12/24/2017 Continuum Mechanics - Tensors

Home

A Brief Introduction to Tensors and their properties

1. BASIC PROPERTIES OF TENSORS

1.1 Examples of Tensors

The gradient of a vector field is a good example of a second-order tensor. Visualize a vector field: at every point in space,
the field has a vector value u(x , x , x ) . Let G = ∇u  represent the gradient of u. By definition, G enables you to
1 2 3

calculate the change in u when you move from a point x in space to a nearby point at x + dx :
du = G ⋅ dx

G is a second order tensor. From this example, we see that when you multiply a vector by a tensor, the result is another
vector.

This is a general property of all second order tensors. A tensor is a linear mapping of a vector onto another vector.
Two examples, together with the vectors they operate on, are:

The stress tensor

t =  n ⋅ σ
where n is a unit vector normal to a surface, σ is the stress tensor and t is the traction vector acting on the surface.

The deformation gradient tensor

dw = F ⋅ dx
where dx is an infinitesimal line element in an undeformed solid, and dw is the vector representing the deformed
line element.

1.2 Matrix representation of a tensor

To evaluate and manipulate tensors, we express them as components in a basis, just as for vectors. We can use the
displacement gradient to illustrate how this is done. Let u(x , x , x ) be a vector field, and let G = ∇u represent the
1 2 3

gradient of u. Recall the definition of G

du = G ⋅ dx

Now, let {e , e , e } be a Cartesian basis, and express both du and dx as components. Then, calculate the components of
1 2 3

du in terms of dx using the usual rules of calculus

∂u 1 ∂u 1 ∂u 1
du 1 = dx 1 + dx 2 + dx 3
∂x 1 ∂x 2 ∂x 3

∂u 2 ∂u 2 ∂u 2
du 2 = dx 1 + dx 2 + dx 3
∂x 1 ∂x 2 ∂x 3

∂u 3 ∂u 3 ∂u 3
du 3 = dx 1 + dx 2 + dx 3
∂x 1 ∂x 2 ∂x 3

We could represent this as a matrix product

http://www.brown.edu/Departments/Engineering/Courses/En221/Notes/Tensors/Tensors.htm 1/10
12/24/2017 Continuum Mechanics - Tensors
∂u ∂u ∂u
1 1 1
⎡ ⎤
∂x ∂x ∂x
du1 1 2 3 dx1
⎡ ⎤ ⎢ ⎥⎡ ⎤
⎢ ∂u
2
∂u
2
∂u
2 ⎥
⎢ du 2 ⎥ = ⎢ ⎥ ⎢ dx 2 ⎥
⎢ ∂x
1
∂x
2
∂x
3 ⎥
⎣ ⎦ ⎢ ⎥⎣ ⎦
du 3 ∂u ∂u ∂u dx 3
3 3 3
⎣ ⎦
∂x ∂x ∂x
1 2 3

Alternatively, using index notation

∂u i
du i = dx j
∂x j

From this example we see that G can be represented as a 3 × 3 matrix. The elements of the matrix are known as the
components of G in the basis {e , e , e }. All second order tensors can be represented in this form. For example, a
1 2 3

general second order tensor S could be written as

S11 S12 S13
⎡ ⎤

S ≡ ⎢ S21 S22 S23 ⎥

⎣ ⎦
S31 S32 S33

You have probably already seen the matrix representation of stress and strain components in introductory courses.

Since S can be represented as a matrix, all operations that can be performed on a 3 × 3 matrix can also be performed on
S. Examples include sums and products, the transpose, inverse, and determinant. One can also compute eigenvalues and
eigenvectors for tensors, and thus define the log of a tensor, the square root of a tensor, etc. These tensor operations are
summarized below.

Note that the numbers S , S , … S depend on the basis {e , e , e }, just as the components of a vector depend on the
11 12 33 1 2 3

basis used to represent the vector. However, just as the magnitude and direction of a vector are independent of the basis,
so the properties of a tensor are independent of the basis. That is to say, if S is a tensor and u is a vector, then the vector
v = S ⋅ u
has the same magnitude and direction, irrespective of the basis used to represent u, v, and S.

