See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/337831066

An Invitation to Graphical Tensor Methods: Exercises in Graphical Vector and

Tensor Calculus and More

Preprint · January 2020

CITATIONS READS

0 2,543

3 authors:

Joon-Hwi Kim Maverick Sang-Hyun Oh

Seoul National University University of California, Merced
5 PUBLICATIONS   8 CITATIONS    6 PUBLICATIONS   10 CITATIONS   

SEE PROFILE SEE PROFILE

Keun-Young Kim
Gwangju Institute of Science and Technology
114 PUBLICATIONS   2,444 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Graphical Tensor Calculus View project

Yi-Sang Poems View project

All content following this page was uploaded by Joon-Hwi Kim on 04 June 2020.

The user has requested enhancement of the downloaded file.

 An Invitation to Graphical Tensor Methods
Exercises in Graphical Vector and Tensor Calculus and More

Joon-Hwi Kim∗
Department of Physics and Astronomy, Seoul National University, Seoul, 08826, Republic of Korea

Maverick S. H. Oh† and Keun-Young Kim‡

Department of Physics and Photon Science, Gwangju Institute of Science and Technology, Gwangju, 61005, Republic of Korea
(Dated: May 23, 2020)
This is the supplementary material for “Boosting Vector Calculus with the Graphical Notation.”
It is written for students and introduces instructive and practical examples regarding graphical
vector and tensor algebra and calculus in the physics context. Notable topics, including the ones
mentioned in the main article, are as follows. Graphical proof of the Jacobi identity is discussed
in P6a and its sequels in the following two sections. Next, most importantly, graphical proof of
first and second derivative vector calculus identities are covered in Sections I B 1 and I B 2. Then,
the strength of the graphical technique of tensor calculus is elaborated throughout Section I C.
Graphical notation for general multi-index quantities is introduced in Section I C 2. Do not miss
the next section, Section I C 3, to learn the “arrow pushing” that enables to efficiently denote and
explain the invariance property of tensorial expressions. Finding the oblivious δ (3) (~r) term and
the calculus of multipolar fields are discussed throughout Sections I C 5 and I C 6. Section I A 4
introduces permutation symmetry of indices. Refer to the comments in A67b for an instructive
example of diagrammatic perturbation in statistical mechanics. Lastly, Section I D invites the
reader into the world of Feynman diagrams.

Difficulty Levels P4 Use the economy of the graphical notation to show that
~ × B)
(A ~ · (C~ × D)
~ = A ~ · (B
~ × (C
~ × D)):
~ that is, show
Without mark: essential or elementary.
With [?] : intermediate level, adequate for advanced mathematical that the both sides of the equation are just two different
physics courses. readings of an identical diagram.
With [ ! ] : challenge only if you find yourself relishing graphical P5 Graphically represent the following plaintext equations.
reasoning or you want to be a true “virtuoso.”
~ × B)
(a) (A ~ × (C ~ × D)
~ = (D ~ × C)
~ × (A ~ × B)
~
I. Problems (b) ikj jli Ak Bl = −ijl ijk Al Bk
(c) ijk Ai Bj Ck = ijk Aj Bk Ci
A. Graphical Vector Algebra P6 Translate the following graphical equations into the plain-
text notation. (You don’t need to prove them.)
1. Basic Translation Tasks

A possible calculation strategy is to translate the plaintext (a) + + =0

equations into graphical notation, proceed by graphical manip- A B C B C A C A B
ulations, then return to the plaintext notation if the answer is
required to be written in the plaintext notation. If you become
a true “bilingual,” translating one to the other will not bother (b) = −
you anymore. The problems in this section is for practicing such A B B A A B
translation tasks.
B
P1 Draw corresponding diagrams for each of following plain-
A B C B
text expressions, and then give at least one alternative A
A B A
reading in index-free plaintext notation.
~ · (B
(a) A ~ × C)
~ ~ × (B
(b) A ~ × C)
~ (a) (b) (c) (d)
~ × (B
(c) (A ~ × C))
~ ·D~
A D A D A D
P2 Draw corresponding diagrams for each of following plain-
text expressions, and then give at least one alternative B C C B C
reading in the index notation. E E
(a) Ai Bl Cm ijk klm (b) Am Bk δij δml lkj (e) i. ii. iii.
(c) Ai Bj Ck imj lkm (d) Al Bm ilj jmi
(e) Ai Bk Cn ijm kln jlm
P3 Translate diagrams in Table I into plaintext (index-free or (f) (g) (h) (i)
index) notation. Attach index markers to the diagrams
adequately if needed. TABLE I. Diagrams for P3.
 2

~ A
P7 Prove that A× ~ = 0 in the graphical notation, and then in loops, it is a “tree-level diagram.” Then we have “one-loop
the plaintext notation. Show the correspondence between diagrams,” “two-loop diagrams,” and so on. Classify your
two proofs by pairing each step. diagrams by their loop numbers.
P8 Prove the following graphical identity for a unit vector P15 According to the terminology for Feynman diagrams, “one-
~n (i.e., ~n · ~n = 1) in the graphical notation, then in the particle irreducible diagrams” (abbreviated as “1PI dia-
plaintext notation. Show the correspondence between two grams”) are diagrams that cannot be split into two pieces
proofs by pairing each step. [Note: this is a linear map by cutting an internal line. For example, is not a 1PI
that projects out the component parallel to ~n.] diagram ( or ), but is. Find all 1PI diagrams
that can be built from three cross product machines and
− = − n n (1) reduce them into lower-epsilon terms until it becomes im-
possible to break them down further by the two-epsilon
n n
identities in P12 (because the two-epsilon identities re-
duces a two-epsilon subdiagram into a zero-epsilon subdi-
P9 Confirm that the above ~n-projector is idempotent by dia- agram, there will remain one epsilon in each term of the
grams. That is, if we call Eq. (1) by “ ,” show that final expression).
= .
How do three-epsilon networks play a role in real use? First,
consider the following three-terminal one-loop diagram, which
2. See the Bones, Attach the Flesh Pieces
you may have obtained in P15.
P10 Show that (A×~ B)·(
~ C ~ × D)
~ = (A· ~ C)(
~ B ~ · D)−(
~ ~ D)(
A· ~ B~ · C)
~
~ × (B~ × C)
~ = B(~ A ~ · C)
~ − C(
~ A
~ · B)
~ = − (2)
and the BAC-CAB rule A
are daughters of the same tensorial structure, = − .
List all the algebraic identities that originate from = P16 The proof of the Jacobi identity in P11a proceeds with
− , in the plaintext notation. breaking down all the terms into expressions of lower-
epsilon level (expressions that has less cross product ma-
P11 P6a is called the “Jacobi identity.”
chines). The resulting six terms cancels one another. In-
(a) Prove the Jacobi identity using = − . Reduce stead of that, however, start from and rather “climb
each term into smaller units, and show that they can-
cel out one another. up the stairs” so that move on to four-epsilon expression,
then to the desired result. Use Eq. (2) and a two-epsilon
(b) Generate all vector algebraic identities that are iden-
identity during the process. Keep in mind the lesson here:
tical to the Jacobi identity. [Hint: extract the bones
sometimes, going up can provide a shortcut.
first, then attach the flesh pieces one by one.]
Meanwhile, a tree-level three-epsilon diagram also comes into
3. Epsilon Networks (Essential) play in vector algebra.
P17 Construct a vector algebra identity based on the two pos-
Now, we abstract out the flesh pieces and investigate the world
sible expansion of a three-epsilon composition . First,
of bones. First, consider identities involving two cross product
machines. apply = − to the adjacent cross product ma-
chines on the left side, then on the right side. [Answer :
P12 By joining (contracting) the terminals of = − , we ~ B
(A· ~ × C)
~ D~ = A(
~ B~ ·C
~ × D)−
~ ~ A·
B( ~ C~ × D)+
~ ~ A·
C( ~ B
~ × D).]
~
can obtain derived identities.
If you want, you can spend your time playing with higher-epsilon
(a) Join two terminals of = − , and obtain a non-
diagrams. Penrose1 , for example, provides more identities, so
trivial identity. [Answer : = −2 .]
you may want to check it.
(b) Join the two terminals of the identity you obtained
from the previous question. What identity do you
get? [Answer : = −6 .] 4. Epsilon Networks (Advanced)
(c) Prove the identities obtained in P12a and P12b using This section introduces the concept of the antisymmetrizer
the plaintext notation. and the consequences of permutation symmetry of diagrams
Next, let’s move on to identities involving three cross product (Jucys-Levinson-Vanagas theorem or Schur’s lemma) in a con-
machines. cise manner. There are good references2,3 regarding this subject,
so we rather mention only the essentials.
P13 Have some fun playing with the cross product machines. Until now, = − has been utilized as the fundamen-
Take out three cross product machines from the toy box tal identity that enables us to proceed all calculations. It was
and assemble them into various configurations. Find con- instructive to introduce the graphical vector algebra by such
nected nonzero expressions as many as you can. For ex- manner, instead of giving all the details in a rigorous and logi-
ample, exclude diagrams such as or , which are cally preferred order from the beginning. In fact, however, the
disconnected or zero due to the knot at its rightmost part, fundamental two-epsilon identity is not = − but the
respectively. following,
P14 How can you classify the diagrams you obtained in P13?
A possible criterione is to divide them in terms of the num- − = 3! , (3)
ber of loops they contain. If a diagram doesn’t have any
 3


where a thick line over three Kronecker delta lines means that ~ =f∇·A
(a) ∇ · f A ~ + ∇f · A
~
they are totally antisymmetrized, i.e.

(b) ∇ × f A~ =f∇×A~ + ∇f × A
~
 
1 P26 Express the following identity by diagrams.
:= + + − − − (4)
3!
 
1 (∇ × A) ~ = ∂Aj ∂Bj − ∂Aj ∂Bi
~ · (∇ × B) (7)
= +(−)1 +(−)3 , (5) ∂xi ∂xi ∂xi ∂xj
3
 
1
:= − . (6) P27 Prove the following identities by diagrams. The expression
2!

∇˙ A~˙ · · · means that the differentiation operates on A ~ only
The minus sign in Eq. (3), the “master” identity, is due to our 6
(Hestenes’ overdot notation).
convention that reads terminals of a cross product machines



(a) ∇ A ~ · (B~ × C)
~ =∇ ˙ A~˙ · (B
~ ×C
~ +∇ ˙ A ~˙ × C)
~ · (B ~ +
anti-clockwise. 3! is due to normalization convention, which is


optional. In the plaintext notation, Eq. (3) reads “ijk lmn = ∇˙ A ~ · (B ~˙
~ × C)
j k
3!δ[li δm δn] .”4 Antisymmetrizing four Kronecker deltas gives zero ˙ A ~˙ · (B
~ × C)


~ = (B ~ × C)


~ ·∇ A~ + (B
~ × C)
~ × (∇ × A) ~
(b) ∇
in three-dimensions, i.e. = 0, because there cannot exist four
linearly independent vectors. P28 Calculate ∇(~n), where ~n := ~r/r is the unit radial vector.
[Answer : r−1 times the ~n-projector, Eq. (1)] Can you give
P18 [?] Confirm that = − can be derived from Eq. (3) this result a geometric interpretation?
by diagrams. (How is it compared to using the plaintext
notation?) P29 Graphically represent the identity “∇·(~n/r2 ) = 4πδ (3) (~r).”

P19 [?] Confirm that the identity in P17 can be derived from
2. Second Derivatives
= 0.
For later purposes, we define the symmetrizer too. It is defined We already mentioned in the main article that (∇ × ∇) van-
as averaging ishes as an operator identity for well-behaved tensor fields, i.e.,
1
 all permutations of indices. For example, :=
~ = 0, or (∇ × ∇)(any well-behaved ten-
2! + . is defined similarly: just change all minuses into (∇ × ∇)f = 0, (∇ × ∇)A
pluses in Eq. (4). sor field) = 0.
P20 Confirm that + = , but + 6= . P30 The following three are second derivative identities that
2 are commonly mentioned in the literatures. Prove these
One can prove Eq. (3) from the identity in P17, but the
by the graphical notation.
calculation is tedious. Instead of that, the observation that the
three upper terminals of the left hand side of Eq. (3), , are to- ~ =0
(a) ∇ × ∇f = 0, ∇ · (∇ × A)
tally antisymmetric as much as the three lower terminals implies ~ = ∇(∇ · A)
(b) ∇ × (∇ × A) ~ − ∇2 A
~
that it is proportional to .5 Then, the proportionality constant
P31 We are now going to study the Laplacian operator. Prove
can be determined by appropriate contraction of terminals.
the following identities by diagrams.
P21 [?] Prove Eq. (3) by such “appealing to symmetry” method (a) ∇2 (f g) = g ∇2 f + 2∇f · ∇g + f ∇2 g
we just described. ~ =A ~ + 2(∇f · ∇)A
~ ∇2 f + f ∇2 A ~
(b) ∇2 (f A)
P22 [?] This strategy is readily applied to other multi-terminal 2 ~ ~ ~ 2~ ~ 2~ ~ · ∇)A
~ +
(c) ∇ (A · B) = A · ∇ B − B · ∇ A + 2 ∇ · (B
cases. Schur’s lemma says that any two-terminal epsilon

~ × (∇ × A)
B ~
network is proportional to Kronecker delta. Prove =
−2 from Schur’s lemma. P32 [?] Calculate the following expressions by diagrams. Note
P23 [?] Prove Eq. (2) by appealing to permutation symmetry, that ∇2~r = 0.
provided that = 6. (a) ∇2 r, ∇2~n (b) ∇2 z
2 2 2 2
P24 [?] Any one-terminal epsilon network (which is called a (c) ∇ (r ), ∇ (3z −
r ), ∇2 (x2 − y 2 ),
∇2 (xy)
2


“tadpole diagram” in the terminology of Feynman dia- (d) ∇2 z(5z 2 − 3r2 )
, ∇2 x(5z 2 − r2 ) , ∇2 z(x2 − y 2 ) ,
grams) must be equal to zero. Explain why. ∇2 x(x2 − 3y 2 )
Note that students themselves, by doodling the graphical alge- P33 Proving Bernoulli equation for the case of incompressible
bra, can find the fact that “a compound n-terminal object that potential flow involves

the following index gymnastics: (~u ·
has a permutation symmetry can be reduced into a simpler ex- ∇)~u = 21 ∇ ~u · ~u , where ~u(~r) is a vector field that can be
pression of the same symmetry up to a proportionality constant”, written as ~u(~r) = ∇Φ(~r) for some scalar field Φ(~r). Prove
as mentioned in the main article. this by diagrams.
P34 When deriving the vorticity equation, one takes curl to the
Navier-Stokes

equation. Prove

that for a vector field ~u(~r
),
B. Graphical Vector Calculus ∇× (~u·∇)~u = (~u·∇) ∇×~u +(∇·~u)(∇×~u)− (∇×~u)·∇ ~u
using diagrams.
1. First Derivatives
Identities of higher order derivatives can be obtained by com-
P25 There are six first derivative identities, and the missing binations of first and second order identities. Since they do not
ones in the main article are the following two. Prove these illustrate the unique power of graphical calculus much, we choose
by the graphical notation. not to discuss higher-order identities in detail.
 4

P35 Graphically represent the operator identity ∇2 ∇ = ∇ ∇2 i.e., flipping the balloon inside out. When a balloon swallows a
for well-behaved tensor fields. manifold P, it becomes “differentiated:” ∂P. This is a graphical
P36 When studying spherical waves, a useful fact is that if manifestation of the pairing of differential forms and chains in
ψ(~r) satisfies the Helmholtz equation, (∇2 + k 2 )ψ(~r) = 0, a de Rham-theoretic manner, namely, hP, df i = h∂P, f i. When
then ~r × ∇ψ(~r) satisfies the vector Helmholtz equation, a manifold and a tensor field are met, they produce the “con-

traction,” i.e., the value of integration of the tensor field on the
(∇2 + k 2 ) ~r × ∇ψ(~r) = 0. Prove this using the graphical
notation. manifold. Contraction between a manifold and a field is denoted
just by juxtaposing them, where Stokes theorem translates to
flipping the differentiation balloon inside out.
3. Graphical Notation for Integral Calculus What about curl and divergence theorems? To be specific,
Eq. (11) used the following substitution.
According to the graphical notation for R and ∇
R scalars, vectors,
discussed until now, the curl theorem, S d2~a · ∇ × A = ∂S d~l· A,
~ Z i Z
can be written as the following. dli ↔ and ↔ (13)
P P ∂P P

. (8) R R
For the divergence theorem, V d3 x ∇ · A ~ = ~ for a
d2~a · A
R ∂V 3
volume V, one can guess that a substitution V d x ↔ can be
∂ is the boundary operator, and S is a surface. Here, one can made. Then, the left hand side of the divergence theorem can
observe that A is rather a “dummy” vector field and extract be written in a completely graphical form. After that, flipping
the “essence” of the curl theorem as the balloon will generate the right hand side. The result is the
following.
. (9)
A A  
simple
R = (14)
In the plaintext notation, Eq. (9) will be written as S d2~a × version
R
∇ [· · · ] = ∂S d~l[· · · ]. Understanding the curl theorem in this
form is in manyR ways useful. For R
R example, inserting a scalar Then, it follows that ∂V d2 ai ↔ i . In case of the curl
field f (~r) gives S d2~a × ∇ f = ∂S d~l f . Or, insertingR ~r then theorem, substitutions
taking cross product derives the vector area formula S d2~a =
R
1
2 ∂S ~ r × d~l as
:= and := (15)

(10) make Eq. (9) to be represented in the “flipping balloon” form.

