Search Continuum Mechanics Website

Tensor Notation (Advanced) Search

home > basic math > tensor notation (advanced)

Introduction
This page addresses advanced aspects of tensor notation. A key strength of tensor notation is its ability to represent systems of equations with a single
tensor equation. This makes it possible to recognize relationships among tensor terms, and manipulate them, that would otherwise be nearly impossible to
do using matrix notation.

Kronecker Delta and Derivatives of Axis Variables

The Kronecker Delta is related to the derivatives of coordinate axis variables with respect to themselves. Start with an x,y,z coordinate system and ask,
∂x ∂x ∂x
"What are ∂x , ∂y , and ∂z ?" The answers are simple: 1, 0, and 0. Also,

∂y ∂y ∂y
= 0 = 1 = 0
∂x ∂y ∂z

And similarly, the derivatives of z with respect to the three variables are

∂z ∂z ∂z
= 0 = 0 = 1 Close X
∂x ∂y ∂z

Moving abroad
At this point, the pattern should be obvious. It can allisbe
overwhelming.
summarized in tensor notation as
 ∂xi
= xi,j = δij
∂xj

Example

The relationship between the Kronecker delta and derivatives of coordinates can be seen in the following example. First, recall that a derivative
such as dy/dx represents ynew − yold divided by xnew − xold . Next start at some reference point, say (2, 5, 9). This is the "old" point. Now
change the x coordinate by 1 from 2 to 3, the "new" value. So xnew − xold = 1.

But nothing has changed the y value, so it is still equal to 5. (Note that the coordinates are independent, so y is not required to change just
because x did.) So ynew − yold = 5 − 5 = 0 and dy/dx = 0.

dxi
In contrast, dx/dx = (3 − 2)/(3 − 2) = 1 . Clearly dxj
= xi,j = δij .

Kronecker Delta Multiplication

The Kronecker Delta is nicknamed the substitution operator because of the following simple property of multiplication, best explained by example. Multiplying
vi by δij gives

vi δij = v1 δ1j + v2 δ2j + v3 δ3j

where the i subscript has been automatically summed from 1 to 3 because it appeared twice in the term, once in vi and once in δij .

Now arbitrarily select a value for j, say j = 3 . This gives

vi δi3 = v1 δ13 + v2 δ23 + v3 δ33 = v3

Close X
The result is simply v3 (the value chosen for j) because δ33 = 1 while δ13 = δ23 = 0 . In general, vi δij will always equal vj for whatever value is selected for
j. Therefore, the general rule is
 vi δij = vj

hence the nickname of substitution operator.

The rule applies regardless of the complexity of the term. Recall the vector cross-product term ϵijk ωj rk . Multiply it by δim for example.

δim ϵijk ωj rk = ϵmjk ωj rk

It has the sole effect of substituting m for i in the alternating tensor.

Scalar Equations
It is very important to be able to recognize the rank of any tensor term. Quantities such as vi,i , ai bi , Aij Bij , and σij ϵij are all scalar terms in fact, which also
means they would be part of a single scalar equation. For example, the strain energy density, W , of a linear elastic material is a scalar quantity given by

1
W = σij ϵij
2

Although there are several subscripts in the equation, they all expand out (all sum from 1 to 3) because each letter occurs exactly twice. This is amplified by
the fact that on the left hand side (LHS) of the equation, W is undeniably a scalar because it has no subscript(s) at all.

Vector Equations
A simple vector equation is the cross product of an angular rotation vector, ωj , with a position vector, rk , to obtain the velocity vector, vi . (Yes, this is just
v = ω × r.)

vi = ϵijk ωj rk

Close X

This very small tensor equation represents so much. For starters, it represents three equations, not one, because the vi term shows clearly that there is an
equation for each i = 1, 2, and 3. Second, each of the three equations contains nine terms because j and k both sum to three. However, we've already
seen here that ϵijk is zero in most terms, leaving
 And you could

v1 = ω2 r 3 − ω3 r 2 ge t a c
r

Ac c

v2 = ω3 r 1 − ω1 r 3

v3 = ω1 r 2 − ω2 r 1

That is a lot packed into one small tensor equation. Of course, v = ω × r actually represents the same
equations as well, except that the subscripts in the tensor equation explicitly lead one to the three component
equations.

