikman’s variational principle (Hashin
and Shtrikman [1962ab],[1964]), it is now the foundation of an entire compos-
ite material industry.
Less than ten years ago, Lattice Boltzmann method first debuted as a numer-
ical emulation of continuous Boltzmann equation in statistical physics. Today,
Lattice Boltzmann method has become a bona fide computational mesome-
chanics paradigm, and it has been used to solve problems such as turbulence
flow (Qian et al [1992][1993]), combustion, and flow pass through porous me-
Preliminary 3
dia and even cooling of packed flowers (Van der Sman [1997][2000]); In later
1980s, Clementi and his co-workers [1988] initiated the idea of multiscale
modeling, or multiscale simulation, i.e. using super-computers to conduct
large scale computations that combine ab initio modeling, classical molecu-
lar dynamic modeling, and phenomenological modeling in a single simula-
tion. The unified macroscopic, atomistic, ab initio dynamics (MAAD) de-
scription brings all three descriptions together into a seamless union, embrac-
ing all the size scales, from the very small to the very big (e.g. Abraham et al
[1996],[1997ab],[2000]).
The simplest and earlist multi-scale modeling notion is the so-called Cauchy-
Born rule. By combining this concept with the finite element methods, the so-
called quasicontinuum method was developed by Tadmor, Ortiz, Phillips and
their co-workers (Tadmor et al 1996). The Cauchy-Born rule is ensentially a
simplistic “homogenization postulation” in lattice kinematics, and it serves as
passage to link between the molecular dynamics and continuum mechanics.
The Born rule assumes that the continuum energy density W can be computed
using an atomic potential, with the link to the continuum being the deformation
gradient F. To briefly review continuum mechanics, the deformation gradient
F maps an undeformed line segment dX in the reference configuration onto a
deformed line segment dx in the current configuration,
dx = FdX (1.1)
In general, F can be written as
du
F=I+ (1.2)
dX
where u is the displacement vector. If there is no displacement in the contin-
uum, the deformation gradient is equal to unity.
The major restriction and implication of the Cauchy-Born rule is that the
continuum deformation must be homogeneous. This results from the fact that
the underlying atomic system is forced to deform according to the contin-
uum deformation gradient F. By using the Born rule, one may be able to
derive a continuum stress tensor and tangent stiffness directly from the inter-
atomic potential, which allowed the usage of the standard nonlinear finite ele-
ment method. This procedure is now called as the so-called quasi-continuum
method.
Apparently, the contemporary mico-mechanics or nano-mechanics is only
at its infancy. There are many unknown approaches to be explored and many
new phenonmena to be studied. In this lecture notes, we are attempting to
synthesize the most recent research results in the forefront of nano-mechanics
while presenting traditional micro-mechanics in a coherent fashion. By doing
so, we hope that it may serve as a stepping stone for us to reach a new height
in the quest for a multiscale nano-mechanics of our time.
Chapter 2
PRELIMINARY
2.1 Vectors and Tensors

2.1.1 Vectors
Consider a Cartisian coordinate in a three dimensional space with unit vector
basis, {ei }, i = 1, 2, 3. An arbitrary position vector, x, may be expressed as
x = x1 e1 + x2 e2 + x3 e3 = xi ei = (x · ei )ei (2.1)
where Einstein convention is used that the repeated indices indicates summa-
tion from 1 to 3.
Consider two vectors, V = Vi ei and W = Wj ej . The scalar (dot) product
of two vectore, V and W, is defined as

V · W = Vi ei · Wj ej = Vi Wj ei · ej = Vi Wj δij = Vi Wi (2.2)
where
1, i=j
ei · ej = =: δij (2.3)
0, i 6= j
is called Keronecker delta.
A cross product of two vectors, A = Ai ei , B = Bj ej , is defined as
A × B = (Ai ei ) × (Bj ej ) = Ai Bj ei × ej = ekij Ai Bj ek (2.4)
where ei × ej = ekij ek , and ekij is called the permutation symbol,


 1, for an even permutation of 1, 2, 3
eijk = −1, for an odd permutation of 1, 2, 3 (2.5)
0, repeated indices

Preliminary 5
This definition can be explained as a permutation rule that change of any two
adjcent indces of the symbol, there is a negative sign (−1) occurs.
For example, since e123 = 1, then
e132 = (−1)e123 = (−1)(1) = −1
and
e312 = (−1)e132 = (−1)(−1)e123 = (−1)(−1)1 = 1
The cross product of two vectors can also written as
A × B = ekij Ai Bj ek = e1ij Ai Bj e1 + e2ij Ai Bj e2 + e3ij Ai Bj e3

= (A2 B3 − A3 B2 )e1 + (A3 B1 − A1 B3 )e2 + (A1 B2 − A2 B1 )e3
e1 e2 e3
= A1 A2 A3 (2.6)
B1 B2 B3
Therefore
ei × ej = ekij ek , ⇒ ekij = (ei × ej ) · ek (2.7)
Since
e1 e2 e3
ei × ej = δ1i δ2i δ3i (2.8)
δ1j δ2j δ3j
then
δ1k δ2k δ3k δ1i δ2i δ3i
ekij = eijk = (ei × ej ) · ek = δ1i δ2i δ3i = δ1j δ2j δ3j (2.9)
δ1j δ2j δ3j δ1k δ2k δ3k
This provides a link between permutation symbol and Keronecker delta.
Consider the product of two permutation symbols,
δ1i δ2i δ3i δ1r δ2r δ3r
eijk erst = δ1j δ2j δ3j δ1s δ2s δ3s
δ1k δ2k δ3k δ1t δ2t δ3t
δ1i δ2i δ3i δ1r δ2s δ3t
= δ1j δ2j δ3j δ1r δ2s δ3t
δ1k δ2k δ3k δ1r δ2s δ3t
δir δis δit
= δjr δjs δjt (2.10)
δkr δks δkt
One may show that for any second order tensor A,
1 When i = r, eijk eist = δjs δkt − δjt δks ;
2 When i = r and j = s, eijk eij` = 2δk` ;
3 When i = r, j = s, and k = t, eijk eijk = 3! = 6.
which are call e − δ identities.
2.1.2 Tensor Algebra

Consider two vectors, A = Ai ei and B = Bj ej . One can form a second
order tensor, C by using the tensor product

C = A ⊗ B = Ai ei ⊗ Bj ej = Ai Bj ei ⊗ ej (2.11)
The dyad is called the second order tensor 1 , and its basis, ei ⊗ ej , is called
dyadic basis. In this case, the components of the second order tensor are Cij =
Ai Bj .
Figure 2.1. Cartesian Coordinate
In fact, every second order tensor can be expressed in a dyadic basis, such
as
σ = σij ei ⊗ ej (2.12)
= ij ei ⊗ ej (2.13)
1 One may call the vector as the first order tensor.

Preliminary 7
A conjugate of a dyad (second order tensor) is defined as

T
:= ji ei ⊗ ej (2.14)
This is why in linear elasticity we may define the infinitesimal strain tensor as
1 1
= ∇ ⊗ u + (∇ ⊗ u)T = uj,i + ui,j ei ⊗ ej (2.15)
2 2
1
or in component form ij = uj,i + ui,j .
2
In general, a n-th order tensor is a polyads, or has a polyadic representation,
e.g.
C = Cijk` ei ⊗ ej ⊗ ek ⊗ e` (2.16)
is a forth order tensor.
Analogous to the scalar product of vectors, the double contraction of two
tensors are defined as two dot products among of Cartesian tensor bases, i.e. if
A = Aij ei ⊗ ej and B = Bk` ek ⊗ e` , then
A : B = (Aij ei ⊗ ej ) : (Bk` ek ⊗ e` ) = Aij Bk` (ei · ek )(ej · e` )

= Aij Bk` δik δj` = Aij Bij (2.17)
The trace of a second order tensor is defined as
trA := A : 1(2) = Aii = A11 + A22 + A33 (2.18)
In each contraction, there are two bases annihilated. Consider a forth order
tensor C = Cijk` ei ⊗ ej ⊗ ek ⊗ e` and a second order tensor = ij ei ⊗ ej .
There are total six basis vectors. A double contraction between the two will
annihilate four basis vectors and produce a second order tensor, i.e.

σ = C : = Cijk` ei ⊗ ej ⊗ ek ⊗ eell : st es ⊗ et
= Cijk` st ei ⊗ ej δks δ`t = Cijk` k` ei ⊗ ej (2.19)
In component form, σij = Cijk` k` .

We say that a second order tensor is symmetric, if
T
A = A , or in component form Aij = Aji (2.20)
A second order tensor is skew symmetric, if

T
A = − A , or in component form Aij = −Aji (2.21)
In general, an arbitrary second order tensor can be expressed as

1
Aij = Aij + Aji + Aij − Aji = A(ij) + A[ij] (2.22)
2
Denote an arbitrary second order Cartesian basis as
eij = ei ⊗ ej . (2.23)
The second order unit tensor and the forth order unit tensor are constructed
based on the following rules:

1(2) := ei · ej ei ⊗ ej = δij ei ⊗ ej = δij eij (2.24)

1(4) := ei ⊗ ej : ek ⊗ e` ei ⊗ ej ⊗ ek ⊗ e`
= (eij : ek` )eij ⊗ ek` = δik δj` ei ⊗ ej ⊗ ek ⊗ e` (2.25)
The superscript indicates the order. It is interesting to note that the fourth order
unit tensor defined in (2.25) is not symmetric with all indices.
To represent symmetric tensors, it may be expedient to first define symmec-
tric tensor basis. The second order symmetric basis is defined as
1 1
eSij = eij + eTij = ei ⊗ ej + ej ⊗ ei (2.26)
2 2
Any second order symmetric tensor can then be expressed as S = Sij eSij . One
may denote the space of all second order symmetric tensors as
T (2s) = {S S = Sij eSij } (2.27)
The corresponding second order symmetric unit tensor is then defined as

1
1(2s) = ei · ej + ej · ei ei ⊗ ej
2
= δij ei ⊗ ej = 1(2) (2.28)

1
One may also define the second order anti-symmetric tensor as eA
ij = 2 ei ⊗

ej − ej ⊗ ei .
The fourth-order symmetric tensor bases is built upon the second order sym-
metric tensor bases, i.e.
eSijk` = eSij ⊗ eSk` (2.29)
and the fourth-order symmectric tensor space is defined as
T (4s) = {S S = Sijk` eSijk` } (2.30)

Preliminary 9
The corresponding fourth-order unit tensor is defined as

1
1(4s) := eSij : eSk` eSijk` == δik δj` + δi` δjk ei ⊗ ej ⊗ ek ⊗ e` (2.31)
2
It may be noted that the fourth-order unit tensor can be decomposed to sym-
metric part and antisymmetric part in terms of the first and second indices, or
the of the third and forth indices,
(4) 1 1
1ijk` := δik δj` = (δik δj` + δi` δjk ) + (δik δj` − δi` δjk )
2 2
(4s) (4a)
= 1ijk` + 1ijk` (2.32)
One may show that for given second-order tensor, A,
1(4) : A → A (2.33)
1
1(4s) : A → A + AT (2.34)
2
(4a) 1
1 :A → A − AT (2.35)
2
Note that 1(4) 6= 1(2) ⊗ 1(2) .
2.1.3 Inversion formula for fourth-order isotropic tensor

Consider general form of fourth order isotropic tensor,
Q = m1(2) ⊗ 1(2) + 2w1(4s) (2.36)
Let Q−1 be its inverse tensor. According to the well-known Sherman-Morrision
formula (e.g. Dahlquist and Bjorck [1974]),
m 1 (4s)
Q−1 = − 1(2) ⊗ 1(2) + 1 . (2.37)
2w(3m + 2w) 2w
In component form,
Qijk` = mδij δk` + w(δik δj` + δi` δjk ) (2.38)
m 1
Q−1
ijk` = − 2w(3m + 2w) δij δk` + 4w (δik δj` + δi` δjk ) (2.39)
A more straightforward approach to invert an isotropic tensor is to adopt the

following E-basis orthogonal decomposition. Let
1 (2) (1) 1
E(1) := 1 ⊗ 1(2) , Eijk` = δij δk` (2.40)
3 3
1 (2)
E(2) (2)
:= − 1 ⊗ 1 + 1 (4s)
3
(2) 1 1
⇒ Eijk` = − δij δk` + (δik δj` + δi` δjk ) (2.41)
3 2
The E-bases have the following special properties,
E(1) + E(2) = 1(4s)

E(1) : E(1) = E(1) , and E(2) : E(2) = E(2)
E(1) : E(2) = E(2) : E(1) = 0 .
We now use E-basis approach to verify Sherman-Morrison formula. Let,
Q = (3m + 2w)E(1) + 2wE(2) (2.42)
and
Q−1 = hE(1) + vE(2) (2.43)
By definition,
Q : Q−1 = 1(4s) = E(1) + E(2)

(3m + 2w)hE(1) + 2wvE(2) = E(1) + E(2)
which then leads to

1
h = (2.44)
3m + 2w
1
v = (2.45)
2w
Consequently, we can write that
Q−1 = (h − v)E(1) + v(E(1) + E(2) )

3m 1 (4s)
= − E(1) + 1
2w(3m + 2w) 2w
m 1 (4s)
= − 1(2) ⊗ 1(2) + 1
2w(3m + 2w) 2w
Let’s practice more examples.
Example 2.1 Consider an isotropic elastic tensor,
C = λ1(2) ⊗ 1(2) + 2µ1(4s)

= 3KE(1) + 2µE(2)
Since by definition, C : D = 1(4s) , it can be readily shown that

1 (1) 1 (2)
D = E + E
3K 2µ
λ 1 (4s)
= − 1(2) ⊗ 1(2) + 1
2µ(3λ + 2µ) 2µ
Preliminary 11
Example 2.2 For spherical inclusion, the Eshelby tensor is

5ν − 1 (2) 2(4 − 5ν) (4s)
SΩ = 1 ⊗ 1(2) + 1
15(1 − ν) 15(1 − ν)
(1 + ν) (1) 2(4 − 5ν) (2)
= E + E
3(1 − ν) 15(1 − ν)
= s1 E(1) + s2 E(2)
1+ν 2(4 − 5ν)
where s1 = and s2 = .
3(1 − ν) 15(1 − ν)
Then
3(1 − ν) (1) 15(1 − ν) (2)
(SΩ )−1 = E + E
1+ν 2(4 − 5ν)
(1 − ν)(3 − 5ν) (2) 15(1 − ν) (4s)
= 1 ⊗ 1(2) + 1
2(1 + ν)(4 − 5ν) 2(4 − 5ν)
Moreover,
TΩ = 1(4s) − C : SΩ : D
= (E(1) + E(2) ) − (3KE(1) + 2µE(2) ) : (s1 E(1) + s2 E(2) )
1 1 (2)
: E(1) + E
3K 2µ
= (1 − s1 )E(1) + (1 − s2 )E(2)
2.1.4 Tensor analysis

Define gradient operator as
∂
∇= ei (2.46)
∂xi
It is a vector operation.
Applying gradient operator to a scalar function, f ∈ C 0 (Ω), Ω ⊂ IRd , will
result a vector. In other words, the gradient of a scalar function (zero-th order
tensor) is a first order tensor, i.e.
∂ ∂f
grad f := ∇f = ei f = ei (2.47)
∂xi ∂xi
For a vector function, A(x) = Ai (x)ei , its gradient is a tensor product
between the gradient operator and the vector field,
∂ ∂Aj
grad A := ∇ ⊗ A = e i ⊗ Aj e j = ei ⊗ ej (2.48)
∂xi ∂xi
The gradient of a vector field, a first order tensor field, is a second order tensor.
In general, the gradient operation increases the ordero f a tensorial field up to
one order higher.
On the other hand, the scalar product or contraction between a gradient op-
erator and a tensorial field is called divergence operation, which will result a
new tensorial field with reduced order. Consider a vector field, A = Ai ei . Its
divergence is being defined as
∂ ∂A ∂Ai
j
divA := ∇ · A = ei · Aj ej = (ei · ej ) = (2.49)
∂xi ∂xi ∂xi
The cross product between the gradient operator and a tensorial field. A =
Ai ei , is called the Curls or rot of the tensorial field.
∂Aj
CurlA := ∇ × A = ei × ej = eijk ∂i Aj ek = eijk ∂j Ak ek (2.50)
∂xi
In what follows, a few integral transformations are listed.
Suppose that there is a continuous function, f (x) ∈ C 1 (Ω), defined in a
domain Ω ∈ IRd with smooth boundary ∂Ω. A well-known integral theorem is
Z Z
∇f dΩ = f ndS (2.51)
Ω ∂Ω
or in component form
Z Z
∂f
dΩ = f ni dS (2.52)
Ω ∂xi ∂Ω
In general for a smooth tensorial field, A, we have the following statement,

Z Z
∇ ⊗ AdΩ = n ⊗ AdS (2.53)
Ω ∂Ω
Consider a continuous m-order tensorial field, A(x) ∈ [C 1 (Ω)]m × d, the

well known divergence theorem can be expressed in a Cartesian coordinate as
Z Z
∇ · AdΩ = n · AdS (2.54)
Ω ∂Ω
If A is a vector field, i.e. A = Ai ei , the divergence theorem can be expressed

in a component form as
Z Z
∂Ai
dΩ = ni Ai dS (2.55)
Ω ∂xi ∂Ω
Preliminary 13
If we consider the volume integration of a cross product between gradient

operator and the tensorial field, we can have the following integral transforma-
tion, Z Z
∇ × AdΩ = n × AdS (2.56)
Ω ∂Ω
Again, if A is a vector field, we may write its Cartesian component form,
Z Z
∂Ak
eijk dΩ = eijk nj Ak dS (2.57)
Ω ∂xj ∂Ω
2.2 Review of Linear Elasticity Theory

To set the stage, we first review the basic formulations of infinitesimal, linear
elasticity theory.
• Equations of motion
Denote σ = σij ei ⊗ ej as Cauchy stress tensor, and u = ui ei as the in-
finitesimal displacement field, ρ as the density of the continuum, and b = bi ei
as the body force per unity volume. The equation of motion of a material
particle can be expressed in a Cartesian coorinate as ∀x ∈ Ω,
∂2u
∇ · σ + ρb = ρ (2.58)
∂t2
For convenience, we often write the component form
∂ 2 ui
σji,j + ρbi = (2.59)
∂t2
∂uji
where uji,j = .
∂xj
• Geometric relation
The infinitesimal strain field = ij ei ⊗ ej is defined as
1
= ∇ ⊗ u + (∇ ⊗ u)T (2.60)
2
Note that ∇ ⊗ u = uj,i ei ⊗ ej . Hence (∇ ⊗ u)T = ui,j ei ⊗ ej .
Therefore in component form,
1
ij = (ui,j + uj,i ) (2.61)
2
• Constitutive equations
For linear elastic solids, the constitutive equations have the following form,
σ = C : ⇒ σij = Cijkl kl (2.62)

where C = Cijkl ei ⊗ ej ⊗ ek ⊗ el is the elasticity tensor.

For isotropic elastic media, it has the form,
C = λI ⊗ I + 2µ1(4s) (2.63)
where λ, µ are Lame constants. In component form, it reads
Cijkl = λδij δkl + µ(δik δjl + δil δjk ) (2.64)
Inversely, one may write that
= C−1 : σ = D : σ ij = Dijkl σkl (2.65)
where the fourth order tensor, D, is called compliance tensor. For isotropic
materials, it has the form
λ 1
Dijkl = − δij δkl + (δik δjl + δil δjk ) (2.66)
2µ(3λ + 2µ) 4µ
• Compatibility condition
Compatibility conditions for infinitesimal deformation field may be expressed
as (Melvan [1969]),
∇××∇=0 (2.67)
In indicial natation, it reads,
epki eqlj ij,kl = 0 (2.68)
or alternatively
ij,kl + kl,ij − ik,jl − il,jk = 0 (2.69)
• Elastic potential energy
The strain energy density is defined as
Z
0 0
U () = σ( ) : d (2.70)
0
Based on foundamental theorem of calculus, one may find its inverse relation-
ship as
∂U ∂U
= σ, = σij (2.71)
∂ ∂ij
The complementary strain energy density can be obtained via Legendre
transform,
U ∗ (σ) = σ : − U () (2.72)
Or one may define Z σ
0 0
U ∗ (σ) = (σ )dσ (2.73)
0
Preliminary 15
One may derive that

∂U ∗ ∂U ∗
= , or ij = (2.74)
∂σ ∂σij
For linear elastic materials,
∂U ∂2U
Cijkl kl = ⇒ Cijkl = (2.75)
∂ij ∂ij ∂kl
In general, for hyperelastic media, the elastic stiffness tensor can be calculated
based on the formula
∂2U
Cijkl = (2.76)
∂ij ∂kl
Similarly, one may find elastic compliance tensor by calculation
∂2U ∗
Dijkl = (2.77)
∂σij ∂σkl
Change the order of differentiation in Eq.(2.66),
∂2U ∂2U
= (2.78)
∂ij ∂kl ∂kl ∂ij
One may derive that Cijkl = Cklij .
Furthermore since ij = ji and kl = lk , Cijkl = Cjikl = Cijlk = Cjlik .
These are called minor symmetry.
Similar conclusions can be drawn from elastic compliance tensors as well.
Both elastic tensor C and compliance tensor D are positive definite, because
both strain energy density and complementary strain energy density must be
positive, i.e.
1 1
U () = : C : = Cijkl ij kl > 0
2 2
1 1
U ∗ (σ) = σ : D : σ = Dijkl σij σkl > 0
2 2
By definition that a fourth-order tensor, Cijkl , is positive-definite, when
1
Cijkl ij kl > 0, ∀ij (2.79)
2
where equality holds only if ij = 0.
2.2.1 Betti’s reciprocal theorem and Somigliana Identity

Consider two sets of different self-equilibrating states: u(α) , (α) , σ (α) , f (α) ,

α = 1, 2,
∇ · σ (α) + f (α) = 0 (2.80)
Figure 2.2. Two sets of different self-equilibrating states
with boundary conditions,

n · σ(α) = t(α)0 , ∀x ∈ Γ0t (2.81)
(α) (α)0
u = u , ∀x ∈ Γ0u , α = 1, 2 (2.82)
acting in a same object Ω0 .
The Betti’s reciprocal theorem2 states that: the work done by the first set
of self-equilibrating surface traction, t(1) , and body force f (1) in any interior
region Ω ⊂ Ω0 , going through the displacement field, u(2) , of the second
self-equilibrating system, equals the work done by the second set of tractions,
t(2) , and the body force, f (2) , in the same interior region going through the
displacement field, u(1) , of the first self-equilibrating system, i.e.
Z Z Z Z
(1) (2) (1) (2) (2) (1) (2) (1)
fi ui dΩ + ti ui dS = fi ui dΩ + ti ui dS (2.83)
Ω ∂Ω Ω ∂Ω
Proof:
Consider both states being equilibrium states. It has
Z Z
(1) (2) (1) (2)
fi ui dΩ = − σji,j ui dΩ
Ω ZΩ Z
(1) (2) (1) (2)
= − σji nj ui dS + σji ui,j dΩ
Z∂Ω Z Ω
(1) (2) (1) (2)
= − ti ui dS + σji ji dΩ (2.84)
∂Ω Ω
Moving the first term of the right-hand side of (2.74) to the left-hand side yields
Z Z Z
(1) (2) (1) (2) (1) (2)
fi ui dΩ + ti ui dS = σij ij dΩ (2.85)
Ω ∂Ω Ω
2 Precisely speaking, it is the Betti’s second reciprocal theorem.

Preliminary 17
Similarly, one may show that

Z Z Z
(2) (1) (2) (1) (2) (1)
fi ui dΩ + ti ui dS = σij ij dΩ (2.86)
Ω ∂Ω Ω
Consider the fact that the two systems exist in the same material
Z Z Z Z
(1) (2) (1) (2) (1) (2) (1) (2)
σij ij dΩ = Cijkl kl ij dΩ = Cklij kl ij dΩ = kl σkl dΩ
Ω Ω Ω Ω
Compare the both sides of (2.75) and (2.76), the theorem holds.
In addition, the equality
Z Z
(1) (2) (2) (1)
σij ij dΩ = σij ij dΩ (2.87)
Ω Ω
is called Betti’s first reciprocal theorem.

To derive Somigliana identity, we first consider Dirac’s delta function, which
is the limit of the following function, δ(x) = lim→0 δ (x),

 0; x < −/2
δ (x) = lim 1/; −/2 < x < /2 (2.88)
→0 
0; x > /2
A graph of Dirac’s delta function is shown in Fig. 2.3.
Dirac delta function has following properties
Z ∞
(1) δ(x)dx = 1 (2.89)
−∞
Z ∞
(2) δ(x − y)f (y)dy = f (x) (2.90)
−∞
The first property (2.79) can be easily shown by definition that

Z ∞ Z /2
1
δ(x)dx = dx = 1 (2.91)
−∞ −/2
To show the second property, we let x − y = z and dy = −dz. Thus

Z ∞ Z −∞ Z ∞
δ(x − y)f (y)dy = − δ(z)f (x − z)dz = δ(z)f (x − z)dz
−∞ ∞ −∞
Z /2 Z /2
1 1
= f (x − z)dz = f (x − ζ ) dz
−/2 2 −/2
= f (x), as → 0 (2.92)
where −1 < ζ < 1.

Figure 2.3. Dirac’s delta function
Consider an infinitely space filled with homogeneous elastic medium. The

body force is form of concentrated load at a fixed point y,
f = δ(x − y)δmk ek (2.93)
The subscript index m is in the direction of m.
The equilibrium equations then have the form,
∇ · σ m + δ(x − y)δmk ek = 0, ∀x ∈ IR3 (2.94)
The displacement solution of this problem is called foundamental solution
of Navier equation, or the Green’s function for an infinitely extended homoge-
neous elastic domain. Denote the displacement solution as
um = G∞ ∞
m (x, y) = Gmi (x, y)ei (2.95)
The corresponding strain and stress fields are:
G∞ 1 ∞
G∞ G∞
ijm = Gmi,j + G∞ mj,i , σij
m
= Cijkl ijm (2.96)
2
Next, we consider a singly connected finite region Ω ⊂ IR3 . The finite
region Ω is in a self-equilibrating state, i.e., there is a body force distribution:
∇·σ +f = 0, ∀x ∈ Ω, and a traction force distribution: t = n·σ, ∀x ∈ ∂Ω.
Let
f (1) (x) = δ(x − y)δmk ek , u(1) (x) = G∞
mi (x, y)ei (2.97)
G∞
t(1) (x) = σij m (x)nj ei (2.98)
f (2) (x) = fi (x)ei , u(2) (x) = ui (x)ei (2.99)
(2)
t (x) = σij (x)nj ei (2.100)
Preliminary 19
Apply Betti’s reciprocal theorem,

Z Z
G∞
δ(x − y)δmi ui (x)dΩx + nj σjim ui (x)dSx
ZΩ Z ∂Ω
= fi (x)G∞mi (x, y)dΩx + nj σji G∞
mi (x, y)dSx (2.101)
Ω ∂Ω
Considering the property of Dirac delta function, one can obtain:

Z Z
um (y) = fi (x)G∞mi (x, y)dΩ x + ti (x)G∞
mi (x, y)dSx(2.102)
ΩZ ∂Ω
G∞
− ti m (x, y)ui (x)dSx , m = 1, 2, 3
∂Ω
Equation (2.92) is the well-known Somigliana identity.
2.3 Exercises
Probelm 2.1 Let δu be a virtual displacement field and σ be a self-equilibrium
stress field. Show

∇ · σ · δu = ∇ · σ · δu − σ : (∇ ⊗ δu) (2.103)
Probelm 2.2 Assume body force f = 0. The elastostatic equilibrium equa-

tion takes the form:
σji,j = 0, or ∇ · σ = 0 (2.104)
Show Z Z
σ : dΩ = t · udS (2.105)
Ω ∂Ω
where t = n · σ.
(Hint: use Gauss theorem, the divergence theorem.)
Probelm 2.3 Suppose that there are two different solutions of equilibrium
equation,
∇ · σ 1 = 0, ∇ · σ 2 = 0 (2.106)
which satisfy the same boundary conditions,
u1 = u0 ,

∀x ∈ Γu (2.107)
u2 = u0 ;
n · σ 1 = t0 ,

∀x ∈ Γt (2.108)
n · σ 2 = t0 ;
S
where Γu Γt = ∂Ω.
By using the positive-definiteness of elastic tensor and compliance tensor,

show:
∆σ = σ 1 − σ 2 = 0 (2.109)
∆ = 1 − 2 = 0 (2.110)
Probelm 2.4 Show that for a given second-order tensor, A,
1(4) : A → A (2.111)
1
1(4s) : A → A + AT (2.112)
2
1
1(4a) : A → A − AT (2.113)
2
Homogenization I — Classical Averaging Method 21
Chapter 3
HOMOGENIZATION I — CLASSICAL AVERAGING

METHOD
"Curiouser and curiouser!" cried Alice,"Now I’m opening out like the largest
telecope that everwas!"
— Lewis Carroll, Alice in Wonderland
3.1 Representative volume element

One of the foundamental concept in classical micromechanics is the so-
called Representative volume element, or RVE.
The classical micromechanics paradigm is a two-level hierarchical mechani-
cal structure: Macro-level and Micro-level, or it consists of two elements: macr
o-element and micro-element. At macro-level, a continuum is made of many
material points, and each material point is related with a micro-space. A macro
material point is also called a macro-element, or volume element. Its associ-
ated micro-space contains many micro-elements. In fact, it is a microscopic
continuum. If a material is statistically homogeneous at macro-level, to study
material behaviors, we only need to examine material properties at an arbitrary
(typical) macro-point, and the micro-space associated with that macro-point is
called the representative volume element.
An RVE for a material point of a continuum mass is a statistical ensemble of
microscale objects surrounding or constituting the macro material point. This
means that an RVE should contain a very large number of micro-elements such
that it can be a statistically representative of the local continuum properties, or
it is statistically stable.
In essence, the concept of representative volume element in classical mi-
cromechanics is a mathematical paradigm. It has no fixed length scale associ-
ated with each level.
The length scales associated macro-level and microlevel are relative. If you
study effective material properties of a heterogeous metal, the lengthscale of
microlevel maybe from a few nm to µm, and the lengthscale of macrolevel

may be from a few mm to centimeter. If you study the stiffness of a dam, the
lengthscale of microlevel could be from centimeters, whereas the lengthscale
of macro-level could be meters.
In classical mechanics, at macro-level, the material properties are always
assumed to be homogeneous but unknown, whereas at micro-level, i.e., inside
the RVE, the material properties are heterogeneous but known.
At microlevel, the heterogeneous micro-structure is known and physical
laws is known. The task of micromechanics is based on information of mi-
crostructure to find homogeneous material properties at macro-level, which is
often called overall material properties or effective material properties.
The methodology to find effective material properties is called homogeniza-
tion. Homogenization is another word that has been widely used in many
different contexts. In this book, the term "homogenization" is used to mean
statistical averaging. There are mainly two sets of homogenization methods,
mathematical homogenization and mechanical homogenization.
The objectives of micromechanics is to find both material properties at macro-
level, or overall (effective) material properties and physical laws at macro-
level.
The first subject of continuum micromechanics if micro-elasticity. The ba-
sic premises of microelasticity is to assume that inside an RVE, the micro-
constitutive relation of a material is elastic, and in more cases, they are as-
sumed to be linear elastic. In micromechanics, the concept of the RVE is used
to derive material properties due to microstructures. In most cases, the micro-
structures are often independent with gravity or other types of body forces.
Therefore, in micro-continuum mechanics, the body force effect is often ne-
gleted. The equilibrium equations inside an RVE is often written as
∇ · σ = 0 ⇒ σij,j = 0. (3.1)
3.2 Average stress in an RVE

Definition of average operator < · >. Suppose that T(x,X) is a general
tensor field defined in an RVE. Note that here x is the spatial coordinate inside
an RVE for a fixed material point, whereas X is the spatial coordinate of the
material point with respect to a macro-coordinate. If at macro-level, material is
homogeneous, i.e. material properties at macro-level do no change from place
to place, X is often dropped out. We simply write T = T(x), which means that
one RVE is sufficient to represent all the material points in the object that is
under investigation.
To associate a micro-level tensor field with a tensorial quantity at macro-
level is called homogenization. To do so, we first define the so-called average
operator. The average value of the tensor field T(x) at a material point is de-
fined as Z
1
< T >X := T(x, X)dVx (3.2)
V V
If the material is homogeneous at macro-level, we have
Z
1
< T >:= T(x)dVx (3.3)
V V
For instance, if T = σ(x) is a micro-stress field, the macro-stress at a
material point will be Σ =< σ >. Similarly, if T = (x) is a micro-strain
field, the macro-strain at a material point is E =< >.
A very useful average theorem about micro-Cauchy stress tensor may be
stated as follows:
Theorem 3.1 Suppose an RVE is subjected to natural boundary condition,
and the traction on remote boundary of an RVE (∂V ) is generated by a constant
stress tensor, σ 0 . Then the average stress at this material point, or the macro
stress at the material point,
Σ =< σ >= σ 0 (3.4)
Note that the point here is that one only knows the traction distribution on the
remote boundary of the RVE, but one does not know the exact stress distribu-
tion inside the RVE.
Proof
Consider,
∂xi
= δij and σji,j = 0 (3.5)
∂xj
One then can express Cauchy stess inside an RVE as
∂x
j
σij = σik δkj = σik δjk = σik
∂xk
= (σik xj ),k − σik,k xj = (σik xj ),k (3.6)
Therefore,
Z Z
1 1
< σij > = σij dV = σik xj dV
V V V V ,k
I I
1 1
= σik xj nk dS = σ 0 xj nk dS
V ∂V V ∂V ik
0 I
σik 0 Z
σik ∂xj
= xj nk dS = dV
V ∂V V V ∂xk
0 Z
σik σ0 0
= δjk dV = ik δjk V = σij (3.7)
V V V
3.3 Average strain and strain rate

Consider a displacement field, u = ui ei , inside an RVE. Suppose that on
the remote boundary of the RVE, the displacement filed is prescribed,
ui (x) = u0i (x), ∀x ∈ ∂V (3.8)
One can find the average displacement gradient field in terms of boundary data,
i.e., Z Z
1 1
< ui,j >= ui,j dV = nj u0i dS (3.9)
V V V ∂V
Note that you don’t know exact distribution of the displacement field inside the
RVE.
Moreover, one may find the average strain and rotation fields in terms of
boundary displacement data,
I
1 1
< ij >= < ui,j > + < uj,i > = (nj u0i + ni u0j )dS
2 2V ∂V
I
1 1
< ωij >= < ui,j > − < uj,i > = (nj u0i − ni u0j )dS
2 2V ∂V
Remark 3.3.1 in general, the average displacement fields of an RVE can
not be expressed in terms of remote surface data. To see this, one may evaluate
the average displacement field. Using the trick,
∂xi
ui = uk δki = uk δik = uk = (uk xi ),k − uk,k xi
∂xk
Hence
Z Z
1 1
< ui > = ui dV = (uk xi ),k − uk,k xi dV
V V V
I Z
1
0

= uk xi nk dS − uk,k xi dV (3.10)
V ∂V V
It is clear that < ui > can not be expressed in terms of boundary data, unless
uk,k = 0.
However, for incompressible materials, such as rubber or plastic zone of
ductile materials, it is often true that uk,k = 0. Therefore,
Z I
1 1
< ui >= ui dV = u0 xi nk dS (3.11)
V V V ∂V k
An average theorem for infinitesimal strain can be stated as follows.
Theorem 3.2 Suppose that an RVE is only subjected to essential bound-

ary condition. On the remote surface of the RVE, its displacement fields are
prescribed as
u0 = 0 · x, ⇒ u0i = 0ij xj (3.12)
where 0ij is a constant strain tensor. Then, the average strain field of the RVE
equals the constant strain tensor, i.e.
< >= 0 , ⇒ < ij >= 0ij (3.13)
Proof:
First of all, the prescribed essential boundary condition does not necessarily
generate a constant strain field inside the RVE, i.e.
ij (x) 6= 0ij
In fact
ij (x) = 0ij + ˜ij (x), ∀x ∈ V
and the perturbation strain field satisfying ˜ij (x) = 0, ∀x ∈ ∂V .

By definition,
Z Z
1 1
< ij > = ij dV = ui,j + uj,i dV
V V 2V V
I
1
= (u0 nj + u0j ni )dS
2V ∂V i
I
1
= (xk 0ki nj + xk 0kj ni )dS
2V ∂V
I
1
= (0 δkj V + 0kj δki V ) = 0ij (3.14)
2V ∂V ki
One may also show the following identities about average virtual work and
average strain energy density.
I
1
< σ : δ >= t · δudS (3.15)
V ∂V
<σI : > − < σ >:< >

1
= u − x· < ∇ ⊗ u > · n · (σ− < σ >) dS (3.16)
V ∂V
Since σij δij = 12 σij (δui,j + δuj,i ) = σij δui,j ,
Z Z
1 1
σij δij dV = σij δui,j dV
V V V V
Z I
1 1
= σij δui dV = σij δui nj dS
V V ,j V ∂V
I
1
= ti δui dS (3.17)
V ∂V
where ti := nj σji . Hence,(3.15)holds.

To show (3.16), one may write
Z
1
ui − xj < ui,j > nk (σki − < σki > dS
V ∂V
Z
1
= ui nk σki − ui nk < σki > −xj < ui,j > nk σki
V ∂V

+xj < ui,j > nk < σki > dS
Z 1 Z
1
= σki ui,k dV − ui,k dV < σki >
V V V V
Z
1
−δjk < ui,j > σki dV + < ij >< σij >
V V
= < σij ij > − < σij >< ij > (3.18)
3.4 Definition of eigenstrain, eigenstress, and inclusion

’Eigenstrain’ is a generic name to describe a transformation strain field that
can equivalently represent induced strain due to misfit of inhomogeneities,
thermal expansion, plastic strain, residual strain , phase transformation, etc.,
all of which, when homogeneously applied produce a compatible deformation
field without generating stresses. The German word "eigen" means character-
istic. It is believed that any strain field generated by an inhomogeneity distri-
bution may have a one-to-one correspondence to a fictitious eigenstrain field,
which is characteristically equivalent (in the sense of mechanical variables,
such as stress, strain, and displacements) to the induced strain field generated
by the inhomogeneity distribution.
’Eigenstress’ is a generic name given to self-equilibrated transformation
stress (internal) field that can generate equivalent perturbed stress and strain
distributions caused by one or several of there eigenstrains in bodies which are
free from any other external forces and surface constraints. The eigenstress
field is created by the incompatibility of the eigenstrains.
Figure 3.1. Illustration of Eshelby’s equivalent eigenstrain principle. (a)Initial heterogeneous

body, (b) equivalent homogeneous body (V = Ω ∪ M ).
The term inclusion denotes a subdomain in the matrix subjected to trans-

formation strains (eigenstrains), while the inhomogeneity is a subdomain with
properties distinct from those from the matrix.
3.5 Eshelby’s equivalent eigenstrain method I: Traction

boundary condition
Eshelby’s equivalent eigenstrain principle is a homogenization method. It
establishes the equivalency between an eigenstrain (eitenstress) field and an
inhomogeneity distribution, such that distribution of inhomogeneities may be
replaced by the eigenstrain field with the equivalent mechanical effect. This
equivalency mapping process translates the heterogeneity of material into an
added non-uniform strain distribution, while making the material properties
become homogeneous again.
Let’s consider an Elastic solid, V, with elasticity tensor, C, and compliance
tensor, D. Inside the elastic solid, there is an inhomogeneity, a subdomain, Ω,
with different elastic constants, CΩ and DΩ (see Fig. 3.1).
The so-called Eshelby’s equivalent eigenstrain principle, or Mura’s equiva-
lent eigenstrain principle, is to replace the inhomogeneity with a homogenized
inclusion, within which an eigenstrain field is prescribed, such that the homog-
enized field is mechanical equivalent to the original inhomogeneous field.
Consider that the original inhomogeneous solid is subjected to a traction
boundary condition, t = n · σ 0 . The presence of inhomogeneity will produce
stress perturbation and hence the strain field perturbation,
σ = σ0 + σd, = 0 + d .
The stress and strain distributions inside the inhomogeneous solid are
C : (0 + d )

x∈M
σ =
CΩ : (0 + d ) x ∈ Ω
D : (σ 0 + σ d )

x∈M
= Ω 0 d (3.19)
D : (σ + σ ) x ∈ Ω
The Eshelby’s equivalent eigenstrain homogenization method is to choose a
suitable strain field,
0, ∀x ∈ M
= (3.20)
∗ , ∀x ∈ Ω
to superpose with the actual strain field, = 0 + d , such that the total strain
field of homogenized solid is equivalent to the total strain field of inhomoge-
neous solid, i.e.
σ(x) = C : ((x) − ∗ (x))
C : (0 + d ) C : (0 + d ),

x∈M
= 0 d ∗ = (3.21)
C : ( + − ) CΩ : (0 + d ), x ∈ Ω
Consider 0 = D : σ 0 . Under the chosen traction boundary condition, <
σ >= σ 0 , but 0 6=< >.
From (3.21), one may derive that
σ d (x) = C : (d (x) − ∗ (x)), ∀x ∈ V (3.22)
CΩ (0 + d ) = C : (0 + d − ∗ ), ∀x ∈ Ω (3.23)
where Eq.(3.23) is called "stress consistency condition". It is the criterion for
choosing suitable eigenstrain field. Note that 0 + d − ∗ is the total elastic
strain.
Alternatively, Eqs (3.21) and (3.22) can be recast into following forms,
σ = C : ( − ∗ ) ⇒ = D : σ + ∗ (3.24)
σ d = C : (d − ∗ ) ⇒ d = D : σ d + ∗ (3.25)
3.6 Eshelby’s equivalent eigenstress method II:

Displacement boundary condition
Consider the same inhomogeneous solid and following displacement bound-
ary condition
u(x) = 0 · x, ∀x ∈ ∂V (3.26)
The inhomogeneity inside the solid will generate a disturbance stress field,
σ. The total stress field is
D : (σ 0 + σ d )

(x) = (3.27)
DΩ : (σ 0 + σ d )
Figure 3.2. Illustration of Eshelby’s equivalent eigenstress principle. (a) Initial heterogeneous
body, (b)equivalent homogeneous body (V = Ω ∪ M ).
As proved in previous section, under prescribed boundary condition, the aver-

age strain, < >= 0 . On the other hand,< σ >6= σ 0 .
To homogenize the heterogeneous medium, we introduce the following eigen-
stress distribution,
∗ 0, ∀x ∈ M
σ (x) = (3.28)
σ∗, ∀x ∈ Ω
such that
D : (σ 0 + σ d ) D : (σ 0 + σ d ),

x∈M
(x) = = (3.29)
D : (σ 0 + σ d − σ ∗ ) DΩ : (σ 0 + σ d ), x∈Ω
From Eq.(3.29), we can derive that
d (x) = D : (σ d (x) − σ ∗ ), ∀x ∈ V (3.30)

DΩ (σ 0 + σ d ) = D : (σ 0 + σ d − σ ∗ ), ∀x ∈ Ω (3.31)
where Eq.(3.31) is called "strain consistency condition."

Alternatively,
d (x) = D : (σ d (x) − σ ∗ ) ⇒ σ d = C : d + σ ∗ (3.32)
Comparing Eq.(3.32) with (3.25) yield the following identities,
∗ + D : σ ∗ = 0, or σ ∗ + C : ∗ = 0 (3.33)
3.7 Effective material properties via eigenstrain method

In this section, we illustrate how to use equivalent eigenstrain method to
find overal material properties.
Figure 3.3. Illustration of Eshelby’s equivalent eigenstrain principle
We still consider the previous problem: an RVE with only on inhomogene-

ity. Denote the total volume of RVE as V, the volume of the matrix as M, and
the volume of the inhomogeneity as Ω. Assume that the RVE is a heteroge-
neous linear elastic medium and the micro-constitutive relations are:
= D : σ, x ∈ M (3.34)
= DΩ : σ, x ∈ Ω (3.35)
Our objective is to find the constitutive relation at macro-level,i.e.
Σ = C̄ : E ⇒ < σ >= C̄ :< > (3.36)
Note that here we have already assumed that the constitutive relation at macro-
level is also linear elastic. The only unknown is the effective compliance
tensor, or effective elastic tensor. This shows the primitive feature of clas-
sical micro-elasticity. In contemporary micromechanics, one does not know
whether the material behaviors at macro-level is linear elastic or some other
forms. One determines macro behaviors of the material as an outcome of ho-
mogenization.
Apply the traction boundary condition on the remote surface of the RVE,
t = n · σ0
As mentioned before, under such boundary condition, < σ >= σ 0 , neverthe-
less, < >6= 0 , i.e. < 0 + d >6= 0 . Therefore, our goal is to find the
effective elastic compliance tensor such that < >= D̄ : σ 0
Denote the average strain and stress in the matrix and in the inhomogeneity
as
Z Z
1 1
¯M := (x)dV , σ̄ M := σ(x)dV ; (3.37)
M M M M
Z Z
1 1
¯Ω := Ω
(x)dV , σ̄ := σ(x)dV ; (3.38)
Ω Ω Ω Ω
Therefore, ¯M = D : σ̄, and¯Ω = DΩ : σ̄ Ω .

Ω
Consider V = M ∪ Ω and let f := .Then
V
Z Z
1 1
¯ = dV = dV
V V V M ∪Ω
Z Z
1 M Ω M Ω
= dV + dV = ¯M + ¯Ω (3.39)
V M M Ω Ω V V
Hence,
M M
¯ = < > −f¯Ω
V
= D̄ : σ 0 − f DΩ : σ̄ Ω (3.40)
On the other hand,

M M M M 1 Z
¯ = D : σ̄ M = D : σ(x)dV
V V V M M
1 Z
= D: σ(x)dV
V V −Ω
1 Z 1
Z
= D: σ(x)dV − σ(x)dV
V V V Ω

0 Ω
= D : σ − f σ̄ (3.41)
Compare Eqs. (3.40) and (3.41),
D̄ : σ 0 − f DΩ : σ̄ Ω = D : σ 0 − f D : σ̄ Ω (3.42)
Therefore,

D − D̄ : σ 0 = f D − DΩ : σ̄ Ω = f D − DΩ :< σ 0 + σ d >Ω (3.43)
The equaqtion is often referred to as The Basic Equation for Average Stress.
By definition,
σ̄ Ω = CΩ :< 0 + d >Ω (3.44)
From the stress consistency condition, one may obtain

−1
∗ = C−1 : (C − CΩ ) : (0 + d ) = AΩ : (0 + d ) (3.45)
where AΩ = (C − CΩ )−1 : C.
If one can relate the perturbed strain with the eigenstrain, i.e.
d = SΩ : ∗ (3.46)
Eq (3.45) may be rewritten as

−1
0 + d = AΩ : ∗ ⇒ ∗ = AΩ − SΩ : 0 (3.47)
Subsequently,
(x) = 0 + d = AΩ : ∗ = AΩ : (AΩ − SΩ )−1 : 0

= AΩ : (AΩ − SΩ )−1 : D : σ 0 , ∀x ∈ Ω (3.48)
In the literature, we denote AΩ = AΩ : (AΩ −SΩ )−1 as the so-called “concen-

tration tensor”, because it represents the relation ship between the background
strain field and the actual strain field in the inhomogeneity, i.e. how are the
strains concentrated. Suppose both the Eshelby tensor SΩ and tensor AΩ are
constant tensors, then AΩ = const., and
(x) = AΩ : 0 , ∀x ∈ Ω ⇒ ¯Ω = AΩ : 0 (3.49)
Therefore,
σ̄ Ω = CΩ : AΩ : (AΩ − SΩ )−1 : D : σ 0 , ∀x ∈ Ω (3.50)
Substituting the expression (3.50) into (3.43) yields

D̄ − D : σ 0 = f (DΩ − D) : CΩ : AΩ : (AΩ − SΩ )−1 : D : σ 0 (3.51)
Consider
DΩ − D : CΩ = 1(4s) − D : CΩ
and
−1 −1
AΩ = (C − CΩ )−1 : C = C−1 : (C − CΩ )
= 1(4s) − C−1 : CΩ
= 1(4s) − D : CΩ = DΩ − D : CΩ
−1
⇒ DΩ − D : CΩ = AΩ (3.52)
Therefore −1
D̄ − D : σ 0 = f AΩ − SΩ : D : σ0 (3.53)
It is the straightforward to derive

D̄ = 1(4s) + f (AΩ − SΩ )−1 : D
(3.54)
Note that the crucial step of this derivation is the assumption that disturbance
strain field can be related to eigenstrain distribution, i.e. d = SΩ : ∗ , where
the tensor SΩ is called the Eshelby tensor. Chapter 6 will be devoted to derive
Eshelby tensor for specific shapes of inhomogeneities or inclusions.
3.8 Jock Eshelby (I)

John Douglas Eshelby was born in Puddington, Cheshire, On December 21,
1916, the eldest son of Alan Douglas Eshelby. Because of ill health he missed
his formal schooling from the age 13 and ilved at the family home in north
Somerset, where he learned instead from tutors. So, as he used to say, he had
to work many things our for himself, and perhaps this helped to make him
such an original and creative thinker. Ovservant of people and things, he had
a deep physical insight into the workings of nature around him. As a child,
watching his father’s diesel generator, he noticed how a moving belt ratains its
shape when struck; and recently he was to be seen studying the spider’s web
pattern of cracks in broken windows, while he pondered on the limitations of
the present theory of elastic plates.
Through a contact with Professor Mott (now Sir Nevill) he went early to the
University of Bristol and obtained a first in physics there in 1937. During the
second World War he served first at the Admiralty, degaussing ships, and then
in the technical branch of the Royal Air Force, where he reached the rank of
squadron leader. He flew sometimes in Sunderlands out of Pembroke Dock,
and there is in the Science Museum some radar equipment that he helped to
design.
He returned to Bristol in 1946, at an exciting time for solid state physics
when rapid advances were made in the theory of the deformation of crystals.
The opportunity arose for him to take up theoretical research, and here he made
his initial mark in dislocation theory, revealing quite suddently to those around
him a mastery of some of the most difficult problems of the time. he obtained
his Ph.D. in 1950 and two years later spent a year at the University of Illinois.
There followed some ten years at the University of Birmingham, a period
in 1963 as visiting prefessor at the Technische Hochschule, Sturgar, and then
two years at Cambridge, where he became a Fellow and College Lecture at
Churchhill. In 1966 he went to the University of Sheffield, holding a readership
and, from 1971, a personal chair in the theory of materials.
Figure 3.4. Illustration of Eshelby’s equivalent eigenstrain principle
His work was a great part of his life. His general field was the theoretical
physics of the deformation, strength and fracture of engineering materials, and
his principal interests were lattice defects and continuum mechanics.
Though motivated by the desire to understand he kept a firm eye on appli-
cation and had no time for useless erudition, like willard Gibbs his object was
to make things appear simple by "looking at them in the right way". With
a keen discrimination he selected those worthwhile difficult problems whcich
nevertheless had some chance of solution. Entirely unconcerned with personal
advancement, he hoped only of his paper that each would be a "little gem".
And so it is. Many indeed are treasure houses, abounding in undeveloped
asides on which others may later build, for often he did not elaborate. He
regarded himself as a modest "supplier of tools for the trade", and he felt to
others their day to day use. His colleagues everywhere were always consulting
him.
Eshelby was elected a Fellow of the Toyal Society in 1974, being "distin-
guished for his theoretical studies of the micromechanics of crystalline imper-
fections and material inhomogeneities". he made major contributions to the
theory of static and moving dispocations and of point defects. By an elegant
use of the theory of the potential he obtained some remarkable results on the
elastic fields of ellipsoidal inclusions and inhomogeneities.
In 1951 he introduced, in analogy with the Maxwell tensor, the elastic en-
ergy momentum tensor, which yields forces on elastic singularities. During
his later years he was much concerned with this concept and its developments,
which can provide parameters characterizing the singular fields.
In 1968 he published accounts of its application to the calculation of forces
on static and moving cracks inelastic media. Related work, formulated for
application also to plastic-elastic media, was published simultaneously and in-
dependently by J.R.Rice. Many others have made widespread use of these

characterizing parameters in fracture mechanics, sometimes in a way to which
Eshelby did not wholly subscribe.
Eshelby had a wide knowledge of theoretical physics and repeatedly applied
ideas in one discipline to solve problems in another. He drew much inspiration
from masters of the past and liked to regard some of his most important works
as amusing applications of the theorem of Gauss.
But his scholarly interests went far beyond science. He read French, German
and Russian and could find his way about a Chinese dictionary; indeed, he
knew a great deal about languages and the ancient world and enjoyed holding
his own in discussions with professionals in these fields. His dry jokes and
sayings will long be remembered:
"It’s obvious", he would say,"I forget exactly why". One of his great plea-
sures was to find good secondhand books.
Just before his death he was in correspondence with former colleagues about
some implications of recent calculations he had made of forces on defects in
liquid crystals; and also about cracks in metal fatigue. He was also preparing
lectures to be given in California in the new year.
3.9 Exercises
Probelm 3.1 Let
1 x·x
w(x) = exp(− 2 ), (3.55)
πR3 R
representing a Gaussian distribution .
For any smooth vector field, A ∈ IR3 , define weighted average operation,
Z
< A > (x) := 3
w(x − x0 )A(x0 )dΩx0 (3.56)
IR
where dΩx0 := dx01 dx02 dx03 .

Show that
∇· < A >=< ∇ · A > (3.57)
(Hint: Use Gauss theorem (divergence theorem), and the fact that w(x) →
0 as |x| → ∞.)
Probelm 3.2 Use identidy
δir δis δit

eijk erst = δjr δjs δjt (3.58)
δkr δks δkt
show:
eijk eijk = 3! = 6 (3.59)

eijk eij` = 2δk` (3.60)
eijk ei`m = δj` δkm − δjm δk` (3.61)
Probelm 3.3 Prove
SΩ + D : TΩ : C = 1(4) (3.62)
TΩ + C : SΩ : D = 1(4) (3.63)
where SΩ and TΩ are the Eshelby tensor and the conjugate Eshelby tensor
respectively.
Hint: First show that
σ d = C : (d − σ ∗ ) , and σ ∗ + C : ∗ = 0 . (3.64)
Probelm 3.4 Consider eigenstress homogenization problem illustrated in

Fig. (3.2). Suppose that the disturbance stress field, σ d , can be related to the
eigenstress field, σ ∗ , i.e.
σ d = TΩ : σ ∗ , ∀x ∈ Ω (3.65)
where TΩ is the so-called conjugate Eshelby tensor. Show that the effective
elastic tensor is equal to
h i
C̄ = 1(4s) + f (BΩ − TΩ )−1 : C (3.66)
where the tensor, BΩ := (D − DΩ )−1 : D.
Probelm 3.5 Suppose that an RVE (V) is subjected the following pure trac-
tion boundary condition,
n · σ = t̄ = n · σ 0 , ∀x ∈ ∂V (3.67)
Show that
< σ : δ >= σ 0 :< δ > . (3.68)
Green’s function and Fourier transform 37
Chapter 4
GREEN’S FUNCTION AND FOURIER TRANSFORM
To this end, the key problem of micro-elasticity is to find the relationship

between disturbance strain and eigenstrain (transformation strain). In specific,
Find SΩ such that d = SΩ : ∗ (4.1)
or to find the conjugate Eshelby tensor,
Find TΩ such that σ d = TΩ : σ ∗ (4.2)
A systematic and elegant procedure to derive SΩ and TΩ was established by
Jock Eshelby, which is one of the most important contribution in classical elas-
ticity in the twentieth century.
To understand Eshelby’s inclusion/eigenstrain theory, we first review basic
theory of Green’s function and Fourier transform.
4.1 Green’s Function

Suppose L is a general differential operator, i.e.
L[u] = f (x), ∀x ∈ Ω (4.3)
B[u] = h(x), ∀x ∈ ∂Ω (4.4)
Suppose the above boundary value problem (BVP) is well posed. Choose
f (x) = δ(x − y) (Dirac’s delta function). Then, the solution of BVP (4.3)-
(4.4) is called Green’s function, and it is denoted as G(x, y), i.e.
L[G(x, y)] = δ(x − y), ∀x ∈ Ω (4.5)
B[G(x, y)] = h(x), ∀x ∈ ∂Ω (4.6)
Why are we interested in Green’s function, why are we so fond of Green’s
function? What makes it so special?
To answer this question, we first consider a differential operator, L. Suppose

that there exists an inverse operator to L, and it is denoted as L−1 , such that,
LL−1 = L−1 L = I (4.7)
The simplest differential operator is,

Z
d −1
L= · ⇔ L = (·)dx (4.8)
dx
For general differential operator L, its inverse operator may be written as
Z
−1
L (·) = K(x − y)(·)dy
where K is the so-called kernel function. Once the kernel function is deter-
mined, the inverse operator L−1 is determined.
Suppose that we have already known the inverse operator of L in Eqs.(4.3)
and (4.4). We then can solve the differential equation by applying the inverse
operation,
L−1 L[u] = L−1 (f (x))

Z
−1
u(x) = L (f (x)) = K(x − y)f (y)dy (4.9)
Equation (4.9) is usually termed as “the superposition principle”.

Next question: what is the kernel function? Or how to find the kernel func-
tion for a differential operator L?
Since
Z
−1
u(x) = Iu(x) = LL (u(x)) = L K(x − y)u(y)dy
Z
= LK(x − y)u(y)dy (4.10)
Comparing (4.10) with

Z
u(x) = δ(x − y)u(y)dy
one may find that LK(x − y) = δ(x − y). Therefore, one can deduce that the
kernel function of a differential operator L is its Green’s function:
K(x − y) = G(x − y) (4.11)
In principle, if the Green’s function of a BVP has been found, the BVP is
considered to be solved. This is becaruse one can obtain the general solution
of the differential equation L[u] = f (x) via superposition through certain

reciprocal formula.
Example 4.1 We consider Euler-Bernoulli beam equation with clamped bound-
ary conditions
d2 d2 u
L[u] = EI 2 = f (x), ∀x ∈ (0, l) (4.12)
dx2 dx
0 0
u(0) = u(l) = 0, and u (0) = u (l) = 0 (4.13)
Suppose that we have found the Green’s function related to this problem, i.e.
d2 d2 G(x, y)
L[G] = EI = δ(x − y), ∀x, y ∈ (0, l) (4.14)
dx2 dx2
0 0
G(0, y) = G(l, y) = 0, and G (0, y) = G (l, y) = 0 (4.15)
Via integration by parts, one can show that
Z l 2
d d2 v h d d2 v il h du d2 v il
u EI dx = u EI − EI 2
0 dx2 dx2 dx dx2 0 dx dx 0
Z l 2 2
d u d v
+ 2
EI dx (4.16)
0 dx dx2
Let v = G(x, y). We will have the following reciprocal formula
Z l Z l
uL[G]dx − GL[u]dx
0 0
h d d2 G il h du d2 G il
= u EI 2 − EI 2
dx dx 0 dx dx 0
h d d2 u il h dG d2 u il
− G EI 2 + EI 2 (4.17)
dx dx 0 dx dx 0
Consider the fact that both u(x) and G(x, y) satisfy the same homogeneous
essential boundary conditions. A simple reciprocal holds
Z l Z l
uL(G)dx = GL(u)dx (4.18)
0 0
which leads to
Z l Z l
u(y)δ(x − y)dy = G(x − y)f (y)dy (4.19)
0 0
and consequently,
Z l
u(x) = G(x − y)f (y)dy (4.20)
0
In structural engineering, the Green’s function solution represents the concen-

trated load solution, and the Green’s function is called the influence funtion.
Eq.(4.20) is obtained as an argument of superposition.
Example 4.2 In the second example, we consider Poisson’s equation,
∇2 u = f1 (x), and ∇2 v = f2 (x), ∀x ∈ Ω (4.21)
One can derive the following identity via integration by parts,

Z Z
u∇ · (∇v)dΩ = {∇ · (u∇v) − (∇u) · (∇v)}dΩ
Ω Ω
Z Z
∂v
= udS − (∇u) · (∇v)dΩ (4.22)
∂Ω ∂n Ω
Interchange the position of u and v,

Z Z Z
∂u
v∇ · (∇u)dΩ = vdS − (∇v) · (∇u)dΩ (4.23)
Ω ∂Ω ∂n Ω
Subsraction of (4.22) from (4.23) yields the so-called Green’s reciprocal theo-
rem:
Z Z n
∂v ∂u o
u∇2 v − v∇2 u dΩ = u − v dS (4.24)
Ω ∂Ω ∂n ∂n
Let v(x) = G(x, y), f1 (x) = f (x), and f2 (x) = δ(x − y). We can then show
that
Z n Z
∂G ∂u o
u(x) = u−G dSy + G(x, y)f (y)dΩy (4.25)
∂Ω ∂n ∂n Ω
Note that in 4.25, the Green’s function solution does not necessarily have the
same boundary data as unknown function, u(x), as in the previous example.
Often times, the Green’s function in the infinite domain is chosen in a recipro-
cal representation.
4.2 Fourier transform Z ∞

Consider a function, f (x) ∈ L1 (IR), or |f (x)|dx < ∞. We define the
−∞
Fourier transform as
Z ∞
¯ 1
f (ξ) = F[f ] = f (x)exp(−iξx)dx (4.26)
2π −∞
Z ∞
f (x) = F −1 [f¯] = f¯(ξ)exp(iξx)dξ (4.27)
−∞
In generalized Fourier transform, ξ is a complex number. Assume that func-

tion f (x) has the property such that exp(C1 x)|f (x)| → 0 as x → ∞ and
exp(−C2 x)|f (x)| → 0 as x → −∞. The inversion foumula may be expressed
as the following contour integral
Z ∞−iγ
f (x) = f¯(ξ)exp(iξx)dξ (4.28)
−∞−iγ
where C1 > γ > C2 . The integration contour is usually referred as the

Bromwich contour(Thomas John l’Anson Bromwich (1875-1929)).
Lemma 4.3 (Jordan) Suppose that on the circular arc CR shown in Fig.(4.2)
we have f (ξ) → 0 uniformly as R → ∞. Then
lim exp(ixξ)f (ξ)dξ = 0, (x > 0)

R→∞
We note that if x < 0 similar result holds for the contour in lower half space.
Theorem 4.4 (Cauchy-Gousat) if f (z) is an analytical function at each

point within and on a closed contour C, then
I
f (z)dz = 0 (4.29)
C
Theorem 4.5 (Cauchy’s residue theorem) if f (z) is analytical in-

side a closed contour C (taken in the positive sense) except at points, z1 , z2 , · · · , zn ,
where f (z) has singularities, then
I n
X
f (z)dz = 2πi Residue of f (z) at zj (4.30)
C j=1
Now, the question becomes what is a residue and how to calculate it. The
answer involves with the singularity of f (z). For a function of complex varible,
f (z), one may express f (z) in a local region by its Laurent expansion – an
extension of Taylor expansion of real variable. For instance around a fixed
point zj , we may write
∞
X ∞
X
f (z) = an (z − zj )n + a−n (z − zj )−n , 0 < |z − zj | < a (4.31)
n=0 n=1
The residue is defined as the coefficient a−1 .

There are three types of singularities:(1) essential singularity, (2) removable
singularity, and (3 )pole of order n.
• The essential singularity refers to a singularity, or pole of infinity order.

