You are on page 1of 28

1

Vectors and Tensors

1 Elementary properties of vectors


Definition: A real linear vector space consists of of a set V, a mapping (u, v) =⇒ u+v
of V ×V into V , and a mapping (λ, v) =⇒ λv of R×V into V such that the following
conditions are satisfied.

1. V is a commutative group with respect to ’+’

i
2. the associative law λ(µv) = (λµ)v holds

ora
3. the distributive laws (λ + µ)v = λv + µv and λ(u + v) = λu + λv hold

4. 1v = v
Gh
Definition: Elements v1 , v2 , · · · , vn of V are called linearly dependent if there exits
real number λ1 , λ1 , · · · , λn not all zero such that

λ 1 v 1 + λ2 v 2 + · · · + λn v n = 0
:S

If such numbers do not exist, then the elements v1 , v2 , · · · , vn of V are linearly


aft

independent.
Definition: A vector space V is said to be finite dimensional if there exists a finite
set of vectors v̂1 , v̂2 , · · · , v̂n such that
Dr

1. v̂1 , v̂2 , · · · , v̂n are linearly indpendent

2. ∀v ∈ V , there exists λ1 , λ1 , · · · , λn ∈ R such that

v = λ1 v̂1 + λ2 v̂2 + · · · + λn v̂n

The set {v̂1 , v̂2 , · · · , v̂n } is called a basis of V and the above representation is unique.
λ1 , λ1 , · · · , λn are called coordiantes of v w.r.t. the basis {v̂1 , v̂2 , · · · , v̂n }.
Definition: A three dimensional vector space V is said to be Eucledian vector space
if for (u, v) ∈ V × V, we can define the following two operations u · v ∈ R and
(scaler product) u ∧ v ∈ V with the following properties:

1. u · v = v.u

2. (αu + βv) · w = αu · w + βv · w
2

3. u · u ≥ 0 and u · u = 0 iff u = 0

4. u ∧ v = −v ∧ u

5. (αu + βv) ∧ w = α(u ∧ w) + β(v ∧ w)

6. u · (u ∧ v) = 0

7. (u ∧ v) · (u ∧ v) = (u · u)(v · v) − (u · v)2

The norm of a vector u is defined by

i
|u| := (u · u)1/2

ora
The vectors u, v is called orthogonal if u · v = 0.
Definition: The scaler triple product of three vectors u, v, w is defined by
Gh
[u, v, w] = u · (v ∧ w)

Properties:
:S

1. [u, v, w] = [v, w, u] = [w, u, v] = −[v, u, w] = −[w, v, u] = −[u, w, v]

2. [α1 u1 + α2 u2 , v, w] = α1 [u1 , v, w] + α2 [u2 , v, w]


aft

3. [u, v, w] = 0 iff u, v, w are linearly dependent.

In three dimensional vector space with basis vectors e1 , e2 , e3 , a vector u may be


Dr

written as
u = u1 e1 + u2 e2 + u3 e3

where we usually take the basis vectors e1 , e2 , e3 to be mutually orthonormal i.e.


(
1, if i = j
ei · ej =
0, if i 6= j

2 Indicial Notation

2.1 Summation Convention, Dummy Indices


Consider the sum
s = a1 x 1 + a2 x 2 + · · · + an x n (1)
3

which can be written in the compact form


n
X
s= ai x i (2)
i=1

It is obvious that the index i in (2) can be replaced by j or m. The index i is (2) is
a dummy index in the sense that the sum is independent of the letter used.
We can further simplify the writing of (1) if we adopt the Einstein’s summation
convention: Whenever an index is repeated once, it is a dummy index indicating a
summation with index running through the natural numbers 1, 2, · · · , n.
Using this convention, (1) shortens to

i
ora
s = ai x i (3)

It is also clear that


ai x i = aj x j = am x m (4)
Gh
Note that ai bi xi is not defined with this convention (index is repeated more than
once). Thus the expression
n
X
s= ai b i x i
i=1
:S

must retains its summation sign.


Since we usually deal with three dimensional cartesian field, we shall always take
aft

n = 3. Thus

ai x i = a1 x 1 + a2 x 2 + a3 x 3
Dr

aii = a11 + a22 + a33


ui ei = u1 e1 + u2 e2 + u3 e3

The summation convention can be used to express double sum, triple sum, etc. For
example we can write the double sum involving 9 terms as
3 X
X 3
aij xi xj = aij xi xj
i=1 j=1

For beginers, it is better to perform the above expansion in two steps, first, sum
over i and then sum over j. Similarly the triple sum involving 27 terms is given by
3 X
X 3 X
3
aijk xi xj xk = aijk xi xj xk
i=1 j=1 k=1

As before the expression aii xi xj xj or aijk xi xj xk are not defined in the summation
convention.
4

2.2 Free Indices


Consider the following system of three equations

x01 = a11 x1 + a12 x2 + a13 x3


x02 = a21 x1 + a22 x2 + a23 x3
x03 = a31 x1 + a32 x2 + a33 x3

which using the summation convention can be written as

x01 = a1m xm

i
ora
x02 = a2m xm
x03 = a3m xm

which can be shortened into


Gh
x0i = aim xm , i = 1, 2, 3 (5)

An index which appears only once in each term of an equation such as index i in
:S

(5) is called free index. A free index takes on values 1, 2 or 3 one at a time. Also
x0j = ajm xm , j = 1, 2, 3 is the same as (5). However
aft

ai = b j

is a meaningless equation. The free index appearing in every term of the equation
Dr

must be the same. Thus the following are meaning full

ai + b i = c i
ai + bi cj dj = 0

If there are two free indices appearing in an equation such as

Tij = Aim Ajm , i = 1, 2, 3; j = 1, 2, 3 (6)

then the equation is a short-hand writing of 9 equations. Again equations such as