−2 r r .

Now, although it is not necessary, one would feel an “im- = . (16)

pulse” to express Stokes’ theorems completely in diagrammatic
terms. Let us see if we can, anticipating the insight that the
graphical language will provide. One possible way of graphi-
The origin of these substitution rules can be consistently ex-
cally denoting integrals is to treat dummy variables of integrals
plained by employing the framework of differential forms; this
analogous to dummy indices in tensor expressions; the result-
will interest more mathematically sophisticated readers. In fact,
ing notation shares the same ground with the graphical ver-
index gymnastics with , , and so on corresponds to ma-
sion of bra-ket notation.7 However, when it comes to Stokes’
theorems, we would like to suggest a different approach high- nipulating differential forms. Our convention is to represent the
integration over an n-dimensional manifold as a totally antisym-
R between ∇Rand ∂. In this approach, the
lighting the trading
metric n-legged black rectangle. However, the divergence theo-
gradient theorem, P d~l · ∇f = ∂P f , where ∂P = +(endpoint
rem Eq. (14) appears to be an exception, because the integration
of P) − (starting point of P) for a path P, is expressed as
R 3a volume V was earlier denoted as a zero-legged object as
over
V
d x ↔ . To be consistent with Eq. (12) and Eq. (16), the
f f divergence theorem should be written as the following.
= . (11)
P P
= (17)
Note that the scalar field f in Eq. (11) is also a “dummy” field
~ in Eq. (8) did. Omitting f , Eq. (11) becomes
as A
is a totally antisymmetric three-legged object that denotes
R V. To be concrete, its components may be
the integration over
identified with 3! dλ1 dλ2 dλ3 (∂X[i /∂λ1 )(∂Xj /∂λ2 )(∂Xk] /∂λ3 ),
= , (12)
~ 1 , λ2 , λ3 ) is a parametrization of V. Then, when
where ~r = X(λ
P P 1
an identification = 3! is made, index gymnastics show
 5

that Eq. (17), when contracted with A , is equivalent to the P41 [?] Calculate the time-averaged Poynting vector, S ~ =
1 ~ ∗ ~
form first introduced, Eq. (14) (exercise!). One of the equations 2µ0 Re E × B, with Eq. (20) and Eq. (21). This will yield a
that appears during the index gymnastics is the following. huge tensorial expression, which, however, can be reduced
wisely with the guidance of the graphical notation. Decide
which “knot” is the best to cut off first.
A A
− = (18) Before closing this section, we present another electromagnetism
example that involves a considerable use of index gymnastics,
3
while the graphical notation effectively boosts the speed and
helps us grasp the contraction structure in a bird’s eye view (cf.
1
This corresponds to d( 2! ~ 1 ijk dxi ∧
ijk Ak dxi ∧ dxj ) = (∇ · A) the process with the plaintext notation, given in the literature8 ).
3!
dxj ∧ dxk in the calculus of differential forms, where d is the ~ ~r ) and B(t,
P42 [ ! ] Consider electric and magnetic fields E(t, ~ ~r )
exterior derivative. in a homogeneous but anisotropic linear medium that has
a constant permittivity and permeability matrix of com-
ponents εij and µij in the lab frame, respectively. The
C. Graphical Tensor Calculus
auxiliary fields are given by Di (t, ~r ) = εij Ej (t, ~r ) and
Hi (t, ~r ) = (µ−1 )ij Bj (t, ~r ), where (µ−1 )ij is the compo-
1. Symmetric Rank-2 Tensors nent of the inverse of the permeability matrix. Both the
permittivity and the permeability matrices are symmetric.
Welcome to the world of tensors, finally. Graphically, ten-
sors are nothing but multi-terminal objects: monopods, dipods, (a) Graphically write Maxwell’s equations in media with-
tripods, tetrapods, and so on. We already mentioned the graph- out any free charges and currents. Introducing a new
∂
ical notation for symmetric dipods in the main article. Let us graphical notation for ∂t depends on your choice.
do a little stretching first with the inertia tensor. The inertia (b) Consider a plane wave solution and use the substitu-
tensor of a rigid body is defined as follows. tion ∂t ∂
→ −iω and ∇ → i~k. Let the amplitude of
the electric field be E ~ 0 . Let v = k/ω be the phase
velocity. Find the equation that can be used to cal-
 r culate v 2 when E ~ 0 and the direction of propagation,
= dm := I (19) ~
~u := k/k, are known.
r
L (c) Show that the direction of E ~ 0 , can be found by the
ω ω eigenvalue problem Wij (E0 )j + v 2 (E0 )i = 0, when ~u
is known. Find the matrix Wij .
The integral is over infinitesimal mass elements of the body, (d) If you insist on working further without component-
which is labelled by ~r. wise unpacking of the matrices, you will have to cal-
1
P37 Confirm that the inertia tensor is symmetric under swap- culate Wii , 2! ijk imn Wjm Wkn , and the determinant
1
then-yanking of the external arms of the shaded expression 3!  ijk lmn W il W jm Wkn in terms of ~ u and the permit-
in the middle of Eq. (19). tivity and permeability matrices to obtain an analytic
formula of v 2 . If you have some time to spare and have
P38 The mass quadrupole moment R Qij , a rank-2 symmetric

finished Section I C 4, have fun with calculating them.
tensor, is defined as Qij = dm 12 3xi xj − r2 δij . Graph-
ically represent this equation.
2. Graphical Notation for General Multi-index Quantities
P39 Calculate the trace of the inertia tensor and the mass
quadrupole moment, respectively, by diagrams. That is, In the main article, we only introduced the graphical notation
calculate Iii and Qii . for rank-2 symmetric tensors. Now consider a general two-index
P40 Prove that Qij = − 32 Iij + 12 δij Ikk by diagrams. quantity (need not be a tensor), Xij . Suppose that Xij has no
symmetry so that Xji is unrelated to Xij . Then, following the
Now, we move on to electromagnetism. Consider the elec- philosophy of self-explanatory design, its graphical counterpart
tric quadrupole radiation of a monochromatic source ρ(t, ~r) = should be non-symmetrical in its two terminals, which becomes
~ ~r) = j(~r)e−iωt localized in V near the ori-
%(~r)e−iωt and J(t, different when it is ”swap-then-yanked.”
gin. Define
R the electric quadrupole
moment with %(~r), i.e.,
Qij := V d3 x %(~r) 21 3xi xj − r2 δij . An examination of the or-
ders of r and ω reveals that the radiation electromagnetic fields
are given as follows. Xij = i X j (22)

" n # A designer’s choice that suffices these requirements is a bead

µ0 eiω(r/c−t) iω 3 that has a non-symmetrical shape and slides along a line. The
=− 2 Q − n Q (20) bead is rigid so that the holes that lines come in and out cannot
E c 6πr 2
n n be repositioned like X −→ X −→ X . Instead,
" # swap-then-yanking (transposing) works as follows.
1 µ0 eiω(r/c−t) iω 3 Q n
= =− 3 (21)
c c 6πr 2 n Xji = i X j =i X j =i X j (23)
B n E
 6

If Xij = Xji , the designer would advise us to use a symmetric P49 [?] Consider a rank-2 tensor Tij .
shape for X, such as an oval, a rectangle, or a diamond. We
used a square in the previous sections. Diagrams for general n- (a) Show that tr T = Tii is a scalar.
1
index quantities are the same. For instance, a possible graphical (b) Show that T[ij] = 2! (Tij − Tji ) is a tensor and in fact
lm has three independent components so that it can be
representation of a general 5-index quantity, Kijklm , is K . encoded into a form of a vector (yet it is not exactly
i j k a vector). [Hint: Do some combinatorics. Then con-
3. Rotational Symmetry and Contra/Covariance sider ijk Tjk . Is it a vector? ]
(c) For Tij = Ai Bj , calculate tr T and the vector you got
To be called a “tensor,” a multi-index quantity must be en- in the last question. What are they?
dowed with invariance property. Suppose we actively transform
the vector ~r = xi~ei by ~r 0 = Rij xj~ei , where (Rij ) is a spe-
cial orthogonal matrix (i.e., encodes proper rotation) so that P50 [?] Decomposition of a rank-2 tensor into irreducible rep-
(Rij )> = (Rij )−1 and det(Rij ) = +1. In the graphical notation, resentations of SO(3). As you might already have caught
the idea in the previous problem, a rank-2 tensor Tij has
r0 = r , (24) many faces: the scalar times δij (trace-only part), the part
that can be encoded as a vector (antisymmetric part), and
where Rij := i j instead of i R j to avoid clutter. (Let lastly, the remainder, a symmetric traceless rank-2 tensor.
blank triangles always denote the rotation matrix from now on.) Thus, we have decomposed Tij into parts that transforms
differently under rotation. Show that this can be writ-
P43 Rij satisfies Rik Rjk = δij and Rki Rkj = δij . Graphically ten as Eq. (26). Check that the numbers of independent
represent these equations. [Answer : = = components are 1, 3, and 5 for the three parts, respec-
. Opposite arrowheads are pair created or annihi- tively. These numbers may remind you of the degeneracies
lated! ] of spin-0, 1, and 2 angular momentum eigenstates. In fact,
P44 There is one more constraint: det([Rij ]) = +1. we call the three parts “spin-0,” “spin-1,” and “spin-2”
part, respectively.
(a) Show that is totally antisymmetric in its three
terminals so that is equal to . (26)
1
(b) Using det([Rij ]) = 3! ijk lmn Ril Rjm Rkn and Eq. (3),
conclude that = .
(c) Confirm the following “propagation of arrowheads” If you want to distinguish between contravariant and covari-
or “arrow pushing” holds. ant indices, restrict the terminals to be always vertical. Termi-
nals heading up corresponds to contravariant indices and termi-
nals heading down corresponds to covariant indices. Invariant
= = = tensors are denoted as follows: (Euclidean) metric tensor and
i j
its inverse δij = , Kronecker delta δji = ji ,
and δ ij =
i j
the metric volume form and its index-raised version εijk =
P45 [ ! ] This problem is just for fun. Simplify the following i j k i j k

diagrams (minimize the number of cross product machines and εijk = . We follow Penrose9 for the design of εijk
and arrowheads, respectively). and εijk . For three-dimensional Euclidean space, εijk and εijk
can be “heterarchized,” i.e., we can substitute → and
i j k i j k i j k i j k
→− . A rank-2 contravariant tensor, a (20 )-tensor, X ij ,
A transforms as → . If the two indices are lowered by the
X X
metric, we have a rank-2 contravariant tensor, a (02 )-tensor, Xij :
(a) (b) (c) (d)
X
→ X . Contravariant indices get dressed with upward
When r → r , a tensor dresses outgoing arrowheads
arrowheads, and covariant indices get dressed with downward
in its all terminals. For example, the inertia tensor defined by
arrowheads. Graphical calculus with this “up-down hierarchy”
Eq. (19) transforms by the following, so it is indeed a tensor.
indeed has a beautiful and neat syntax; however, we decided not
to unnecessarily pursue the distinction between covariant and
Z Z
0 r0 0 r contravariant indices which might confuse the readers who are
I = dm = dm = I , (25) mathematically unsophisticated. Penrose9 , for example, will be
r0 r
helpful to interested readers.10
P46 Confirm the last equality in Eq. (25). P51 [?] Based on our definition of the inertia tensor, the tensor
P47 Prove that the mass quadrupole moment defined in P38 connecting two vectors L~ and ω~ , what kind

of rank-2 tensor
is a tensor. should the inertia tensor be, among 20 , (11 ), and (02 )?
P48 When A ~ and B
~ are vectors, i.e., A0 = Rij Aj and B 0 = P52 [?] Write Eq. (3), = − , = −2 , and = −6
i i
Rij Bj , show that A ~·B ~ and A ~×B ~ are a scalar and a in the graphical notation that distinguishes contravariant
vector, respectively, using the graphical notation. and covariant indices.
 7

4. Rotation Matrix P58 [?] Show that applying the decomposition Eq. (26) to
∂i (nj /r2 ) reads as the following.
The previous section studied how tensors get dressed with
rotation matrices, Rij = i j , when rotated. Now, let us
focus on the rotation matrix itself. Denote the rotation matrix n 4π (3) 3 n n − (29)
that encodes a rotation with respect to an axis, specified by a = δ (r) −
r2 3 r3
unit vector ~n, with an angle α as i n, α j .

P53 (a) For infinitesimal , show that = + P59 Derive Eq. (27) and Eq. (28) from Eq. (29).
n,
 up to O(1 ). P60 [?] Consider a stationary flow ~u(~r ) of an incompressible
n (∇ · ~u = 0), viscous fluid at low Reynolds number. The
(b) Calculate the commutator between infinitesimal rota- Navier-Stokes equation reads µ∇2 ~u − ∇p = 0. A “unit”
tion matrices, i.e., 0 − 0 , up to O(2 ). nonhomogeneous solution is called Stokeslet: µ∇2 ~u −∇p =
n, n, n, n,
P54 [?] n is called the generator of rotation. f~δ (3) (~r).
(a) It is a rank-2 tensor—what representation does it live (a) Show that the pressure field p(~r ) is determined before
in? [Answer : rank-2, spin-1 ] knowing ~u(~r ) as p(~r ) = p0 + f~ · ∇ 4πr 1
, assuming that
(b) Show that − ex ex − ey ey − ez ez is propor- p(~r ) → p0 as r → ∞.
tional to Kronecker delta. The proportionality con- (b) Show that ~u(~r ) = − 8πµr1
(f~ + ~n ~n · f~ ). [Hint: P32a.]
stant is called the quadratic Casimir.
P55 [?] Generally, a tensor A is called (minus of) the Hodge 6. Harmonic Tensor Fields
dual of a vector A.~ Let us study its algebra. Calculate the
following. (tr of a two-terminal diagram means joining the This section is entirely in the [?] level.
terminals by Kronecker delta: taking a trace.) Eq. (29) is related to dipolar fields; its generalization to arbi-



trary ranks is related to multipolar fields, which are described
(a) tr , tr , tr by the spherical harmonics Y`m (θ, φ). Consider describing scalar
A A B A B C
k
Laplace’s equation ∇2 f (~r) = 0 in spherical coordinates. Spher-

k    ical symmetry tells us that the space of solutions is spanned
(b) := ···
n n n n by functions of the form R` (r)Y`m (θ, φ) with ` = 0, 1, 2, · · · and
P56 [?] Now, let us exponentiate the infinitesimal rotation (ro- m = −`, ` + 1, · · · + `.
tation algebra) to obtain non-infinitesimal ones (rotation 1 1 ∂ 2 ∂ 1 1

