arXiv:math/9909178v1 [math.

QA] 29 Sep 1999

Vertex operator algebras and the zeta

function
J. Lepowsky
Dedicated to Howard Garland on the occasion of his sixtieth
birthday
Abstract
We announce a new type of Jacobi identity for vertex operator
algebras, incorporating values of the Riemann zeta function at negative integers. Using this we explain and generalize some recent work
of S. Blochs relating values of the zeta function with the commutators
of certain operators and Lie algebras of differential operators.

Introduction

Consider the famous classical formula

1+2+3+ =

1
,
12

(1.1)

which has the rigorous meaning

1
.
12
Here is of course the Riemann zeta function
(1) =

(s) =

ns

(1.2)

(1.3)

n>0

(analytically continued), and (1.1) is classically generalized by the formal

equality
X
ns = (s)
(1.4)
n>0

for s = 0, 1, 2, . . .. The classical number theory behind this analytic continuation is well known to be related to the widely-pervasive issue of regularizing
certain infinities in quantum field theory, in particular, in conformal field
theory. Here we shall announce some general principles of vertex operator
algebra theory that elucidate the passage from the unrigorous but suggestive
formula (1.1) to formula (1.2), and the generalization (1.4). In the process,
we shall explain some recent work of S. Blochs involving zeta-values and
differential operators. The work [L2] contains details and related results.
The material that we shall present involves foundational notions of vertex
operator algebra theory, and we shall try to make this writeup accessible to
nonspecialists by reviewing elementary matters.
We were motivated by a desire to understand some very interesting phenomena found by Bloch [Bl] relating the values (n), n = 1, 3, 5, . . ., of the
zeta function at negative odd integers to the commuatators of certain operators on an infinite-dimensional space. We shall begin with some elementary
background and a brief description of this work, then we shall explain how to
recover and somewhat generalize these results using vertex operator algebra
theory, and finally, we shall place these ideas and results into a very general
context and present some new general results in vertex operator algebra theory. These methods serve incidentally to enhance the many already-existing
motivations for vertex operator algebra theory (see [Bo], [FLM]) and its underlying formal calculus (as developed in [FLM] and [FHL]).
One of our main themes is to always use generating functionsto introduce new formal variables and generating functions in order to try to make
complicated things easier and more natural and at the same time, much more
general, as in the corresponding parts of [FLM]. We use commuting formal
variables rather than complex variables because they provide the most natP
ural way to handle the doubly-infinite series such as (x) = nZ xn that
pervade the natural formulations and proofs. Other central themes are to
d
exploit the formal exponential of the differential operator x dx
as a formal
change-of-variables automorphism (again as in [FLM]); to formulate Eulers
interpretation of the divergent series (1.4) by means of the operator product
expansion in conformal field theory; and to place considerations about Lie
algebras of differential operators into the very general context of what we
termed the Jacobi identity [FLM] for vertex operator algebras. There are
some interesting points of contact between the present work and [KR], [M]
and [FKRW].
2

I am very pleased to dedicate this paper to Howard Garland on the occasion of his sixtieth birthday. This work was presented in a talk at Yale
University in fall, 1997 at a seminar in his honor. I would like to mention
here that it was from Howard Garland that Robert Wilson and I, in 1980,
first learned about the idea of using formal delta-function calculus, which
was also used in [DKM]; cf. [FLM]. This was one among many of Howards
insights that have influenced us.
This work was also presented in a talk at the May, 1998 Conference on
Representations of Affine and Quantum Affine Algebras and Related Topics
at North Carolina State University. I would like to thank Naihuan Jing and
Kailash Misra, the organizers, for a stimulating conference.
I am very grateful to Spencer Bloch for informing me about his work and
for many valuable discussions.
This work was partially supported by NSF grants DMS-9401851 and
DMS-9701150.

Background

Consider the commutative associative algebra C[t, t1 ] of Laurent polynomials in an indeterminate t, and consider its Lie algebra d of derivations:
d

= Der C[t, t1 ],

(2.1)

the Lie algebra of formal vector fields on the circle, with basis {tn D|n Z},
where
d
D = Dt = t .
(2.2)
dt
(A preview of one of our main themes: The homogeneous differential operator Dt , rather than dtd , will be the appropriate form of differentiation for
our considerations, and we shall be using it for various variables as well as
t.)
Consider also the Virasoro algebra v, the well-known central extension
0 Cc v d 0,

(2.3)

where v has basis {L(n)|n Z} together with a central element c; the bracket
relations among the L(n) are given by
1
[L(m), L(n)] = (m n)L(m + n) + (m3 m)m+n,0 c
(2.4)
12
3

1
(m3 m), being oneand L(n) maps to tn D in (2.3).
The number 12

half the binomial coefficient m+1
, is closely related to a third derivative,
3
which becomes visible when we use generating functions to write the bracket
relations (2.4), as we review below. The Virasoro algebra is naturally Zgraded, with deg L(n) = n and deg c = 0.
The following classical realization of the Lie algebra v is well known: We
start with the Heisenberg Lie algebra with basis consisting of the symbols
h(n) for n Z, n 6= 0 and a central element 1, with the bracket relations

[h(m), h(n)] = mm+n,0 1.