Tij = Tik

have no meaning.
5

2.3 Kronecker Delta


The Kronecker delta, denoted by δij , is defined as
(
1, if i = j
δij = (7)
0, if i 6= j
Thus the matrix [δij ] is the identity matrix. The following properties are easy to
prove

1. δii = 3

2. δim am = ai

i
ora
3. δim Tmj = Tij

Also the the relations δim δmj = δij and δim δmj δjn = δin are easy to prove. If e1 , e2 , e3
are mutually orthonormal vectors then
Gh
ei · ej = δij

2.4 Permutation Symbol


:S

The permutation symbol (also known as alternator), denoted by ijk , is defined by



 +1 if(ijk) form an even permuation of (123)
aft



ijk = −1 if(ijk) form an odd permuation of (123) (8)


 0 otherwise
Dr

Thus we have
ijk = jki = kij = −jik = −ikj = −kji
The permutation symbol allows us to write determinant of an 3 × 3 matrix in a very
compact form

T11 T12 T13



T21 T22 T23 = pqr T1p T2q T3r = pqr Tp1 Tq2 Tr3


T31 T32 T33
If {e1 , e2 , e3 } be an orthonormal basis for Eucledian vector space V, then the fol-
lowing relation holds

e2 ∧ e3 = ±e1
e3 ∧ e1 = ±e2
e1 ∧ e2 = ±e3 ,
6

where the + or − sign holds together. This can be written as

ei ∧ ej = ±ijk ek

For any vector a = ai ei ∈ V, we have

a · ej = ai ei · ej = ai δij = aj

Thus we have a = (a · ei )ei . Using this we get e2 ∧ e3 = [e1 , e2 , e3 ]e1 . Similarly


e3 ∧e1 = [e1 , e2 , e3 ]e2 and e1 ∧e2 = [e1 , e2 , e3 ]e3 . Now using property 7 of Eucledian
vector space (take u = e2 and v = e3 ) we get [e1 , e2 , e3 ] = ±1. We usually take

i
{e1 , e2 , e3 } to be right hand triad. In this case [e1 , e2 , e3 ] = 1.

ora
Now we can write vector product in terms of components.

u ∧ v = (ui ei ) ∧ (vj ej ) = ui vj ei ∧ ej
Gh
= ijk ui vj ek
= (u2 v3 − u3 v2 )e1 + (u3 v1 − u1 v3 )e2 + (u1 v2 − u2 v1 )e3

This is formally written as


:S

e e e

1 2 3

u ∧ v = u1 u2 u3
aft


v1 v2 v3

Similarly we can write scaler triple product as


Dr

[u, v, w] = u · (v ∧ w) = uk (v ∧ w)k = uk (ijk vi wj ) = kij uk vi wj = pqr up vq wr

This is equivalent to the determinant



u u u u v w

1 2 3 1 1 1

v1 v2 v3
or u2 v2 w2


w1 w2 w3 u3 v3 w3

2.5 Manipulation with the Indicial Notations


Substitution: If ai = Uim bm and bi = Vim cm , then ai = Uim Vmn cn .
Multiplication: If p = am bm and q = cm dm , then pq = am bm cn dn .
Factoring: Let Tij nj −λni = 0. Then using Kronecker delta, we can write ni = δij nj .
So we get
(Tij − λδij )nj = 0
7

Contraction: The operation of identifying two indices and summing on them is


known as contraction. For example, Tii is the contraction of Tij . Thus if

Tij = λθδij + 2µEij ,

then
Tii = λθδii + 2µEii = 3λθ + 2µEii .

3 Second Order Tensors

i
Definition: A second order tensor A is a linear transformation of three dimensional

ora
vector space V into itself. Thus

A(αu + βv) = α(Au) + β(Av),


Gh ∀u, v ∈ V, α, β ∈ R (9)

If A and B are two second order tensors, we can define the sum A + B, difference
A − B and scaler multiplication αA by the following rules

(A ± B)u := Au + Bu (10)
:S

(αA)u := α(Au) (11)


aft

There are two particularly important second order tensors (or simply tensors) ,
identity tensor I and zero tensor O which are defined by

Iu = u and Ou = 0
Dr

Using the definitions (10) and (11), we see that the set of all tensors (i.e. second order
tensors) is a linear vector space. We shall denote this space by CT (2). Subsequenly
we shall prove that this is finite dimensional and hence any tensor can be expressed
as a linear combination of simple tensors.
Definition: The tensor (or dyadic) product of two vector u and v denoted by u ⊗ v
and represents the linear transformation defined by

(u ⊗ v)w = (w · v)u ∀w ∈ V (12)

This satisfies the following properties

1. (αu + βv) ⊗ w = α(u ⊗ w) + β(v ⊗ w)

2. w ⊗ (αu + βv) = α(w ⊗ u) + β(w ⊗ v)


8

3. In general u ⊗ v 6= v ⊗ u

Note: If u ⊗ w = v ⊗ w holds for every w 6= 0 then u = v. (Prove by operating


both sides on w).
Proposition: If B = {e1 , e2 , e3 } is a basis of V, then B̂ = {ei ⊗ ej |1 ≤ i, j ≤ 3} forms
a basis of CT (2). Hence dim(CT (2)) = 9 and any A ∈ CT (2) can be represented
as A = Aij ei ⊗ ej .
Proof: Let us first check the independence of B̂. Let

αij ei ⊗ ej = O, αij ∈ R

i
ora
Let p ∈ {1, 2, 3} be fixed but arbitrary. Then

0 = αij (ei ⊗ ej )ep = αij (ej ⊗ ep )ei = αip ei


Gh
Since B is linearly independent, we have αip = 0 for i = 1, 2, 3. Now since p is
arbitrary, we let αip = 0 for p = 1, 2, 3. Thus B̂ is independent. Now we want to
prove that every tensor A ∈ CT (2) can be expressed as linear combination of B̂.
Let us write
:S

A = Aij ei ⊗ ej , Aij ∈ R

Let p, q ∈ {1, 2, 3} are fixed but arbitrary. Now


aft

Aeq = Aij (ei ⊗ ej )eq = Aiq ei


Dr

which in turn implies


ep · Aeq = Apq

This shows that the coefficients Aij can be determined uniquely.


