a

r
X
i
v
:
h
e
p
-
t
h
/
9
4
1
1
0
2
8

v
1



3

N
o
v

1
9
9
4
NSF-ITP-94-97
hep-th/9411028
WHAT IS STRING THEORY?
Joseph Polchinski
1
Institute for Theoretical Physics
University of California
Santa Barbara, CA 93106-4030
ABSTRACT
Lectures presented at the 1994 Les Houches Summer School Fluctuating Geome-
tries in Statistical Mechanics and Field Theory. The rst part is an introduction
to conformal eld theory and string perturbation theory. The second part deals
with the search for a deeper answer to the question posed in the title.
1
Electronic address: [email protected]
Contents
1 Conformal Field Theory 5
1.1 The Operator Product Expansion . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2 Ward Identities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3 Conformal Invariance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.4 Mode Expansions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.5 States and Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.6 Other CFTs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.7 Other Algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
1.8 Riemann Surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
1.9 CFT on Riemann Surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2 String Theory 44
2.1 Why Strings? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.2 String Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.3 The Spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
2.4 The Weyl Anomaly . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
2.5 BRST Quantization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
2.6 Generalizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
2.7 Interactions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
2.8 Trees and Loops . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3 Vacua and Dualities 79
3.1 CFTs and Vacua . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
3.2 Compactication on a Circle . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
2
3.3 More on R-Duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
3.4 N = 0 in N = 1 in . . .? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
3.5 S-Duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
4 String Field Theory or Not String Field Theory 98
4.1 String Field Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
4.2 Not String Field Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.3 High Energy and Temperature . . . . . . . . . . . . . . . . . . . . . . . . . . 109
5 Matrix Models 113
5.1 D = 2 String Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
5.2 The D = 1 Matrix Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
5.3 Matrix Model String . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
5.4 General Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
5.5 Tree-Level Scattering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
5.6 Spacetime Gravity in the D = 2 String . . . . . . . . . . . . . . . . . . . . . 130
5.7 Spacetime Gravity in the Matrix Model . . . . . . . . . . . . . . . . . . . . . 134
5.8 Strong Nonlinearities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
5.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
3
While I was planning these lectures I happened to reread Ken Wilsons account of his
early work[1], and was struck by the parallel between string theory today and quantum
eld theory thirty years ago. Then, as now, one had a good technical control over the
perturbation theory but little else. Wilson saw himself as asking the question What is
quantum eld theory? I found it enjoyable and inspiring to read about the various models
he studied and approximations he tried (he refers to clutching at straws) before he found
the simple and powerful answer, that the theory is to be organized scale-by-scale rather than
graph-by-graph. That understanding made it possible to answer both problems of principle,
such as how quantum eld theory is to be dened beyond perturbation theory, and practical
problems, such as how to determine the ground states and phases of quantum eld theories.
In string theory today we have these same kinds of problems, and I think there is good
reason to expect that an equally powerful organizing principle remains to be found. There
are many reasons, as I will touch upon later, to believe that string theory is the correct
unication of gravity, quantum mechanics, and particle physics. It is implicit, then, that the
theory actually exists, and exists does not mean just perturbation theory. The nature of
the organizing principle is at this point quite open, and may be very dierent from what we
are used to in quantum eld theory.
One can ask whether the situation today in string theory is really as favorable as it was
for eld theory in the early 60s. It is dicult to know. Then, of course, we had many
more experiments to tell us how quantum eld theories actually behave. To oset that,
we have today more experience and greater mathematical sophistication. As an optimist,
I make an encouraging interpretation of the history, that many of the key advances in
eld theoryWilsons renormalization group, the discovery of spontaneously broken gauge
symmetry as the theory of the electroweak interaction, the discovery of general relativity
itselfwere carried out largely by study of simple model systems and limiting behaviors,
and by considerations of internal consistency. These same tools are available in string theory
today.
My lectures divide into two partsan introduction to string theory as we now under-
stand it, and a look at attempts to go further. For the introduction, I obviously cannot in
ve lectures cover the whole of superstring theory. Given time limitations, and given the
broad range of interests among the students, I will try to focus on general principles. I
will begin with conformal eld theory (2.5 lectures), which of course has condensed mat-
ter applications as well as being the central tool in string theory. Section 2 (2.5 lectures)
introduces string theory itself. Section 3 (1 lecture), on dualities and equivalences, covers
4
the steadily increasing evidence that what appear to be dierent string theories are in many
cases dierent ground states of a single theory. Section 4 (1 lecture) addresses the question
of whether string eld theory is the organizing principle we seek. In section 5 (2 lectures)
I discuss matrix models, exactly solvable string theories in low spacetime dimensions.
I should emphasize that this is a survey of many subjects rather than a review of any
single subject (for example R-duality, on which I spend half a lecture, was the subject of a
recent review [2] with nearly 300 references). I made an eort to choose references which
will be useful to the studenta combination of reviews, some original references, and some
interesting recent papers.
1 Conformal Field Theory
Much of the material in this lecture, especially the rst part, is standard and can be found
in many reviews. The 1988 Les Houches lectures by Ginsparg [3] and Cardy [4] focus on
conformal eld theory, the latter with emphasis on applications in statistical mechanics.
Introductions to string theory with emphasis on conformal eld theory can be found in
refs. [5]-[9]. There are a number of recent books on string theory, though often with less
emphasis on conformal techniques [10]-[14] as well as a book [15] and reprint collection [16]
on conformal eld theory and statistical mechanics. Those who are in no great hurry will
eventually nd an expanded version of these lectures in ref. [17]. Finally I should mention
the seminal papers [18] and [19].
1.1 The Operator Product Expansion
The operator product expansion (OPE) plays a central role in this subject. I will introduce
it using the example of a free scalar eld in two dimensions, X(
1
,
2
). I will focus on two
dimensions because this is the case that will be of interest for the string, and I will refer to
these two dimensions as space though later they will be the string world-sheet and space
will be something else. The action is
S =
1
8
_
d
2

_
(
1
X)
2
+ (
2
X)
2
_
. (1.1.1)
The normalization of the eld X (and so the action) is for later convenience. To be specic
I have taken two Euclidean dimensions, but almost everything, at least until we get to
5
nontrivial topologies, can be continued immediately to the Minkowski case
1

2
i
0
.
Expectation values are dened by the functional integral
< T[X] > =
_
[dX] e
S
T[X], (1.1.2)
where T[X] is any functional of X, such as a product of local operators.
2
It is very convenient to adopt complex coordinates
z =
1
+i
2
, z =
1
i
2
. (1.1.3)
Dene also

z
=
1
2
(
1
i
2
),
z
=
1
2
(
1
+i
2
). (1.1.4)
These have the properties
z
z = 1,
z
z = 0, and so on. Note also that d
2
z = 2d
1
d
2
from
the Jacobian, and that
_
d
2
z
2
(z, z) = 1. I will further abbreviate
z
to and
z
to

when
this will not be ambiguous. For a general vector, dene as above
v
z
= v
1
+iv
2
, v
z
= v
1
iv
2
, v
z
=
1
2
(v
1
iv
2
), v
z
=
1
2
(v
1
+iv
2
). (1.1.5)
For the indices 1, 2 the metric is the identity and we do not distinguish between upper and
lower, while the complex indices are raised and lowered with
3
g
z z
= g
zz
=
1
2
, g
zz
= g
z z
= 0, g
z z
= g
zz
= 2, g
zz
= g
z z
= 0 (1.1.6)
The action is then
S =
1
4
_
d
2
z X

X, (1.1.7)
and the equation of motion is

X(z, z) = 0. (1.1.8)
The notation X(z, z) may seem redundant, since the value of z determines the value of z,
but it is useful to reserve the notation f(z) for elds whose equation of motion makes them
1
Both the Euclidean and Minkowski cases should be familiar to the condensed matter audience. The
former would be relevant to classical critical phenomena in two dimensions, and the latter to quantum
critical phenomena in one dimension.
2
Notice that this has not been normalized by dividing by < 1 >.
3
A comment on notation: being careful to keep the Jacobian, one has d
2
z = 2d
1
d
2
and d
2
z
_
| det g| =
d
1
d
2
. However, in the literature one very frequently nds d
2
z used to mean d
1
d
2
.
6
analytic in z. For example, it follows at once from the equation of motion (1.1.8) that X is
analytic and that

X is antianalytic (analytic in z), hence the notations X(z) and

X( z).
Notice that under the Minkowski continuation, an analytic eld becomes left-moving, a
function only of
0
+
1
, while an antianalytic eld becomes right-moving, a function only
of
0

1
.
Now, using the property of path integrals that the integral of a total derivative is zero,
we have
0 =
_
[dX]

X(z, z)
_
e
S
X(z

, z

)
_
=
_
[dX] e
S
_

2
(z z

, z z

) +
1
2

z
X(z, z)X(z

, z

)
_
= <
2
(z z

, z z

) > +
1
2

z
< X(z, z)X(z

, z

) > (1.1.9)
That is, the equation of motion holds except at coincident points. Now, the same calculation
goes through if we have arbitrary additional insertions . . . in the path integral, as long as
no other elds are at (z, z) or (z

, z

):
1
2

z
< X(z, z)X(z

, z

) . . . > = <
2
(z z

, z z

) . . . > . (1.1.10)
A relation which holds in this sense will simply be written
1
2

z
X(z, z)X(z

, z

) =
2
(z z

, z z

), (1.1.11)
and will be called an operator equation. One can think of the additional elds . . . as
preparing arbitrary initial and nal states, so if one cuts the path integral open to make an
Hamiltonian description, an operator equation is simply one which holds for arbitrary matrix
elements. Note also that because of the way the path integral is constructed from iterated
time slices, any product of elds in the path integral goes over to a time-ordered product
in the Hamiltonian form. In the Hamiltonian formalism, the delta-function in eq. (1.1.11)
comes from the dierentiation of the time-ordering.
Now we dene a very useful combinatorial tool, normal ordering:
: X(z, z)X(z

, z

) : X(z, z)X(z

, z

) + ln[z z

[
2
. (1.1.12)
7
The logarithm satises the equation of motion (1.1.11) with the opposite sign (the action
was normalized such that this log would have coecient 1), so that by construction

z
: X(z, z)X(z

, z

) : = 0. (1.1.13)
That is, the normal ordered product satises the naive equation of motion. This implies
that the normal ordered product is locally the sum of an analytic and antianalytic function
(a standard result from complex analysis). Thus it can be Taylor expanded, and so from the
denition (1.1.12) we have (putting one operator at the origin for convenience)
X(z, z)X(0, 0) = ln[z[
2
+ : X
2
(0, 0) : + z : XX(0, 0) : + z : X

X(0, 0) : +. . . . (1.1.14)
This is an operator equation, in the same sense as the preceding equations.
Eq. (1.1.14) is our rst example of an operator product expansion. For a general expec-
tation value involving X(z, z)X(0, 0) and other elds, it gives the small-z behavior as a sum
of terms, each of which is a known function of z times the expectation values of a single
local operator. For a general eld theory, denote a complete set of local operators for a eld
theory by /
i
. The OPE then takes the general form
/
i
(z, z)/
j
(0, 0) =

k
c
k
ij
(z, z)/
k
(0, 0). (1.1.15)
Later in section 1 I will give a simple derivation of the OPE (1.1.15), and of a rather broad
generalization of it. OPEs are frequently used in particle and condensed matter physics as
asymptotic expansions, the rst few terms giving the dominant behavior at small z. However,
I will argue that, at least in conformally invariant theories, the OPE is actually a convergent
series. The radius of convergence is given by the distance to the nearest other operator in
the path integral. Because of this the coecient functions c
k
ij
(z, z), which as we will see
must satisfy various further conditions, will enable us to reconstruct the entire eld theory.
Exercise: The expectation value
4
< X(z
1
, z
1
)X(z
2
, z
2
)X(z
3
, z
3
)X(z
4
, z
4
) > is given by the
sum over all Wick contractions with the propagator ln[z
i
z
j
[
2
. Compare the asymptotics
as z
1
z
2
from the OPE (1.1.14) with the asymptotics of the exact expression. Verify that
the expansion in z
1
z
2
has the stated radius of convergence.
4
To be precise, expectation values of X(z, z) generally suer from an infrared divergence on the plane.
This is a distraction which we ignore by some implicit long-distance regulator. In practice one is always
interested in good operators such as derivatives or exponentials of X, which have well-dened expectation
values.
8
The various operators on the right-hand side of the OPE (1.1.14) involve products of
elds at the same point. Usually in quantum eld theory such a product is divergent and
must be appropriately cut o and renormalized, but here the normal ordering renders it
well-dened. Normal ordering is thus a convenient way to dene composite operators in
free eld theory. It is of little use in most interacting eld theories, because these have
additional divergences from interaction vertices approaching the composite operator or one
another. But many of the conformal eld theories that we will be interested in are free, and
many others can be related to free eld theories, so it will be worthwhile to develop normal
ordering somewhat further.
For products of more than 2 elds the denition (1.1.12) can be extended iteratively,
: X(z, z)X(z
1
, z
1
) . . . X(z
n
, z
n
) : X(z, z) : X(z
1
, z
1
) . . . X(z
n
, z
n
) : (1.1.16)
+
_
ln[z z
1
[
2
: X(z
2
, z
2
) . . . X(z
n
, z
n
) : + (n 1) permutations
_
,
contracting each pair (omitting the pair and subtracting ln[z z
i
[
2
). This has the same
properties as before: the equation of motion holds inside the normal ordering, and so the
normal-ordered product is smooth. (Exercise: Show this. The simplest argument I have
found is inductive, and uses the denition twice to pull both X(z, z) and X(z
1
, z
1
) out of
the normal ordering.)
The denition (1.1.16) can be written more formally as
: X(z, z)T[X] : = X(z, z) : T[X] : +
_
d
2
z

ln[z z

[
2

X(z

, z

)
: T[X] :, (1.1.17)
for an arbitrary functional T[X], the integral over the functional derivative producing all
contractions. Finally, the denition of normal ordering can be written in a closed form by
the same strategy,
: T[X] : = exp
_
1
2
_
d
2
z d
2
z

ln[z z

[
2

X(z, z)

X(z

, z

)
_
T[X]. (1.1.18)
The exponential sums over all ways of contracting zero, one, two, or more pairs. The operator
product of two normal ordered operators can be represented compactly as
: T[X] : : ([X] : = exp
_

_
d
2
z

d
2
z

ln[z

[
2

F
X(z

, z

G
X(z

, z

)
_
: T[X] ([X] :,
(1.1.19)
9
where
F
and
G
act only on the elds in T and ( respectively. The expressions : T[X] : : ([X] :
and : T[X] ([X] : dier by the contractions between one eld from T and one eld from (,
which are then restored by the exponential. Now, for T a local operator at z
1
and ( a local
operator at z
2
, we can expand in z
1
z
2
inside the normal ordering on the right to generate
the OPE. For example, one nds
: e
ik
1
X(z, z)
: : e
ik
2
X(0,0)
: = [z[
2k
1
k
2
: e
ik
1
X(z, z)+ik
2
X(0,0)
:
[z[
2k
1
k
2
: e
i(k
1
+k
2
)X(0,0)
: , (1.1.20)
since each contraction gives k
1
k
2
ln[z[
2
and the contractions exponentiate. Exponential op-
erators will be quite useful to us. Another example is
X(z, z) : e
ikX(0,0)
:
ik
z
: e
ikX(0,0)
: , (1.1.21)
coming from a single contraction.
1.2 Ward Identities
The action (1.1.7) has a number of important symmetries, in particular conformal invariance.
Let us rst derive the Ward identities for a general symmetry. Suppose we have elds

()
with some action S[], and a symmetry

() =

() +

(). (1.2.1)
That is, the product of the path integral measure and the weight e
S
is invariant. For a path
integral with general insertion T[], make the change of variables (1.2.1). The invariance of
the integral under change of variables, and the invariance of the measure times e
S
, give
0 =
_
d
2

<

()

()
T[] > < T[] > . (1.2.2)
This simply states that the general expectation value is invariant under the symmetry.
We can derive additional information from the symmetry: the existence of a conserved
current (Noethers theorem), and Ward identities for the expectation values of the current.
Consider the following change of variables,

() =

() +()

(). (1.2.3)
10
This is not a symmetry, the transformation law being altered by the inclusion of an arbitrary
function (). The path integral measure times e
S
would be invariant if were a constant, so
its variation must be proportional to the gradient
a
. Making the change of variables (1.2.3)
in the path integral thus gives
0 =
_
[d

] e
S[

_
[d] e
S[]
=
i
2
_
[d] e
S[]
_
d
2
j
a
()
a
(). (1.2.4)
The unknown coecient j
a
() comes from the variation of the measure and the action, both
of which are local, and so it must be a local function of the elds and their derivatives.
Taking the function to be nonzero only in a small region allows us to integrate by parts;
also, the identity (1.2.4) remains valid if we add arbitrary distant insertions . . ..
5
We thus
derive

a
j
a
= 0 (1.2.5)
as an operator equation. This is Noethers theorem.
Exercise: Use this to derive the classical Noether theorem in the form usually found in
textbooks. That is, assume that S[] =
_
d
2
L((),
a
()) and ignore the variation of
the measure. Invariance of the action implies that the variation of the Lagrangian density is
a total derivative, L =

under a symmetry transformation (1.2.1). Then the classical

result is
j

= 2i
_
L

_
. (1.2.6)
The extra factor of 2i is conventional in conformal eld theory. The derivation we have
given is the quantum version of Noethers theorem, and assumes that the path integral can
in fact be dened in a way consistent with the symmetry.
Now to derive the Ward identity, take any closed contour C, and let () = 1 inside C
and 0 outside C. Also, include in the path integral some general local operator /(z
0
, z
0
) at a
point z
0
inside C, and the usual distant insertions . . .. Proceeding as above we obtain the
operator relation
1
2
_
C
(d
2
j
1
d
1
j
2
) /(z
0
, z
0
) = i/(z
0
, z
0
). (1.2.7)
5
Our convention is that . . . refers to distant insertions used to prepare a general initial and nal state
but which otherwise play no role, while F is a general insertion in the region of interest.
11
This relates the integral of the current around any operator to the variation of the operator.
In complex coordinates, the left-hand side is
1
2i
_
C
(dz j d z

j) /(z
0
, z
0
), (1.2.8)
where the contour runs counterclockwise and we abbreviate j
z
to j and j
z
to

j. Finally, in
conformal eld theory it is usually the case that j is analytic and

j antianalytic, except for
singularities at the other elds, so that the integral (1.2.8) just picks out the residues. Thus,
i/(z
0
, z
0
) = Res
zz
0
j(z)/(z
0
, z
0
) + Res
z z
0

j( z)/(z
0
, z
0
). (1.2.9)
Here Res and Res pick out the coecients of (z z
0
)
1
and ( z z
0
)
1
respectively. This
form of the Ward identity is particularly convenient in CFT.
It is important to note that Noethers theorem and the Ward identity are local properties
that do not depend on whatever boundary conditions we might have far away, not even
whether the latter are invariant under the symmetry. In particular, since the function ()
is nonzero only in the interior of C, the symmetry transformation need only be dened there.
1.3 Conformal Invariance
Systems at a critical point are invariant under overall rescalings of space, z z

= az; if the
system is also rotationally invariant, a can be complex. These transformations rescale the
metric,
ds
2
= d
a
d
a
= dzd z dz

d z

= [a[
2
ds
2
. (1.3.1)
Under fairly broad conditions, a scale invariant system will also be invariant under the larger
symmetry of conformal transformations, which also will play a central role in string theory.
These are transformations z z

(z, z) which rescale the metric by a position-dependent

factor:
ds
2

2
(z, z)ds
2
. (1.3.2)
Such a transformation will leave invariant ratios of lengths of innitesimal vectors located
at the same point, and so also angles between them. In complex coordinates, it is easy to
see that this requires that z

be an analytic function of z,
6
z

= f(z), z

=

f( z). (1.3.3)
6
A antianalytic function z

= f( z) also works, but changes the orientation. In most string theories,

including the ones of greatest interest, the orientation is xed.
12
A theory with this invariance is termed a conformal eld theory (CFT).
The free action (1.1.7) is conformally invariant with X transforming as a scalar,
X

(z

, z

) = X(z, z), (1.3.4)

the transformation of d
2
z osetting that of the derivatives. For an innitesimal transforma-
tion, z

= z + g(z), we have X = g(z)X g( z)

X, and Noethers theorem gives the

current
j(z) = ig(z)T(z),

j( z) = i g( z)

T( z), (1.3.5)
where
7
T(z) =
1
2
: XX:,

T( z) =
1
2
:

X

X: . (1.3.6)
Because g(z) and g( z) are linearly independent, both terms in the divergence

j

j must
vanish independently,

j =

j = 0, (1.3.7)
as is indeed the case. The Noether current for a rigid translation is the energy-momentum
tensor, j
a
= iT
ab

b
(the i from CFT conventions), so we have
T
zz
= T(z), T
z z
=

T( z), T
z z
= T
zz
= 0. (1.3.8)
With the vanishing of T
z z
= T
zz
, the conservation law
a
T
a
b
= 0 implies that T
zz
is analytic
and T
z z
is antianalytic; this is a general result in CFT. By the way, T
zz
and T
z z
are in no sense
conjugate to one another (we will see, for example, that in they act on completely dierent
sets of oscillator modes), so I use a tilde rather than a bar on

T. The transformation of X,
with the Ward identity (1.2.9), implies the operator product
T(z)X(0) =
1
z
X(0) + analytic,

T( z)X(0) =
1
z

X(0) + analytic. (1.3.9)

This is readily veried from the specic form (1.3.6), and one could have used it to derive
the form of T.
7
Many students asked why T automatically came out normal-ordered. The answer is simply that in this
particular case all ways of dening the product (at least all rotationally invariant renormalizations) give the
same result; they could dier at most by a constant, but this must be zero because T transforms by a phase
under rotations. It was also asked how one knows that the measure is conformally invariant; this is evident
a posteriori because the conformal current is indeed conserved.
13
For a general operator /, the variation under rigid translation is just
a

a
/, which de-
termines the 1/z term in the T/ OPE. We usually deal with operators which are eigenstates
of the rigid rescaling plus rotation z

= az:
/

(z

, z

) = a
h
a

h
/(z, z). (1.3.10)
The (h,

h) are the weights of /. The sum h +

h is the dimension of /, determining its
behavior under scaling, while h

h is the spin, determining its behavior under rotations.

The Ward identity then gives part of the OPE,
T(z)/(0, 0) = . . . +
h
z
2
/(0, 0) +
1
z
/(0, 0) +. . . (1.3.11)
and similarly for

T. A special case is a tensor or primary operator O, which transforms
under general conformal transformations as
O

(z

, z

) = (
z
z

)
h
(
z
z

h
O(z, z). (1.3.12)
This is equivalent to the OPE
T(z)O(0, 0) =
h
z
2
O(0, 0) +
1
z
O(0, 0) +. . . , (1.3.13)
the more singular terms in the general OPE (1.3.11) being absent. In the free X CFT, one
can check that X is a tensor of weight (1, 0),

X a tensor of weight (0, 1), and : e
ikX
: a
tensor of weight
1
2
k
2
, while
2
X has weight (2, 0) but is not a tensor.
For the energy-momentum tensor with itself one nds for the free X theory
T(z)T(0) =
1
2z
4
+
2
z
2
T(0) +
1
z
T(0) + analytic, (1.3.14)
and similarly for

T,
8
so this is not a tensor. Rather, the OPE (1.3.14) implies the transfor-
mation law
T(z) =
1
12

3
z
g(z) 2
z
g(z)T(z) g(z)
z
T(z). (1.3.15)
8
One easily sees that the T

T OPE is analytic. By the way, unless otherwise stated OPEs hold only at
non-zero separation, ignoring possible delta functions. For all of the applications we will have the latter do
not matter. Occasionally it is useful to include the delta functions, but in general these depend partly on
denitions so one must be careful.
14
More generally, the TT OPE in any CFT is of the form
T(z)T(0) =
c
2z
4
+
2
z
2
T(0) +
1
z
T(0) + analytic, (1.3.16)
with c a constant known as the central charge. The central charge of a free boson is 1; for
D free bosons it is D. The nite form of the transformation law (1.3.15) is
(
z
z

)
2
T

(z

) = T(z) +
c
12
z

, z, (1.3.17)
where f, z denotes the Schwarzian derivative,
f, z =
2
3
z
f
z
f 3
2
z
f
2
z
f
2
z
f
z
f
. (1.3.18)
The corresponding form holds for

T, possibly with a dierent central charge c.
1.4 Mode Expansions
For an analytic or antianalytic operator we can make a Laurent expansion,
T(z) =

m=
L
m
z
m+2
,

T( z) =

m=

L
m
z
m+2
. (1.4.1)
The Laurent coecients, known as the Virasoro generators, are given by the contour integrals
L
m
=
_
C
dz
2i
z
m+1
T(z),

L
m
=
_
C
d z
2i
z
m+1

T( z), (1.4.2)
where C is any contour encircling the origin. This expansion has a simple and important
interpretation [20]. Dening any monotonic time variable, one can slice open a path integral
along the constant-time curves to recover a Hamiltonian description. In particular, let time
be ln[z[, running radially outward from z = 0. This may seem odd, but is quite natural
in CFTin terms of the conformally equivalent coordinate w dened z = e
iw
, an annular
region around z = 0 becomes a cylinder, with Im(w) being the time and Re(w) being a spatial
coordinate with periodicity 2. Thus, the radial time slicing is equivalent to quantizing the
CFT on a nite periodic space; this is what will eventually be interpreted as the quantization
of a closed string. The +2s in the exponents (1.4.1) come from the conformal transformation
15
z=0
C C
1 2 3
C
2
C
C
1
C
3
z
2
a) b)
Figure 1: a) Contours centered on z = 0. b) For given z
2
on contour C
2
, contour C
1
C
3
is
contracted.
of T, so that in the w frame m just denotes the Fourier mode; for an analytic eld of weight
h this becomes +h.
In the Hamiltonian form, the Virasoro generators become operators in the ordinary sense.
Since by analyticity the integrals (1.4.2) are independent of C, they are actually conserved
charges, the charges associated with the conformal transformations. It is an important fact
that the OPE of currents determines the algebra of the corresponding charges. Consider
charges Q
i
, i = 1, 2:
Q
i
C =
_
C
dz
2i
j
i
. (1.4.3)
Then we have
Q
1
C
1
Q
2
C
2
Q
1
C
3
Q
2
C
2
= [Q
1
, Q
2
]C
2
(1.4.4)
The charges on the left are dened by the contours shown in g. 1a; when we slice open
the path integral, operators are time-ordered, so the dierence of contours generates the
commutator. Now, for a given point z
2
on the contour C
2
, we can deform the dierence of
the C
1
and C
3
contours as shown in g. 1b, with the result
[Q
1
, Q
2
]C
2
=
_
C
2
dz
2
2i
Res
zz
2
j
1
(z)j
2
(z
2
) (1.4.5)
16
Applying this to the Virasoro generators gives the Virasoro algebra,
[L
m
, L
n
] = (mn)L
m+n
+
c
12
(m
3
m)
m+n,0
. (1.4.6)
The

L
m
satisfy the same algebra with central charge c. For the Laurent coecients of an
analytic tensor eld O of weight (h, 0), one nds from the OPE (1.3.13) the commutator
[L
m
, O
n
] = ([h 1]mn)O
m+n
. (1.4.7)
Note that commutation with L
0
is diagonal and proportional to n. Modes O
n
for n > 0
reduce L
0
and are termed lowering operators, while modes O
n
for n < 0 increase L
0
and are
termed raising operators. From the OPE (1.3.13) and the denitions, we see that a tensor
operator is annihilated by all the lowering operators,
9
L
n
O = 0, n > 0. (1.4.8)
For an arbitrary operator, it follows from the OPE (1.3.11) that
L
0
/ = h/,