1.3 The difference between a matrix and a tensor

If a tensor is a matrix, why is a matrix not the same thing as a tensor? Well, although you can multiply the three
components of a vector u by any 3 × 3 matrix,
b1 a 11 a 12 a 13 u1
⎡ ⎤ ⎡ ⎤⎡ ⎤

⎢ b2 ⎥ = ⎢ a 21 a 22 a 23 ⎥ ⎢ u2 ⎥
⎣ ⎦ ⎣ ⎦⎣ ⎦
b3 a 31 a 32 a 33 u3

the resulting three numbers (b , b , b ) may or may not represent the components of a vector. If they are the components
1 2 3

of a vector, then the matrix represents the components of a tensor A, if not, then the matrix is just an ordinary old matrix.

To check whether (b , b , b ) are the components of a vector, you need to check how (b , b , b ) change due to a change
1 2 3 1 2 3

of basis. That is to say, choose a new basis, calculate the new components of u in this basis, and calculate the new matrix
in this basis (the new elements of the matrix will depend on how the matrix was defined. The elements may or may not
change – if they don’t, then the matrix cannot be the components of a tensor). Then, evaluate the matrix product to find a
new left hand side, say (β , β , β ). If (β , β , β ) are related to (b , b , b ) by the same transformation that was used to
1 2 3 1 2 3 1 2 3

calculate the new components of u, then (b , b , b ) are the components of a vector, and, therefore, the matrix represents
1 2 3

the components of a tensor.

1.4 Formal definition

Tensors are rather more general objects than the preceding discussion suggests. There are various ways to define a tensor
formally. One way is the following:

A tensor is a linear vector valued function defined on the set of all vectors

More specifically, let S(v) denote a tensor operating on a vector. Linearity then requires that, for all vectors v, w and
scalars α
· S (v + w) = S (v) + S (w)

http://www.brown.edu/Departments/Engineering/Courses/En221/Notes/Tensors/Tensors.htm 2/10
12/24/2017 Continuum Mechanics - Tensors

· S (αv) = αS (v)

Alternatively, one can define tensors as sets of numbers that transform in a particular way under a change of coordinate
system. In this case we suppose that n dimensional space can be parameterized by a set of n real numbers x . We could i

change coordinate system by introducing a second set of real numbers x (x ) which are invertible functions of x . ′
i k i

Tensors can then be defined as sets of real numbers that transform in a particular way under this change in coordinate
system. For example

· A tensor of zeroth rank is a scalar that is independent of the coordinate system.

∂xj
· A covariant tensor of rank 1 is a vector that transforms as v ′
i =
∂x
′
vj
i
′

· A contravariant tensor of rank 1 is a vector that transforms as v

∂x
′i i j
= v
∂xj

∂xj
· A covariant tensor of rank 2 transforms as S
∂xi
′
ij = Skl
∂x′ k ∂x′ l
′ ′

· A contravariant tensor of rank 2 transforms as S

∂x ∂x
′ kl i i ij
= S
∂xk ∂xl
′ ∂x
·
∂x
A mixed tensor of rank 2 transforms as S ′
i

j
=
∂x
k j

∂x′
S
l
k

i l

Higher rank tensors can be defined in similar ways. In solid and fluid mechanics we nearly always use Cartesian tensors,
(i.e. we work with the components of tensors in a Cartesian coordinate system) and this level of generality is not needed
(and is rather mysterious). We might occasionally use a curvilinear coordinate system, in which we do express tensors in
terms of covariant or contravariant components – this gives some sense of what these quantities mean. But since solid
and fluid mechanics live in Euclidean space we don’t see some of the subtleties that arise, e.g. in the theory of general
relativity.

1.5 Creating a tensor using a dyadic product of two vectors.

Let a and b be two vectors. The dyadic product of a and b is a second order tensor S denoted by
S = a ⊗ b          Sij = a i bj .
with the property
S ⋅ u = (a ⊗ b) ⋅ u = a(b ⋅ u)            Sij uj = (a i bk )uk = a i (bk uk )

for all vectors u. (Clearly, this maps u onto a vector parallel to a with magnitude |a| (b ⋅ u) )

The components of a ⊗ b in a basis {e 1, e2 , e3 } are

a 1 b1 a 1 b2 a 1 b3
⎡ ⎤

⎢ a 2 b1 a 2 b2 a 2 b3 ⎥
⎣ ⎦
a 3 b1 a 3 b2 a 3 b3

Note that not all tensors can be constructed using a dyadic product of only two vectors (this is because (a ⊗ b) ⋅ u always
has to be parallel to a, and therefore the representation cannot map a vector onto an arbitrary vector). However, if a, b,
and c are three independent vectors (i.e. no two of them are parallel) then all tensors can be constructed as a sum of scalar
multiples of the nine possible dyadic products of these vectors.