(2.5)

For convenience we adjoin an additional central basis element h(0), so that

the relations (2.5) hold for all m, n Z. This Lie algebra acts irreducibly on
the polynomial algebra
S = C[h(1), h(2), h(3), . . .]

(2.6)

as follows: For n < 0, h(n) acts as the multiplication operator; for n > 0,

h(n) acts as the operator n h(n)

; h(0) acts as 0; and 1 acts as the identity
operator. Then v acts on S by means of the following operators:
c 7 1,
L(n) 7

(2.7)

1X
h(j)h(n j) for n 6= 0,
2 jZ

(2.8)

1X
h(|j|)h(|j|).
2 jZ

(2.9)

L(0) 7

It is important to observe that in the case of L(0), the absolute values

make the formal quadratic operator well defined, while for n 6= 0 the operator
is well defined as it stands, since [h(j), h(nj)] = 0. Thus the operators (2.8)
and (2.9) are in normal-ordered form, that is, the annihilation operators
h(n) for n > 0 act to the right of the creation operators h(n) for n < 0.
Using colons to denote normal ordering (the appropriate reordering of the
factors in the product if necessary), we thus have
L(n) 7

1 X
h(j)h(n j)
2 jZ
4

(2.10)

for all n Z.
It is an instructive and not-so-trivial (classical) exercise to verify by direct
computation that the operators (2.10) indeed satisfy the bracket relations
(2.4). (This exercise and the related constructions are presented in [FLM],
for example, where the standard generalization of this construction of v using
a Heisenberg algebra based on a finite-dimensional space of operators h(n)
for each n is also carried out.)
Vertex operator algebra theory and conformal field theory place this exercise into a very general, natural setting (among many other things), with
conceptual approaches and techniques (cf. [FLM]). It is standard procedure
to embed operators such as h(n) and L(n) into generating functions and to
compute with these generating functions, using a formal calculus, and to systematically avoid computing with the individual operators. Doing this vastly
simplifies computations that would otherwise be complicated or sometimes
almost impossible. In fact, we shall be using a number of generating-function
ideas below.
The space S carries a natural Z-grading, determined by the rule deg h(j) =
j for j < 0. Then S is in fact graded by the nonpositive integers, and the
v-module S is a graded module. It turns out to be appropriate to use the
negative of this grading, that is, to define a new grading (by conformal
weights) on the space S by the rule wt h(j) = j for j > 0. One reason
why this is natural is that for each n 0, the homogeneous subspace of S of
weight n coincides with the eigenspace of the operator L(0) with eigenvalue
n, as is easy to see. For n Z (or n 0) we define Sn to be the homogeneous
subspace of S of weight n, and we consider the formal power series in the
formal variable q given by
dim S =

(dim Sn )q n

(2.11)

n0

(the graded dimension of the graded space S). Clearly, from the definitions,
dim S =

(1 q n )1 .

(2.12)

n>0

Here are the main points about these classical considerations that we
want to emphasize: As is well known in conformal field theory, removing the
normal ordering in the definition of the operator L(0) introduces an infinity
5

which formally equals 12 (1): The unrigorous expression

1X

h(j)h(j)
L(0)
=
2 jZ

(2.13)

formally equals (by the bracket relations (2.5))

1
L(0) + (1 + 2 + 3 + ),
2

(2.14)

which itself formally equals

1
1
L(0) + (1) = L(0) .
2
24

(2.15)

We rigorize L(0)
by defining it as:

and we define

L(0)
= L(0) + (1),
2

(2.16)

L(n)
= L(n) for n 6= 0,

(2.17)

to get a new basis of the Lie algebra v. (We are identifying the elements of
v with operators on the space S.) The brackets become:

[L(m),
L(n)]
= (m n)L(m
+ n) +

1 3
m m+n,0 ;
12

(2.18)

that is, m3 m in (2.4) has become the pure monomial m3 .

