Tensor Properties of Materials

Lecture Notes
– 2023 –

Bálint Koczor

email: [Link]@[Link]
Acknowledgments

I would like to thank Professor G. A. D. Briggs who previously lectured this course in Oxford
and provided helpful advice. These lecture notes build on his previous notes.

Contents

1 Linear properties of materials 3

1.1 Linear response in materials . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

1.2 Isotropic and anisotropic properties . . . . . . . . . . . . . . . . . . . . . . . 5

2 Tensors 6

2.1 Thought experiment in anisotropic materials . . . . . . . . . . . . . . . . . . 6

2.2 Einstein summation convention . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.3 What is a tensor? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

3 Coordinate system rotations 12

3.1 Basic rotations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

3.2 Orthogonal and rotation matrices . . . . . . . . . . . . . . . . . . . . . . . . 14

3.3 Passive rotation of vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

3.4 Rotations of tensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

3.5 Principal axis system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

4 Visualisation and Crystal Symmetries 22

4.1 Implicit equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

4.2 Visualising tensors as surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . 23

4.3 Examples of surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

1
 4.4 Rotational properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

4.5 Effect of crystal structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

5 Applications of tensors 30

5.1 Stress tensor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

5.2 Stress tensor in equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

5.3 Principal stresses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

5.4 Strain in 1D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

5.5 Strain in 3D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

5.6 Pure shear stress and strain . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

5.7 Strain tensor via thermal expansion . . . . . . . . . . . . . . . . . . . . . . . 37

5.8 Hooke’s law and linear elasticity . . . . . . . . . . . . . . . . . . . . . . . . 38

Appendix 39

A Constructing rotation matrices . . . . . . . . . . . . . . . . . . . . . . . . . 39

B Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

B.1 Solution of Question 5 . . . . . . . . . . . . . . . . . . . . . . . . . . 39

B.2 Solution of Question 7 . . . . . . . . . . . . . . . . . . . . . . . . . . 40

2
1 Linear properties of materials

1.1 Linear response in materials

In physics we often use linear equations to quantitatively describe that some external cause
results in a proportional response of the system. These relationships can be written formally
as a linear equation of the form

effect = proportionality factor × cause.

Here the cause is an external ‘force’ that we apply to our system (or material), while the
proportionality factor is a constant property of our material. Given knowledge of these two
quantities we can predict exactly how much is the resulting ‘effect’. To be more concrete, let
us consider now a few examples. First, recall that an electrical conductor has the property
that if we apply an electric field E to it then a current density J will flow through it whose
magnitude is proportional to E. The proportionality factor is the conductivity as

J
|{z} = σ
|{z} × E
|{z} .
current density conductivity electric field

We could of course similarly write inverse relations, such as E = ρJ which uses the resistivity
ρ = 1/σ as the (inverse) proportionality factor. Another example could be the electric
susceptibility χ of a material which describes the proportionality between the electric field
E and the electric polarisation P as

P = ϵ0 χE.

Yet another example is the thermal expansion coefficient α which expresses the proportion-
ality between the extension of a material ϵ and the temperature difference ∆T we apply to
the material as
ϵ = α∆T.
Recall that the extension is actually a strain ϵ and Hook’s law relates it to the stress σ
via the proportionality factor C that quantifies the stiffness of the material resulting in yet
another linear relationship as
σ = Cϵ. (1)
The above equation is also often expressed in an inverse form as ϵ = sσ where s = 1/C is
the elastic compliance.

These equations are, however, only idealisations as in reality the response of a material
is never perfectly linear. A response curve of a material is illustrated in Fig. 1(left): When
the strain in the material is small then we see a perfectly linear strain-stress relationship
as expected from Hook’s law in Eq. (1). Depending on the properties of the material large

3
Figure 1: The cause and effect relationship is linear in the region where the material is only
subject to a small cause. (left) Stress in different materials as a function of strain. The size
of the linear region depends on properties of the material while the slope of the linear region
is the stiffness of the material. (right) Current density as a function of the applied electric
field. The proportionality factor is the electric conductivity σ and is the slope of the linear
region. The material with the steeper slope σ2 is the better conductor.

stresses may cause plastic deformations and we may see a deviation from Hook’s law sooner
or later. Similarly, we can imagine that if we induce a large enough current density in a
conducting material then it might lead to a heating up. This heating up of the material then
leads to a ‘proportionality constant’ that depends on the temperature ultimately leading to
a non-linear response. If the current density is too large than heating may even lead to a
complete breakdown.

We can mathematically formalise these observations using a Taylor expansion. In com-

plete generality we can state that the effect f (x) is a function of the cause x. For example,
the current density is a function of the electric field as f (x) = J(E). If we assume such a
response function is smooth then we can express the dependence using a Taylor series as

f (x) = f (0) + f ′ (0)x + 21 f ′′ (0)x2 + . . . .

When the cause, here expressed as the variable x, is ‘small’ then the effect is almost perfectly
linear as in the second third term the square x2 of the small x can be negligibly small. Thus,
the slope in the linear region in Fig. 1 is determined by the derivative f ′ (x) = σ and this
derivative is the proportionality in our cause-effect relations.

4
1.2 Isotropic and anisotropic properties

In Section 1.1 we considered several physical phenomena where a material is subject to an

external cause which results in a linear response. However, both the cause and effect were
expressed as a number and not as a vector. In a real, three-dimensional world a physical
quantity may have both magnitude and direction and we need to use vectors to express
them. It is often the case that the direction of the cause and effect are parallel, for example
J⃗ = σ E,
⃗ and the property of the material σ does not depend on the direction – resulting in
the same cause-effect relationship from whichever direction we apply the electric field. For
this reason, the proportionality factor, for example the conductivity σ, is a global property
of the material and we can describe it just using a scalar. A material in which macroscopic
properties are independent of the direction is called an isotropic material. Isotropic materials
include the following.

• amorphous materials, for example glasses

• small-grained polycrystalline materials: even if the individual, small crystals are not
isotropic, their properties average out on a macroscopic scale and polycrystalline ma-
terials are thus isotropic

• some crystals are isotropic due their high symmetry, for example linear response in
cubic crystals is isotropic

On the other hand, if a material is anisotropic then both the magnitude and the direction
of the effect depends on the directionality of the cause. In such a case we need to describe
physical variables in terms of vectors and the proportionality factors in terms mathematical
objects that can take into account directionality. We are interested in anisotropy because it
is the key to so many technologies and applications, such as the following.

• Liquid crystals: appear in many applications, such as LCD screen, displays, digital
watches

• fibre composites: many construction materials, such as wood, reinforced concrete

• single crystals: most crystal lattices

• Piezoelectricity: quartz oscillators are very commonly used, for example, for generating
the clock frequency in computers. They are also used in gas lighters, stepper motors,
high-precision positioning, sensors, but also in microphones

• many optical devices use anisotropy, such as polarisers, beam splitters

5
Figure 2: Thought experiment: We apply an electric field E ⃗ to an anisotropic material from
the three coordinate directions and measure the resulting components of the current density
⃗
vector J.

2 Tensors

2.1 Thought experiment in anisotropic materials

In the previous section we considered isotropic materials which have the property that a suffi-
⃗ results in a proportional
ciently small external cause, for example an external electric field E,
⃗
response that is independent of the direction of E. We now consider an anisotropic material in
which the response, in this example the current density, is not parallel with the electric field.
We use vectors to describe both the effect J⃗ = (Jx , Jy , Jz ) and the cause E⃗ = (Ex , Ey , Ez )
and derive a higher mathematical object that correctly describes the anisotropy of the con-
ductivity, which was previously just a proportionality factor σ.

Suppose we apply an electric field to a conductor such that E ⃗ is parallel with the x
direction of our coordinate system. We measure the resulting current density. We illustrate
in Fig. 2(left) that the vector J⃗ that we measure has non-zero x, y and z vector entries and
⃗ Assuming that the electric field is small enough such that the
it is thus not parallel with E.
response of the material is linear, we can write down the following relation between cause
and effect as
⃗ = (Ex , 0, 0)
E → Jx = σxx Ex , Jy = σyx Ex , Jz = σzx Ex . (2)

The proportionality factor between cause (E) ⃗ and effect (J)

⃗ is the conductivity of the ma-
terial and we needed three such proportionality factors σxx , σyx and σzx for the following
reasons. First, if we repeat the experiment and apply an electric field twice as large Ex′ = 2Ex ,
we should measure a current density twice as large |J| ⃗ ′ = 2|J|
⃗ due to linearity. Second, the
two vectors in the two different experiments, J⃗′ and J, ⃗ should point to the same direction
(since the electric field had the same direction in both experiments) and only their magnitude

6
should be different by a factor of 2. It is easy to verify that both properties are satisfied by
Eq. (2). This is the reason why the proportionality factor must depend on two indexes, for
example, σyx denotes the linear response in the y direction when an electric field is applied
in the x direction.