For instance, for the pole z = 0,
1 1 1 1
cos =1− + − + ···
z 2!z 2 4!z 4 6!z 6
z = 0 is an essential singularity.
• The removable singularity is an unsubstantial singularity, i.e. the alleged
singularity disappears in Laurent expansion. For instance, at z = 0,
sin z z2 z4 z6
f (z) = =1− + − + ···
z 3! 5! 7!
• Pole of order n: Consider the function,
1 1
f (z) = +
z + 1 (z − 1)3
This function has two singularities at z = −1 and z = 1. For singularity at
z = −1, its order is one, and it is called a pole of order one. For singularity at
z = 1, its order is three, and it is called a pole of order 3.
The formula to calculate the residue for a pole, zj , of order n is
1 dn−1 h n
i
Residue at (z = zj ) = lim (z − z j ) f (z) (4.32)
(n − 1)! z→zj dz n−1
We call the pole of order one as simple pole. For simple pole,
Residue of a simple pole at (z = zj ) = lim (z − zj )f (z) (4.33)
z→zj
If f (z) = p(z)/q(z), one may also write

p(zj )
Residue of a simple pole at (z = zj ) = (4.34)
q 0 (zj )
Figure 4.1. Contour integral and the count of residue
Example 4.6 In this example, we apply Cauchy’s residue theorem to evalu-

ate the following line integral.
Z ∞
exp(ikt)
dt
−∞ (t − x)(t − ia)
where k > 0 and a > 0.

Since k > 0, based on Jordan’s lemma, we can use the following contour
integral to replace the line integral,
Z ∞ Z Z
exp(ikt) exp(ikt) exp(ikt)
dt = dt + dt
−∞ (t − x)(t − ia) C∞ (t − x)(t − ia) C (t − x)(t − ia)
where the contour integral is a half circle. Thus,

Z ∞
exp(ikt)
dt = 2πi Residue(f (ia)) + πi Residue(f (x))
−∞ (t − x)(t − ia)
exp(−ka)(x + ia) exp(ikx)(x + ia)
= −2πi 2 2
+ πi
x +a x2 + a2
(4.35)
The simple pole at x is only counted for half of the residue is because that it
has only half circle.
Theorem 4.7 (Cauchy’s Integral Formula) Let f (z) be analytical

interior to and on a simple closed contour C. Then at any interior point z
I
1 f (ζ)
f (z) = dζ (4.36)
2πi C ζ − z
Theorem 4.8 (Convolution) If f (x), g(x) ∈ L1 (IR) ∩ L2 (IR), the fol-

lowing identity holds
Z ∞ Z ∞
¯ 1
f (ξ)ḡ(ξ) exp(iξx)dξ = g(x − y)f (y)dy (4.37)
−∞ 2π −∞
Proof:
by definition,
Z ∞ Z ∞h 1 Z ∞ i
f¯(ξ)ḡ(ξ) exp(iξx)dξ = ḡ(ξ) f (y) exp(−iξy)dy exp(iξx)dξ
−∞ 2π −∞
Z−∞∞ h 1 Z ∞ i
= f (y) ḡ(ξ) exp(iξ(x − y))dξ dy
−∞ 2π −∞
Z ∞
1
= g(x − y)f (y)dy (4.38)
2π −∞
In 3D, we have
Z ∞ Z ∞
1
f¯(ξ)ḡ(ξ) exp(iξ · x) = g(x − y)f (y)dy (4.39)
−∞ (2π 3 ) −∞
Example 4.9 Consider Heaviside function,

1 x>0
H(x) = (4.40)
0 x<0
Note that at x=0 Heaviside function is not defined.
To find the Fourier transform of the Heaviside function,
Z ∞
1
H̄(ξ) = H(x) exp(−iξx)dx
2π −∞
Z ∞
1 1 (−1) ∞
= exp(−iξx)dx = exp(−iξx)
2ξ 0 2π iξ 0
1
= (4.41)
2πiξ
The result implies that exp(−iξ∞) → 0, which requires that Im(ξ) < 0.
Lighthill showed that in the sense of generalized function,
πi 1
1 ξ>0
H̄(ξ) = exp − sgn(ξ) , where sgnξ := (4.42)
2 2π|ξ| −1 ξ < 0
Note that H(x) ∈ / L1 (IR). Therefore,

Z ∞ Fourier transform of Heaviside function
1
does not really exit for f ∈ L . |f (x)| < ∞ is a very stringent condition.
−∞
It is why many functions that has Laplace transform do not possess Fourier
transform, which is the reason why sometimes we use Laplace transform in-
stead of Fourier transform. By the way, if ξ is taken as a complex number,
Fourier transform is equivalent to bilateral Laplace transform.
Example 4.10 To find the Fourier transform of the Dirac’s delta function,
Z ∞
1 1
δ̄ = δ(x) exp(−ixξ)dx = (4.43)
2π −∞ 2π
Inversely,
Z ∞ Z ∞
1
δ(x) = δ̄(ξ) exp(iξx)dξ = exp(iξx)dξ (4.44)
−∞ 2π −∞
Example 4.11 On the other hand, consider the inversion formula,

Z ∞
δ(ξ) exp(iξx)dξ = exp(i0x) = 1, ⇒ 1̄(ξ) = δ(ξ) (4.45)
−∞
Hence Z ∞
1
1̄(ξ) = δ(ξ) = exp(−iξx)dx (4.46)
2π −∞
In three-dimensional space, we have the identity,

Z ∞
1
δ(ξ) = exp(−iξ · x)dx (4.47)
(2π)3 −∞
Combining (4.44) and (4.47), one may draw conclusion that
Z ∞
1
δ(ξ) = cos(ξ · x)dx (4.48)
(2π)3 −∞
Example 4.12 The Fourier transform of f (x) is
1 ia
f¯(ξ) = 2
(4.49)
2π ξ(ξ − iaξ − a)
Find f (x)?
f¯(ξ) has three poles in the complex plane:
r
ia a2 ia p
ξ1 = 0, and ξ2,3 = ± a− = ± β, β := a − a2 /4 (4.50)
2 4 2
Therefore,
Z ∞−iγ
f (x) = f¯(ξ) exp(iξx)dξ
−∞−iγ
I
1 ia exp(iξx)
= dξ
C 2π (ξ − 0)(ξ − ξ2 )(ξ − ξ3 )
X3
= πiResidue of ξ at ξ1 + 2πi Residue of ξ at ξj
j=2
n exp(iξ x) exp(iξ2 x) exp(iξ3 x) o
1
= ia + +
ξ2 ξ3 ξ2 (ξ2 − ξ3 ) ξ3 (ξ3 − ξ2 )
ia ia
n 1 exp[ix( + β)] exp[ix( − β)]
= (−a) + 2r − 2r
−a ia a2 ia a2
+β 2 a− −β 2 a−
xa 2 4 2 4
exp − n ia ia o
= 1− r 2 β− exp(iβx) + β + exp(−iβx)
a2 2 2
2 a−
xa4
exp −
= 1− r 2 2β cos x + a sin βx (4.51)
a2
2 a−
4
4.3 Examples of Green’s Function

Example 4.13 Find the Green’s function of two-dimensional Poission’s equa-
tion in infinite domain,
∇2 G(x, y) + δ(|x − y|) = 0, ∀x ∈ IR2 (4.52)
1 d d 0
Use the polar coordinate ∇2 = r and denote x = x − y. We have
r dr dr
1 d d 0 0
r G = −δ(x1 )δ(x2 ) (4.53)
r dr dr
and
Z 2π Z r Z
1 d 0 d 0 0 0 0 0 0
r 0 G r dr dθ = − δ(dx1 )δ(dx2 )dx1 dx2 (4.54)
0 0 r0 dr0 dr Ω
0 p
where r is the dummy variable and r = |x−y| = (x1 − y1 )2 + (x2 − y2 )2 .
The integration domain is a circular region centered at x = y and with the
radius r.
Therefore,
d d 1 1
2π r G = −1, ⇒ G = − (4.55)
dr dr 2π r
Finally, we find that
1
G(x − y) = − lnr (4.56)
2π
Example 4.14 Consider one dimensional Helmhotz equation,
d2 u
+ k 2 u = δ(|x − y|) (4.57)
dx2
Apply Fourier transform,
Z ∞
1
ū(ξ) = u(x) exp(−iξx)dx
2π −∞
d2 u Z ∞ 2
1 d u
F = exp(−iξx)dx = −ξ 2 ū(ξ)
dx2 2π −∞ dx2
Z ∞
1 1
δ̄(|x − y|) = δ(x − y) exp(iξx)dξ = exp(−iξy) (4.58)
2π −∞ 2π
and
1 1
ū(ξ) = exp(−iξy) (4.59)
2π k 2 − ξ 2
Therefore,
Z ∞
u(x) = ū(ξ) exp(iξx)dξ
−∞
Z ∞
1 1
= exp(iξ(x − y))dξ
2π −∞ (k + ξ)(k − ξ)
2
exp(iξ(x − y))
I
1 iX
= dξ = Residues of ξ at ξi
2π R (k + ξ) k − ξ) 2
i=1
i n 1 1 o
= − exp(ik(x − y)) − exp(−ik(x − y))
2 2k 2k
i n o
= − cos k(x − y) + i sin k(x − y) − cos k(x − y) − i sin k(x − y)
4k
i 1
= − 2i sin k(x − y) = sin k(x − y) (4.60)
4k k
Figure 4.2. Inversion paths of Fourier transform
Example 4.15 Find Green’s function for three-dimensional Poisson’s equa-

tion,
0
∇2 G + δ(x − x0 ) = 0 ⇒ G,ii + δ(x − x ) = 0 (4.61)
∂2
where ∇2 = , i = 1, 2, 3 and δ(x − x0 ) = δ(x1 − x01 )δ(x2 − x02 )δ(x3 −
∂xi ∂xi
x03 )
Consider the fact that
1 1
δ̄(x − x0 ) = exp −iξ · x 0
⇒ δ(x − x0
) = exp iξ · (x − x0
) dξ
π3 2π 3
Therefore, based on definition,
Z ∞
0
G(x − x ) = Ḡ(ξ) exp iξ · x dξ
−∞
one may derive that
Z ∞
0
G,ii (x − x ) = − Ḡ(ξ)ξi ξi exp(iξ · (x − x0 ))dξ (4.62)
−∞
and
1 1 1
Ḡ(ξ)ξi ξi = ⇒ Ḡ(ξ) = (4.63)
(2π)3 (2π)3 ξi ξi
Figure 4.3. Inversion of three-dimensional Fourier transform

0 0 0
Let ξ 2 := ξi ξi , r = x − x , r := |x − x |, and ξ · (x − x ) = ξr cos θ. Then,
Z ∞
0 1 1 0
G(x − x ) = 3 2
exp(iξ · (x − x ))dξ
(2π ) −∞ ξ
Z ∞ Z π Z 2π
1 1 0
= 3 2
exp(iξ · (x − x ))ξ 2 dξ sin θdθdφ
(2π ) 0 0 0 ξ
Z ∞ Z −1 Z 2π
1
= exp(iξr cos θ)dξ(−d cos θ)dφ
(2π 3 ) 0 1 0
Z ∞Z 1
1
= exp(iξrt)dξdt
(2π 2 ) 0 −1
Z ∞ Z 1h
1 i
= dξ cos(ξrt) + i sin(ξrt) dt (4.64)
(2π 2 ) 0 −1

Z 1
1 1 2 sin ξr
cos(ξrt)dt = sin(ξrt) = (4.65)
−1 ξr −1 ξr
Z 1
sin(ξrt)dt = 0 (4.66)
−1
Hence
Z ∞
0 1 sin ξr
G(x − x ) = 2
dξ
2π 0 ξr
Z ∞
1 sin ξr 1
= 2
d(ξr) = 2 Si(∞) (4.67)
2π r 0 ξr 2π r
Z x
sin t π
where Si(x) := dt and Si(∞) = . Finally, we have
0 t 2
0 1 1
G(x − x ) = (4.68)
4π |x − x0 |
4.4 Static Green’s function for 3D linear elasticity

The Green’s function for static, linear, isotropic elasticity was derived by
Lord Kelvin (1882). The derivation shown below employs the Fourier integral
transform, which is a systematic and elegant procedure to find Green’s function
for partial differential equations. Consider the Navier equation,
σji,j + fi = 0 (4.69)
Denote Green’s function vector of the displacement field as
∞
um
i (x, y) = Gmi (x, y) (4.70)
We let
G∞ ∞
σij m = Cijkl G
kl (4.71)
fim = δ(x − y)δmi (4.72)
where δ(x − y) := δ(x1 − y1 )δ(x2 − y2 )δ(x3 − y3 ), and the integer m is a
free index, which indicates the direction of the concentrated load.
Then,
G∞ ∞ ∞ G∞∞
σij m = Cijkl G
kl = Cijkl Gmk,l → σij,j = Cijkl Gmk,lj
m
Then Green’s function for an infinite linear elastic medium is the solution of
the following equatin,
Cijkl G∞
mk,lj + δ(x − y)δmi = 0 (4.73)
Figure 4.4. The unit sphere S 2 in the ξ-space. Green’s function at point z is expressed by a
line integral along S 1 which lies on the plane perpendicular to z
Apply Fourier integral transform,

Z ∞
∞
Gmk (x − y) = Ḡ∞
mk (ξ) exp(iξ · (x − y))dξ (4.74)
−∞
Z ∞ Z Z Z ∞
where = , and dξ = dξ1 dξ2 dξ3 .
−∞ −∞
Consider
Z ∞
G∞
mk,lj (x − y) = − Ḡ∞
mk (ξ)ξl ξj exp(iξ · (x − y))dξ (4.75)
−∞
Z ∞
1
δ(x − y) = exp(iξ · (x − y))dξ (4.76)
(2π)3 −∞
We obtain the following algebraic equations in Fourier space,

1
Cijkl Ḡ∞
mk (ξ)ξl ξj = δim (4.77)
(2π)3
Let
1
Kik = Cijkl ξj ξl ⇒ Kik Ḡ∞
mk = δim (4.78)
(2π)3
Consider Laplace expansion,
Nji (ξ)Kik (ξ) = D(ξ)δjk (4.79)
where Nji is the cofactor of Kji and D(ξ) = det{Kij (ξ)}.

Multiplying (4.78) with Nji yields
1
Nji (ξ)Kik (ξ)Ḡ∞
mk (ξ) = Nji (ξ)δim (4.80)
(2π)3
1
D(ξ)δjk Ḡ∞
mk (ξ) = Njm (ξ) (4.81)
(2π)3
which leads to
1 Njm (ξ)
Ḡ∞
jm (ξ) = (4.82)
(2π)3 D(ξ)
Change indices j ↔ i and m ↔ j. Via inverse Fourier transform, one may
find that
Z ∞
1 Nij (ξ)
G∞
ij (x − y) = 3
exp(iξ · (x − y))dξ (4.83)
(2π) −∞ D(ξ)
For linear isotropic material, one may find that

Nij (ξ) = µξ 2 (λ + 2µ)δij ξ 2 − (λ + µ)ξi ξj (4.84)
D(ξ) = µ2 (λ + 2µ)ξ 6 (4.85)
Let z = x − y. We have
Z ∞
∞ 1 1
2

Gij (z) = (λ+2µ)δ ij ξ −(λ+µ)ξ i ξj exp(iξ·z)dξ
(2π)3 −∞ µ(λ + 2µ)ξ 4
(4.86)
2
To integrate (4.86), we donote S as a unit sphere where |ξ| = 1, and denote
S 1 as a unit circle on the surface of S 2 , where S 2 is intersected by a plane
perpendicular to vector z.
Apply Radon decompositon,
dξ = dVξ = dξ1 dξ2 dξ3 ⇒ dVξ = ξ 2 dξdS (4.87)

where ξ 2 = ξ12 + ξ22 + ξ32 and dS is the surface element on the unit sphere S 2
in ξ-space. Imagine that the ξ-space is a expanded spherical balloon.
Denote ξ̄ = ξ¯i eξi as a unit vector pointing from the origin to the surface of
2
S along ξ direction and denote z̄ = z̄i ezi as another unit vector point from the
origint to thepsurface of S 2 along z direction. Therefore, ξ = ξ ξ̄ and z = zz̄
where ξ = ξ12 + ξ22 + ξ32 and z = z12 + z22 + z32 . Obviously, ξ¯i = ξi /ξ
p
and z̄i = zi /z.

Then Eq.(4.86) can be written as
Z ∞ Z
1 1
¯i ξ¯j

Gij (z) = dξ (λ + 2µ)δ ij − (λ + µ) ξ
(2π)3 0 S 2 µ(λ + 2µ)
· exp{iξz ξ̄ · z̄}dS(ξ̄) (4.88)
Consider the symmetry property (change ξ → −ξ of Eq.(4.86)). We may

also have
Z ∞ Z
1 1
¯ ¯

Gij (z) = dξ (λ + 2µ)δ ij − (λ + µ)ξ ξ
i j
(2π)3 0 S 2 µ(λ + 2µ)
· exp{−iξz ξ̄ · z̄}dS(ξ̄) (4.89)
Change the scalar ξ → −ξ. Eq.(4.89) yields

Z 0 Z
1 1
¯ ¯

Gij (z) = dξ (λ + 2µ)δij − (λ + µ)ξi ξj
(2π)3 −∞ S 2 µ(λ + 2µ)
· exp{iξz ξ̄ · z̄}dS(ξ̄) (4.90)
Combining (4.88) with (4.90) yields

Z ∞ Z
1 1
¯ ¯

Gij (z) = dξ (λ + 2µ)δ ij − (λ + µ)ξ ξ
i j
2(2π)3 −∞ S 2 µ(λ + 2µ)
· exp{iξz ξ̄ · z̄}dS(ξ̄) (4.91)
since Z ∞
exp(iξz ξ̄ · z̄)dξ = 2πδ(z ξ̄ · z̄) (4.92)
−∞
one has
[(λ + 2µ)δij − (λ + µ)ξ¯i ξ¯j ]
I
1
Gij (z) = δ(z ξ̄ · z̄) dS(ξ̄) (4.93)
2(2π)2 S2 µ(λ + 2µ)
To integrate (4.93), one has to evaluate the following two integrals:
Z Z
δ(z ξ̄ · z̄)dS? and ξ¯i ξ¯j δ(z ξ̄ · z̄)dS?
S2 S2
Consider ξ̄ · z̄ = cos θ, d cos θ = − sin θdθ. One may decompose the sur-
face element
on S 2 into:
dS(ξ̄) = sin θdθdφ = −d(ξ̄ · z̄)dφ, where θ →
[0, π] cos θ → [1, −1] and φ → [0, 2π]. If we let t = ξ¯ · z̄,
Z Z 1 Z 2π
2π
δ(z ξ̄ · z̄)dS = δ(zt)dt dφ = (4.94)
S2 −1 0 z
On the other hand,
Z Z 1 Z 2π
ξ¯i ξ¯j δ(z ξ̄ · z̄)dS = δ(zt)ξ¯i ξ¯j dtdφ (4.95)
S2 −1 0
Consider the projection of vector ξ̄,
P rojz̄k ξ̄ = cos θz̄ = cos θz̄i ei (4.96)

P rojz̄⊥ ξ̄ = sin θb = sin θ(cos φa1 + sin φa2 ) (4.97)
Considering,
a1 = (a1 · ei )ei ; a2 = (a2 · ei )ei
one has
ξ̄ = xi ei = cos θz̄ + sin θb

= cos θz̄i ei + sin θ cos φa1i + sin φa2i ei (4.98)
Thereby,
ξ¯i = cos θz̄i + sin θ(cos φa1i + sin φa2i )

⇒ ξ¯i ξ¯j = cos θz̄i + sin θ(cos φa1i + sin φa2i )

· cos θz̄j + sin θ(cos φa1j + sin φa2j )
h
= cos2 θz̄i z̄j + sin θ cos θ z̄i (cos φa1j + sin φa2j )
i
+z̄j (cos φa1i + sin φa2i )
+ sin2 θ(cos φa1i + sin φa2i )(cos φa1j + sin φa2j )
p h
= t2 z̄i z̄j + t 1 − t2 z̄i (cos φa1j + sin φa2j )
i
+z̄j (cos φa1i + sin φa2i )
+(1 − t2 )(cos φa1i + sin φa2i )(cos φa1j + sin φa2j )(4.99)
where t = cos θ.

Z 1
t2 δ(zt)dt = 0
−1
Z 1 p
t 1 − t2 δ(zt)dt = 0 (4.100)
−1
We have
I Z 1 Z 2π
δ(zt)ξ¯i ξ¯j dS = δ(zt) {cos2 φa1j a1i
S2 −1 0
+ cos φ sin φ(a1j a2i + a1i a2j ) + sin2 φa2j a2i }dtdφ
π π
= (a1i a1j + a2i a2j ) = δij − z̄i z̄j (4.101)
z z
because a1i a1j + a2i a2j + z̄i z̄j = δij . Note that a1 , a2 , and z̄ form a triads.
Let Q1i = a1i , Q2i = a2i and Q3i = z̄i . From Qik QTkj = Qik Qjk = δij , one
derives that a1i a1j + a2i a2j + z̄i z̄j = δij .
Consequently,
1 1 h 2π(λ + 2µ)δij − π(λ + µ)(δij − z̄i z̄j ) i
G∞ij (z) =
(2π)2 2z µ(λ + 2µ)
1 1 1 (λ + µ) λ + 3µ
n o
= δij + z̄i z̄j
8π z µ (λ + 2µ) λ + µ
1 n (xi − yi )(xj − yj ) o
= (3 − 4ν)δij +
16πµ(1 − ν)|x − y| |x − y|2
(4.102)
4.5 Variation in a Theme: Radon Transform

Let x = (x1 , x2 , x3 ) be the positoin vector of a spatial point in IR3 and
consider a regular function f (x) (image density) defined in IR3 . The Radon
transform of f (x) is defined as
Z ∞
ˆ
f (s, n) = R{f (x)} = f (x)δ(s − n · x)dx (4.103)
−∞
fˆ is the projectin of f (x) on the plane n · x = s, where n is a unit vector, and

s is the distance from the plane to the origin of the coordinate (see Fig. (4.5)).
The integral is the integration of image density, f (x), along the plane. The
collection of all fˆ(s, n) for all unit vector n is called the Radon transform.
The inverse Radon transform is carried out by two steps:
1. f˜(s, n) = ∂s2 fˆ(s, n) (4.104)
Z
1
2. f (x) = R−1 (f˜) = − 2 f˜(n · x, n)dS(n) (4.105)
8π |n|=1
Figure 4.5. Projection plane of three-dimensional Radon transform
The Radon transform has the following properties:

1 fˆ(s, n) is an even and homogeneous, of order -1, function, i.e.
fˆ(αs, αn) = |α|− 1fˆ(s, n);
2 linearity: R(c1 f + c2 g) = c1 fˆ + c2 ĝ;
3 transform of derivatives:
R(∂i f (x)) = ni ∂s fˆ(s, n)
R(∂i ∂j f (x)) = ni nj ∂ 2 fˆ(s, n)
s
Example 4.16 Consider an image density function, g(x, y). The two-dimensional
Radon transform may be defined as
Z ∞Z ∞
ĝ(ρ, θ) = g(x, y)δ(ρ − x cos θ − y sin θ)dxdy (4.106)
−∞ −∞
which is identical to the following line integral

Z ∞
ĝ(ρ, θ) = g(ρ cos θ + t sin θ, ρ sin θ − t cos θ)dt (4.107)
−∞
where parameter, t, is the length of straight line cos θx + sin θy = ρ. It is

shown in Fig (4.6) that
x = ρ cos θ + t sin θ, and y = ρ sin θ − t cos θ (4.108)
In Fig. (4.6), it can be seen that two very bright spots are found in the
Radon transform, and the postion shown the parameters of the lines in the real
physical image.
(a)
(b) (c)
Figure 4.6. Two-dimensional Radon transform: (a) prjectinline, (b) image in the physical
space, and (c) image in the Radon transform space
Example 4.17 Let f (x) = δ(x). The Radon transform of Dirac’s delta
function is
Z ∞
δ̂(s, n) = R(δ) = δ(x)δ(s − n · x)dx = δ(s) (4.109)
−∞
where s = ni xi .
Subsequently,
δ̃(s, n) = δ 00 (s) (4.110)
and the inverse Radon transform is
Z
1
δ(x) = − δ 00 (nk xk )dS (4.111)
8π 2 S2
One can verify this by considering the identity (4.94), i.e.

Z
2π
δ(nk xk )dS = (4.112)
S 2 |x|
Figure 4.7. Two-dimensional Radon transform: (a) projection line, (b) image in the physical
space, and (c) image in the Radon transform space
∂2
Applying the harmonic operator ∇2 = to the above identity and con-
∂xi ∂xi
sidering Example (4.15) (Eq.(4.68)) yields
Z 1
δ 00 (nk xk )ni ni dS = 2π∇2 = −8π 2 δ(x) (4.113)
S2 |x|
Now we use the Radon transform to derive 3D static Green’s function of a
linear elastic medium. Consider the concentrated load is acting at the origin of
the coordinat (y = 0).
Cijkl Gkm,lj + δ(x)δim = 0 (4.114)
Assume that the Green’s function can be written as a form of inverse Radon
transform, Z
∞ 1
Gkm (x) = − 2 G˜∞ (ξ¯n xn )dS (4.115)
8π S 2 km
Then Z
∞ 1 00
Gkm,lj (x) = − 2 G˜∞
km (ξ¯n xn )ξ¯l ξ¯j dS (4.116)
8π S 2
On the other hand,
Z
−1
1
δ(x) = R δ̃(s) = − 2 δ 00 (ξ¯n xn )dS (4.117)
8π S 2
We then obtain
00 00 ¯
Cijkl ξ¯j ξ¯l G˜∞ ¯
km (ξn xn ) = −δim δ (ξn xn ) (4.118)
which leads to
¯ −1 ¯ ¯
G˜∞
ij (ξ) = −Kij (ξ)δ(ξn xn ) + C1 ξn xn + C0 (4.119)
where
¯
Nij (ξ)
−1
Kri Cijkl ξ¯l ξ¯j = δrk , or Kij = (4.120)
D(ξ)
Note that C1 = C0 = 0 because it is required that G∞
ij (x) → 0, as x → ∞.
For isotropic materials,
−1 1h (λ + µ)ξ¯i ξ¯k i
Kij (ξ̄) = δij − (4.121)
µ (λ + 2µ)
and, correspondingly,
Z
1 −1
G∞
ij (x) = 2 Kij (ξ̄)δ(ξ¯n xn )dS (4.122)
8π S2
and subsequently,
1 h δij (λ + µ) i
G∞
ij (x) = − |x|,ij (4.123)
4πµ |x| 2(λ + 2µ)
4.6 Joseph Fourier(I)

Joseph Fourier was born in 1768 in Auxerre, the ninth child of a master
tailor. Although the death of his father left him an orphan at the age of ten, his
intelligence gained him a free place at the local Benedictine school. At the end
of a brilligent school career he applied to enter the artillery only to be informed
that such a profession was only open to those of noble blood and was closed to
him ’even if he were a second Newton’.
Fourier began to prepare to enter the Benedictine teaching order but, what-
ever his plans may have been, the course of his life was violently altered by
the outbreak of the French Revolution, .... The situation of the new Republic
called for ruthless measures which the government, conscious of its own revo-
lutionary virtue, was well prepared to take. Treachery was fought by a political
terror in which opponents both to the left and right were executed and, as the
definition of treachery was extended, it became clear that no one was safe.
Fourier himself was arrested, released and then rearrested. A deputation from
Auxerre which, with considerable courage, went to Paris to plea his case, was
told-’Yes, he speaks well, but we nolonger have any need of musical patriots.’
Only the fall of Robespierre saved Fourier’s head.
However Fourier’s release did not mark the end of his troubles. As coup
d’etate follows coup d’eta, and the revolution swung erractically to the right he
Figure 4.8. Joseph Fourier
would remain a marked man. No one had been executed in Auxerre but Fourier
had been an agent of the terror there. His arrest was on a charge of H’ebertism
and the H’ebertists were to the left of Robespierre. The word ’terrorist’ then,
like ’Trotskyist’ now, denoted a defeated yet feared opponent.
Luckily an opportunity to leave Auxerre now presented itself. A new col-
lege (the Echole Normale) was being set up in Paris to help train teachers and
Fourier could now study under men like Lagrange, Monge and Laplace and
excape his terrorist past. Fourier’s talents were soon noted, but the college was
not successful and its closure was followed by further problems for Fourier.
’We shudder when we think that the pupils of the Ecole Normale were cho-
sen under the reign of Robespierre and his proteges. It is only too true that
Balme and Fourier, pupils of the department of Yonne have long prefessed the
atrocious principles and infernal maxims of the tyrants. Nevertheless they pre-
pare to become teachers of our children. Is it not to vomit their poison in the
bosim of innocence (From an address to the National Convention, quoate by
Herivel)’
Fourier was again arrested, released, rearrested and finally, following yet
another political swing, released to become a teacher at the new Ecole Poly-
technique.
Here Fourier remained for three years. That his talent was recognized is
shown by the fact that he succeeded Lagrange in the Chair of Analysis and
Mechanics. The quiet interlude was ended by a gonernment order to join the
invasion of Egypt. Ostensible intended to liberate Egypt from the Turks and to
threaten the British position in India, the expedition may have been seen by the
government as a way of keeping a troublesome general as far away as possible
and by the general (Napoleon) as the first step toward becoming Emperor of
the East. Fourier wa one of a ghroup of scientists and intellectualls intended to
form part of the immense cultural benefits that France was to bestow on Egypt.
Both before and after Napolean’s departure, Fourier occupied several im-
portant administrative and political posts in Egypt. When the French expedi-
tion finally surrended in 1801 and Fourier was repatriated, Napoleon offered
him the post of Prefect of the Department of the Isere centred round Grenoble
(France had been divided into 83 Departments and each Prefect governed his
Department of behalf of the central government.)
Although he could have continued a Professor at the Polytechnique, Fourier
accepted the offer. Herivel suggests that Egypt had given him a taste for admin-
istration and that he hoped to rise higher. Herivel also accounts that Fourier’s
close association with Kleber after Napoleon’s departure account for the fact
that these hopes were not fulfilled.
Fourier seems to have been popular and efficient Prefect. His greatest achieve-
ment during his 14 years of office was by reconciling the conflicting interests
of some forty communities to enable the swamps of Bourgion to be drained.
The draining of twenty thousand acres of swamps resulted in major economic
and health benefits and was achieved during a period morenoted for grandiose
paper plans than for concrete achievements. Fourier’s other administrative
memorial was a new road across the Alps (now Route 91).
Apart from his perfectorial duties Fourier helped organize the Description
of Egypt. This work written by the intellectuals attached to the Egyptian ex-
pedetion did much to inspire European interest in Egypt and was thus one of
the two permanent results of the expediton. (The other was the discovery of
the Rosetta Stone, atrilingual inscription which was to provide the key to the
deciphering of hieriglyphics.)
Fourier’s main contribution was the general introduction – a survey of Egyp-
tian history up to modern times. An Egyptologist with whom I discussed this
described the introduction as a masterpiece and a turning point in the subject,
was surprised to hear that Fourier also had a reputation as a mathematician!
–T.W.Korner From Fourier Analysis

4.7 Exercises
Probelm 4.1 Find the Green’s function for a both end clamped Euler-Bernoulli
beam, i.e.
d2 d2 G(x, y)
EI = δ(x − y), ∀x, y ∈ (0, `) (4.124)
dx2 dx2
and
G(0, y) = G(`, y) = 0, G0 (0, y) = G0 (`, y) = 0 . (4.125)
Probelm 4.2 For isotropic materials, elasticity tensor has the form
Cijk` = λδij δk` + µ(δi` δjk + δik δj` ) (4.126)
Show
1.
Kik (ξ) = Cijk` ξj ξ` = (λ + µ)ξi ξk + µδik ξj ξj (4.127)
2. (Hint : use eijk eimn = δjm δkn − δjn δkm .)
1
Nij (ξ) = eik` ejmn Kkm K`n
2
= µξ 2 ((λ + 2µ)δij ξ 2 − (λ + µ)ξi ξj ) (4.128)
3.
D(ξ) = µ2 (λ + 2µ)ξ 6 (4.129)
Probelm 4.3 The Green’s function, G∞ (x, x0 ), satisfies the 2D Laplace

equation,
∇2 G∞ (x, x0 ) + δ(x − x0 ) = 0, ∀x ∈ IR2 (4.130)
∂2 ∂2 ∂2
where ∇2 = 2 + 2 = , α = 1, 2. And δ(x − x0 ) = δ(x1 −
∂x1 ∂x2 ∂xα ∂xα
x01 )δ(x2 − x02 ). Use Fourier transform method to derive
1
G∞ (x − x0 ) = − ln |x − x0 | . (4.131)
2π
Hints
Z ∞ Z ∞
1
δ(x − x0 ) = exp iξ · (x − x0 ) dξ (4.132)
(2π)2 −∞ −∞
and
Z ∞ Z ∞
exp(i(ξ1 x1 + ξ2 x2 ))
dξ1 ξ2
−∞ −∞ ξ12 + ξ22
= −2πlnR (4.133)
q
0 0
where R = (x1 − x1 )2 + (x2 − x2 )2 .
Probelm 4.4 In isotropic materials, the static Green’s function of linear

elasticity is
1 δij 1 ∂2
G∞ 0
ij (x, x ) = − |x − x0 | (4.134)
4πµ |x − x0 | 16πµ(1 − ν) ∂xi ∂xj
Let x̄ = x − x0 and x̄ = |x̄| = |x − x0 |. Show that for isotropic materials,

−1 δmi x̄n + δni x̄m − δmn x̄i x̄m x̄n x̄i
Cj`mn Gij,` = (1 − 2ν) +3
8π(1 − ν) x̄3 x̄5
(4.135)
where ν is the Poisson ratio, and µ, λ are the Lam«e constants with
2µν λ(1 − 2ν) λ
λ= , µ= , ν= (4.136)
1 − 2ν 2ν 2(λ + µ)
Hint: (Cj`mn = λδj` δmn + µ(δjm δ`n + δjn δ`m )).

Eigenstrain Theory 63
Chapter 5
EIGENSTRAIN THEORY
There are mainly two homogenization meghods used in engineering appli-

cations today. The first one is Eshelby’s, or Mura’s eigenstrain theory. The
central part of the theory is Eshelby’s eigenstrain solution for ellipsoidal inclu-
sion. The theory has been further refined, detailed and articulated by Professor
Mura and his co-workers. Today, it is called eigenstrain theory, and it has
widespread applications.
5.1 Fundamental equations of micro-elasticity

Consider equilibrium equation in an RVE
σji,j = 0 (5.1)
After homogenization, inhomogeneities are replaced by a eigenstrain distribu-
tion ∗ ij (x). Assuming that material is linear elastic, and the total strain is the
sum of elastic strain and eigenstrain,
ij = eij + ∗ ij (5.2)
1
The total strain is defined as ij = (ui,j + uj,i ). And elastic strain is related
2
with Cauchy stress by Hooke’s law
σij = Cijk` (k` − ∗ k` ) = Cijk` (uk,` − ∗ k` ) (5.3)
The equilibrium equation then takes a form
Cijk` ui,`j − Cijk` ∗ k`,j = 0 (5.4)
Note that one interprets the effect of eigenstrain distribution as a type of body
force, fi = −Cijk` ∗ k`,j , and the original equilibrium equation has the form
σji,j + fi = 0.
Let,
Z ∞
uk (x) = ūk (ξ) exp(iξ · x)dx
Z−∞
∞
= ūk (ξ) exp(iξm xm )dx (5.5)
Z−∞
∞
∗ k` (x) = ¯∗k` (ξ) exp(iξm xm )dx (5.6)
−∞
Hence
Z ∞
uk,`j (x) = − ūk ξ` ξj (ξ) exp(iξm xm )dx (5.7)
−∞
Z ∞
∗
k`,j (x) = i ¯∗k` (ξ)ξj exp(iξm xm )dx (5.8)
−∞
Substituting (5.7) and (5.8) into (5.4) yields

Z ∞
(Cijk` ūk ξ` ξj + iCijk` ¯∗k` (ξ)ξj ) exp(iξm xm )dx = 0 (5.9)
−∞
which leads to
Cijk` ξj ξ` ūk = −iCijk` ¯∗k` (ξ)ξj (5.10)
Denote
Kik (ξ) = Cijk` ξj ξ` (5.11)

f¯i = −iCijk` ¯∗k` ξj (5.12)
They are related by
f¯1
    
K11 K12 K13 ū1
 K21 K22 K23   ū2  =  f¯2  (5.13)
K31 K32 K33 ū3 f¯3
We find that
Nij (ξ) ¯ −1 ¯
ūi (ξ) = fj = Kij fj (5.14)
D(ξ)
where
1
Nij (ξ) = eik` ejmn Kkm K`n (5.15)
2
D(ξ) = emn` Km1 Kn2 K`3 (5.16)

n o
K(ξ) = ξ · C · ξ = ξ · λ1(2) ⊗ 1(2) + 2µ1(4s) · ξ

= λξ ⊗ ξ + µ ξ ⊗ ξ + |ξ|2 1(2)
= (λ + µ)ξ ⊗ ξ + µ|ξ|2 1(2) (5.17)
Denote
Q(ξ) = K−1 (ξ) (5.18)
Q must be an isotropic second order tensor in Fourier space as well. Assume
that
Q(ξ) = {ξ · C · ξ}−1 = Aξ ⊗ ξ + B1(2) (5.19)
then
h i h i
(λ + µ)ξ ⊗ ξ + µ|ξ|2 1(2) · Aξ ⊗ ξ + B1(2) = 1(2) (5.20)
subsequently,
h i
A(λ + 2µ)|ξ|2 + B(λ + µ) ξ ⊗ ξ + Bµ|ξ|2 1(2) = 1(2) (5.21)
One can then determine the constant A and B,

(λ + µ)
A = − (5.22)
µ(λ + 2µ)|ξ|4
1
B = (5.23)
µ|ξ|2
Hence,
−1 |ξ|−2 n (λ + µ) o
(2)
Q(ξ) = ξ · C · ξ = − ξ ⊗ ξ + 1 (5.24)
µ µ(λ + 2µ)|ξ|2
or in component form,
−1 |ξ|−2 n (λ + µ) o
Qij = Kij = − ξi ξj + δ ij (5.25)
µ µ(λ + 2µ)|ξ|2
Consider
Nij (ξ)
ūi (ξ) = Qij (ξ)f¯j = −iCj`mn ¯∗mn ξ` (5.26)
D(ξ)
Applying Fourier inverse transform,
Z ∞
Nij (ξ)
ui (x) = −i Cj`mn ¯∗mn (ξ)ξ` exp(iξ · x)dξ (5.27)
−∞ D(ξ)
Z ∞
Nij (ξ)
= − f¯j (ξ) exp(iξ · x)dξ (5.28)
−∞ D(ξ)
Consequences of (5.27) are

1 ∞
Z
ij (x) = Ck`mn ¯∗mn (ξ)ξ` Nik (ξ)ξj + Njk (ξ)ξi D−1 (ξ)
2 −∞
· exp(iξ · x)dξ (5.29)
nZ ∞
σij (x) = Cijk` Cpqmn ¯∗mn (ξ)ξq ξ` Nkp (ξ)D−1 (ξ)
−∞
o
· exp(iξ · x) dξ − ∗ k` (x) (5.30)
5.2 Method of Green’s Functions

Consider
Z ∞
1
Gij (x − y) = Nij (ξ)D−1 (ξ) exp(iξ · (x − y))dξ (5.31)
(2π)3 −∞
Based on convolution theorem and according to (5.28) and (5.29), one can
derive that
Z ∞
ui (x) = − Cj`mn ∗ mn (y)Gij,` (x − y)dy (5.32)
−∞
Z ∞
ui (x) = Gij (x − y)fj (y)dy (5.33)
−∞
The corresponding expressions for stress and strain are
1 ∞
Z
ij (x) = − Ck`mn ∗ mn (y){Gik,`j (x − y) (5.34)
2 −∞
+Gjk,ì (x − y)}dy (5.35)
nZ ∞
σij (x) = −Cijk` Cpqmn ∗ mn (y)Gkp,q` (x − y)dy
−∞
o
+∗ k` (x) (5.36)
Eq.(5.37) is rewritten by Mura(1963) as the following form
Z ∞
σij (x) = Cijk` esth e`nh Cpqmn Gkp,qt (x − y)∗ sm dy (5.37)
−∞
To prove the equivalenct between (5.37) and (5.38), we use the identity esth e`nh =
δs` δtn − δsn δt` to expand (5.38),
Z ∞
σij (x) = Cijk` Cpqmn δs` δtn − δsn δtl Gkp,qt (x − y)∗ sm dy
Z−∞∞
= Cijk` Cpqmn Gkp,qn (x − y)∗ `m − Gkp,q` (x − y)∗ nm dy
−∞
(5.38)
The first term of the integrand is

Cpqmn Gkp,qn (x − y) = Gmnpq Gkp,qn (x − y) = −δmk δ(x − y) (5.39)
Therefore,
Z ∞
Cijk` Cpqmn Gkp,qn (x − y)∗ `m dy
−∞
Z ∞
= −Cijk` δ(x − y)∗ k` dy = −Cijk` ∗ k` (5.40)
−∞
We then recover (5.37).

Recall,
Z
1
G∞
ij (x − y) = 2 δ (x − y) · ξ Qij (ξ)dS (5.41)
8π S2
where Qij (ξ) = Nij (ξ)/D(ξ).

Substitute (5.42) into (5.34),
Z ∞ Z
1
ui (x) = − 2 δ (x − y) · ξ Qij (ξ)dS fj (y)dy
8π −∞ S 2
1
Z hZ ∞ i
= − 2 Qij (ξ) fj (y)δ(s − ym ξm )dy dS
8π S 2 −∞
Z
1
= − 2 Qij (ξ)fˆ(s, ξ)dS (5.42)
8π S 2
where s = xm ξm and
Z ∞
fˆj (s, ξ) = fj (y)δ(s − ym ξm )dy
−∞
is the Radon transform of fj (y).

Example 5.1 Assume that a linearly distributed eigenstrain is prescribed in
a spherical ball (|x| ≤ a).
1
(
∗ (ck x` + c` xk ) |x| ≤ a
k` = 2 (5.43)
0 |x| > a
Hence
1
∗ k`,j = ck δ`j + c` δkj (5.44)
2
and for isotropic materials
1
fi = −Cijk` ∗ k`,j = Cijkj ck + Cijj` c` = −(λ + 4µ)ci
2
The area of intersection of the plane ξm xm = s with the sphere of radus a is

π(a2 − s2 ), if |s| ≤ a and zero otherwise. Thus
Z ∞
ˆ
fj (s, ξ) = − (λ + 4µ)ci δ(s − xm ξm )dy
Z−∞
= − (λ + 4µ)ci dS = −(λ + 4µ)ci π(a2 − s2 )
S a ∩{ξm xm =s}
= −(λ + 4µ)ci π(a2 − (ξm xm )2 ) (5.45)
Therefore, the induced displacement field inside the sphere is
Z
(λ + 4µ)
ui (x) = Qij (ξ)cj (a2 − (ξm xm )2 )H(a2 − xm xm )dS (5.46)
8π 2 S2
where H(·) is the Heaviside function, ξm ξm = 1, and

1h (λ + µ)ξi ξj i
Qij (ξ) = δij −
µ (λ + 2µ)
(a) (b)
Figure 5.1. Illustraions of dislocations: (a)edge dislocation, and (b) screw dislocation
5.3 Application I: Dislocation problems

A dislocation is a distorted region among substantially perpect crystal lattice
environment. In other words, a dislocation is a linear defect around which
some of the atoms are misaligned or crystal lattice being distorted. There are
two types of dislocations: (1) edge dislocation, and (2) screw dislocation (see
Fig. 5.1). Use of eigenstrain theory to describe the effect of dislocations and
their induced disturbance mechanical fields is a success. Eigenstrain theory has
been an important approach in the development of dislocation theory. Here we
only introduce some simple examples.
Consider a straight screw dislocation on a half space. There is a jump or
discontinuity in displacement at x2 = 0 and −∞ < x1 < 0, with the magni-
tude of b(burgers vector). A ficticious eigenstrain field is prescribed on the slip
plane to mimic the mechanical effect of dislocation,

1
∗
23 = 2 bδ(x2 )H(−x1 ), x ∈ Ω (5.47)
0. x ∈ IR3 /Ω
where the slip surface may be described as
n o
Ω = (x1 , 0, x3 ) x1 < 0, −∞ < x3 < ∞
and H(·) is the heaviside function.
The eigenstrain field may be considered as the consequence of the displace-
ment field,
u∗3 (x) = bH(x2 )H(−x1 ) (5.48)
since
1 ∂u∗3 ∂u∗2 b
∗ 23 = + = δx2 H(−x1 )
2 ∂x2 ∂x3 2
(Question: what about ∗ 31 ?)
Apply Fourier transform
Z ∞
1
¯∗23 (ξ) = ∗ 23 (x) exp(−iξ · x))dx
(2π)3 −∞
Z ∞
1 b
= 3
δ(x2 )H(−x1 ) exp(−iξ · x)dx (5.49)
(2π) −∞ 2
Consider
Z ∞
δ(x2 ) exp(−iξ2 x2 )dx2 = 1 (5.50)
−∞
Z ∞ Z 0
H(−x1 ) exp(−iξ1 x1 )dx1 = exp(−iξ1 x1 )dx1
−∞ −∞
i
= Im(ξ1 ) < 0
ξ1
Z ∞
1
exp(−iξ3 x3 )dx3 = δ(ξ3 ) (5.51)
2π −∞
Therefore,
1 b i
¯∗23 = δ(ξ3 ) (5.52)
(2π)2 2 ξ1
Substituing (5.53) into the general formula of micro-elasticity,
Z ∞
ui (x) = −i Cj`mn ¯∗mn ξ` Qij (ξ) exp(iξ · x)dξ
−∞
Z ∞
= −2i Cj`23 ¯∗23 ξ` Qij (ξ) exp(iξ · x)dξ
−∞
2b Z ∞ δ(ξ )
3
= Cj`23 Qij (ξ) exp(iξ · x)dξ (5.53)
2(2π 2 ) −∞ ξ1
where the factor 2 is due to the presence of ∗ 32 , if the minor synmmetry is

being considered. For isotropic materials,
Figure 5.2. A screw dislocation
Cj`23 = λδj` δ23 + µ(δj2 δ`3 + δj3 δ`2 )

= µ(δj2 δ`3 + δj3 δ`2 )
The only non-zero components are C2323 = µ and C3223 = µ. Therefore,
b Z ∞ δ(ξ )
3
u1 (x) = C Q
2323 12 (ξ)ξ3 + C Q
3223 13 (ξ)ξ2
(2π)2 −∞ ξ1
exp(iξ · x)dξ
b Z ∞ δ(ξ )
3
u2 (x) = C 2323 Q22 (ξ)ξ3 + C3223 Q23 (ξ)ξ2
(2π)2 −∞ ξ1
exp(iξ · x)dξ
b Z ∞ δ(ξ )
3
u3 (x) = C 2323 Q32 (ξ)ξ3 + C3223 Q33 (ξ)ξ2
(2π)2 −∞ ξ1
exp(iξ · x)dξ
in which,
(λ + µ) ξ1 ξ2
Q12 (ξ) = −
µ(λ + 2µ) ξ 4
[(λ + 2µ)ξ 2 − (λ + µ)ξ22 ]
Q22 (ξ) =
µ(λ + 2µ)ξ 4
(λ + µ) ξ1 ξ3
Q13 (ξ) = −
µ(λ + 2µ) ξ 4
(λ + µ) ξ2 ξ3
Q23 (ξ) = −
µ(λ + 2µ) ξ 4
Q32 (ξ) = Q23 (ξ)
[(λ + 2µ)ξ 2 − (λ + µ)ξ32 ]
Q22 (ξ) = (5.54)
µ(λ + 2µ)ξ 4
Obviously,
Z ∞
δ(ξ3 )Q12 (ξ)ξ3 dξ3 = 0
Z−∞
∞
δ(ξ3 )Q13 (ξ)ξ2 dξ3 = 0
Z−∞
∞
δ(ξ3 )Q22 (ξ)ξ3 dξ3 = 0
Z−∞
∞
δ(ξ3 )Q23 (ξ)ξ2 dξ3 = 0
Z−∞
∞
δ(ξ3 )Q32 (ξ)ξ3 dξ3 = 0
Z−∞
∞
1 ξ2
δ(ξ3 )Q33 (ξ)ξ2 dξ3 =
−∞ µ (ξ12 + ξ22 )
Thereby, u1 (x) = u2 (x) = 0, and
Z ∞Z ∞
b ξ2
u3 (x) = exp i(ξ1 x1 + ξ2 x2 ) dξ1 dξ2
(2π)2 −∞ −∞ ξ1 (ξ12 + ξ22 )
b x
2
= tan−1 (5.55)
π x1
according to the inverse Fourier transform (Mura’s book page 17),
Z ∞Z ∞
ξ2
−1 x2

2 2 exp i(ξ x
1 1 + ξ x
2 2 ) dξ dξ
1 2 = 2π tan
−∞ −∞ ξ1 (ξ1 + ξ2 ) x1
5.4 Application II: Stress intensity factor for a flat

ellipsoidal crack
In late 1960s, John Willis used eigenstrain method solving a class of crack
and contact problems in anisotropic space.
In the following, we illustrate Willis’ solution procedure in the case of a 3D
ellipsoidal crack in an isotropic space.
Consider an ellisoidal crack embbeded in an infinite space. Suppose that the
crack region Ω is:
x21 x22
Ω: + 2 ≤ 1, and x3 = 0 . (5.56)
a21 a2
Figure 5.3. A three-dimensional ellipsoidal crack
For simplicity, we assume that the crack opening has the following form:
s
x21 x22
[u3 ] = b 1− − 2 χ(Ω) (5.57)
a21 a2
where parameter b is the Burger’s vecter, and χ(Ω) is the characteristic func-
tion of crack region, which can be defined as interpreted as

1, ∀x ∈ Ω
χ(Ω) = H(Ω − x) = (5.58)
0, ∀x ∈ IR3 /Ω
where H(·) is the Heavyside function.

This is equivalent to prescrib the following eigenstrain on the crack region,
s
x21 x22
∗33 = b 1− − 2 δ(Ω − x) . (5.59)
a21 a2
Therefore,
Z Z Z ∞ 0
∗ (x0 ) exp −iξ · (x − x0 dx =
−∞
s
x02 x02
Z Z
0
b 1− 1
− 2
exp(−iξ 3 x3 − iξ · (x − x ) dx01 dx02 (5.60)
Ω a21 a22
where in the second line, all vectors become 2D vectors, i.e. ξ = ξ1 e1 + ξ2 e2
and x = x1 e1 + x2 e2 .
Employ the fundamental formula of micro-elasticity,
Z ∞ Z ∞
i
ui (x) = Cj`mn ∗ (x0 )ξ` Nij (ξ)D−1 (ξ)
(2π)3 −∞ −∞
o
exp −iξ · (x − x0 ) dξ dx0 (5.61)
Changing the dummy indices i → k, j → p, m → 3, n → 3, ` → q, we have

s
Z ∞Z 0
ib x0 2 x22 ξq Nkp (ξ)
uk (x1 , x2 , 0) = Cpq33 1 − 2 − 2
(2π)2 −∞ Ω a1 a2 D(ξ)
0
· exp −iξ · (x − x )dΩx0 dξ (5.62)
and
s
Z ∞Z 0
b x0 2 x22 ξq ξ` Nkp (ξ)
uk,` (x1 , x2 , 0) = Cpq33 1 − 2 − 2
(2π)2 −∞ Ω a1 a2 D(ξ)
0
· exp −iξ · (x − x )dΩx0 dξ (5.63)
subsequently,
Z ∞
b Cijk` Cpq33 ξq ξ` Nkp (ξ)
σij = Cijk` uk,` =
(2π)3−∞ D(ξ)
Z s 02 02
x x 0
0 0
1 − 2 + 2 exp −iξ · (x − x ) dx1 dx2 (5.64)
Ω a1 a2
We first calculate the inverse Fourier transform along ξ3 , i.e. evaluating the
following integral,
Z ∞
Cijk` Cpq33 ξq ξ` Nkp (ξ) 0
exp(−ξ · (x − x )dξ3 . (5.65)
−∞ D(ξ)
Nkp (ξ) [(λ + 2µ)δkp ξ 2 − (λ + µ)ξk ξp ]
= (5.66)
D(ξ) µ(λ + 2µ)ξ 4
where the denominator may be decomposed into

2 q 2 q 2
ξ 4 = ξ12 + ξ22 + ξ32 = ξ3 − i ξ12 + ξ22 ξ3 + i ξ12 + ξ22 (5.67)
Since the problem is symmetric, we only consider the upper halp space (x3 >
0). Because the convergence requirement of Fourier transform, we are only
interested in the root with a negative imaginary part, i.e.
q
ξ3N = −i ξ12 + ξ22 (5.68)
which is a double root as shown in Eq. (5.67).

Suppose zj is a n-th pole of f (z), its residue is then
1 dn h n
i
Residue at (z = zj ) = lim (z − z j ) f (z) (5.69)
(n − 1)! z→zj dz n−1
Therefore, the integrand inside (5.65) is

∂ N 2 ξq ξ` Nkp (ξ) 0
Fijm = Cijk` Cpq33 (ξ3 − ξ3 ) exp(−iξ · (x − x )
∂ξ3 D(ξ)
(5.70)
After some tedious calculation, we find that at x3 = 0,
µ(λ + µ)
q
F333 = −i ξ12 + ξ22 . (5.71)
(λ + 2µ)
Hence,
Z Z s
bµ(λ + µ) x02 x02
σ33 (x1 , x2 , 0) = − 2 1 − 12 − 22
4π (λ + 2µ) Ω a1 a2
Z ∞ Z ∞
ξ exp(−iξ · (x − x0 )dξ1 dξ2 dx01 dx02 (5.72)
−∞ −∞
p
where ξ = ξ12 + ξ22 .
Let y1 = x1 /ap
1 , y2 = x2 /a2 ; ζ1 = a1 ξ1 , ζ2 = a2 ξ2 ; and η1 = ζ1 /ζ, η2 =
ζ2 /ζ, where ζ = ζ12 + ζ22 . Then
ξ · (x − x0 ) = ζ · (y − y0 ) (5.73)
dx0 dx0 dξ1 dξ2 = dy10 dy20 dζ1 dζ2 (5.74)
s 1 2
0 0
x12 x22
q p
0 0
1− 2 − 2 = 1 − y12 − y22 = 1 − y 02 (5.75)
a1 a2
s
η12 η22
q
ξ = ξ12 + ξ22 = ζ + (5.76)
a21 a22
Thus in Eq. (5.72)

Z ∞Z ∞
ξ exp(−iξ · (x − x0 )dξ1 dξ2
−∞ −∞
Z ∞Z ∞ s 2
η1 η22
0

= ζ + exp −iζη · (y − y ) dζ1 dζ2
−∞ −∞ a21 a22
Z 2π Z ∞ r
ζ1 2 ζ2 2 0

= ζ2 + exp −iζη · (y − y ) dζdφ (5.77)
0 0 a1 a2
Denote g = −η · y. The above integral becomes
Z 2π Z ∞ r
ζ1 2 ζ2 2 0

ζ2 + exp −iζη · (y − y ) dζdφ
0 0 a1 a2
2 Z 2π r
∂ ζ1 2 ζ2 2
= − 2 + exp(iζ(g + η · y0 )dηdφ
∂g 0 a1 a2
ζ1 2 ζ2 2 ∂ 2 ∞
Z 2π r
∂2
Z
= − 2 + − 2 exp(iζ(g + η · y0 )dζ dφ
∂g 0 a1 a2 ∂g 0
Z 2π r 2
ζ1 2 ζ2 2 ∂ −i
= + 2 g + η · y0
dφ (5.78)
0 a 1 a 2 ∂g
Denote η · y0 = y 0 cos(θ − φ) and consider following integral identity,
Z 2π
d(θ − φ) 2π
0 cos(θ − φ)
=p . (5.79)
0 g + y g − y02
2
p ( 0
)
2π
y 0 1 − y 02 dy
∂2
Z r Z
ibµ(λ + µ) η1 2 η2 2 0 0
σ33 = + 0 dy1 dy2
x3 =0 2π(λ + µ) 0 a1 Ω g + y cos(θ − φ)
a2 ∂g 2
( Z 1 0p 0
)
ibµ(λ + µ) 2π η1 2 η2 2 ∂ 2
r
y 1 − y 02 dy
Z
= + (5.80)
∂g 2 0
p
2π(λ + µ) 0 a1 a2 g2 − y02
Let p 0
1
y0 1 − y 02 dy
Z
I= p
0 g2 − y02
Change of variable
0 g2 − 1 1 2
y2 =1− w− (5.81)
4 w
One can show that
∂2I 1 g+1 g
= − ln + . (5.82)
∂g 2 2 g − 1 g2 − 1
Interior solution (y < 1):

When y < 1 x3 = 0, it is crack region. Obviously |g| = |η · y| < 1. Since
g+1 1 + g 1 + g
=− = exp(−iπ)
g−1 1−g 1−g
then,
∂2I 1 1+g π g
= − ln −i + . (5.83)
∂g 2 2 1−g 2 g2 − 1
Both ln |(1 + g)/(1 − g)| and g/(g 2 − 1) are odd function of φ, whereas
1/2
cos2 φ/a21 + sin2 φ/a22 is an even function of φ.
Hence when y < 1
bµ(λ + µ) 2π cos2 φ sin2 φ 1/2

Z
σ33 (x1 , x2 , 0) = − + dφ
4(λ + 2µ) 0 a21 a22
bµE(k)
= − (5.84)
2a2 (1 − ν)
where
π/2
a21 − a22
Z
E(k) = (1 − k 2 sin2 φ)1/2 dφ, k 2 := . (5.85)
0 a21
If
0 bµE(k)
σ33 (Ω) = −σ33 =− (5.86)
2a2 (1 − ν)
it then links the Burgers’ vector with the prescribed stress on the crack surfaces,
0
2(1 − ν)a2 σ33
b= . (5.87)
µE(k)
This suggests that the type of prescribed eigestrain is equivalent to prescribed
constant stress on crack surfaces.
Exterior solution:
We are only interested the asymptotic solution, i.e. y → 1. When y → 1,
the term |g/(g 2 − 1)| > ln|(g + 1)/(g − 1)| → ∞ is the leading term of
asymptotic expansion. Therefore
ibµ(λ + µ) 2π η1 2 η2 2 gdφ
Z r
σ33 (x1 , x2 , 0) = + + O(1) (5.88)
2π(λ + µ) 0 a1 a2 g 2 − 1
Let
η2 η22 1/2 y
1
f (η) = g + , and ŷ = . (5.89)
a21 a22 y
Figure 5.4. The shortest distance between the crack surface and a point
ibµ(λ + µ) 2π f (η − f (ŷ)
Z Z 2π
1 f (ŷ)
σ33 (x1 , x2 , 0) = dφ + dφ
2π(λ + µ) 0 g2 − 1 2π 0 g 2 − 1
Z 2π
1 f (ŷ)
= dφ + O(1) (5.90)
2π 0 g 2 − 1
Assume that g = −η · y = y cos ψ. Then
Z 2π Z 2π
dφ d(φ − ψ) −2π 2iπ
2 2 2
=p =p . (5.91)
0 g − 1 0 y cos (φ − ψ) − 1 1−y 2 y2 − 1
and
bµ(λ + µ) ŷ · y ŷ1 ŷ2 1/2
σ33 (x1 , x2 , 0) = + (5.92)
(λ + 2µ) y 2 − 1 a21 a22
p
y→ŷ
The stress intensity factor is defined as

k1 := lim (2πr)1/2 σ33 (5.93)
r→0
For an ellipsoidal crack,
(y − 1)y 2
r= 2 (5.94)
x1 x22 1/2
+ 2
a21 a2
and
√ √
2πbµ(λ + µ) y − 1 x21 x22 1/4
k1 = + 4 , y→1 (5.95)
y 2 − 1 a41
p
(λ + 2µ) a2
0 /(µE(k)) into the above expression, one has
Substituting b = (2(1 − ν)a2 σ33
√ 0 x2
πa2 σ33 1 x22 1/4
k1 = + . (5.96)
E(k) a41 a42
5.5 Isotropic inclusion-Eshelby’s solution

From 1957 to 1961, J.D.Eshelby published three landmark scientific papers
systematically solving inclusion problem in an elastic medium.
Eshelby’s ellipsoidal inclusion problem is stated as follows: Find induced
displacement, strain, and stress fields by an ellipsoidal incluseion, Ω, embed-
ded in an isotropic unbounded elastic medium, in which a uniform eigenstrain
is prescribed, i.e. ∗
∗ ij , x ∈ Ω
ij (x) = (5.97)
0, x ∈ IR3 /Ω
Using the fundamental formula of micro-elasticity,
Z ∞
ui (x) = − Cj`mn ∗ mn (y)G∞ ij,` (x − y)dΩy
−∞
Z
= −∗ mn Cj`mn G∞ij,` (x − y)dΩy
Ω
For isotropic elastic materials,

−1 n δmi zn + δni zm − δmn zi z m z n zi o
Cj`mn G∞
ij,` (z) = (1 − 2ν) + 3
8π(1 − ν) z3 z5
gimn (`)
= (5.98)
8π(1 − ν)|z|2
where z = x − y and ` = −z/|z|, and
gimn (`) = (1 − 2ν)(δmi `n + δni `m − δmn ì ) + 3`m `n ì (5.99)
5.5.1 Interior solution

Consider x ∈ Ω. Let z = |z. Take a radon decomposition centering around
a the point x dΩy = dzdS = z 2 dzdw, where dw is the volume angle on S 2 .
We can rewrite displacement field as
−∗ mn
Z
dΩy
ui (x) = gimn (`)
8π(1 − ν) Ω |x − y|2
−∗ mn
Z rZ
= gimn (`)dzdw
8π(1 − ν) 0 S 2
−∗ mn
Z
= r(`)gimn (`)dzdw (5.100)
8π(1 − ν) S 2
where vector r = y − x, y ∈ ∂Ω and the scalar r(`) is the distance between
the point x and a point y on the surface of the ellipsoidal in the direction of
r. In other words, r(`) is the distance between x and the interseption point of
straight line y = x + r, y ∈ IR3 and the surface of the ellipsoidal. To find such
Figure 5.5. An ellipsoidal inclusion
interception point along a fixed direction of `. We assume that the interception

point is marked as x0 . Since it must be on both the straight line, x0 = x + r,
i.e.  0
 x1 = x1 + r`1
x0 = x2 + r`2 (5.101)
 20
x3 = x3 + r`3
and on the surface of the ellipsoidal
0 0 0
x12 x22 x32
+ 2 + 2 =1 (5.102)
a21 a2 a3
One can substitute (5.139) into (5.140). For fixed point x and a fixed direction
`, it yields a quadratic equation,
(x1 + r`1 )2 (x2 + r`2 )2 (x3 + r`3 )2
+ + =1 (5.103)
a21 a22 a23
of unknown variable, r(`). More explicitly,
`2 `22 `23 x ` x2 `2 x3 `3
1 1
r2 1
+ + + 2r + + 2
a21 a22 a23 a21 a22 a3
h x2 x2 x2 i
+ 1
+ 2
+ 3
− 1 = 0, ⇒ gr2 + 2rf − e = 0 (5.104)
a21 a22 a23
where
`22
`2 `23
1
g := ++ (5.105)
a21
a22 a23
x ` x2 `2 x3 `3
1 1
f := + + 2 (5.106)
a21 a22 a3
x2 x2 x2
e := 1 − 21 + 22 + 23 (5.107)
a1 a2 a3
Eq. (5.142) has two roots,
f f2 e 1/2
r(`) = − ± 2 + (5.108)
g g g
f2 e 1/2
Since + is even in `, while gimn (`) is odd in `,
g2 g
Z 2
f e 1/2
2
+ gimn (`)dw = 0 (5.109)
S2 g g
Let λ1 = `1 /a21 ,λ2 = `2 /a22 and λ3 = `3 /a23 . We have
∗ mn
I
f
ui (x) = gimn (`)dw
8π(1 − ν) S 2 g
∗ mn
I
x` λl
= gimn (`)dw
8π(1 − ν) S 2 g
∗ mn x`
I
λ`
= gimn (`)dw (5.110)
8π(1 − ν) S 2 g
Then
∗ mn δ`j
I
λ`
ui,j (x) = gimn (`)dw
8π(1 − ν) S 2 g
∗ mn
I λ
j
= gimn (`)dw (5.111)
8π(1 − ν) S 2 g
One can find induced elastic strain field by symmetrizing the elastic distor-
tion,
∗ mn
I
1 λi gjmn + λj gimn
ij = (ui,j + uj,i ) = dw (5.112)
2 16π(1 − ν) S 2 g
ì
where λi = is the component of the normalized vector λ = λi ei and
a2i
g = λ · λ = λ2 .
Figure 5.6. Illustration of integration scheme over an ellipsoidal
Consider
gijk (`) = (1 − 2ν)(δij lk + δi` `k − δjk ì ) + 3ì `j `k = gik` (`) (5.113)
The last two indices of the third order tensor gijk is symmetric. We can then
define a fourth order symmetric tensor,
I
Ω 1 λi gjmn + λj gimn
Sijmn := dw (5.114)
16π(1 − ν) S2 g
This leads to the long anticipated result,

ij (x) = or dij (x) = Sijmn
Ω
∗ mn (5.115)
It is obvious that
Ω Ω Ω
Sijmn = Sijnm = Sjimn
where the superscript indicates that the Eshelby tensor is for induced strain
field inside the ellipsoidal, Ω.
Remark 5.5.1 The most amazing fact of this result is that the induced strain
field and stress field inside the inclusion are uniform, and the Eshelby tensor
for any ellipsoidal shape of inclusion is a constant tensor.
Define the following elliptic integrals

