Wess-Zumino-Witten Models

YRISW PhD School in Vienna

Lorenz Eberhardt, ETH Zürich

eberhardtl@itp.phys.ethz.ch

February 19, 2019

Abstract
We introduce the principal chiral model in two dimensions and its extension by
the Wess-Zumino term. We discuss the symmetries, corresponding currents and
their quantum algebra. We explain the Sugawara construction to demonstrate that
the Wess-Zumino-Witten model defines a conformal field theory. We then move on
and discuss representations, characters, fusion and modular invariants by following
the example of su(2). Finally, we briefly discuss the coset construction.

Contents
1 Introduction 2

2 The classical theory 2

2.1 Principal chiral model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2.2 WZW action . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

3 The quantum theory 7

3.1 Current algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.2 Affine Kac-Moody algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.3 The Sugawara construction . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.4 The central charge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.5 Free field constructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.6 Representations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.7 Characters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.8 Fusion rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.9 Modular invariants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

4 Cosets 31
4.1 The coset construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.2 Representations and characters . . . . . . . . . . . . . . . . . . . . . . . . 33

5 Outlook 36

1
1 Introduction
In these lectures, we will introduce Wess-Zumino-Witten (WZW)-models. They are a
prime example of rational conformal field theories and are completely solvable. They
feature an extended chiral algebra: The Virasoro algebra gets extended to an affine Lie
algebra, due to the existence of additional symmetries in the theory.
We start by the classical formulation of the theory. Wess-Zumino-Witten models
can be viewed as non-linear sigma models, where the target space is a group manifold.
To make them conformal, one introduces an additional term in the action. We discuss
the action and derive from it the quantum symmetry algebra. It is given by an affine
generalization of finite-dimensional Lie algebras:
a
[Jm , Jna ] = kδ ab δm+n,0 + if abc Jm+n
c
. (1.1)
This algebra forms the chiral algebra of the model. We explain how to construct out of
it the holomorphic component of the energy-momentum tensor, thereby demonstrating
conformal invariance of the theory. A study of the representations will show that they
are severely restricted, in particular there are only finitely many possible representations.
This is very much in contrast to finite-dimensional Lie algebras.
We then discuss the structure of the complete theory. A study of the fusion rules gives
us information about possible three-point functions of the theory. We will not explore
three- and four-point functions any further. They can be computed via the Knizhnik-
Zamolodchikov (KZ)-equations. We discuss the constraint of modular invariance, which
is a consistency condition arising when putting the theory on the torus. We explain the
classification of modular invariants of the SU(2) WZW-model.
Finally, we introduce the coset construction, which corresponds to gauged WZW-
models. It accounts for a wealth of known rational conformal field theories. In particular,
we show how to retrieve the unitary minimal models as a coset.
The material presented here has become standard. Good general references are [1, 2].
The ‘big yellow book’ [3] is a very complete treatment of all the topics discussed in these
lectures. The material is however stretched over several hundred pages. For these lectures,
we require only a basic knowledge of CFT in two dimensions.
There are exercises scattered throughout the text. The reader is very much urged to
solve those exercises.

2 The classical theory

2.1 Principal chiral model
We consider a quantum field theory on the Riemann sphere, viewed as the one-point
compactification of the two-dimensional plane. Throughout this lecture, we will work in
Euclidean signature. This will provide us with the tools of complex analysis. The model
we will discuss in this lecture is a principal chiral model on a Lie-group G.1 We will use
a matrix notation for group and Lie algebra elements. In particular, for two Lie algebra
elements X and Y , we can define a non-degenerate invariant product as


hX, Y i = tr XY . (2.1)
1
Technically, we need to assume this Lie group to be semi-simple, i.e. (locally) a product of simple
Lie groups. This condition is equivalent to the existence of a non-degenerate invariant trace on the
corresponding Lie algebra g = Lie(G).

2
We will normalise the trace such that tr (ta tb ) = 12 δ ab for two generators of the Lie algebra
in the fundamental representation.2
The field of the model is denoted by g(z, z̄). We think of them as matrices, i.e. we
choose some (faithful) representation of G. g(z, z̄) is hence a map

g : S2 −→ G . (2.2)

Consider the action

Z
1
d2 z tr g −1 ∂µ g g −1 ∂ µ g .


S0 = 2 (2.3)
4λ S2

Since g is an element in the Lie group, g −1 ∂µ g defines an element of the Lie algebra.3
The action has a global G × G-symmetry, given by

g(z, z̄) 7−→ gL g(z, z̄) gR−1 (2.4)

for some constant group elements gL , gR ∈ G.

This theory is classically conformally invariant. Computing the β-function shows that
this is no longer true at the quantum level. One finds a negative β-function. Hence
conformal invariance is broken at the quantum level and the theory is asymptotically
free. This is why we will modify the action (2.3) to obtain the action of a conformal field
theory, which is referred to as Wess-Zumino-Witten (WZW) model.
Let us compute the classical equations of motion by varying the action w.r.t. g:
Z
1
d2 z tr (−g −1 δgg −1 ∂µ g + g −1 δ∂µ g)g −1 ∂ µ g


δS0 = 2 (2.5)
2λ S2
Z
1
d2 z tr ∂µ (g −1 δg) g −1 ∂ µ g


= 2 (2.6)
2λ S2
Z
1
d2 z tr g −1 δg∂µ (g −1 ∂ µ g) .


=− 2 (2.7)
2λ S2
Here, we used the cyclicity of the trace and integrated by parts. Since this holds for every
variation, we deduce the classical equations of motions:

∂µ (g −1 ∂ µ g) = 0 , (2.8)

hence the Lie algebra valued current J µ = g −1 ∂ µ g is conserved. This current is in fact
the associated current to the right multiplication symmetry g 7→ ggR−1 . The current for
the left-multiplication symmetry g 7→ gL g is given by J˜µ = ∂ µ gg −1 and is also conserved.
Indeed, we can rewrite the equations of motion (2.8) as

∂µ (g −1 ∂ µ g) = 0 ⇐⇒ g∂µ (g −1 ∂ µ g)g −1 = ∂µ (∂ µ gg −1 ) = 0 . (2.9)

In complex coordinates, current conservation reads

¯ z=0,
∂Jz + ∂J (2.10)
2
A fundamental representation does not exist in some cases, but the following analysis still goes
through.
3
Mathematically, g −1 dg is the pullback of the Maurer-Cartan form to S2 under the map defined by g.

3
where here and in the following ∂ ≡ ∂z and ∂¯ ≡ ∂z̄ . In conformal field theory, we would
like to have (anti)holomorphic quantities, in other words both terms in this equation
should vanish separately. Indeed, the situation is entirely analogous to the conservation
of the stress energy tensor in a CFT. We have
¯ zz̄ = 0 ,
∂Tz̄z̄ + ∂T ¯ zz = 0 .
∂Tz̄z + ∂T (2.11)

Tracelessness of the energy-momentum tensor in a CFT reads Tzz̄ = Tz̄z = 0 in complex

coordinates. Hence T ≡ Tzz and T̄ ≡ Tz̄z̄ are holomorphic (antiholomorphic) quantities.
In the WZW-case, one of the components of the currents vanishes identically and hence
the other component will be (anti)holomorphic. This is precisely achieved by adding the
WZ-term to the action, as we shall see below.
But for the principal chiral model, not both terms can vanish separately. Indeed, J µ
can also be seen to be a gauge potential for the gauge group G. By definition, it is of
pure gauge and hence its field strength vanishes, so

∂µ Jν − ∂ν Jµ + [Jµ , Jν ] = 0 . (2.12)

If in (2.10) both terms would vanish separately, then this would imply that the dual
current µν J ν would also be conserved. Here µν is the totally antisymmetric tensor.
However, we have
1 1
∂µ (µν Jν ) = µν (∂µ Jν − ∂ν Jµ ) = − µν [Jµ , Jν ] = −µν Jµ Jν . (2.13)
2 2
This expression vanishes only for an abelian Lie algebra,4 so the currents are not sep-
arately conserved. Thus, the theory will not contain (anti)holomorphic currents. As a
consequence, we do not expect this to be a two-dimensional conformal field theory at the
quantum level.
Note that our task would be complete for an abelian Lie algebra. It turns out that in
this case, this indeed defines a conformally invariant theory and the Wess-Zumino term we
define below can be seen to vanish. In fact, the WZW-model for the abelian Lie algebra
Rn gives the theory of n free bosons. In the following, we want to focus on the non-abelian
case.

2.2 WZW action

To rectify the issue of non-holomorphicity in the non-abelian case, we will modify the
action and will add to it the Wess-Zumino term, which is given by
Z
i
d3 y αβγ tr g −1 ∂ α gg −1 ∂ β gg −1 ∂ γ g


Γ[g] ≡ − (2.14)
12π B
Z
i
tr g −1 dg ∧ g −1 dg ∧ g −1 dg .


=− (2.15)
2π B
This needs some explanation. The field g is defined on the Riemann sphere S2 , which can
be thought of being the boundary of a three-dimensional ball B. The fields are smooth
maps

g : S2 −→ G . (2.16)
4
An abelian Lie algebra is a Lie algebra, where all commutators vanish.

4
Such maps are classified up to homotopy by the second homotopy group π2 (G). It is a
mathematical fact that the second fundamental group of every Lie group vanishes, i.e.
π2 (G) = 0. This implies that every map as in (2.16) is homotopic to the constant map.
Clearly the constant map can be continued to the interior of S2 , and hence so can any
map g. We denote this extension of g also by g, i.e. g maps now from B to G and on
∂B = S2 it is equal to the original map. Of course this extension is not unique. Let us
analyse how changing the relevant extension will modify the value of Γ. If we have two
different extensions, we can glue them together along the common boundary (where they
agree by assumption). This provides a map
g : (B t B)/∂B ≈ S3 −→ G , (2.17)
where ≈ denotes homeomorphism. Such maps are classified up to homotopy by the third
homotopy group π3 (G). It is also a mathematical fact that π3 (G) ∼ = Z for every compact
3
simple Lie group. In fact, every continuous mapping of S 7→ G is homotopic to a mapping
into a SU(2)-subgroup.5 So, we may for the moment assume that G = SU(2) ∼ = S3 . Thus,
gluing the two different extensions together provides us with a map S3 → S3 ⊂ G. Then
the integer Z simply classifies how many times S3 wraps around itself.
The WZ-term (2.15) features the extension of the field g. In order for this action to be
well-defined, we have to check, that the result is independent of the extension we choose.
First, let us check that the action is invariant under small perturbation of the extension g
(technically under homotopies of g relative ∂B = S2 ). Under this variation, the integrand
will be exact, so we can use Stokes’ Theorem:
Z
i
tr (−g −1 δgg −1 dg + g −1 dδg) ∧ g −1 dg ∧ g −1 dg


δΓ = − (2.18)
4π B
Z
i
d tr g −1 δgg −1 dg ∧ g −1 dg


=− (2.19)
4π B
Z
i
tr g −1 δgg −1 dg ∧ g −1 dg


=− (2.20)
4π S2
Z
i
tr g −1 δg d(g −1 dg) .