We can repeat this thought experiment but now applying the electric field in the y
direction as illustrated in Fig. 2(middle). We measure the following current densities as

⃗ = (0, Ey , 0)
E → Jx = σxy Ey , Jy = σyy Ey , Jz = σzy Ey . (3)

Here we had to introduce three more proportionality constants as the conductivities σxy , σyy
and σzy to describe the material’s response when the external electric field is applied only
from the y direction. Finally we repeat the thought experiment for an electric field applied in
the z direction as illustrated in Fig. 2(right). To describe the linear response of the material
we need to introduce yet again three more conductivities as
⃗ = (0, 0, Ez )
E → Jx = σxz Ez , Jy = σyz Ez , Jz = σzz Ez . (4)

What happens when the electric field is applied from a direction that does not coincide
with one of the axis vectors? For example, when the electric field vector is described by
some arbitrary vector as E⃗ = (Ex , Ey , Ez )? The answer is simple when the cause and effect
are small enough so that the response of the material is linear. In this case all contributions
in the x, y and z directions add up independently of each other. For example, we can
sum all three contributions to the x component of the current density from the above three
equations as Jx = σxx Ex + σxy Ey + σxz Ez . We can similarly do so for all components of the
current density vector and find the complete response of the material is described by the
linear system of equations as

Jx = σxx Ex + σxy Ey + σxz Ez , (5)

Jy = σyx Ex + σyy Ey + σyz Ez ,
Jz = σzx Ex + σzy Ey + σzz Ez .

2.2 Einstein summation convention

Eq. (5) describes a linear cause-effect relationship of an anisotropic material. In the

following we want to write such equations more compactly and thus we need a way to distil
the meaning of Eq. (5). We first introduce a more compact notation for axis vectors and
corresponding vector entries.

7
 Reminder 1. We denote the axis vectors of a right-handed Cartesian coordinate system
as ⃗xi where the free index i can take up any value between 1 and 3 as ⃗x = ⃗x1 , ⃗y = ⃗x2
and ⃗z = ⃗x3 . We index the corresponding vector entries of vectors using numbers, for
example Ji = J⃗ · ⃗xi , where again the index i can take up any values between 1 and 3 as
Jx = J1 , Jy = J2 and Jz = J3 .

This indexing scheme allows us to re-write Eq. (5) into the more compact form as
3
X
Ji = σij Ej . (6)
j=1

We can even further simplify this equation by using the Einstein summation convention: We
drop the summation sign and write Eq. (6) compactly as Ji = σij Ej because by convention
we imply that the repeated indexes are summed over. For example, in the expression
σij Ej the index j appears twice and thus we interpret this expression as
3
X
σij Ej = σij Ej = σi,1 E1 + σi,2 E2 + σi,3 E3
j=1

Furthermore, the free index i can take up any value between 1 and 3 and thus both sides of
the equation Ji = σij Ej refer to 3 entries of a vector, for example, the left-hand side is Ji
⃗
refers to any of the three vector entries of J.

Another way of compactly writing Eq. (5) is by recognising that the 9 numbers σij can
be arranged into a matrix of size 3 × 3. By convention the first index i indexes the columns
of the matrix whereas the second index j indexes the rows. As such, the right-hand side of
Eq. (5) expresses a matrix multiplying a column-vector as
    
J1 σ11 σ12 σ13 E1
J2  = σ21 σ22 σ23  E2  ,
J3 σ31 σ32 σ33 E3
which we can more compactly write as J⃗ = σ E
⃗ where the matrix σ is denoted with bold
symbols.

Summary 1. The conductivity in an anisotropic material (as illustrated in Fig. 2)

can be compactly described using either a matrix-vector multiplication or an Einstein
summation as

matrix/vector: J⃗ = σ E,
⃗ Einstein summation: Ji = σij Ej .

We denote column vectors as E ⃗ and refer to their vector entries as Ei while we denote
matrices as σ and refer to their matrix entries as σij .

8
Figure 3: Tensors are rank-r arrays of numbers. For small ranks we can identify them with
scalars (r = 0), vectors (r = 1) and matrices (r = 2). For ranks 3 and higher tensors become
more complex to visualise.

2.3 What is a tensor?

We have introduced the Einstein summation convention in the previous section that allowed
us compactly write equations where we sum over products of vector entries Ei and matrix
entries σij . While we can refer to the collection of matrix entries σij as a matrix, it is
important to note that matrices are not just a collection of numbers as they have some
special properties. For example we can make sense of a matrix-matrix multiplication or the
inverse of a matrix. If we forget about these special properties then σij is nothing more than
just a 3 × 3 array of numbers.

Simply put, a tensor, such as σij , is an array of numbers and its rank refers to the number
of free indexes it has which we illustrate in Fig. 3. For example, the collection of vector entries
of a vector, such as Ti , is a rank-1 tensor. Similarly, the collection of matrix entries Tij of a
matrix is a rank-2 tensor. It is however important that matrices are not equivalent to rank-2
tensors as matrices have special properties (matrix multiplication, matrix inverse or matrix
determinant) that tensors do not have. In another sense though tensors can be viewed as
generalisations of vectors and matrices because we can define tensors with higher rank. For
example, a three-dimensional array of numbers is a rank-3 tensor which then has 3 free
indexes as Tijk .

Why are we interested in using tensors? The reason is they can be a very useful set
of tools in so many different applications. For example, we have already seen that we can
express linear but anisotropic properties of materials using tensors of rank-2 (matrix) and
rank-1 (vector). Furthermore, we will see that some properties of materials are described
by higher rank tensors, for example in Section 5.8 we will use rank-4 tensors to describe the
stiffness of materials.

As an aside, tensors are also very useful in applications not related to materials. For

9
 explicit Einstein conv. matrix notation type

a = 3i=1 vi wi
P
a = vi wi a = ⃗v · w
⃗ vector/vector

ai = 3j=1 Tij vj
P
ai = Tij vj ⃗a = T · ⃗v matrix/vector

Rik = 3j=1 Sij Tjk

P
Rik = Sij Tjk R=S·T matrix/matrix

Rij = 3k,l=1 Sijkl Tkl

P
Rij = Sijkl Tkl

Table 1: Examples of equations involving tensors. We can explicitly write out a summation
(left) that involves products of tensor entries. These can be simplified using the Einstein
convention by dropping the summation signs – we assume all indexes are summed over that
appear more than once. When the tensors involved are rank-1 or rank-2 then we can use
convenient vector and matrix multiplications. When the rank of any of the tensors involved
is higher than 2 then we need to resort to Einstein summations.

example, a widespread machine-learning software is called Tensor Flow. The reason is that
weights in neural networks are modelled using tensors and Google even developed bespoke
hardware for manipulating tensors (called tensor processing units) that have been used, for
example, for text recognition in Street View. Tensors are also extremely useful in atomistic
simulations of physical or chemical systems, for example, in approximating the wavefunction
of a molecule. All these applications actually use tensors of rank much higher than 2.

In this lecture we are primarily motivated to go beyond matrices (rank-2) because certain
properties of materials are described using higher-rank tensors. It was very convenient in
Section 2.2 to use compact and intuitive matrix and vector multiplications, however, for
ranks larger than 2 we must resort to tensors and Einstein summations. We will particularly
focus on so-called Cartesian tensors. The reason is that in materials science and in physics we
work in a three-dimensional, orthogonal, Cartesian coordinate system spanned by the axis
vectors ⃗x1 , ⃗x2 and ⃗x3 as introduced in Reminder 1. For example, the number E1 refers to the
electric field vector entry that we would measure along the ⃗x1 coordinate axis. Furthermore,
we used the tensor entry σ13 to refer to the conductivity as the proportionality factor between
the electric field that we applied in the direction ⃗x3 and the resulting current density that we
would measure in the direction ⃗x1 . In general, all indexes of a Cartesian tensor (for example
i, j and k of Tijk ) refer to the corresponding coordinate axes of a Cartesian coordinate system
1
. We can use the Einstein summation to conveniently write various different matrix and
vector operations as we summarise in Table 1.
1
See the Wikipedia page for more information on Cartesian tensors.

10
Summary 2. A Cartesian tensor of rank r is an r-dimensional array of real numbers
with r free indexes that each can take up values 1, 2 and 3 as they refer to the three
axis vectors of a Cartesian coordinate system. When multiple tensors appear in an
expression then we assume a summation over any index that appears more than once.
For example, in the equation Rij = Sabi Tabj we sum over the repeated indexes a and b
as
3 X
X 3
Rij = Sabi Tabj = Sabi Tabj .
a=1 b=1

We have two free indexes, i and j, thus the result is a rank-2 tensor which we denoted
as Rij .

Question 1. As discussed, the expression σij Ej refers to a matrix multiplying a column

vector. What does the analogous expression σij Ei refer to? Hint: we can use the
transpose of the matrix (tensor) as (σ T )ij = σji .

11
Figure 4: (left) A rotation around the z coordinate axis. By convention, rotations are right-
handed. (right) the effect of a z rotation on a vector that is parallel with the x coordinate
axis.