N
0= r R` (r) + 2 m ∇ 2 m
° Y` (θ, φ) (30)
group); = lim α = exp α . Calcu- 2
R` (r) r ∂r ∂r r Y` (θ, φ)
n,α N →∞ n, N n
late this and obtain the Rodrigues’ rotation formula. Here, we defined ∇ ° 2 as Laplacian on unit sphere. Since
P57 [?] Given a rotation matrix Rij , how can one figure out its −∇ ° Y` (θ, φ) = `(` + 1)Y`m (θ, φ), R` (r) = r` or 1/r`+1 . The
2 m

rotation angle and rotation axis? former gives solid harmonics that are not singular at the origin

(“regular” solid harmonics).
(a) Show that tr = 1 + 2 cos α.
n,α While dealing with spherical harmonics, working in Cartesian
(b) As you might discovered while playing with P45, coordinates can provide shortcuts in calculations, so we will de-
velop such Cartesian technology in several following problems.
= . What does this mean? From dimensional considerations, we expect that r` Y`m (θ, φ) can
be written as a `th -order polynomial of z, x, and y, which are
(c) The previous two problems are related to the spin-0 ~r · ~ez , ~r · ~ex , and ~r · ~ey , respectively. This means that there ex-
and spin-1 parts of , respectively. What is its ists a converter (a bunch of Kronecker deltas) that connects `
n,α
spin-2 part? ~r ’s with ~ez ’s, ~ex ’s, and ~ey ’s of total number ` so that such con-
tractions make r` Y`m (θ, φ). Denote this “converter” as .
5. Finding the Oblivious Delta Term ` 0 `
For example, r Y` (θ, φ) ∝ r P` (~n · ~ez ) has no φ dependence so
In electromagnetism, the fields that Gilbert and Ampère (or, that ~n and ~ez are the only players; therefore, define by
electric and magnetic) dipole moments (denoted p~ and m ~ below) Eq. (31). Then, allowing ~ex and ~ey to join the game, we get
generate contain delta function terms as the following. the spherical harmonics algebraically, rather than the analytic
  approach—solving partial differential equation on the sphere for
p~ · ~n 3 ~n ~n · p~ − p~ 1 (3) nonzero m.
−∇ = − p~ δ (~r) (27) 
4πr2 4πr3 3
  e z ez ez
~ × ~n
m 3 ~n ~n · m~ −m ~ 2
∇× = + m ~ δ (3) (~r) (28)
4πr2 4πr3 3 P` (~n · ~ez ) = . (31)
Undergraduate education used to avoid tensorial expressions so n n n
that somewhat long approaches are required to derive these
identities.13,14 However, tensor calculus allows one to alge- However, Eq. (31) does not determine uniquely: an anti-
braically obtain the delta function terms in dipolar fields, and symmetric component vanishes when identical vectors are con-
we discuss its graphical form. tracted in its indices. Thus, we demand the upper indices to
 8

be totally symmetric and so do the lower indices in addition to P66 [?] i~ is the “spin operator” (of direction ~ei ) acting on
i
Eq. (31). For example, P1 (~n · ~n0 ) = ~n · ~n0 =⇒ = , and vectors. Why?
P2 (~n·~n ) = 32 (~n ·~n0 )(~n·~n0 )− 12 (~n·~n)(~n0·~n0 ) =⇒
0
= 32 − 21 . (a) Show that (i~)2 − (i~)2 = (i~)2 .
i j j i i j
Note that 32 − 12 , 23 − 21 , or other possibilities such as How does this coincide with the spin operator you
1.7
+ 1.3 − 12 also gives 32 (~n ·~n0 )2 − 12 when contracted with have learned in quantum mechanics?
2 2
0
~n in its lower indices and ~n in its upper indices but we choose (b) Show that i~ em = (−)m ~ em , where m =
e0
“the” totally symmetric one, 32 − 12 . +1, 0, −1. Thus, em ’s are the eigenvectors of z-
axis rotation.
P61 The converter, , is denoted by a horizontally sym- (c) Show that Y2+1 (θ, φ) transforms like a vector, partic-
metrical shape. Can you explain why would the designer ularly, like ~e +, under infinitesimal z-axis rotation.
 
choose such shape? 2
(d) Calculate (i~) + +2 and explain
 

P62 [?] Taking Laplacian to r` = should give its meaning.

n n n r r r P67 [ ! ] Some diagram gymnastics. Calculate the following by
zero. From this, show that is “traceless,” i.e., it diagrams. (You may wonder where these expressions ap-
vanishes when any of its two lower indices are contracted pear. The solution to this problem contains an answer for
(the same applies to the upper indices). such curiosity.)
P63 [?] From the generating function of Legendre polynomi-
als (the multipole expansion), prove the following. By p m 4
“−(traces),” we mean subtracting terms proportional to 9
and so that the entire right hand side is traceless in n n n n
the upper indices and so do in the lower indices.
(a) (b) (c)
 

! 1
= (-) - traces (32) D. Addendum
r +1 r
n n n
1. Quantum Mechanics with the Graphical Notation

P64 [?] So far, we have found that , regarding only its The last application of graphical notation to be mentioned is
upper indices, is a rank-` totally symmetric and traceless quantum mechanics. In quantum mechanics, physical quantities
tensor (i.e., it lives in the spin-` representation of SO(3)). are promoted to non-commuting operators. To denote operators
The same applies to the lower indices. in the graphical notation, simply place the boxes in a row and
demand the relative positions to be not changed. For example,
(a) Check that tracelessness holds for ` = 2 and 3 by the canonical commutation relation [x̂i , p̂j ] = i~δij 1̂, where xi
explicit calculation, given that P2 (ξ) = 32 ξ 2 − 12 and and pi are the components of the position vector operator ~r and
P3 (ξ) = 25 ξ 3 − 32 ξ. the momentum vector operator p~, is denoted as follows.
(b) Also, do the inverse direction: find the coefficients aj
P4 x̂i p̂j − p̂j x̂i = i~δij 1̂
where P4 (ξ) = j=0 aj ξ j , using the fact that
is a rank-4 totally symmetric and traceless tensor. l
j j j (33)
(Fix the overall coefficient by P4 (1) = 1.) i i i
− = i~
Note that taking Laplacian

to Eq. (32) gives the equality be- r p p r
tween ∇2 r−`−1 Y`m (~n) and an `th -order, (` − 2)th -order, · · ·
derivatives of delta

function, i.e., point multipole sources. Thus, For convenience, introduce the conventional square bracket no-
∇2 r−`−1 Y`m (~n) is not zero in physicists’ sense but raises mul- tation for commutator, instead of inventing a new graphical no-
tipolar singularities at the origin (it is “almost zero”). tation. For example, the commutator of two vector operators A ~ˆ
P65 [?] Define ~e + := − √12 (~ex + i~ey ), ~e − := + √12 (~ex − i~ey ), ~ˆ [Âi , B̂j ] = Âi B̂j − B̂j Âi , is represented as the following.
and B,
and ~e 0 := ~ez (the spherical basis vectors). Comparing  
withRefer to the spherical harmonics table, and check that A , B A B B A
q e+ := − (34)
+1 3
the first few ones can be obtained by Y1 = 4π , i j i j i j
e +
e +
e +
e 0 n
q q Remember that by taking a commutator you should only alter
Y2+2 = 2
3
15
8π , Y2+1 = 2
3
15
4π , etc. the order of boxes, keeping the endpoints of the lines (vector
n n n n indices) fixed.
 9

The orbital angular momentum operator is graphically repre- been always discussing about. The spin operators satisfy the
following commutation relations.
sented as = .
L r p  
S S S S S S S
P68 Prove the following by diagrams. , = − = i~ (36)
i j i j i j i j
(a) ~rˆ · p~ˆ = p~ˆ · ~rˆ + 3i~1̂ (b) ~rˆ × p~ˆ = −p~ˆ × ~rˆ
As usual, Pauli matrices can be employed to represent the spin
P69 Graphically represent the so(3) = su(2) commutation rela- b
operators as (Si )a = 12 (σi )a .
b

tions, i.e., [L̂i , L̂j ] = i~ijk L̂k . Graphically prove that the Note that H-lines (spin-1/2 lines, labelled by a, b, · · · ) are be-
orbital angular momentum satisfies such relation. ing differently denoted from so(3)-lines (spin-1 lines, labelled by
P70 Angular momentum operator identities can be quickly de- i, j, · · · ), because the two vector spaces are different. We cannot
rived using the graphical notation. One of the examples connect the two ends of a spin-1 line and a spin-1/2 line directly
~ˆ · L
~ˆ = r̂2 p̂2 − (~rˆ · p~ˆ)2 + i~~rˆ · p~ˆ. Prove this identity such as A . Also, there is no invertible “basis change” ma-
are L̂2 = L trix that connects one spin-1 and one spin-1/2 line. Even the
by diagrams. dimensions are different—3 and 2. We must use appropriate con-
There are also other miscellaneous identities, such as [L̂2 , L̂i ] verters (invariant symbols). What we expect from such convert-
" # ers is a “proper propagation of arrowheads.” For example, ro-
1 
= 0, r , L =
r
, = + = 2i~ , tating an SO(3)-vector translates into unitary transformation of
i~ r , L r L L r r
the corresponding spinor in the spin-1/2 language; that is, there
 
[~rˆ, L̂2 ], L̂2 = 2~2 (~rˆL̂2 + L̂2~rˆ ), etc. One may benefit from using is a correspondence between a rotation matrix Rij = i j
b
graphical notation while deriving these identities. We shall omit and a unitary matrix Ua = a b . The translator in charge is
the details. b
the spin operator (Si )a = a S b .
i

2. Intertwining Different Types of Lines

S S S
−→ =
The graphical notation for operators leads us to an interesting (37)
i i i
discussion. First, let us hear the designer’s alternative choice for a
denoting noncommutative numbers. Our previous notation can χ̄a (Si )a χb
b
−→ b
χ̄c U † c (Si )a Ub d χd = Rij χ̄a (Sj )a χbb

be made more clear by turning on an “order marker,” like the

It is not necessary to introduce the concept of representation
following.
converters in the context of quantum mechanics (we did because
i j we wanted to start with objects that are familiar from undergrad-
Âi B̂ Ĉj ↔ (35) uate education). Objects that intertwines two different kinds of
A B C lines naturally appear in group theory: generators of group ac-
tion, Clebsch-Gordan coefficients, converters between two rep-
Noncommutativity made explicit: the order of loading boxes resentations (e.g., differential operators and SO(3) rotation or
on the “conveyor belt” matters. A vector operator Âi is a spin-3/2 and spin-1/2 representations of SU(2)), etc. Further-
i more, it is also possible to consider objects entailed with multiple
box with a vector tail i and a pair of thick grey lines: .
A
It intertwines two kinds of lines. Note that thick lines are types and numbers of indices, such as , , or .
“left-right hierarchical,” that is, they must be drawn horizon- In group theory, there exists a systematic procedure of ob-
tally and lines stemming left and right are distinguished (cf. taining tensorial identities.3 We choose not to reproduce it here;
“up-down hierarchy” in Section I C 3). Meanwhile, what is most of the essentials of such procedure is already illustrated
that grey conveyor belt anyway? The answer can be found by in a heuristic manner throughout the development of this arti-
inferring what objects with a thick grey line represent. They cle. It is all about imposing constraints on the group algebra
are transformed by objects with one input and one output thick from demanding the “proper propagation of arrows.” For ex-
i
grey lines, such as ; they are wavefunctions, i.e., Hilbert ample, the invariance of δij under infinitesimal rotations imply
A i that the rotation generators are antisymmetric. The invariance
space vectors! For example, hφ| Âi |ψi ↔ φ A ψ . of generators themselves (cf. Eq. (37)) boils down to the commu-
tation relation studied in P53b, where ijk being the structure
Vectors living in the Hilbert space H are represented as the constant of the so(3) algebra. The invariance of the structure
objects that have one thick grey line. The reason for choosing constant then implies the Jacobi identity (cf. Eq. (64)). One
a thick line is to imply that it is infinite-dimensional, i.e., wave- thing that needs an extra care is the analog of = − for
functions ψ(~r) living in H are labelled by a “continuous index” ~r. a general Lie group. In case of the group SO(3), the invariant
Meanwhile, we also have seen finite-dimensional Hilbert spaces in symbol ijk , the rotation generators, and the structure constant
undergraduate quantum mechanics, such as spin states. For ex- all coincide. However, this is not true for general Lie groups so
ample, the spin part of an electron wavefunction lives in a Hilbert that the analog of = − can be not unique.
space of complex dimension two. The state vectors are called
spinors: χa = a ∈ H (a = 0, 1). (χ0 , χ1 ) = (1, 0) is the “up” P71 [?] In this section, we are broadening our horizons to mul-
state, while (0, 1) is the “down” state. Also, there are spin op- tiple types of lines. To illustrate the “proper propagation
b of arrowheads” condition for invariant symbols, consider a
erators (Si )a = a S b acting on spinors ( → S ),
i i hypothetical algebra that has three types of lines: black,
where i = 1, 2, 3. The index i here is an SO(3)-index we have red, and blue. Suppose there exists an invariant symbol
 10
µ
entailed with all the types of lines: σaµḃ = . The (c) We stacked two i-lines and then made the I-world
a ḃ
condition for this object to be invariant under infinitesimal that reproduces the syntax of its mother. We can
transformations is the following (infinitesimal version of) repeat this procedure to build, say, the I-world,
“propagation of arrowheads.” where I ↔ [IJ], that has the same syntax with the
I-world so with the i-world.
S = + (38)
(43)
Suppose that the basic grammar of this algebraic system
is given by the following. Furthermore, we can repeat this again and again and
construct an infinite tower—fractals! On the other
= , = = 0, = = d. (39) hand, can we step down? An interesting observation
by Penrose17 is that it is possible. What he calls
“binors” have the following property.
Find the expression for the generator for red lines, ,
S 1
provided that is known. [Answer : d S
.] = d, + + = 0. (44)

P72 [ ! ] The idea of collective indices is to regard a bunch The relation between binors and the i-world par-
of indices as one group. Consider a grouping I ↔ [ij]. allels that of the i-world and the I-world. That
Then, an antisymmetric tensor Tij = T[ij] is written is, when i ↔ [ab] and j ↔ [cd], δij = i j ↔
√
“TI ” in the collective index notation. We can regard i δa[c δd]b = ab c
d . Show that ↔ ± 2
and I as “different types of lines” and investigate the
correspondence between the i-world and the I-world. It is reproduces = − . In addition, find the
instructive because one can derive the algebra of I-tensors value of d that reproduces = −2 , i.e.,
from the lower level syntax, the algebra of i-tensors. In √ 2
(± 2) = −2 . [Answer : d = −2.]
the graphical notation, the converter between the two
worlds is drawn as the following. These diagrams for collective indices
reminds the author of this picture of
−2
i j a collision between two dipolar vort-
(MI )ij = (40) ices.18 Two vortices interact by int-
I erchanging half of their body with
each other; a big particle comprised
Do you see the antisymmetrizer inside the “sheath?” If of two smaller elements interact with
we peel off the sheath, that is, when I ↔ [mn], another one by interchanging one of
their constituents. It is analogous to
−2
i j the case of binors where two spin-1/2
(MI )ij ↔ = −2δi[m δn]j . (41) “particles” made a spin-1 “particle.”
mn When the “small” particles are appro-
priately grouped, the syntax of the
In this problem, i, j, · · · are SO(3)-vector indices as usual. small world reproduces the syntax of the big world. How-
ever, the “big syntax” does not necessarily tell everything
(a) Prove the commutation relation (MI )ik (MJ )kj − about the “small syntax.” The vortices may not remain in
(MJ )ik (MI )kj = fKIJ (MK )ij , where fKIJ = a dipolar shape after collision in some circumstances, for
−2(δk[m δn][r δs]l − δl[m δn][r δs]k ) when I ↔ [mn], J ↔ example; there are binor identities that do not correspond
[rs], and K ↔ [kl]. to any i-world expressions. The flow from the small to the
−2 −2 −2 −2 −2
big world is one-sided; information is lost. This implies a
possibility of multiple small worlds corresponding to the
−4 (42) big world; a number of “small syntax” can implement a
- =
“universal syntax.” Repeating the grouping procedure,
microscopic details will be completely washed out so that
a fixed point will be achieved; recall the “self-similarity”
(b) The I-version of Kronecker delta is the antisymmetric of SO(3)-tensors under antisymmetric grouping. The
projector; δIJ ↔ δm[r δs]n when I ↔ [mn] and J ↔ “renormalization group flow” flows from binors to the
[rs]. In the graphical notation, δIJ = I J . This i-world which is the fixed point.
satisfies δIJ δJK = δIK . What is the I-version of ijk
then? The requirement is that δIJ and IJK have the 3. Connection to Feynman Diagrams
same syntax with δij and ijk . (After constructing
δIJ and IJK and establishing their syntax that is We just have opened up new possibilities to our toy box: ver-
derivable from the lower level structures (the i-world), tices that intertwine different kinds of lines. Now, one may be
we can solely work on the I-world when dealing with curious about what is going on in Feynman diagrams, where var-
I-world identities, not referring to the “microscopic ious kinds of lines standing for different kinds of particles appear:
±1
implementations.”) [Answer : i IJK := √ f
2 IJK
.] straight, wiggly, squiggly, dotted, · · · . Regarding the structural
 11