L
0
/ =

h/, L
1
/ = /,

L
1
/ =

/. (1.4.9)
Note that L
0
+

L
0
is the generator of scale transformations, or in other words of radial time
translations. It diers from the Hamiltonian H of the cylindrical w coordinate system by an
additive constant from the non-tensor behavior of T,
H = L
0
+

L
0

c + c
24
. (1.4.10)
Similarly, L
0

L
0
measures the spin, and is equal to the spatial translation generator in the
w frame, up to an additive constant.
For the free X CFT, the Noether current of translations is (iX(z), i

X( z)). Again,
the components are separately analytic and antianalytic, which signies the existence of an
enlarged symmetry X X +y(z) + y( z). Dene the modes
iX(z) =

m=

m
z
m+1
, i

X( z) =

m=

m
z
m+1
. (1.4.11)
9
Often one deals with dierent copies of the Virasoro algebra dened by Laurent expansions in dierent
coordinates z, so I like to put a between the generator and the operator as a reminder that the generators
are dened in the coordinate centered on the operator.
17
From the OPE
iX(z) iX(0) =
1
z
2
+ analytic, (1.4.12)
we have the algebra
[
m
,
n
] = m
m+n,0
, (1.4.13)
and the same for
m
. As expected for a free eld, this is a harmonic oscillator algebra for
each mode; in terms of the usual raising and lowering operators
m

ma,
m

ma

.
To generate the whole spectrum we start from a state [0, k) which is annihilated by the
m > 0 operators and is an eigenvector of the m = 0 operators,

m
[0, k) =
m
[0, k) = 0, m > 0,
0
[0, k) =
0
[0, k) = k[0, k) . (1.4.14)
The rest of the spectrum is generated by the raising operators
m
and
m
for m < 0. Note
that the eigenvalues of
0
and
0
must be equal because X is single valued,
_
(dz X +
d z

X) = 0; later we will relax this.
Inserting the expansion (1.4.11) into T(z) and comparing with the Laurent expansion
gives
L
m

1
2

n=

mn
. (1.4.15)
However, we must be careful about operator ordering. The Virasoro generators were dened
in terms of the normal ordering (1.1.12), while for the mode expansion it is most convenient
to use a dierent ordering, in which all raising operators are to the left of the lowering
operators. Both of these procedures are generally referred to as normal ordering, but they
are in general dierent, so we might refer to the rst as conformal normal ordering and the
latter as creation-annihilation normal ordering. Since conformal normal order is our usual
method, we will simply refer to it as normal ordering. We could develop a dictionary between
these, but there are several ways to take a short-cut. Only for m = 0 do non-commuting
operators appear together, so we must have
L
0
=
1
2

2
0
+

n=1

n
+A
L
m
=
1
2

n=

mn
, m ,= 0. (1.4.16)
for some constant A. Now use the Virasoro algebra as follows
(L
1
L
1
L
1
L
1
)[0, 0) = 2L
0
[0, 0) = 2A[0, 0). (1.4.17)
18
All terms on the left have
m
with m 0 acting on [0, 0) and so must vanish; thus,
A = 0. (1.4.18)
Thus, a general state

m
1
. . .
mp

m

1
. . .
m

q
[0, k) (1.4.19)
has
L
0
=
1
2
k
2
+ L,

L
0
=
1
2
k
2
+

L (1.4.20)
where the levels L,

L are the total oscillator excitation numbers,
L = m
1
+. . . +m
p
,

L = m

1
+. . . +m

q
. (1.4.21)
One needs to calculate the normal ordering constant A often, so the following heuristic-
but-correct rules are useful:
1. Add the zero point energies,
1
2
for each bosonic mode and
1
2
for each fermionic.
2. One encounters divergent sums of the form

n=1
(n), the arising when one considers
nontrivial periodicity conditions. Dene this to be

n=1
(n ) =
1
24

1
8
(2 1)
2
. (1.4.22)
I will not try to justify this, but it is the value given by any conformally invariant renormal-
ization.
3. The above is correct in the cylindrical w coordinate, but for L
0
we must add the non-
tensor correction c/24.
For the free boson, the modes are integer so we get one-half of the sum (1.4.22) for = 0,
that is
1
24
, after step 2. This is just oset by the correction in step 3. The zero-point sum
in step 2 is a Casimir energy, from the nite spatial size. For a system of physical size l we
must scale H by 2/l, giving (including the left-movers) the correct Casimir energy /6l.
For antiperiodic scalars one gets the sum with =
1
2
and Casimir energy /12l.
To get the mode expansion for X, integrate the Laurent expansions (1.4.11). Dene rst
X
L
(z) = x
L
i
0
lnz +i

m=0

m
mz
m
, X
R
( z) = x
R
i
0
ln z +i

m=0

m
m z
m
, (1.4.23)
19
with
[x
L
,
0
] = [x
R
,
0
] = i. (1.4.24)
These give
X
L
(z)X
L
(z

) = ln(z z

) +analytic, X
R
( z)X
R
( z

) = ln( z z

) +analytic. (1.4.25)
Actually, these only hold modulo i as one can check, but we will not dwell on this.
10
In
any case we are for the present only interested in the sum,
X(z, z) = X
L
(z) +X
R
( z), (1.4.26)
for which the OPE X(z, z)X(0, 0) ln[z[
2
is unambiguous.
1.5 States and Operators
Radial quantization gives rise to a natural isomorphism between the state space of the CFT,
in a periodic spatial dimension, and the space of local operators. Consider the path integral
with a local operator / at the origin, no other operators inside the unit circle [z[ = 1, and
unspecied operators and boundary conditions outside. Cutting open the path integral on
the unit circle represents the path integral as an inner product
out
[
in
), where [
in
) is the
incoming state produced by the path integral at [z[ < 1 and [
out
) is the outgoing state
produced by the path integral at [z[ > 1. More explicitly, separate the path integral over
elds into an integral over the elds outside the circle, inside the circle, and on the circle
itself; call these last
B
. The outside integral produces a result
out
(
B
), and the inside
integral a result
in
(
B
), leaving
_
[d
B
]
out
(
B
)
in
(
B
) . (1.5.1)
The incoming state depends on /, so we denote it more explicitly as [
A
). This is the
mapping from operators to states. That is, integrating over the elds on the unit disk, with
xed boundary values
B
and with an operator / at the origin, produces a result
A
(
B
),
which is a state in the Schrodinger representation. The mapping from operators to states
is given by the path integral on the unit disk. To see the inverse, take a state [) to be an
eigenstate of L
0
and

L
0
. Since L
0
+

L
0
is the radial Hamiltonian, inserting [) on the unit
10
But it means that one sometimes need to introduce cocycles to x the phases of exponential operators.
20
Q
Q

A
a) b)
A
Figure 2: a) World-sheet (shaded) with state [
A
) on the boundary circle, acted upon by
Q. b) Equivalent picture: the unit disk with operator / has been sewn in along the dotted
line, and Q contracted around the operator.
circle is equivalent to inserting r
L
0

L
0
[) on a circle of radius r. Taking r to be innitesimal
denes a local operator which is equivalent to [) on the unit circle.
Exercise: This all sounds a bit abstract, so here is a calculation one can do explicitly. The
ground state of the free scalar is e

m=1
mXmXm/2
, where X
m
are the Fourier modes of X
on the circle. Derive this by canonical quantization of the modes, writing them in terms of
X
m
and /X
m
. Obtain it also by evaluating the path integral on the unit disk with X xed
on the boundary and no operator insertions. Thus the ground state corresponds to the unit
operator.
Usually one does not actually evaluate a path integral as above, but uses indirect argu-
ments. Note that if Q is any conserved charge, the state Q[
A
) corresponds to the operator
Q /, as shown in g. 2. Now, in the free theory consider the case that / is the unit operator
and let
Q =
m
=
_
C
dz
2
z
m
X, m 0. (1.5.2)
With no operators inside the disk, X is analytic and the integral vanishes for m 0. Thus,

m
[
1
) = 0, m 0, which establishes
1 [0, 0) (1.5.3)
21
as found directly in the exercise. Proceeding as above one nds
: e
ikX
: [0, k), (1.5.4)
and for the raising operators, evaluating the contour integral (1.5.2) for m < 0 gives
i
1
(k 1)!

k
X
k
, k 1 , (1.5.5)
and in parallel for the tilded modes. That is, the state obtained by acting with raising oper-
ators on (1.5.4) is given by the product of the exponential with the corresponding derivatives
of X; the product automatically comes out normal ordered.
The state corresponding to a tensor eld O satises
L
m
[
O
) = 0, m > 0. (1.5.6)
This is known as a highest weight or primary state. For almost all purposes one is interested
in highest-weight representations of the Virasoro algebra, built by acting on a given highest
weight state with the L
m
, m < 0.
The state-operator mapping gives a simple derivation of the OPE, shown in g. 3.
Consider the product /
i
(z, z)/
j
(0, 0), [z[ < 1. Integrating the elds inside the unit cir-
cle generates a state on the unit circle, which we might call [
ij,z, z
). Expand in a complete
set,
[
ij,z, z
) =

k
c
k
ij
(z, z)[
k
). (1.5.7)
Finally use the mapping to replace [
k
) on the unit circle with /
k
at the origin, giving
the general OPE (1.1.15). The claimed convergence is just the usual convergence of a com-
plete set in quantum mechanics. The construction is possible as long as there are no other
operators with [z

[ [z[, so that we can cut on a circle of radius [z[ +.

Incidentally, applying a rigid rotation and scaling to both sides of the general OPE
determines the z-dependence of the coecient functions,
/
i
(z, z)/
j
(0, 0) =

k
z
h
k
h
i
h
j
z

h
k

h
i

h
j
c
k
ij
/
k
(0, 0). (1.5.8)
From the full conformal symmetry one learns much more: all the c
k
ij
are determined in terms
of those of the primary elds.
22

A
a)
b)
A
c)
A
i
j
(0,0)
(z,z)
ij,z,z
ij,z,z
Figure 3: a) World-sheet with two local operators. b) Integration over elds on the interior
of the disk produces boundary state [
ij,z, z
). c) Sewing in a disk with the corresponding
local operator. Expanding in operators of denite weight gives the OPE.
For three operators, /
i
(0)/
j
(1)/
k
(z), the regions of convergence of the z 0 and z 1
OPEs ([z[ < 1 and [1z[ < 1) overlap. The coecient of /
m
in the triple product can then
be written as a sum involving c
l
ik
c
m
lj
or as a sum involving c
l
jk
c
m
li
. Associativity requires
these sums to be equal; this is represented schematically in g. 4.
A unitary CFT is one that has a positive inner product [ ); the double bracket is to
distinguish it from a dierent inner product to be dened later. Also, it is required that
23

i
k j
m
l
=
l

k
j
m i
l
l
Figure 4: Schematic picture of OPE associativity.
L

m
= L
m
,

L

m
=

L
m
. The X CFT is unitary with
0, k[0, k

) = 2(k k

) (1.5.9)
and

m
=
m
,

m
=
m
; this implicitly denes the inner product of all higher states.
Unitary CFTs are highly constrained; I will derive here a few of the basic results, and
mention others later.
The rst constraint is that any state in a unitary highest weight representation must
have h,

h 0. Consider rst the highest weight state itself, [O). The Virasoro algebra gives
2h
O
O[O) = 2O[L
0
[O) = O[[L
1
, L
1
][O) = |L
1
[O)|
2
0, (1.5.10)
so h
O
0. All other states in the representation, obtained by acting with the raising
generators, have higher weight so the result follows. It also follows that if h
O
= 0 then
L
1
O =

L
1
O. The relation (1.4.9) thus implies that O is independent of position;
general principle of quantum eld theory then require O to be a c-number. That is, the unit
operator is the only (0,0) operator. In a similar way, one nds that an operator in a unitary
CFT is analytic if and only if

h = 0, and antianalytic if and only if h = 0.
Exercise: Using the above argument with the commutator [L
n
, L
n
], show that c, c 0 in
a unitary CFT. In fact, the only CFT with c = 0 is the trivial one, L
n
= 0.
24
1.6 Other CFTs
Now we describe briey several other CFTs of interest. The rst is given by the same
action (1.1.7) as the earlier X theory, but with energy-momentum tensor [21]
T(z) =
1
2
: XX: +
Q
2

2
X,

T( z) =
1
2
:

X

X: +
Q
2

2
X. (1.6.1)
The TT operator product is still of the general form (1.3.16), but now has central charge
c = 1 + 3Q
2
. The change in T means that X is no longer a scalar,
X = (gX + g

X)
Q
2
(g +

g). (1.6.2)
Exponentials : e
ikX
: are still tensors, but with weight
1
2
(k
2
+ ikQ). One notable change
is in the state-operator mapping. The translation current j = iX is no longer a tensor,
j = gj jg iQ
2
g/2. The nite form is
11
(
z
z

)j
z
(z

) = j
z
(z)
iQ
2

2
z
z

z
z

. (1.6.3)
Applied to the cylinder frame z

= w = i lnz this gives

1
2i
_
dwj
w
=
0
+
iQ
2
. (1.6.4)
Thus a state [0, k) which whose canonical momentum (dened on the left) is k corresponds
to the operator
: e
ikX+QX/2
: . (1.6.5)
Note that i
0
just picks out the exponent of the operator, so
0
= k iQ/2.
The mode expansion of the m ,= 0 generators is
L
m
=
1
2

n=

mn
+
iQ
2
(m + 1)
m
, m ,= 0, (1.6.6)
11
To derive this, and the nite transformation (1.3.17) of T, you can rst write the most general form
which has the correct innitesimal limit and is appropriately homogeneous in z and z

indices, and x the

few resulting constants by requiring proper composition under z z

.
25
the last term coming from the
2
X term in T. For m = 0 the result is
L
0
=
1
8
(2
0
+iQ)
2
+
Q
2
8
+

n=1

,n
=
1
2
k
2
+
Q
2
8
+

n=1

,n
. (1.6.7)
The constant in the rst line can be obtained from the z
2
term in the OPE of T with the
vertex operator (1.6.5); this is a quick way to derive or to check normal-ordering constants.
In the second line, expressed in terms of the canonical momentum k it agrees with our
heuristic rules.
This CFT has a number of applications in string theory, some of which we will encounter.
Let me also mention a slight variation,
T(z) =
1
2
: XX: +i

2
X,

T( z) =
1
2
:

X

X: i

2
X, (1.6.8)
with central charge c = 1 3
2
. With the earlier transformation (1.6.2), the variation of
X contains a constant piece under rigid scale transformations (g a real constant). In other
words, one can regard X as the Goldstone boson of spontaneously broken scale invariance.
For the theory (1.6.8), the variation of X contains a constant piece under rigid rotations (g
an imaginary constant), and X is the Goldstone boson of spontaneously broken rotational
invariance. This is not directly relevant to string theory (the i in the energy-momentum
tensor makes the theory non-unitary) but occurs for real membranes (where the unitarity
condition is not relevant because both dimensions are spatial). In particular the CFT (1.6.8)
describes hexatic membranes,
12
in which the rotational symmetry is broken to Z
6
. The
unbroken discrete symmetry plays an indirect role in forbidding certain nonlinear couplings
between the Goldstone boson X and the membrane coordinates.
Another simple variation on the free boson is to make it periodic, but we leave this until
section 3 where we will discuss some interesting features.
Another family of free CFTs involves two anticommuting elds with action
S =
1
2
_
d
2
z b

c +

b c. (1.6.9)
12
I would like to thank Mark Bowick and Phil Nelson for educating me on this subject.
26
The equations of motion are

c(z) =

b(z) = c( z) =

b( z) = 0, (1.6.10)
so the elds are respectively analytic and antianalytic. The operator products are readily
found as before, with appropriate attention to the order of anticommuting variables,
b(z)c(0)
1
z
, c(z)b(0)
1
z
,

b( z) c(0)
1
z
, c( z)

b(0)
1
z
. (1.6.11)
We focus again on the analytic part; in fact the action (1.6.9) is a sum, and can be
regarded as two independent CFTs. The action is conformally invariant if b is a (, 0)
tensor, and c a (1 , 0) tensor; by interchange of b and c we can assume positive. The
corresponding energy-momentum tensor is
T(z) =: (b)c: : (bc) : . (1.6.12)
One nds that the TT OPE has the usual form with
c = 3(2 1)
2
+ 1. (1.6.13)
The elds have the usual Laurent expansions
b(z) =

m=
b
m
z
m+
, c(z) =

m=
c
m
z
m+1
, (1.6.14)
giving rise to the anticommutator
b
m
, c
n
=
m+n,0
. (1.6.15)
Also, c
m
, c
n
= b
m
, b
n
= 0. Because of the m = 0 modes there are two natural ground
states, [ ) and [ ). Both are annihilated by b
m
and c
m
for m > 0, while
b
0
[ ) = 0, c
0
[ ) = 0. (1.6.16)
These are related [ ) = c
0
[ ), [ ) = b
0
[ ). With the antianalytic theory included, there are
also the zero modes

b
0
and c
0
and so four ground states[ ), etc.
27
The Virasoro generators in terms of the modes are
L
0
=

n=1
n(b
n
c
n
+c
n
b
n
)
( 1)
2
L
m
=

n=
mnb
n
c
mn
, m ,= 0 . (1.6.17)
The ordering constant is found as before. Two sets (b and c) of integer anticommuting modes
give
1
12
at step 2, and the central charge correction then gives the result above.
The state-operator mapping is a little tricky. Let be an integer, so that the Laurent
expansion (1.6.14) has no branch cut. For the unit operator the elds are analytic at the
origin, so
b
m
[
1
) = 0, m 1 , c
m
[
1
) = 0, m . (1.6.18)
Thus, the unit state is in general not one of the ground states, but rather
[
1
) = b
1
b
2
. . . b
1
[ ), (1.6.19)
up to normalization. Also, we have the dictionary
b
m

1
(m)!

m
b, c
m

1
(m+ 1)!

m+1
c. (1.6.20)
Thus we have, taking the value = 2 which will be relevant later,
[ ) = c
1
[
1
) c, [ ) = c
0
c
1
[
1
) c c. (1.6.21)
The bc theory has a conserved current j = : cb :, called ghost number, which counts the
number of cs minus the number of bs. In the cylindrical w frame the vacua have average
ghost number zero, so
1
2
for [ ) and +
1
2
for [ ). The ghost numbers of the corresponding
operators are 1 and , as we see from the example (1.6.21). As in the case of the
momentum (1.6.4), the dierence arises because the current is not a tensor.
For the special case =
1
2
, b and c have the same weight and the bc system can be split
in two in a conformally invariant way, b = (
1
+i
2
)/

2, c = (
1
i
2
)/

2, and
S =
1
2
_
d
2
z b

c =
1
4
_
d
2
z
_

1
+
2

2
_
. (1.6.22)
28
Each theory has central charge
1
2
. The antianalytic theory separates in the same way. We
will refer to these as Majorana (real) fermions, because it is a unitary CFT with

m
=
m
.
Another family of CFTs diers from the bc system only in that the elds commute. The
action is
S =
1
2
_
d
2
z

. (1.6.23)
The elds and are analytic by the equations of motion; as usual there is a corresponding
antianalytic theory. Because the statistics are changed, some signs in operator products are
dierent,
(z)(0)
1
z
, (z)(0)
1
z
. (1.6.24)
The action is conformally invariant with a weight (, 0) tensor and a (1 , 0) tensor.
The energy-momentum tensor is
T(z) = : () : : () : . (1.6.25)
The central charge has the opposite sign relative to the bc system because of the changed
statistics,
c = 3(2 1)
2
1. (1.6.26)
All of the above are free eld theories. A simple interacting theory is the non-linear sigma
model [22]-[24], consisting of D scalars X

with a eld-dependent kinetic term,

S =
1
4
_
d
2
z
_
G

(X) + iB

(X)
_
X

, (1.6.27)
with G

= G

and B

= B

. Eectively the scalars dene a curved eld space, with

G

the metric on the space. The path integral is no longer gaussian, but when G

(X)
and B

(X) are slowly varying the interactions are weak and there is a small parameter.
The action is naively conformally invariant, but a one-loop calculation reveals an anomaly
(obviously this is closely related to the -function for rigid scale transformations),
T
z z
=
_
2R

+
1
2
H

_
X

. (1.6.28)
Here R

is the Ricci curvature built from G

(I am using boldface to distinguish it from

the two-dimensional curvature to appear later),

denotes the covariant derivative in this

metric, and
H

. (1.6.29)
29
To this order, any Ricci-at space with B

= 0 gives a CFT. At higher order these con-

ditions receive corrections. Other solutions involve cancellations between terms in (1.6.28).
A three dimensional example is the 3-sphere with a round metric of radius r, and with
H

=
4q
r
3

(1.6.30)
proportional to the antisymmetric three-tensor. By symmetry, the rst two terms in T
z z
are
proportional to G

and the third vanishes. Thus T

z z
vanishes for an appropriate
relation between the constants, r
4
= 4q
2
. There is one subtlety. Locally the form (1.6.30)
is compatible with the denition (1.6.29) but not globally. This conguration is the analog
of a magnetic monopole, with the gauge potential B

now having two indices and the eld

strength H

. Then B

must have a Dirac string singularity, which is invisible to the

string if the eld strength is appropriately quantized; I have normalized q just such that
it must be an integer. So this denes a discrete series of models. The one-loop correction
to the central charge is c = 3 6/[q[ + O(1/q
2
). The 3-sphere is the SU(2) group space,
and the theory just described is the SU(2) Wess-Zumino-Witten (WZW) model [25]-[27] at
level q. It can be generalized to any Lie group. Although this discussion is based on the
one-loop approximation, which is accurate for large r (small gradients) and so for large q,
these models can also be constructed exactly, as will be discussed further shortly.
For c < 1, it can be shown that unitary CFTs can exist only at the special values [18], [28]
c = 1
6
m(m + 1)
, m = 2, 3, . . . . (1.6.31)
These are the unitary minimal models, and can be solved using conformal symmetry alone.
The point is that for c < 1 the representations are all degenerate, certain linear combinations
of raising operators annihilating the highest weight state, which gives rise to dierential
equations for the expectation value of the corresponding tensor operator. These CFTs have
a Z
2
symmetry and m2 relevant operators, and correspond to interesting critical systems:
m = 3 to the Ising model (note that c =
1
2
corresponds to the free fermion), m = 4 to the
tricritical Ising model, m = 5 to a multicritical Z
2
Ising model but also to the three-state
Potts model, and so on.
This gives a survey of the main categories of conformal eld theory, including some CFTs
that will be of specic interest to us later on. It is familiar that there are many equivalences
between dierent two dimensional eld theories. For example, the ordinary free boson is
30
equivalent to a free fermion, which is the same as the bc system at =
1
2
. The bosonization
dictionary is
b : e
iX
: , c : e
iX
: (1.6.32)
In fact this extends to the general bc and X CFTs, with complex Q = i(12). The reader
can check that the weights of b and c, and the central charge, then match. There is also a
rewriting of and in terms of exponentials, which is more complicated but useful in the
superstring. All of these subjects are covered in ref. [19]. As a further example, the level 1
SU(2) WZW model, which we have described in terms of three bosons, can also be written in
terms of a single free boson; the level 2 SU(2) WZW model can be written in terms of three
Majorana fermions; these will be explained further in the next section. The minimal models
are related to the free X theory with Q such as to give the appropriate central charge [29],
but this is somewhat indirect.
1.7 Other Algebras
The Virasoro algebra is just one of several important innite dimensional algebras. Another
is obtained from T(z) plus any number of analytic (1, 0) tensors j
a
(z). The constraints
obtained at the end of section 1.5 imply that if the algebra is to have unitary representations
the jj OPE can only take the form
j
a
(z)j
b
(0)
k
ab
z
2
+i
f
ab
c
z
j
c
(0). (1.7.1)
The corresponding Laurent expansion is
j
a
(z) =

m=
j
a
m
z
m+1
, (1.7.2)
and the corresponding algebra
[j
a
m
, j
b
n
] = mk
ab

m+n,0
+if
ab
c
j
c
m+n
. (1.7.3)
This is known variously as a current algebra, an ane Lie algebra, or sometimes as a Kac-
Moody algebra; for general references see [30], [26] and [27]. The m = n = 0 modes form
an ordinary Lie algebra g with structure constants f
ab
c
. The latter must therefore satisfy
the Jacobi identity; another Jacobi identity implies that k
ab
is g-invariant. The energy-
momentum tensor can be shown to separate into a piece built from the current (the Sugawara
31
construction) and a piece commuting with the current. The CFT is thus a product of a part
determined by the symmetry and a part independent of the symmetry.
For a single Abelian current we already have the example of the free X theory. The next
simplest case is SU(2),
[j
a
m
, j
b
n
] = mk
ab

m+n,0
+i

2
abc
j
c
m+n
. (1.7.4)
The value of k must be an integer, and non-negative in a unitary theory.
Exercise: Construct an SU(2) algebra containing (j
1
+ij
2
)
1
and (j
1
ij
2
)
1
, and use it to
show that k is an integer.
The Sugawara central charge is c = 3k/(k + 2). The SU(2) WZW model just discussed has
k = [q[. The case k = 1 can also be realized in terms of a single free scalar as
j
1
=

2 : cos

2X: , j
2
=

2 : sin

2X: , j
3
= iX . (1.7.5)
The case k = 2 can also be realized in terms of three Majorana fermions, j
a
= i
abc

c
/

2.
The energy momentum tensor together with a weight (
3
2
, 0) tensor current (supercurrent)
T
F
form the N = 1 superconformal algebra [31], [32]. The T
F
T
F
OPE is
T
F
(z)T
F
(0)
2c
3z
3
+
2
z
T(0); (1.7.6)
TT
F
has the usual tensor form (1.3.13). A simple realization is in terms of a free scalar X
and a Majorana fermion ,
T
F
= iX, T =
1
2
(: XX: + : : ). (1.7.7)
With the Laurent expansions
T
F
(z) =

r=
G
r
z
r+3/2
, (z) =

r=

r
z
r+1/2
, (1.7.8)
the algebra is
G
r
, G
s
= 2L
r+s
+
c
12
(4r
2
1)
r+s,0
. (1.7.9)
The central charge must be the same is in the TT OPE, by the Jacobi identity. Note that for
r running over integers, the elds (1.7.8) have branch cuts at the origin, but the correspond-
ing elds in the cylindrical w frame are periodic due to the tensor transformation (w/z)
h
.
32
This is the Ramond sector. Antiperiodic boundary conditions on and T
F
in the w frame
are also possible; this is the Neveu-Schwarz sector. All of the above goes through with r
running over integers-plus-
1
2
, and the elds in the z frame are single valued in this sector.
Exercise: Work out the expansions of L
m
and G
r
in terms of the modes of X and . [An-
swer: the normal-ordering constant is 0 in the Neveu-Schwarz sector and
1
16
in the Ramond
sector.]
The operators corresponding to Ramond-sector states thus produce branch cuts in the
fermionic elds, and are known as spin elds. They are most easily described using bosoniza-
tion. With two copies of the free representation (1.7.7), the bosonization is (
1
i
2
)/

2 =
: e
iX
: . There are two Ramond ground states, which correspond to the operators : e
iX/2
:.
Observe that this has the necessary branch cut with
1
,
2
, and also that its weight,
1
8
= 2
1
16
,
agrees with the exercise.
The energy-momentum tensor with two (
3
2
, 0) tensors T