=− (2.21)
4π S2
Hence, if the variation of g vanishes on S2 , then Γ is invariant.
It remains to check the behaviour of Γ under taking topologically different extensions.
In fact, for the classical theory to be well-defined, the action does not have to be invariant
under topologically inequivalent extensions. After all, we can still get equations of motion
just by varying the action. However, this has an important consequence for the quantum
theory. As seen above, Γ is invariant under homotopies and g maps without loss of
generality into an SU(2) subgroup, so to analyse the behaviour of Γ under change of
homotopy class, we can just take g to be the identity, i. e.
g(y) = y 0 − iy k σk , y ∈ S3 ⊂ R4 . (2.22)
Then S3 wraps the target S3 exactly once, so we compute the difference of Γ for two
extensions with neighbouring homotopy classes. Then g −1 ∂ k g = −iσ k and using the form
(2.14) of the action, we obtain
Z
i
d3 y αβγ tr g −1 ∂ α gg −1 ∂ β gg −1 ∂ γ g


∆Γ = − (2.23)
12π B
5
This is assuming that the group as an SU(2)-subgroup. This may fail for global reasons as for SO(3).

5
 i(−i)3 2 X
=− 2π ijk tr (σ i σ j σ k ) (2.24)
12π i,j,k
π X
= ijk tr ([σ i , σ j ]σ k ) (2.25)
12 i,j,k
π X
= 2i ijk tr (ij` σ` σk ) (2.26)
12 i,j,k,`
iπ X iπ
= 2tr (σ k σk ) = 12 = 2πi . (2.27)
6 k,` 6

We used that the integrand is a constant and inserted the volume of the unit 3-sphere.
Let us now set

S[g] ≡ S0 [g] + kΓ[g] (2.28)

for some real number k. As we have discussed, this is a perfectly well-defined classical
theory for any value of k. However, in the quantum theory, we will have an additional
constraint on k. Indeed, let us compute path integrals of the form
Z
hOi = Dg O[g] e−S[g] (2.29)

for some operators O[g]. In order for this procedure to be well-defined, the exponential
e−S[g] has to be single-valued. This means that S[g] has to be well-defined up to the
addition of 2πi times an integer. Since ∆Γ = 2πi, we conclude that this is the case if

k∈Z. (2.30)

Thus, for compact Lie groups, the number k has to be an integer for topological reasons.6
It is referred to as the level of the model. Notice that for non-compact Lie groups, there
is no such quantization condition.
Let us again consider the classical model.

Exercise 1 Derive the equations of motion for the complete action. Verify that
the equations of motion still reflect the conservation of the global G × G-symmetry.
Express the result in complex coordinates. You should obtain

λ2 k λ2 k ¯ −1
   
−1 ¯
1+ ∂(g ∂g) + 1 − ∂(g ∂g) = 0 . (2.31)
π π

Solution 1 We already derived the variation of the two pieces of the action
S0 + kΓ, see eqs. (2.7) and(2.21). So our task consists in translating the terms into
complex coordinates. We have

z = x + iy , z̄ = x − iy , ∂ = 21 (∂x − i∂y ) , ∂¯ = 21 (∂x + i∂y ) . (2.32)

6
This analysis relied on the fact that we have a SU(2)-subgroup inside G. For SO(3), no such subgroup
exists and the quantization condition reads instead k ∈ 2Z.

6
 The measure reads

d2 z = dx ∧ dy = 1
4i
( dz + dz̄) ∧ ( dz − dz̄) = 2i dz ∧ dz̄ . (2.33)

Thus, the variation (2.7) becomes

Z
i ¯ + ∂(g
¯ −1 ∂g) .
dz dz̄ tr g −1 δg ∂(g −1 ∂g)


δS0 = − 2 (2.34)
4λ S2

For the term (2.21), we compute

Z
i ¯ dz̄)
tr g −1 δg d(g −1 ∂g dz + g −1 ∂g


∂Γ = − (2.35)
4π S2
Z
i ¯ dz ∧ dz̄
¯ −1 ∂g) dz̄ ∧ dz + ∂(g −1 ∂g)
tr g −1 δg ∂(g



=− (2.36)
4π S2
Z
i ¯ −1 ∂g) + ∂(g −1 ∂g)
¯
dz dz̄ tr g −1 δg − ∂(g



=− . (2.37)
4π S2

Adding the two contributions yields (2.31).

By a suitable choice of λ2 and k, we can achieve for one of the two brackets to vanish. We
consider k ∈ Z>0 and λ2 = π/k.7 This choice defines the WZW-model action. We obtain
¯
an antiholormophic current J¯ ≡ kg −1 ∂g:
¯ =0,
∂(g −1 ∂g) (2.38)

This implies also that the current J ≡ −k∂gg −1 is holomorphic:

¯
∂(∂gg −1 ¯ −1 = 0 .
) = g∂(g −1 ∂g)g (2.39)

Thus, the theory possesses now the desired holomorphic and antiholomorphic currents.
Notice however the asymmetry in the definitions of the holomorphic and the antiholomor-
phic current.
(Anti)holomorphicity arises from an enhanced symmetry of the action. The global
G × G symmetry extends to a local G(z) × G(z̄) symmetry acting as

g(z, z̄) 7→ gL (z)g(z, z̄)gR−1 (z̄) . (2.40)

Here gL (z) is an arbitrary holomorphic map S2 7→ G and gR (z̄) is an arbitrary antiholo-

morphic map.

3 The quantum theory

3.1 Current algebras
So far, the analysis was mostly classical. In particular, we have not yet shown that the
WZW-model is conformal on the quantum level. The usual way to move to the quantum
level is to compute the Poisson brackets and then canonically quantize by replacing them
7
For k < 0, the action would not be positive definite and path-integrals do not converge.

7
with commutators. This also fixes the OPE’s of the currents. Instead of taking this route,
we will take a shortcut. In fact, we will see that possible OPE’s are highly constrained.
So, let us start with the holomorphic current J a (z) (or rather its components). Here
and in the following, indices a, b, c, . . . always denote adjoint indices. By dimensional
analysis, this current has conformal weight one. Thus, J(z) = −k∂gg −1 has conformal
dimension one (at this point, we should rather only say dimension). Holomorphicity
protects the operator from acquiring an anomalous dimension in the quantum theory.
Similarly, the right-moving conformal weight is zero, and so the spin is h − h̄ = 1, as
expected. In fact, this is true in general — any conserved current of a CFT has conformal
weight (1, 0) or (0, 1). Hence, when writing out the OPE in the form
X Xp (w)
J a (z)J b (w) ∼ , (3.1)
p
(z − w)p

the holomorphic field Xp (w) has conformal weight 2 − p. Unitarity imposes severe re-
strictions on the possible field content of a CFT. In particular, there are no operators of
negative conformal dimensions. Thus, the highest possible pole order is a second order
pole. Furthermore, there is a unique field of conformal dimension zero, namely the iden-
tity. The field X1 (w) is of conformal dimension one and is hence a current itself! Thus,
the OPE between J a (z) and J b (w) has to take the form

κab if abc J c (w)

J a (z)J b (w) ∼ + (3.2)
(z − w)2 z−w

for some constants κab and f abc .

Symmetry under exchange of J a (z) and J b (w) implies κab = κba and f abc = −f bac .
Finally, associativity of the OPE gives the following additional conditions:

κcd f abd = κbd f cad = κad f bcd , f abd f dce + f bcd f dae + f cad f dbe = 0 . (3.3)

In other words, f abc are the structure constants of a Lie algebra and κab is a symmetric
invariant tensor. Since the currents reflect the symmetry of the model, the relevant Lie
algebra is g ≡ Lie(G). For simple Lie algebras, there is a unique invariant 2-tensor — the
Killing form. We may choose the basis of the Lie algebra such that it becomes proportional
to the identity. In this case, we have κab = kδ ab for some constant k. Thus, we finally
have the OPE structure
kδ ab if abc J c (w)
J a (z)J b (w) ∼ + , (3.4)
(z − w)2 z−w

where f abc are the structure constants of g. The parameter k is in fact identified with the
level k appearing in the action.
The OPE structure (3.4) is called current algebra. We have similarly an antiholomor-
phic copy satisfying
kδ ab if abc J c (w̄)
J¯a (z̄)J¯b (w̄) ∼ + . (3.5)
(z̄ − w̄)2 z̄ − w̄
This algebraic structure is the main organizing principle of WZW-models.

8
3.2 Affine Kac-Moody algebras
In this subsection, we reformulate the OPEs in terms of modes. Their equal-time com-
mutator gives an equivalent way of characterizing the algebraic structure. Since J a (z) is
holomorphic (or in fact meromorphic), we can consider its Laurent expansion around the
origin:
X
J a (z) ≡ Jna z −n−1 . (3.6)
n∈Z

where the −1 in the exponent is motivated by the fact that J a (z) is a field of conformal
weight 1. Note in particular that with this definition, J0a are the conserved charges of the
symmetry current, since I
J0a = dz J a (z) (3.7)

and the contour integral goes over an equal-time surface. We can compute the commuta-
tion relations by contour deformation:
I I I I !
a 1
[Jm , Jnb ] = dz dw − dz dw z m wn J a (z)J b (w) (3.8)
(2πi)2
|z|>|w| |z|<|w|

kδ ab if abc J c (z)
I I  
1 m n
= dw dz z w + (3.9)
(2πi)2 0 w (z − w)2 z−w
I
1
dw kmδ ab wm+n−1 + if abc J c (w)wm+n


= (3.10)
2πi 0
= kmδ ab δm+n,0 + if abc Jm+n
c
. (3.11)

This Lie algebra is called an affine Kac-Moody algebra. We note in particular that the zero
modes (the conserved charges) satisfy the algebra g. The additional term kmδ ab δm+n,0 is
a central term, in a similar way that the Virasoro algebra is a central extension of the
Witt algebra. In fact, to any simple Lie algebra g, we can associate the corresponding
affine Kac-Moody algebra, which is often denoted by b gk or simply gk . We stress that
this algebra is not a symmetry algebra of the theory. In fact, we shall see below that
a
all modes Jm with m 6= 0 do not commute with the Hamiltonian. Only the zero-modes
corresponding to the conserved charges form a symmetry algebra. For this reason, the
affine Kac-Moody algebra is sometimes referred to as spectrum-generating algebra.

