PHY680 (Fall 2017) Peter van Nieuwenhuizen

1 The Dirac Equation in Flat Space

Soon after quantum mechanics was constructed in 1925 and 1926, the question arose of a relativistic
equation for the electron. Schrödinger and others had taken in 1926 the wave equation for a free
complex one-component field φ
 
1 2  mc 2
∂x2 + ∂y2 + ∂z2 − 2 ∂t − φ=0 (1.1)
c ~

but this seemed to lead to a fatal problem1 : there was a currrent

jµ ∼ φ? ∂µ φ − (∂µ φ? )φ (1.2)
. .
which was on-shell conserved (∂ µ jµ = 0 if (1.1) holds). Believing (incorrectly) that 1c j 0 = ~c12 φ? φ − φ? φ
was the probability density for an electron to be at a point x, it would imply that there could be
negative probabilities. That made of course no sense. Later it was realized that ejµ is the electric
current, and ec j 0 is the charge density which can, of course, be negative. But at the time, the wave
equation did not seem correct, and one was searching for a better equation.
∂
Dirac studied this problem in 1928, and believed that if he could get rid of the ∂t in j 0 , the
density problem could be cured. He therefore considered a field equation for a relativistic electron
which was linear in time derivatives, and then the theory of special relativity of 1905 also required
that the field equation be linear in space derivatives. Thus he considered the following equation

µ mc 
γ ∂µ + ψ=0 (1.3)
~
where mc ~
is the Compton wavelength of the electron which A.H. Compton had encountered in his
studies of Compton scattering in 1924. What was the meaning of the objects γ µ ? Dirac took them
to be dimensionless matrices which would act on a multi-component ψ. In order that the relativistic
mass relation cpµ · cpµ + (mc2 )2 = 0 would follow from his field equation, he multiplied the latter by
γ ν ∂ν − mc
~
and required that the result should be equal to the wave equation. This led to a constraint
on the matrices γ µ

{γ µ , γ ν } = 2η µν (1.4)
1
Another equally fatal problem with this equation for the electron in the hydrogen atom was that it did not yield
the correct hydrogen spectrum. However, this is less relevant for our discussion of the origin of the Dirac equation.

1
Such a relation between 2 × 2 matrices had been used by Pauli in 1927 for his nonrelativistic two-
component spinor theory, but Dirac claimed later that he was in 1928 unaware of it. But he may have
seen such an equation before because in 1843, Hamilton had discovered quaternions with imaginary
units I, J, K which satisfy {I, J} = {J, K} = {K, I} = 0, {I, I} = {J, J} = {K, K} = −2, and there
was much interest in quaternions in England. Later, a 2 × 2 matrix representation for quaternions
was found in terms of “Pauli matrices”2 (I = −ı̇σ1 , J = −ı̇σ2 , K = −ı̇σ3 ) which indeed satisfy
(1.4) (and more: σ1 σ2 = ı̇σ3 and cyclic relations). In fact, the more general problem of finding a
solution of the equation {γ µ , γ ν } = 2δ µν for any number of γ µ in terms of matrices had been studied
by the English mathematician William Clifford in 1878. Clifford had seen the work of 1844 of the
German mathematician Hermann Grassmann who had studied anticommuting numbers (ab = −ba
so {a, b} = 0) but he added the requirement that (γ µ )2 = 1. Thus there were some mathematical
studies available for someone who would like to solve (1.4).
Dirac just tried. He could not use the 3 Pauli matrices because he needed 4 of them. He may first
have tried 3 × 3 matrices but that does not work. Then he tried 4 × 4 matrices. That did work, and
the rest is history.3 We shall not need the explicit form of the Dirac matrices γ m , but we mention
here the representation most often used in 4-dimensional Minkowski space4
! !
0 −iσ k 0 −iI
γk = for k = 1, 2, 3; γ0 = (1.5)
iσ k 0 −iI 0

It is obvious that they satisfy the Clifford algebra if one writes them as σ k ⊗ σ 2 and −iI ⊗ σ 1 .
As we shall discuss, the 4-component spinors transform under Lorentz transformation as δψ =
1 mn 1
ψ where 12 γmn = 41 (γm γn −γn γm ) are the Lorentz generators in the spinor representation.


