Tensors: Appendix B
Tensors: Appendix B
Tensors: Appendix B
Tensors
Scalar and vector fields are often not sufficient to describe the real physical
world, so that we have to introduce what can be thought as an extension of these
tools: the tensor fields. In the simplest form a Cartesian tensor can be defined
as
T = ei τij ej = τij ei ej = τij , i, j = 1, 2, 3.
In vector notation, a tensor will be denoted by uppercase calligraphic characters
and its components with the same characters already used for scalars. In Carte-
sian index notation a tensor is represented directly by its components. Thus, the
expression τij is sufficient to denote a tensor.
Clearly, this is an object characterized by nine components that can be ar-
ranged as a matrix
τ11 τ12 τ13 τxx τxy τxz
T = τ21 τ22 τ23 = τyx τyy τyz .
τ τ32 τ33 τzx τzy τzz
31
The tensor is symmetric if τij = τji , and antisymmetric if τij = −τji . The trace
Tr T = τii
vii
viii Franco Mattioli (University of Bologna)
of a tensor, obtained for contraction of the indices, is defined as the sum its
diagonal terms. The determinant of a tensor
In both cases the result is a vector which, however, is different in the two cases,
unless the tensor is symmetric. Writing the tensor in matrix form, the two
products correspond to the multiplication by a row vector on the left and by a
column vector on the right. Thus a tensor can be seen as an operator able to
generate a vector starting from another vector.
An expression of the kind
c=T ·b or c=b·T
Here, in the expression uv, the symbols of the two vectors are simply placed near
each other, without any intermediate punctuation mark. This particular tensor
applied to any vector furnishes a vector parallel to u.
so that
a0l e0l = ai ei .
If we multiply scalarly this expression by e0l , we obtain
a0l = ai e0l · ei .
ai = a0l ei · e0l .
xii Franco Mattioli (University of Bologna)
we have
a0l = uli ai , and ai = uil a0l .
The information about the change of the system of coordinates is contained
in the matrix U. Since the length of a vector is a scalar quantity, then
It follows that
ukl ulm = δkm .
The matrix U is therefore unitary, i.e., equal to its inverse.
The first property we have discussed about tensors is that a tensor applied
to a vector (i.e., multiplied scalarly by a vector) must result in another vector.
Let T be the tensor. In two systems of reference, rotated with respect to each
other, the initial and final vectors must be the same. If
and
ci = τij bj , and c0p = τpq
0 0
bq ,
it follows
0 0
τpq bq = c0p = uipci = uip τij bj = uip τij ujq b0q .
This expression can be rewritten as
0
τpq − uipτij ujq b0q = 0.
Problem B.9 Prove that A = bc is a tensor on the basis of the present definition.
Solution.
a0pq = b0p c0q = uipbi ujq cj = uip ujq bi cj = uip ujq aij .
Principles of Fluid Dynamics (www.fluiddynamics.it) xiii
Problem B.10 Prove that the identity tensor is the same in any reference system.
Solution.
0
δpq = uip ujq δij = uip uiq = δpq = δij .
Comment. A tensor whose components are the same in all the reference systems is said
to be isotropic.
At this point we can reverse the argument and define the tensor as an ensemble
of nine components arranged as a 3x3 matrix subject to the transformation rule
(B.3). Reversing the above demonstration, we conclude that a tensor is able to
derive a vector from another vector by scalar multiplication. Thus this latter
property becomes the proof that a certain arrangement of nine components are
in effect the components of a tensor. This procedure is called quotient rule and
is the easiest way to detect tensors in the framework of a physical context.
D · a = λa = λIa.
for three independent real values of λ. The three normal direction are called
principal directions.
We have to solve the set of equations
(dij − λδij ) ej = 0.
λ3 − Θλ2 + Φλ − Ψ = 0,
where
Θ = d11 + d22 + d33
is the trace of the tensor,
Φ = d22 d33 − d23 d32 + d33 d11 − d31 d13 + d11 d22 − d12 d21
xiv Franco Mattioli (University of Bologna)
and
Ψ = d11 d22 d33 + d12 d23 d31 + d13 d21 d32 − d13 d22 d31 − d11 d23 d32 − d12 d21 d33
is the determinant of the tensor. The three solutions for λ are real, while the three
correspondent vectors ej , normalized to unity, are called characteristic vectors.
It turns out that such vectors are normal with respect to each other.
Since, the principal directions of a tensor are the same independently of the
reference system, it follows that the coefficients Θ, Φ and Ψ are invariant quan-
tities, i.e., scalar quantities.
Problem B.11 Prove directly that the trace of a tensor is invariant under rotation in
index notation.