3 Coordinate system rotations

Properties of isotropic media do not change when we rotate the material or the external
coordinate system. Anisotropic media, on the other hand, do depend on directionality. For
example, when we apply an electric field E ⃗ the resulting current density J⃗ is not necessarily
parallel in an anisotropic material and thus we need to treat these quantities as vectors.
However, as opposed to scalar quantities in isotropic materials, the vector entries of both E ⃗
⃗
and J are always relative to a coordinate system. As such, even when the physical vectors are
fixed, their vector entries do change as we change our coordinate system. We actually have
complete freedom in choosing the coordinate system and our main motivation for doing so is
that we can simplify the description of anisotropic materials when we view them in a special
coordinate system. For this reason in this section we develop the necessary mathematical
tools for compactly describing rotations using rotation matrices.

3.1 Basic rotations

Matrices are mathematical objects that we can view as a collection of rules how to take
an ‘input’ vector and map it to an ‘output’ vector – more precisely, a matrix multiplies a
vector V⃗ ′ = LV⃗ and transforms its vector entries so that we obtain the vector V⃗ ′ . We now
specifically focus on rotations of vectors and we will denote their transformation matrices as
L.

To illustrate, let us take a simple example of a rotation around the z coordinate axis. We

12
will use the right-hand rule as illustrated in Fig. 4(left) and a variable rotation angle that we
denote with the symbol θ. We now want to analyse how the rotation matrix L transforms
the entries of a vector via the matrix-vector multiplication LV⃗ .

Assume we are given a vector and for the sake of argument we assume this vector is
parallel with the ⃗x1 axis vector as V1⃗x1 with length V1 . Fig. 4 illustrates the rotation of this
vector around the z axis using the right-hand rule with a rotation angle θ. From Fig. 4(right)
it is clear that the new vector after the rotation has the vector entries
   
V1 cos θ
L  0  = V1  sin θ  . (7)
0 0

This is of course a quite special case because our vector was chosen to be parallel with the
⃗x1 axis. For this reason we repeat the previous argument but now the vector we rotate is
chosen such that is parallel with the ⃗x2 axis as V2⃗x2 . We find the coordinates after rotation
as    
0 − sin θ
L V2  = V2  cos θ  . (8)
0 0
We could also repeat the previous argument with a vector that is parallel with the ⃗x3 axis,
however, it would be unchanged under our rotation around the z axis. For this reason we
can write    
0 0
L 0 = V3 0 .
   (9)
V3 1
But how about the rotation of any vector? For example an arbitrary vector that we can
write as V⃗ = (V1 , V2 , V3 )? We actually already have all ingredients to answer this question
and we just need to combine the above equations together. First, we denote the vector that
we obtain after the rotation with a prime as V⃗ ′ = LV⃗ . Second, we write our original vector
V⃗ as a sum of its components as V⃗ = V1⃗x1 + V2⃗x2 + V3⃗x3 and apply the rotation to the
individual components
     
V1 0 0
⃗ ′ ⃗
V = LV = L 0 + L V2 + L 0 .
     (10)
0 0 V3

We have already evaluated the rotation of the first component in Eq. (7), the rotation of
the second component in Eq. (8), and the rotation of the third component in Eq. (9). We
just copy back the right-hand sides of those equations and find our desired matrix-vector
equation as      
cos θ − sin θ 0
⃗ ′
V = V1 sin θ + V2 cos θ + V3 0
     (11)
0 0 1

13
One can show (we will discuss this in the next section) that due to Eq. (10), the column
vectors in Eq. (11) must be the column vectors of the transformation matrix L, and thus L
must have the form  
cos θ − sin θ 0
L =  sin θ cos θ 0 ≡ Rz .
0 0 1
This is actually the elementary z rotation matrix which most textbooks use as a mathematical
definition of the rotation around the z axis, see for example the Wikipedia page. While above
we evaluated the matrix of rotations around the z axis, similarly we could repeat our analysis
and rotate vectors around the x or y axes to find the other two basic rotation matrices.

Summary 3. The three basic rotation matrices rotate any column vector by an angle
of θ around one of the three Cartesian coordinate axes as
     
1 0 0 cos θ 0 sin θ cos θ − sin θ 0
Rx = 0 cos θ − sin θ , Ry =  0 1 0 , Rz =  sin θ cos θ 0 .
0 sin θ cos θ − sin θ 0 cos θ 0 0 1
(12)

We discuss in Appendix A that we can construct any rotation matrix, for example, a rotation
around a tilted vector, by multiplying at most three basic rotations together.

Question 2. Verify the x and y basic rotation matrices by drawing their effect on
vectors as in Fig. 4.

Question 3. Suppose the homogeneous electric field between two large parallel charged
⃗ = (0.1, 0.5, 0.9). We rotate the two plates by 120◦
plates is described by the vector E
⃗ ′ after the rotation of the plates.
around the ⃗x3 axis. What is the electric field vector E

3.2 Orthogonal and rotation matrices

In the previous section we introduced basic rotations that rotate around one of the standard
axes, for example, around the x axis. How about a rotation around an arbitrary axis? For
example, a rotation around an axis that is halfway between the z and x axes? It turns out we
can express any rotation in terms of a 3 × 3 rotation matrix L. Conversely, any 3 × 3 matrix
is a rotation matrix if it satisfies the two properties: First, its row and column vectors must
be mutually orthogonal – such a matrix is called an orthogonal matrix. Second, a rotation

14
matrix L has determinant one. Both properties are important as we can use the rotation
matrix to rotate a set of coordinate axes (basis vectors) as ⃗x′1 = L⃗x1 , ⃗x′2 = L⃗x2 , and ⃗x′3 = L⃗x3
into a new set of coordinate axes (we have denoted the new axes with a prime).

Note 1. The matrix entries Lij of the transformation matrix are actually cosines of
angles between the axes ⃗xi and ⃗x′j . The reason is the following. The coordinate axis ⃗x′j
is the j th column vector of L and the ith vector entry of ⃗x′j is given by the scalar product
⃗xi · ⃗x′j . This scalar product is actually given by the cosine of the angles between the old
axes ⃗xi and new axes ⃗x′j and thus Lij = cos θij . For example, θ12 is angle between the
old axis ⃗x1 and the new axis ⃗x′2 .

Property 1 (Determinant): The property that L must have determinant 1 ensures that
a right-handed basis remains right handed (with determinant −1 it could be changed to a
left-handed one).

Property 2 (Orthogonality): Without loss of generality, we can assume that ⃗xi are just
the standard coordinate axes in which case the new coordinate axes ⃗x′i are column vectors
of the rotation matrix. For example, ⃗x′2 is the second column vector as
    
L11 L12 L13 0 L12
′
⃗x2 = L⃗x2 = L21 L22 L23
   1 = L22  .
 
L31 L32 L33 0 L32
In some sense we have already verified this property: in the example in Eq. (7) we evaluated
how the z rotation acts on the vector V1⃗x1 and we found it is rotated into V1 times the
first column vector of the rotation matrix. Similarly, in Eq. (8), and in Eq. (9) we obtained
vectors proportional to the second and third column vectors of L.

In general, the column vectors of L form a coordinate system and it is therefore necessary
that they be mutually orthogonal. For example, we want that the scalar product between
the first and second axes (column vectors) are zero as
Li1 Li2 = ⃗x′1 · ⃗x′2 = 0,
while we also want the axes (column vectors) be normalised as
Li1 Li1 = ⃗x′1 · ⃗x′1 = 1.
We can summarise all equation of this kind very compactly using our Einstein convention as
Lij Lik = δjk , where the Kronecker delta symbol is a rank-2 tensor that evaluates to 1 when
j = k and otherwise it evaluates zero. The equation Lij Lik expresses the scalar product
between the basis vector ⃗x′j and ⃗x′k , which is indeed 1 if j = k and zero otherwise. In fact,
the Kronecker delta symbol is the identity matrix 1 in matrix notation and we can re-write
the equation as a matrix-matrix product using that Lij Lik = (LT )ji Lik which allows us to
finally state the defining property of orthogonal matrices.

15
 Summary 4. Orthogonal matrices have mutually orthogonal column and row vectors
and thus satisfy the property

LT L = 1, Lij Lik = δjk ,

using matrix notation and using tensor notation, respectively.

Most importantly, the above property immediately guarantees us that the transpose is the
inverse operation LT = L−1 . We will make use of this property, for example to undo the
effect of a rotation on a vector as LT V ′ = V .

Whenever we rotate vectors with a rotation matrix as LV⃗ we change its vector entries.
This can be interpreted as an active rotation that turns V⃗ into another vector V⃗ ′ . We now
detail in the next section that, at the same time, the transformation of the vector entries
by LV⃗ can be interpreted such that V⃗ remains unchanged and only the coordinate axes
are being transformed (via the inverse rotation LT ) into a new set of coordinate axes as
⃗x′i = LT ⃗xi . This is a passive rotation and it changes the vector entries because the vector
entries are always relative to a coordinate system.