Notes:

1. We can write v = Au in the component from as

v1 = e1 · v = e1 · A(u1 e1 + u2 e2 + u3 e3 = A11 u1 + A12 u2 + A13 u3

Similarly

v2 = A21 u1 + A22 u2 + A23 u3


v3 = A31 u1 + A32 u2 + A33 u3
9

Thus bi = Aij uj . This can be written in the matrix form as


    
v1 A11 A12 A13 u1
    
 v2  =  A21 A22 A23   u2 
 
   
v3 A31 A32 A33 u3
The matrix  
A11 A12 A13
 
 A21 A22 A23 
 
A31 A32 A33
is called the matrix of the tensor A with respect to B = {e1 , e2 , e3 }. Note that

i
the components in the first column of the matrix are the components of the

ora
vector Ae1 etc. It should be noted that the components of A depend on the
coordinate system being used through the basis B = {e1 , e2 , e3 } just like the
components of a vector. A vector does not depend on any coordiante system
Gh
even though its components do. Similarly a tensor does not depend on any
coordinate system even though its components do.

2. Returning to the tensor product of two vectors, we have


:S

(u ⊗ v)ij = ((up ep ) ⊗ (vq eq ))ij = up vq (ep ⊗ eq )ij

which further on simplification gives


aft

(u⊗v)ij = up vq ei ·(ep ⊗eq )ej = up vq ei ·(ej ·eq )ep = up vq δjq ei ·ep = up vq δjq δip = ui vj

Thus u ⊗ v can be written in the matrix form as


Dr

 
u1 v 1 u1 v 2 u1 v 3
 
 u2 v 1 u2 v 2 u2 v 3 
 
u3 v 1 u3 v 2 u3 v 3

3. The identity tensor I admits the following representation

I = ep ⊗ ep

Indeed, we have
(ep ⊗ ep )u = u = Iu
Thus the components of I are given by
 
1 0 0
 
 0 1 0 
 
0 0 1
10

Definition: Let A ∈ CT (2).

1. The transpose of A denoted by AT ∈ CT (2) is defined by

u · (AT v) = v · (Au)

2. A is said to be symmetric if AT = A and skew-symmetric if AT = −A.

3. The product of two tensors A and B denoted by AB are defined by

(AB)u = A(Bu), ∀u ∈ V

i
ora
4. A is said to be positive semi-definite if u · (Au) ≥ 0 ∀u ∈ V. And it is called
positive definite if u · (Au) > 0 ∀0 6= u ∈ V.

5. A is said to be invertible if there exits A−1 ∈ CT (2) such that


Gh
AA−1 = A−1 A = I

6. A is said to be orthogonal if
:S

AT A = AAT = I
aft

Observations:

1. From the definition of transpose of a tensor, we have


Dr

ei · (Aej ) = ej · (AT ei )

Thus
Aij = ATji

Thus a symmetric tensor has 6 independent components and a skew-symmetric


tensor has 3 independent components.

2. Any tensor A ∈ CT (2) can be written as the sum of a symmetric and a


skew-symmetric tensors,
1 1
A = (A + AT ) + (A − AT )
2 2
11

3. If A ∈ CT (2) is a skew-symmetric tensor, then there exits a vector, called the


dual or axial vector of A, denoted by tA , such that

Au = tA ∧ u, ∀u ∈ V

Proof: Since A is skew-symmetric, we can write

A = A23 (e2 ⊗ e3 − e3 ⊗ e2 ) + A13 (e1 ⊗ e3 − e3 ⊗ e1 ) + A12 (e1 ⊗ e2 − e2 ⊗ e1 )

Then using the definition of tensor product of vectors we have

Au = A23 u3 e2 − A23 u2 e3 + A13 u3 e1 − A13 u1 e3 + A12 u2 e1 − A12 u1 e2

i
ora
= (A12 u2 + A13 u3 )e1 + (A23 u3 − A12 u1 )e2 + (−A13 u1 − A23 u2 )e3

Thus if tA = (tA A A
1 , t2 , t3 ), then

tA
1 = −A23 , tA
2 = A13 , tA
3 = −A12
Gh
Observation: If Au = tA ∧ u, ∀u ∈ V then A is skew-symmetric.
Proof: For any x, y ∈ V we have
:S

x · (Ay) = x · (tA ∧ y)
= −y · (tA ∧ x)
aft

= −y · (Ax)

Hence AT = −A
Dr

4. The transpose of a tensor in CT (2) can easily be shown to satisfy

1. (αA + βB)T = αAT + βB T


2. (AB)T = B T AT
3. (AT )T = A

5. (u ⊗ v)T = v ⊗ u
Proof: Let A = (u ⊗ v). Then

y · (Ax) = x · (AT y), ∀x, y ∈ V

Now

y · (Ax) = y · [(u ⊗ v)x] = y · [(x · v)u]


= (y · u)(x · v) = x · [(y · u)v] = x · [(v ⊗ u)y]
12

which implies (u ⊗ v)T = v ⊗ u


Properties of ⊗:

(u ⊗ v)(w ⊗ x) = (v · w)(u ⊗ x) (13)


A(u ⊗ v) = (Au) ⊗ v (14)
(u ⊗ v)A = u ⊗ (AT v) (15)

Proof: Let S = (u ⊗ v)(w ⊗ x) and T = (v · w)(u ⊗ x). Note that both T


and S are in CT (2). To prove the first result, we need to show that T a = Sa
for all a ∈ V. Now

i
ora
T a = [(u ⊗ v)(w ⊗ x)]a = (u ⊗ v)[(w ⊗ x)a] = (u ⊗ v)[(a · x)w]
= (a · x)(v · w)u

and
Gh
Sa = (v · w)(u ⊗ x)a = (v · w)(a · x)u.

