INFINITE DIMENSIONAL ALGEBRAS

Notes taken from a course of E. Frenkel, Fall 2002, Berkeley

Edward Frenkel,
Mathematics department, Evans Hall, UC Berkeley, CA 94720, U.S.A.
e–mail: [email protected]
home page: www.math.berkeley.edu/~frenkel

Alex Barnard,
Mathematics department, Evans Hall, UC Berkeley, CA 94720, U.S.A.
e–mail: [email protected]
home page: www.math.berkeley.edu/~barnard

Contents

1 Introduction 3

1.1 Affine Kac–Moody Algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.2 Representations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

1.3 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

1.4 Universal Enveloping Algebras . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2 Vertex Algebras 13

2.1 More on Locality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

2.2 A vertex algebra associated to an AKMA . . . . . . . . . . . . . . . . . . . . 18

2.3 Associativity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

3 The AKMA vertex algebra 31

3.1 The Virasoro Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

3.1.1 Detour: Why do central extensions keep appearing? . . . . . . . . . . 36

3.2 Lie algebras associated to vertex algebras . . . . . . . . . . . . . . . . . . . . 37

1
 3.3 The centre of Ũκ (ĝ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

3.4 Jet schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

3.5 Proof of the theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

3.6 Back to the centre . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

4 Geometric Interpretations 52

4.1 Twisting by torsors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

4.2 Action of coordinate changes . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

4.3 Projective structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

4.4 Opers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

4.5 Miura Opers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

2
1 Introduction

This course is mainly about the representation theory of certain infinite dimensional Lie
algebras (for example Affine Kac–Moody algebras). So, we start off by defining some of the
more basic objects that will occur in the course.

A Lie algebra is a vector space g with a bilinear form (the bracket)

[·, ·] : g ⊗ g −→ g

which satisfies two additional conditions

• [x, y] = − [y, x]

• [x, [y, z]] + [y, [z, x]] + [z, [x, y]] = 0

The last of these conditions is called the Jacobi identity.

A Lie group is a smooth manifold which also has a group structure. We need the group
structure to be compatible with the smoothness (i.e. multiplication by an element and taking
inverses should be smooth maps).

These appear to be very different objects, however they are closely related. The tangent
space at the identity element of a Lie group naturally has the structure of a Lie algebra.
So, thinking about Lie algebras is a way of linearising problems from Lie groups. What is
remarkable is that it is often possible to completely answer questions about Lie groups from
the related facts about Lie algebras.

The easiest Lie algebras to deal with are the simple finite–dimensional complex Lie algebras.
Let us first explain what these terms mean. Simple means that the algebra has no non–
trivial ideals (in other words the Lie algebra can not be broken down into smaller pieces).
Finite–dimensional means that the vector space is finite dimensional. Complex means
that we are dealing with vector spaces over the complex field. These Lie algebras have been
completely classified and fit into 4 infinite families and 6 exceptional groups. The families
are called An , Bn , Cn and Dn ; the exceptional Lie algebras are called E6 , E7 , E8 , F4 and G2 .

For example, the Lie algebra An is the Lie algebra also called sln+1 C (this is the vector space
of (n + 1) × (n + 1) matrices with zero trace and bracket given by [A, B] = AB − BA). If
you are wondering why this is something to do with the special linear group (which would
appear to want a determinant 1 condition rather than a trace condition) recall that the Lie
algebra should be the elements of the tangent space. To find these we usually write elements
of SLn+1 (C) in the form I + !M and see what restriction this puts on M for “infinitessimal”
!. In this case it is easy to check that this gives a trace 0 condition.

3
1.1 Affine Kac–Moody Algebras

Let’s begin with a motivational discussion for the definition of affine algebras.

If we have a Lie algebra g and a commutative and associative algebra A then g ⊗ A is a Lie
algebra if we use the bracket
[g ⊗ a, h ⊗ b] = [g, h] ⊗ ab
What exactly is A? One can usually think of commutative, associative algebras as functions
on some manifold. This suggests that the simplest case of this construction might be when
we choose our manifold to be the circle. Now

g ⊗ Func(M ) = Func(M, g)

so we are thinking about maps from the circle into the Lie algebra. We, of course, have to
decide what type of maps we want to deal with. Smooth maps would lead us into analysis
whereas we would like to stick to algebra — so we use algebraic (polynomial) maps. This
gives us the loop algebra
Lg = g ⊗ C[t, t−1 ]
with the bracket given as above.

It turns out that this is not quite the correct object to be thinking about. There is much
more structure and theory about centrally extended loop algebras (these facts will become
clearer as the course progresses). The centrally extended Lie algebras ĝ fit into

0 −→ CK −→ ĝ −→ Lg −→ 0

It turns out that these central extensions are classified by the second cohomology group
H 2 (Lg, C), and for simple Lie algebras these are all known to be 1 dimensional. In fact, it is
easy to write down the general construction for such an extension.

Let (·, ·) be an inner product on g which is invariant1 :

([x, y] , z) + (y, [x, z]) = 0

Then the central extension bracket is given by

[A ⊗ f, B ⊗ g] = [A, B] ⊗ f g − (A, B) Res(f g " )K

Here the residue is the coefficient of t−1 in the polynomial.

We will sometimes place the inner product as a subscript to ĝ to make explicit what central
extension we are using. For example, if the inner product is denoted by κ then we will denote
the central extension by ĝκ .
1
This is the infinitessimal verson of (gx, gy) = (x, y)

4
There is a formal version of the above construction where we replace the Laurent polynomials
by Laurent series. In other words we take the central extension

0 −→ CK −→ ĝf −→ g ⊗ C((t)) −→ 0

As before, these central extensions are parametrised by a second cohomology group and this
is a one dimensional space. The central extensions are again given by specifying a bilinear
form (the bracket is given by the same formula as before — note that the residue is still well
defined). If we want to make the bilinear form explicit we may also write ĝκ,f .

1.2 Representations

We want to study the category of representations of both the formal and non-formal affine
Kac–Moody algebras. There are, however, far too many representations of these algebras
without placing extra restrictions on what kind of representation we deal with. This extra
restriction is the choice of dealing only with smooth representations.

A representation of a Lie algebra g is a pair (V, ρ) of a vector space V and a map ρ : g →

End(V ) such that
ρ([x, y]) = ρ(x)ρ(y) − ρ(y)ρ(x)
We will often abuse notation and call the representation V rather than specifying the homo-
morphism ρ, however ρ will always be implicit.

A smooth representation (V, ρ) of ĝ has the following additional property: For any vector
v ∈ V there is a positive integer N such that2

g ⊗ tN C[[t]] · v = 0

Note that, for N ! 0, g ⊗ tN C[[t]] is a Lie subalgebra of ĝ.

By Schur’s lemma we know that the centre of the Lie algebra should act as multiplication
by a scalar on the representation (well, a choice of scalar for each irreducible component of
the representation — but we will mainly be studying the irreducible representations). We
therefore chose the representation to have the central element K act as the identity.

Note now, that these smooth representations of ĝ extend canonically to smooth representa-
tions of ĝf — applying a Laurent series to a vector v only leads to a finite sum because the
smoothness condition removes all sufficiently large powers of t. So, from now on we will only
consider smooth representations of ĝf .

We denote by Oκ the category of smooth representations of ĝκ,f . The objects in this

category are the representations defined above and the morphisms are the obvious maps
between representations.
2
The smoothness condition can also be thought of as a continuity condition on ρ, however this isn’t so
important

5
1.3 Motivation

The Langlands Program is a program of study of representations of groups such as GLn (Qp )
and GLn (Fq ((t))). There is a corresponding definition of a smooth representation for these
groups (in fact, it is in this context that this definition first appeared).

Theorem 1.1 (Roughly) The (isomorphism classes of ) irreducible smooth represenations of

these groups are in correspondence with the (isomorphism classes of ) n dimensional repre-
sentations of the Galois group Gal(F̄ /F ) (where F is either Qp or Fq ((t)) and ‘bar’ denotes
separable closure).

This was conjectured by Langlands and has recently been proved.

For a general reductive group G the conjecture is more complicated and still open.

Conjecture 1.2 (Roughly) The (isomorphism classes of ) irreducible smooth representations

of G(F ) (F as above) are in correspondence with the (isomorphism classes of ) homomor-
phisms
Gal(F̄ /F ) −→ L G
where L G is the Langlands dual group of G.

The idea of this course is to study what happens if we replace the field Fq ((t)) by C((t)). This
gives a formal loop group and we can study the Lie algebra instead. As mentioned before, it
will be better to study the affine algebra rather than just the loop algebra. So we have made
the transition:
G(Fq ((t))) " G(C((t))) " g((t)) " ĝ
The philosophy indicated above suggests that the represenations should be parametrised by
something to do with the Galois group and the Langlands’ dual group. However, the Galois
group of C((t))/C((t)) is far too simple an object

Gal(C((t))/C((t))) = lim Z/nZ = Ẑ

←

The important idea is that the Galois group should be thought of as some kind of fundamental
group.

Imagine an algebraic curve X over k. It has function field given by k(X). If we take Y to be
any finite cover of X then we have that k(Y ) is an algebraic extension of k(X). Automor-
phisms of Y preserving X are deck transformations and these correspond to automorphisms
of k(Y ) preserving k(X) — in other words to an element of the Galois group. Deck transfor-
mations are, of course, transformations of the cover induced by lifting non–trivial loops on
the base space. So, if we take Y to be the universal cover of X (which will correspond to

6
the algebraic closure of the funciton field) then we see that the fundamental group of X is
strongly related to the Galois group of the algebraic closure.

So we want homomorphisms π1 (X) −→ L G (if we wish to deal with reductive groups other
than the general linear group). These are basically the same as local systems.

A L G local system on X is a L G bundle on X with a given (transitive) local identification

of fibres. Note that all bundles are by definition locally trivial (i.e. have a local identification
of fibres) however the identification is not usually given or canonical; in the case of a local
system it is fixed3 .

It is clear how a local system gives rise to a homomorphism π1 (X) −→ L G (simply follow
the identification around the loop, it will be independent of homotopy class because of the
transitive local identification of fibres).

To get a local system from a given homomorphism is also easy. Let φ : π1 (X) → L G be such
a homomorphism and form the fibre product Pφ = X ×π1 (X) L G. This is a bundle over X
with fibre isomorphic to L G. By definition it has a given local identification of fibres and the
correct monodromy.

1.4 Universal Enveloping Algebras

A standard tool in the study of Lie algebras and their representations is the universal en-
veloping algebra. This is an associative algebra that can be constructed from any Lie algebra
and has the property that any representation of the Lie algebra can be regarded as a repre-
sentation of the corresponding associative algebra.

As we want to be able to multiply elements of g together in the associative algebra it makes

sense to look at the tensor algebra of g

T · (g) = C ⊕ g ⊕ g⊗2 ⊕ g⊗3 ⊕ · · ·

If we have a representation (V, ρ) of g we clearly want it to be unchanged on the g piece of

the tensor algebra. We also want it to extend to a representation of an associative algebra,
so we want
ρ(g ⊗ h) = ρ(g)ρ(h)
However, the Lie algebra representations satisfy one additional relation

ρ([g, h]) = ρ(g)ρ(h) − ρ(h)ρ(g)

This suggests that we should identify [g, h] and g ⊗ h − h ⊗ g in the enveloping algebra and,
indeed, this is what we do.
3
This is the same thing as a flat connection on a L G bundle over X in caseX is compact.

7
The universal enveloping algebra U (g) is the quotient of the tensor algebra by the two
sided ideal generated by elements of the form

g ⊗ h − h ⊗ g − [g, h]

Note that we have a natural map g → U (g).

It is called universal because it has the following universal property: If A is any associative
algebra (regarded naturally as a Lie algebra) and f : g → A is a Lie algebra homomorphism
then the following diagram can be uniquely completed to a commutative diagram
U (g)
" .. ..
..
..!
..
. .
..
!
f # A
g

There is a natural filtration on the universal enveloping algebra coming from the gradation
on the tensor algebra. The filtration is denoted by U (g)!i.

The structure of the universal enveloping algebra is very easy to work out in the case that g
is abelian. In this case the generators for the ideal are simply g ⊗ h − h ⊗ g (as the bracket is
zero). This is exactly the definition of the symmetric algebra generated by g. Hence we get

U (g) ∼
= Sym· (g)

In the case where g is not abelian something similar can be done, but the construction is a
little more complicated.

Given any filtered algebra

0 = F!−1 ⊂ F!0 ⊂ F!1 ⊂ · · · ⊂ F

we define the associated graded algebra to be

!
gr F = F!i /F!i−1
i"0

This is naturally a graded algebra and considered as vector spaces we have gr F ∼

= F . Taking
the associated graded algebra of a quotient is reasonably easy:

gr(F/I) = gr(F )/Symb(I)

where Symb is the symbol map: Let x be an element of a filtered algebra F . There is an
integer i such that x ∈ F!i but x )∈ F!i−1 . The symbol of x is defined to be the image of x
in the quotient F!i /F!i−1 . In other words, the symbol map is picking out the piece of x of
highest degree.

8
So, for the universal enveloping algebra we have

gr U (g) = gr T (g)/Symb(I)

As the tensor algebra is already graded, gr T (g) ∼

= T (g).

Lemma 1.3 The ideal Symb(I) is generated by the symbols of the generators for the ideal I.

The proof of this fact is the heart of most proofs of the Poincaré–Birkhoff–Witt theorem. It
fundamentally relies on the fact that the bracket satisfies the Jacobi identity4 . The symbols
of the generators are easy to work out and are just g ⊗ h − h ⊗ g. Hence we have shown that

gr U (g) ∼
= Sym(g)

for an arbitrary Lie algebra g. In other words, as vector spaces, U (g) and Sym(g) are
isomorphic. This is often quoted in the form

Theorem 1.4 (Poincaré–Birkhoff–Witt) Let g1 , g2, . . . be an ordered basis for g (as a

vector space), then the universal enveloping algebra has a basis given by elements of the form
g1a1 g2a2 · · · , where ai ! 0.

One of the main reason for studying the universal enveloping algebra is that the category of
representations for U (g) is equivalent to the category of representations for g (this is just the
statement that representations of g can be lifted to representations of U (g)). However, there
may be many more central elements in U (g) than in g and as central elements act by scalars
this can place a strong restriction on the possible representation. To give some examples of
this we will look at the centre Z(g) of the universal enveloping algebra of a finite dimensional
simple Lie algebra g.

Theorem 1.5 Z(g) is a polynomial ring C[Pi ] generated by elements Pi . The polynomial
ring has the same rank as the Lie algebra g, and the elements Pi have degrees di + 1, where
di are the exponents of g.

Many terms in the previous theorem have not been defined, however they can be found in
most books on Lie algebras. We are only trying to illustrate the rough structure of Z(g) (i.e.
it is a polynomial ring).

It is always true that the first exponent of a simple Lie algebra g is 1, so there is always a
quadratic element in the centre of U (g). This element is called the Casimir element and
can be constructed as follows. Let {J a } be a basis for g as a vector space, fix any non–zero
4
If we were using an arbitrary bilinear form to define the ideal I (e.g. if we used the ideal generated by
a ⊗ b − b ⊗ a − B(a, b)) there would be elements in Symb(I) not obtainable from the symbols of the generators.

9
inner product on g and let {Ja} be the dual basis to {J a }. The Casimir element is then given
by
dim
"g
C= J a Ja
a=1
It is a good exercise compute this element for the case of g = sl2 and check that it is central.

After we have defined the universal enveloping algebra for the affine algebras we are looking
at we will prove a similar structure theorem for Z(g).

The obvious choice for the universal enveloping algebra of ĝκ,f is simply the one defined
above, namely U (ĝκ,f ). However, this is not the correct choice. We are only dealing with
smooth representation whose centre acts as the identity so we only need these representations
to lift to the enveloping algebra (for the one we just wrote down, all representations will lift
and hence there will not be an equivalence of categories). This means we should expect to
use a bigger enveloping algebra than U (ĝκ,f ). Because we know that g ⊗ tN C[[t]] acts as zero
on any vector v for sufficiently large N we can expect the enveloping algebra to allow infinite
sums of such elements (because the action on a representation will only be a finite sum by
the previous property).

Define a topology on U (ĝκ,f ) by using a basis of neighbourhoods for 0:

{BN = U (ĝκ,f ) · g ⊗ tN C[[t]] : N ! 0}
Define
Ũ (ĝκ,f ) = lim(U (ĝκ,f )/BN )
←
By definition this is an algebra in which we can take the infinite sums that we wanted (in
other words, infinite sums where the power of t is bounded below are in Ũ ).

