Einstein Summation, Levi-Civita, and Dirac Deltas

Joseph McGowan
Sep 17, 2021

1 Einstein Summation and the Levi-Civita Symbol

The study of electricity and magnetism involves a number of vector quantities, not least the electric and
magnetic fields. Before we jump into the main body of the class, we’re going to add a couple of important
tools to our toolbox for doing calculations with vectors. The first of these is the idea of dealing with the
components of vectors rather than the vectors themselves, and the second is the Levi-Civita symbol, which
has many uses but here will be used to simplify cross products.
We are all familiar with usual 3d vectors, for example a = (a1 , a2 , a3 ). Here, I have written the vector
using vector notation as well as writing it as a list of components. I could also write it as a sum of its
components and the associated unit vectors, a = a1 ex + a2 ey + a3 ez . However, I can also represent this
vector simply by writing ai , where i = 1, 2, 3 is understood. Of course, this requires knowing beforehand
how many dimensions are in the space in which a lives, but for our purposes this will always be 3. Naturally,
you could also represent higher-dimensional objects this way: for example, a matrix could be written as Aij ,
where both i and j can be 1,2,3. There are 9 possible combinations of i and j, giving the 9 entries of the
3x3 matrix A.
What is the reason for doing this? Well, for one, it is notationally compact. It requires fewer symbols
than writing out the unit vectors, and I don’t have to type “\vb” or “\vec” every time. But the real reason
is that, for a given i and j, ai and Aij are scalars (i.e. numbers), and we can deal with them in all the
usual numbery ways. We can multiply them together commutatively and distributively, usual derivative
rules apply, and so forth. This is in contrast with vector operations, which are not always commutative (the
cross product, remember, is anticommutative!).
There are two things that we use to help this translation between the world of vectors and the world of
scalars. For one, sometimes we do operations with all components of a vector. Here we adopt what is called
the Einstein summation convention. It says this: any time you see a repeated index, you sum over it. That
is, if we see for example ai bi for two vectors a and b, we should understand that to mean
3
X
ai bi = a1 b1 + a2 b2 + a3 b3 .
i=1

This should look familiar; in fact,

ai bi = a · b .
It’s good to note that since i is being summed over (albeit implicitly), it is effectively a dummy variable. We
could rename i to j or λ or anything else and it would change nothing.
So representing dot products this way is simple. For cross products, we need our second tool, an object
known as the Levi-Civita symbol. The Levi-Civita symbol is usually written εijk , and it is always equal to
either 0, 1, or -1 depending on the values of i, j, and k, as follows:

1
 1. If any of i, j, and k are equal, εijk = 0. For example, ε112 = 0.
2. If the indices go in ascending order or are cycled from ascending order, εijk = 1. For example,
ε123 = ε312 = 1.
3. If the indices go in descending order or are cycled from descending order, εijk = −1. For example,
ε321 = ε213 = −1.
It’s rare that you will ever need to figure out what a given symbol is equal to. The key fact that is useful
when using the Levi-Civita symbol is that swapping two adjacent indices always gives you a minus sign.
That is, εijk = −εikj = εkij etc. This kind of manipulation, along with cycling the indices (which does not
change the symbol), is what will be useful.
With the Levi-Civita symbol in hand, we can now write down the cross product in component form. The
cross product of two vectors a and b is given by

a × b = εijk ei aj bk .

Equivalently, the ith component of the cross product can be written

[a × b]i = εijk aj bk .