Perhaps the most important (and also well-known) thing accomplished
by this formal removal of the normal ordering is the following: We define a
new grading of the space S by using the eigenvalues of the modified operator

L(0)
in place of L(0), so that the grading of S is shifted from the previous
1
grading by conformal weights by the subtraction of 24
from the weights. We
let (S) be the corresponding graded dimension, so that
(S) =
where

(q) = q 24

1
,
(q)

(1 q n ).

n>0

(2.19)

(2.20)

The point is that (q) has important (classical) modular transformation propQ
erties, unlike n>0 (1 q n ), when viewed as a function of in the upper
half-plane via the substitution q = e2i ; (q) is Dedekinds eta-function.
Bloch [Bl] extended this classical story in various ways, in particular, the
following: Instead of the Lie algebra d, we consider the larger Lie algbebra
of formal differential operators, spanned by
{tn D m |n Z, m 0}

(2.21)

or more precisely, we restrict to m > 0 and further, to the Lie subalgebra D +

spanned by the differential operators of the form D r (tn D)D r for r 0, n Z.
Then we can construct a central extension of D + using generalizations of the
normally-ordered quadratic operators (2.10) above:
L(r) (n) =

1X r
j h(j)(n j)r h(n j) for n 6= 0,
2 jZ

(2.22)

1X
(j)r h(|j|)j r h(|j|),
2 jZ

(2.23)

1X r
j h(j)(n j)r h(n j)
2 jZ

(2.24)

L(r) (0) =
that is,
L(r) (n) =

for n Z. Direct computation of the commutators among these operators

[Bl] shows that they provide a central extension of D + such that
L(r) (n) 7 (1)r+1 D r (tn D)D r

(2.25)

(cf. [KP]). (It is not surprising in retrospect that these operators L(r) (n)
are related to differential operators, because the generating function of these
operators as n ranges through Z is based on D r , as we discuss below.)
A central point of [Bl] is that the formal removal of the normal-ordering
procedure in the definition (2.23) of L(r) (0) adds the infinity (1)r 12 (2r
P
1) = n>0 n2r+1 (generalizing (2.13)(2.15)), and if we correspondingly
define
(r) (0) = L(r) (0) + (1)r 1 (2r 1)
L
(2.26)
2
(r) (n) = L(r) (n) for n 6= 0 (generalizing (2.16) and (2.17)), the commuand L
tators simplify in a remarkable way: As direct computation [Bl] shows, the
7

 (r) (m), L
(s) (m)] reduces to
complicated polynomial in the scalar term of [L
a pure monomial in m, by analogy with, and generalizing, the passage from
m3 m to m3 in (2.18). The precise formulas can be found in [Bl], along
with further results; for instance, in [Bl], these considerations and results are
generalized to Dirichlet L-series in place of the zeta function.

First explanation and generalization

Our goal is to present two layers of explanation and generalization of the

results of [Bl] sketched above. First we need some elementary formal background:
What does (2r 1) mean, for a nonnegative integer r?
It is a well-known classical fact that for k > 1,
(k + 1) =

Bk
,
k

(3.1)

where the Bk are the Bernoulli numbers, defined by the generating function
ex

X Bk
x
=
xk ,
1 k0 k!

(3.2)

where x is a formal variable. This formal power series in x is understood to

be computed (on the left-hand side) by expanding ex 1 as the formal series
2
x + x2! + and performing the division of formal power series to obtain
a formal power series with constant term 1; this of course corresponds to
expanding a complex function in a certain domain, but we are operating
purely formally.
Why does Bk defined in this way have anything to do with the formal
P
series n>0 nk1 ? We recall Eulers heuristic interpretation of such formal
sums as 1 + 2 + (cf. the Preface of [Hi]); actually, we give a variant of
Eulers interpretation adapted to the main theme that we shall introduce:
Consider the expansion (from (3.2))
X Bk
1
=
xk1 .
x
1e
k!
k0

(3.3)

Expand the left-hand side unrigorously as the formal geometric series

1 + ex + e2x + = 1 +

1k k X 2k k
x +
x + .
k0 k!
k0 k!
X

(3.4)

For k > 1, the coefficient of xk1 in this formal expression is

1
(1k1 + 2k1 + ),
(k 1)!

(3.5)

1
(k +1). Also, the coefficient of x0 in (3.4) is formally
which looks like (k1)!
1 + 0!1 (10 + 20 + ), which we formally view as 1 + (0) (and not as (0)).
Thus, formally equating the coefficients of xl for l 0 in (3.2) explains
(3.1) and the fact that (0) = B1 1 (= 12 ); now we know what (1.4)
says.
The key point here is the interplay between the formal geometric series
expansion (in powers of ex ) and the expansion in powers of x.
Now, how do we interpret all of this via vertex operator algebra theory?
First note that the expressions (2.22)(2.24) above for L(r) (n) suggest
th
r derivatives. We have already mentioned that a basic theme in vertex
operator algebra theory is to always use appropriate generating functions (as
we just did, incidentally, in the heuristic discussion above). First we put our
individual operators into generating functions. Using a formal variable x, we
define
X
h(x) =
h(n)xn
(3.6)
nZ

and
L(r) (x) =

L(r) (n)xn ,

(3.7)

nZ
d
and using Dx to denote the operator x dx
(recall the comment after (2.2)),
we observe that
1
(3.8)
L(r) (x) = (Dxr h(x))2 ,
2
where the colons, as always, denote normal ordering (recall (2.10)). (For
other purposes, other versions of these generating functions are used, in parP
ticular, h(x) = nZ h(n)xn1 , as in (4.14) below, in place of (3.6), but we
have chosen the appropriate generating functions for our purposes.)