Note that we could have performed this construction on the polynomial version of the affine
algebra ĝκ,p . It is easy to check that

Lemma 1.6 Ũ (ĝκ,p) and Ũ (ĝκ,f ) are isomorphic as topological algebras.

So, we can forget about whether or not we were using the formal or polynomial version of
the affine algebra.

Finally, we want to impose the condition that K acts as the identity. This is done by
quotienting out by the ideal generated by K − 1. This leads us to define the following two
algebras
Uκ (ĝf ) = U (ĝκ,f )/(K − 1)
Ũκ (ĝ) = Ũ (ĝκ,f )/(K − 1)
There is now an equivalence of categories between the smooth representations of ĝκ,f and the
continuous representations of Ũκ (ĝ)5 .
5
The notion of a continuous representation is the obvious one as the algebra Ũ is a topological algebra

10
As the Casimir operator was particularly easy to construct in the finite dimensional case we
start by attempting to define an equivalent operator in the affine case. Let g be any element
of g and n an integer. Then g ⊗ tn is an element of g ⊗ C((t)) and hence of ĝ. We denote
this element by gn . We collect all of the element associated to an element of g into a single
power series: "
g(z) = gn z −n−1
n∈Z

The choice of exponent for z may seem a little strange at first but it is the correct one. For
example, we get the following formula

gn = Res(g(z)z n)

which we can also write in integral notation6 as

#
gn = g(z)z n dz

An obvious guess now for an equivalent to the Casimir operator is

dim
"g
J a (z)Ja(z)
a=1

There are however many problems with this. If we extract the coefficients of this sum we see
that they are two–way infinite sums and hence do not even belong to the completed universal
enveloping algebra Ũ (ĝ). For example, if we use g = sl2 with standard generators
$ % $ % $ %
0 1 1 0 0 0
e= , h= , f=
0 0 0 −1 1 0

and bilinear form

*a, b+ = tr(ab)
Then the Casimir element is
1
C = ef + f e + h2
2
This would give our potential Casimir element as
1
C(z) = e(z)f (z) + f (z)e(z) + h(z)h(z)
2
The coefficient of z −2 in this sum is
" 1
ei fj + fi ej + hi hj
2
i+j=0

6
This is just formal notation which is meant to be reminiscent of Cauchy’s integral formula. None of the
power series we use actually have to converge in any analytic way.

11
It is easy to see from this sum that the element is not in Ũ . However, we might be lucky
and have that this element makes sense on any smooth representation despite not being in
the universal enveloping algebra. Even this is not true as simply evaluating this element on
a representation will lead to infinite sums of the form
"
v ,→ i·v
i"1

which makes no sense to us.7

If we examine where this infinite sums come from we see that they occur due to the piece of
the Casimir element which is not allowed in the completed universal enveloping algebra. For
example we have things like
" " "
hi hj = h−i hi + h−i hi
i+j=0 i"0 i<0

The first term on the right hand side is well defined on any smooth representation, however
the second terms is not. When computing with the second term we are lead to swap the
order of the element using the commutation relations

h−i hi = hi h−i + [h−i , hi ]

and it is the bracket term which is contributing the infinite sums. So we can get around
this problem if these terms were already written this way around (meaning we wouldn’t need
to use the commutation relations and so wouldn’t get the infinite sums). This may seem
like a completely useless definition trick, however it is the beginnings of an operation that
physicists call normally ordered product. Before we go on to rigorously define what things
like normal ordering are we will introduce the concept of a vertex algebra.
7 1
There are methods of renormalisation in physics which would replace the above sum by − 12 however we
are not going to use this approach.

12
2 Vertex Algebras

Let R be an algebra over C; a formal power series over R in variables z1 , . . .zn is a sum
of the form "
Ai1 ···in z1i1 · · · znin
i1 ,...,in ∈Z

The set of all such formal power series is denoted by R[[z1±1 , . . . , zn±1 ]].

Note carefully the difference between formal power series R[[z ±1 ]] (which can have arbitrarily
large and small powers of z), Taylor power series R[[z]] (which have no negative powers of z)
and Laurent power series R((z)) (which have negative powers of z bounded below).

What operations can be perform on formal power series? We can certainly add them, dif-
ferentiate them, multiply them by polynomials. However, we cn not multiply them by other
formal power series. The reason for this is that the coefficients of the product will consist of
infinite sums, e.g.
& '& '  
" " " "
An z n Bm z m = zn  Ai Bj 
n m n i+j=n

We can multiply two formal power series if the variables they are in are disjoint, so f (z)g(w)
makes sense as a formal power series in the two variables z and w.

A particularly important example of a field is the formal delta function. This is denoted
by δ(z − w)8 and is defined by
"
δ(z − w) = z n w −n−1
n∈Z

It has the following easy to check properties:

1. A(z)δ(z − w) = A(w)δ(z − w)

2. (z − w)δ(z − w) = 0

3. (z − w)n+1 ∂wn δ(z − w) = 0

The first of these properties tells us that

#
A(z)δ(z − w)dz = Res (A(z)δ(z − w)) = A(w)

8
It is not a function of z − w, this is just notation which will make a couple of formulæ involving the formal
delta function correspond closely with properties of the usual delta function

13
which is a property we would expect the delta function to have.

We can make the analogy between C[[z ±1 ]] and distributions precise in the following way.
Given a formal power series A in z, define a distribution φA on the polynomials C[z, z −1 ] by
#
φA (f (z)) = A(z)f (z)dz = Res (A(z)f (z)dz)

Recall that the product A(z)f (z) is well defined as f (z) is a polynomial and so only finite
sums turn up in the product. Given a distribution φ on C[z, z −1 ], define a formal power series
Aφ by "
Aφ (z) = φ(z n )z −n−1
n

It is easy to see that these two operations are inverse to each other. Hence C[[z ±1 ]] is exactly
the distributions on the polynomials C[z, z −1 ].

We can now think of δ(z − w) as being a formal power series in the variable z with w ∈ C×
any non–zero complex number. Under the above identification it is easy to see that we get a
delta function in the usual sense.

A field is a special type of formal power series. Let V be a vector space over C, so End(V )
is an algebra over C. A field is a formal power series in End(V )[[z ±1 ]]. We write the field as
follows "
A(z) = An z −n
n∈Z

The power of z chosen is one that will later make much of the notation simpler. Fields have
the following additional property: For each v ∈ V the is an integer N ! 0 such that An ·v = 0
for all n ! N .9
,
If the vector space V is graded (i.e. V = n Vn ) then we have the usual concept of homoge-
neous elements of v as well as homogeneous endomorphisms: φ ∈ End(V ) is homogeneous
of degree m if φ(Vn ) ⊂ Vn+m for all n. In the case of vertex algebras it is common to call
the homogeneous degree the conformal dimension.

A vertex algebra consists of the following objects

,
1. A Z+ –graded vector space V = n"0 Vn where dim(Vn) < ∞ (the state space).

2. A choice of vector in V denoted by |0+ (the vacuum vector).

3. An endomorphism T : V → V of degree 1 (the time translation)

4. A linear map Y (·, z) : V → End(V )[[z ±1 ]] sending homogeneous elements of degree m

to fields of degree m (the state–field correspondence).
9
We can rephrase the final condition for a field as saying that A(z) · v is a Laurent polynomial.

14
These satisfy the following axioms

1. Y (|0+ , z) = idV

2. Y (a, z) |0+ = a + O(z)

3. [T, Y (a, z)] = ∂z Y (a, z)

4. For any two fields Y (a, z) and Y (b, w) there is an integer N such that

(z − w)N [Y (a, z), Y (b, w)] = 0

It follows from the axioms for a vertex algebra that T has the following representation

T (v) = v(−2) |0+

A particularly simple example of a vertex algebra can be constructed from a commutative

associative unital algebra with a derivation of degree 1 (note, however, that this example is
too simple to illustrate most of the important features of a vertex algebra).

Let V be a Z+ –graded commutative associative unital algebra with a derivation T of degree

1. We define the vertex algebra structure as follows

|0+ = 1V
" zn
Y (A, z) = mult(T n A) = mult(eT z A)
n!
n"0
T = T

Here the operators mult(A) in the power series are left multiplication by A.

It is an easy exercise to check that this is a vertex algebra structure. The vertex algebra
structure is particularly simple because the axiom of locality has become a form of commu-
tativity
[Y (A, z), Y (B, w)] = 0
Any vertex algebra with this property is called commutative. Another property that this
vertex algebra structure has that is unusual is that the formal power series that occur have
only non–negative powers of z. This property is known as being holomorphic. It turns out
that these two properties are equivalent.

Lemma 2.1 A vertex algebra is commutative if and only if it is holomorphic

Proof. If V is commutative then

Y (A, z)Y (B, w) |0+ = Y (B, w)Y (A, z) |0+

15
Expanding these in powers of w and taking the constant coefficient we see that Y (A, z)B ∈
V [[z]] for any A and B. This shows that V is holomorphic.

If V is holomorphic then Y (A, z)Y (B, w) ∈ End(V )[[z, w]]. Locality then says that

(z − w)N Y (A, z)Y (B, w) = (z − w)N Y (B, w)Y (A, z)

As (z − w)N has no divisors of zero in End(V )[[z, w]] it follows that V is commutative. !

So we have seen that any Z+ –graded commutative associative unital algebra with degree
1 derviation gives rise to a commutative vertex algebra. It is easy to see that the above
construction can be run in the other direction and so these two categories are equivalent.

2.1 More on Locality

We have just seen that the idea of commutativity in a vertex algebra is very restrictive. It is
therefore not a surprise that we have a more general axiom such as locality. In this section
we will look closely at the locality axiom and try to give some insight into its meaning.

Let v ∈ V and φ : V → C. Given A, B ∈ V we can form the two functions

*φ, Y (A, z)Y (B, w)v+ and *φ, Y (B, w)Y (A, z)v+

These two function belong to C((z))((w)) and C((w))((z)) (respectively). These spaces are
different: The first consists of bounded below powers of w but arbitrary powers of z; the
second consists of bounded below powers of z but arbitrary powers of w. These two spaces
can be represented by the following diagram

Y(A,z)Y(B,w)v w

Y(B,w)Y(A,z)v

16
We see that the intersection of the two spaces consists of function in z and w which have
bounded below powers in both z and w. So
C((z))((w)) ∩ C((w))((z)) = C[[z, w]][z −1, w −1 ]
So, within each field C((z))((w)) and C((w))((z)) we have a sub–ring C[[z, w]][z −1, w −1 ],
therefore within each of the fields we have the fraction field of C[[z, w]][z −1, w −1 ] (note now
that the embeddings into each field will now be different). The fraction field is denoted by
C((z, w)) and consists of ratios f (z, w)/g(z, w) where f, g are in C[[z, w]].

The embeddings into C((z))((w)) and C((w))((z)) are easy to describe: we simple take Lau-
rent expansions assuming one of the variables is “small”. We will illustrate this using the
element (z−w)
1
which is in C((z, w)).

Assume that w is the small variable. We can then expand

1 1 " - w .n
−1
= = z
z −w z(1 − wz ) z
n"0

We can expand in positive powers of w/z because w is the “small” variable. Note that the
result will have bounded below powers of w and so will lie in C((z))((w)).

Assume that z is the small variable. We can then expand

1 1 " - w .n
−1
=− = −z
z−w w(1 − wz ) z
n<0

We can expand in negative powers of w/z because z is the “small” variable. Note that the
result will have bounded below powers of z and so will lie in C((w))((z)).

So we can think of the two rings C((w))((z)) and C((z))((w)) as representing functions in two
variables which have one of their variables much smaller than the other; this is illustrated by
the following diagram

w Y(A,z)Y(B,w)v

Y(B,w)Y(A,z)v

17
So we have now seen that elements in C((w))((z)) and C((z))((w)), although they can look
very different can infact be representing the same function. This is very similar to the idea
of analytic continuation from complex analysis. When these two different elements come
from the same rational function in z and w we could think of them as “representing the
same function” (we could even think of the rational function as being the fundamental object
rather than the individual representations).

Locality is telling us something about the functions

*φ, Y (A, z)Y (B, w)v+ and *φ, Y (B, w)Y (A, z)v+

representing the same rational function in z and w. In fact, due to the appearence of (z − w)
in the locality axiom, it is telling us that these two functions are representation of the same
rational function in C[[z, w]][z −1, w −1 , (z − w)−1 ].10 So, we can rephrase the definition of
locality as11 :

For A, B ∈ V and any v ∈ V , φ : V → C the functions

*φ, Y (A, z)Y (B, w)v+ and *φ, Y (B, w)Y (A, z)v+

are expansions of the same function in C[[z, w]][z −1, w −1 , (z − w)−1 ]. Moreover, as v and φ
vary there is a universal bound on the power of (z − w) that occur (the powers of z and w
can be arbitrarily negative).

2.2 A vertex algebra associated to an AKMA

We now construct an exaple of a non–commutative vertex algebra. This will be based on

the elements of an affine Kac–Moody algebra. All interesting properties a vertex algebra can
have will be visible in this example.

Let g be a finite dimensional complex simple Lie algebra with ordered basis {J a } where
a = 1, . . . , l and l is the rank of the algebra g. Recall that the affine Kac–Moody algebra ĝ
has a basis consisting of elements {Jna } ∪ {K} where a varies as before and n ∈ Z.

Previously, we grouped the elements associated to J a into a formal power series

"
J a (z) = Jna z −n−1
n∈Z

this gives us a hint about what some of the fields in the vertex algebra should be.

Before thinking about fields we should describe the vector space V on which the vertex
algebra lives. We know that we need a special vector |0+ to be the vacuum vector. We also
10
It may seem slightly strange that we only allow these three types of singularities, however these are the
only coordinate free singularities and we are currently working towards a coordinate free description of vertex
operators and algebras.
11
It is not difficult to see that this is equivalent to the old definition of locality.

18
know that we want the elements Jna for n ! 0 to annihilate the vacuum. Lastly we want K
to act as the identity operator.

Notice that the set of Lie algebra elements for which we have defined the action on |0+
form a subalgebra b (in fact it is a particularly special type of subalgebra called a Borel
subalgebra). b has an obvious one–dimensional representation on the space spanned by
|0+ given by the relations mentioned in the previous paragraph. We can therefore get a
representation of g by induction:
V (ĝ) = Indgb(W )
where W is the one–dimensional representation.

The structure of this induced representation is easy to describe. It is spanned by elements of

the form
Jna11 · · · Jnaii |0+
where the nk are strictly negative. If we multiply on the left by an element of the form Jna
with n < 0 we get exactly what we expect. If we multiply on the left by Jna with n ! 0 then
we use the commutation rules in the Lie algebra to move this term through to the vacuum
vector. Jna with n ! 0 is defined to annihilate the vacuum.

The Z+ –grading for the vertex algebra is given by assigning

Jna11 · · · Jnaii |0+

with nk < 0, weight

−n1 − n2 − · · · − ni
Note that the unique elements of weight 0 are multiples of the vacuum vector.

We can illustrate the structure of V in the case of the affine Lie algebra obtained from sl2 (C).
The elements J a are now denoted by e, f, h and the commutation relations are

[h, e] = 2e, [h, f ] = −2f, [e, f ] = h

An element of the representation is

e−1 f−2 |0+
If we multiply on the left by h1 we get

h1 e−1 f−2 |0+ = ([h1 , e−1 ] + e−1 h1 ) f−2 |0+

= (2e0 + e−1 h1 ) f−2 |0+
= 2 ([e0 , f−2 ] + f−2 e0 ) |0+ + e−1 ([h1 , f−2 ] + f−2 h1 ) |0+
= 2h−2 |0+ + 0 − 2e−1 f−1 |0+ + 0
= 2 (h−2 − e−1 f−1 ) |0+

So, at each step we are simply moving the annihilation operators closer and closer to the
vacuum until we have made them all disappear.

19
We now have the vector space V and vacuum vector |0+ for our vertex algebra. We still need
to define the translation operator T and the state–field correspondence Y (·, z).

Looking at the axioms for the translation operator we see from [T, A(z)] = ∂z A(z) that
/ 0
T, A(n) = −nA(n−1)

In our case this means that we need T to satisfy

T |0+ = 0, [T, Jna] = −nJn−1

a

These conditions uniquely specify the action of T on the vector space V .