2. OPERATIONS ON SECOND ORDER TENSORS

Tensor components.

Let {e , e , e } be a Cartesian basis, and let S be a second order tensor. The components of S in
1 2 3 {e1 , e2 , e3 } may be
represented as a matrix
S11 S12 S13
⎡ ⎤

⎢ S21 S22 S23 ⎥

⎣ ⎦
S31 S32 S33

where

http://www.brown.edu/Departments/Engineering/Courses/En221/Notes/Tensors/Tensors.htm 3/10
12/24/2017 Continuum Mechanics - Tensors

S11 = e1 ⋅ (S ⋅ e1 ) ,            S12 = e1 ⋅ (S ⋅ e2 ) ,             S11 = e1 ⋅ (S ⋅ e3 ) ,

S21 = e2 ⋅ (S ⋅ e1 ) ,            S22 = e2 ⋅ (S ⋅ e2 ) ,             S21 = e2 ⋅ (S ⋅ e3 ) ,

S31 = e3 ⋅ (S ⋅ e1 ) ,            S32 = e3 ⋅ (S ⋅ e2 ) ,             S31 = e3 ⋅ (S ⋅ e3 ) ,

The representation of a tensor in terms of its components can also be expressed in dyadic form as
3 3

S = ∑ ∑ Sij ei ⊗ ej

j=1 i=1

This representation is particularly convenient when using polar coordinates, or when using a general non-orthogonal
coordinate system.

Addition

Let S and T be two tensors. Then U = S + T is also a tensor.

Denote the Cartesian components of U, S and T by matrices as defined above. The components of U are then related to
the components of S and T by
U11 U12 U13 S11 + T 11 S12 + T 12 S13 + T 13
⎡ ⎤ ⎡ ⎤

⎢ U21 U22 U23 ⎥ = ⎢ S21 + T 21 S22 + T 22 S23 + T 23 ⎥

⎣ ⎦ ⎣ ⎦
U31 U32 U33 S31 + T 31 S32 + T 32 S33 + T 33

In index notation we would write

Uij = Sij + Tij

Product of a tensor and a vector

Let u be a vector and S a second order tensor. Then

v = S ⋅ u

is a vector.

Let (u , u , u ) and (v , v , v ) denote the components of vectors u and v in a Cartesian basis

1 2 3 1 2 3 {e1 , e2 , e3 } , and denote
the Cartesian components of S as described above. Then
v1 S11 S12 S13 u1 S11 u1 + S12 u2 + S13 u3
⎡ ⎤ ⎡ ⎤⎡ ⎤ ⎡ ⎤

⎢ v 2 ⎥ = ⎢ S21 S22 S23 ⎥ ⎢ u2 ⎥ = ⎢ S21 u1 + S22 u2 + S23 u3 ⎥

⎣ ⎦ ⎣ ⎦⎣ ⎦ ⎣ ⎦
v3 S31 S32 S33 u3 S31 u1 + S32 u2 + S33 u3

Alternatively, using index notation

v i = Sij uj

The product
v = u ⋅ S

is also a vector. In component form

S11 S12 S13 u 1 S11 + u 2 S21 + u 3 S31
⎡ ⎤ ⎡ ⎤

[ v1 v2 v3 ] = [ u1 u2 u3 ] ⎢ S21 S22 S23 ⎥ = ⎢ u 1 S12 + u 2 S22 + u 3 S32 ⎥

⎣ ⎦ ⎣ ⎦
S31 S32 S33 u 1 S13 + u 2 S23 + u 3 S33
or
v i = uj Sji

Observe that u ⋅ S ≠ S ⋅ u (unless S is symmetric).

Product of two tensors

Let T and S be two second order tensors. Then U = T ⋅ S is also a tensor.