There are a few things to notice here. First, I want to point out that the first equation contains the unit
vectors ei . This is necessary because the left hand side, the cross product, is a vector, and so the right hand
side must also be a vector. Of course, since the indices i, j, and k are all repeated, all three get summed
over, so it is really a sum of vectors (which is obviously also a vector). In contrast, the lower equation is an
equality of two scalars: a single component of a vector on the left, and a product of three numbers on the
right. Another thing to note is that in the upper equation, there are no free indices on either side. On the
right, all three are repeated, therefore summed over, therefore dummy indices. In the lower, there is a free
i on both sides. These can be good sanity checks when performing these kinds of calculations. Finally, it is
worth stressing that the important thing is the order in which the ijk and the various components appear.
If we swap the order of the indices on ε, for instance swapping the last two to get εikj , we would get an extra
minus sign. We can swap the order of aj and bk freely (they are scalars so they commute no problem), and
we would get that [a × b]i = −εikj bk aj = −[b × a]i . And this checks out, since a × b = −b × a.
Is this equality really true? You could just take my word for it, but it’s not too hard to check. Pick
your favourite method of computing, say, the x-component of a cross product and compare it to the lower
equation. If i = 1, then the other two indices must be different, and the x-component indeed only depends
on the y- and z-components of the two vectors. The only two nonzero Levi-Civita symbols with i = 1 are
ε123 and ε132 , and there is a minus sign between them. Indeed, each component of the cross product has
one thing subtracted from another. You can verify the rest of the details if you like, but I hope it is at least
somewhat intuitive from here.
Now, armed with these weapons, let’s try them out on a few exercises.
1. Show that δij δij = 3 where δij is the Kronecker delta, which is 1 if i = j and 0 otherwise.
Solution: Both indices here are being summed over. Let’s start by doing the sum over j. For a given i,
δij is only nonzero when j = i. So these are the only terms in the sum that survive. Therefore, by doing
the sum over j, we can replace all the js with is and remove one of the δs. This is a common trick.
After that, we are left with just δii and an implied sum over i. There are three terms in the sum and
they are each equal to 1, so of course we get that δij δij = δii = 3.
2. Show that ∇ · (A × B) = (∇ × A) · B − A · (∇ × B)
Solution: This is like a version of the product rule but for vector derivatives. You could explicitly
evaluate each product and show the two sides are equal for each component, but we will exploit the
techniques we learned above to do this in just a few lines.

2
 ∇ · (A × B) = ∂i [A × B]i
= ∂i (εijk Aj Bk )

ε is just a number, so it can be pulled out of the derivative. Aj and Bk are just scalars (in general, they
are functions of x, y, z) and ∂ just a normal partial derivative, so we can just use the normal product
rule.

∇ · (A × B) = εijk (Bk ∂i Aj + Aj ∂i Bk )
= εijk Bk ∂i Aj + εijk Aj ∂i Bk
= εkij Bk ∂i Aj − εjik Aj ∂i Bk

In the last line, I made use of the two kinds of manipulation with Levi-Civitas that I mentioned earlier.
Now, we just need to collapse these two terms back into cross products. Remember: the indices that
are summed over are dummy indices, so we can rename them whatever we want to pattern-match with
the boxed line above. All that is important is that the order of the ε indices matches the order of the
indices on the components of the cross product.

∇ · (A × B) = Bk [∇ × A]k − Aj [∇ × B]j
= (∇ × A) · B − A · (∇ × B)

3. Show that ∇ × (∇ × A) = ∇(∇ · A) − ∇2 A.

Solution: We’ll start from the outside in. We have a cross product on the left, so let’s expand it as
before.

[∇ × (∇ × A)]i = εijk ∂j [∇ × A]k

= εijk ∂j (εk`m ∂` Am )
= εkij εk`m ∂j ∂` Am

Here we’ll use a handy identity for the product of two εs. Even if you didn’t know such an identity
existed, you might suspect something would be happening with them, since we need to end up with two
terms somehow, and look up a list of identities to see if one would help. That’s what I did! The identity
looks like this:
εijk εi`m = δj` δkm − δjm δk` .
Pattern-matching appropriately, we can continue, remembering the trick with Kronecker deltas that I
mentioned in the first exercise.

[∇ × (∇ × A)]i = (δi` δjm − δim δj` )∂j ∂` Am

= δi` δjm ∂j ∂` Am − δim δj` ∂j ∂` Am
= δjm ∂j ∂i Am − δj` ∂j ∂` Ai
= ∂j ∂i Aj − ∂j ∂j Ai
= ∂i ∂j Aj − ∂j ∂j Ai
= ∂i (∇ · A) − [∇2 A]i
= [∇(∇ · A)]i − [∇2 A]i
= [∇(∇ · A) − ∇2 A]i

Since the two sides are equal for each component i, the expressions inside the brackets are equal.