2
λ γ
2 mn
Using the explicit representations given the (1.5) one find that the 6 matrices γmn = (γkl , γk γ0 ) are
2
These matrices were already known to the mathematician Felix Klein around 1850.
3
The Dirac equation was spectacularly successful: spin came out instead of having to be put in by hand, and also
the famous (or notorious) Thomas factor 2 for spin-orbit coupling came out. But of course the new equation also led
to new problems. For example, how did the 4-component ψ α (~x, t) transform under Lorentz rotations and boosts? In
mc
γ ∂µ + ~ ψ = 0, the ∂x∂ µ transformed as a vector, but the γ µ were constant matrices and thus did not transform,
µ

so how could the equation still be Lorentz-covariant? Clearly, the four component ψ α did not transform as a vector,
but were there other four-component objects which transformed in another but still well-defined way? What was
that way? Ehrenfest in Leiden was confused, and asked van der Waerden (a Dutch mathematician in Göttingen with
a knack for helping physicists with mathematics). He asked if there was a “spinor calculus” similar to the vector
calculus of relativity. van der Waerden took up the challenge and created in 1929 the nowadays standard “dotted and
undotted” spinor formalism with two-component spinors ψ a and χ.a .
4
In flat space there is no distinction between curved indices µ and flat indices m, so in flat space one uses them
interchangeably.

2
given by
! !
σkl 0 σk 0
γkl = ; γk0 = (1.6)
0 σkl 0 −σk

where σkl = iklm σ m . Hence the set of matrices γmn are block-diagonal, and thus the spinors Dirac
found form (in d = 3 + 1 dimensions) a reducible representation of the Lorentz group.
The Dirac equation in (1.3) can be obtained from an action by the Euler-Lagrange variational
R
principle. This action is called the Dirac action, and is given by S = d3 xdt L with
 mc µ
L = −ψ γ ∂µ + ψ (1.7)
~

Here ψ = ψ † iγ 0 . One needs the factor iγ 0 for Lorentz invariance of the action. If one varies with
respect to ψ one retrieves the Dirac equation, and if one varies with respect to ψ one finds the Dirac
equation for ψ
h
µ mc i
(∂µ ψ)γ − ψ =0 (1.8)
~
The Dirac action is hermitian
h  mc  i†  ← − mc  0
µ †
ψ γ ∂µ + ψ = ψ γµ ∂ µ + iγ ψ
~ ~ 
 ←
− mc
= ψ −γ µ ∂ µ + ψ
~
 mc
= ψ γ µ ∂µ + ψ + total derivative (1.9)
~

(We used a useful notation: (γ µ )† = γµ , which contains both (γ k )† = γ k = γk and (γ 0 )† = −γ 0 = γ0 .)

ie
Minimal electromagnetic coupling is introduced by replacing ∂µ by ∂µ − ~c Aµ , and together with the
Maxwell action, this yields the action for QED. We shall study minimal gravitational coupling, and
see that one needs in this case more than only adding a gravitational term to ∂µ .

Group Theory for Dirac matrices

We shall study whether the solution Dirac found for the matrices γ m is unique, and also discuss similar
questions in higher (and lower) dimensions, both for Euclidean, and Minkowski spaces, in fact for
spaces with s space coordinates and t time coordinates. We shall also study the reality properties

3
of the Dirac matrices. For that we shall introduce (and prove) the Frobenius-Schur theorem for the
reality properties of irreps of finite groups. To define Majorana spinors we must impose a reality
condition on spinors, and for that we need the charge conjugation matrix (actually, matrices in even
dimensions).
What has all this to do with group theory? The connection is as follows: One can form products
of γ µ , and products of these products, etc., simplifying the expressions we get by using the Clifford
algebra, until the process closes (until we find no more new elements). This yields the multiplication
table of all these elements. The set of all these elements forms a finite group (apparently not known
to Dirac, but certainly to Pauli) which we shall call the Dirac group.
We begin in Euclidean space, with n objects γ µ satisfying γ µ γ ν +γ ν γ m u = 2δ µν I. We take I as the
unit element e of the group and γ µ as the first n elements. The products γ µ γ ν for µ < ν are 21 n(n−1)
new elements, and γ ν γ µ = −γ µ γ ν for µ 6= ν are different eelmeents from +γ µ γ ν . (In an algebra one
can add and multiply, but in a group one can only multiply, thus −γ µ γ ν is a different element from
γ µ γ ν . For simplicity though, we keep writing −γ µ γ ν instead of inventing new names like (γ µ γ ν )0 .)
Having obtained the element γ 1 γ 2 , we can multiply it with γ 1 in two ways: γ 1 γ 1 γ 2 = (γ 1 γ 1 )γ 2 = γ 2
and γ 1 γ 2 γ 1 = γ 1 (−γ 1 γ 2 ) = −γ 2 . Thus we find also a new set of n elements −γ µ . Going on in this
way we get the following set of elements