Finally, we need to define the state–field correspondence. For the vacuum this is determined
by the axioms:
Y (|0+ , z) = id
The elements of the next weight are of the form J−1 a |0+. To guess the form of the vertex

operators we either recall that the fields J (z) naturally turned up before and so would
a

probably be a nice choice; or, we look at the commutative case to see that the fields associated
to the equivalent vectors were basically the same as J a (z). So, we guess12
"
Y (J−1
a
|0+ , z) = Jna z −n−1 = J a (z)
n∈Z

It is easy to see that these satisfy the axioms

Y (A, z) |0+ = A + O(z), [T, Y (A, z)] = ∂z Y (A, z)

We should also check that these operators satisfy the locality axiom. To do this we use the
commutation relations in ĝ to evaluate the commutator of J a (z) and J b (w):
1 2 1 2
J a (z), J b (w) = J a , J b (w)δ(z − w) − κ(J a , J b)∂w δ(z − w)

Now, recalling that (z − w)2 annihilates both δ(z − w) and its derivative, we see that the
locality axiom is satisfied. We now need to define the vertex operators corresponding to the
more complicated elements of V .

Before doing this we’ll take a slight detour. It follows from the axioms for a vertex algebra
that13
Y (A, z) |0+ = ezT (A)
12
Notice the abuse of notation going on here. On the left hand side, Jna refers to an element of the universal
enveloping algebra. On the right hand side, Jna refers to the operator of left multiplication in the representation
V.
13
We have already seen this for commutative vertex algebras

20
Thus

Y (T A, z) |0+ = ezT (T A)
" zn
= T n (T A)
n!
n"0
" zn
= T n+1 (A)
n!
n"0

= ∂z ezT (A)
= ∂z Y (A, z) |0+

There is a general theorem (Goddard’s uniqueness theorem) which we will see later
which tells us (roughly) that if a field acts on the vacuum identically to some Y (A, z) then
it is equal to Y (A, z). So the above calculation shows us that Y (T A, z) = ∂z Y (A, z).

This allows us to guess the correct formula for Y (J−k

a |0+ , z):

1
Y (J−k
a
|0+ , z) = ∂ k−1 J a (z)
(k − 1)! z

To simplify notation we introduce the following short hand:

1 n
T (n) is defined to be T
n!
So the previous formula becomes

Y (J−k
a
|0+ , z) = ∂z(k−1)J a (z)

For a vertex operator of the form Y (AB, z) we know that it must satisfy

Y (AB, z) |0+ = ezT (AB)

= ezT (A) · ezT (B) because T is a derivation

This immediately suggests something like Y (AB, z) = Y (A, z)Y (B, z) however we know from
before that this product is in general not defined. We mentioned before that the problem
with this product was that the annihilating operators didn’t appear to the right and if we
stipulated this as a rule we got something well defined. To make this precise we define, for
a formal power series f (z), f (z)+ to be the piece with positive powers of z and f (z)− to be
the piece with strictly negative powers of z. The normally ordered product of two fields
A(z) and B(z) is then defined to be

: A(z)B(z) : = A(z)+ B(z) + B(z)A(z)−

It is an easy exercise (left for the reader) to check that : A(z)B(z) : is again a field (although
the coefficients of z may be infinite sums, when applied to a vector they become finite sums).

21
When we have more than two fields, the normally ordered product is defined from right–to–
left:
: A(z)B(z)C(z) : =: A(z) (: B(z)C(z) :) :

Another way to write the definition for normally ordered product is to use residues. For a
function in two variables F (z, w) define F|z|'|w| to be the expansion assuming z is the “large”
variable (and similarly for F|w|'|z| ). For example
3 4 ∞ - .
" 3 4 ∞ - .
"
1 −1 w n 1 −1 z n
=z , = −w
z−w |z|'|w| z z−w |w|'|z| w
n=0 n=0

Then we can represent the normally ordered product as

&3 4 3 4 '
1 1
: A(z)B(z) := Res A(w)B(z) − B(z)A(w)
w→0 w − z |w|'|z| w − z |z|'|w|

This identity is easy to see from the following formulæ

& 3 4 '
1
Res A(z) = A(w)+
z→0 z − w |z|'|w|
& 3 4 '
1
Res A(z) = −A(w)−
z→0 z − w |w|'|z|

and these can be proved by simple calculation.

So now we try the definition Y (AB, z) = : Y (A, z)Y (B, z) : for the field associated to AB.
Putting these two choices together gives the guess
a1 ai
Y (J−n 1
· · · J−n i
|0+ , z) = : ∂z(n1 −1) J a1 (z) · · ·∂z(ni −1) J ai (z) :

Theorem 2.2 The above definitions satisfy the axioms for a vertex algebra structure on V .

Proof. We have Y (|0+ , z) = id by definition. We need to check Y (A, z) |0+ = A + O(z)

to see that the vacuum axiom holds. This is clearly true when A = |0+, so we proceed by
induction assuming it to hold for Y (B |0+ , z) = B(z)

Y (A−n B |0+ , z) |0+ = : ∂z(n−1)A(z)B(z) : |0+

- . - .
= ∂z(n−1)A(z) B(z) |0+ + B(z) ∂z(n−1) A(z) |0+
−
- .+
= ∂z(n−1)
A(z) B(z) |0+ + 0
+
= (A−n + O(z)) (B |0+ + O(z))
= A−n B |0+ + O(z)

22
So, the vacuum axioms hold by induction.

The translation covariance axioms are also easy

[T, : A(z)B(z) :] = [T, A(z)+B(z)] + [T, B(z)A(z)−]

= [T, A(z)+] B(z) + A(z)+ [T, B(z)] + [T, B(z)] A(z)− + B(z) [T, A(z)−]
= (∂z A(z)+ ) B(z) + A(z)+ ∂z B(z) + (∂z B(z)) A(z)− + B(z) (∂z A(z)− )
= ∂z : A(z)B(z) :

So, the axiom holds by induction.

We will prove locality from a more general lemma which will follow the end of this theorem.
This lemma will show us that normally ordered products of local fields will remain local. It
is clear that derivatives of local fields are still local. So, all we need to do is check that the
generating fields J a (z) are local. From the commutation relations we get
1 2 1 2
J a (z), J b (w) = J a , J b (w)δ(z − w) − κ(J a , J b)∂w δ(z − w)

From this it is clear that 1 2

(z − w)2 J a (z), J b(z) = 0

and so we have the required locality.

Hence all the conditions for a vertex algebra hold. !

Lemma 2.3 (Dong) If A(z), B(z), C(z) are mutually local fields then : A(z)B(z) : and
C(z) are mutually local too.

Proof. The result will follow (by taking Resx→0 ) if we can show that the following two
expressions are equal after multiplying by a suitable power of (y − z)
3 4 3 4
1 1
F = A(x)B(y)C(z) − B(y)A(x)C(z)
x − y |x|>|y| x − y |y|>|x|
3 4 3 4
1 1
G = C(z)A(x)B(y) − C(z)B(y)A(x)
x − y |x|>|y| x − y |y|>|x|

As A, B and C are mutually local we know there is an integer N such that

(x − y)N A(x)B(y) = (x − y)N B(y)A(x)

(y − z)N B(y)C(z) = (y − z)N C(z)B(y)
(x − z)N A(x)C(z) = (x − z)N C(z)A(x)

We will now show that

(y − z)4N F = (y − z)4N G

23
The binomial identity gives
3N 3
" 4
3N
(y − z) 4N
= (y − x)3N −n (x − z)n (y − z)N
n
n=0

So, if 0 # n # N the power of (y − x) is large enough that we can swap A(x) and B(y); the
two terms in F (and G) then cancel, so these do not contribute. For n > N the powers of
(x − z) and (y − z) are large enough that we can swap A(x), C(z) as well as B(y), C(z). This
allows us to make terms in F the same as those in G. Hence

(y − z)4N F = (y − z)4N G

and we are done. !

The above theorem for constructing a vertex algebra works in more generality than simply the
case of an affine Kac–Moody algebra. In this form it is called the reconstruction theorem.

Theorem 2.4 (Easy reconstruction) Let V be a vector space, |0+ a vector of V and T an
endomorphism of V of degree 1. Let {aα (z)}α∈I be a collection of fields such that

1. [T, aα(z)] = ∂z aα (z)

2. T |0+ = 0, aα(z) |0+ = aα + O(z)

3. aα(z) and aβ (z) are mutually local

4. The vectors aα(−j

1
1 −1)
· · · aα(−j
n
n −1)
|0+ with js ! 0 form a basis for V

Then the formula

Y (aα(−j
1
1 −1)
· · · aα(−j
n
n −1)
|0+ , z) =: ∂ (j1 ) aα1 (z) · · · ∂ (jn ) aαn (z) :

defines a vertex algebra structure on V such that |0+ is the vacuum vector, T is the infinites-
imal translation operator and
Y (aα, z) = aα (z)

To prove this we simply repeat the proof used for the AKMA case.

The reconstruction theorem is actually true in a much less restrictive case (we don’t have to
assume that the vectors in 4 form a basis, mearly that they span). And even in this case it
is possible to deduce that the resulting vertex algebra structure is unique.

Theorem 2.5 (Hard reconstruction) Let V be a vector space, |0+ a vector of V and T
an endomorphism of V of degree 1. Let {aα (z)}α∈I be a collection of fields such that

24
 1. [T, aα(z)] = ∂z aα (z)

2. T |0+ = 0, aα(z) |0+ = aα + O(z), where aα are linearly independent

3. aα(z) and aβ (z) are mutually local

4. The vectors aα(−j

1
1 −1)
· · · aα(−j
n
n −1)
|0+ with js ! 0 span V

Then the formula

Y (aα(−j
1
1 −1)
· · · aα(−j
n
n −1)
|0+ , z) =: ∂ (j1 ) aα1 (z) · · · ∂ (jn ) aαn (z) :

gives the unique vertex algebra structure on V such that |0+ is the vacuum vector, T is the
infinitesimal translation operator and

Y (aα, z) = aα (z)

2.3 Associativity

We have already seen that locality axiom for a vertex algebra is telling us that the two formal
power series
Y (A, z)Y (B, w)C and Y (B, w)Y (A, z)C
are expansions of the same function from V [[z, w]][z −1, w −1 , (z − w)−1 ] in two different
domains. One of these domains, V ((z))((w)) corresponds to w being “small”; the other,
V ((w))((z)) corresponds to z being small. If we think of the points z and w as being complex
numbers (in other words elements of the Riemann sphere) we can think of these two domains
as being w very close to 0 and w very close to ∞. There is now an obvious third choice, w
is very close to z, which we have not discussed.

The space corresponding to “w is very close to z” is V ((w))((z − w)) (or alternatively,

V ((z))((z − w)), these are actually identical). The expression, in terms of vertex algebras,
that we expect to live in this ring is

Y (Y (A, z − w)B, w)C

We should clearly expect from the previous results that this be an expansion of the same
function in V [[z, w]][z −1, w −1 , (z − w)−1 ] but assuming z − w to be “small”. This, as we shall
show, is true. We start with a couple of basic but useful results about vertex algebras.

Lemma 2.6 Suppose that U is a vector space, f (z) ∈ U [[z]] is a power series and R ∈
End(U ) is an endomorphism. If f (z) satisfies

∂z f (z) = Rf (z)

then it is uniquely determined by the value of f (0).

25
Proof. A simple induction shows that f (z) must by of the form

R2 R3
f (z) = K + R(K)z + (K)z 2 + (K)z 3 + · · ·
2! 3!
Which completes the proof. !

This gives us another way to prove the relationship

Y (A, z) |0+ = ezT A

mentioned previously. We can now simply show they satisfy the same differential equation
with identical initial conditions.

Lemma 2.7 In any vertex algebra we have

ewT Y (A, z)e−wT = Y (A, z + w)

where negative powers of (z + w) are expanded as power series assuming that w is “small”
(i.e. in powers of w/z).

Proof. In a Lie algebra the following is a formal identity

" wn
ewT Ge−wT = (ad T )n G
n!
n"0

So, in our case

" wn
ewT Y (A, z)e−wT = (ad T )n Y (A, z)
n!
n"0
" wn
= ∂zn Y (A, z)
n!
n"0

= ew∂z Y (A, z)

The fact that ew∂z Y (A, z) = Y (A, z + w) is the formal version of the Taylor expansion. !

This lemma tells us that exponentiating the infinitesimal translation operator really does give
us a translation operator z ,→ z + w.

Theorem 2.8 (Skew Symmetry) In any vertex algebra we have

Y (A, z)B = ezT Y (B, −z)A

as an equality of formal power series.

26
Proof. By locality we know that there is a large integer N such that
(z − w)N Y (A, z)Y (B, w) |0+ = (z − w)N Y (B, w)Y (A, z) |0+
This is actually an equality in V [[z, w]] (note that there are no negative powers of w on the
left and no negative powers of z on the right). Now, by the above results we compute
(z − w)N Y (A, z)Y (B, w) |0+ = (z − w)N Y (B, w)Y (A, z) |0+
⇒ (z − w)N Y (A, z)ewT B = (z − w)N Y (B, w)ezT A
⇒ (z − w)N Y (A, z)ewT B = (z − w)N ezT Y (B, w − z)A
⇒ (z)N Y (A, z)B = (z)N ezT Y (B, −z)A
⇒ Y (A, z)B = ezT Y (B, −z)A
In the fourth line we have taken w → 0, this is allowed as there are no negative powers of w
in the above expressions. !

In terms of the Fourier coefficients we can write the skew symmetry property as
- .
A(n) B = (−1)n+1 B(n) A − T (B(n−1) A) + T (2)(B(n−2) A) − T (3)(B(n−3) A) + · · ·

We now have enough to prove the associativity property

Theorem 2.9 In any vertex algebra the expressions

Y (Y (A, z − w)B, w)C and Y (A, z)Y (B, w)C and Y (B, w)Y (A, z)C
are all expansions of the same function from V [[z, w]][z −1, w −1 , (z − w)−1 ].

Proof. We already know this for the latter two expressions, so we only need to show it for
the first two. We compute
Y (A, z)Y (B, w)C = Y (A, z)ewT Y (C, −w)B
= ewT Y (A, z − w)Y (C, −w)B
In this final expression we expand negative powers of (z − w) assuming that w is “small”.
This means that the expression lives in V ((z))((w)) (note that it is okay to multiply on the
left by the power series ewT as it has no negative powers of w). As mentioned before we can
also think of this ring as V ((z − w))((w)), so we can think of this calculation as holding in
V ((z − w))((w)).

We can also compute

& '
"
Y (Y (A, z − w)B, w)C = Y A(n) B(z − w)−n−1 , w C
n
"
= Y (A(n)B, w)C(z − w)−n−1
n
"
= ewT Y (C, −w)A(n)B(z − w)−n−1
n
= ewT Y (C, −w)Y (A, z − w)B

27
This calculation holds in V ((w))((z − w)).

By locality we know that Y (C, −w)Y (A, z − w)B and Y (A, z − w)Y (C, −w)B are expansion
of the same function under two different assumptions. So, we have shown that Y (Y (A, z −
w)B, w)C and Y (A, z)Y (B, w)C are expansions of the same function. !

It is helpful (although wrong) to think of the associativity property as saying

"
Y (A, z)Y (B, w) = Y (Y (A, z − w)B, w) = Y (A(n) B, w)(z − w)−n−1
n

so that it gives a way to represent the product of two vertex operators as a sum of vertex
operators. Note that none of the above expressions really converge. They do converge when
we apply the terms to a vector C ∈ V , but even then the expressions are not equal as they are
expansions of a common function in two domains. Written in the above form, the equality
is known as the operator product expansion (or OPE) and originally turned up in the
Physics literature.

We now look at consequences of the associativity law. In particular, we will see that our
previous definitions of normally ordered product and the formula for the vertex operators in
the AKMA case are basically unique.

Lemma 2.10 Suppose that φ(z) and ψ(w) are two fields. Then the following are equivalent

N
" −1
1. [φ(z), ψ(w)] = γi (w)∂w(i)δ(z − w)
i=0

N
" −1 3 4
1
2. φ(z)ψ(w) = γi (w) + : φ(z)ψ(w) :
(z − w)i+1 |z|'|w|
i=0
N
" −1 3 4
1
ψ(w)φ(z) = γi (w) + : φ(z)ψ(w) :
(z − w)i+1 |w|'|z|
i=0

where the γi (w) are fields, N is a positive integer and : φ(z)ψ(w) : is the obvious generalisation
of normally ordered product to two variables.