F
plus a (1, 0) current j form the
N = 2 superconformal algebra [33]
T
+
F
(z)T

F
(0)
2c
3z
3
+
1
z
2
j(0) +
2
z
T(0) +
1
2z
j(0),
j(z)T

F
(0)
1
z
T

F
(0), (1.7.10)
with T
+
F
T
+
F
and T

F
T

F
analytic. This can be generalized to N supercurrents, leading to an
algebra with weights 2,
3
2
, . . . , 2
1
2
N. From the earlier discussion we see that there are no
unitary representations for N > 4. There is one N = 3 algebra and two distinct N = 4
algebras [33],[34].
These are the algebras which play a central role in string theory, but many others arise in
various CFTs. I will briey discuss some higher-spin algebras, which have a number of in-
teresting applications (for reviews of the various linear and nonlinear higher spin algebras see
refs. [35],[36]). The free scalar action (1.1.7) actually has an enormous amount of symmetry,
but let us in particular pick out
X(z, z) =

l=0
g
l
(z)(X(z))
l+1
, l = 0, 1, . . . (1.7.11)
The Noether currents
V
l
(z) =
1
l + 2
: (X(z))
l+2
: (1.7.12)
33
have spins l + 2. Making the usual Laurent expansion, one nds the w

algebra
[V
i
m
, V
j
n
] =
_
[j + 1]m[i + 1]n
_
V
i+j
m+n
. (1.7.13)
The l = 0 generators are just the usual Virasoro algebra. There are a number of related
algebras. Adding in the l = 1 generators (which are just the modes of the translation
current) denes the w
1+
algebra. Another algebra with the same spin content as w

but a
more complicated commutator is W

; w

can be obtained as a limit (contraction) of W

.
These algebras have a simple and useful realization in terms of the classical and quantum
mechanics of a particle in one dimension:
V
i
m

1
4
(p +x)
i+m+1
(p x)
im+1
. (1.7.14)
The Poisson bracket algebra of these is the wedge subalgebra (m i + 1) of w

; the
commutator algebra is the wedge subalgebra of W

. In the literature one must beware of

diering notations and conventions.
All of these algebras have supersymmetric extensions, with generators of half-integral
spins. There are also various algebras with a nite number of higher weights. One family is
W
N
, closely related to W

, with weights up to N. The commutator of two weight-3 currents

in W

contains the weight-4 current. In W

3
this is not an independent current but the
square of T(z), so the algebra is nonlinear.
1.8 Riemann Surfaces
Thus far we have focussed on local properties, without regard to the global structure or
boundary conditions. For string theory we will be interested in conformal eld theories on
closed manifolds. The appropriate manifold for a two-dimensional CFT to live on is a two-
dimensional complex manifold, a Riemann surface. One can imagine this as being built up
from patches, patch i having a coordinate z
(i)
which runs over some portion of the complex
plane. If patches i and j overlap, there is a relation between the coordinates,
z
(j)
= f
ij
(z
(i)
) (1.8.1)
with f
ij
an analytic function. Two Riemann surfaces are equivalent if there is a mapping
between them such that the coordinates on one are analytic functions of the coordinates on
the other. This is entirely parallel to the denition of a dierentiable manifold, but it has
34
more structurethe manifold comes with a local notion of analyticity. Since in a CFT each
eld has a specic transformation law under analytic changes of coordinates, the transition
function (1.8.1) is just the information needed to extend the eld from patch to patch.
A simple example is the sphere, which we can imagine as built from two copies of the
complex plane, with coordinates z and u, with the mapping
u =
1
z
. (1.8.2)
The z coordinate cannot quite cover the sphere, the point at innity being missing. All
Riemann surfaces with the topology of the sphere are equivalent. For future reference let us
note that the sphere has a group of globally dened conformal transformations (conformal
Killing transformations), which in the z patch take
z

=
z +
z +
, (1.8.3)
where , , and are complex parameters which can be chosen such that = 1.
This is the M obius group.
The next Riemann surface is the torus. Rather than build it from patches it is most
convenient to describe it as in g. 5 by taking a single copy z of the complex plane and
identifying points
z

= z + 2

= z + 2, (1.8.4)
producing a parallelogram-shaped region with opposite edges identied. Dierent values of
in general dene inequivalent Riemann surfaces; is known as a modulus for the complex
structure on the torus. However, there are some equivalences: , , + 1, and 1/
generate the same group of transformations of the complex plane and so the same surface
(to see the last of these, let z = z

). So we may restrict to Im() > 0 and moreover

identify
+ 1
1

. (1.8.5)
These generate the modular group

=
a +b
c +d
, (1.8.6)
where now a, b, c, d are integers such that ad bc = 1. A fundamental region for this is
given by [[ 1, Re()
1
2
, shown in g. 6. This is the moduli space for the torus: every
35
z
2
2
a a'
b
b'
0
2(+1)
Figure 5: The torus by periodic identication of the complex plane. Points identied with
the origin are indicated. Edges a and a

are identied, as are edges b and b

1
2

I I'
II II'
i
1
2

Figure 6: The standard fundamental region for the modulus of the torus. Identifying
boundaries I and I

, and II and II

, produces the moduli space for the torus. Note that this
is a closed space except for the limit Im() .
36
Riemann surface with this topology is equivalent to one with in this region. The torus also
has a conformal Killing transformation, z z +.
Notice the similarity between the transformations (1.8.3) and (1.8.6), diering only in
whether the parameters are complex numbers or integers. You can check that successive
transformations compose like matrix multiplication, so these are the groups SL(2, C) and
SL(2, Z) respectively (2 2 matrices of determinant one).
13
We will meet SL(2, Z) again in
a dierent physical context.
Any closed oriented oriented two-dimensional surface can be obtained by adding h handles
to the sphere; h is the genus. It is often useful to think of higher genus surfaces built up
from lower via the plumbing xture construction. This essential idea is developed in many
places, but my lectures have been most inuenced by the approach in refs [37]-[40]. Let z
(1)
and z
(2)
be coordinates in two patches, which may be on the same Riemann surface or on
dierent Riemann surfaces. For complex q, cut out the circles [z
(1)
[, [z
(2)
[ < (1 )[q[
1/2
and
identify points on the cut surfaces such that
z
(1)
z
(2)
= q, (1.8.7)
as shown in g. 7. If z
(1)
and z
(2)
are on the same surface, this adds a handle. The genus-h
surface can be constructed from the sphere by applying this h times. The number of complex
parameters in the construction is 3h, being q and the position of each end for each handle,
minus 3 from an overcounting due to the M obius group, leaving 3h 3 which is the correct
number of complex moduli. An index theorem states that the number of complex moduli
minus the number of conformal Killing transformations is 3h 3, as we indeed have in each
case.
Note that for q < 1 the region between the circles [z
(1)
[ = 1 and [z
(2)
[ = 1 is conformal
to the cylindrical region
0 < Im(w) < 2 ln(1/[q[), w

= w + 2, (1.8.8)
which becomes long in the limit q 0.
Just as conformal transformations can be described as the most general coordinate trans-
formation which leave dz invariant up to local multiplication, there is a geometric interpre-
tation for the superconformal transformations. The N = 1 algebra, for example, can be de-
scribed in terms of a space with two ordinary and two anticommuting coordinates (z, z, ,

)
13
To be entirely precise, ipping the signs of , , , or a, b, c, d gives the same transformation, so we
have SL(2, C)/Z
2
and SL(2, Z)/Z
2
respectively.
37
a)
b)
Figure 7: Plumbing xture construction. Identifying annular regions as in (a) produces the
sewn surface (b).
as the space of transformations which leave dz + id invariant up to local multiplication.
Super-Riemann surfaces can be dened as above by patching. The genus-h Riemann surface
for h 2 has 3h 3 commuting and 2h 2 anticommuting complex moduli.
1.9 CFT on Riemann Surfaces
On this large subject I will give here only a few examples and remarks that will be useful
later. A tensor O of weight (h,

h) transforms as O
(u)
= O
(z)
z
2h
z
2

h
from the z to u patch on
the sphere. It must be smooth at u = 0, so in the z frame we have
< O
z
(z, z) . . . >
S
2
z
2h
z
2

h
, z (1.9.1)
38
An expectation value which illustrates this, and will be useful later, is (all operators implicitly
in the z-frame unless noted)
<: e
ik
1
X(z
1
, z
1
)
: : e
ik
2
X(z
2
, z
2
)
: . . . : e
iknX(zn, zn)
:>
S
2
= 2(k
1
+. . . +k
n
)

i<j
[z
i
z
j
[
2k
i
k
j
, (1.9.2)
obtained but summing over all graphs with the propagator ln[z
i
z
j
[
2
. Using momentum
conservation one nds the appropriate behavior [z
i
[
2k
2
i
as z
i
.
Another example involves the bc system, specializing to the most important case = 2.
Consider an expectation value with some product of local operators, surrounded by a line
integral
1
2i
_
dzj
z
(1.9.3)
where j =: cb :. Contracting C down around the operators and using the OPE gives N
c
N
b
for the contour integral, counting the total net ghost number of the operators. We have
previously noted that j is not a tensor, and one nds that the contour integral above is equal
to
1
2i
_
duj
u
+ 3. (1.9.4)
Now, we can contract contour C to zero in the u patch, so it must be that N
c
N
b
= 3 for
a nonvanishing correlator.
Exercise: Using the construction of the genus h surface from the sphere vis h plumbing
xtures, show that N
c
N
b
must be 3h 3.
The simplest one is
< c(z
1
)c(z
2
)c(z
3
) c(z
4
) c(z
5
) c(z
6
) >
S
2
= z
12
z
13
z
23
z
45
z
46
z
56
, (1.9.5)
which is completely determined, except for normalization that we x by hand, by the re-
quirement that it be analytic, that it be odd under exchange of anticommuting elds, and
that it go as z
2
i
or z
2
i
at innity, c being weight (1, 0) and c being (0, 1).
Now something more abstract: consider the general two-point function
< /

i
(, )/
j
(0, 0) >
S
2
=
i
[
j
) = (
ij
(1.9.6)
where I use a slightly wrong notation: (, ) denotes the point z = (u = 0) but the
prime denote the u frame for the operator. Recall the state-operator mapping. The operator
/
j
(0, 0) is equivalent to removing the disk [z[ < 1 and inserting the state [
j
); the operator
39
/

i
(, ) is equivalent to removing the disk [u[ < 1 ([z[ > 1) and inserting the state [
j
).
All that is left of the sphere is the overlap of the two states. It is also useful to regard this
as a metric (
ij
on the space of operators, the Zamolodchikov metric. Note that the path
integral (1.9.6) does not include conjugation, so if there is a Hermitean inner product [ )
these must be related

i
[
j
) =

i
[
j
) , (1.9.7)
where

is some operation of conjugation. For the free scalar theory, whose Hermitean
inner product has already been given,

just takes k k and conjugates explicit complex
numbers.
Similarly for the three-point function, the operator product expansion plus the deni-
tion (1.9.6) give
< /

i
(, )/
k
(z, z)/
j
(0, 0) >
S
2
=
i
[/
k
(z, z)[
j
)
= z
h
l
h
k
h
j
z

h
l

h
k

h
j
(
il
c
l
kj
= z
h
l
h
k
h
j
z

h
l

h
k

h
j
c
ikj
. (1.9.8)
This relates the three-point expectation value on the sphere to a matrix element and then
to an OPE coecient.
The torus has a simple canonical interpretation: propagate a state forward by 2Im()
and spatially by 2Re() and then sum over all states, giving the partition function
< 1 >
T
2
()
= Tr
_
e
2i(L
0
c/24)
e
2i (

L
0
c/24)
_
, (1.9.9)
where the additive constant is as in eq. (1.4.10). This must be invariant under the modular
group. Let us point out just one interesting consequence [41]. In a unitary theory the
operator of lowest weight is the (0, 0) unit operator, so
< 1 >
T
2
()
e
(c+ c)Im()/12
, i. (1.9.10)
The modular transformation 1/ then gives
< 1 >
T
2
()
e
(c+ c)/12Im()
, i0. (1.9.11)
The latter partition function is dominated by the states of high weight and is a measure of
the density of these states. We see that this is governed by the central charge, generalizing
the result that c counts free scalars.
40
A
i
i
A

i
a)

i
b)
A
i
i
A
Figure 8: a) Path integral on sewn surface of g. 7 written in terms of a sum over intermediate
states. b) Each state replaced by disk with local operator.
We have described the general Riemann surface implicitly in terms of the plumbing
xture, and there is a corresponding construction for CFTs on the surface (again, I follow
refs. [18], [37]-[40]). Taking rst q = 1, sewing the path integrals together is equivalent to
inserting a complete set of states. As shown in g. 8, each can be replaced with a disk plus
vertex operator. Including the radial evolution for general q we have
< . . .
1
. . .
2
>
M
=

ij
q
h
i
q

h
i
< . . .
1
/
i
>
M
1
< . . .
2
/
i
>
M
2
, (1.9.12)
where / is sewn from /
1,2
, the operators are inserted at the origins of the z
(1,2)
frames,
and indices are raised with the inverse of the metric (1.9.6).
14
14
The one thing which is not obvious here is the metric to use. You can check the result by applying it to
41
a) b)
c)
Figure 9: Some sewing constructions of the genus-two surface with four operators. Each
vertex is a sphere with three operators, and each internal line represents the sewing con-
struction.
By sewing in this way, an expectation value with any number of operators on a general
genus surface can be related to the three-point function on the sphere. For example, g. 9
shows three of the many ways to construct the genus-two surface with four operators. There
is one complex modulus q for each handle, or 7 in all here, corresponding to the 3h 3 = 3
moduli for the surface plus the positions of the four operators. As a consequence, the OPE
coecients c
ij
k
implicitly determine all expectation values, and two CFTs with the same
OPE (and same operator identied as T(z)) are the same. However, the c
ij
k
are not arbitrary
because the various methods of constructing a given surface must agree. For example, the
constructions of g. 9a and 9b dier only by a single move described earlier corresponding to
associativity of the OPE, and by further associativity moves one gets g. 9c. The amplitudes
must also be modular invariant. In g. 9c we see that the amplitude has been factorized into
a tree amplitude times one-loop one-point amplitudes, so modular invariance of the latter is
sew two spheres together, with . . .
1
and . . .
2
each being a single local operator, to get the sphere with two
local operators.
42
sucient. It can be shown generally that all constructions agree and are modular invariant
given two conditions [40]: associativity of the OPE and modular invariance of the torus with
one local operator (which constrains sums involving c
ij
j
).
The classication of all CFTs can thus be reduced to the algebraic problem of nding
all sets c
ij
k
satisfying the constraints of conformal invariance plus these two conditions. This
program, the conformal bootstrap [18], has been carried out only for cases where conformal
invariance (or some extension thereof) is sucient to reduce the number of independent c
ij
k
to a nite numberthese are known as rational conformal eld theories.
My description of higher-genus surfaces and the CFTs on them has been rather implicit,
using the sewing construction. This is well-suited for my purpose, which is to understand
the general properties of amplitudes. For treatments from a more explicit point of view see
refs. [42], [43].
Unoriented surfaces, and surfaces with boundary, are also of interest. In particular,
CFTs with boundary have many interesting condensed matter applications. I do not have
time for a detailed discussion, but will make a few comments about boundaries. Taking
coordinates such that the boundary is Im(z) = 0 and the interior is Im(z) > 0, the condition
that the energy-momentum be conserved at the boundary is
T(z) =

T(z), Im(z) = 0. (1.9.13)
It is convenient to use the doubling trick, extending T into the lower half-plane by dening
T(z) =

T( z), Im(z) < 0. (1.9.14)
The boundary conditions plus conservation of T and

T are all implied by the analyticity
of the extended T. Then T and

T together can be expanded in terms of a single Virasoro
algebra, the rst of equations (1.4.1). The doubling trick is also useful for free elds. For
free scalar with Neumann boundary conditions,
n
X = 0, the mode expansion is
X(z, z) = X
L
(z) +X
L
( z), (1.9.15)
with
X
L
(z) =
x
2
+i
0
lnz i

m=0

m
mz
m
. (1.9.16)
There is a factor of 2 dierence from the earlier (1.4.23); with this, x is the mean value of
X at time [z[ = 1. The commutator [x,
0
] is found to be 2i, so
0
= 2k where k is the
conjugate to x. The leading surfaces with boundary are the disk and annulus.
43
a) b) c)
Figure 10: a) Two particles propagating freely. b) Correction from one-graviton exchange.
c) Correction from two-graviton exchange.
2 String Theory
2.1 Why Strings?
The main clue that leads us to string theory is the short-distance problem of quantum gravity.
Figure 10 shows some process, say two particles propagating, and corrections due to one-
graviton exchange and two-graviton exchange. The one graviton exchange is proportional
to Newtons constant G
N
, which with h = c = 1 has units of length
2
or mass
2
: G
N
=
M
2
P
where the Planck mass M
P
= 1.2 10
19
GeV. The dimensionless ratio of the one-
graviton correction to the original amplitude must then be of order E
2
/M
2
P
, where E is the
characteristic energy of the process. This is thus an irrelevant coupling, growing weaker
at long distance, and in particular is negligible at particle physics energies of hundreds of
GeV. By the same token, the coupling grows stronger at high energy and at E > M
P
perturbation theory breaks down. This shows up as the nonrenormalizability of the theory:
the two-graviton correction (c) is of order
M
4
P
_

dE

E
3
(2.1.1)
where E

is the energy of the virtual intermediate state, and so diverges if the theory is
extrapolated to arbitrarily high energies.
There are two main possibilities. The rst is that the theory has a nontrivial ultraviolet
xed point and is ne at high energy, the divergences being an artifact of naive perturbation
44
theory. The second is that there is new physics at some energy and the extrapolation of the
low energy theory beyond this point is invalid.
The existence of a nontrivial xed point is hard to determine. One of the usual tools,
Monte Carlo simulation, is extremely dicult because of the need to retain coordinate in-
variance in the discretized theory. Expansion around the critical dimension d = 2 indicates
a nontrivial UV xed point when gravity is coupled to certain kinds of matter, but it is
impossible to say whether this persists to d = 4.
1
The more common expectation, based in part on experience (such as the weak inter-
action), is that the nonrenormalizability indicates a breakdown of the theory, and that at
short distances we will nd a new theory in which the interaction is spread out in space-
time in some way that cuts o the divergence. At this point the condensed matter half of
the audience is thinking, OK, so put the thing on a lattice. But it is not so easy. We
know that Lorentz invariance holds to very good approximation in the low energy theory,
and that means that if we spread the interaction in space we spread it in time as well,
with consequent loss of causality or unitarity. Moreover we know that we have local coordi-
nate invariance in naturethis makes it even harder to spread the interaction out without
producing inconsistencies.
In fact, we know of only one way to spread out the gravitational interaction and cut o
the divergence without spoiling the consistency of the theory. That way is string theory, in
which the graviton and all other elementary particles are one-dimensional objects, strings,
rather than points as in quantum eld theory. Why this should work and not anything
else is not at all obvious a priori, but as we develop the theory we will see that if we try
to make a consistent Lorentz-invariant quantum theory of strings we are led inevitably to
include gravity [45], [46], and that the short distance divergences of eld theory are no longer
present.
2
Perhaps we merely suer from a lack of imagination, and there are many other consistent
theories of gravity with a short-distance cuto. But experience has shown that the divergence
problems of quantum eld theory are not easily resolved, so if we have even one solution we
1
The idea that the divergence problems of quantum gravity might be solved by a resummation of pertur-
bation theory has been examined from many points of view, but let me mention in particular the approach
ref. [44] as one that will be familiar to the condensed matter audience.
2
There is an intuitive answer to at least one common question: why not membranes, two- or higher-
dimensional objects? The answer is that as we spread out particles in more dimensions we reduce the
spacetime divergences, but encounter new divergences coming from the increased number of internal degrees
of freedom. One dimension appears to be the one case where both the spacetime and internal divergences
are under control. But this is far from conclusive: just as pointlike theories of gravity are still under study,
so are membrane theories, as we will mention in section 3.5.
45
should take it very seriously. In the case of the weak interaction, for example, there is only
one known way to spread out the nonrenormalizable four-fermi theory consistently.
3
That
way is spontaneously broken Yang-Mills theory, which did indeed turn out to be the correct
theory of the weak interaction. Indeed, we are very fortunate that consistency turns out to
be such a restrictive principle, since the unication of gravity with the other interactions
takes place at such high energy, M
P
, that experimental tests will be dicult and indirect.
So what else do we nd, if we pursue this idea? We nd that string theory ts very nicely
into the pre-existing picture of what physics beyond the Standard Model might look like.
Besides gravity, string theory necessarily incorporates a number of previous unifying ideas
(though sometimes in transmuted form): grand unication, Kaluza-Klein theory (unication
via extra dimensions), supersymmetry and extended supersymmetry. Moreover it unies
these ideas in an elegant way, and resolves some of the problems which previously arose
most notably diculties of obtaining chiral (parity-violating) gauge interactions and the
renormalizability problem of Kaluza-Klein theory, which is even more severe than for four-
dimensional gravity. Further, some of the simplest string theories [47] give rise to precisely
the gauge groups and matter representations which previously arose in grand unication.
Finally, the whole subject has a unity and structure far nicer than anything I have seen or
expect to see in quantum eld theory. So I am strongly of the opinion, and I think that
almost all of those who have worked in the subject would agree, that string theory is at least
a step toward the unication of gravity, quantum mechanics, and particle physics.
In this lecture and the next I will try to sketch our current understanding of the answer
to the question posed in the title. Given limits of time and the nature of the audience, I
will focus on broad dynamical issues, especially the mechanics by which string theory cuts
o gravity in a consistent way. Most notably, spacetime supersymmetry and the superstring
will be underemphasized.
There is one graph, g. 11, that I want to show you before I launch into the introduction
to string theory. It shows how the three dimensionless gauge couplings and the dimensionless
gravitational coupling E
2
/M
2
P
depend on energy. The gauge couplings evolve slowly (loga-
rithmically); the big news of recent years is that in the minimal supersymmetric extension
of the Standard Model they meet to high accuracy at a common scale, of order 10
16
GeV,
3
During the lecture, Prof. Zinn-Justin reminded me that there is evidence from the large-N approximation
that some four-fermi theories have nontrivial xed points. So perhaps there is more than one way to smooth
the weak interactionbut perhaps also we should take this as an indication that, given a choice between
new physics and a nontrivial ultraviolet xed point, nature will choose the former. At any right, given my
understanding of renormalizable eld as an eective theory that emerges at long distance, I would nd the
xed point resolution very unappealing.
46
10 10 10 10 10 10
3 6 9 12 15 18
10
10
10
10
0
-2
-32
-34
GeV

1
2
3
G

G
Figure 11: Energy dependence of dimensionless gauge couplings
1,2,3
(running logarithmi-
cally) and
G
= E
2
/M
2
P
.
giving evidence for supersymmetric grand unication [48]. The gravitational coupling starts
much smaller but grows as a power and so is just a bit late for its meeting with the oth-
ers, missing by two orders of magnitude or perhaps a little less. These extrapolations are
sensitive to assumptions about the spectrum, so perhaps all four couplings meet at a single
energy, a very grand unication. Or perhaps there is a small hierarchy of scales near the
Planck scale. But the near meeting in this minimal extrapolation suggests that nature may
have been kind and put little new physics between current energies and the Planck scale.
With the thorough exploration of the weak interaction scale in coming years, and hopefully
the discovery of supersymmetry, we will have several additional extrapolations of the same
sort and so several handles on physics near the Planck scale. Also, proton decay reaches into
the same region, and if we are lucky it will occur at a rate that will one day be seen.
47
2.2 String Basics
We want to describe the dynamics of one-dimensional objects. The rst thing we need is an
action, and the simplest that comes to mind is the Nambu-Goto action,
4
S =
1
2

(Area of world-sheet)
=
1
2

_
d
2

_
det
a
X

b
X

. (2.2.1)
This generalizes the relativistic action for a point particle, which is minus the mass times
the invariant length of the world-line. For a static string, this action reduces to minus
the length of the string times the time interval times 1/2

, so the latter is the string

tension. Note that in the second line we are describing the world-sheet by X

(
1
,
2
), using
a parameterization
a
of the world-sheet, but the action is independent of the choice of
parameterization (world-sheet coordinate invariant). This will play an important role soon.
In quantum eld theory we are familiar with a variety of one-dimensional objects
magnetic ux tubes in superconductors and other spontaneously broken gauge theories,
color-electric ux tubes in QCD. Also, the classical statistical mechanics of membranes is
given by a sum over two-dimensional surfaces, and so is closely related to the quantum-
mechanical path integral for the string. In all of these cases the leading term in the action is
the tension (2.2.1). But these are all composite objects, with a thickness, and so there will
be higher-dimension terms in the action, such as a rigidity term, multiplied by powers of the
thickness. The strings I am talking about, the fundamental strings which give rise to gravity,
are exactly one-dimensional objects, of zero thickness. Composite strings also have a large
contact interaction when they intersect; fundamental strings do not. Fundamental strings are
thus simpler than the various composite strings, simpler in particular than the hypothetical
string theory of QCD.
5
Nevertheless there has been a great deal of cross-fertilization between
the theories of fundamental and composite strings.
It is useful to rewrite the action (2.2.1) in a form which removes the square root from
the derivatives. Add a world-sheet metric g
ab
() and let
S
P
=
1
4

_
d
2

gg
ab

a
X

b
X

, (2.2.2)
4
Indices , = 0, 1, . . . D1 are raised and lowered with the at-space metric

= diag(, +, +, . . . , +).
5
In the coming sections I will discuss ideas that strings are in some sense composite, but not composites
of ordinary gauge elds.
48
where g = det g
ab
. This is commonly known as the Polyakov action because he emphasized
its virtues for quantization [49]. The equation of motion for the metric determines it up to
a position-dependent normalization
g
ab

a
X

b
X

; (2.2.3)
inserting this back into the Polyakov action gives the Nambu action.
6
Now, the Polyakov
action makes sense for either a Lorentzian metric, signature (, +), or a Euclidean metric,
signature (+, +).
7
Much of the development can be carried out in either case. These are
presumably related by a contour rotation in the integration over metrics, since the light-cone
quantization (Lorentzian) gives the same theory as the Euclidean Polyakov quantization that
I will describe. The relation between path integrals over Lorentzian and Euclidean metrics is
a complicated and confusing issue in four-dimensional gravity. It seems to work out simply
in two dimensions, though I dont have a simple explanation of whythe demonstration
of the equivalence is rather roundabout [51]. Perhaps it is simply that there is enough
gauge symmetry to remove the metric entirely.
8
In any case I will take a Euclidean metric
henceforth as dening the theory.
In addition to the two-dimensional coordinate invariance mentioned earlier (di invari-
ance for short),
X

) = X(),

a

d
g

ab
(

) = g
cd
() , (2.2.4)
the Polyakov action has another local symmetry, Weyl invariance, position-dependent rescal-
ings of the metric,
g

ab
() = e
2()
g
ab
(). (2.2.5)
To proceed with the quantization we need to remove the redundancy from the local
symmetries, to x the gauge. Noting that the metric has three components and there are
three local symmetries (two coordinates and the scale of the metric), it is natural to do this
by conditions on the metric, setting
g
ab
() =
ab
. (2.2.6)
6
So these are classically equivalent. How about quantum-mechanically, say in a path integral? The glib
answer is that the Nambu action is hard to use in a path integral, so the way to dene it is via the Polyakov
path integral. On the the other hand, ref. [50] shows an example of a composite string where the Nambu
description is more natural than the Polyakov.
7
Though the Lorentzian case needs an overall minus sign and one in the square root.
8
I would like to thank M. Natsuume for this suggestion.
49
This is always possible at least locally. The Polyakov action (2.2.2) then reduces to D copies
of the earlier scalar action (1.1.1),
9
provided that we choose units such that

, which has
units of length-squared, is equal to 2.
It is not an accident that the gauge-xed action is conformally invariant. Fixing the
at metric does not fully determine the local coordinate system. From the discussion of
conformal invariance we know that if two coordinate systems are related by

1
+i
2
= f(
1
+i
2
) (2.2.7)
for analytic f, the metric changes only by a position-dependent rescaling, so that a Weyl
transformation restores it to its original form. In other words, the coordinate transforma-
tion (2.2.7) combined with the appropriate Weyl transformation leaves the metric in at
gauge and so is a conformal symmetry of the at world-sheet action.
10
So the two-dimensional spacetime of the previous section is now the string world-sheet,
while spacetime is the eld space where the X

live, the target space of the map X

:
world-sheet spacetime.
2.3 The Spectrum
For a closed string, where the spatial coordinate
1
is periodic, we can immediately use the
earlier results to write down the spectrum. We have D sets of harmonic oscillators,
[

m
,

n
] = m
m+n,0

, [

m
,

n
] = m
m+n,0

, (2.3.1)
the covariant generalization of the earlier commutator, as well as the D momenta