aspects only, Feynman diagrams are just special categories of the so that the “right-to-left hierarchy” reading rule cannot be ap-
graphical equations we have been discussing. plied anymore.
Let us consider Feynman diagrams in continuum quantum Eq. (46) is in fact an equal-time Feynman diagram! Indices
field theory as an example. Field theory studies about physical p~, ~k, ~q are momentum labels, indicating momentum flows. Now,
quantities distributed on spacetime, such as electromagnetic po- let us move on to operator fields at different times. A diagram
~ ~x)) or displacement field of vibrating atoms ~
tential (V (t, ~x), A(t, corresponding to an expression O2 q~~k (t2 , ~y )O1 kp~ (t1 , ~x), for t2 >
~ ~x). Typically, field variables take values
in a solid lattice, ξ(t, t1 , would be the following.
at an infinite or large number (i.e., an order of Avogadro num-
q
ber) of places—electromagnetic fields take values at every point t2 y k
O2
in spacetime, and the displacement field takes values at each (47)
t1 O1
site on the lattice. In this sense, field theory is a study of in- x
p
finite or large degrees of freedom. It is typical to approximate
~ ~x) as continuum fields. For instance,
discrete fields such as ξ(t, As conventional space-time diagrams do, the flow of time is de-
~
for ξ(t, ~x), this is done by regarding the atomic spacing to be picted as vertical direction, while the spatial dimension is de-
small. Quantum field theory studies quantum fluctuations of picted horizontally. This diagram shows a particle moving in
fields. Let φ(t, ~x) be an anonymous field (the letter “φ” for de- spacetime. Recall that the total space of invariants was spanned
grees of “freedom”). When ~ is turned on, the value of φ(t, ~x) by rank-0, 1, 2, · · · tensors in the case of finite-dimensional ten-
gets uncertain as a particle’s position variables do in quantum sors. Analogously, the whole possibilities of quantum fluctua-
mechanics; that is, it becomes an operator, φ̂(t, ~x). The result tions come in zero-particle state, one-particle states, two-particle
states, and so on: diagrams with 0, 1, 2, · · · lines.
is various operator fields: φ̂2 (t, ~x), φ̂3 (t, ~x), and so on (cf. x̂2 ,
Finally, if another type of Hilbert space H0 (labelled by primed
x̂3 , · · · in quantum mechanics).19 Generally, φ and its conjugate
indices p~ 0 , ~q 0 , · · · ) participates in this game, a different kind of
momentum π together comprise all kinds of fields in a theory.
grey line comes into play. For example, there can be an operator,
Think of some local operator Ô(t, ~x). When the Hamilto- 0
say, Ap~ q~p~ (t, ~x), that intertwines two H indices and one H0 index:
nian Ĥ of the system is given, operators evolve through time
according to the time evolution law in the Heisenberg picture: p  q
Ô(t, ~x) = Û −1(t ← t0 ) Ô(t0 , ~x) Û (t ← t0 ) with A

Û (t ← t0 ) being t x p . (48)
1
the time-evolution operator exp i~ (t − t0 ) Ĥ . Let us first con-
sider operators at a certain equal-time slice, thus suppressing Then, we can assemble such operators to build diagrams such as
the time label. How can Ô(~x) be represented graphically? It
is achieved by simply replacing the discrete index i in the for- l p  q
A q O p t2 y
mer of Section I D 2 into a continuous index ~x: t1 A , (49)
x k A p
i x
:= Oq~ p~ (~x). The continuous indices p~, ~q, · · · label H-lines. (A ~ 0
which denotes Ap~ 0 l~k (t2 , ~y )Ap~ q~p~ (t1 , ~x): two “ ” particles are
“state” in a quantum field theory corresponds to a probability
scattered via exchanging a “ ” particle. Different 0types of
profile of field configurations at a certain time slice, so the state
Hilbert spaces means different kinds of particles. Ap~ q~p~ (t, ~x),
space H is infinite-dimensional.) Here, instead of being denoted
dressed with various types of indices, functions as a converter
by using the hat symbol, an operator is seen as a tensor that has
between Hilbert spaces: an interaction vertex between particles.
one input H-line and one output H-line: it takes a state in the
The lines are particles: objects, nouns. The vertices are inter-
Hilbert space and returns another state!
actions: morphisms, verbs. Eq. (49) is a typical diagram you
Now, successive application of such operators would be de-
would meet in quantum electrodynamics, up to extra structures
noted as the following.
including polarization and spinor information.
k The significance of Feynman diagrams in physics is perhaps
~
O2 q~~k (~y ) O1 kp~ (~x) = q O2 O1 p
(45) that it explicates the particle interpretation. The deduction
of Feynman diagrams above may be vague in the standards of
y x
physics. For example, are those grey lines really particles, while
~k is a dummy index. The designer now pursues minimalism they come from the graphical representation of operator fields
and suggests us to put boxes at position ~x rather than writing Ô1 (t, ~x), Ô2 (t, ~x), · · · that are arbitrary? This is because the
“~x ” at its side—abolishing the abstraction inherent in the gap discussion was somewhat formal. A typical explanation is op-
between letters and its meaning and returning to the primitive, posite to discussions here: one arrives at birdtracks from Feyn-
“hieroglyphic” level. man diagrams by leaving only the group-theoretical part. To
make the physical content clearer, it is recommended to start
q k with concrete physical examples, e.g., analyzing the grammar
~
O2 q~~k (~y ) O1kp~ (~x) = y
O2 O1 (46) of a Feynman diagram in quantum mechanics. Studying how
x p
Feynman diagrams arise in the Hamiltonian framework of quan-
The position vectors ~x and ~y (indicated by mint arrows) are tum field theory is also helpful. Further, Lagrangian (functional)
three-dimensional; we regret that we could not use a 3D printer formalism20 helps to comprehend the “input and output” nature
to print the diagram Eq. (46) in three dimensions. Instead of of Feynman diagram elements. The worldline formalism will also
that, we expressed the spatial position as horizontal displace- provide insights by directly associating the lines in Feynman di-
ment. We drew small arrows alongside the grey lines to clarify agrams with first-quantized Hilbert space. Depending on the
the operator ordering, as the placement of operators now de- formalism, the specific interpretation of Feynman diagrams may
pends on the position labels rather than the operator ordering appear a little differently. For example, the diagram Eq. (47) can
 12

be seen as depicting a path integration with operator insertions the hairs would spiral around the origin. We believe that the
Ô1 (t, ~x) and Ô2 (t, ~x) presented in a form of “gluing formula,” abstraction of the hairs into “ ” is acceptable enough.) We
rather than a particle process; Eq. (50) schematically shows this just “physically implemented” or “materialized”23 the differen-
idea, where red shade represents path integration and the bumps tiation balloon notation by a topological surface operator. Such
at the ends depict the boundary conditions of the path integral. an inversion is intriguing. A Feynman diagram notates an S-
matrix that corresponds to a scattering scenario. At the same
β β β β time, scattering of particles itself also “notates” the Feynman
O2 O2 O2
diagram; it implements the choreography (computes the dia-
~ ~ ~ O2 (50) gram). Drawing hands.24 Mathematical structures “represent”
O1 O1 O1 O1
the physical reality and vice versa. Physical reality, or materials,
α α α α also serves as a notation. Among various notations, notations
that are in the primitive, “hieroglyphic” level may be likely to
Starting from diagrams of finite-dimensional tensors, our
descriptively display the physical reality—if we assume that the
graphical calculus has undergone a rapid expansion in the past
physical reality is written in geometric terms. Or, to be more
two sections: the addition of infinite-dimensional lines and in-
conservative, graphical notations provide a way to give a phys-
tertwiners carrying multiple kinds of lines. What is next? Fol-
ical interpretation of mathematical structures of a theory (the
lowing purely graphical reasoning, one can think of braided dia-
“notational realism”). Think of the way how Feynman diagrams
grams (where there exists a notion of “passing over” and “pass-
give an answer to a question such as “What is a photon?”
ing under” when two lines meet), ribbon diagrams (lines become
A final comment: having acquainted
stiff ribbons so that they can be twisted), “straw” diagrams (a
with group theory before meeting
cross-section of lines become a loop rather than a point), two-
Feynman diagrams, we can ask a ques-
categories (faces also comes into play in addition to vertices and
tion: can we continue the habits from
lines), and so on. In fact, the first three appear when considering
group theory to quantum field theory?
anyons, non-commutative field theory, and string theory. How-
Can we make a parallel between the
ever, in general, whether such extended diagrammatic systems
two? Can we study the interactions Edward Tufte, Torqued
have connections with actual physics is not certain. Some exotic
of particles by applying the systematic Space-Time Feynman diag-,
diagrammatic systems may fail to find physical significance yet
procedure of demanding “proper prop- ram, 2012. Stainless steel.
remaining as a mathematical possibility. Whether an extended
agation of arrows?” It seems that modern amplitude tech-
diagrammatic syntax can be derivable from some Lagrangian is
niques implement this idea. In the story of color-kinematics
also not certain, yet a “radical” scenario is to build a quantum
duality,38,39 it turns out that the kinematic factors of a Feynman
field theory without reference to Lagrangian by taking the “dia-
diagram satisfy a relation analogous to the Jacobi identity.
grammar” as its definition. Nonetheless, there are always rooms
Moreover, this can be explicated by carrying out the idea of
for extended objects.
“S-matrices as momentum-index tensors” as a diffeomorphism
One interesting extended graphical element that comes from
Lie algebra for self-dual Yang-Mills theory.40–43 In the spinor-
standard quantum field theory is the topological surface opera-
helicity formalism,37 scattering amplitudes are investigated
tor. Let us demonstrate it briefly in the case of Lorentz-covariant
from symmetry principles. Imagine if you can touch a Feynman
scalar field theory. Classically, energy-momentum is locally con-
diagram (should be felt like a steel wire or something else).
served: ∂µ T µν [φ](x) = 0, where T µν [φ](x) is the stress-energy
Grab one of its arms and twist it one direction as if you are
tensor
R for a field φ(x). This leads to a global conservation:
twiddling a knob. Then, the amplitude somehow “reacts” to
− d3 x T 0ν [φ](x0 , ~x) is a conserved quantity R (remains the same this transformation (called the little group action). Doing a
as x0 changes). In fact, P ν [φ](S) := − S d3 Σµ T µν [φ](x) for
kind of dimensional analysis on the little group weight of each
an arbitrary exact hypersurface S equals to zero. Now, the
arm, you can deduce the form of the scattering amplitude! It
quantum version of the local conservation of energy-momentum
would be interesting to see how these approaches will change
is the Ward identity22 ∂x∂ µ hT µν (x) O1 (x1 ) · · · On (xn )i =
Pn our understandings on quantum field theory.
j=1 −δ
(4)
(x − xn ) ∂x∂jν hO1 (x1 ) · · · On (xn )i, where the symbol
h i stands for averaging over all quantum fluctuations. Then, Bonus Cuts
it turns out that hP ν(∂W)O(x)i = ∂ ν hO(x)i holds for any vol- Artwork Inspired by Graphical Notations
ume W that contains x. (When multiple operators are inserted,
the differentiation only applies for operators put inside W.) To
“draw” this identity, we have the following.
 
ν
= ∂ (51)

Wait, this is the “differentiation balloon!” (If you feel the tail
“ ” somewhat artificial, we could have drawn the picture more
“descriptively”:
   
= − . (52)

The “hairs” on the surface are the Killing vector field ξ µ ∂µ = Joon-Hwi Kim, A Pleasant Dream II, 2016.
µ
δνµ ∂µ . If we were considering rotations ξ µ ∂µ = −2(x[α δβ] )∂µ , Digital printing on canvas.
 13

II. Solutions
A5 (a) = .
A B C D D C A B
A. Graphical Vector Algebra
(b) =− A B .
A B
A
A1 (a) . Depending on grouping the three boxes, you A C
B C (c) = .
~ C×~ B C A B
can read off this diagram various ways, such as B·(
~ and C
A) ~ · (A~ × B).
~ Reading in clockwise direction ~ × (B
A6 (a) A ~ × C)
~ +B ~ × (C~ × A)
~ +C ~ × (A
~ × B)
~ = 0.
~ ~ ~
gives −A×(C × B), −B ~ ×(A×
~ C),
~ and −C ~ ×(B~ × A).
~ (b) imj Am jnk Bn = Ak Bi − δik Al Bl .
C A7 In the graphical notation, do the “clank” procedure. In
(b) . Reading the two cross product machines
A B the plaintext notation, rearrange terms, relabel dummy in-
anti-clockwise or clockwise, this can be read off by dices, and then permute two indices of the epsilon tensor.
~ × (B
A ~ × C),
~ −(B ~ × C)
~ × A,~ −A ~ × (C
~ × B),
~ and
~ × B)
(C ~ × A.
~ = = − = 0
A A A A A A
D C ijk Ak Aj (53)
(c) . Different groupings give different readings. ijk Aj Ak = = −ijk Aj Ak = 0

=
A B ikj Aj Ak
~ and D
A ~ versus B~ and C~ gives (D~ × A)
~ · (B
~ × C).
~ A ~


~ (B
versus the others gives A· ~ × C)×
~ D ~ . B~ versus the A8 Use − = − and ~n · ~n = 1.
others, C~ versus the others, and D ~ versus the others
are also possible. In addition, reading the cross prod- − = −
uct machines anti-clockwise or clockwise gives further n n n n n n
disguises of an identical expression. = − n n (54)
A C −ilk nl kmj nm = δij δlm nl nm − δim δlj nl nm
A2 (a) ; jlk Ak (−lmn Cm Bn ).
j B = δij − ni nj
A
(b) ; Al Bk lki . A9 As = − n n , = −2 n n +
B i
n n n n = − n n = ; is idempo-
B C tent.
(c) ; Bi (−ijk (jlm Cm )Ak ).
A l A10 From the given “bones” = − , we find Eqs. (55)
and (56).
(d) ; (−ijl Al )(−jkm Bm )δki .
A B
B A B A B A
B = −
(e) A ; C D C D C D
C (55)
l
Ai imn jnm jkl Ck Bl , ijk Ci Bj (klm mln An ), etc.
~ B)·(
(A× ~ C ~ × D)
~ = A·
~ C~B~ ·D
~ − A·
~ D~B
~ ·C
~
A3 (a) i A B j = Ai Bj .
(b) Ak Bk δij , labelling the two ends of the Kronecker delta A A A
= −
i and j, respectively. B C B C B C
~ · (B
~ × C),
~ etc. (56)
(c) Ai Bj Ck ijk or A l
(d) Ai Bl ijk ljk , etc. ~ B
A×( ~ × C)
~ = B(
~ A·
~ C)
~ − C(
~ A·
~ B)
~
~ ~ ~
(e) i. Aj Bk Cm DnEo ijk iol lmn , or [(( A × B) × E) ×
~ · D,
C] ~ E ~ · (C~ × D)
~ × (A ~ × B)
~ , etc. Eq. (55) is a scalar equation, while Eq. (56) is a vector
equation. We can “peel off” to obtain identities of two,
ii. Aj Cm Dn Eo ijk iol lmn , labelling the end of the three, and four free terminals. The resulting four-terminal
diagram by k. Or, (E ~ × (C~ × D))
~ × A,~ etc. equation is just = − . The three-terminal equation
iii. Aj Bk Cm Dn ijk iol lmn , labelling the end of the is what we get when a single A ~ at one of the terminals of
diagram by o. Or, (C ~ × D)
~ × (A ~ × B),
~ etc.
= − is attached, which is not that interesting. In
(f) δij δij , δii , etc. case of two-terminal equations, there are two possibilities.
(g) ijk δjk , labelling the end of the diagram by i.
= B A − A B (57)
(h) ijk (−ijk ), etc.
A B
(i) ijk jki , etc.
A D A D = B A − (58)
A4 Different groupings of a diagram , and A B A B
B C B C
A D
~ × B)
, gives (A ~ · (C
~ × D)
~ and A
~ · (B
~ × (C
~ × D)),
~ The former relates antisymmetrization with epsilon
B C tensors. The later describes matrix product of Hodge dual
respectively. of two vectors.
 14

A11 (a) Extract the “bones” of the Jacobi identity. Then, (2, 1, 1), because it is equal to zero. Now, the remaining
what we have to prove is that ones are , , and .