Using standard elementary techniques, we could directly compute the

brackets [L(r) (x1 ), L(s) (x2 )] of these generating functions, for r, s 0, where
x1 and x2 are independent commuting formal variables. (As always in vertex
operator algebra theory or conformal field theory, when we consider such
operations as brackets of generating functions, we need independent commuting formal variables; the expression [L(r) (x), L(s) (x)], with the variable
x repeated, would be meaningless.) But this computation, which might be
carried out as a more complicated variant of the argument on pp. 224226
of [FLM], for example, would not be simple. It would of course recover the
information of the brackets [L(r) (m), L(s) (n)] computed in [Bl].
The best use of generating functions in this context is instead to also introduce suitable generating functions over the number of derivatives. Consider
the elementary formal Taylor theorem equating the application of a formal
d
exponential of a formal multiple of dx
with a formal substitution operation:
d

ey dx f (x) = f (x + y),

(3.9)

where f (x) is an arbitrary formal series of the form n an xn , and where it

is understood that each binomial expression (x + y)n is to be expanded in
nonnegative integral powers of y. Here n is allowed to range over something
very general, like Z or even C, say, and the an lie in a fixed vector space;
f (x) is very definitely not necessarily the expansion of an analytic function.
Formula (3.9) is proved by direct formal expansion of both sides (cf. [FLM],
Proposition 8.3.1; Taylors theorem in this kind of generality is heavily exd
ploited in Chapter 8 of [FLM], for instance). Now dx
is of course a formal
infinitesimal translation (as (3.9) states), but for our present purposes we
want the following formal multiplicative analogue of (3.9):
P

eyDx f (x) = f (ey x),

(3.10)

with f (x) as above (again cf. [FLM], Proposition 8.3.1), which expresses the
fact that Dx is a formal infinitesimal dilation.
Now (Dxr h(x))2 (recall (3.8)) is hard to put into a good generating
function over r, but we can make the problem easier by making it more
general: Consider independently many derivatives on each of the two factors
h(x) in h(x)2 , use two new independent formal variables y1 and y2 , and
form the generating function
1
1
L(y1 ,y2 ) (x) = (ey1 Dx h(x))(ey2 Dx h(x)) = h(ey1 x)h(ey2 x)
(3.11)
2
2
10

(where we use (3.10)), so that L(r) (x) is a diagonal piece of this generating
function in the sense that it is (r!)2 times the coefficient of y1r y2r in L(y1 ,y2 ) (x).
Using formal vertex operator calculus techniques (generalizing the argument
on pp. 224-226 of [FLM], for example), we can calculate
[ h(ey1 x1 )h(ey2 x1 ) , h(ey3 x2 )h(ey4 x2 ) ].

(3.12)

Then, a nontrivial, and in fact quite tricky, vertex operator computation

recovers Blochs formulas, in somewhat generalized form, as we explain next.
Here are the main points:
The expression h(x)2 is not rigorous (as we observe for instance by trying
to compute the coefficient of any fixed power of x in this expression), while
the normal-ordered expression h(x)2 is of course rigorous. The expression
h(ey1 x)h(ey2 x) is still not rigorous (even though the expressions ey1 x and
ey2 x are distinct), as we see by (for example) trying to compute the constant
term in the variables y1 and y2 in this expression. The lack of rigor in this
expression in fact corresponds exactly to the occurrence of formal sums like
P
r
n>0 n with r > 0, as we have been discussing above.
However, we have
h(x1 )h(x2 ) = h(x1 )h(x2 ) + x2

x2 1 x2 /x1

(3.13)

(an exercise using elementary vertex operator techniques), and it follows that
h(ey1 x1 )h(ey2 x2 ) = h(ey1 x1 )h(ey2 x2 ) + x2

1
.
y
x2 1 e 2 x2 /ey1 x1

(3.14)