0
=

0
=
k

. Starting from the states [0, k) which are annihilated by the m < 0 modes, we build the
spectrum by acting any number of times with the m > 0 modes. By choice of conformal
gauge, the string thus separates into a superposition of harmonic oscillators.
9
To be precise, because of the Minkowski signature, the action for X
0
has the opposite sign and gives a
divergent gaussian path integral. This is not problem; the path integral is implicitly dened by the Euclidean
rotation X
0
iX
D
. This is similar to the treatment of Grassman path integralswe dont have to take
them seriously as integrals, as long as they have certain key properties, most notably factorization (so we
can cut them open to get a Hamiltonian formalism) and the integral of a derivative vanishing (so we can
derive equations of motion).
10
Of course there will be some global conditions that x most or all of this residual invariance, as we will
discuss further later. This is not relevant now: we noted earlier that to derive Noethers theorem and the
Ward identities we only need a symmetry transformation to be dened in a region.
50
But there is one more point to deal with. The index on the oscillators (2.3.1) runs
over D values. A string stretched out in the X
1
-direction should by able to oscillation in the
D 2 transverse directions = 2, . . . , D 1, but oscillation along the X
1
-direction leaves
the world-sheet unchangedaccording to the earlier discussion it is just an oscillation of the
parameterization . The same is true of oscillation in the X
0
-direction. It is essential that
this be true, because from the oscillator algebra we have the inner product
0, k[

1
[0, k

) = (2)
D

D
(k k

. (2.3.2)
The timelike oscillation

0
[0, k

) thus has a negative norm and so had better not be in the

Hilbert space of the theory.
The point is that when we x g
ab
we lose the equations of motion we get from varying
g
ab
, and we have to restore them as constraints. Varying the metric gives
T
ab
= 0. (2.3.3)
In fact T
z z
vanishes as a consequence of the X

equation of motion, but T

zz
and T
z z
do not.
The equation of motion does imply that if they vanish at one time they vanish for all times;
this is a general feature of such missing equations of motion. Classically, then, we impose
these equations on the initial values; quantum mechanically we impose them on the states.
In either case the constraint is then preserved by the dynamics.
Later we will discuss a general and powerful way to implement the constraints, the
BRST quantization, but it is useful to proceed rst by a bit of trial and error (so-called old
covariant quantization [52]). Going to the Laurent modes, we could try to impose L
n
[) = 0
for all n. But this is inconsistent with the Virasoro algebra, since it would imply that
0 = [L
m
, L
m
][) =
c
12
(m
3
m)[). Instead we require physical states to satisfy
(L
0
a)[) = (

L
0
a)[) = 0
L
n
[) =

L
n
[) = 0, n > 0 , (2.3.4)
allowing a possible ordering constant in the L
0
condition, which will turn out to be necessary.
This implies that matrix elements of (2.3.3) between physical states vanish for all n,
[(L
n
a
n,0
)[

) = [(

L
n
a
n,0
)[

) = 0, (2.3.5)
the n < 0 generators annihilating the bra since L
n
= L

n
.
51
There is one more provision. A state of the form
L
n
[) +

L
n
[ ), n > 0 (2.3.6)
for any [), [ ) is orthogonal to all physical states and so is called spurious. A physical state
which is also spurious is called null. All physical amplitudes involving such a state vanish, so
it is physically equivalent to the zero state. Thus we dene an equivalence relation between
physical states
[)

= [

) if [

) [) = L
n
[) +

L
n
[ ). (2.3.7)
The observable Hilbert space is the set of equivalence classes, physical states modulo null
states.
Let us see how this works for some of the lowest levels. We focus on the open string
because it has only one set of modes. At the rst level are the states [0, k), with all internal
oscillators in their ground states. The physical state conditions (2.3.4) for n > 0 all involve
lowering operators and so hold. There remains
0 = (L
0
a)[0, k) = (2k
2
a)[0, k). (2.3.8)
Thus we obtain a mass shell condition,
M
2
= k
2
=
a
2

a
2
2

=
a

. (2.3.9)
where we have restored

by dimensional analysis.
At the next level the states are
[e, k) = e

1
[0, k). (2.3.10)
for some polarization vector e

. The nontrivial physical state conditions are

0 = (L
0
a)[e, k) = (2k
2
+
1

1
a)[e, k) = (2k
2
+ 1 a)[e, k)
0 = L
1
[e, k) = 2k
1
[e, k) = 2k e[0, k). (2.3.11)
These give a mass shell condition and a transversality condition on e,
M
2
= k
2
=
1 a

, k e = 0. (2.3.12)
There is a spurious state,
L
1
[0, k) = 2k
1
[0, k). (2.3.13)
52
Thus [e, k) is spurious for e

. According to the conditions (2.3.12), this is physical,

and so null, only if k k = 0, which is the case only if a = 1.
There are now three cases:
i) If a < 1, the mass-squared is positive. Going to the rest frame, k

= (m, 0, 0, . . . , 0), the

physical state condition k e = 0 removes the negative norm timelike polarization and leaves
the D1 spacelike polarizations. There are no null states, and the spectrum consists of the
D 1 positive-norm states of a massive vector particle.
ii) If a = 1, the mass-squared is zero. Going to a frame in which k

= (, , 0, 0, . . . , 0), the
physical states are e

plus the D 2 transverse polarizations = 2, . . . , D 1. The

state e

is null, leaving the D 2 positive-norm transverse states of a massless vector

particle.
iii) If a > 1, the mass-squared is negative and we can go to a frame k

= (0, k
1
, 0, . . . , 0).
The physical state condition removes a positive-norm spacelike polarization. There is no null
state, so we are left with D2 positive-norm spacelike polarizations and one negative-norm
timelike polarization.
Case (iii) is obviously unacceptable, but either (i) or (ii) seem satisfactory so far. It is
case (ii) that agrees with the BRST quantization, and also with the light-cone quantization,
a dierent gauge in which the number of oscillators is reduced to D 24 from the start. In
fact, there is no consistent way known to introduce interactions in case (i).
The result at the next level, states
f

2
[0, k) +f

1
[0, k), (2.3.14)
is quite interesting. It depends on the constant a and also on the spacetime dimension D.
If a > 1 or D > 26, there are negative-norm states. If a 1 and D 26 the OCQ spectrum
has positive norm. For the particular case a = 1 and D = 26 the observable spectrum is the
same as the BRST and light cone spectra; otherwise there are extra states. The derivation
is left to the reader.
This pattern persists at all higher levels as well: the observable spectrum has only positive
norm states provided a 1 and D 26, and if a = 1 and D = 26 it is the same as the
BRST and light cone spectra. This is the no-ghost theorem for the spectrum. This is our
rst encounter with the critical dimension; later we will understand it more deeply.
For the value a = 1, the second level is a massless vector particle, a gauge boson. This
implies a spacetime gauge symmetry. In amplitudes this symmetry appears as spacetime (not
53
world-sheet) Ward identities, to the eect that the unphysical polarization is not produced,
and the amplitudes for equivalent polarizations e

= e

+ck

are equal. We will see how this

works when we discuss interactions. Since null and unphysical states appear at all levels, this
means that ordinary gauge symmetry is only one piece of some much large gauge symmetry
in string theory. We will discuss the form of this later.
Notice also that for a = 1 the lowest state (2.3.9) is a tachyon, M
2
= 1/

. Since the
potential for a scalar eld is
1
2
M
2

2
, this means that the ground state is unstable. (One way
to think about this is as resulting from the negative Casimir energy of the X

s.) We are
using the bosonic string only as a toy model, and are indeed expanding around an unstable
state. The superstring does not have a tachyon, though unfortunately time will not permit
me in these lectures to give a detailed treatment of this.
The closed string spectrum is just the tensor product of two copies of the above, one
right-moving and one left-moving. The lowest state [0, k) satises
11
M
2
=
4a

. (2.3.15)
The next level
[e, k) = e

1
[0, k) (2.3.16)
satises
M
2
=
4(1 a)

, k

= k

= 0. (2.3.17)
The correct values are again a = 1, D = 26, so these states are massless, and there are null
states leading to the equivalence relation
e

+k

(2.3.18)
with k = k

= 0. Again letting k

= (, , 0, 0, . . . , 0), a complete set of (D 2)

2
observable states is obtained from transverse e

, , 2, . . . , D 1. This set can be de-

composed under the SO(D2) transverse rotation group, into a traceless symmetric tensor,
antisymmetric tensor and invariant. These are respectively the graviton, antisymmetric ten-
sor, and dilaton. Again there are spacetime gauge invariances associated with the graviton
and antisymmetric tensor.
11
I am using the same notation for open and closed string states. It should always be clear which is meant.
54
2.4 The Weyl Anomaly
The fact that strings can be consistently quantized only in the critical dimension is due to
an anomaly, a quantum violation of the local world-sheet symmetries. To see the anomaly,
we will work in a more general gauge in which the gauge symmetry is used to x the metric
to some general form g
ab
(), not necessarily at. The result should be independent of what
g
ab
() we choose; let us see if it is. I should warn you that the next lecture or so will get
steadily more technical, but then things will get better again. I will try to highlight the main
results that we will need later on.
We start by examining the path integral over X

in the xed metric g

ab
. The action
is di Weyl invariant, but we have to dene the path integration. It is easy to preserve
the di invariance. For example, expand X

() in a complete set of eigenfunctions of the

invariant Laplacian
2
, and put a cuto on the eigenvalues. Any cuto will refer to world-
sheet lengths, so the Weyl invariance is not automatically preserved and there may be a
Weyl anomaly [53, 49]. We must check the Weyl invariance by explicit calculation. This
is one virtue of the Polyakov actionany possible anomaly appears in the Weyl symmetry,
which is somewhat easier to work with than an anomaly in the di invariance.
Let us expand around the at metric in the plane, g
ab
() =
ab
+h
ab
(). The variation
of the action with respect to the metric is the energy-momentum tensor, so to second order
in h
z z
the path integral is
< 1 >
O(h
2
)
=
1
8
2
_
d
2
z d
2
z

h
z z
(z, z)h
z z
(z

, z

) < T
zz
(z)T
zz
(z

) > + local. (2.4.1)

The local term comes from the second order variation of the action; we will not need its
explicit form. The expectation value is evaluated on the at world-sheet, and from the OPE
we know it to be
< T
zz
(z)T
zz
(z

) > =
c
2(z z

)
4
=
c
12

4
z
ln[zz

[
2
=
c
6

4
z
(
z

z
)
1

2
(zz

, z z

), (2.4.2)
where c = D for the X theory. Thus the second order term (2.4.1) becomes
< 1 >
O(h
2
)
=
c
48
_
d
2
z d
2
z

2
h
z z
(z, z)
1

2
h
z z
(z

, z

c
96
_
d
2
d
2

gR()
1

2
(,

gR(

) . (2.4.3)
55
The second line, using the curvature scalar built from g
ab
, is the unique coordinate invariant
form to order h
2
with the given nonlocal (z z

)
4
term. Now, under a Weyl transformation
g

ab
= e
2
g
ab
, the curvature changes
_
g

gR 2
2
. (2.4.4)
Thus

Weyl
< 1 >
O(h
2
)
=
c
24
< 1 >
_
d
2

gR. (2.4.5)
Although derived to second order in the background, the result (2.4.5) is in fact the full
answer. Given that the Weyl variation must be local (since it comes from the violation of the
symmetry by the cuto) and that the theory is conformally invariant on the at world-sheet,
eq. (2.4.5) is the only possible form.
12
This goes through for any CFT, so we have the result
that if we put a CFT on a curved world-sheet the Weyl-dependence is determined entirely
by the central charge of the at world-sheet theory,

Weyl
ln < . . . > =
c
24
_
d
2

gR, (2.4.6)
where as usual the transformation properties of the insertions . . . are a separate issue. Also,
we can integrate the Weyl anomaly to give the full result
13
< 1 > = exp
_
c
96
_ _
d d

R
2
R
_
. (2.4.7)
To complete the determination of the Weyl anomaly we need to carry out the gauge
xing carefully, taking into account the Fadeev-Popov determinant. That is, we write the
integral over metrics as in integral over the gauge group times a Jacobian, and divide by the
gauge volume. Under a small di Weyl transformation, the change in the metric is
g
ab
= (
a

b
+
b

a
g
ab
) g
ab
(2 + )
= (P
1
)
ab
2g
ab

. (2.4.8)
12
For example, terms with more derivatives would be suppressed by powers of the cuto. A possible Weyl
variation
_
d
2

g can be removed by a counterterm

_
d
2

g in the action.
13
But what if we have a CFT with c = c? Then we cannot extend the result as in the second line of
eq. (2.4.3) to a di invariant: the h
2
zz
and h
2
z z
terms are inconsistent. That is, there is in this case an
anomaly in the two-dimensional coordinate invariance [54]. The necessary and sucient condition for this
to be absent is c = c.
56
In the rst line we have separated the variation into a traceless part and a part proportional
to the metric. In the second line we have dened the dierential operator P
1
taking vectors
to traceless symmetric tensors, and absorbed the term in a shift of . Short circuiting
some formal steps that are parallel to the standard treatment of non-Abelian gauge theories,
we get
_
[dg] det(P
1
)V
diWeyl
. (2.4.9)
The

part being purely local gives a trivial Jacobian, equivalent to local counterterms.
Thus,
1
V
diWeyl
_
[dX dg] e
S
P
[g]
= det(P
1
)
_
[dX] e
S
P
. (2.4.10)
In order to expose the general structure of amplitudes it is useful to rewrite the determi-
nant as a path integral over anticommuting Fadeev-Popov ghost elds. From the denition
of Grassman integration we have
_
db dc e
bMc
=
_
db dc (1 bMc) = M (2.4.11)
for two anticommuting variables b, c. This generalizes to
_

i
(db
i
dc
i
) e
b
i
M
ij
c
j
= det M , (2.4.12)
as is evident by diagonalizing M. This applies to functional determinants as well, so the
path integral (2.4.10) becomes
_
[dX db dc] e
S
P
Sg
(2.4.13)
where
S
g
=
1
2
_
d
2

g b
ab
(P
1
c)
ab
. (2.4.14)
To see the conformal transformation properties, consider a conformally at metric
g
ab
() = e
()

ab
. (2.4.15)
The action (2.4.14) becomes
S
g
=
1
2
_
d
2
z
_
b
zz

z
c
z
+b
z z

z
c
z
_
. (2.4.16)
57
We have used a convenient trick here. The metric (2.4.15) is not at, so the covariant
derivatives are nontrivial. But for the special case of a z derivative acting on a tensor with
only z indices, or vice versa, the covariant derivatives reduce to the ordinary ones. We have
raised and lowered indices so as to take advantage of this. In the form (2.4.16) it is evident
that the action is Weyl-invariant with c
z
, c
z
, b
zz
, and b
z z
neutral under the Weyl transform.
The conformal transformation of these elds comes only from the coordinate transformation,
so c
z
is a (1, 0) tensor, c
z
a (0, 1) tensor, b
zz
a (2, 0) tensor, and b
z z
a (0, 2) tensor.
The ghosts thus are a = 2 bc system. Referring back to the earlier result (1.6.13), the
ghost system has central charge 26. As we have seen, this determines the Weyl transfor-
mation properties, so for the combined X and ghost system we have

Weyl
ln < . . . > =
D 26
24
_
d
2

gR. (2.4.17)
So the Weyl invariance is anomalous except in the critical dimension D = 26 [49]!
Is there any possibility for making sense of the theory without Weyl invariance? One
possibility is to make some choice for the scale factor of the metric, but there is no natural
way to do this. For example, the rather natural choice

g = 1 is not di invariant and just
moves the anomaly into that symmetry. The other possibility is to integrate over the scale
factorthat is, to treat the theory as one that has only coordinate invariance, and use this
to x only two of the three components of the metric [49]. This is a rather large change in
the theory, introducing a new degree of freedom. Such non-critical string theories are of
great interest, and we will return to them later.
For now, let us just point out the following. We can use the coordinate invariance to
bring the metric to the form
g
ab
() = e
()
g
ab
(), (2.4.18)
where () is to be integrated. Now, this theory has a fake Weyl invariance, under which
g
ab
is neutral but
g
ab
() e
2()
g
ab
(), () () 2(). (2.4.19)
But this is indistinguishable from a theory in which g
ab
is the real metric and is another
eld. So we can regard this as a Weyl-invariant theory with an extra degree of freedom.
Strominger and Verlinde will be interested in two-dimensional coordinate invariance without
Weyl invariance, and will make use of this trick to use techniques from CFT. It is interesting
58
to apply this to a theory whose Weyl-variation comes only from the central charge. Inserting
the metric (2.4.18) into the action (2.4.3) and using the relation (2.4.4) gives the -dependent
c
96
_
d
2

_
g
_
g
ab

b
+ 2

R
_
. (2.4.20)
The rst term is the same as the action for one of the coordinates X, so this looks like an
extra dimension. The second term is not translationally invariant, so it means that this
theory may have some cosmological interpretation, but is not relevant to the translationally
invariant surroundings that we nd ourselves in.
It may seem odd that a theory without Weyl invariance can be regarded as having it.
This is the rst example of a general theme which will arise again, both on the world-sheet
and in spacetime. This is that gauge invariance is in the end just a redundancy, though a
useful one. We can always add redundant elds and sometimes it is useful to do so, as here
where it will enable us to apply critical string methods to the noncritical string.
2.5 BRST Quantization
Representing the Fadeev-Popov determinant in terms of ghosts gives us an even larger Hilbert
space, and we need to identify the observable states. This leads us to BRST quantization,
which is a general method for quantizing systems with gauge symmetries and is indispensable
for understanding the general structure of string amplitudes.
14
One way to motivate it is to
imagine some arbitrary small change in the gauge-xing condition. Thus far we have taken
a gauge in which g
ab
() was xed. That is, g
ab
() g
ab
() = 0 for some xed function
g
ab
(). In the previous section we in eect checked that the result was invariant under a
Weyl transformation of g
ab
. But we could imagine a much more general change of gauge
g
ab
() g
ab
() F
ab
(g, X) = 0 . (2.5.1)
For example, the manifestly unitary light-cone gauge places conditions on X as well as
g
ab
; to interpolate between the conformal and light-cone gauges one would have to consider
deformations of the form (2.5.1).
In order to derive the full invariance condition, it is useful to take a more general and
abstract point of view. Consider a path integral with a local symmetry. The path integral
14
My treatment is similar to that in many modern eld theory texts. For other points of view, see
refs. [55], [56].
59
elds are denoted
i
, which in the present case would be X

() and g
ab
(). Here we use
a very condensed notation where i labels the eld, the component, and also the coordinate
. The gauge invariance is

, where again labels component and also coordinate.

By assumption the gauge parameters

are real, since we can alway separate a complex

parameter into its real and imaginary parts. The gauge transformations satisfy an algebra
15
[

] = f

. (2.5.2)
Now x the gauge by conditions
F
A
() = 0, (2.5.3)
where again A includes the coordinate. Following the usual Fadeev-Popov procedure, the
path integral becomes
_
[d
i
]
V
gauge
e
S
1

_
[d
i
dB
A
db
A
dc

] e
S
1
S
2
S
3
, (2.5.4)
where S
1
is the original gauge invariant action, S
2
is the gauge-xing action
S
2
= iB
A
F
A
(), (2.5.5)
and S
3
is the Fadeev-Popov action
S
3
= b
A
c

F
A
(). (2.5.6)
We have introduced the eld B
A
to produce an integral representation of the gauge-xing
(F
A
).
There are two things to notice about this action. The rst is that it is invariant under
the Becchi-Rouet-Stora-Tyupin (BRST) transformation [57],

i
= ic

i
,
B
B
A
= 0,

B
b
A
= B
A
,
B
c

=
i
2
c

. (2.5.7)
Note that the transformation mixes commuting and anticommuting objects, so that must
be taken to be anticommuting. The original action S
1
is invariant by itself, because the
action of
B
on
i
is just a gauge transformation with parameter c

. The variation of S
2
15
This is not the most general gauge symmetry possible, but is sucient for our application to the string.
60
cancels the variation of b
A
in S
3
, while the variations of

F
A
and c

in S
3
cancel. The
second key property is that

B
(b
A
F
A
) = i(S
2
+S
3
). (2.5.8)
Now consider a small local change F in the gauge-xing condition. The change in the
gauge-xing and ghost actions gives
[

) = i[
B
(b
A
F
A
)[

) = [Q
B
, b
A
F
A
[

), (2.5.9)
where we have written the BRST variation as an anticommutator with the corresponding
conserved charge Q
B
. For physical states the amplitude must be independent of the gauge
condition. In order that this hold for arbitrary F, it must be that
Q
B
[) = Q
B
[

) = 0, (2.5.10)
with Q

B
= Q
B
. This is the essential condition: physical states must be BRST invariant.
There is one more key idea. In order to move around in the space of gauge choices,
the BRST charge must remain conserved. Thus it must commute with the change in the
Hamiltonian,
0 = [Q
B
, Q
B
, b
A
F
A
]
= Q
2
B
b
A
F
A
Q
B
b
A
F
A
Q
B
+Q
B
b
A
F
A
Q
B
b
A
F
A
Q
2
B
= [Q
2
B
, b
A
F
A
]. (2.5.11)
In order for this to vanish for general changes of gauge, we need
Q
2
B
= 0. (2.5.12)
That is, the BRST charge is nilpotent. You can check that acting twice with the BRST
transformation (2.5.7) does indeed leave all elds invariant.
The nilpotence of Q
B
has an important consequence. A state of the form
Q
B
[) (2.5.13)
will be annihilated by Q
B
for any and so is physical. However, it is orthogonal to all
physical states including itself:
[
_
Q
B
[)
_
=
_
[Q
B
_
[) = 0 (2.5.14)
61
if Q
B
[) = 0. All physical amplitudes involving such a null state thus vanish. Two physical
states which dier by a null state,
[

) = [) +Q
B
[) (2.5.15)
will have the same inner products with all physical states and are therefore indistinguish-
able. So we again identify the observable Hilbert space as a set of equivalence classes, the
cohomology of Q
B
. Following terminology from cohomology, a BRST-invariant state is also
called closed, and a null state exact.
Applying this to string theory [58] gives the BRST transformation
16

B
X

= i(c + c

)X

B
c = i(c + c

)c
B
c = i(c + c

) c

B
b = i(T
X
+T
g
)
B

b = (

T
X
+

T
g
), (2.5.16)
where the energy-momentum tensor has been divided into matter and ghost parts. Noethers
theorem gives the BRST current
j
B
= cT
X
+
1
2
: cT
g
:= cT
X
+ : bcc:, (2.5.17)
and correspondingly for

j
B
. (A total derivative can be added to make this a tensor.) This
form is rather general: the c-ghost times the matter gauge current plus half the ghost gauge
current.
Based on our previous experience, we expect that something will go wrong outside the
critical dimension. The problem is with Q
2
B
. We can calculate this from the j
B
j
B
OPE, but
the following gives a slight shortcut. The OPE
j
B
(z)b(0)
1
z
2
: bc(0) : +
1
z
(T
X
(0) +T
g
(0)) (2.5.18)
gives the commutator for the corresponding charge,
Q
B
, b
m
= L
X
m
+L
g
m
. (2.5.19)
16
A few details are left to the student: we have kept only the ghosts associated with the coordinate
part, because the Weyl ghosts are auxiliary elds (no derivatives) and can be integrated out. And we have
implicitly integrated B
ab
to give the delta-function on the metric, so the equation of motion has been used
for B
ab
.
62
Exercise: Using this anticommutator and the Jacobi identity, show that [Q
B
, L
X
m
+L
g
m
], b
n

vanishes if and only if the total central charge D 26 vanishes. This implies in fact that
[Q
B
, L
X
m
+ L
g
m
] = 0, because by ghost number any term in the commutator would have to
contain at least one c mode. Now extend this to [Q
B
, Q
B
, b
n
]: it vanishes, and so also does
Q
B
, Q
B
, if and only if D = 26.
So the Weyl anomaly shows up as an anomaly in the nilpotence Q
2
B
= 0 [58], which is
necessary for the consistency of the formalism.
To see how all the formalism works, let us look again at the lowest levels of the open
string. There is one more condition that must be imposed on the states, namely
b
0
[) = 0. (2.5.20)
The way I think about this is that when we work out the interactions later on we will nd that
the string propagator always includes a factor b
0
, which projects onto states satisfying (2.5.20)
because b
0
b
0
= 0. Recalling the notation (1.6.16), this means that we are interested in states
built with raising operators acting on the ghost vacuum [ ). The condition (2.5.20) implies
in turn that physical states satisfy
Q
B
, b
0
[) = (L
X
0
+L
g
0
)[) = 0, (2.5.21)
where L
X
0
+ L
g
0
= 2k
2
+ L 1, with L = L
X
+ L
g
the total ghost plus X excitation level,
while the total normal ordering constant is taken from the earlier result (1.6.17). The
condition (2.5.21) thus relates the mass of a string state to its level of excitation,
M
2
=
L 1
2
(2.5.22)
In terms of the modes,
Q
B
=

n=
c
n
L
X
n
+

m,n=
(mn)
2
(c
m
c
n
b
mn
)
CA
c
0
, (2.5.23)
where the subscript CA denotes creation-annihilation normal ordering. The c
0
term comes
from the normal ordering constant in L
g
0
, as follows from Q
B
, b
0
= L
X
0
+L
g
0
.
At the lowest level, L = 0, the states are [ , 0, k), denoting the ghost vacuum, X vacuum
and momentum. Then
0 = Q
B
[ , 0, k) = (2k
2
1)c
0
[ , 0, k), (2.5.24)
63
giving the same mass shell condition as in the old covariant quantization. In this formalism,
the constant a = 1 is from the ghosts. There are there are no exact states at this level, so
we have the same tachyon states as before.
At the next level, N = 1, there are 26+2 states,
[) = (e
1
+b
1
+c
1
)[ , 0, k), (2.5.25)
depending on a 26-vector e

and two constants and . The BRST condition is

0 = Q
B
[) = 2(k
2
c
0
+k ec
1
+k
1
)[ , 0, k), (2.5.26)
so an invariant state satises k
2
= 0 and k e = = 0. There are 26 linearly independent
states left. A general [) is of the same form (2.5.25) with constants e

, so the general
BRST-exact state at this level, with k
2
= 0, is
Q
B
[) = 2(k e

c
1
+

k
1
)[ , 0, k). (2.5.27)
Thus the ghost state c
1
[0; k) is BRST-exact, while the polarization is transverse with the
equivalence relation e

= e

+ 2

. This leaves the 24 positive-norm states expected for a

massless vector particle.
Exercise: Do the rst massive level.
This pattern is general: with the ghosts there are 26 + 2 oscillators at each level. The
BRST condition eliminates two of these, and two others are exact, leaving 26 2 oscilla-
tors. The BRST quantization is equivalent to the old covariant and light-cone quantiza-
tions [58], [59]. In fact, the BRST quantization reduces to the old covariant quantization if
we consider only states where the ghosts are in their ground state. The no-ghost theorem
states that every cohomology class includes a state of this form. The generalization to the
closed string is again straightforward, with k

/2.
This derivation of the BRST formalism makes it look like a consequence of gauge xing,
but should be thought of as more fundamental. It carries the full information of the original
gauge symmetry; in eect, we increase the redundancy of the theory by the addition of the
ghosts, but the BRST principle then singles out the physical observables. When we try to
generalize string theory, it is generally easiest to generalize it within the BRST formalism,
rather than the locally invariant formalism.
It is worth noting that in most familiar circumstances, Ward identities are important
but are only part of the storyone also has dynamics. But string theory has coordinate
64
invariance, both on the world-sheet and in spacetime, so that time itself is a gauge degree of
freedom. As a consequence the dynamics becomes part of the constraints. As an example,
the mass shell condition (which is the Fourier transform of a Klein-Gordon equation in
spacetime) arose from the physical state condition.
2.6 Generalizations
We have developed the importance of the Weyl invariance, and its relation to the central
charge and the nilpotence of the BRST operator. Now we are in a position to state one
answer to the question, What is the most general consistent string theory? We will restrict
attention to theories for which the action for the embedding time takes the Polyakov form