Proof. Assuming 2 we see the commutator [φ(z), ψ(w)] is a difference of expansions of

(z − w)−i−1 in the domains |z| 1 |w| and |w| 1 |z|. These differences are easily seen to be
derivatives of the formal delta function. So 2 ⇒ 1. Finally, for 1 ⇒ 2 we compute
φ(z)ψ(w) = (φ(z)+ + φ(z)− ) ψ(w)
= (φ(z)+ ψ(w) + ψ(w)φ(z)−) + (φ(z)− ψ(w) − ψ(w)φ(z)−)
= : φ(z)ψ(w) : + [φ(z)− , ψ(w)]
= : φ(z)ψ(w) : + [φ(z), ψ(w)]−

28
Expanding the final term we see that it takes the negative powers of z from 1/(z − w)i+1 .
This is easily seen to be equivalent to expanding in the domain |z| 1 |w|. The same proof
works for the product ψ(w)φ(z). !

It is easy to see that the commutator [Y (A, z), Y (B, w)] satisfies the first condition above —
the only non–obvious part is that the coefficients of the sum should be fields in w only, but
this follows from expansion of Y (Y (A, z − w)B, w). Thus, we can apply the above lemma to
the case of vertex operators and compare coefficients to see

γi (w) = Y (A(i)B, w) for i ! 0

Comparing coefficients and using Taylor’s formula to expand : Y (A, z)Y (B, w) : as a power
series in w and z − w we see

: ∂w(i−1)Y (A, w)Y (B, w) := Y (A(−i) B, w)

Thus the acioms of a vertex algebra uniquely specify the normally ordered product and the
vertex operators associated to elements of the form A(−i) B. In particular, if we pick B to be
the vacuum
Y (T A, w) = ∂w Y (A, w) = [T, Y (A, w)]
The second equality here is an axiom, however the first requires all of the structure of a vertex
algebra to deduce.

If we now use the formulæ for γi (w) in the expression above for the commutator we can
extract Fourier coefficients to give a formula for the commutators of the A(n) and B(m)
3 4
/ 0 " n 5 6
A(n) , B(m) = A(k) B (n+m−k)
k
k"0

Several important remarks should be made about this formula: It shows that the collection
of all Fourier coefficients of all vertex operators form a Lie algebra; and the commutators of
Fourier coefficients depend only on the singular terms in the OPE.

If we want to keep one of the elements in the bracket as a field then we get
3 4
/ 0 " n 5 6
A(n) , B(w) = A(k)B (w)w n−k
k
k"0

We will usually simplify the notation in the above lemma, writing

N
" −1 3 4
1
φ(z)ψ(w) = γi (w) + : φ(z)ψ(w) :
(z − w)i+1 |z|'|w|
i=0
as
N
" −1
γi (w)
φ(z)ψ(w) ∼
(z − w)i+1
i=0

29
In the physics literature this is what is usually written down for an OPE. What we have
basically done is to remove all the terms which are non-singular at z = w. We can do this
because the structure of all the commutation relations depends only on the singular terms.

30
3 The AKMA vertex algebra

We had already found that the commutation relations for the fields J a (z) were
1 2 1 2
J a (z), J b (w) = J a , J b (w)δ(z − w) − κ(J a , J b)∂w δ(z − w)

In the above notation this becomes the OPE

/ a b0
J , J (w) κ(J a , J b )
J a (z)J b (w) ∼ −
(z − w) (z − w)2
Our possible generalisation of the Casimir element was (now that we know normal ordering)
1" a
S(z) = : J (z)Ja (z) :
2 a
where Ja is a dual basis to J a under some fixed inner-product κ0 . We know from our previous
discussions about normally ordered product that this is a well defined field, so we would now
like to work out the commutation relations of the S(n) and the J(n) a .

By the definition of the vertex algebra structure, we have the following alternative expression
for S(z) & '
1" 1 "
S(z) = Y (J(−1)Ja,(−1) |0+ , z) = Y
a a
J(−1) Ja,(−1) |0+ , z
2 a 2 a
We define
1" a
S= J J |0+
2 a (−1) a,(−1)
so that S is the element of V generating the field S(z), in other words
Y (S, z) = S(z)
We need to compute
" Y (J(n)
b S, w)
Y (J b , z)Y (S, w) ∼
(z − w)n+1
n"0
Which means we need to compute the element of V we obtain from
1 b " a
J J J |0+
2 (n) a (−1) a,(−1)

By the grading we know that we must have n # 2 (otherwise the resulting element is of
negative grade, and there are no such elements).

n=2 In this case, repeatedly using the commutation relations quickly shows that the result
is
1" - b a .
κ J , [J , Ja] |0+
2 a
If we now choose J a to be orthonormal, so that J a = Ja we see that this is zero.

31
n=1 We get
1" 1" b a
κ(J a , Ja)J(−1)
a
|0+ + [[J , J ], Ja](−1) |0+
2 a 2 a

The first of these terms is the decomposition of J b using the basis J a , however we have
to be careful that κ is not necessarily the same inner product as κ0 . However, they are
multiples of each other. So, the first term is just κκ0 J(−1)
b
|0+.
The second term is the action of the Casimir element on J b in the Lie algebra g. This
is a central element and so it acts as a scalar, λ, say. Recall that any representation V
of g gives a natural inner product by

κV (x, y) = tr(ρV (x)ρV (y))

In particular κg is the inner product from the adjoint representation (also known as the
Killing form). If we pick κ0 = κg it is easy to see that the saclar λ should be 1/2. So,
κ
the second term is 12 κg0 J(−1)
b |0+.

n=0 We get
1 "- a .
J(−1) [J b , Ja ](−1) + [J b , J a](−1)Ja,(−1)
2 a
This corresponds to taking the commutator with the Casimir element and hence is zero
(as the Casimir is central).

Hence the OPE is & '

κ + 12 κg J b (w)
J b (z)S(w) ∼
κ0 (z − w)2
κ
The ratio 12 κg0 is related to the dual Coxeter number — if κ0 is chosen so that the maximal
root has squared length 2 then this ratio is h∨ , the dual Coxeter number.

So, there is a critical value for the inner product κ which causes these two fields to commute.
The critical value is κc = − 12 κg. In this case, the operators from Fourier coefficients of S(w)
will be central elements in the vertex algebra.

Even in the case when we are not at the critical value we see that the commutation relations
are very nice. From now on we normalise S(w) so that the factor at the front is 1, i.e.

J b (w)
J b (z)S(w) ∼
(z − w)2

Often, this is written with the J and S in the opposite order. This can be achieved by using
locality, swapping z and w and using Taylor’s formula:

J b (w) ∂w J b (w)
S(z)J b (w) ∼ +
(z − w)2 (z − w)

32
From the first form of the OPE we immediately see

[J(n) , S(m)] = nJ(n+m−1)

Define Sn to be S(n+1) (in other words we shift the grading by 1; this should not be a surprise
because S(z) is a homogeneous field of grade 2 rather than grade 1). Then the commutation
relations become
[Sn , J(m)] = −mJ(n+m)
These commutation relations are very suggestive. Recall that J(n) stood for the element
J ⊗ tn in ĝ. Thus, the operations given by ad(Sn ) look very much like

ad(Sn ) = −tn+1 ∂t

One can now ask if the ad(Sn ) actually behave like the algebra of these operators. It will
turn out that this isn’t quite true (we should actually take a central extension).

We now compute the OPE for S(z)S(w) (this calculation will immediately show the fact that
the Sn are elements of a central extension). To perform the calculation we need to compute
the element of V we obtain from
1 κ0 "
a
S(n) J(−1) Ja,(−1) |0+
2 κ − κc a

By the grading we know that we must have n # 3 (otherwise the resulting element is of
negative grade, and there are no such elements).

n=3 In this case, repeatedly using the commutation relations quickly shows that the result
is
1 κ0 "
κ (J a , Ja) |0+
2 κ − κc a

Hence we get 1 κ
2 κ−κc dim(g) |0+.
n=2 We get
1 κ0 " a
[J , Ja ](−1) |0+
2 κ − κc a
which is zero by picking J a = Ja .

n=1 We get
1 κ0 " a
2J(−1)Ja,(−1) |0+
2 κ − κc a
which is just 2S.

n=0 We get
1 κ0 " - a .
J(−1) Ja,(−2) + J(−2)
a
Ja,(−1) |0+
2 κ − κc a
which is just T (S).

33
So we see that the OPE is
κ
κ−κc dim(g)/2 2S(w) ∂w S(w)
S(z)S(w) ∼ + +
(z − w)4 (z − w) 2 (z − w)
We denote the constant occuring in the first terms as c/2 (later we will see that c is called
the central charge). If we use the OPE to compute the commutation relations among the
S(n) we get
n(n − 1)(n − 2)
[S(n) , S(m)] = (n − m)S(n+m−1) + cδn+m=2
12
As before we shift the grading by 1 and use the operators Sn instead. This gives
n3 − n
[Sn , Sm ] = (n − m)Sn+m + cδn,m
12
The final term is the one coming from the central extension (hence part of the reason for
calling c the “central” charge).

3.1 The Virasoro Algebra

The Virasoro algebra is the central extension of the algebra C((t))∂t. The central element
is denoted by C and we have
0 −→ CC −→ V ir −→ C((t))∂t −→ 0
A basis for the elements of C((t)) is given by Ln = −tn+1 ∂t . The bracket in the Virasoro
algebra is given by
n3 − n
[Ln , Lm] = (n − m)Ln+m + δn+m C
12
We can construct a vertex algebra associated to the universal enveloping algebra of the
Virasoro algebra in a similar way the affine Kac-Moody case.

Consider the Lie algebra generated by {L−1 , L0 , L1 , . . .} and C (it is easy to check this is
indeed a Lie algebra). This has a one–dimensional representation where the Li act as zero
and the central element acts as the scalar c ∈ C. This representation is denoted Cc . Inducing
this representation to the Virasoro algebra gives an infinite dimensional representation Virc .
A basis for this representation is given by expressions of the form
Ln1 Ln2 · · · Lnm |0+
where n1 # n2 # · · · # nm < −1 and |0+ is a generating vector for the representation Cc .

The operator L−1 is −∂t and so we choose it to be T . We define

"
Y (L−2 |0+ , z) = Ln z −n−2
n

note that it is −n − 1 rather than −n − 1. It is easy to check that this field is local with repect
to itself and so we can apply the reconstruction theorem to give a vertex algebra structure.

34
Theorem 3.1 Virc can be given a vertex algebra structure. The defining OPE is

c/2 T (w) ∂w T (w)

T (z)T (w) ∼ + +
(z − w) 4 (z − w) 2 (z − w)

The number c is known as the central charge.

A vertex algebra homomorphism is a map φ : V → V " , where V and V " are vertex
algebras, such that

1. φ is a homogeneous linear map sending |0+ to |0+"

2. φ intertwines T and T " , i.e. φ ◦ T = T " ◦ φ

3. φ intertwines Y and Y " , i.e.

φ ◦ Y (A, z)B = Y " (φ(A), z)φ(B)

A vertex algebra V is called conformal of central charge c if there is a non-trivial vertex

algebra homomorphism φ : Virc → V such that if ω = φ(L−2 |0+) then ω(−1) is the gradation
operator and ω(−2) = T .

An obvious example of a conformal vertex algebra is Virc itself.

Lemma 3.2 If V is a conformal vertex algebra then, with the notation as above, the ω(n)
satisfy the relations of the Virasoro algebra.

Proof. One of the properties of φ being a vertex algebra homomorphism is that

φ(Y (L−2 |0+ , z) · L−2 |0+) = Y (ω, z)ω

Now use the OPE formula to evaluate both sides. !

We will see that being a conformal vertex algebra means that there is an action of the Virasoro
operator. This means that our vertex algebras naturally have symmetries generated by the
Virasoro algebra and that these symmetries extend to the level of the representations.

Proposition 3.3 V is a conformal vertex algebra of central charge c if and only if there is
a vector ω ∈ V2 such that "
Y (ω, z) = LVn z −n−2
n

• LV−1 = T

35
 • LV0 is the gradation operator

• LV2 ω = 2c ω

Proof. The conditions tell us that the OPE will be of the form
c/2 Y (L1 ω, w) 2Y (ω, w) ∂w Y (ω, w)
Y (ω, z)Y (ω, w) ∼ + + +
(z − w) 4 (z − w)3 (z − w)2 z −w

So we need to show that the term in (z − w)−3 is zero. Switching z and w should give the
same thing (by locality); what we get is

c/2 Y (L1 ω, z) 2Y (ω, z) ∂z Y (ω, z)

Y (ω, w)Y (ω, z) ∼ − + −
(z − w)4 (z − w)3 (z − w)2 z−w

Now use Taylor’s formula to expand the Y (∗, z) terms and we see that the required term
vanishes. So, ω does generate a Virasoro field. We now need to construct a homomorphism
from Virc to V , this is easy

Ln1 · · · Lnm |0+ ,→ LVn1 · · · LVnm |0+

This works because we only need the commutation relations in any computation containing
these vectors and we already know they satisfy the same commutation relations. !

Recall that we have a vertex algebra Vκ (g) associated to every central extension of a finite
dimensional Lie algebra g (the central extension being indexed by an inner product κ). We
have now seen that this vertex algebra is conformal provided that κ )= κc .14

3.1.1 Detour: Why do central extensions keep appearing?

It seems to be true from what we have done so far that all Lie algebras associated to the disk
(or punctured disk) come with a canonical central extension. It is therefore natural to ask
whether there is something general happening.
7 by the Lie algebra of ∞ × ∞ matrices with finitely many non–zero diagonals. This
Let gl ∞
is bigger than the naı̈ve definition of gl∞ as an inverse limit of the finite gln .

Let g be a lie algebra, then we can embed g ⊗ C[t, t−1 ] in gl 7 ∞ by regarding g as a vector space
with basis v1 , . . . , vl . Then g ⊗ C[t, t−1 ] is isomorphic to V ⊗ C[u, u−1 ] for V an l–dimensional
vector space and this is clearly isomorphic to C[u, u−1 ]. So the action of the Lie algebra is
given by endomorphisms of the space C[u, u−1 ]. These endomorphisms can be thought of as
∞ × ∞ matrices and the usual rules for the action of g shows that we do not get infinitely
many non–zero diagonals.
14
In fact, it is not conformal if κ = κc

36
Now, gl∞ has a canonical central extension. If we write the matrix A as
3 4
A−+ A++
A−− A+−
Then the cocycle for the extensions is given by
γ(A, B) = tr(A++ B−− − B++ A−− )
It is easy to check that this is finite. Since we can regard our Lie algebras as embedded inside
7 ∞ we obtain a central extension for each of them by using the central extension of gl
gl 7 ∞.
If we work out what this extension is in our cases it turns out the be exactly the central
extension we have been using.

3.2 Lie algebras associated to vertex algebras

There is a natural Lie algebra associated to any vertex algebra, it is the subspace of End(V )
spanned by all the Fourier coefficients of all vertex operators of V . This is a Lie algebra due
to the formula
" 3n4
[A(n) , B(m)] = (A(k)B)(m+n−k)
k
k
which clearly shows that our chosen space is closed under the Lie bracket.

Unfortunately, this Lie algebra is sometimes too small for us. For example, because all the
Fourier coefficients are viewed as endomorphisms it is possible that some of them will act as
zero even though they came from a non-zero vertex operator. In order to get around this
we try to define an intrinsic Lie algebra associated to the Fourier coefficients. To do this we
define formal symbols A[n] for each Fourier coefficient and stipulate that they satisfy
" 3n4
[A[n], B[m] ] = (A(k)B)[m+n−k]
k
k

Now we have removed the possibility that things act as zero when they are non-zero them-
selves; unfortunately we no longer know that the structure we have defined is a Lie algebra!
The structure we have defined above is denoted U " (V ) (we will also consider a completion of
this, but this won’t matter for now). To be more precise, U " (V ) is the space
V ⊗ C[t, t−1 ]/im(∂) where ∂ = T ⊗ 1 + 1 ⊗ ∂t
All this quotient construction means is that things in U " (V ) satisfy the obvious relations
(A + B)[n] = A[n] + B[n] and (T A)[n] = −nA[n−1]

Define U " (V )0 to be the subspace of U " (V ) given by sending A[n] to zero if n )= 0. So it is

generated by the elements A[0] only. As above we can regard U " (V )0 as being
V /im(T )

37
Lemma 3.4 The algebra U " (V )0 is a Lie algebra under the defined bracket.

Proof. The bracket is now the much simpler

[A[0], B[0]] = (A(0)B)[0]

We need to show this is antisymmetric and satisfies the Jacobi identity. Recall the identity

Y (A, z)B = ezT Y (B, −z)A

Looking at the z −1 coefficient gives

A(0) B = −B(0) A + T (· · · )

Taking the [0] part of this and recalling that things of the form T (∗) are zero we see that

(A(0)B)[0] = −(B(0) A)[0]

In other words, the bracket is antisymmetric. What we want to show for the Jacobi identity
is
[C[0] , [A[0], B[0]]] = [[C[0], A[0]], B[0]] + [A[0], [C[0], B[0]]]
The right hand side of this gives

((C(0)A)(0)B)[0] + (A(0)(C(0)B))[0]

But, from the commutation relations for the Fourier coefficients we know that

[C(0), A(0)] = (C(0)A)(0)

Substituting this in shows the Jacobi identity. !