Solution. Let d0pq the representation of the tensor dij in the rotated frame. It follows
But the main result is that any tensor can be diagonalized. In other words we
can always associate to a tensor a reference system in which it can be represented
by a diagonal matrix. Seen from this particular reference system the physical
structure of the tensor is very simple.
All these properties are not intuitive, and appear rather strange in a physical
context. The geometry of the space impose strong constraints to the physical
quantities of our world.
is
λ2 − T 2 = 0
and vanishes for λ(1) = T and λ(2) = −T . With respect to the principal directions the
tensor takes the form
0
T 0
T = .
0 −T
Principles of Fluid Dynamics (www.fluiddynamics.it) xv
In correspondence to the first eigenvalue, we derive the first eigenvector e1 from the
system
−T T e1x
T = 0.
−T e1y
√ √ √
It follows e1x√
= 1/ 2 and e1y = 1/ 2, after normalization. Analogously, e2x = −1/ 2
and e2y = 1/ 2. Therefore, the principal directions are rotated by 45◦ with respect to
the original axes.
Comment. In a symmetric tensor with a trace equal to zero the diagonal and off-diagonal
terms have the same importance. Changing the reference frame we can have only diagonal
terms or only off-diagonal terms, whose magnitude is of the same order.
Problem B.13 The addition of an isotropic tensor to a given tensor will change its prin-
cipal directions?
Solution. The principal directions of the new tensor remain the same as those of the old
one because an isotropic tensor cannot change the direction of the eigenvectors of the
original tensor. In the reference system of the principal directions the diagonal terms of
the final tensor are given by the sum of the diagonal terms of the two tensors.
Comment. The diagonal terms of the tensor behave differently with respect to the other
off-diagonal terms, because they are made of two components. A component is invariant
under rotations, while the other component mixes with the off-diagonal components as
in a tensor with a vanishing trace.
For an antisymmetric tensor (for which eigenvalues and eigenvector are com-
plex) the three invariant quantities correspond to a zero trace, the square of
the length of the associated vector and to a zero determinant, respectively (see
problem [B.3]).
For comparison, consider that a vector has only one invariant quantity, its
length.
Certain operations defined between second-order tensor can give rise to higher-
order tensors. For example the fourth-order tensor F can be obtained by the two
second-order tensors B and C on the basis of the following definition
fijpq = bij cpq .
So, while joining two vectors gives rise to a second-order tensor, joining two
second-order tensors gives rise to a fourth-order tensor. In general, by joining
two tensor of order m and n we obtain a tensor of order m + n.
In particular, the application a fourth-order tensor F to a second-order tensor
B gives rise to a second-order tensor C
cij = fijpq bpq ,
similarly to the scalar product between second-order tensors and vectors.
A tensor is said to be isotropic if its components remain unchanged under
a rotation of the reference frame. All the scalars are isotropic. There are no
isotropic vectors. The tensors proportional to the identity tensor δij are the only
second-order isotropic tensors and those proportional to the permutation symbol
ijk are the only third-order isotropic tensors. It is possible to show that there are
many fourth-order isotropic tensors H. They can be identified by the expression
hijpq = c1 δij δpq + c2 δip δjq + c3 δiq δjp , (B.4)
where the ci are arbitrary constants.
Problem B.14 Find the structure of the most general linear and isotropic transformation
between two symmetric tensors.
Solution. The most general linear relation between two symmetric tensors can be written
τij = fijpq dpq , (B.5)
where fijpq is a fourth order tensor. If the medium is isotropic, then F must be isotropic
as well. According to (B.4) we have
lijpq = hijpq = c1 δij δpq + c2 δip δjq + c3 δiq δjp .
Let us rewrite this expression as
lijpq = λδij δpq + µ (δip δjq + δiq δjp ) + ν (δip δjq − δiq δjp ) ,
where λ = c1 , µ = (c2 + c3 )/2 and ν = (c2 − c3 )/2. Then (B.5) becomes
τij = lijpq dij = λδij dpp + µ (dij + dji) = 2µdij + λδij dpp . (B.6)
Observe that the term in ν gives no contribution.
Principles of Fluid Dynamics (www.fluiddynamics.it) xvii
Since scalars, vectors and tensors are physical entities, they can be defined in
whatever reference system. So far we have considered only Cartesian systems of
reference, because they are the simplest ones. But for particular problems other
coordinate systems can become more convenient or intrinsically necessary.
For example, the study of motions involving the whole earth’s surface can be effi-
ciently faced using spherical coordinates.
In the present text only right-handed frames of reference are considered, so
that no difference is made between polar or axial vectors.
In some old text of oceanography left-handed frames of references have been used,
because the space of interest was in the sole lower half-space. This complicates the
mathematical expressions that sometimes are the same, but other times are the contrary
of the analogous expressions for right-handed systems.