Note that x2 x 2 can be replaced by y 1 in the last expression (and this is

one illustration of the naturalness of our emphasis on the homogeneous

differential operator Dx = x x
rather than x
). The expression 1ey2 x12 /ey1 x1
came from, and is, a geometric series expansion (recall (3.13)).
Now we try to set x1 = x2 (= x) in (3.14). The result of this procedure is unrigorous on the left-hand side, as we have pointed out above, but
the result has rigorous meaning on the right-hand side, because the normalordered product h(ey1 x)h(ey2 x) is certainly well defined, and the expression
y 1 1ey1 1 +y2 can be interpreted rigorously as in (3.2) and (3.3); more precisely (the role of x in (3.3) being played here by y1 + y2 ), we take 1ey1 1 +y2
11

to mean the formal (Laurent) series in y1 and y2 of the shape

1
= (y1 y2 )1 F (y1 , y2),
1 ey1 +y2

(3.15)

where (y1 y2 )1 is understood as the binomial expansion (geometric series)

in nonnegative powers of y2 and F (y1 , y2) is an (obvious) formal power series
in (nonnegative powers of) y1 and y2 . This motivates us to define a new
normal-ordering procedure
+
+

h(ey1 x)h(ey2 x) ++ = h(ey1 x)h(ey2 x)

1
,
y1 1 ey1 +y2

(3.16)

with the last part of the right-hand side being understood as we just indicated. Again compare this with the heuristic discussion above; this expression came from a geometric series, but it becomes rigorous only when we
expand in the new way (actually, we might alternatively replace the binomial
expansion (y1 y2 )1 by the different expansion of the same formal expression in nonnegative powers of y1 rather than of y2 , but it is more natural to
make the choice that we did).
Formula (3.16) and its indicated interpretation give a natural explanation of the zeta-function-modified operators defined in (2.26): We use (3.16)
to define the following analogues of the operators (3.11):
(y1 ,y2 ) (x) = 1 + h(ey1 x)h(ey2 x) + ,
L
+
2+

(3.17)

(r) (n) is exactly (r!)2 times the

and it is easy to check that the operator L
coefficient of y1r y2r xn
in (3.17); the significant case is the case n = 0.
0
We are now ready to formulate the result mentioned just after (3.12)
above. With the new normal ordering (3.16) replacing the old one, remarkable cancellation occurs in the commutator (3.12), and the result is:
Theorem 3.1 With the formal delta-function Laurent series (x) defined as
(x) =

xn ,

(3.18)

nZ

and with independent commuting formal variables as indicated, we have:

(y1 ,y2 ) (x1 ), L
(y3 ,y4 ) (x2 )]
[L
12

1 (y1 +y2 +y3 ,y4 )

ey1 x1
=
L
(x2 ) y3
2 y1
e x2
 y1 
e
x
1
(y1 +y2 +y4 ,y3 ) (x2 )
+L
ey4 x2

 y2 
1 (y1 y2 +y3 ,y4 )
e x1

L
(x2 ) y3
2 y2
e x2
 y2 
e
x
1
(y1 y2 +y4 ,y3 ) (x2 )
.
+L
ey4 x2


(3.19)

Actually, hidden in the right-hand side of (3.19) are formal expressions

(suitably expanded) of the type (y1 y2 y3 + y4 )3 because of the formal
pole y1 = y2 in (3.16), and these expressions, multiplied by the formal deltafunction expressions, are the source of the pure monomials in m that we set
out to explain (recall the end of Section 2). Indeed, the delta-function expression (ey1 x1 /ey3 x2 ), for instance, can be written as ey1 Dx1 ey3 Dx2 (x1 /x2 ), and
when we extract and equate the coefficients of the monomials in the variables
y1r y2r y3s y4s on the two sides of (3.19), we get expressions like (D j ) (x1 /x2 ),
whose expansion, in turn, in powers of x1 and x2 clearly yields a pure monomial analogous to and generalizing the expression m3 in (2.18). In this way,
we have an explicit generalization and explanation of Blochs formula for
(r) (m), L
(s) (n)] in terms of the operators L
(j) (m + n) and a monomial in
[L
m.
We have been working all along with a Heisenberg algebra based on a
one-dimensional spacethat is, a Heisenberg algebra with only one dimension of operators, spanned by the element h(n), for each n, but all of these
considerations hold equally well in the more general situation where we start
with a Heisenberg algebra based on a finite-dimensional space.
All of this has been an interesting (and nontrivial) special case of vertex
operator algebra theory, but what is really happening?

Second explanation and generalization

It turns out that Theorem 3.1 is an extremely special case of a something

that can be formulated and proved for an arbitrary vertex operator algebra
(and indeed this gives another motivation for the general theory). We recall
the definition of the notion of vertex (operator) algebra from [Bo], [FLM] and
13

[FHL]; the principles that we have found are based heavily on the Jacobi
identity as formulated in [FLM] and [FHL]:
Definition 4.1 A vertex operator algebra (V, Y, 1, ), or simply V (over C),
is a Z-graded vector space (graded by weights)
V =

V(n) ; for v V(n) , n = wt v;

(4.1)

nZ

such that
dim V(n) < for n Z,

(4.2)