Z ∞
`2i dw
Z
ds
II (0) = 2g = 2πa a a
1 2 3 2 + s)∆(s) (5.116)
S 2 a i 0 (aI
`2i lj2 dw
Z
IIJ (0) = 3 2 2
S 2 ai aj g
Z ∞
ds
= 2πa1 a2 a3 (5.117)
0 (aI + s)(a2J + s)∆(s)
2
JIJ (0) = a2I IIJ − IJ (5.118)

p
where ∆(s) = (a21 + s)(a22 + s)(a23 + s) and argument (0) indicating the
lower limit of the elliptic integrals are zero.
One can show that Eshelby tensor can be explicitly expressed by these inte-
grals through the following identity,
Ω
8π(1 − ν)Sijk` = δij δk` (2νII (0) + JIK (0)) + (δik δk` + δjk δi` )

(1 − ν)(Ik (0) + IL (0)) + JIJ (0) (5.119)
where the upper case indices are not summed with lower case indices.
Ω , we consider
Example 5.2 To compute S1111
Ω
8π(1 − ν)S1111 = 2νI1 (0) + J11 (0) + 2(1 − ν)2I1 (0) + 2J11 (0)
= (4 − 2ν)I1 (0) + 3J11 (0)(a21 I11 (0) − I1 (0))
= (1 − 2ν)I1 (0) + 3a21 I11 (0) (5.120)
which leads to
Ω 3a21 (1 − 2ν)
S1111 = I11 (0) + I1 (0) (5.121)
8π(1 − ν) 8π(1 − ν)
The integral II (0) and IIJ (0) can be expressed in terms of standard elliptic
integrals. For example, assuming a1 > a2 > a3 , we have
4πa1 a2 a3
I1 (0) = {F (θ, k) − E(θ, k)}
(a21 − a22 )(a21 − a22 )1/2
4πa1 a2 a3 n a (a2 − a2 )1/2 o
2 1 3
I3 (0) = − E(θ, k)
(a22 − a23 )(a21 − a23 )1/2 a1 a3
where
Z θ
dt
F (θ, k) =
0 (1 − k 2 sin2 t)1/2
Z θ
E(θ, k) = (1 − k 2 sin2 t)1/2 dt (5.122)
0
h i1/2
and θ = sin−1 (1 − a23 /a21 )1/2 , k = (a21 − a22 )/(a21 − a23 ) .
In applications, the following invariant formulas are very useful,
I1 (0) + I2 (0) + I3 (0) = 4π

3I11 (0) + I12 (0) + I13 (0) = 4π/a21
3a21 I11 (0) + a22 I12 (0) + a23 I13 (0) = 3I1
I12 (0) = (I2 (0) − I1 (0))/(a21 − a22 )
When the ellipsoidal becomes a sphere, Eshelby tensor become simple num-
bers. Let a1 = a2 = a3 = a. We have
4π
IIs =
3
s 4π 1
II,J =
5 a2
s 8π
JIJ = −
15
and hence
5ν − 1 2(4 − 5ν)
Ω
Sijk` = δij δk` + (δik δj` + δjk δi` ) (5.123)
15(1 − ν) 15(1 − ν)
A remarkable property of the Eshelby tensor for spherical inclusion is that

it does not depend on its size, i.e. it does not depend on its radius a. This
implies that no matter how large or how small spherical inclusions are, they
share the same Eshlby tensor. In other words, there is no embeded length
scale or scaling factor for spherical inclusion. This property will lead to some
remarkable consequences in ensuing homoginization process.
For other specified shape of ellipsoidal inclusions, readers may consult Mura’s
book for detailed information. A systematic documentation on Eshelby’s ten-
sor in various cases can be found in Mura [1987].
5.6 Exterior Solution of Ellipsoidal Inclusion

For x ∈/ Ω, the exterior disturbance displacement and strain fields due to
eigenstrain distribution had been also found by Eshelby, though evaluation of
the induced exterior displacement fields and strain fields are often difficult.
Suppose that eigenstrain distribution inside the ellipsoid is constant. For any
point x ∈ IR3 , we have
Z
ui (x) = −Cjkmn ∗ mn Gij,k (x − x0 )dΩx0 (5.124)
Ω
where
−1
Cj`mn G∞ 0
ij,` (x − x ) = ·
8π(1 − ν)
n δmi (xn − x0n ) + δni (xm − x0m ) − δmn (xi − x0i )
(1 − 2ν)
|x̄|3
(xm − xm )(xn − xn )(xi − x0i ) o
0 0
+3
|x̄|5
−1 n ∂3 h ∂ δ
mi
= |x̄| − 2(1 − ν)
8π(1 − ν) ∂xi ∂xm ∂xn ∂xn |x̄|
∂ δni i ∂ 1 o
− 2νδmn (5.125)
∂xm |x̄| ∂xi |x̄|
Introduce the following potential functions,
Z
ψ(x) = |x − x0 |dΩx0 (5.126)
ZΩ
1
φ(x) = 0
dΩx0 (5.127)
Ω |x − x |
where ψ(x) is the biharmonic potential, whereas φ(x) is the Newtonian poten-
tial. This is because of the fact

 −8π x ∈ Ω
∇4 ψ = 2∇2 φ = (5.128)
0 x ∈ IR3 /Ω

To verify Eq. (5.166), one can show first
∂2
Z
2
∇ ψ = |x − x0 |dΩx0
∂x2` Ω
(x` − x0` )(x` − x0` ) o
Z n
δ``
= − dΩx0
Ω |x̄| |x̄|3
Z
2
= dΩx0 = 2φ(x) (5.129)
Ω |x̄|
Subsequently,
∇4 ψ = ∇2 ∇2 ψ = 2∇2 φ
Z Z
1 1
= 2 ∇ dΩ = 8π ∇2 dΩ
Ω |x̄| Ω 4π|x̄|
Z
= 8π ∇2 GL (x − x0 )dΩx0 (5.130)
Ω
where GL (x − x0 ) is the Green’s function for three-dimensional Laplace equa-

tion, i.e.
∇2 GL (x − x0 ) + δ(x − x0 ) = 0 (5.131)
Consequently,
Z
4 2
∇ ψ = 2∇ φ = 8π δ(x − x0 )dΩx0
 Ω
 −8π x ∈ Ω
= (5.132)
0 x ∈ IR3 /Ω

We can then express induced displacement as

Z
ui (x) = − ∗ mn Cj`mn Gij,` (x − x0 )dΩx0
Ω
∗ mn n ∂3 ∂ ∂
= ψ − 2(1 − ν) δmi + δni φ
8π(1 − ν) ∂xi ∂xm ∂xn ∂xn ∂xm
∂ o
−2νδmn φ (5.133)
∂xi
Similarily for elastic distortion field and strain field,
∗ mn
ui,j (x) = ψ,mnij − 2(1 − ν)(δmi φ,nj + δni φ,mj )
8π(1 − ν)

−2νδmn φ,ij (5.134)
1 ∗ mn
ij (x) = (ui,j + uj,i ) = {ψ,mnij − 2νδmn φ,ij
2 8π(1 − ν)
− (1 − ν)(δmi φ,nj + δni φ,mj + δmj φ,ni + δnj φ,mi } (5.135)
One can rewrite the above expression in a succinct manner,

∞
dij (x) = Sijk` (x)∗ k` , ∀x ∈ IR3 /Ω (5.136)
∞ (x).
which defines the exterior Eshelby tensor, Sijk`
∞ 1
Sijk` (x) = ψ,ijk` (x) − 2νδk` φ,ij (x)
8π(1 − ν)
−(1 − ν)(δki φ,`j (x) + δì φ,kj (x)

+δkj φ, ì(x) + δ`j φ,ki (x)) (5.137)
It depends on where the tensor is being evaluated.

The derivatives of Newtonian potential and biharmonic potential can be also

expressed by elliptic integrals. For instance,
φij (x) = −δij II (λ) − xi IIJ (λ) (5.138)
ψ,ijk` (x) = −δij (xk IIK (λ)),` + (xi xj IIJ (λ)),k` (5.139)
where
Z ∞
ds
II (λ) = 2πa1 a2 a3 (5.140)
λ (a2I+ s)∆(s)
Z ∞
ds
IIJ (λ) = 2πa1 a2 a3 (5.141)
λ
2
(aI + s)(a2J + s)∆(s)
JIJ (λ) = a2I II J(λ) − IJ (λ) (5.142)
where λ is zero when x ∈ Ω and λ is the largest positive root of the following
equation,
x21 x22 x23
+ + =1 (5.143)
(a21 + λ) (a22 + λ) (a23 + λ)
∞ (x) with elliptic integrals is
A very useful identity that related Sijk`
∞ Ω
8π(1 − ν)Sijk` (x) = 8π(1 − ν)Sijk` (λ)
h
+(1 − ν) δi` xk IK,j (λ) + δk` IK,i (λ)
i
+δik IL,j (λ) + δjk x` IL,i (λ)
δij xk JIK,` (λ) + (δik xj + δjk xi )JIJ,` (λ)
(δi` xj + δj` xi )JIJ,k (λ)
+xi xj JIJ,k` (λ) (5.144)
where
Ω
8π(1 − ν)Sijk` = δij δk` (2νII (λ) + JIK (λ)) + (δik δk` + δjk δi` ) ·

(1 − ν)(Ik (λ) + IL (λ)) + JIJ (λ) (5.145)
when x ∈ Ω, Eq. (5.144) becomes (5.145). Ju and Chen [1994] developed a
more simple and explicit way to evaluate exterior Eshelby tensor. From
Z
ui (x) = − Cjkmn Gij,` (x − y)∗ mn (y)dΩy (5.146)
Ω
one may derive that
Z
1
ij (x) = − Ck`mn Gik,`j (x − y) + Gjk,ì (x − y) ∗ mn (y)dΩy
2 Ω
Z
= Gijmn (x − y)∗ mn (y)dΩy (5.147)
Ω
where
1
Gijmn (x − y) = − Ck`mn Gik,`j (x − y) + Gjk,ì (x − y)
2
1 h
= (1 − 2ν)(δim δjn + δin δjm − δij δmn )
8π(1 − ν)r3
+3ν(δim `j `n + δin `j `m + δjm ì `n + δjn `j `m )
i
+3δij `m `n + 3(1 − 2ν)δmn ì `j − 15ì `j `m `n (5.148)
where Gijmn is called the fourth order Green’s function (the second derivative
of the Green’s function).
If ∗ mn (x) is constant inside the inclusion, the exterior Eshlby tensor can
be defined as
Z
∞
Ḡijmn (x) := Gijmn (x − y)dΩy = Sijmn (5.149)
Ω
For a spherical inclusion (a1 = a2 = a3 = a), one may find that

4πa3 4πa3 a2
φ= , and ψ = |x| + . (5.150)
3|x| 3 5|x|
The exterior Eshelby tensor can then be obtained by straighttfoward differen-
tiation,
ρ3 h
Ḡijmn (x) = (3ρ2 + 10ν − 5)δij δmn
30(1 − ν)
+(3ρ2 − 10ν + 5)(δim δjn + δim δjn )
+15(1 − ρ2 )δij `m `n + 15(1 − 2ν − ρ2 )δmn ì `j
+15(ν − ρ2 )(δim `j `n + δin `j `m + δjm ì `n + δjn ì `m )
i
∞
+15(7ρ2 − 5)ì `j `m `n = Sijmn (5.151)
where ρ = a/r. Note that when r → a, Sijmn ∞ Ω

6→ Sijmn , which indicates that
both disturbance strain field is not continuous across the interface of the matrix
and inclusion.
In fact, when ρ → 1,
[ex] [in] ∞
[ij ] := ij − ij = (Sijmn Ω
− Sijmn )∗ ij
−1 h 1
= νδmn ì `j + δim `j `n + δin `j `m
(1 − ν) 2
i
+δjm ì `n + δjn ì `m − ì `j `m `n mn (5.152)
which is the weak discontinuity at the interface between matrix and inculsion.
5.7 Jock Eshelby (II): Lessons from J.D.Eshelby

The measure of your education is what you remember 15 years afterward,
says one wiseacre. Well, it’s been a little more than 15 years, and I don’t think
that I learned anything at the time, but the lectures I had from Professor J.D.
(Jock) Eshelby still leave a mark.
Undergraduate students in materials science at Sheffield University were
barely aware of the towering stature of this man, in the intellectual sense any-
way. If you don’t know who he was or what contributions he has made, then
you probably have some serious holes in your own materials education, but
you can still read on. A few Britishisms must be explained, though. First,
the term "Jock" is used in the United Kingdom not for an athlete, as in the
United States, but is a nickname commonly accorded to Scotsmen living in
England; the U.S. sense could never apply to Jock Eshelby. Second the term
"Faculty" in England is equivalent to a college in a U.S. University. Third, a
professorship in the United Kingdom is a distinguished academic rank that has
almost no equivalent in the United States. The closest would be a "leading
professorship".
Way back then, Sheffield had a Faculty of Materials, with departments of
Metallurgy, Ceramics, Glasses, Polymers, and the theory of materials. The
department of the theory of materials was arguably a little top heavy. It had
two professors, Eshelby and B.A.Bilby (whose name you should also know),
one other lecturer, and a computer programmer. In a good year it had one
undergraduate student.
Eselby taught courses in elasticity and solid state bonding to the undergrad-
uates in all of the departments, and his lecturing style was not particularly
student-friendly. He did not work from notes. He would walk into the lecture
hall, apparently already half-way through this lectur, pick upthe chalk, and
start writing on the board. Whether he was trying to show us how to solve
Schrodinger’s equation or develp the strain compatibility relations, the tech-
nique was always the same. He would clear a patch of board and start deriving
a theorem. Running out of space, he would clear another patch, not neces-
sarily connected with the first, and fill that up. Eventually, small pieces of
the theorem would be scattered more-or-less at random across the chalkboard,
stochastically mixed with the detritus of the previous lecture, and with random
parts missing–erased to make space for more. It did not help that his writing
was arocious, and his speech sounded as though he had filled his cheeks with
marbles before starting. On one occation, one of my classmates managed to
ge the professor’s attention (a challenge) and asked him if he could possibly
write a little more clearly. For a few lines, the writing was four times as large,
but still as ilegible as before. Several lectures ended with Eshelby’s discovery
that he had misderived the theorm in question–a significant risk if you try to
do it without notes, even if you are a bona fide genius. When this happened,
he would stand back and survey the board. Agter a few moments, he would
announce something like, "Well, there’s a sign error there. You can correct it
and work through to the result for yourselves." As if.
As time went by, our horror at his teaching style gave way to an understand-
ing that the man was, in fact, a genius. Eccentric, yes but a genius. Apparently
addicted to cheap cigars, he would smoke them down to the smaalest butt, then
draw a cherry pipe out of his pocket, and stuff the remains of the cigar into it,
tob e smoked until not a scrap of tobacco was left. He cared little for what peo-
ple thought of hom, I think, and did not pay much attention to the politics of
academia and the scientific community. This resultd in anunconscionalbe de-
lay in his being elevated to the rank of Fellow of the Royal Society, which does
seem to have been a sore point. In one memorable lecture, he described all of
the current theories on a particular topic, listing the names of their authors on
an uncharacteristically cleared chalkboard. He then described what was wrong
with each of their work, condemning the weak-mindedness of these "so-called
scientists" in quite direct terms. Having disposed of their failed logic, he then
wrote the magical letters "FRS" after each of the names. He was elected an
FRS himself that year and did not repeat the performance as far as I can gather.
Eshelby’s impact on material sience is far, far out of proportion to the num-
bers of his publications. In total, he published less than 20 papers over his
entire career (This is not true by the way. Eshelby published alomost 50-to-
60 papers in his lifetime, but the point is valid: this days, you can see a lot
of mediocre people published hundreds of junks, and good papers can not be
published–Li’s comment), but each of them is a classic. A fine demonstraion
of the futility of today’s obsessiion with publication-counting as a means of
career assessment. Eshelby’s work is characterized by real physical insight,
complemented by elegant mathematical analysis (He was a professor of ap-
plied mathematics at Sheffield, in addition to being a professor of the theory of
materials.) In contrast with his lectures, his written work is a modl of clarity.
Although he was a powerful mathematician, he felt that we should only engage
in "mathematical weightlifting" if we could not reason our way to the desired
result through simple physical logic. Goodness knows what he would have
made of today’s computer simulation techniques. I think he would probably
have thought of them as the last desperate resort after both physical reasoning
and mathematical analysis failed.
An insight into Eshelby’s motivations was provided to us in an informal mo-
ment one day, sitting in the small but splendid museum of glassware belonging
to the Faculty of Materials, in a traditional British tea break. The usually unap-
proachable Esheby was unusually affable that day–perhaps he had just receibed
word of his FRS election–but we fell into conversation and one undergraduate
student adked him what had led to his being a "pure theoretician". He told
us the story of a formative experience in his life. It seems that as a young
teenager he had made a calculation of the thermal shock resistance of a piece

of glass. This resulted from his mother’s always using a thick cork pad beneath
a coffee table. She explained the reason to him and he set to work calculationg
the effect of the anticipated thermal shock. A short while later, he came to his
mother and announced that he had completed his analysis, and that table would
withstand a sudden local rise to the boiling point of water. His mother, being a
wise woman, advised him that the obvious experiment would not be forthcom-
ing and that he was forbidden from performing it himself. Well, curiosity and
the budding scientific mind got the better of his youthful judgement one day
when he was alone in the house. He boiled a pan of water and place it at the
center of the prized coffee table. In his own words, "Well, cracks flew in ev-
ery direction, and I suddently received a discouragement that from performing
experiments that has lasted me the rest of my life."
True to the creed of the theoretician, however, he refused to allow that the
analysis was flawed, and instead blamed the experiment. " Of course, I knew
immediately what was wrong. The d***d thing hadn’t been annealed properly.
It was FULL of residual stress!"
By all accounts, this attack on the quality of the prized table did not endear
him to his mother. Let all theorists beware of blaming the experiment lest they
suffer similarly.
—By Alex King(From MRS Bulletin, July, 1999)
5.8 Exercises
Probelm 5.1 Show that the integral
r
π 3 J3/2 (η)
Z
exp{iξ · x}dVx = 4π a (5.153)
V0 2 η 3/2
p
where V0 is a sphere with radius a; η = a|ξ|, and |ξ| = ξ12 + ξ22 + ξ32 .
Hint:
(1)Consider the identity

∇x exp iξ · x = iξ exp iξ · x
(2) Z 1
t sin(a|ξ|t)dt = Γ(1)(a|ξ|)−1/2 J3/2 (a|ξ|)
0
r
π
where Γ(1) = , J (η) is the Bessel function of the first kind.
2 3/2
Probelm 5.2 Derive the displacement field inside an inclusion in which pre-
scribed eigenstrain is a linear functin of coordinates, i.e. Example 5.1.
Probelm 5.3 Derive Green’s functin for plane strain problem by solving the
following Navier equations,
σβα,β + δ(x − y)δαγ = 0 (5.154)
where γ is the direction that the concentrated force point at.
Assume that the 2D elastic tensor is
Cαβζη = λδαβ δζη + µ(δαζ δβη + δαη δβζ ), α, β, ζ, η = 1, 2 (5.155)
define 2D permutaion symbol
eαβ : e11 = 0, e12 = 1, e21 = −1, e22 = 0 (5.156)
The corresponding e-δ identities are:
δα1 δα2
(1) eαβ =
δβ1 δβ2
(2) eαζ eβη = δαβ δζη − δαη δβζ
(3) eαη eβη = δαβ (5.157)
(4) eαη eαη = δαα = 2! (5.158)
Hints:
Z ∞ Z ∞
exp(i(ξ1 x1 + ξ2 x2 ))
dξ1 ξ2 = −2π ln R (5.159)
ξ12 + ξ22
Z−∞ −∞
∞ Z ∞
ξα ξβ xα xβ
4
exp(iξ · x)dξ = −πδαβ ln R − π 2 (5.160)
−∞ −∞ ξ R
p
where R = x21 + x22 .
Probelm 5.4 Let Ω be the half plane (x2 = 0, x1 < 0), and ∗21 be pre-
scribed as
b
∗21 (x) = δ(x2 )H(−x1 ) (5.161)
2
Show
b x
2 b 1 x1 x2
u1 (x) = tan−1 + (5.162)
2π x1 4π 1 − ν x21 + x22
where ν is the Poisson’s ratio.
Hint: (Mura’s book page 17)
Z ∞Z ∞
ξ2 −1 x2

exp{i(ξ1 x1 + ξ2 x2 )}dξ1 dξ2 = 2π tan
ξ1 (ξ12 + ξ22 ) x1
Z−∞∞ Z−∞∞
ξ1 ξ2 πx1 x2
2 2 2 exp{i(ξ1 x1 + ξ2 x2 )}dξ1 dξ2 = − x2 + x2
−∞ −∞ (ξ1 + ξ2 ) 1 2
Figure 5.7. A straight edge dislocation
Probelm 5.5 Verify the following Hilbert transform formulas

1 x
H = (5.163)
(x2 + a2 ) a(x2 + a2 )
H(sin(bx)) = − cos(bx) (5.164)
Hints: use Cauchy’s residue theorm.
Probelm 5.6 Derive Eqs. (5.131), (5.132) and (5.134). Start from (5.111).
Hints:
Hirth and Lothe [1992] Theory of Dislocations, Reprint Edition, Krieger
Publishing Co. pages 228,235-237
Cottrell, A.H. [1953] Dislocations and plastic flow in crystals, Oxford Uni-
versity Press. pages 62-64, 98
Probelm 5.7 The 2D Green’s function for plane strain problem is
( )
(x − x 0 )(x − x0 )
1 1 α α β β
Gαβ (x−x0 ) = − (3 − 4ν)δαβ ln R α, β = 1, 2
8π µ(1 − ν) R2
(5.165)
where R = (x1 − x01 )2 + (x2 − x02 )2 .
p
Consider the following elliptical inclusion problem,

 ∗
 αβ ; ∀ x∈Ω
∗
αβ (x) = (5.166)
0; ∀ x ∈ IR2 /Ω

x21 x22

where ∗αβ is a constant tensor, and Ω := x + 2 ≤1 .
a21 a2
Figure 5.8. 2D elliptical inclusion
Find the Eshelby tensor for interior problem (x ∈ Ω). Hint (see Li (2000)
pages 5606-5607 ).
Probelm 5.8 Consider a spherical inclusion with radius a. Use identities
4πa2
I
ì `j dS = δij (5.167)
S2 3
4πa2
Z
ì `j `m `n dS = (δij δmn + δim δjn + δin δjm ) (5.168)
S2 15
to show that
I
Ω 1 λi gjmn + λj gimn
Sijmn = dS
16π(1 − ν) S 2 g
5ν − 1 2(4 − 5ν)
= δij δmn + (δim δjn + δjm δin ) (5.169)
15(1 − ν) 15(1 − ν)
where gijk = (1 − 2ν)(δij `k + δik `j − δjk ì ) + 3ì `j `k , g = ì ì /a2 = a−2 ,

and λi = ì /a2 .
Probelm 5.9 Show that

1
G(x − y) = − Cklmn Gik,lj (x − y) + Gjk,li (x − y)
2
1 h
= (1 − 2ν)(δim δjn + δin δjm − δij δmn )
8π(1 − ν)r3
+3ν(δim lj ln + δin lj lm + δjm li ln + δjn lj lm )
i
+3δij lm ln + 3(1 − 2ν)δmn li lj − 15li lj lm ln (5.170)
where Gijmn is called the fourth order Green’s function (the second derivative
of the Green’s function).
Effective Elastic Modulus 95
Chapter 6
EFFECTIVE ELASTIC MODULUS
We now present Eshebly’s equivalent eigestrain theory and its related engi-
neering homogenization methods.
6.1 Effective elastic moduli for composites of dilute

suspension
First, we apply the engineering homogenization theory to composites whose
second phase concentration or other phase concentrations are small in compar-
ison with the concentration of the matrix. In literature, we usually refer this as
the composite with inhomogeneities of dilute suspension.
6.1.1 Basic equations for average stress and strain

Consider a solid with multiple phases of inhomogeneities, α = 1, 2, · · · , n.
The elastic tensor and compliance tensor in the matrix is denoted as C and D,
and the elastic tensors and compliance tensors in the heterogeneous phases are
denoted as Cα and Dα where α = 1, 2 · · · , n.
Define the averge stress and average strain in the matrix and in the inclu-
sions,
Z Z
1 1
< σ >M := σdV, < >M := dV (6.1)
M M M M
Z Z
1 1
< σ >α := σdV, < >α := dV (6.2)
Ωα Ωα Ωα Ωα
By definition,
Z Z
1 1
<σ> = σdV = σdV
V V V M ∪Ωα
Z n Z
1 hM X Ωα i
= σdV + σdV
V M M Ω
α=1 α Ωα
X
= f0 < σ >M + fα < σ >α (6.3)
α
Therefore,
X
f0 < σ >M = <σ>− fα < σ >α
α
X
= C̄ :< > − fα Cα :< >α (6.4)
α
On the other hand,

hM 1 Z i
f0 < σ >M = f0 C :< >M = C : dV
V M M
h1 Z i
= C: dV
V V /∪Ωα
h1 Z X Ωα 1 Z i
= C: dV − dV
V V α
V Ωα Ωα
X
= C: <>− fα < >α (6.5)
α
Combining Eqs. (6.4) and (6.5) yields

X
C̄ − C :< >= fα Cα − C :< >α (6.6)
α
If the prescribed displacement boundary condition is applied, it may be also

written X
C̄ − C : 0 = fα Cα − C :< >α (6.7)
α
Following a similar steps, one can show that

Z Z
1 1
<> = dV = dV
V V V M ∪Ωα
X
= f0 < >M + fα < >α (6.8)
α
Therefore,
X
f0 < >M = <>− fα < >α
α
X
= D̄ :< σ > − fα Dα :< σ >α (6.9)
α
and
hM 1 Z i
f0 < >M = f0 D :< σ >M = D : σdV
V M M
h1 Z X Ωα 1 Z i
= D: σdV − σdV
V V α
V Ωα Ωα
X
= D: <σ>− fα < σ >α (6.10)
α
Combining Eqs. (6.9) and (6.10) yields

X
D̄ − D :< σ >= fα Dα − D :< σ >α (6.11)
α
If the traction boundary condition is applied, it may be written

X
D̄ − D : σ 0 = fα Dα − D :< σ >α (6.12)
α
We name Eqs. (6.6) and (6.11) as the basic equations of average stress/strain
fields.
6.1.2 Homeogenization: Equivalent stress/strain conditions

Consider the prescribed macro stress boundary condition,
t = n · σ 0 , ∀ x ∈ ∂V
Based on the averaging theorem, < σ >= σ 0 .

One may note that the remote background strain as
0 = D : σ 0 = D :< σ >6=< > (6.13)
Similarly, for prescribed macro-strain boundary condition,
u(x) = x · 0 , x ∈ ∂V
the averaging theorem asserts that in this case
0 =< > .
The background stress,

σ 0 = C :< >6=< σ > . (6.14)
Suppose that there are α = 1, 2, · · · , n distinct inhomogenous phases. ∀x ∈
Ωα , the stress and strain equivalent conditions are
Cα : (0 + d ) = C : (0 + d − ∗ ) (6.15)
or
Dα : (σ 0 + σ d ) = C : (σ 0 + σ d − σ ∗ ) (6.16)
Then one can find the average stress and strain fields inside each inclusion,
< >α = Aα : ∗ (6.17)
< σ >α = Bα : σ ∗ (6.18)
where
Aα = (C − Cα )−1 : C (6.19)
Bα = (B − Bα )−1 : B (6.20)
Since the inclusion population is small, one can neglect the interaction among
inclusions. The disturbance field inside each inclusion can then be related to
eigenstrain fields,
d = S̄α : ∗ , ∀x ∈ Ωα (6.21)
σ d = T̄α : σ ∗ , ∀x ∈ Ωα (6.22)
Subsequently, one can decide how much the eigenstrain or eigenstress have to
be prescribed by the following conditions,
∗ = (Aα − Sα )−1 : 0 (6.23)
σ ∗ = (Bα − Tα )−1 : σ 0 (6.24)
Therefore the average strain/stress inside the α-th phase inclusion may be
expressed by eigenstrain/eigenstress, i.e.
< >α = Aα : ∗ = Aα : (Aα − Sα )−1 0 (6.25)
< σ >α = Bα : σ ∗ = Bα : (Bα − Tα )−1 σ 0 (6.26)
Subsequently, one can relate the average strain and average stress in the α-th
inclusion (inhomogeneity) with the background strain and background stress
through the so-called concentration tensors,
< >α = Aα : 0 (6.27)
< σ >α = B α : σ 0 (6.28)
where the concentration tensors are defined as
Aα = Aα : (Aα − Sα )−1 (6.29)

B α = Bα : (Bα − Tα )−1 (6.30)
Since by definition < σ >α = Cα :< >α and < >α = Dα :< σ >α , one
can rewrite Eqs. (6.27) and (6.28) as
α
C : Aα : D : σ 0
< σ >α = (6.31)
Bα : σ0
or
Dα : B α : C : 0

< >α = (6.32)
Aα : 0
Suppose that prescribed macro-stress boundary condition is applied. Sub-
stituting both expressions in Eq. (6.31) into the basic average equation (6.11)
yields,
 α
Xn  C : Aα : D̄ : σ 0
(D̄ − D) : σ 0 = fα (Dα − D) : (6.33)
 α
α=1 B : σ0
Therefore, for prescribed traction boundary condition, we have the following

estimate on effective compliance tensor,
 n
X
fα (Dα − D) : Cα : Aα : D̄

D +





 α=1
D̄ = (6.34)

 X n
fα (Dα − D) : B α

 D+



α=1
By considering the identities,
(Aα )−1 = (Dα − D) : Cα , and Bα = (D − Dα )−1 : D (6.35)
Finally, we obtain
 n
X
fα (Aα − Sα )−1 : D

D +





 α=1
D̄ = (6.36)

 n
X
fα D : (Bα − Tα )−1




 D −
α=1
If prescribed macro-strain boundary condition is applied, one may substitute

the both expressions of (6.58) into the basic average equation (6.6). It leads to
 α
X n  D : B α : C : 0
0 α
(C̄ − C) : = fα (C − C) : (6.37)
 α 0
α=1 A :
The following estimate on effective elastic tensor may be obtained,

 n
X
fα (Cα − C) : Dα : B α : C

C+





 α=1
C̄ = (6.38)

 X n
fα (Cα − C) : Aα

 C+



α=1
Using the identities,
(Bα )−1 = (Cα − C) : Dα , and Aα = −(Cα − C)−1 : C (6.39)
we have the following estimate on effective elastic stiffness tensor

 n
X
fα (Bα − Tα )−1 : C

 C+




 α=1
C̄ = (6.40)

 Xn
fα : C : (Aα − Sα )−1

 C−



α=1
Note that the index α starts from 1, and each α is an inhomogeneous phase.
One of the drawback of dilute distribution homogenization is
D̄ : C̄ 6= 1 or D̄ 6= C̄−1 .
This can be shown for α = 1:

D̄ : C̄ = 1(4s) + fα (Aα − Sα )−1 : D : C : 1(4s) − fα (Aα − Sα )−1
= 1(4s) − fα2 (Aα − Sα )−1 : (Aα − Sα )−1
= 1(4s) + O(fα2 ) 6= 1(4s) . (6.41)
Obviously, the effective elastic stiffness is not consistent with the effective
compliance tensors.
6.1.3 Elastic moduli in isotropic case

Suppose that there are n different phases of inhomogeneities in a solid.
For prescribed traction boundary condition, Eshelby’s equivalent strain method
yields,
n
( )
X
α α −1
D= 1+ fα (A − S ) :D
α=1
where Aα is defined as
Aα = (C − Cα )−1 : C
Here C is the elastic tensor of the matrix, which is assumed to be isotropic, i.e.
C = 3KE(1) + 2µE(2) . We can then calculate
C − Cα = 3(K − K α )E(1) + 2(µ − µα )E(2)
and
Aα = (C − Cα )−1 : C
1 (1) 1 (2)

(1) (2)

= E + E : 3KE + 2µE
3(K − K α ) 2(µ − µα )
K µ
= α
E(1) + E(2)
K −K µ − µα
Since the composite is isotropic, we use the Eshelby tensor of spherical inclu-
sions, For spherical inclusion, the Eshelby tensor is
5ν − 1 (2) 2(4 − 5ν) (4s)
SΩ = 1 ⊗ 1(2) + 1
15(1 − ν) 15(1 − ν)
(1 + ν) (1) 2(4 − 5ν) (2)
= E + E
3(1 − ν) 15(1 − ν)
= s1 E(1) + s2 E(2)
1+ν 2(4 − 5ν)
3(1 − ν) 15(1 − ν)
Then
K µ
Aα − Sα = − sα
1 E(1)
+ − sα
2 E(2)
(K − K α ) (µ − µα )
and
−1 1 1
Aα − Sα = E(1) + µ E(2)
K α
− s2
− sα1 (µ − µα )
(K − K α )
Hence
n
X
D̄ = 1+ fα (Aα − Sα )−1 : D
α=1


 n
 X fα
= E(1) + E(2) + E(1)
K
− sα1

 α=1
(K − K α )



fα (2)
 1 1 (2)
+ µ E : E(1) + E
− sα  3K 2µ
2 
(µ − µα )
Finally,
 
n
1  X fα  (1)
D̄ = 1 + E
3K K α
α=1 − s1
 K − Kα 
n
1  X fα  (2)
+ 1 + µ E (6.42)
2µ − sα
α=1 2
µ − µα
Assume that fα << 1,

 −1
n
K̄ X fα
= 1 +
 
K K α

α=1 − s1
K − Kα
n
X fα
= 1− + O(fα2 )
K
α=1 − sα1
K − Kα
and
 −1
n
µ̄ X fα
= 1 +
 
µ µ α

α=1 − s2
µ − µα
n
X fα
= 1− µ + O(fα2 )
α
α=1 − s2
µ − µα
Similarly, by considering remote traction boundary condition, we have the

estimate of effective elastic modulus for solids with dilute suspension of inho-
mogeneities,
n
( )
X
C= 1+ fα (Bα − Tα )−1 :C
α=1
where Bα is defined as
Bα = (D − Dα )−1 : D
1 (1)
Here D is the elastic compliance tensor of the matrix, i.e. D = 3K E +
1 (2)
2µ E . We can then calculate
Bα = (D − Dα )−1 : D
1 (2) −1

1 1 1 (1) 1 1
= − α E + − E :D
3 K K 2 µ µα
K − K (1) µα − µ (2) −1 1 (1)
α
1 (2)
= E + E : E + E
3KK α 2µµα 3K 2µ
Kα (1) µα
= − E − E(2)
K − Kα µ − µα
Subsequently,
Kα µα
Bα − T α = − − (1 − s α
1 ) E (1)
+ − − (1 − sα
2 ) E(2)
(K − K α ) (µ − µα )
K µ
α (1) α (2)
= − − s1 E − − s2 E
K − Kα µ − µα
and
−1 1 1
Bα − Tα =− E(1) − µ E(2)
K α
− s2
− sα1 (µ − µα )
(K − K α )
Finally
n
X
C̄ = 1+ fα (Bα − Tα )−1 : C
α=1
 
 
 n n 
 X fα X fα 
= 1− E(1) + 1 − µ E(2)
K α
− s2
− sα1
 
 α=1 α=1 

(K − K α ) (µ − µα ) 
: (3KE(1) + 2µE(2) )
n n
X fα
(1)
X fα
= 3K 1 − E + 2µ 1 − µ E(2)
K α α
− s2
α=1 − s1 α=1
(µ − µα )
(K − K α )
Therefore
n
K̄ X fα
=1−
K K
α=1 − sα1
K − Kα
and
n
µ̄ X fα
=1− µ
µ − sα2
α=1
µ − µα
It is obviously that these results are different from the results obtained from
prescribed traction boundary condition. They are only agreeable to the first
order of the volume fraction. In other words, these two results (the results
obtained from prescribed stress b.c. and the results obtained from prescribed
strain b.c.) are not consistent in the homogenization scheme for dilute inhomo-
geneity distribution.
6.2 Self-consistent method

As shown above, effective elastic tensor and compliance tensor obtained
via homogenization of inhomogeneities of dilute distribution are not recipro-
cal to each other as supposed to be. As the volume fraction of inhomogeneity
increases, the accuracy of dilute suspension homogenization schemes deterio-
rates, because the interaction among inhomogeneities become strong.
To take into account the interaction among inhomogeneities, a so-called
self-consistent homogenization method is proposed, which is largely attributed
to a series papers by Hill ([1962],[1963],[1964]). Rodney Hill is a highly intel-
lectual individual, whose writing style is very close to mathematics literature,
which is rigorous, terse, and often esoteric.
The following presentation is mainly adopted from Nemat-Nasser and Hori,

and it is blended with authors own interpretation, which is more engineering
oriented.
There are two main differences between self-consistent homogenization and
dilute suspension homogenization.
The first difference is in the treatment of remote (background) strain and
stress.
Consider the prescribed macro stress boundary condition,
t = n · σ 0 , ∀ x ∈ ∂V
Based on the averaging theorem, < σ >= σ 0 . In self-consistent homogeniza-

tion, we define the remote background strain as
0 = D̄ : σ 0 = D̄ :< σ > (6.43)
Therefore in this case,
0 = D̄ :< σ >=< > .
Similarly, for prescribed macro-strain boundary condition,
u(x) = x · 0 , x ∈ ∂V
the averaging theorem asserts that in this case
0 =< > .
If σ = C̄ : , the background stress will be the average stress,
σ 0 = C̄ :< >=< σ > . (6.44)
The second main difference between the self-consistent method and dilute
suspension method is that Eshelby’s equivalent inclusion principle is applied
with respect to the homogenized solid, instead of matrix. Suppose that there
are α = 1, 2, · · · , n distinct inhomogenous phases. ∀x ∈ Ωα ,
Cα : (0 + d ) = C̄ : (0 + d − ∗ ) (6.45)
or
Dα : (σ 0 + σ d ) = C̄ : (σ 0 + σ d − σ ∗ ) (6.46)
Moreover, the disturbance field generated by eigenstrain is also calculated with
respect to homogenized solid, i.e.
d = S̄α : ∗ , ∀x ∈ Ωα (6.47)
σ d = T̄α : σ ∗ , ∀x ∈ Ωα (6.48)
Therefore the average strain/stress inside the α-th phase inclusion may be
expressed by eigenstrain/eigenstress, i.e.
< >α = Āα : ∗ (6.49)
< σ >α = B̄α : σ ∗ (6.50)
where
Āα = (C̄ − Cα )−1 : C̄ (6.51)
B̄α = (B̄ − Bα )−1 : B̄ (6.52)
Subsequently, one can relate the average strain and average stress in the α-th
inclusion (inhomogeneity) with the background strain and background stress
by concentration tensors,
< >α = Āα : 0 (6.53)
< σ >α = B̄ α : σ 0 (6.54)
where the concentration tensors are defined as
Āα = Āα : (Āα − S̄α )−1 (6.55)
B̄ α = B̄α : (B̄α − T̄α )−1 (6.56)
Since by definition < σ >α = Cα :< >α and < >α = Dα :< σ >α , one
can rewrite Eqs. (6.53) and (6.54) as
α
C : Āα : D̄ : σ 0
< σ >α = (6.57)
B̄ α : σ 0
or α
D : B̄ α : C̄ : 0
< >α = (6.58)
Āα : 0
Note that the relationships 0 =< > and σ 0 =< σ > are used.
Suppose that prescribed macro-stress boundary condition is applied. Sub-
stituting Eqs. (6.57) and (6.58) into the basic average equation (6.11) yields,
 α
Xn  C : Āα : D̄ : σ 0
0 α
(D̄ − D) : σ = fα (D − D) : (6.59)
 α
α=1 B̄ : σ 0
Therefore, self-consustent method gives the following estimate on effective
compliance tensor,
 n
X
fα (Dα − D) : Cα : Āα : D̄

D +





 α=1
D̄ = (6.60)

 Xn
fα (Dα − D) : B̄ α

 D+



α=1
If prescribed macro-strain boundary condition is applied, one may substitute

Eqs. (6.57) and (6.58) into the basic average equation (6.6). It leads to
 α
X n  D : B̄ α : C̄ : 0
0 α
(C̄ − C) : = fα (C − C) : (6.61)
 α 0
α=1 Ā :
Hence self-consustent method gives the following estimate on effective elas-
tic tensor, 
Xn
fα (Cα − C) : Dα : B̄ α : C̄

C +





 α=1
C̄ = (6.62)

 Xn
fα (Cα − C) : Āα

 C+



α=1
Note that the index α starts from 1, and each α is an inhomogeneous phase.
We now show that
D̄ : C̄ = 1 or D̄ = C̄−1 .
Consider
D = D : 1 = D : C̄ : C̄−1
h Xn i
= D: C+ fα (Cα − C) : Āα : C̄−1
α=1
n
X
= C̄−1 + fα D : (Cα − C) : Aα : C̄−1 (6.63)
α=1
Since,
D : (Cα − C) = D : Cα − 1
= −1 + D : Cα
= −(Dα − D) : Cα
The last line of (6.63) may be rewritten as

n
X
D = C̄−1 − fα (Dα − D) : Cα : Aα : C̄−1 (6.64)
α=1
which leads to
n
X
C̄−1 = C−1 + fα (Dα − D) : Cα : Āα : C̄−1 (6.65)
α=1
Compare (6.65) with the first line of Eq. (6.60). One can conclude that
C̄−1 = D̄ (6.66)
Similar arguments can be made to show that D̄−1 = C̄.
Example 6.1 For isotropic composites, the effective moduli obtained from
self-consistent scheme can be further simplified.
Consider
Xn
C̄ = C + fα (Cα − C) : Āα (6.67)
α=1
Step 1.
C = 3KE(1) +2µE(2) , and (Cα −C) = 3(K α −K)E(1) +2(µ(α) −µ)E(2)

Step 2:
Āα = (C̄ − Cα )−1 : C̄

1 (1) 1 (2)

= E + E : (3K̄E(1) + 2µ̄E(2) )
3(K̄ − K α ) 2(µ̄ − µα )
K̄ µ̄
= E(1) + E(2)
K̄ − K α µ̄ − µα
Then,
Āα = Āα (Āα − S̄α )−1
h K̄ µ̄ ih K̄ −1
(1) (2)
= E + E − s̄1 E(1)
K̄ − K α µ̄ − µα K̄ − K α
µ̄ −1 i
(2)
+ − s̄2 E
µ̄ − µα
K̄ µ̄
= E(1) + E(2)
K̄ − (K̄ − K α )s̄1 µ̄ − (µ̄ − µα )s̄2
Therefore,
C̄ = 3K̄E(1) + 2µ̄E(2)
Xn
= C+ fα (Cα − C) : Āα
α=1
X (K α − K)K̄ (1)
= 3 K+ fα E
α
K̄ + (K̄ − K α )s̄1
X (µα − µ)µ̄ (2)
+2 µ + fα E
α
µ̄ + (µ̄ − µα )s̄2
Figure 6.1. Schematic illustration of Mori-Tanaka lemma
which lead to
n Kα
K̄ X Kα −1
= 1+ −1 1+(
fα − 1)s̄1 (6.68)
K K K̄
α=1
n µα
µ̄ X µα −1
= 1+ fα −1 1+( − 1)s̄2 (6.69)
µ µ µ̄
α=1
3K − 2µ
Note that ν = .
2(3K + µ)
6.3 Mori-Tanaka methods

6.3.1 Tanaka-Mori lemma
In 1972, a less than two-page technical note was published in Journal of
Elasticity by Tanaka and Mori (Tanaka and Mori [1972]), which revealed an
importance consequence of the scalability of the Eshelby tensor.
That result is the well-known Tanaka-Mori lemma, and it then leads a very
effecient homogenization procedure called Mori-Tanaka method. Today, the
Mori-Tanaka method is one the most popular homogenization methods used in
composite industry. Its applications include abraided composite, nano-composites,
and reinforce fiber composites.
We start with the Tanaka-Mori lemma first.
Lemma 6.2 (Tanaka and Mori) Consider two coaxial, similar ellipsoidal
domains, Ω0 , Ω (Ω0 ⊂ Ω),
x21 x22 x23

Ω0 = x + + ≤ 1
a21 a22 a23
x21 x22 x23

Ω = x 2 + 2 + 2 ≤1 (6.70)
b1 b2 b3
where
a1 a2 a3
+ + =k
b1 b2 b3
Assume that a uniform eigenstrain state, ∗ij (x), is prescribed in the smaller
ellipsoidal region, i.e.
∗
∗ ij x ∈ Ω0
ij (x) =
0 x ∈ IR3 /Ω0
The the average disturbance strain field is zero, i.e
Z
1
< >Ω−Ω0 = ij (x)dΩ = 0 . (6.71)
Ω − Ω0 Ω−Ω0
Proof:
Suppose that there are three coaixial, similar ellipsoidals, Ω0 ⊂ Ω1 ⊂ Ω2 in
an infinite homogeneous medium, and a uniform eigenstrain is presecibed in
Ω0 , i.e. ∗
∗ ij x ∈ Ω0
ij (x) =
0 x ∈ IR3 /Ω0
The disturbance displacement field can be then written as
Z
ui (x) = − ∗mn Ck`mn Gik,` (x − x0 )dx0 (6.72)
Ω0
and the disturbance strain field is

Z
Ck`mn
ij (x) = − ∗mn Gik,`j (x − x0 ) + Gjk,ì (x − x0 ) dx0 (6.73)
Ω0 2
where Ck`mn is the elastic tensor, Gik (x − x0 ) is the Green’s function in the
infinite domain, and
Z
Ck`mn
Sk`mn = − Gik,`j + Gjk,ì dx0
Ω0 2
 Ω
 Sk`mn0
, x ∈ Ω0
= (6.74)
 ∞
Sk`mn , x ∈ IR3 /Ω0
Figure 6.2. Schematic illustration of the Proof of Mori-Tanaka lemma
is the Eshelby tensor.

Now consider the average strain in the region Ω1 − Ω2 .
Z Z Z
h Ck`mn i
ij (x)dx = ∗mn − Gik,`j (x−x0 )+Gjk,ì (x−x0 ) dx0 dx
Ω2 −Ω1 Ω2 −Ω1 Ω0 2
Since x ∈ Ω2 − Ω1 , the integrand does contain singularity in either integration

domains, Ω0 and Ω2 − Ω1 . We can then change the order of the integration,
Z Z
h
∗ Ck`mn i
mn − Gik,`j (x − x0 ) + Gjk,ì (x − x0 ) dx0 dx
2
ZΩ2 −Ω 1
Z Ω0
h Ck`mn i
= ∗mn − Gik,`j (x − x0 ) + Gjk,ì (x − x0 ) dx0 dx
2
ZΩ0 h ZΩ2 −Ω1
Ck`mn i
= ∗mn − Gik,`j (x − x0 ) + Gjk,ì (x − x0 ) dx0 dx
Ω0 2
Z h ΩZ2
Ck`mn i
− ∗mn − Gik,`j (x − x0 ) + Gjk,ì (x − x0 ) dx0 dx
Ω0 Ω1 2
Z h i h i
= ∗mn Ω2
Sk`mn Ω1
− Sk`mn dx0 = ∗mn Ω0 Sk`mn
Ω2 Ω1
− Sk`mn (6.75)
Ω0
Since Eshelby tensor only depends on the material property and the aspect ratio
of the ellipsoidals,
h i
∗mn Ω0 Sk`mn
Ω2 Ω1
− Sk`mn =0 (6.76)
if Ω2 , Ω1 are similar. Hence,

Z
ij (x)dΩx = 0 . (6.77)
Ω2 −Ω1
Let Ω1 → Ω0 and Ω2 → Ω. We have the desired result,

Z Z
ij (x)dΩx = ij (x)dΩx = 0 . (6.78)
Ω2 −Ω1 Ω−Ω0
♣
Remark 6.3.1 1 It is also true that the average disturbance stress field is
also zero Z
σij dΩx = 0 . (6.79)
Ω−Ω0
2 Eq. (6.75) is valid as long as Ω1 ⊂ Ω2 . They don’t need to be confocal, but

they definitely need to be similar, and they may need to be coaxial (some
people questioned nececity of this requirment too, the real issue is : does
Eshelby tensor depend on coordinates ?).
3 This result can be generalized into the cases that the inclusion Ω0 is not
ellipsoidal and the eigenstrain distribution in Ω0 is not uniform.
6.3.2 Mori-Tanaka’s two-phase model

In this section, we present a straightforward application of Tanaka-Mori
lemma for a two-phase double inclusion problem.
We assume that there are only two phases in an RVE, and both the RVE and
the inhomogeneity have the shape of ellipsoidal. The are coaxial and similar
in shape.
Suppose in the far field, there are constant stress and strain fields, σ 0 and
0
. Due the presence of inhomogeneity, the total strain and stress fields consist
of two parts: constant far fields and perturbed fields, i.e.
(x) = 0 + d (x), ∀x ∈ V (6.80)
σ(x) = σ 0 + σ d (x), ∀x ∈ V (6.81)
Inside the inclusion, x ∈ Ω, the disturbance field may be expressed in terms

of eigenstrain
d = SΩ : ∗ , ⇒ (x) = 0 + SΩ : ∗ , ∀x ∈ Ω (6.82)
Therefore,
< >Ω = 0 + SΩ : ∗ (6.83)
Figure 6.3. Schematic illustration of two-phase model
Recall that the homogenization condition (Eshelby’s equivalent principle),
CΩ : (0 + d ) = C : (0 + d − ∗ ), (6.84)
let to
0 + d = A Ω : ∗ , ∀ x ∈ Ω (6.85)
where AΩ = (C − CΩ )−1 : C. Combining with d = SΩ : ∗ , one can find
that
∗ = (AΩ − SΩ )−1 : 0 (6.86)
Substitute (6.86) back to (6.83). We finally have

< >Ω = 1(4s) + SΩ : (AΩ − SΩ )−1 : 0 (6.87)
The average stress inside the inclusion can be also evaluated by considering
homogenization condition and (6.86)

< σ >Ω = C : 0 + d − ∗ = C : 0 + (SΩ − 1(4s) )∗

= C : 1(4s) + (SΩ − 1(4s) (AΩ − SΩ )−1 : 0 . (6.88)
One the other hand, by the Tanaka-Mori lemma, the average strain in the
matrix is
< >M =< 0 + d >M = 0 (6.89)
and hence
< σ >M = C : 0 . (6.90)
Let f be the volume fraction of the inhomogeneity. We then have the following
balance equations for average strain and stress
< >V = (1 − f ) < >M +f < >Ω (6.91)

< σ > = (1 − f ) < σ >M +f < σ >Ω (6.92)
One can readily find that

< >V = (1 − f )0 + f (0 + SΩ : ∗
= 0 + f SΩ (AΩ − SΩ ) : 0

= 1(4s) + f SΩ (AΩ − SΩ )−1 : 0 (6.93)
and

< σ >V = (1 − f )C : 0 + f C : 1(4s) + (SΩ − 1(4s) (AΩ − SΩ )−1 : 0 .

= C : 1(4s) + f (SΩ − 1(4s) (AΩ − SΩ )−1 : 0 . (6.94)
By definition,
< σ >V = C :< >V (6.95)
It leads to

C : 1(4s) +f (SΩ −1(4s) )(AΩ −SΩ )−1 : 0 = C̄ : 1(4s) +f SΩ (AΩ −SΩ )−1 : 0
Finally, the effective elastic tensor is obtained

−1
C̄ = C : 1(4s) +f (SΩ −1(4s) )(AΩ −SΩ )−1 : 1(4s) +f SΩ (AΩ −SΩ )−1
(6.96)
6.3.3 Mori-Tanaka mean field theory

In previous homogenization procedures, the disturbance strain and stress
fields due to an inhomogeneity are approximated by Eshelby’s single inclusion
solution in an infinte space.
In real applications, an RVE is finite, and it is subjected with remote bound-
ary conditions, e.g. prescribed traction condition or prescribed displacement
condition, i.e.
u = x · 0 , x ∈ ∂V (6.97)
or
t = n · σ 0 , x ∈ ∂V (6.98)
Let pt and σ pt representing perturbed strain and stress fields due to Es-
helby’s single inclusion solution in an infinite medium. If we let
(x) = 0 + d = 0 + pt (6.99)

σ(x) = σ 0 + σ d = σ 0 + σ pt (6.100)
Obviously,
σ 0 + σ pt 6= σ 0 , or 0 + pt 6= 0 , ∀ x ∈ ∂V (6.101)
Therefore, either boundary condition (6.98) and (6.97) will not be satisfied.
This is because a finite size RVE will cause additional interaction between
matrix and inclusions, interaction between the boudary and inclusions, and
interaction among inclusions themself. Note that pt , σ pt , → 0 only when
|x| → ∞.
To take into account the effects of a finite size RVE, additional stress and
strain fields are need to faithfully represent total stress and strain distribution
in an RVE, i.e.
σ = σ 0 + σ̃ + σ pt (6.102)
= 0 + ˜ + pt (6.103)
where σ̃ and ˜ are the so-called image stress and image strain.
In literature, especially literatures on dislocations, additional stress and strain
fields that accommodate the stress solution of a infinite space to satisfy bound-
ary conditions are called image stress and image strain fields, because in prac-
tice some of these stress and strain fields are found by placing certain image
external sources to achieve their objectives.
Nevertheless, the homogenization problem in a finite REV becomes com-
plicated, because in general it is very difficult to know the precise distribution
of image stress and image strain fields. To circumvent this difficulty, Mori
and Tanaka [1973] proposed the following mean field assumption, which is an
ingenous and very successful method.
Mori & Tanaka’s theory was later refined in a landmark paper by G. Weng
(Weng [1990]). The following presentation is an adaption of Weng’s formula-
tion. Suppose that in an RVE there are many inhomogeneities, or the density
of inhomogeneities are statistically stable. Then the strain or stress field in the
matrix may be written as
(x) = 0 + d , ∀x ∈ M ⇒ < >M = 0 + < d >M ;

σ(x) = σ 0 + σ d , ∀x ∈ M ⇒ < σ >M = σ 0 + < σ d >M ;
In general we don’t know the precise disturbance fields in a matrix, i.e., dM or
σ dM .
Consider the matrix is the dominate phase in a composite. We denote the

average field in the matrix, < >M or < σ >M , as the mean field, which
include boundary effects and effects of interactions of many other inclusions.
Now we add an inclusion into the average ensemble—the RVE. After the
inclusion is added, we call the field as the new field in contrast with the old
field before the inclusion is being added. Therefore, in the matrix,
< new >M =< old >M + < pt >M + < im >M , ∀x ∈ M (6.104)
where pt and im are the inclusion solution for infinite space and the corre-
sponding image strain solution due to the finite RVE.
By the Tanaka-Mori lemma, < pt >M = 0. Mori and Tanaka then further
argued that since there have been so many inclusions inside the RVE, the aver-
age effects of the image strain or image stress field for a single inclusion may be
negligible without alter the mean field of value of the RVE, i.e. < im >M = 0,
which is the essence of Mori-Tanaka mean field theory. Note that < old >M
does take into account the average effects of the image stress/strain fields all
other inclusions.
Therefore, we have
< new >M =< old >M =< >M , ∀x ∈ M . (6.105)
Inside the inclusion, we still neglect the effects of image strain or image
stress field of the newly added inclusion, we then have
< >Ω = < >M + < pt >Ω + < im >Ω
= < >M + < pt >Ω
= < >M +SΩ : ∗ , ∀x ∈ Ω (6.106)
Similarly, for the stress field,
< σ new >M = < σ old >M =< σ >M , x ∈ M
< σ >Ω = < σ >M +TΩ : σ ∗ , x ∈ Ω (6.107)
Based on Eshelby’s equivalence homogenization conditions,

CΩ :< >Ω = C : < >Ω −∗ (6.108)
or

DΩ :< σ >Ω = D : < σ >Ω −σ ∗ (6.109)
One may obtain

< >Ω = AΩ : ∗ ⇒ < >M + < pt >Ω = AΩ : ∗
or < σ >Ω = BΩ : σ ∗ ⇒ < σ >M + < σ pt >Ω = AΩ : σ ∗
where AΩ := (C − CΩ )−1 : C and BΩ := (D − DΩ )−1 : D.

Subsequently, one can obtain that
< >Ω = Adil

Ω :< >M
dil
or < σ >Ω = BΩ :< σ >M (6.110)
according to different boundary conditions or different homonization schems.

In passing, we note that that the concentration tensors may be written in
different forms,
i−1
Ω −1
h
Ω −1
Adil
Ω = A Ω
: (A Ω
− S ) = (A Ω
− SΩ
) : A
−1 −1
h i
= 1 − SΩ : AΩ
h i−1
= 1 − SΩ : C−1 : (C − CΩ )
h i−1
= 1 + PΩ : (CΩ − C) (6.111)
and
−1 −1
h i
dil
BΩ = BΩ : (BΩ − TΩ )−1 = (BΩ − TΩ ) : AΩ
−1 −1
h i
= 1 − TΩ : BΩ
h i−1
= 1 − TΩ : D−1 : (D − DΩ )
h i−1
= 1 + QΩ : (DΩ − D) (6.112)
where
PΩ = SΩ : C−1 (6.113)
QΩ = TΩ : D−1 (6.114)
are called polarization tensors.