+ + (59) A14 and are one-loop diagrams; is a tree-level dia-

gram.
is equal to zero. Following the instruction given in A15 Tree-level diagrams cannot be a 1PI diagram. Also, it
the problem, we find can be easily checked that all epsilon networks that have
self-connections are not 1PI diagrams. Therefore, among
− + − + − (60) the connected nonzero diagrams we have obtained in P13,
we only have as a 1PI diagram. On the other hand,
= − + − + − (61) (2, 1, 1), , which vanishes (equals zero) so that we have
=0. (62) excluded in P13, is also a 1PI diagram. Thus, the answer
is and . The first one reduces into − as
(b) We can obtain zero, one, two, and three-terminal
= − = − . The later reduces into 0 , as it is
identities by attaching the “flesh pieces.” The
equal to − = 0 − 0.
one-terminal identity is P6a itself, that is,
A~ × (B
~ × C)~ +B ~ × (C ~ × A)
~ +C ~ × (A~ × B)
~ = 0. Con-
A16 − =− = .
tracting this with a vector D~ gives the zero-terminal
 
identity: A~ · (B~ × C)~ ×D ~ +B ~ · (C~ × A)
~ ×D ~ + A17 From = − = − , we obtain
 
C~ · (A ~ × B)
~ ×D ~ = 0. In case of two-terminal − + − + = 0. As we have practiced in
identity, it is written graphically as Section I A 2, we can obtain zero to five terminal identities
by attaching the “flesh pieces.” The most convenient to
− = . (63) denote in the plaintext notation is zero and one-terminal
A B A B A B identities. The one-terminal identity reads as follows.
You will later meet this identity again: this is the C B C B C B C B
commutation relation of the so(3) algebra. The D A D A D A D A
= − + (65)
three-terminal identity reads as follows.

A A
+ + = 0 (64) This is translated to the plaintext notation as
A ~ B
(A· ~ × C)
~ D~ = A(
~ B
~ ·C
~ × D)
~ − B(~ A·
~ C ~ × D)
~ + C(~ A·
~ B~ × D).
~
Contracting this with another vector E, ~ we obtain
This will be interpreted as the invariance of ijk with
also a zero-terminal identity: (D ~ · E)(
~ A ~ ·B~ × C)~ =
respect to infinitesimal rotation.
~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~
(A · E)(B · C × D) − (B · E)(A · C × D) + (C · E)(A · B × D).
A12 (a) = − = −2 (this identity is in fact men- A18 Using diagrams,
tioned in the main article). Note that self-contracting  
a cross product machine gives a trivial identity, 0 = 0. 1
− = 3! − − (66)
3
(b) = − = 3 − 32 = −6.  
1
(c) Starting from ijk mnk = δim δjn − δin δjm , we found = 3! 3 − − (67)
3
that ijk mjk = δim δjj − δij δjm = 3δim − δim = 2δim
and ijk ijk = 2δii = 6. = 2! = − . (68)
A13 The original purpose of this problem is to let students ex- k i j
ploratively find tensor identities while playing with dia- In the plaintext notation, kmn kij = 3!δ[k δm δn] =
j k i j i j j
grams; however, we approach this problem in a top-down 3! 13 (δkk δ[m
i
δn] + δm δ[n δk] + δnk δ[k δm] ) = 3! 13 (3δ[m
i
δn] +
manner. Consider three cross product machines gathered: i j i j i j i j
δ[n δm] + δ[n δm] ) = 2!δ[m δn] = δm δn − δni δm
j
. (Or, it can be
. We would like to assemble these in a single connected
piece. Let the number of connecting lines between two dif- worked out by expanding all the antisymmetrizers.) It is
ferent cores to be a, b, and c, respectively. Surely, any of often quick to use the graphical notation, especially when
a + b, b + c, and c + a should not exceed three, because one we have to permute many indices.
cross product machine only has three arms. Without loss of A19 By the basic property of antisymmetrizers, =
generality, let a ≥ b ≥ c. Then, the possible combinations h i
1
are (a, b, c) = (3, 0, 0), (2, 1, 1), (2, 1, 0), (2, 0, 0), (1, 1, 0), 4 − + − . The left hand side is zero.
(1, 0, 0), and (0, 0, 0): , , , ··· . (There can be Adjoining three upper-right
h arms by the crossiproduct ma-
also self-connections other than mutual connections: such chine gives = 14 − + − .
as , etc. However, self-connections always give zero,
so they are not our interest.) Rule out (3, 0, 0), (1, 0, 0), The left hand side is zero as mentioned. Also, note
and (0, 0, 0), because they are disconnected. Rule out also that the cross product machine is already totally an-
 15

tisymmetric so that antisymmetrization of its three in- diagram vanishes.

dices has no effect on it: = . This leads to This is a consequence of rotational symmetry. An epsilon
− + − = 0 , which is identical to the network must be rotationally invariant because it is com-
posed of cross product machines, which are rotationally in-
“bone” identity in A17.

variant. (“Rotationally invariant” means that the “arrow-
A20 + = , but + = 13 + + 6= . Instead, heads” propagate properly along the network; refer to Sec-
= + 43 + 43 + ; you may want to check it. This tion I C 3.) However, there are no (nonzero) rotationally
is the identity for irreducibly decomposing rank-3 tensors. invariant tensors that have one terminal: if it existed, it
means that it discriminates a particular direction in space.
A21 Based on the presented argument, we only need to fix Thus, no nonzero tadpole epsilon networks can exist.
the proportionality constant c, where = c . To find

the value of c, contract all indices as = c .

B. Graphical Vector Calculus
Then, we have −6 at the left hand side and c/3! times the
following at the right hand side.
A25 (a) f f A + f A
A =
+ + − − − (69) (74)
l
3 2 2 2
=3 +3+3−3 −3 −3 =6 (70) ~ =
∇ · (f A) ~
∇f · A + ~
f ∇· A

Thus, −6 = (c/3!) × 6 = c.
(b)
A22 From rotational invariance and permutation symmetry, = +
The two-terminal expression must be proportional to f A f A f A
, so let =c . Connecting the two terminals (75)
l
yields = c ; −6 = 3c; c = −2.
~ ) = ∇f × A
∇ × ( fA ~ + f ∇× A
~
A23 First, we should identify what permutation symmetry
has. Swapping two of its indices, we find
A26 A A A

= − = −(−)2 = (−)3 . (71) = −

B B B (76)
To flip the whole diagram , we need to flip three cross
l
product machines individually; thus the (−)3 factor. ∂Aj ∂Bm ∂Aj ∂Bj ∂Aj ∂Bi
Considering its cyclic symmetry, we conclude that is ∂xi ∂xl ijk lmk = ∂xi ∂xi − ∂xi ∂xj
totally antisymmetric in its three indices. Therefore, it
must be proportional to the “unit” totally antisymmetric
rank-3 tensor: C B C B C B C B
A27 (a) A = A + A + A .
:= c . (72)

Attaching another cross-product machine gives (b) Using the = − trick,

= c = −6c . (73) C C C
A A A
B = B − B (77)

Provided that = 6, we can prove that c = −1. (Note:

the value of is equal to the number of ways to color C C
the edges of the graph with three colors so that distinct A A
= B + B . (78)
colors meet at each vertex.17 To see why, call the three
colors “x,” “y,” and “z” and remind that there lies ijk at
each vertex with xyz = 1.)
n 1 r 1 1 n 1
A28 = r = r − r2
r = r .
A24 Tadpole diagrams are diagrams such as , , etc. Note
that at the tail of tadpole diagrams there always sits a ~n at a position ~r does not change when ~r makes a radial
cross product machine. displacement. When ~r is infinitesimally varied along a tan-
Consider a tadpole diagram. It can be split into one cross gential direction, i.e., the infinitesimal displacement d~r is
product machine that carries its tail and the rest of it, perpendicular to ~r, ~n rotates by an angle 1r |d~r | so that it
which has two free terminals so that is proportional to changes by d~n = 1r d~r. Therefore, dni = 1r Pij dxj , where
Kronecker delta by Schur’s lemma. Thus, any tadpole di- Pij , the ~n-projector δij − ni nj , ensures that ~n only react
agram is proportional to , which is zero: any tadpole to tangential displacement.
 16

A29 n = 4πδ (3) (~r) .

r 2 A B = A B − A B (82)

A30 (a) As mentioned in the main article, = 0 holds as

To sum up, ∇2 (A ~ · B)
~ is equal to A
~ · ∇2 B
~ −B ~ · ∇2 A~
an operator identity. Feeding a scalar field f (~r) and plus the double of Eq. (82)
 which, as a total, result in

~ r) gives
a vector field A(~ ~ 2B
A·∇ ~ −B·∇
~ 2 A+2
~ ~ · ∇)A
∇· (B ~ −B ~ × (−∇ × A) ~ .
This proves the equation given by the problem.
f A
= 0 and = 0, (79) A32 Recall that r = and r = n . Also,
1


r vanishes, and n = r − n n
respectively. The former proves the first identity

1
of the problem. The latter is rarely mentioned in := r .
the literature, but if we connect the two ends with
1 1 2
A A (a) r = n = r = r (3 − 1) = r .
Kronecker delta, we have 0 = = ,
The fact that the trace of is equal to 2 reflects
which translates into the plaintext notation as its dimensionality: it projects vectors to a two-
~ = 0. Note that if we use a cross
∇ · (∇ × A) dimensional subspace.
product machine instead of Kronecker delta, we find
For ∇2~n, write n = 1
r
r and then apply
A ~
that = 0 , which translates into ∇(∇ · A)−
the Leibniz rule:
∂i ∇(Ai ) = 0.
1 1
=2 r
r + r
r (83)
A A A A A
(b) = − = − . (3) ((((
= − r22 n
(
−(
4πδ
(((~ r) r . (84)

Thus, −r2 ∇2~n = 2~n . This “2” originates from

A31 (a) f g = f g + f g
`(` + 1), ` = 1.
(b) ∇2 z = ∇2 (~ez · ~r), and ~ez , a constant vector, can per-
= f g +2 f g + f g . meate through the differentiation loop. Therefore,
since ∇2~r = 0, ∇2 z = 0 .
(b) f A = f A + f A
ez

= f A +2 f A + f A . ez r = r =0 (85)

~ · B)
(c) ∇2 (A ~ = A B = A B + A B .
(c) Use the same strategy with the previous problem.
One can proceed in the same way as the former First, consider
problems to get

r r . (86)
A B +2 A B + A B . (80)

In addition, let us try to obtain an expression that is Contracting the two terminals of this expression with
also tractable in the index-free plaintext notation. gives ∇2 (r2 ). How about using ex ey ? This
From  
gives ∇2 (xy). ex ex − ey ey gives ∇2 (x2 − y 2 ).
 
A B = A B − A B , (81) Lastly, 3 ez ez − gives ∇2 (3z 2 − r2 ). Hence,
calculating Eq. (86) will provide all answers for
this question. It proceeds by applying the Leibniz
~ · B)
~ = A rule. Since ∇2~r = 0, the only surviving term is the
∇2 ( A B − A B +2 A B . The
cross terms (i.e., “one differentiation per one ~r ”),
last term can be written differently employing the 2 r , which is equal to 2 .
r
= − trick.
 17

Thus, ∇2 (r2 ) = 2 = 6 . (This “6” originates from A34 The proof proceeds in two steps. First,
`(` + 1), ` = 2.)
 For the cases of contracting
 Eq. (86)

e
with x y ,e e x e e
x − y e y e
, and 3 z z −e ,
u u u
you can easily check that they are zero! Therefore, = − (99)
∇2 (3z 2 − r2 ), ∇2 (x2 − y 2 ), ∇2 (xy) = 0 .
u u u
(d) In the same way, we are interested in

1
r r r = − u , (100)
, (87) 2 u2
ω

which is equal to ~ := ∇ × ~u. Then,

where ω

2 r r r +2 r r r +2 r r r − = − (101)
u ω u ω u ω


= 6 average of all permutations of r
= + − − . (102)
u ω u ω u ω u ω
=6 r
. (88)


Then,
we can obtain ∇2 z(5z 2 − 3r2 ) , ∇2 x(5z 2
−r2 ) , etc. by plugging in appropriate combination A35 = .
··· ···
of basis vectors.


∇2 z(5z 2 − 3r2 ) (89)
! A36 Provided that ψ = −k 2 ψ , the proof proceeds as the
ez ez ez ez
=6 5 −3 (90) following.
r r
!
ez ez ez ez ez
=6 5 − −2 (91)
r r r r ψ = r ψ +2 r ψ + r ψ

= 6 5z − 3z − 2z) = 0 (92)
(103)
Untying symmetrizers into all possible permutations
gives transition from Eq. (90) to Eq. (91). Note that =0+2 ψ + r = −k 2 r ψ (104)
ψ
ez ez ez ez ez ez
= . Similarly,

∇2 x(5z 2 − r2 ) (93)
!
ex ez ez ex C. Graphical Tensor Calculus
=6 5 − (94)
r r
! A37 When the two arms of the shaded expression in the mid-
ez ez ex ex ex
=2 5 − −2 (95) dle of Eq. (19) is swapped as I → I then yanked, it
r r r
rotates by 180°. As the resulting expression is identical to
= 2(5x − 3x − 2x) = 0 . (96)
the expression before swap-then-yanking, the inertia ten-
2 2 2

sor is symmetric. Or, one can expand the expression by
In fact, you can
2 2 2

check that ∇ z(x − y ) and = − then apply the swap-then-yanking.
∇ x(x − 3y ) are also zero. Thus, the answers to  
Z
this problem is 0 . 1 r
A38 dm 3 − r2 .
A33 Letting ~u = ∇Φ, 2 r
Z
A39 The trace of the inertia tensor reads: dm r r =
Z
= (97)
u u Φ Φ dm 2r2 (refer to P55a). For the case of mass quadrupole
Z  
1 r
1 moment, it is traceless: dm 3 − r 2 =
= = . (98) Z   2 r
Φ Φ 2 Φ Φ
1
dm 3 r2 − 3 r2 = 0.
2
 18
     
1 r 1 r 1
A40 3 − r2 = 3 − 3 r2 + 2 r2 . The first
2 r 2 r 2 →
− µ H =0
∇ · µH = 0 ↔ , (107)
term is equal to − 32 r
r and give − 32 Iij
when integrated
R
by dm. The second term translates into r2 δij , and gives ∂ → − →
−
R ∂t µ H +∇× E =0
1 2
2 δij Ikk when integrated, as dm 2r = Iii shown in the
l
previous problem. (108)
A41 It surely is an “enormous” tensorial expression, but can µ E
H + = 0.
be wisely reduced further using the graphical notation.
First, observe that we can write the bracketed term in
n Then, consider the remaining two. Letting ρfree = 0
→
−
Eq. (20) as . This is easily verified by = − and J free = 0,
n Q n
~ is equal to
and the fact that ~n is a unit vector. Then, S
6 2 4 −2 →
− ε E = 0,
(µ0 ω /288π c )r times the real part of ∇ · εE = 0 ↔ (109)

∂ → − →
−
− ∂t ε E +∇×H =0
n l (110)
= n n n . (105)
Q n H
I II III − ε E + = 0.
n Q∗ n Q∗ Q
n n n