V(n) = 0 for n sufficiently small,

(4.3)

equipped with a linear map V V V [[x, x1 ]], or equivalently,

V (End V )[[x, x1 ]]
X
v 7 Y (v, x) =
vn xn1 (where vn End V ),

(4.4)

nZ

Y (v, x) denoting the vertex operator associated with v, and equipped also with
two distinguished homogeneous vectors 1 V(0) (the vacuum) and V(2) .
The following conditions are assumed for u, v V : the lower truncation
condition holds:
un v = 0 for n sufficiently large
(4.5)
(or equivalently, Y (u, x)v involves only finitely many negative powers of x);
Y (1, x) = 1 (1 on the right being the identity operator);

(4.6)

the creation property holds:

Y (v, x)1 V [[x]] and lim Y (v, x)1 = v
x0

(4.7)

(that is, Y (v, x)1 involves only nonnegative integral powers of x and the
constant term is v); with (x) as in (3.18) and with binomial expressions
understood (as above) to be expanded in nonnegative powers of the second
variable, the Jacobi identity (the main axiom) holds:
x1
0

x2 x1
x1 x2
Y (u, x1 )Y (v, x2 ) x1
Y (v, x2 )Y (u, x1 )
0
x0
x0


x1 x0
= x1
Y (Y (u, x0 )v, x2 )
(4.8)
2
x2


14

(note that when each expression in (4.8) is applied to any element of V , the
coefficient of each monomial in the formal variables is a finite sum; on the
right-hand side, the notation Y (, x2 ) is understood to be extended in the
obvious way to V [[x0 , x1
0 ]]); the Virasoro algebra relations hold (acting on
V ):
[L(m), L(n)] = (m n)L(m + n) +

1
(m3 m)n+m,0 (rank V )1
12

(4.9)

for m, n Z, where
L(n) = n+1 for n Z, i.e., Y (, x) =

L(n)xn2

(4.10)

nZ

and
rank V C;
L(0)v = nv = (wt v)v for n Z and v V(n) ;
d
Y (v, x) = Y (L(1)v, x)
dx

(4.11)
(4.12)
(4.13)

(the L(1)-derivative property).

Note that as in Theorem 3.1, the Jacobi identity encapsulates higher
derivatives of delta-function expressions, since the expression ((x1 x2 )/x0 ),
say, can be expanded by means of (3.9). The use of the three formal variables,
rather than complex variables (which would require changes in the formulas),
allows the full symmetry of the Jacobi identity to reveal itself (see [FLM] and
[FHL]).
The polynomial algebra S (recall (2.6)) carries a canonical vertex operator
algebra structure of rank 1 with vacuum vector 1 equal to 1 S and with the
operators L(n) agreeing with the operators defined in Section 2 (cf. [FLM]).
We will not describe the vertex operators Y (v, x) here, except to say that
Y (h(1), x) = x1 h(x) =

h(n)xn1

(4.14)

nZ

(recall (3.6)) and that the element is 12 (h(1))2 S.

If we omit the grading and the Virasoro algebra from Definition 4.1 and
adjust the axioms appropriately, we have the notion of vertex algebra as
15

introduced in [Bo], except that Borcherds used certain special cases of the
Jacobi identity instead of (4.8). The identity (4.8) is the canonical maximal
axiom: It contains the full necessary information in compact form; it is
analogous to the classical Jacobi identity in the definition of the notion of
Lie algebra; and it is invariant in a natural sense under the symmetric group
on three letters (see [FLM] and [FHL]).
There are also minimal axioms, stemming from the fact that the (suitably formulated) commutativity of the operators Y (u, x1 ) and Y (v, x2 )
implies associativity (again suitably formulated) and hence the Jacobi
identity (see [FLM] and [FHL]; cf. [BPZ] and [G]). The simplest minimal axiom, as found in [DL] (actually, in the greater generality of abelian
intertwining algebras) states that for u, v V , there exists n 0 such that
(x1 x2 )n [Y (u, x1 ), Y (v, x2 )] = 0

(4.15)

(see [DL], formula (1.4)). However, it is still a nontrivial matter to construct

examples, even relatively simple ones like S, of vertex operator algebras,
partly because one has to extend the condition (4.15) from generators to
arbitrary elements of V . A general and systematic approach and solution to
this and related problems was first found by Li (see [Li1], [Li2]).
The commutativity condition asserts that for u, v V ,
Y (u, x1 )Y (v, x2 ) Y (v, x2 )Y (u, x1 ),

(4.16)

where denotes equality up to a suitable kind of generalized analytic

continuation, and the associativity condition asserts that
Y (u, x1 )Y (v, x2 ) Y (Y (u, x1 x2 )v, x2 ),

(4.17)

where the right-hand side and the generalized analytic continuation have to
be understood in suitable ways (again see [FLM] and [FHL] and cf. [BPZ]
and [G]); the right-hand side of (4.17) is not a well-defined formal series in
x1 and x2 .
On the level of these basic principles, for any vertex operator algebra V
we shall now conceptually formulate and considerably generalize the normalordering procedure (3.16) and we shall formulate a new general Jacobi identity which implies Theorem 3.1 in the very particular case of the vertex
operator algebra S and very special vertex operators. Ideas in Zhus work
[Z1], [Z2] enter into our considerations.
16

Formally replacing x1 by ey x2 in (4.17), we find (formally and unrigorously) that

Y (u, ey x2 )Y (v, x2 ) Y (Y (u, (ey 1)x2 )v, x2 ).