Denote the components of U, S and T by 3 × 3 matrices. Then,

http://www.brown.edu/Departments/Engineering/Courses/En221/Notes/Tensors/Tensors.htm 4/10
12/24/2017 Continuum Mechanics - Tensors

U11 U12 U13 T11 T12 T13 S11 S12 S13

⎡ ⎤ ⎡ ⎤⎡ ⎤

⎢ U21 U22 U23 ⎥ = ⎢ T21 T22 T23 ⎥ ⎢ S21 S22 S23 ⎥

⎣ ⎦ ⎣ ⎦⎣ ⎦
U31 U32 U33 T31 T32 T33 S31 S32 S33

T11 S11 + T12 S21 + T13 S31 T11 S12 + T12 S22 + T13 S32 T11 S13 + T12 S23 + T13 S33
⎡ ⎤

      = ⎢ T T21 S12 + T22 S22 + T23 S32 ⎥

21 S11 + T22 S21 + T23 S31 T21 S12 + T22 S22 + T23 S32
⎣ ⎦
T31 S11 + T32 S21 + T33 S31 T31 S12 + T32 S22 + T33 S32 T31 S13 + T32 S23 + T33 S33

Alternatively, using index notation

Uij = T ik Skj

Note that tensor products, like matrix products, are not commutative; i.e. T ⋅ S ≠ S ⋅ T

Transpose

Let S be a tensor. The transpose of S is denoted by S and is defined so that T

T
u ⋅ S = S ⋅ u

Denote the components of S by a 3x3 matrix. The components of S

T
are then
S11 S21 S31
⎡ ⎤
T
S ≡ ⎢S S22 S32 ⎥
12

⎣ ⎦
S13 S23 S33

i.e. the rows and columns of the matrix are switched.

Note that, if A and B are two tensors, then

T T T
(A ⋅ B) = B ⋅ A

Trace

Let S be a tensor, and denote the components of S by a 3 × 3 matrix. The trace of S is denoted by tr(S) or trace(S), and
can be computed by summing the diagonals of the matrix of components
trace (S) = S11 + S22 + S33

More formally, let {e 1

, e2 , e3 } be any Cartesian basis. Then
trace (S) = e 1 ⋅ S ⋅ e 1 + e 2 ⋅ S ⋅ e 2 + e 3 ⋅ S ⋅ e 3

The trace of a tensor is an example of an invariant of the tensor – you get the same value for trace(S) whatever basis you
use to define the matrix of components of S.

In index notation, the trace is written S kk

Contraction.

Inner Product: Let S and T be two second order tensors. The inner product of S and T is a scalar, denoted by S : T .
Represent S and T by their components in a basis. Then
S : T = S11 T 11 + S12 T 12 + S13 T 13

           + S21 T 21 + S22 T 22 + S23 T 23

           + S31 T 31 + S32 T 32 + S33 T 33

In index notation S : T ≡ S ij T ij

Observe that S : T = T : S , and also that S : I = trace(S), where I is the identity tensor.

Outer product: Let S and T be two second order tensors. The outer product of S and T is a scalar, denoted by S ⋅ ⋅T.
Represent S and T by their components in a basis. Then
S ⋅ ⋅T = S11 T 11 + S21 T 12 + S31 T 13

           + S12 T 21 + S22 T 22 + S32 T 23

           + S13 T 31 + S23 T 32 + S33 T 33

In index notation S ⋅ ⋅T ≡ S ij T ji

http://www.brown.edu/Departments/Engineering/Courses/En221/Notes/Tensors/Tensors.htm 5/10
12/24/2017 Continuum Mechanics - Tensors

Observe that S ⋅ ⋅T = S T
: T

Determinant

The determinant of a tensor is defined as the determinant of the matrix of its components in a basis. For a second order
tensor
S11    S12     S13
⎡ ⎤
det S = det ⎢ S21    S22    S23   ⎥

⎣ ⎦
S31    S32    S33

               = S11 (S22 S33 − S23 S32 ) + S12 (S23 S31 − S21 S33 ) + S13 (S21 S32 − S31 S22 )

In index notation this would read

1
det(S) = ∈ijk ∈lmn Sli Smj Snk  
6

Note that if S and T are two tensors, then

T
det(S) = det (S )               det(S ⋅ T) = det(S) det(T)

Inverse

Let S be a second order tensor. The inverse of S exists if and only if det(S) ≠ 0, and is defined by
−1
S ⋅ S = I
where S −1
denotes the inverse of S and I is the identity tensor.