G = {±e, ±γ µ , ±γ µ γ ν , ±γ µ1 γ µ2 γ µ3 , · · · , ±γ 1 γ 2 · · · γ n } (1.10)
µ<ν µ1 <µ2 <µ3

Products with more than n γ-symbols can be reduced to expressions with less than (or equal to) n
γ-symbols using the Clifford relations. This set of elements forms a group, with e the unit element.
Closure holds by construction. Since (γ µ )2 = +1, (γ µ )−1 is also equal to γ µ (so the group element γ µ
has order 2). We shall look for matrix representations, and matrix multiplication is associative, so
we also assume that the multiplication is associative. Thus we have a group and it is of order 2n+1

Order G = 2n+1 .

The first question to be settled is: what are the classes of this group? Since γ 1 γ 2 γ 1 = −γ 2 , a
class with γ 2 contains also −γ 2 . Going on in this way one finds that all classes, except the classes
with ±γ 1 γ 2 · · · γ n , contain two elements

+γ µ1 · · · γ µk and − γ µ1 · · · γ µk for µ1 < µ2 < · · · < µk .

4
The exceptions are the maximal elements ±γ 1 · · · γ n . Here it matters whether n is even or odd. If n
is even one has for any γ k that γ k γ 1 · · · γ n γ k = −γ 1 · · · γ n , hence both elements form one class. But
if n is odd, one has γ k γ 1 · · · γ n γ k = γ 1 · · · γ n . (Example: if n = 2 one has γ k γ 1 γ 2 γ k = −γ 1 γ 2 for any
k = 1, 2, but for n = 3 one has γ k γ 1 γ 2 γ 3 γ k = γ 1 γ 2 γ 3 for any k = 1, 2, 3.) Hence the class structure
is as follows

Even n : G = {e, −e, ±γ µ , ±γ µ γ ν , · · · , ±γ 1 · · · γ n } : 2n + 1 classes

Odd n : G = {e, −e, ±γ µ , · · · , γ 1 · · · γ n , −γ 1 · · · γ n } : 2n + 2 classes

The next question to be settled is: what is the commutator subgroup C(G)? Since all group
elements square to ±e one has γ k γ l (γ k )−1 (γ l )−1 = γ k γ l γ k γ l = −γ k γ l γ l γ k = −e. Hence C(G)
contains only two elements

C(G) = {e, −e} (1.11)

orderG n+1
Then order C(G)
= 2 2 = 2n , and group theory tells us that there are 2n one-dimensional irreps of
the group. These one-dimensional irreps can not satisfy γ µ γ ν + γ ν γ µ = 0 for µ 6= ν, hence they are
not a representation of the Clifford algebra. Only the higher-dimensional irreps can be used
in the Dirac equation.
i 2
P
Using another group formula, order G = i (dim R ) , where Ri are the dimensions of the
irreducible matrix representations (irreps) of G, and a third group formula: the number of irreps is
equal to the number of classes, we find

Even n : 2n+1 = |12 + ·{z

· · + 1}2 +x2
2n squares

Odd n : 2n+1 = |12 + ·{z

· · + 1}2 +x2 + y 2
n−1
2 2 squares

n−1
There is only one solution: x = 2n/2 for even n and x = y = 2 2 for odd n. Thus for even n = 2k
there is a unique 2k × 2k matrix representation (up to equivalence), but for odd n = 2k + 1, there
are two inequivalent irreps, both of dimension 2k × 2k . It is actually easy to find these two irreps in
odd dimensions from the unique irrep in even dimensions. Denoting in even dimensions the matrix
αγ 1 · · · γ n by γc (c for chiral; for n = 4 one often calls this γ5 ) and fixing α (up to a sign) by requiring
that γc2 = +1, we get the two irreps of the next odd dimension by adding +γc or −γc to the set of
n = even matrices γ µ . The set {γ µ , −γc } is inequivalent to the set {γ µ , γc } as we shall shortly show.