(4.18)

Now we observe that while the left-hand side of (4.18) is not a well-defined
formal series in the formal variables y and x2 , the right-hand side of (4.18) is
in fact a well-defined formal series in these formal variables. By replacing x1
by ey x2 we have made the right-hand side of (4.18) rigorous (and the left-hand
side unrigorous). This situation should be compared with our motivation for
introducing the normal-ordering procedure (3.16) above.
Next, instead of the vertex operators Y (v, x), we want the modified vertex
operators defined for homogeneous elements v V by:
X(v, x) = xwt v Y (v, x) = Y (xL(0) v, x),

(4.19)

as in [FLM], formula (8.5.27) (recall that L(0)-eigenvalues define the grading

of V ); the formula X(v, x) = Y (xL(0) v, x) works for all v V (not necessarily
homogeneous). For instance, the operator h(x) (3.6) is exactly X(h(1), x)
(cf. (4.14)).
What we really want is a rigorous expression that equals the unrigorous
expression X(u, ey x2 )X(v, x2 ); this will considerably generalize our interpretation of the unrigorous expression h(ey x2 )h(x2 ) (cf. (3.14) and (3.16)). So
L(0)
we replace u by (ey x2 )L(0) u and v by x2 v in (4.18) and using basic techniques we get:
X(u, ey x2 )X(v, x2 ) X(Y (eyL(0) u, ey 1)v, x2 ),

(4.20)

and this right-hand side is still rigorous (and the left-hand side still unrigorous). But Y (eyL(0) u, ey 1) is exactly Zhus operator Y [u, y] in [Z1], [Z2], so
that
X(u, ey x2 )X(v, x2 ) X(Y [u, y]v, x2).
(4.21)
By Zhus change-of-variables theorem, x 7 Y [u, x] defines a new vertex
operator algebra structure on the same vector space V under suitable conditions; this theorem was a step in Zhus vertex-operator-algebraic proof of
the modular-invariance properties of characters (cf. the comments surrounding (2.19) above). There have been two subsequent treatments of this
change-of-variables theorem, in [L1] and in [H1], [H2]; in the latter works,
17

Huang considerably generalized this result (and removed a hypothesis of

Zhus) using his geometric analysis of the Virasoro algebra structure in a
vertex operator algebra. The formal relation (4.21) generalizes to products
of several operators.
So we want to bracket the (rigorous) expressions X(Y [u, y]v, x), which
are the correct natural generalization of the expression (3.16) above (at least
with y2 = 1 in (3.16), but this restriction is a minor issue since y2 can easily be
restored). Keep in mind that the formal relation (4.21) naturally implements
the formal relation (1.4) in a foundational way from the viewpoint of vertex
operator algebras.
But just as in [FLM], where the Jacobi identity for vertex operator algebras was the most natural general formula, here we find that the most
natural thing to do is to formulate and prove a new Jacobi identity, based
on the operators X(Y [u, y]v, x), in the general setting of an arbitrary vertex
(operator) algebra, rather than just to bracket the operators. It turns out
that delta-function expressions of the type appearing on the right-hand side
of (3.19), and that in turn explained the phenomenon of pure monomials
in m (as discussed above), arise naturally in complete generality, and when
we ask for a Jacobi identity rather than just a commutator formula in general, we find that delta-function expressions of this same type appear on the
left-hand side as well as the right-hand side. This is another instance of the
naturalness of Jacobi identities, which have symmetries that commutator
formulas do not have. We state our result for the operators X(v, x) rather
than X(Y [u, y]v, x) (i.e., the case where u is the vacuum vector):
Theorem 4.2 In any vertex (operator) algebra V , for u, v V we have:


y21
x1
0 e

x1
y12 x2
X(u, x1)X(v, x2 ) x1
X(v, x2 )X(u, x1)
0 e
x0
x0


1
y01 x1
= x2 e
X(Y [u, y01 ]v, x2 ),
(4.22)
x2


where
y21

x1
x0
x2
, y12 = log 1
, y01 = log 1
.
= log 1
x1
x2
x1


(4.23)

If we want the commutator [X(u, x1), X(v, x2 )], we simply extract the
coefficient of x1
0 (the formal residue in the variable x0 ) on both sides, and it
18

turns out that the resulting right-hand side can be put into an elegant form.
If we replace u and v by expressions of the shape Y [u, y]v, we obtain naturally
a formula that generalizes formula (3.19) (Theorem 3.1) to arbitrary elements
of arbitrary vertex (operator) algebras. That is, interesting as they are, the
phenomena that we have been discussing concerning central extensions of
Lie algebras of differential operators form extremely special cases of general
vertex-operator-algebraic phenomena. The detailed formulations and proofs,
and generalizations, are found in [L2]. Also, my student Antun Milas has
generalized some of these results in a number of directions.