3
 4. Write down an expression using the ESC for the trace of a matrix Aij .
Solution: The trace of a matrix is the sum of its diagonal elements. For a 3x3 matrix, for example,

Tr(Aij ) = A11 + A22 + A33 .

By comparing this to the expanded sum form of the dot product we saw earlier, it should be simple to
see that
Tr(Aij ) = Aii .

2 Dirac Deltas
The Dirac delta function δ(x − a) is like the continuous version of the Kronecker delta. It is a “function”
which is equal to 0 everywhere except at a, where it is infinite. When I took PDEs, my prof insisted on using
such quotation marks, since a function whose value is infinite is not a mathematically rigorous statement.
In this section I will make a few statements which are not proofs in the strictest sense but which handwave
some of the mathematical details. But, after all, I am a physicist and not a mathematician, and this is a
physics class. I’ll therefore entrust the finicky details to the mathematicians, because the rough version will
suffice for our purposes. For more rigorous details, we would need to discuss things called distributions, also
known as generalized functions.
The defining property of the delta “function” is how it behaves inside an integral; for a function f (x)
which is defined at a,
Z
dx δ(x − a)f (x) = f (a) ,

which is hopefully somewhat familiar. (If I don’t write the bounds of integration, they should be assumed
to be from −∞ to +∞.) This, by the way, is also connected to how one can make it more mathematically
rigorous. The important takeaway here is this: whenever you see a Dirac delta, your mind should be
screaming out “I should probably be integrating this somehow.” Much like the Kronecker delta above was
used to remove a sum and replace one index with another, Dirac deltas inside multiple integrals can also be
used to remove one integral and replace one integration variable with another.
R
1. Show that the delta “function” has unit area dx δ(x − a) = 1.
Solution: This is simple to see. δ(x − a) = 1 · δ(x − a). For f (x) = 1, f (a) = 1 no matter what a is.
dH(x)
2. Show δ(x) = dx , where H(x) is the Heaviside step function.
Solution: Well, the Heaviside step function is discontinuous. It is flat everywhere but at 0, so its
derivative should be 0, but we can’t define the derivative at 0 in the usual sense. So, as I said, we really
need to make sense of this equality inside an integral. We’ll use integration by parts, which is a pretty
common trick when dealing with Dirac deltas.
∞
Z Z
dH(x)
dx f (x) = f (x)H(x) − dx f 0 (x)H(x)
dx −∞
Z ∞
= f (∞) − dx f 0 (x)
0
= f (∞) − (f (∞) − f (0))
= f (0)

In the second line, we used that H(x) is 1 for positive x and 0 for negative x. In the third line, we used
the fundamental theorem of calculus. Of course, it is a bit odd to evaluate functions at infinity, so if it
makes you more comfortable you could instead integrate from −M to M for some bound M > 0, and
then take the limit M → ∞ in the last line.

4
 So we see that, inside an integral, the derivative of the Heaviside function does the same thing that the
Dirac delta does, i.e. pick out the value of f at 0. It is in this sense that they are equal.
1
3. Show that δ(ax) = |a| δ(x)
Solution: Again, we need to do this inside an integral. We’ll start with a substitution u = ax so that
du = a dx:
Z Z
1
dx f (x)δ(ax) = du f (u/a)δ(u)
a
1
= f (0/a) = f (0)/a.
a
(Of course we assume that a 6= 0.) So where does the absolute value come in? Well, this is all fine as
long as a > 0. However, if a < 0, we run into a problem with the bounds of the integral. If a > 0, then
when x = ∞, u = ∞. But if a < 0, when x = ∞, u = −∞ and vice versa. We would need to pick up
an extra minus sign swapping the bounds of the integral in order to evaluate it using the delta, and we
would get Z
dx f (x)δ(ax) = −f (0)/a.