1
8
_
d
2

gg
ab

a
X
0

b
X
0
. (2.6.1)
That is, the theory is stationary, and (jumping ahead a bit) the spacetime metric has
G
0
=
0
. Even in eld theory, non-stationary situations are more complicated. There is
no longer a distinguished zero-particle vacuum state, and the meaning of a particle becomes
ambiguous. This subject is much less developed in string theory, and I will avoid it here.
Now, the immediate problem is the one we have discussed, that the quantization of the
action (2.6.1) leads to negative norm states, and we need the observable spectrum to have a
positive norm. The general solution is simply stated: it depends only on the local symmetry,
or BRST invariance, being preserved. That is, the no-ghost theorem holds if we replace the
25 spatial X

elds with any (c, c) = (25, 25) unitary CFT. In particular, the calculation of
Q
2
B
goes precisely as before, with T
X
replaced by the T of the CFT. The unitary condition
is necessary because we do not want any other negative norm states besides those from X
0
.
Needless to say this generalization has something to do with the idea of compactifying some
of the spatial dimensions, but we will discuss this in section 3. We have not yet described
string interactions, but this is also the condition under which we can introduce consistent
interactions (the CFT must satisfy the two conditions discussed earlier, OPE associativity
and one-loop one-point modular invariance).
I should emphasize that all of these theories have tachyons. Whatever the spatial CFT is,
it contains the unit operator. The state [0, k
0
) from the X
0
theory times [1) from the spatial
theory is in the physical spectrum for 1 = L
0
=
1
2
(k
0
)
2
, which is unstable (tachyonic). To
eliminate the tachyon we have to generalize in a dierent way, enlarging the world-sheet gauge
65
symmetry. Everything that we have done generalizes readily to the N = 1 superconformal
algebra [31], [32]. A simple N = 1-invariant extension of the bosonic string action uses D
copies of the theory (1.7.7),
S =
1
4
_
d
2
z
_
X

_
. (2.6.2)
To be precise this is (1, 1) supersymmetric, with one analytic and one antianalytic supercur-
rent. To build a BRST charge we need a bc or theory for each gauge current, with the b
or ghost having the same weight and opposite statistics from the current. So in this case
we need a = 2 bc system and a =
3
2
system. The BRST current follows the same
pattern as (2.5.17),
j
B
= cT
X
+T
X
F
+
1
2
: (cT
g
+T
g
F
) : . (2.6.3)
This is nilpotent in the critical dimension, which is again where the total central charge
vanishes. The ghosts contribute c = c = 11, while each

contributes
1
2
, for
c = c =
3
2
D 26 + 11 D
crit
= 10. (2.6.4)
Both the
0
and
0
oscillators create negative norm states, which in the critical dimension
are removed by the constraints. The spectrum is equivalent to the light-cone gauge, in which
only D2 transverse sets of modes remain. The same generalization as above can be made:
replace the 9 spatial X theories with any (c, c) = (
27
2
,
27
2
) unitary (1, 1) super-CFT.
The details of the spectrumconstraints from modular invariance, the GSO projection,
absence of tachyons, and spacetime supersymmetrymake a long story that we do not have
time for. Let me just mention that in the Ramond sector (integer modes) the Ramond
generator G
0
is
G
0
= k

+. . . , (2.6.5)
terms with lowering operators being omitted, and

0
,

0
=

. (2.6.6)
The algebra (2.6.6) is the same as that of the Dirac matrices, and its unique representation
is the 2
D/2
-dimensional Dirac spinor representation. The G
0
constraint then gives the Dirac
equation, just as the L
0
constraint gave the Klein-Gordon equation.
66
The analytic and antianalytic constraint algebras need not be the same [60]. The (1, 0)
heterotic string combines an analytic N = 1 superconformal symmetry with antianalytic
conformal symmetry. The ghosts have central charge (c, c) = (15, 26). For D = 10 the
action (2.6.2) with

omitted has central charge (15, 10), so an additional (0, 16) is needed
from another unitary CFT. The constraints of modular invariance are strong, and the only
solutions are an E(8)E(8) or SO(32) level 1 current algebra [60], [61]. The generalization is
to combine the (c, c) = (
3
2
, 1) CFT from = 0 with the ghosts plus any (c, c) = (
27
2
, 25) theory
with (1, 0) superconformal symmetry. This string is the most promising for a unied theory,
the superconformal side eliminating the tachyon and producing spacetime supersymmetry,
and the side with the current algebra giving rise to the gauge symmetries.
Moving on, the N = 2 superconformal algebra needs = 2 and = 1 bc systems plus two
=
3
2
systems for total central charge 262+2 11 = 6. The basic free representation
consists of one complex X and one complex (= two real) for c = 3, so the critical theory
has two such representations. As usual the constraints remove two sets of modes, leaving in
this case zero transverse dimensions. The only degree of freedom is a scalar from the center
of mass motion of the string. Nevertheless, there is some interesting structure [62].
Higher-N extended superconformal algebras have zero or negative critical dimension.
These are the only possibilities as long as we are restricted to half-integer spins 2. Both
the fractional case [63] and the higher spin case [64] (particularly W
3
) are under investigation.
In the fractional case, the OPE has branch cuts so a projection is needed to get well-dened
amplitudes. For both cases the product of two representations is not a representation, owing
to the nonlocality in the fractional case and the nonlinearity in the W
3
case. This has made
it hard to nd theories with a spacetime interpretation, and also makes it hard to construct
the BRST operator (since the ghosts and matter do not separate). For W
3
, a theory with a
spacetime interpretation has been found and the BRST operator constructed, but the result
is disappointing: it turns out to be a special case of the bosonic string. So the theories
described above are for now the most general known.
17
String theories are also distinguished by the world-sheet topologies allowed: closed only,
or with boundaries, and oriented only, or unoriented also. Boundaries must be consistent
with the constraint algebra, as in eq.(1.9.13) for the conformal case.
18
In particular, the
17
I have restricted attention to algebras which have nontrivial unitary representations, because this is
necessary to get a spacetime interpretation. This implicitly excludes topological string theories, about which
I have little to say.
18
There are also additional conditions generalizing the OPE associativity and modular invariance [65].
67
a) b)
Figure 12: a) World-sheet of closed string splitting. b) Time slices of this process.
constraint algebra must be left-right symmetric, so that boundaries are not possible in the
heterotic case. Unoriented surfaces are allowed only if the full world-sheet action is invariant
under world-sheet parity (exchange of z and z). Theories without boundaries have closed
strings only; theories with boundaries have both closed and open strings. There are no
string theories with open strings alonea surface with boundary can be cut open along
a curve running from boundary to boundary, corresponding to open string intermediate
states, or along a closed curve, corresponding to closed string intermediate states. Inclusion
of unoriented surfaces has the eect of projecting the spectrum onto states of even parity.
2.7 Interactions
For composite strings, there are a variety of contact and long-range interactions. In the
fundamental string, such interactions cannot be introduced without spoiling the local sym-
metries, and are not possible.
19
The only interactions that are possible are those that are
already contained in the sum over all surfaces. For example, g. 12 shows a world-sheet in
which one closed string splits into two, or the time-reversed process in which two join into
one. Figure 13 shows two strings scattering by exchange of one or two strings. Since the
string spectrum includes the graviton, these amplitude include the gravitational processes
discussed earlier. But now there is no short-distance divergence because the interaction is
spread out. I will be describing in somewhat more detail how this works.
Now it is perhaps time to mention an additional term that can appear in the Polyakov
19
For example, a contact interaction
_
d
2
d
2

_
g()
_
g(

)(X() X(

)) is not Weyl invariant

(Fourier transforming the delta function, one obtains exponentials with a continuum of weights.)
68
a)
b)
Figure 13: String analogs of g. 10bc: closed strings scattering by exchange of one string (a)
or two strings (b).
action. This is
0
, with
0
a parameter and
=
1
4
_

gR, (2.7.1)
R again being the world-sheet curvature. This is not merely di Weyl invariant, it is
topologically invariant. For a closed surface with h handles is the Euler number = 22h.
Adding a handle, as in going from the one-string to the two-string exchange in g. 13, the
path-integral weight e

changes by e
2
0
. So while this term does not aect anything local
like the world-sheet equations of motion, it does aect the relative weights of surfaces of
dierent topologies. Adding a handle is like adding two trilinear closed string interactions,
so the closed string coupling g
c
depends on
0
as e

0
. For world-sheets with boundaries, as
in open string theory, di Weyl invariance requires also a surface term
=
1
4
_
d
2

gR+
1
2
_
boundary
ds k (2.7.2)
69
where k is the geodesic curvature of the boundary; in terms of the tangent and normal
vectors, k = t
a
n
b

a
t
b
. Again this is purely topological,
0
times the Euler number. For a
compact surface with h handles and b boundaries the Euler number is 2 2h b. For the
open string amplitudes analogous to g. 13, with a pair of open strings exchanging one or
two open strings, adding a strip increases b by one and so the path integral weight changes
by e

0
. Thus the coupling g
o
of three open strings goes as e

0
/2
. This will play an important
role at a later point, so let me emphasize it: in theories with both open and closed strings,
the couplings are related
g
c
g
2
o
. (2.7.3)
For both particle physicists and condensed matter physicists, the natural thing to try do
at this point is to calculate a Greens function, a propagator. Since we are talking about
one-dimensional objects, this would be the amplitude to propagate from a given initial
conguration to a given nal conguration, the sum over all surfaces bounded by these two
loops. However, while this seems like a very natural thing to do it is actually extremely
hard to carry out consistent with the di Weyl invariance. One already sees this in that
the physical state conditions require strings to be on-shell, k
2
= m
2
. Any local source in
spacetime would couple to all momenta, not just those on the mass shell. This should not be
a surprise, because this theory has spacetime gravity, so we have to have observables which
are spacetime coordinate-invariant. It is extremely clumsy to try to describe the position
and shape of a loop in a coordinate-invariant way.
What can be easily dened in an invariant way is the S-matrix, for scattering from some
set of incoming strings to some set of outgoing strings. Eectively the sources have been
taken to innity. Each of the external strings is a semi-innite cylinder in a world-sheet
coordinate w. We are familiar with the mapping z = e
iw
which takes this to the interior
of the disk, with i mapped to the origin. As shown in g. 14, this mapping leaves a
compact surface. It may seem that I am being careless in identifying a long cylinder in
spacetime with a long cylinder in the world-sheet coordinate, but we will see that long-
distance propagation in spacetime in fact comes from the latter. The source which creates
the incoming or outgoing string state is then a local operator, known as a vertex operator 1.
We are already familiar with the mapping between states and operators. Thus, the tachyon
[0, k) with k
2
= 2 is created by 1 = e
ikX
, the graviton, antisymmetric tensor, and dilaton
states

1
[0, k) with k
2
= 0 are created by 1 = X

X

e
ikX
, and so on.
In order to make the vertex operator di-invariant we need to integrate over the world-
70
a) b)
Figure 14: a) 2 2 scattering process. b) Conformally equivalent picture, the cylinders
reduced to small holes.
sheet,
V = 2e

0
_
d
2

g1 , (2.7.4)
where we have included the coupling constant. In at gauge this becomes
e

0
_
d
2
z 1 . (2.7.5)
Under conformal transformations the measure d
2
z transforms as a (1, 1) tensor, so con-
formal invariance requires that 1 be a (1, 1) tensor.
20
This is precisely the OCQ physical
state condition (2.3.4), which gives us a simple way to understand the value a = 1: it makes
1 a (1,1) tensor.
20
By the way, any vertex operator which is conformally invariant on the at world-sheet can be made di
Weyl invariant by appropriate coupling to the metric.
71
We are led to the following expression for the S-matrix,
S =

compact
topologies
e
(n)
0
_
[dX dg]
V
diWeyl
e
S
P
n

i=1
_
d
2

i
1
i
. (2.7.6)
The product runs over the vertex operators incoming and outgoing states, these being dis-
tinguished only by the sign of k
0
. (For a general CFT S
P
is replaced by the appropriate
action). We now need to understand how gauge xing works globally. Locally the number
of metric degrees of freedom (three) matches the number of gauge degrees of freedom, but
globally there is a small mismatch. In fact, the space of equivalence classes, metrics modulo
di Weyl, is identical to a space described earlier in these lectures, the space of Riemann
surfaces. Specifying the metric up to Weyl transformations singles out a family of complex
coordinates, namely those in which the metric is proportional to dzd z. This is the denition
of a Riemann surface. In the opposite direction, given a Riemann surface we can construct
a metric by taking dzd z in each patch and smoothing between patches (this can always be
done). So this is an isomorphism.
For example, take for the torus the xed coordinate region 0
1
2, 0
2
2 with
periodic boundary conditions, so the metric is a doubly periodic function g
ab
(
1
,
2
). Then
by coordinate transformations which preserve the periodicity and Weyl transformations we
can bring the metric to the form
ds
2
= [d
1
+d
2
[
2
. (2.7.7)
for some . These are the same metrics described earlier in terms of a xed metric and
-dependent coordinate region. So after gauge-xing we are left with an integral over the
moduli space of Riemann surfaces, of complex dimension 0 for h = 0, 1 for h = 1 and 3h3
for h 2. For h = 0, 1 there is the further complication of conformal Killing vectors, gauge
symmetries which remain after xing the metric. These can be xed by xing the positions
of some vertex operators, 3 at h = 0 or 1 at h = 1. In all, if we have n vertex operators on
a genus h surface, the total number of complex moduli (for the metric and the positions) is
3h +n 3.
To carry out the Fadeev-Popov procedure we trade the original integral [dg]

i
d
2

i
for
[d d] d

i
d
2

i
where

t are the moduli for the surface and the prime on the product
denotes the omission of any xed vertex operators. The Fadeev-Popov procedure can be
carried out as before, giving a mixture of functional and nite-dimensional determinants,
which again can be expressed in terms of a path integral over ghosts. I will quote here only
72
the result. The S-matrix for n external strings is given by a sum and path integral
S =

compact
topologies
e
(n)
0
_
[dX db dc] e
S
P
Sg
. . . . (2.7.8)
The insertions are of three types [19], [66], [8], [17]:
For each modulus (now divided into real parts t
r
), the b-ghost insertion
1
4
_
d
2

gb
ab
g
ab
t
r
. (2.7.9)
For each vertex operator which is xed, the insertion
c c1
i
. (2.7.10)
For each vertex operator which is integrated, the insertion
_
d
2

g1
i
. (2.7.11)
This can also be expressed in terms of the data which dene the Riemann surface, the
transition functions. It is convenient to x every vertex operator, put all the moduli in
the transition functions. The vertex operators are then c c1
i
as above, while the b-ghost
insertions are
1
2i

(mn)
_
Cmn
_
dz
m
f
nm
t
r

zn
b
zmzm
d z
m

f
nm
t
r

zn
b
zm zm
_
, (2.7.12)
where the sum runs over all pairs of overlapping patches and the integral runs along any
curve separating the patches.
Notice that for OCQ-type vertex operators the ghosts are in their vacuum state [ ),
which translates into c c. So it is the xed vertex operators which are given by the state-
operator mapping. The integrated vertex operators can be understood as arising from the
b-ghost insertions for the position. The insertion for translating a little patch containing the
vertex operator is

b
1
b
1
, giving

b
1
b
1
c c1
i
= 1
i
, (2.7.13)
which is the integrated form. The rules in terms of Riemann surfaces, by the way, apply to
all BRST-invariant vertex operators, not just the OCQ-type.
73
2.8 Trees and Loops
To conclude this subject I work out one example and discuss some general principles. For
four closed string tachyons on the sphere, fully xing the gauge invariance leaves one position
integrated and three xed,
S(k
1
, k
2
, k
3
, k
4
) = e
2
0
C
S
2
_
d
2
z
4
(2.8.1)
< cce
ik
1
X
(z
1
, z
1
) cce
ik
2
X
(z
2
, z
2
) cce
ik
3
X
(z
3
, z
3
)e
ik
4
X
(z
4
, z
4
) >
S
2
,
where the integral runs over the complex plane, and C
S
2
is a numerical normalization factor
for the path integral on the sphere. The expectation value can be obtained from our earlier
results, giving
21
S(k
1
, k
2
, k
3
, k
4
) = ie
2
0
C
S
2
(2)
26

26
(

i
k
i
)
_
d
2
z
4
[z
12
[
2
[z
13
[
2
[z
23
[
2

i<j
[z
ij
[
2k
i
k
j
ie
2
0
C
S
2
(2)
26

26
(

i
k
i
)
_
d
2
z
4
[z
4
[
2k
1
k
4
[1 z
4
[
2k
2
k
4
. (2.8.2)
In the second line we have used the fact that the result is independent of the xed positions
(as can be shown by a M obius transformation) to move z
1
0, z
2
1, z
3
. The
integral can be related to gamma functions with the result
S(k
1
, k
2
, k
3
, k
4
) = ie
2
0
C
S
2
(2)
26

26
(

i
k
i
) (2.8.3)
(s/2 1)(t/2 1)(u/2 1)
(s/2 t/2 2)(t/2 u/2 2)(u/2 s/2 2)
,
where s = (k
1
+k
2
)
2
, t = (k
1
+k
3
)
2
, and u = (k
1
+k
4
)
2
, and u +t +s = 8.
This is the Virasoro-Shapiro amplitude. It has a single pole when s, t, or u takes a value
2, 0, 2, 4, . . . . These are the masses of string states and correspond to processes where two
of the external strings join into one; the coecient of the pole is related to the square of the
three-point amplitude, which determines C
S
2
= 8
2
. The massless pole, for example, has
the correct form expected from graviton plus dilaton exchange (the antisymmetric tensor
doesnt contribute due to world-sheet parity symmetry).
21
The factor of i, needed for unitarity, can be attributed to the analytic continuation needed to dene the
X
0
integral.
74
1
2
3 4
Figure 15: One loop four-string amplitude.
The notable feature of this amplitude is its soft high-energy behavior. Using Stirlings
approximation for the exponentials, one nds that at large s and xed (and so xed t/s,
u/s) it behaves as
e
sf()
(2.8.4)
This is in contrast to similar amplitudes in pointlike eld theory, which fall as powers of s,
and arises from spread-out nature of the string (more on this later).
Now we turn to the one loop amplitude. The one-loop four-point amplitude shown in
g. 15 contains the two-graviton exchange described earlier, which had the severe divergence
in eld theory. Let us look at the integration over the momentum running around the loop.
From the earlier discussion (1.9.9), the path integral translates into a Hamiltonian expression
involving
e
2i(L
0


L
0
)
. (2.8.5)
For large loop momenta the dominant term here is
e
2k
2
Im()
. (2.8.6)
After Wick rotation of k
0
, this is a convergent gaussian at xed
2
. One might also worry
about a divergence from the sum over the string states running around the loop, but in
spite of the large number of states the sum is handily convergent at xed owing to the
exponential suppression factor (2.8.5). The region Im() 0 is the potential danger, both
for the integral over momentum and for the sum over states. But here we run into the happy
circumstance that this is not in the range of integration, the moduli space of the torus,
75
1
4
2
3
1
4
2
3
1
4
2
3
a) c)
b)
Figure 16: Three pictures of the same limit of moduli space. a) z
4
z
1
on the sphere. b)
Annular region around z
1,4
conformally transformed to a long cylinder. c) Long cylinder
conformally transformed to a pinched cylinder.
shown in g. 6. There is a lower bound on Im(), so the integral over momenta is gaussian,
as would be expected from the high energy behavior (2.8.4). In contrast, in eld theory one
could write the loop integral in a Schwinger parameterization,
_

0
ds e
(k
2
+m
2
)s
, (2.8.7)
with s being analogous to Im(), but here the integral does indeed run down to s = 0,
leading in the end to a divergent momentum integral. (Cutting o the s integral doesnt
work, as I will explain shortly).
So in this example we see that the would-be ultraviolet region of moduli space is missing.
This is a general principle in string theory: all the limits of moduli space can be interpreted
as infrared limits. This can already be seen in the Virasoro-Shapiro amplitude. The moduli
space has three limits, z
4
z
1
, z
4
z
2
, or z
4
z
3
. The rst is shown g. 16a; it
is conformally equivalent to the long cylinder in the g. 16b, and to the plumbing-xture
construction in g. 16c with q z
4
z
1
0. In the last form we can use the sewing
formula (1.9.12) to express the asymptotics in terms of a sum over intermediate string states.
76
The contribution of state i is proportional to
_
0
d
2
q [q[
k
2
+m
2
i
2
. (2.8.8)
The integral runs over some neighborhood of the origin. The behavior as q 0 is dominated
by the lightest states. The integral converges when k
2
+m
2
i
is positive for all states and can
be dened elsewhere by analytic continuation.
22
The result has a pole proportional to
1
k
2
+m
2
i
. (2.8.9)
This is the origin of the series of poles in the Virasoro-Shapiro amplitude. These poles
correspond to long-distance propagation in spacetime, so as in the discussion of g. 14 this
comes from a degenerating (long) cylinder.
We could extract the asymptotics in this limit z
4
z
1
directly from the OPE; the sewing
formula in this case reduces to the OPE. But the sewing formula is more general. The surface
formed by sewing general surfaces /
1
and /
2
is conformally equivalent to a small copy of
/
1
inserted into /
2
(and vice versa). The sewing formula related this to a sum of local
operators inserted in /
2
, generalizing the OPE.
All limits of moduli space are of the same type as this one, with one or more handles
degenerating.
23
They can thus be analyzed by means of sewing in the same way.
24
The
asymptotics are dominated by the lightest states, and the divergences all arise from inter-
mediate states being on the mass-shell. This corresponds to long-distance propagation in
spacetime, and so is an infrared eect. There are no short-distance divergences. As an exam-
ple, consider the limit of the one-loop four-point amplitude in which all the vertex operators
come together, g. 17a. This looks as though it could produce a short-distance divergence.
But, in analogy to the previous g. 16, this is conformal to g. 17b in which the vertex
operators are at the end of a long cylinder, and to g. 17c where a sphere and torus are
sewn together. There may indeed be a divergence, but it is an infrared eect, dominated by
the lightest states. The momentum owing through the pinch must be zero by momentum
22
The divergence for k
2
+m
2
i
< 0 is uninteresting, an artifact of the Schwinger-like integral representation,
but the pole at k
2
+m
2
i
= 0 is meaningful.
23
Although this is well-known in the mathematics literature, I nd the most useful discussion for the
purposes of physics to be that in section 6 of ref. [67], where moduli space is explicitly decomposed into a
Feynman diagram-like sum. We will return to ths in section 4.1.
24
By the way, the b-ghost insertion for the moduli q and q is b
0

b
0
. This projects onto states which are
annihilated by b
0
and

b
0
, as claimed in our discussion of the BRST cohomology.
77
a)
b)
c)
Figure 17: Three pictures of the another limit of moduli space. a) Four vertex operators
coming together on the torus. b) Annular region around operators conformally transformed
to a long cylinder. c) Long cylinder conformally transformed to a pinched cylinder.
conservation, so the pole (2.8.9) is 1/0 for any massless state. The interpretation is that
the one-loop one-point function produces a source for the massless eld, so it is necessary
to expand around a solution to the loop-corrected equations. This cancels the divergence,
and is known as the Fischler-Susskind mechanism [68], [39]. More generally the principle is
that all divergences are from long distances and so, as in eld theory, go away if one asks
the right questions.
Finally a few comments about the general structure of amplitudes. The interactions
must respect the BRST invariance. In particular, the amplitude for an exact state Q
B

1

must be zerothis is equivalent the the gravitational Ward identity, together with the
corresponding Ward identities for all the higher levels. To see the decoupling of the null
state, write Q
B
as a contour integral around 1

. Expand the contour C and contract it

down around the other insertions in the path integral. The other vertex operators are BRST
invariant and give zero. However, there is a nonzero residue at the b insertions, which by the
OPE (2.5.18) is T
total
. In either form (2.7.9) or (2.7.12), this gives a total derivative of the
78
path integral with respect to the moduli. Upon integration over moduli space this becomes a
surface term[19], [66], [8], [69]. The surface term vanishes under the same condition that the
(analytically continued) integral is nitethat is, unless we are sitting on a pole (in which
case the Fischler-Susskind mechanism restores conformal invariance).
Notice that if we mutilate the theory by cutting out parts of moduli space, the total
derivative no longer integrates to zero and the theory is inconsistent. The same thing happens
in eld theory if we try to cut o the Schwinger parameter integral (2.8.7). String theory
manages to cut o the modular integral while leaving a smooth moduli space without spurious
internal boundaries.
One can also arrive at the rules for the string S-matrix from considerations of BRST
invariance, rather than via gauge-xing. First, as found in the exercise below eq. (1.9.4),
ghost number conservation requires the path integral to contain insertions with net N
b
N
c
=
3h3; our rules for the string S-matrix are consistent with this. These insertions are BRST
invariant only up to a total derivative with respect to the moduli, so the path integral must
be integrated over moduli space as found from gauge xing. In trying to generalize string
theory it is often easiest to use this strategy directly to determine the form of the amplitudes,
rather than gauge-xing from a locally-invariant form. For example, superstring amplitudes
can in this way be written as an integral over supermoduli space.
This concludes our survey of the current perturbative answer to the question, what is
string theory? In particular we have seen what are the ingredients that make the pertur-
bation theory consistent.
3 Vacua and Dualities
3.1 CFTs and Vacua
In the last section we described the most general string theory as dened by the world-sheet
gauge algebra, by the world-sheet topologies allowed, and by the particular world-sheet CFT
or super-CFT.
1
In this section we will try to get a better understanding of the nature of the
space of theories. The main theme is that what we have called dierent string theories are
in many (and possibly all) cases the same theory expanded around a dierent vacuum state.
Moreover, in some cases what appear to be dierent theories are in fact the same theory
1
This was nicely laid out in the review article [70].
79
in the same vacuum. Some of these connections are well-developed, while others are highly
conjectural.
Let us start by trying to understand better the idea of replacing the spatial X

s with
a more general CFT. I will focus on the bosonic string, but the ideas all generalize to the
superstring. For example, replace the Polyakov action with the non-linear sigma model [22]-
[24]
S

=
1
8
_
d
2

_
_

gg
ab
G

(X) +i
ab
B

(X)
_

a
X

b
X

+ 2

gR(X)
_
, (3.1.1)
which is the most general coordinate invariant action we can make with two derivatives. This
is the same theory (1.6.27) that we have already discussed on a at world-sheet, and when the
X-dependence is slow the theory can be studied perturbatively (this is now perturbatively
in the world-sheet couplings). This gives rise to a consistent string theory, to leading order
in the world-sheet perturbation theory, if
2
2R

+ 4

1
2
H

= 0

+ 2

= 0
D 26 6
2
+ 12

1
2
H

= 0. (3.1.2)
This conformal eld theory has an obvious interpretation. The functions G

(X), B

(X)
and (X) correspond to nontrivial backgrounds of the string graviton, antisymmetric tensor
and dilaton. One might have expected something stringier-looking, since the graviton, etc.,
are supposed to be strings. But we found in our discussion of scattering amplitudes that
strings could be created by local vertex operators; the sigma model action can then be
thought of from exponentiating the vertex operators, corresponding to a coherent state of
strings.
3
Notice in particular that dierent values of the string coupling e