3.3 Passive rotation of vectors

In Section 3.1 we interpreted the rotation matrix L such that it acts on a vector V⃗ and
rotates it into another vector V⃗ ′ but we described the vector entries of both vectors using
the same coordinate system. This is called an active rotation. At the same time, rotation
matrices also express coordinate axis transformations. For example, if we take a vector that
⃗ then we would expect this vector is
represents something physical, such as an electric field E,
unchanged as a physical quantity should not change when we change our coordinate system.
We can repeat the thought experiment in Section 3.1 but now rotating the coordinate system
with a negative angle while keeping the vector unchanged V⃗ as illustrated in Fig. 5 (right) –
surprisingly we find the process results in exactly the same transformation of vector entries.
This is called a passive rotation.

We can show mathematically that the rotation of a vector and the (inverse) rotation of
a coordinate system has the same effect on the vector entries as we always specify vector
entries relative to a coordinate system 2 . First, a passive rotation rotates the coordinate
vectors with an inverse rotation as, for example, ⃗x′1 = LT ⃗x1 . We can calculate the entries of
2
See also the discussion in Sec 4.2 in: Goldstein, H., C. Poole and J. Safko, Classical Mechanics, 3rd
Edition, San Francisco: Addison Wesley, 2002

16
Figure 5: (left) an active rotation L keeps the coordinate system unchanged and rotates
the vector V⃗ into another vector V⃗ ′ = LV⃗ . (right) A passive rotation inversely rotates the
coordinate axes as ⃗x′i = LT ⃗xi but leaves the vector unchanged. Passive and active rotations
have the same effect on the vector entries: The vector entries of the unchanged vector V⃗ in
the new coordinate system are exactly those of V⃗ ′ .

the unchanged vector J⃗ in the new coordinate system using the scalar product, for example

J1′ = J⃗ · ⃗x′1 = J⃗ · LT ⃗x1 .

passive rotation: (13)

On the other hand, if we interpret the transformation as an active rotation as we did in

Section 3.1 then our transformation matrix formally rotates the vector J⃗ into a new vector
J⃗′ = LJ⃗ whose vector entries can be calculated in the unchanged coordinate system as
 
active rotation: J1 = J · ⃗x1 = LJ⃗ · ⃗x1 .
′ ⃗′
(14)

We can now immediately confirm that the vector entry J1′ is the same in case of both active
and passive rotations: in a scalar product it does not matter whether a matrix multiplies
the vector on the
left or the transpose of the matrix multiplies the vector on the right as
(L⃗u) · ⃗v = ⃗u · LT ⃗v . This confirms that indeed the right-hand sides of Eq. (13) and Eq. (14)
are equivalent.


Question 4. Show the equivalence of scalar products (A⃗u)·⃗v = ⃗u · AT ⃗v for any matrix
A and any vector ⃗u and ⃗v . Hint: rewriting the equation in the Einstein convention helps.

Since in this lecture we are interested in how the entries of a vector or a tensor change
under a transformation, rather than how the coordinate system is changed, we use the con-
vention of active rotations. This is also a more natural interpretation of rotation matrices:
a matrix acts on a vector and transforms it into another vector. We can of course inter-
changeably refer to the rotation matrix either as an active rotation or as an inverse a passive

17
rotation. However, by making one interpretation default any instruction, such as “apply a
right-handed x rotation with an angle π/2”, unambiguously identifies an active rotation of
a vector or the passive rotation with an angle −π/2 and not the other way around.

The interpretation as an active transformation is indeed the more common one, especially
in programming, scripting and mathematical software, see for example the Wikipedia article,
or the relevant wolfram and mathematica and Matlab pages.

Summary 5. A rotation matrix L satisfies LT = L−1 and detL = 1 and expresses a

transformation of vector entries due to a change of basis: Changing the coordinate axes
⃗x′i = LT ⃗xi transforms the vector entries of all vectors. This transformation has the same
effect on the vector entries as an active rotation of the vector as

J⃗′ = LJ,
⃗ Ji′ = Lij Jj . (15)

By convention, L multiplies column vectors from the left.

Question 5. Calculate the matrix L of the rotation that actively rotates column vectors
by an angle π/2 around the axis that is halfway between the ⃗x2 and ⃗x3 axes (the rotation
axis is in the direction [011]). See solution in Appendix B.1.

3.4 Rotations of tensors

We are now equipped with all tools to derive how tensors transform under a change of basis.
As in the case of vectors, we will be interested in how the entries of a tensor transform due
to a change of coordinate system and we thus derive the active rotation of a tensor.

Suppose a tensor that expresses the relation between a cause and an effect, for example
the relation between the current density J⃗ and electric field E.
⃗ Imagine we know all tensor
entries σij because we determined them in the experiment in Section 2.1 so we can write the
linear relationship
J⃗ = σ E.
⃗ (16)
When we determined these tensor entries in Section 2.1 we used a specific coordinate system
– we can refer to it as our original coordinate system. However, imagine that for some
reason we would like to re-write the above cause and effect in a different coordinate system,
for example, because we hope this way Eq. (16) will simplify – indeed we sill see later that in
certain coordinate systems the equation greatly simplifies. As we discussed in the previous
section, when we rotate our coordinate system then the vector entries of all vectors change
which we expressed using a rotation matrix that acts on the original vectors J⃗′ = LJ⃗ and

18
⃗ ′ = LE,
E ⃗ formally actively rotating them into new vectors. We can apply the inverse
rotation to transform the new vectors back to the original ones as J⃗ = LT J⃗ ′ and E
⃗ = LT E
⃗′
and substitute back the expressions for the original vectors in Eq. (16) as

LT J⃗ ′ = σLT E
⃗ ′.

Let us now multiply both sides of the above equation with L from left to obtain

LLT J⃗′ = LσLT E

⃗′

We can simplify the left-hand side of the equation by using the orthogonality of rotation
matrices as LLT = 1 to finally obtain the expression

J⃗′ = LσLT E
⃗ ′.

If here we denote the matrix σ ′ = LσLT then it is immediately clear that the above equation
expresses a cause and effect relation just like Eq. (16) but expressed in a new coordinate
system as
J⃗′ = σ ′ E
⃗ ′.

Summary 6. Changing coordinate axes transforms vector entries according to a trans-

formation matrix LJ⃗ as detailed in Summary 5. Entries of a rank-2 tensor (matrix) are
then transformed by rotating the tensor (matrix) using the same rotation matrix as

σ ′ = LσLT , σij′ = Lik σkl Ljl .

So far, we only made an informal statement that a tensor is an “array of numbers” but
have not yet precisely stated what is a tensor. The generalisation of the above transformation
rule is actually the main defining property of a general tensor.

Note 2. A Cartesian tensor is a rank-r array of real numbers Ti1 i2 ...ir with the following
property. Given a transformation of the Cartesian coordinate system that transforms
entries of vectors as Vi′ = Lij Vj , a tensor must transform according to (see the Wikipedia
page)
Ti′1 i2 ...ir = Li1 j1 Li2 j2 · · · Lir jr Tj1 j2 ...jr .

Question 6. Show that if we apply the general definition of a tensor in Note 2 to the
special case of a rank-2 tensor, we obtain our formula in Summary 6 as expected. Hint:
substitute r = 2 and substitute the indexes, for example i1 = i, i2 = j, etc.

19
3.5 Principal axis system

In the previous section we saw that rank-2 tensors transform according to the equation
σ ′ = LσLT when we change the coordinate system. Here a rotation L is always an orthogonal
transformation and maps a rank-2 tensor σ onto another rank-2 tensor σ ′ . We now want
to find a special rotation that maps our tensor to a diagonal form. This will help us to
better understand properties of the material because the rotated tensor reveals important
information about, for example, symmetries of a crystal. But, let us first recollect a very
important property of orthogonal matrices, namely, that they can be used to bring any
symmetric matrix into a diagonal form.

Reminder 2. Recall that a symmetric matrix M has real entries and satisfies MT =
M. Such matrices have real eigenvalues λi and arranging eigenvectors of M into the
orthogonal matrix L as columns allows us to write M as a “rotated” diagonal matrix
 
λ1 0 0
M = L  0 λ2 0  LT .
0 0 λ3

As a consequence of Reminder 2 we can diagonalise any rank-2 tensor that satisfies the sym-
metry σij = σji , or equivalently in matrix from σ T = σ, by rotating the coordinate system.
The rotation transforms our coordinate system into a special one, called the principal axis
system, in which the tensor becomes diagonal as
 pas 
σ1 0 0
LT σL =  0 σ2pas 0  ≡ σ P AS . (17)
pas
0 0 σ3

Here the column vectors of L are the eigenvectors ⃗xipas which we refer to as the principal
axes. The eigenvalues σipas are the principal components, for example, here σipas are principal
conductivities. These principal axes are eigenvectors of σ in the sense that they satisfy the
eigenvalue equation
σ⃗xipas = σipas⃗xipas . (18)
Interestingly, it is a direct consequence of this eigenvalue equation that along the principal
directions we find a parallel cause and effect. To show this we first write our general cause
and effect equation J⃗ = σ E ⃗ assuming that the cause, the electric field, is parallel with one
of the principal axes and thus substitute E ⃗ = ⃗x pas . We find the equation
i

J⃗ = σ⃗xipas ,

20
which is exactly the left-hand side of the eigenvalue equation in Eq. (18). For this reason we
substitute back the right-hand side of the eigenvalue equation in Eq. (18) as

J⃗ = σipas⃗xipas .