3.3 The Sugawara construction

To show that the theory is conformal, we have to show that the energy-momentum tensor
satisfies the Virasoro algebra. In terms of OPE’s,
c 2T (w) ∂T (w)
T (z)T (w) ∼ 4
+ 2
+ . (3.12)
2(z − w) (z − w) z−w
This is the statement that the conformal algebra acts on the Hilbert space. To understand
this, we first have to identify the energy-momentum tensor of the model.
Classically, it is given by
1 1 a
T (z) = δab J a (z)J b (z) = J (z)J a (z) . (3.13)
2k 2k

9
A corresponding quantum version will need to implement normal ordering, in order for
the vacuum expectation value to be finite. It turns out that the prefactor is corrected in
the quantum theory and hence we will denote it for now by γ. Thus, we define

T (z) ≡ γ(J a J a )(z) , (3.14)

where normal ordering is defined as the constant part of the J a (z)J b (w) OPE. Hence all
short-distance singularities are subtracted. It is convenient to extract the constant part
via a contour integral:
I
a a 1 dx a
(J J )(z) ≡ J (x)J a (z) . (3.15)
2πi z x − z
As it should be, our energy-momentum tensor defined in this way is indeed holomorphic,
even in the quantum theory.
As a first step, we compute the OPE of T (z) with J a (w). We are only interested in
the singular part of the OPE, which we denote by a contraction.
I
a γ dx a
J (z) J b (x)J b (w)


J (z)T (w) = (3.16)
2πi w x − w
I
γ dx
J a (z)J b (x)J b (w) + J b (x)J a (z)J b (w)


= (3.17)
2πi w x − w
kδ ab if abc J c (x) b
I  
γ dx
= + J (w)
2πi w x − w (z − x)2 z−x
 ab ab c
!
kδ if c J (w)
+ J b (x) + . (3.18)
(z − w)2 z−w

Now, we have to evaluate the remaining JJ-OPE’s using (3.4). Here the singular and the
regular terms appear. However, since the structure constants are totally antisymmetric
in this basis and since the Kronecker symbol is symmetric, the second order pole does not
contribute.

kδ ab J b (w) kδ ab J b (x)
I
a γ dx
J (z)T (w) = +
2πi w x − w (z − x)2 (z − w)2
if abc if cbd J d (w)
 
c b
+ + (J J )(w)
z−x x−w
!
if abc if bcd J d (w)

+ + (J b J c )(w) (3.19)
z−w x−w
2kδ ab J b (w) f abc f cbd J d (w)
 
=γ − . (3.20)
(w − z)2 (w − z)2

We used that the normal ordered expressions vanish after integration, they cancel out by
(anti)symmetry in b and c.
We recall the definition of the dual Coxeter number h∨ , the quadratic Casimir of the
adjoint representation:

f abc f bcd = 2h∨ δ ad . (3.21)

10
Indeed, since f abc f bcd is an invariant 2-tensor, it has to be proportional to the Killing
form, which we chose to be the identity. With this definition, we finally obtain

T (z)J a (w) = J a (w)T (z) (3.22)

a
J (z)
= 2γ(k + h∨ ) (3.23)
(w − z)2
 a
∂J a (w)

∨ J (w)
= 2γ(k + h ) + . (3.24)
(z − w)2 z−w
In order for J a (z) to have conformal weight one, we define
1
γ≡ . (3.25)
2(k + h∨ )
Then we obtain
J a (w) ∂J a (w)
T (z)J a (w) ∼ + . (3.26)
(z − w)2 z−w
This is indeed the expected OPE of the stress-energy tensor with a primary field of
dimension one.
Now we still need to demonstrate that T (z)T (w) obeys the Virasoro algebra. We are
particularly interested in the central charge of the model. We compute
I
1 dx
T (z)T (w) = ∨
(T (z)J a (x)J a (w) + T (z)J a (x)J a (w)) (3.27)
4πi(k + h ) x−w
 a
∂J a (x) a
I 
1 dx J (x)
= + J (w) (3.28)
4πi(k + h∨ ) x−w (z − x)2 z−x
 a a
!
J (w) ∂J (w)
+ J a (x) + (3.29)
(z − w)2 z−w
I
1 dx k dim g 2k dim g
= −
4πi(k + h∨ ) x − w (z − x)2 (x − w)2 (z − x)(x − w)3
!
k dim g 2k dim g
+ +
(x − w)2 (z − w)2 (z − w)(x − w)3
(J a J a )(w) (∂J a J a )(w)
I
1 dx
+ +
4πi(k + h∨ ) x−w (z − x)2 z−x
!
(J a J a )(w) (J a ∂J a )(w)
+ + (3.30)
(z − w)2 z−w
(3 − 2 + 0 + 0)k dim g 2T (w) ∂T (w)
= + + . (3.31)
2(k + h∨ )(z − w)4 (z − w)2 z−w
Thus, the OPE gives indeed the Virasoro algebra with central charge
k dim g
c≡ . (3.32)
k + h∨
This finally demonstrates that the conformal algebra is also realized on the quantum level
and hence WZW-models indeed define CFTs. This construction of the energy-momentum
tensor is called Sugawara construction [4].

11
Exercise 2 Rederive the same result using the commutation relations (3.11). In
terms of modes, the normal ordering reads
X X X 
a a a a a a
Lm = γ : Jn Jm−n := γ Jn Jm−n + Jm−n Jn . (3.33)
n∈Z n≤−1 n≥0

You should obtain

a
[Lm , Jna ] = −nJm+n , (3.34)
c
[Lm , Ln ] = m(m2 − 1)δm+n,0 + (m − n)Lm+n , (3.35)
12
where the central charge is again given by (3.32).

Solution 2 We start with [Lm , Jna ]. We have

" #
X X
[Lm , Jna ] = γ Jrb Jm−r
b
+ b
Jm−r Jrb , Jna (3.36)
r≤−1 r≥0
(
X
Jrb k(m − r)δ ab δm+n−r,0 + if bac Jm+n−r
c


=γ
r≤−1

b 
+ krδ ab δr+n,0 + if bac Jr+nc
Jm−r
X
b
krδ ab δr+n,0 + if bac Jr+n
c


+ Jm−r
r≥0
)

b
+ k(m − r)δ ab δm+n−r,0 + if bac Jm+n−r
c
Jr . (3.37)

Let us first simply the central terms. This yields

( )
X X
[Lm , Jna ]central = γ k(m − r)δ ab Jrb δm+n−r,0 + krδ ab Jm−r
b
δr+n,0 (3.38)
r∈Z r∈Z
a
= −2γknJm+n . (3.39)

The other terms are a little more complicated. The trick is to try to bring the terms
back to normal ordering. Two of the terms are already correctly normal ordered,
the others need correction terms. This yields
(
X X
[Lm , Jna ]non-central = γi f bac Jrb Jm+n−r
c
+ f bac Jm+n−r
c
Jrb
r≤−1 r≥0
)
X X
+ b
f bac Jrc Jm+n−r + f bac Jm+n−r
b
Jrc (3.40)
r≤n−1 r≥n

12
 n−1
X
= if bac (J b J c )m+n + if bac (J c J b )m+n + if bac [Jrc , Jm+n−r
b
] (3.41)
r=0
n−1
X
d
if bac krδ bc δm+n,0 + if cbd Jm+n


=γ (3.42)
r=0
= γnf abc f cbd Jm+n
d
(3.43)
= −2γh∨ nJm+n a
. (3.44)

Here, we used the definition of the dual Coxeter number. The two normal ordered
terms cancel by renaming b ↔ c in one of them and by the antisymmetry of the
structure constants. In total, we hence obtain

[Lm , Jna ] = −2γ(k + h∨ )nJm+n

a a
= −nJm+n (3.45)

by definition of γ.
Next, we compute the Virasoro algebra:
" #
X X
[Lm , Ln ] = γ Lm , Jra Jn−r
a
+ a
Jn−r Jra (3.46)
r≤−1 r≥0
(
X
a a
− (n − r)Jra Jm+n−r
a


=γ − rJm+r Jn−r
r≤−1
)
X
a a a a


+ − rJn−r Jm+r − (n − r)Jm+n−r Jr (3.47)
r≥0
(
X X
(m − r)Jra Jm+n−r
a
(n − r)Jra Jm+n−r
a


=γ −
r≤m−1 r≤−1
X X
a
Jra a
Jra


+ (m − r)Jm+n−r − (n − r)Jm+n−r (3.48)
r≥m r≥0
! m−1
X X X
= γ(m − n) Jra Jm+n−r
a
+ a
Jm+n−r Jra +γ (m − r)[Jra , Jm+n−r
a
]
r≤−1 r≥0 r=0
(3.49)
m−1
X
= (m − n)Lm+n + γ (m − r)rkδ aa δm+n,0 (3.50)
r=0
k
= (m − n)Lm+n + γ dim(g) m(m2 − 1)δm+n,0 . (3.51)
6
Thus, the Virasoro algebra with the correct central charge

k dim(g)
c = 2kγ dim(g) = (3.52)
k + h∨
is satisfied.

The Sugawara construction shows that each affine Kac-Moody algebra has naturally
a Virasoro algebra contained in its universal enveloping algebra. The combined algebra

13
consisting of the affine Kac-Moody modes and the Virasoro modes forms a semidirect
product, as (3.34) shows that the affine Kac-Moody algebra is a Lie ideal inside the
combined algebra.
When g is only semisimple, i.e.
j
M
g= gj , (3.53)
i=1

we can simply define

j
X
Tg (z) = Tgi (z) . (3.54)
i=1

This again satisfies the Virasoro algebra with central charge

j
X
cg = cg i . (3.55)
i=1

In the semisimple case, we can choose k independently for all the individual factors.

3.4 The central charge

We now assume again that g is simple and consider the expression (3.32) in some detail.
We first note that the central charge is in general not a (half)integer, hence we are dealing
with a truly interacting theory.
The values of the dual Coxeter numbers for the simple Lie algebras and their dimen-
sions are listed in table 1.

g An Bn Cn Dn E6 E7 E8 F 4 G2
h∨ (g) n+1 2n − 1 n+1 2n − 2 12 18 30 9 4
dim(g) n2 + 2n 2n2 + n 2n2 + n 2n2 − n 78 133 248 52 14

Table 1: The dual Coxeter number for the simple Lie algebras. We are always using the
compact real form of these complex Lie algebras. The subscript always refers to the rank
of the Lie algebra. To physicists, the regular series are better known as An = su(n + 1),
Bn = so(2n + 1), Cn = sp(2n) and Dn = so(2n). We also tabulated the dimensions.

One can show that for a positive integer k, the central charge satisfies the bound

rank g ≤ cg < dim g . (3.56)

The lower bound is saturated precisely if g is simply-laced (i.e. of type A, D or E) and

k = 1. In particular, for these values, the central charge becomes an integer suggesting
that these theories are free-field theories in disguise. This turns out to be true and is
referred to as the Frenkel-Kac-Segal construction [5, 6]. The free theory consists of rank g
bosons on the Cartan torus of the group. For the algebras su(n) (actually u(n)) and
so(2n), there is alternatively a construction in terms of free fermions.

14
 We also notice that there is also another canonical value, where the central charge
attains a half-integer value, namely k = h∨ :
h∨ dim g 1
c(h∨ ) = ∨ ∨
= dim g . (3.57)
h +h 2
This suggests that we can represent this algebra in terms of dim g fermions. We shall see
this below.

3.5 Free field constructions

Let us now work out the free-field representations we mentioned above. We start with
so(n)1 , whose central charge is c = 12 n. Thus, we suspect that this theory might be
equivalent to n free (Majorana-Weyl) fermions. Indeed, the theory has Lagrangian
Z
S= d2 z ψ̄ i γ µ ∂µ ψ i , (3.58)
S2

where ψ i , i = 1, . . . , n are the fermion fields. This action clearly has an SO(n)-symmetry
acting by rotating the fermions. The corresponding currents are
1
J a (z) = Tija (ψ i ψ j )(z) , (3.59)
2
where Tija are generators of so(n) in the fundamental representation. By our general
argument, these currents have to satisfy an affine Kac-Moody algebra, which turns out to
have level one.
Similarly, we can start with dim g fermions and construct a current algebra via
i
J a (z) = f abc (ψ b ψ c )(z) . (3.60)
2
The level turns out to be h∨ .

Exercise 3 Based on the basic OPE

δ ij
ψ i (z)ψ j (w) ∼ , (3.61)
z−w
show that the current 12 Tija (ψ i ψ j )(z) indeed satisfies the so(n)1 current algebra.
Similarly show that starting from the OPE

δ ab
ψ a (z)ψ b (w) ∼ , (3.62)
z−w

the current J a (z) = 2i f abc (ψ b ψ c )(z) satisfies the algebra gh∨ .