5
Thus there are the two inequivalent irreps for odd n. For example, if for n = 2 we have γ 1 , γ 2 (for
example σ 1 , σ 2 or σ 1 , σ 3 ) then γc = iγ 1 γ 2 is a new abstract group element for n = 3, and one gets two
irreps for n = 3, namely the irrep generated by {γ 1 , γ 2 , γc } and the irrep generated by {γ 1 , γ 2 , −γc }.
Why are they inequivalent? Suppose they were equivalent, then there should be a matrix S such
that γ k = Sγ k S −1 for k = 1, · · · , n (even n) and Sγc S −1 = −γc . But since γc = αγ 1 · · · γ n , one
gets Sγc S −1 = α(Sγ 1 S −1 ) · · · (Sγ n S −1 ) = αγ 1 · · · γ n = +γc . Thus they are inequivalent. In the
same way one may prove that the irreps generated by {γ 1 , · · · , γ n , γc } and {−γ 1 , · · · , −γ n , −γc } are
inequivalent.

The charge conjugation matrices C+ and C−

In order to define real (Majorana) fermions in certain spaces, one needs the charge conjugation
matrices. The name is only of historical origin, they have nothing to do with charge. They are
defined by the following equation

C+ γ µ C+−1 = +γ µ,T ; C− γ µ C−−1 = −γ µ,T (1.12)

Are there always such matrices?

1. The Clifford algebra reads {γ µ , γ ν } = 2δ µν (we are still in Euclidean space) and taking the
transpose one obtains {γ µ,T , γ ν,T } = 2δ µν . Hence γ µ,T satisfies the same Clifford algebra as γ µ ,
and since the irrep in even dimensions is unique (up to similarity transformations) this implies
that γ µ,T = C+ γ µ C+−1 in even dimensions. Also −γ µ,T satisfies the Clifford algebra, so also
−γ µ,T = C− γ µ C−−1 (in even dimensions). Given C+ in even dimensions (even n), one obtains
C− as C− = C+ γc . This is clear from

C− γ µ C−−1 = C+ γc γ µ γc C+−1 = −C+ γ µ C+−1 = −γ µ,T (1.13)

Thus there exist both a C+ and C− in even dimensions, and C− = C+ γc .

2. In odd dimensions the situation is different. It is still true that γ µ,T and −γ µ,T satisfy the
Clifford algebra, but there are two irreps in odd dimensions which we denote by γIµ and γII µ
.
µ µ µ µ,T
Then all we know is that +γI is equivalent to either γI or γII , and −γI to the other. The
µ,T
same holds for ±γII . We shall now show that in in odd dimensions there always is a
C matrix, but only a C+ or a C− but not both. To see this consider γc = αγ 1 · · · γ n
where α is chosen such that γc2 = I, and evaluate C+ γc C+−1 where C+ is the C+ is the C+ of

6
 the even dimension below the odd dimension. Since C+ γc C+−1 = C− γc C−−1 (because C− = C+ γc
and γc2 = I), we find

C+ γc C+−1 = α(C+ γ 1 C+−1 ) · · · (C+ γ n C+−1 )

= C− γc C−−1 = α(C− γ 1 C−−1 ) · · · (C− γ n C−−1 )
(
+αγ n,T · · · γ 1,T = +γcT for n + 1 = 5, 9, · · ·
= αγ 1,T · · · γ n,T (n is even) =
−αγ n,T · · · γ 1,T = −γcT for n + 1 = 3, 7, · · ·

(n−1)
If C± γc C±−1 = +γcT we choose C (n) = C+ (the charge conjugation matrix C+ in n dimensions
−1 (n−1)
for γ , · · · , γ ), while if C± γc C± = −γc we choose C (n) = C−
1 n T
(the charge conjugation
matrix C− in n dimensions) (for γ , · · · , γ ). Then in the first case C+ γµ C+−1 = γµT for all Dirac
1 n
(D)
matrices while in the second case C− γ µ C−−1 = −γµ . So there is always a C+ or a C− matrix
in odd dimensions, but in 5, 9, · · · dimensions there is only a C+ , and in 3, 7 · · · dimensions
there is only a C− .

3. Now consider chiral or antichiral spinors in n = even dimensions. If one uses as Dirac matrices
the set Γi = γ i ⊗ τ2 and I ⊗ τ1 , where γ i are the Dirac matrices in n − 1 odd dimensions, then
C± is either block diagonal (in d = 4, 8) or block anti-diagonal (in d = 2, 6).