⃗
We now recollect that the electric field was parallel with the principal direction ⃗xipas = E
and thus find
J⃗ = σipas E.
⃗

In the above equation the anisotropic tensor property appears to behave like an isotropic
property (parallel cause and effect) – but note that this is only true along the principal direc-
⃗ is parallel with one of the three principal axes, and we would see anisotropic
tions, when E
behaviour for vectors of any other direction.

Question 7. Recall from Reminder 2 that we can write any symmetric matrix M
as a rotated diagonal matrix as M = LDLT . Although L must be an orthogonal
matrix, it is not necessarily a proper rotation. A proper rotation satisfies det L = 1.
Show that we can always choose L with det L = 1 and thus Reminder 2 can always
be interpreted as a proper rotation of a diagonal matrix. Hint: Using the properties
of the determinant as det(AB) = det(A) det(B) and det(AT ) = det(A) helps. See the
solution in Appendix B.2.

21
4 Visualisation and Crystal Symmetries

4.1 Implicit equation

By now we have developed all necessary mathematical tools to work with tensors and to anal-
yse their properties in the principal axis system. In this section we consider how properties of
materials determine tensor properties, such as symmetries of crystals. We will first develop
visualisation tools that are quite intuitive as they use simple geometric objects rather than
abstract matrices or vectors. The primary advantage of using these visualisation surfaces
is that they clearly reflect symmetry properties of tensors. We will see that single crystals
can have high symmetry which then manifests in tensor properties that we can directly infer
from their visualisations.

The geometric objects we will use are three-dimensional surfaces. Such surfaces can be
specified using so-called implicit equations or implicit functions. But first, what are implicit
functions? Usually we use explicit functions, such as f (x, y) = ax + b + y for variables x
and y. This equation tells us explicitly that the function value f (x, y) is computed for any
argument x and y just by multiplying x with a and adding b and y to the result. In contrast,
in an implicit function we relate the variables x and y to each other via an implicit equation
that sets constraints that the solution must satisfy. For example, a two-dimensional circle
of unit radius is defined by a set of points that satisfy the constraint
2D circle: x2 + y 2 = 1.
Indeed, this constraint is only satisfied by points that are of unit distance from the origin
which is the defining property of a unit circle as illustrated in Fig. 6(b). Similarly, we could
consider a circle whose radius is larger or smaller than 1 by rescaling both variables with
the desired radius r as x2 /r2 + y 2 /r2 = 1. Following a similar logic, we can write down an
implicit equation for an ellipse: we similarly rescale the x and y variables but we use two
different scaling factors a and b. The resulting implicit equation is
2D ellipse: x2 /a2 + y 2 /b2 = 1,
and it defines a 2D ellipse of width 2a and height 2b as illustrated in Fig. 6(b). In order to
specify surfaces in three dimensions we use three variables x, y and z. The generalisation
of the unit circle is then the unit sphere which we can specify similarly using the implicit
equation
3D sphere: x2 + y 2 + z 2 = 1,
which is illustrated in Fig. 6(c). The ellipsoid is the generalisation of the ellipse in three
dimensions and rescaling the variables x, y, and z, yields its implicit equation as
3D ellipsoid: x2 /a2 + y 2 /b2 + z 2 /c2 = 1.
In the following we want to extract an implicit equation from a tensor.

22
 (a) x2 + y 2 = 1 (b) x2 /a2 + y 2 /b2 = 1
1.0 1.0

0.5 0.5

0.0 0.0

-0.5 -0.5

-1.0 -1.0
-1.0 -0.5 0.0 0.5 1.0 -1.0 -0.5 0.0 0.5 1.0

(c) x2 + y 2 + z 2 = 1 (d) x2 /a2 + y 2 /b2 + z 2 /c2 = 1

Figure 6: Examples of implicit equations and visualisation of their solutions. (a) 2D circle
(b) 2D ellipse (c) 3D sphere (d) 3D ellipsoid.

4.2 Visualising tensors as surfaces

Assume we are given a fixed, rank-2 tensor as a symmetric matrix T. We want to extract a
3D surface from T that is characteristic to the tensor. We specify this surface as an implicit
equation that depends on the three variables x, y and z. First, we arrange the variables into
the coordinate (column) vector ⃗x = (x, y, z)T . Second, we multiply this column vector with
our tensor as ⃗v = T⃗x, similarly to how we described physical cause-effect relations. Finally,
we define our implicit equation using the scalar product ⃗x ·⃗v between the original coordinate
vector ⃗x and ⃗v . This scalar product is actually the length of the parallel component between
⃗v and ⃗x as illustrated in Fig. 7. We can now express our implicit equation mathematically

23
Figure 7: (left) The representation surface is the collection of points ⃗x that result in
the parallel component ⃗x · ⃗v = 1. Suppose a coordinate-dependent electric field ⃗x = E: ⃗
the representation surface of the conductivity tensor σ can be though of as the parallel
component ⃗x · ⃗v = J⃗ · E
⃗ between the current density and the electric field due to the cause-
⃗ ⃗ (right) In isotropic materials J⃗ is always proportional to the
effect relation ⃗v = J = σ E.
⃗ hence the representation surface is a
electric field with the constant proportionality as σ E,
sphere.

as
all ⃗x that satisfy: ⃗xT T⃗x = ⃗x · (T⃗x) = 1 for fixed T.
| {z }
⃗v

This equation is only satisfied by points ⃗x that produce a parallel component ⃗x · ⃗v of unit
length – and these form points on a surface in 3D that we will refer to as the representation
surface of the tensor T. The meaning of the scalar product ⃗x · ⃗v is illustrated in Fig. 7 using
the example of a cause-effect relationship whereby the vector ⃗x = E ⃗ is an electric field and
the tensor is a conductivity as T = σ. The vector ⃗v is then given by the current density
J⃗ = σ E
⃗ and the scalar product then refers to the length of the parallel component between
⃗ and effect (J).
the cause (E) ⃗

Summary 7. The representation surface of a symmetric, rank-2 tensor is the 3D surface

that consist of points ⃗x = (x, y, z)T that satisfy the implicit equation ⃗xT T⃗x = 1.

24
4.3 Examples of surfaces

Let us first consider the simplest case of an isotropic material. Recall that in an isotropic
material a tensor T can be treated as a scalar (or equivalently as a scalar multiple of the
identity matrix as T 1). In this case our representation surface can be simplified considerably
as
1 = ⃗x · (T⃗x) = T ⃗x · ⃗x.
Here the scalar product of the vector with itself is actually the distance of the corresponding
point frompthe origin as ⃗x · ⃗x = x2 + y 2 + z 2 . We thus find the implicit equation of a sphere
of radius 1/T as

Isotropic material: 1/T = x2 + y 2 + z 2

Let us now consider a tensor that is given with respect to a coordinate system that is
aligned perfectly with the Principal Axis System. In this case the tensor is diagonal with
three non-zero entries as T = diag(λ1 , λ2 , λ3 ) and thus our vector can be simplified as
⃗v = T⃗x = (λ1 x, λ2 y, λ3 z)T . Let us further assume that the principal components are all
positive resulting in the implicit equation

principal components > 0: 1 = λ 1 x2 + λ 2 y 2 + λ 3 z 2

This is the equation for an ellipsoid whose parameters a, b and c are determined by the
−1/2 −1/2 −1/2
inverse square roots of the principal components a = λ1 , b = λ2 , c = λ3 .

But what happens when the tensor has negative principal components (eigenvalues)?
Although we do not expect a conductivity to be negative, we will consider several tensors
later that can have negative principal components. We obtain surfaces that are more general
than ellipsoids: Generally any equation of the form 1 = ⃗xT T⃗x for any symmetric matrix T

25
is called a quadric and when the coordinate system aligns with the principal axes then the
quadric equation simplifies to 1 = λ1 x2 + λ2 y 2 + λ3 z 2 . The difference to an ellipsoid is that
the principal components here can be either negative or zero and we can imagine implicit
equations, such as 1 = λ1 x2 − λ2 y 2 via the eigenvalues (λ1 , λ2 , λ3 ) = (1, −1, 0).

Let us consider the case when we have a zero eigenvalue via the following example:
imagine that we have a series of tensors Tn which we index by n and imagine that each has
principal components (λ1 , λ2 , λ3 ) = (1, 1, 1/n) that depend on the index n. As all eigenvalues
are positive, the corresponding representation surfaces √ are all ellipsoids. However, each
−1/2
successive ellipsoid in the series has a height c = λ3 = n that is larger than the previous
one. As we keep increasing the index n we find that the height of the ellipsoid grows to
infinity. Indeed, we find that the limit corresponds to the eigenvalues limn→∞ (1, 1, 1/n) =
(1, 1, 0) whose representation surface is a cylinder as

one p. comp. zero (λ1 , λ2 , λ3 ) = (1, 1, 0)

Similarly, we could also consider the case when one of the eigenvalues is negative. While
we can think about a cylinder (zero eigenvalue) as an infinite number of identical 2D circles
stacked on top of each other, a negative eigenvalue (λ1 , λ2 , λ3 ) = (1, 1, −1) corresponds to
an infinite number of 2D circles stacked on top of each other but the radius of these circles
grows as we increase the distance from the origin along the z coordinate. This surface is
called a hyperboloid of one sheet as

one p. comp. negative (λ1 , λ2 , λ3 ) = (1, 1, −1)

Question 8. In the case when the tensor has two negative eigenvalues as, e.g., (λ1 , λ2 , λ3 ) =
(1, −1, −1), we obtain a hyperboloid of two sheets. Draw this surface or use computer
software to visualise it.