Solution 3 We will show the following more general statement: For fermions
transforming in any representation labelled by a highest weight state λ (see also
next section), the bilinears J a = 21 Tija (ψ i ψ j ) give a Kac-Moody algebra of g at level

C(λ)dim(λ)
k= . (3.63)
2dim(g)

15
Here, C(λ) is the Casimir of the representation and dim(λ) its dimension. The ex-
ercise asks then about the cases where λ describes the fundamental representation
and the adjoint representation respectively. (For general representations, it is how-
ever not true that the Sugawara tensor of the current algebra coincides with the
Sugawara tensor of the free fermions.)
The computation can be either done in modes or OPEs, whatever one prefers.
Let us show it in the OPE language. The normal ordered product is again defined
by I
a 1 a dx
J (z) = Tij ψ i (x)ψ j (z) . (3.64)
2 z 2πi(x − z)
We then compute first the OPE between J a (z) and ψ ` (w):
I
a ` 1 a dx
J (z)ψ (w) ∼ Tij ψ i (x)ψ j (z)ψ ` (w) (3.65)
2 z 2πi(x − z)
 i` j
δ ψ (z) δ j` ψ i (x)
I 
1 a dx
∼ Tij − + (3.66)
2 z 2πi(x − z) x−w z−w
1  
∼ − T`ja ψ j (z) + Ti`a ψ i (w) (3.67)
2(z − w)
1  
∼ − T`ja ψ j (w) + Ti`a ψ i (w) (3.68)
2(z − w)
T a ψ i (w)
∼ − `i . (3.69)
z−w
Here, one has to be attentive to the minus sign due to the fermionic statistics. We
also used that SO(n) representation are antisymmetric matrices. This means that
the fermions will transform in the respective representation of the current, see eq.
(3.83) below. Next, we compute the OPE between the currents:
I
a b 1 b dx
J (z)J (w) ∼ Tij J a (z)ψ i (x)ψ j (w) (3.70)
2 w 2πi(x − w)
!
−Ti`a ψ ` (x)ψ j (w) −Tj`a ψ i (x)ψ ` (w)
I
1 b dx
∼ Tij + (3.71)
2 w 2πi(x − w) z−x z−w
Ti`a δ `j Tj`a δ i`
I
1 b dx
∼ Tij − −
2 w 2πi(x − w) (z − x)(x − w) (z − w)(x − w)
!
Ti`a (ψ ` ψ j )(w) Tj`a (ψ i ψ ` )(w)
− − (3.72)
z−x z−w
Tijb Tija Tijb Ti`a (ψ ` ψ j )(w) Tijb Tj`a (ψ i ψ ` )(w)
∼− + − (3.73)
2(z − w)2 2(z − w) 2(z − w)
a b a b j `
tr (T T ) [T , T ]j` (ψ ψ )(w)
∼ 2
+ (3.74)
2(z − w) 2(z − w)

Now let us evaluate the representation theoretic quantities. In the first order pole,
we have
[T a , T b ] = if abc T c , (3.75)

16
 this is the characterising property of a representation. For the second order pole,
we recall the definition of the quadratic Casimir:

(T a T a )ij = C(λ)δij . (3.76)

By taking the trace of this relation, we learn

C(λ)dim(λ) = tr (T a T a ) . (3.77)

On the other hand, tr (T a T b ) has to be an invariant tensor of g and hence propor-

tional to δ ab . By taking the trace we hence conclude that

C(λ) dim(λ) ab
tr (T a T b ) = δ . (3.78)
dim(g)

The constant is sometimes called the index of the representation. Thus, we conclude
that the current algebra is satisfied with level

C(λ)dim(λ)
k= . (3.79)
2 dim(g)

In particular, for so(n), we have dim(so(n)) = 21 n(n − 1). The Casimir of the
fundamental representation is n − 1 and hence we have

(n − 1)n
kfund = =1. (3.80)
n(n − 1)

For the adjoint representation we have instead

C(adj)
kadj = = h∨ . (3.81)
2

We learn that seemingly different conformal field theories may actually become equivalent
at the quantum level. The actions (2.28) and (3.58) look certainly very different, but turn
out to describe the same quantum theory! In particular, one action is bosonic and the
other is fermionic. The simplest incarnation of this fact is the observation that one
free boson (on a circle with a particular radius) is equivalent to two fermions. This
phenomenon is called non-abelian bosonization [7].
In general, for any CFT with a holomorphic conserved current, the theory will contain
a subsector with a Kac-Moody symmetry.

3.6 Representations
We have established that a Kac-Moody algebra acts on the Hilbert space of the theory.
Hence all states of the CFT will transform in representations of the algebra gk . Thus, we
will now discuss representations of these algebras. For this, we again restrict to simple
compact Lie groups. The non-compact case is much more complicated and much less
understood.
As for the Virasoro algebra, only highest weight representations of the Kac-Moody
algebra are physically relevant, since these have bounded energy spectrum from below.

17
Thus, there is a highest weight state |λi satisfying
Jna |λi = 0 for n > 0 . (3.82)
Furthermore, the zero-modes J0a satisfy the original Lie algebra g. In particular, we require
|λi to be a highest weight state of a highest weight representation of g, which is labelled
by λ.8 This then characterizes the representation completely.

Exercise 4 The Sugawara construction may be used to derive the analog of

the BPZ-equations for WZW-models. Show first that the primary condition (3.82)
translates in OPE language to

ta Φλ (w)
J a (z)Φλ (w) ∼ , (3.83)
z−w
where ta are the matrices of the representation of the zero-modes and Φλ is the field
corresponding to the primary state |λi. By inserting the identity
1
L−1 − (J a J a )−1 = 0 (3.84)
k + h∨
in a correlator of affine primary fields, show that the correlator obeys the Knizhnik-
Zamolodchikov equation [8]:
!
1 X tai ⊗ taj
∂zi − ∨
hΦλ1 (z1 ) · · · Φλn (zn )i = 0 . (3.85)
k + h i6=j zi − zj

Here, the subscripts in the generators tai indicate on which field they are acting.
Thus, the existence of the affine algebra constrains the n-point functions severely.
Solving this differential equation is challenging, but the four-point functions may be
computed in this fashion.

Solution 4 We have in general for an OPE:

∞ a
a λ
X V (Jp−1 |λi , w)
J (z)Φ (w) ∼ , (3.86)
p=1
(z − w)p

a a
where V (Jp−1 |λi , w) is the field associated to the state Jp−1 |λi via the operator-
state correspondence. Since |λi is primary, this is only non-vanishing for p = 1, in
which case V (J0a |λi , w) = ta V (|λi , w) = ta Φλ (w), hence (3.83) is the corresponding
statement for the OPEs.
Let us now consider the correlator
   
λ1 1 a a λi λn
0 = Φ (z1 ) · · · L−1 − (J J )−1 Φ (zi ) · · · Φ (zn ) . (3.87)
k + h∨

We have L−1 Φλi (zi ) = ∂zi Φλi (zi ). Let us now compute the bilinear current term.
By definition, (J a J a )−1 Φλi (zi ) is the first order pole of the OPE between (J a J a )(z)

8
λ can be understood as the Dynkin labels of the representation.

18
 and Φλi (zi ). Thus, we have
I
a a λi dz a a
(J J )−1 Φ (zi ) = (J J )(z)Φλi (zi ) (3.88)
2πi
Izi I
dz dx
= J a (x)J a (z)Φλi (zi ) . (3.89)
zi 2πi z 2πi(x − z)

Now note that we are computing the correlators on the Riemann sphere. Instead of
encircling z, we can let x encircle the complement of z. Possible singularities come
then from all zi ’s. Thus, we have

Φλ1 (z1 ) · · · (J a J a )−1 Φλi (zi ) · · · Φλn (zn ) (3.90)

n I I
X dz dx
=− J a (x)J a (z)Φλ1 (z1 ) · · · Φλn (zn ) (3.91)
j=1 zi
2πi zj 2πi(x − z)
n
taj
I I
X dz dx
=− J a (z)Φλ1 (z1 ) · · · Φλn (zn ) (3.92)
j=1 zi
2πi zj 2πi(x − z) x − zj
n
tai ⊗ taj
I
X dz
=− Φλ1 (z1 ) · · · Φλn (zn ) (3.93)
j=1 zi
2πi (z j − z)(z − z i )
n
X tai ⊗ taj
= Φλ1 (z1 ) · · · Φλn (zn ) . (3.94)
j=1
zi − zj

Let us exemplify the meaning of primary fields at the simplest example of g = su(2). In
a suitable basis, the affine Kac-Moody algebra reads9

3 k
[Jm , Jn3 ] = mδm+n,0 , (3.95)
2
3
[Jm , Jn± ] = ±Jm+n
±
, (3.96)
+ − 3
[Jm , Jn ] = kmδm+n,0 + 2Jm+n . (3.97)

A highest weight state can now be denoted by |`i and satisfies

Jna |`i = 0 for n > 0 , (3.98)

J0+ |`i = 0 , (3.99)
J03 |`i = ` |`i . (3.100)

Hence, the ground state transforms in the spin ` representation. We can now build up the
representation by acting with as many of the other oscillators as desired. More precisely,
this defines a representation
3 n32 + n+ − n− 3 n31 + n+ 0
V = · · · (J−2 ) 1 (J0− )n− |`i n+ − 3

) (J−2 ) 2 (J−1 ) 1 (J−1 ) (J−1 i , ni , ni ≥ 0 . (3.101)

This space is called the Verma-module of the representation. However, this representation
is typically not an irreducible representation, because of the existence of null-vectors.
9
This is not the basis we have been using so far in which the Killing form is proportional to the
identity. However, this basis is much more useful for understanding the representation theory.

19
To talk about such null-vectors, we first have to introduce the canonical norm on this
vectorspace. It is induced from the following hermitian properties:

(Jn+ )† = J−n
−
, (Jn3 )† = J−n
3
. (3.102)

This specifies a real form of the Kac-Moody algebra and implies that we are indeed
considering the compact form su(2).
A simple such null-vector arises already from the ground states, (J0− )2`+1 |`i = 0. But
there is a second null-vector taking the form
+ k+1−2`
|N i = (J−1 ) |`i = 0 . (3.103)

Since this fact is of central importance in what follows, we shall demonstrate it. We even
show the stronger property of being singular, meaning that this vector is again a highest
weight state of an affine representation. This implies in particular that it is null. To start,
J0+ annihilates this singular vector, since [J0+ , J−1
+
] = 0. We now show that the vector is
also annihilated by all modes J1 . This is obvious for J1+ , since [J1+ , J−1
a +
] = 0 and hence
+
we can simply commute the oscillator J1 through, where it hits the highest weight state
|`i. Next, we consider J13 . We again commute J13 through all oscillators:
k−2`
X
J13 (J−1
+ k+1−2`
) |`i = [J13 , (J−1
+ k+1−2`
) ] |`i = + m +
(J−1 + k−2`−m
) J0 (J−1 ) |`i = 0 , (3.104)
m=0

since J0+ can be commuted through, where it hits the highest weight state. Thus, we only
have to show that J1− also annihilates the state. This will complete the proof, since we
can obtain any other positive mode as a repeated commutator of J1a ’s. We calculate

J1− (J−1
+ k+1−2`
) |`i = [J1− , (J−1
+ k+1−2`
) ] |`i (3.105)
k−2`
X
+ m
= (J−1 ) (k − 2J03 )(J−1
+ k−2`−m
) ] |`i (3.106)
m=0
k−2`
X
+ m + k−2`−m
= (J−1 ) (k − 2(k − ` − m))(J−1 ) ] |`i (3.107)
m=0
k−2`
X
+ k+1−2`
= (2` + 2m − k)(J−1 ) |`i = 0 . (3.108)
m=0

+ N
Exercise 5 Show the following generalization of this result. The vector (J−1 ) |`i
has norm
N
Y
h`| (J1− )N (J−1
+ N
) |`i = n(k + 1 − n − 2`) , (3.109)
n=1

where we assume that h`|`i = 1. In particular, this norm is positive for N < k+1−2`
and zero for N ≥ k + 1 − 2`, provided that k ∈ Z.