Proposition 3.5 The algebra U " (V ) is a Lie algebra.

Proof. We will show this by constructing a bigger vertex algebra W such that U " (V ) is
U " (W )0 (from which the previous lemma gives that it is a Lie algebra).

As C[t, t−1 ] is a commutative, associative, unital algebra with derivation (T = −∂t ) we can
make it into a vertex algebra. We claim that the vertex algebra V ⊗ C[t, t−1 ] is the W we
want. To do this we need to look at the commutation relations we get for the [0] part of W .

[(A ⊗ tn )[0], (B ⊗ tm )[0]] = ((A ⊗ tn )(0)(B ⊗ tm ))[0]

Suppose that f (t) is an element of C[t, t−1 ], then the vertex operator associated to f (t) is

Y (f (t), z) = mult(ezT f (t)) = mult(f (t + z))

38
where necessary expansions are performed in the domain |z| 4 |t|. In our case this means
we need the z −1 coefficient of
Y (A, z) ⊗ (t + z)n
which is
" 3n4
A(k) ⊗ tn−k
k
k
Putting this together gives
" 3n4
[(A ⊗ t )[0], (B ⊗ t )[0]] =
n m
(A(k)B) ⊗ tn+m−k
k
k

which are the commutation relations we want for U " (V ). !

The algebra U (V ) is defined to be the quotient

V ⊗ C((t))/im(∂)

with the same bracket relations as U " (V ). This can be thought of as a completion of the
algebra U " (V ). As such, it carries a Lie algebra structure too.

We would like to think of U (∗) as a functor from the category of vertex algebras to the
category of Lie algebras. To do so we will need to work out what U (∗) does to morphisms.

Given a vertex algebra homomorphism φ : V1 → V2 we define U (φ) to be the map

U (φ) : A[n] → φ(A)[n]

Lemma 3.6 This is a Lie algebra homomorphism, hence we can think of U (∗) as a functor.

Proof. Denote U (φ) by φ̃. We need to check that

φ̃([A[n], B[m] ]) = [φ̃(A[n] ), φ̃(B[m] )]

Expand the left hand side using the definition of the bracket. Applying the definition of φ̃ to
each side shows we need
φ(A(n)B) = φ(A)(n) φ(B)
but this follows from the fact that φ is a vertex algebra homomorphism. !

We now have almost enough technology to answer the original question we had: describe the
centre of Ũκ (ĝ). We have already seen that there are some vertex operators (the Segal–
Sugawara operators S(z)) which are good candidates for central elements. However, just
because they were central as elements of the vertex algebra does not imply that they will
be central as elements of Ũκ (ĝ). The reason for this is that the coefficients of S(z) are
interpreted as endomorphisms of the vertex algebra, they can be thought of as elements in

39
the universal enveloping algebra Ũκ (ĝ) but it is not clear that things which act as the zero
endomorphism will act as zero on the universal enveloping algebra. Therefore we can not tell
if the coefficients of S(z) are truely central or simply just acting as zero when thought of as
endomorphisms.

Restrict attention now to V = Vκ (g). Elements of U (V ) are Fourier coefficients from elements
like
: ∂zn1 J a1 (z) · · ·∂znm J am (z) :
Thus the coefficients can be viewed as infinite sums of coefficients from the basic fields
J a (z). The property of being a normally ordered product means that we keep the terms
corresponding to negative powers of z to the right; notice that this is exactly the definition of
the neighbourhoods of zero in Ũκ (ĝ). Hence we can sensibly think of the Fourier coefficients
of fields in V as elements of the universal enveloping algebra. Therefore we have a map

U (V ) −→ Ũκ (ĝ)

Lemma 3.7 This map is injective.

Proof. None in class.

Proposition 3.8 The map is a Lie algebra homomorphism.

Proof. Define three formal power series

"
Y (A, z) = A(n) z −n−1
n

which is a vertex operator, so the coefficients are endomorphisms of V ,

"
Y [A, z] = A[n] z −n−1
n

where the coefficients are elements of the Lie algebra U (V ),

"
Ỹ [A, z] = Ã[n] z −n−1
n

where the coefficients are the A[n] regarded as elements of the universal enveloping algebra
Ũκ (ĝ). The map above is the one which sends

Y [A, z] −→ Ỹ [A, z]

Consider first the case when A = J(−1)

a |0+. In this case, the OPE for the Y (A, z) encodes
exactly the commutation relations that the elements Jna satisfy in the Lie algebra ĝκ . These

40
relations are conserved in the universal enveloping algebra and are the definition of the
commutation relations in U (V ). Hence the map is a homomorphism on these elements.

In particular, we deduce that the formal power series in Ũκ (ĝ) of the basic fields are local.
Looking back at Dong’s lemma, we see that its proof did not need that the coefficients of
the fields were endomorphisms. Hence, Dong’s lemma holds in our current situation and we
have that all formal power series Ỹ [A, z] are local with respect to each other. For the same
reasons as before the commutation relations satisfied by the Fourier coefficients of the Ỹ [A, z]
are encoded by the singular terms in the OPE. So, we have
" Cn (w)
Ỹ [A, z]Ỹ [B, w] ∼
(z − w)n+1
n"0

and we need to show that

Cn (w) = Ỹ [A(n) B, w]
Define the nth product of two series as
A(w)(n)B(w) = Res [A(z), B(w)](z − w)n
z→0

Then what we need to show is

Ỹ [A, w](n)Ỹ [B, w] = Ỹ [A(n) B, w]
As our fields are constructed by taking normally ordered products we need some way of
seeing how the nth product behaves on normally ordered products. This is called the non-
commutative Wick formula
A(z)(n) : B(z)C(z) : = : (A(z)(n)B(z))C(z) : + : B(z)(A(z)(n)C(z)) : +
" 3n4 5
n−1
6
+ A(z)(k) B(z) (n−k−1) C(z)
k
k=0

This is proved by using the residue formulation for the normally ordered product.

So, we would like to compute

Ỹ [A, z](n)Ỹ [B(−1) C, z] = Ỹ [A, z](n) : Ỹ [B, z]Ỹ [C, z] :
which by the non-commutative Wick formula is
: (Ỹ [A, z](n)Ỹ [B, z])Ỹ [C, z] : + : Ỹ [B, z](Ỹ [A, z](n)Ỹ [C, z]) : +
" 3n4 -
n−1 .
+ Ỹ [A, z](k)Ỹ [B, z] Ỹ [C, z]
k (n−k−1)
k=0

By induction on the order of A, B and C (i.e. linear, quadratic, . . . ) this becomes

: Ỹ [A(n)B, z]Ỹ [C, z] : + : Ỹ [B, z]Ỹ [A(n) C, z] : +
" 3n4
n−1
+ Ỹ [A(k)B, z](n−k−1) Ỹ [C, z]
k
k=0

41
We want to show that this is equal to Ỹ [A(n)B(−1) C, z]. By the commutation relations for
the Fourier coefficients we see

A(n) B(−1) C = [A(n), B(−1) ]C + B(−1) A(n) C

"n 3 4
n 5 6
= A(k)B (n−1−k) C + B(−1) A(n) C
k
k=0
" 3n4 5
n−1
6 5 6
= A(k) B (n−1−k) C + A(n) B (−1) C + B(−1) A(n) C
k
k=0

Taking the vertex operators of this gives

Ỹ [A(n)B(−1) C, z] = : Ỹ [A(n) B, z]Ỹ [C, z] : + : Ỹ [B, z]Ỹ [A(n) C, z] : +

" 3n4 15
n−1
6 2
+ Ỹ A(k) B (n−k−1) C, z
k
k=0

So, we would be finished if we could make our induction hypothesis apply to A(k) B and
C. Assume that A and C are of order at most l and B is linear (order 1). Then by skew-
symmetry we can rewrite A(k) B in terms of elements like B(m) A for m # k. We can rewrite
these as
[B(m) , A(−1)] |0+ + A(−1) B(m) |0+
The last term here is zero as B(m) anihilates the vacuum, the first is of order at most l.15
Hence the induction hypothesis applies.

So far we have shown the result for A of order at most l and B of order at most l + 1.
However, we could repeat the above arguement with A as the compound field to allow us to
increase the order of A. Hence we get the result for A and B of order at most l + 1. Thus,
induction on the order shows the result. !

Corollary 3.9 When κ = κc the elements Sn are central in Ũκ (ĝ). When κ )= κc the
elements Ln give an embedding of the Virasoro algebra into Ũκ (ĝ).

Proof. This follows from the injectivity of the above map and the fact that the elements Jna
topologically generate Ũκ (ĝ). !

3.3 The centre of Ũκ (ĝ)

So, for a special value of the inner product κ we have found some elements of the centre of
Ũκ (ĝ). We should now see if there are any more we can construct. In order to do this we
should think about why the elements we constructed worked.
15
The bracket of an element of order a and an element of order b is of order at most a + b − 1

42
Lemma 3.10 The Fourier coefficients of Y (A, z) will be central if A is anihilated by all
operators J(n)
a
for n ! 0.

Proof. This follows immediately from the fact that the J a (z) generate the vertex algebra
and the OPE formula. !

This leads us to define the centre of a vertex algebra to be

z(V ) = {A ∈ V : [Y (A, z), Y (B, w)] = 0 for all B ∈ V }

By the OPE formula, an equivalent definition of the centre is

z(V ) = {A ∈ V : Y (A, z)B ∈ V [[z]] for all B ∈ V }

This is clearly non-empty because the vacuum |0+ is contained in z(V ).

Lemma 3.11 The centre z(V ) of a vertex algebra V is a vertex algebra itself.

Proof. We already know that |0+ is in the centre. A simple calculation shows that T maps
z(V ) into z(V )
Y (T A, z)B = ∂z Y (A, z)B ∈ V [[z]] for all B ∈ V
Finally we need that the Fourier coefficients of Y (A, z), for A ∈ z(V ), are endomorphisms of
z(V ). Let A, B ∈ z(V ) be central, then we need to show

Y (A(n)B, z)C ∈ V [[z]] for all C ∈ V

For n ! 0 we know that A(n) B = 0 (by the OPE formula) so this is automatic. For n < 0 we
use - .
Y (A(n)B, z) =: ∂z(−n−1) Y (A, z) Y (B, z) :
Using the formula for normally ordered products : AB := A+ B + BA− we easily see that
applying this to C gives something in V [[z]]. !

So the plan is now to study the centres of the vertex algebras associated to AKMA’s and
then to show that we get all central elements for Ũκ (ĝ) from the elements in z(Vκ (g)).

Proposition 3.12 If κ )= κc the centre of Vκ (g) is trivial, i.e.

z(Vκ(g)) = C · |0+

Proof. Suppose that A is a homogeneous element in the centre. Let ω be the vector
generating the Virasoro field "
Y (ω, z) = Ln z −n−2
n

43
recall that we need κ )= κc for this as we needed to divide by this factor. The OPE formula
shows that
Y (A, z)ω ∈ V [[z]] ⇐⇒ Y (ω, z)A ∈ V [[z]]
Taking the coefficient of z −2 in the latter expression shows

L0 (A) = 0

but L0 is the grading operator so we see that A must have grade 0, under our assumptions
this means it is a multiple of the vacuum vector. !

So, we can restrict our attention to the case when κ = κc , because this is the only case when
our vertex algebra methods will generate non-trivial elements of the centre. As we will only
be dealing with one special value for κ we will often omit it.

As we saw above, the OPE formula shows us that the centre z(Vκc (g)), which we will denote
by z(ĝ), is the set vectors of V (g) anihilated by left multiplication by J(n)
a
for n ! 0. Hence

z(ĝ) = V (g)g[[t]]

There is a natural filtration on V (g) (we used it in the induction proof that the map U " (V ) →
Ũ (ĝ) was a homomorphism) given by the order of the element. Denote this filtration by
V (g)!i. We can now look at the associated graded algebra gr(V (g)).

What does gr(V (g)) look like? It is the symmetric algebra with generators given by the J(n)
a

with n < 0. Recalling the definition of ĝ we see that the associated graded is exactly

gr(V (g)) = Sym(t−1 g[[t−1 ]])

As Sym(V ) = Fun(V ∗ ) naturally (where Fun denotes the polynomial functions) we see that
(using the residue pairing)
gr(V (g)) = Fun(g∗ [[t]]dt)
Finally, if we pick a choice for t and an inner product on g we can (un-naturally) identify

gr(V (g)) ∼
= Fun(g[[t]])

Proposition 3.13 Using the induced filtration on z(ĝ) we have

gr(z(ĝ)) ,→ (Fun(g[[t]]))g[[t]]

In other words, we can think of the graded version of the centre as invariant functions on
g[[t]].

Proof. Suppose that A ∈ V (g)!i, denote the projection of A onto the ith piece in gr(V (g))
by
Symbi (A) ∈ V (g)!i /V (g)!(i−1)

44
It is easy to see that (any commutators that occur on the left hand side are removed when we
apply the symbol map; by definition these commutators don’t even occur on the right hand
side)
Symbi (x · A) = x · Symbi (A) for x ∈ g[[t]]
If A ∈ z(g), x · A is always zero and so x · Symbi (A) is always zero. In particular, this means
that the usual symbol map of an element of z(g) is invariant under the action of g[[t]]. As
the symbols generate the image we see the result. !

We need now to study the invariant functions on g[[t]], we denote these by

Inv(g[[t]]) = Fun(g[[t]])g[[t]]

If we can work out the size of this algebra of functions then we will be able to place an upper
bound on the size of z(ĝ) and in the case of the critical level this will (hopefully) be small
enough to allow us to conclude that the elements we constructed using vertex algebras give
all of the centre of Ũκ (ĝ).

In the finite case, the structure of Inv(g) is known (see Theorem 1.5). We will use the elements
Pi to create a large number of elements in Inv(g[[t]]). Let the linear elements in Fun(g) be
denoted by J¯a . In this case, if an inner product κ0 is given then J¯a : g → C is given by

J b ,→ κ0 (J a , J b )

Denote by J¯a (t) the elements "

J¯a (t) = J¯na t−n−1
n<0

whose coefficients J¯na are in Fun(g[[t]]) with the action given by the residue pairing. Suppose
that we write Pi as a polynomial in the linear elements, Pi (J¯a ). Define a set of elements Pi,n
by "
Pi (J¯a (t)) = Pi,n t−n−1
n<0

Each of the elements Pi,n is in Fun(g[[t]]) (note that there are no convergence problems as all
the sums are finite).

Lemma 3.14 The elements Pi,n are in Inv(g[[t]])

Proof. Exercise.

So we have a homomorphism

C[Pi,n ] i=1,...,l −→ Inv(g[[t]])

n<0

we will show that this is actually an isomorphism, but in order to do this we need some more
technology.

45
3.4 Jet schemes

A Jet scheme is a representation of a functor X −→ JX which associates to a scheme X of

finite type, an scheme JX of infinite type which should be thought of as the space of formal
paths on X.

Affine space AN is given by Spec(C[x1, . . . , xN ]), in other words its complex points are just
n-tuples of complex numbers (a1 , . . . , aN ). The formal path space of this should be given by
replacing the complex numbers by formal power series, i.e.