For three negative eigenvalues we do not have a real solution of the implicit equation but

26
in such an instance we could just take −T and visualise this tensor. In most usual physical
phenomena we do not expect to come across negative principal components, in fact, some
physical phenomena can only have positive proportionality constants, such as the electrical
conductivity. However, in some instances we do have negative proportionality constants,
such as in the case of the magnetic susceptibility of a diamagnetic material. Moreover, we
will later use tensors to describe physical variables rather than proportionality constants,
and those variables can indeed have either negative or zero principal components.

4.4 Rotational properties

How do these surfaces look when the coordinate axes don’t exactly align with the principal
axes? In other words, what happens when the tensor is not diagonal? In such an instance
the equation ⃗xT T⃗x = 1 becomes really complicated as it is a sum of many terms.

Question 9. When we expand the equation ⃗xT T⃗x = 1 for a diagonal tensor T =
diag(λ1 , λ2 , λ3 ) and a position vector ⃗x = (x, y, z)T it is only a sum of three terms as
1 = λ1 x2 + λ2 y 2 + λ3 z 2 . Expand the equation ⃗xT T⃗x = 1 assuming that T is a full matrix
with matrix entries, T1,1 , T1,2 , T1,3 etc. How many terms does the result have?

We can actually understand how these surfaces look without trying to handle these
complicated equations. The reason is that we can diagonalise any symmetric matrix just by
rotating the coordinate system due to the property summarised in Reminder 2. We thus
take our representation quadric 1 = ⃗xT T⃗x and substitute the tensor T with its diagonalised
form as T = LDLT . Here D is a diagonal matrix and L is a rotation. The resulting implicit
equation is
1 = (LT ⃗x)T D(LT ⃗x),
but it is not obvious the above is still a representation quadric. Here, we can interpret the
expression LT ⃗x such that it is a rotation of the coordinate vector. We can thus introduce
the notation LT ⃗x = ⃗x′ for a position vector ⃗x′ in a rotated coordinate system. Indeed, the
result is now clearly a representation quadric in a rotated coordinate system
1 = (⃗x′ )T D⃗x′ .
Most importantly, this is the representation quadric of a diagonal tensor that we already know
how to interpret as we discussed in Section 4.3. Fig. 8 illustrates that the representation
quadric of any symmetric tensor T (Fig. 8(left)) is obtained as follows. First, we diagonalise
T, second, we visualise the diagonal matrix D and third, by rotating the coordinate system
we obtain the representation quadric of the general matrix T.

Importantly, the converse holds too: Just by looking at Fig. 8(left) we see that the
representation quadric of T is a cylinder of radius 1 whose axis of symmetry is 45◦ off

27
    
1 0 0 1 0 0
T = L 0 1 0 LT D = 0 1 0
0 0 0 0 0 0

Figure 8: The representation quadric of any symmetric tensor T can be understood by

transforming the tensor into a diagonal form via a rotation: Here L is a rotation of 45◦
around ⃗x1 . (left) As we rotate the tensor into the Principal Axis System it becomes diagonal
and we can immediately infer its representation quadric based on the eigenvalues (diagonal
entries). For example, here they are (λ1 , λ2 , λ3 ) = (1, 1, 0) and therefore the representation
quadric of D is a cylinder of radius 1.

between the y and z axes. From this description we can immediately tell that T must have
eigenvalues (λ1 , λ2 , λ3 ) = (1, 1, 0) since it is a cylinder of radius 1 and we can also tell that
its eigenvectors (principal directions) must be 45◦ off from the coordinate axes.

 
0 0 −1
Question 10. What is the representation quadric of the tensor  0 1 0 ? Hint:
−1 0 0
diagonalising the matrix first helps to determine the representation quadric.

4.5 Effect of crystal structure

Properties of a crystalline material depend on symmetries of the crystal. As such, if a

property of a material is described by a tensor then the tensor, as well as its representation
surface, must have the same symmetry properties as the crystal. This is summarised by
Neumann’s principle.

28
 cryst. syst. quadric orientation params. tensor
 
T 0 0
cubic sphere 1 0 T 0
0 0 T
tetragonal, 
T1 0 0


hexagonal, symm. around x3 2 0 T1 0

trigonal 0 0 T3
x1 , x2 , x3
 
T1 0 0
orthorhombic 3 0 T2 0
parallel to diads 0 0 T3
 
T11 0 T13
monoclinic x2 parallel to diad 4  0 T22 0 
T13 0 T33
 
T11 T12 T13
triclinic - 6 T12 T22 T23 
T13 T23 T33

Table 2: Tensor properties of a crystalline material must contain symmetries of the crys-
tal. Symmetry properties of the 7 crystal systems constrain what principal directions and
principal components the corresponding tensors can have.

Neumann’s principle: Any physical property of a crystal must include the symmetry ele-
ments of the point group of the crystal.

The symmetry of the tensor may actually be higher but must include at least the symme-
try elements of the crystal. Furthermore, the crystal symmetry axes determine the principal
axis system (the principal directions). Table 2 summarises symmetry properties of the 7
crystal systems and the possible eigenvalues (principal components) of the corresponding
tensors.

• A cubic crystal has a high symmetry and it only has 1 degree of freedom. The corre-
sponding tensors must be proportional to the identity matrix and the material must
thus be isotropic.

• tetragonal, hexagonal and trigonal lattices have an axis of symmetry around the ⃗x3
axis and thus their tensors must have two identical eigenvalues. This also means that
their representation surfaces must have a rotational symmetry around the ⃗x3 axis, such
as a symmetric ellipsoid with axes a = b.

29
5 Applications of tensors

We are now equipped with all necessary mathematical tools to work with tensors. So far
our motivation to using tensors was to describe the linear response in anisotropic materials.
In these applications the tensors typically express properties of materials, for example, the
electric conductivity is a proportionality factor in the equation J⃗ = σ E.
⃗ Here we consider
some more advanced applications of tensors. In these applications the tensors don’t always
express properties of the materials but rather they are used to express a physical variable,
such as mechanical stress and strain.

5.1 Stress tensor

As we apply forces to a solid body it becomes deformed. Stress is used to describe these
forces in a way that the description is independent of the size or the shape of the body.
Stress is defined as the force divided by the area it is acting on. For example, if we apply a
tension force F to stretch a wire of radius r, the stress is defined as the tension relative to
the cross-sectional area as
F
σ = 2. (19)
πr
Indeed, as we apply a multiple of the tension, for example 4F to a wire of radius 2r, then
the stress remains the same σ as the same forces pull the atoms in the wire apart.

In three dimensions, we model stress by taking a infinitesimally small cube of the material
and analyse what forces act upon it – and we assume that the stress is homogeneous and all
parts of the material experience the same forces. The three edges of the cube are parallel
with our coordinate axes ⃗x1 , ⃗x2 and ⃗x3 as illustrated in Fig. 9. Each of the six faces of the
cube can be described by an axis vector ⃗x1 , −⃗x1 , ⃗x2 , −⃗x2 , and ⃗x3 , −⃗x3 . We only need to
take into account what forces act on the three faces of the cube that correspond to the three
axis vectors ⃗x1 , ⃗x2 and ⃗x3 , while the forces acting on the opposite faces must be equal and
opposite as the total force acting on the cube in equilibrium must be zero. Consider the force
acting on the face that corresponds to the axis ⃗x1 : this force has some component that is
parallel with the axis σ11 but it can also have components in the other two directions which
we denote as σ21 and σ31 as illustrated in Fig. 9(left). We similarly repeat this analysis for
the other two faces and find the 9 tensor entries that describe the stress tensor as forces
acting on the three faces of the cube
 
σ11 σ12 σ13
σ = σ21 σ22 σ23  . (20)
σ31 σ32 σ33

Since stress is described by a rank-2 tensor, it relates two vectors to each other. For
this reason, we can define the traction vector T⃗ as a force per unit area that acts on the

30
 positive: tensile shear stress rotation of the cube
negative: compressive conserves ang. mom. zero in equilibrium

Figure 9: We analyse stress by taking a small cube of the material such that the edges of
the cube are parallel with the coordinate axes. (left) Tensile and compressive stresses (red
arrows) correspond to diagonal entries of the stress tensor. (middle) Shear stresses (blue
arrows) correspond to off-diagonal entries of the symmetrised stress tensor as σ s . (right)
Off diagonal entries of the antisymmetrised stress tensor σ a try to rotate the cube (blue
arrows). The stress tensor is thus symmetric due to the conservation of angular momentum
(no rotation).

plane represented by an area vector ⃗n. Here, an area vector is a vector whose direction is
perpendicular to the plane and its length is identical to the area of the plane. The stress
tensor relates these two vector quantities to each other as

T⃗ = σ⃗n, Ti = σij nj . (21)

This equation expresses the relationship between two rank-1 tensors (vectors) using a rank-2
tensor, just like in the case of cause and effect relations. As such it also guarantees the
rotational property σ ′ = LσLT from Summary 6. The important conceptual difference here
is that the stress tensor is not a property of the material, it is rather a physical variable.