20
 Solution 5 Let us proceed by induction over N . The case N = 0 is trivial. To
show N → N + 1, we compute

h`| (J1− )N (J−1

+ N
) |`i = h`| (J1− )N −1 [J1− , (J−1
+ N
) ] |`i (3.110)
N
X −1
h`| (J1− )N −1 (J−1
+ m

+ N −1−m
= ) k − 2J03 (J−1 ) |`i (3.111)
m=0
N
X −1
h`| (J1− )N −1 (J−1
+ m

+ N −1−m
= ) k − 2(` + N − 1 − m) (J−1 ) |`i
m=0
(3.112)
N
X −1
k − 2(` + N − 1 − m) h`| (J1− )N −1 (J−1
+ N −1


= ) |`i (3.113)
m=0
N
Y −1
= N (k + 1 − N − 2`) n(k + 1 − n − 2`) (3.114)
n=1
N
Y
= n(k + 1 − n − 2`) , (3.115)
n=1

where we used the induction hypothesis in the penultimate line.

It turns out that these two vectors together with their descendants are the only null-
vectors in the representation. Thus, we can obtain the corresponding irreducible rep-
resentation by dividing the Verma-module by the null-vector relations. We denote the
resulting space by M` .
The existence of this null-vector constrains the possible representations severely. Since
WZW-models (on compact Lie groups) are unitary, it is vital that no negative-norm
states are part of the representation. For this to be the case, the exercise shows that
two conditions have to be met. First, we see algebraically that k ∈ Z>0 . Indeed, if k
is not a positive integer, we cannot have k + 1 − 2` ∈ Z>0 for any value of ` ∈ 12 Z≥0 .
Hence these theories could not have any (unitary) representations. Second, this null-
vector has to occur at a positive level and hence k + 1 − 2` ≥ 1. So we conclude that only
representations with
k
`≤ (3.116)
2
define consistent unitary representations of the affine Kac-Moody algebra. In particular
there are only finitely many representations. A CFT with only finitely many representa-
tions is called rational and WZW-models are prime examples of rational CFTs. A similar
statement is true for any WZW-model based on a compact Lie group. Only finitely many
representations of g lift to unitary representations of gk . We will denote the modules in
general by Mλ , where λ ∈ R labels the allowed representations of the model.
Let us compute the conformal weight of the Kac-Moody highest weight state for an
arbitrary Lie group:
X 
1 a a
X
a a
L0 |λi = J J + J J |λi (3.117)
2(k + h∨ ) n≤−1 n −n n≥0 −n n

21
 1
= J a J a |λi (3.118)
2(k + h∨ ) 0 0
C(λ)
= |λi . (3.119)
2(k + h∨ )
Here, C(λ) denotes the value of the quadratic Casimir of the zero-mode representation
specified by λ. In particular for su(2), we have
`(` + 1)
h(|`i) = . (3.120)
k+2
Knowing the conformal weight of the highest weight state fixes the conformal weights of
all states in the representation. Indeed, from (3.34), we conclude that
m

a1 am
 C(λ) X
h J−n1 · · · J−nm |λi = + np . (3.121)
2(k + h∨ ) p=1

In other words, the conformal weight of a state equals the conformal weight of the highest
weight state plus the total mode number of oscillators applied to the ground state.

3.7 Characters
Characters are a convenient way of summarizing all states in a given representation. Let
us recall the character of a conformal representation. We define q = e2πiτ , where τ ∈ H
is in the upper half plane. Then the character of a given representation measures the
contribution of this representation to the partition function of the theory. We define
c

χ` (τ ) = trM` q L0 − 24 . (3.122)
Thus the character keeps track of all states in the representation and counts them with
their corresponding multiplicity.
Recall that the collection (χλ )λ∈R of characters of a rational conformal field theory
with allowed representations R transform into each other under modular transformations
as follows:
X
χλ (τ + 1) = Tλµ χµ (τ ) , (3.123)
µ∈R
X
χλ (− τ1 ) = Sλµ χµ (τ ) . (3.124)
µ∈R

Here T and S are finite-dimensional constant matrices.

In the context of the su(2)k WZW-model, it is convenient to refine the notion of
characters slightly by keeping track of the su(2)-charge as well. Thus, we define
c 3

χ` (z; τ ) = trM` q L0 − 24 y J0 , (3.125)
where y = e2πiz .10 z is usually called ‘chemical potential’, since it represents the chem-
ical potential associated to the U(1)-symmetry. The characters defined in this way still
transforms covariantly under modular transformations:
X
χλ (z; τ + 1) = Tλµ χµ (z; τ ) , (3.126)
µ∈R

10
Geometrically, z should be understood as a coordinate on the torus with modular parameter τ .

22
 kz 2
X
χλ ( τz ; − τ1 ) = e 2τ Sλµ χµ (z; τ ) . (3.127)
µ∈R

For the su(2)k WZW-model, the characters take the following form. Let us first define
the level-k theta functions
2
X
Θ(k)
m (z; τ ) ≡ q kn y kn . (3.128)
m
n∈Z+ 2k

Then su(2)k -characters are given by

(k+2) (k+2)
Θ2`+1 (z; τ ) − Θ−2`−1 (z; τ )
χ` (z; τ ) = (2) (2)
(3.129)
Θ1 (z; τ ) − Θ−1 (z; τ )
(k+2) (k+2)
Θ2`+1 (z; τ ) − Θ−2`−1 (z; τ )
= 1 1 1 Q . (3.130)
q 8 (y 2 − y − 2 ) ∞
n=1 (1 − q n )(1 − yq n )(1 − y −1 q n )

To go from the first to the second line, we used Jacobi’s triple product identity
∞
n n+ 21 21 (n+ 21 )2 1 1
− 12
X Y
(−1) y q = q (y − y )
8 2 (1 − q n )(1 − yq n )(1 − y −1 q n ) . (3.131)
n∈Z n=1

Such an identity exists for any simple Kac-Moody algebra, where it is called the Kac-Weyl
denominator formula.
Let us unpack the character. For this, we use the version (3.130). The infinite product
3 ±
in the denominator comes about as follows. We have three oscillators J−n and J−n of which
we can apply arbitrary many on the ground states to generate new states. Since these
oscillators are bosonic, they contribute

1 + q n + q 2n + q 3n + · · · = (1 − q n )−1 (3.132)

to the character. Two of the oscillators also have a charge, which explains the appearance
of the additional factors of y and y −1 . The rest of the expression accounts for the fact
that there is more than one ground state and null vectors in the module. Let us look at
the lowest order (in q) terms of the numerator. We have
1 )2
(`+ 2 3 )2
(k−`+ 2
(k+2) (k+2) 1 1
3 3

Θ2`+1 (z; τ ) − Θ−2`−1 (z; τ ) = q y `+ 2 − y −`− 2 − q k+2 y k−`+ 2 − y −k+`− 2 + · · ·
k+2

(3.133)
Let us also recall the finite dimensional su(2)-character. It is given by
` 1 1
su(2)
X
m y `+ 2 − y −`− 2
χ` (z) = y = 1 1 . (3.134)
m=−`, m+`∈Z y 2 − y− 2

Thus, the oscillator contribution is multiplied by

(`+ 1 2 (k−`+ 3 2
2) −1 su(2) 2) −1 su(2)
q k+2 8 χ` (z) − q k+2 8 χk+1−` (z) + · · ·
k `(`+1)
= q − 8(k+2) +
su(2) su(2)
(z) − q k+1−2` χk+1−` (z) + · · · . (3.135)


k+2 χ`
c
We recognize the prefactor as q h− 24 , where the central charge is given in (3.32) and the
conformal weight of the ground state in (3.120). We also understand the second term

23
 + k+1−2`
in the series as subtracting out the null-vector (J−1 ) |`i, which we have discussed
before at length. The next terms in the series correspond to the fact that we have also
subtracted all the descendants of the null-vector. However, some descendants are actually
not there and have to be put in again into the character. This pattern continues and
yields an alternating sum.
We now investigate the behaviour under modular transformations. Under the T-
modular transformation, q → qe2πi and the character picks up a phase. We conclude
`(`+1)


k
2πi k+2 − 8(k+2)
T``0 = e δ`,`0 . (3.136)

The behaviour under the S-modular transformation is much more interesting.

Exercise 6 By using the Poisson ressumation formula

X 2 +bn 1 X − π (m− b )2
e−πan =√ e a 2πi , (3.137)
n∈Z
a m∈Z

show first that

r k
−iτ 2πikz2 X 2πimm0 (k)
Θ(k) z 1
e−


m τ ; −τ = e 4τ Θm0 (z; τ ) . (3.138)
2k
2k m0 =−k+1

Deduce then that the S-matrix in (3.124) is given by

r
π(2` + 1)(2`0 + 1)
 
2
S``0 = sin . (3.139)
k+2 k+2

Check that the S-matrix is unitary and symmetric.

Solution 6 For completeness, let us first prove the Poisson ressumation formula.
Let f (x) be a (reasonable nice) function on the real line. The consider
X
F (x) = f (x + n) . (3.140)
n∈Z

By construction, this is a 1-periodic function. Hence we can compute its Fourier

series. The Fourier coefficients are given by
Z 1
am = dx F (x)e−2πimx (3.141)
0
XZ 1
= dx f (x + n)e−2πim(x+n) (3.142)
n∈Z 0
Z ∞
= dy f (y)e−2πimy (3.143)
−∞

= fˆ(m) . (3.144)

24
Here, fˆ(p) denotes the Fourier transform of f (x). Thus, we have
X X
F (x) = f (x + n) = fˆ(m)e2πimx . (3.145)
n∈Z m∈Z

Specialising to x = 0, we obtain the general form of the ressumation formula:

X X
f (n) = fˆ(m) . (3.146)
n∈Z m∈Z

The given formula is simply the special case of

2 +bx
f (x) = e−πax . (3.147)

Its Fourier transform is

Z ∞
2 +bx−2πipx 1 π b 2
fˆ(p) = dx e−πax = √ e− a (p− 2πi ) . (3.148)
−∞ a

Thus, the result follows.