The inverse of a tensor may be computed by calculating the inverse of the matrix of its components. Formally, the inverse
of a second order tensor can be written in a simple form using index notation as
−1 1
S = ∈ipq ∈jkl Spk Sql
ji 2 det(S)

In practice it is usually faster to compute the inverse using methods such as Gaussian elimination.

Change of Basis.

Let S be a tensor, and let {e 1

, e2 , e3 } be a Cartesian basis. Suppose that the components of S in the basis {e 1
, e2 , e3 } are
known to be
(e) (e) (e)
⎡ S11 S
12
S
13
⎤

(e) ⎢ (e) (e) (e) ⎥

[S ] = ⎢S S S ⎥
⎢ 21 22 23 ⎥
(e) (e) (e)
⎣ ⎦
S S S
31 32 33

Now, suppose that we wish to compute the components of S in a second Cartesian basis, {m 1
, m2 , m3 } . Denote these
components by
(m) (m) (m)
⎡ S11 S
12
S
13
⎤

(m) ⎢ (m) (m) (m) ⎥

[S ] = ⎢S S S ⎥
⎢ 21 22 23 ⎥
(m) (m) (m)
⎣ ⎦
S S S
31 32 33

To do so, first compute the components of the transformation matrix [Q]

m 1 ⋅ e 1       m 1 ⋅ e 2       m 1 ⋅ e 3
⎡ ⎤

[Q] = ⎢ m 2 ⋅ e 1       m 2 ⋅ e 2       m 2 ⋅ e 3 ⎥

⎣ ⎦
m 3 ⋅ e 1       m 3 ⋅ e 2       m 3 ⋅ e 3

(this is the same matrix you would use to transform vector components from {e 1, e2 , e3 } to {m 1 , m2 , m3 } ). Then,

http://www.brown.edu/Departments/Engineering/Courses/En221/Notes/Tensors/Tensors.htm 6/10
12/24/2017 Continuum Mechanics - Tensors
(m) (e) T
[S ] = [Q][S ][Q]

or, written out in full

(m) (m) (m) (e) (e) (e)
⎡ S11 S
12
S
13
⎤ m 1 ⋅ e 1       m 1 ⋅ e 2       m 1 ⋅ e 3 ⎡ S11 S
12
S
13
⎤ m 1 ⋅ e 1       m 2 ⋅ e 1       m 3 ⋅ e 1
⎡ ⎤ ⎡ ⎤
⎢ (m) (m) (m) ⎥ ⎢ (e) (e) (e) ⎥
⎢S S S ⎥ = ⎢ m 2 ⋅ e 1       m 2 ⋅ e 2       m 2 ⋅ e 3 ⎥ ⎢ S S S ⎥ ⎢ m 1 ⋅ e 2       m 2 ⋅ e 2       m 3 ⋅ e 2
⎢ 21 22 23 ⎥ ⎢ 21 22 23 ⎥
⎣ ⎦ ⎣ ⎦
⎣ (m) (m) (m) ⎦ m 3 ⋅ e 1       m 3 ⋅ e 2       m 3 ⋅ e 3 ⎣ (e) (e) (e) ⎦ m 1 ⋅ e 3       m 2 ⋅ e 3       m 3 ⋅ e 3
S S S S S S
31 32 33 31 32 33

To prove this result, let u and v be vectors satisfying

v = S ⋅ u
Denote the components of u and v in the two bases by u
(e)
,     u
(m)
and v (e)
,     v
(m)
, respectively. Recall that the vector
−−− −−−− −−− −−−−
components are related by
(m) (e) (e) Q (m)
u = [Q]u                u = [ u
−−−− −−− −−− ] −
− −−

(m) (e) (e) Q (m)

v = [Q]v                v = [ v
−−−− −−− −−− ] −
− −
−
Now, we could express the tensor-vector product in either basis
(m) (m) (m) (e) (e) (e)
v = [S ]u                        v = [S ]u   
−−−− −−−− −−− −−−
Substitute for u (e)
,     v
(e)
from above into the second of these two relations, we see that
−−− −−−
T (m) (e) T (m)
[Q] v = [S ] [Q] u  
−−−− −−−−
Recall that
T (m) (m)
[Q] [Q] = [I ]                     [I ] v = v
−−−− −−−−
so multiplying both sides by [Q] shows that
(m) (e) T (m)
v = [Q] [S ] [Q] u  
−−−− −−−−
so, comparing with the first of equation (1)
(m) (e) T
[S ] = [Q][S ][Q]

as stated.