It is important to recognize that tensor notation often provides the freedom to write any tensor term in many forms. In this case, alternative examples include

vi = ωj ϵijk rk vi = rk ϵijk ωj vi = rk ωj ϵijk vk = ϵijk ωi rj

All are equivalent, and technically correct, because the multiplication details are dictated by the subscripts, not the order of the factors. However, it is
customary for readability considerations to write ϵijk first in a term, and write the remaining factors in approximately the same order they would appear in
vector or matrix notation. This leads to the original form of vi = ϵijk ωj rk as being preferred.

It is also important to recognize forms that are incorrect, and why. One incorrect form is vm = ϵijk ωj rk . This is wrong because the free indices on the two
sides of the equation are different: m on the LHS and i on the RHS. While one might guess the author's intent in such a case, such a mismatch of free
indices should be avoided.

A similar bad form is the addition of ai + bj . This is simply meaningless.

A second incorrect form is vi = ϵijj ωj rj . This is wrong because j appears more than twice in the RHS term.

2nd Order Tensor Equations

A very common 2nd order tensor equation is Hooke's Law, relating stress to strain in linear elastic materials. It is written in matrix notation as
Close X

1
ϵ = [(1 + ν)σ − ν I tr(σ)]
E
where: σ is the stress tensor,
ϵ is the strain tensor,
E is the elastic modulus,
ν is Poisson's ratio,
I is the identity matrix,
tr(...) is the trace of a tensor, in this case, σ11 + σ22 + σ33

Hooke's Law is written in tensor notation as

1
ϵij = [(1 + ν)σij − ν δij σkk ]
E

This shows clearly that each term is 2nd rank because of the i and j indices. Note that σkk is just a scalar quantity because k is summed from 1 to 3 since it
appears twice. So σkk is the trace of σ. Another key insight is that the entire ν δij σkk term is nonzero only when i = j. This corresponds to the normal
stresses and strains and is the classic Poisson's Effect. On the other hand, the entire term is zero when i ≠ j because of the presence of δij .

Writing all the matrices out gives

ϵ11 ϵ12 ϵ13 ⎧

⎪ σ11 σ12 σ13 1 0 0 ⎫
⎡ ⎤ ⎡ ⎤ ⎡ ⎤⎪
1
⎢ ϵ21 ϵ22 ϵ23 ⎥ = ⎨(1 + ν) ⎢ σ21 σ22 σ23 ⎥ − ν (σ11 + σ22 + σ33 ) ⎢ 0 1 0 ⎥⎬
E ⎩
⎪ ⎭
⎣ ⎦ ⎣ ⎦ ⎣ ⎦⎪
ϵ31 ϵ32 ϵ33 σ31 σ32 σ33 0 0 1

So what are the individual component equations represented by the tensor notation equation, and how many are there? Since there are two free indices,
there would normally be 3 ∗ 3 = 9 equations. But it turns out that all terms are symmetric in Hooke's Law, so there are only six independent equations. They
are initially

1 1
ϵ11 = [(1 + ν)σ11 − ν (σ11 + σ22 + σ33 )] ϵ12 = [(1 + ν)σ12 ] Close X
E E

1 1
ϵ22 = [(1 + ν)σ22 − ν (σ11 + σ22 + σ33 )] ϵ13 = [(1 + ν)σ13 ]
E E
 1 1
ϵ33 = [(1 + ν)σ33 − ν (σ11 + σ22 + σ33 )] ϵ23 = [(1 + ν)σ23 ]
E E

but then clean up into more concise forms

1 (1 + ν)
ϵ11 = [σ11 − ν (σ22 + σ33 )] ϵ12 = σ12
E E

1 (1 + ν)
ϵ22 = [σ22 − ν (σ11 + σ33 )] ϵ13 = σ13
E E

1 (1 + ν)
ϵ33 = [σ33 − ν (σ11 + σ22 )] ϵ23 = σ23
E E

Again, this is an amazing amount of information packed into a single tensor equation.