(Or, if you want, using the plaintext notation, Si = →

− −→ →
− →
−
(µ0 ω 6 /288π 2 c4 )r−2 ijl jhk kab lde nh na nc nd nf Re Q∗bc Qef .) (b) For a plane wave solution E = E0 e−iωt+i k · x and
→
− −→ →
− → −
Here, the asterisk in Q means that it is complex conju- H = H0 e−iωt+i k · x , Eqs. (107) and (109) reads
gated. One may wonder why we rewrote the bracketed
term in Eq. (20) involving two cross products, increasing k ε E0 = 0 and k µ H0 = 0 , (111)
the net number of cross product machines. In fact, it
turns out that it was a wise prescription that effectively respectively. Next, Eqs. (108) and (110) reads as
reduces the required steps. We have three options for follows, respectively.
applying the “ = − ” identity; among options I,
II, and III marked in Eq. (105), what “knot” is the best to − µ iω H0 = E0 , (112)
cut off first? It is the option III, because the “ ” term
i k
vanishes by ~n × ~n = 0.
ε iω E0 = H0 . (113)
n i k
n
It is evident that Eq. (111) comes from Eqs. (112)
and (113) by contracting with k . Therefore, one
See the two ~n’s that are plugged into a cross product ma- can focus only on Eqs. (112) and (113). Using the
chine? Thus, only the “ ” term survives, and we obtain 1
definitions u = k k and v = ω/k, they boil
down to the following eigenvalue problems.

n n n n n ε−1 µ−1 E0 = −v 2 E0 , (114)

− = − n . (106) u u
Q∗ Q Q∗ Q
n n n n µ−1 ε−1 H0 = −v 2 H0 . (115)
u u
(One may further expand the here.) Therefore, the
answer is When E ~ 0 and ~u := ~k/k, are known, v 2 can be
obtained using Eq. (114); contracting the entire
(µ0 ω 6 /288π 2 c4 )r−2 times the real part of Eq. (106).
equation with E0 ,
A42 It is not necessary to write equations in a completely graph-
ical manner. However, if some want to do so, they can E0 ε−1 µ−1 E0 = −v 2 E0 E0 . (116)
denote ∂/∂t as a balloon, of which shape is distinguished u u
from that of spatial derivative ∇, such as .
Or, for a more symmetric form, the permittivity ten-
(a) Start with the equations that do not couple with sor can be moved to the right hand side of Eq. (114)
→
− −→ →
− −→
sources. Substituting D = εE and B = µH,25 (by ε ε−1 = ) before the contraction:
 19

notation,
u E0  
µ−1 W W W
E0 u 1 1
v = 2
. (117) = − (123)
E0 ε E0 −2! W 2 W W
 
1
2
= W − W W (124)
2
(c) From Eq. (114),
1

2

= tr W − tr W 2 . (125)
2
Wij = i ε−1 µ−1 j . (118)


u u Writing tr W 2 in terms of →−u and the permittiv-
ity and permeability matrices is straightforward.
Note: if
u
Nij := i ε−1 j and Mij := i µ−1 j , (119) −1

µ ε−1
u u tr W 2 = u u (126)
ε−1 µ−1
W = N M (i.e., Wij = Nik Mkj ). W gives the u
−→
−
→
eigenvalue problem of E0 : W + v 2 1 E0 = 0. On
1
the other hand, one can find the matrix that gives iii. det W = 3! ijk lmn Wil Wjm Wkn .
−→
the eigenvalue problem of H0 = M N . Provided that
As W = ε−1 ?u µ−1 ?u,
N is invertible, W = N M and M N are related by
similarity transformation so that they have the same


2

eigenvalues. det W = det ?u / det εµ . (127)



1 1 However, det ?u = 0, so det W = 0. This is
(d) Wii , 2! ijk imn Wjm Wkn , and 3! ijk lmn Wil Wjm Wkn →
− →
−
appear as coefficients in the characteristic equation. because the map V 7→ ?u V is non-invertible, as
→
− →
− →
−
?u u = u × u = 0. Or, by explicit calculation,
We refer interested readers to Peterson’s article.27,28

u u u


i. tr W = Wii = = − (128)
u u u u u u
ε−1
ε−1 µ−1
= u u . (120) = 0 − 0 = 0. (129)
−1
u u µ

A43 Rik Rjl = ji k

. Joining k and l gives = .
This is simple enough; however, if you want to l

write Joining i and j gives = . These two corre-

 this in the index-free plaintext notation, spond to Rik Rjk = δij and Rjk Rjl = δkl , respectively.
tr ε−1 ?u µ−1 ?u can be a choice (refer to
A49b for the ? notation). Or, try regarding it A44 (a) Observe that = , if you “slide” the arrowheads
→
−
as a matrix “ β ” sandwitched with two u ’s : upwards. Generally, attaching identical rank-2 ten-
←
− → −
u β u . The matrix β can be massaged as sors and then permuting the indices gives the same
result with permuting the indices first and then at-
ε taching identical rank-2 tensors: e.g., = =
1 ε

, (121) − . Any two indices are antisymmetric in the same
−2! det ε µ−1
way. Thus, is totally antisymmetric so that is
using the expression for the inverse matrix equal to .
((ε−1 )ij = det1(ε) 2!
1
ijk lmn εjm εkn ). Untying the (b) By Eq. (3), = 1
= − 3! . By the definition
in both sides by = − , we have
 of determinant, this is equal to .
1  

εµ−1 ε − ε tr µ−1 ε , (122) (c) Provided that = , =
det ε
= , = = ,
8
which appears in Taouk’s work. and = = . The given

1
graphical equation demonstrates these four situa-
ii. adj(W ) il = 2! ijk lmn Wjm Wkn is called the tions.
adjugate matrix of W . We want to calculate
1 A45 (a) = −2 = −2 = 4 .
its trace, 2! ijk imn Wjm Wkn . In the graphical
 20

rank-1 tensor, i.e., a vector. Therefore, an antisym-

(b) = = = − . metric rank-2 tensor, say, Aij , can be encoded in a
1
A A A A vector (?A)i := 2! ijk Ajk without loss of information
so that it can be inverted as Aij = ijk (?A)k (the
1 1
(c) = = = − . normalization factors 2! and 1! are conventional).
In other words, we can always convert a bivector (a
directed area) into a vector (a directed line) by “the
(d) = − right hand” (ijk or ?) and vice versa.
ANTISYMMETRIC
= − = − .
1
2! 1
=  (132)
A A − 2! A
A46 = = .

(c) If T = A B , = A B = A B =
A47 Provided that the integration measure is invariant, the fol- T
lowing confirms that the mass quadrupole moment is a
~ B,
A· ~ and 1
= 1 ~ B.
= 12 A× ~ These are the
2! 2!
r r T A B
r
r r r “invariant products” of two vectors: scalar and vec-
tensor: 3 r
− −→ 3 r
− = 3 r
− .
r tor products. (Therefore, the invariant products are
r r
in general irreducible parts of direct product of two
objects. Then, we can also think about the “sym-
A48 A B → A B = A B ; metric traceless product” of two vectors, say, A ~ ∩ B,~
~ B)
defined as (A∩ ~ ij := Ai Bj − 1 δij Ak Bk . ∇∩~u(~r) will
A B
→ A B
= A B
. 3
be related to the “shear”29 of a vector field ~u(~r).30 )
A49 (a) When T → T , tr T = → A50 First, a rank-2 tensor Tij can be divided into its an-
T tisymmetric part and the symmetric part. The former
gives Eq. (130), which has three independent components.
= = . It does not change; it is
T T T
The latter, , can be further decomposed into irre-
a scalar.
ducible parts, because we can extract out a scalar from it:
(b) When T → T ,
1
= = tr T . Let the scalar projector be N . A
1
projector must be idempotent; as N2 = 3 N12 , N = 3.
→ = (130)
T T T Hence, we have the decomposition Eq. (26).32
A51 Our definition of the inertia tensor is a machine that has
(permuting indices and dressing indices with the same one input for an angular velocity vector and one out-
matrices commute: see A44a). It dresses outward put for giving the corresponding angular momentum vec-
arrowheads at each of its two indices; it is a rank-2 tor (Eq. (19)). Thus, it is a (11 )-tensor. In the graphi-
tensor. Since it is antisymmetric (T[ij] = −T[ji] ), cal notation that distinguishes contravariant and covariant
3!
it has only (32 ) = 2!1! = 3 independent components indices,
(number of ordered pairs (i, j) such that i < j and
i, j ∈ {1, 2, 3}): T[23] , T[31] , and T[12] .
r
Meanwhile, a vector also has (31 ) = 3 independent I = dm . (133)
components. This suggests that an antisymmetric r
rank-2 tensor can be reformatted into a vector,
preserving the contained information. This can be
achieved by employing the epsilon tensor, because A52 . = = 3! (134)
it is antisymmetric in its two indices and has three
indices so that can convert a two-index quantity = = 3! = 1!2! (135)
to a one-index quantity and vice versa. Use the
substitution = − 12 . Then, Eq. (130) reads:
= 3! = 2!1! (136)

1 1 1 = 3!0! (137)
− =− =− . (131)
2 2 2
T T T A53 (a) A geometric reasoning for a moment reveals that a
vector ~r is transformed into ~r 0 = ~r + ~n × ~r under
Note that although the whole expression is a rank-2 an infinitesimal rotation of angle  with respect to
tensor, the purple-shaded area transforms like a axis ~n. From Rij xj = x0i = xi +  ijk nj xk + O(2 ),
 21

we obtain Rij = δij +  ijk nj , which translates into

(b) = = = = = = · · · (the la-
= + + O(2 ).
n, n

(b) Substituting the result obtained in the previous bel “n, α” omitted to avoid clutter); this little one-
problem, 0 is equal to dimensional creature living on a golden braided knot
n, n,

proliferates eternally!12 This means that is a dis-
+ + 0
+ 2 0
(138)
n n n n tinguished vector: an eigenvector with eigenvalue one
of the rotation , i.e. it remains the same after being
up to O(2 ). We can see that infinitesimal ro-
tation matrices are additive up to O(1 ). Then, rotated. Therefore, we conclude that it is parallel to
− 0 is order of 2 : the rotation axis.
0
n, n, n, n,

Using Rodrigues’ formula, we can explicitly calcu-
2 −1
 0
− 0
= 2 , (139) late it in terms of ~n and α: n =
2 sin α
n n n n

n n0 (or, = − = 3 − tr =
2
3 − (1 + 2 cos 2α) = 4 sin α.) This provides a formula
where we used the Jacobi identity, Eq. (63), at the for finding the corresponding rotation axis when a ro-
last step. tation matrix is given.
A54 (a) Since it is an antisymmetric rank-2 tensor, it lives

in the rank-2, spin-1 (antisymmetric) representation: (c) Since tr = 1 + 2 cos
α, the spin-2 part of
generators of SO(3), the elements of so(3), are real is equal to 12 + − (1 + 2 cos α)/3 =


antisymmetric matrices. (1 − cos α) n n − .
(b) From completeness, we know that ex ex +
ey ey + ez ez = . Thus, the given A58 At r > 0,
equation is equal to − = 2 ; the
quadratic Casimir is 2. = −3r−4 + r−3 (144)
r−3 r n r

B  
B
A55 (a) A = 0; B A = −2 B A ; A =− A . = −r−3 3 − . (145)
C n n
C

0
1
2
(b) = , and = . = This confirms Eq. (29) except for the delta function term.
n n n n
(Note that Eq. (145) is traceless and symmetric; it is
− , as calculated in P8. Now, recall that spin-2. There is no spin-1 part because ∇ × (~n/r2 ) = 0.
, a shorthand for − n n , selects Recall that ∂[i (nj] /r2 ) and ∇ × (~n/r2 ) are compatible, as
components that are orthogonal to ~n; therefore, explained in A49b.) However, to examine whether there

3
= − = − . This is proportional to sits a delta function at the origin, we have to investigate
n n n the behavior of ~n/r2 at r → 0, while it is well-known that

1
: the sequence is periodic. Hence, ∇ · (~n/r2 ), i.e., trace of ∂i (nj /r2 ), is 4πδ (3) (~r ). Assume
n
that the singular term of ∂i (nj /r2 ) is cij δ (3) (~r ), where

4k+1
4k+3
=− = (140) cij is a constant tensor. Then, cij should not have a
n n n particular preferred direction, because ~n/r2 is an isotropic

4k+4
4k+2 vector field. Its components must be unchanged under
=− = (141)
n n rotation, i.e., it must be a scalar. Thus, cij only has the
trace-only part, 31 δij ckk . ∇ · (~n/r2 ) = 4πδ (3) (~r ) confirms

0
for any integer k ≥ 0, and = . that ckk = 4π. Therefore, the delta function term is
4π (3)
n
3 δij δ (~r ). To sum up, ∂i (nj /r2 ) has spin-0 and spin-2
A56 From the results of the previous problem, parts but no spin-1 part, and the two parts are the delta
function term and Eq. (145), respectively.


= exp α (142)
n,α n A59 Since p~ and m
~ are constant vectors, they can be freely
α α 3 2 4 overpass the differentiation loop. Proofs of Eq. (27) and
= + ( 1! − 3! + · · ·) + (− α2! + α
4! + · · ·) Eq. (28) proceed as follows.
n
= + sin α + (cos α − 1)
n p p p p
= cos α + sin α + (1 − cos α) n n . (143) 1 n 1 (3) 3 n n −
n − r2
= − δ (r) + (146)
4π 3 4πr3




A57 (a) tr = cos α tr + sin α tr + (1 − p
n,α n

1 (3)
cos α) tr n n = 3 cos α + 0 + (1 − cos α) = 1 + 3 n n − p
= − δ (r) p + (147)
2 cos α. 3 4πr3
 22

"  #
m `(` + 1)
1
2
−∇ −
1 m n 1 n
r2 r
n n n
r2
=− r2
(148)  (152)
4π 4π
(3)
= 4πδ (~r ) .
1 m 1 m m
= − δ (3) (r) − 3
3 − n n n
3 4πr n n A61 Since P` (~n0 · ~n) = P` (~n · ~n0 ) for arbitrary unit vectors ~n
2 (3) 1 m m m and ~n0 , should be the same if the upper terminals
= δ (r) m − 3 − 3 +2 and the lower terminals are swapped. The self-explanatory
3 4πr3 n n n n
design of this is to choose a horizontally symmetric shape.
m 
A62 Consider ` symmetrized ~r ’s: . Taking Lapla-
2 3n n −m
= δ (3) (r) m − (149) r r r
3 4πr3
cian, , to this, as r = 0, we get

A60 Note that, for a scalar field Φ(~r), if, −∇2 Φ(~r ) = δ (3) (~r )  
1
and Φ(~r ) → 0 as r → ∞, then Φ(~r ) = 4πr . Let us
1 (3) 1
denote this as “Φ(~r ) = −∇2 δ (~r ) = 4πr ” in shorthand. 2 r r
+2 r r
+ ··· (153)
If the boundary condition is different, we formally write
“Φ(~r ) = −∇1 (3) 1 1 
2δ (~r ) + −∇ 2 0,” where −∇2 0 means an
2
element in ker[−∇ ] (i.e., a term that vanishes when being ∝ . (154)
subjected to −∇2 ) that is responsible for matching the r r
boundary conditions.
Refer to the calculations in A32. Note that the equality
from Eq. (153) to Eq. (154) holds because the ` lines are
(a) Taking divergence to µ − = f δ (3) (~r ) symmetrized so that all terms in Eq. (153) are

identical.
u p
`
(The proportionality constant will be 2 2 , since the


gives − p = f δ (3) (~r ) , since u = 0. Taking number of ways to connect two lines out of ` is 2` .)

1 1 1 Note that the ` ends of are by definition totally

−∇2 to this equation gives p = f 4πr
+ −∇2 0 = 
1 1 1
− 4πr 2 f n + −∇2 0 .
Since − 4πr
f n → 0 when 2 symmetric. Therefore, taking Laplacian to will
1
r → ∞, the homogeneous solution −∇2 0 should be p0 give the following. r r r
provided that the given boundary condition holds.