However, if a < 0, then a = −|a|, which cancels with the minus sign from switching the bounds of
integration. Thus, both cases are captured by writing δ(ax) = δ(x)/|a|.
1
4. Show that δ(g(x)) = Σi |g0 (x i )|
δ(x − xi )
Solution: First, it should be noted that this only is true if all the xi are distinct. Otherwise, this
proceeds basically just as the above. The integral can be split into one integral near each zero, with
bounds in a small interval around the zero, giving the sum. For each integral, Taylor expand f (x) to
first order and use the previous result.
5. Show that dx f (x)δ 0 (x − a) = −f 0 (a).
R

Solution: This one is almost exactly like the Heaviside function example.
∞
Z Z
0
dx f (x)δ (x − a) = f (x)δ(x − a) − dx f 0 (x)δ(x − a)
−∞
0
= 0 − 0 − f (a)
= −f 0 (a)

You can also use this technique multiple times to show that
Z
dx f (x)δ (n) (x − a) = (−1)n f (n) (a).

5
 2 2
6. For the Gaussian Ga (x) = 1
√
a π
e−x /a , show that Ga→0 (x) = δ(x).
Solution: For this one, we’ll do it in a bit of a rough way. First, note that the integral is constant for
any a:
Z Z
1 2 2
dx Ga (x) = √ dx e−x /a
a π
1 √
= √ a π
a π
=1
√
Next, note that Ga (0) = 1/a π. This of course diverges to +∞ as a → 0. Last, note that for x 6= 0, we
have the following:
1
lim Ga (x) = lim √ x2 /a2
a→0 a→0 a πe

Let b = x/a
b
lim Ga (x) = lim √ b2
a→0 b→∞ x πe
1
= lim b2
b→∞ be
=0

We used L’Hôpital’s rule to evaluate the limit, remembering that x is a constant for the purposes of the
limit.
So since the Gaussian has constant area, and it goes to 0 everywhere but x = 0 where it diverges, we can
argue that it goes to a Dirac delta in the limit. For a more precise argument, we would need to perform
this limit inside an integral using an ε − δ style argument.
7. Show that, in spherical coordinates, ∇2 1/r = −4πδ(r).
Solution: Again, we’ll take a less-than-mathematically-rigorous tack here. First, note that everywhere
but at r = 0, the gradient of 1/r is given by
   
1 ∂ 1 r̂ r
∇ = = − 2 = − 3.
r ∂r r r r
The other components of the gradient only contain derivatives with respect to other variables, which
of course vanish. (I highly recommend having the wikipedia page “Del in cylindrical and spherical
coordinates" bookmarked for these occasions! It saved my life countless times in undergrad.)
Now, let’s get this bad boy into an integral. We’ll integrate over a sphere of radius a around 0, which
I’ll denote by B(0, a).
ZZZ   ZZZ   
2 1 1
∇ dV = ∇· ∇ dV
B(0,a) r B(0,a) r

Here, we will ignore the fact that the gradient is not well-defined at 0, and we will use the divergence
theorem. In order to make this work, we shall wave our hands vigorously as we do so.
ZZZ   ZZZ  
1 1
∇2 dV = ∇ · dS
B(0,a) r B(0,a) r
ZZ
r
= − 3 · dS
B(0,a) r

6
For a sphere centered at 0, the r̂ component of the differential area element dS is equal to r2 sin θ dθ dϕ
(also from the same wiki page!). We can proceed:
ZZZ   ZZ
2 1 r
∇ dV = − 3 · dS
B(0,a) r B(0,a) r
Z 2π Z π
r
= − 3 · r̂r2 sin θ dθ dϕ
0 0 r
Z 2π Z π
=− sin θ dθ dϕ
0 0
= −4π

Since this result does not depend on the radius of the ball, we conclude that integrating around any ball
containing the origin will give us the result of −4π, just as we would expect −4πδ(r) to do. Again, this
is a bit lean on rigor, and if you wanted to do it properly you should integrate with a function, use the
divergence theorem where it is valid, and use a different technique in the region containing the origin.
But this is a sufficient demonstration of the idea.