Now, we apply this to the Theta-function. We have

X − 2πik (n+ m )2 + 2πizk (n+ m )
Θ(k) z 1
m τ ; −τ = e τ 2k τ 2k (3.149)
n∈Z
r r
−iτ 2πikz2 2k X − 2πik (n+ m − z )2
= e 4τ e τ 2k 2 . (3.150)
2k −iτ n∈Z

We already have the desired prefactor. The rest is precisely the right hand side of
the Poisson ressumation formula with a = −iτ
2k
and b = πi(z − m k
). Thus, we have
r
2k X − 2πik (n+ m − z )2 X πiτ r2 +πi(z− m )r
e τ 2k 2 = e 2k k . (3.151)
−iτ n∈Z r∈Z

We now write
r = 2kn (3.152)
0
where n ∈ Z + m 2k
and m0 ∈ {−k + 1, . . . , k}. This change of variables clearly still
covers every integer value of r precisely once. Thus:
k
πiτ r 2
+πi(z− m 2 +2πi(kz−m)n
X X X
)r
e 2k k = e2πiτ n (3.153)
r∈Z m0 =−k+1 n∈Z+ m0
2k
k
2πimm0
X (k)
= e− 2k Θm0 (z; τ ) . (3.154)
m0 =−k+1

This shows the first part.

25
 Now we work out the transformation behaviour of the su(2)k characters. The
denominator becomes:
2
√ πim0
 
πiz 2
(2) z (2) z
X (2)
1 1



Θ1 τ ; − τ − Θ−1 τ ; − τ = i −iτ e τ sin − Θm0 (z; τ ) (3.155)
m0 =−1
2
√ πiz 2 (2) (2)


= −i −iτ e τ Θ1 z; τ ) − Θ−1 z; τ . (3.156)

Thus, the denominator only receives a prefactor under the S-modular transforma-
tion. Now, let us look at the numerator:
r k+2
πi(2` + 1)m0
 
(k+2) (k+2) −2iτ 2πi(k+2)z2 X
Θ2`+1 τz ; − τ1 Θ−2`−1 τz ; − τ1



− = −i e 4τ sin
k+2 m0 =−k−1
k+2
(k+2)
× Θm0 (z; τ ) . (3.157)

Next, we observe that the right hand side is identically equal to zero if m0 = 0 or
m0 = k + 2. Furthermore, we can pair up m0 with −m0 and rename m0 = 2`0 + 1, `0
then runs over half-integer from 0 to k2 . Thus,

r k
2
πi(2` + 1)(2`0 + 1)
 
(k+2) (k+2) −2iτ 2πi(k+2)z2 X
Θ2`+1 τz ; − τ1 Θ−2`−1 τz ; − τ1



− = −i e 4τ sin
k+2 `0 =0
k+2
(k+2) (k+2)

× Θ2`0 +1 (z; τ ) − Θ−2`0 −1 (z; τ ) .
(3.158)

Finally, we can combine numerator and denominator and conclude

r k
2
πi(2` + 1)(2`0 + 1)
 
z 1

2 2πikz 2
X
χ` τ
; −τ = e 4τ sin χ`0 (z; τ ) . (3.159)
k+2 `0 =0
k+2

This proves that the S-matrix has indeed the form (3.139).
The S-matrix is obviously symmetric, so let us show that it is unitary. It is
convenient to set m = 2` + 1, which then runs from 1 to k + 1. We have
k
k+1
2
πmm00 πmm0
   
X † 2 X
S`` 00 S``0 = sin sin (3.160)
`=0
k + 2 m=1 k+2 k+2
k+2
πmm00 πmm0
   
1 X
= sin sin (3.161)
k + 2 m=−k−1 k+2 k+2
k+2
1 X  πm(m0 +m00 ) πm(−m0 −m00 )
=− e k+2 +e k+2
4(k + 2) m=−k−1
πm(m0 −m00 ) πm(−m0 +m00 )

−e k+2 −e k+2 . (3.162)

26
 We extended the range of summation, which does not change the result. The remain-
ing terms are now simple geometric series, which impose m0 = m00 or m0 = −m00 .
The latter cannot be by assumption and hence we conclude
k
2
X † 1
S`` 00 S``0 = − × 2 × (−2(k + 2))δm0 ,m00 = δ`0 ,`00 . (3.163)
`=0
4(k + 2)

Thus, the S-matrix is unitary.

3.8 Fusion rules

A conformal field theory has an important invariant: the fusion ring. If we bring two fields
φλ (z) and φµ (w) close together, their product again transforms in a certain representation.
In the context of WZW-models, the fusion rules express simply the Clebsch-Gordan coef-
ficients of affine Lie algebras. One may think of the fusion product as some sort of tensor
product. We denoted it as ×. We thus have a commutative, associative ring structure on
the set of allowed modules (Mλ )λ∈R of the theory:
M
Mλ × Mµ = Nλµ ν Mν . (3.164)
ν∈R

The coefficients Nλµ ν encode the multiplicity of the representation Mν appearing in the
fusion product. The Verlinde-formula establishes a surprising relation between the fusion
rules of the theory and its modular properties [9]. The S-matrix diagonalizes the fusion
rules and one can deduce X Sλσ Sµσ S ∗
νσ
Nλµ ν = . (3.165)
σ∈R
S0σ
Let us apply this again to the case of su(2)k . By using the explicit form of the S-matrix
(3.139), we can derive the fusion rules. One obtains
min(`1 +`2 ,k−`1 −`2 )
M
M`1 × M`2 = M` . (3.166)
`=|`1 −`2 |, `+`1 +`2 ∈Z

As a comparison, we recall the Clebsch-Gordan coefficients of two su(2)-representations.

They read
`M
1 +`2

(`1 ) ⊗ (`2 ) = (`) , (3.167)

`=|`1 −`2 |, `+`1 +`2 ∈Z

where (`) denotes the spin-` representation of su(2). The fact that the affine representa-
tions are cut off at ` = k2 leads to a folding-back of the representations appearing in the
tensor product.
Notice in particular
M` × M k = M k −` . (3.168)
2 2

A representation with the property that the fusion with any other representation yields
only a single representation is called a simple current. Simple currents are useful in
constructing modular invariants (see next section). The simple current Mk/2 has conformal
weight h = k4 and will lead to an extended modular invariant if k is divisible by four.

27
3.9 Modular invariants
Given that we know all possible irreducible modules of the su(2)k WZW-model, we now ask
how to consistently combine them to a full-fledged theory. This gives also the opportunity
to discuss conformal embeddings.
We declare the full Hilbert space of the theory to have the form
M
H= Mλµ Mλ ⊗ Mµ . (3.169)
λ∈R

Here, we have been general for a rational conformal field theory. λ runs over all allowed
representations, we denote the modules by Mλ . We should not forget that we actually
have two copies of affine Kac-Moody algebras in the model, one left-moving and one
right-moving. Correspondingly, the full Hilbert space is a linear combination of represen-
tations of the form Mλ ⊗ Mµ . The matrix Mλµ encodes the multiplicity of the respective
representation.
What consistency condition should Mλµ obey? We want to impose the following
conditions:
(a) Integrality. Mλν should be a matrix of non-negative integers.

(b) Uniqueness of the vacuum. The vacuum representation should appear exactly once
in the model. Hence M00 = 1, where λ = 0 labels the vacuum representation.

(c) Modular invariance. As we have discussed before, the partition function should be
invariant under modular transformations. This amounts to the conditions

S †M S = M ⇔ [S, M ] = 0 , (3.170)
T †M T = M ⇔ [T, M ] = 0 . (3.171)

Every conformal field theory has one obvious solution to these constraints, the so-called
diagonal modular invariant. It consists of taking M to be the identity, i.e.

Mλµ = δλµ . (3.172)

For the su(2)k WZW-model, there exists a complete classification of modular invariants
due to Cappelli, Itzykson and Zuber [10, 11]. There are three types of modular invariants:
(a) A-type. The A-type modular invariant corresponds to the diagonal modular invari-
ant. It exists for every level k ∈ Z≥1 .

(b) D-type. This modular invariant exists for even level and takes the form
n−1
M
k = 4n : H = 2 Mn ⊗ Mn ⊕ (M` ⊕ M2n−` ) ⊗ (M` ⊕ M2n−` ) , (3.173)
`=0, `∈Z
2n−1
M
k = 4n − 2 H = Mn− 1 ⊗ Mn− 1 ⊕ M` ⊗ M`
2 2
`=0, `∈Z
n− 32
M
M` ⊗ M2n−1−` ⊕ M2n−1−` ⊗ M` .


⊕ (3.174)
`= 12 , `∈Z+ 12

28
 (c) E-type. These are three exceptional invariants existing at the levels k = 10, k = 16
and k = 28. They take the form

k = 10 : H = (M0 ⊕ M3 ) ⊗ (M0 ⊕ M3 ) ⊕ (M 3 ⊕ M 7 ) ⊗ (M 3 ⊕ M 7 )
2 2 2 2

⊕ (M2 ⊕ M5 ) ⊗ (M2 ⊕ M5 ) , (3.175)

k = 16 : H = (M0 ⊕ M8 ) ⊗ (M0 ⊕ M8 ) ⊕ (M2 ⊕ M6 ) ⊗ (M2 ⊕ M6 )
⊕ (M3 ⊕ M5 ) ⊗ (M3 ⊕ M5 ) ⊕ M4 ⊗ M4
⊕ M4 ⊗ (M1 ⊕ M7 ) ⊕ (M1 ⊕ M7 ) ⊗ M4 , (3.176)
k = 28 : H = (M0 ⊕ M5 ⊕ M9 ⊕ M14 ) ⊗ (M0 ⊕ M5 ⊕ M9 ⊕ M14 )
⊕ (M3 ⊕ M6 ⊕ M8 ⊕ M11 ) ⊗ (M3 ⊕ M6 ⊕ M8 ⊕ M11 ) . (3.177)

The terminology ADE comes from the fact that the problem can be mapped to the
classification of simply-laced Lie algebras.
The modular invariants have the following physical interpretation. The A-type mod-
ular invariant defines the SU(2) WZW-model at level k. It contains every representation
exactly once. The D-type modular invariant corresponds to the SO(3) WZW-model. In-
deed, as we have remarked in footnote 6, this is exactly the quantization condition on the
level. We will comment on the meaning of the three exceptional theories below.
Let us have a closer look at the k = 4 D-type modular invariant. Explicitly, we have
by (3.173)
H = (M0 ⊕ M2 ) ⊗ (M0 ⊕ M2 ) ⊕ 2 M1 ⊗ M1 . (3.178)
Thus the vacuum M0 is combined with the spin-2 representation M2 . The conformal
weight of the spin-2 representation is

2(2 + 1)
h(|2i) = =1. (3.179)
4+2
Thus, there are in fact 5 more holomorphic spin-1 fields in the model, which by our
previous discussion constitute five more conserved currents. Thus, we have in total 8 con-
served currents, which turn out to generate su(3). The D-type level-4 modular invariant
has hence in fact su(3)-symmetry! We thus conclude that

D-type su(2)4 = su(3)k . (3.180)

What should be the level k of the su(3)-theory? This can be determined by comparing
central charges and yields k = 1. In the su(3)1 -language, the modular invariant (3.178)
becomes either one of

H = M1 ⊗ M1 ⊕ M3 ⊗ M3 ⊕ M3̄ ⊗ M3̄ , (3.181)

H = M1 ⊗ M1 ⊕ M3 ⊗ M3̄ ⊕ M3̄ ⊗ M3 . (3.182)

The three allowed modules of the su(3)1 theory are the vacuum module M1 and the mod-
ule based on the fundamental (antifundamental) representation M3 (M3̄ ). Both Hilbert
spaces are possible, the first corresponds to the diagonal modular invariant, the second
to the charge conjugated modular invariant.
This phenomenon is called modular extension. For a modular extension, we choose a
modular invariant in which the vacuum gets combined with some other representations
and we obtain a bigger chiral algebra. This happens for all D-type invariants of level

29
k = 4n and the three exceptional invariants. If the new fields have spin-1, these modular
extensions are associated to conformal embeddings. A conformal embedding of two Kac-
Moody algebras is an embedding hk0 ⊂ gk such that the two Sugawara-tensors become
identified. A necessary and sufficient condition for this to happen [12] is the requirement
of equal central charges
c(hk0 ) = c(gk ) . (3.183)
We have seen above the example su(2)4 ⊂ su(3)1 , where c(su(2)4 ) = c(su(3)1 ) = 2.