Since C − CM = 0 and D − DM = 0, it is easy to see that both
Adil M = 1 and B dil M = 1 . (6.115)
By definition,
< > = (1 − f ) < >M +f < >Ω (6.116)

< σ > = (1 − f ) < σ >M +f < σ >Ω (6.117)
From (6.116) and (6.117), we may find that

h i−1
< >M = (1 − f )1 + f Adil Ω :< >
h i−1
= fM Adil M + fΩ Adil Ω :< >= Ã0 :< >(6.118)
h i−1
< σ >M = fM B dil M + fΩ B dil Ω :< σ >= B̃0 :< σ >(6.119)
where
h i−1
Ã0 := fM Adil M + fΩ Adil Ω (6.120)
h i−1
B̃0 := fM B dil M + fΩ B dil Ω (6.121)
Accordingly,
< >Ω = Adil Ω :< >M = Adil Ω : Ã0 :< > (6.122)
< σ >Ω = B dil Ω :< σ >M = B dil Ω : B̃0 :< σ > (6.123)
Therefore,
< σ > = fM < σ >M +fΩ < σ >Ω

= fM C0 < >M +fΩ CΩ < >Ω
= fM C0 < >M +fΩ CΩ Adil Ω < >M

= fM C0 + fΩ CΩ Adil Ω Ã0 < >
= C̄ :< > (6.124)
and
< > = fM < >M +fΩ < >Ω

= fM D0 < σ >M +fΩ DΩ :< σ >Ω
= fM D0 < σ >M +fΩ DΩ : B dil Ω :< σ >M

= fM D0 + fΩ DΩ : B dil Ω B̃0 :< >
= D̄ :< σ > (6.125)
Recall that Adil M = B dil M = 1. We have

−1
C̄ = fM C0 : Adil M + fΩ CΩ : Adil Ω : fM Adil M + fΩ Adil Ω
−1
D̄ = fM D0 : B dil M + fΩ DΩ : B dil Ω : fM B dil M + fΩ B dil Ω
(a) (b)
(c) (d)
(e) (f)
Figure 6.4. Comaprison of effective bulk modulus among various homogenization methods:
dilute distribution (DD & DT), self-consistent, and Mori-Tanaka
In general, for a solid with n+1 phases (from α = 0 to α = 0), the Mori-Tanaka
mean field theory gives the following estimates,
n
X n
X −1
C̄ = fα Cα : Adil α : fα Adil α
α=0 α=0
n
X n
X −1
D̄ = fα Dα : B dil α : fα B dil α (6.126)
α=0 α=0
where the pahse α = 0 represents the matrix, and non-zero α represents the
inhomogeneous phases.
Figure 6.5. Rodney Hill
6.4 Rodney Hill

Rodney Hill was born on the 11th June 1921 at Stourton, near Leeds, in
Yorkshirt. He comes from a family with deep roots in the practical and culture
tradtions of the West Riding, although with no known mathematical ability
in an earlier generation. Rodney’s father, Harold Harrison Hill, had been an
only child and he was educated at the University of Leeds, gaining an M. A.
for postgraduate work in history. He also took an external London degree in
economics. After wartime service in the Royal Navy he became a schoolmas-
ter, and was eventually senior History Master at Leeds Boy’s Modern School.
Rodney’s mother had been a student at Leeds School of Art. Rodney himself
was also an only child, in an immediate home background which encouraged
scholarship and self-sufficiency.
Rodney entered Leeds Grammar School with a scholarship in 1932, and
there gave regular prize-wining evidence of all-round intellectual ability not
only in mathematics, but equally in art, English literature, and other Arts sub-
jects. During this period he taught himeself to play the piano, and became
proficient at chess in which he was later to represent Cambridge University
and town. Thus were developing the powers of accurate observation and anal-
ysis to be brought to bear on the mathematics and physics which became his
formal specialism from the age of 15. The customary large-team games did
not attract him as school, but Rodney enjoyed the one-to-one sports of squash,
fencing, and golf. He left school as Head of House, and in December 1938
he was awarded an Open Major Scholarship at Pembroke College, Cambridge.
However, it needed the State and County Scholarships gained in the preceding
summer to make a financially independent undergraduate.
Hill went up to Cambridge to read Mathematics in October 1939, againt a
background of external events which must have seemed the least auspicious
since the very founding of the University. Major Scholars were expected to
take Part II of the Tripos in two years instead of three by omitting all first-year
courses. This imposed a heavy workload, to be carried under spartan condi-
tions created by wartime restrictions such as blackout and rationing combined
with antique College plumbing. For example, there was no running hot wa-
ter, the nearest bath was courts away, and the winter allocation of one sack of
coal per week fuelled a fire in one’s room only in the evenings. Hill was not
deflected by the adverse general situation from his aim of a first-class honours
degree, and he became a Wrangler in June 1941. This entitled him to take
Part III of the mathematical Tripos, in the applied mathematical part of which
quantum mechanics figured prominently at the time. However, he felt obliged
to war-work, and so lost the opportunity for advanced training which those
lecture courses would have provided.
........
Problems brought to the Theoretical Research Branch were distributed ini-
tially according to specialisms of the more senior members, some of whom
had acquired relevant experience at Woolwich Arsenal. Those problems which
were quite new in context tended to go to the young inexperienced graduates
newly arrived from university. This was indeed a baptism of fire for them,
but it was a test which was to reveal Hill’s true metier. One of his initial as-
sigments was the deep penetration of very thick armour by Munoroe jets and
high-velocity shells with tungsten-carbide cores. This required a mechanics
of plastic deformation with unlimited magnitude, and thus was aroused Hill’s
interest in the field in which he later became perhapse the foremost world au-
thority. At this stage, however, he had no prior knowledge of the physics and
metallurgy of plasticity, and little of stress, strain or the tensors which the
mathematics would eventually require. There was no useful textbook, but G.
I. Taylor had written one or two helpful reports on shaped charges and Munros
jets. Nevertheless, working at first with Mott and Pack, Hill was soon able
to show, for example, that penetration by a tungsten–carbide core with pure
ogival head would be seriously degraded if too much of the tip were ground
conical (the British practice for manufacturing convenience). The demonstra-
tion was achieved not only theoretically, but also in field trials planned by Hill
in collaboration with an experimental group under Dr. Charles Sykes, F.R.S.
The problems at Fort Halstead called for simple but effective mathemat-
ics guided by physical intuiation and a willingness to communicate with oth-
ers, including non-mathematicians and experimentalists. There was not time
for complicated mathematics, there were no electronic computers to assist it,
and the experimental data were ususally too crude to warrant it anyway. He
acquired a lasting taste for a pragmatic blend of rigour, elgance, and simple
realism in the application of mathematics.
The sense of purpose discovered at this time was noticed by colleagues as
a cheerful and sparking earnestness. Popular relaxations among the group at
Cambridge had included music, books, and lightning chess. At Fort Halstead
ballroom dancing was added as a consuming passion for some, and Hill was
not slow to find that he had medal-winning ability in this new enthusiasm. He
met his future wife, Jeanne Wickens, early in 1945. She had been transferred
to work at Fort Halstead from the bombing range at Shoeburyness. Previously
she had trained as a dancer and teacher of ballet, but war cut short a promising
career. They were married in Cambridge in 1946, and they have one duaghter,
born in 1955. The strength of his wife’s support could already be detected in
the Preface to Hill’s first book, and the passage of years has happily reinforced
this bond.
By this time the applied mechanics of both solid and fluids was being forced
to push the boat out onto a sea of nonlinear problems, and away from the haven
linearity in which much pre-war work had lingerd. The trend was evident not
only in England, of course, but in other countries too. Hill found himself in
demand as the sole adviser on continuum plasticity in England, not only con-
cerning problems arising from the interests at Fort Halstead, but also fot new
theories of metal-working processes needed by engineers in the steel indus-
try. He obtained a Cambridge Ph.D. in 1948 for a Thesis entitled “Theoretical
studies of the plastic deformation of metals”. From the Ph.D. Thesis grew a
much more extensive monograph on “The Mathematical Theory of Plasticity”,
published at the Clarendon Press, Oxford, in 1950. This very rapidly estab-
lished Hill as an international authority. The final draft was written in his spare
time, i.e. in the evenings and weekends. He was then still only in his 28th year,
and it is timely to recall a remark from the review of the book in Engineering:
“The author has done his work so well that it is difficult to see how it could be
bettered. The book should rank for many years as an authoritative source of
reference.” This prognostication was fully borne out. The book was in print at
Oxford for 21 years, Japanese and Russian translations have been made, and
total sales currently approach 13,000.
The Journal of Mechanics and Physics of Solids was launched with the en-
couragement of the infan Pergamon Press in 1952. Hill suggested the title and
the general aim of a forum for effective applied mathematics, linked with ex-
perimenation, in engineering science. From the onwards the Journal has been
regarded as among the foremost in its general field, and unique in flavor. Hill
served as Eidtor-in-Chief untill handing over in 1968 to H.G. Hopkins.
The University of Nottinggham had received its Chater, and independence
from London, in 1948, and was shortly to embark on two decades substantial
expansion. Professor H. R. Pitt was appointed in 1950 to head the existing
Mathematics Department, and he was soon instrumental in securing the cre-
ation of a new Chair of Applied Mathematics. Rodney Hill applied, and was
offered the post in 1953 while still on 31. It was his responsibility to modernize
the teaching of applied mathematics. Hill took over some existing course him-
self, and instigated new ones with the aim of encouraging research students.
His undergraduate lectures were characterized by conciseness and tendency to
brevity. He would never exceed the time limit. But those stidents who took
the trouble to write down what he said, in addition to what was written on the
blackboard, found after reflection that they had a first-calss and substantial set
of notes.
It may only have been a coincidence that emergence of interest in the so-
called rational continuum mechanics was taking place in some American and
British universities at this time. Hill’s writings demonstrate an independent
view of these development, and no taste at all for axiomatics. He was beginning
to lay down the basis of general studies of non-uniqueness and instability in
continua which were to prov highly influential over the next two decades, and
which in due course brought further students and able collaborators.
The University of Cambridge conferred the degree of Sc. D. upon Rodney
Hill in 1959. The highest honour to which any British scientist aspires followed
in 1961, when he was elected a Fellow of the Royal Society. This gave much
pleasure to his colleagues at Nottingham and to his friends elsewhere.
In 1963 Hill was elected to a Berkeley Bye-Fellowship at Gonville and Caius
College, Cambridge. This he held for 6 years until the University conferred a
personal Readership in Mechanics of Solids. Thus he became a member of
the teaching staff of the Department of Applied Mathematics and Theoretical
Physics, and in 1972 a personal Professorship was conferred.
During this Cambridge period (he is still at Cambridge under semi-retirement—

Li’s comment), properties of heterogeneous media (including fibre compos-
ites), single crystals, continuum plasticity, and an independent reformulation
of rubber elasticity were explored, .....
His standards of scholarship and intellectual honesty are the highest. He is
ready in his appreciation of the good work of others; and he has been sharp
in candid criticism of misguided thinking or slack presentation (especially by
those mature enough to know better) if he thought the subject-matter would be
best served thereby—as some celebrated footnotes and book reviews testify.
The outward character of the man is not unlike his papers: physically tall
and slim, with the long fingers of a pianist, and having a quiet but compelling
presence. His unusally deep reserve has meant that casual social gatherings
and conferences have held less interest and been less rewarding for him than
for others.
—– By Geoffery Hopkins and Michael Sewell
From Mechanics of Solids Pergamon Press
6.5 Exercises
Probelm 6.1 Consider a n-phase composite material, and each phase has
its own elastic tensor Cα , compliance tensor Dα ; and matrix has elastic ten-
sor, C, and compliance tensor, D. Assume that in the representative volume
element (RVE), each phase only appears as one ellipsoidal inclusion. Under
dilute distribution assumption, the corresponding Eshelby tensor and conju-
gate Eshelby tensor for each phase are Sα and Tα respectively. Denote
Aα = (C − Cα )−1 : C (6.127)
Bα = (D − Dα )−1 : D (6.128)
Show
Cα : Aα : (Aα − Sα )−1 : D = Bα : (Bα − Tα )−1 (6.129)
Dα : Bα : (Bα − Tα )−1 : C = Aα : (Aα − Sα )−1 (6.130)
Probelm 6.2 For an isotropic two phase material. Assume the inhomogene-
ity phase is random distributed spherical cavities (µI = 0; KI = 0), and
the matrix is an incompressible masterial (K → ∞). Use the self-consistent
scheme,
n Kα
K̄ X Kα −1
= 1+ fα −1 − 1)s̄11+( (6.131)
K K K̄
α=1
n µα
µ̄ X µα −1
= 1+ fα −1 1+( − 1)s̄2 (6.132)
µ µ µ̄
α=1
where
1 + ν̄
s̄1 = (6.133)
3(1 − ν̄)
2(4 − 5ν̄)
s̄2 = (6.134)
15(1 − ν̄)
to find the effective bulk modulus, K̄, and the effective shear modulus, µ̄.
Hint:
J.R. Willis, “Variational and related methods for the overall properties of
composite”, in Advance in Applied Mechanics, Edited by C.-S. Yih (pages 45-
46), (1981), Academic Press, New York.
B. Budiansky, “On the elastic moduli of some heterogeneous materials”,
Journal of Mechanics and Physics of Solids, Vol. 13, (1965), pages 223-227.
Probelm 6.3 Assume that in an RVE there are n+1 phases, α = 0, 1, · · · , n
Mori-Tanaka mean theory states that
X n n
X −1
dil
D̄ = fα Dα : B α : fα B dil α (6.135)
α=0 α=0
Xn n
X −1
C̄ = fα Cα : Adil α : fα Adil α (6.136)
α=0 α=0
Show that Mori-Tanaka scheme is self-consistent, i.e.
C̄ = D̄−1 (6.137)
Hint: First show that
Cα : Adil α = B dil α : C0 (6.138)
Dα : B dil α = Adil α : D0 (6.139)
Probelm 6.4 Consider a two-phase composite with randomly distributed
spherical inclusions. The ratios of material constants between inhomogeneity
and matrix are
KΩ
= 25, and K Ω = 750M Pa (6.140)
K
νΩ
= 4, and ν Ω = 0.4 (6.141)
ν
K̄ µ̄ ν̄
Plot the ratio of , , and verses the volume fraction of inhomogeneity,
K µ ν
f , by using homogenization methods under the assumption of dilute suspen-
sion (both prescribed traction and prescribed displacement), self-consistent
method, and Mori-Tanaka mean field method.
Figure 6.6. Definition of the Volterra dislocation

Introduction of Dislocation Theory 127
Chapter 7
INTRODUCTION OF DISLOCATION THEORY
In material science, a dislocation may be defined as a disturbed region be-

tween two substantially perfect parts of a crystal. In elasticity theory, a dislo-
cation is defined as the strong discontinuity of the displacement field. In this
Chapter, we shall first study dislocation theory within the framework of linear
elasticity, and then we shall examine dislocation theory by considering lattice
structure, i.e. we shall study the Peierls-Nabarro model and a screw dislocation
solution in the framework of molecular dynamics. At the end of this Chapter,
we shall discuss one of the most important applications of dislocation theory:
dislocations in thin films.
7.1 Screw dislocation

A multiply-connected region is defined as a region that it at least contains
one irreducible circuit, i.e. a closed curve that can not be contracted to a single
point without passing out of the region (see Fig. 6.6). Consider a multiply-
connected region V. A Volterra dislocation is defined as the displacement or
rotation discontinuity over the line segment S (2D) or surface S (3D), i.e.
h i
u = u(P+ ) − u(P− ) = b + d × x
h i
ω = ω(P+ ) − ω(P− ) = d (7.1)
where b is the Burgers vector that can be defined as

I h iT
b = E(y) + (x − y) × ∇ × E(y) dy (7.2)
C
and I T
d=− ∇ × E(y) dy (7.3)
C
and E is the strain tensor.
Figure 7.1. Illustraions of dislocations: (a)edge dislocation, and (b) screw dislocation
Historically, there is another type of dislocation: the Somigliana disloca-

tions that are defined as
h i
u = u+ − u− = b, ∀x ∈ S (7.4)
h i
t = t+ − t− = 0, ∀x ∈ S (7.5)
That is the traction is required to be continuous across the slip plane. However,
the solution of such boundary-value problem is difficult, and people have not
found any important applications of such dislocation model.
7.1.1 The solution of screw dislocation

We first derive the solution for the screw dislocation. The kinematics of the
screw dislocation belong to that of anti-plane problem:
u1 = 0, u2 = 0, and u3 = w(x, y) . (7.6)
All the strain components are zero, except the out-plane shear strains
1 ∂w 1 ∂w
xz = , yz = . (7.7)
2 ∂x 2 ∂y
The corresponding non-zero shear stresses are
∂w
σxz = µ (7.8)
∂x
∂w
σxy = µ (7.9)
∂y
The non-trivial equilibrim equation
∂σxz ∂σyz ∂σzz
+ + =0 (7.10)
∂x ∂y ∂z
leads to the governing equation

∂2w ∂2w
+ = ∇2 w = 0 . (7.11)
∂x2 ∂y 2
We denote the displacement jump in w at y = 0 and x > 0 as bz i.e. b = bz ez ,
and the jump condition may be expressed as

lim w(x, −η) − w(x, η) = [w(x, 0)] = bz , η > 0 (7.12)
η→0,x>0
Use the polar coordinate,

∂2 1 ∂ 1 ∂2
∇2 w = + + w=0. (7.13)
∂ 2 r ∂r r2 ∂θ2
Separation of variables and let
w(r, θ) = f (r)g(θ) (7.14)
we have
r 2 d2 f 1 df 1 d2 g
+ + =0. (7.15)
f (r) dr2 r dr g(θ) dθ2
We then end with two ordinary differential equations,
 2
d f 1 df n2 f
 dr2 + r dr − r2 = 0



(7.16)
2
 d g + n2 g(θ) = 0



dθ2
If n = 0, one may find that
g(θ) = A + Bθ (7.17)
f (r) = C ln r + D (7.18)
For n 6= 0,
g(θ) = Cn cos nθ + Dn sin nθ (7.19)
f (r) = En rn + Fn r−n (7.20)
This is true because
d2 1 d n2 n
2
+ − r = n(n − 1) + n − n rn−2 ≡ 0 . (7.21)
dr2 r dr r2
Because the displacement, w, has to be finite, we can only consider the case
n = 0. Again, because the convergence requirement for displacement field,
C = 0; and because of jump condition, A = 0.
By absorbing the constant D into the constant B, the displacement field is
w(r, θ) = Bθ (7.22)
Use the jump condition,
w(r, 2π) − w(r, 0) = b (7.23)
one may find that 2πB = b and hence

b
B= (7.24)
2π
Finally,
θb b y
w(r, θ) = = arctan (7.25)
2π 2π x
and
∂w b y b sin θ
= − 2 2
=− (7.26)
∂x 2π x + y 2πr
∂w b x b cos θ
= = (7.27)
∂y 2π x2 + y 2 2πr
Consequently, the non-zero stress components are
bµ y
σxz = − (7.28)
2π x2 + y 2
bµ x
σyz = (7.29)
2π x2 + y 2
In the cylindrical coordinate,
     
σrr σrθ σrz cos θ sin θ 0 0 0 σxz cos θ − sin θ 0
 σθr σθθ σθz  =  − sin θ cos θ 0   0 0 σyz   sin θ cos θ 0 
σzr σzθ σzz 0 0 1 σzx σzy 0 0 0 1
The non-zero stress components are
σrz = cos θσxz + sin θσyz = 0 (7.30)

bµ
σθz = − sin θσxz + cos θσyz = . (7.31)
2πr
In the following, we calculate the self-energy of the screw dislocation in a
hollow cylinder with inner radius r0 and outer radius R. Note that the self-
energy of a dislocation is defined as the strain energy contribution from stress-
strain field of the dislocation solution in an unbounded region.
Assume that the length of the hollow cylinder is L. The energy per unit
length in z-direction is,
2
1 L 2π R σzθ 2
Z Z Z Z
W 1 σzθ
= dV = rdrdθz.
L L V 2µ V 0 0 r0 2µ
b2 µ R d b2 µ R
Z
= = ln . (7.32)
4π r0 r 4π r0
First, as R → ∞, W/L → ∞. This shows that the self-energy of the dis-
location depends on the size of the crystal. On the other hand, for a finite
size crystal, the dislocation solution of unbounded domain does not hold true
because the image stress caused by the boundary.
Assume that the dislocation is far away from the boundary, the boundary
effecrts are abated inside, one may choose the dimension of the crystal, say `
as R; in polycrystallines, one may choose the size of a grain as R, where the
dislocation resides.
Second, as r0 → 0, W/L → −∞. This abnormality is due to the limitation
of linear elasticity model. Within five atomic spacing of a dislocation core, the
linear elasticity model is no longer valid. In general, the length of the Bergurs
vector is close to the lattice spacing. Therefore, in practice, we usually choose
r0 = 5b or r0 = b/α, 0 < α < 1 such that the elastic self-energy equals to
W µb2 ` W µb2 α`
= ln , or = ln . (7.33)
L 4π 5b L 4π b
By defnition, the self-energy should include the core energy, i.e.
W self = W elas + W core (7.34)
The core energy is relatively small, but may not be negligible, because it is
10% to 20 % of the elastic self-energy. It may be relatively small, but can not
be neglected. Overall, the linear elasticity theory gives a good estimate of self-
energy. In Sec. 4 of this Chapter, we shall discuss the Peierls-Nabarro model,
which provides a means to estimate the core energy.
7.1.2 Image stress of a screw dislocation in a half space

Consider a crystal occupying a half space x ≤ 0. Consider a screw disloca-
tion located at the position x = −` (see Fig. 7.2). The screw dislocation in an
unbounded space gives the following stress distrubution,
∞ bµ y
σxz (x, y) = − (7.35)
2π (x + `)2 + y 2
∞ bµ (x + `)
σyz (x, y) = . (7.36)
2π (x + `)2 + y 2
Figure 7.2. An image screw dislocation
This solution does not satisfy the traction-free boundary condition at x = 0,

because
∞ bµ y
σxz (0, y) = − 6= 0 . (7.37)
2π ` + y 2
2
To enforce the traction-free boundary condition, we place a fictitious screw

0
dislocation with the Bergurs vector, b = −b, at the position x = `, and it
generates the following so-called image stress distribution:
I bµ y
σxz (x, y) = (7.38)
2π (x − `)2 + y 2
I bµ (x − `)
σyz (x, y) = − . (7.39)
2π (x − `)2 + y 2
The total stress distribution is then the superposition of the solution in the infi-
t = σ∞ + σI ,
nite space and the solution of of image stress distribution, i.e. σij ij ij
where the superscript, t, ∞, and I denote the total stress solution, the solution
obtained in the infinite space, and the image stress solution.
By anti-symmetry, the traction-free boundary condition at x = 0 is then
enfored,
t ∞ I by y by y
σxz (0, y) = σxz (0, y)+σxz (0, y) = − + ≡ 0 . (7.40)
2π `2 + y 2 2π `2 + y 2
Remark 7.1.1 1. Note that the image stresses at x = −` and y = 0, i.e. the
position of the real dislocation, are
I I bµ
σxz (−`, 0) = 0, σyz (−`, 0) = . (7.41)
4π`
2. When |x|, |y| >> `,
t t
σxz (x, y) ≈ 0, and σyz (x, y) ≈ 0, (7.42)
which means that outside the region of {(x, y) (x + `)2 + y 2 ≤ `2 }, the total
stress is almost negligible.
7.1.3 Eshelby’s twist: screw dislocation in a finite whisker

Consider a screw dislocation in a finite cylinder (whisker). One may find
that the solution of a single screw dislocation in an infinite space actually sat-
isfies the lateral boundary conditions of the problem:
µb
σzθ = , ∀r ≤ R (7.43)
4πr
σrr = σrθ = σrz = 0, 0 ≤ r ≤ R (7.44)
However, there is one problem there are resulting moments or torques at the
two open ends of the cylinder, i.e.
Z R Z 2π
Mz = rσθz rdrdθ
0 0
µb R µbR2
Z
= 2π rdr = . (7.45)
2π 0 2
To negate the end moment, we superpose two ends moments with the oppo-
0
site direction of Mz = −Mz such that the total moments at the two ends of
the cylinder become zero, and then based on Saint-Venatet’s principle we can
declare the validity of the solution.
The superposed two-end moments will result the following stress distribu-
tion that can be calculated by the elementary torsion formula,
0
0 M r µbr
σθz = z = − 2 (7.46)
J πR
In the last equation, we used the fact that the polar moment of a circular region
is J = πR4 /2.
Then the stress distribution in a whisker is
µb µbr
σθz = − . (7.47)
2πr πR
where the extra term −(µbr)/R may be viewed as an equivalent image stress
steming from the superposed boundary moment.
7.2 Edge dislocation

The edge dislocation problem can be solved as a plane strain problem.
Introduce the Airy stress function, such that
∂2ψ ∂2ψ ∂2ψ

σxx = , σ yy = , and σ xy = − . (7.48)
∂y 2 ∂x2 ∂x∂y
The in-plane equilibrium equation,
∂σxx ∂σyx
+ = 0, (7.49)
∂x ∂y
leads to the following bi-harmonic equation,
∇2 ∇2 ψ = 0 . (7.50)
Let φ = σxx + σyy = ∇2 ψ. Then
∇2 ∇2 ψ = ∇2 φ = 0 , and in the polar coordinate :

∂2 1 ∂ 1 ∂2
+ + φ=0. (7.51)
∂r2 r ∂r r2 ∂θ2
Based on the general solution obtained in the previous subsection, φ has the
following form,
∞
X
φ(r, θ) = (α0 + β0 ln r) + αn rn + βn r−n sin nθ
n=1
∞
X
+ γn rn + δn r−n cos nθ (7.52)
n=1
Because the defect configuration, for an edge dislocation, the region right
above around the dislocation core should be in compression, whereas the re-
gion right below the dislocation core should be in tension, i.e.
φ(r0 , π/2) = φmin , and φ(r0 , −π/2) = φmax . (7.53)
In consideration with the convergence at remote region, i.e. (φ → 0, r → ∞),

the right choice of the solution should be n = 1 and
φ = β1 r−1 sin θ . (7.54)
Then,
∂2 1 ∂ 1 ∂2
+ + ψ = β1 r−1 sin θ . (7.55)
∂r2 r ∂r r2 ∂θ2
Let ψ = h(r) sin θ. One may find that

d2 1 d 1 d 1 d
+ − h = (rh) = β1 r−1 . (7.56)
dr2 r dr r2 dr r dr
By straightforward integration, one can verify that a particular solution is
β1 β1 y
ψe = r sin θ ln r = ln(x2 + y 2 ) . (7.57)
2 4
Consider the jump condition,
Z ∞h i
lim − xx (x, η) − xx (x, −η) dx = b . (7.58)
η→0 −∞
One can determine the constant β1 ,

µb νby
β1 = − ⇒ ψe = − ln(x2 + y 2 ) . (7.59)
π(1 − ν) 4π(1 − ν)
One can then find stress components
µb y(3x2 + y 2 )
σxx = − (7.60)
2π(1 − ν) (x2 + y 2 )2
µb y(x2 − y 2 )
σyy = (7.61)
2π(1 − ν) (x2 + y 2 )2
µb x(x2 − y 2 )
σxy = , and (7.62)
2π(1 − ν) (x2 + y 2 )2
σzz = ν(σxx + σyy ) (7.63)
or in the polar coordinate
µb sin θ
σrr = σθθ = − (7.64)
2π(1 − ν)r
µb cos θ µbν sin θ
σrθ = σzz = ν(σrr + σθθ ) = − . (7.65)
2π(1 − ν)r π(1 − ν)r
It is then easy to find the strain fields by simply applying Hooke’s law of
plane strain condition,
by (µy 2 + (2λ + 3µ)x2
xx = (7.66)
2π (λ + 2µ)(x2 + y 2 )2
by ((2λ + µ)x2 − µy 2 )
yy = − (7.67)
2π (λ + 2µ)(x2 + y 2 )2
b x(x2 − y 2 )
xy = − (7.68)
2π(1 − ν) (x2 + y 2 )2
Figure 7.3. An image edge dislocation
By neglecting all the integration constants, a straightforward integration of the

above strain components gives
b h −1 y λ+µ xy i
u(x, y) = − tan + (7.69)
2π x λ + 2µ x2 + y 2
b h µ λ+µ y2 i
v(x, y) = − − ln(x2 + y 2 ) + (7.70)
2π 2(λ + 2µ) λ + 2µ x2 + y 2
7.2.1 Image stress for an edge dislocation

The solution of the image stress distribution for an edge dislocation is more
complicated than that of a screw dislocation.
Consider an edge dislocation being placed at x = −` inside a half space
(x < 0). The solution obtained from the unbounded space,
∞ µb y(3(x + `)2 + y 2
σxx = − (7.71)
2π(1 − ν) ((x + `)2 + y 2 )2
∞ µb y((x + `)2 − y 2
σyy = (7.72)
2π(1 − ν) ((x + `)2 + y 2 )2
∞ µb (x + `)((x + `)2 − y 2
σxy = (7.73)
2π(1 − ν) ((x + `)2 + y 2 )2
will not satisfy the traction-free boundary condition at x = 0 i.e. σxx (0, y) 6= 0
and σxy (0, y) 6= 0.
If we place a fictitous dislocation at x = ` with the opposite Burgers vector.
The induced image stress fields,
I µb y(3(x − `)2 + y 2
σxx = (7.74)
2π(1 − ν) ((x − `)2 + y 2 )2
I µb y((x − `)2 − y 2
σyy = (7.75)
2π(1 − ν) ((x − `)2 + y 2 )2
I µb (x − `)((x − `)2 − y 2
σxy = (7.76)
2π(1 − ν) ((x − `)2 + y 2 )2
∞ (0, y)+σ I (0, y) =
will cancel the normal stress on traction-free surface, i.e. σxx xx
0, but it can not cancel the shear stress at x = 0. In fact,
∞ I µb `(`2 − y 2 )
σxy (0, y) + σxy (0, y) = 6= 0 . (7.77)
π(1 − ν) (`2 + y 2 )2
To cancel the shear stress on traction-free surface, one has to superpose another
stress field, such that the third stress fields satisfy the condition,
000 000 µb `(`2 − y 2 )

σxx (0, y) = 0, and σxy (0, y) = − . (7.78)
π(1 − ν) (`2 + y 2 )2
Consider the Airy stress function, Ψ(x, y), which satisfies the bi-harmonic
equation,
∇2 ∇2 Ψ = 0 . (7.79)
Introduce the Fourier-sine and the Fourier-cosin transforms,
1 ∞
Z Z ∞
f¯s (ξ) = f (y) sin(ξy)dy, f (y) = f¯s (ξ) sin(ξy)dξ; (7.80)
π −∞ 0
1 ∞
Z Z ∞
¯
fs (ξ) = f (y) cos(ξy)dy, f (y) = f¯c (ξ) cos(ξy)dξ .(7.81)
π −∞ 0
Since σxy must be even in y, the Airy stress function, Ψ, is anti-symmetric in

y. We apply the Fourier-sine transform to Eq. (7.79), and it yields a ordinary
differential equation,
d4 Ψ̄s 2
2 d Ψ̄s
− 2ξ + ξ 4 Ψ̄s = 0. (7.82)
dx4 dx2
Solving (7.82) yields the following solution,
Ψ̄s (x, ξ) = (a0 (ξ) + a1 (ξ)) exp(ξx) + (b0 (ξ) + b1 (ξ)) exp(−ξx) . (7.83)
The boundary conditions,
1. x → −∞, Ψ̄s → 0, ⇒ b0 = b1 = 0; (7.84)

2. x = 0, σxx (0, y) = 0, ⇒ a0 = 0 . (7.85)
Therefore, Ψ̄s (x, ξ) = a1 (ξ)x exp(ξx), and
1 ∞
Z
Ψ(x, y) = a1 (ξ)x exp(ξx) sin(ξy)dξ . (7.86)
π ∞
Using the boundary condition for the shear stress,

∂2Ψ Z ∞
000
−σxy (0, y) = = a1 (ξ)ξ cos(ξy)dy
∂x∂y x=0 0
µb `(`2 − y 2 )
= (7.87)
π(1 − ν) (`2 + y 2 )2
and the definition of the Fourier-cosin transform, one may find that
∞
`(`2 − y 2 )
Z
1 µb
a1 (ξ)ξ = 2 2 2
cos(ξy)dy
π −∞ π(1 − ν) (` + y )
Z ∞
µb `(`2 − y 2 )
= exp(iξy)dy . (7.88)
π 2 (1 − ν) −∞ (`2 + y 2 )2
∞
`(`2 − y 2 )
Z
The last line is because of 2 2 2
sin(ξy)dy = 0.
−∞ (` + y )
Use the residue theorem to evaluate the integra,
∞
`(`2 − y 2 )
Z X
2 2 2
exp(iξy)dy = 2πi Res F (yN )
−∞ (` + y ) yN =i`
iξ`
= 2πi − exp(−ξ`) = πξèxp(−ξ`) . (7.89)
2
Wer then find that
µb`
a1 (ξ) = exp(−ξ`) (7.90)
π(1 − ν)
so that
Z ∞
µb`
Ψ(x, y) = x exp ξ(x − `) sin ξydξ
π(1 − ν) 0
µb`xy
= (7.91)
π(1 − ν)[(x − `)2 + y 2 ]
Figure 7.4. A virtual displacement of a dislocation loop
and
000 µb` (`2 − x2 )y 2 y 2 (3x2 − (y + `)2
σxy = − + (7.92)
π(1 − ν) [(x − `)2 + y 2 ]2 [(x − `)2 + y 2 ]3
000 2µb`xy
σxx = − [3(` − x)2 − y 2 ] (7.93)
π(1 − ν)r6
Indeed, it can be found that
000 µb` `2 − y 2 000

σxy (0, y) = 2 2 2
, and σxx (0, y) = 0 . (7.94)
π(1 − ν) (` + y )
000
Moreover, since σxy (`, 0) = 0, the shear stress acting on the real dislocation
due the traction-free boundary is the stress applied by the image dislocation
(the second dislocation), i.e.
I 000 µb
σxy (−`, 0) + σxy (−`, 0) = . (7.95)
4π(1 − ν)
7.3 The Peach-Koehle force

Consider a dislocation loop undergoing a virtual displacement δη (see Fig.
7.4). An infinitesimal dislocation line segment, dX will sweep through an area,
dA = dX × δη . (7.96)
Note that the direction of dA is its out-normal.

All the atoms on this area will be sujected a discontinuous jump with the
direction and the magnitude of the local Burgers vector, b. The traction forces
on the infinitesimal area can be expressed as σ · dA. Be precise, it is
σ · dA = σ · (dX × δη) (7.97)
If we assume that the work done by the stresses relates to the decreases of the
potential energy of the dislocation,
d(δE) = −b · σ · (dX × δη) (7.98)
The change of the total energy due to the virtual displacement field is
Z Z
δE = − b · σ · (dX × δη) = − (σ · b) × td` · δη (7.99)
L L
where dX = td`.
By definition, the decrease of the potential energy under the virtual displace-
ment field is the external virtual work done along the dislocation loop, i.e.
Z
δE = −F · η = − F` d` · δη , (7.100)
L
where F` is the force per unit length along the dislocation loop.
Hence, we derived the celebrated Peach-Koehle equation,
Z
F= σ · b × td`, and F` = σ · b × t . (7.101)
L
where F` is the force per unit length. In the case of straight dislocation line,
F
we often denote it as .
L
Now, let’s look at a few examples.
To simplify the computation, we denote
g := σ · b. (7.102)
Then the Peach-Koehle force formula can be conveniently written into a matrix
form,
e1 e2 e3
F` = g × t = g1 g2 g3 . (7.103)
t1 t2 t3
Example 7.1 This example is illustrated in Fig. 7.5. We are examing the
external forces exerted on a straight screw dislocation.
Let x = 1, y = 2, z = 3. In this case, the unit vector of the dislocation line
is t = ez , the Burgers vector is b = bez , and the stresses other than self-stress
are
σ = σxz ex ⊗ ez + σzx ez ⊗ ex + σyz ey ⊗ ez + σzy ez ⊗ ey . (7.104)
Figure 7.5. A straight screw dislocation.
and
gx = σxz b, gy = σyz b, gz = 0 . (7.105)
Hence
ex ey ez
F` = g × u = σxz b σyz b 0 = σyz bex − σxz bey . (7.106)
0 0 1
To interprete the meanings of this expression, we would say that the shear
stress, σxy , moves the dislocation line to +x direction, whereas shear stress,
σxz moves the dislocation line towards the negative direction of Y-axis, i.e. -Y
direction.
Example 7.2 In the second example, we consider a straight edge disloca-

tion. This example is illustrated in Fig. 7.6. In this example, again u = ez , but
b = bex , and
σ = σxx ex ⊗ ex + σxy ex ⊗ ey + σyz ey ⊗ ex . (7.107)
Thus,
gx = σxx b, gy = σyx b, and gz = 0 , (7.108)
and
ex ey ez
F` = g × u = σxx b σxy b 0 = σxy bex − σxx bey . (7.109)
0 0 1
Figure 7.6. A straight edge dislocation.
Figure 7.7. Interactions of two parallel screw dislocations
This is to say that the shear stress, σxy , will move the dislocation line along
the slip plane in the positive direction of X-axis. On the other hand, the normal
stress, σxx , will make the dislocation line tranlating along its own direction.
This is an unconservative motion, because if the motion is addmissible, one
has to remove material at one end of dislocation line and add material (atoms)
at the other end of the dislocation line. In literature, we refer such dislocation
movement as “climbing”.
From Eq. (7.109), one may find that if σxx < 0, which means the material is
under compression, the Peach-Koehle force will squeeze the dislocation line up
in Y-axis, and when σxx > 0 it will pull the material apart and let dislocation
line climbing down.
Example 7.3 In this example, we consider the interactions between two

parallel screw dislocations along the Z-axis, t = ez , S1 and S2 . They have
different Burgers vectors, i.e. b1 = b1 ez and b2 = b2 ez . For the dislocation,
S1 , the stress field is
I µb1 sin θ I µb1 cos θ

σxz =− , σyz = ; (7.110)
2π r 2π r
and for the dislocation, S2 , the stress field is
II µb2 (y − y0 )
σxz = − , (7.111)
2π (x − x0 )2 + (y − y0 )2
II µb2 (x − x0 )
σyz = . (7.112)
2π (x − x0 )2 + (y − y0 )2
In this case, the Peach-Koehle force equation is
F` = σyz ex − σxz ey . (7.113)
1. Calculate the force, F1→2
` , which is the force exeretd on the dislocation, S2 ,
by the dislocation, S1 . Let r = r0 and θ = θ0 in (7.110) and substitute them
into (7.113). We have
F1→2
`
I
= σyz I
b2 ex − σxz b 2 ey
x0 ,y0 x0 ,y0
µb1 b2 cos θ0 µb1 b2 sin θ0
= ex + ey
2π r0 2π r0
µb1 b2 µb b
1 2
= cos θ0 ex + sin θ0 ey = r̄0 , (7.114)
2πr0 2πr0
where r̄0 = r0 /|r0 | is the unit vector in r0 direction.
2. Calculate the force exerted on the dislocation S − 1 by the dislocation
S2 . In this case, we let x = 0, y = 0 in (7.111) and (7.112) and substitute them
into (7.113),
F2→1
`
II
= σyz II
b1 ex − σxz b 1 ey
0,0 0,0
µb1 b2 cos θ0 µb1 b2 sin θ0
= − ex − ey
2π r0 2π r0
µb1 b2 µb1 b2
= − cos θ0 ex + sin θ0 ey = − r̄0 . (7.115)
2πr0 2πr0
It is obvious that F1→2
` = −F2→1
` (see Fig. 7.7).
We then conclude that when b1 and b2 are along the same direction, the two
screw dislocation repel each other, if b1 b2 < 0, i.e. b1 and b2 are in opposite
direction, then the two screw dislocations attract to each other.
Remark 7.3.1 [Biot-Savart analogy]

In electro-magnetics, if there are two parallel wires having electric current
passing through, the interaction force between the two wires can be calcualted
by the well-known Boit-Savart law,
Ii
Fi` = t × Bj , i 6= j and i, j = 1, 2 (7.116)
c
where Fi` is the force exerted on the wire i by the magentic field generated
by the wire j; Ii is the electric current density in the wire i, while Bj is the
magnetic induction flux density generated by the wire j, and c is the light speed
in the medium.
In the Peach-Koehle equation, if we define Gj = σ j · t, then
g = σ j · bi = σ j · tbi = Gj bi . (7.117)
We can rewrite the Peach-Koehle force as
Fi` = −bi t × Gj . (7.118)
It has a similar form with the Biot-Savart law. Since bi is the analogy of Ii /c,
we may call the strength of a Burgers vector as the dislocation current density.
By the same token, we may call the stress projection due to the dislocation line
Ej , j = 1, 2 as the stress induction flux.
The only difference between (7.116) and (7.117) are is the minus sign in
(7.117). This is because in electro-maganetics. Two wires with the same (oppo-
site) electric current direction attract (repel) to each other, whereas two screw
dislocation lines having the same (opposite) dislocation current direction repel
(attract) to each other.
7.4 Configuration force: Eshelby’s energy-momentum

tensor
Assume that if the solid that contains the edge dislocation (b = bex ) is
under external hydrostatic pressure, σ11 = σ22 = σ33 = −p, this will cause
the edge dislocation climbing. While an edge dislocation climbs, it does not
produce volumetric strain, thus, σ11 never does work any work in the process.
Therefore, there is actually no real force acting on the dislocation.
Therefore, there is no actual force acting on the dislocation. Then the “vir-
tual force” 1 defined as the decrease of the potential energy change due to the
change of the dislocation position,
∂W
Fη = − , (7.119)
∂η
1 Do not confusion this with the statically admissible virtual forces in continuum mechanics.
Figure 7.8. Eshelby’s argument on configuration force
is really a force due to the change of material’s configuration.

Configuration mechanics has been an active research subject since Eshelby’s
pineeor contribution on configuration force study. In this section, we outline
the basic theory of configuration mechanics, and introduce Eshelby’s energy-
momentum tensor.
In order to evaluate the configuration force acting on a defect, we first cal-
culate the change of potential energy due to the change of configuration.
To do this, we follow the Eshelby’s famous thought experiment. The set-
ting of Eshelby’s thought experiment is a solid that is subjected external forces
or displacement constraints at boundary. Inside the solid, there is a point de-
fect denoted as D, and we link the defect D with its local configuration by
embedding it into an arbitrarily chosen local volume V . We define the local
configuration as the relative position of D inside V . We denote the boundary
of the local volume as L = ∂V (see Fig. 7.8(a)).
The basic idea of Eshelby’s thought experiment is to change the global con-
figuration or the defect position, while comparing the energy change in a local
configuration.
The following is the adaptation of Eshelby’s imaginary operation, which
mainly consists of four steps (I reshuffled the order):
(1) We first change the global configuration, or the position of the defect
by amount of δX in the material configuration. We denote the original local
0
volume containing D as V . When the defect, D, moves its new materials
position, we still choose the same local struction, or local configuration (but
a different sets of material points), to identify it, i.e. we surround the defect
0
D with local volume V , which has the same local configuration as V . It
means the relative position of D is the same with respect to V as it was before
0
with respect to V . The comparison is made under the same local structure,
Eshelby called the local configuration of V is a replica of the original local

0
configuration V .
Under this condition, the material virtual displacement field represents a
change of configuration. One may observe this in Fig. 7.8(b).
0
(2) Before calculating the difference of the energy stored inside V and V ,
we would like to clearify the following point: since the defect changes its
position +δX, this may change the self-stress field as well as image stress
field of the defect, and consequently the energy density at each point. How-
ever, the change of energy density due to the defect movement is at order
δXi δXi ∼ (δX)2 , and it is a second order effect that can be neglected if
δX is infinitesimal. Therefore, we can calculate strain energy stored inside V
0
and V without taking into account the effects of the defect’s movement.
(3) We then calculate the energy difference in two local volume V 0 and
V , which have the same local structure with respect to the defect, due to the
variation in global material location,
Z Z
δE1 = W dV − W dV . (7.120)
V0 V
From Fig. 7.8, one may observe that the area difference between V 0 and V is
ω1 − ω2 , i.e. adding the area ω1 and removing the area ω2 . Hence the stored
strain energy difference is
Z Z
δE1 = W dV − W dV . (7.121)
ω1 ω2
Since δX is infinitesimal,
Z Z
ω1 − ω2 = dA = −δX · dsn
ω1 −ω2 L
where dA = −δX · nds . shown in Fig. 7.9. Therefore,

Z Z
δE1 = −δX · W d`n = −δX` W dsn` . (7.122)
L L
Note that in this step, all the operations are performed in the material config-
uration. We are comparing the energy difference between two adjacent local
material volumes differing a translation.
(3) During a configuration change, the defect moves +δX from its original
material position to the new material position, it will cause the relative material
virtual displacement,
∂ui
δui = δXj . (7.123)
∂Xj
Figure 7.9. Eshelby’s imaginary operation
0
This is to say that if there is no displacements along ∂V , the displacement on
∂ui
∂V is δui = δXj ∀X ∈ L. Then the difference of the work done to the
∂Xj
environment of the two local configurations is:
Z Z Z
ext
δW = 0 · Ti ds − δui Ti ds = − δui σij nj ds
0
LZ L L
= − ui,k σij nj dsδXk , (7.124)

L
which will cause the decrease of the potential energy of the local configuration,
i.e. δE2 = −δW ext .
Then the total variation due to the change of configuration is,
I
δE = δE1 + δE2 = −δX` (W n` − ui,` σij nj )ds
L
I
= = −δX` W δ`k − ui,` σik nk ds
L
To honor the tradtion, the force on the defect is defined to be minus the rate of
increase of the total potential energy of the system, i.e.
∂E
δE = −Finh · δX = δX` (7.125)
∂X`
Therefore the force acting on the inhomogeneity is

I
inh
F` = W δ`k − ui,` σik nk ds . (7.126)
L
In two-dimensional space, the special case, ` = 1, is Rice’s celebrated J-

integral, I
Finh
1 = J = W dx 2 − u σ n
i,1 ik k ds , (7.127)
L
which can be interpreted as the driving force of a crack that grows along x-axis.
The integrand of (7.126) is Eshelby’s another celebrate tensor: the energy-
momentum tensor. The name comes from the fact that the tensor is obtained by
tranlating or giving a motion to the energy of a local configuration. We denote
it as
P`k = W δ`k − ui,` σik . (7.128)
Just like the Peach-Koehle force, Eshelby’s energy momentum tensor was in-
spired by an electromagnetic analogy as well. As Eshelby pointed out, “the
archetypal energy-momentum tensor is Maxwell’s stress tensor in electromag-
netics.” We juxtapose the two for comparison,
PE = W 1(2) − E ⊗ D (7.129)
PM = W 1(2) − ∇u ⊗ σ . (7.130)
where the supercripts, E and M , denote mechanical and electrical energy-

momentum tensors respectively.
In the following, we show that the energy-momentum tensor is divergence-
free in homogeneous solid, which is in essence the path-independence of the
J-integral.
The straightforward differentiation gives,
∂P`k ∂W ∂mn
= δ`k − ui,`k σik − ui,` σik,k
∂xk ∂mn ∂xk
= σmn um,nk δ`k − ui,`k σik
= σmn um,n` − σik ui,k` = 0 . (7.131)
Therefore, for homogenous solids,

I I
F` = P`k nk ds = W δ`k − ui,` σik )nk ds = 0 . (7.132)
L L
For inhomogeneous solids, the above statement is no longer true, this is

because,
∂Cijmn (x)
6= 0,
∂xk
and
∂W
6= σmn um,nk .
∂xk
Suppose that there is a defect at a material point ξi , we assume that this may
be captured by an equivalent inhomogeneous elastic stiffness tensor Cijk` (X−
ξ), i.e. 0
Cijk` , ∀X 6= ξ
Cijk` (X − ξ) = (7.133)
Cijk` (ξ), ∀X = ξ
where
∂2W ∗ 1
0
Cijk` (ξ) = Cijk` − and W ∗ = C ijk` ∗ij ∗k` (7.134)
∂ij ∂k` 2
and ∗ij is the character eigenstrain of the defect.

Therefore, the total strain energy of the inhomogeneous body is
Z
1
E= Cijk` (X − ξ)ij k` dV (7.135)
2 V
By the definition,
Z
∂E 1 ∂Cijk`
Finh
n = − =− ij k` dV
∂ξ 2 V ∂ξn
Z n Z
1 1
= Cijk`,m (δmn − um,n )ij k` dV ≈ Cijk`,n ij k` dV
2 V 2 V
Z h
1 i
= Cijk` ij k` − 2Cijk` ui,j uk,`n dV
2 V ,n
Consider Cijk` ui,j = σk` and integration by parts for the second term of the
integrand.
Z h
inh 1 i
Fn = Cijk` ij k` − (σk` uk,n )` + σk,`` uk,n dV
2 ,n
ZV h
1 i
= Cijk` ij k` − (σk` uk,n ),` dV
2 ,n
IV I
= W δn` − uk,n σk` n` ds = Pn` n` ds . (7.136)
L L
Example 7.4 The asymptotic stress fields for a mode III crack is
KIII θ KIII θ
σ13 = − √ sin , σ23 = √ cos . (7.137)
2πr 2 2πr 2
Figure 7.10. Contour for J-integral around a crack tip
We choose the integration contour Γ : x1 = r cos θ, x2 = r sin θ, − π ≤ θ ≤

π.
The J-integral reads as follows,
I
∂ui
J = W dx2 − σik nk ds
∂x1
ZΓπ
∂u3
= W r cos θ − (σ31 n1 + σ32 n2 )rdθ (7.138)
−π ∂x1
∂u3 σ31 2 /(4µπr).
Consider n1 = cos θ, n2 = − sin θ, = 231 = , and W = KIII
∂x1 µ
Z π 2 2 Z π
KIII KIII θ θ θ
J = cos θdθ − sin2 cos θ − sin cos sin θ dθ
−π 4µπ 2πµ −π 2 2 2
2 Z π
KIII θ θ θ θ
= 2 sin2 cos2 θ − sin2 cos2 − sin2 dθ
2πµ −π 2 2 2 2
2 Z π
KIII K2
θ
= sin2 dθ = III (7.139)
2πµ −π 2 2µ
7.5 Continuum theory of dislocation

One of the popular meso-scale simulations in solids is the discrete disloca-
tion dynamics, which is often referred in the literature as DD. Since Kubin and
Devincre’s pioneer work, numerical simulations of dislocation dynamics has
become an indispensible part of multiscale simulations. The current trend is to
develop con-current multiscale simulations to couple the atomistic molecular
dynamics (MD) simulations with continuum based dislocation dynamics (DD)
simulations. In this section, we shall briefly introduce the basic concepts and
theories of dislocation dynamics.
7.5.1 Volterra and Mura’s formulas

We begin the discussions with the displacement and the stress fields of the
curved dislocations. The general theory of curved dislocations in anisotropic
media was developed by Volterra [1907], De Wit [1960, and Mura [1963,1968].
The special case of curved dislocation in an isotropic medium was attributed
to Burgers [1924] and Peach & Koehler [1950]. The presentation in this book
is an adaptation of Mura’s work with contemprorary flavor.
Before we proceed to derive the Volterra and Mura’s formulas, it is expe-
dient to lay out some useful formulas. Consider a simply connected region,
Ω ∈ IR3 , with a smooth boundary. Define a characteristic function,

1, x ∈ Ω
χ(x) = (7.140)
0, x ∈ /Ω
Consider a (slip) plane S that is characterized by its normal n and its distance
to the origin of the coordinate, s. The Radon transform of χ(x) will be
Z ∞ Z
0 0 0
χ(x )δ(s − n · x )dx = dS (7.141)
−∞ S∩Ω
if Ω = IR3 , we have
Z ∞ Z ∞ Z
0 0 0 0 0
χ(x )δ(s − n · x )dx = δ(s − n · x )dx = dS (7.142)
−∞ −∞ S
Conceptually, we can generalize the Radon projection formula to a two-

dimensional curved surface (2D manifold), S, i.e.
Z Z
0 0 0
f (x )δ(s − n · x )dx = f (x0 )dS 0 (7.143)
Ω S∩Ω
Z ∞ Z
f (x0 )δ(s − n · x0 )dx0 = f (x0 )dS 0 (7.144)
−∞ S
or
Z Z
0 0 0
f (x )δ(S − x )dx = f (x0 )dS 0 (7.145)
Ω S∩Ω
Z ∞ Z
f (x0 )δ(S − x0 )x0 = f (x0 )dS 0 (7.146)
−∞ S
where δ(S − x) is an abbrieviation of δ(dist(S, x)) and dist(S, x) = inf{|x −

y|, ∀y ∈ S}.
Now we consider the following integral,
Z
δ(x − x0 )dS 0 (7.147)
S
where δ(x − x0 ) is Dirac’s delta function in three-dimensional space. Based

on 7.146, we have
Z Z ∞
δ(x − x0 )dS 0 = δ(x − x0 )δ(S − x0 )dx0 = δ(S − x) (7.148)
S −∞
Figure 7.11. Curved dislocation loop L and the Burgers circuit C.
Assume that there is a dislocation loop embedded in an elastic continuum.

To define a dislocation line, we take the tangent at a position x on the dislo-
cation loop, t, as the local direction of the dislocatin. Obviously, t lies on the
tangent plane at point x. We denote the tangent plane at x as S. S is also the
local slip plane. It is assumed that the upper plane of S (denoted by S + ) slips
a distance b relative to its lower plane S − . Choose a circuit around the vector
t in a plane that is perpendicular to t (or t is the normal of the plane). Circle
the circuit (the Burgers circuit) in a direction that makes tas a right-handed
rotation vector.
In this definition, both the tangent vector t and the local Burgers vector, b
could depend on the spatial location, though in the rest of the presentation, we
assume that b is a constant vector. Note that the real slip plane may not be the
tangent plane at x, it could be a curved surface, but the tangent plane of the slip
surface at the interception of Burgers circuit should coincide with the tangent
plane of the dislocation loop at point x.
To homogenize such dislocation field, one may assume that the total dis-
placement gradient can be written as two parts,
∗
ui,j = βij + βij (7.149)
where βij is elastic distortion, and β ∗ is equivalent eigen-distortion, or plastic

distortion. 2
The total strain, ij , elastic strain, eij , and eigenstrain, ∗ ij can be expressed
as
1
ij = ui,j + uj,i (7.150)
2
1
eij = βij + βji (7.151)
2
1 ∗

∗ ij = β + βji ∗
(7.152)
2 ij
where the eigen-distortion is prescribed as
∗
βji = −bi nj δ(S − x) (7.153)
where the normal vector, n, is pointing from S + to S − .
The eigen-distortion caused by slip bi of plane S + may be wretten as
∗
βji (x) = −bi nj δ(S − x) (7.154)
(Question: why is there a minus sign?) Therefore,
1
∗ ij = − bi nj + bj ni δ(S − x) (7.155)
2
Therefore,
Z ∞
ui (x) = − Cj`mn ∗ mn (y)Gij,` (x − y)dy
Z ∞−∞
= Cj`mn ∗ mn (y)δ(S − y)Gij,` (x − y)dy
−∞
Z
= Cj`mn bm nn Gij,` (x − y)dSy (7.156)
S
The above expression was derived by Volterra, and it is called Volterra formula
(Volterra [1907]).
Differentiating (7.156) yields
Z
ui,j (x) = Cj`mn bm nn Gij,`j (x − y)dSy (7.157)
S
and the elastic distortion becomes

Z
βji (x) = Cj`mn bm nn Gij,`j (x − y)dSy + bi nj δ(S − x) (7.158)
S
2 There are many attempts to derive plasticity theory from this formualtion.
Mura showed (Mura [1963]) that the above surface integration can be written
as a line integration,
I
βji (x) = ejnh Cpqmn Gip,q (x − y)bm th d`y (7.159)
L
which is termed as Mura’s formula.

To prove the equivalency between (7.159) and (7.158), we first consider
Stokes’ theorem of a third order tensor field, A = Ajih ej ⊗ eh .
Z I
n · (∇ × A)dS = t · Ad` (7.160)
S
Z I
ek`h nk Ajih,` dS = th Ajih d` (7.161)
S
Let Ajih = ejnh Cpqmn bm Gip,q . We have

I
ejnh Cpqmn bm Gip,q (x − y)th d`y
LZ

= − ek`h nk ejnh Cpqmn bm Gip,q` (x − y) dSy (7.162)
S
∂
where Gip,q` = − Gip,q . Utilizing the identity ek`h ejnh = δkj δ`n − δkn δ`j ,
∂x0`
one can obtaion
Z
− (δkj δ`n − δkn δ`j )nk bm Cpqmn Gip,q` (x − x0 )dS 0
ZS
= − nj bm Cpqm` Gip,q` (x − x0 ) − nn bm Cpqmn Gip,qj (x − x0 ) dS
Z S
= nj bm δim δ(x − x0 ) + nn bm Cpqmn Gip,qj (x − x0 ) dS 0
ZS Z
= nj bi δ(S − x0 )δ(x − x0 )dx0 + nn bm Cpqmn Gip,qj (x − x0 )dS 0
Ω S
Z
= nj bi δ(S − x) + nn bm Cpqmn Gip,qj (x − x0 )dS 0 (7.163)
S
Finally, we showed that (7.158) is equivalent to (7.159).
7.5.2 The Burgers formula

For isotropic materials, the Volterra formula can be simplified and explicited
expressed in terms of elementary line integrals, which are instrumental in con-
temporary discrete dislocation dynamics formulations.
To derive the Burgers formula, we start from the Volterra formula,

Z
0
um (x) = bi Cijk` G∞ km,` (x − x )dSj
0
(7.164)
S
where the surface S is the dislocation surface, which is a cap of dislocation

line C = ∂S, and dSj0 := nj dS.
For isotropic materials, both the elastic tensor and the Green’s function are
quite amieble
Cijk` = λδij δk` + µ(δik δj` + δi` δjk ) (7.165)
1 h λ + µ i
G∞
km (x) = δkm r,pp − r,km . (7.166)
8πµ λ + 2µ
q
0 0 0 0
Denote R = x − x and R = |x − x | = (xi − xi )(xi − xi ).
Then,
1 h
Cijk` G∞
km,` (R) = (λδ δ
ij k` + µ(δ δ
ik j` + δ δ
i` jk )) δkm R,pp`
8πµ

λ+µ i 1 λµ
− R,km` = δij R,ppm
λ + 2µ 8πµ λ + µ
λ+µ
+µ(δim R,ppj + δjm R,ppi ) − 2 µR,mij (7.167)
λ + 2µ
Utilizing the identity,
λ (λ + µ)
=2 −1,
λ + 2µ λ + 2µ
one may find that
1
bi Cijk` G∞
km,` (R) = {µbm R,ppj + µ(b` R,pp` δjm − bj R,ppm )
8πµ
λ+µ
+ 2 µ bj R,ppm − bi R,mij (7.168)
λ + 2µ
Changing the dummy variable, we can then write
Z Z
1 0 1 0 0

um (x) = bm R,ppj dSj + b` R,pp` dSm − b` R,ppm dS`
8π S 8π S
Z
1 λ+µ 0
+ bj (R,pmp dSj − R,jmp dSp ) . (7.169)
4π λ + 2µ S
Consider Stoke’s theorem,

Z I
(∇ × A) · dS = A · d` . (7.170)
S ∂S
Let,
∂
∇= em , A = A, ··· en , dS = dSk ek , and d` = tk dèk = dxk ek .
∂xm
A special case of the Stoke’s theorem is,
Z I
∂A,···
mnk dSk = A,··· dxn . (7.171)
S ∂xm ∂S
Change the free-index, n → k,

Z I
∂A,···
− mnk dSn = A,··· dxk . (7.172)
S ∂xm ∂S
We then have
Z I
−ijk mnk A,···m dSn = ijk A,··· dxk
S ∂S
Z I
−(δim δjn − δin δjm ) A,···m dSn == ijk A,··· dxk (7.173)
S ∂S
which eventually leads to the desired form,

Z I
A,···j dSi − A,···i dSj = ijk A,··· dxk . (7.174)
S ∂S
In (7.169), we may view R,pp as A,pp in the second integral and R,mp as
A,mp in the third integral and then apply the Stoke’s theorem (7.174) to (7.169),
Z Z
0 0 0 0
b` R,pp` dSm − R,ppm dS` = −b` R,pp`0 dSm − R,ppm0 dS`
S S
I
0
= −b` m`k R,pp dxk
C
Z Z
0 0 0 0
bj R,pmp dSj − R,pmj dSp = −bj R,pmp0 dSj − R,pmj 0 dSp
SI S
0
= −bj jpk R,pm dxk
C
We derive the Burgers formula,

Z Z
1 0 1 0
um (x) = bm R,ppj dSj − b` m`k R,pp dxk
8π S 8π C
Z
1 0
− bj jpk R,mp dxk . (7.175)
8π(1 − ν) C
λ+µ 1
In the last line, the identity = is used. Consider the fact that
λ + 2µ 2(1 − ν)
0
xj − xj Rj δmp Rm Rp
R,j = = , and R,mp = −
R R R R3
hence
2 −2Rj
R,pp = and R,ppj = .
R R3
Therefore,
Z I
1 bm R j 0 1 m`k b` 0
um (x) = − 3
dSj − dxk
4π S R 4π C R
I
1 ∂ Rp 0
− pjk bj dxk (7.176)
8π(1 − ν) C ∂xm R
which can be put into an elementary vector form, i.e. the Burgers formula
0 0
b × d` b × R · d`
Z I
b 1 1
u(x) = − Ω − − ∇ . (7.177)
4π 4π C R 8π(1 − ν) C R
0 0
In (7.177), d` = tk dèk = dxk ek , and Ω is the so-called solid angle, which
is defined as the surface area Ω of a unit sphere covered by the surface’s pro-
jection onto the sphere. In this case, the angle is subtended by the dislocation
surface, S, i.e.
0 0
Rj dSj n · dS
Z Z
Ω= = (7.178)
S R3 S R2
where n := R/R is a unit vector from the point x to the dislocation surface,
S.
If the surface is a sphere, dS = R2 dω and
R2 n · dω
I I
Ω = = n · dω
S2 R3 S2
I
= ni ni dω = 4π . (7.179)
S2
7.5.3 Peach-Koehler stress formula for dislocation loop

The objective of this section is to express stress field of a dislocation loop in
terms of line integral. Take derivative of the Bergurs’ displacement formula,
Z I
1 0 1 0
um,` = bm R,ppj` dSj − mnk bn R,pp` dxk
8π S 8π C
I
1 0
= − jpk bj R,mp` dxk (7.180)
8π(1 − ν) C
In the above equation, only the first term is not a line integral. Nevertheless,
we claim that
Z I
0 0
bm R,pp`j dSj = −8πδ(S − x)bm n` − bm j`k R,ppj dxk .
S C
Proof:
Apply Stokes’ theorem,
I Z h i
0 0
ijk φdxk = φ,j dSi − φ,i dSj (7.181)
C S
to the above expression,
I Z
0 0 0
i`k R,pp dxk = R,pp`0 dSj − R,ppj 0 dS`
C ZS
0 0

= R,ppj dS` − R,pp` dSj (7.182)
S
Therefore,
I Z h
∂ 0 0 0
i
j`k R,pp dxk = R,ppjj dS` − R,pp`j dSj (7.183)
∂xj C S
Since
0 1 0
GP (x − x ) = , and ∇2 Gp = −δ(x − x ),
4πR
we then have
2 0 0 0
R,pp = = 8πGP (x−x ) and R,ppjj = 8π∇2 GP (x−x ) = −8πδ(x−x ).
R
Consequently,
I Z Z
0 0 0 0
bm j`k R,ppj dxk = −8πbm δ(x − x )dS` − bm R,pp`j dSj
C S S
Use Radon transformation,
Z Z
0 0 0
δ(x − x )dS` = δ(x − x )n` dS
ZS S
0 0 0
= δ(x − x )n ` δ(S − x )dΩ = δ(S − x)n` (7.184)
IR3
Hence, we verfied the claim.
∗ = −8πb n δ(S − x), we again recover Mura’s formula
Note that βm` m `
I
∗ 1 0
βm` = um,` − βm` =− j`k bm R,ppj dxk
8π C
I I
1 0 1 0
− mnk bn R,pp` dxk − jpk bj R,mp` dxk (7.185)
8π C 8π(1 − ν) C
Shifting the dummy indices, one may find that

I
1 1 1
eij = (βij + βji ) = − jk` bi R,` + ik` bj R,`
2 8π C 2

1 0
−jk` b` R,i − ik` b` R,j + mnk bn R,ijm dxk (7.186)
1−ν
Repeatly using the e-δ identity pij pmn = δim δjn − δin δjm , one has
jk` (bi R,` − b` R,i ) = jk` (δis δ`t − δ`s δit )bs R,t = jk` i`p stp bs R,t
= pst jk` ip` bs R,t = pst (δji δkp − δjp δki )bs R,t
= (kst δji − jst δki )bs R,t (7.187)
Similarly, one may find,
ik` (bj R,` − b` R,j ) = (kst δij − ist δkj )bs R,t (7.188)
which enable us to write
I
1 h 1 1 i
eij = −bs R,ppt kst δij − ist δkj − jst δki
8π C 2 2

1 0
+ mnk bn R,ijm dxk (7.189)
1−ν
For linear isotropic elastic materials,
σij = Cijk` ek` , and Cijk` = λδij δk` + µ(δik δj` + δi` δjk ) (7.190)
Finally, one can obtain the Peach-Koehler formula for stress field of a disloca-
tion loop,
I
µ bn 0 0
σij = R,mpp + (jmn dxi + imn dxj )
4π C 2
bn 0

+ kmn (R,ijm − δij R,ppm )dxk (7.191)
1−ν
Considering,
2Rm ∂ 2
R,ppm = − 3 =
R ∂xm R
0

R,ijm = ∇m · ∇i ⊗ ∇j R) (7.192)
One can re-write the Peach-Koehler formula in a vector form,

I I
µ 0 1 0 µ 0 0 1
σ = (b × ∇ ) ⊗ d` + d` ⊗ (b × ∇ )
4π C R 4π C R
I
µ 0 0
= − ∇ · (b × d` ) · (∇ ⊗ ∇ − 1∇2 )R . (7.193)
4π(1 − ν) C
7.6 Discrete Dislocation Dynamics (DD)

The first discrete dislocation dynamics simulation was attempted in late
1980s by Lepinous and Kubin [1987] and Ghoniem and Amodeo [1988]. The
simulations were conducted then were the interactions among infinitely long
straight dislocations. Since 1990s, more realistic DD simulations have been
proposed in situations that are involved with more complicated micro-structures.
In the following, we shall outline one of the latest formulations of DD simula-
tions.
7.6.1 Galerkin weak form formulation

The Galerkin weak form formulation is proposed by Ghoniem and Sun and
their co-workers.
The following presentation is mainly based on a series papers by Ghoniem
et al [1990] [2000], and [2004].
In this approach, the formulation focus on simulating one dislocation loop
among many different dislocation loops.
To formulate the discrete dislocation dynamics, we employ the virtual work
principle. For a given virtual displacement field, δx, the virtual work will be
balanced on the dislocation loop considered.
The internal virtual work consists of the virtual work done by all the stresses
acting on the dislocation loop, which includes the stress fields of all other dis-
location loops and the stress field due to external loads, the virtual work done
by the self-stress field. The external virtual work is mainly the virtual work
done by the friction forces that resist the motion of the dislocation loop.
We first consider the virtual work due to all other internal stresses except
the self-stress,
I I h i
δWP K = dFP K · δx = (b · Σ) × d` · δx
IC
C
I
= b · Σ × t d` · δx = (ijk Σjm bm tk δxi )d` ,(7.194)
C C
where b is the Burgers vector, t is the tangential vector along the dislocation
loop, and
I e
Σij = σij + σij (7.195)
I are the stress fields of all other dislocation loops inside the solid,
Here σij
which can be expressed as
I
I µ h1 0 0
σij = bn R,mpp (jmn dxi + imn dxj )
4π C 2
1 i 0
+ kmn (R,ijm − δij R,ppm ) dxk (7.196)
1−ν
e is the stress field due to externally applied loads.
and σij
Denote
fiP K = ijk Σjm bm tk . (7.197)
One may write I
δWP K = fiP K d`δxi . (7.198)
C
In principle, the virtual work done by the self stress field can be also ex-
pressed by Eq. (7.196). However, in that case, Eq. (7.196) would become
a singular integral, which can be evaluated in the sense of Cauchy principal
value.
Since the core of a dislocation loop has specific physical meanings, it would
be appropriate to treat the virtual work of self-stress field separately. Gavazza
and Barnett [1976] expressed the virtual work of the self-stress field of planar
curved dislocation loop in terms of a single integral expression,
I h 8 i
00
δWself = E(t) − E(t) + E (t) ln κ − J(L, p) n · δxd`
C κ
+[dU ]core (7.199)
where E(t) = 12 σij (t)bi nj , is related to the core size, κ is the curvature of
the dislocation line, J(L, p) is a non-local interaction term, and [dU ]core is the
virtual work contribution from the core of the dislocation loop. Since [dU ]core
is related to the dislocation mobility, this term may be absorbed into the friction
force.
Let,
8 i
self 00
E = E(t) − E(t) + E (t) ln κ − J(L, p) (7.200)
κ
and
fiself = Eni (7.201)
The total active forces acting on a dislocation loop are
fiT = fiP K + fiself (7.202)
In many cases, it has to include the change of chemical potential induced Os-
motic force. Since the change in chemical potnetial per vacncy or interstitial
will cause the dislocation loop climbing, or causing the none-conservative dis-
location loop movement, the Osmotic force is usually responsible for the dis-
location loop climb (see Hirth, Rhee, and Zbib [1996]).
When a dislocation loop starting to move, it has to overcome the friction
forces that resist its motion. The friction forces consist of (1) extrinsic resis-
tances due to alloying, impurity atoms, Peierls stress (this part of force coming
from [dU ]core ), etc., and (2) Intrinsic friction forces that are due to the atom-
istic bond force in a surface separation (fracture) process. Empirically, one
can always assume that the friction forces are proportional to the dislocation
velocity, such that
I I
f riction
δW = Cik Vk d`δxi = C · V)d` · δx (7.203)
C C
where
dx
V= (7.204)
dt
and C is called the resistivity matrix, which has three independent components
in an isotropic medium (two for glide motion and one for climb motion),
 
C1 0 0
[Cik ] =  0 C2 0  (7.205)
C1 0 C3
Then the principle of virtual reads
I
int f ric
δW − δW = 0, ⇒ fiT − Cik Vk d`δxi = 0 . (7.206)
C
7.6.2 Finite element implementation

Truncating the dislocation loop into Ns segments, and mapping each seg-
ment into a one-dimensional parametric space, i.e., NI : [xI−1 , xI ] → u ∈
[0, 1]. Thereby, for x ∈ NI ,
r
∂xi ∂xi
d` = du (7.207)
∂u ∂u
Consider the finite element discrettization,
N
X DF
xhi (u, t) = Nim (u)qm (t) (7.208)

m=1
where Nim (u) is the finite element shape function. The discreteized velocity
field is
N
X DF
h h
Vi = xi,t = Nim (u)qm,t (t) . (7.209)
m=1
Denote the gradient of FEM shape function as Bim (u) := Nim,u (u). The
line integration element will be
NX
DF 1/2
d` = (x` x` )1/2 du = qp qs B`p (u)B`s (u) du (7.210)
p,s=1
Figure 7.12. Simulations of Discrete Dislocation Dyanmics
We can evaluate the internal stresses acting on the dislocation loop by quadra-
ture integration, i.e.
Nloop Ns Qmax
I µ X X X h1
σij = bn wα R,mpp (jmn xi,u + imn xj,u )
4π 2
γ=1 β=1 α=1
1 i
+ kmn (R,ijm − δij R,ppm )xk,u (7.211)
1−ν
where Nloop is the total number of dislocation loops, Ns is the total number of
segments in each dislocation loop, and Qmax is the total number of quadrature
point in a segment, and wα is the quadrature weight.
Denote each segment of the dislocation loop as Lj . The discretized weak
formulation is
Ns QX
X max N
X DF h N
X DF i
Nim (u)δqm fiT − Cik Nkn q̇n
j=1 α=1 m=1 n=1
NX
DF 1/2
× qp qs B`p B`s wα = 0 . (7.212)
p,s=1
Define the generalized force vector,

QX
max NX
DF 1/2
h
fm = fiT Nim (u) qp , qs B`p B`s wα (7.213)
α=1 p,s=1
and the resistivity matrix {γmn }, in which

QX
max NX
DF 1/2
γmn = Nim (u)Cik Nkn (u) qp , qs B`p B`s wα (7.214)
α=1 p,s=1
Then, we can put the dislocation loop weak form into a matrix form,
Ns h
X h dq i iT h i
[f ]j − [γ]j δq = 0 , (7.215)
dt j j
j=1
which leads to the global matrix formulation,
dQ i T h i
h i h ih
F − Γ δQ = 0 , (7.216)
dt
where
h i h i1×NDF
F = AN s
j=1 f (7.217)
j
h i h iNDF ×NDF
Γ = AN s
j=1 γ (7.218)
j
Sovling (7.216) yields,

h dQ i h i−1 h i
= Γ F (7.219)
dt
Employing any desirable time stepping algorithm, one find the updated dis-
location loop configuration or position by
h i h i h i−1 h i
Q = Q + Γ F ∆t (7.220)
n+1 n n+α n+α
where 0 ≤ α ≤ 1.
This is the state of the art discrete dislocation dynamics formulation.
7.7 The Peierls-Nabarro Model

7.7.1 Hilbert transform
The Hilbert transform is a particular case of the Cauchy integral transforms.
Let L be a closed smooth contour and φ(ζ) be an arbitrary Holder continuous
function specified on L and vanishing at infinity. Cauchy integral transforms

are the following pair of mutually invertible integrals (e.g. Zhdanov [1984]),
Z
1 φ(ζ)
ψ(ζ0 ) = dζ (7.221)
πi L ζ − ζ0
Z
1 ψ(ζ)
φ(ζ0 ) = dζ (7.222)
πi L ζ − ζ0
One special case of great value for applications is that L real axis, Im(ψ(ζ)) =
g(x), Re(ψ(ζ)) = 0, Re(φ(ζ)) = f (x), and Im(φ(ζ)) = 0. That is φ(ζ) =
f (x) + i0 and ψ(ζ) = 0 + ig(x). Here f (x) and g(x) are real functions of
a real variable x satisfying the Holder condition for any finite x and vanishing
at infinity. This special case of Cauchy integral transforms is the so-called the
Hilbert transforms:
1 ∞ f (t)dt
Z
g(x) = H(f (x)) = (7.223)
π −∞ x − t
1 ∞ g(t)dt
Z
f (x) = −H(g(x)) = − (7.224)
π −∞ x − t
Note the position between x and t and position between ζ and ζ0 .