4. The matrices C± are (anti)symmetric. This follows from (γ µ,T )T = γ µ . Indeed,

(γ µ,T )T = (±C± γ µ C±−1 )T = ±C±−1,T (±C± γ µ C±−1 )C±T (1.14)

from which it follows that C±−1 C±T commutes with γ µ . Then Schur’s lemma tells us that
C±−1 C±T = σI, or C±T = σC± . Since (C±T )T = C± , we get σ 2 = 1, so σ = ±1.

5. C± can be scaled to be unitary. Using a unitary irrep for γ µ (any irrep of a finite group
can always be made unitary by a suitable similarity transformation) the γ µ are hermitian
(γ µ )† = (γ µ )−1 = γ µ . Thus

γ µ = (γ µ )† = ±(C± γ µ C±−1 )∗ = ±C±∗ γ µ,T C±−1,∗ = ±C±∗ (±C± γ µ C±−1 )C±−1,∗ (1.15)

This time C±∗ C± commutes with γ µ , so C±∗ C± = αI. Using C±T = σC± we get C±† C± = ασI.
Since C±† C± is positive definite, ασ is positive (ασ > 0), so we can take the square root of ασ,
and scale C± to be unitary.

7
 6. C± γ µ are (anti)symmetric.

(C± γ µ )T = γ µ,T C±T = (±C± γ µ C±−1 )σC± = ±σ(C± γ µ ) (1.16)

7. C± are invariant tensors of SO(N ). If C± γµ C±−1 = ±γµT then C± γµν C±−1 = γ[µ T T
γν] =
T T
−(γµν ) , so C± γµν + γµν C± = 0. This can be written as the transformation law of a tensor
Cαβ :

0 0
δC±,αβ = C±α,β 0 (γµν )β β + (γµν
T
)α α C±,α0 β
0 0
= C±α,β 0 (γµν )β β + C±α0 ,β (γµν )α α = 0 (1.17)

Thus δC+αβ = 0: it is an invariant tensor of SO(N ).

8. The Dirac matrices in n Euclidean dimensions, are invariant tensors of SO(N ).

Consider γµ α β as a tensor with 3 indices. If for a vector vµ we have δvµ = λµ ν vν , then
transformation of the 3 matrices of γµ α β yields

1 0 0
δγµ α β = λµ ν γν α β + λρσ (γρσ )α α0 γµ α β − γµ α β 0 14 λρσ (γρσ )β β
4
ν 1 ρσ
= λµ γν + 4 λ [γρσ , γµ ] (suppressing spinor indices)
= λµ ν γν + λρσ γρ δσµ
= λµ ν γν + λρ µ γρ = 0 (1.18)

9. The Dirac matrices in Minkowski spacetime are obtained by multiplying one of them by +i or
−i. This changes, of course, the reality properties of the set of Dirac matrices, but the charge
conjugation matrix (matrices) remains the same.

This concludes the list of properties of C± . We shall give examples below, but first we want to
discuss the reality properties of the Dirac matrices in γ µ .

Real, pseudoreal and complex representation

We consider a representation D(g) of a finite group, and (can) assume that the representation is
unitary. We then define

1. a rep is real if for some matrix R (not necessarily a group element) the matrices RD(g)R−1 =

8
 D(g)real are real (have real entries for all g). One may assume that D(g)real is again unitary5
(and hence orthogonal).

2. a rep is pseudoreal if D(g)∗ is similar to D(g), D(g)∗ = SD(g)S −1 , but cannot be made real.

3. a rep is complex if D(g)∗ is not equivalent to D(g). (Since for a real rep one has RD(g)R−1 =
D(g)real = R∗ D(g)∗ R−1 , and R = R∗ , the rep D(g) is similar to D(g)∗ . So this covers all
possibilities.)

Theorem (Froebenius-Schur): A unitary irrep of a finite group is

 
real 
 X  order G

2
pseudoreal if χ(g ) = −order G
 
g
complex 0
 

Before proving this theorem we consider some examples.