To prove the second properties, let S = A(u ⊗ v) and T = (Au) ⊗ v. Now


proceeding as before we have
:S

Sa = [A(u ⊗ v)]a = A[(u ⊗ v)a] = A[(a · v)u] = (a · v)Au

and
aft

T a = [(Au) ⊗ v]a = (a · v)Au


To prove the third relation, again let us take S = (u⊗v)A and T = u⊗(AT v).
Dr

Now

Sa = [(u ⊗ v)A]a = (u ⊗ v)(Aa) = [v · (Aa)]u = [a · (AT v)]u

and
T a = [u ⊗ (AT v)]a = [a · (AT v)]u

Application: Let us find the components of the second order tensor T = AB. We
have A = Aij ei ⊗ ej and B = Bij ei ⊗ ej . Now

T = AB = (Aij ei ⊗ ej )(Bpq ep ⊗ eq ) = Aij Bpq (ei ⊗ ej )(ep ⊗ eq )


= Aij Bpq (ej · ep )(ei ⊗ eq ) = Aij Bpq δjp (ei ⊗ eq )
= Aip Bpq (ei ⊗ eq ) = Aip Bpj (ei ⊗ ej )

Thus we have
T = Tij ei ⊗ ej , with Tij = Aip Bpj
13

3.1 Change of basis


Vectors and tensors remain invariant upon change of basis, they are said to be
independent of coordinate system. However their respective components do depend
on the coordinate system being used.
Let V denote the usual Eucledian vector space and take a right-handed orthonor-
mal basis vectors e1 , e2 , e3 so that any vector u can be written as

u = ui ei

Now choose a second orthonormal basis ê1 , ê2 , ê3 and we may write u in this new

i
ora
basis as
u = ui êi

We can express êi as linear combination of e1 , e2 , e3 . Thus


Gh
êi = Qij ej , (i = 1, 2, 3)

Thus we have
êi · êj = Qij
:S

which are called the direction cosines of êi relative to ej . Now we have
aft

δij = êi · êj = Qip ep · êj = Qip Qjp

Also we can write


Dr

ei = (ei · êp )êp = Qpi êp

and using δij = ei · ej we get


Qki Qkj = δij

In matrix notation the above two can be expressed as

[Q][Q]T = [Q]T [Q] = [I]

Thus [Q] is an orthogonal matrix. Now we have det[Q] = ±1. If det[Q] = 1 then
[Q] is a proper orthogonal matrix representing rotation. If det[Q] = −1, then [Q] is
improper orthogonal which represents a rotation combine with a reflection.
We now have the inverse transformation as

Qip êi = Qip Qij ej = δpj ej = ep


14

Thus we have the important relations

êi = Qij ej and ei = Qji eˆj

Also since
u = uj ej = uj Qij êi , u = ûi êi

we get
ûi = Qij uj

and similarly
ui = Qji eˆj

i
ora
Next we consider the transformation law for tensors. We have

A = Aij ei ⊗ ej ,
Gh A = Âij êi ⊗ êj

Now
Aij = ei · Aej , and Âij = êi · Aeˆj

Thus
Âij = (Qip ep ) · A(Qjq eq ) = Qip Apq Qjq
:S

and similarly
aft

Aij = Qpi Qqj Âpq

Using the matrix notation we have

[Â] = [Q][A][Q]T , [A] = [Q]T [Â][Q]


Dr

Alternate tensor definition: A cartesian tensor of order n is a mathematical en-


tity A which has components Ai1 i2 i3 ···in relative to rectangular cartesian basis B =
{e1 , e2 , e3 } and transforms like

Âi1 i2 i3 ···in = Qi1 j1 Qi2 j2 · · · Qin jn Aj1 j2 j3 ···jn

under a change of basis to B̂ = {ê1 , ê2 , ê3 } given by êi = Qij ej with [Q] a proper
3 × 3 orthogonal matrix. The set of all cartesian tensor of order n is a linear vector
space denoted by CT (n).
Note: In the definition of tensors using the alternated definition, we insist on proper
orthogonal transfomration. Since there are pseudo tensors which transforms accord-
ing to
Âi1 i2 i3 ···in = det(Q)Qi1 j1 Qi2 j2 · · · Qin jn Aj1 j2 j3 ···jn
15

For example consider the vector p = a ∧ b. In the transformed coordinate we get

p0i = a0 ∧ b0 = ijk a0j b0k


= ijk Qjp Qkq ap bq
= rpq ap bq det(Q)Qir = det(Q)Qir pr

Where we have used

ijk Qjp Qkq Qir = rpq det(Q)


=⇒ ijk Qjp Qkq Qir Qmr = rpq det(Q)Qmr
mjk Qjp Qkq = rpq det(Q)Qmr

i
ora
Definition: If the components of A ∈ CT (n) are unchanged under arbitrary rota-
tions of orthonormal basis, then A is said to be isotropic.
Examples:
Gh
1. The set of all scalers CT [0] are isotropic.

2. The zero vector 0 is the only isotropic vector. To satisfy the definition of
isotropy, we must have
:S

ûi = Qij uj = ui
for all rotations Qij . Let us choose
aft


0 1 0


[Qij ] = −1 0 0 

Dr


0 0 1

which represents a rotation of π/2 about e3 . This gives u1 = u2 = 0. Taking


any other choice would lead to u3 = 0. Thus u = 0.