Hilbert transform table is available in many mathematics handbooks. In
general, one can find Hilbert transform via Cauchy’s residue theorem.
The following are a few examples:
1 ∞
Z
1 dt
H = = δ(x − b) (7.225)
π(b − x) π −∞ π(b − t)(x − t)
1 ∞
Z
1 dt x
H = = (7.226)
(x2 + a2 ) π −∞ (t2 + a2 )(x − t) a(x2 + a2 )
1 ∞ sin(bt)dt
Z
H sin(bx) = = − cos(bx) (7.227)
π −∞ (x − t)
7.7.2 The Peierls-Nabarro dislocation model

In the early development of dislocation theory, scientists were concerned
with two important issues: (1) What is the size of a dislocation for a given
Burgers vector? (2)How much force is needed to move a dislocation out of its
stable position?
The second question is the so-called dislocation mobility, which is central
to the understanding of the ductile material strength. The Peierls-Nabarro dis-
location model tries to answer this question.
Before we discuss Peierls-Nabarro model, we first examine the mechanical
fields of a straight edge dislocation (displacement fields are given up to a rigid
Figure 7.13. The Peierls-Nabarro Model
body displacement) ,
b h −1 y xy i
ux = tan + (7.228)
2π x 2(1 − ν)(x2 + y 2 )
b 1 − 2ν
h x2 i
uy = − ln(x2 + y 2 ) + (7.229)
2π 4(1 − ν) 2(1 − ν)(x2 + y 2 )
µb y(3x2 + y 2 )
σxx = − (7.230)
2π(1 − ν) (x2 + y 2 )2
µb y(x2 − y 2 )
σyy = (7.231)
2π(1 − ν) (x2 + y 2 )2
µb x(x2 − y 2 )
σxy = (7.232)
2π(1 − ν) (x2 + y 2 )2
As evident from the above equations, the stress fields are singular at the ori-
gin. Therefore the analytical solution presented above is no longer accurate
near the core of the dislocation. To remove this singularity inside the dislo-
cation core, Peierls [1940] and Nabarro [1947] included the discrete atomic
nature of the material and proposed the following lattice correction model.
The Peierls-Nabarro model(PN model) for a straight edge dislocation is de-
scribed using two semi-infinite simple cubic crystals as shown in Fig. 5.4. The
formal glide plane is y = 0. The two elastic half spaces are terminated on the
planes y ≥ d/2 and y ≤ −d/2. At the middle of glide plane, a non-Hookean
slab of width d (atomic spacing) joins the two half spaces. The symmetrical
configuration indicated in Fig. 5.4 suggests that this is done by cutting the
perfect crystal into two halves along the y = 0 plane, and inserting an addi-
tional layer of atoms in the upper half of the crystal space, which displaces the
upper half crystal moving rigidly a distance 0.5b in both positive and negative
x-direction, and we then re-weld the two half crystals.
Before the "re-welding", the initial dis-registry (misalignment) in x-direction
of two vertical atom layers with respect to the upper and lower half crystal
spaces is

b
 2, x > 0



−
φ0x (x) := Xm+
− Xm = m = ±1, ±2, · · · ± ∞ (7.233)

 b
 − , x<0

2
After the re-welding, the misalignment, or the discontinuity, between the atom
layer in the upper part of crystal and the same atom layer (m) of the lower part
of the crystal becomes
φx (x) = x+ − x− + + − −
m = Xm + u (x) − (Xm + u (x))
m
b + −
 2 + u (x) − u (x), x > 0



φx (x) =
 b
 − + u+ (x) − u− (x), x < 0


 2
b
 2ux (x) + 2 , x > 0



=

 b
 2ux (x) − , x < 0

2
By antisymmetry, we assume that ux (x) = u+ (x) = −u− (x).
At the remote boundary, dis-registry is enforced to be zero, i.e. there is no
discontinuity at the remote boundary
b
φx (x) → 0, when x → ±∞ ⇒ 2ux (x) ± = 0, x → ±∞ (7.234)
2
b
Therefore, ux (±∞) = ∓ . This implies that the total displacement along the
4
interface should be
Z ∞
dux b
ux (∞) − ux (−∞) = dx0 = − (7.235)
−∞ dx x=x 0 2
Based on Eshelby’s interpretation (Eshelby [1949]), one may think that
Peierls-Nabarro model deploys a continuous edge dislocation distribution along
the cohesive interface with its local Burgers vector density as b0 (x0 ) to replace
a single dislocation with a Burgers vector b. To make sure that these two
dislocation systems are equivalent, we enforce the following condition on net
Burgers vector equality,
Z ∞ Z ∞
dux 0
−2 dx = b0 (x0 )dx = b (7.236)
−∞ dx x=x 0
−∞
From the above relation, one may derive that the distribution density of Burgers
dux 0
vector should be b0 (x0 ) = −2 (x ).
dx
The strains near the dislocation core are large, and therefore use of Hooke’s
law for the stresses is unappropriate. One the other hand, it is relevant to use
the periodicity of the lattice, which implies σxy to be a periodic function of
φ(x). We therefore assume that,
2πφ
x
σxy (x, 0) = C sin (7.237)
b
2πφx (x)
When φx (x) << 1,σxy (x, 0) ∼ C . Under small deformation limit,
b
it is assumed that the cohesive law should comply to Hooke’s law as well (is
this a good assumption?), i.e.
µφx (x) 2πφx (x)

σxy (x, 0) = 2µxy = =C (7.238)
d b
µb
which determines the constant C = . Note that the shear strain inside the
2πd
cohesive interface is (see Fig. 5.4)
φx (x)
γxy = (7.239)
d
Thereby, one obtain that
µb 4πux (x) µb 4πu (x)

x
σxy (x, 0) = sin ±π + =− sin (7.240)
2πd b 2πd b
One can calculate the shear stress inside the cohesive strip due the continu-
ously distributed dislocation via superposition. At y = 0,
Z ∞ 0
µ b (t)dt
σxy (x, 0) =
2π(1 − ν) −∞ x − t
Z ∞
µ (dux /dx)x=t dt
= − (7.241)
π(1 − ν) −∞ x−t
One may also derive the above integral equation based Boussinesq solution of
linear elastic half space (e.g. Timoshenko and Goodier [1972]).
Apparently, σxy (x, 0) is proportional to the Hilbert transform of dux /dx.
Thereby the inverse Hilbert transform gives
(1 − ν) ∞ σxy (t, 0)dt
Z
dux
= (7.242)
dx µ −∞ x−t
Integrating this yields,
∞
(1 − ν)
Z
u(x) = σxy (t, 0) ln |t − x|dt (7.243)
µ −∞
Using ((7.240)) and ((7.241)), one can obtain the well-known Peierls-Nabarro
integral equation for unknown displacement field, ux (x),
Z ∞
(dux /dx)x=t dt b(1 − ν) 4πux
= sin (7.244)
−∞ x−t 2d b
which is a singular, nonlinear integral equation with unknown function ux (x).
Luckily, the solution of the above integral equation can be found in closed
form 3 ,
b x
ux (x) = − tan−1 (7.245)
2π rc
where rc = d/2(1 − ν), which is a parameter that characterizes the size of the
dislocation core. When |x| < rc , the dis-registry φx (x) > b/4. At x = rc ,
ux (rc ) = −b/8 and φx (rc ) = b/4.
Substituting ((7.245)) into ((7.240)) and utilizing the trigonometry identity
y
tan−1 (y) = sin−1 p
1 + y2
one can find that
µb x
σxy (x, 0) = (7.246)
2π(1 − ν) x + rc2
2
On the other hand, by virtue of (7.245) the displacement gradient in x-direction

is du
x b rc
=− (7.247)
dx x=t 2π t + rc2
2
and the Hilbert transform of the above expression is

du br 1 b x
x c
H =H − =− (7.248)
dx 2π x2 + rc2 2π x2 + rc2
3 My guess is that the reason why they took sine function as the cohesive law was to match the exact solution
of this particular integral equation, which people had known before.

where the following Hilbert transform formula is used,

1 1 x
H 2 =
x + rc2 rc x2 + rc2
Based on ((7.241)),
µ du
x µb x
σxy (x, 0) = − H = (7.249)
(1 − ν) dx 2π(1 − ν) x + rc2
2
which is the same as the expression obtained above.
7.7.3 Misfit Energy and the Peierls Force

As we mentioned before, one of the motives to discuss the Peierls-Nabarro
dislocation model is to find the critical stress needed in order to move a dislo-
cation from its stable position. This question can not be answered by analyzing
a Volterra dislocation.
To find the critical stress to move a dislocation, we first examine the stored
elastic energy due to an edge dislocation. The total elastic energy stored in-
duced by an edge dislocation may be divided into two parts: the energy stored
inside the elastic crystal and the energy stored inside the cohesive layer. Since
the two crystal half spaces maintain substantially perfect lattice structure, most
of shear deformation is confined within the cohesive layer. For this reason, we
call the energy stored inside the cohesive layer as the misfit energy.
The shear strain, in fact that it is the eigen shear strain because it is the
“shear strain” caused by the local jump, inside the cohesive zone is,
φx (x) 2ux (x) + (b/2)
γxy = = , x>0 (7.250)
d d
The misfit energy for a pair of atomic planes is,
1 γxy 0
Z
0
∆W = − σxy (x, 0)dγxy b · d
2
Z ux 0
= σxy dux b · d (7.251)
−b/4
The factor of half is introduced in calculating the misfit energy because it is

getting shared between two planes. Note that when u(x) = −b/4 → γxy = 0.
Therefore,
µb2 ux µb3
Z 4πu 4πu ux
x x
∆W (x) = sin dux = 2 cos
2πd −b/4 b 8π d b −b/4
µb 3 4πux
= 2
1 + cos (7.252)
8π d b
Substitute,
b x
ux = − tan−1 (7.253)
2π rc
to obtain the misfit energy for a pair of atomic planes as,
µb3 x
∆W = 2 1 + cos 2 tan−1 ( ) (7.254)
8π d rc
Let the distance of the center of the dislocation from the nearest position of
symmetry be ξ = αb, where α is a variable. Then the position of all the atoms,
on the two faces of the slip plane are defined by

 2m b

the upper half crystal
xm = 2 (7.255)
b
 (2m − 1)
 the lower half crystal
2
and m = 0, ±1, ±2, ±3, · · · (see Fig. 7.14).
Then the total misfit energy is the summation,
∞
X
W = ∆W (2m) + ∆W (2m − 1)
m=−∞
··· +∞
X µb3 X
−1 b
= 1 + cos 2 tan (α + 0.5n)( )
8π 2 d n=−∞ rc
n=0,±2,±4
··· +∞
X µb3 X
−1 b
+ 1 + cos 2 tan (α + 0.5n)( ) (7.256)
8π 2 d n=−∞ rc
n=±1,±3
which can be combined into a single expression, i.e. x = (α + 0.5n)b and

n = 0, ±1, ±2, .... Therefore summing up over all the atomic planes we get
the total misfit energy as
+∞ +∞
X µb3 X
−1 b
W = f (n) = 1+cos 2 tan (α+0.5n)( ) (7.257)
n=−∞
8π 2 d n=−∞ rc
This may be transformed using the Poission’s summation formula in Har-

monic analysis:
+∞
X +∞ Z +∞
X
f (n) = f (x) exp(−i2πxn)dx, (7.258)
n=−∞ n=−∞ −∞
where f (x) is an even function, it reads

+∞
X +∞ Z +∞
X
f (n) = f (x)cos(2πxn)dx, (7.259)
n=−∞ n=−∞ −∞
Figure 7.14. The Nabarro counting scheme
where we have used the fact that the function f (n) is even in n. We can rewrite
the above relation as,
+∞
X Z +∞ ∞ Z
X +∞
f (n) = f (x)dx + 2 f (x)cos(2πxn)dx, (7.260)
n=−∞ −∞ n=1 −∞
Therefore we can rewrite the total misfit energy from the equation ((7.257))
as,
+∞
µb3
Z
W = (1 + cos(2 tan−1 z))dx
8π 2 d −∞
+∞ Z
µb X +∞
3 dz
+ (1 + cos(2 tan−1 z))cos 2πn − 2α dx
4π 2 d (1 − ν)b
n=1 −∞
(7.261)
where z = (α + x2 ) rbc = 2(1 − ν)(α + x2 ) db . Therefore dx

dz
= (1 − ν) db
and dx = (1−ν)b dz. Using these transformations and that cos(2 tan−1 z) =
d
2
1+z 2
− 1, we get,
+∞
µb2
Z
1
W = dz
4π 2 (1 − ν) −∞ 1 + z
2
+∞ Z +∞
µb2 X dz dz
+ cos 2πn − 2α
2π 2 (1 − ν) (1 − ν)b 1 + z2
n=1 −∞
(7.262)
The first integral above can be calculated using the Cauchy residual theorem,
that is we use the result:
Z +∞
1 1
2
dz = 2πiRe( )=π
−∞ 1 + z 1 + z2
where Re(.) denotes the residual. Therefore the first term of the total misfit en-
µb2
ergy as 4π(1−ν) . The second term in equation ((7.262)) can be further reduced
to,
+∞ Z +∞
µb2 X 2πnzd dz
cos(4πnα) cos
2π 2 (1 − ν) −∞ (1 − ν)b 1 + z 2
n=1
2πnd
To evaluate this term we again use Cauchy residual theorem. Say k = (1−ν)b ;
then the integral in the above equation is equal to,
Z +∞ ikz
e
2
dz
−∞ 1 + z
which is equal to πe−k . Therefore we obtain the total misfit energy as,
+∞
µb2 µb2 X −4πrc n
W = + 2 πe b cos(4πnα) (7.263)
4π(1 − ν) 2π (1 − ν)
n=1
The term in n = 1 dominates the sum, therefore we have,
µb2 µb2 4πr

c
W (α) = + exp − cos 4πα (7.264)
4π(1 − ν) 2π(1 − ν) b
The corresponding force acting on dislocation is given by,
1 dW (α)
F =− (7.265)
b dα
Note that the dislocation moves a distance −αb.
dW (α)/dα reaches to maximum when sin 4πα = 1. From the relation that
σxy = F (b × 1)(unit thickness in z-direction), the critical shear stress to move
the dislocation by one lattice site is
2µ 4πr
c
σ= exp − (7.266)
(1 − ν) b
where F is called the Peierls force and σ is called the Peierls stress, which are
required to move a dislocation over a Peierls barrier.
A more physically realistic restoring stress is obtained if we use relative
displacement (of the two half planes) instead of the lattice displacement in the
above discussion. In the following, a more recent treatment of the PN model is
outlined (Joós and Duesbery, 1997) which considers the relative displacement
instead of the independent lattice displacements in two half planes. We restrict
our attention to the case of a straight edge dislocation. The new model predicts
a Peierls stress which differs from the above mentioned expression by a factor
of two in both the exponential and the coefficient of the exponential. This
approach is also valid for the case of narrow dislocations. By f (x) we define
the displacement of the upper half of the crystal with respect to the lower half.
If c is a constant, then f (x − c) corresponds to a dislocation translated by
c. For a discrete lattice this can be understood like this: If the dislocation is
introduced at c, then the atomic planes at a position mb in the upper half of the
crystal will experience a displacement of f (mb − c) along the Burgers vector.
The total misfit energy in this case can be written as:
+∞
µb3 X
−1 mb − c

W (c) = 1 + cos 2 tan ( ) (7.267)
4π 2 d m=−∞ rc
Note the difference of factor of half in the expression of W from the earlier dis-
cussion. This is because we are no longer treating the two half planes indepen-
dently, but we are using a relative displacement. Using further manipulations
and substituting Γ = rc /b and y = c/b we have,
+∞
µb2 X Γ
W (y) = (7.268)
4π (1 − ν) m=−∞ Γ + (m − y)2
2 2
W (y) is an even periodic function of period 1. Using this information we can

express the energy as the sum,
+∞
a0 X
W (y) = + an cos 2πny (7.269)
2
n=1
Where we can calculate the Fourier coefficients in the usual manner. After
substituting the value of these Fourier coefficients, we get the expression for
the total misfit energy as,

+∞
µb2 µb2 X
W (y) = + e−2πnΓ cos 2πny (7.270)
4π(1 − ν) 2π(1 − ν)
n=1
For the limit of wide dislocations (Γ 1), only the first exponential term is
kept. Then in the limit of wide dislocations we have,
µb2 −2πrc 2πc

W (c) = 1 + 2e b cos (7.271)
4π(1 − ν) b
n o
From which we obtain, (using the relation σ = max 1b dW
dc )
µ 2πr
c
σ= exp − (7.272)
(1 − ν) b
Note the difference between the above stress and the one obtained in the equa-
tion ((7.266)).
Figure 7.15. Paul Dirac (left), Wolfgang Pauli (middle) and Rudolf Peierls (right) in discus-
sion at the international Conference on Nuclear Physics, Birmingham, 1948
7.7.4 Story of the Peierls-Nabarro Model

The following is an account on the discovery of Peierls-Nabarro model,
which was given by the late Professor, Egon Orowan, of Massachusetts In-
stitute of Technology, who was a well-known physicist and material scientist
at the time.
"1937 I was invited to work at the University of Birmingham, in the Physics Department
which had just taken over by M. L. E. Oliphant (now Sir Mark Oliphant). I felt that it
would be urgent to know the width of the dislocation belt and the stress required to
move it. The simplest assumption about this was the one made by Taylor, that the
stress was zero; however, the extremely high yield stress of many hard materials such
as diamond (which could be remarkably free from imperfections and thus could not
contain too many dislocations) indicated that the most frequent cause of the hardness
of crystalline materials was the high shear stress required to move a dislocation. I found
that the width of the dislocation and the stress for moving it could be calculated, with
a crude approximation, simply enough by assuming that the shearing force between
the opposite shores of the slip plane in a dislocation was a sine function of the relative
shear displacement (the initial tangent of the sine, of course, was given by the elastic
modulus).
One the other hand, displacement and shear traction at the surface of a half-space were
connected by the equations of Boussinesq; equating the stresses and displacements of
the sine approximation with those of Boussinesq led to an integral equation which was
the solution of the problem It would have taken me days or weeks of study to solve
it; fortunately I was a daily guest in the hospitable house of the brilliamnt theoretical
physicist Rudolf Peierls. He solved the equation, if I remember well, within a few
hours, and he also drove me to a conference at Bristol University in 1939 where I gave
a paper and he gave another on the problem he had just solved.
The calculation of the width of the dislocaiton and of the Peierls-Nabarro stress required
for moving it was repeated and improved by Nabarro in 1947. The result was puzzling
at first: the width calculatied by Nabarro amounted to a few atomic spacings while
Peierls obtains an order of magnitude of thousands of spacings. After some research in
Birmingham and in Cambridge (where I was wat the time) I discovered the sheet with
Peierls’s calculations in my desk; Peierls checked it and found that a factor of 2π was
accidentally omitted in an exponent, which amounted to a factor of about 1000 in the
result.
Of course, the calculation with the sinusoidal approximation is useless in most interest-
ing cases of directinal bonds, in transition metals and the hard non-metallic crystals."
From The Sorby Centennial Symposium on the History of Metallugy, MSC,
Vol. 27, 1963, pages 368-369.
7.8 Dislocations in the epitaxial thin film

The thin film is the basic configuration structure for integrated circuits, com-
puter memories (RAM), and various sensors, filters, and other electronic de-
vices. Study the mechanical, chemical, and electrical properties of the thin
films has particular significance for nano-technologies.
The ancient Greek word πι (epi–placed or resting upon) and the word
τ αξιζ (taxis – arrangement) are the root of the modern word epitaxy, which de-
scribes an extremely important phenomenon exhibited by thin films. Epitaxy
refers to a single-crystal film formation on top of a crystalline substrate and
both have the exactly the same crystal structure as the thin film. 90 % of thin
films used in semi-conductor and computer industry, communication industry,
and sensor and information industry are epitaxial thin films. To grow various
Figure 7.16. An epitaxial thin film.
defect-free epitaxial thin films has been the main challenge in semi-conductor
industry in the past half century.
In this section, we shall introduce the two basic dislocation models in thin-
film mechanics.
7.8.1 Frenkel & Kontorova model and Frank & van der
Merwe model
The Frenkel & Kontorova dislocation model is a one-dimensional disloca-
tion model, which was proposed in 1937. This model was studied in detailed
by Frank and van der Merwe [1950ab], and they applied it to study thin film
mechanics or epitaxial thin film mechanics.
In Frenkel & Kontorova model, the thin film is modeled as one dimensional
monolayer with lattice spacing af , and the substrate is modeled as large slab
with lattice spacing as , and as 6= af and the lattice misfit is ∆ = af − as (see
Fig. 7.17).
The row of atoms in the thin film are under combined influence of harmonic
forces between the nearest neighbours in the monolayer and non-linear inter-
action forces from substrate. Since the substrate is assumed much larger in
dimension than the thin film, it is assumed to be rigid. The interaction between
the thin film and substrate, or the force exerted on the thin film by the sub-
Figure 7.17. Frank-van der Merwe dislocation thin film model
strate is characterized by a sinusoidal potential with the amplitude 21 W (see

Fig. 7.17).
Make the position (the open circle in Fig. 7.17) of the m-th atoms in the
un-strained monolayer as
Xm = maf , m = 0, ±1, ±2, · · · (7.273)
After attach the thin film onto the substrate, the thin film will be stretched to
the position
xim = mas = Xm + umis

m , m = 0, ±1, ±2, · · · . (7.274)
where xrm is denoted as the reference position of the m-th atom with respect to
the aubstrate, and umis
m is the displacement of the atom due to the lattice misfit,
umis
m = m(as − af ) .
During actual deformation, the spatial position the m-th atom is
xm = Xm + umis e
m + um (7.275)
or
um = xm − xm = umis e
m + um (7.276)
where uem is the elastic deformation of the atom.
The relative displacement between the two atoms is now
um+1 − um = (uem+1 − uem ) − (af − as ) . (7.277)

The total potential energy of the system is
2πuem

1X e e 2
Π= µ(um+1 − um − (af − as )) + W [1 − cos ] (7.278)
2 m a
Let,
uem af − as
ξm = , and f = . (7.279)
as as
Hence
1 X 2
Π= µa (ξm+1 − ξm − f )2 + W [1 − cos(2πζm )] (7.280)
2 m
The equilibrium equation is derived from the stationary condition

dE
= 0, n = 0, ±1, ±2, · · · ⇒
dξn
−µa2 (ξn+1 − ξn + f ) + µa2 (ξn − ξn−1 + f ) + W π sin 2πξn = 0,
(7.281)
i.e.
π
∆2n ξ = (ξn+1 − 2ξn + ξn−1 ) = sin 2πξn (7.282)
2`20
p
where `0 = µa2 /2W .
The dynamics version of Eq. (7.282) is the finite-difference sine-Gordon
equation,
mn d2 ξn π
∆2n ξ − = 2 sin 2πξn (7.283)
µ dt2 2`0
If `0 >> 1, one may use continuous approximation to replace the finite
difference equation with a differential equation,
d2 ξn 2 2 d4 ξ 4 d2 ξ
∆2n ξ = a + a + O(a6
) = + O(a4f ) (7.284)
dXn2 f 4! dXn4 f f
dn2
Therefore, if we only consider static deformation, we have the following non-
linear ordinary differential equation
d2 ξ π
2
= 2 sin 2πξ . (7.285)
dn 2`0
Consider the following boundary conditions,

dξ
= , and ξ =0. (7.286)
dn n=n0 n=n0
One can integrate (7.282),

dξ 2 1
− 2 = (1 − cos 2πξ) , (7.287)
dn 2`20
which can be re-arranged as
dξ 2 (1 + `20 2 cos2 πξ
= 1 − (7.288)
dn `20 1 + `20 2
Change variable
π
φ = πζ − and k = (1 + `20 2 )−1/2 . (7.289)
2
One may transfer into the standard form of differential equations that can be
solved by using elliptic functions and integrals,
dφ π
=± (1 − k 2 sin2 φ)1/2 (7.290)
dn `0 k
Solutions of FKV model:
1. Consider boundary condition
= 0, and k = 1. (7.291)
In this case, Eq. (7.288) is simplified to

dξ 1
= sin πξ (7.292)
dn `0
Assume at n = 0, ξ(0) = 0.5, and then
π n
Z Z ξ
dζ
dp = π (7.293)
`0 0 0 sin πζ
which yields the solution

πn πξ
= ln tan (7.294)
`0 2
Or inversely,
2 h πn i
ξ= tan−1 exp (7.295)
π `0
This solution represents a single dislocation far away from the remote bound-
ary. We plot the positive solution in Fig. 7.18. One may find that at ξ = 1/2,
dξ 1
= (7.296)
dn `0
Figure 7.18. A single dislocation solution of FKV model
Since a unit change of ξ means a relative displacement of one lattice spacing as ,

it then implies that in a region of length `0 number of troughs is one more than
the number of atoms, i.e. there is extra plane of atoms in the substrate, which
forms a edge dislocation. We call `0 as the effective length of the dislocation
region.
2. General solution
The general static solution of sine-Gordon equation can be expressed by
elliptic function,
π Z φ
= (1 − k 2 sin2 ψ)−1/2 dψ = F (φ, k) (7.297)
`0 k 0
where the upper limit φ is called the amplitude. The inverse relation of the
above elliptic function is πn
φ = am (7.298)
`0 k
or
1 1 πn
ξ = + am (7.299)
2 π `0 k
and
dξ 1 πn 1
= dn = (1 − k 2 cos2 πξ)1/2 (7.300)
dn `0 k `0 k `0 k
At ξ = ξ(0) = 1/2,
dξ 1
= (7.301)
dn `0 k
i.e. `0 k is now the effective dislocation length.
dξ
Figure 7.19. The general solution of static sine-Gordon equation ( dn ≥ 0).
Assume that ξ(p) = 1.5. The general solution of FKV model is depicted on
Fig. 7.19. Obviously, the number is the atoms per dislocation,
2`0 kE(k)
p= (7.302)
π
where E(k) is the following elliptic integral,
Z π/2
E(k) = (1 − k 2 sin2 ψ)1/2 dψ (7.303)
0
The general solution indicates that there are many dislocation occuring simu-
tanelously along the chain in periodic fashion. In Fig. 7.20, we show the
dislocation pattern created by the general solution.
It would be interesting to examin the stability of Frenkel-Kontorova system.
The potential energy of one dislocation
p−1 p−1
X W X
Π = W `20
(ξn+1 − ξn − f ) + 2
1 − cos 2πξn )
2
n=0 n=0
Z P Z P
2 dξ 2
= W `0 − f dn + W sin2 πξdn (7.304)
0 dn 0
Consider
dξ 1
= 2 2 (1 − k 2 cos2 πξ)1/2 (7.305)
dn `0 k
Figure 7.20. Dislocation pattern for p = 3.
dξ
Figure 7.21. Dislocation pattern for dn
≤ 0.
One can write the potential energy per dislocation as

2(1 − k 2 )K(k)

2 4E(k) 2
Π = W `0 − − 2f + pf (7.306)
πk`0 πk`0
where Z π/2
K(k) = (1 − k 2 sin2 ψ)−1/2 dψ
0
One may find that the potential energy consists of contribution from both lattice
misfit and dislocation misfit.
To examine the stability, let,
∂Π
= W `20 (2 − 2pf ) = 0 . (7.307)
∂f
We find the critical lattice misfit,
1 π
fcr = = (7.308)
p 2`0 kK(k)
Figure 7.22. Matthews & Blackeslee Model
When k = 1,
1 2 W 1/2
fcr = = (7.309)
p π µa2 /2
It is beleived that when lattice misfit f > fcr , dislocations will spontaneous-
lly enter or depart from the monolayer chain.
7.8.2 Matthews & Blackesless’s equilibrium theory

In 1974, Matthews and Blackeslee proposed their equilibrium theory of dis-
location relaxation mechanism for thin film growth. It was an immediate suc-
cess, and it was soon received widespread attentions. Today, the Matthews
theory has become the foundamental theory for epitaxial thin film growth in
semi-conductor industry, and it is now viewed an early and integrated part of
nano-mechanics.
In the following, we outlined a simple version of the Matthews theory based
on Nix’s presentation.
Assume that the thin film is under homogeneous bi-axial palne stress load,
E
i.e. in the film, x = y = and σx = σy = . The homogeneous misfit
1−ν
strain is due to the lattice misfit, i.e.
as − af as − af
= or = . (7.310)
af as
The deformation of the substrate may be neglected. For a coherent thin film-
substrate system, the strain energy per unit thin film area is (see Fig. 7.22)
2µ(1 + ν) 2
E= h = M 2 h . (7.311)
(1 − ν)
When the lattice misfit increases, it is energetically favorable to have dislo-
cations present to relaxe the lattice misfit strain.
Consider a simplist sceenario that there is periodically distributed edge dis-

locations distributed along the interface between the thin film and the substrate.
The homogeneous distributed lattice misfit strain will be reduced to f − b/S
where S is the spacing between two edges dislocations. Then the elastic energy
due to homogeneous deformation is
b 2
Eh = M − h (7.312)
S
Since there are two edge dislocations in an area S × 1, the strain energy due to
dislocation is
µb2 βh 2
Ed = ln (7.313)
4π(1 − ν) b S
The total energy is the summation of Eh and Ed ,
b 2 µb2 βh 2
E =M − h+ ln (7.314)
S 4π(1 − ν) b S
The two competing effects will yield an equilibrium point at the bottom of
energy well as shown in Fig. 7.24. We are seeking to find an equilibrium
state that is defect-free, i.e. we are intereted in an equilibrium state at which

b/S = 0.
Consider the stationary condition,
∂E b µb2 βh
= −2M h − b + ln =0. (7.315)
∂ S1 S 2π(1 − ν) b h=hcr
We can find a critical thickness, hcr , of the thin film below which the thin film
will stay in a coherent state with the substrate that is the thin film is defect-free.
From (7.315), one can find that the critical thickness can be determined from
the following non-linear equation,
hcr µb
βh = (7.316)
cr 4π(1 − ν)M
ln
b
Exercise
Probelm 7.1 Consider cuboidal region of inelastic strain (eigenstrain) due
to solute segregation forming cuboidal precipitates. The precipitate subdomain
(or inclusion) has the dimension 2a × 2a × 2a, and the unit cell (U) has the
dimension 2L × 2 : ×2L. The eigenstrain is assumed to have a constant value
within each inclusion, and be zero outside the inclusion,

δij ε, ∀ x ∈ Ω;
ε∗ij = (7.317)
0; ∀ x ∈ U/Ω
where
n o
U =x − L ≤ xi ≤ L, i = 1, 2, 3 (7.318)
n o
Ω = x − a ≤ xi ≤ a, i = 1, 2, 3 , and a < L (7.319)
Find the disturbed displacement field u1 (x). (Hint:Mura pages: 20-21).

Figure 7.25. Distribution of periodic precipitates

Chapter 8
COMPARISON VARIATIONAL PRINCIPLES
8.1 Review of Variational Calculus

Consider a functional, which is a map,
I[y] : H 1 ([x0 , x1 ]) → IR (8.1)
where I[y] is the following integral a map

Z x1 h i
0
I[y] = p(x)(y )2 + q(x)y 2 + 2f (x)y dx (8.2)
x0
with prescribed boundary conditions,
y(x0 ) = y0 , y(x1 ) = y1 (8.3)
Assume that p(x), q(x), and f (x) are given continuous functions, i.e. p(x), q(x),
and f (x) ∈ C 0 [x0 , x1 ], and p(x) > 0, q(x) > 0. Let,
ỹ(x) = y(x) + αη(x) (8.4)
as a function that is very close to function, y(x).

◦
We require that y(x) ∈ V and η(x) ∈ V, and
n o
V := y(x) y ∈ H 1 ([x0 , x1 ]), y(x0 ) = y0 and y(x1 ) = y1 (8.5)
◦ n o
V := eta(x) η ∈ H 1 ([x0 , x1 ]), η(x0 ) = 0 and η(x1 ) = 0 (8.6)
We usually call y as the trial function and αη(x) as the test function.
Comparison Variational Principles 189
In order to find the function y(x) that yields the extreme value of I[y], we
consider the value of I[ỹ],
Z x1 n
0 0
I[y(x) + αη(x)] = p(x)[y (x) + αη (x)]2 + q(x)[y(x) + αη(x)]2
x0
+2f (x)[y(x) + αη(x)]} dx
Z x1 h i
0
= p(x)(y (x))2 + q(x)y 2 (x) + 2f (x)y(x) dx
x0
Z x1 h i
0 0
+ 2α p(x)y (x)η (x) + q(x)y(x)η(x) + f (x)η(x) dx
x0
Z x1
2 0
+α p(x)(η (x))2 + q(x)η 2 (x) dx (8.7)
x0
Thereby,
α2 2
∆I = I[y(x) + αη(x)] − I[y(x)] = αδI + δ I (8.8)
2!
where
Z x1
0 0
δI = 2 [p(x)y (x)η (x) + q(x)y(x)η(x) + f (x)η(x)]dx (8.9)
Zx0x1 h i
0
δ2I = 2 p(x)(η (x))2 + q(x)η 2 (x) dx (8.10)
x0
We say that
I[y] is stationary at y = y(x) if δI = 0. Since both p(x), q(x) > 0
y=y(x)
and δ 2 I > 0, I[y] will reach a minimum at y = y(x).
The first order variation illustrated above is in the sense of Gateaux. The
definition of the Gateaux variation is in terms of the so-called Gateaux deriva-
tive
I(y + αη) − I(y) d
δG I = DG Iη = lim = I(y + αη) (8.11)
α→0 α dα α=0
Remark 8.1.1 One may compare this with the so-called Fr«echet derivative,
DF I[y]η, which is defined as a linear functional such that
I(y + η) − I(y) − DF I(y) · η
⇒ 0, as kηkV → 0 . (8.12)
kηkV
Gateaus derivative coincides with Fre’chet derivative, if δF I is linear in η and
uniformly continuous in η, i.e. |δI(y, η)−δI(y0 , η)| → 0, as y → y0 uniformly
∀y ∈ B(y0 ).
In general, the n-th order Gateaux variation is defined as
n dn
δG I= I(y + αη) , ∀n ≥ 1 (8.13)
dαn α=0
such that
α2 2 α3 3 α4 4
∆I = I(y + αη) − I(y) = αδG I + δG I + δG I + δG I + · · · (8.14)
2! 3! 4!
In the rest of the book, we omit the subscript G in variation operator. Let
α = 1. We have
1 2 1 1
∆I = I(y + η) − I(y) = δI + δ I + δ3I + δ4I + · · · (8.15)
2! 3! 4!
One nice thing about the Gateaux variation is that it is defined based on a
scaler differentiation operation. In other words, the variation operation follows
the same rule as the differentiation operation in elementary calculus.
This can be seen by examining the first order variation of I[y],
Z x1 h i
0 0
δI = 2 p(x)y (x)η (x) + q(x)y(x)η(x) + f (x)η(x) dx (8.16)
x0
Let η(x) = δy. The Gateaux variation becomes,

Z x1 h i
0 0
δI = 2 p(x)y (x)δy + q(x)y(x)δy + f (x)δy dx
x
Z0 x1 h i
0 2 2
= δ p(x)(y (x)) + q(x)(y(x)) + f (x)y(x) dx
x0
= δI .
This is to say that one can find the first variation of a functional, I[y], by
simply differentiating (taking G-derivative) the unknown function according
to the same rule of differentiation in calculus. The only difference is: dy is
replaced by δy, which is the variation of the unknown function, or in general,
◦
a test function satisfying homogeneous boundary conditions, i.e. δy ∈ V.
Consider the first term in (8.16). Integration by parts yields,
Z x1 Z x1 Z x1
0 0 0 x1 0 0 0
p(x)y η dx = [p(x)y η]x0 − (p(x)y )ηdx = − (p(x)y ) ηdx
x0 x0 x0
Therefore,
Z x1 h i
0 0
δI = 2 −[p(x)y (x)] + q(x)y(x) + f (x) η(x)dx = 0 (8.17)
x0
◦
Since this equation must holds for any η(x) ∈ V, the integrand must vanish,
i.e. the solution of the following differential equation
0 0
−[p(x)y (x)] +q(x)y(x)+f (x) = 0, y(x0 ) = y0 and y(x1 ) = y1 . (8.18)
is a minimizer of the functional I[y]. Eq. (8.18) is called the Euler-Lagrange

equation.
Note that the solution of (8.18), y ∗ (x) may not be the only minimizer of
the functional I[y]. In fact, y ∗ ∈ C 1 ([x0 , x1 ]), and hence Eq. (8.18) is called
strong form of the Euler-Lagrange equation. On the other hand, a necessary
minimizer only requires that y ∈ H 1 ([x0 , x1 ]), since
Z x1 h i
0
I=2 p(x)(y (x))2 + q(x)y(x)2 η(x) + f (x)y(x) dx (8.19)
x0
and for this purpose we call a function that makes I[y] stationary, but not nec-
essarily satisfy the Euler-Lagrange equation, i.e.,
Z x1 h i
0 0
δI = 2 p(x)y (x)δy + q(x)y(x)δy + f (x)δy dx (8.20)
x0
as the weak solution, since C 1 ([x0 , x1 ]) ⊂ H 1 ([x0 , x1 ]).

In general, consider a functional of the following form,
Z x1
0
I[y] = F (x, y, y )dx, y(x0 ) = y0 and y(x1 ) = y1 . (8.21)
x0
Its first variation is

Z x1
∂F ∂F 0
δI = δy + 0 δy dx
x0 ∂y ∂y
Integration by parts yields

Z x1 h Z x1
∂F i ∂F x1 ∂ h ∂F i
δI = δy dx + 0 δy − 0 δydx
x0 ∂y ∂y x0 x0 ∂x ∂y
Z x1 h
∂F ∂ ∂F i
= − δydx (8.22)
x0 ∂y ∂x ∂y 0
One obtains the Euler-Lagrange equation,
∂F ∂ ∂F
E[F ]y = − =0. (8.23)
∂y ∂x ∂y 0
8.2 Extreme variational principles in linear elasticity

8.2.1 Minimum potential enery principle
Consider a linear elastic solid, V . The total potential energy of the elastic
solid is
Z Z Z
1
Π(ui , ui,j ) = σij ij dV − fi ui dV − t0i ui dS
2 V V Γt
Z Z Z
1
= Cijk` ui,j uk,` dV − fi ui dV − t0i ui dS
2 V V Γt
The solid is subjected to the following boundary conditions,
ui = u0i = xj 0ij , ∀x ∈ Γu (8.24)

ti = nj σij = t0i = 0
nj σij , ∀x ∈ Γt (8.25)
where the displacement boundary conditions are essential boundary conditions

for ensuing variational principles, because they are the constraints on primary
variables ui and the space of the trial function. Consider trial function ui ∈ V,
n o
V := yi (x) yi (x) ∈ H 1 (V ), and yi = xj 0ij ∀x ∈ Γu (8.26)
◦
and test function δui ∈ V where,
◦ n o
V := ηi (x) ηi (x) ∈ H 1 (V ), and ηi (x) = 0, ∀x ∈ Γu (8.27)
which is equivalent to δui ∈ Hc1 (V ). When ui (x) ∈ V, we say ui (x) is

kinematically addmissible.
A necessary condition that Π(ui , ui,j ) reaches to an extreme is the stationary
condition of its first variation, i.e.
Z Z Z
δΠ[ui , ui,j ] = Cijk` ui,j δuk,` dV − fi δui dV − t0i ni dS = 0 (8.28)
V V Γt
which is often called virtual displacement principle in solid mechanics. By

the way, the stationary condition in mechanics terms is equilibrium condition.
Any y satisfies virtual displacement principle is an equilibrium solution.
On the other hand, Eq.(8.28) is called as the weak formulation of Navier
equations in computational mechanics. This can be easily seen via integration
by parts,
Z Z Z
δΠ = σij δui,j dV − fi δui dV − t0i δui dS
ZV V
ZΓt Z
= (σij δui ),j − σij,j δui dV − fi δui dV − t0i δui
V V Γt
Z Z Z
= σij nj δui dS − (σij,j + fi )δui dV − t0i δui dS
Z∂V V
Z Γt
Z
0
= (σij nj − ti )δui dS − (σij,j + fi )δui dV + σij nj δui dS
Γt V Γu
which yields the Navier equation

Cijk` uk,`j + fi = 0, (8.29)
and the natural boundary conditions,
σij nj = t0i = σij
0
n j , ∀ x ∈ Γt . (8.30)
Examine the perturbance of the potential energy ∆Π(ui , ui,j ) around an
equilibrium configuration,
∆Π = Π(ui + δui , ui,j + δui,j ) − Π(ui , ui,j )
Z Z
1
= Cijk` (ui,j + δui,j )(uk,` + δuk,` )dV − fi (ui + δui )dV
2 V V
Z
− t0i (ui + δui )dV
Γt
Z Z Z
1
− Cijk` ui,j uk,` dV − fi ui dV − t0i ui dV
2 V
Z Z V Z Γt
= Cijk` ui,j δuk,` dV − fi δui dV − t0i δui dV
V V Γt
Z
1
+ Cijk` δui,j δuk,` dV
2 V
1
= δΠ + δ 2 Π (8.31)
2!
For the equilibrium solution δΠ = 0, ∆Π = 2!1 δ 2 Π > 0.
This means that for all the kinematically admissible vector fields, u = ui ei ,
ui (x) ∈ V the equilibrium solution (real solution ? is the solution unique
? weak solution = strong solution) is the minimizer of total potential energy
Π(ui , ui,j ).
Theorem 8.1 (Minimum potential energy principle) Among all (in-
finitesimal) kinematically admissible displacement fields, that which is also
statically admissible (real solution) render the potential energy Π an absolute

minimum.
That is Π(ũ, ∇ · ũ) ≤ Π(u, ∇ · u) ∀u ∈ V. Or
Π(ũ, ∇ · ũ) = inf Π(u, ∇u) (8.32)

u∈V
If macros strain boundary condition is applied on entire boundary ∂V ,
u = x · 0 , x ∈ ∂V (8.33)
Then Γt = ∅ and Π(u, ∇ · u) = V W (∇ · u), where

Z
1
W (∇u) := Cijk` ij k` dV (8.34)
2V V
The minimum potential energy principle reads as
W (˜) = inf W () (8.35)

u∈V
For the real solution, ũ,

Z
1 1
W (˜) = σ : dV = < σ̃ >:< ˜ >
2V V 2
1 1 0
= < σ̃ >: 0 = : C̄ : 0
2 2
On the other hand,
Z
1 1
W () = < σ >:< ˜ >
σ : dV =
2V V 2
n
1 0 1 0 X α
= < σ̃ >: = : C :< >α
2 2
α=0
Since 0 ∈ V, we can choose α = 0 . Then we have

n
1 0 0 1 0 X α 0
: C̄ : ≤ : C :
2 2
α=0
which then leads to

n
X
C̄ ≤ fα Cα . (8.36)
α=0
8.2.2 Minimum complementary potential energy principle

Consider the following complementray potential energy,
Z Z
1
Πc (σij ) = Dijk` σij σk` dV − u0i σij nj dS (8.37)
2 V Γu
which is a map,
Πc : S → IR (8.38)
where S is the trial function space
n o
S = σij σij ∈ H 1 (V ), σij,j = 0 and nj σij = t0i , ∀x ∈ Γt (8.39)
and the test function space is

◦ n o
S = σij σij ∈ H 1 (V ), σij,j = 0 and nj σij = 0, ∀x ∈ Γt (8.40)
Note that in this variational statement, the essential boundary condition be-
comes
nj σij = t0i , ∀x ∈ Γt (8.41)
whereas the natural boundary condition becomes
ui = ūi , ∀x ∈ Γu . (8.42)
To study extreme property, we examine complementary potential energy

perturbance,
∆Πc = Πc (σij + δσij ) − Πc (σij )

h1 Z Z i
= Dijk` (σij + δσij )(σk` + δσk` )dV − u0i (σij + δσij )nj dS
2 V Γu
h1 Z Z i
− Dijk` σij σk` dV − u0i σij dS
2 V Γu
Z Z
= Dijk` σij δσk` dV − u0i δσij nj dS
V Γu
| {z }
=δΠc
Z
1
+ Dijk` δσij δσk` dV
2 V
| {z }
=δ 2 Πc
The necessary condition for Πc (σij ) attaining extreme value is the stat-
tionary condition,
δΠc = 0 .
Hence
1 2 c
∆Πc = δ Π >0 (8.43)
2!
since Dijk` is positive definite. Thus, Πc (σij ) reaches a minimum value at
σij = σ̃ij , where σ̃ij renders stationary condition δΠc (σ̃ij ) = 0. This fact is
the so-called minimum complementray potential energy principle.
Theorem 8.2 (Minimum Complementary Energy Principle) Among
all statically admissible stress fields, the actual stress field (whose correpond-
ing strain field satisfies compatibility condition) rensers Πc an absolute mini-
mum, i.e.
Πc (σ̃) ≤ Πc (σ), ∀σ ∈ S (8.44)
or
Πc (σ̃) = inf Πc (σ) (8.45)
σ ∈S
The stationary condition of complementary energy has well-known names,
e.g. virtual force principle in continuum mechanics, or the weak form of com-
patibility condition in computational mechanics,
Z Z
c
δΠ (σ̃ij ) = Dijk` σ̃ij δσk` dV − u0i δσij nj dS = 0 (8.46)
V Γu
The above equation can be rewriten as

Z Z
1
˜ij δσij dV − ui,j δσij + uj,i δσij dV
VZ 2 V
Z
+ ui,j δσij dV − u0i δσij nj dS = 0
V Γu

Z Z
1
˜ij − (ui,j + uj,i )δσij dV + ui δσij nj dS
V 2 ∂V
Z Z
− ui δσij,j dV − u0i δσij nj dS = 0
V Γu
| {z }
=0
Z Z
1
⇒ ˜ij − (ui,j + uj,i ) δσij dV + (ui − u0i )δσij nj dS = 0 .
V 2 Γu
which leads to the Euler-Lagrange equation,

1
˜ij
= Dijk` σ̃k` = (ui,j + uj,i ) (8.47)
2
⇒ ˜ij,k` + ˜k`,ij − ˜ik,j` − ˜j`,ik = 0 . (8.48)
and the natural boundary condition
ui = u0i , ∀x ∈ Γu (8.49)
Consider prescribed macro-stress boundary condition, n·σ = t0 , ∀x ∈ ∂V ,

Γu = ∅. In this case, Γu = ∅. Therefore,
Z
c 1
Π = Dijk` σij σk` dV = Wc (σ)V (8.50)
2 V
where Z
1
Wc := Dijk` σij σk` dV (8.51)
2V V
is the complementary energy density.
The minimum complementary potential energy principle then gives
Wc (σ̃) = inf Wc (σ) (8.52)

σ ∈S
Recall,
Z
1
< σ : > − < σ >:< >= u−x· < ∇⊗u > n·(σ− < σ >) dS
V ∂V
The real complementary energy density becomes

1 1
Wc (σ̃) = < σ̃ : ˜ >= < σ̃ >:< ˜ >
2 2
1 0 1
= σ : D̄ : σ 0 = < σ̃ >: D̄ :< σ̃ > (8.53)
2 2
Note that under prescribed remote stress boundary condition,
< σ >= σ 0 , ∀ σ ∈ S .
Choose σ = σ 0 ∈ S,
Z Z
1 1
Wc (σ) = σ : dV = σ 0 : dV
2V V 2V V
n
Z X
1
= σ0 : Dα : σ α dV
2V V α=0
n
1 0 X Ωα α
= σ : D : σ0
2 V
α=0
n
1 0 X
= σ : fα Dα : σ 0
2
α=0
Therefore,
n
X
σ 0 : D̄ : σ 0 < σ 0 : fα Dα : σ 0 (8.54)
α=0
Since D̄ : C̄ = 1(4s) and both D̄ and C̄ are positive definite, we then have
n
X −1
fα C−1
α ≤ C̄ (8.55)
α=0
which is called the Reuss bound. It is a lower bound for elastic moduli.
Assume that
Cα = 3K α E(1) + 2µα E(2)

1 1
Cα−1 = α
E(1) + α E(2)
3K 2µ
One can derive that
n
X n
X n
X
α (1)
fα C = 3 fα Kα E +2 fα µα E(2)
α=0 α=0 α=0
n −1
X 3 2
fα Cα−1 = n E(1) + n E(2)
α=0
X fα X fα
Kα µα
α=0 α=0
Combining Reuss bound with the Voigt bound, we have

n
X n
X
fα Cα−1 < C̄ < fα Cα
α=0 α=0
and consequently,
n
1 X
n < K̄ < fα Kα
X fα
α=0
Kα
α=0
n
1 X
n < µ̄ < fα µα
X fα
α=0
µα
α=0
One can see that the Voigt bound is in fact an arithmetic average and the Reuss
bound can be viewed as a geometric average or the harmonic average.
8.3 Hashin-Shtrikman variational principles

In order to narrow the gap between the Voigt bound and the Reuss bound,
we need new mathematical tools. One of powerful such tools is the celebrated
Hashin-Shtrikman (HS) variational principle. The essence of the HS varia-
tional principles is that they are the variational principles specifically designed
for composites, or inhomogeneous solids. To measure the differences between
homogeneous solids and inhomogeneous solids, a comparison homogenous
solid is used to identify the inhomogeneous fields.
Let’s first consider a boundary value problem of the original composite
(RVE),
σij,j = 0,
σij = Cijk` (x)k` ,
1
U () = Cijk` ij k` , and W () =V
2
ui = ūi , ∀x ∈ Γu , (Γt = ∅, Γu = ∂V ).
Consider a second BVP in a comparison solid,

(0)
σij,j = 0,
(0) (0) (0)
σij = Cijk` (x)k` ,
1 (0) (0) (0)
U (0) ((0) ) = C , and W0 ((0) ) =V
2 ijk` ij k`
(0)
ui = ūi , ∀x ∈ Γu , (Γt = ∅, Γu = ∂V ).
To relate the two BVPs, we introduce the following decomposition in strain

field and stress field,
(0)
ui = ui + udi (8.56)
(0)
ij = ij + dij (8.57)
and
(0)
σij = pij + Cijk` k`
(0) (0)
= pij + Cijk` (ijk` + dijk` ) (8.58)
where udi is the disturbance displacement field and pij is called polarization
stress.
A better definition of stress polarization is
(0) (0)
pij = σij − Cijk` k` = (Cijk` − Cijk` )k` (8.59)
which indicates that stress polarization is due to inhomogeneouness of the

composite.
Furthermore, since
(0)
ui = ūi , ∀x ∈ ∂V and ui = ūi , ∀x ∈ ∂V
it leads to homogeneous boundary conditions for displacement disturbance
field
udi = 0, ∀x ∈ ∂V (8.60)
In passing, we note that because udi = 0, ∀x ∈ ∂V it can be readily to show
that the average work done by the disturbance field over any self-equilibrium
stress field will be zero, that is
Z Z
d
σij ij dV = σij udi,j dV
V V
Z Z
d
= ui nj σij dS + udi σij,j dV = 0 . (8.61)
∂V V
On the other hand, since

(0)
σij,j = 0, σij,j = 0 ,
one has
(0) (0)
σij,j = σij,j + pij,j + Cijk` dk` =0
,j
We can see that the stress field can be divided into the homogeneous (or com-
(0)
parison) stress field, σij , and the inhomogeneous stress field,
0 0
σij = σij + tij , where tij = pij + Cijk` dk` (8.62)
0 , and inhomogeneous stress field, t satisfy
Both homogeneous stress field, σij ij
equilibrium equations, i.e.
(0)
σij,j = 0, tij,j = 0 . (8.63)
In literature, the inhomogeneous equilibrium equation

(0)
tij,j = Cijk` dk` + pij,j = 0 (8.64)
,j
is often called “the subsidiary condition.”

Theorem 8.3 (Hashin-Shtrikman) Let udi ∈ U and pij ∈ S where
n o
U = ui ui ∈ H 1 (V ), ui = 0, ∀x ∈ ∂V (8.65)
n o
S = σij σij ∈ L2 (V ) (8.66)
Consider the following functional,
Π : S × U → IR,
where
Z
1 (0) (0) (0) −1 (0)

Π(pij , dij )
= d
Cijk` ij k` − ∆Cijk` pij pk` + pij ij + 2pij ij dV
2
 V
(0)
 ∆Cijk` = Cijk` − Cijk`

where pij = ∆Cijk` k` (8.67)
 d
 (0)
ij = ij − ij
We have the following variational statements:

1. The functional Π is stationary, i.e. δΠ = 0, if the inhomogeneous equi-
librium equation (subsidiary condition) is satisfied,

(0)
Cijk` dk` + pij,j = 0 ; (8.68)
,j
2.
δ 2 Π > 0, if ∆C < 0, Π → M inimum (8.69)

δ 2 Π < 0, if ∆C > 0, Π → M aximum (8.70)
Proof:
∆Π = Π(pij + δpij , dij + δdij ) − Π(pij , dij )

Z
1 −1 (0)

= −2∆Cijk` pij δpk` + pij δdij + δpij dij + 2δpij ij dV
2 V
Z
1 −1
1
+ −∆Cijk` δpij δpk` + δpij δdk` dV = δΠ + δ 2 Π
2 V 2!
We first show that the first statement is true.
1 Z
−1 (0)
δΠ = − 2∆Cijk` pk` δpij − 2ij δpij − ij δpij − pij δdij dV
2 V
1 Z
−1 (d)
= − 2∆Cijk` pk` δpij − 2 (ij − ij ) δpij − dij δpij − pij δdij dV
2 V | {z }
(0)
=ij
1 Z
−1
= − 2 (∆Cijk` pk` − ij ) δpij + dij δpij − pij δdij dV
2 V | {z }
=0
1 Z
= − dij δpij − pij δdij dV (8.71)
2 V
If the subsidiary condition is satisfied, i.e.

(0)
Cijk` dk` + pij,j = 0, or tij,j = 0 . (8.72)
,j
which leads to
(0)
δtij = δpij + Cijk` δdk` , and δtij,j = 0 . (8.73)
Substituting (8.72) and (8.73) into (8.71) yields
1 Z
(0) (0)
δΠ = − dij (δtij − Cijk` δdk` ) − δdij (tij − Cijk` dk` ) dV
2 V
 
1 Z
 d
(0) (0)
 ij δtij − tij δdij − dij Cijk` δdk` + δdij Cijk` dk`

= −  dV
2 V  | {z } 
=0, because C(0) has major symmetry
1 Z
= − udi,j δtij − tij δudi,j dV
2 V
Considering the facts
Z Z Z
δtij udi,j dV = δtij nj udi dS − δtij,j udi dV ≡ 0
V
Z Z∂V ZV
tij δudi,j dV = tij nj δudi dS − tij,j δudi dV ≡ 0,
V ∂V V
we just proved that δΠ = 0, if tij,j = 0 holds.

(0)
Now we examin the extreme conditions. Substituting δpij = δtij −Cijk` δdk`
into the second order variation,
1 Z
2 −1
δ Π = −∆Cijk` δpij δpk` + δpij δdij dV
2 V
1 Z
−1 (0)
= − −∆Cijk` δpij δpk` + Cijk` δdij dk` − δtij dij dV
2 V
Agan, the last term Z
δtij dij dV = 0.
V
Therefore, we have
1 Z
2 −1 (0)
δ Π= − ∆Cijk` δpij δpk` + Cijk` δdij dk` dV (8.74)
2 V
Obviously if ∆C > 0, ∆Π = δ 2 Π < 0, therefore, Π achives a maximum
value.
On the other hand, if ∆C < 0, the judgement is not straightforward.