0
are now seen as
corresponding to expanding around dierent backgrounds in a single theory, namely dierent
2
To relate this to the at world-sheet discussion, note that on the at world-sheet the term in the action
is trivial. However, it aects the energy-momentum tensor, which is given by a derivative with respect to
the metric. When the equations (3.1.2) are satised with a nontrivial , T
z z
given by eq. (1.6.28) does not
vanish, but adding an improvement term produces a conserved energy-momentum tensor with T
z z
= 0.
The last of line of eq.(3.1.2) is the perturbatively-corrected matter plus ghost central charge.
3
The dilaton actually involves uctuations both of and the diagonal part of G

. The curvature term

in its vertex operator is of course not evident on the at world-sheet, but arises on a curved world-sheet
when the operator is renormalized in a coordinate-invariant way.
80
constant values of the dilaton eld . The eld equations (3.1.2) arise from a spacetime
action
S =
1
2
_
d
D
X

Ge
2
_

(D26)
3
+R
1
12
H

+ 4

_
. (3.1.3)
Higher orders in the world-sheet perturbation expansion produce terms with more derivatives
in the eld equations and action.
Each static solution corresponds to a possible ground state. For example, a background
with four at dimensions and the remainder curved on a suciently small distance scale l
c
would look like what we see around us.
4
Even before string theory this was a promising idea
for unication. It unites elds of dierent spins into a single higher-dimensional eld (Kaluza-
Klein theory), and provides a natural origin for the fermion generations as states with the
same gauge quantum numbers but dierent wavefunctions in the compact dimensions. By
the way, we see from the loop correction to the central charge in the last line of eq. (3.1.2)
that the number of compact dimensions need not be 22. Actually the expansion which gives
rise to eqs. (3.1.2) breaks down when the corrections are of order 1, but exact CFTs with
D ,= c are known. An example is the WZW model, correpsonding to strings propagating on
a group manifold [72], [27].
For more general CFTs, the interpretation in terms of compactication is not so clearcut.
But given the many equivalences between two-dimensional eld theories (the most familiar
being bosonization), it is quite possible that all can be understood in terms of background
elds, perhaps including massive string elds. As one example, the minimal models were rst
constructed in an abstract way, but now several Lagrangian formulations are known. One in
particular, the Landau-Ginsburg theory, can be regarded as the string moving in a tachyon
background [73]. So the general assumption is that dierent CFTs should be regarded as
dierent vacua.
Let us make a few general observations. Just as we have noted that the bosonic string
always has a tachyon, it also has a graviton, antisymmetric tensor, and dilaton. The vertex
operator e

e
ikX
, with e

and k

lying in the noncompact at directions, is always

a (1, 1) vertex operator for k
2
= 0. The (1, 1) and (1, 0) superstrings also always include the
graviton and tachyon.
4
Direct experimentthe non-observation of the enormous number of states with nontrivial dependence
on the compact dimensionsrequires l
c
to be less than 10
16
cm. Other considerations, however, require it to
be right at the string scale 10
32
cm. In particular, if there is a large hierarchy between the string and com-
pactication scales, the theory at intermediate scales is a D > 4 eld theory; this is badly nonrenormalizable
and breaks down well before the string scale [71].
81
If the compact theory has a current algebra, the vertex operator X

j
a
e
ikX
or j
a

e
ikX
is (1, 1) for k
2
= 0 and corresponds to a gauge boson. So the spacetime gauge symmetries
are in one-to-one correspondence with the global symmetries (current algebras) of the com-
pact CFT. A similar result, a bit less straightforward to derive, is that there is spacetime
supersymmetry if and only if the N = 1 superconformal constraint algebra of the superstring
is embedded in a larger N = 2 symmetry of the CFT, with a quantization condition on the
U(1) charge of the N = 2 algebra [74].
The number of CFTs or super-CFTs of appropriate central charge is enormous, even
restricting to those with exactly four noncompact at directions. We are now interpret-
ing these as dierent backgrounds within a single theory, but the eective four-dimensional
physics is dierent in each. The problem of nding the right theory is thus transmuted in
string theory into the dynamical problem of nding the right background. For backgrounds
with spacetime supersymmetry, it is a theorem that they remain stable to all orders of per-
turbation theory, but it is also known, from the understanding of supersymmetry breaking
in eld theory, that most are destabilized by non-perturbative eects. Spacetime supersym-
metry of the eective eld theory below the string scale gives a great deal of information
about the dynamics, a subject currently in active development. But it seems likely that
dynamics at the string scale will also play an important role. It may also be that there are
in the end many stable ground states of string theory, so that the choice between them will
be determined in part by the initial conditions.
3.2 Compactication on a Circle
A at spacetime with one dimension periodic,
X
1

= X
1
+ 2R (3.2.1)
is the simplest compactication of the bosonic string, but is quite instructive. (Equivalently,
let the periodicity be 2 but G
11
= R
2
). The periodicity has two eects. The rst is that
the string wavefunctions must respect it, so
k
1
=
n
R
(3.2.2)
is quantized. This is the same as for a eld theory on this space. The second eect is unique
to string theory: a string can be wound around the periodic dimension, so that
X
1
(2) = X
1
(0) + 2mR. (3.2.3)
82
Referring back to our mode expansion (1.4.23) for the free scalar, we see that this means
that the eigenvalues k
1
L,R
of
0
and
0
are no longer equal, k
1
L
k
1
R
= mR. The total
momentum (3.2.2) is
1
2
(k
1
L
+k
1
R
), so
k
1
L
=
mR
2
+
n
R
, k
1
R
=
mR
2
+
n
R
. (3.2.4)
Also from the expansion (1.4.23), the vertex operator for such a state will be proportional
to
e
ik
L
X
L
(z)+ik
R
X
R
( z)
. (3.2.5)
Although the X
R
X
R
and X
L
X
L
OPEs contain branch cuts, the OPE of any two vertex
operators (3.2.5) in the spectrum (3.2.4) is single-valued.
The L
0

L
0
physical state conditions thus become
M
2
= k
2
0

25

=2
k
2

=
m
2
R
2
4
+
n
2
R
2
+ L +

L 2
mn + L

L = 0 . (3.2.6)
Looking at the massless spectrum, the states

1
[0, m = n = 0), (3.2.7)
with no compact momentum or winding, remain massless. For and both in the range
0, 2 . . . , 25, these are just the graviton, dilaton, and antisymmetric tensor of the 25-dimen-
sional theory. When either = 1 or = 1 the state is a vector in the noncompact dimensions,
a gauge boson. The corresponding vertex operators are
(X

X
1
X
1

)e
ikX
. (3.2.8)
The plus sign comes from the 26-dimensional metric, the Kaluza-Klein mechanism. The
minus sign comes from the 26-dimensional antisymmetric tensor, a generalization of the
Kaluza-Klein mechanism; call this an H gauge boson. The operator product of the gauge
boson vertex operator (3.2.8) with a general vertex operator of momentuml is proportional to
l

(l
1
L
l
1
R
). The Kaluza-Klein and H gauge bosons thus couple to the compact momentumand
winding number respectively. Finally, the state (3.2.7) with = = 1 is a 25-dimensional
scalar, and is the metric component corresponding to the radius R of the compact dimension.
83
So far this is the same as would be found just from the low energy eld theory (3.1.3),
but this simple theory has some interesting stringy physics. Consider the four sets of states
with [m[ = [n[ = 1 and L +

L = 1,

1
[0, m = n = 1),

1
[0, m = n = 1). (3.2.9)
Their masses are
M = [R
2
2[/2R. (3.2.10)
Precisely at the radius R =

2 these states are massless, and so are gauge bosons. At this

radius, the spectrum (3.2.4) includes the four currents
: e
iX
1
L
(z)

2
:, : e
iX
1
R
( z)

2
: . (3.2.11)
Together with the two U(1) currents
X
1
(z),

X
1
( z), (3.2.12)
these form an analytic and an antianalytic SU(2) current algebra.
5
This SU(2) SU(2)
symmetry has nothing to do with the Lorentz invariance of the at-spacetime theory. Its
emergence at the critical radius is an example of the large symmetry that is hidden in string
theory, almost completely broken. By the way, the mass of the gauge boson, as R moves
away from the critical value, comes from the ordinary Anderson-Higgs mechanism. The
vertex operator for a small change in R,
X
1

X
1
, (3.2.13)
is built out of the SU(2) currents (3.2.12), and so transforms as the z-component of a vector
under each SU(2), breaking SU(2) SU(2) down to the Kaluza-Klein and H U(1) U(1).
There is another stringy phenomenon here. As R , the states with m ,= 0 go to
innite mass, while the states with m = 0, n ,= 0 form a continuum. This is simply because it
costs an energy of order R to wind around the large dimension, while the momentum in that
direction becomes continuous in the limit. This is all as in eld theory. But look at R 0.
States with n ,= 0 become very massive because of the large compact momentum, just as in
5
This SU(2) is likely familiar to those in condensed matter physics who have looked at the Luttinger liquid
or other one-dimensional quantum systems. At general radii, these currents have dimension h+

h = 1 +M
2
and spin h

h = 1, and are no longer conserved. See ref. [75] for a recent example of a (dirty) system
where R ows to

2 at long distance, producing an SU(2) symmetry in the long distance theory which is
not present in the underlying theory.
84
eld theory. But the states with n = 0, m ,= 0 are now forming a continuum, something
which has no analog in eld theory. In fact, the spectrum is invariant under R-duality, also
known as T-duality,
R
2
R
, m n, (3.2.14)
which interchanges large and small radius, and interchanges compact momentum with wind-
ing number [76]; for recent reviews see [2], [77].
This is not only a symmetry of the spectrum, but also of the interactions. R-duality
takes k
1
L
k
1
L
, k
1
R
k
1
R
. If we extend this to

1
m

1
m
,
1
m

1
m
, (3.2.15)
so that
X
1
L
X
1
L
, X
1
R
X
1
R
, (3.2.16)
it is a symmetry of the OPE, and so by the sewing principle holds for all Riemann sur-
faces. Finally, a transformation of the dilaton is needed to make the loop expansions the
same. After integrating the spacetime action (3.1.3) over X
1
, the eective 25-dimensional
action is weighted by 2Re
2(X)
. In order that this be invariant we need 2R

e
2

(X)
=
2Re
2(X)
[78],

(X) = (X) ln(R/

2). (3.2.17)
The R and 2/R theories are then physically identical.
6
This can also be seen in a very dierent, and deeper, way [79]. Notice that the self-dual
radius is also the point of enlarged gauge symmetry. We have observed that the vertex
operator corresponding to a change in the radius transforms non-trivially under SU(2)
SU(2). In fact, a rotation by around the x-axis of one of the SU(2)s takes this operator
into minus itself. So increasing R is gauge-equivalent to decreasing it. This implies that
R-duality is a symmetry not only in perturbation theory (which is all that we can conclude
from the argument above) but is in fact an exact symmetry. We can say this, even though we
know nothing about non-perturbative string theory, because we do know that any violation
of gauge symmetry would make the low energy theory inconsistent. We can also say that
the R and 2/R vacua are not just identical states, they are the same state. This is not just
a semantic distinction. It means that there can exist defects in spacetime, such that as one
encircles them R changes continuously from its original value to the dual value.
6
To restore units, the statement is that R

/R, where

is again of order 10
32
cm.
85
Duality is a striking indication that strings do not sense spacetime in the same way as
particles, and that our notions of spacetime geometry and even topology break down at short
distance.
7
Thus, we want to think of X
1
(z, z) = X
1
L
(z)+X
1
R
( z) as the location of some world-
sheet point in spacetime, while the dual coordinate X
1
d
(z, z) = X
1
L
(z) X
1
R
( z) is a much
more complicated and nonlocal object. Yet if the theory is compactied at some radius near

2 these are equally physical, with the R physics being simple in terms of X
1
, and the
R 0 physics simple in terms of X
1
d
. Duality suggests that there is a minimum spacetime
length scalewe can restrict to R >

2.
8
So if spacetime breaks down at the string scale,
what is to replace it? A given CFT may have many dierent Lagrangian representations,
each giving a dierent picture of spacetime. It is the OPE coecients c
k
ij
which as we have
discussed are common to all representations of the theory, and which determine the string
amplitudes. But these seem to me rather abstract to be the fundamental description.
3.3 More on R-Duality
Duality has been a source of great fascination, and has many extensions, discussed in the
reviews [2], [77]. I will mention only a few. First, it can be extended to any transla-
tionally invariant direction [78]. Consider a world-sheet action involving elds V
a
, , and
X
0
, X
2
, . . . X
25
,
S =
1
4
_
d
2
z
_
G
1
(X)V

V +(

V

V ) +. . .
_
. (3.3.1)
Here, the ellipsis stands for terms involving the other world-sheet elds X but not or V ,
and G(X) is an arbitrary function. Integrating out V by completing the square leaves
S

=
1
4
_
d
2
z
_
G(X)

+. . .
_
. (3.3.2)
On the other hand, integrating out forces

V

V = 0, so V
a
is a gradient
V =
d
,

V =

d
(3.3.3)
for some function
d
. The action then becomes
S

=
1
4
_
d
2
z
_
G
1
(X)
d

d
+. . .
_
. (3.3.4)
7
We have already seen another example of this in the level 1 SU(2) WZW model, which can be described
in terms of three coordinates or in terms of one.
8
Though see ref. [80] for a situation where it appears that some dimensions are actually becoming much
smaller than this.
86
The two actions with reciprocal kinetic terms are thus equivalent. Noting that the equation
of motion for V is
V =
_
G
1
(X)
_
,

V =

_
G
1
(X)
_
, (3.3.5)
this reduces for constant G(X) to the earlier duality transformation, with a constant times
X
1
. In particular, a careful treatment of the measure produces the transformation of the
dilaton,

(X) = (X) ln
_
G(X). The above is readily extended to nonzero G
1
and
nonzero B
1
.
9
In this form, duality can be applied to many interesting string backgrounds.
With more than one periodic dimension there is much more structure. I briey summarize
the case of two compact directions X
1
and X
2
[81]. Let each be periodic with period 2, with
G
11
, G
12
, G
22
, and B
12
constants. Parameterize these four elds in terms of two complex
parameters =
1
+i
2
and =
1
+i
2
as
ds
2
=

2

2
[dX
1
+dX
2
[
2
B
12
= 2
1
. (3.3.6)
There is a large discrete group of equivalences. The reparameterization
X
1
= X
1

d +X
2

b, X
2
= X
1

c +X
2

a (3.3.7)
for a, b, c, d integers such that ad bc = 1 preserves the periodicity. The transformed
background is

=
a +b
c +d
,

= . (3.3.8)
Notice that this is exactly the same as the modular transformation (1.8.6) of the torus, but
now acting on spacetime rather than the world-sheet. This equivalence is just a change of
basis vectors for the spatial periodicity. It is not at all stringyit also holds for a eld theory
in this spacetime.
There are other, stringier, equivalences. Consider the term in the action involving B
12
,
B
12
4
_
d
2
z
_
X
1

X
2

X
1
X
2
_
. (3.3.9)
9
A loose end: the constraint from only forces V
a
to be a gradient locally, so around a closed curve
d
need not be single-valued. Thus we have related the theory at R = to the
d
theory at R = 0. With
careful attention to surface terms one can extend this to nite R.
87
The integrand is a total derivative, (X
1

X
2
)

(X
1
X
2
). This would seem to imply that
the theory is independent of B
12
, but we have to be careful because X
1,2
need not be periodic
on the world-sheet. Consider a toroidal world-sheet wound once on the toroidal spacetime,
X
1
= Re(z), X
2
= Im(z)/Im(
w
) (
w
being the world-sheet modulus). The action (3.3.9)
becomes iB
12
, so the path integral weight becomes e
iB
12
. This is invariant under discrete
shift
B
12
B
12
+ 2 + 1 . (3.3.10)
In addition there is simultaneous duality on the X
1
and X
2
. This takes E

= G

+iB

to its inverse. In terms of the parameterization (3.3.6) this is simply 1/. Again this
and the shift + 1 generate the full SL(2, Z)

=
a +b
c +d
,

= . (3.3.11)
There are a few other transformations. Duality on the separate axes takes (, ) (, ).
Spacetime parity, X
1
X
1
takes (, ) ( , ), and world-sheet parity takes (, )
(, ). (The last two are not symmetries of the heterotic string). In all, the full set of
dualities is SL(2, Z) SL(2, Z), up to some Z
2
factors. The space of backgrounds, which for
the single dimension was the half-line R

2, is here given by two copies of the modular

region of the torus, with some additional Z
2
identications.
Besides duality, the other equivalence of CFTs which has attracted a great deal of
attention is mirror symmetry [82]. This is an equivalence of N = 2 super-CFTs arising from
compactication on smooth manifolds, with the distinguishing feature that it ips the sign
of the U(1) current of the N = 2 algebra.
10
The equality of the Yukawa couplings (OPE
coecients) on a manifold and its mirror give relations between previously unconnected
mathematical structures. The most interesting physical phenomenon is the existence of
examples where a continuous change in the background elds of one manifold (and so a
continuous change in the CFT) maps to its mirror passing through a singular conguration
and changing topology.
The open string cannot wind around a periodic dimension, so one does not expect it to
be dual [84]. The normal open string has the Neumann boundary condition n
a

a
X

= 0.
10
Mirror symmetry is not the same as R-duality, except in special cases [83]. In general, the Lagrangians
of mirror-symmetric theories cannot be directly transformed into one another in the way we have done for
R-duality.
88
The duality transformation is
a
X
1
=
a
b

b
X
1
d
, so the boundary condition is
n
a

a
b

b
X
d
= t
b

b
X
d
. (3.3.12)
The tangential derivative vanishes, so X
1
d
is constant along the boundary. Moreover if we
consider points p
1
and p
2
on two dierent boundaries, we have
X
1
d
(p
1
) X
1
d
(p
2
) =
_
C
(dz X
1
d
+d z

X
1
d
) =
_
C
(dz X
1
d z

X
1
) = 4k
1
=
4n
R
= 2R
d
n.
(3.3.13)
Here C is any curve connecting p
1
and p
2
. We can imagine cutting the path integral open
along C in terms of the open string Hilbert space. The mode expansion (1.9.16) then relates
this to the momentum k
1
in the Neumann picture. The result is that the points dier by
precisely a multiple of the periodicity of the dual spacethey are at the same point. It also
follows from the mode expansion (1.9.16) that ipping the sign of the right-moving part of
X converts Neumann to Dirichlet boundary conditions. Taking R 0, and so R

, we
have a space in which all string endpoints are constrained to move on a hyperplane of xed
X
1
d
. Open strings are thus found only at this hyperplane, while the closed strings (which
as we have noted are present in any open string theory) are free to move everywhere. This
hyperplane is actually a dynamical object, the D(irichlet)-brane [84], [85]. The open string
state
1
1
[0) is a massless excitation along the D-brane, whose couplings are just those of a
transverse ripple of the hyperplane.
In the classication of string theories by CFT, world-sheet topology, and constraint al-
gebra, we have explored the idea that the rst of these corresponds to the vacuum of the
theory. Dierent world-sheet topologies, however, would seem to be truly distinct theories.
Remarkably, duality suggests that this is not the case: translating the D-brane o to innity,
one obtains in the limit a theory of closed strings only. A similar result holds for unoriented
theoriesthe dual theory has an extended object (the orientifold), away from which there
are only oriented world-sheets [84].
3.4 N = 0 in N = 1 in . . .?
The third part of the Schwarz classication is the world-sheet constraint algebra. A recent
argument of Berkovits and Vafa [86] suggests that this too is determined by the vacuum. Let
T
m
be the c = 26 CFT for any bosonic string theory. Add in a =
3
2
bc system (b
1
, c
1
), which
89
has central charge 11. Then the following energy-momentum tensor and supercurrent form
a c = 15 N = 1 super-CFT:
T = T
m
: b
1
c
1
:
1
2
: b
1
c
1
: +
1
2

2
(c
1
c
1
)
T
F
= b
1
+c
1
T
m
+ : c
1
c
1
b
1
: +
5
2

2
c
1
. (3.4.1)
This super-CFT can be used as the matter CFT for the N = 1 superstring. It is not in the
general class described in section 2.6 because the non-unitary part is not of the standard
X
0

0
form (the b
1
c
1
theory must be nonunitary because the central charge is negative). But
it might arise from backgrounds of elds with nontrivial time components.
Now, what seems remarkable is that the BRST cohomology of this N = 1 theory is
identical to that of the N = 0 theory based on T
m
, the constraints from T
F
removing the
b
1
c
1
degrees of freedom. Moreover, the amplitudes of the two theories are the same, the
b
1
c
1
path integral canceling the
1

1
path integral in a nontrivial way [86].
11
So the bosonic
string theories would indeed seem to be vacua (in the broad sense of super-CFTs) of the
N = 1 string.
This can be carried furthera general N = 1 string can in this way be embedded in
N = 2, which can be embedded in N = 3, and so on indenitely [87]. It can also be embedded
in the W
3
string and generalizations [88]. Another chain of embeddings is possible in a series
of linear higher spin algebras w
N
[89]. This is true even though the N > 2 superconformal
algebras and N > 2 w
N
algebras do not have at spacetime realizations of the ordinary sort.
This begins to seem like too much of a good thing.
Indeed, the following example may help to put this in perspective. Consider a eld theory
with some complex scalars
i
(x) and a Lagrangian density L(
i
) of no special symmetry.
We will make this look as though it has a local U(1) symmetry
i
e
iq
i
(x)

i
for arbitrary
choice of q
i
. First, add a scalar eld (x) and dene
L

(
i
, ) = L(
i
e
iq
i

)
1
2

. (3.4.2)
This is invariant under the global symmetry

(x) = (x) +,

i
(x) = e
iq
i

i
(x). (3.4.3)
11
According to Distler (private communication) there may be a problem on higher-genus surfaces.
90
Now add a gauge eld A

without a kinetic term,

L

(
i
, ) = L(
i
e
iq
i

)
1
2
(

)(

) . (3.4.4)
This is invariant under the local symmetry

(x) = (x) +(x),

i
(x) = e
iq
i
(x)

i
(x), A

(x) = A

(x) +

(x) (3.4.5)
But in fact L

describes the same theory as the original L. By the gauge choice (x) = (x)
we can set

= 0. Then A

decouples and has a trivial gaussian path integral, leaving L.

One has to think that this symmetry does not mean anything, since it is completely
independent of the original theory. What we have done is to gauge a nonlinearly realized
symmetry. Non-linear in this context refers to the term (x) in the transformation of ,
which is zeroth order in the elds; nonlinear terms of order eld-squared and higher would
not have the same eect. We see that a gauged nonlinear symmetry is like no symmetry
at all. This is the second time (out of three) that we encounter this theme: that a gauge
symmetry is after all just a redundancy, though sometimes a very useful one, and we can
always be more redundant.
Supersymmetry also has nonlinear realizations (in fact they were discovered quite early
in the subject). Kunitomo [90] argues the above procedure for world-sheet supersymmetry,
applied to the bosonic string, gives the Berkovits-Vafa construction (there are some quantum
corrections in the currents (3.4.1) that have to be found by hand). Indeed, the linear b
1
term
in T
F
means that the superconformal symmetry is nonlinearly realized on c
1
. It is then not
so surprising that one can make the bosonic string look like it has all these extra symmetries.
It is not clear what the moral is. This makes the embedding seem rather trivial, but it is
a reminder that the amount of gauge symmetry in a theory can be somewhat arbitrary, and
so the classication by gauge algebra is not so absolute. It would be telling if one could reach
the theory (3.4.1) by turning on background elds from some more familiar vacuum of the
superstring. By the way, it is somewhat odd to expect the N = 0 string as a ground state
of the N = 1 string, since the latter has fewer degrees of freedom (smaller matter central
charge), and even odder to get N = 1 from N = 2 since the latter has just a scalar. Earlier
attempts to go the other way, getting N = 1 as a ground state of N = 0, did not seem to
lead anywhere.
91
3.5 S-Duality
The last equivalence we have to discuss is S-duality, a conjectured equivalence between
weakly coupled and strongly coupled string theory [91]. Since the string coupling is e

, this
is to say that the vacua and

are actually the same state, where

runs from to
as does the reverse.
This idea is much more far-reaching than anything we have discussed thus far. All the
previous equivalences that we discussed held order-by-order in perturbation theory. S-duality
certainly does not hold order by orderit relates the perturbative expansion, around zero
coupling, to an expansion around innite coupling. Thus it is a statement about the exact
amplitude, and involves nontrivial relations among all orders of perturbation theory. My
purpose in the second half of these lectures was to try to go beyond string perturbation
theory, and S-duality is our rst example. It is a subject of great current interest, several
major papers having appeared just since I began writing these lectures. Unfortunately it
is a hard subject to present, because it is rather intricate and because it is not one I have
worked on in detail. So I will just try to summarize some of the main ideas.
12
To begin, consider the free Maxwell equations,

= 0

d
= 0, (3.5.1)
where F

d
=
1
2

. These are invariant under

F

d
, (3.5.2)
which interchanges the electric and magnetic elds. Note that the rst equation is an equa-
tion of motion, derived from the action, while the second is a Bianchi identity, which follows
from F

independent of the action. One can instead write F

d
as the curl of
a dual vector potential A
d
, and the equation of motion and Bianchi identity again change
roles.
The classical Maxwell equations remain invariant if we add both electric and magnetic
sources. This can be extended to the quantum theory provided the Dirac quantization
condition is satised [93]. That is, if a particle of electric and magnetic charge (Q
e
, Q
m
)
exists, and another of charges (Q

e
, Q

m
), then
Q
e
Q

m
Q
m
Q

e
2Z . (3.5.3)
12
A recent review by Sen [92] and a seminar given by Je Harvey at UCSB were very helpful.
92
The simplest solution to this is that states be restricted to the lattice
Q
e
= en
1
, Q
m
=
2
e
n
2
(3.5.4)
for integer n
1
and n
2
. A theory with an electrically charged eld (e, 0) and a magnetically
charged eld (0, 2/e) is thus invariant under the electric-magnetic duality (3.5.2) if we also
take e 2/e, interchanging weak and strong coupling. Unfortunately this is rather formal,
because either the electric or magnetic coupling is always strong, and we do not know how
to make sense of the resulting quantum eld theory.
The charge lattice (3.5.4) is not the most general solution to the Dirac quantization
condition. If a -parameter is added for the gauge eld, the electric charges shift by an
amount proportional to the magnetic charge [94],
Q
e
= en
1
+en
2

2
, Q
m
=
2
e
n
2
. (3.5.5)
This is invariant under +2, with n
1
n
1
n
2
. Electric-magnetic duality generalizes
to nonzero . It is useful to form the combination
=

2
+i
2
e
2
. (3.5.6)
Under electric-magnetic duality and + 2 we have respectively

1

, n
1
n
2
+ 1, n
1
n
1
n
2
, n
2
n
2
. (3.5.7)
These do not commute, and they generate the familiar SL(2, Z), found earlier as the modular
symmetry (1.8.6) of the torus and in section 3.3 as a duality symmetry with two compact
dimensions.
The allowed spectrum (3.5.5) is invariant under the SL(2, Z) duality, but it is a stronger
statement for the actual spectrum to be dual. In the example above, for example, only the
states (n
1
, n
2
) = (1, 0) and (0, 1) appeared and the spectrum was invariant under electric-
magnetic duality (in this Abelian example we are putting in elds by hand, but in the non-
Abelian case below the dynamics will determine the spectrum). Now consider what happens
to the magnetic state (0, 1) as we increase and so its electric charge. One scenario is that at
93
= , where its electric charge is
1
2
, it becomes degenerate with the state (n
1
, n
2
) = (1, 1)
with charge
1
2
, and that at somewhat larger it is unstable to decay into the latter state
plus the (n
1
, n
2
) = (1, 0) electric state. The SL(2, Z), if a symmetry, relates the theory at
any to that at some in the standard fundamental region. In the case just described,
there are two stable states everywhere, but at the boundaries of the standard fundamental
region there is a phase transition in the spectrum and the quantum numbers of the stable
states change.
To be precise, the discussion assumed that the state (n
1
, n
2
) = (1, 0) was massless at
= . If it is extremely light there is a narrow region where both magnetic states are stable,
and the phase transition separates into two. On the other hand, if it is heavy enough both
states are stable for all . In fact, in the supersymmetric case described below, the latter
situation holds and states cannot become unstable, at least for n
1
and n
2
relatively prime.
The states (n
1
, n
2
) = (0, 1) and (1, 1) must both be stable everywhere, as well as a large
number of others related by further SL(2, Z) transformationsnamely, all states with n
1
and n
2
relatively prime.
Now let us go on to the case of interest, the pure non-Abelian gauge theory. The gauge
eld is itself electrically, not magnetically, charged, which seems to introduce an essential
asymmetry. This shows up in the eld equations and Bianchi identities,
D