(a1 , . . . , aN ) " (a1 (t), . . . , aN (t))

As there were no conditions on the ai ’s there should be no conditions on the formal power
series. We can equivalently think of the collection of N formal power series as a collection of
all of the Fourier coefficients. Hence the coordinates on JAN are

{ai,n }, i = 1, . . . , N, n < 0

Hence - .
JAN = Spec C[xi,n ] i=1,...,N
n<0

Suppose we have an affine scheme X given by

X = Spec (C[x1 , . . . , xN ]/ *R1 , . . . , RM +)

Where the Ri are polynomials in the variables x1 , . . ., xN . This means that the complex
points of X are n-tuples of complex numbers (a1 , . . . , aN ) satisfying the conditions

Ri(a1 , . . ., aN ) = 0 for all i

As before, the jet scheme of X should be obtained by replacing the coordinates by formal
power series. So we have coordinates (a1 (t), . . . , aN (t)) but now we have the constraints

Ri(a1 (t), . . ., aN (t)) ≡ 0 for all i

We get the constraints on the Fourier coefficients by reading off the coefficients of tn . A nice
way to do this is as follows: Define a derivation of C[xi,n ] by

T : xi,n ,→ −nxi,n−1

Replace xi by xi,−1 in the defining polynomials Ri and call these Ri again. The conditions
that the coordinates of the jet scheme have to satisfy are then given by T n Ri. Hence
3 4
JX = Spec C[xi,n ] i=1,...,N / *T Rj + j=1,...,M
m
n<0 m"0

Finally, any scheme is going to be decomposable into a number of affine schemes which are
glued together in some way. To define the jet scheme on these we simply need to work out

46
how to glue together the jet schemes for the affine pieces. For example P1 can be decomposed
into two affine schemes A1 ∪ A1 where these are the complements of ∞ and 0 respectively.
Therefore they are given by Spec(C[x]) and Spec(C[y]) with identification given by x ↔ y −1 .
We already know how to get the jet schemes of these affine components. The obvious way to
generalise the glueing conditions is to substitute formal power series into them, i.e.

x(t) ←→ y(t)−1

In order to write y(t)−1 as a formal power series it is only necessary to invert y−1 which is the
coordinate corresponding to the original coordinates on the scheme P1 . This is true in general
and hence the glueing is possible (the only difficulties come from the inverted variables, but
these are compatible as they were compatible for the original scheme). Therefore we have
constructed the scheme JX associated to any scheme (of finite type) X.

There are a couple of other ways to think about the construction of JX. It is meant to be
the space of maps from the formal disk D into the scheme X. That is maps

Spec(C[[t]]) −→ Spec(R)

which is the set of homomorphisms

R −→ C[[t]]
So, we can think of J as a functor from the category of algebras over C to the category of
sets which assigns to an algebra R the set of homomorphisms R → C[[t]]. The scheme JX is
then a representation of this functor.

We can define the nth order neighbourhood of a point to be the scheme

8 9
Dn = Spec(C[[t]]/ tn+1 )

So that the formal disk can be represented as an inductive limit

D = lim Dn
←

We can construct nth order jet schemes Jn X which represent the maps from the nth
order neighbourhood of a point into X. Due to the natural maps Dn → Dn−1 we have maps
Jn X → Jn−1 X and so we can form

JX = lim Jn X
←

This gives a realisation of the jet scheme as an inductive limit of schemes of finite type. This
is very useful because the schemes of finite type have very nice properties which the schemes
of infinite type do not.

Lemma 3.15 If a morphism φ : X → Y is smooth and surjective then the corresponding

morphism of jet schemes Jφ : JX → JY is smooth and surjective.

47
3.5 Proof of the theorem

We now use the notion of jet scheme to prove the theorem about the invariant functions on
g[[t]].

Theorem 3.16 Inv(g[[t]]) is isomorphic to C[Pi,n ].

Proof. Inv(g[[t]]) is the space of g[[t]] invariant functions on g[[t]]. If we could integrate the
action to one of the group G[[t]] (where G is the connected, simply-connected group with Lie
algebra g) then we would have

Inv(g[[t]]) = Fun(g[[t]])G[[t]]

This can be shown by considering

1 −→ G(1) −→ G[[t]] −→ G −→ 1

where G(1) consists of the elements of G[[t]] congruent to the identity modulo t. The map
G[[t]] → G is given by sending t to 0. Now, G(1) is a prounipotent group (a projective limit
of unipotent groups). For a unipotent group we know that the exponential map exists and
is surjective (the power series for the exponential will truncate as the Lie algebra will be
nilpotent). Taking the projective limit gives an exponential map from g(1) to G(1) which will
be surjective. So we have invariance of our functions under the large group G(1). To finish
we need to show the invariance under G (this will follow by the standard finite dimensional
arguement). Hence we have our desired invaiance under the group.

In terms of jet schemes, we can rewrite this as

Inv(g[[t]]) = Fun(g[[t]])G[[t]] = Fun(Jg)JG

We would like to think of JG–invariant functions on Jg as the quotient Jg/JG, unfortunately

this is not a variety (consider the C× action on the affine line A1 ); the same problem occurs
even at the level of G–invariant functions on g. Let us see how to deal with this difficulty for
the g/G case.

Define P to be Spec(C[Pi]). We have a map g → P but the fibres of the map do not always
consist of single G–orbits. For example, in the SLn case the inverse image of 0 consists of
the nilpotent cone (all nilpotent matrices in sln ). The orbits of nilpotent matrices are
parameterised by the Jordan forms. There is one big orbit consisting of matrices similar to
 
0 1
 0 1 
 
 .. .. 
 . . 
0

48
and several much smaller orbits corresponding to different Jordan block forms. It is these
smaller orbits that are causing the problems. So, we throw them away

greg = {g ∈ g : Z(g) has dimension l}

where Z(g) is the centraliser of g, l is the rank of g.16 It is clear from the definition that the
set of regular elements is open and dense in g.

Proposition 3.17 The map obtained by restriction

greg −→ P

is smooth, surjective and has fibres consisting of single orbits.

Taking jet schemes we obtain a map

Jgreg −→ JP

which is smooth and surjective. We want to show that this map also has fibres consisting of
single orbits. Consider the map

G × greg −→ greg ×P greg

This is smooth and its surjectivity is equivalent to the fact that the fibres of greg −→ P
consist of single orbits. Taking jet schemes gives

JG × Jgreg −→ Jgreg ×JP Jgreg

is smooth and surjective. This shows that the fibres of Jgreg −→ JP consist of single orbits.

Since the scheme JP is smooth we have that the JG–invariant functions on Jgreg are given
by Fun(JP). The same is therefore true for JG–invariant functions on Jg as the regular
elements are open and dense (and hence uniquely determine the functions). Fun(JP) is
exactly the definition of C[Pi,n ]. !

3.6 Back to the centre

So far we know that there is an injection from the graded centre gr(z(ĝ)) into the invariant
functions Inv(g[[t]]) (Proposition 3.13). The invariant functions have just been shown to be
a polynomial ring C[Pi,n ]. So, in the case of sl2 (C) we can deduce

Proposition 3.18 The centre z(ĝ) of the vertex algebra corresponding to the Lie algebra
sl2 (C) is C[Sn ] |0+. Here, the Sn are the Fourier coefficients of the Segal–Sugawara operator
(we can restrict to n < −1 because the other Sn anihilate the vacuum).
16
This is the minimal possible dimension the centraliser can have

49
Proof. We already know that C[Sm ] |0+ is a subset of the centre of the vertex algebra. To
show that this gives all the elements we just need to show that the symbols of the Sm are
given in terms of the P1,n defined above. Recall that Sn is a normally ordered product of two
fields so it can be written in the form
1" "
Sn = : Jja Ja,k :
2 a
j+k=n

Some of these terms will have both j and k negative and the rest will have at least one of j
and k positive; the former will act

V (g)!i −→ V (g)!(i+2)

(because they are creation operators) and the rest act

V (g)!i −→ V (g)!(i+1)

(becuase they have at least one anihilation operator). Hence the latter terms will be removed
when looking at the symbol. The remaining terms are exactly those which occur in P1,n+1 .!

Corollary 3.19 A similar theorem holds for the completed universal enveloping algebra Ũκ (ĝ):
The centre is topologically generated by C[Sn ] (except this time all the Sn are used).

Proof. This is done by considering a series of vertex algebras constructed from inducing a
one-dimensional representation from g ⊗ tN C[[t]] ⊕ CK to ĝ. Similar theorems to the above
hold for each N and as the completed universal enveloping algebra is an inverse limit of these
vertex algebras we get the result. !

Note that we now have two centres around: we will refer to the centre of the universal
enveloping algebra Ũκ (ĝ) as Z(ĝ) and the centre of the vertex algebra associated to V (ĝ) as
z(ĝ).

As can probably be guessed, both of the above results remain true in the case of a more
general Lie algebra g. Unfortunately there is not a direct proof of this by constructing
“Segal–Sugawara” fields corresponding to the higher order Casimir elements (the leading
order terms are obvious but it is finding the lower order terms that causes lots of problems).

There is however another problem with our current approach that wouldn’t be solved by
finding explicit formulae for the centre generating elements. Everything we have been doing
has been coordinate dependent. We started with a Lie algebra g and constructed the algebra
ĝ from g ⊗ C((t)). But there are many fields which are isomorphic to C((t)) without being
naturally isomorphic. For example: Let X be a smooth algebraic curve (e.g. a Riemann
surface) and x any point on X. The choice of x gives rise to a valuation | · |x on the rational
function C(X). The valuation is defined to be the order of vanishing of f at x (this is
coordinate independent). The completion of the algebra of functions under this valuation is

50
denoted Kx and is isomorphic to C((t)) (pick any function vanishing to order 1 at x) but is
certainly not naturally isomorphic (pick any other function...).

Now, suppose we were given a smooth algebraic curve X and a simple Lie algebra g. We
may want to associate a natural affinisation of g to each point x on X. This can be done by
considering
0 −→ CK −→ ĝx −→ g ⊗ Kx −→ 0
The central extension is given by the cocycle

c(A ⊗ f, B ⊗ g) = κ(A, B) Res(f dg)

This is coordinate independent as f dg is. We can define a coordinate independent vertex

algebra associated to each point as

Indĝg⊗O
x
x ⊕CK
C

where Ox is the ring of integers of Kx (and hence coordinate independent). Now our task
would be to describe the centre of the vertex algebra Vx(g) just defined and as everything so
far has been natural we should try to do this in a coordinate independent way. Therefore we
would like to find our what sort of geometrical object the centre of Vx (g) is — this is where
torsors and opers will come in.

51
4 Geometric Interpretations

We start off by doing something which looks coordinate dependent, however it will eventually
allow us to do things in a coordinate independent way.

4.1 Twisting by torsors

Let O = C[[t]] and K = C((t)) be fixed topological algebras. We study the group of continuous
automorphisms of O which we denote by Aut(O). As t is a topological generator for the
algebra O it suffices to define the automorphism by its action on t.17 So we write

ρ(t) = a0 + a1 t + a2 t2 + · · ·

In order for this to be an automorphism it is necessary and sufficient that a0 = 0 and a1 )= 0.

Let Aut+ (O) be the subgroup of Aut(O) which has a1 = 1. We have an exact sequence

1 −→ Aut+ (O) −→ Aut(O) −→ C× −→ 1

We can use this to work out the structure of Aut(O).

The Lie algebra of Aut+ (O) is denoted Der+ (O). It is easy to see that

Der+ (O) = t2 C[[t]]∂t

Which shows that Der+ (O) is a pronilpotent Lie algebra and hence Aut+ (O) is a prounipotent
group.

If G is a group then a G torsor is a non-empty set S with a right action of G which is simply
transitive. By choosing any point in S we have an obvious isomorphism between G and S,
however ther is no natural choice of such point.

Given a G torsor S and a representation M of G we can twist by S by defining

M = S ×G M

where this is the usual fibre product.

Now take G to be the group Aut(O). Let x be any non-singular point on a variety X. Define
Autx to be the set of all formal coordinates at x (a formal coordinate is a topological
generator of the maximal ideal of Ox ). This is clearly an Aut(O) torsor (let changes of
coordinate act on the formal coordinates).
17
We are using continuity here. We want to pass from information about the action on t to information
about the action on a power series in t. This is possible only becuase we are thinking about continuous
automorphisms.

52
We can now twist the standard versions of the affine algebra, vertex algebra, and centre to
get the intrinsic versions attached to each complex point x of X. For example

Ox = Autx ×Aut(O) C[[t]]

Kx = Autx ×Aut(O) C((t))
ĝκ,x = Autx ×Aut(O) ĝκ
V (ĝκ )x = Autx ×Aut(O) V (ĝκ )
Z(ĝ)x = Autx ×Aut(O) Z(ĝ)
z(ĝ)x = Autx ×Aut(O) z(ĝ)

So, to fully understand the intrinsic nature of these objects we need to know how they behave
under changes of coordinates. We already know how the torsor changes, so we need to study
the Aut(O) action on the vertex algebra.

4.2 Action of coordinate changes

We have a the short exact sequence of automorphism groups

1 −→ Aut+ (O) −→ Aut(O) −→ C× −→ 1

and the short exact sequence of their Lie algebras

0 −→ Der+ (O) −→ Der(O) −→ C −→ 0

As Aut+ (O) is a prounipotent group we know that the exponential map from Der+ (O) to
Aut+ (O) is well defined and onto. We would like to use this to lift a representation from
Der+ (O) to one on Aut+ (O). This will be possible if the action of Der+ (O) is nilpotent (for
then the exponential form will always terminate and hence be well defined). To lift the action
from C to C× is a little more complicated.

C has representations parametrised by any complex number whereas C× only has represen-
tations parametrised by integers. So, although the exponential map from C to C× is always
well defined the representations do not necessarily lift. To ensure that they lift we need the
elements in C to act semisimply and with integral eigenvalues.

Luckily, in our case we have a Z–grading on the vertex algebra which guarantees the above
two conditions will hold. We know that the elements of Der(O) are generated by
∂
Ln = −tn+1 , n = 0, 1, 2, . . .
∂t
The elements of Der+ (O) are generated by the Ln with n ! 1. We have previously shown
that these are elements with positive degree. The Lie algebra C is generated by L0 which
is exactly the grading operator on the vertex algebra. Hence the action of C is semisimply

53
with integral eigenvalues because this is true for the grading operator. It is also clear from
our definition of vertex algebra that the grade is bounded above by 0. This means that any
operator of positive degree must be nilpotent. Hence both the conditions we required hold in
the case we are studying. Hence we only need to describe the Der(O) action on the various
objects.

Restrict for now to the case when g = sl2 . Define the inner product

κ0 (A, B) = trC2 (AB)

and
κ = kκ0
This gives a one-to-one correspondence between inner products and non-zero real numbers
k. The critical inner product κc corresponds to k = −2 under this identification.

We want to study the action of Der(O) on the centre of the vertex algebra. Luckily, for
k )= −2 we have the Segal–Sugawara operators which gave a vertex operator which simulated
the action of Der(O). Specifically, we have a vertex operator S̃(z) whose Fourier coefficients
S̃n acted on the generating elements Jna in exactly the way that elements from Der(O) would.
Here, S̃(z) is defined by (note it isn’t defined when k = −2)

1
S̃(z) = S(z)
k+2
We found that
n3 − n dim(g)k
[S̃n , S̃m] = (n − m)S̃n+m + δn,−m
12 k+2
Multiplying through by k +2 and noting that we can replace ad(S̃n ) with Ln and dim(sl2 ) = 3
we see
k
Ln ◦ Sm = (n − m)Sn+m + (n3 − n)δn,−m
4
This formula is true for all k )= −2 and taking the limit we see

n3 − n
Ln ◦ Sm = (n − m)Sn+m − δn,−m
2
Note that the limit of the Sm ’s does exist (they are the elements generating the centre in
the case of sl2 . We now want to work out what sort of geometrical object gives rise to
transformation laws like this.

As a warm up we will consider what sort of geometrical object gives rise to the simpler objects
a . These transform according to
Jm
<
−mJn+m a
if n + m # −1
Ln ◦ Jm =
a
0 otherwise

54
Lemma 4.1 The object "
a −m−1
J a (t) = Jm t dt
m!−1

is a canonical 1–form on the formal disk D (i.e. it is independent of t).