Consider a positive integral,
Z
(0) −1
I := Cijk` δpij δpk` dV > 0 (8.75)
V
(0)
Substitute δpij = δtij − Cijk` δdk` into (8.75). It can be readily shown that
Z
(0) −1

(0)
I = Cijk` δtij δtk` − 2δtij δdk` +Cijk` δdij δdk` dV
V | {z }
=0
Z
(0) −1

(0)
= Cijk` δtij δtk` + Cijk` δdij δdk` dV
V
A direct consequency is
Z Z
(0) −1 (0)
Cijk` δpij δpk` dV > Cijk` δdij dk` dV (8.76)
V V
which leads the following inequality,

1 Z
−1 (0)
δΠ = − ∆Cijk` δpij δpk` + Cijk` δdij dk` dV
2 V
1 Z
(0) −1

−1
≥ − ∆Cijk` + Cijk` δpij δpk` dV
2 V
Consider
−1 −1
∆C−1 + C(0) = ∆C−1 + C(0) : (C − C(0) ) : (C − C(0) )−1
−1
= ∆C−1 + C(0) : C : ∆C−1 − ∆C−1
−1
= C(0) : C : ∆C−1 .
One can write that

1 Z −1
2
δ Π> − p : C(0) : C : ∆C−1 : pdV (8.77)
2 V
It is clear now that if ∆C−1 < 0, δ 2 Π > 0 and hence Π has a global minimum.
To sum up, we have the following extreme conditions,
δ 2 Π < 0, if ∆C > 0, Π → maximum ;

δ 2 Π > 0, if ∆C < 0, Π → minimum .
♣
0 are self-equlibrium stress field,

Since both σij and σij
Z Z
σij dij dV = σij udi,j dV = 0
Z V ZV
(0) d (0)
σij ij dV = σij udi,j dV = 0
V V
because udi
= 0, ∀x ∈ ∂V .
Therefore the total potential energy of a kinematically admissible field, ui ∈
V, can be written as
Z Z
1 1 (0)

Π() = σij ij dV = σij ij − σij dij dV
2 V 2 V | {z }
=0
Z
1 (0)
= σij ij dV
2 V
Consider

(0) (0) (0) (0)
σij ij = σij + pij + Cijk` dk` ij
(0 (0) (0) (0) (0) (0) (0)
= σij ij + pij ij + Cijk` dk` ij + + pij ij − pij ij
| {z }
=0
(0) (0) (0) d (0) (0)
= Cijk` k` ij + Cijk` k` ij + +2pij ij − pij (ij − dij )
(0) (0) (0) (0) (0) (0)
= Cijk` k` ij + Cijk` ij dk` +2pij ij − pij ij + pij dij
| {z }
=0
Therefore under prescribed remote strain boundary condition,

Z
1
Π() = σij ij dV = W ()V
2 V
Z
1 (0) (0) (0) −1 (0)

= Cijk` ij k` − ∆Cijk` pij pk` + pij dij + 2pij ij dV
2 V
Z
(0) (0) 1 −1 (0)

= W ( )V + −∆Cijk` pij pk` + pij dij + 2pij ij dV
2 V
= W (0) ((0) )V + Rπ V
1
R −1

d + 2p (0) dV .
where Rπ := 2V V −∆C ijk` p ij p k` + p ij ij ij ij
Based on Hashin-Shtrikman principle, if ∆C > 0 Π has a global minimum,
W (0) ((0) ) + Rπ ; whereas if ∆C < 0, Π has a global maximum, W (0) ((0) ) +
R¯pi . Therefore, the Hashin-Shtrikman principle provides the following bound,
Rπ (p̃, ˜d ) ≤ W () − W (0) ((0) ) ≤ R¯π (p̃, ˜d ) (8.78)

8.4 Review of Functional Analysis and Convex Analysis

Definition 8.4 (Vector Space (Linear Space)) Let F be a field, whose
elements are referred to as scalars. A vector space over F is a nonempty set
V, whose elements are referred to as vectors, together with two operations.
The first operation, called addition and denoted by +, assignes to each pair
(u, v) ∈ V × V of vectors in V a vector u + v in V. The second opera-
tion, called multiplication and denoted by juxtaposition, assigns to each pair
(r, u) ∈ F × V a vector rv ∈ V . Furthermore, the following properties must
be satisfied,
1 Associativity of addition
u + (v + w) = (u + v) + w, ∀u, v, w ∈ V
2 Commutivity of addition
u + v = v + u, ∀u, v ∈ V
3 Existence of a zero vector, 0 ∈ V such that
0 + u = u + 0 = u, ∀u ∈ V
4 Existence of additive inverse: i.e. ∀u ∈ V , ∃ − u ∈ V , such that
u + (−u) = (−u) + u = 0
5 Properties of scalar multiplication. ∀r, s ∈ F and u, v ∈ V ,
r(u + v) = ru + rv
(r + s)u = ru + rv
rsu = r(su)
1u = u
Remark 8.4.1 1 The first four properties in the definitions of vector space
can be summarized that V is an abelian group under addition;
2 Any expression of the form
r1 v1 + r2 v2 + · · · + rn vn
where ri ∈ F and vi ∈ V ∀i = 1, 2, · · · , n is called a linear combination

of the vectors v1 , v2 , · · · , vn , and
r1 v1 + r2 v2 + · · · + rn vn ∈ V
3 The addition operation
V × V → V : (u, v) → u + v ∈ V
4 and the scalar multiplication operation,
F × V → V : (α, u) → αu ∈ V
are closed.
5 When the operations
f : (u, v) → u + v ∈ V
g : (α, u) → αu ∈ V
are continuous, the vector space is called topological vector space.
Example 8.5 Let F = IR. The set of all ordered n-tuples, i.e.
u = (u1 , u2 , · · · , un ), ui ∈ IR
with addition and scalar miltiplication defined component-wise,
(a1 , · · · , an ) + (b1 , · · · , bn ) = (a1 + b1 , · · · , an + bn )
and
α(a1 , · · · , an ) = (αa1 , · · · , αan )
is a vector space, and it is denoted as IRn . Note that in general vector space
(a mathematical concept) is still a primitive set. It may have some algebraic
structures, but it does not have topologival structures, or geometric structures,
such as distance between two elements.
Example 8.6 Let F = IR. The set of all continuous function, C 0 (IR), i.e.
∀f ∈ C 0 (IR)
f : X ⊂ IR → Y ⊂ IR
and
dY (f (x), f (y)) < , ∀dX (x, y) < δ, ∀δ > 0 .
is a vector space under the operations of addition and scalar multiplication,
i.e.
(f + g)(x) = f (x) + g(x), f, g ∈ C 0 (IR)
and
αf (x) = αf (x), ∀α ∈ IR, f ∈ C 0 (IR)
Definition 8.7 (Bilinear form) Let X be a vector space and X ∗ is its

dual space. A mapping g of X × X ∗ into IR is called a bilinear functional or
a bilinear form if
1 For fixed y, g(x, y) is a linear functional in x, i.e.
g(αx + βy, z) = αg(x, z) + βg(x, z), ∀x, y ∈ X, z ∈ X ∗
2 For fixed x, g(x, y) is a linear functional in y, i.e.
g(x, αy + βz) = αg(x, y) + βg(x, z), ∀x ∈ X, y, z ∈ X ∗
A bilinear form is denoted as
g(x, y) :=< x, y >
Definition 8.8 (Inner product) Choose X ∗ = X. The bilinear form

of X × X is called inner product, denoting < ·, · > as (·, ·), such that
(·, ·) : X × X → IR
with properties:
1 (x, x) ≥ 0, ∀x ∈ X and (x, x) = 0 iff x = 0;
2 Symmetry (x, y) = (y, x);
3 Linearity
(αx + βy, z) = α(x, z) + β(y, z),
and
(x, αy + βz) = α(x, y) + β(x, z) ∀x, y, z ∈ X and αβ ∈ IR.
Example 8.9 Space E n . Let X = IRn . For x = (x1 , x2 , · · · , xn ) and

y = (y1 , y2 , · · · , yn ) ∈ IRn , we define an inner product
n
X
(x, y) = xi yi
i=1
This particular inner product space is denoted as En = {IRn , (·, ·)}. It gener-
ates a norm,
Xn 1/2 p
kxk`2 := xi xi = (x, x)
i=1
This norm is called Euclidean norm on IRn . The space is therefore a normed
space as well — called n-dimensional Euclidean space, En = {IRn , k · k`2 }.
One can show that
(i) kxk`2 ≥ 0, ∀x ∈ En
kxk`2 = 0, ⇐⇒ x = 0 ;
(ii) kαxk`2 = |α|kxk`2 , ∀xEn , α ∈ IR
(iii) kx + yk`2 ≤ kxk`2 + kyk`2 ← triangle inequality;
(iii) k(x, y)k`2 ≤ kxk`2 kyk`2 ← Cauchy − Schwartz inequality;
Based on the `2 -norm, one can measure the distance between two vectors in
En ,
ρ(x, y) := kx − yk`2 ;
One can also show that
(i) ρ(x, y) = ρ(y, x);
(ii) ρ(x, y) > 0, and ρ(x, y) = 0, iff x = y;
(iii) ρ(x, y) ≤ ρ(x, z) + ρ(z, y), ∀x, y, z ∈ En
The distance function ρ(x, y) is called a metric, and the associated vector
space is called metric space.
Figure 8.1. Banach space and Hilbert space
Remark 8.4.2 1 A normed space or a metric space is not necessarily an

inner product space, but an inner product vector space is a normed space,
becauce inner product can generate a norm, not vice versa.
2 A complete normed vector space is called Banach space and a complete
inner product space is called Hilber space.
Note that the term completeness means that: A metric space, V, is called
complete if every Cauchy sequence {fi } of V has a limit f ∈ V . For a metric
space, a Cauchy sequence is one such that kvj − vk k → 0, as j, k → ∞.
Example 8.10 (L2 Space) Consider a real value function f (x), x ∈ [a, b].
Define an inner product,
Z b
(f, g) = f (x)g(x)dx
a
We call the set that contains all f (x) such that

s
Z b
f 2 (x)dx < +∞
a
as space L2 ([a, b]), where L2 norm is defined as

s
p Z b
kf kL2 ([a,b]) = (f, f ) = f 2 (x)dx (8.79)
a
Therefore, L2 ([a, b]) is an inner product vector space, and of course, normed
space (metric space).
Example 8.11 (Lebesgue Space (Lp (Ω))) Let Ω be an open set in IRn .
For 1 < p < ∞, one can define a Lp -norm for a measurable function f ,
Z 1/p
kf kLp (Ω) := |f (x)|p dx
Ω
and a Lebesgue space is defined as

n o
Lp (Ω) := f kf kLp (Ω) < ∞
It has the following properties,
(i) kf kLp (Ω) ≥ 0, kf kLp (Ω) = 0, ⇒ f = 0 almost everywhere;

(ii) kcf kLp (Ω) ≤ |c|kf kLp (Ω) , ∀f ∈ Lp (Ω), c ∈ IR
(iii) kf + gkLp (Ω) ≤ kf kLp (Ω) + kgkLp (Ω) ← Minkowski0 s inequality
1 1
(iv) For 1 ≤ p, q ≤ ∞, such that + = 1,
p q
if f ∈ Lp (Ω) and gLq (Ω), then for finite Ω, f, g ∈ L1 (Ω), and
kf gkL1 (Ω ≤ kf kLp (Ω) kgkLq (Ω) , ← Holder0 s inequality
In particular, p = q = 2, then f · g ∈ L1 (Ω) because

Z
|f (x)g(x)|dx ≤ kf kL2 (Ω) kgkL2 (Ω)
Ω
Note that in general Lp (Ω) is not an inner product space, except p = 2.

Lp (Ω) is, nevertheless, a complete normed space, therefore, a Banach space
and L2 (Ω) is a Hilber space.
Example 8.12 (Sobolev Space) Define Soblev norm
k
X 1/p
kf kWpk (Ω) = kDα f kpLp (Ω)
α=0
Note that the Sobolev norm is not generated by an inner product in general.
A Sobolev space is defined as
Wpk (Ω) = {f kf kWpk (Ω) < ∞}
For p = 2, Sobolev spaces become inner product spaces. In particular,
1 For p = 2, k = 0, W20 (Ω) = L2 (Ω),

Z
(f, g)L2 (Ω) = f (x)g(x)dV
Ω
2 For p = 2, k = 1, W21 (Ω) = H 1 (Ω),

Z h i
(f, g)H 1 (Ω) = f (x)g(x) + ∇f (x) · ∇g(x) dV
Ω
and sZ
h i
kf kH 1 (Ω = f (x)2 + ∇f (x) · ∇f (x) dV
Ω
3 For p = 2, k = 2, W22 (Ω) = H 2 (Ω),

Z h i
(f, g)H 2 (Ω) = f (x)g(x)+∇f (x)·∇g(x)+∇⊗∇f (x) : ∇⊗∇g(x) dV
Ω
and
sZ
h i
kf kH 2 (Ω = f (x)2 + ∇f (x) · ∇f (x) + ∇ ⊗ ∇f (x) : ∇ ⊗ ∇f (x) dV
Ω
Figure 8.2. Convex set and non-convex set in IR2
8.4.1 Concept of convexity

Definition 8.13 Let U be a linear vector space over IR. A subset (sub-
space) K ⊂ U is said to be convex, if it contains the line segment between any
two of its elements, i.e.
θu + (1 − θ)v ∈ K, ∀u, v ∈ K
where θ ∈ [0, 1].
Example 8.14 Let U = IR × IR, and K ∈ U . We say K is convex, when
u1 = (x1 , x2 ), u2 = (y1 , y2 ) ∈ K, then θu1 + (1 − θ)u2 ∈ K, θ ∈ [0, 1].
We say K is not convex, for any u1 , u2 ∈ K, if ∃uθ ∈ θu1 + (1 − θ)u2 but
uθ 6∈ K. A graphic illustration is demonstrated in Fig. (8.2).
Definition 8.15 (Convex and concave functionals) 1 A functional
P : U → IR is said to be convex on U if
P (θu1 + (1 − θ)u2 ) ≤ θP (u1 ) + (1 − θ)P (u2 ), ∀u1 , u2 ∈ U, ∀θ ∈ [0, 1]
whenever the right-hand side is defined.
2 P is said to be strictly convex if the strict form of the inequality holds for
any u1 6= u2 ;
3 P is said to be concave if −P is convex.
Example 8.16 Let U = IR and P (x) = (x − a)2 .
Example 8.17 Consider a 1D elastic string, I = [0, `]. Let U = E and
U ∗ = S where
du
E{ ∈ Lα (I), = }
dx
dσ
S{σ σ ∈ Lβ (I), = 0}
dx
1 1
1 < α, β < ∞, and + = 1.
α β
Figure 8.3. An example of convex function.
Figure 8.4. Strain energy density and complementray strain energy
Define
Z
U : E → IR, U () =
σ(˜
)d˜
0
Z σ
c ∗ c
U : E = S → IR, U (σ) = (σ̃)dσ̃
0
Both strain energy density and complementary strain energy density are con-
vex, and they are plotted in Fig. (8.4).
8.4.2 G^
ateaux variation and convex functional
The Gâteaus variation of a functional in a linear space is the generalized
directional derivative of a real-value function in vector calculus.
Definition 8.18 (Gâteaux variation) 1 Let P : U → IR be a real-

valued functional and Ua ⊂ U a subspace. For a given ū ∈ Ua , if the
limit,
P (ū + λu) − P (ū)
δU (ū, u) := lim , ∀u ∈ Ua
λ→0 + λ
exists as λ → 0+ (i.e. λ → 0, λ > 0), then δP (ū; u) ∈ IR is called the
Gâteaus variation of P at ū in the direction of u.
2 If the Gâteau variation is a linear operator in u such that
δP (ū, u) =< u, DP (ū) >, ∀u ∈ Ua
we say that P is Gâteaux differentiable at ū. The linear operator DP (ū) :

Ua → U ∗ , which generally depends on ū, is called the Gâteaux derivative
of P at ū.
3 The functional P : U → IR is said to be Gâteaux differentiable on Ua if it
is Gâteaux differentiable at each u ∈ Ua .
Note that
d
δP (ū, u) = P (ū + λu)
dλ λ=0
δP
:= DP (ū)
δu
Question: why are convex functionals so special ? The following theorem
answers this question:
Theorem 8.19 If P : Uk ⊂ U → IR is Gâteaux differentiable, then, the
following statements are equivalent to each other
(S1) P : Uk ⊂ U → IR is convex;
(S2) P (v) − P (u) ≥< v − u, DP (u) >, ∀v, u ∈ Uk
(S3) < v − u, DP (v) − DP (u) >≥ 0 , ∀v, u ∈ Uk
Remark 8.4.3 The statement (S3) shows that Gâteaux derivative of a con-
vex function is a monotone operator of U into U ∗ . By the mean value theorem,
< v − u, DP (v) − DP (u) >=< v − u, D2 P (ū) · (v − u) >≥ 0
where ū = v + θ(v − u), θ ∈ [0, 1].

Hencea, a sufficient condition for P being convex on U is that
D2 P (u) ≥ 0, ∀u ∈ Uk
Recall the total potential energy for a linear elastic solid is

Z Z
Π(u, ∇u) = U ()dV − t0i ui dS
V Γt
Z Z
∂U
δΠ(u, ∇u) = δij dV − t0i δui dS
V ∂ ij Γt
Z
∂ 2U Z
2
δ Π(u, ∇u) = δij δk` dV = Cijk` δij δk` dV ≥ 0 .
V ∂ij ∂k` V
This is to say that if elastic tensor is positive definite, the elastic potential
energy is convex. Similar statement can be made for complementary potential
energy, if the compliance tensor is positive definite.
8.4.3 Primal variational problems

We consider the following primal variational problems:
Let P : Uκ ⊂ U → IR be a given functional.
1 The infimum (or inf) primal variational problems is to find a global mini-
mizer ũ ∈ Uκ such that

Pinf : P (ũ) = inf P (u), ∀u ∈ Uκ
2 The supremum (or sup) primal problem is to find a global maximizer ũ ∈

Uκ such that

Psup : P (ũ) = sup P (u), ∀u ∈ Uκ
3 The stationar (or sta) primal variational problem is to find a stationary point
u ∈ Uκ such that

Psta : P (ũ) = sta P (u), ∀u ∈ Uκ
Remark 8.4.4 1 A stationary point is also called critical point. The criti-
cal point condition,
δP (ũ, u) = 0, ∀u ∈ Uκ
leads to the Euler-Lagrange equation.

2 The problem (Pinf ) is called realisable if there exists a vector ũ ∈ Uκ such

that the infimum of P is achieved at ũ and is not +∞. Then ũ is called the
minimizer of (Pinf ) and we write P (ũ) = min P (u).
u∈Uκ
Similarly, a vector ũ ∈ Uκ is called the maximizer of (Psup ) if the super-
mum is achieved at ũ and is not +∞. We write P(ν̃) = maxu∈Uκ (u).
Example 8.20 The real-value function, P (x) = exp(x) is convex on U =
IR and
inf P (x) = 0, sup P (x) = +∞
x∈U
Howeverm on the closed interval, Uκ = [a, b] with −∞ < a < b < +∞, the
two inf − and sup− problmes are realisable and
inf P (x) = min P (x) = P (a) = ea ,
x∈Uκ x∈Uκ
sup P (x) = min P (x) = P (b) = eb .

x∈Uκ x∈Uκ
8.5 Legendre Transformation and Duality

In continuum mechanics, for a given stored-energy density U () such that
∂U
the strain-stress relation σ = is invertible, then one can define so-called
∂ c
complementary energy density of U (σ) by
U c (σ) = σ : (σ) − U ((σ)) (8.80)
Note that here
U = U () : E → IR (8.81)
U = U c (σ) :
c
S → IR (8.82)
< , σ >= σ : : E × E ∗ → IR (8.83)
where the space S may be viewed as E ∗ .
In mathematics, this is the well-known Legendre transformation. Generally
speaking, the classical Legendre transformation can be viewed as a conversion
of one continuous real-valued function into another one. If the transforma-
tion is reversible, then we say that each function is the dual of the other. The
reversible Legendre transformation is also called the Legendre conjugate trans-
formation, or simply the Legendre transformation.
Let E = IRn = E ∗ . The element = {i } ∈ E and σ = {σi } ∈ E ∗ , (i =
1, 2, · · · , n) are vectors in IRn . The bilinear form
n
X
< , σ >= · σ = i σ i (8.84)
i=1
Figure 8.5. Duality between the pole and polar
is then the inner product on IRn .

Let U : E → IR be a real-valued function. Its graph,
{(, X) ∈ IRn+1 X = U ()}
is a manifold (or hypersurface) in IRn+1 .

Let any particular point (σ, Y ) ∈ IRn+1 be called the pole. Then the linear
function
X() = · σ − Y (8.85)
is called the polar, which is a hyperplane in IRn+1 .
Thus, given a pole at a finite point, the polar is well-defined by (8.85), Con-
versely, given a polar of finite slope, a finite pole can be read off from Eq.
(8.85). This correspondence is called the duality between points and planes.
The duality comes to live when the graphi of a paraboloid is blended into
the picture.
Theorem 8.21 (Duality between the pole and polar) (T1)
If the pole is outside the paraboloid, the points of contact of tangents drawn from the
pole to the paraboloid lie on the polar.
(T2) If the pole is inside of the paraboloid, the polar lies outside it.
Proof:
We only prove the theroem in IR2 , which has the full flavor of a rigorous
proof.
We first show (T1). The tangential vector from the pole to the paraboloid is
t = (σ̄ − , Y (σ̄) − X)
1
the normal vector of graph G = U − 2 = 0 is
2
∂G ∂G
n= , = (−, 1)
∂ ∂U
We want show that the contact point is in the polar : X() = σ̄ − Y (σ̄).
Consider the condition t · n = 0.
t · n = (σ̄ − , Y − X)(−, 1)
= −σ̄ + 2 + Y − X
= −σ̄ + 2X + Y − X = −σ̄ + X + Y = 0
We just showed that X = σ̄ − Y .

We now show (T2). Suppose the pole is inside the paraboloid. We want to
show that the polar is outside the paraboloid region.
Assume that part of the polar is inside or no the paraboloid, i.e.
1
X ≥ 2
2
Since the pole is also inside the paraboloid, i.e.
1
Y (σ̄) > σ̄ 2
2
Therefore,
1 2
X + Y (σ̄) > σ̄ + 2
2
1 2
σ̄ > σ̄ + 2
2
1 2 1
0 > σ̄ − 2σ̄ + 2 = (σ̄ − )2 > 0
2 2
which leads to contradiction. Hence, polar must be outside the paraboloid, if
the pole is inside the paraboloid. ♣
Definition 8.22 (Regular point and regular domain) Let U : E →

IR be a piecewise C 2 function.
(D1) A regular point of the function U () is a point ∈ E where the deter-
∂2U
minant of the Hessian matrix D2 U = { } satisfies,
∂i ∂j
2
∂ U
det 6= 0, or ± ∞
∂i ∂j
(D2) A regular domain, denoted by Er is a continuous subset of regular
points.
Now we let U c : IRn → IR be a given continuous function such that the

graph,
GU c = {(σ, Y ) ∈ IRn Y = U c (σ), σ ∈ IRn }
of U c is a continuous surface in IRn+1 .

When the pole, (σ, Y ), moves on the graph of U c , each point on GU c is cor-
responding to a polar hyperplane. The collective of these polars hyperplanes
will envelop another continuous surface, the graph of X = U (), described
as U : IRn → IR, which is the conjugate Legendre pair of U c (σ). This is the
geometric interpretation of Legendre transformation. In other words, the cor-
respondence between the functions U () and U c (σ) is called Legendre trans-
formation.
Now we state the important Legendre Dulaity theorem.
Theorem 8.23 (Legendre Duality Theorem) Let U () ∈ C 2 (E).

If Er ⊂ E is an open, finite subset of the regular domain of U and Er∗ ⊂ IRn
is the range of the mapping DU : Er → E ∗ . Then there exists a unique C 2
function U c on E ∗ , which is dual to U on Er in the sense that the Legendre
duality relates
U () + U c (σ) = σ · ⇔ σ = ∂U (), ⇔ = ∂U c (σ)
hold. Moreover, for (, σ) ∈ Er × Er∗ satisfying above relationship,
∂2U ∂2U c
= δij .
∂i ∂k ∂σk ∂σj
The proof of this theorem is basically application of implicit function theo-
rem. It is omitted here. The readers who are interested in the proof may consult
Gao [2000].
Now we move to the essentail technical ingradient of convex analysis.
Theorem 8.24 (Duality between the regular manifolds) Let U
and U c be Legendre dual functions over the duality domain E and E ∗ respec-
tively.
Figure 8.6. Geometric interpretation of Legendre transformation
(S1) If U is convex on E, U c is convex on E ∗ and
U c (σ) = max{σ · − U ()}

∈E
(S2) If U is concave on E, U c is concave on E ∗ and
U c = min{σ − U ()}
∈E
Proof;
For simplicity, we only prove it for case E ⊂ IR, which contains the enssen-
tial substance of a general, rigorous proof.
∂U
Since σ = , by Taylor expansion,
∂
∂U ∂2U
σ= + 2 ∗ ( − ¯) (8.86)
∂ =¯ ∂
where
∂2U ∂2U
∗ =
∂2 ∂2 =¯
+θ∆
and 0 ≤ θ ≤ 1.
Eq. (8.86) can be rewritten as
∂2U
(σ − σ̄) = + ∗ ( − ¯) (8.87)
∂2
∂U c
By the same token, because of = , one can have
∂σ
∂2U
( − ¯) = + 2 ∗ (σ − σ̄) (8.88)
∂
where
∂2U c ∂2U
∗ =
∂σ 2 ∂σ 2 =σ̄+θ∆σ
and 0 ≤ θ ≤ 1. Therefore,
∂2U
(σ − σ̄)( − ¯) = ∗ ( − ¯)2
∂2
∂2U c
= ∗ (σ − σ̄)2 (8.89)
∂σ 2
∂2U ∂2U c
Eq. (8.89) indicates that if 2
∗ is positive definite, 2
∗ is also
∂ ∂σ
∂2U ∂2U c
positive; whereas if ∗ is negative definite, ∗ is also negative
∂2 ∂σ 2
definite, or both being indefinite.
To prove the Legendre inequality, we consider a special 1D example, U () =
1 2
k0 , k0 > 0.
2
For a given point ¯ on horizontal axis, the associated stress σ̄ = k0 ¯ is the
slope of the polar, the straight line X = σ̄ − Y , which is tangent to the graphy
of U at ¯ (see Fig. (8.7).
Therefore, point (¯ ) is in both polar X = σ̄ − Y and on U = 1/2k0 2 ,
, U (¯
which is to say that X(¯ ) = U (¯ ) and
) =: U c (σ̄)
− U (¯
Y = σ̄¯
For any given ∈ Er , we define a continuous function,
y() = σ̄ − U ()
we want to show that Y = U c (σ̄) ≥ y().

Since the polar X() is always below the parabola (U () ≥ X(),
U () − X ≥ 0 ⇒ U () − (σ̄ − Y ) ≥ 0

⇒ Y ≥ σ̄ − U ()
Figure 8.7. Legendre transformation
∂2y
Since U () is convex, y() is then concave because < 0. It then takes its
0
∂2
maximum value at ¯ because y (¯
) = 0. That is
Y = U c (σ̄) = max {σ̄ − U ()} (8.90)

∈Er
One drop the bar on σ, because domain of σ̄ is the same as σ.

Similarly, for concave function, one can show that
U c (σ) = min {σ − U ()}

∈Er
Remark 8.5.1 In the infinite-dimensional space E, Eq. (8.90) is called

Legendre-Fenchel transformation, and it reads as
U ∗ (σ) = sup {σ · − U ()}

∈E
where the superscript ∗ replaces the superscript c meaning as the dual func-
tion.
Accordingly, if U is concave, its Legendre-Fenchel conjugate is defined as
U ∗ (σ) = inf {σ · − U ()}

∈E
The reason we add the name Fenchel is because when U is defined as

[
U : E → IR {+∞}
the transformation
U ∗ (σ) = sup {σ · − U ()}
∈E
is called the Fenchel transformation.
8.6 Legendre-Fenchel transformation in linear elasticity

In a classical paper (Hill [1965]), Hill illustrated the Legendre-Fenchel trans-
formation in linear elastic system and extend the use of classical minimum
potential energy principle and minimum complementary energy principle to
micromechanics.
Consider the prescribed displacement boundary condition (prescribed macro
strain condition),
u0 = x · 0 , ∀x ∈ ∂V
Under such condition, we have shown previously that
0 =< >=< ˜ >, ∀ ∈ E
where E is the space of compatible strain.
Therefore, the potential energy and complementary energy take the form
Z Z
1
Πc = Dijk` σij σk` dV − xk 0ki σij nj dS
2 Ω
Z Z∂Vh
1 i
= Dijk` σij σk` dV − δkj 0ki σij + xk 0ki σij,j dV
2 Ω V | {z }
=0
Z Z
1
= Dijk` σij σk` dV − 0ij σij dV
2 Ω V
Based on minimum complementary energy principle, for any statically admis-
sible stress field, ∀σ ∈ S,
Z Z
1
Πc (σ) ≥ Dijk` σ̃ij σ̃k` − u0i σ̃ij nj dS
2 V
Z Z∂V
1
= Dijk` σ̃ij σ̃k` − 0ij σ̃ij dV
2 V ∂V
Z
1
= − Cijk` ˜ij ˜k` dV
2 C
where σ̃ and ˜ is the real solution. In the last line the equality under prescribed
macros strain, Z Z
0ij σ̃ij dV = ˜ij σ̃ij dV
V V
is used.
Therefore,
Z Z Z
1 1
Cijk` ˜ij ˜k` dV ≥ σij 0ij dV − Dijk` σij σk` dV
2 V V 2 V
which is essentially
) = sup 0 :< σ > −W c (σ)

W (˜ (8.91)
{σ ∈S}
where
Z
1
W () = Cijk` ij k` dV
2V V
Z
1
W c (σ) = Cijk` σijσk` dV
2V V
One may further tighten the bound
0
W (˜
) = sup :< σ > −W c (σ̃) (8.92)
{<σ >:σ ∈S}
Remark 8.6.1 1. Note that Eq. (8.91) looks like Legendre-Fenchel trans-
formation. However, there is a subtle difference.
If W is a convex functional of ∈ E, the Legendre-Fenchel transformation
assures that
W c (σ) = sup {σ : − W ()}
{∈E}
If the space E = E ∗∗ is reflexive (all the Lp (V ) spaces are reflexive, see

Rudin [1991]), the inverse Legendre-Fenchel transformation exists,
W () = (W c )c () = W cc () = sup { : σ − W c (σ)}

{σ ∈S}
2. Choose
n Z n
X 1 α 0
X
< σ >= C : dV = fα C : 0 .
V Vα
α=0 α=0
One can show that

( n
1 0 X
: C̄ : 0 ≥ 0 : fα Cα
2
α=0
n n n
)
1 X α
X X
− fα C : fα Cα : fα Cα : 0
2
α=0 α=0 α=0
Hence
n n n
( )
X X X
C̄ ≥ 2 fα Cα : 1 (4s)
− α
fα C : fα Cα
α=0 α=0 α=0
which is referred to as the Sachs bound.
8.7 Talbot-Willis variational principles

In a series papers (Talbot and Willis [1985],[1987]), Talbot and Willis gen-
eralized Hashin-Shtrikman variational principles to well-behaviored nonlinear
media.
Consider a composite with nonlinear strain potential energy density, U (),
∇ · σ = 0,
σ = ∂ U,
1
= (∇ ⊗ u + (∇ ⊗ u)T )
2
u = x · ¯, ∀x ∈ ∂V (Γt = ∅)
Consider a homogeneous composite,
∇ · σ 0 = 0, (8.93)
σ 0 = ∂0 U 0 , (8.94)
1
0 = (∇ ⊗ u0 + (∇ ⊗ u0 )T ) (8.95)
2
u0 = x · ¯, ∀x ∈ ∂V (Γt = ∅) (8.96)
Compare the differences in potential energy density, U() = U () − U 0 ().
We define
Up () := U () − U 0 (), ∂2 U > 0 (8.97)
U p () := U () − U 0 (), ∂2 U < 0 (8.98)
Assume the following kinematic decomposition,
u = u(0) + u(d) (8.99)
= (0) + d (8.100)
Assume that the stress and strain fields in the comparison solid are uniquly
determined by the boundary condition. The total potential energy difference is
a functional of d , i.e.
Z
1
Πp (d ) = W (d ) − W0 (d ) = Up (d )dV (8.101)
V V
Z
1
Πp (d ) = W (d ) − W0 (d ) = U p (d )dV (8.102)
V V
where
Z
1
W (d ) = U (d )dV
2V V
Z
1
W0 (d ) = U0 (d )dV
2V V
Obviously, Πp () is convex and Πp () is concave.
Define stress polarization
∂U
pij = d (8.103)
∂ij
Subsequently, we can form the following Legendre-Fenchel transformation,
n o
Π∗p = sup −Πp (d ) (8.104)
d ∈E
n o
Πp∗ = inf −Πp (d ) (8.105)
d ∈E
where Z
d 1
= p : d dV
V V
and

1 ◦
2
E := ij ij ∈ L (V ), ij = (ui,j + uj,i ), and ui ∈ V
2
n o
V := ui ui ∈ L2 (V ), W (ui,j ), W0 (ui,j ) < ∞, ui = xj 0ij , ∀x ∈ ∂V
◦ n o
V := ui ui ∈ L2 (V ), W (ui,j ), W0 (ui,j ) < ∞, ui = 0, ∀x ∈ ∂V
In fact, in plain terms, Eqs. (8.104) and (8.105) are just

n o
∗ ∗ d d
Πp (p) = (W − W0 ) (p) = sup −(W − W0 )( ) ,(8.106)
{d ∈E}
when ∂ 2 U > 0, and (W − W 0 ) is convex,

n o
Πp∗ (p) = (W − W0 )∗ (p) = inf −(W − W0 )(d ) ,(8.107)
{d ∈E}
when ∂ 2 U < 0, and (W − W 0 ) is concave.
(1.) Assume ∂ 2 U > 0. From Eq. (8.106)

Π∗p (p) ≥ −W (d ) + W0 (d )
⇒ W (d ) ≥ { +W0 (d )} − Π∗p (p)
Take an infimum through the both sides of the inequality,

inf W (d ) ≥ inf { +W0 (d )} − Π∗p (p) (8.108)
{d ∈E} d ∈E
(2.) Assume ∂ 2 U < 0. From Eq. (8.107)

Πp∗ (p) ≤ −W (d ) + W0 (d )
⇒ W (d ) ≤ { +W0 (d )} − Π∗p (p)
Take an infimum through the both sides of the above inequality
inf W (d ) ≤ inf { +W0 (d )} − Πp∗ (p) (8.109)
{d ∈E} {d ∈E}
The prime variational principle is

(The primal problem) P : inf W (d )
{d ∈E}
Combining Eqs. (8.108) and (8.109), we have the original form of Talbot-
Willis variational princinple
inf { +W0 (d )} − Π∗p (p)
{d ∈E}
≤ inf W (d ) ≤
{d ∈E}
inf { +W0 (d )} − Πp∗ (p) (8.110)

{d ∈E}
which is the generalization of Hashin-Shtrikman principle.

If both the original composite and the comparison solid are linear elastic
materials, we easily calculate,
Z
1 1
Π∗p (p) or Πp∗ (p) = dij pij − ∆Cijk` ij k` dV
V V 2
Z
1 1
= (ij − 0ij )pij − pij ij dV
V V 2
Z
1
= ij pij − 20ij pij dV
2V V
Z
1 −1

= ∆Cijk` pij pk` − 20ij pij dV
2V V
Denote
inf { +W0 (d )} − Π∗p (p)
I(d , p) = (8.111)
d ∈E
¯ , p) = inf { +W0 (d )} − Πp∗ (p)
I( d
(8.112)
d ∈E
We can find that

Z
¯ = 1 1 0
I (or I) pij dij + Cijk` k` (0ij + dij )
V V 2
1 −1

− ∆Cijk` pij pk` + 0ij pij dV
2Z
1
0

= pij + Cijkl k` dij dV
2V V
| {z }
=0
Z
1
+ C 0 (0 + dk` )0ij dV
2V V ijk` k`
Z
1 1 d 1 −1

+ ij pij − ∆Cijk` pij pk` + 0ij pij dV
V V 2 2
Z
1
= C 0 0 d dV
2V V ijk` k` ij
| {z }
=0
Z
1 1 0 0 0 1 1 −1

+ Cijk` ij k` + dij pij − ∆Cijk` pij pk` + 0ij pij dV
V V 2 2 2
Hence
Z
¯ = 1 1 0 0 0 1 d 1 −1 0

I, (or I) C + pij − ∆Cijk` pij pk` + ij pij dV
V V 2 ijk` ij k` 2 ij 2
0
= W0 ( ) + Rπ (or R̄π )
where
Z
1 −1

Rπ , (orR̄π ) = −∆Cijk` pij pk` + pij dij + 2pij 0ij dV
2V V
We then recover the Hashin-Shtrikman variational principle
Rπ (p, d ) ≤ inf W (d ) − W0 (0 ) ≤ R̄π (p, d )

d ∈E
8.8 Exercises
Probelm 8.1 Consider a functional
P : H 1 ([a, b]) → IR
where Z bq
P (u) = 1 + [u0 (x)]2 dx .
a
with essential boundary condition u(a) = ūa and u(b) = ūb .
Find the first variation, second variation, and Gâteaux derivative. Derive
associated the Euler-Lagrange equation.
Probelm 8.2 Let Γu = ∅, ∂V = Γt , and fi = 0. Assume that the RVE has

the prescribed traction boundary condition,
n · σ̄ = t0 (x), ∀x ∈ ∂V (8.113)
where σ̄ > 0 is a constant tensor.

Show that
n o
W c (σ̃) = sup σ̄ : hi − W̃ (< >) (8.114)
{<> ∈E}
where E := {ij ij,kl + kl,ij − ik,jl − jl,ik = 0, and ij ∈ L2 (V )},

Z Z
1 1
W c (σ̃) := Dijkl σ̃ij σ̃kl dV = ˜ij σ̄kl dV (8.115)
2V V 2 V
Z
1
W̃ (< >) := inf Cijkl ij kl dV (8.116)
2V { V1
R
V dV =<>, ∈E} V
Note that σ̃ij and ˜ij are the real solutions.
Probelm 8.3 Let Γu = ∅ and ∂Ω = Γt . Consider the following the boundary-

value problem,
σij,j = 0, ∀x ∈ Ω (8.117)
nj σij = t0i , ∀x ∈ Γt , and Γu = ∅ (8.118)
1
ij = ui,j + uj,i (8.119)
2
∂Uc 1
ij = , Uc (σ) := Dijk` σij σk` . (8.120)
∂σij 2
0
Consider a comparison elastic solid with compliance tensor, Dijk` and
(0)
σij,j = 0, ∀x ∈ Ω (8.121)
(0)
nj σij = t0i , ∀x ∈ Γt , and Γu = ∅ (8.122)
(0) 1 (0) (0)

ij = ui,j + uj,i (8.123)
2
(0)
(0) ∂U0 (0) 1 (0) (0) (0)
ij = , U0 (σ) := D σ σ . (8.124)
∂σij
(0) 2 ijk` ij k`
Let
(0)
d
σij = σij + σij (8.125)
(0)
ij = Dijk` σk` + qij (8.126)
where σijd is called disturbance stress, and q is called polarization strain

ij
(eigenstrain).
They are connected by the following subsidary conditions: 1. the weak form
of subsidiary condition (complementary virtual work principle),
Z
d
ij σij dΩ = 0 (8.127)
Ω
or 2. the strong form of subsidiary condition

0 (0)
d 0 0 0 0 0
:= Dijk` σk` + qij , C(ij ) = ij,k` + k`,ij − ik,j` − i`,jk = 0, ∀x ∈ Ω
(8.128)
Consider the following variational problem
(The primal problem :) P : inf Πc (σ d ) (8.129)

σ d ∈S(Ω)
or
(The primal problem :) P : inf Wc (σ d ) (8.130)
σ d ∈S(Ω)
where
Z Z
d 1 (0)d (0) d
Wc (σ ) := Dijk` σij σk` dΩ = Dijk` (σij + σij )(σk` + σk` )dΩ,
2|Ω| Ω Ω
(8.131)
Πc (σ d ) = ΩWc (σ d ) and
n o
S := σ nj σij = 0, ∀x ∈ Γt , and σij ∈ C 0 (Ω) (8.132)
Derive Hashin-Shtrikman variational principle.

Hints:
Z. Hasin and S. Shtrikman [1962], “On some variational principles in anisotropic
and nonhomogebeous elasticity,” Journal of Mechanics and Physics of Solids,
10, pp. 335-342.
D. R. S. Talbot and J. R. Willis [1985], “Variational principles for inhomo-
geneous non-linear media,” IMA Journal of Applied Mathematics, 35, 39-54.
Chapter 9
BOUNDS ON EFFECTIVE PROPERTIES
9.1 Hashin-Shtrikman bounds

Consider prescribed macro strain boundary condition for both the composite
and the comparison solid,
u = ū = x · ¯, ∀x ∈ ∂V (Γt = ∅)
u0 = ū = x · ¯, ∀x ∈ ∂V (Γt = ∅)
by the averaging theorem ¯ =< >.

Under such condition, Hashin-Shtrikman variational principles are
I¯
I ≤ inf W (d ) ≤ |{z} (9.1)
|{z} d ∈E
∆C>0 ∆C<0
where ∆C = C − C(0) , and

Z h

¯

0 1 −1
i
I or I = W0 ( ) − ∆Cijk` pij pk` − pij dij − 2pij 0ij dV (9.2)
2V V
Assume that there are n-phase in the composite (including the matrix). In
each phase (inclusion), the elastic tensor as well as stress polarization tensor is
constant, i.e.
n
X
C(x) = Cr H(Ωr ) (9.3)
r=1
n
X
p(x) = pr H(Ωr ) (9.4)
r=1
Bounds on Effective Properties 231
where H(·) is the Heaviside function, and Ωr is the domain of each phase,

 1, ∀x ∈ Ωr
H(Ωr ) =
0, ∀x 6∈ Ωr

We now calculate each term in (9.1).

1
Z
d 1 1
inf W ( ) = σ : dV = < σ >:< >
d ∈E 2V V 2
1 1
= < >: C̄ :< >= ¯ : C̄ : ¯ (9.5)
2 2
2
Z
1 1
W0 (0 ) = σ 0 : 0 dV = < σ 0 >:< 0 >
2V V 2
1 1
= < 0 >: C0 :< 0 >= ¯ : C0 : ¯ (9.6)
2 2
3
Z n Z
1 −1 1X 1
r
p : ∆C : pdV = pr : C−1 r
r : p dV
2V V 2 V Ωr
r=1
n
1X
= fr pr : ∆C−1
r :p
r
(9.7)
2
r=1
4
Z 1 Z n
1 0
X
p : dV = pdV : ¯ =: ¯ = fr pr : ¯ (9.8)
V V V V r=1
5 Z n
1 d 1X
p : dV = − fr pr : Pr : pr − (9.9)
2V V 2
r=1
where Z
P := r
Γ∞ (x0 − x)dVx0
Ωr
and
1 ∞
Γ∞
ijk` := − Gki,j` (x0 −x)+G∞
kj,i` (x0
−x)+G ∞
ì,jk (x0
−x)+G ∞
`j,ik (x0
−x)
4
How to integrate Z
1
p : d dV =? (9.10)
2V V
Consider the subsidiary condition,
(0)
Cijk` udk,`j + pij,j = 0 (9.11)
We solve udk in terms of pij by using Green’s function method. Conisder the
Green’s function of the comparison solid in an infinite medium, i.e.
(0)
Cijk` G∞ 0 0
km,`j + δim δ(x − x ) = 0, ∀ x, x ∈ IR
3
Multiplying Gim (x0 − x) with (9.11) and integrating it over V, one has
Z h i
(0)
Cijk` udk,` + pij G∞ 0
im (x − x)dVx0 = 0
V ,j
(0)
Let tij = Cijk` udk,` . Integration by parts yields,
Z h i
(0) d
G∞
im (x 0
− x) C u
ijk` k,` +p ij nj dS
∂V | {z }
tij
Z
∂ ∞ 0 h
(0) d
i
− G
0 im (x − x) C u
ijk` k,` + p ij dV
V ∂xj
Z Z h
∞ 0 ∂ ∞ 0 ih
(0) d
i
= Gim (x − x )[tij + pij ]nj dS − G
0 im (x − x) C u
ijk` k n ` dS
∂V ∂V ∂xj | {z }
=0
∂2
Z h Z
∞
ih
(0) d
i ∂ ∞ 0 0
+ 0 0 Gim Cijk` uk n` dV − 0 Gij (x − x)pij (x )dV
V ∂x j ∂x ` V ∂x j
Z Z
∂
= G∞ 0
im (x − x)[tij + pij ]nj dS −
∞ 0 0
0 Gij (x − x)pij (x )dV
∂V V ∂xj
Z
(0)
+ Cijk` G∞ 0 d 0
km,j` (x − x) ui (x )dV
V | {z }
−δim δ(x0 −x)
because of major symmetry of C(0) , one can interchange indices k → i and

j → `.
Therefore,
Z Z
∞ 0 0 0
d
um (x) = G (x − x)[tij (x ) + pij (x )]nj dS − G∞ 0 0
im,j (x − x)pij (x )dV
∂V V
(9.12)
Since ud = 0, ∀ x ∈ ∂V , tij oscillate around zero. Then its average

(0)
< Cijk` udk,` >∂V along the boundary should be very small. We assume that
(0)
< Cijk` udk,` >∂V ≈ 0
Now the only term remaining is

Z
G∞ 0 0
im (x − x)pij (x )dS
∂V
To essence of the additional manipulation is to modify the volume integral

in (9.12) in order to drop out the surface integral in (9.12). To do accomplish
this goal, we consider identity,
Z
< pij >,j = 0 ⇒ < pij >,j G∞ 0
im (x − x)dV = 0
V
Integration by parts yields,

Z Z
∞ 0
< pij >,j Gim (x − x)dV = < pij > nj G∞ 0
im (x − x)dS
VZ ∂V
− G∞ 0
im,j (x − x) < pij > dV = 0 (9.13)
V
Thus substracting (9.13) from (9.12) will be affect the value of (9.12),
Z
0 0
d
um (x) = G∞ 0
im (x − x)[tij (x ) + (pij (x )− < pij >)]nj dS
∂V
Z
0
− G∞ 0
im,j (x − x)(pij (x )− < pij >)dV (9.14)
V
Now pij − < pij > also oscillates around zero, since its mean is zero, i.e.
< pij − < pij >>= 0. We can then neglect the boundary term, and finally we
have
Z
d
um (x) ≈ − G∞ 0 0
im,j (x − x)(pij (x )− < pij >)dVx0 (9.15)
V
The gradient of the disturbance displacement field is

Z
0
udm,` (x) = G∞ 0
im,j` (x − x)(pij (x )− < pij >)dVx0
V
Hence
Z h
1 i 0 0

dm` (x) = G∞
im,j` + G ∞
i`,jm (x − x) p ij (x )− < pij > dVx0 (9.16)
2 V
Since pij is symmetric, we can also write that

Z h
1 i 0
dm` (x) = G∞ + G ∞
+ G ∞
+ G ∞
j`,im (x − x)
4 V im,j` i`,jm jm,i`
0

· pij (x )− < pij > dVx0
Z
0 0
= − Γ∞mìj (x − x) p ij (x )− < pij > dVx0
ZV
0
0

= − Γ∞ (x − x) : p(x )− dVx0 (9.17)
V
where
1h ∞ i
Γ∞
mìj (y − x) := − Gim,j` + G∞ ∞ ∞
i`,jm + Gjm,i` + Gj`,im (y − x) (9.18)
4
Consider a bounded and simply-connected region, Ω ∈ V . We define a new

tensor, P,
Z
PΩ (x) := Γ∞ (y − x)dVy , ∀x ∈ Ω (9.19)
Ω
and in components form,

Z
Ω
Pijk` (x) = Γ∞ijk` (y − x)dVy
Ω
Z h
1 i
= − G∞ + G ∞
+ G ∞
+ G ∞
j`,im (y − x)dVy
4 Ω im,j` i`,jm jm,i`
(9.20)
One may verify that when Ω is an ellipsoidal PΩ is constant. In fact, if one

recalls the general definition of Eshelby tensor, for x ∈ Ω,
Z
Ω
Sijk` = Gijk` (y − x)dVy (9.21)
Ω
Z
1 h i
= − Cmnk` G∞ ∞
im,nj + Gjm,ni (y − x)dVy
2 Ω
Z
1 h ∞ i
= − Gim,nj + G∞ ∞ ∞
jm,ni + Gin,mj + Gjn,mj (y − x)Cmnk` dVy
4 Ω
Z
= Γ∞ijmn (y − x)Cmnk` dVy
Ω
Ω
= Pijmn Cmnk` (9.22)
Now we come back to evaluate (9.10). Let stress polarization p(x) is piece-
wise constant, i.e.
Xn
p(x) = pr H(Ωr )
r=1
n
X
 = fr pr
r=1
Therefore,
Z Z X n
1 1
p : d dV = pr H(Ωr ) :
2V V 2V V
r=1
Z 0
h 0
i
− Γ∞ (x − x) : p(x )− dVx0 dVx
V0
Consider x ∈ Ωs . ps − is constant inside Ωs . Thus,
Z
0
Γ∞ (x − x) : pr − dVx0
V0
Z Z
∞ 0 0
= Γ (x − x)dVx0 + Γ∞ (x − x)dVx0 : pr − dVx0
Ωs V 0 −Ωs
Assume that the RVE is a gigantic spherical ball and all Ωr are spherical inclu-
sions. By Mori-Tanaka lemma,
Z
0
Γ∞ (x − x)dVx0 = 0
V 0 −Ωs
In fact, for x ∈ Ωs
Z Z
∞ 0 0
Γ (x − x)dVx0 = Γ∞ (x − x)dVx0
V Ωs
because the integral over a spherical ball does not dependent on the size of
inlcusion (recall P = S : D).
Hence,
Z n n Z
1 d 1 XX
p : dV = − pr H(Ωr (x)
2V V 2V Ωr
r=1 s=1
Z
∞ 0
: Γ (x − x)dVx : ps H(Ωs (x) dVx
0
Ωs
n Z
1 X
+ pr H(Ωr (x)
2V Ωr
r=1
Z
∞ 0
: Γ (x − x)dVx0 : dVx
Ωs
Consider 
 1 r=s
H(Ωr (x))H(Ωs (x)) = (9.23)
0 r 6= s

and let Z
0
r
P := Γ∞ (x − x)dVx0
Ωr
We then have
Z n Z
1 d 1 X
p : dV = − dVx pr : Pr : pr
2V V 2V
r=1 Ωr
n Z
1 X
+ dVx pr : Pr :
2V Ωr r=1
n
1X
= − fr pr : Pr : (pr − )
2
r=1
n
1X
= − fr prij Pijk`
r
prk` − < pk` >
2
r=1
Pn r
where < pk` >= r=1 fr pk` .
Remark 9.1.1 Recall that by using Radon transform one can write,
Z
1 00
δ(x) = − 2 δ (ξn xn )dS
8π |ξ |=1
and consequently,
Z
1 −1
G∞
ij (x) = 2 Kij (ξ)δ(ξn xn )dS
8π |ξ |=1
and for isotropic materials,

−1 1h (λ + µ)ξi ξj i
Kij (ξ) = δij −
µ (λ + 2µ)
Therefor, Z
1 −1 00
G∞
ij,k` (x) = 2 Kij (ξ)ξk ξ` δ (ξn xn )dS
8π |ξ |=1
By definition,
1h ∞ i
Γ∞
ijk` (x − y) := − Gik,j` + G∞
i`,jk + G ∞
jk,i` + G ∞
j`,ik (x − y)
4 Z
1 −1 00

= − 2 Kij (ξ)ξk ξ` δ ξn (xn − yn ) dS
8π |ξ |=1
−1
because indices i & j and k & ` are symmetric (Kij (ξ) is symmetric).
To this end, we are in a position to establish Hashin-Shtrikman bounds. Be-

fore proceeding to derive Hashin-Shtrikman bound, we first evaluate P tensor,
which can be written as
P = S : D(0)
For spherical inclusion,
(0) (0)
S = s1 E(1) + s2 E2
where
1 + ν (0) 2(4 − 5ν (0) )
s1 = , s 2 =
3(1 − ν (0) ) 15(1 − ν (0) )
and for isotropic comparison solid,
1 1
D(0) = (0)
E(1) + E(2)
3K 2G(0)
Therefore,
(0)
s1 (1) s(0) (2)
P = E + E
3K (0) 2G(0)
1 + ν (0) (1) (4 − 5ν (0) )
= (0) (0)
E + (0) (0)
E(2)
9K (1 − ν ) 15G (1 − ν )
( )
1 1 (2) 2(4 − 5ν (0)
= − 1 ⊗ 1(2) + 1(4s) (9.24)
2G(0) (1 − ν (0) ) 15 15
3K (0) − 2G(0)
Consider ν (0) = . One can also have
2(3K (0) + G(0) )
1 (1) 3(K (0) + 2G(0) )

P= E + E(2)
3K (0) + 4G(0) 5G(0) (3K (0) + 4G(0) )
For simplicity, we only illustrate Hashin-Shtrikman bound for a two-phase
composite. Consider a two-phase well order composite, which implies that
K2 > K1 and G2 > G1 .
Step 1. Let
K0 = K1 , K = K2 , and G0 = G1 , G = G2 .
Obviously that
∆C = C − C(0) = 3(K2 − K1 )E(1) + 2(G2 − G1 )E(2) > 0

Choose a special stress polarization distribution,

(1) (2)
pij = 0, and pij = pδij .
and remote macro strain distribution
¯ij = ¯δij
We now calculate each terms in I.
1
1
inf W (d ) = C̄ijk` ¯δij ¯δk`
d ∈E 2
1h (1) (2)
i
= 3K̄Eijk` + 2ḠEijk` (¯ )2 δij δk`
2
9 2
= K̄¯
2
(1) (2)
Note that Eijk` δij δk` = 3 and Eijk` δij δk` = 0.
2
1 (0)
W0 (0 ) = C (¯ δij )(¯
δk` )
2 ijk`
1h (1) (2)
i
= 3K1 Eijk` + 2G2 Eijk` (¯ )2 δij δk`
2
9
= K1 ¯2
2
3 Z
1
p : (0) dV = f1 p1 : ¯ + f2 p2 : ¯ = 3f2 p¯

V V
(1) (2)
4 Because pij = 0 and pij = pδij ,
Z 2
1 1X
p : ∆C−1 : pdV = fr pr : ∆C−1r : pr
2V V 2
r=1
1 f2 f2

= E(1) + E(2) p2 δij δk`
2 3(K2 − K1 ) 2(G2 − G1 )
f2 p 2
=
2(K2 − K1 )
5 Because < pk` >= f2 pδk` ,

Z 2 2
1 d 1X r r r 1X r
pij ij dV = − fr Pijk` pij pk` + fr Pijk` prij < prk` >
2V V 2 2
r=1 r=1
f2 3p2 1 3f 2 p2
2
= − +
2 3K1 + 4G1 2 3K1 + 4G1
1 3f1 f2 p2 1 f1 f2 p2
= − =−
2 3K1 + 4G1 2 4
K1 + G1
3
Therefore, when ∆C > 0,
9 f2 p2 1 f1 f2 p2 9 2
I(p) = K1 ¯2 − − ≤ K̄¯
+ 3f2 p¯ (9.25)
2 2(K2 − K1 ) 2 K1 + 43 G1 2
To find minp I, we check the stationary condition,
∂I f2 p f1 f2 p
=0 ⇒ − − + 3¯
f2 = 0
∂p (K2 − K1 ) K1 + 43 G1
3¯

⇒ psta = (9.26)
1 f1
+
K2 − K1 K1 + 43 G1
Substituting (9.26) into (9.25) yields a lower bound on bulk modulus
f2
K̄ ≥ K1 + (9.27)
1 f1
+
K2 − K1 K1 + 34 G1
Step 2: Let
K0 = K2 , K = K1 , and G0 = G2 , G = G1
and choose
(1) (2)
pij = pδij , pij = 0 .
One can find an upper bound,
¯ = 9 K2 ¯2 − f1 p2 1 f1 f2 p2 9 2
I(p) − 4 ≥ K̄¯
+ 3f1 p¯ (9.28)
2 2(K1 − K2 ) 2 K2 + 3 G2 2
¯
To find the maximum value of I(p), we examine the stationary condition,
∂ I¯ 3¯

= 0, ⇒ psta = (9.29)
∂p 1 f2
+
K1 − K2 K2 + 43 G2
Figure 9.1. Variational Bounds for Bulk Modulus: (a) Medium one, and (b) Medium two.
Figure 9.2. Variational Bounds for Shear Modulus: (a) Medium one, and (b) Medium two.
Substituting (9.29) into (9.28), one will find that

f2
K̄ ≤ K2 + (9.30)
1 f2
+
K1 − K2 K2 + 34 G2
By combining (9.27) and (9.30), we will have the Hashin-Shtrikman bound on
bulk modulus,
f2 f1
K1 + ≤ K̄ ≤ K2 + (9.31)
1 f1 1 f2
+ +
K2 − K1 K1 + 43 G1 K1 − K2 K2 + 34 G2
It is readily to show that the following Hashin-Shtrikman bounds are held for
shear modulus,
f2 f1
G1 + ≤ Ḡ ≤ G2 +
1 6(K1 + 2G1 )f1 1 6(K2 + 2G2 )f2
+ +
G2 − G1 5(3K1 + 4G1 )G1 G1 − G2 5(3K2 + 4G2 )G2
(9.32)
Figure 9.3. A compsite with n-phases
9.2 Microstructure Characterization

9.2.1 Preliminary
In this section, a few important concepts about statistical evaluate of a ran-
dom heterogeneous material shall be discussed, or formally defined. First, we
assume that any sample of a random heterogenous material is a realization of
a specific random or stochatic process. Mathematically speaking, a realization
is an event, α, that belongs to a sample space, S. Second, an ensemble is the
collection of all the possible realizations of a random medium generalized by
a specific stochastic process.
Consider a sample space S over which a probability density function, p(α),
is defined, α ∈ S. Then any particular property, f , of a composite (such as
mass density, volume fraction density) is a function of α, and its ensemble
average can defined as
Z
< f >= f (α)p(α)dα (9.33)
S
Of particular interest is the indicator function, Suppose that there is a n-

phase ramdom medium (composite), V ∈ IRd . The total volume of V is par-
tition into n-disjoint random sets or phases. The phase 1 occupies the set V1 ,
and, in general, the phase r occupies the Vr , r = 1, 2, · · · , n. The measure of
set Vr is denoted as volume fraction, fr = meas(Vr ). Obviously, the set {Vr }
is a subdivision, i.e.
n
[
Vr (α) = V,
r=1
Vi ∩ Vj = ∅, if i 6= j
The indicator function for the phase, r, is defined as


 1, if x ∈ Vr (α)
I (r) (x, α) = (9.34)
0, otherwise

The indicator function is a partition of unity,

n
X
I (r) (x, α) = 1 .
r=1
In many mathmatical literature, the indicator function is also called as chara-

teristic function.
The expectation or probability of finding phase R ar a chosen point, x, is
then denoted as
Z n o
r (r)
S1 (x) :== I (r) (x, α)p(α)dα = P I (r) (x) = 1 (9.35)
S
(r)
In the literature, the function, S1 , is referred to as the one-point probability
function for phase, r, since it gives the probability to find phase r at position x.
It is also referred to as the one-point correlation function for the phase indicator
function, I (r) .
In general, the expectation, or probablity, to find the phase, r, at different n
points simulatenously is referred to as the n-point probability function, which
is defined as
Sn(r) (x1 , x2 , · · · , xn ) := (9.36)
Here the subscript, n, indicates that this is a n-point probability function, and
the superscript, (r), denotes that this is a n-point correlation function for phase
r.
One can further generalize the above concept of correlation function to the
probability of finding any subset of points ni of the n points in phase i and
another subset of points nj of the n points in phase j as
Sn(ij) (x1 , x2 , · · · , xn ) :=
(9.37)
For instance, a two-point correlation function that represents the probability
to find the phase, r, in x1 and the phase, s, in x2 is defined as
(rs)
S2 (x1 , x2 ) := (9.38)
Consider a n-phase composite. Its mass density can be expressed as
n
X
ρ(x) = ρr I (r) (x) (9.39)
r=1
Figure 9.4. Examples of statistically inhomogeneous mdeia
Then the expectation of the density function is

n
Z X
< ρ(x) > = ρr I (r) (x, α)p(α)dα
S r=1
n
(r)
X
= ρr S1 (x)
r=1
The expectation of the product of ρ(x1 ) and ρ(x2 ) is

n
Z X n
X
< ρ(x1 )ρ(x2 ) > = ρr I (r) (x1 , α) ρs I (s) (x2 , α) p(α)dα
S r=1 s=1
n n
(rs)
XX
= ρ r ρs S 2 (x1 , x2 )
r=1 s=1
9.2.2 Symmetry and Ergodicity

(r)
If a n-point probability function, Sn depends on the absolute positions,
x1 , x2 , · · · , xn , explicitely, i.e.
Z
Sn(r) = Sn(i) (x1 , x2 , · · · , xn ) = I (i) (x1 , α)I (i) (x2 , α) · · · I (i) (xn , α)p(α)dα,
S
we say that the medium is strictly statistically inhomogeneous. Examples of

statistically inhomogeneous media are shown in Fig. (9.5)
We say that a system is statically homogeneous, or when a stochastic spatial
(i)
distribution is homogeneous, if Sn (x1 , x2 , · · · , xn ) is invariant under trans-
lation, i.e. ∀ y ∈ IRd ,
Sn(i) (x1 , x2 , · · · , xn ) = Sn(i) (x1 + y, x2 + y, · · · , xn + y)

= Sn(i) (x12 , x13 , · · · , x1n ) (⇐ y = −x1 ) (9.40)
where xjk = xk − xj . Obviously, in this case, V = IRd and x1 , x2 , · · · , xn ∈

IRd .
When a system is statistically homogeneous, or when a stochastic spatial
distribution is homogeneous, one can relate ensemble (time) average to the
volume (spatial) average. This is because that material properties in every
regions of the space are similar, and hence any realization of a statistical en-
semble must contain the all statistical information or details as the rest of other
realizations do, provided that the spatial realization space is large enough to
render a stable statistical interpretation.
This suggests an ergodic hypotheis: The result of averaging over all re-
alizations of the ensemble is equivalent to averaging over the volume of one
realization in an infinite-volume limit.
Under the ergodic assumption, the complete probabilistic information can
be obtained from a single realization of an infinite domain. By letting
1
α = y, p(α) = , and dα = dVy
V
the ergodic hypothsis enables us to replace ensemble averaging with volume
averaging in the limit that the volume tends to infinity, i.e.
Z
(i) 1
Sn = lim I (i) (y)I (i) (y + x12 ) · · · I (i) (y + x1n )dy
V →∞ V V
We refer to such systems as ergodic media.
Remark 9.2.1 Ergodicity is a mathematics term, meaning “ space filling”.

Ergodic theory has its origins from the work of Boltzmann in statistical physics.
Ergodic theory in statistical mechanics refers to where time- and space-distribution
averages are equal. Steinhaus (1983, pp. 237-239) gives a practical analogy
to ergodic theory as to keeping one’s feet dry ("in most cases," "stormy weather
excepted") when walking along a shoreline without having to constantly turn
one’s head to anticipate incoming waves. The mathematical origins of ergodic
theory are due to von Neumann, Birkhoff, and Koopman
In practice, instead of using the infinite spatial space, if a domain is much

larger than a basic spatial mechanical element, we usually take it as the spatial
sampling space that is the so-called representative volume element (RVE).
One can see that for statistically homogeneous media, the n-point proba-
bility function do not depend on their absolute positions, but on their relative
(a) (b)
Figure 9.5. Examples of homogeneous isotropic (a) and homogeneous anisotropic media.
displacement. Therefore, there is no preferred origin in the system. In Eq.