= 0
D

d
= 0, (3.5.8)
where D

is a covariant derivative, containing A

and not A
d
. These equations can no
longer be written in terms of a dual vector potential A
d
, and any attempt to rewrite the
action in terms of a dual vector potential seems to lead rapidly to a mess.
But maybe we just havent looked hard enough. Non-Abelian theories do have magnetic
monopole congurations, some of which become stable when the symmetry is spontaneously
broken to an Abelian group.
13
There is some circumstantial evidence for duality, especially
in N = 4 supersymmetric gauge theories (and in some N = 2 theories with matter). Namely,
1. The lattice of allowed electric and magnetic charges is invariant under duality (or in some
cases is related to that of a dual gauge group) [95], [96].
Also, when the non-Abelian symmetry is spontaneously broken to an Abelian symmetry,
13
As in the Ising model, duality would be expected to exchange such topological objects with the funda-
mental ones.
94
2. The low energy (Abelian) theory is invariant under duality.
3. The spectrum of massive gauge bosons and massive monopoles is invariant under duality.
4. The long-ranged forces between electric and magnetic charges (from gauge elds and also
massless scalars) are invariant under duality.
It may seem remarkable that one can make such statements, since we know so little
about strongly coupled theories. Actually, point 1 is just based on topology and should not
depend on the interactions. Points 2 to 4 use the supersymmetry in an essential way. The low
energy theory consists of Abelian gauge bosons and neutral scalars (and their superpartners)
without any renormalizable interactions, so is free at low energy even though the underlying
theory is strongly coupled. Moreover its form is xed by the N = 4 supersymmetry, so we
dont need to solve the underlying theory to nd it. What is left is the familiar Abelian
duality.
Points 3 and 4 use the supersymmetry in a richer way [97].
14
Let us consider the following
algebra,
Q
1
, Q

1
= H +Q, Q
2
, Q

2
= H Q
Q
i
, Q
j
= Q
i
, H = Q
i
, Q = Q

i
, Q

j
= Q

i
, H = Q

i
, Q = 0. (3.5.9)
Here Q
i
are two supercharges, whose anticommutator includes the Hamiltonian H (as always
in supersymmetry) and also a conserved charge Q (as is often the case). You can think of
this example as living in zero space dimensions so that there is no momentum or spin to
worry about. Going to eigenspaces of H and Q with eigenvalues h and q, the supercharges
form two fermionic oscillators, with the standard representation in terms of 2
2
states,
Q
1
[ ) = Q
2
[ ) = 0, Q

1
[ ) =
_
h +q [ ), Q

2
[ ) =
_
h q [ ), (3.5.10)
and so on, with the normalization factors hq coming from the algebra. Note the inequality
h +q = [Q
1
, Q

1
[) = |Q

1
[)|
2
+|Q
1
[)|
2
0, (3.5.11)
and the same for h q, so
h [q[. (3.5.12)
Now, when the inequality is saturated, one of the terms in (3.5.10) vanishes and a small
representation with 2 states becomes possible. This idea plays a key role throughout super-
14
As an aside, spacetime supersymmetry and its breaking are key issues in nonperturbative string theory,
but ones that I will have to largely neglect because of restrictions of time and emphasis. This is the one
point where I will discuss supersymmetry in any detail.
95
h
q
Figure 18: Allowed states in the h-q plane (shaded). Small representations occur on the
boundary.
symmetry. Now note that if we have a small representation and we make continuous changes
in the parameters in the Hamiltonian, the energy of the states cannot changeit must stay
at h = [q[ because to move to h > [q[ we would need four states.
15
The N = 4 supersymmetry algebra has regular representations of 2
8
states and small
representations of 2
4
states. The masses, electric and magnetic charges satisfy the Bogo-
mol nyi-Witten-Olive bound [97]
M
2
v
2
(Q
2
e
+Q
2
m
), (3.5.13)
where v is the symmetry-breaking expectation value (this is for a single U(1) group). Small
representations saturate the bound. In the perturbative, small g, theory, both the charged
gauge bosons and the stable monopole solutions are in small representations, and moreover
one nds the same spin spectrum in each case. Now vary the coupling g. Small represen-
tations have to stay small, and saturate the bound, so by the time we get to g

4/g
the spin and mass spectrum is the same as we started with but with electric and magnetic
charges interchanged.
I have never known what to make of thisis it a simple consequence of supersymmetry
restricted to the Bogomol nyi sector, or is it evidence for a duality of the full theory? I have
always been skeptical, again because the duality seems to lead to a mess at the Lagrangian
15
This would be possible if there happened to be another small representation, of opposite fermion number,
at the same value of q.
96
level. However, notice that while the BWO bound determines the allowed spectrum of small
representations, but it is a further non-trivial fact that the actual spectrum found is dual [98].
The nal bit of evidence for duality, the long-ranged force, is also a consequence of the BWO
bound.
Everything I have just said was known before 1980 for the electric-magnetic duality. Sen
recently observed that the SL(2, Z) would have additional implications ([92] and references
therein). In particular, it requires stable states of monopole charge n
2
greater than one, which
must appear as bound states of the n
2
= 1 states. Some of these have now been found [99].
Recently, long papers by Seiberg and Witten [100], [101] and Vafa and Witten [102] on
N = 2 and N = 4 supersymmetric gauge theories have appeared, which seem to present
further evidence for duality, though I have not absorbed these. All of this evidence refers in
a sense only to the Bogomol nyi sector of the spectrum, but so much is accumulating that it
is harder to believe that it is not a symmetry of the full spectrum.
Thus far the discussion has involved only eld theory, N = 4 supersymmetric Yang-Mills.
For the heterotic string compactied on a six-torus (that is, = 4, . . . 9 periodic) the low
energy theory contains N = 4 Yang-Mills, and the conjecture is that this string theory also
is self-dual [91], [92].
16
The circumstantial evidence is of the same type as the above. Let
me just note a few important dierences. The coupling constant g
2
is now a eld, being
proportional to e

. But so also is the topological angle : it is proportional to the axion

eld a obtained from the antisymmetric tensor.
Exercise: Show that for the two-index potential B

, if one interchanges the Bianchi identity

for the eld strength (vanishing curl) with the equation of motion (vanishing gradient), in
four dimensions one obtains a massless scalar eld. Show that in ten dimensions one obtains
a 7-index antisymmetric tensor eld strength which is the curl of a 6-index potential, and
which is invariant under a 5-index gauge transformation.
The low energy eld theory for the dilaton, axion, and gauge elds, and their supersymmetric
partners, is SL(2, Z) invariant.
The other dierence is that the BWO sector, which in the eld theory case consisted of
a few small representations, is now very much larger, with innite numbers of electrically
charged string states and magnetically charged soliton states, of various types. The duality of
the allowed states follows from supersymmetry as above, but duality of the actual spectrum
implies much more, including stringy monopoles which involve elds not in the low energy
16
For the heterotic string theory on backgrounds of lower symmetry, S-duality might also change the
background.
97
eective eld theory. At this point it is much less clear whether the actual spectrum is
S-dual. Incidentally, Sen has also argued that S-duality, like R-duality, must be a gauge
symmetry in string theory.
Finally, another conjecture is that string theory is dual to a theory of fundamental ve-
dimensional objects (ref. [103] and further references in ref. [92]). The evidence is of the
same type as the aboveduality of the low energy eld theory, and interchange of the
spectra of fundamental objects and solitonic ones. These arguments are made even though
it is not known if it is possible to quantize the fundamental ve-branethe world-sheet
theory is a six-dimensional eld theory and so non-renormalizable. They are made on the
basis of presumed low-energy eld theory of the ve-brane (it couples naturally to a six-index
antisymmetric tensor gauge eld; see the above exercise), on its classical soliton solutions,
on scaling arguments based on the world-sheet and spacetime theories, and on the possible
winding states of a ve-dimensional object. An interesting connection between this and
S-duality is that
S
string
= (5-brane string)(R
5-brane
)(string 5-brane) , (3.5.14)
so string/5-brane duality would imply S-duality, given the more straightforward R-duality.
4 String Field Theory or Not String Field Theory
Thus far, we have dened string theory only through its perturbation expansion, the analog
of Volume One of Bjorken and Drell. Now we would like to nd Volume Two. We know that
in eld theory there are many important phenomena that cannot be seen in perturbation
theory, and string theory will have all of these and probably more. Also, we have found
that string theory contains an enormous amount of spacetime gauge invariance, in the form
of spacetime Ward identities satised by the scattering amplitudes. But again, we know
from Yang-Mills theory and general relativity that this is a very clumsy way to think about
spacetime gauge invariance. These theories of course have a geometrical interpretation which
is essential to understanding the physics, and which is disguised in the perturbation theory
but is evident when they are written as eld theories. So it seems that we should try
something similar in string theory, to introduce some sort of string eld , to nd an action
with the appropriate gauge symmetries, and then to recover the perturbation theory from a
gauge-xed path integral. This section consists of a few assorted remarks about this idea,
focusing rst on some very attractive features, and then on some indications that it may not
be quite the right thing to do.
98
4.1 String Field Theory
A closed string eld will create or destroy a string along some closed curve in spacetime,
so it is a functional [X] of such paths. An open string eld will be a functional of open
curves. In the earlier discussion, the closed and open string wavefunctions [X] were similarly
functionals of the conguration of the string in spacetime, so this is like second quantization,
promoting the one-particle wavefunction to an operator (or, in the path integral formalism,
to a variable of integration). It turns out that things work very much more nicely if one rst
goes to the BRST-invariant form of the theory, and then second quantizes [104], [105].
1
In
this form the wavefunction also includes the state of the ghosts. Since the b and c ghosts are
conjugate, we can for example regard b as the momentum and c are the coordinate, so the
wavefunction would be a functional [X, c], and the corresponding eld a functional [X, c].
Actually, it is still convenient to use the bra-ket notation, writing the string eld as [),
which is just the abstract notation for the functional [X, c] = X, c[). We would have to
use a dierent notation for the states of the theory, something like |). That is, |) is a
state with any number of strings, and [X, c] or [) act as an operators on these states.
We can think about this another way by expanding the functional [) in terms of a
complete set of such functionals. For the open string, for example,
[) =
_
d
26
k
(2)
26
_
T(k)[0, k) +iA

(k)

1
[0, k) +B(k)b
1
c
0
[0, k) +. . .
_
. (4.1.1)
The functions T(k), A

(k), B(k), . . . or their Fourier transforms T(X), A

(X), B(X), . . .
are the arbitrary coecients in the expansion. There is an obvious interpretation of T(X)
as the spacetime tachyon eld and A

(X) as the spacetime gauge eld, while B(X) will turn

out to be an auxiliary eld.
2
Now we need an invariance principle. Since wavefunctions Q
B
[) are equivalent to zero,
the natural guess is
[) [) +Q
B
[) (4.1.2)
for any [). Taking
[) =
_
d
26
k
(2)
26
_
(k)b
1
[0, k) +. . .
_
, (4.1.3)
1
String eld theory is a large subject. I will make a few appropriate references, but the reader should
consult the reviews [106], [107], [67] for extensive references.
2
In the expansion (4.1.1), I have for brevity kept only terms with the same ghost number as the ground
state; more on this later.
99
(so that Q
B
[) has the same ghost number as [)) one nds from the expansion (2.5.23) of
Q
B
that the invariance (4.1.2) becomes
A

(X) = 2

(X) (4.1.4)
which is indeed the linearized gauge invariance of electromagnetism.
There is an obvious free eld equation,
Q
B
[) = 0, (4.1.5)
which is invariant under (4.1.2) because Q
2
B
= 0. In components, this becomes

2
T(X) =
1
2
T(X), B(X) =

(X)

2
A

(X) =

B(X) =

(X). (4.1.6)
We obtain the appropriate Klein-Gordon equation for the tachyon and the gauge-invariant
free Maxwell equation for A

. Note that the eld equation (4.1.5) has the same form as the
physical state condition which earlier gave us the mass-shell conditions. The one dierence
is that we required b
0
to annihilate physical states, which at this level leads to B(X) = 0
and so
2
A

(X) = 0. So the b
0
condition is a stringy generalization of Feynman gauge.
We can readily write down an invariant free action as well. It is simply
S
open
=
1
2
[Q
B
[). (4.1.7)
The ghost number works out so that the action for the elds (4.1.1) is non-zero. Writing
the inner product in terms of the path integral on the disk, the vertex operators 1

each
have ghost number 1, as does Q
B
, adding up to 3 as required by the same calculation (1.9.4)
as on the sphere. Not surprisingly, the action for A

is the free Maxwell action, after B is

integrated out. We have used the bilinear inner product [ ), but in fact the string eld
must satisfy a reality condition
[ = [. (4.1.8)
This is necessary for unitarity, generalizing the familiar fact that the metric and Yang-Mills
elds must be real.
This all seems very beautiful to me. The familiar spacetime gauge invariance is embed-
ded as the lowest component of a much larger symmetry acting on all higher levels of the
100
string, and the BRST formalism allows this to be done in an extremely compact and elegant
way. In fact, the action (4.1.7) has the same structure as the Abelian Chern-Simons action
_
d
3
X AdA [105], where Q
B
is analogous to the exterior derivative d: note that both are
nilpotent. Incidentally, the components of other ghost number, not written in the expan-
sion (4.1.1), just play the role of Fadeev-Popov ghosts for the spacetime gauge symmetry
after the gauge is xed, and one obtains a BRST-invariant string eld theory. To be precise,
there are several complications relative to the earlier discussion of BRST symmetry on the
world-sheet, so a generalization of BRST, the Batalin-Vilkovisky formalism, emerges [106].
This all generalizes to the closed string, with a slight complication. It is necessary to
impose conditions (b
0

b
0
)[) = (L
0

L
0
)[) = 0 on the closed string eld and on the gauge
parameter [), and the action is
S
closed
=
1
2
[(c
0
c
0
)Q
B
[). (4.1.9)
The ghost number is then correct. The invariance (4.1.2) is now an innite-dimensional
generalization of linearized coordinate invariance.
Now the question is whether we can nd an appropriate nonlinear generalization of the
string gauge symmetry and action. Let us transform this into a related question. After
gauge xing, the path integral over the string eld can be expanded perturbatively in terms
of Feynman diagrams built from string propagators and vertices. Is it possible to represent
the amplitudes, which were described earlier in terms of a sum over world-sheets, in this
way?
I am going to make a distinction here between what I will call an eective string eld
theory and a bare string eld theory, and describe the former rst. What I am call an eective
string theory is described in ref. [67]. It is always possible by brute force to write the string
amplitudes in terms of Feynman diagrams. Let us introduce an arbitrary three-closed-string
vertex, as depicted in g. 19a. Figure 19b shows a four-string amplitude built from two
such vertices. The propagators are cylinders, with lengths integrated from zero to innity
(generalizing the Schwinger representation of the propagator in eld theory). There are three
such graphs, from the three channels. Now recall from the Virasoro-Shapiro amplitude (2.8.1)
that this amplitude is supposed to be given by an integral of one vertex operator position z
4
over the complex plane, the other three operators being xed. The graphs of g. 19b cover
three round regions centered on the three xed positions, g. 20a, but this inevitably leaves
101
a) b)
Figure 19: a) Three-string vertex. b) Four string amplitude.
1
z
4
a) b)
Figure 20: a) Graph of g. 19b covers three round regions in the complex plane. b) Four-
string vertex needed to cover unshaded region in g. 20a.
a region between uncovered.
3
So it is necessary to introduce a four-string vertex as shown
in g. 20b. An integration over shape in included in the denition of the vertex so as to
cover exactly the missing region of g. 20a. Going on to higher amplitudes, one must further
introduce n-string vertices for all n. At tree level the vertices have the topology of spheres
3
It also could be that the round regions overlap and double-cover some part of the z
4
-plane; one can
avoid this by making the stubs on the three-string vertex long enough. Also, by taking a more complicated
propagator it may be possible to cover the four-point amplitude correctly, but this will fail for the ve-point
amplitude.
102
with n holes. However, when one goes on to loop amplitudes, the tree level action again
does not cover moduli space, so it is necessary to add additional vertices containing internal
loops. This procedure can be carried out iteratively, and in the end the full perturbation
series is written in terms of Feynman graphs. Ref. [67] gives an explicit construction of one
possible set of vertices. All of this can also be applied to theories of open plus closed strings.
Now one can work backwards, writing the n-string vertex as a term in the action with a
product of n string elds. The resulting action does indeed have a non-linear generalization of
the string gauge symmetry. This is not surprising for the following reason. In the discussion
of spacetime Ward identities in section 2.7, I emphasized that these would hold if and only if
the integration ran over the correct moduli space. The iterative construction of the vertices
does this, so the spacetime Ward identities hold and a corresponding invariance should be
present in the action. The condition that moduli space be properly covered is equivalent to
a set of identities for the vertices, and these same identities imply the nonlinear invariance
of the action [67].
4.2 Not String Field Theory
I have called the above construction an eective string eld theory because it is very similar
to a Wilsonian eective eld theory. The long-distance physics is explicitly represented,
while the short-distance physics is already integrated out. (A similar analogy was made in
refs. [108]). In particular, we have seen that long-distance propagation in spacetime, pro-
ducing the poles in string amplitudes, comes from long cylinders or strips at the boundaries
of moduli space. In the eective string eld theory, the vertices include an integration over
moduli, but only an interior region of moduli space, as in the example of g. 20. The de-
generating cylinders appear explicitly as propagators. On the other hand, we will see at
two points later on that stringy physics comes from the interior of moduli space, which are
hidden inside the vertices.
Eective string eld theory is very useful for some purposes. By describing explicitly the
boundaries of moduli space it is a useful tool for demonstrating the niteness and unitarity
of string perturbation theorysee the discussion at the end of section 2. But it would not
seem to be the right tool for studying nonperturbative string theory, since the stringy physics
is already integrated out. Can we do better? For the open string we certainly can [105].
The simple Witten vertex shown in g. 21, in which three open strings are joined by gluing
their halves together in pairs, exactly covers moduli space. The interacting action thus has
103
Figure 21: Witten vertex for three open strings.
a nonlinear generalization of the string gauge invariance, and has the same structure as the
non-Abelian Chern-Simons action. I will call this a bare string theory, since the action has
a closed form without corrections from all orders in the loop expansion.
The perturbative formulations of open and closed string theories were quite parallel. For
example, the spectrum of the closed string was essentially the tensor product of two copies of
the open string spectrum. But now we run into a real asymmetry, for there does not seem to
be any covariant closed string eld theory simpler than the eective theory that I described
earlier.
4
The existence of n-point tree level interactions would not have been surprising, since
general relativity is non-polynomial while Yang-Mills theory is polynomial. But the fact that
the action receives corrections from all orders in the quantum loop expansion means that a
great deal of non-trivial physics is already integrated out.
There is another, rather remarkable, asymmetry between open and closed strings. The
Witten open string theory covers the moduli space of Riemann surface with boundaries. But
as I have emphasized at the end of section 2.5, this will include processes with intermediate
closed strings. The simplest example of this is in g. 22, the one loop vacuum amplitude,
with one propagator and no vertices. This has a modulus t, the ratio of the circumference to
the length, which runs from 0 to . For large t, g. 22a, this looks like a vacuum uctuation
consisting of a pair of open strings. But for small t, g. 22b, it looks like a single closed
string appearing and then disappearing. Indeed, the string graph contains both processes.
The same thing happens at for higher order open string amplitudes. While we can think of
the Witten theory as a path integral over open string elds, it contains the full open plus
4
I exclude here the non-covariant light cone string eld theory. Possibly this is the right approach, but
there are many things about it that I do not understandsubtleties with contact terms, and with the vacuum
structure. Also, there was some development of light-cone-like covariant string eld theories, but these seem
to have had diculties.
104
a) b)
Figure 22: a) Annulus at large t. Cutting on dashed line gives an open string pair. b)
Annulus at small t. Cutting on dashed line gives a single closed string.
closed string physics [109].
It is not clear what the logical relation is between the closed and open strings here. It
is not that the closed strings are bound states of open stringsthere are no interactions in
g. 22. In some sense the closed strings are singular congurations of the open string eld.
Siegel has suggested recently that the relationship is similar to bosonization, though I do
not grasp all his arguments [110]. Perhaps the lesson is that open string eld theory is all we
need, with the closed string emerging as in g. 22. This idea was pursued for a while, but
it is not clear where to go with it [111]. There is also the diculty that the most promising
theory, the heterotic string, does not have an open string version.
The graph in g. 22 is the same as the one-loop vacuum amplitude in eld theory, with
t the Schwinger parameter, and it corresponds to summing
1
2
times the spacetime volume
over all open string modes. It is interesting to contrast this with closed string theory.
Figure 22 is obtained by taking an open string propagator, a strip of length/width= t,
gluing the ends and integrating over t. The analogous construction in closed string theory
would be to take a cylinder, gluing the ends, and integrating over lengths and also twist
angles. This would correspond to integrating over tori, with the modulus running over
the full region Re() <
1
2
, Im() > 0. This is not the integration region, g. 6, for closed
string theory. So the closed string vacuum amplitude is not given by summing
1
2
over closed
string frequencies, a strong indication that closed string theory is not a eld theory [112].
The open string vacuum amplitude is given by summing
1
2
over open string frequencies,
105
though it includes also the closed string process of g. 22b.
There is one more important non-eld-theoretic property of closed strings, having to do
with the large-order behavior of the perturbation series. Consider a quantum eld theory
where the coupling constant g appears only as an overall factor g
2
in the action, so that g
2
is
the loop counting parameter. Rather generally, the large order behavior of the perturbation
series

h=0
g
2h
T
h
(reviewed in ref. [113]) is
T
h
h!h
A
C
h
(4.2.1)
with A and C constants. The dominant factor here is the factorial. Generically this arises
simply because the number of Feynman graphs with h loops is of order h!if one cuts o the
momentum integrations in the IR and UV the propagator and vertex factors just enter into
the constants A and C.
5
The perturbation series is thus divergent. The ratio of successive
terms is
g
2h
T
h
g
2h2
T
h1
hC
1
g
2
, (4.2.2)
so the smallest term is for

h Cg
2
. From Stirlings approximation,
A

h
e
C/g
2
. (4.2.3)
This is the smallest term, so represents the maximum accuracy of perturbation theory as an
asymptotic series. Indeed, many non-perturbative eects, eects which do not occur at any
order of perturbation theory, are of this magnitude, including connement, chiral symmetry
breaking, supersymmetry breaking in supersymmetric gauge theories, instantons, even BCS
superconductivity.
Open string theory is much the same [114]. The Witten eld theory represents moduli
space as a sum of Feynman graphs, so the factorial growth of the number of graphs im-
plies a factorial volume of moduli space. The integrand can be bounded in the interior of
moduli space, so again barring infrared divergences the same estimate (4.2.1) holds, and
nonperturbative eects of order e
C/g
2
o
are expected.
The same argument does not apply to closed string theory, because the eective string
eld theory has complicated vertices containing integrals over moduli space. Shenker esti-
mates the large order behavior as follows [115]. He uses the fact that Wittens open string
5
But in some cases there are also contributions of order h! coming from the momentum integrations in a
small subset of graphs. I do not know if there is any deep reason why two such dierent sources give eects
of the same order.
106
eld theory generates the amplitudes of the full open plus closed string theory. He then
argues that the purely closed-string amplitudes are a non-negligible portion of the full am-
plitudes, so that the perturbation theory for them grows as rapidly. Nonperturbative eects
are then of order
e
C/g
2
o
= e
C/gc
. (4.2.4)
In terms of the closed string perturbation theory, this translates into large order behavior
proportional to (2h)!, as had been discovered in the matrix models.
These e
C/gc
eects have no analog in eld theory, and their nature is not known. We
will see what they are in the matrix model, but it has been hard to guess how to generalize
the result. At small g
c
these stringy eects are larger than the familiar and important e
C/g
2
c
nonperturbative eects in the low energy eld theory. Moreover, they are likely to involve
phenomena which are unique to string theory. Of all the things I am covering in these
lectures, this is the one where I most wanted to be able to say something new. I have made
one observation [116], which is that if one includes boundaries with Dirichlet conditions,
X

=constant, these act as instantons but with a weight of the desired form e
C/gc
. The
conjecture is that the stringy nonperturbative eects make their appearance as a sum over
various kinds of boundaries; I am trying to test this in the matrix model.
In lieu of anything solid to say, let me make a conjecture: that closed string eld theory is
simply wrong nonperturbatively, and that closed strings themselves are collective excitations
of some other degrees of freedom, in terms of which the theory should be formulated. Let
me list some evidence for this.
The rapid growth of perturbation theory, suggesting that the perturbative description
is rather far from the exact formulation.
The fact that closed string eld theory does not seem to work, except as an eective
theory with the stringy physics integrated out. Note that the rapid growth of pertur-
bation theory implies that the interior of moduli space, the part which does not look
like Feynman graphs and is integrated into the vertices, is very large.
The appearance of closed strings in open string eld theory.
Matrix models, where closed strings are collective (bosonized) excitations of free fer-
mions.
The closed string is almost the product of right- and left-moving theories, suggesting
that it might be useful to regard it as a bound state of right-moving and left-moving
107
strings. In the at space theory, there are two things which glue the two sides together.
The rst is the equality of the zero modes eigenvalues,
0
=
0
. The second is the global
structure of moduli space. Locally, the moduli space has a natural complex structure
corresponding to the right-left separationfor example, the modulus of the torus
appears as L
0