In index notation, we would write

(m) (e)
S = Qik S Qjl                Qij = m i ⋅ e j
ij kl

Another, perhaps cleaner, way to derive this result is to expand the two tensors as the appropriate dyadic products of the
basis vectors
(m) (e)
S mk ⊗ ml = S ek ⊗ el
kl kl

(m) (e)
⇒ m i ⋅ [S mk ⊗ ml ] ⋅ mj = mi ⋅ S e k ⊗ e l ⋅ mj
kl kl

(m) (e)
⇒ S = (m i ⋅ e k )S (e l ⋅ m j )
ij kl

Invariants

Invariants of a tensor are scalar functions of the tensor components which remain constant under a basis change. That is
to say, the invariant has the same value when computed in two arbitrary bases {e , e , e } and {m , m , m } . A 1 2 3 1 2 3

symmetric second order tensor always has three independent invariants.

Examples of invariants are

1. The three eigenvalues
2. The determinant
3. The trace
4. The inner and outer products

These are not all independent – for example any of 2-4 can be calculated in terms of 1.

http://www.brown.edu/Departments/Engineering/Courses/En221/Notes/Tensors/Tensors.htm 7/10
12/24/2017 Continuum Mechanics - Tensors

In practice, the most commonly used invariants are:

I1 = trace(S) = Skk

1 S 1
I2 = (trace( −S ⋅ ⋅S) = (Sii Sjj − Sij Sji )
2 ) 2

1
I3 = det(S) = ∈ijk ∈pqr Sip Sjq Skr =∈ijk Si1 Sj2 Sk3
6

Eigenvalues and Eigenvectors (Principal values and direction)

Let S be a second order tensor. The scalars λ and unit vectors m which satisfy
S ⋅ m = λm
are known as the eigenvalues and eigenvectors of S, or the principal values and principal directions of S. Note that λ may
be complex. For a second order tensor in three dimensions, there are generally three values of λ and three unique unit
vectors m which satisfy this equation. Occasionally, there may be only two or one value of λ . If this is the case, there are
infinitely many possible vectors m that satisfy the equation. The eigenvalues of a tensor, and the components of the
eigenvectors, may be computed by finding the eigenvalues and eigenvectors of the matrix of components.

The eigenvalues of a symmetric tensor are always real, and its eigenvectors are mutually perpendicular (these two results
are important and are proved below). The eigenvalues of a skew tensor are always pure imaginary or zero.

The eigenvalues of a second order tensor are computed using the condition det(S − λI) = 0. This yields a cubic
equation, which can be expressed as
3 2 2
λ − I1 λ + I2 λ − I3 = 0           I1 = trace(S),     I2 = (I − S ⋅ ⋅S)/2    I3 = det(S)
1

There are various ways to solve the resulting cubic equation explicitly – a solution for symmetric S is given below, but the
results for a general tensor are too messy to be given here. The eigenvectors are then computed from the condition
(S − λI)m = 0.

The Cayley-Hamilton Theorem

Let S be a second order tensor and let I 1 = trace(S),     I 2 = (I

2
1
− S ⋅ ⋅S)/2    I 3 = det(S) be the three invariants.
Then
3 2
S − I1 S + I2 S − I3 = 0

(i.e. a tensor satisfies its characteristic equation). There is an obscure trick to show this… Consider the tensor
S − αI (where α is an arbitrary scalar), and let T be the adjoint of S − αI , (the adjoint is just the inverse multiplied by

the determinant) which satisfies

3 2
T(S − αI) = det(S − αI)I = (−α + I1 α − I2 α + I3 )I

Assume that T= α 2
T1 + αT2 + T3 . Substituting in the preceding equation shows that
T1 = I         T1 S − T2 = I1 I            T3 − T2 S = I2 I               T3 S = I3 I
Use these to substitute for I 1
, I2 , I3 into
3 2 3 2
S − I1 S + I2 S − I3 = S − (S − T2 )S + (T3 − T2 S)S − T3 S = 0