References
[BPZ]

A. A. Belavin, A. M. Polyakov and A. B. Zamolodchikov, Infinite conformal symmetries in two-dimensional quantum field theory, Nucl. Phys. B241 (1984), 333380.

[Bl]

S. Bloch, Zeta values and differential operators on the circle, J.

Algebra 182 (1996), 476500.

[Bo]

R. E. Borcherds, Vertex algebras, Kac-Moody algebras, and the

Monster, Proc. Natl. Acad. Sci. USA 83 (1986), 30683071.

[DKM]

E. Date, M. Kashiwara and T. Miwa, Vertex operators and

functionstransformation groups for soliton equations II, Proc.
Japan Acad. Ser. A Math. Sci. 57 (1981), 387392.

[DL]

C. Dong and J. Lepowsky, Generalized Vertex Algebras and Relative

Vertex Operators, Progress in Math., Vol. 112, Birkhauser, Boston,
1993.

[FKRW] E. Frenkel, V. Kac, A. Radul and W. Wang, W1+ and W (glN )

with central charge N, Comm. Math. Physics 170 (1995), 337357.
[FHL]

I. B. Frenkel, Y.-Z. Huang and J. Lepowsky, On axiomatic approaches to vertex operator algebras and modules, preprint, 1989;
Memoirs Amer. Math. Soc. 104, 1993.

19

[FLM]

I. B. Frenkel, J. Lepowsky and A. Meurman, Vertex Operator Algebras and the Monster, Pure and Appl. Math., Vol. 134, Academic
Press, Boston, 1988.

[G]

P. Goddard, Meromorphic conformal field theory, Infinite Dimensional Lie Algebras and Groups, Advanced Series in Math. Physics,
Vol. 7, ed. V. Kac, World Scientific, Singapore, 1989, 556587.

[Hi]

H. Hida, Elementary Theory of L-functions and Eisenstein Series,

London Math. Soc. Student Texts, Vol. 26, Cambridge University
Press, Cambridge, 1993.

[H1]

Y.-Z. Huang, Applications of the geometric interpretation of vertex

operator algebras, Proc. 20th International Conference on Differential Geometric Methods in Theoretical Physics, New York, 1991, ed.
S. Catto and A. Rocha, World Scientific, Singapore, 1992, 333343.

[H2]

Y.-Z. Huang, Two-dimensional Conformal Geometry and Vertex

Operator Algebras, Progress in Math., Vol. 148, Birkhauser, Boston,
1997.

[KP]

V. Kac and D. Peterson, Spin and wedge representations of infinitedimensional Lie algebras and groups, Proc. Natl. Acad. Sci. USA
78 (1981), 33083312.

[KR]

V. Kac and A. Radul, Quasifinite highest weight modules over the

Lie algebra of differential operators on the circle, Comm. Math.
Physics 157 (1993), 429457.

[L1]

J. Lepowsky, Remarks on vertex operator algebras and moonshine, Proc. 20th International Conference on Differential Geometric Methods in Theoretical Physics, New York, 1991, ed. S. Catto
and A. Rocha, World Scientific, Singapore, 1992, 362370.

[L2]

J. Lepowsky, A Jacobi identity for vertex operator algebras related to zeta-function values, to appear.

[Li1]

H. Li, Local systems of vertex operators, vertex superalgebras and

modules, preprint, 1993; J. Pure Appl. Alg. 109 (1996), 143195.
20

[Li2]

H. Li, Local systems of twisted vertex operators, vertex superalgebras and twisted modules, Contemporary Math. 193 (1996), 203
236.

[M]

Y. Matsuo, Free fields and quasi-finite representations of W1+ ,

Physics Lett. B 326 (1994), 95100.

[Z1]

Y. Zhu, Vertex operators, elliptic functions and modular forms,

Ph.D. thesis, Yale University, 1990.

[Z2]

Y. Zhu, Modular invariance of characters of vertex operator algebras, J. Amer. Math. Soc. 9 (1996), 237307.

Department of Mathematics, Rutgers University, Piscataway, NJ

08854

E-mail address: [email protected]

21