3. Scaler multiple of δij are the only isotropic tensor of CT (2). We have

Aˆij = Qip Qjq Apq = Aij

for all rotations [Q]. The choice of [Q] of the previous examples gives

A11 = A22 , A13 = A23 = 0

Now taking 
1 0 0


[Qij ] = 0 0 1 
0 −1 0

16

gives
A22 = A23 , A21 = A31 = 0

Next choose [Q] representing rotations around e2 etc. Thus we have

Aij = A11 δij

where A11 is a scaler.

4. Scaler multiples of ijk are the only isotropic tensor in CT (3).

5. The three basic idependent isotropic tensors in CT (4) are

i
ora
δij δkl , δik δjl , δil δjk

Thus a general isotropic tensor in CT (4) has components of the form

αδij δkl + βδik δjl + γδil δjk


Gh
Principal Invariants: Corresponding to arbitrary tensor A ∈ CT (2), there are
scalers IA1 , IA2 , IA3 such that ∀u, v, w ∈ V
:S

[Au, v, w] + [u, Av, w] + [u, v, Aw] = IA1 [u, v, w], (16)


[Au, Av, w] + [Au, v, Aw] + [u, Av, Aw] = IA2 [u, v, w], (17)
aft

[Au, Av, Aw] = IA3 [u, v, w] (18)

These IA ’s are called pricipal invariants of A. Let u = ui ei , v = vj ej , w = wk ek


Dr

then

[Au, v, w] + [u, Av, w] + [u, v, Aw]


= ui vj wk {[Aei , ej , ek ] + [ei , Aej , ek ] + [ei , ej , Aek ]}
= ijk ui vj wk {[Ae1 , e2 , e3 ] + [e1 , Ae2 , e3 ] + [e1 , e2 , Ae3 ]}
= [u, v, w] ([Ae1 , e2 , e3 ] + [e1 , Ae2 , e3 ] + [e1 , e2 , Ae3 ])

The third relations follows from the fact that if any two i, j, k are equal, then the
sum is zero, etc. Now

[Ae1 , e2 , e3 ] + [e1 , Ae2 , e3 ] + [e1 , e2 , Ae3 ]


= [Ai1 ei , e2 , e3 ] + [e1 , Ai2 ej , e3 ] + [e1 , e2 , Ak3 ek ]
= Ai1 [ei , e2 , e3 ] + Aj2 [e1 , ej , e3 ] + Ak3 [e1 , e2 , ek ]
= (A11 + A22 + A33 ) [e1 , e2 , e3 ] = (A11 + A22 + A33 )
17

Thus we have IA1 = (A11 + A22 + A33 ) = tr(A).


Proceeding same way as before we have

[Au, Av, w] + [Au, v, Aw] + [u, Av, Aw]


= ui vj wk {[Aei , Aej , ek ] + [Aei , ej , Aek ] + [ei , Aej , Aek ]}
= ijk ui vj wk {[Ae1 , Ae2 , e3 ] + [Ae1 , e2 , Ae3 ] + [e1 , Ae2 , Ae3 ]}
= [u, v, w] ([Ae1 , Ae2 , e3 ] + [Ae1 , e2 , Ae3 ] + [e1 , Ae2 , Ae3 ])

Again

[Ae1 , Ae2 , e3 ] + [Ae1 , e2 , Ae3 ] + [e1 , Ae2 , Ae3 ]

i
ora
= [Ai1 ei , Aj2 ej , e3 ] + [Ai1 ei , e2 , Ak3 ek ] + [e1 , Aj2 ej , Ak3 ek ]
= Ai1 Aj2 [ei , ej , e3 ] + Ai1 Ak3 [ei , e2 , ek ] + Aj2 Ak3 [e1 , ej , ek ]
= (A11 A22 − A12 A21 ) [e1 , e2 , e3 ] + (A22 A33 − A23 A32 )[e1 , e2 , e3 ]
Gh
+ (A11 A33 − A13 A31 )[e1 , e2 , e3 ]

Thus

A
11 A12
A
22 A23
A
11 A13
1
IA2 = (trA)2 − tr(A2 )

+ + =

:S

A21 A22 A32 A33 A31 A33 2

Lastly, we have
aft

[Au, Av, Aw] = ui vj wk [Aei , Aej , Aek ]


= ui vj wk [Api ep , Aqj eq , Ark er ]
Dr

= ui vj wk Api Aqj Ark [ep , eq , er ]


= ui vj wk Api Aqj Ark pqr
= ijk det(A)ui vj wk
= det(A)[u, v, w]

Thus we have IA3 = det(A)


Properties of trace: Trace is a linear mapping from from CT (2) to R i.e.

tr(A + B) = tr(A) + tr(B)


tr(αA) = αtr(A)

It also satisfies the following properties

tr(AB) = tr(BA)
tr(AT ) = tr(A)
18

Properties of determinant: The determinant of a tensor A ∈ CT (2) satisfies the


following properties

det(αA) = α3 det(A)
det(AB) = det(A)det(B)
det(I) = 1
det(A−1 ) = (det(A))−1
det(AT ) = det(A)

The invariants of IA1 , IA2 , IA3 can be proved by easily by calculating these components

i
ora
in different basis. For example we can prove that

A0ii = Ajj
Gh
where the left hand side is with respect to basis {e01 , e02 , e03 } and the right hand side
is with respect to {e1 , e2 , e3 }.
Definition: A scaler λ is an eigen value of A ∈ CT (2) if there exists a non-zero
vector p which transforms into a vector parallel to inself. Thus
:S

Ap = λp, or (A − λI)p = 0
aft

and p is called an eigen vector. The above is a system of homogeneous linear


equations which for non-zero vector p must satisfy
Dr

det(A − λI) = 0

(Note the above relation can be proved other way also. Take q and r such that
[p, q, r] 6= 0. Using A − λI for A and p, q and r for u, v and w we also get
det(A − λI) = 0.)
Now we can directly carry out the expansion of the determinant or employ the
invariance properties to expand the determinant.