(9.40), x = x1 is chosen as the origin of the coordinate.
For one-point probability function (or one-point correlation function), we
then have
Z Z Z
(r) 1 (r) 1 1
S1 := I (x, y)dVy = H(Vr )dVy = dVy = fr (9.41)
V V V V V Vr
which is the volume fraction of the phase r.
If the n-point probability function of a medium is both translation and rota-
tion invariant, the medium is called isotropic homogeneous. It means that the
n-point correlation function only depend on the distance among the particles.
For instance,
(r) (r)
S2 (x1 , x2 ) = S2 (x12 )
(r) (r)
S3 (x1 , x2 , x3 ) = S2 (x12 , x13 )
where xkj = kxj − xk k.
9.2.3 Applications
Example 9.1 Consider Voigt bound and Reuss bound,
X n Xn
fr Cr−1 ≤ C̄ ≤ fr Cr
r=1 r=1
Both these two bounds only require information of volume fraction of each
phase. Since volume fraction,
(r)
fr = S1 (x),
is, by definition, the one-point probability function (or correlation function),

both Voigt bound and Reuss bound are called as one-point bound.
Example 9.2 To evaluate Hashin-Shtrikman bound, we may let

n
X
p(x) = pr I (r) (x)
r=1
where pr is a constant second order tensor.

Then
n n n
(r)
X X X
= pr = pr S1 (x) = fr pr
r=1 r=1 r=1
Therefore,
Z Z X n
1 d 1
p : dV = pr I (r) (x) :
V V 2V V
r=1
Z h 0
i
− Γ∞ (x0 − x) : p(x − dVx0 dVx
V0
n
Z X Z
1 00
= − pr I (r)
(x) : Γ∞ (x )
2V V r=1 V 00
n
hX n i
00
X
: ps I (s) (x + x ) − fs ps dVx00 dVx
s=1 s=1
n X
n Z
1 X 00

= − pr : Γ∞ (x )dVx00
2 V 00
r=1 s=1
Z
1 (r) (s) 00 (r)

: I (x)I (x + x ) − I fs ps dVx
V V
n n Z
1 XX 00

= − pr : Γ∞ (x )dVx00
2 V 00
r=1 s=1
00

(rs)
: S2 (x, x + x ) − fr fs ps
Assume that the composite possesses no long-range interaction. The mathe-

matical implication is that
(rs) 0 (r) (s) 0 0
S2 (x, x ) = S1 (x)S1 (x ), when kx − x k >> 1
because the probability of two independent events occur simulatenously should

equal to the product of the probability of two single events.
0
One the other hand, when kx − x k ≤ Rr or Rs . There can be only one
phase exists within the region, hence
(rs) 0 (r)
S2 (x, x ) = S1 δrs
To sum up  0
 fr δrs kx − x k ≤ Rr
(rs) 0
S2 (x, x ) =
0
fr fs kx − x k > Rr

Again, we end with the relationship,

Z n n
1 1 XX r r
p : d dV = − pij Pijk` fr δrs − fr fs psk`
2V V 2
r=1 s=1
n n
1 X X
= − fr prij Pijk`
r
(δrs psk` − fs psk`
2
r=1 s=1
n
1X
= − fr prij Pijk`
r
prk` − < pk` >
2
r=1
which was derived previously by using the argument of Mori-Tanaka theorem.

As shown above, the evaluation of Hashin-Shtrikman bounds is intimately
related with the evaluation of two-point probability function, or two-point cor-
(rs)
relation function, S2 . It is this reason that Hashin-Shtrikman bounds are
called two-point bounds.
9.2.4 Ergodic principle

The intuistive concept of Ergodicity was popularized by Hugo Steinhaus.
Steinhaus wrote in his well-known book Mathematical Snapshots,
“When strolling along a sandy beach in shores most people choose the wet
strip left by retreating waves, which is hard and smooth enough to make the
walk more comfortable than the dry part of the beach. On the other hand, to
avoid their shoes and socks being soaked they must constantky watch the play
the surf licking the strip. This steady twisting of the neck becomes disagreeable
after a few minutes. There is, however, a remedy. Instead of looking sidewise
one keeps looking straight ahead; in every instant he sees the instantaneous
water edge and he directs his steps tangentially; he walks along a line touching
the edge in a single point without cutting contact lies far enough away to render
the variations small and easily accounted for: neither looking to the left, nor
sudden jumping to the right is necessary.
The background for the behavior I recommend here (after having tried it)
is the ‘ ergodic principle’: the distribution of water tongues licking the shore
in a fixed point observed during a long time is the same as the distributions
shown in a fixed moment by a long portion of the water edge — the principle
involved is the identity of time-distribution and space-distribution. To apply
it here the walker has to limit his observation to the part of the shore he will
cover in the next minute — in most cases such tactics keep him on the safe side
without leading him out of the wet strip of the beach. · · · · · · ”
I thought that some explanation may be needed to correctly understand
Steinhaus’ analogy:
What Steinhaus was trying to say is that consider an infinite set of good
weather day, if a person comes to a beach every afternoon at 2:00 clock he may
find that at a particular spot (fixed spatial location) the sea water line on the
beach is a stochastic event and all the measurement on water line on each day
consist of a statistic ensemble. We assume that there is a statistical average
value for the sea water line on that spot, which is the average in time. The
ergodic principle suggests that if a system is both homogeneous in space and
in time, one can then find that average without measuring water line at 2:00
pm on infinite days. Instead, he can just walk along a path that is tangential
to the water (shore) line on the beach, which is also assumed to “infinite”. By
doing so, the average position along his path on the beach may be equal to the
statistical average of the time ensemble.
Note that we do not consider the the surge or recede of sea water line due to
the effect of tide. Hence, the person who is in charge the measurement has to
come to the beach every afternoon at the same time (e.g. 2:00 pm), provided
that the weather is always good.
9.3 Exercises
Probelm 9.1 Show that for a spherical inlusion, Ω ⊂ V ,
Z
P := Γ∞ (y − x)dVy
Ω Z
1 ∞
= Γ̃ (ξ)dS (9.42)
4π |ξ |=1
Probelm 9.2 Consider a well-order two phase composite (K2 > K1 and
G2 > G1 ). Derive the Hashin-Shtrikman bounds for shear modulus,
f2 f1
G1 + ≤ Ḡ ≤ G2 +
1 6(K1 + 2G1 )f1 1 6(K2 + 2G2 )f2
+ +
G2 − G1 5(3K1 + 4G1 )G1 G1 − G2 5(3K2 + 4G2 )G2
(9.43)
Assume that K1 = 8GPa & G1 = 5GPa and K2 = 20.0GPa & G2 = 18GPa .
Plot the Voigt bound, Ruess bound, Mori-Tanaka, and Hashin-Shtrikman bounds
for both bulk modulus and shear modulus for comparison.
Hints:
Hashin, Z. and Shtrikman, S. [1961], “Note on a variational approach to
the theory of composite elastic materials,” The Frabklin Institute Laboratories,
pp. 336-341.
Hashin, Z. and Shtrikman, S. [1962a], “On some variational principles
in anisotropic and non-homogeneous elasticity,” Journal of Mechanics and
Physics of Solids, Vol. 10, pp. 335-342.
Hashin, Z. and Shtrikman, S. [1962b], “A variational approach to the the-
ory of the elastic behavior of polycrystals,” Journal of Mechanics and Physics
of Solids, Vol. 10, pp. 343-352.
Example 9.3 Consider a two-phase fiber reinforced composite as shown in

Figure (9.6) . Use two-dimensional Hashin-Shtrikman bounds to find the in-
plane (or transverse) bulk modulus and shear modulus.
Hints:
Hashin, Z. [1965] “On elastic behaviour of fibre reinforced materials of
arbitrary transverse phase geometry,” Journal of Mechanics and Physics of
Solids, Vol. 13, pp. 119-113.
Figure 9.6. Cylindrical fibre-reinforced composite
Torquato, S. [2002] Random Heterogeneous Materials, Springer, New York,

pp. 328-337.
Christensen, R. M. [1979],
Mechanics of Composite Materials, Chapter III;
Periodic Microstructure 251
Chapter 10
PERIODIC MICROSTRUCTURE
In engineering applications, often times, we encounter situations where ma-

terials have periodic structure. Such examples are various composites with pe-
riodic structure, reticulated structures (see Fig. (10.1), DNA, masonary struc-
tures, so forth. In fact, at very fine scale, most metals may be regarded as
composites with periodic structure because of their lattice structures. There
are mainly two types of methodologies in analysis: (1) equivalent eigenstrain
Figure 10.1. An example of periodic reticulated structure

approach, and (2) asymptotic homogenization. We first start with equivalent

eigenstrain approach.
10.1 Unit cell and Fourier series

Conisder a rectangular unit cell defined as
n o
Y := x −aj ≤ xj ≤ aj , j = 1, 2, 3 (10.1)
where aj is the half length of the unit cell in j-th direction.

For materials with periodic structures, material properties should be periodic
functions, i.e.
C(x + d) = C(x)
3
X
where d = 2mj aj ej , j = 1, 2, 3. Here mj are arbitrary integers. The
j=1
vector, d, is not the minimum periodicity, unless mj = 1.
Under certain conditions, it is possible that displacement field may be peri-
odic as well, i.e.
u(x + d) = u(x)
An immediate consequence is that strain field is periodic,
(x + d) = (x)
Nevertheless, periodic strain field does not necessarily produce periodic dis-
placement field. For instance, a constant strain field is periodic,
(x + d) = (x) = 0 , ∀d ∈ IR3 ,
but it does not generate a periodic displacement field, instead u(x) = x · 0 ,
and u(x + d) 6= u(x).
A convenient mathematical tool to treat periodic functions is Fourier series.
Define a vector,
nj π
ξ = ξj ej , and ξj = , nj = 0, ±1, ±2, · · · , · · ·
aj
and a countable set,

nj π
Λ = ξ = ξj ej ξj , nj = 0, ±1, ±2, · · · , · · · , (10.2)
aj
For any real function, f (x) ∈ C 1 (Y ), f (x) can be expanded into Fourier
series, X √
f (x) = F[f ](ξ) exp(iξ · x), i = −1, (10.3)
ξ ∈Λ
where the Fourier coefficient is

Z
1
F[f ](ξ) = u(x) exp(−ix · ξ)dVx
|Y | Y
where |Y | is the volume of the unit cell. For a rectangular unit cell, |Y | =
8a1 a2 a3 .
Recall the definition of Fourier series in an 1D interval, [−a, a],
∞
X nπ
f (x) = F[f ](ξ) exp i x , n = 0, ±1, ±2, · · · ,
n=−∞
a
Z a
1 nπ
F[f ] = f (x) exp(−i x)dx
2a −a a
and the orthonormal condition
Z a
1
exp(ixξm ) exp(−ixξn )dx = δmn
2a −a
nπ mπ
where ξn = and ξm = .
a a
Accordingly, 3D orthonormal condition is
Z
1 1 ξ=ζ
exp(ix · ξ) exp(−ix · ζ)dVx =
|Y | Y 0 ξ 6= ζ
where ξ, ζ ∈ Λ, i.e.
nj π nk π
ξ = ξj ej = ej and ζ = ζk ek = ek .
aj ak
10.1.1 Fourier transform of displacement field and strain

field
Suppose that displacement field is periodic. We may exoand displacement
field into Fourier series
X
u(x) = F[u](ξ) exp(ix · ξ) (10.4)
ξ ∈Λ
where Z
1
F[u](ξ) = u(x) exp(−ix · ξ)dVx
|Y | Y
Z
1
F[ui ](ξ) = ui (x) exp(−ix · ξ)dVx
|Y | Y
Remark 10.1.1 In literature, the following expression is often used,

X
u(x) = F[u](ξ) exp(ix · ξ)
0
ξ ∈Λ
where
0 nj π
Λ = ξ = ξj ej ξj = , j = ±1, ±2, · · · , · · ·
aj
0
Note that the difference between index set Λ and Λ is that nj 6= 0, or ξ 6= 0.
When ξ = 0, Z
1
F[u](0) = u(x)dVx
|Y | Y
which is the average displacement field.
On the other hand, if the composite undergoes a rigid body translation,
u(x) = u0 , which is not periodic, one may find that
F[u](0) = u0
Obviously, u = u0 6∈ L1 (IR) nor u = u0 ∈ L2 (IR). Convergence issue may

rise in mathematical manipulation. Anyway, rigid body translation is a trivial
physical motion, we neglect its contribution in Fourier transform by restricting
0
ξ∈Λ.
Now, we consider the Fourier transform of displacement gradient,

X
∇ ⊗ u(x) = F[∇ ⊗ u](ξ) exp(ix · ξ) (10.5)
ξ ∈Λ
and Z
1
F[∇ ⊗ u](ξ) = ∇ ⊗ u(x) exp(−ix · ξ)dVx
|Y | Y
On the other hand, from (10.4), one may find that

X
∇ ⊗ u(x) = ∇ exp(ix · ξ) ⊗ F[u](ξ) (10.6)
ξ ∈Λ
X
= i ξ ⊗ F[u](ξ) exp(ix · ξ) (10.7)
ξ ∈Λ
Comparing (10.5) with (10.7), we have
F[∇ ⊗ u](ξ) = iξ ⊗ F[u](ξ) .

Moreover, we may write Fourier series transform of strain field as

i X
(x) = ξ ⊗ F[u](ξ) + F[u](ξ) ⊗ ξ exp(ix · ξ) (10.8)
2
ξ ∈Λ
From (10.8), we can deduce that
i
F[](ξ) = ξ ⊗ F[u](ξ) + F[u](ξ) ⊗ ξ
2
Hence Z
1
F[](0) = (x)dVx = 0
|Y | Y
which implies that the average of a periodic strain field is a null field.
10.1.2 Fourier series transform of stress field

Consider a periodic elastic stiffness tensor, C(x + d) = C(x), which may
be expanded into Fourier series,
X
C(x) = F[C](ξ) exp(ix · ξ) (10.9)
ξ ∈Λ
where Z
1
F[C] = C(x) exp(−ix · ξ)dVx
|Y | Y
The corresponding stress field may then be written as
σ(x) = C(x) : (x)
   
X  X 
= F[C](ξ) exp(ix · ξ) : F[](ζ) exp(ix · ζ)
   
ξ ∈Λ ζ ∈Λ
Let η = ξ + ζ or ξ = η − ζ. We have
 
X X
σ(x) =  F[C](η − ζ) : F[](ζ) exp(ix · η)
η ∈Λ ζ ∈Λ
and it is straightforward that
X
F[σ](η) = F[C](η − ζ) : F[](ζ)
ζ ∈Λ
If C = C0 is a constant fourth order tensor,
F[C](η − ζ) = C0 , η = ζ, and F[C](η − ζ) = 0, η 6= ζ,
There is only term left,

F[σ](η) = F[C](0) : F[](ζ) = C0 : F[](η) when η = ζ.
Therefore,
X
σ(x) = C0 : F[](η) exp(ix · η)
η ∈Λ
i X 0
= C : η ⊗ F[u](η) + F[u](η) ⊗ η exp(ix · η)
2
η ∈Λ
Last, we evaluate Fourier expansion,
X
∇·σ = F[∇ · σ](ξ) exp(ix · ξ)
ξ ∈Λ
Via integration by parts,
Z
1
F[∇ · σ](ξ) = ∇ · σ(x) exp(−ix · ξ)dVbx
|Y | Y
Z n
1 o
= ∇ · σ(x) exp(−ix · ξ) − σ · ∇ exp(−ix · ξ) dVx
|Y | Y
Z
1
= n · σ(x) exp(−ix · ξ)dS
Y ∂Y
Z
+iξ σ(x) exp(−ix · ξ)dVx
Z Y
= iξ σ(x) exp(−ix · ξ)dVx
Y
because Z
n · σ(x) exp(−ix · ξ)dS = 0
∂Y
by periodicity. In particular, when ξ = 0,
Z
n · σ(x)dS = 0
∂Y
which stems from the fact that unit cell is in equilibrium.
10.2 Eigenstrain homogenization

Let CM and DM be elastic stiffness and compliance tensors in the matrix,
CΩ , Ω
D be the effective stiffness and compliance tensors in the second phase,
which is assumed to be distributed periodically in the composite. We are look-
ing for finding effective stiffness and compliance tensors, C̄ and D̄.
Consider prescribed macro-strain boundary condition,
= x · 0 , ∀x∂V
The total strain may be written as
ij = 0ij + dij , ∀ x ∈ V
The stress fields in the matrix and in the second phase are
M M
σij = Cijk` (0ij + dij ), ∀x ∈ M = Y /Ω
Ω Ω
σij = Cijk` (0ij + dij ), ∀x ∈ Ω
They satisfy the equilibrium equations,

M
σij,j = = 0, ∀ x ∈ M (10.10)
Ω
σij,j = 0, ∀ x ∈ Ω (10.11)
and continuity condition at interface,
ud+ d−
i = ui , ∀ x ∈ ∂Ω
Consider a eigenstrain field,
∗ij (x) = ∗ij (x)H(Ω)
Eshelby’s equivalent inclusion principle reads as

Ω
σij Ω
= Cijk` M
(0k` dk` ) = Cijk` (0k` + dk` − ∗k` ) (10.12)
Substituting (10.12) into (10.11) yields

M
Cijk` (0k` + dk` − ∗k` ),j = 0, ⇒ M
Cijk` M ∗
udk,`j = Cijk` k`,j (10.13)
Let,
X X
∗k` (x) = F[∗k` ](ξ) exp(iξ · x) = ˆ∗k` exp(iξ · x) (10.14)
0 0
ξ ∈Λ ξ ∈Λ
where
Z Z
1 1
ˆ∗k` = ∗k` exp(−iξ · x)dVx = ∗k` exp(−iξ · x)dVx
Y Y Y Ω
and
X
ui (x) = F[ui ](ξ) exp(ix · x) exp(iξ · x) = ûi (ξ) exp(iξ · x) (10.15)
0
ξ ∈Λ
where Z
1
ûi (ξ) = ui (x) exp(−iξ · x)dVx
|Y | Y
Note that uniform eigenstrain is excluded because it induces a divergent
displacement field, i.e.
u∗i (x) = ∗0

ij xj → ∞ as x → ∞
Substituting (10.14) and (10.15) into (10.13), we have

M M ∗
−Cijk` ûk ξ` ξj = iCijk` ˆk` ξj (10.16)
M ξ ξ and K −1 (ξ) = N (ξ)/D(ξ).

Denote Kik (ξ) = Cijk` ` j ik ik
Nik (ξ) M
F[ui ](ξ) := ûi (ξ) = −i C ∗ ξ` (10.17)
D(ξ) k`mn mn
Recall,
i X
dij (x) = ξi F[uj ](ξ) + F[udi ](ξ)ξj exp(iξ · x)
2
ξ ∈Λ0
One can write

X 1 Njk (ξ) M Nik (ξ) M ∗
dij = ξi ξ` Ck`mn + ξj ξ` C ˆ exp(iξ · x)
2 D(ξ) D(ξ) k`mn mn
ξ ∈Λ0
X
= gijmn (ξ)ˆ∗mn exp(iξ · x)
ξ ∈Λ0
Z
1 X 0 0
= gijmn (ξ) ∗mn (x ) exp(−iξ · x )dVx0 exp(iξ · x)
|Y | Y
ξ ∈Λ0
where a new fourth order tensor gijmn is defined as
1 CM ξ
`
gijmn (ξ) = ξi Njk (ξ) + ξj Nik (ξ) k`mn (10.18)
2 D(ξ)

1 h i
gijk` (ξ) = ξ (δ ξ
j i` k + δ ξ
ik ` ) + ξ (δ ξ
i j` k + δ ξ
jk ` )
2ξ 2
1 ξi ξj ξk ξ` ν ξi ξj
− + δk` (10.19)
1 − ν ξ4 1 − ν ξ2
Consider the dilute homoegenization scheme,
CΩ : (0 + d ) = CM : (0 + d − ∗ ) .
We have
0 + d = (CM − CΩ )−1 : CM : ∗
and subsequently,
0 = AΩ : ∗ (x) − d (x)
This leads to the following integral equation,
∗
0ij − AΩ ijmn mn (x)
Z
X 1 0 0
+ gijmn (ξ) ∗mn (x ) exp(i(x − x ) · ξ)dVx0 = 0(10.20)
.
0
|Y | Ω
ξ ∈Λ
Z
1
This equation is difficult to solve. Calculate the average (10.20)dVx
|Ω| Ω
in the inclusion. One has
X 1 Z
0 = AΩ : ¯∗ − g(ξ) : exp(iξ · x)dVx
|Ω| Ω
∈Λ0
1 Z
0 0
· ∗ (x ) exp(−iξ · x )dVx0
|Y | Ω
Define a scalar function,
Z
1
g0 (ξ) = exp(iξ · x)dVx (10.21)
|Ω| Ω
The eigenstrain integral equation may be written as

X
0ij − AΩ ¯∗mn +
ijmn gijmn (ξ)g0 (ξ)
ξ ∈Λ0
1 Z 0 0

· ∗mn (x ) exp(−iξ · x )dVx0 = 0 . (10.22)
|Y | Ω
For prescribed macros stress boundary condition, one may be able to show
that
X
∗
¯ij − AΩ
¯
ijmn mn + gijmn (ξ)g0 (ξ)
0
ξ ∈Λ
1 Z 0 0

· ∗mn (x ) exp(−iξ · x )dVx0 = 0 . (10.23)
|Y | Ω
M σ0 .
where ¯ij = Dijmn mn
The simplest approach to solve (10.22) is to replace ∗ (x) by its volume
average, i.e., ∗ (x) ≈ ¯∗ . Therefore,
X 1 Z 0

∗
0
= A : ¯ − Ω
g(ξ)g0 (ξ) exp(−iξ · x )dVx0 : ¯∗
|Y | Ω
ξ ∈Λ0
X
= AΩ : ¯∗ − f g0 (ξ)g0 (−ξ)g(ξ) : ¯∗
ξ ∈Λ0
X
= AΩ : ¯∗ − f G(ξ)g(ξ) : ¯∗
ξ ∈Λ0
where G(ξ) = g0 (ξ)g0 (−ξ).

Define Eshelby tensor for periodic inhomogeneities,
X
Ω
Sijmn = f G(ξ)gijmn (ξ) (10.24)
0
ξ ∈Λ
We recover the relationship between remote strain and eigenstrain ( average
eigenstrain be more precise),

0ij = AΩ
ijmn − S ijmn ¯∗mn
To this end, the homogenization of a composite with periodic microstructure

can follow the same route as the homogenization of a composite with randomly
distributed inhomogeneities, if one can find the corresponding Eshelby tensor.
The key to evaluate Eshelby tensor is to find function, G(ξ).
Example 10.1 Calculate G(ξ) for a one-dimensional periodic unit cell as

shown in Fig. (10.2).
One can show that
Z a
1
g0 (ξ) = exp(iξx)dx
2a −a
1 1 a
= exp(iξx)
2a iξ −a
1 h i
= cos(ξa) + i sin(ξa) − cos(ξa) − i sin(ξa)
2aξi
1
= sin(ξa)
aξ
It is obvious that
g0 (−ξ) = g0 (ξ)
Figure 10.2. An 1D model for a nanowire with periodic structure
Figure 10.3. Periodic distribution of spherical percipitates.
Hence
1
G(ξ) = sin2 (ξa)
a2 ξ 2
Example 10.2 In this example, we consider a spherical percipitate distri-

bution in a cubic lattice as shown in Fig. (10.3). The unit cell in this case is a
2L × 2L × 2L cubic region. There is a spherical ball with radius r = a inside
the unit cell.
Recall
J3/2 (η)
Z
exp(−iξ · x)dΩ = (2π)3/2 a3 3/2
Ω η
where
q
η = a|ξ| = a ξ12 + ξ22 + ξ32
r
n1 π 2 n2 π 2 n3 π 2
= a + +
L L L
πa πa
q
= n21 + n22 + n23 = |n|
L L
Considering,
2 1/2 r
−1 2 1
J3/2 (η) = (η sin η − cos η) = (sin η − η cos η)
πη π η 3/2
one may write

Z
1 3
exp(−iξ · x)dΩ = (sin η − η cos η)
|Ω| Ω η3
and
9 h i2
G(ξ) = sin(a|ξ|) − a|ξ| cos(a|ξ|) .
a6 |ξ|6
One may find that for bcc precipitate cluster,
3
g0 (−ξ) = (sin η − η cos η) 1 + exp(−iξ · c)
η3
and for fcc precipitate cluster,
3
g0 (−ξ) = (sin η−η cos η) 1+exp(−iξ·c 1 )+exp(−iξ·c 2 )+exp(−iξ·c 3 )
η3
as shown in Fig. (10.4)
10.3 Introduction to Asymptotic Homogenization

The asymptotic method of homogenization is a systematic tool to find effec-
tive material properties or effective coefficients of a homogenized differential
equation.
The main technique of asymptotic homogenization is the use of multiple-
scale expansion. Often times, it involves with singular purturbation technique.
(a) (b)
Figure 10.4. Cluster of peripitates in various unit cells: (a) b.c.c. cluster, and (b) f.c.c cluster .
10.3.1 One-dimensional model problem

Consider an 1D model,
d du
E = 0, 0 < x < L (10.25)
dx dx
This equation can be viewed as either the deformation of 1D elastic bar, or 1D
steady-state heat diffusion, etc.
Assume that the medium has periodic micro-structure that is varying at mi-
croscale, `, which is the characteristic length of a unit cell. Therefore, the
coefficient, E, is a periodic function of spatial variable. We also assume that
at the interface of two different media in the unit cell the following continuity
conditions hold,
h du i
[u] = 0, E =0.
dx
This 1D model problem has a very simple differential equation. An exact
solution is possible. In general, for multiple dimension problems or nonlinear
problems, analytical solutions may not be possible.
An important characteristics of this problem is the existence of two vastly
different length scales: the microscale `, which characterizes the dimension of
the unit cell, and the macroscale L, which characterizes the global variations
of external force or boundary data.
Suppose that one is more interested in the average variation over a region
which is much greater than the typical period and less interested in the detailed
variation over a local region. One may ask oneself that
Can one bypass the details to find an equation governing the variations on the global
scale L ?
`
We define a small paramter = . Obviously, << 1. To separate the
L
effect of two scales, we introduce two coordinates: a fast coordinate and a slow
coordinate, which are defined as
y and x = y (10.26)
You may suggest that the slow coordinate is slowed by small parameter, . Or
vice versa,
x
x and y = (10.27)

1
You may suggest that the fast coordinate is speed up by a large paramter .

Then, the field variable u may be expressed in a two-scale representation:
u = u(x, y) By using chain rule, we may write
d ∂ ∂
= + (10.28)
dy ∂y ∂x
or vice versa,
d ∂ 1 ∂
= + (10.29)
dx ∂x ∂y
One can then rewrite Eq. (10.25) as
d du
E(y) = 0, 0 < y < L (10.30)
dy dy
It is clear that the coefficent has to be a periodic fundtion of fast coordinate,

i.e. E = E(y).
Consider the following muti-scale expansion,
u(x, y) = u0 (x, y) + u1 (x, y) + 2 u2 (x, y) + · · · (10.31)
where ui (x, y) represents activity at i-th scale.

Applying (10.28) to (10.30) leads to the following partial differential equa-
tion,
∂ ∂ h ∂u h ∂u ∂u1 i h ∂u ∂u2 i
0 0 1
+ E(y) + + + 2 +
∂y ∂x ∂y ∂x ∂y ∂x ∂y
i
······ =0
A complete equilibrium implies that equilibrium holds in each scale,

∂ h ∂u0 i
0 : E(y) = 0;
∂y ∂y
∂ h ∂u
0 ∂u1 i ∂ 2 u0
1 : E(y) + + E(y) = 0;
∂y ∂x ∂y ∂x∂y
∂ h ∂u ∂u2 i ∂2u ∂u2 u1
1 0
2 : E(y) + + E(y) + = 0;
∂y ∂x ∂y ∂x2 ∂x∂y
······
We first solve the zero-th order equation,

∂ ∂u0
E(y) =0 (10.32)
∂y ∂y
which only involves with the lowest scale field variable, u0 (x, y).
Integrate (10.32) once,
∂u0
E(y) = A1 (x)
∂y
where A1 (x) is a integration constant.
Integrating second time, we have
Z y
dỹ
u0 (x, y) = A1 (x) + A2 (x)
y0 E(ỹ)
Since u0 (x, y) is periodic,

Z y0 +`
dỹ
u0 (x, y0 ) = u0 (x, y0 + `) ⇒ A2 (x) = A1 (x) + A2 (x)
y0 E(ỹ)
which implies that A1 (x) = 0.

This suggests that the leading-order displacement field only depends on the
macro-scale variable,
u0 = A2 (x) = u0 (x) (10.33)
Now let’s examine the first order differential equation,
∂ h ∂u
0 ∂u1 i ∂ 2 u0
E(y) + + E(y) =0 (10.34)
∂y ∂x ∂y ∂x∂y
Based on (10.33), the last term in (10.34) vanishes.
To solve (10.34), we introduce the following partial separation of variable,
∂u0
u1 (x, y) = Q(x, y) + ū1 (x)
∂x
where Q(x, y) is an unknown function.

Substitute the above expression into (10.34),

∂ ∂u
0 ∂Q ∂u0
E(y) + =
∂y ∂x ∂y ∂x

∂u0 ∂ ∂Q
E(y) 1 + =0.
∂x ∂y ∂y
This leads to the so-called inhomogeneous canonical cell problem for unknown
function, Q(x, y),

∂ ∂Q
E(y) 1 + = 0, ∀y ∈ (y0 , y0 + `) (10.35)
∂y ∂y
h ∂Q i
[Q] = 0, and E(y) 1 + = 0 , ∀x at interface. (10.36)
∂y
Integrate (10.35) once,
∂Q
E(y) 1 + = D1 (x)
∂y
∂Q D1 (x)
or = −1 +
∂y E(y)
where D1 (x) is an integration constant.
Integrate second times,
Z y
dỹ
Q(x, y) = −y + D1 (x) + D2 (x) (10.37)
y0 E(ỹ)
where D2 (x) is another integration constant.

Since Q(x, y) is y-periodic,
Q(x, y0 ) = Q(x, y0 + `)
It leads to
Z y0 +`
dỹ
−y0 + D2 (x) = −(y0 + `) + D1 (x) + D2 (x) (10.38)
y0 E(ỹ)
Eq. (10.38) is called the solvability condition for inhomogeneous problem for
Q or u1 .
We then find that
1
D1 (x) = Z y0 +` (10.39)
1 dỹ
` y0 E(ỹ)
and hence Z y
dỹ
y0 E(ỹ)
Q(x, y) = −y + Z y0 +`
+ D2 (x) (10.40)
1 dỹ
` y0 E(ỹ)
Therefore,
y
 Z 
dỹ

y E(ỹ)  ∂u
 0
u1 (x, y) = −y + Z y00 +` + D2 (x) + ū1 (x)(10.41)

 1 dỹ  ∂x
` y0 E(ỹ)
∂u1 ∂u0 1 ∂u0
= − + 1 Z y0 +` dỹ ∂x (10.42)
∂y ∂x
E(y)
` y0 E(ỹ)
Next, we consider the differential equation at the second scale,
∂ h ∂u ∂u2 i ∂2u ∂u2 u1
1 0
2 : E(y) + + E(y) + =0. (10.43)
∂y ∂x ∂y ∂x2 ∂x∂y
Consider
∂ 2 u1 ∂ 2 u0 1 ∂ 2 u0
=− 2 + 1 y0 +` dỹ ∂x2
∂x∂y ∂x
Z
E(y)
` y0 E(ỹ)
Eq. (10.43) becomes
∂ h ∂u
1 ∂u2 i 1 ∂ 2 u0
E(y) + + Z y0 +` =0 (10.44)
|
∂y ∂x
{z
∂y
} 1 dỹ ∂x2
f unction of y ` y0 E(ỹ)
| {z }
f unction of x
Hence,
1 ∂ 2 u0
1 y0 +` dỹ ∂x2 = 0
Z
` y0 E(ỹ)
or  

 

 
∂  1 ∂u0 
∂x  1 Z y0 +` dỹ ∂x  = 0 . (10.45)

 

` y0 E(ỹ)
 
Figure 10.5. One-dimensional unit cell
This is the homoegenized differential equation that governs the macroscale

variation of the mean displacement field.
Compare the mean-field differential equation to the original differential equa-
tion,
d du
E(y) =0
dy dy
We conclude that the effective coefficient for the homogenized differential
equation is
1 D 1 E−1
Ee = Z y0 +` = (10.46)
1 dỹ E
` y0 E(ỹ)
which is the harmonic mean of E(y), or the estimate from Reuss bound.
Consider the unit cell shown in Fig. (10.5). One may find that
y0 +` Z 1 − f`
1 f ` dt
Z Z
1 dt 2 2 dt
= +
` y0 E(t) ` 0 E1 ` 0 E2
` − f` 1 f (1 − f )E2 + f E1
= + =
` E1 E2 E1 E2
and
1 E1 E2
Ee = y0 +`
= (10.47)
(1 − f )E2 + f E1
Z
1 dt
` y0 E(t)
The homogenized differential equation is,
d du0
Ee =0. (10.48)
dx dx
To sum up, asymptotic homogenization consists of the following steps:
Summary of Asymptotic Homogenization
1 The objective of the homogenization is to find the average coefficients of

the homogenized differential equation and find its solution;
2 Identify the micro- and macroscales;
3 Introduce multiple-scale variables and expansions, and deduce cell boundary-
value problems (BVPs) at successive orders. The leading-order cell prob-
lem is homogeneous, i.e. u0 = u0 (x);
4 Use linearity (or separation of variables) to express the next-order solu-
tion in terms of the leading-order solution and deduce an inhomogeneous
canonical cell BVP;
5 Require the solvability of the inhomogeneous cell problem;
6 Find the differential equation that governs the macro-scale variation of the
mean displacement or the evolution of the leading-order solution which
includes the constitutive coeffocoients of the differential equation.
10.3.2 A multiple dimension example

Consider a 3D example,
A u = f, ∀x ∈ Ω (10.49)
u = 0, ∀x ∈ ∂Ω (10.50)
where
∂ x ∂
A =− aij ( )
∂xi ∂xj
where x = (x1 , x2 , x3 ).
Define the fast coordinate,
x
y=

1
as if y is speed-up by the large paramter . We then can express the field

variable as a function of two independent scales, u (x) = u(x, y).
From chain rule, we have
∂ ∂ ∂ ∂yi ∂ 1 ∂
= + = +
∂xi ∂xi ∂yi ∂xi ∂xi ∂yi
We can then expand the differential operator, A , as
Figure 10.6. Illustration of multiscale phenomena
∂ 1 ∂ h ∂ 1 ∂ i
A = − + aij (y) +
∂xi ∂yi ∂xi ∂yi
h ∂ ∂ i h ∂ ∂ ∂ ∂ i
= −−2 aij − −1 aij (y) + aij (y)
∂xi ∂yj ∂xi ∂yj ∂yi ∂xi
h ∂ ∂ i
−0 aij (y)
∂xi ∂xi
= −2 A1 + −1 A2 + 0 A3 (10.51)
where
h ∂ ∂ i
A1 = − aij
∂xi ∂yj
h ∂ ∂ ∂ ∂ i
A2 = − aij (y) + aij (y)
∂xi ∂yj ∂yi ∂xi
h ∂ ∂ i
A3 = − aij (y)
∂xi ∂xi
Now we consider multiple scale expansion,
u (x) = u0 (x, y) + u1 (x, y) + 2 u2 (x, y) + · · · (10.52)
which decomposes or separates the activities at different scales.

Substituting both (10.52) and (10.51) into (10.49), we have

−2 A1 + −1 A2 + 0 A3 u0 + u1 + 2 u2 + · · · = f

−2 A1 u0 + −1 A1 u1 + A2 u0 + 0 (A1 u2 + A2 u1 + A3 u0 )
+··· = f (10.53)
The total state equilibrium is equivalent to equilibrium states in each every

scale. That is
−2 : A1 u0 = 0; (10.54)
−1 : A1 u1 + A2 u0 = 0; (10.55)
0 : A1 u 2 + A2 u 1 + A3 u 0 = f (10.56)
······
If one can solve differential equations at each scale, one can find out both local
detailed information as well as global information.
As far as homogenization concern, we are looking for a homogenized dif-
ferential equation that carries the overall information of fine scale.
Before we proceed further, we prove the following lemma.
Lemma 10.3 If the differential equation,
A1 u = F, ∀y ∈ Y
has a unique Y -periodic solution, the following equation holds
Z
1
< F >= F (y)dVy = 0 (10.57)
|Y | Y
where y = (y1 , y2 , y3 ).
Proof:
By the assumption, one can assume that both u and F are Y-periodic, and
X
F (y) = F[F ](ξ) exp(iξy) (10.58)
ξ∈Λ
X
u(x, y) = F[u](ξ) exp(iξy) (10.59)
ξ∈Λ
Hence
∂ ∂
A1 u = − aij (y) u
∂yi ∂yj
X ∂aij
= − ξj + iξi F[u](ξ) exp(iξy)
∂yi
ξ∈Λ
Based on A1 u = F , one has

X ∂aij X
− iξj + iξi F[u] exp(iξy) = F[F ](ξ) exp(iξy)
∂yi
ξ∈Λ ξ∈Λ
∂a
ij
⇒ F[F ](ξ) = −iξj + iξi
∂yi
Therefore,
Z
1
F[F ] = 0 ⇒ F[F ](0) = F dVy = 0 .
|Y | Y
♣
To this end, we start to solve differential equations at each scale. At scale
−2 , we have
∂ ∂
A1 u 0 = − aij (y) u0 = 0
∂yi ∂yj
We claim that
u0 = u0 (x) .
That is the leading-order expansion is only the function of slow scale variable.
Since u0 is Y-peridoic, we have
X
u0 = F[u0 ](ξ) exp(iξy) .
ξ∈Λ
Consequently,
∂a
ij
X
A1 u 0 = 0 ⇒ − iξj + iξi F[u](ξ) exp(iξy) = 0 .
∂yi
ξ∈Λ
Then for ξ 6= 0, it is necessary
F[u](ξ) = 0 . (10.60)
Assume that
u0 = c(x)Q(y) + ū0 (x)
Eq. (10.60) becomes
Z
1
F[u](ξ) = c(x)Q(y) + ū0 (x) exp(−iξy)dVy
|Y | Y
Z
1
= c(x)Q(y) exp(−iξy)dVy = 0 (10.61)
|Y | Y
Z
because ū0 (x) exp(−iξy)dVy = 0 when ξ 6= 0.
Y
The only possibility that (10.61) holds is that Q(y) = 1 or Q(y) = 0. In
either case, u0 = u0 (x). We proved our claim.
Next, we consider the differential equation at scale −1 :
A1 u 1 + A2 u 0 = 0 .
One can show that

h ∂ ∂ ∂ ∂ i ∂aij ∂u0
A2 u 0 = − aij (y) + aij (y) u0 (x) = −
∂xi ∂yj ∂yi ∂xj ∂yi ∂xj
Hence
∂aij ∂u0
A1 u 1 = (10.62)
∂yi ∂xj
This suggets the following separation of variable,
∂u0
u1 (x, y) = Uk (y) + ū1 (x) (10.63)
∂xk
and subsequently,
∂u
0
A1 u 1 = A1 Uk (y)
∂xk
∂ ∂U ∂u
k 0
= − aij (y) (10.64)
∂yi ∂yj ∂xk
Combining (10.62) and (10.64), we find the canonical equation for a unit cell
problem,
∂aik ∂ ∂U
k
+ aij (y) =0. (10.65)
∂xi ∂yi ∂yj
with the possible boundary conditions at interface of different phases,
h i h ∂Uk i
Uk = 0, and aik + aij ni = 0 (10.66)
∂xj
We now consider the differential equation at 0 scale,
A1 u 2 + A2 u 1 + A3 u 0 = f
which can be rewritten as

A1 u 2 = f − A2 u 1 + A3 u 0 (10.67)
The condition that equation (10.67) has a unique periodic solution is that
< f − (A2 u1 + A3 u0 ) >= 0
That is Z
1
A2 u1 + A3 u0 dy = f (10.68)
|Y | Y
Consider
u0 = u0 (x)
∂u0
u1 = Uj + ū1 (x)
yj
One can show that
∂ 2 u0
A3 u0 = −aij (10.69)
∂xi ∂xj
h ∂ ∂ ∂ ∂ i ∂u0
A2 u1 = − aij (y) + aij (y) Uk (y) + ū1
∂xi ∂yj ∂yi ∂xj ∂xk
∂Uk ∂ u0 2 ∂ 2
∂ u0
= −aij − aij (y)Uk (y)
∂yj ∂xi ∂xk ∂yi ∂xj ∂xk
∂ ∂ ū
1
− aij (y) (10.70)
∂yi ∂xj
Change the dummy indices j ↔ k in the first term of (10.70). We can write
that
∂Uj ∂ 2 u0 ∂ ∂ 2 u0
A2 u1 + A3 u0 = − aij + aik − aij (y)Uk (y)
∂xk ∂xi ∂xj ∂yi ∂xj ∂xk
∂ ∂ ū1
− aij (y)
∂yi ∂xj
Via divergence theorem,
∂ 2 u0
Z Z
1 1 ∂Uj
(A2 u1 + A3 u0 )dy = − aij + aik dVy
|Y | Y |Y | Y ∂xk ∂xi ∂xj
h i h i
− (aij (y)Uk (y)u0,jk (x) ni − aij (y)ū1,j ni
By periodicity, the boundary terms will vanish. We then have

∂ 2 u0
Z
1 ∂Uj
− aij + aik dVy =f
|Y | Y ∂xk ∂xi ∂xj
Denote the effective coefficients as
Z
1 ∂Uj
āij = aij + aik dVy (10.71)
|Y | Y ∂xk
and homogenized differential operator
∂ ∂
AH = − āij (10.72)
∂xi ∂xj
Figure 10.7. An unilateral composite with periodic straucture.
we finally derived the homoegenized boundary-value problem,
AH u0 = 0, ∀x ∈ Ω (10.73)
u0 = 0, ∀x ∈ ∂Ω (10.74)
Example 10.4 Consider a 2D steady-state heat transfer problem (see Fig.

(10.7)),
∂ x ∂T
1
λαβ = 0, ∀x ∈ D (10.75)
∂xα ∂xβ
where T (x) is temperature field
n and λαβ are heat conduction coefficients.oWe
assume that the region D = (x1 , x2 ) 0 ≤ x1 ≤ `1 , and 0 ≤ x2 ≤ `2 is
thermally insulated in horizontal boundaries, i.e.
∂T
q2 = λ2β = 0, ∀x2 = 0, and x2 = `2 (10.76)
∂xβ
Along the vertical boundaries of the region D, the heat flows are prescribed,
∂T
q1 = λ1β = ∓q0 , ∀x1 = 0, and x1 = `1 (10.77)
∂xβ
Consider multiple expansion,
T (x) = T0 (x) + T1 (x, y1 ) + · · ·
and the following separation of variable,
∂T0 (x)
T1 (x, y1 ) = Uα (y1 ) , α = 1, 2
∂xα
Note that first we assume that the mean temperature at this scale is zero, i.e.
T̄1 (x) = 0; and Uα (y1 ) are Y-periodic functions that are the following 1D
canonical cell problem,
d dUα (y1 ) dλ1α

− λ11 (y1 ) = , ∀ y1 ∈ Y (10.78)
dy1 dy1 dy1
h i h dUα i
Uα = 0, and λ11 = 0, ∀y1 at interface. (10.79)
dy1
Integrate (10.78),
dUα (y1 )
−λ11 (y1 ) = λ1α (y1 ) − Cα
dy1
dUα (y1 ) λ1α (y1 ) Cα
⇒ =− +
dy1 λ11 (y1 ) λ11 (y1 )
where Cα are constants (note that they are not functions of x ! ).

Integrate second time,
Z y1 Z y1
Uα (y1 ) = − λ1α (ξ)λ−1
11 (ξ)dξ + Cα λ−1
11 (ξ)dξ + Dα
0 0
Note that we choose Dα = 0, because the average temperature at scale −1 is

assumed to be zero.
The solvability condition of the canonical cell problem requires Uα (y1 ) as
a Y-periodic function, i.e.
Uα (0) = Uα (`)
This condition allows us to determine the constants Cα ,

Z `
λ1α (ξ)λ−1
11 (ξ)dξ
0
Cα = R ` −1 (10.80)
0 λ11 (ξ)dξ
In specific,
Z ` −1
C1 = λ−1
11 (ξ)dξ
0
Z `
λ12 (ξ)λ−1
11 (ξ)dξ
0
C2 = Z `
λ−1
11 (ξ)dξ
0
Consequently, we find the closed form solution for canonical cell problem,
Z y1
λ−1
11 (ξ)dξ
U1 (y1 ) = −y1 + Z0 ` (10.81)
λ−1
11 (ξ)dξ
0
Z y1
U2 (y1 ) = − λ12 (ξ)λ−1
11 (ξ)dξ
0
Z `
λ12 (ξ)λ−1
11 (ξ)dξ Z y1
0 −1
+ Z ` λ 11 (ξ)dξ (10.82)
0
λ−1
11 (ξ)dξ
0
Define the effective heat conduction coefficients,
Z
1 ∂Uj
λ̄ij := aij + aik dy .
|Y | Y ∂xk
It is easy to find that
1 `
Z
∂U1
λ̄11 = λ11 (ξ) + λ11 (ξ) (ξ) dξ
` 0 ∂y1
1 `
Z 1
= λ11 − λ11 + C1 dy = C1
` 0 `
Z `
1 −1
= ( λ−1 (ξ)dξ
` 0 11
and
1 `
Z
∂U2
λ̄12 = λ12 (ξ) + λ11 (ξ) (ξ) dξ
` 0 ∂y1
1 `
Z
= λ12 (ξ) − λ12 (ξ) + C2 dξ
` 0
1 `
Z
λ12 (ξ)λ−1
11 (ξ)dξ
` 0
= = λ̄21
1 ` −1
Z
λ (ξ)dξ
` 0 11
and
1 `
Z
∂U2
λ̄22 = λ22 (ξ) + λ21 (ξ) (ξ) dξ (10.83)
` 0 ∂y1
1 `
Z
= λ22 (ξ) − λ212 λ−1
11 (ξ) + C 2 λ 12 λ −1
11 (ξ) dξ (10.84)
` 0
1 ` 1 `
Z Z
= λ22 (ξ)dξ − λ12 (ξ)λ−1
11 (ξ)dξ
` 0 ` 0
Z ` 2
λ12 (ξ)λ−1
11 (ξ)dξ
1 0
+ Z ` (10.85)
`
λ−1
11 (ξ)dξ
0
and the homogenized partial differential equation becomes

∂ 2 T0 ∂ 2 T0 ∂ 2 T0
λ̄11 + 2 λ̄ 12 + λ̄ 22 =0.
∂x21 ∂x1 ∂x2 ∂x22
10.4 Variational Characterization

Recall the homogenization of conduction problem,
A u = f, ∀x ∈ Ω
u = 0, ∀x ∈ ∂Ω
Assume that
∂u(0) (x)
u(1) (x, y) = Uk (y) (10.86)
∂xk
One can derive the following governing equations for the canonical cell prob-
lem,
∂ ∂Uj
akj + ak` = 0, ∀y ∈ Y (10.87)
∂yk ∂y`
with the proper interface and periodic conditions.

Subsequently, one can derive the effective coefficients for homogenized dif-
ferential equation,
Z Z
1 ∂Uj 1 ∂Uj
āij = aij + ai` dy = ai` δ`j + dy (10.88)
|Y | Y ∂y` |Y | Y ∂y`
Based on (10.87), one may find that
Z
1 ∂Uj
− akj (y) + ak` (y) Ui (y)dy = 0
|Y | Y ∂y`
Z Z
1 ∂Uj 1 ∂Uj ∂Ui
− akj + ak` Uj (y)nk dS + akj + ak` dy
|Y | ∂Y ∂y` |Y | Y ∂y` ∂yk
Z
1 ∂Uj ∂Ui
= ak` δ`j + dy = 0 (10.89)
|Y | Y ∂y` ∂yk
Adding (10.89) to (10.88), one may find that
Z
1 ∂Uj ∂Ui
āij = δ`j + ai` (y) + ak` (y) dy
|Y | Y ∂y` ∂yk
Z
1 ∂Ui ∂Uj
= ak` (y) δik + δ`j + dy (10.90)
|Y | Y ∂yk ∂y`
Eq. (10.90) links the effective coefficients of the homogenized equation with
the variational characters of unit cell problem, which plays a significant role in
Tartar’s variational principle.
Consider constant vector, ξ = ξi ei , or a flux vector of macro-scale variable.
We can form the following quadratic form,
Z
1 ∂Ui ∂Uj
āij ξi ξj = ξi ξj ak` (y) δik + δ`j + dy
|Y | Y ∂yk ∂y`
Z
1 ∂Ui ξi ∂Uj ξj
= ak` (y) ξk + ξ` + dy (10.91)
|Y | Y ∂yk ∂y`
Eq. (10.91) suggests that there exists a functional,
Z
1 ∂Uk ξk ∂U` ξ`
J(U) = aij (y) ξi + ξj + dy (10.92)
Y Y ∂yi ∂yj
such that
āij ξi ξj = min J U (10.93)
1 (Y )
U∈H#
where the function space H# 1 (Y ) 1 is defined as H 1 (Y ) space of Y-periodic
functions, i.e.
n o
1
H# (Y ) := u u is Y − periodic, and u ∈ H 1 (Y )
that is Z
(u2 + |∇u|2 )dy < +∞
Y
To show this, we first show that the Euler-Lagrange equation of J(U) is the
governing equation of canonical cell problem.
Assume that aij is symmetric and real. It subsequently implies that aij is
positive definite. Therefore,
Z
1 ∂δU ξ
k k ∂U` ξ` ∂Uk ξk ∂δU` ξ`
δJ = aij (y) ξj + + ξi + dy
|Y | Y ∂yi ∂yj ∂yi ∂yj
Z
2 ∂Uk ξk
= aij (y) ξi + δU` ξ` dS
|Y | ∂Y ∂yi
Z
2 ∂ ∂Uk ξk
− aij (y) ξi + δU` ξ` dy = 0
|Y | Y ∂yi ∂yi
By periodic conditions
Z
2 ∂Uk ξk
aij (y) ξi + δU` ξ` dS = 0,
|Y | ∂Y ∂yi
it then leads to
Z
2 ∂ ∂Uk
δJ = − aij (y) δik + δU` ξk ξ` dy = 0
|Y | Y ∂yi ∂yi
and hence
∂ ∂Uk
− aij (y) δik + δU` = 0 .
∂yi ∂yi
Consider Uk = 0 ∈ H# 1 (Y ). One can find an upper bound for effective
coefficient, āij , i.e.

1 Z
0 < āij ξi ξj ≤ aij (y)dy ξi ξj (10.94)
|Y | Y
or Z
1
āij ≤ aij (y)dy (10.95)
|Y | Y
1 In music, the sign # is used to indicate that a note is to be raised by a half tone. Similar meaning implies
here as well, i.e. a “half level higher” H 1 space.

This is the arithmetic mean or the so-called Voigt bound.

To find the lower bound, we have to enlarge the space H# 1 (Y ). Consider
function ζi ∈ L2# (Y ) and the mean value of ζi is zero, i.e.

Z
ζi (y)dy = 0 .
Y
It is obvious that
āij ξi ξj ≥ min Jc (ζ) (10.96)
ζ ∈L2# (Y ) and
Y ζ (y)dy=0
R
where
Z
1
Jc (ζ) := aij (ξi + ζi (y))(ξj + ζj (y))dy
|Y | Y
Z
−2Ck ζk (y)dy − 0 (10.97)
Y
where Ck are Lagrange multipliers.

To find the minimizer in L2# (Y ), we calculate the first variation of the func-
tional, Jc (ζ),
Z Z
2 1
δJc = aij (y)(ξi + ζi )δζj dy − 2δCj ζj (y)dy
|Y | Y |Y | Y
Z
1
−2Cj δζj (y)dy
|Y |
Z Y Z
2 1
= aij (y)(ξi + ζi ) − Cj δζj dy − 2δCj ζj (y)dy = 0
|Y | Y |Y | Y
which yields Euler-Lagrangian equation and the constrain condition,
aij (ξj + ζj ) = Ci (10.98)

Z
ζj (y)dy = 0 . (10.99)
Y
Solving (10.98), we have

ξi + ζi = a−1
ij Cj (10.100)
Average the above expression over the unit cell and considering the constraint
condition (10.99),
ξi =< a−1ij (y) > Cj (10.101)
which solves Cj in terms of ξi , i.e.
Cj =< a−1
ji (y) >
−1
ξi (10.102)
The minimizer in L2# (Y ) under the constraint is then

Z
1
min Jc (ζ) = aij (ξi + ζi )(ξj + ζj )dy
ζ ∈L2# (Y ) and |Y | Y
Y ζ (y)dy=0
R
Z
1
= Cj (ξi + ζi )dy = Cj ξj
|Y | Y
= < a−1 −1
ji >Y ξi ξj
1 Z −1
= a−1
ij (y)dy ξi ξj
|Y | Y
From the above estimate, we find a lower bound for effective coefficient, āij ,
i.e.
1 Z −1
āij ≥ a−1 (y)dy . (10.103)
|Y | Y ij
which is the so-called Reuss bound.
10.5 Multiscale Finite Element Method

10.5.1 Asymptotic homogenization of linear elasticity
Consider a composite material with periodic structure and its elastic stiff-
ness tensor satisfies the relation,
x
Cijk` ξij ξk` = Cijk` (y)ξij ξk` ≥ αξij ξij

where α > 0.
Consider the following boundary value problem,

∂σij
+ fi = 0, ∀x ∈ Ω (10.104)
∂xj

σij = Cijk`
uk,` = Cijk` ek` (10.105)
1 ∂u ∂u`
ek` = + (10.106)
2 ∂x` ∂xk

σij nj = t0i , ∀x ∈ Γt (10.107)
ui = ūi , ∀x ∈ Γu (10.108)
Consider multiple scale expansion,

(0) (1) (2) x
ui (x) = ui (x, y) + ui (x, y) + 2 ui (x, y) + · · · , y :=

Hence
∂ ∂ 1 ∂ 0
uk,` = uk = + uk + u1k + 2 u2k + · · ·
∂x` ∂x` ∂y`
= −1 eY k` (u(0) ) + 0 (eXk` (u(0) + eY k` (u(1) )) +
+1 (eXk` (u(1) ) + eY k` (u2 )) + · · · (10.109)
and

σij (x, y) = Cijk` (y)uk,`
h
(0) (1) (2)
= Cijk` (y) −1 uY k,` + 0 (u0Xk,` + u1Y k,` ) + (uXk,` + uY k,` )
i
+···
(0) (1) (2)
= −1 σij + 0 σij + 1 σij + · · · (10.110)
In each scale, the constitutive relations are

(0) (0)
−1 : σij = Cijk` (y)uY k,` ;
(1) (0) (1)
0 : σij = Cijk` (y)(uXk,` + uY k,` );
(1) (1) (2)
1 : σij = Cijk` (y)(uXk,` + uY k,` );
···
To derive equilibrium equation at different scales, one may write

∂σij ∂ 1 ∂
= + σ + fi = 0
∂xj ∂xj ∂yj ij
∂ 1 ∂ −1 (0)
(1) (2)
= + σij + 0 σij + 1 σij + · · · + fi = 0
∂xj ∂yj
Consequently,
(0)
−2
∂σij
: = 0; (10.111)
∂yj
(0) (1)
−1
∂σij ∂σij
: + = 0; (10.112)
∂xj ∂yj
(1) (2)
∂σij ∂σij
0 : + + fi = 0; (10.113)
∂xj ∂yj
(s) (s+1)
∂σij ∂σij
s−1 : + = 0; s = 2, 3, · · · (10.114)
∂xj ∂yj
and the boundary conditions are

(0)
−1 σij + 0 σij
1
+ 1 σij
2
+ · · · nj = t0i , ∀x ∈ Γt (10.115)

(0) (1) (2)
ui + 1 ui + 2 ui + · · · = 0, ∀x ∈ Γu (10.116)
The boundary conditions in different scale are

(0)
−1 : σij nj = 0;
(1)
0 : σij nj = t0i ;
(2)
∀x ∈ Γt (10.117)
1 : σij nj = 0;
······
and
(0)
0 : ui = ūi ;
(1)
1 : ui = 0; ∀x ∈ Γu (10.118)
(2)
2 : ui = 0;
······
We first examine the leading order equilibrium equation and boundary con-
dition,
(0)
∂σij
=0
∂yj
This yields
(0) (0)
σij = σij (x)
On the other hand
(0)
(0) ∂u
σij = Cijk` (y) k
∂y`
To commodate both conditions, we have to set
(0)
σij = 0 . (10.119)
and
(0) (0)
ui = ui (x) (10.120)
To solve the second order boundary-value problem, the follwing separation
of variable is adopted
(0)
(1) k` ∂uk (1)
ui (x, y) = χi (y) (x) + ūi (x) (10.121)
∂x`
where the unknown vector function, χk` i (y)ei , is often referred to as the char-
acteristic displacement field. We further assume that
(1) k` ∂u0k
σij (x, y) = σ̂ij (y) (x) (10.122)
∂x`
Consider
(1) (0)
∂ui ∂χk`
i ∂uk
= (10.123)
∂yj ∂yj ∂x`
and

(1) (0) (1) (0) (1)
σij = Cijk` uXk,` + uY k,` = Cijk` eXk` + uY k,` . (10.124)
We find that
(1)

mn ∂χmn
k

(0)
σij = Cijk` Tk` + uXm,n (10.125)
∂y`
mn = 1 δ

where Tk` mn (0) (0)
km δ`n + δkn δ`m , because Tk` uXm,n = eXk` .
2
Accordingly,
mn

mn ∂χmn
k

σ̂ij = Cijk` Tk` +
∂y`
Then the equilibrium equation on second scale (−1 ) provides the governing
equation for the canonical cell problem,
(1) mn (0) mn
∂σij ∂ σ̂ij ∂um ∂ σ̂ij
= 0, ⇒ = 0, ⇒ =0. (10.126)
∂yj ∂yj ∂xn ∂yj
More explicitely, the governing equation for canonical cell problem is
∂ h
mn ∂χmn
k
i
Cijk` Tk` + = 0, ∀y ∈ Y
∂yj ∂y` (10.127)
The related interface continuity conditions and periodic conditions are omited
here.
Consider the equilibrium equation at third scale (0 ). We have
(2) (1)
∂σij ∂σij
= − fi + = Fi , ∀ y ∈ Y
∂yj ∂xj
The Fredholm alternative condition requires that
Z
1
Fi (y)dy = 0 .
|Y | Y
This can be shown from the fact that

(2)
∂σij
Z Z
1 1 (2)
dy = σ nj dS = 0 .
|Y | Y ∂yj |Y | ∂Y ij
Thereby,
(1)
∂σij
Z
1 ∂ (1)
fi + dy = 0, ⇒ fi + < σij >Y = 0 .
|Y | Y ∂xj ∂xj
where
(0) (0)
(1) k` ∂uk h ∂uk
< σij >Y =< σ̂ij (y) >Y = Cijk` (10.128)
∂x` ∂x`
and the homegenized elastic stiffness tensor is determined by the solution of
the canonical cell problem,
∂χk`
Z
h 1 h
k` m
i D E
k`
Cijk` = Cijmn (y) Tmn + dy = σ̂ij . (10.129)
|Y | Y ∂y` Y
The homogenized BVP is,

< σij >,j +fi = 0, ∀x ∈ Ω (10.130)
< σij > nj = ti , ∀x ∈ Γt (10.131)
(0)
ui = ūi , ∀x ∈ Γu (10.132)
10.5.2 Finite element formulation

Choose vi ∈ H# 1 (Y ). Multipling v with the leading order equilibrium
i
equation (10.111) and integrating it over Y, we have
(0)
∂σij
Z
1
vi dΩy = 0, ∀vi ∈ H# (Y )
Y ∂yj
Integration by parts yields,
Z Z
(0) (0) ∂vi
σij nj vi dS − σij dVy
Y Y ∂yj
(0)
∂uk ∂vi
Z Z
(0) ∂vi
= − σij dVy = − Cijk` dΩy = 0 .
Y ∂yj Y ∂y` ∂yj
(0)
Let vi (x, y) = ui (x, y). We have
(0) (0)
∂uk ∂ui
Z
Cijk` dΩy = 0 . (10.133)
Y ∂y` ∂yj
Since Cijk` (y) is positive definite,

(0)
∂ui (0) (0)
= 0 , ⇒ ui = ui (x)
∂yj
(0)
and consequently σij = 0, as we have derived before.
Multiply Eq. (10.127) with a test function, vi ∈ H# 1 (Y ), and integrate them
over Y. Integration by parts yields,

∂χmn
Z
∂ h
mn k
i
Cijk` Tk` + vi dVy
Y ∂yj ∂x`
∂χmn ∂χmn
Z h i Z ∂v
mn k mn k i
= Cijk` Tk` + nj vi dSy − Cijk` Tk` + dVy
∂Y ∂x` Y ∂x` ∂yj
∂χmn
Z ∂v
mn k i
= − Cijk` Tk` + dVy = 0 .
Y ∂x ` ∂y j
Consider the following parametric vector,
Pmn = ym δnk ek = Pkmn ek (10.134)
One can show that
mn 1 ∂P`mn ∂Pkmn mn
Tk` = + = P(k,`)
2 ∂yk ∂y`
Therefore, the weak formulation for the canonical cell problem can be written
as Z
1
mn

Cijk` (y) P(k,`) + χmn
(k,`) v(i,j) dVy = 0 . (10.135)
|Y | Y
Define the bilinear form
Z
1
aY (u, v) = Cijk` (y)u(i,j) v(k,`) dVy (10.136)
|Y | Y
The finite element formulation of canonical cell problem is:

Find χmn ∈ H# 1 (Y ), such that
aY (Pmn + χmn , v) = 0, ∀ v ∈ H#
1
(Y ) (10.137)
Once χmn
k,` being determined, the effective elastic stiffness tensor can then
be calculated based on definition
Z
H 1 k`
Cstk` = Cstmn (y)(Pm,n + χk`
m,n (y))dVy (10.138)
|Y | Y

ij ij 1
Tst = P(s,t) = δsi δtj + δsj δti
2
It is readily to show that
H ij H 1
H
Cstk` Tst = Cstk` δsi δtj + δsj δti == Cijk` (10.139)
2
and
Z
H 1
k` ij
Cijk` = Cstmn Pm,n + χk`
m,n )Tst dVy
|Y | Y
Z
1
k` ij
= Cstmn Pm,n + χk`
m,n )P(s,t) dVy
|Y | Y

= aY (Pk` + χk` , Pij dVy (10.140)
Finally, we define another function space,

VΩ = v(x), x ∈ Ω v(x) ∈ [H 1 (Ω)]d , d = dim{Ω}, and , v(x) =0
Γu
The weak formulation for the following macro-level BVP,

(1)
∂ < σij >Y
+ fi = 0, (10.141)
∂xj
(1)
H (0)
where < σij > = Cijk` u(k,`) (10.142)
∂ h
(0)
i
and CH u + fi = 0, ∀x ∈ Ω (10.143)
∂xj ijk` (k,`)
(1)
< σij > nj = t0i , ∀ x ∈ Γt (10.144)
(0)
ui = ūi , ∀ x ∈ Γu (10.145)
is:
Find u(0) (x) ∈ VΩ such that
Z Z Z
H (0)
Cijk` u(k,`) v(i,j) dVx = fi vi dVx + t0i vi dS, ∀ v ∈ VΩ . (10.146)
Ω Ω Γt
where v = vi ei .
Summary of Multiscale Finite Element Method

1 Solve the canonical cell problem on Y first, i.e. find χk` (y) ∈ H#
1 (Y ) by
solving
aY Pk` + χk` , v = 0, ∀v ∈ H# 1
(Y )
2 Calculate macro-scale elastic stiffness tensor

Z
H

ij ij k`
1
Cijk` = aY P +χ , P and aY (u, v) := Cijk` (y)ui,j vk,` dVy
Y Y
3 Solve the macro displacement field, u(0) (x) ∈ VΩ ,

D E D E Z
aH
Ω (u(0)
, v) = f , v + t 0
, v where aH
Ω (u, v) := H
Cijk` ui,j vk,` dVy
Ω Γt Ω
where v is any function in VΩ ;
4 Calculate the fine (local) scale stress distribution,
∂u(0)
(1) mn mn m
σij (x, y) = Cijk` (y) Tk` + χ(k,`) (y)
∂xn
10.6 G-, H-, and Γ- convergence
Various notions of convergence are introduced in relation to asymptotic ho-
mogenization theory, such as Γ-convergence of De Giorgi [1975][1984], the
G-convergence of Spagnolo [1968][1976], and the H-convergence of Tartar
[1978]. These abstract mathematical notions provide powerful tools to analy-
sis various numerical simulations of homogenization.
The question we would like to answer is: what is the limit in a homogeniza-
tion process when micro-scale approaches to zero (Fig. (10.8) ? does upscale
homogenizations will eventually converge to that limit ?
To answer this questions, we have to first define what do we mean by con-
vergence, or convergence in what sense.
10.6.1 Strong convergence and weak convergence

We first discuss the notion of strong convergence and weak convergence of
functions in Banach spaces.
Let Ω be an open set in IRd . For 1 ≤ p ≤ +∞, the Lebesgue space Lp (Ω) of
all measurable functions u in Ω is a Banach space endowed with the following
norm,
Z 1/p
kukLp (Ω) = |u|p dx , ∀1 ≤ p < +∞
Ω
When p = ∞, we define the so-called essential supremum

kukL∞ (Ω) = ess sup |u(x)| := inf sup |u(x)|
x∈Ω Z∈Ω x∈Ω−Z
µ(Z)=0
Figure 10.8. Notion of convergence in homogenization
Note that the physical meaning of L∞ (Ω) space is that its occupant func-
tions satisfying the condition |u(x)| < ∞ almost everywhere in Ω.
We use the short-handed notation, → 0 to denote a limit process of a
sequence = {1 , 2 , · · · n , · · · · · · }, and n → 0 as n → ∞.
The strong convergence of a function sequence, u := {u1 , u2 , · · · , un , · · · },
is measured by the distance in the particular normed space, i.e. a sequence, u ,
is said to converge strongly in Lp (Ω) to a limit u0 , if
lim ku − u0 kLp (Ω) = 0 .