Exercise 7 Show that the existence of the E-type modular invariants at level
k = 10 and k = 28 are explainable by the conformal embeddings

su(2)10 ⊂ so(5)1 , su(2)28 ⊂ (g2 )1 . (3.184)

The data of table 1 might be useful. Recast the structure of the Hilbert space in
terms of so(5)1 and (g2 )1 -modules. so(5)1 has the following allowed modules:

M1 , M4 , M5 , (3.185)

corresponding to the modules based on the trivial, the spinor and the vector repre-
sentation. (g2 )1 has only two allowed modules

M1 , M7 , (3.186)

based on the trivial and the vector representation.

Solution 7 We are given the two conformal embeddings (3.184), whose existence
we assume. Let us first discuss so(5)1 , which has the three modules M1 , M4 and
M5 . These three modules have to decompose into su(2)10 modules in some way. We
know that su(2)10 has the following modules M` for ` = 0, 21 , . . . , 5. Their conformal
weights are
1 1 5
h(M0 ) = 0 , h(M 1 ) = 16
, h(M1 ) = 6
, h(M 3 ) = 16
, (3.187)
2 2
1 35 21
h(M2 ) = 2
, h(M 5 ) = 48
, h(M3 ) = 1 , h(M 7 ) = 16
, (3.188)
2 2
5 33 5
h(M4 ) = 3
, h(M 9 ) = 16
, h(M5 ) = 2
. (3.189)
2

Let us assume the decomposition

5
M
M1 = N` M` (3.190)
`=0

of the vacuum module of the so(5)1 theory, where N` are non-negative integers.
Since the vacuum representation of so(5)1 contains only integer conformal weights
by (3.121), the only modules which can appear on the right hand side are actually
M0 and M3 . How many times do they appear? Since the vacuum appears precisely
once, the su(2)10 vacuum representation has to appear precisely once. Thus

M1 = M0 ⊕ N3 M3 (3.191)

30
 for some integer N3 . N3 is fixable by counting the h = 1 states. These correspond
by definition the Kac-Moody fields of so(5)1 , so M1 contains 21 5 × 4 = 10 h = 1
states, whereas su(2)10 contains only 3. Now we observe that M3 contains precisely
7 h = 1 states (since the spin 3 representation is 7-dimensional). Thus we have to
set N3 = 1 in order to get a matching. We conclude

M1 = M0 ⊕ M3 , (3.192)

and this is indeed the combination which appears in the modular invariant. For the
other factors, one can argue similarly and one concludes

M4 = M 3 ⊕ M 7 , (3.193)
2 2

M5 = M2 ⊕ M5 . (3.194)

Thus, the E6 modular invariant takes in so(5)1 language the diagonal form

H = M1 ⊗ M1 ⊕ M4 ⊗ M4 ⊕ M5 ⊗ M5 . (3.195)

Similarly, one can argue in the (g2 )1 case to obtain simply

H = M1 ⊗ M1 ⊕ M7 ⊗ M7 . (3.196)

This explains at least the physical origin of the k = 10 and k = 28 E-type invariants.
The k = 16 E-type invariant is closely related to the k = 16 D-type invariant and can be
obtained by a permutation of fields.
A similar classification exists for su(3)k WZW-models. In addition to the A-type and
D-type series and their charge conjugated versions, there are five exceptional invariants.
Some of them can be explained by the conformal embeddings

su(3)5 ⊂ su(6)1 , su(3)9 ⊂ (E6 )1 , su(3)21 ⊂ (E7 )1 . (3.197)

A general classification for arbitrary Lie algebras is not known and is an important open
problem.

4 Cosets
In this final chapter, we will briefly introduce the coset construction. Cosets are one of
the main examples of rational conformal field theories.

4.1 The coset construction

A coset is a gauged WZW-model, where a subgroup of G is gauged. While this can be
described on the level of the action, we will focus on the algebraic construction. We will
illustrate the concepts with the following two theories:
su(2)k
Parafermions : , (4.1)
u(1)
su(2)k ⊕ su(2)1
Minimal models : . (4.2)
su(2)k+1

31
As one can see, we usually denote coset theories as quotients of two affine Kac-Moody al-
gebras. For the parafermion theory, the denominator is embedded as the Cartan-generator
of the numerator theory. In the minimal model series, the denominator is embedded as
the diagonal subalgebra of su(2)k ⊕ su(2)1 . The names of the two theories are traditional
and will become clear in the following.
The coset construction starts with a WZW-model gk and the subalgebra hk0 ⊂ gk to
be gauged. The chiral algebra (i.e. the holomorphic fields of the theory) consists of all
holomorphic fields in the gk WZW model which have regular OPE with the Kac-Moody
fields of hk0 .
This is indeed an algebra for the following reason. Let X(z) and Y (z) be two fields
with trivial OPE with the Kac-Moody fields of hk0 , which we denote by K a (z). Then we
have to check that the normal-ordered product of X with Y has also trivial OPE with
K a (z). We compute
I
a 1 dx
K (z)(XY )(w) = K a (z)X(x)Y (w) . (4.3)
2πi w x − w
But the OPE of K a (z) with both X(x) and Y (w) is regular and hence the complete OPE
is regular.
We illustrate this by computing the vacuum module of the parafermion theory. Let
us assume that the level k is large to avoid the complication of null-vectors. The chiral
algebra consists of all fields having trivial OPE with J 3 (z). By definition, neither J − (z),
J + (z), nor J 3 (z) have this property. So, we have to look at composite fields. There is
exactly one bilinear field with this property and two trilinear fields. They are given by
(J + J − ) − ∂J 3 2(J 3 J 3 )
T = − , (4.4)
k+2 k(k + 2)
(∂J + J − ) + (J + ∂J − ) − ∂ 2 J 3 4(∂J 3 J 3 )
∂T = − , (4.5)
k+2 k(k + 2)
4(J 3 (J 3 J 3 )) k k∂ 2 J 3
W (3) = − (J 3 (J + J − )) + (∂J + J − ) − (J + ∂J − ) + (∂J 3 J 3 ) +


.
3k 4 12
(4.6)
We gave the spin-2 field the suggestive name T , since it will be the energy-momentum
tensor of the coset theory. One spin-3 field is the derivative of the energy-momentum
tensor, whereas the other one is a primary (with respect to the coset energy-momentum
tensor T ) spin-3 field, which we denoted by W (3) . Continuing like this, we obtain one
primary field at every spin. For finite k, the k-th field turns out to be a null field. Thus the
chiral algebra is generated by k − 1 fields of spin 2, . . . , k. It is also known as Wk -algebra.
The fact that this model is constructed via a coset is somewhat obscured, in particular,
there is no trace of an affine Kac-Moody symmetry left in the chiral algebra.
Repeating the same calculation for the theory (4.2) shows that there is no spin-3
primary field. It turns out that the chiral algebra of the coset is given only by the
Virasoro algebra. For this reason, it gives a concrete realisation of the minimal-model
series.
Let us show that there is always an energy-momentum tensor in the coset. For this,
we denote by T g the energy-momentum tensor in the numerator theory and by T h the
energy-momentum tensor in the denominator theory. Then we claim that

T g/h = T g − T h (4.7)

32
is the energy-momentum tensor of the coset. For this, we first have to check that its
OPE with any Kac-Moody generator K a (z) of h is non-singular. By definition, K a (z) is a
primary field of spin 1 with respect to the energy-momentum tensor T h . But since K a (z)
is in particular also in the numerator algebra gk , it is also a primary field of spin 1 with
respect to the energy-momentum tensor T g . Hence, T h and T g have identical OPEs with
K a (z), so the OPE of T g/h with K a is non-singular. Hence, T g/h defines indeed a field in
the coset. Next, we have to check that (4.7) satisfies the Virasoro algebra. For this, we
note that by the same argument as for the currents themselves, the OPE between T g/h
and T h is non-singular. Thus, we obtain

T g/h (z)T g/h (w) = (T g (z) − T h (z))T g (w) (4.8)

= T g (z)T g (w) − T h (z)T h (w) − T h (z)T g/h (w) (4.9)
cg − ch 2T g/h (w) ∂T g/h (w)
∼ + + . (4.10)
2(z − w)4 (z − w)2 z−w
This shows that the coset theory again defines a CFT with central charge c = cg − ch . For
the central charges of the theories (4.1) and (4.2), we conclude in particular
 
su(2)k 3k 2(k − 1)
c = −1= , (4.11)
u(1) k+2 k+2
 
su(2)k ⊕ su(2)1 3k 3·1 3(k + 1) 6
c = + − =1− . (4.12)
su(2)k+1 k+2 1+2 k+3 (k + 2)(k + 3)
Observe that in the special case of k = 2, the parafermion theory has central charge
c = 12 and turns out to be equivalent to the free-fermion theory (with a suitable modular
invariant). This is where the name parafermion stems from. On the other hand, the
central charge of the minimal model coset is always smaller than one. The minimal model
coset gives a concrete realization of the minimal Virasoro models.

4.2 Representations and characters

After having determined the chiral algebra of the theory, we study its representations.
Constructing them is quite easy. For the chiral algebra, we considered fields of the numer-
ator algebra in the vacuum representation which transformed in the trivial representation
of the numerator algebra. We can expand this to fields in any representation of the nu-
merator algebra with some fixed representation of the denominator algebra. Concretely,
let us consider some module Mgλ of the parent theory. Then we decompose it into primary
representations of the denominator algebra:
M g/h
Mgλ = M(λ,µ) ⊗ Mhµ . (4.13)
µ∈Rh

g/h
M(λ,µ) forms then a module of the coset algebra. In fact, all modules are constructed in
this way. Thus, representations are labelled by a pair of indices (λ, µ) ∈ Rg × Rh .
Not all pairs (λ, µ) might be allowed. This is formalised by the so-called selection
rules. They express the fact that not all representations of hk0 might appear in a given
representation of gk . Indeed, an example of this is given by the parafermion theory.
Integer-spin representations of su(2)k contain only integer u(1)-charges, whereas half-
integer-spin representations contain only half-integer u(1)-charges.