Proof. This is equivalent to saying that the diagonal action of Der(O) on J a (t) is zero. The
action of Ln on t−m−1 dt is

Ln ◦ (t−m−1 dt) = (m − n)tn−m−1 dt

So
" "
a −m−1
Ln ◦ J a (z) = (Ln ◦ Jm )t dt + a
Jm (Ln ◦ t−m−1 dt)
m!−1 m!−1
" "
= a
−mJm+n t−m−1 dt + a
Jm (m − n)tn−m−1 dt
m!−1 m!−1
n+m!−1
" "
= (n − k)Jka tn−k−1 dt + (m − n)Jm
a n−m−1
t dt
k!−1 m!−1
= 0

Hence we are done. !

Now, in the case of the centre we have the following transformation laws under Ln

 (n − m)Sn+m if n + m # −2
Ln ◦ S m = − 12 (n3 − n) if n + m = 0

0 otherwise

In order to explain the geometrical object that this transformation corresponds to we need
to introduce projective connections

A projective connection on D = Spec(C[[t]]) is a 2nd order differential operator

ρ : Ω−1/2 −→ Ω3/2

such that

• The symbol is 1
• The operator is self adjoint

Let us explain what all these things are. Ω is the set of 1–forms; Ωλ is the set of λ–forms
(things which look like f (t)(dt)λ ). In terms of the action of a vector field ξ(t)∂t an element
f (t)(dt)λ transforms as

ξ(t)∂t ◦ f (t)(dt)λ = (ξ(t)f " (t) + λf (t)ξ " (t))(dt)λ

55
Locally, we can regard Ωλ as functions (by identifying f (t)(dt)λ with f (t)). Under this
identification a 2nd order differential equation is something that can be written in the form
∂2 ∂
ρ(t) = v2 (t) 2
+ v1 (t) + v0 (t)
∂t ∂t
Note here that v0 (t) is not a function (this is due to our identifications). The symbol of the
differential operator is the coefficient v2 (t). So the first condition on projective connections
means that ρ is of the form
∂2 ∂
ρ(t) = 2 + v1 (t) + v0 (t)
∂t ∂t
The adjoint to ρ should be a map Ω3/2∗ → Ω−1/2∗. However there is the residue pairing

Ωλ ⊗ Ω1−λ −→ C

so the adjoint to ρ is again a map Ω−1/2 → Ω3/2. In the above notation the adjoint will be
of the form
∂2 ∂
ρ∗ (t) = 2 − v1 (t) + v0 (t)
∂t ∂t
So the second condition is equivalent to v1 (t) = 0. Therefore a projective connection is a
differential equation of the form
∂
ρ(t) = 2 − v0 (t)
∂t
Note we have changed the sign of v0 (t).

How do projective connections transform under the action of a vector field? As we know the
action on Ω−1/2 and Ω3/2 and that ρ act between these spaces we just need to compute
- . - .
ξ(t)∂t ◦ (∂t2 − v(t)) ◦ f (t)(dt)−1/2 = ξ(t)∂t ◦ (f "" (t) − v(t)f (t))(dt)3/2
3
= ξ(t)(f """(t) − v " (t)f (t) − v(t)f " (t))+
4
3
+ (f "" (t) − v(t)f (t))ξ "(t) (dt)3/2
2
and
- . 3 4
−1/2 1
(∂t2 − v(t)) ◦ ξ(t)∂t ◦ f (t)(dt) = (∂t − v(t)) ξ(t)f (t) − f (t)ξ (t) (dt)−1/2
2 " "
2
3
3 "
= ξ (t)f "" (t) + ξ(t)f """ (t) − v(t)ξ(t)f " (t)−
2
4
1 """ 1 "
− f (t)ξ (t) + v(t)f (t)ξ (t) (dt)3/2
2 2
The action on the projective connection will be the difference of these. This gives
1
ξ(t)∂t ◦ (∂t2 − v(t)) = −ξ(t)v " (t) − 2v(t)ξ "(t) + ξ """ (t)
2

56
Note there is no ∂t2 term here becuase we were acting by an element of Der(O) which is the
infinitessimal version of Aut(O) and the coefficient in front of ∂t2 is constantly 1.

It is possible to exponentiate up the action of Der(O) to give the action of Aut(O) on projec-
tive connections. Suppose that we have two formal coordinates z1 and z2 which are related
by a change of coordinates z1 = f (z2 ). In the first coordinates, the projective connection ρ
has the form
∂2
ρ(z1 ) = 2 − v1 (z1 )
∂z1
and in the second coordinate
∂2
ρ(z2 ) = − v2 (z2 )
∂z22
Then the relationship between v1 and v2 is given by
3 4
∂f 2 1
v2 (z2 ) = v1 (f (z2 )) − {f, z2 }
∂z2 2
where 3 42
f """ 3 f ""
{f, z} = " −
f 2 f"
is the Schwartzian derivative of f .

Lemma 4.2 The object

"
ρ(t) = ∂t2 − v(t) = ∂t2 − Sm t−m−2
m!−2

is a canonical projective connection on the formal disk D (i.e. it is independent of t).

Proof. It is certainly a projective connection, so all we need to do is check that the diagonal
action of Der(O) is zero. The action of Ln = −tn+1 ∂t on the t part of ρ(t) is
1
Ln ◦ ρ(t) = tn+1 v " (t) + 2(n + 1)tn v(t) − (n3 − n)tn−2
2
" 5 6 1
= −(m + 2)Sm t n−m−2
+ 2(n + 1)Sm tn−m−2 − (n3 − n)tn−2
2
m!−2
" 1
= (2n − m)Sm tn−m−2 − (n3 − n)tn−2
2
m!−2

and the action on the Sm part is

" " 1
Ln ◦ ρ(t) = − (n − m)Sn+m t−m−2 + (n3 − n)t−m−2
m!−2 m!−2
2
n+m!−2 n+m=0
" 1
= − (2n − k)Sk tn−k−2 + (n3 − n)tn−2
2
k!−2

57
Summing these gives zero. !

@ 2 vertex algebra at the critical level is geometrically given by

Hence the centre of the sl
functions on the space of projective connections on the formal disk. Twisting this by the
torsor Autx we see

Theorem 4.3 The centre z(sl @ 2 )x attached to a point x on a curve X is canonically isomor-
phic to Fun(P rojx ), where P rojx is the space of projective connections on the formal disk
attached to the point x.

Corollary 4.4 The centre Z(sl @ 2 )x attached to a point x on a curve X is canonically iso-
morphic to Fun(P rojx ), where P rojx× is the space of projective connections on the punctured
×

formal disk attached to the point x.

As the centre of the universal enveloping algebra has been identified with functions on the
space of projective connections on the punctured formal disk, to any ρ ∈ P roj × we obtain a
central character
@ 2 ) −→ C
χρ : Z(sl
by evaluation at ρ. Thus, if M is a smooth representation we can attempt to force the centre
to act by this character
Mρ = M/span{f · m : f ∈ ker(χρ ), m ∈ M }
Of course, it is possible that this may end up being the trivial module.

@ 2 ) acts by the character ρ.

Lemma 4.5 On Mρ the centre Z(sl

Proof. Take z in the centre, m ∈ Mρ and consider

(z − χρ (z)) · m ∈ Mρ
It is clear that z − χρ (z) is in the kernel of χρ , hence this element must be zero. !

@ 2 ), then we have a natural map

Now, take M to be the vacuum module Vκc (sl
@ 2 ) −→ End b Vκ (sl
Z(sl @ 2)
sl2 c

@ 2 ) due to the fact that all Sn with n ! −1

However, this map certainly factors through z(sl
anihilate the vacuum. Thus we have
@ 2)
Z(sl # End b Vκ (sl @ 2)
sl2
c

$ %
&
$ %
$ %
$ % α
!
$ %
@2)
z(sl

58
Lemma 4.6 The map α is an isomorphism.

Proof. The inverse is given by φ ,→ φ(|0+). !

Because the central elements factor through the vertex algebra centre, in any Vρ we must have
the elemtents Sn with n ! −1 acting as zero. This means that the projective connection is
actually a projective connection on the formal disk (rather than the punctured formal disk).

Proposition 4.7 Suppose that ρ is a projective connection on the formal disk

"
ρ(x) = ∂t2 − un t−n−2
n!−2

@ 2 on which the element Sn acts as the scalar

Then Vρ is a non–zero, irreducible module of sl
un .

@ 2,κ modules such that

Theorem 4.8 Let C0 be the category of (smooth) sl c

• sl2 [[t]] acts locally finitely

@ 2 ) acts through z(sl
• Z(sl @2)

and M0 the category of quasi–coherent sheaves on P roj(D) (which is the category of modules
@ 2 )). Then C0 is equivalent to M0 where the equivalences are given as
over z(sl

F0 : C0 → M0 ; M ,→ Homsl @
b 2 (Vκc (sl2 ), M ) = M
sl2 [[t]]

@ 2) ⊗ F
G0 : M0 → C0 ; F ,→ Vκc (sl
b 2)
z(sl

4.3 Projective structures

If X is a Riemann surface then a projective structure on X is given by covering X by an

atlas of charts Uα with local coordinates zα such that the transition functions

zβ = fαβ (zα)

are Möbius transformations.

Proposition 4.9 There is a bijection between the projective structures on X and the projec-
tive connections on X.

59
Proof. It is easy to check that the Schwartzian derivative of a Möbius transformation is
zero. Hence, given a projective structure we can define a projective connection by assigning
the second order operator ∂z2α on each chart Uα . These transform correctly.

Given a projective connection we can consider the solution to the homogeneous equation
5 2 6
∂zα − vα φ(zα) = 0
on Uα . This has a 2–dimensional space of solutions, φ1,α and φ2,α. Choosing the cover to
be fine enough we can assume that φ2 is never 0 and the Wronskian of the solutions is never
zero. Define
φ1,α
µα = : Uα −→ C
φ2,α
This is well defined and has non–zero derivative (the derivative is basically the Wronskian).
Hence, near to the origin it gives a complex coordinate on Uα . In a different basis it is clear
that the µ’s will be related by a Möbius transformation. !

If X is a Riemann surface of genus > 1 then the uniformisation theorem tells us that X ∼
= H/Γ
for some Γ a group of real Möbius transformations. This clearly gives us a very special
type of projective structure on X — namely one whose transition functions are real Möbius
transformations.

Theorem 4.10 The vector space of C–projective structures on a Riemann surface X of

genus g has dimension
0 if g = 0
1 if g = 1
3g − 3 otherwise

We now describe another way to think about projective sturctures (which is therefore another
way to think about projective connections). This way will be the one that we will be able to
generalise to a more general situation.

We can think of a Möbius transformation as giving an element of PGL2 (C). Hence, if we

have a projective structure on X then to two charts Uα and Uβ which overlap we have a map
φαβ : Uα ∩ Uβ −→ PGL2 (C)
Provided we know the terminology it is clear that this is giving us a principal PGL2 (C)–bundle
with flat connection over X.

4.4 Opers

A principal G–bundle over a manifold X is a manifold P which is fibred over X such that
there is a natural right–G action on P making each fibre into a G–torsor. In addition, given

60
a small open set U ⊂ X there is a trivialisation of the bundle P|U

tU : P|U −→ U × G

commuting with the obvious right–G actions.

Given a principal G–bundle and a manifold V with left G–action we can form the associated
bundle P ×G V which is then a bundle over X with fibres isomorphic to V and transition
functions given by elements of G.

Given a principal G–bundle and two small overlaping charts Uα and Uβ there are trivialisa-
tions tα and tβ . We therefore get maps

tα ◦ t−1
β : (Uα ∩ Uβ ) × G −→ (Uα ∩ Uβ ) × G

As this will preserve fibres and the action of G is transitive we can think of these as maps

φαβ : Uα ∩ Uβ −→ G

These maps satisfy the obvious condition on triple overlaps

φαγ = φαβ ◦ φβγ

In fact, given any maps satisfying these conditions we can use them to “glue” together copies
of G to form a principal G–bundle.

If the maps φαβ are constant functions then we say that the bundle P has a flat connection.
In this case, there is a canonical identification of nearby fibres. This gives us a way to
“transport” elements of a fibre to elements of a different fibre by moving then along a path
(this is what a connection is). The transportation depends only on the image of the path
in π1 (X) (this is what flat means).

Flat connections are closely related to first order differential equations. Given a principal
G–bundle with flat connection there are canonical local sections near any given point x: Pick
a point in the bundle above x and above nearby points x" pick the identified points. Under
a local trivialisation these sections will also be the solutions to a set of first order linear
differential equations:
∂ ∂
s(x) = 0, . . . , s(x) = 0
∂x1 ∂xn
This means that flat connections locally give sections which are solutions to first order linear
ODE’s.

Continuing with this, we can think of connections as giving a way to “differentiate” sections
of a vector bundle. Suppose that X is a real manifold and E some vector bundle over X.
Then we have the sheaf OX of smooth functions on X, the sheaf T of smooth vector fields
on X and the sheaf E(E) of smooth endomorphisms of E. A connection on E is a map

∇ : T −→ E(E)

61
which takes a tangent field ξ to an endomorphism ∇ξ . It has the following additional prop-
erties

• It is OX linear: ∇f ξ = f ∇ξ for f ∈ OX .

• It satisfies Leibnitz: ∇ξ (f φ) = f ∇ξ (φ) + φ · ξ(f ) for f ∈ OX and φ ∈ E(E).

A flat connection has the additional property that ∇ is a Lie algebra homomorphism

[∇ξ , ∇χ] = ∇[ξ,χ]

If we pick a small enough chart we can pick coordinates x1 , . . . , xm on the manifold X and
trivialise the bundle. Then the connection takes the form
3 4
∂ ∂
∇ = + Ai (x)
∂xi ∂xi

where Ai is an endomorphism of Rn (n is the dimension of the bundle). This is exactly the

type of differential equation we saw before. The flatness condition then takes the form
∂Aj ∂Ai
− + [Ai , Aj ] = 0
∂xi ∂xj
So, for a very concrete description of a flat connection in the real case we can simply specify
matrices Ai locally, satisfying the above equations and such that they transform in the correct
way under changes of coordinates.

Suppose now that X is a complex manifold. A holomorphic vector bundle E over X

is a vector bundle over X together with an action of the anti–holomorphic vector fields on
sections of X. This action of anti–holomorphic vector fields gives us half of a flat connection.
So, to specify a flat connection on E we only need to define an action of the holomorphic
tangent fields on holomorphic sections of E (all the other conditions follow automatically).
Therefore, we locally have 3 4
∂ ∂
∇ = + Ai (z)
∂zi ∂zi
where Ai is a holomorphic function. The corresponding anti–holomorphic piece is
3 4
∂ ∂
∇ =
∂ z̄i ∂ z̄i
As before we can use the Ai to define connection.

What about connections on G–bundles? So far we have only given the differential equa-
tion formulation of connections on vector bundles. Given a principal G–bundle P and a
representation V of G we can form the associated vector bundle

VP = P × V
G

62
Any local identification of fibres on P is clearly going to give a local identification of fibres
on VP . Hence, a principal G–bundle with flat connection defines a functor from the category
of representations of G to the category of flat vector bundles on X. Both these categories
are tensor categories and it is clear that the above functor is actually a tensor functor. By
restricting the corresponding vector bundle to a point x ∈ X we have a tensor map from the
representations of G to vector spaces. This is a fibre functor and so the Tannakian recon-
struction theorem allows us to obtain G from only this data. This allows us to reconstruct
the bundle P from the functor it defines. Hence the connection on P must locally be of the
form 3 4
∂ ∂
∇ = + ai (x)
∂xi ∂xi
where now the ai must act as first order differential operators on all representations of G.
Hence, ai ∈ g.

So, now we have seen that a projective structure on X is defining a principal PGL2 (C)–bundle
with flat connection. However, we have not yet used the fact that we had coordinates zα
around...

If we form the associated P1 –bundle P ×PGL2 (C) P1 (which is possible because there is a natural
left action of PGL2 (C) on P1 as there is a natural action of GL2 (C) on C2 ). Now, if we have a
chart Uα which is small enough so that the bundle becomes trivial we can use the coordinate
zα to define a local section: We simply pick the section to take value zα (x) at the point x.
This is actually a global section because the coordinates transform between each other by
exactly the element of PGL2 (C) which is used to transform between overlaping charts. As zα
is a coordinate it has non–vanishing derivative at all points. Therefore the coordinates are
giving us a global section of the associated P1 –bundle which has non–vanishing derivative at
all points.

We define a PGL2 (C)–oper on X to be a principal PGL2 (C)–bundle P over X with flat

connection such that the associated P1 –bundle has a globally defined section with nowhere–
vanishing derivative (this is the same as the section being transverse to all locally defined
horizontal sections).