[(A − λI)u, (A − λI)v, (A − λI)w] = det(A − λI)[u, v, w]

Thus we have

0 = [Au − λu, Av − λv, A − λw]


= [Au, Av, Aw] − λ ([Au, Av, w] + [Au, v, Aw] + [u, Av, Aw])
+ λ2 ([Au, v, w] + [u, Av, w] + [u, v, Aw]) − λ3 [u, v, w]
19

Since [u, v, w] is arbitrary we obtain, after using the invariant properties and re-
moving out the factor [u, v, w], the characteristic equation for A as

λ3 − IA1 λ2 + IA2 λ − IA3 = 0

Properties:

1. The eigen values of a positive definite matrix are positive.


Proof: We have Ap = λp. Taking scaler product with p we get

p · Ap = λp · p

i
ora
Since both p · Ap and p · p are positive, we conclude that λ > 0

2. If A is symmetric, then its charateristic equation has three real eigen values.
Proof: Since the characteristic equation is of degree three and of real coeffi-
Gh
cients, one of the eigen value is real and the corresponding eigen vector is real
too. Let n1 corresponds to the eigen values λ1 . Now we choose n2 and n3 such
that {n1 , n2 , n3 } is a right handed triad. If we shift to the new coordinate
system, then we have
:S

 
λ1 0 0
aft

 
[A]n1 ,n2 ,n3 = 
 0 a022 a023 

0 0
0 a23 a33

The eigen values remains same and hence are given by λ = λ1 and the remaing
Dr

eigen values are roots of

λ2 − (a022 + a033 )λ + a022 a033 − (a023 )2 = 0

and are given by


p
a022 + a033 ± (a022 − a033 )2 + 4(a023 )2
λ1 , λ2 =
2
Since the discriminant is non-negative, the eigen values are real.

3. If A is symmetric then the eigen vectors corresponding to two distinct eigen


values are orthogonal.
Proof: Let λ, µ be the eigen values and u, v the corresponding eigen vectors.
Now we have
v · Au − u · Av = (λ − µ)u · v
20

Due to symmetry, the left hand side is zero and thus since λ 6= µ, we have
u · v = 0.

Cayley-Hamilton Theorem: Every tensor A satisfies its own characteristic equations


i.e
A3 − IA1 A2 + IA2 A − IA3 I = O
Spectral Representation: For a symmetric second order tensor A, there is an or-
thonormal basis for V consisting of entirely eigen vectors of A. For any such basis
{n1 , n2 , n3 }, the corresponding eigen values are λ1 , λ2 , λ3 and A can be written as

i
A = λ1 n1 ⊗ n1 + λ2 n2 ⊗ n2 + λ3 n3 ⊗ n3

ora
The orthonormal basis {n1 , n2 , n3 } and corresponding eigen values λ1 , λ2 , λ3 are
called principal directions and principal values respectively.
Proof: We have already proved that if the eigen values λ1 , λ2 , λ3 are distinct, then
Gh
the corresponding eigen values u, v, w are mutually orthogonal. We normalize the
eigen vectors to unit length and then let n1 = u, n2 = v, n3 = w. Thus
 
λ1 0 0
:S

 
[A]n1 ,n2 ,n3 = 
 0 λ 2 0 

0 0 λ3
aft

Now suppose that λ1 6= λ2 = λ3 are the eigen values. Let n1 corresponds to the
eigen values λ1 . Now we choose n2 and n3 such that {n1 , n2 , n3 } is a right handed
triad. If we shift to the new coordinate system, then we have
Dr

 
λ1 0 0
 
[A]n1 ,n2 ,n3 =  0 0 
 0 A 22 A 23 
0 A23 A033
0

Since the tensor has eigen values λ1 , λ2 = λ3 we must have A022 = A033 = λ2 and
A023 = 0. Thus  
λ1 0 0
 
[A]n1 ,n2 ,n3 = 
 0 λ2 0 
0 0 λ3
In this case any vector in the plane of n2 , n3 is an eigen vector.
(Proof: Let u = αn2 + βn3 . Now Au = αAn2 + βAn3 . Now let An2 = γ1 n1 +
γ2 n2 + γ3 n3 . Since n1 · An2 = n2 · An1 = λ1 n1 · n2 = 0 due to symmetry we have
γ1 = 0. Also n3 · An2 = A023 = 0 and hence γ3 = 0 etc. Thus γ1 = λ2 etc.)
21

Finally, let λ1 = λ2 = λ3 = λ = λ. Proceeding as in the case of double root we


have  
λ 0 0
 
[A]n1 ,n2 ,n3 = 
 0 λ 0 

0 0 λ
In this case every vector is an eigen vector and so we choose any set of right hand
triad as the basis {n1 , n2 , n3 }.
Observation:

a. If A is invertible symmetric matrix then we have

i
ora
A = λ1 n1 ⊗ n1 + λ2 n2 ⊗ n2 + λ3 n3 ⊗ n3

Hence
A−1 = λ−1 −1 −1
1 n1 ⊗ n1 + λ2 n2 ⊗ n2 + λ3 n3 ⊗ n3
Gh
b. The principal values of a tensor A contains the maximum and minimum values
that the diagonal elements of any matrix of A can have. Consider any unit
vector n = αn1 + βn2 + γn3 . Now
:S

Ann = n · An = λ1 α2 + λ2 β 2 + λ3 γ 2

Without loss of generality we assume λ1 ≥ λ2 ≥ λ3 . Now since α2 +β 2 +γ 2 = 1,


aft

we have
λ1 = λ1 (α2 + β 2 + γ 2 ) ≥ λ1 α2 + λ2 β 2 + λ3 γ 2 = Tnn
Dr

Similarly

Tnn = λ1 α2 + λ2 β 2 + λ3 γ 2 ≥ λ3 (α2 + β 2 + γ 2 ) = λ3

c. Cayley-Hamilton Theorem (for symmetric tensor A)