33
 There is a further subtlety, which is in some sense complementary. While not all pairs
(λ, µ) might be allowed, some might be equivalent. This is referred to as field identifica-
tions. The reason for this is the existence of outer automorphisms of the parent theory,
which act trivially on the coset algebra. In the parafermion theory, the corresponding
field identification is (`, m) ∼
= ( k2 − `, k2 − m).11
We can apply characters to (4.13). This implies that we have the character identity
X g/h
χgλ (τ ) = χ(λ,µ) (τ )χhµ (τ ) . (4.14)
µ∈Rh

Thus, given the characters of the numerator and denominator theory, one can work out
the characters of the coset theory. Similarly, the modular transformation behaviour can
be deduced. We claim that the S-matrix of the coset theory is given by
g/h g h −1


S(λ,µ)(λ0 ,µ0 ) = Sλλ 0 S
µµ0
. (4.15)

To show this, we simply have to apply the S-modular transformation to both sides of (4.14)
and check that they indeed agree. This then also implies that modulo field identifications,
modular invariants of the coset theory are given by products of modular invariants of the
numerator theory and the denominator theory. Similarly, we conclude via the Verlinde
formula that the fusion rules of the coset theory are the product of the fusion rules of
numerator and denominator theory.
For illustration, we work this out for the minimal model theory. Representations
are in principle labelled by three spins (`1 , `2 ; j), where 0 ≤ `1 ≤ k2 , 0 ≤ `2 ≤ 21 and
0 ≤ j ≤ k+12
. There is again a selection rule, similar to the parafermion theory. We have
`1 + `2 + j ∈ Z and hence `2 is determined in terms of the two other labels. For this
reason, it is customary to suppress `2 and use only two labels (` = r−1 2
, j = s−1
2
). Using
(4.7), the representation (r, s) has conformal weight
r−1 r−1

s−1 s−1

(
+ 1 + 1 n r − s ∈ 2Z ,
∆(r,s) = 2 2
− 2 2
+ 1
, (4.16)
k+2 k+3 n + 4 r − s ∈ 2Z + 1

where n is a non-negative integer. It can appear, since the representation j = s−1

2
does
r−1
not necessarily appear in the ground state of the representation ` = 2 , but higher up
in the Verma module. For an appropriate choice of n, this becomes

2
r(k + 3) − s(k + 2) − 1
∆(r,s) = , (4.17)
4(k + 2)(k + 3)
which matches the conformal dimensions of primary fields in the unitary minimal models.
We deduce the minimal model fusion rules:
min(r1 +r2 −1,2k+3−r1 −r2 ) min(s1 +s2 −1,2k+5−s1 −s2 )
M M
Mr1 ,s1 × Mr2 ,s2 = Mr3 ,s3 . (4.18)
r3 =|r1 −r2 |+1, s3 =|s1 −s2 |+1,
r1 +r2 +r3 +1∈2Z s1 +s2 +s3 +1∈2Z

Similarly, we can reproduce the existence of the D-series minimal models by the D-series
modular invariant of the su(2)k WZW-model.

11
We mean the extended u(1)k algebra here, so there is an identification m ∼ m + k of u(1)k -
representations.

34
Exercise 8 Reproduce the spectrum of the D-series minimal models you learned
about it Sylvain’s lectures from the D-series su(2)k modular invariants. Find then all
modular invariants of minimal models, including the exceptional modular invariants.

Solution 8 We have p = k +2 and q = k +3 (or vice versa). The D-series minimal

models only exist for p even, which means also k even. So we use the D-modular
invariant for the numerator theory and the A-modular invariant for the denominator
theory. There are two cases, corresponding to the two cases (3.174) and (3.173).
k = 4n: The modular invariant reads by definition
q−1 2n−1
!
M M
2H = 2 M2n+1,s ⊗ M2n+1,s ⊕ (Mr,s ⊕ M4n−r+2,s ) ⊗ (Mr,s ⊕ M4n−r+2,s )
s=1 2
r=1
(4.19)
q−1 4n+1 4n+1
!
M M M
= Mr,s ⊗ Mr,s ⊕ Mr,s ⊗ M4n−r+2,s (4.20)
s=1 2 2
r=1 r=1
p−1 q−1
M M
Mr,s ⊗ Mr,s ⊕ Mr,s ⊗ Mp−r,s .


= (4.21)
2 s=1
r=1

2
Here, r = 1 to 4n − 1 means that r runs over all odd numbers.
k = 4n − 2: The modular invariant reads in this case
q−1 4n−1
M M
2H = M2n,s ⊗ M2n,s ⊕ Mr,s ⊗ Mr,s
s=1 2
r=1
2n−2
M 
Mr,s ⊗ M4n−r,s ⊕ M4n−r,s ⊗ Mr,s


⊕ (4.22)
2
r=2
q−1 4n−1 4n−2
!
M M M
= Mr,s ⊗ Mr,s ⊕ Mr,s ⊗ M4n−r,s (4.23)
s=1 2 2
r=1 r=2
p−1 q−1 p−2 q−1
MM MM
= Mr,s ⊗ Mr,s ⊕ Mr,s ⊗ M4n−r,s . (4.24)
2 s=1 2 s=1
r=1 r=2

Thus, in total we obtain

p−1 q−1 q−1
1 MM 1 M M
H= Mr,s ⊗ Mr,s ⊕ Mr,s ⊗ M4n−r,s . (4.25)
2 2 s=1 2 p s=1
r≡ 2 mod 2
r=1
1≤r≤p−1

This reproduces equation (4.10) in Sylvain’s lecture notes.

The modular invariants are in general of type (A, A), (D, A), (A, D), (E, A) or
(A, E), where the first entry denotes the numerator theory and the second entry the

35
 denominator theory. At least one entry has to be of A-type since not both the levels
of the numerator as well as of the denominator can be even.

We have seen that there are surprising identifications of CFTs at the quantum level. We
mention here the somewhat surprising level-rank duality. It states the equivalence of the
conformal field theories
su(M + N )k ∼ su(k)M ⊕ su(k)N
= . (4.26)
su(M )k ⊕ su(N )k ⊕ u(1) su(k)M +N

We have seen the simplest instance of this equivalence, namely for k = 2 and M = N = 1,
where both sides describe the Ising model. Let us check at least the equality of central
charges:

M (k 2 − 1) N (k 2 − 1) (M + N )(k 2 − 1)
 
su(k)M ⊕ su(k)N
c = + − (4.27)
su(k)M +N k+M k+N k+M +N
k((M + N )2 − 1) k(M 2 − 1) k(N 2 − 1)
= − − −1 (4.28)
 k + M + N k + M k + N
su(M + N )k
=c . (4.29)
su(M )k ⊕ su(N )k ⊕ u(1)

5 Outlook
Let us mention a few topics we did not have time to mention or elaborate on.

(i) Affine Kac-Moody algebras. We have introduced affine Kac-Moody algebras in

passing. There is a mathematical theory of these algebras, developing their rep-
resentations, characters, etc. from a purely mathematical point of view. This is
described in detail in chapter 14 of [3] or chapter 3 of [1].

(ii) Correlation functions. We have mentioned correlation functions only in passing.

Their form is fixed by the Knizhnik-Zamlodochikov equations [8].

(iii) Non-compact WZW-models. We have almost exclusively discussed the compact

case. For a non-compact Lie group, the story is much more complicated: The level
is no longer quantised, infinite-dimensional ground state representations appear and
spectrally flowed representations appear. Moreover, they typically contain indecom-
posable modules and fall into the realm of logarithmic conformal field theory. Good
general overviews of logarithmic conformal field theories are [13, 14]. Arguably
the simplest and most important non-compact WZW-model is the SL(2, R) WZW-
model. An introduction can be found in the concluding chapter of [15], a more
extensive review in [16].

(iv) Supersymmetry. We have not explained how to make WZW-models supersymmet-

ric. Every WZW-model can be made N = 1 supersymmetric by adding fermions
transforming in the adjoint representation. Some cosets are moreover N = 2 (or
even large N = 4) supersymmetric. The better known case of N = 2 is referred to
as Kazama-Suzuki models [17].

36
 (v) Relation to string theory. While WZW-models provide interesting CFTs in their
own right, they have important applications to string theory. The addition of the
WZ-term to the action corresponds to adding NS-NS flux to the background. Thus,
all pure NS-NS background string theories on group manifolds can be described by
WZW-models. This includes in particular the important case of AdS3 -backgrounds.
For instance, the maximally supersymmetric background AdS3 × S3 × T4 can be
described by the worldsheet theory sl(2, R)k+2 ⊕ su(2)k−2 ⊕ u(1)4 ⊕ 10 free fermions.
The level k has then the interpretation of the amount of B-field flux.

(vi) W-algebras. Many W-algebras can be constructed via cosets and Casimir algebras,
thereby providing an explicit definition. A general overview can be found in [18].

(vii) Admissible levels. While we have defined WZW models at positive integer levels,
they can actually be defined also at other levels (in which case they define a non-
unitary theory). A particularly interesting class of levels is given by the so-called
admissible levels, which have the property that they have still null vectors in their
Verma modules. A good introduction can be found in the yellow book [3].

References
[1] P. Goddard and D. I. Olive, Kac-Moody and Virasoro Algebras in Relation to
Quantum Physics, Int. J. Mod. Phys. A1 (1986) 303.

[2] R. Blumenhagen and E. Plauschinn, Introduction to conformal field theory, Lect.

Notes Phys. 779 (2009) 1.

[3] P. Di Francesco, P. Mathieu and D. Senechal, Conformal Field Theory, Graduate

Texts in Contemporary Physics. Springer-Verlag, New York, 1997.

[4] H. Sugawara, A Field theory of currents, Phys. Rev. 170 (1968) 1659.

[5] I. B. Frenkel and V. G. Kac, Basic Representations of Affine Lie Algebras and Dual
Resonance Models, Invent. Math. 62 (1980) 23.

[6] G. Segal, Unitarity Representations of Some Infinite Dimensional Groups,

Commun. Math. Phys. 80 (1981) 301.

[7] E. Witten, Nonabelian Bosonization in Two-Dimensions, Commun. Math. Phys. 92

(1984) 455.

[8] V. G. Knizhnik and A. B. Zamolodchikov, Current Algebra and Wess-Zumino

Model in Two-Dimensions, Nucl. Phys. B247 (1984) 83.

[9] E. P. Verlinde, Fusion Rules and Modular Transformations in 2D Conformal Field

Theory, Nucl. Phys. B300 (1988) 360.

[10] A. Cappelli, C. Itzykson and J. B. Zuber, Modular Invariant Partition Functions in

Two-Dimensions, Nucl. Phys. B280 (1987) 445.

[11] A. Cappelli, C. Itzykson and J. B. Zuber, The ADE Classification of Minimal and
(1)
A1 Conformal Invariant Theories, Commun. Math. Phys. 113 (1987) 1.

37
[12] P. Goddard, A. Kent and D. I. Olive, Virasoro Algebras and Coset Space Models,
Phys. Lett. 152B (1985) 88.

[13] M. R. Gaberdiel, An Algebraic approach to logarithmic conformal field theory, Int.

J. Mod. Phys. A18 (2003) 4593 [hep-th/0111260].

[14] T. Creutzig and D. Ridout, Logarithmic Conformal Field Theory: Beyond an

Introduction, J. Phys. A46 (2013) 4006 [1303.0847].

[15] S. Ribault, Conformal field theory on the plane, 1406.4290.

[16] W. McElgin, Notes on the SL(2, R) CFT, 1511.07256.

[17] Y. Kazama and H. Suzuki, New N = 2 Superconformal Field Theories and

Superstring Compactification, Nucl. Phys. B321 (1989) 232.

[18] P. Bouwknegt and K. Schoutens, W symmetry in conformal field theory, Phys.

Rept. 223 (1993) 183 [hep-th/9210010].

38