! For anyone diving directly into this page here, note that ϵ12 , ϵ13 , and ϵ23 are one-half of the normal shear quantities, i.e., ϵ12 = γ12 /2 .

Tensor Notation Algebra - Inverting Hooke's Law

So far, tensor notation has not actually provided any capabilities beyond matrix notation. After all, the matrix form of Hooke's Law does contain all
the same information that is available in the tensor equation. However, the great power of tensor notation over matrix notation becomes evident
when one starts to manipulate tensor equations. The example here will demonstrate how to invert Hooke's Law from strain-as-a-function-of-stress
to stress-as-a-function-of-strain. Such an inversion is all but impossible using matrix notation.

Returning to the fundamental equation...

1
ϵij = [(1 + ν)σij − ν δij σkk ]
Close X
E

It is easy to solve for σij to get

 1
σij = [E ϵij + ν δij σkk ]
(1 + ν)

But there is a problem, a major problem. There is still a stress term, σkk , on the RHS that must be resolved. It cannot be combined with σij on the
LHS because they are different animals. σij is an individual component of the stress tensor, σkk is the trace of the stress tensor,
(σ11 + σ22 + σ33 ).

This hurdle is overcome in a couple steps by first multiplying the equation through (both sides of course) by δij as follows.

1
δij σij = [E δij ϵij + ν δij δij σkk ]
(1 + ν)

At this point, several advanced properties of tensor notation kick in. First, note that the entire equation has been transformed from a 2nd rank
tensor equation to a scalar equation because there are no longer any free indices. Each term now has i and j occurring twice in it, so both are
automatically summed from 1 to 3.

Second, the terms containing δij σij and δij ϵij can be simplified by recalling the substitution property of the Kronecker Delta. Since δij = 1 only
when i = j, then

δij σij = σii = σjj = σ11 + σ22 + σ33

and

δij ϵij = ϵii = ϵjj = ϵ11 + ϵ22 + ϵ33

So choose, arbitrarily, to write each term with j subscripts as σjj and ϵjj .

The final interesting term contains δij δij . This equals 3 because

δij δij = δii = δ11 + δ22 + δ33 = 1 + 1 + 1 = 3

Close X

So the now-scalar equation reduces to

 1
σjj = [E ϵjj + 3 ν σkk ]
(1 + ν)

It's now time for another round of insight into tensor notation. Note that σjj and σkk are in fact equal because both expand to σ11 + σ22 + σ33 . So
they can be combined. The σjj is simply rewritten as σkk , producing

1
σkk = [E ϵjj + 3 ν σkk ]
(1 + ν)

Now the σkk terms on the LHS and RHS can be combined to obtain

E ϵjj
σkk =
(1 − 2ν)

which is the simple clean equation we need to insert back into

1
σij = [E ϵij + ν δij σkk ]
(1 + ν)

to obtain

E ν
σij = [ϵij + δij ϵjj ]
(1 + ν) (1 − 2ν)

Except this is incorrect because j now appears three times in the last term. But the solution is simple. Simply rewrite ϵjj as ϵkk . This is completely
acceptable because both forms expand out to give the same trace of the strain tensor.
Close X

E ν
σij = [ϵij + δij ϵkk ]
(1 + ν) (1 − 2ν)
 And we are finally done.

So what happened? The example is typical of equations that contain tensors occurring in two different forms. In this case, the two different forms were σij
and σkk . The first is a component of the stress tensor while the second is the trace of the same stress tensor. (Both involve the same stress tensor, but they
are in fact different quantities.) When this occurs, the usual procedure is to multiply the equation through by δij .

The second key aspect of the example was the demonstrated freedom to change indices when they occur twice. This is because any indices occurring twice
are automatically expanded out so it does not matter what letter they are. This was demonstrated by changing σjj to σkk and ϵkk to ϵjj during the example.