Let f (λ) be the characteristic polynomial of A. Let u, v, w be the independent
eigen vectors of A with corresponding eigen values λ1 , λ2 , λ3 . Let x be any
arbitrary vector. Thus
x = αu + βv + γw

Now
f (A)x = αf (λ1 )u + βf (λ2 )v + γf (λ3 )w = ~0

Hence by definition
f (A) = 0
22

Square Root Theorem: Let A ∈ CT (2) be a symmetric positive tensor. Then there
is unique positive definite symmetric tensor B ∈ CT (2) such that B 2 = A.
Proof: (Existence) Since A is symmetric we may write

A = λ1 n1 ⊗ n1 + λ2 n2 ⊗ n2 + λ3 n3 ⊗ n3 ,

where λi are the eigen values of A. Since A is positive definite, λi > 0. Define B
by
√ √ √
B= λ1 n1 ⊗ n1 + λ2 n2 ⊗ n2 + λ3 n3 ⊗ n3 ,

Then

i
√ √ √
BT =

ora
λ1 n1 ⊗ n1 + λ2 n2 ⊗ n2 + λ3 n3 ⊗ n3 = B

Hence B is symmetric. Also


3
! 3
! 3 X
3
X √ X √ X
2
B = λ i ni ⊗ ni
Gh
λj nj ⊗ nj =
p
λi λj (ni ⊗ ni )(nj ⊗ nj )
i=1 j=1 i=1 j=1
3 X
X 3 3 X
X 3
p p
= λi λj (ni · nj )(ni ⊗ nj ) = λi λj δij ni ⊗ nj
i=1 j=1 i=1 j=1
:S

3
X
= λi ni ⊗ ni = A
i=1
aft

(Uniqueness) Suppose C is another symmetric positive definite matrix and

B2 = C 2 = A
Dr

Let u an eigen vector of A with λ > 0 the corresponding eigen value. Then letting

β = λ,
0 = (B 2 − λI)u = (B + βI)(B − βI)u

Let
v = (B − βI)u,

then
Bv = −βv

and v must vanish, for otherwise −β will be an eigen value of B, an impossibility


since B is positive definite and β > 0. Hence

Bu = βu
23

and similarly
Cu = βu

Thus Bu = Cu for every eigen vector v of A. Since we can form a basis of eigen
vectors of A, B and C must concide.
Theorem: A symmetric tesnor A is positive definite iff each of its principal values
is strictly positive.
Proof: Let A is positive definite and λ and u denote a principal value and direction
respectively. Now since u is unit vector we have

λ = (u, Au) > 0

i
ora
due to A being positive deinite.
Next let the principal values are positive. Let n1 , n2 , n3 denote the principal
axes. Now we have
Gh
A = λ1 (n1 ⊗ n1 ) + λ2 (n2 ⊗ n2 ) + λ3 (n3 ⊗ n3 )

Now
:S

Au = (λ1 (n1 ⊗ n1 ) + λ2 (n2 ⊗ n2 ) + λ3 (n3 ⊗ n3 ))u


= λ1 un1 + λ2 un2 + λ3 un3
aft

Now since u = un1 n1 + un2 n2 + un3 n3 , we have


Dr

(u, Au) = λ1 u2n1 + λ2 u2n2 + λ3 u2n3

which is greater than or equal to zero. Let (u, Au) = 0, then uni = 0 and hence
u=0
Polar Decomposition Theorem: An arbitrary invertible tensor A ∈ CT (2) can be
expressed uniquely in the forms

A = RU = V R

where R is an orthogonal tensor and U , V are symmetric positive definite tensors.


Here RU and V R are respectively called the left (right) polar decomposition of A.
Proof: Since AAT and AT A are symmetric positive definite. Hence using square
root theorem there exists positive definite symmetric U , V ∈ CT (2) such that

U 2 = AT A, V 2 = AAT
24

Now let us define R = AU −1 . Now R is orthogonal since

RT R = I

Since det(U ) > 0, both the det(A) and det(R) have the same sign. Since U is
unique, R must be unique.
Proceeding in the same way we find the A = V P where P = V −1 A is orthogonal
and V is unique. Next we show that P = R. We have

A = RU = IA = (P P T )V P = P (P T V P )

i
Now P T V P is symmetric. Let u be an arbitray vector in V. Now

ora
u · P T V P = (P u) · V (P u) > 0

since P u 6= ~0 and V is positive definite. Since RU and P (P T V P ) represnt the


Gh
unique right polar decomposition of A, we have P = R and U = RT V R.
Observation: If U has eigen values λi with eigen vectors pi , then λi > 0 and λi are
also the eigen values of V with eigen vectors q i = Rpi
:S

Proof: λi > 0 is obvious. Now

V (Rpi ) = (V R)pi = (RU )pi = R(U pi ) = λi (Rpi )


aft

Double contraction: This is denoted by doble dots and for second order tensors it
results in a scaler and is defined by
Dr

A : B = tr(AT B)

Note that this operation is commutative. The norm of a tensor A denoted by |A|
is defined by
|A| = (A : A)1/2 = (Aij Aij )1/2 ≥ 0

Properties of double contraction:

a. I : A = tr(A) = A : I

b. A : (BC) = (B T A) : C = (AC T ) : B

c. A : (u ⊗ v) = u · Av = (u ⊗ v) : A

d. (u ⊗ v) : (w ⊗ x) = (u · w)(v · x)
25

4 Vector and Tensor calculas

4.1 Differentiation
In this section we introduce a notion of differentiation sufficiently general to include
scaler, point, vector, or tensor functions whose arguments are scalars, points, vectors
or tensors. Let E denote the point space. A point space can be associated with a
vector space V since the difference of two points is a vector. Also the sum of a point
and a vector is a point. We can associate norm with R, V and L(V), where L(V) is
the set of second order tensors.