→0
The strong convergence is denoted by an arrow, namely,
u → u0 , in Lp (Ω) strongly
On the other hand, the weak convergence is measured by a so-called weighted

residual distance, which is associated with a weighting function, or test func-
tion in the dual space of the original norm space.
For the weak convergence in Lebesgue space Lp (Ω), the test function is in
0
its dual space Lp (Ω) with
1 1
+ 0 =1.
p p
Therefore, the formal statement of weak convergence in Lp (Ω), 1 ≤ p < +∞
is as follows: a sequence u is said to converge weakly in Lp (Ω) to a limit u0 ,
0
if for any test function φ ∈ Lp (Ω), it satisfies
Z Z
lim u (x)φ(x)dx = u(x)φ(x)dx
→0 Ω Ω
The weak convergence is denoted by a harpoon, namely
u * u0 in Lp (Ω) weakly .
The main interest of the weak convergence is that it is sequentially rela-

tive compact on bounded set. This means that for all the bounded sequence,
ku kLp (Ω) ≤ C, there exists a subsequence u 0 and a limit u0 such that
>0
u 0 converges weakly to u0 in Lp (Ω), 1 0
strong convergence.
Intuitively speaking, the strong convergence is more or less the usual point-
wise convergence, while the weak convergence is a notion of convergence “in
average” (up to a fluctuation of zero-mean).
If Ω is finite, we may choose test function
1 0
φ(x) = ∈ Lp (Ω)
Ω
then u (x) * u0 (x) requires that
Z Z Z
1 1
lim u = u (x)dx = u0 (x)dx
→0 Ω Ω Ω Ω Ω
That is lim→0 Ω =< u0 >Ω .

We state (without proof) the connection between strong convergence and
pointwise convergence. This statement is false for weakly convergence.
Theorem 10.5 1 Let Ω be a bounded open set in IRd . Let u be a sequence

converging strongly to a limit u0 in Lp (Ω), 1 ≤ p ≤ +∞, i.e.
u (x) → u0 (x)
Then there exists a subsequence, u0 ⊂ u , and a function h(x) ∈ Lp (Ω)

such that,
lim u0 (x) = u0 (x), almost everywhere in Ω

0 →0
|u (x)| ≤ h(x), almost everywhere in Ω
2 Assume that the sequence u (x) is bounded in Lp (Ω) (1 < p ≤ ∞), and
lim u (x) = u0 (x), almost everywhere in Ω

→0
Then
u (x) → u0 (x) in Lq (Ω) (1 ≤ q < p) strongly .
To feel the differences between strong convergence and weak convergence,

we consider the following example.
x
Example 10.6 Let u (x) = sin , p = 2, and Ω = (1, 0). Choose test

function φ(x) = 1. We have
Z 1 Z 1
x
u (x)φ(x)dx = sin dx
0 0
x 1 1
= − cos = 1 − cos
0
As → 0, u * 0, weakly in L2 (Ω), i.e. the weak limit of the sequence u (x)
is zero.
On the other hand, it seems that u (x) has no strong limit in L2 (Ω). This is
because
s s
Z 1 Z 1 x
ku kL2 (Ω) = u2 (x)dx = sin2 dx
0 0
s
1 1
Z 2x
= 1 − cos dx
2 0
r
1 2
= 1 − sin2 (
2 2
Suppose u → f (x) and f (x) ∈ L2 (Ω). Therefore,
Z 1 Z 1 Z 1
x 2 x x
lim sin − f (x) dx = sin2 dx − 2 sin f (x)dx
→0 0 0 0
Z 1 Z 1
2 1
+ f (x)dx = + f 2 (x)dx 6= 0 .
0 2 0
0
because f (x) ∈ (L2 ) (Ω).
Moreover, the fact that
Z 1 x 1
lim sin2 dx =
→0 0 2
also indicates that the product of two weakly convergence sequences does not
converge to the product of their weak limits. Otherwise,
Z 1 x
lim sin2 dx = 0
→0 0

because both sin x * 0 in L2 ([0, 1]).
It is worth noting that the product of two strong convergence sequence does
converge to the product of the two limits strongly, but it may be in a different
Lebesgue space in general.
For instance, if both u → u0 in L2 (Ω) strongly and v → v0 in L2 (Ω)

strongly, then
ku v − u0 v0 kL2 (Ω) = k(u − u0 )(v − v0 ) + (u − u0 )v0 + (v − v0 )u0 kL2 (Ω)
1/2 1/2
≤ ku − u0 kL2 (Ω) kv − v0 kL2 (Ω)
1/2
1/2
+kv0 kL2 (Ω) ku − u0 kL2 (Ω)
1/2
1/2
+ku0 kL2 (Ω) kv − v0 kL2 (Ω)
Hence
u v → u0 v0 in L2 (Ω) strongly .
Unfortunately, the same is not true for the weakly convergent sequences. In
our previous example,
x
u (x) = sin → 0 in L2 (Ω) weakly

but for u (x) = v (x) = sin x
1
u (x)v (x) * ! in Lp (Ω) 1 ≤ p < +∞ .
2
Moreover, in practice, if u * u0 in Lp (Ω), and J(u) is a nonlinear func-
tional, say quadratic functional, J : Lp (Ω) → IR.
It is usually
J(u ) 6* J(u0 ) in any sense !
10.6.2 G- Convergence
Consider our model homogenization BVP,
x
L u = f, x ∈ Ω, where L = −∇ · A( ) · ∇

u = ū, ∀x ∈ ∂Ω
∂Ω
where the heat conduction (or diffusion) coefficient Aij (y) are Y-periodic func-
tions.
Suppose that solution of the above BVP can be found as
−1
u (x) = L f,
Obviously, u ∈ H 1 (Ω) and f ∈ H −1 (Ω).

Recall the definition of Green’s function. We have

−1 Z

u (x) = L f= G (x − y)f (y)dy
Ω
Suppose that there exists a weak limit u0 (x) in H 1 (Ω) such that
u (x) * u0 (x) in H 1 (Ω) weakly
and the weak limit u0 (x) has the representation,
Z −1
u0 (x) = G0 (x − y)f (y)dy =: L0 f
Ω
Therefore, the weak convergence of u (x), i.e. u * u0 (x), implies that

Z Z
G (x − y) − G0 (x − y) f (y)dydx = 0, → 0 (10.147)
Ωx Ωy
Change the order of integration, (10.147) yields

Z Z
f (y) G (x − y) − G0 (x − y) dx = 0, as → 0 . (10.148)
Ωy Ωx
Equation (10.148) suggests that the weak convergence of Green’s function,

i.e. G * G0 , which implies a special type of convergence of the differential
operator sequence L = −∇ · A · ∇. We call the convergence of differential
operator sequence L as the G-convergence,
G
L → L0 (10.149)
in the sense of
G ∗ f * G0 ∗ f, in H 1 (Ω) weakly .
Note that the symbol ∗ denotes the standard convolution.
In fact, the convergence of the differential operator sequence, L = −∇·A ·
∇, may be viewed as the convergence of matrix sequence, Aij , to its G-limit
A0ij , or
x
G
Aij → A0ij

The following definition of G-convergence is provided by Allaire.
Let Msd be the linear space of symmetric real matrices of order d. For any
two positive constants α > 0 and β > 0, we define a subspace of Msd made of
coercive matrices with coercive inverse, namely,
Msα,β := {Mij } ∈ Msd , such that αξ 2 ≤ Mij ξi ξj

o
and βξ 2 ≤ Mij−1 ξi ξj , ∀ξ ∈ IRd
Let Ω be a bounded open set in IRd and define the space L∞ (Ω; Msα,β ) of
admissible symmetric coefficient matrices.
We have the following definition of G-convergence,
Definition 10.7 A sequence of symmetric matrices, A ∈ L∞ (Ω, Msα,β )

is said to be G-convergence to an homogenized, or G-limit, matrix A0 ∈
L∞ (Ω, Msα,β ), if, for any f ∈ H −1 (Ω)., the sequence solution u (x) of the
following model problem
−∇ · A ∇u = f, x ∈ Ω
u = ū, ∀x ∈ ∂Ω
converges weakly in H 1 (Ω) to the solution of the homogenized BVP,
−∇ · A0 · ∇u0 = f, x ∈ Ω
u0 = ū, ∀x ∈ ∂Ω
This definition makes sense because the following compactness theorem,

Theorem 10.8 For any sequence A ∈ L∞ (Ω; Mα,β ) of symmetric matri-
0
ces, there exits a subsequence, A ⊂ A , and a limit A0 ∈ L∞ (Ω; Mα,β )
such that A0 G-converges to A0 .
In the following examples, we want to show the differences between strong
convergence, weak convergence, and G-convergence.
Example 10.9 In this example, suppose that we have two objects with the
same macroscopic dimensions but different checkerborad microscopic struc-
ture.
The diffusitity matrix coefficients are assumed to be
Aij = aδij
We denote the diffusitivity in the white region as a1 and the diffusitivity in the
black region as a2 , and a2 > a1 .
We denote the first micro-structure as S1 and the second micro-structure as
S2 .
Obviously, the first sequence A (S1 ) and the second sequence A (S2 ) have
the same G-limit, i.e.
a0 (S1 ) = a0 (S2 ) .
As one can see that there is no pointwise convergence possibility, because for
a fixed spatial point,
|a0 (S1 ) − a0 (S2 )| = a2 − a1 > 0 .

Figure 10.9. The difference between strong convergence and G-convergence
Nevertheless, in this example, indeed, the weak convergence limit of the two
layouts are the same
< A (S1 ) >Ω =< A (S2 ) >Ω
Example 10.10 In this second example, we would like to show a case that
there are two micro-structure layouts with the same weak convergence limits,
but different G-limits.
In this example, we assume that in each unit cell, the black and white areas
are the same, therefore the volume fraction of the two phases are the same.
In the layout A, all the “good” material are connected, therefore it is a bet-
ter arrangement for heat conduction, whereas in the layout B, all the “good”
materials are isolated, disconnected, or insulated, it should be very hard for
heat to diffuse from one point to another point.
Based on this argument, the two layouts should have different G-limit, and
a0 (S1 ) > a0 (S2 ) .
On the other hand,

< a(S1 ) >=< a(S2 ) >
Figure 10.10. The difference between weak convergence and G-convergence
Figure 10.11. The difference between weak convergence and G-convergence
as indicated above.
Example 10.11 In the third example, we would like to show a case in which
two microstructure layouts have the same G-limit but different weak conver-
gence limits.
In this example, we fix the second layout of the previous example. Therefore,
we know that the G-limit of the second layout will be bounded by Vogit upper
bound and Reuss lower bound, i.e.
2a1 a2 1
Reuss bound = ≤ a0 S2 ≤ (a1 + a2 )
a1 + a2 2
We know change the first layout by increase the volume fraction of insolated
white phase, f1 such that f1 ∈ [0.5, ) and f1 → 1. Therefore, the G-limit of
the first layout will be bounded by
1
≤ a0 (S1 ) ≤ f1 a1 + (1 − f1 )a2 (10.150)
f1 1 − f1
+
a1 a2
Initially when f1 = 0.5 we have,
2a1 a2 1
< a0 S2 < a0 S1 < (a1 + a2 )
a1 + a2 2
If a1 << a2 .
The Reuss bound for the second layout is almost ≈ 2a1 . From Eq. (10.150),
one can see that as f1 → 1, the Reuss bound (lower bound) of the first layout
will become
1
→ a1 , as f1 → 1 .
f1 1 − f1
+
a1 a2
This suggests that at certain volume fraction, 0.5 < f1 = fw < 1.0, the
G-limits of the two layouts will be the same, i.e.
a0 (S1 ) = a0 (S2 ) .
At that moment, since fw > 0.5 6= f2 , the weak convergence limits of the
two layouts will not be the same, i.e.
< a0 (S1 ) >= fw a1 + (1 − fw )a2 6=< a0 (S2 ) >= 0.5(a1 + a2 ) .
10.6.3 H- Convergence
H-convergence is a generalization of G-convergence, in which, the differ-
ential operator A , or its coefficient matrix, does not require to be symmetric
anymore.
Definition 10.12 (Definition of H-Convergence) A sequence of ma-

trices A in L∞ (Ω, Mα,β ) is said to converge in the sense of homogeniza-
tion, or simply H-convergence, to an homogenized limit, or H-limit, matrix
A0 ∈ L∞ (Ω, Mα,β ) if, for any right hand side f ∈ H −1 (Ω), the sequence u
of solution of
−∇ · A · ∇u = f (x), ∀x ∈ Ω (10.151)

u = ū, ∀x ∈ ∂Ω (10.152)
satisfies
u (x) → u0 (x) weakly in H1 (Ω) (10.153)
h iN
A · ∇u → a∗ · ∇u0 weakly in L2 (Ω) (10.154)
where u0 is the solution of the homogenized equation,

−∇ · A0 · ∇u0 = f (x), ∀x ∈ Ω (10.155)
u0 = ū, ∀x ∈ ∂Ω (10.156)
10.6.4 Γ- Convergence
For a large class of elliptical BVPs, each BVP under consideration has
one-to-one correspondence to a variational principle. The well-known Lax-
Milgram theorem guarantees the equivalence between the two.
Therefore, the convergence of differential operators may imply a possible
convergence of the corresponding functional in the related function spaces.
Definition 10.13 (Definition of Γ-Convergence) Let X be a func-
tional space endowed with a norm k·kd . Let be a sequence of positive indexes
which goes to zero. Let F be a sequence of functional defined on X with val-
ues in IR. The sequence F is said to Γ-convergence to a limit functional F0 if,
for any function x ∈ X,
1 all sequences x converging to x satisfy
F0 (x) ≤ lim inf F (x )
→0 x∈X
and
2 there exists at least one sequence x converging to x, such that
F0 (x) = lim F (x )
→0
Example 10.14 (An Example of Γ-Convergence) Consider the fol-

lowing diffusion problem, with diffusion coefficient matrix, A is symmetric
and Y-periodic,
x
−∇ · A ∇u = f, ∀x ∈ Ω (10.157)

u (x) = 0, ∀ x ∈ ∂Ω (10.158)
The BVP (1) and (2) is equivalent to the following variational problem:
Find u ∈ H01 (Ω) such that
1 Z x Z
inf J(u) = inf ∇u · A · ∇udx − f udx
u∈H01 u∈H01 2 Ω Ω
Therefore. the Γ-convergence of J (u) (with respect to the strong topology

of L2 (Ω)) is equivalent to the homogenization of the PDE (1)-(2).
10.7 Exercises
Probelm 10.1 Show that for isotropic materials the fourth-order tensor,
1 h i
gijk` (ξ) = ξj (δ i` ξk + δ ik ξ` ) + ξi (δ j` ξk + δ jk ξ` )
2ξ 2
1 ξi ξj ξk ξ` ν ξi ξj i
− + δk` . (10.159)
1 − ν ξ4 1 − ν ξ2
Probelm 10.2 Consider cuboidal region of inelastic strain (eigenstrain)
due to solute segregation forming cuboidal precipitates. The precipitate sub-
domain (or inclusion) has the dimension 2a×2a×2a, and the unit cell (U) has
the dimension 2L × 2L × 2L. The eigenstrain is assumed to have a constant
value within each inclusion, and be zero outside the inclusion,

δij ε, ∀x ∈ Ω;
ε∗ij = (10.160)
0, ∀x ∈ U/Ω,
where
n o
U x −L ≤ xi ≤ L, i = 1, 2, 3
= (10.161)
n o
Ω = x −a ≤ xi ≤ a, i = 1, 2, 3 , and a < L (10.162)
Find :
(a) the disturbed displacement field u1 (x) (Hint: Mura’s book pages: 20-
21).
(b) G(ξ) = g0 (ξ)g0 (−ξ).
Probelm 10.3 Consider the followin boundary-value problem in a medium
with periodic structure,
∂ 2 u
− = f, ∀x ∈ Ω (10.163)
∂xi ∂xi
u = 0, ∀x ∈ ∂Ω (10.164)
∂u
= 0, ∀x ∈ Γ (10.165)
∂n
where Γ is the interface between the matrix and inhomogeneous phase.
Show that the homogenized differential equation is
∂ 2 u0
−qik = f, ∀x ∈ Ω
∂xk ∂xk
Figure 10.12. Distribution of periodic precipitates
with effective coefficients qik defined as

Z
1 ∂Uk
qik = δik + 2 dy
|Y | Y ∂yi
and the associated canonical cell problem is
∂ 2 Uk
= 0, ∀y ∈ Y (10.166)
∂yi ∂yi
∂Uk
ni = nk , ∀y ∈ S (10.167)
∂yi
10.8 Toshia Mura

This is the biography sketch of Professor Toshio Mura, the sole author of
our second text book, " Micromechanics of Defects in Solids". The biography
sketch was written more than 10 years ago by Professor Mori (who also made
some contributions in micromechanics as well, the Mori-Tanaka theory, for
instance, bears his name). Before I copy the biography sketch, I would say few
things about professor Mura myself. For the past four and five years, I have
the opportunity to study and work with Professor Mura, and I have stayed with
Figure 10.13. Toshio Mura
him in the same office for almost four years (I was a postdoctal fellow then and
he was an emeritus professor).
Almost every week, he took me to lunch (because he insisted to pay ev-
erytimes, so we can not go out everyday), and I learned a lot of things from
Professor Mura, and had many good conversations as well as good memories.
Last year, Professor Mura received the Japanese Imperial model—the highest
honor bestow by Janpanes emperor and Royal family to scientists and other
citizens—for his contribution in micromechanics. I remembered back in 1997,
in his retirement party, professor Jan Achenbach said that Professor Mura is
one of the “seven samurai” (an international renowed Japanese moive, samurai
in Japanese means warrior, previously in Northwstern there were seven fa-
mous Mechanics professors: Achenbach, Belytschko, Dundurs, Keer, Mura,
Nemat-Nasser, and Bazant). Professor Mura is a theoretician, and has a very
“romantic” outlook of the world, (romantic is opposed to the “down-to-earth”
mentality of experimentalist) he believes that you are at your most creative
stage, when you are in your dream.
Biography sketch of Toshio Mura.
“ Toshio Mura, second son of Shinzo and Chie Fujii, was born in Ono, a
small port village of Kanazawa, the capital of Ishikawa Prefecture, Japan, on
December 7, 1925. Among the locals, the Fujiis are well known as brewers
having a long history in the area. Kanazawa is an old city on the coast of the
Sea of Japan, where traditional culture is proudly maintained and apprecaited.

.....
In 944, during the most difficult time of the war, Mura went to the Impe-
rial University of Tokyo to read Aeronautical Engineering. After the war, his
department was dissolved and changed to the Department of Applied Mathe-
matics at the University of Tokyo. ....
The title of his Ph.D. dissertation was “Study on Thermal Stresses”. His
work in the dissertation turned out to be one of the earliest papers on the dy-
namic wave of thermal stresses.
As a graduate student, Mura also began his teaching career as a mathemat-
ics professor at Meiji University, where he met and worked with his lifelong
friend, Nobuo Kinoshita. Their joint paper, “On the boundary value problem of
elasticity,” which was published during his tenure at Mriji University (1956),
agitated some Russian mathematicians in the field of integral equations. Had
this work been extended, it would have led to the powerful computational tech-
nique now known as the boundary element method. .....
At the graduate school, Mura was introduced to his future wife, Sawa, by her
sister, Sumi, who had worked in the Department of Aeronautical Engineering.
During the courtship, Mura often visited the Ozaki’s and Sumi fondly recalls
that he praised Sawa’s cooking. They married in 1953 and their first daughter,
Miyako, was born in 1955.
In 1958, Mura went to Northwestern University’s Department of Materi-
als Science, Evanston, Illinois, to work with John O. Brittain. While at this
department, Mura conceived the idea of the Periodic Distribution of Disloca-
tions, which was documented in a paper and published later in the Proceddings
of the Royal Society of London as a communication by A. H. Cottrell and R.
E. Peierls (1964). In this paper, for the first time, the Fourier method was used
to obtain the elastic field of dislocations. As seen in his later publications, the
Fourier method became Mura’s favorite tool to analyze elastic fields.
In 1961 Mura jointed the department of civil engineering at Northwestern
University as an assistant professor. The pleasant but stimulating atmosphere,
brewed by his colleagues, John Dundurs and Leon Keer, also encouraged him.
Dundurs and Mura obtained the elastic fields of dislocations parallel to a cylin-
drical inhomegeneity (1964). Keer and Mura analyzed a penny-shaped crack
with a plastic zone by solving an integral equation, Mura’s first paper con-
cerned with a crack (1963).
In 1963, Mura succeeded in expressing the elastic field of a curved disloca-
tion in a line integral, now known as Mura’s Formula (1963). The line integral
is along the dislocation and contains only the state quantities that character-
ize the dislocation. This solution was later extended by John R. Willis, who
gave the field of a dislocation segment in the form algebraic equations, wh-
cih equired the solution of sextic equation (1970). .... The paper in 1963 is
also noteworthy for introducing the concept of a dislocation flux tensor, which
is yseful when the dynamic motion of dislocations is examined. The period,
during which Mura’s Formula was found, coincided with his promotion to As-
sociate Professor of Civil Engineering.
....
The dislocation density and flux tensors were applied to continuum plastic-
ity theory. Believing that a stress appearing within the framework of continuum
plasticity was the sum of external and dislocation stresses, Mura published a
series papers, in the late 1960s, along these lines that emphasized the distribu-
tion and stress of dislocation.
In 1967 Mura became Professor of Civil Engineering. At that time Mura
nad J. G. Kunag, his student, obtained the solutions for a pile-up of edge dis-
locations against the interfacial boundary between different materials.
The pioneering work of J. D. Eshelby, his beloved peer, appears to have
inspired and stimulated Mura, as seen in his studies of static and dynamic
fields of dislocations in anisotropic media and in dislocation pile-ups. As can
be inferred from the preface to his book, Micromechanics of Defectcs in solids,
Mura regards Eshelby’s work on inclusions and inhomogeneities as being the
most important and fundamental.
To Mura the evaluation of the disturbance in elastic fields due to elastic in-
homogeneities is the most interesting application of the theory of inclusions.
For example, Z. A. Moschovides and Mura solved the stress field caused by
two inhomogeneities by applying the equivalent inclusion method with poly-
nomial eigenstrains. A computer program, performing the numerical calcula-
tions, complained that the matrice involved for linear equations were singular.
Moschovides looked for the bugs that might have caused this complaint, but no
bugs were found. The linear equations were carefully examined analytically
and the cause of the complaint was found. There existed certain distributions
of eigenstrains that yields no elastic field. Rozo Furuhashi, a visiting scholar,
and Mura later generalized this finding and showed that impotent inclusions
exist in a general sense. The impotent inclusions have eigenstrains defined by
derivatives of a continuous vector (displacement) that vanished at the bound-
ary of the inclusions. This anecdote illustrates Mura’s teachings: "study and
examine a specific subject carefully. If there is anything strange and exciting,
you can later generalize it in a broader sense.”
Mura also interacted with experimentalists, who eagerly sought his advice
and aid on issues of mathematics and mechanics. In particular, Morris E. Fine,
and his students in Northwestern’s Department of Materials Science and Engi-
neering, benefited from this interaction in their studies of the fatigue of alloys.
Mura also gained insight into material properties and structures by the interac-
tions with these materials scientists.
........
In 1986, Mura was elected to membership in the National Academy of Engi-

neering, U.S.A. with the citation, ‘For initiating and promoting micromechnics
to bridge the gap between metal physics and engineering mechanics.’ During
the same year, he was appointed Walter P. Murphy Professor in the Technolog-
ical Institute at Northwestern University.
.......”
Chapter 11
MICROMECHANICS THEORY OF VOID GROWTH
Damage theory of void growth is central to failure mechanism of ductile ma-

terials. In late 1960’s and early 1970’s, pioneer contribution have been made by
several authors, Mclintock [1968], Rice and Tracy [1969], and Gurson [1972],
using micro-mechanics techniques to develop damage theory in constitutive
modeling of ductile materials.
The homogenization result obtained by Gurson marks a significant mile-
stone in the development of micromechanics, because the outcome of the ho-
mogenization is foundamentally different from that of micro-elasticity theory.
In micro-elasticity theory, the homogenized consititutive relations are virtually
the same as the constitutive relation in micro-scale, i.e., linear elastic constitu-
tive relations or generalized Hook’s law. The only differences in constitutive
laws at different scales are the magnitude and the spatial distribution of elastic
constants. Whereas, in the Gurson model, a completely new constititive rela-
tion at macro-level emerges from the homogenization, which represents a new
philosophy:
finding new physical laws and new mechanics by doing homogenization.
This notion is so attractive, and it has remained the very ideal and ultimate
objective of contemporary micromechanics and multiscale simulations.
11.1 Void Growth in Linear Viscous Solids

Consider a linear viscous RVE, whose constitutive behaviors at microscale
can be described as the following rate dependent expression,
σij = Cijk` ˙k`
The viscous coefficients resemble to that of linear elastic tensor,
2ην
Cijk` = δij δk` + η(δik δj` + δi` δjk )
1 − 2ν
Micromechanics Theory of Void Growth 307
Figure 11.1. A spherical void in the middle of an RVE
In the case of incompressible viscous media,

h1 1 i 2
Cijk` = 2η δik δj` + δi` δjk − δij δk` + ηδij δk`
2 3 3
Consider a spherical void, Ω, inside an RVE with a radius, R = a. A
uniform triaxial stress state is imposed at the remote boundary of the RVE, i.e.
∞
ti = σij nj , ∀x ∈ ∂V
∞ = Tδ .
where σij ij
Applying Eshelby’s equivalent eigenstrain principle, the stress inside the
void may be written as

σij = Cijk` ˙∞ d ∗
k` + ˙k` − ˙k`
−1
Note that ˙∞ ∞
ij = Dijk` σk` and Dijk` = Cijk` .
Since inside the void, there is no stress σij = 0, we have
˙ij = ˙∞ d ∗
ij + ˙k` = ˙ij
This means that eigenstrain rate should be the same as the actual strain rate,
which gives the physical meaning for eigenstrain rates. That is the prescribed
eigenstrain rate should be the expansion rate of the void.
Moreover, one can find that

∞
σij = Cijk` (˙∗k` − ˙dk` )
By Eshelby’s single inclusion solution, one can write
dij = Sijk` ˙∗ij
Therefore,
σ = C : (1(4s) − S) : ˙ .
Denote
Q := C : (1(4s) − S) .
The remote stress can be related with volumetric strain rate of the void, i.e.
σii∞ = Qii11 ˙∗11 + Qii22 ˙∗22 + Qii33 ˙∗33
Consider,
C = 2η1 + ν + 1 − 2νE(1) + 2ηE(2)

S = s1 E(1) + s2 E(2)
(1) (2)
Eii11 = 1, and Eii11 = 0 ,
1+ν 2(4 − 5ν)
3(1 − ν) 3(1 − ν)
One may find that
(1 + ν)
Qii11 = Qii22 = Qii33 = 8η .
3(1 − ν)
By symmetry, it is easy to see that
˙∗11 = ˙∗22 = ˙∗33 = ˙
Consequently, we have
8 η(1 + ν)
T = ˙
3 (1 − ν)
Since the volume of the void is,
4π 3
V = a ⇒ V̇ = 4πa2 ȧ,
3
The relative void growth rate will be
V̇ ȧ
= 3 = 3˙
V a
where ˙ = ȧa is the strain rate in radial direction.

Finally, we link the magnitude remote stress with the void growth rate,
8 η(1 + ν) V̇
T =
9 1−ν V
The above solution was obtained by Budiansky et al in 1981, almost ten
years after publication of the McClintock solution and the Gurson model.
Figure 11.2. A solid with traction-free defect
11.1.1 Averaging theorems for soilds with traction-free

defects
Consider an RVE, V , containing a traction-free defect, Ω. That is the trac-
tion force ti = σij nj = 0, ∀x ∈ ∂Ω. Suppose that on the remote boundary
condition ∂V the prescribed traction boundary condition is imposed
ti = σij nj = Σij nj ∀x ∈ ∂V
where Σij is a contant tensor, and it is often denoted as the macro-stress tensor.
The following averagy theorems hold in the RVE,
1. < σij >V = Σij (11.1)

(add)
2. < ˙ij >V = Ėij + ˙ij (11.2)
Z
(add) 1
where ˙ij = u̇i nj + u̇j ni dS (11.3)
2V ∂Ω
and Ėij = Dijk` Σk` for the linear viscous solid.

Expressions (11.2) and (11.3) are called additional strain rate formulas 1 .
We first show (11.1),
Z Z
1 1
< σij > = σij dV = σip xj dV
V V V V ,p
Z Z
1
= σip np xj dS − σip np xj dS
V ∂V
| ∂Ω {z }
=0, because σip np =0, ∀x∈∂Ω
Z
1
= Σip np xj dS = Σij
V ∂V
We know that under the prescriber traction boundary condition,
< ˙ij >6= Ėij
To prove the additional strain rate formula, we use the so-called reciprocal
theorem of virtual power. Consider two sets of traction boundary conditions
and the corresponding velocity fields on the same ilinear viscous RVE, V , the
following equality holds,
Z Z
(1) (2) (2) (1)
t i u̇ i dS = ti u̇i dS
∂Ω− ∂Ω−
S S
∂V ∂V
Let the traction b.c. for the first state be

[
t(1) = n · δΣ, ∀x ∈ ∂V ∂Ω−
which yields the following trivial solution,
{u̇(1) , ˙ (1) , σ (1) } = {x · δ Ė, δ Ė, δΣ}
where δ Ė = D : δΣ.
Let the traction b.c. for the second state as

(2) n · Σ, ∀x ∈ ∂V
t =
0, ∀x ∈ ∂Ω−
and it correspondes to the real solution,
{u̇(2) , ˙ (2) , σ (2) } = {u, ,

˙ δσ}
1A similar expression is hold for infinitesimal strain as well.

The reciprocal theorem gives,

Z Z Z Z
(1) (2) (1) (2) (2) (1) (2) (1)
ti u̇i dS + ti u̇i dS = ti u̇i dS + ti u̇i dS
∂Ω− ∂Ω− ∂Ω−
S
∂V ∂V
| {z }
=0
Z Z Z
n · δΣ u̇dS + (n · δΣ) · u̇dS = (n · Σ) · (x · δ Ė)dS
∂V ∂Ω− ∂V
Notice the following facts:
1
1
n · δΣ · u̇ = δΣ : (u̇ ⊗ n + n ⊗ u̇)
2
2
n · Σ · x · δ Ė = δΣ : D : (x ⊗ n) · Σ
We then have
Z Z Z
1
δΣ : { D : (x⊗n)·ΣdS − n⊗ u̇dS − n⊗ u̇dS} = 0 . (11.4)
V ∂V ∂V ∂Ω
Consider
1 1 Z
D: x ⊗ ndS · Σ = Ė; (11.5)
V ∂V
2 Z Z
1 1
Sym n ⊗ u̇dS = ˙
dV =< ˙ > (11.6)
V ∂V V V
3
Z Z
1 1
Sym n ⊗ u̇dS = − n ⊗ u̇dS
V ∂Ω− V ∂Ω
Z
1
= − n ⊗ u̇ + u̇ ⊗ n dS (11.7)
2V ∂Ω
Substitution (11.5)–(refeq:cond3) into (11.4) gives the following additional

formula for strain rate
< ˙ >= Ė + ˙(add)
where Z
1
˙ (add) = u̇ ⊗ n + n ⊗ u̇ dS
2V ∂Ω
Figure 11.3. A cylindrical void in an inelastic RVE
11.2 The McClintock solution

The McClintock solution is the classic result of void growth in an inelastic
RVE, which has been served as the bench mark example in many homogeniza-
tions of inelastic solids.
The basic premises of McClintock solution are two: (1) at micro-level, the
RVE behaves as a rigid-plastic material, and (2) the RVE is incompressible.
Consider the following flow rule,
∂f
˙pij = λ̇
∂sij
The yield surface is described by J2 criterion (von Mises criterion),
Y2 1 Y2
f = J2 − = sij sij − =0
3 2 3
where sij is the deviatoric stress tensor,
1
sij = σij − σii
3
One can then rewrite the flow rule as
∂f
˙pij = λ̇ = λ̇sij (11.8)
∂sij
where the proportionality λ̇ can be determined by contracting the flow rule

with plastic strain rate, i.e.
1 1 Y2
˙ij ˙ij = λ̇2 sij sij = λ̇2
2 2 3
One can then solve for λ̇,
q
p p
3 ˙ij ˙ij
p
r
3 2 q 0 p 1/2 3 ¯˙
λ̇ = = √ I2 (˙ij ) =
2 Y 2Y 3 2Y
where
0 1 p p
I2 := (11.9)
2 ij ij
p 2 0
¯˙ = √ I2 (pij ) (11.10)
3
Therefore, the constitutive relation at micro-level are,
p
3 ¯˙
˙pij = sij
2Y
In the cylindrical coordinate,
h2 i1/2
p
¯˙ = (˙pr )2 + (˙pθ )2 + (˙pz )2
3
Consider the problem is axisymmetry and independent on z coordinate. The
equlibrium equation becomes,
dσr σr − σθ
+ =0. (11.11)
dr r
Assume that the velocity field is
ur = u(r), uθ = 0, and ˙z = ˙ = constant .
Hence,
du̇
˙r = (11.12)
dr
u̇
˙θ = (11.13)
r
The incompressible condition yields,
du̇ u̇
˙r + ˙θ + ˙z = + + ˙z = 0 .
dr r
Rewrite the above expression as

du̇ d
r + u̇ + r˙z = 0 ⇒ ru̇ = −r˙z
dr dr
Integrate over the radial direction from the surface of the void to the interior
of the RVE, Z r Z r

d ρu̇(ρ) = − ρ˙z dρ
b b
Note that variable ρ is the dummy variable.
Considering ˙z = ˙ = const., we have
r ρ2 r
ρu̇(ρ) =− ˙z
b 2 0
Consequently,
˙
z
ru̇(r) − bḃ = − r2 − b2
2
˙z 2
⇒ ru̇ = bḃ + r − b2
2
Finally
b2 ḃ ˙z ˙z r
u̇(r) = + − (11.14)
r b 2 2
Let,
1
σ = (σr + σθ + σz ) .
3
We have
sr = σr − σ (11.15)
sθ = σθ − σ (11.16)
The components of the flow rule in an axisymmetric plane are
p p
3 ¯˙ 3 ¯˙
˙r = ˙pr = sr = (σr − σ) (11.17)
2 Y 2 Y
p p
3
¯˙ 3 ¯˙
˙θ = ˙pθ = sθ = (σθ − σ) (11.18)
2 Y 2 Y
(11.17) - (11.18) leads to
p
3 ¯˙
˙θ − ˙r = (σθ − σr ) (11.19)
2 Y
Utilizing (11.19), it can be found that
σθ − σr 2Y ˙θ − ˙r
=
r 3r ¯˙p
Therefore the equilibrium equation becomes

dσr σr − σθ dσr 2Y (˙r − ˙θ )
+ = + p =0. (11.20)
dr r dr 3r ¯˙
Integrating over the radius direction,
1 ∞ 2 ∞ (˙θ − ˙r ) dρ
Z Z
dσr = p
Y b 3 b ¯˙ ρ
Z ∞
1 2 (˙θ − ˙r ) dρ
⇒ [σr (∞) − σr (b)] = p
Y 3 b ¯˙ ρ
Consider the traction boundary condition,
σr (b) = 0, and σr (∞) = σ ∞ (11.21)
We have ∞
(˙θ − ˙r ) dρ
Z
σr (∞) 2
= p (11.22)
Y 3 b ¯˙ ρ
p
To integrate (11.22), one has to evalute ¯˙ first. Since
b2 ḃ ˙z ˙z r
u̇(r) = + − ,
r b 2 2
direct calculation gives
du̇r b2 ḃ ˙z ˙z
˙r = =− 2 + − (11.23)
dr r b 2 2
u̇r b2 ḃ ˙z ˙z
˙θ = = 2 + − (11.24)
r r b 2 2
In cylindrical coordinate, the effective strain rate is
h2 i1/2
p
¯˙ = (˙pr )2 + (˙pθ )2 + (˙pz )2
3
i 1/2
( )
2 h b2 ḃ ˙z ˙z 2 b2 ḃ ˙z ˙z 2 2
= + + + 2 + − + ˙z
3 r2 b 2 2 r b 2 2
h 4 ḃ 1 2 b2 2 i1/2
= |˙z | + + 1
3 b˙z 2 r2
Define
b2 2 ḃ 1 b2
x := √ + = α (11.25)
r2 3 b˙z 2 r2
where
2 ḃ 1
α := √ + (11.26)
3 b˙z 2
Subsquently,
¯˙p = ˙x (1 + x2 )1/2 (11.27)
and
b2 ḃ ˙z √ 2 b2 ḃ 1 √
˙θ − ˙r = 2 + = 3˙ z √ + = 3˙z x (11.28)
r2 b 2 3 r2 b˙z 2
Since
2b2 2 ḃ 1 2
dx = − 3
√ + dr = − xdr,
r 3 b˙z 2 r
dr 1 dx
=− .
r 2 x
Make change of variable,
b2
x=α ,
r2
and
r = b, x → α; r → ∞, x → 0 .
We can then integrate (11.22)
√
2 ∞ (˙θ − ˙r ) dρ 2 ∞
Z Z
σ∞ 3˙ x dr
= p = √ z
Y 3 b ˙
¯ ρ 3 b ˙z 1 + x2 r
Z 0 Z α
1 dx 1 dx
= −√ √ =√ √
3 α 1+x 2 3 0 1 + x2
1 α 1
= √ arcsinhx = √ arcsinh(α)
3 0 3
The inverse expression of the above result is
2 ḃ 1 h √3σ i
∞
√ + = sinh (11.29)
3 b˙z 2 Y
Based on uniaxial tension test, one can measure
Y
q
0
τ0 = J2 = √
3
We obtain the relationship between void growth rate and remote stress value,
√
ḃ 3 hσ i 1
∞
= ˙z sinh − ˙z (11.30)
b 2 τ0 2
A few comments about the McClintock solution are as follows:
1 McClintock solution is the only (essential) exact solution available for void
growth in nonlinear viscous media;
2 McCintock solution reveals an exponential increase in the void growth rate
under the positive remote stress load.
To illustrate the fact, we consider a finite cylindrical void with a heigh, H,
and radius b. The volume of the cylinder is
Ω = πb2 H ⇒ Ω̇ = 2πbḃH + πb2 Ḣ
Thereby,
Ω̇ ḃ
= 2 + ˙z
Ω b
and hence √
Ω̇ 3 h √3σ i
∞
= ˙z sinh (11.31)
Ω 2 Y
Compare (11.31) with Budiansky et al’s linear viscous void solution,
Ω̇ 9 1−ν
= σ∞
Ω 8 η(1 + ν)
One may appreciate the significant difference between the two.
3 At the remote boundary, x ∈ ∂V ,
1
˙ ˙r = ˙θ = − ˙
˙z = ,
2
Hence the macro equivalent strain rate is
h2 i1/2 h21 1 i1/2
˙∞
eq = ˙∞ ∞
ij ˙ij = ˙2 + ˙2 + ˙2 = ˙ (11.32)
3 3 4 4
Bi-axial stress state is applied at the remote boundary, ∂V , i.e.

1
Σ11 = Σ22 = σ∞ , Σ33 = T, and Σm = 2Σ11 + Σ33
3
The von Mises criterion becomes

h3 i1/2
Σeq = Σij Σij
2
h3 i1/2
= (Σ11 − Σm )2 + (Σ22 − Σm )2 + (Σ33 − Σm )2
2
= |Σ33 − Σ11 | ≤ Y
The yield surface is |Σ33 − Σ11 | = Y .

Under such condition, we can rewrite the void growth rate equation as
Ω̇ √ √3σ √ √3Σ
∞ 11
= 3 sinh = 3 sinh . (11.33)
Ω˙ Y |Σ33 − Σ11 |
4 Let the total volume of the RVE be

V = Ω + Vmatrix
and
dV dΩ dVmatrix dΩ
= V̇ = + =
dt dt dt dt
dVmatrix
because the matrix is incompressible, = 0.
dt
Define the volume fraction of the void as
Ω
f= .
V
Then
Ω̇ Ω Ω̇ V − Ω
f˙ = − 2 V̇ =
V V V V2
Ω̇ Ω̇
= (1 − f ) = f (1 − f )
V Ω
Finally, we can express the rate of volume fraction as
√ √3Σ
˙ 3f (1 − f ) 11

f= sinh
˙eq |Σ33 − Σ11 |
11.3 The Gurson model

The significance of McClintock solution it that it links the remote stress, or
macro stress, with the void growth rate, and it reveals that in a perfectly plas-
tic RVE, the void growth rate is expenonetially related with the macro-stress.
Although, it can be argued that the notion representative volume element is em-
ployed in McClintock solution, it does provide new constitutive representation
at macro-level.
Not long after the publication of McClintock solution, a young scientist at
the time, A. L. Gurson, realized that there is more in the cylindrical void model
analyzied by McClinktock. In fact, one can derived the plastic potential at
macro-level by homogenized (meaning averaging in space) micro-stress distri-
bution. It was eaxctly what Gurson did his Ph.D. thesis, which has become one
of most cited papers in inelastic constitutive modeling and micromechanics.
11.3.1 Gurson’s homogenization of cylindrical void in a

rigid perfectly-plastic RVE
Figure 11.4. A cylindrical void in a rigid-perfectly plastic von Mises RVE
The objective of the Gurson model is to find macroscopic yield potential

function in terms of macro-stress and volume fraction of void in an RVE, i.e.,
we are looking for
F (Σeq , Σm , f ) = 0
where
r
3 0 0 0 1
Σeq = Σij Σij , Σij = Σij − Σm , and Σm = Σii
2 3
Again, the governing equations in the RVE are,
1 Equilibirum equations:
dσrr σrr − σθθ
+ =0.
dr r
2 von Mises flow rule:
2 σy
sij = ˙ij
3 ˙eq
3 incompressible condition of the matrix:

˙rr + ˙θθ + ˙zz = 0 .
Consider axisymetric remote (macro-stress) loading,
σ11 = Σ11 , σ22 = Σ22 , and Σ11 = Σ22 (11.34)

∂V ∂V
σ33 = Σ33 (11.35)

∂V
Under axisymmetric loading condition,

r h
1 i
Σeq = (Σ11 − Σ22 )2 + (Σ33 − Σ11 )2 + (Σ33 − Σ22 )2
2
= |Σ33 − Σ11 |, (11.36)
1 1
Σm = (Σ11 + Σ22 + Σ33 ) = Σαα + Σ33
3 3
1 1 1
= Σ11 + Σ33 − Σ11 = Σαα + Σeq (11.37)
3 2 3
1
where Σαα = Σ11 + Σ22 = 2Σ11 , or Σ11 = Σ22 = Σαα . Therefore, we are
2
essentially looking for the yeilding effects due to Σ11 and Σ33 − Σ11 .
Consider the following axisymmetric kinemetic pattern,
u̇r = u̇(r), u̇z (z) = Ė33 z .
Strain rate components are
du̇ u̇
˙rr = , ˙θθ = , ˙zz = Ė33 .
dr r
Since the matrix is incompressible,
du̇ u̇
˙rr + ˙θθ + ˙zz = + + Ė33 = 0,
dr r
one has Z Z
Ė33 A
d(ru̇) = − Ė33 rdr ⇒ u̇(r) = − r+
2 r
where A is an unknown constant.
Subsequently,
du̇ Ė33 A
˙rr = =− − 2 (11.38)
dr 2 r
u̇ Ė33 A
˙θθ = =− + 2 (11.39)
r 2 r
In fact, the constant A has a clear physical interpretation. Consider a cylin-

drical void with finite height,
Ω = πa2 H
The void growth rate and relative void growth rate are
dΩ
= 2πaȧH + πa2 Ḣ (11.40)
dt
Ω̇ ȧ Ḣ
= 2 + (11.41)
Ω a H
Since,
ȧ Ḣ
= ˙rr (a) and = Ė33 ,
a H
one may find that
Ω̇ Ė
33 A 2A
=2 − − 2 + Ė33 = − 2
Ω 2 a a
which leads to
a2 Ω̇
A=− . (11.42)
2 Ω
That is: A is proportional to the relative void growth rate.
Since the matrix is a rigid-perfectly plastic von-Mises material, it obeys the
following flow rule,
2 σy
sij = ˙ij
3 ˙eq
where the effective strain rate can be explicitely expressed as
2 1/2 h2 i1/2
˙eq = ˙ij ˙ij = ˙2rr + ˙2θθ + ˙2zz
"3 3 #
2 Ė33
h A 2 Ė
33 A 2 2
i
= + 2 + − 2 + Ė33
3 2 r 2 r

2 4 A2 1/2
2 a
4 1/2
= Ė33 + = Ė 33 1 + α (11.43)
3 r4 r
where the parameter, α, is defined as
2 |A| Ω̇ 1
α := √ 2
= √ (11.44)
3 Ė33 a2 Ω 3Ė33
Therefore, we can write,

2 σy 2 σy Ė
33 A
srr =
˙ rr = − −
3 a 4 1/2 3 a 4 1/2 2 r2
Ė33 1 + α2 Ė33 1 + α2
r r
2 σy 2 σy Ė
33 A
sθθ =
˙ θθ = + −
3 a 4 1/2
3 a 4 1/2
2 r2
Ė33 1 + α2 Ė33 1 + α2
r r
2 σy 2 σy
szz = a 4 1/2 Ė33 = 3 a 4 1/2
3
Ė33 1 + α2 1 + α2
r r
We can then find that
2 σy A A
sθθ − srr = +
3 Ė33 (1 + α2 (a/r)4 )1/2 r2 r2
4 σy A
=
3 Ė33 (1 + α2 (a/r)4 )1/2 r2
= σθθ − σrr
and
1 2 σy
szz − (srr + sθθ ) =
2 3 (1 + α (a/r)4 )1/2
2
12 σy
− (−Ė33 )
2 3 Ė33 (1 + α2 (a/r)4 )1/2
σy
=
(1 + α2 (a/r)4 )1/2
1
= σzz − (σrr + σθθ )
2
To this end, we are in a position to link the macro-stresses, Σ11 , Σ33 − Σ11 ,
and void volume fraction, f , together in a macro yield potential.
We first link Σ11 and |Σ33 − Σ11 | with remote strain rate, Ėij .
Consider the traction boundary conditions on the surface of the void and the
surface of the RVE,
1
σrr (a) = 0, and σrr (b) = Σαα = Σ11
2
note that Σrr (b) = Σθθ (b) = 12 Σαα .
1. Integrating equilibrium equation along the radius direction yields,
Z b Z b
dσrr σθθ − σrr
Σ11 = σrr (b) − σrr (a) = dr = dr
a dr a r
Since,
σθθ − σrr sθθ − srr 4 σy A
= =
r r 3 Ė33 (1 + α2 (a/r)4 )1/2 r3
we have Z b
4 A dr
Σ11 = σy (11.45)
3 a Ė33 (1 + α (a/r) ) r3
2 4 1/2
2. Consider the fact that σ11 + σ22 = σrr + σθθ , and Σ11 = Σ22 = 12 Σαα ,
Z
1 1 1
Σ33 − Σ11 = Σ33 − (Σ11 + Σ22 ) = σzz − (σxx + σyy ) dV
2 V V 2
Z
1 1
= σzz − (σrr + σθθ ) dV
V V 2
Z
1 1
= szz − (srr + sθθ ) dV
V V 2
Z
1 1
= szz − (srr + sθθ ) dV
V VM 2
Recall that
1 σy
szz − (srr + sθθ ) = a 4 ,
2
(1 + α2 )1/2
r
and dV = rdrdθdz. We have
Z b
2πH σy
Σ33 − Σ11 = a 4 1/2 rdr
πb2 H a

1 + α2
r
Z b
2σy rdr
= (11.46)
b2 a
a 4 1/2
1 + α2
r
Make change of variable,
a 2
x=α : x → [α, f α], when r → [a, b] .
r
a2 Ω
where f = 2 = .
b V
Therefore,
a2 4A 2A
dx = −2α 3
dr = − √ dr ⇐ α = √
r 3Ė33 r3 3Ė33 a2
and
√
Adr 3
= − dx, (11.47)
Ė33 r3 4
r4 a2 α dx a2
rdr = − 2 dx = − , ⇐ x = α (11.48)
2a α 2 x2 r2
√
Adr 3
Reconsider (11.45) and =− dx,
Ė33 r 3 4
Z b
1 4 1 Adr
Σ11 = Σαα = σy 2 1/2
2 3 a (1 + x ) Ė33 r3
√
4 3 fα Z
dx
= − σy √
3 4 α 1 + x2
Thereby, Z α
Σαα σy dx
=√ √ (11.49)
2 3 f α 1 + x2
We then find that the in-plane hydrostatic stress can be written as
√
3 Σαα h α + √1 + α 2 i
= log p
2 σy f α + 1 + (f α)2 (11.50)
αa2 dx
Reconsider Eq. (11.46) and rdr = − ,
2 x2
2σy b
Z
rdr
Σ33 − Σ11 = 2 2 1/2
b a (1 + x )
2σ αa2 Z f α dx
y
= 2
− √
b 2 α x 1 + x2
2
Z α
dx
= f ασy √
2 1 + x2
fα x
Carrying the integration, we have

hp p i
Σ33 − Σ11 = σy 1 + α2 f 2 − f 1 + α2
We can then link the deriatoric macro-stress with macro-strain rate and void
volume fraction,
Σeq p p
= 1 + f 2 α2 − f 1 + α2
σy (11.51)
Denote that
√
3 Σαα
A1 =
2 σy
Σeq
A2 =
σy
p
A3 = α + 1 + α2
p
A4 = f α + 1 + f 2 α2
Then results (11.50) and (11.51) can be rewritten as

A3
A1 = log (11.52)
A4
A2 = A4 − f A 3 (11.53)
We want to connect A1 and A2 by elminating A3 and A4 .

Rewrite (11.52) and (11.53) as
A3
exp A1 = (11.54)
A4
A4 = A2 + f A 3 (11.55)
Substituting (11.55) into (11.54) leads to an equation of A1 , A2 , and A3 ,

A3
exp A1 =
A2 + f A 3
which expresses A3 in terms of A1 and A2 ,
A2 exp(A1 )
A3 =
1 − f exp(A1 ) (11.56)
Substituting (11.57) back into (11.55) yields an equation among A1 , A2 ,

and A4 ,
(1 − f exp(A1 )A2 + f A2 exp(A1 )
A4 = A2 + f A 3 =
1 − f exp(A1 )
Solving this equation yields
A2
A4 =
1 − f exp(A1 ) (11.57)
Consider the identities,

p p
A23 = (α + 1 + α2 )2 = α2 + 2α 1 + α2 + (1 + α2 )
p
= 2α(α + 1 + α2 ) + 1 = 2αA3 + 1
and
p
A24 = A23 (f α) = 2f α(f α + 1 + f 2 α2 ) + 1 = 2f αA4 + 1 (11.58)
We may find that
A23 − 1
2α = (11.59)
A3
A24 − 1
2f α = (11.60)
A4
Combining (11.59) and (11.60), we may find that the following expression,
A23 − 1 A2 − 1
2α = = 4
A3 f A4 (11.61)
Substituting
A2 exp(A1 )
A3 =
1 − f exp(A1 )
A2
A4 =
1 − f exp(A1 )
into (11.61), we obtain the following identity,
A23 − 1 A2 − 1 A2 exp(2A1 ) − (1 − f exp(A1 ))2 A2 exp(A1 )
= 4 ⇒ 2 =
A3 f A4 A22 − (1 − f exp(A1 ))2 f A2
Rewrite the above equation,
f A22 exp(2A1 ) − f (1 − f exp(A1 ))2
= A22 exp(A1 ) − exp(A1 )(1 − f exp(A1 ))2
⇒ A22 exp(A1 )(1 − f exp(A1 )) = (1 − f exp(A1 ))2 (exp(A1 ) − f )
which leads to
A22 = (1 − f exp(A1 ))(1 − f exp(A1 ))
h i
= 1 + f 2 − f exp(A1 ) + exp(−A1 )
= 1 + f 2 − 2f cosh A1
We finally link A1 and A2 in a single equation.
Substituting the expressions of A1 and A2 into the above equation, we have

the desired result,
Σ2eq √3 Σ
αα
F (Σeq , Σαα , f ) = 2 + 2f cosh − (1 + f 2 ) = 0 .
σy 2 σy (11.62)
On the other hand, if we rewrite (11.50) as,

√
3 Σαα h α + √1 + α 2 i
= log p = Arcsinh(α) − Arcsinh(f α)
2 σy f α + 1 + f 2 α2
p p
= Arcsinh(α 1 + α2 f 2 − f α 1 + α2 )
Therefore,
√3 Σ
αα
p p
sinh = α( 1 + f 2 α2 − f 1 + α2 ) (11.63)
2 σy
Consider
Ω̇ 1
α = √ (11.64)
Ω 3Ė33
Σeq p p
= 1 + f 2 α2 − f 1 + α2 (11.65)
σy
Eq. (11.63) can be rewritten as
√3 Σ Ω̇ 1 Σeq
αα
sinh = √
2 σy Ω 3Ė33 σy
or
Ω̇ √ σ √3 Σ
y αα
= 3Ė33 sinh
Ω Σeq 2 σy
Ω̇
Considering the fact f˙ = f (1 − f ), we recover the McClintock solution,
Ω
√ σ √3 Σ
y αα
f˙ = 3f (1 − f )Ė33 sinh
Σeq 2 σy (11.66)
11.3.2 Gurson-Tvergaard-Needleman model

σ
eq
Φ= (11.67)
σy
11.4 Exercise
References
Ablowitz, M. and Fokas, A. S. (1997), Complex Variables: Introduction and Applications, Cam-
bridge University Press.
Cottrell, A. H. (1953), Dislocations and Plastic Flow in Crystals, Oxford University Press.
Eshelby, J. D. (1957) “The determination of the elastic field of an ellipsoidal inclusion and
related problems,” Proceedings of Royal Society of London, A 241, pp. 376-396.
Eshelby, J. D. (1959), “The elastic field outside an ellipsoidal inclusion,” Proceedings of Royal
Society of London, A 252, pp. 561-569.
Hashin, Z. and Shtrikman, S. [1961], “Note on a variational approach to the theory of composite
elastic materials,” The Franklin Institute Laboratories, 271, pp. 336-341.
Hashin, Z. and Shtrikman, S. [1962], “On some variational principlrd in anisotropic and nonho-
mogeneous elasticity,” Journal of Mechanics and Physics of Solids, 10, pp. 335-342.
Hashin, Z. and Shtrikman, S. [1962], “The elastic moduli of heterogeneous materials,” Journal
of Applied Mechanics, pp. 143-148.
Hashin, Z. (1965), “On elastic behaviour of fibre reinforced materials of arbitrary transverse
phase geometry,” Journal of Mechanics and Physics of Solids, 13, pp. 119-134.
Hirth, J. P. and Lothe, J. (1982). Theory of Dislocations, Krieger
Love, A. E. H. (1926). A Treatise on the Mathematical Theory of Elasticity, Fourth Edition,
Dover Publication, New York.
Malvern, L. E. (1969). Introduction to the Mechanics of a Continuous Mdedium, Prentice-Hall.
Mura, T. (1987). Micromechanics of Defects in Solids, Second, Revised Edition, Kluwer Aca-
demic Pub. Dordrecht/Boston/London
Nemat-Nasser, S. and Hori, M. (1999), Micromechanics: overall properties of heterogeneous
materials, Elsevier, Amsterdam-Lausanne-New York
Sneddon, I. N. (1950), Fourier Transforms, Dover Publication, Inc. New York.
Sokolnikoff, I. S. (1956). Mathematical Theory of Elasticity, 2n edition, McGraw-Hill
Talbot, D.R.S. and Willis, J.R. [1985], “Variational principles for inhomogeneous non-linear
media,” IMA Journal of Applied Mathematics, 35, pp. 39-54.
Talbot, D.R.S. and Willis, J.R. [1998], “Upper and lower bounds for the overall response of an
elastoplastic composite,” Mechanics of Materials, 28, pp. 1-8.
Timoshenko, S. and Goodier, J. N. (1951). Theory of Elasticity, 2nd ed. McGraw-Hill
Titchmarsh, E. C. (1948). Introduction to the Theory of Fourier Integrals, Oxford.
REFERENCES 331
Zhdanov, M. (1988). Integral Transforms in Geophysics, Springer-Verlag,

Berlin / Heidelberg / New York

高等固体力学

Uploaded by

Copyright:

Available Formats

高等固体力学

Uploaded by

Document Information

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

高等固体力学

Uploaded by

Copyright:

Available Formats

INTRODUCTION TO MICROMECHANICS

8.8 Exercises 227

What is micromechanics ? Generally speaking, micromechanics is a scien-

composite/synthetic materials, e.g. composite structures, cementitious materi-

2.1 Vectors and Tensors

A × B = (Ai ei ) × (Bj ej ) = Ai Bj ei × ej = ekij Ai Bj ek (2.4)

where ei × ej = ekij ek , and ekij is called the permutation symbol,

e132 = (−1)e123 = (−1)(1) = −1

A × B = ekij Ai Bj ek = e1ij Ai Bj e1 + e2ij Ai Bj e2 + e3ij Ai Bj e3

ei × ej = ekij ek , ⇒ ekij = (ei × ej ) · ek (2.7)

1 When i = r, eijk eist = δjs δkt − δjt δks ;

2 When i = r and j = s, eijk eij` = 2δk` ;

3 When i = r, j = s, and k = t, eijk eijk = 3! = 6.

which are call e − δ identities.

2.1.2 Tensor Algebra

Figure 2.1. Cartesian Coordinate

1 One may call the vector as the first order tensor.

A conjugate of a dyad (second order tensor) is defined as

A : B = (Aij ei ⊗ ej ) : (Bk` ek ⊗ e` ) = Aij Bk` (ei · ek )(ej · e` )

The trace of a second order tensor is defined as

trA := A : 1(2) = Aii = A11 + A22 + A33 (2.18)

In component form, σij = Cijk` k` .

A second order tensor is skew symmetric, if

In general, an arbitrary second order tensor can be expressed as

T (2s) = {S S = Sij eSij } (2.27)

The corresponding second order symmetric unit tensor is then defined as

T (4s) = {S S = Sijk` eSijk` } (2.30)

The corresponding fourth-order unit tensor is defined as

2.1.3 Inversion formula for fourth-order isotropic tensor

A more straightforward approach to invert an isotropic tensor is to adopt the

The E-bases have the following special properties,

E(1) + E(2) = 1(4s)

We now use E-basis approach to verify Sherman-Morrison formula. Let,

Q = (3m + 2w)E(1) + 2wE(2) (2.42)

Q : Q−1 = 1(4s) = E(1) + E(2)

which then leads to

Q−1 = (h − v)E(1) + v(E(1) + E(2) )

C = λ1(2) ⊗ 1(2) + 2µ1(4s)

Since by definition, C : D = 1(4s) , it can be readily shown that

Example 2.2 For spherical inclusion, the Eshelby tensor is

2.1.4 Tensor analysis

In general for a smooth tensorial field, A, we have the following statement,

Consider a continuous m-order tensorial field, A(x) ∈ [C 1 (Ω)]m × d, the

If A is a vector field, i.e. A = Ai ei , the divergence theorem can be expressed

If we consider the volume integration of a cross product between gradient

2.2 Review of Linear Elasticity Theory

σ = C :  ⇒ σij = Cijkl kl (2.62)

where C = Cijkl ei ⊗ ej ⊗ ek ⊗ el is the elasticity tensor.

One may derive that

2.2.1 Betti’s reciprocal theorem and Somigliana Identity

Figure 2.2. Two sets of different self-equilibrating states

with boundary conditions,

2 Precisely speaking, it is the Betti’s second reciprocal theorem.

Similarly, one may show that

is called Betti’s first reciprocal theorem.

The first property (2.79) can be easily shown by definition that

To show the second property, we let x − y = z and dy = −dz. Thus

where −1 < ζ < 1.

Figure 2.3. Dirac’s delta function

Consider an infinitely space filled with homogeneous elastic medium. The

In component form, σij = Cijk` k` .

σ = C : ⇒ σij = Cijkl kl (2.62)

< >= 0 , ⇒ < ij >= 0ij (3.13)

ij (x) 6= 0ij

and the perturbation strain field satisfying ˜ij (x) = 0, ∀x ∈ ∂V .

<σI : > − < σ >:< >

Since σij δij = 12 σij (δui,j + δuj,i ) = σij δui,j ,

d (x) = D : (σ d (x) − σ ∗ ), ∀x ∈ V (3.30)

d (x) = D : (σ d (x) − σ ∗ ) ⇒ σ d = C : d + σ ∗ (3.32)

Therefore, ¯M = D : σ̄, and¯Ω = DΩ : σ̄ Ω .

(x) = 0 + d = AΩ : ∗ = AΩ : (AΩ − SΩ )−1 : 0

(x) = AΩ : 0 , ∀x ∈ Ω ⇒ ¯Ω = AΩ : 0 (3.49)

σ d = C : (d − σ ∗ ) , and σ ∗ + C : ∗ = 0 . (3.64)