(b) We have determined the pressure field, so sub-

∝ = (155)
stitute it into the original equation: µ u r r r r
r r r r
(3) 1
=− f n 1
4πr 2
+ f δ (~r ). Now, take −∇2 to both
1 Therefore, connecting the first and second lower indices
sides. The delta-function term becomes 4πr f . How
about the first term in the right hand side? Note of gives zero; so do any of its two lower indices; so
2 1 1 do any of its upper indices. (But not for one upper index
that −r2 ∇2~n(= −∇ ° ~n) = 2~n; −∇2 (~n/r2 ) = 2 ~n.
and one lower index.) Thus, it is “traceless.”
1 1
Thus, it becomes − 8π f n = − 8πr f . Adding −1
A63 Multipole expansion of |~r − ~r 0 | for |~r 0 | < |~r| gives

the two, we obtain the following. (The tensor in the ∞ ∞ r r r
shaded area is called the Oseen tensor.) X r0` X 1
0
`+1
P` (~
n · ~
n ) = `+1
, (156)
r r
`=0 `=0 n n n
 
1 1 n
u = +
n (150) where r := |~r|, r0 := |~r 0 |, ~r := r~n, and ~r 0 := r0~n0 .
µ 8πr −1
Meanwhile, from Taylor expansion, |~r − ~r 0 | is equal to
f f 
r r r
2
A side note: we found −∇ ° ~n = 2~n fruitful. Its ∞
X (−)`
generalization to arbitrary ` would be 1/r , (157)
"  #  `!
`=0
2
°
−∇ = `(` + 1) (151) −1
n n n n n n i.e., shifting |~r | = 1/r by ~r → ~r − ~r 0 (thus the factor
(−)` ). At first glance, Eq. (158) may be concluded by com-
so that paring Eqs. (156) and (157) according to the order of ~r 0 :
 23

  q e+ e+ q

• 2 15
= 2 15 3
e + · ~n)2
2 (~ − 12 (
~e +
· ~e+)(~n · ~n)

3 8π 3 8π
! 1 n n
= (-) .- traces
(158) q q
r +1 r
= 15 x+iy 2 15
sin2 θ e2iφ = Y2+2 (θ, φ);
8π (− 2r ) =
√
n n n 32π

q e+ e0
However, there is a caveat here: this argument only works
for ~r 6= 0. There can sit singularities at r = 0. To see this, • 2
3
15
4π
consider the case ` = 2. We have n n
q

2 15 3
= e + · ~n)(~e 0 · ~n)
2 (~ − 21 (
~e +
· ~e 0)(~n · ~n)

3 4π
q
2! 3 n n − = 15 x+iy z
4π (− 2r ) r
√
1/r = = , (159)
r3 n n
r3 q
=− 15
8π sin θ cos θ eiφ = Y2+1 (θ, φ).
which is Eq. (29) without the delta function term. In
One can obtain further spherical harmonics in a similar q
fact, Eq. (158) is an “illegal” tensor equation, because
3
the left hand side is totally symmetric and traceless (as manner. Note that the normalization constants — 4π ,
q q
we have shown in P62), while the right hand side is 2 15 2 15
totally symmetric but not traceless. Therefore, we need 3 8π , 3 4π , and all that — can be found algebraically
“−(traces).” (rather than doing integrals R in a typical
R manner) by the
following identity, where d2 Ω = sin θdθ dφ is the solid
  angle integral.
Z
! 1
= (-)
1
- traces (160) hni1 ni2 · · · ni2` i := d2 Ω ni1 ni2 · · · ni2` (161)
r +1 r 4π
n n n 1
= δ(i i · · · δi2`−1 i2` ) (162)
2` + 1 1 2
While ∇2 (1/r) = −4πδ (3) (~r), the “−(traces)” contains
(` − 2)th -order, (` − 4)th -order, · · · derivatives of 4πδ (3) (~r). That is, the spherical average of ni1 ni2 · · · ni` is equal to
For example, for ` = 2, the trace term we should subtract the average of “all possible contractions” of indices i1 to
i2` by Kronecker delta divided by 2` + 1. For example,
from 1/r is 2!31
= 1/r =,
n n is
3which − equal to − 4π
3 δ
(3)
(~r).
r3 r3 1
n n
hni nj i = δij , (163)
3 1
3 
A64 (a) For ` = 2, = 2 − 2 is traceless, as 1 1
3 1 3 1 hni nj nk nl i = (δij δkl + δli δjk + δik δjl ) . (164)
2 − 2 = 2 − 23 = 0. 5 3

For ` = 3, we have = 5
2 − 3
2 and Eqs. (163) and (164) show up in standard textbooks.33,34
5 3 5 35 In the graphical notation,
2 − 2 = 2 − 23 = 0, as 3 =
+ + =3 + + =5 . * + 1
n n = , (165)
3
(b) Let = a4 + a2 + a0 . (Note
* + 1
that a3 and a1 terms cannot appear.) Then, trace- = . (166)
n n n n 5
lessness implies a4 + a2 + a0 = 0.
Doodling diagrams for a moment, one can see that  

1
4
= 5 + 2 + . Using this, one obtains Note that = 3 + + . Then,
2

(a4 + 76 a2 ) + ( 16 a2 + 10
6 a0 ) = 0. Therefore, demanding hY`m ∗ Y`m i = 1
4π ,
a4 : a2 : a0 = 35 : −30 : 3. Since P4 (~n · ~n) = 1, * e+ e+
+ e+ e+
a4 + a2 + a0 = 1. In conclusion, a4 = 35 30
8 , a2 = − 8 , • 1
= |c|
2
=
2
|c| 13 = |c|
2 1
3 4π 3
and a0 = 8 . n n
3 1
q
A65 Recall that = and = − . Also, note =⇒ c := 3
4π .
2 2
+ + − −
that ~e · ~e = 0 and ~e · ~e = 0.
Here, ~e+ = − √12 (~ex − i~ey ) is a dual vector; numer-
q e+ q q ically, ~e+ := (~e+ )∗ . Similarly, ~e0 := (~e 0)∗ = ~ez and
• 3
= 3 +
e · ~n 3 x+iy ~e− := (~e −)∗ = √12 (~ex + i~ey ). These dual basis vectors
4π ~ = 4π (− 2r )
√
4π 0 0
n q satisfy ~e m · ~em = δm
m
(m, m0 = +, 0, −).
=− 3
8π sin θ eiφ = Y1+1 (θ, φ); For Y2+2 and Y2+1 , note that the following holds because
 24

is traceless and symmetric. that we have discussed in P6a, P11, and P16.
Multiplying (i~)2 gives the proposed equation. Then,
* + S
what is its interpretation? Let i~ := .
(167) i i
n n n n Then, we have
" #
1
 S S S S S
=  + + − = i~ (169)
15 
i j j i i j

2
= (168) This is the familiar “[Ŝi , Ŝj ] = i~ijk Ŝk !”
15
S
One can explicitly assure that is the spin op-
3 1 i
Then, using Eq. (168) with = 2 − 2 ,
erator for spin-1 states by calculating its components
* e+ e+ e+ e+ + (“matrix elements”). Sandwiching with ~em0 and ~em ,
2
• |c|
em0 S em
n n n n (Si )m0 m := (170)
i
e+ e+ e+ e+ e+ e+ e+ e+
em0 em
= |c|
2 2
= |c|
2 2 33 = i~ = i~(~e m×~em0 )i . (171)
15 15 2 2 i
q Displaying these numbers in a matrix format,
2 2 3 2 2 15
= |c| 15 ( 2 ) =⇒ c := 3 8π ; m0 = +, 0, − labelling rows and m = +, 0, − labelling
* e+ e0 e+ e0 + columns, we get the following.
2
• |c|  + 0 − 
n n n n   + 1
(S3 )m0 m = ~ 0  0  (172)
e+ e0 e+ e0 e+ e0 e+ e0
2 2 2 2 33 − −1
= |c| 15 = |c| 15 2 2
 + 0 − 
e+ e0 e+ e0   ~ + 1
2 2 3 2 1 2 2 3 21 (S1 )m0 m = √ 0 1 1 (173)
= |c| 15 ( 2 ) 2 = |c| 15 ( 2 ) 2 2 − 1
q
2 15  + 0 − 
=⇒ c := 4π .
3
  ~ + −i
(S2 )m0 m = √ 0  i −i  (174)
Note that the identity Eq. (162) enables us to do spher- 2
ical integrals in a coordinate-free way. It turns out − i
to

be useful in other 
calculations too. For example,
cos4 θ sin2 θ sin2 φ = (~ez · ~n)4 (~ey · ~n)2 = 17 average of (Notice: zeros sitting in the blank spaces, omitted

 1 3
to avoid clutter.) These matrices should be familiar
all possible
 contractions of ~ez ,~ez ,~ez ,~ez ,~ey ,~ey = 17 6! · 1
1
from undergraduate quantum mechanics.
2!4! = 35 , counting the number of ways for the grouping


2 4 2 2
 (b) Explicit calculation with ~e + = − √12 (~ex + i~ey ),

(~ey ,~e y ), (~
e z ,~
e z ), (~
e z ,~
e z) ; cos
 θ sin θ
 sin φ cos φ =
2 2 2 1 1 3
(~ez ·~n) (~ex ·~n) (~ey ·~n) = 7 6! · 3! 2! = 105 , counting 1 ~e 0 = ~ez , and ~e − = √12 (~ex − i~ey ) reveals that
the number of ways for the grouping (~ex ,~ex ), (~ey ,~ey ), i~~ez × ~e + = +~~e +, i~~ez × ~e 0 = 0~~e 0, and

i~~ez × ~e − = −~~e −; ~e +, ~e 0, and ~e − are the spin-1
(~ez ,~ez ) .
eigenstates |1, 1i, |1, 0i, and |1, −1i, respectively.
A66 By “operator,” we mean that i~ maps a vector into (Note that their eigenvalues with respect to the
i
another vector through its terminals drawn horizontally. Casimir operator Ŝ 2 = (i~)2 = 2~2
2
It is one of the possible realizations of the spin operator: (refer to P54b) are 2~ .)
spin operator for spin-1 states (for example, for spin-1/2 “i~~ez ×”= i~ ez is the generator of z-axis ro-
states, we cannot use i~ ).
i tation, while the rotation matrix corresponding
The sudden appearance of ~ here may evoke an impres- to z-axis rotation of
rotation angle α is

given by
α
sion that we are making a stark transition to the quantum = exp α = exp i~ i~ . ~e +, ~e 0,
ez ,α ez ez
world, but in fact it is not a necessary one; one could have
and ~e − are the eigenvectors of the generator of
defined , not i~ , as a spin operator. Usually,
i i z-axis rotation and so are the eigenvectors of z-axis
physicists prefer to carry i~ to let the generators be Hermi- rotation. They gain phase factors eα/i , 1, and e−α/i
tian (observables). In a sense, we are not really introducing under , respectively.
quantum mechanics but doing representation “practice” ez ,α q e+ e0
+1 2 15
(cf. representation theory) of the rotation group. (c) From P65, we know that Y2 (θ, φ) = 3 4π ,
n n
(a) − = is the Jacobi identity where (ni ) = (sin θ cos φ, sin θ sin φ, cos θ). Ro-
i j j i i j
 25

tating Y2+1 (θ, φ) by an infinitesimal angle  with Contracting this with δij and substituting
respect to axis ~ez , we get Y2+1 (θ, φ − ). (Or, in = i~ , we obtain Ŝ 2 = Ŝi Ŝi = δij Ŝi Ŝj
the differential operator language, we are acting i i


(1− ~ez · ~r × ∇) = 1 + ~ ez
i~ · (−i~~ r × ∇) to applied to T .
Y2+1 (θ, φ)) This corresponds to substituting n
 
into n = n . The fact that is a (i~) 2
+ +2 (178)
ez ,− ez , T T T
tensor allows us to “push the arrowheads” as the
following way, where the labels “ez , ” are omitted to In general, we cannot determine the value of Eq. (178)
avoid clutter. i j
for an arbitrary rank-2 tensor Tij = . However,
e+ e0 T
e+ e0 e+ e0
2
its traceless and symmetric part, 3 , is an eigen-
= = (1 − i) (175) T
n n n n state of Ŝ 2 having an eigenvalue 2(2 + 1)~2 = 6~2 .
n n To see this, it suffices to show that the answer of this
Here we used em = em + em +O(2 ) = problem is 6~2 .
ez , ez


1 + (−)m /i em + O(2 ). Thus, Y2+1 (θ, φ)  
+ 2
transforms alike to ~e , as it gains the phase factor (i~) + +2 (179)
(1 − i) when subjected to infinitesimal z-rotation.
 
In fact, one can see this by the explicit formula
= ~2 2 +2 +2 (180)
of Y2+1 (θ, φ) that is proportional to eiφ so that
eiφ → ei(φ−) = (1 − i + O(2 ))eiφ . However,  
the strength of tensor language is that we can = ~2 4 +2 −2 (181)
easily see that this is generalized to arbitrary  
Y`m . By the “arrow pushing,” the transformation
n → n was equivalent to dressing all
= ~2 6 (182)
ez ,
m
~e ’s with outward arrowheads but leaving ~n’s as
they were. Therefore, Y`+1 (θ, φ) must have the same This justifies to call the symmetric traceless part of
+2 a rank-2 tensor by “spin-2 part!”
transformation property with e+ , Y` (θ, φ)
must have the same transformation property with How about the trace-only part? Surely, there are no
e+ rooms for attaching to a scalar, so it will have
, and so on.
e+ i
an eigenvalue 0 for Ŝ 2 . For the antisymmetric part,
(d) Consider an SO(3) transformation of an arbitrary h i
rank-2 tensor T . Each index will be dressed (i~)2 + +2 (183)
with outward arrowheads (rotation matrices): h i
= ~2 2 +2 +2 (184)

→ . If = + i~ + O(2 ) for a h i h i
T T n = ~2 4 +2 +2 = ~2 2 (185)
given fixed unit vector ~n, that is,
  We obtain the eigenvalue 1(1 + 1)~2 ; thus, it is indeed
ni i i
− = + . (176) a “spin-1 part.” Note that rank-1 tensors (vectors)
T T i~ T T
also have an eigenvalue 1(1 + 1)~2 for Ŝ 2 ; the secret
Setting this to zero gives the infinitesimal version underlying here is the arrow pushing = − 12 .
of rotational invariance. (We apologize for using a Considering the infinitesimal version of this arrow
different symbol for denoting the generators of pushing, one can confirm that applying Ŝ 2 in the
i
S upper two indices of = − 12 is identical to
rotation, i~ , instead of the earlier for the
i i applying Ŝ 2 in the middle internal line, which will
sake of visual brevity.) One can think the bracketed
term in the right hand side of Eq. (176) as a result give (i~)2 = 2~2 .
of acting Ŝi to T , because it measures how T
Lastly, if you are interested, try applying Ŝ 2 to the
deviates from itself when infinitesimally rotated.
spin-` representation with δii = = d.
Acting Ŝi then Ŝj to T , we get This will give the Casimir of spin-` representation of
SO(d), `(` + d − 2).
j j j j
i +i +i +i . (177)
T T T T A67 Before we proceed, it is fruitful to consider the following
subdiagram first.
 26

n n  n n n n 
3 3 3 1 3
3 (b) = 2 = 2 2 − 2 = 2 .
= (186)
2 n n n n n n n n n n

Consider two point dipoles µ ~ 1 and µ~ 2 in thermal

Eq. (186) is easily confirmed if one understands that
equilibrium at an inverse temperature β. The dipole
2
3 is the “symmetric traceless projector” (so that moments can orient all directions and the separation
between the two is fixed at R ~ := RN~ . The canonical
2
= ( 23 )2 ) and = . Or, one partition function is equal to the following, where
3
K is a constant that measures the coupling between
can see it by direct calculation: two dipoles (2 µ0 4π
µ1 µ2
if the dipoles are magnetic),
µ ~ 2 := µ2~n0 .
~ 1 := µ1~n, and µ
33 31 31 11  
− − + n0
22 22 22 22 * +
   N 
=
3 3
−
1
. (187) Z = (4π)2 exp
 −βK N