Example 1. For the Dirac metric in d = (2, 0) (n = 2 Euclidean case) we can take γ 1 = σ 1 and
γ 2 = σ 3 , which is evidently a real representation (the products of real matrices are real matrices).
One finds

G = {±I, ±σ 1 , ±σ 3 , ±iσ 2 }
{g 2 } = {I, I, I, −I}
X
=⇒ χ(g 2 ) = 6 × 2 − 2 × 2 = 8 = order G
g

This agrees with the theorem. If one would instead start from γ 1 = σ 1 and γ 2 = σ 2 , the matrices are
not real, but one still finds g χ(g 2 ) = order G, and this is clear because one can find a similarity
P

transformation which maps σ 1 into σ 1 but σ 2 into σ 3 .

Example 2. d = (1, 1) (the Minkowski case with n = 2) with γ 1 = σ 1 and γ 2 = −iσ 2 . Again
this is a real representation. One gets the Minkowski case by multiplying one of the γ µ by −i. Often
5
If RD(g)R−1 = D(g)real is not unitary but real, one can make ]
Pa similarity transformation such that D(g) =
SD(g)real S −1 is unitary. The key idea is to construct the matrix T = g D(g)real (D(g)real )T . This is a real symmetric
matrix hence it can be diagonalized by a real orthogonal matrix O, namely OT OT = Λ with Λ diagonal and positive.
] = Λ−1/2 OD(g)real OT Λ1/2
We can then repeat the usual steps to unitarize, but with O instead of U , to show that D(g)
is real and unitary (and hence orthogonal).

9
one defines a matrix γ 0 by γ 0 = −iγ 2 . Then

G = {±I, ±σ 1 , ±iσ 2 , ±σ 3 }
{g 2 } = {I, I, −I, I}
X
=⇒ χ(g 2 ) = 8 = order G
g

So there is a real irrep in d = (1, 1). Indeed, σ 1 and −iσ 2 are real matrices.

Example 3. d = (4, 0). The Dirac group in 4 Euclidean dimensions has the following group
elements

G = {±I, ±γ µ , ±γ µ γ ν , ±γ µ γ5 , ±γ5 } (1.19)

µ<ν

where we used that γ5 = γ 1 γ 2 γ 3 γ 4 and

γ 1 γ 2 γ 3 = γ 1 γ 2 γ 3 γ 4 γ 4 = γ5 γ 4 (1.20)

Since {g 2 } = {I, I, −I, −I, I} we get

X
χ(g 2 ) = 4(2 + 8 − 12 − 8 + 2) = −32 = −order G (1.21)
g

Hence this irrep is pseudoreal. For a particular representation such as the one used in supersymmetry
!
µ 0 −iσ µ σ µ = (~
σ , iI)
γ = (1.22)
iσ µ 0 σ µ = (~
σ , −iI)

one may determine the matrix S such that (γ µ )∗ = Sγ µ S −1 (in this case (γ 1 )∗ = −γ 1 and (γ 3 )∗ = −γ 3
and S = γ 1 γ 3 .)

Example 4. d = (3, 1). In Minkowski space one finds

G = {±I, ±γ k and ± γ 0 , ±γ k γ l and ± γ k γ 0 , ±iγ k γ5 and ± iγ 0 γ5 , ±iγ5 } (1.23)

One needs the factors of i since γ5 = γ 1 γ 2 γ 3 γ 4 = γ 1 γ 2 γ 3 iγ 0 . Then, for example, G contains γ 1 γ 2 γ 0

10
which can be written as γ 3 (γ 1 γ 2 γ 3 iγ 0 )(−i) = −iγ 3 γ5 . Similarly since γ 1 γ 2 γ 0 = γ 1 γ 2 (γ 3 γ 3 )γ 0 (i)(−i),
G contains γ 1 γ 2 γ 3 γ 0 = −iγ5 . One now finds
X
χ(g 2 ) = 4 × 2[1 + (3 − 1) + (−3 + 3) + (3 − 1) + (−1)] = 32 = order G (1.24)
g

Thus there should be a set of real Dirac matrices in 4-dimensional Minkowski space (a so called
Majorana representation), and one can then use real spinors.
! It is not difficult
! to come up with 4
~
0 σ I 0
such real matrices: α1 , α3 , β and iα2 where α
~ = and β = are the matrices Dirac
~ 0
σ 0 −I
in his original paper.