3 SPECIAL TENSORS

Identity tensor The identity tensor I is the tensor such that, for any tensor S or vector v
I ⋅ v = v ⋅ I = v

S ⋅ I = I ⋅ S = S
In any basis, the identity tensor has components
1 0 0
⎡ ⎤

⎢0 1 0⎥
⎣ ⎦
0 0 1

Symmetric Tensor A symmetric tensor S has the property

T
S = S
The components of a symmetric tensor have the form
http://www.brown.edu/Departments/Engineering/Courses/En221/Notes/Tensors/Tensors.htm 8/10
12/24/2017 Continuum Mechanics - Tensors

S11 S12 S13

⎡ ⎤

⎢ S12 S22 S23 ⎥

⎣ ⎦
S13 S23 S33

so that there are only six independent components of the tensor, instead of nine. Symmetric tensors have some nice
properties:

· The eigenvectors of a symmetric tensor with distinct eigenvalues are orthogonal. To see this, let u, v be two
eigenvectors, with corresponding eigenvalues λ , λ . Then u v

⋅ v] = u ⋅ [S ⋅ v] ⇒ v ⋅ λ u = u ⋅ λ v ⇒ (λ − λ )u ⋅ v = 0 ⇒ u ⋅ v = 0 .
T
v ⋅ [S ⋅ u] = u ⋅ [S u v u v

· The eigenvalues of a symmetric tensor are real. To see this, suppose that λ, u are a complex
eigenvalue/eigenvector pair, and let λ ¯
¯ denote
, ū their complex conjugates. Then, by definition
¯. And hence ū¯ ⋅ [S ⋅ u] = λ ū¯ ⋅ u,       u ⋅ [S ⋅ ū¯] = λu ⋅ ū¯. But note that for a
¯ ¯
S ⋅ u = λu ⇒ S ū¯ = λ ū

symmetric tensor ū¯ ⋅ [S ⋅ u] = u ⋅ [S ⋅ ū¯] = u ⋅ [S ⋅ ū¯]. Thus λ ū¯ ⋅ u = λ

T ¯
¯ ⇒ λ = λ.
u ⋅ ū
¯

The eigenvalues of a symmetric tensor can be computed as

−
−− −
−−
I1 −p 3q −3 2(k−1)π
1 −1
λk = + 2√ cos { cos ( √ ) − }              k = 1, 2, 3
3 3 3 2p p 3

3
2I −9I1 I2 +27I3
1 2 1
p = I2 − I           q = −
3 1 27

1 2
I1 = trace(S)          I2 = (I − S : S)          I3 = det(S)
2 1

The eigenvectors can then be found by back-substitution into [S − λI] ⋅ m = 0. To do this, note that the matrix equation
can be written as
S11 − λ S12 S13 m1 0
⎡ ⎤⎡ ⎤ ⎡ ⎤

⎢ S12 S22 − λ S23 ⎥ ⎢ m2 ⎥ = ⎢ 0 ⎥

⎣ ⎦⎣ ⎦ ⎣ ⎦
S13 S23 S33 − λ m3 0

Since the determinant of the matrix is zero, we can discard any row in the equation system and take any column over to
the right hand side. For example, if the tensor has at least one eigenvector with m ≠ 0 then the values of m , m for 3 1 2

this eigenvector can be found by discarding the third row, and writing
S11 − λ S12 m1 S13
[ ][ ] = −m3 [ ]
S12 S22 − λ m2 S23

· Spectral decomposition of a symmetric tensor Let S be a symmetric second order tensor, and let {λ i
, ei } be the
three eigenvalues and eigenvectors of S. Then S can be expressed as
3

S = ∑ λi ei ⊗ ei

i=1

To see this, note that S can always be expanded as a sum of 9 dyadic products of an orthogonal basis.
λ k     m = k
S = Sij ei ⊗ ej . But since e are eigenvectors it follows that e
i m ⋅ (Sij ei ⊗ ej ) ⋅ ek = Smk = {
0       m ≠ k