Note 3. The expression in Eq. (21) is also commonly stated in terms of row vectors
as T⃗ T = ⃗nT σ row , see for example the Wikipedia page3 . The transpose of this tensor is
equivalent to the one in Eq. (21) as σ row = σ T .

3
The generalisation of row and column vectors are covariant and contravariant tensors – while one trans-
forms according to LT the other transforms according to L. Here we only work in orthogonal, Cartesian
coordinate systems and all tensor entries are real numbers – we referred to these as Cartesian tensors in
Summary 2 and in Note 2. For Cartesian tensors there is no distinction between covariance and contravari-
ance, see the Wikipedia page. For example, the tensors σ row and σ obey the same rotational transformation
rule

31
5.2 Stress tensor in equilibrium

The red arrows in Fig. 9 (left) visualise the tensor entries σ11 , σ22 and σ33 which correspond
to forces that act perpendicular to the faces of the cube. Such a force can be either tensile
(positive value) or compressive (negative value). For example, if a tensile force σ11 acts on
a face ⃗x1 then a compressive force −σ11 must act on the opposite face −⃗x1 such that these
forces cancel each other out in equilibrium.

The other tensor entries are forces that act parallel to the faces. For example, the
tensor entry σ12 is the force that acts on the ⃗x2 face but in the ⃗x1 direction. It simplifies
the following argument if we first decompose our tensor into symmetric and antisymmetric
parts. Recall that a symmetric matrix is unchanged under a transposition (σ s )T = σ s while
an antisymmetric matrix picks up a negative sign under a transposition (σ a )T = −σ a . We
can actually write any matrix as a sum of a symmetric and an antisymmetric matrix as
σ = σs + σa.

Question 11. We can define the symmetric part of a matrix as 2σ s = σ + σ T and the
antisymmetric part as 2σ a = σ − σ T . Verify that by transposing these matrices they
only pick up a sign ±1.

For these reasons, we can write the stress tensor as a sum of a diagonal matrix, a symmetric
matrix with zeros in the diagonal and an antisymmetric matrix with zeros in the diagonal
as      s s
  a a

σ11 σ12 σ13 σ11 0 0 0 σ12 σ13 0 σ12 σ13
s s  a a 
σ21 σ22 σ23  =  0 σ22 0  + σ12 0 σ23 + −σ12 0 σ23 . (22)
s s a a
σ31 σ32 σ33 0 0 σ33 σ13 σ23 0 −σ13 −σ23 0
| {z } | {z } | {z }
tensile shear rotation

The first part, the diagonal matrix corresponds to the aforementioned stresses that act
perpendicular to the faces and they are illustrated as red arrows in Fig. 9(left). The second
matrix in Eq. (22) is off-diagonal and symmetric and it corresponds to so-called shear stresses.
These shear stresses are illustrated in Fig. 9(middle) as blue arrows. The third matrix in
Eq. (22) is antisymmetric and off-diagonal and these terms correspond to a rotation of the
cube. The blue arrows in Fig. 9(right) form a circle and correspond to forces that try to
rotate the cube. In equilibrium, due to Cauchy’s conservation of angular momentum, these
terms must be zero and thus the stress tensor must be symmetric σ ≡ σ s .

5.3 Principal stresses

We saw that in equilibrium the stress tensor is symmetric as σ = σ T and from Reminder 2
we know that any such matrix can be diagonalised. After diagonalisation we obtain the

32
 stress type principal comps. example tensor
uniaxial 1 non-zero σk pulling wire diag(σ1 , 0, 0)

biaxial 2 non-zero σk force on thin plate diag(σ1 , σ2 , 0)

triaxial 3 non-zero σk - diag(σ1 , σ2 , σ3 )

hydrostatic 3 identical σk < 0 pressure p in fluid diag(−p, −p, −p)

pure shear special biaxial rod torsion diag(−σ, σ, 0)

Table 3: Examples where the shape and quality of a material determine what principal
components the stress tensor can have.

principal components and principal vectors. Note that σ is not a property of material but
rather a physical variable and thus the principal components and principal axes are not
related to symmetry properties of the crystal.

In Table 3 we summarise a few scenarios when the shape and quality of the object
determine properties of the stress tensor.

• Imagine a wire that is being stretched. As we set up our coordinate system such that
the wire is parallel with the ⃗x1 coordinate axis, the tensor becomes diagonal. This is
called a uniaxial stress and it only has a single non-zero principal component as shown
in Table 3.

• A biaxial stress can be, for example, due to a force that pulls a thin plate and the
corresponding stress tensor then only has two non-zero principal components.

• Hydrostatic pressure in a liquid has a stress tensor that is proportional to the identity
matrix as −p1. The proportionality factor is −p and its negative sign reflects the fact
that the pressure is compressive.

• We can consider a pure shear stress as illustrated in Fig. 9(middle) with the blue
arrows. As we will see later we can apply a 45◦ rotation to this tensor to obtain the
diagonal form diag(−σ, σ, 0).

In general the stress tensor represents a physical variable, more concretely a force per
unit area. But given a stress tensor, what is the ‘magnitude’ of this physical variable? For
example, when a physical variable is described by a vector then we can conveniently use the
norm of the vector to express the magnitude of the physical property. Unfortunately, there
is no obvious metric of magnitude for matrices.

33
 Note 4. We expect that the ‘magnitude’ of a physical property must not depend on
what coordinate system we use. In order to quantify the norm of a tensor property,
we need to use matrix invariants, such as the trace Tr(A), the determinant or so-called
matrix norms, for example Tr(AT A). These matrix invariants do not change when we
rotate our matrix.

5.4 Strain in 1D

If we apply force or stress to a material, it becomes deformed. We refer to this deformation as

strain and express it as the ratio of the extension of the body relative to the original distance.
The extension of the body is measured as the change in distance between two points in the
body. We first consider a 1-dimensional example, such as a wire that is being stretched.
Suppose we select and analyse a small interval of this wire, for example, the interval (x, x +
∆x) which has a length ∆x. This is illustrated in in the top part of Fig. 10(left). Imagine
that after stretching the wire, the interval has been increased to (x + u, x + ∆x + u + ∆u)
which is illustrated in the bottom part of Fig. 10(left). We can define the extension per unit
length as
increase in length
strain = = ∆u/∆x.
original length
This expression is, however, only valid if the extension of the material is homogeneous which
is illustrated with the red linear curve in Fig. 10(right).

The black curve in Fig. 10(right) shows the profile of a material in which the strain is
not homogeneous. In such an instance we need to consider small intervals ∆x, in fact the
limit when the interval becomes very small as
∆u du
strain = lim = . (23)
∆x→0 ∆x dx
The strain ϵ is thus the derivative of the displacement u(x) with respect to the position x.
Indeed, when the strain is homogeneous, such as in the red curve in Fig. 10(right), then then
the extension u(x) is a linear function of x whose derivative, the strain, is a constant. In
such a case the Taylor expansion only contains the first two terms which lead to the equation
u(x) − u(0) = ϵx.

5.5 Strain in 3D

In the previous example we only considered a material in 1 dimension where we could describe
the coordinate with a number x. As we want to treat materials in three dimensions we need

34
Figure 10: (left) A 1D wire before and after stretching. The interval of length ∆x is
extended to an interval of length ∆x + ∆u due to stretching. Strain is the increase in length
∆u relative to the original length ∆x. (right) The extension u(x) might be a non-linear
function of the coordinate x and thus the strain is more generally the derivative du
dx
.

to describe the coordinate as a column vector ⃗x = (x1 , x2 , x3 )T and similarly the extension
as a column vector ⃗u = (u1 , u2 , u3 )T . The strain in 3D is then a direct generalisation of
the 1-dimensional derivative in Eq. (23): in 3D we have to consider each of the three vector
entries of ⃗u and compute partial derivatives with respect to each vector entry of ⃗x as
∂ui
ϵ̃ij = .
∂xj

These partial derivatives form a tensor as ϵ̃ij . We have seen in Section 5.2 that the anti-
symmetric part of the matrix as (ϵ̃ij − ϵ̃ji )/2 tries to rotate the material. For this reason,
we only consider the symmetric part of the above tensor and define the strain tensor as the
rank-2, symmetric tensor
∂ui ∂uj
ϵij = ( + )/2.
∂xj ∂xi
Similarly as in Section 5.2 we can decompose this symmetric strain tensor as a sum of a
diagonal matrix and a purely off-diagonal matrix as
   
ϵ11 0 0 0 ϵ12 ϵ13
ϵ= 0 ϵ22 0  + ϵ12 0 ϵ23  .
0 0 ϵ33 ϵ13 ϵ23 0
| {z } | {z }
uniaxial extension shear

The diagonal entries ϵ11 , ϵ22 , ϵ33 are the uniaxial extensions per unit length. The off-diagonal
entries in the matrix are shear strains – these of course depend on the coordinate system
and, for example, in the principal axis system all off-diagonal entries are zero. When the
strain is homogeneous, then the extension vector ⃗u(x) is a linear function of the coordinate
vector and we can thus write ⃗u(⃗x) − ⃗u(0) = ϵ⃗x.