Proof of the Frobenius-Schur theorem

Let us now prove the Frobenius-Schur theorem. For a real or pseudoreal irrep we have according to
the definition of real or pseudoreal irreps

D(g)∗ = SD(g)S −1 for all g (1.25)

We have proven that for finite groups we may assume that all D(g) are unitary matrices. Then
D(g)∗ = D(g)−1,T , and thus D(g)−1,T = SD(g)S −1 , which can be written as

D(g)T SD(g) = S (1.26)

Taking the transpose

D(g)T S T D(g) = S T (1.27)

and then the inverse

D(g)−1 S T,−1 D(g)T,−1 = S T,−1 (1.28)

we find upon multiplying (1.26) and (1.28)

D(g)T SS T,−1 D(g)T,−1 = SS T,−1 (1.29)

11
So S −1 S T D(g) = D(g)S −1 S T , from which we conclude by using Schur’s lemma that S −1 S T = σI or
S T = σS. Thus (S T )T = σ 2 S, so σ = ±1, and S is either symmetric or antisymmetric

S T = ±S (1.30)

S can also taken to be unitary when D(g) are unitary. This follows from D(g)T SD(g) = S; taking
the hermitian conjugate yields D(g −1 )S † D(g −1 )T = S † , and multiplying both expressions yields
S † SD(g) = D(g)S † S from which we conclude with Schur’s lemma that S can be scaled to be unitary.
We now first show that if S is symmetric, then the irrep is real, and next we show that if the
irrep is real, then S is symmetric. This implies that for a pseudoreal irrep S is antisymmetric. With
these results we then shall prove the Frobenius-Schur theorem.
If S is symmetric and unitary, we can take a square root

S = T2 (1.31)

where T is again symmetric and unitary. The proof uses some nice simple matrix algebra.
Proof: S = S T and S ∗ S = I. Define S = A + iB and A = 21 (S + S ∗ ) and B = 2i1 (S − S ∗ ). Both A
and B are real and symmetric, and they commute because S ∗ = S −1 . Hence they can be diagonalized
simultaneously by an orthogonal matrix O

OT (A + iB)O = OT SO = eiφ (1.32)

where φ is a diagonal real matrix. Hence

i
S = Oeiφ OT ; T = Oe 2 φ OT (1.33)

and S = T 2 where T is symmetric and unitary. We now see that

D(g)∗ = SD(g)S −1 = T 2 D(g)T −2 (1.34)

so, upon multiplying by T −1 from the left and T from the right,

T D(g)T −1 = T −1 D(g)∗ T = T ∗ D(g)∗ T ∗,−1 (1.35)

because T −1 = T † = T ∗ . Thus T D(g)T −1 is real, and this is the definition that D(g) is a real
representation. Hence, indeed, if S is symmetric then D(g) is a real irrep.

12
 Let us conversely assume that D(g) is a real irrep, and prove that S is symmetric. We begin with
RD(g)R−1 = D(g)real , hence R∗ D(g)∗ R∗,−1 = D(g)real . Then D(g)∗ is given by

D(g)∗ = R∗,−1 RD(g)R−1 R∗ = SD(g)S −1 , (1.36)

with S = R∗,−1 R. Using that both D(g) and D(g)real can be taken to be unitary (as we showed
before), also R is unitary.6 Then R∗,−1 = RT and S = RT R, which is indeed symmetric.
Consider now the orthogonality relations for irreps. For real or pseudoreal irreps R
b we get
X X X
χ(g 2 ) = D(g)µ ν D(g)ν µ = D(g)µ ν S νρ D(g)∗ρ σ Sσµ
−1

g g g
!
X
−1
= (S νρ Sσµ ) D(g)µ ν D(g)∗ ρ σ
g

−1 µ σ order G
= (S νρ Sσµ )δρ δν (1.37)
dim Rb

For real irreps S is symmetric, and then the first factor yields unity times dim R,
b but for pseudoreal
irreps S is antisymmetric, and then the first factor yields minus unity times dim R.b In both cases
the theorem is proven. Finally, when the irrep is complex we get
X X
χ(g 2 ) = D(g)µ ν D(g)ν µ
g g
X
= D(g)µ ν (D(g)∗ )∗ν µ = 0 (1.38)
g

because D(g) and D(g)∗ are inequivalent. This also proves the theorem for complex irreps.

6
The proof that R is unitary follows from RD(g)R−1 = D(g)real and R−1,† D(g −1 )R† = D(g −1 )real . Multiplying
both expressions gives R−1,† D(g −1 )R† RD(g)R−1 = I, or R† R commutes with D(g). Then R† R = λI and R can be
scaled to be unitary.

13