Skew Tensor. A skew tensor S has the property

T
S = −S
The components of a skew tensor have the form
0 S12 S13
⎡ ⎤

⎢ −S12 0 S23 ⎥
⎣ ⎦
−S13 −S23 0

Every second-order skew tensor has a dual vector w that satisfies

S ⋅ u = w × u
for all vectors u. You can see this by noting that S = −w 12 3     S13 = w2    S23 = −w1 and expanding out the tensor and
cross products explicitly. In index notation, we can also write
Sij = − ∈ijk wk                     wi = −
1

2
∈ijk Sjk .
http://www.brown.edu/Departments/Engineering/Courses/En221/Notes/Tensors/Tensors.htm 9/10
12/24/2017 Continuum Mechanics - Tensors

Orthogonal Tensors An orthogonal tensor R has the property

T T
R ⋅ R = R ⋅ R = I

−1 T
R = R

An orthogonal tensor must have det(S) = ±1; a tensor with det(S) = +1 is known as a proper orthogonal tensor.
Orthogonal tensors also have some interesting and useful properties:
· Orthogonal tensors map a vector onto another vector with the same length. To see this, let u be an arbitrary
vector. Then, note that |R ⋅ u| = [R ⋅ u] ⋅ [R ⋅ u] = u ⋅ R R ⋅ u = u ⋅ u = |u|
2 T 2

· The eigenvalues of an orthogonal tensor are 1, e for some value of θ. To see this, let u be an eigenvector,
±iθ

with corresponding eigenvalue λ . By definition, R ⋅ u = λu. Hence,

[R ⋅ u] ⋅ [R ⋅ u] = λu ⋅ λu ⇒ u ⋅ u = λ u ⋅ u ⇒ λ = 1 . Similarly, λλ = 1 . Since the characteristic equation
2 ¯ 2

is cubic, there must be at most three eigenvalues, and at least one eigenvalue must be real.

Proper orthogonal tensors can be visualized physically as rotations. A rotation can also be represented in several other
forms besides a proper orthogonal tensor. For example
· The Rodriguez representation quantifies a rotation as an angle of rotation θ (in radians) about some axis n
(specified by a unit vector). Given R, there are various ways to compute n and θ. For example, one way would
be find the eigenvalues and the real eigenvector. The real eigenvector (suitably normalized) must correspond to
n; the complex eigenvalues give e . A faster method is to note that
iθ

T
trace(R) = 1 + 2 cos θ       2 sin θn = dual(R − R )

· Alternatively, given n and θ, R can be computed from

R = cos θI + W ⋅ W(1 − cos θ) + W sin θ

where W is the skew tensor that has n as its dual vector, i.e. W ij = − ∈ijk nk . In index notation, this formula is
R ij = cos θδij + ni nj (1 − cos θ) − sin θ ∈ijk nk

Another useful result is the Polar Decomposition Theorem, which states that invertible second order tensors can be
expressed as a product of a symmetric tensor with an orthogonal tensor:
T T T
A = R ⋅ U = V ⋅ R                       RR = I           U = U      V = V
Moreover, the tensors R, U, V are unique. To see this, note that
· A A is symmetric and has positive eigenvalues (to see that it’s symmetric, simply take the transpose,
T

and to see that the eigenvalues are positive, note that dx ⋅ (A T

⋅ A) ⋅ dx > 0 for all vectors dx).
· Let λ
2
k
and m be the three eigenvalues and eigenvectors of A
k
T
. Since the eigenvectors are orthogonal, we
A

can write A T
A = ∑ λ mk ⊗ mk
2
k
.
k=1

· We can then set U = ∑ λ k mk ⊗ mk and define R = AU . U is clearly symmetric, and also U

−1 2
= A
T
A .
k=1

To see that R is orthogonal note that: R R = U A AU = U U U = I.T −T T −1 −T 2 −1

· Given that U and R exist we can write R ⋅ U = [R ⋅ U ⋅ R ] ⋅ R so if we define V = RUR then T T

A = VR . It is easy to show that V is symmetric.

· To see that the decomposition is unique, suppose that A = ˆ

R U for some other tensors R , U . Then
ˆ ˆ ˆ

A
T ˆ
A = U . But A T
A has a unique square root so ˆ
U = U . The uniqueness of R follows immediately.

http://www.brown.edu/Departments/Engineering/Courses/En221/Notes/Tensors/Tensors.htm 10/10