35
 shear stress shear strain

Figure 11: Pure shear stress (left) and strain (right) in a cube of material. (left) Shear
tensors can be diagonalised by a 45◦ rotation. (right) The angle of deformation of the
material is approximately the matrix entry ϵ when ϵ is small.

5.6 Pure shear stress and strain

Suppose we have a material undergoing pure shear stress or strain. Pure shear stress is
illustrated in Fig. 11(left) while pure shear strain is illustrated in Fig. 11(right). The corre-
sponding stress and strain matrices are
   
0 σ 0 0 ϵ 0
σ = σ 0 0 , ϵ =  ϵ 0 0 .
0 0 0 0 0 0

Above both matrices can be diagonalised by a 45◦ rotation of the coordinate system, which
can also be thought of as an active rotation of the material as illustrated in Fig. 11(left).
After diagonalization, both matrices have only two non-zero eigenvalues as ±σ and ±ϵ.
Fig. 11(left) illustrates that after the diagonalising rotation the principal stresses act along
the ⃗x1 and ⃗x2 coordinate axes with compressive and tensile stresses as ±σ, respectively.

Shear strain deforms the cubic material in Fig. 11(right) into a rhombic material. Using
the definition of the strain, we can derive the angle of deformation as
∂u1 ∂u2
ϵ = ϵ12 = ( + )/2 = tan(θ) ≈ θ. (24)
∂x2 ∂x1

In the third equation we used that the derivative ∂u1

∂x2
expresses the tangent of the angle θ and
in the last equation we used the Taylor expansion of the tangent function as tan(x) = x + . . .
for small x.

36
Figure 12: A sphere of material is heated up uniformly such that its temperature is raised
by a small amount ∆T . The sphere is deformed into an ellipsoid.

5.7 Strain tensor via thermal expansion

Suppose a material is heated up uniformly by a small amount such that its temperature
increases by a small ∆T – we assume that all parts of the material were initially of the same
temperature and all parts are heated up by exactly the same change of temperature. Under
these assumptions the extension of the material caused by the heating is proportional to
the temperature and thus the components of the strain tensor (ϵij ) are proportional to the
thermal expansion coefficients (αij ) as

ϵ = α∆T, ϵij = αij ∆T.

Here ϵij is a symmetric tensor and since ∆T is just a scalar multiplication factor, αij must
also be a symmetric tensor too. Since the tensor αij is a property of the material, the
principal components and principal directions (eigenvalues and eigenvectors) are determined
by the crystal symmetry according to von Neuamnn’s principle. For example, the principal
components ϵi = αi ∆T express the expansion along the crystal directions. These principal
components αi are typically positive, but in some instances they can be negative too. In the
most typical case when αi are positive, the corresponding representation surface of ϵij and
αij are ellipsoids as we discussed in Section 4.3.

As we heat up a ball of material, it gets deformed into an ellipsoid. The expansion along
the principal directions can be described in terms of the principal components as ϵi = αi ∆.
Fig. 12 illustrates that initially the sphere had a radius r and after heating up the material,
the sphere is transformed into an ellipsoid with axes

a = r(1 + α1 ∆T ), b = r(1 + α2 ∆T ), c = r(1 + α3 ∆T ). (25)

While the sphere had a volume V = 34 πr3 the resulting ellipsoid has a volume V ′ = 34 πabc
that depends on the magnitude of the three axes. We can simplify the expression for the

37
new volume V ′ assuming that the change of temperature ∆T is small as

V ′ = V + (α1 + α2 + α3 )V ∆T + . . . , (26)

where we have neglected all terms that contain (∆T )2 or (∆T )3 . Finally, the increase of
volume can be calculated for small ∆T as

∆V ≈ (α1 + α2 + α3 )V ∆T. (27)

Notice that here the sum of principal components is actually the trace of the matrix α1 +
α2 + α3 = Trα. The trace is actually a matrix invariant as noted in Note 4 and it can thus
be used as a valid measure of the magnitude of the tensor α. In conclusion, the volume
expansion of a material is approximately proportional to the trace of its thermal expansion
tensor – and the trace is not dependent on the coordinate system used.

5.8 Hooke’s law and linear elasticity

Recall Hooke’s law that describes the force in a 1-dimensional spring as

F = kx.

The generalisation of Hooke’s law relates stress and strain in an elastic material. As both
the stress and strain are rank-2 tensors (they are physical variables), in order to relate them
to each other we need to use rank-4 tensors. The stiffness tensor cijkl is the analogue of the
spring constant k from the above equation and it relates stress and strain via the Einstein
summation
σij = cijkl ϵkl .
So far, we have used convenient matrix and vector notations to describe how a rank-2 tensor
relates two vectors to each other. In contrast, here we need to resort to Einstein summations
as rank-4 tensors are more general than matrices as we discussed in Section 2.3.

Since every index in the tensor cijkl can take up 3 values, overall there are 34 = 81 ten-
sor entries in a rank-4 Cartesian tensor. While for rank-2 tensors (matrices) we only had
9 entries, we could even reduce them to 6 entries since the matrices we considered were
symmetric. Similarly, we can exploit symmetry properties of the stiffness tensor: In general
we only have 21 independent tensor entries in cijkl due to the following symmetries. The
symmetry of the stress and strain tensors σij = σji and ϵkl = ϵlk implies the symmetry of
the stiffness tensor as cijkl = cjikl and cijkl = cijlk , respectively. Furthermore, in crystals
the strain tensor ϵkl inherits symmetry properties of the crystal which then further reduces
the number of independent components in the stiffness tensor. For example, in orthorhom-
bic crystals the stiffness tensor has 9 independent entries, in hexagonal crystals it has 5
independent entries, and in cubic crystals it only has 3 independent entries.

38
Appendix

A Constructing rotation matrices

Here we discuss how we can take a physically defined operation, such as a “rotation around
an axis that is halfway between the x and z axes” and calculate its rotation matrix L.

a if the physical operation is a rotation around one of the axis vectors than we can
immediately use one of the three elementary rotation matrices from Eq. (12).

b if the physical operation is specified as a series of elementary rotations, e.g, “rotate

first around the x axis, then around the y axis . . . ” then we can simply construct L as a
matrix product of the elementary rotation matrices. In fact, this is the conventional way
to construct any rotation from the composition of only 3 elementary rotations that rotate
around only 2 axes. This is called the Euler angle convention and is detailed, for example,
in the Wikipedia article.

c We may need to find the rotation matrix that corresponds to a rotation around an
angle that is not one of the standard axes and then we cannot use the above two techniques.
In such a case the simplest way to evaluate the rotation matrix is by analysing how the
standard basis vectors are rotated. This is explained in more detail in Appendix B.1.

B Solutions

B.1 Solution of Question 5

We can calculate the active rotation matrix by evaluating how it acts on the standard basis
vectors as discussed in Section 3.1: For example, the vector ⃗x′1 = L⃗x1 will be the first column
vector of L etc.

By drawing how the three coordinate axes are rotated we can determine column vectors
of L. For example, the first column vector ⃗x′1 = L⃗x1 is obtained by looking at the rotation
axis from the top and performing the rotation that maps √ the ⃗x1 axis to a vector that is
halfway between the ⃗x2 and −⃗x3 axes as L⃗x1 = (0, 1, −1)T / 2 as

39
 axis of rotation
.

We could repeat this similarly and draw the rotations for the other two axes and find the
vector entries as
   1   1
0
   
1 0 − √2 0 √
2
1 1
√
L 0 =  2  , L 1 =  2  , L 0 =  12  . (28)
1 1 1
0 − √2 0 2
1 2

We finally arrange these vectors as column vectors into the transformation matrix L as
 
0 − √12 √12
 √1 1 1 
 2 2 2 .
− √12 1
2
1
2

B.2 Solution of Question 7

We compute the determinant of both sides of the equation M = LDLT as

det M = det(LT DL),
and use the property of determinants that det(AB) = det(A) det(B) as
det(LT σL) = det(LT ) det(D) det(L)
Since the determinant is the product of eigenvalues, and M has the same eigenvalues as D,
we know their determinants must be the same as det M = det D. This allows us to write
that
1 = det(LT ) det(L).
Using the property of determinants that det(LT ) = det(L) we find that the determinant of
the orthogonal matrix is allowed to be either plus or minus one as det(L) = ±1. We can thus
always choose an orthogonal transformation L that is a proper rotation with determinant
one.

40