So, we have seen that a projective structure on X gives us a PGL2 (C)–oper on X. It

is clear that this identification is reversable (use the section to define coordinates and the
principal G–bundle transition functions to define the Möbius transformations). Hence we see
that projective structures (and hence projective connections) on X are the same thing as
PGL2 (C)–opers on X. When we manage to generalise the notion of oper we will be able to
give the geometric definition of the centre of the universal enveloping algebra of a general
affine simple Lie group.

In order to generalise the concept of oper we need to work out what P1 has to do with the
group PGL2 (C) and what the transverse section is telling us.

It is well known that P1 can be represented as a homogeneous space of PGL2 (C) given by

63
PGL2 (C)/B where B is the Borel subgroup (of upper triangular matrices). The Borel
subgroup is a concept that easily generalises to more general groups so G/B should be the
correct generalisation of P1 .

Now, an element of P1 can be regarded as defining a right coset of B in G. Hence, above

each point x ∈ X we have (using the section of the P1 bundle) a right coest of B. This means
that the section of the P1 bundle is giving us a sub–bundle of the principal G–bundle which
is a principal B–bundle.

A reduction to B is a subbundle of a principal G–bundle which is a B–bundle where the

action of B agrees with the action restricted from G and the G–bundle can be recovered by
inducing the B–bundle.

So, an oper is (so far) a flat principal G–bundle with a reduction to the Borel B. We still
have to interpret the transversality of the section in these terms.

The transversality condition is basically telling us that the flat connection on the G–bundle
does not preserve the subbundle. Locally, we know that we can write the connection in the
following form 3 4
a(x) b(x)
∇ = ∂x +
c(x) d(x)
where the matrix is in the Lie algebra for G (in our case this is sl2 (C) and so a(x)+d(x) = 0).
Here we will identify the first component of the trivialisation of the G–bundle with the
B subbundle. The transversallity condition then becomes that c(x) is nowhere vanishing.
Changing coordinates changes the connection by conjugation

∇ −→ f ∇f −1

If we want to preserve the first component as being the B subbundle we can only change
coordinates by matrices from the Borel subgroup. This allows us to put the connection into
the form 3 4
0 v(x)
∇ = ∂x −
1 0
In fact, all of the freedom available in B is used in order to obtain this form.

As is well known from the theory of matrix differential equations, the above differential
equation is basically equivalent to the second order equation

∂x2 − v(x) = 0

which looks like a projective connection. Of course, we need to check that v(x) transforms
correctly (in other words, is v(x) a map Ω−1/2 −→ Ω3/2?).

We have a smooth curve X and a collection of bundles 0 ⊂ P1 ⊂ P2 = P where P is the

principle G–bundle and P1 is the B–bundle. We also have a connection ∇ : P → P ⊗ Ω

64
(rememeber that Ω and T are dual by the obvious pairing). The transversality condition
tells us that
∇(P1 ) ⊂ P2 ⊗ Ω
and the induced map
P1 −→ (P2 /P1 ) ⊗ Ω
is an isomorphism. Although the determinant line bundle of P is not well defined, it is defined
“up to tensoring with some line bundle”. So, det(P) ∼= OX up to some line bundle. From
the induced isomorphism we can therefore deduce that P1 is Ω1/2; it follows from this that
P2 /P1 is isomorphic to Ω−1/2. Hence we have
0 −→ Ω1/2 −→ P −→ Ω−1/2 −→ 0
It is known that
Ext1 (Ω−1/2, Ω1/2) ∼
= H 1 (Ω) ∼
=C
so that there are only two isomorphism classes of extensions of Ω−1/2 by Ω1/2 (the split and
non–split extensions). If the extension were split then there would be an induced connection
on the line bundle Ω1/2; this bundle has degree g − 1 and so can not have a connection if
g )= 1. So, in the case where g )= 1 we have a unique P. The bundle P can be thought of as
the space of 1–jets of sections of the line bundle Ω−1/2 .

We now see that the element v(x) of the connection is actually a map
Ω−1/2 ∼
= P2 /P1 −→ P1 ⊗ Ω ∼
= Ω3/2
So, our formulation of opers certainly gives us projective connections.

In the more general case of a G–oper, we have a principle G–bundle F and a reduction to FB
(where B is a Borel subgroup of G). The Lie algebra g of G has a well known decomposition
g = n− ⊕ h ⊕ n+
The Lie algebra for B can be taken to be b = h ⊕ n+ . The space of positive roots n+ is
spanned by simple roots e1 , . . . , el and the space of negative roots is spanned by simple roots
f1 , . . . , fl . The transversality condition we want is that the connection takes the form
l
"
∇ = ∂t + ψi(t)fi + v(t)
i=1

where v(t) ∈ b[[t]] and the ψi (t) are nowhere vanishing. By scaling we can actually assume
that the ψi (t) are always 1.

One nice thing about the PGL2 case was the canonical form for the connection (in the above
paragraph we could assume even more about the function v(t)). We can get a nice canonical
form for the G–oper as follows: Let
l
"
p−1 = fi
i=1

65
and choose p0 and p1 so that the collection is an sl2 –triple with p0 consisting of elements of
h and p1 a sum of the simple roots e1 , . . . , el . For example, in the sln (with rank l = n − 1)
case we get
 
0
 1 0 
 
 1 0 
p−1 =  
 1 0 
 
.. ..
. .
 
l
 l−2 
 
 l−4 
p0 =  
 l−6 
 
..
.
 
0 l
 0 2l − 2 
 
 0 3l − 6 
p1 =  
 0 4l − 12 
 
.. ..
. .

The operator 12 ad p0 defines the principle gradation on b and in this gradation p−1 acts
from bi+1 to bi . We can therefore find a space Vi given by

bi = [p−1 , bi] ⊕ Vi

This space is known to be non–zero only if i is an exponent of the Lie algebra g and its
dimension is the multiplicity of this exponent (almost always 1).

We can decompose the ad p1 invariants of n+ using this decomposition and pick elements pj
spanning the relevant spaces. We can then write the connection in the form
l
"
∇ = ∂t + p−1 + vi (t)pi
i=1

We can examine the action of changes of coordinates on this canonical form. The result is

Lemma 4.11 ∂t2 − v1 transforms as a projective connection and the vi transform as (di + 1)–
forms (where di is the exponent). Hence the space of opers is
l
!
OpG (D) ∼
= P roj × Ω⊗(di +1)
i=2

66
When the group G is the adjoint18 form of the Lie algebra g we will also write Opg for the
space of G opers.

4.5 Miura Opers

For sl2 we have seen that we can write an oper connection in the form
3 4
0 v(t)
∂t +
1 0
Instead of these we could look at connections of the form
3 4
a(t) 0
∂t +
1 −a(t)
We know that we can gauge transform connections of this form into the standard oper form.
If we do this calculation we see that this is acomplished by letting
v(t) = a(t)2 + a(t)"
This equation is known as the Ricatti equation and in this context is the Miura trans-
formation.

We might hope that connections of this form would be a little easier to understand. In
particular, if we remove the ones below the diagonal then this is a connection on a very nice
H–bundle (where H is the subgroup of diagonal matrices). The generalisation of H to a more
general group G is obvious (it is the Cartan subgroup); what is slightly more difficult is
generalising the H–bundle.

A character of H is a group homomorphism α : H −→ C× . Given a character of H and a

principle H–bundle FH we can form the associated line bundle
F H × Cα
H

where C has an action of H via the character α. There are characters corresponding to the
positive simple roots, we denote these by α1 , . . . , αl . Then there is a special principle G–
∨
bundle, denoted Ωρ which has the property that its twist by any positive simple character
is just the canonical line bundle Ω. This bundle is the correct generalisation of the bundle
from the sl2 case.
∨ ∨
We now have the space Conn(Ωρ )D of connections on the bundle Ωρ over the formal disk
D. There is an obvious map from here to the space of opers given by
l
"
∂t + u(t) −→ ∂t + u(t) + fi
i=1
18
The adjoint form for a Lie algebra is the connected Lie group with minimal centre for which it is a Lie
algebra. It is the image of the map Ad : G −→ g, hence the name adjoint.

67
We have to take gauge transformations of the right hand side if we want to bring it into the
canonical form (just like the sl2 case).

In the sln case the oper connection corresponding to the above map is of the form
 
u1
 1 u2 
 
 1 u3 
∂t +  
 . . . . 
 . . 
1 ul

This has the additional property that it is a connection which preserves the lower triangular
matrices. This suggests that we can think of these Miura opers as opers with an extra
reduction which is preserved by the connection.

A Miura oper is a quadruple (F , FB , ∇, FB" ) where the first three form an oper and FB" is
a B reduction of F which is preserved by B.

Recall that we could think of a reduction from G to B as defining a section of F ×G G/B.

We know that FB gives a transversal section and FB" a horizonatal section. Hence, FB" is
determined by its value at 0 (the origin of the disk). The section FB gives us a second point
above 0. There is a notion of “generic relative position” for two points in G/B.

Assume that we have two points x and y in G/B. We may assume that x is simply the
identity element (by the action of G). We have the Bruhat decomposition
A
G/B = BwB
w∈W

So, the element y corresponds to a unique element w of the Weyl group. The Bruhat de-
composition has a unique large open orbit corresponding to the longest element w0 ∈ W .
If y corresponds to this element of W then we say that x and y are in generic relative
position. In the case of P1 this corresponds to the very simple condition that x is not equal
to y.

We denote by MOpg(D) the set of Miura opers for which the elements above 0 determined
by the two reductions to B are in generic relative position.

∨
Proposition 4.12 The spaces MOpg(D) and Conn(Ωρ )D are isomorphic.

There is an obvious forgetful map from MOpg(D) to Opg(D). The group NFB,0 (which is
N twisted by FB,0 and this is the fibre above 0 of FB ) acts on (G/B)F0 and the subset of
generic reductions is the open orbit. This orbit is an NFB,0 torsor, hence the forgetful map
gives MOpg(D) is an NFB,0 –bundle over the space of opers Opg(D).

68
Looking at the function spaces we now have an injection
∨
Fun(Opg(D)) ,→ Fun(Conn(Ωρ ))

The image is clearly NFB,0 –invariant. At the level of Lie algebras the image will be n–
invariant. Hence it will be anihilated by each of the operators corresponding to the positive
simple roots e1 , . . . , el .

The action of ei is given by an infinitessimal gauge transformation

B C
"
δu(t) = xi (t)ei , ∂t + fi + u(t)
i

where xi (t) ∈ C[[t]], xi (0) = 1 and is such that the infinitessimal gauge transformation
always remains in h[[t]]. These conditions uniquely determine xi (t): multiplying out we see
that being in h[[t]] requires
∂t xi (t) = −ui (t)xi (t)
Hence & '
" ui,−m
xi (t) = exp − tm
m
m>0

where we expanded ui (t) in a power series. This means that the action of ei on uj (t) is given
by aij xi (t) where aij is the ij–entry in the Cartan matrix for g. Thus ei acts by the derivation
l
" " ∂
δi = aij xi,n+1
∂ui,n
j=1 n<0

where xi,n is the coefficient of tn in xi (t).

Proposition 4.13 The image of Fun(Opg(D)) is isomorphic to the intersection of the ker-
nels of the operators δi .

This proposition is simply the statement at the top of this page about Fun(Opg(D)) being
the annihilated by the operators ei .

To any finite dimensional simple Lie algebra g we can associate the Weyl algebra. This has
generators given by
∂
aα,n = , a∗α,n = yα,−n α ∈ ∆+ , n ∈ Z
∂yα,n
and non–trivial relations / 0
aα,n, a∗β,m = δα,β δn,−m
This gives rise to a vertex algebra in the usual way. We define the an to kill the vacuum if
n ! 0 and the a∗n to kill the vacuum if n > 0. This vertex algebra is denotecd Mg.

69
We want a vertex algebra homomorphism Vκc (g) −→ Mg. The existence of this is equivalent
to a map ĝκc −→ U (Mg). Now, the universal enveloping algebra is like differential operators
on n((t)). In the finite dimensional case this would be like looking for a map

g −→ Diff(n)

This is easy to obtain using the flag manifold. n is isomorphic to the open orbit in G/B.
Thus we have a map from g to vector fields on n. As we can lift vector fields to differential
operators we get the required map. This is possible becuase the following exact sequence is
split in the finite dimensional case

0 → Fun(N) → Diff !1 (N ) → Vect(N) → 0

The splitting is given by lifting a vector field to a differential operator which anihilates the
function 1. In the infinite dimensional case we will not be able to do this (because the function
1 will not exist). So, we will not be able to lift our map from loops on g to vector fields on
N ((t)) to one on differential operators. Fortunately, we will be able to get a map from the
affine algebra to the differential operators precisely in the case that we are at the critical
level.

In the sl2 case these differential operators are

d d d
e ,→ − , h ,→ −2y , f ,→ y 2
dy dy dy
where y is the coordinate on N .

If we want to generalise the finite dimensional case we would look at the space D× −→ G/B
(which should be the loop space of G/B). Unfortuantely, this is very hard to define (it is
obvious what to do on affine pieces, unfortunately how to glue these back together is messed
up by the loop construction). Fortunately, we only need to do this on the open orbit (which
is affine). So we look at
D× −→ N w0 B
which is isomorphic to N ((t)) and n((t)).
ˆ 2 case we get
In the sl

e(z) ,→ −a(z), h(z) ,→ −2a∗ (z)a(z), f (z) ,→ a∗ (z)2 a(z)

where a and a∗ are fields from Msl2 .

Unfortuately, these formulae do not involve normal ordering and so can not be well defined on
Mg. What we have actually produced is well defined on Fun(C((t))). In the finite dimensional
case we would be able to obtain from this a map from sl2 ((t)) to U (Msl2 ) by using the splitting
of the diagram

0 # U (M )!0 # U (M )!1 # τloc #0

70
However, such a splitting does not occur in our case. As a replacement we can try to look
for a map from the affine Lie algebra to the differential operators. An obvious guess about
how to do this would be to normally order the operators defined above. Unfortunately, this
doesn’t work. By degree considerations we see that we can not modify the definitions of e(z)
and h(z) but we can modify the definition of f (z) by adding multiples of ∂z a∗ (z). If we try
to do this we get the following

e(z) ,→ −a(z), h(z) ,→ −2 : a∗ (z)a(z) :, f (z) ,→: a∗ (z)2 a(z) : +2∂z a∗ (z)

which gives a vertex algebra homomorphism

Vκc (sl2 ) −→ Msl2

However, the choice of this homomorphism is not unique. Any two homomorphisms will differ
by a map
φ : sl2 ((t)) −→ U (Msl2 )!0
which must be a one–cocycle.

Lemma 4.14 The space of such cocycles is isomorphic to h((t))∗ ∼

= h∗ ((t))dt.

So, we can pick some b(t)dt in h∗ ((t))dt and set

e(z) ,→ −a(z), h(z) ,→ −2 : a∗ (z)a(z) : +b(z), f (z) ,→: a∗ (z)2 a(z) : +2∂z a∗ (z) − b(z)a∗ (z)

This gives a vertex algebra homomorphism

Vκc (sl2 ) −→ Msl2 ⊗ π

where π is a commutative vertex algebra generated by h∗ ((t)).

Since π can be thought of as functions on one–forms we have a map

Vκc (sl2 ) −→ Msl2 ⊗ Fun(1 forms)D

but we wanted a map to functions on connections. Fortunately, the space of connections

∨
ConnL H (Ωρ )D is a torsor over L h[[t]]dt. So, we just need to notice that all our homomor-
phisms were not Aut(O) invariant. To make them invariant we need to replace the one–forms
by connections to get

Vκc (sl2 ) # Msl ⊗ FunConnL (Ωρ∨ )D

2 H
" "

z(sl2 ) # Fun ConnL (Ωρ∨ )D

H

71
We want to show that the image of the bottom arrow is anihilated by all the operators δi
(because then it will be exactly the space of functions on opers that we want).

The above all generalizes to g a semisimple Lie algebra.

To prove the result we construct “screening operators” S̃i,κc such that the image of z(ĝ)
is contained in the kernel of all these operators. Then we show that these operators act
identically to the δi when restricted to π. This shows that z(ĝ) is contained inside the space
we want. To show it is the whole space we need to do a “dimension count”. Since everything
is infinite dimensional we can not do this in the obvious way, however we can look at the
characters of the spaces. If this is all done carefully we get

Theorem 4.15 z(ĝ) is isomorphic to Fun OpL g(D).

72