Epsilon-Delta Identity
The alternating tensor and the Kronecker delta are related to each other through the following identity

∣ δil δim δin ∣

∣ ∣
ϵijk ϵlmn = ∣ δjl δjm δjn ∣
∣ ∣
∣ δkl δkm δkn ∣

= δil (δjm δkn − δjn δkm ) + δim (δjn δkl − δjl δkn ) + δin (δjl δkm − δjm δkl )

which is too complex to be of much use to anyone! However, multiplying through by δil reduces the above equation to something much more manageable
and useful. (Proving this yourself is an excellent homework problem.)

ϵijk ϵimn = δjm δkn − δjn δkm

This is a 4th rank tensor equation because there are four free indices, j, k, m, and n. The i index is repeated twice on the LHS, so it is summed from 1 to 3.
So the equation could be expanded to give

Close X
ϵijk ϵimn = ϵ1jk ϵ1mn + ϵ2jk ϵ2mn + ϵ3jk ϵ3mn = δjm δkn − δjn δkm

The terms can be evaluated once values are chosen for each of the four free indices.
The identity is used when two alternating tensors are present in a term, which usually arises when the term involves cross products. The benefit of employing
it is that once the epsilons are transformed into the deltas, then the substitution property of the Kronecker Deltas can be used to simplify the equation. The
following example demonstrates the usefulness of this identity.

Epsilon-Delta Identity Example

Recall from this discussion of cross products that the area of a triangle is given by

1
Area = | a × b|
2

which can be written in tensor notation as

1
Area = √ϵijk aj bk ϵimn am bn
2

At this point, the epsilon-delta identity can be employed to obtain

1
Area = √(δjm δkn − δjn δkm )aj bk am bn
2

Multiplying through gives

1
Area = √δjm δkn aj bk am bn − δjn δkm aj bk am bn
2

And the substitution property of the delta operator can now be exploited to obtain

Close X
1
Area = √ a m a m bn bn − bm a m a n bn
2
 But am am is simply the dot product of the a vector with itself, as is bn bn . And am bm and an bn are both the same dot product of a with b.

1
2
Area = √(a ⋅ a)(b ⋅ b) − (a ⋅ b)
2

Which offers an alternative calculation to the cross product approach. For what it's worth, the equation can also be written as a determinant.

1 ∣ (a ⋅ a) (a ⋅ b) ∣
Area = √∣ ∣
2 ∣ (a ⋅ b) (b ⋅ b) ∣

And here is a second example...

Second Epsilon-Delta Identity Example

Prove the following identity, where v is a vector.

2
∇ × (∇ × v) = ∇(∇ ⋅ v) − ∇ v

(∇ × v) is ϵijk vk,j and ∇ × (∇ × v) is ϵmni ϵijk vk,jn . Apply the identity to transform the equation to

ϵmni ϵijk vk,jn = (δmj δnk − δmk δjn )vk,jn

Multiply through to get

Close X
ϵmni ϵijk vk,jn = δmj δnk vk,jn − δmk δjn vk,jn

As before, apply the substitution properties to obtain

 ϵmni ϵijk vk,jn = vk,mk − vm,nn

Recognize that vk,mk is the gradient of the divergence of v. This can be made clearer if the term is written more explicitly as (vk,k ),j . So
vk,jk = ∇(∇ ⋅ v) .

The second term is just the Laplacian of v, i.e., ∇2 v. This proves

2
∇ × (∇ × v) = ∇(∇ ⋅ v) − ∇ v

Close X
10 Comments 
1 Login

Join the discussion…

LOG IN WITH OR SIGN UP WITH DISQUS ?

Name

 1 Share Best Newest Oldest

Armadillo − ⚑
5 years ago

Following the substitution operator rule examples given in the Section “Kronecker Delta Multiplication”, and then applying the rule to HW2, problem 2, to the
LHS of the equation,

\delta_{il} \epsilon_{ijk} \epsilon_{lmn}

Would we not swap the subscript {i} in the term \epsilon_{ijk} with the subscript {l} given by the \delta_{il} term?