 (192)
2 2 2 ~
n ~
n0
n

h· · ·i~n and h· · ·i~n0 denote spherical averaging with

p m 3 p 3
p respect to solid angle measure corresponding to ~n
(a) = 2 = 2 and ~n0 , respectively. From this partition function, we
can obtain the average force between the two dipoles,
m which will result in Lenard-Jonnes potential.
m
What would be the leading nontrivial term? A
p p typical calculation that employs coordinates θ, φ,
33 31
= 22 − 22 θ0 , and φ0 goes pages long,35 yet obtaining the
next term in such way by hand will be much more
m m
complicated. However, with the aforementioned
= 21 (0 + 33
2 2m
~ × p~) − 31
2 2p
~ ×m
~ = 15
8 m
~ × p~ . algebraic (or combinatoric) spherical integrating and
the graphical notation, even a general formula for
Consider a ball of radius R, uniformly polarized with terms of arbitrary orders is easily obtainable by a
total electric and magnetic dipole moments p~ and “hairband cutting game.”
m.
~ What would be the electromagnetic momentum
of this system?13 The angular integration of this First, consider the β 2 term in the expansion of the
problem can be worked out algebraically by tensor exponential Eq. (192). (Notice that terms of odd
methods. Contribution from
the inside of the ball powers of β vanishes.)
µ0 c2 2µ0 µ0 1
is µ01c2 (− 4πR 3 p
~ ) × ( ~ 4π
4πR3 m)
3
3 R = πR3 6 m ~ × p~.
* n0 n0
For the outside of the ball, where the electric and
2
+
magnetic fields are given by 2 (βK) N N
(4π) (193)
2! N N ~
n ~
n0
p m n n
µ0 c2 µ0
and , (188)
2πr3 2πr3
n n n n !
2
2 (βK) 1 N N
=(4π) (194)
respectively (Eqs. (27) and (28)), we have 2! 32 N N

p
We have already calculated the bracketed term in
m
" Z∞ #" Z # Eq. (194) in the main problem: it was 32 . However,
µ0 1 1 let us revisit it in terms of the “hairband cutting
r2 dr d2 Ω (189)
π r6 4π n n n n
N 3 N
R game.” Recall that = 2 N − 12 . Substituting
N

" # this, the term within the parentheses in Eq. (194)

µ0 1 2 p m expands into four terms. Three of them “breaks the
= (190)
π 3R3 15
N N N N
hairband:” , , and . The remaining
N N N N
µ0 1
 2 15
 µ0 1
= π 3R3 ~ × p~
15 8 m = ~ × p~.
πR3 12 m (191) one does not. If a “hairband” is broken, its value
is 1 (N ~ ·N
~ or (N~ ·N~ )2 ). If it is not, its value is 3
Eq. (168) is used when obtaining Eq. (190). Finally,
µ0 1
the answer is πR 1 µ0
~ × p~ = 4πR ~ × p~. (δii ). Thus, the bracketed term in Eq. (194) is equal
3 ( 6 + 12 )m 3m
 27

−1 3 3 −1
to 1 · ( 23 32 + 2 2 + 2 2 ) + 3 · ( −1 −1
2 2 )=
3+3
4 = 32 .
p r p r p p r r p r
The graphical notation effectively helps doing this = + i~ (196)
“hairband combinatorics”—counting the possibilities i j i j i j
of getting broken or intact hairbands—for arbitrary p p r r p r
orders of β. The resulting formula is the following. = + i~
i j i j
* n0
2`  + p r p r p r p r
N 1 + 21−2` p` (3) = − i~ + i~
= , (195)
N ~
n ~ 0
(2` + 1)2 i j i j i j
n
n p r p r p r p r
 = − i~ + i~
 1  (` = 1)
i j i j i j
p` (d) := P
`−1
` 2j−1
1 + d (d − 2j) (` > 1) p r p r p r
j=1 j = + i~ . (197)
i j i j
p2 (3) = 7, p3 (3) = −71, and so on; Eq. (195) is equal
1 −78
to 212 32 , 214 30 1 6
25 = 24 5 , and 26 49 for ` = 1, 2, and This completes the proof. The last step is due to the
3, respectively. Meanwhile, obtaining the general Jacobi identity (P16).
formula of higher-order terms by integration intro-
ducing angular coordinates is virtually an impossible
task by hands. This example serves as a instructive A70 Connecting the two ends of Eq. (196) by a Kronecker delta,
demonstration for diagrammatic perturbation.

= − (198)
(c) 22
= 2 r p r p r p r p r p r p
33 3

  = r − r − p r − i~
2 3 1 2 r p p i~ p r p r p
= 3 2 − 2 = 32 ( 32 3 +3
2 − 12 3) = 5 .
= r̂2 p̂2 − 2i~ ~rˆ · p~ˆ − p r r p
This is the dimensionality of the spin-2 projector,
2
(recall the combinatorics in P50). = r̂2 p̂2 − 2i~ ~rˆ · p~ˆ − r p r p + i~ r p
3

= r̂2 p̂2 − 2i~ ~rˆ · p~ˆ − (~rˆ · p~ˆ)2 + 3i~ ~rˆ · p~ˆ
= r̂2 p̂2 − (~rˆ · p~ˆ)2 + i~ ~rˆ · p~ˆ (199)
D. Addendum
A71 This instructive problem is based on the so(1, 3) = ∼
su(2) ⊕ su(2) algebra. For more precise treatment that dis-
A68 The canonical commutation relation, [xi , pj ] = i~δij , reads tinguishes upper and lower indices, see, e.g., Srednicki.36
 
r , p = i~ . Let us employ this in a form of Eq. (33), Use completeness first.

= + i~ . S = + = + (200)
r p p r

Then, take the blue trace to eliminate the blue generator.

(a) = + i~ = + 3i~. S = + = d +0 (201)

r p p r p r

Massaging the both sides, we get the answer.

(b) = + i~ =− + 0.
r p p r p r 1
= S
(202)
d
h i
L , L L −2 −2 −2 −2 −2 −2
A69 = i~ .
A72 (a) = 4 , and = =
i j i j

The orbital angular momentum, L̂i = ijk x̂j p̂k = 4 = 4 = 4 . Subtracting the
p r 2 −2
−ijk p̂j x̂k = , is an implementation of this: two gives 4 = −4 .
i
 28

(b) The lesson from P49b is that an I-line can be Eq. (208) is equal to
converted to an i-line by the cross product machine
1
and vice versa. Refer to Eq. (132). The converters − = − ↔ − . (209)
1
were 2! and − ; this is a standard convention c2
in tensor calculus distinguishing contravariant and √ √
covariant indices (up-down hierarchy). Meanwhile, Therefore, c2 = 2; c = 2 or c = − 2. Next, let
i i
a different normalization, √1!2! and √1!2! is us demand that c2 = −2 . Using only
more apt for the “heterarchized” calculus. = d and the definition of the antisymmetrizer,
Then, ijk with each of its indices converted to we have
I-indices perfectly copies the syntax of ijk to the  
1
I-world. As demonstrated in Eq. (205) to Eq. (207), = −2 + (210)
an I-world expression is translated into an i-world 4
epsilon network by the converters and brought back d−2
= . (211)
to the I-world after simplification. 4
√ Therefore, the equation c2 (d − 2) = −8 must be
( √i2 )3 = −( √i2 )3 = 2 2i ; (203) satisfied. In the previous problem, c2 = −8 so that
d = 3 (δii = 3). In case of binors, c2 = 2 so that
d = −2!
( −1
α3 )
2
= ( −1 2 2 5
α3 ) (α ) (204) As d = δaa , it seems that binors are tensors living in
a (−2)-dimensional space: SO(−2)-tensors. In this
= α4 (−2) (205) sense, Penrose17 introduced the idea of “negative
  dimensional tensors.” The reader might find it
= −2α4 − (206) uncomfortable to accept. Practically, this negative
dimensionality can be understood as a consequence
  of “heterarchizing” spinors. The binor language
2
=−2α
 − , (207) can be repackaged into spinor algebra, where the
converter between binors and SO(3)-vectors being
where α := √i2 . The rightmost expression in the Pauli matrices. Also, the distinction between
contravariant and covariant indices is restored, such
Eq. (203) is the answer for this problem with its com- a
plex conjugate, because we could have used √−i as → − = − = −ab ab = ab ba and
1!2! b

and √−i as converters. The resulting IJK sat- = − 12 → = + 21 (ab cd = δca δdb − δda δcb ).
1!2! ba
isfies total antisymmetry IJK = −JIK = · · · and Since ab  = −2, we have a negative d. Meanwhile,
IJK KLM = δIL δJM − δIM δJL . (Also IJK LM N while restoring hierarchy, one observes that sym-
= 3!δ[LI J K
δM δN ] ; check it!) metrization and antisymmetrization changed their
  role. Cvitanović and Kennedy16 discussed on this
1 point and clarified the meaning of tensors of negative
Note that = 2 − not in gen-
dimensions.
eral is proportional to Eq. (207). While transforming Antisymmetrizing two d-dimensional indices, we
Eq. (205) into Eq. (207), we used the fact that is

obtain a collective index of dimensionality d2 . d = 3
equal to 2−1
so that it can be “factorized” into two

produces the self-similar infinite tower: 3, 32 = 3,
tripod totally antisymmetric tensors, which is not (32)
= 3, · · · . (It is the “critical point.”) If we start
true in dimensions other than three. 2
−2

One more comment: we could have determined from binors, we have −2, −2 2 = 3, ( 22 ) = 3, · · · .
Once we entered the three dimensions, we will always
IJK by first setting it to be c then finding the
stay there. Thus, problems (b) and (c) complete the
value of c. Then, following the same calculation in exploration of the stairs of the recursive tower.
Eq. (205) to Eq. (207), the condition for = −
to be reproduced turns out to be c2 = −8. We Bonus Cuts
employ this approach in the next problem. Listen to Tensors

(c) Let ↔ c . c is a constant to be deter-

mined. First, let us impose the condition that it
reproduces = − . Note that
 
1
− = 2 (208)
2

holds generally (cf. the internal wiring in Eq. (43)).

Then, from the binor identity +2 = 0, https://soundcloud.com/joonhwi-kim/listen-to-tensors
 29

20
Riemann Tensor For a clear construction of Feynman diagrams in the functional
calculus language, refer to N. Beisert’s lecture notes.21
21
N. Beisert, Quantum Field Theory II, Lecture Notes, (2014).
Retrieved from http://edu.itp.phys.ethz.ch/fs13/qft2/.
22
D. Simmons-Duffin, arXiv e-prints, (2016), arXiv:0910.1362 [hep-
th].
23
For another example, think about doing braided group calculations
by anyons (“topological quantum computing”).
24
M. C. Escher, Drawing Hands, 1948. Lithograph.
:=

25
This is the “arrowheads as abstract indices” notation that the first
author suggests.26 It is a compromise plan between graphical and
←
−
plaintext notations. One-forms and vectors are denoted as W and
− −
→ ←−−→
V , respectively, and their contraction is written as W V (cf. the
contraction of a bra and a ket hβ|αi). A (11 )-tensor will be denoted
as T , which has “one output and input terminals.” It acts on a
−
→ −
→ −→ → i j
References vector V to give a vector T V = T V = − e i T j V . This notation is
∗ used exclusively in this problem.
joonhwi.kim@gmail.com 26
† J.-H. Kim and J. Nam, arXiv e-prints, (2019), arXiv:1912.11485
maverick.sh.oh@gmail.com
‡ [physics.ed-ph].
https://sites.google.com/view/gctpb202/; fortoe@gist.ac.kr 27
1 E. Peterson, arXiv e-prints, (2009), arXiv:0910.1362 [math.HO].
R. Penrose, in Quantum Theory and Beyond: Essays and Discus- 28
E. Peterson, arXiv e-prints, (2007), arXiv:0712.2058 [math.HO].
sions Arising from a Colloquium, edited by T. Bastin (Cambridge 29
J. D. Romano and R. H. Price, American Journal of Physics, 80,
University Press, 1971) Chap. 14, pp. 151–180.
2
pp. 519 (2012).
G. E. Stedman, Diagram techniques in group theory (Cambridge 30
Moreover, further invariant products between a symmetric trace-
University Press, Cambridge, 2009). less rank-2 tensor and another invariant object (a scalar, a vector,31
3
P. Cvitanović, Group theory: birdtracks, Lie’s, and exceptional or symmetric traceless rank-2 tensor, etc.) can be defined. Do-
groups (Princeton University Press, 2008) p. 273. ing this requires irreducible decomposition formulae of higher-rank
4
We do not distinguish the SO(3) contravariant and covariant in- tensors, such as the decomposition in A20 and the techniques in
dices in this article; the use of upper indices here is just for nota- Section I C 6.
31
tional convenience. M. Lazar, Journal of Applied Mathematics and Mechanics, 96,
5 11, pp. 1291–1305 (2016).
In fact, we also need the fact that the only rotationally invariant
32
rank-2 tensor is the Kronecker delta. In fact, one can take this decomposition as a starting point then
6
See footnote 18 of the main article. construct the invariant symbols, figuring out what projector of
7
B. Coecke, in AIP Conference Proceedings (AIP, 2006) pp. 81–98. an invariant subspace satisfies the “proper propagation of arrows”
8 condition and identifying it with a projector made by generators
H. Taouk, Wave propagation in general anisotropic media, Master’s
of SO(3), as demonstrated in Cvitanovic̀’s Group Theory.3 The
thesis, Ohio University (1986).
9
validity of Eq. (26) can be confirmed without prior knowledge of
R. Penrose, The road to reality: A complete guide to the physical invariant symbols by the orthogonality of the spin-0, 1, and 2 pro-
universe, vintage ed. (Vintage Books, New York, 2007). jectors in Eq. (26) and the fact that their dimensionalities sum to
10
If you are sad that the “hierarchical” design makes the previously 1 + 3 + 5 = 3 × 3 (cf. A67c).
dancing diagrams “stiffer,” try another way: assigning a “hook” or 33
J. D. Jackson, Classical Electrodynamics (Wiley, 1999), p. 808.
an “eye”11 to each end to the lines. For example, Bi = B i, 34
D. J. Griffiths and D. F. Schroeter, Introduction to quantum me-
k
i
A = i
i
A , Bi A = B A , εi
jk
= , etc. Hooks rep- chanics (Cambridge University Press, 2004).
35
i j S. B. Cahn, G. D. Mahan, and B. E. Nadgorny, A Guide to
resent output lines; eyes represent input lines. (This is equivalent Physics Problems: Part 2: Thermodynamics, Statistical Physics,
to making the lines directed.) and Quantum Mechanics (Springer Science & Business Media,
11
I. V. Lindell, Differential forms in electromagnetics (John Wiley 1997), pp. 131–133.
36
& Sons, 2004). M. A. Srednicki, Quantum Field Theory (Cambridge University
12 Press, 2007), p. 641.
The word “golden” is put just for linguistic playfulness, as a remi-
37
niscence of D. R. Hofstadter’s book Gödel, Escher, Bach: an eter- H. Elvang and Y.-T. Huang, Scattering Amplitudes in Gauge The-
nal golden braid. ory and Gravity (Cambridge University Press, 2015).
38
13
D. J. Griffiths, Introduction to electrodynamics (Prentice Hall, Z. Bern and J. J. M. Carrasco, J. Henrik, Physical Review D, 78,
1999), p. 576. 8: 085011 (2008).
39
14
D. J. Griffiths, American Journal of Physics, 50, p. 698 (1982). N. Arkani-Hamed and Y. Bai, S. He and G. Yan, Journal of High
15 Energy Physics, 8, p. 40 (2018).
Two lines of the same type related by a basis change can be con- 40
N. E. J. Bjerrum-Bohr, P. H. Damgaard, R. Monteiro, and D.
nected by the matrix that corresponds to the basis change.
16
O’Connell. Journal of High Energy Physics, 6, 61 (2012).
P. Cvitanovic and A. D. Kennedy, Physica Scripta 26, p. 5 (1982). 41
C.-H. Fu, Y.-J. Du, and B. Feng. Journal of High Energy Physics,
17
R. Penrose, Combinatorial mathematics and its applications, pp. 3, 50 (2013).
221 (1971). 42
C.-H. Fu, P. Vanhove, and Y. Wang. Journal of High Energy
18
P. Ball, Flow: Nature’s patterns: a tapestry in three parts (OUP Physics, 9, 141, 2018.
Oxford, 2009). 43
R. Monteiro and D. O’Connell. Journal of High Energy Physics,
19
Non-local fields can be also considered, in principle. 7, 7 (2011).

View publication stats