i
Let U and W denote such normed vector spaces and consider g : U → W. Let

ora
g be defined in a neighbourhood of zero in U and have values in W. We say that
g(u) approaches zero faster than u, and write
Gh
g(u) = o(u) as u → 0

if
||g(u)||
lim =0
u→0,u6=0 ||u||
:S

Similarly
f (u) = g(u) + o(u)
aft

implies that
f (u) − g(u) = o(u)
Dr

The last definition makes sense if f and g have values in E, in that case f − g has
values in the vector space V.
As an example let φ(t) = tα . Then φ(t) = o(t) iff α > 1.
Let g be a function whose values are scalers, points, vectors, or tensors and
whose domain is an open interval D of R. The derivative ġ(t) of g at t, if it exists,
is defined by
d 1
ġ(t) = g(t) = lim [g(t + α) − g(t)]
dt α→0 α

or equivalently
g(t + α) = g(t) + αġ(t) + o(α)

Clearly αġ(t) is linear in α; thus g(t + α) − g(t) is equal to a term linear in α


plus a term that approaches zero faster than α. In dealing with functions whose
domains lie in spaces of dimension greater than one the most useful definition of
26

derivative is based on this result: we define the derivative to be the linear map
which approximates g(t + α) − g(t) for small α.
Thus let U and W be finite dimensional normed vector spaces and let D be a
domain in U. Consider
g:U →W

We say that g is differentiable at x ∈ D if the difference

g(x + u) − g(x)

is equal to a linear function of u plus a term that approaches zero faster than u.

i
ora
Thus g is differentiable at x if there exits a linear transformation

Dg(x) : U → W (19)

such that
Gh
g(x + u) = g(x) + Dg(x)[u] + o(u), as u → 0

If Dg(x) exists, then it is unique; since



1 d
:S

Dg(x)[u] = lim [g(x + αu) − g(x)] = g(x + αu)
α→0,α∈R α dα α=0

We call Dg(x) the derivative of g at x. If the domain of g is E, then we replace


aft

U in (19) by the vector space V associated with E. Similarly if g have values in E,


then we replace W in (19) by V.
Dr

4.2 Gradient, Divergence and Curl


We now consider functions defined over a domain D of three dimensional eucledian
point space. A function on D is called a scaler, point, vector, or tensor fields
according as its values are scalers, points, vectors or tensors.
Let φ be a smooth scaler field on D. Then for each x ∈ D, Dφ(x) is a linear
mapping of V into R. By the representation theorem for linear forms there exists
a vector w(x) such that Dφ(x)[u] is the inner product of w(x) with u. We write
∇φ(x) for w(x), so that
Dφ(x)[u] = ∇φ(x) · u

We call ∇φ(x) the gradient of φ at x. In this case we have

φ(x + u) = φ(x) + ∇φ(x) · u + o(u)


27

Similarly if v is smooth vector or point field on D, then Dv(x) is a linear trasfor-


mation from V to V and hence a tensor. In this case we use the standard notation
∇v(x) for Dv(x) and write

Dv(x)[u] = ∇v(x)u

The tensor ∇v(x)u is the gradient of v at x and we write

v(x + u) = v(x) + ∇v(x)u + o(u)

Given a smooth vector field v on D, the scaler field

i
ora
div v = tr∇v

is called the divergence of v. Sometimes div v is denoted by ∇ · v.


Gh
Divergence of a smooth tensor field A is the unique vector field div A defined
by
(div A) · a = div (Aa),

for every constant a ∈ V.


:S

The curl of v, denoted by curl v, is the unique vector field with the property
aft

{∇v − (∇v)T }a = (curl v) ∧ a, ∀a ∈ V

Thus curl v is the axial vector corresponding to the skew tensor ∇v − (∇v)T
Example: Let Φ be a smooth scaler, vector, or tensor field. Then
Dr

1
DΦ(x)[ei ] = lim [Φ(x + αei ) − Φ(x)]
α→0 α

If x has components (x1 , x2 , x3 ), then

∂Φ(x)
DΦ(x)[ei ] =
∂xi
From this we deduce
∂φ
∇φ = (∇φ)i ei = ei := φ,i ei
∂xi
(∇u) = ui,j ei ⊗ ej

∇ · u = ui,i

curl u = ijk uk,j ei .


28

and
(div A) = Aij,i ej

Divergence Theorem: Let D ∈ R3 be a bounded regular region. Let φ, v and A


be smooth scaler, vector and tensor fields. Then
Z Z
φn dS = ∇φ dv
Z ∂D ZD

v · n dS = ∇ · v dv
Z ∂D ZD
AT n dS = ∇ · A dv

i
∂D D

ora
Let us prove the third relation. Taking dot product with a constnt vector a we get
Z Z Z Z Z
T T
a· A n dS = a·A n dS == Aa·n dS = ∇·(Aa) dv = a· ∇·A dv
∂D ∂D
Gh ∂D D D

Since a is arbitrary, we have the desired relations.


Stoke’s Theorem: Whereas divergence theorem relates an integral over a close
volume to an integral over its bounding surface, Stoke’s theorem relates an integral
:S

over an open surface to a line integral around the bounding bounding curve of the
surface. Le C be the bounding curve to the open surface S and dx be the tangential
aft

vector to C. If n̂ is the unit outward normal and v is any vector field defined on S
and C, Stoke’s theorem asserts that
Dr

Z Z
n̂ · (∇ ∧ v) dS = v · dx
S C

Alternative Divergence Theorem: We have seen that


Z Z
φn dS = ∇φ dv,
∂D D

Taking component wise we get


Z Z
φnp dS = φ,p dv
∂D D

We take φ = Aijk··· where Aijk··· is the components of an arbitrary tensor. Then we


have the generalized Gauss divergence theorem as
Z Z
Aijk··· np dS = Aijk··· ,p dv
∂D D

You might also like