0 0 Reply Share ›

Thompson Lemuel > Armadillo

− ⚑
4 years ago

yeah i think it will still give you the same result whether you substitute {i} for {l} or substitute {l} for {i}. in the end what matters is that the kronecker
delta resolves to 1 by setting {i}={l}. Whether you end up with \delta_{ii}\ or \delta _{ll}\ they are both equal to one which means that you simply
replace which ever subscript of the delta function you swaped for the other with same in the alternating function. So with \delta_{ii}\ the alternation
function becomes \epsilon_{ikj} epsilon_{imn}\. with \delta_{ll}\ you get \epsilon_{ljk} epsilon_{lmn}\. Both are the same since the repeated {i} or {l} in
the epsilon terms simply implies automatic addition from 1 to 3 and hence can be any subscript at all provided they are the same.

0 0 Reply Share › Close X

Armadillo − ⚑
4 years ago
 0 0 Reply Share ›

Phạm Thiên − ⚑
3 years ago

I couldn't get this step, please help ⛺ View — uploads.disquscdn.com

0 0 Reply Share ›

BobMcG Moderator > Phạm Thiên

− ⚑
3 years ago

Start for example with the term: delta_mj delta_nk v_k,jn. The delta_mj means that one can change either an "m" to a "j" or a "j' to an "m". There is no
"m" to change, but there is a "j" in the v_k,jn term. So change the "j' to "m" to obtain v_k,mn. Likewise, the delta_nk term means that the "n" in v_k,mn
can be changed to a "k". So that gives v_k,mk and this is the resulting term. See the section near the top of the page titled "Kronecker Delta
Multiplication" for details.

0 0 Reply Share ›

Someone − ⚑
a year ago

If v_k,jk represents differentiation with respect to k first and then j as pointed out in the proof of the last example, how come the curl of the curl of v involves
v_k,jn? Shouldn't it be v_k,nj since v_k is differentiated with respect to j first?

0 0 Reply Share ›

BobMcG Moderator > Someone − ⚑

a year ago

It's all the same because derivatives can be taken in any order. So v_k,jn is the same as v_k,nj.
Recall how (d^2 f / dx dy) gives the same result as (d^2 f / dy dx).

1 0 Reply Share ›

glynnec2008 . − ⚑
8 months ago edited

Close
Hooke's Law can be inverted using pure matrix notation. Simply calculate the trace of both sides of the fundamental equation X solve the resulting linear
and
scalar equation for Trace(stress) in terms of Trace(strain). Substitute this back into the fundamental equation and isolate the stress on the LHS with
everything on the RHS written in terms of the strain.

0 0 Reply Share ›
 Diego Andres Alvarez Marin > glynnec2008 .
− ⚑
7 months ago

Usually, I do not perform algebra calculations by hand. I almost always use either Maxima or Python (Sympy) to do the dirty job. Shorter, error proof
and intuitive.

0 0 Reply Share ›

Diego Andres Alvarez Marin − ⚑

7 months ago

I have studied a lot of books using tensor notation. And I really hate it. Some reasons:

1) Tensor notation requires meticulous attention to subindices A small mistake can lead to incorrect calculations This complexity can be frustrating

Thank You Math Chapter - $4.99 Entire Website - $12.95

Thank you for visiting this webpage. Feel For $4.99, you receive two formatted PDFs For $12.95, you receive two formatted PDFs
free to email me if you have questions. (the first for 8.5" x 11" pages, the second for (the first for 8.5" x 11" pages, the second for
tablets) of the complete math chapter. tablets) of the entire website.
Also, please consider visiting an advertiser
above. Doing so helps to cover website Click here to see a sample page in each of Click here to see a sample page in each of
the two formats. the two formats.
hosting fees.

Email address to receive PDFs Email address to receive PDFs

Bob McGinty
[email protected]

Copyright © 2012 by Bob McGinty

Close X

Table of Contents
Tensor Notation (Basics) Coordinate Transformations
Close X