L
0
. But the region of integration is not any sort of product.
Too much gauge symmetry. Here I am being contrary, since earlier I told you that it
was wonderful that string theory embedded the spacetime gauge symmetries in a much
larger structure. There I was following the usual particle theory paradigm that local
symmetry is holy, and that as one goes to higher and more fundamental energies one
expects to see more and more of it, as in SU(3)U(1) SU(3)SU(2)U(1) SU(5)
of the Standard Model and GUTS. But this need not be the case. As I have tried
to emphasize, gauge symmetry is just a useful redundancy, and there are examples
where one emerges at low energy even though there is no sign of it in the underlying
theory. The familiar example from particle theory is the CP(n) sigma model [117].
In condensed matter physics, this has been proposed to occur in theories of strongly
coupled electrons, where the electron separates into a spinon and holon,

e
(x, t) =
s
(x, t)(x, t). (4.2.5)
This decomposition is redundant, the transformation

s
(x, t) = e
i(x,t)

s
(x, t),

(x, t) = e
i(x,t)
(x, t) (4.2.6)
leaving the physical eld, the electron, invariant. It is plausible that under some
conditions this redundancy is elevated to a dynamical symmetry (for a review see
ref. [118]). As far as I know there is no reason in principle that this cannot happen.
6
So
perhaps the short distance theory, rather than exhibiting an enormous gauge symmetry,
should be formulated entirely in terms of invariants.
Recall from the discussion in section 3.4 that one way to distinguish a useless from a
useful redundancy is to see whether the elds transform inhomogeneously. The string
gauge symmetry is
[) [) +Q
B
[) +O([)) (4.2.7)
The linear term is as before, eq. (4.1.2). We do not know the higher terms in any
simple form but we do not need them, because it is precisely the term Q
B
[), of zeroth
6
Ref. [119] shows that in some circumstances a global symmetry of the underlying theory cannot be pro-
moted to a local symmetry at long distance, but it imposes no restriction on the promotion of a redundancy,
which acts trivially in the underlying theory.
108
order in [), that is relevant here. If this vanishes (at least for some momentum), the
gauge symmetry is real. So we are looking for solutions of
Q
B
[) = 0. (4.2.8)
This looks like the physical state condition, but it is dierent because the [) have
ghost number one less than the states, so it is a dierent cohomology. In fact, the only
solutions for a at background are = X

c or c

, which are just the translations,

so the only real gauge symmetries are those of the graviton and antisymmetric tensor.
But this need not be conclusivethere are other backgrounds where other parts of the
string gauge symmetry are unbroken, as we will see in section 5, so it may be that it
is useful to keep the full redundancy.
The observation that holes in the world-sheet naturally give eects of order e
C/gc
[116], suggesting a breakdown of the world-sheet.
This is the set of ideas I play with, though it does not yet add up to anything coherent.
The idea that we should look for more fundamental degrees of freedom in string theory
was put forward in ref. [120], in a study of the high-temperature behavior of string theory.
This discusses a number of other non-eld-theoretic properties of closed strings, and also
suggests breakdown of the world-sheet by way of holes. Another idea [121], [122] is that the
string should be thought of as a collection of bits; the string picture breaks down when the
density of bits becomes large. It may be that these ideas are connected, holes appearing in
the world-sheet because neighboring bits unbind.
4.3 High Energy and Temperature
In particle physics the traditional way to nd out what things are made of is to bang them
together. I will briey describe three dierent high-energy regimes, each of which gives a
dierent picture. We might hope that in one limit or another the theory will simplify enough
to allow us to go beyond perturbation theory.
The rst limit is scattering at high center-of-mass energy, E = s
1/2
, and xed angle.
This is where Rutherford found the atomic nucleus, and where SLAC found the partonic
constituents of hadrons, quarks and gluons. In relativistic eld theory this process probes
distances of order E
1
. We have already seen in the tree-level Virasoro-Shapiro amplitude
109
that the amplitude is very soft in this limit. To get further insight[123] consider the path
integral over X

from which it was obtained,

_
[dX] e
S
P
+i

4
i=1
k
i
X(
i
)
, (4.3.1)
the second term in the exponent being from the tachyon vertex operators. The limit is
essentially obtained by scaling up the k
i
uniformly. The action S
P
is quadratic in X, so the
path integral is determined by a saddle point X

cl
E, and the amplitude is e
O(E
2
)
. This
same saddle point dominates if the tachyons are replaced with other states of the string; in
particular, the masses of the external states drop out in the limit. Also the z
4
integral in
the resulting amplitude (2.8.2) is dominated by a saddle point, at
z
4
=
k
1
k
4
(k
1
+k
2
) k
4
. (4.3.2)
Note that this is in the interior of moduli space, another example of the idea that stringy
behavior comes from the interior.
In contrast to eld theory, the size of the interaction region grows with energy [123].
Combining the low energy eld theoretic behavior with this high energy result, the eective
uncertainty is
X
1
E
+

E. (4.3.3)
This is further evidence for an eective minimum distance in string theory, as found earlier
from duality. It appears that the string can carry only of order one string unit of energy
per unit length, so to transfer a much larger energy E many bits are neededthe eective
number of partons is proportional to the energy.
Gross and Mende were able to nd the dominant saddle point at every order of pertur-
bation theory. It is just an n-fold cover of the tree-level saddle point, scaled by a factor n
1
,
with n1 = h being the number of loops. The exponential in the scattering amplitude (2.8.4)
is suppressed by n,
c
n
e
sf()/n
(4.3.4)
so at large s high orders dominate. This can be interpreted in terms of dividing the scattering
into n softer scatterings of angle /n. There has been some discussion of the summation of
this series [124], and of a large symmetry in the high energy limit [125], assuming that it is
correct simply to sum the leading behavior from each order.
7
It is interesting to contrast
7
There has also been a complementary study, of the all-orders summation of small-angle scattering [126].
110
this with the earlier discussion of the large-order behavior. The amplitude (4.3.4) comes
from a single saddle point in moduli space, so is unrelated to the large volume noted earlier.
However, it is enhanced by the kinematics and actually grows much faster than the volume
of moduli space, c
n
being proportional to (9h)!.
The second limit is soft scattering, high energy and small angle, holding xed the mo-
mentum transfer q (t)
1/2
. Whereas the hard scattering is an occasional rare process
where many bits of string move together, the soft scattering is a picture of the typical string
conguration. From Stirlings approximation, the Virasoro-Shapiro amplitude is of order
A
VS
s
2+t
t
1
=
s
2
t
e
q
2
(lns)
. (4.3.5)
This is the gravitational amplitude s
2
/t modied by a form-factor which corresponds to an
object with a size of order

lnE [127]. There is another way to see this same result. Let us
calculate the root-mean-square size of the string ground state. The mode expansion gives
0[(X
1
() X
1
)
2
[0) =

m=1
1
m
2
0[(
m

m
+
m

m
)[0) = 2

m=1
1
m
. (4.3.6)
This is divergent but has no direct physical signicance. A measurement on a time-scale t
is sensitive only up to modes of frequency less than t
1
, essentially mt < 1. So the log
divergence becomes ln(t
1
), and the size is the square root of this [127]. This agrees with
the form factor, where the soft scattering probes time scales t 1/E.
Just as the hard scattering (4.3.3) is interpreted as a stringy uncertainty principle,
Susskind interprets this root-log-E growth as a stringy Lorentz transformation, transverse
sizes not being constant as in classical physics. A similar calculation shows that the longitu-
dinal size of the string goes to a constant of order the string scale, not the usual contraction
as 1/E.
The square root of a logarithm is a very slow function, and ordinarily would be of little
importance. But there is one situation where enormous boosts are encountereda black hole.
From Strominger and Verlinde you have heard two very dierent points of view on whether
short distance physics, such as string theory, can be relevant to the information problem.
What makes this so controversial is that in a black hole there are no large local invariants,
such as would ordinarily be needed for the low energy eective eld theory to break down,
but there is a very large nonlocal invariant, the relative boost between an infalling and an
asymptotic observer. The external observer sees the infalling observer slow down and sit
111
forever (or until the hole decays) on the horizon. One tick of the infalling clock takes longer
and longer as seen from the outside, the ratio going as e
t/4M
with M the mass of the hole
and t the time measured by the external observer. So the external observer is seeing the
internal motion of infalling strings slowed down, a xed time scale t
ext
corresponding to
a time interval e
t/4M
t
ext
for the infalling object. As time goes on, the external observer
sees more and more of the modes of the string, and it appears to grow, as
_
ln(e
t/4M
t
1
ext
) t
1/2
. (4.3.7)
Before the black hole evaporates, the string grows to macroscopic sizes, and low energy eld
theory no longer applies. Also, because of its longitudinal behavior the string, rather than
contracting closer and closer to the horizon, remains a nite thickness above it. Again, this
would have a profound eect on the Hawking radiation.
However, I have put sees in quotes, because to see one needs light, and the black hole
is black. Of course there is Hawking radiation, but the conventional wisdom is that this
is produced outside the horizonthe relevant modes are in their ground states near the
horizonand so cannot reveal all of the marvelous things that the string is doing, and the
string physics is irrelevant. At least, I have not been able to imagine a real dynamical
calculation which shows the stringy eects to be relevant. But this picture of the essentially
diusive growth of the string is very simple and appealing [128], and the stringy properties do
dovetail very nicely with what would be needed to solve the information problem. It may be
that the low energy eld theoretic treatment is internally consistent but not correct, because
the large nonlocal invariant makes string theory important even before the low energy eld
theory breaks down internally due to a large local invariant. This would be a satisfying
resolution of the various points of view; the problem is to nd the right calculation!
The previous limits involved two strings, each in a state of low excitation but with large
center of mass energy. The nal limit I will discuss is that of a single string in a very high
state of excitation. The density of single-string states per unit energy n() is related to the
free energy in the noninteracting limit,
F() =
_

0
d n()e

, (4.3.8)
where I have used the fact that e

will be much less than 1 for the relevant states. The

density of states grows exponentially, as e
c
, so the integral converges at low temperature but
diverges at <
c
. It is tempting to interpret this as a transition (the Hagedorn transition)
112
to a phase where the conjectured fundamental degrees of freedom will be evident. As yet,
there is little understanding of the high-temperature phase.
There is a large literature on this subject, but here I just wish to note that there is a
very simple picture of the typical high-energy string state, which gives a surprisingly good
quantitative account of the density of states [129]-[132]. First let me tell you more precisely
what the density of states is. In the limit , at xed volume, it is
n() =
1

e
c
. (4.3.9)
This is true for any nite volume spatial background, such as a torus, or a group manifold.
If instead one takes the linear size to innity faster than
1/2
, one nds
n() =
V

(1+D)/2
e
c
. (4.3.10)
where D is the number of spacetime dimensions and V the volume.
The results (4.3.9) and (4.3.10) follow if one assumes that the typical highly excited string
is a random walk, with its length proportional to its energy. First, the exponential factor
comes from the fact that the number of steps is proportional to the length, with a choice to
be made at each step. The Hagedorn transition thus arises from a competition between the
energy and entropy of a long string. The power-law prefactor is important for the details
of the transition. In the density (4.3.10), the factor of V is counts the number of places a
random walk might start. We must then divide by the typical volume of the random walk,
because we can make a closed string only if the nal point and initial point coincide; this
gives a factor of
(1D)/2
. Finally, this overcounts by a factor of the length of the string, since
we can start at any point along the closed string, so we need an additional factor of
1
,
giving the result (4.3.10). In nite volume, the random walk will eventually ll the space
so that its volume is just V , giving instead (4.3.9). Remarkably, the density of states is
completely independent of the size and shape of the space in this limit. With such a simple
picture of the free theory we might hope that the interacting theory will be tractable.
5 Matrix Models
113
5.1 D = 2 String Theory
Returning to the theme of my introduction, one of the things I learned from Wilsons lecture
is the extent to which he was inuenced by the pion-nucleon static model. This is an
essentially one-dimensional free pion eld in interaction with a static nucleon with some
internal states.
1
In spite of having only one interacting degree of freedom, this model has
the essence of quantum eld theory, a nontrivial renormalization group ow.
String theory has an enormous number of degrees of freedom and a rather rigid structure.
It is not easy to nd models which are simple enough to deal with and yet retain the essential
features of the theory. In the previous section we mentioned various high energy limits where
things seem to simplify a bit, but as yet it has not been possible to do anything beyond
perturbation theory. The one limit which has turned out to be solvable is the limit of small
numbers of spacetime dimensions, 1 D 2 (for reviews see refs. [133], [134]).
A simple solution to the background eld equations (3.1.2) for arbitrary D is [135]
G

, B

= 0, = QX
1
/2 (5.1.1)
with
Q
2
=
26 D
3
. (5.1.2)
The eld equations (3.1.2) were derived in world-sheet perturbation theory and are generally
true only when the gradients of the background elds are small, but the solution (5.1.1) is a
special case because the path integral remains gaussian. In fact, we have already constructed
the CFT. On a at world-sheet the world-sheet action reduces to the ordinary free X

, but
with energy-momentum tensor as in eq. (1.6.1),
T =
1
2
: X

:
1
2
Q
2
X
1
(5.1.3)
giving c = D + 3Q
2
= 26.
What is the physics of this background? Recall that e

plays the role of the string

coupling, so the coupling is position-dependent, as e
QX
1
/2
. For X
1
the coupling goes
to zero, but for X
1
the coupling diverges and string perturbation theory breaks down.
1
To the condensed matter audience this will sound very similar to the Kondo model. In fact it was only
some time later that Wilson learned about the Kondo model and realized that his ideas would apply.
114
Adding a tachyon background produces a theory in which string perturbation theory is valid.
The vertex operator e
X
1
has weight
h =

h =
(2 Q)
2
+Q
2
8
, (5.1.4)
so for Q
2
8, which is D 2, there are two positive real solutions to the mass-shell condition
h =

h = 1. It is the lesser,
1
Q/2, which is appropriate,
2
T(X
1
) e

1
X
1
. (5.1.5)
Adding this to the action gives
S =
1
8
_
d
2

gg
ab

a
X

b
X

+Q

gRX
1
+e
X
1
_
. (5.1.6)
The X
1
action is the Liouville theory. The tachyon background (Liouville wall) suppresses
the path integral at large X
1
where the coupling is strong, the exponential factor in the
acion dominating the linear factor from the growth of the coupling.
This string theory is solvable at D = 1 or 2, or for one Liouville dimension plus a
minimal model (recall from eq. (1.9.11) that the central charge counts the number of degrees
of freedom, so a minimal model is like less than one whole dimension). I will concentrate
on D = 2, which has the most physics. As we have discussed, the physical state conditions
remove two sets of oscillators, leaving in this case none. Only the center of mass motion, the
tachyon, remains.
3
For real momentum the vertex operator (1.6.5) is e
ikX+QX
1
/2
. The factor
e
QX
1
/2
we now recognize as the string coupling, which multiplies the tachyon wavefunction.
The weight, given that Q
2
= 8 in D = 2, is 1 +
1
2
k
2
, so the mass-shell condition is k
2
= 0.
The tachyon is misnamed in D = 2: it is massless!
Exercise: Show in the same way that M
2
=
1
12
(D 2) in general D. This can be
interpreted as the negative Casimir energy of the D 2 transverse modes.
2
This is a subtle point [136], and I do not want to go o on this tangent, but I would like in this footnote
to state my understanding [137]. The tachyon has a second order eld equation, and the background is
a linear combination of both solutions. At large X
1
, the nonlinearities due to the tachyon self-interaction
become large and the linearized solution no longer holds. The appropriate linear combination of the two
solutions is determined by a condition of nonsingularity in the nonlinear region. In the linear region, it is
the more slowly decaying solution, the lesser value of
1
, which dominates. Incidentally, for D = 2 the two
roots are equal and the dominant term is X
1
e
X
1

2
. The linear part is important in understanding details
of the amplitudes.
3
Actually, this counting breaks down at discrete momenta, where there are extra physical states, an
important point to which I will return.
115
L
i
o
u
v
i
l
l
e

W
a
l
l
w
e
a
k

c
o
u
p
l
i
n
g
Figure 23: Two dimensional spacetime with dilaton and tachyon backgrounds. The S-matrix
describes transitions from a state with m incoming tachyons to a state with n outgoing
tachyons.
The theory thus reduces to a massless scalar eld moving in one time and one space
dimension. The spatial dimension is not translation-invariant. Rather, the coupling goes
to zero at X
1
, and grows with X
1
until we reach the Liouville wall, the tachyon
background which cuts o further propagation [138]. The basic physics is to throw tachyons
in from ; near the wall they interact with each other and reect back to . This is
about as rich as throwing pions at a static nucleon. Just as the latter proved to contain the
essence of eld theory, the D = 2 string has taught us at least one important lesson about
the full theory, the existence of the large nonperturbative eects. I will review the solution,
what has been learned, and what more we might hope to learn.
116
5.2 The D = 1 Matrix Model
The method by which the D = 2 string theory was solved is completely orthogonal to what
we have discussed thus far. It begins with the quantum mechanics of an N N Hermitean
matrix M(t) [139]. The action is
S = N
_
dt
_
1
2
Tr (

M
2
) + Tr V (M)
_
. (5.2.1)
Here is a parameter and V a general potential. You have already heard about this sort of
thing from one or more of Ambjorn, Ginsparg, and Moore, so I will go through the solution
without dwelling too much on the details. A graph with given numbers of propagators P,
vertices V , and loops L will depends on and N as
n
V
n
P
N
n
V
n
P
+n
L
=
n
L
(N)

, where
is Euler number of the surface we get by lling in the index loops in the graphs. Thus, for
any amplitude
A(, N) =

n
L
,
A
n
L
,

n
L
(N)

, (5.2.2)
where A
n
L
,
is the contribution from graphs with given numbers of loops and given topology.
The Hamiltonian is
H =
1
2N

i,j

M
ij

M
ji
+NTr V (M). (5.2.3)
The action is invariant under the U(N) which takes M UMU
1
. We will only be
interested in singlet stateswhen one cuts open a graph, the states will always be traces of
Ms, and so invariant. A singlet state is a function only of the eigenvalues
i
. Acting on
such a state one nds (use rst and second order perturbation theory to take the derivatives
of eigenvalues)
H() =

k
_

1
2N

2
k
+
1
N

l=k
1

k
+NV (
k
)
_
()
=
1
()H

()() (5.2.4)
where
H

= N

k
_

1
2
2
N
2

2
k
+V (
k
)
_
, (5.2.5)
and () =

k<l
(
k

l
). We will absorb the factor of into the wavefunction and work
with () = ()(). The inner product works out in a simple way: after the angular
117
integrations,
_
d
N
2
M

()()
_
d
N

()(), (5.2.6)
so is the probability amplitude for the eigenvalues.
The Hamiltonian (5.2.5) describes N decoupled coordinates. The factor of in makes
the wavefunction antisymmetric, so this is a system of N free fermions [139]. It can also be
written in terms of a second-quantized spinless fermion eld (),
H

= N
_
d
_
1
2
2
N
2

+V ()

_
. (5.2.7)
From the N-dependence (5.2.2), the N limit is dominated by graphs of spherical
topology. This limit is easily taken in the free-fermion form. Although we set the original
h to 1, we see that if we dene H

= NH

then N appears in H

precisely the way 1/ h

appears in the Hamiltonian for nonrelativistic quantum mechanics. The N limit at
xed is then classical. Each fermion occupies a volume 2 h = 2/N in phase space.
Levels with single-particle energy E

=
1
2
p
2
+V <
F
are lled. The total number of fermions
is N, so
1

=
N
N
=
_
dp d
2

_

F

p
2
2
V ()
_
. (5.2.8)
Now let us, to be specic, take
V () =
1
4

2
(2 )
2
. (5.2.9)
For large , the phase space integral (5.2.8) is small, and the fermions sit near the quadratic
minimum at = 0. One nds
F

1/2
. There are no fermions in the second minimum
at = 2: the Feynman rules represent perturbation theory around M = 0, and the second
minimum is invisible in perturbation theory. Figure 24 shows the potential, with occupied
states shaded. As decreases, the occupied phase space volume grows. At some critical
value
c
the Fermi level reaches the local maximum,
c
=
1
4
at = 1, and there is a phase
transition (see [140] for a review). Quantities such as
0[H

[0) = N
2
_
dp d
2
_
p
2
2
V ()
_

F

p
2
2
V ()
_
(5.2.10)
118
V
x
Figure 24: Potential energy for fermions, with lled states shaded.
are analytic for >
c
and have a singularity at
c
.
Exercise: Show that in terms of =
c

F
,

c
ln, 0[H[0)
2
ln, (5.2.11)
up to terms analytic in [141]. This takes a bit of care, because the analytic terms in these
two quantities are larger than the singular ones, though they are irrelevant because they
do not come from large surfaces. Show from this that the contribution A
n
L
,2
dened in
eq. (5.2.2) goes as n
3
L

n
L
c
for 0[H[0) at large n
L
.
Now recall, from the counting (5.2.2), that
1
is the loop-counting parameter for per-
turbation theory. So for a potential with a local maximum, the perturbation series has a
nite radius of convergence, diverging beyond
1
=
1
c
. The non-analyticity in
c
then
arises from large orders of perturbation theory, large graphs. Here is where the connection
with string theory comes in. Assign a length a to each propagator. As n
L
, one might
hope to take a 0 in such a way that the sum over graphs approaches a sum over smooth
surfaces, so the critical behavior in
c
is given by a string theory [141]. There are other
possibilitiesthe typical graph might instead look like a branched polymerwhich can arise
in the limit of large discrete surfaces. But in this case we are lucky, and the limit is a string
theory. We can verify this by comparing various quantities that can be calculated both in
the matrix model and in the continuum theory, and then use the matrix model to do many
calculations that we cannot yet do in the continuum.
We have taken the large N limit, so we have only surfaces of spherical topology. Thus,
119
the ground state energy (5.2.10) is explicitly O(N
2
). We can do better [142]. Let us take
N and
c
together, holding xed
= N(
c

F
). (5.2.12)
This is the double scaling limit.
4
Since (N)
1
plays the role of h, the splitting between
fermionic levels is of order 1/N, and the ratio of this splitting to the distance
c

F
from
the critical Fermi level remains constant. The splitting thus remains non-negligible in the
double scaling limit. Since the splitting is a 1/N eect, vanishing in the spherical limit,
the double scaling limit is keeping nontrivial contributions from all topologies, with Euler
number being weighted by

. That is, each additional handle brings in

2
.
The double-scaling limit is easily taken directly in the Hamiltonian. Dene
1 = (N)
1/2
x, = (N)
1/4
. (5.2.13)
Thus, for xed x, the eigenvalue approaches the local maximum as N . Only the
quadratic behavior near the maximum survives in this limit. The Hamiltonian becomes
H

N
F
=
_
dx
_
1
2

x

x
2
2

_
. (5.2.14)
The result is simply a Fermi sea in an inverted oscillator potential.
5.3 Matrix Model String
What do we expect for the continuum string theory corresponding to this critical behavior?
The number of links separating two vertices gives a measure of distance, so we expect a
world-sheet metric g
ab
(). And, each vertex occurs at some time t, so this should become a
world-sheet eld t() in the continuum limit. The most relevant action for these elds is
S
nc
=
_
d
2

g
_
c
1
g
ab

a
t
b
t +c
2
+c
3
R
_
. (5.3.1)
There is no reason to expect Weyl-invariance. Consequently the world-sheet cosmological
constant term c
2
has been included. We have earlier discussed briey the world-sheet theories
without Weyl invariance. We can go to the conformal gauge
g
ab
() = e
()
g
ab
(), (5.3.2)
4
This limit was carried out rst for the D = 0 matrix model [143]. Although this theory has fewer degrees
of freedom, the solution and the double scaling limit are perhaps a bit more dicult.
120
in which the metric is reduced to the degree of freedom (). Taking into account the Fadeev-
Popov determinant and the Weyl anomaly from the t-integration, the action becomes [49]
_
d
2

_
g
_
c
1
g
ab

a
t
b
t +c
2
e

+c
3

R +
25
96
( g
ab

a
b + 2

R)
_
. (5.3.3)
We started with one spacetime dimension, but as we have noted the Weyl factor in the
metric behaves very much like an extra embedding dimension. So we can think of this as a
non-critical string in one dimension with metric g
ab
, or a critical string in two dimensions
with metric g
ab
. The former is appropriate if one is interested in two-dimensional quantum
gravity on the world-sheet; the latter is appropriate if one is interested in two-dimensional
string theory in the embedding space. Ambjorns lectures focus on the former, and mine on
the latter. Although the theories are the same, one tends to ask dierent questions in the
two cases. In the former case, the focus is on various measures of the world-sheet geometry,
while in the latter it is on scattering processes in the embedding space.
Comparing the result (5.3.3) with our expectation (5.1.6), we see precisely the same
terms with X
1
, but the coecients cannot be made to match upthe 25 needs to be
24 and the exponent in the cosmological term is wrong. This has a simple explanation.
The derivation of (5.3.3) implicitly dened the path integral using the metric g
ab
, while
the latter used the metric g
ab
, and the dierence is accounted for by the corresponding
renormalization of the measure [144]. This is presumed now to be well-understood [145] and
is not something I want to dwell on, so I use the action in the form (5.1.6) with the relative
normalization of X
0
(in the string theory) and t (in the matrix model) yet to be determined.
To summarize, we expect that the free fermion theory (5.2.14) is equivalent to D = 2
string theory in the dilaton and tachyon background (5.1.5). In the string theory we had
tachyons bouncing o a Liouville wall. In the matrix model we have fermions bouncing
o the inverted harmonic potential. So things look very close, the string theory being a
bosonization of the matrix model.
Let us carry out the bosonization in detail [146]-[148]. In the classical limit we can
describe the collective motions of the fermions in terms of a time dependent Fermi surface,
separating the lled and empty phase space regions. Fermions on the surface on the surface
move freely in the inverted harmonic potential,
D
t
p = x, D
t
x = p, (5.3.4)
where D
t
denotes the co-moving derivative, following one phase-space point on the surface.
121
The solution is
p = a sinh(t b), x = a cosh(t b) (5.3.5)
with integration constants a and b. In contrast, the position p of the Fermi surface at xed
x satises
t
p[
x
= D
t
p
x
p[
t
D
t
x = x
x
p[
t
p. For perturbations that are not too large
(and we will return to these later), the Fermi sea can be described by the positions of its
upper lower surfaces at each time, p

(x, t), and each satises

t
p

(x, t) = x p

(x, t)
x
p

(x, t). (5.3.6)

For example, from the Hamiltonian (5.2.14) it follows that the static Fermi level is given by
1
2
(p
2
x
2
) + = 0 or
p

(x)
static
=
_
x
2
2 , (5.3.7)
which satises the equation of motion (5.3.6). The Hamiltonian in terms of p

is obtained
by integrating the single-particle energy over the Fermi sea,
H

=
1
2
_

dx
_
p+
p
dp
1
2
(p
2
x
2
)
=
1
2
_

dx
_
1
6
(p
3
+
p
3

)
x
2
2
(p
+
p

)
_
. (5.3.8)
Also, from the equation of motion and the Hamiltonian one can deduce the commutator
[p

(x), p

(y)] = 2i
x
(x y), [p
+
(x), p

(y)] = 0. (5.3.9)
To write this in terms of a massless scalar, dene the coordinate q = ln(x), which
runs from to as x runs from to 0, and dene a scalar eld o(q, t).
p

(x, t) = x
1
x

(q, t)

1/2

(q, t) =
S
(q, t)
q
o(q, t). (5.3.10)
The bar is to distinguish this scalar from a slightly dierent one to appear later. The
Hamiltonian takes the form
H

=
1
2
_

dq
_

2
S
+ (
q
o)
2
+e
2q
O(o
3
)
_
. (5.3.11)
122
x
p
p
+

p
Figure 25: Fermi sea: lled states in the x-p plane are shaded. The state of the sea is given
by the upper and lower surfaces p

(x, t), shown at a representative value of x. Points on the

Fermi surface move along hyperbolic orbits.
The quadratic part is the canonically normalized free scalar Hamiltonian; the equation of
motion is (
2
t

2
q
)o = e
2q
O(o
2
).
The collective motion of the matrix model fermions has the same qualitative behavior
as the D = 2 string theory discussed earlier. Asymptotically it is a free massless scalar,
with interactions that grow exponentially. In the string theory, strings are reected from
the Liouville wall before the coupling diverges, while in the matrix model they are reected
from the end of the eigenvalue distribution. Thus, our guess that the D = 2 critical string
describes the critical behavior of the D = 1 matrix model appears to be correct. The coupling
goes as e

2X
1
in string theory and e
2q
in the matrix model, which determines the relative
normalization of these. Also, the velocity is 1 in both the q-t and X
1
-X
0
planes, so we have
X
1
q

2, X
0
t

2. (5.3.12)
We had earlier concluded that
2
is the loop-counting parameter, which would make
1
the three-string coupling. We can now check this. From the static distribution (5.3.7), the
123
0 0
b) a)

Figure 26: a) Incoming pulse on top of the static Fermi sea. For clarity the variable
+
(q, t)
is shown. b) Later outgoing pulse, with shape distorted by dispersion. Fermions closer to
[p[ = [x[ travel more deeply into the potential and take longer to return.
turning point in the eigenvalue distribution is at x
t

1/2
. The dominant interactions
occur at the Liouville wall x
t
, where the coupling g
s
goes as e
2qt
= x
2
t

1
as expected.
More detailed comparisons are possible as well, and in particular the relation between the
string coordinate X
1
and the matrix model eigenvalue q is more subtle than the scaling
argument (5.3.12) suggests. We will return to this, after discussing some general issues.
Two asides before we go on. First, it is interesting to contrast the bosonization we have
just done with relativistic bosonization. The latter relates the collective motion of a free
massless relativistic fermion to a free massless boson. The dierence is that the relativistic
fermion has a linear dispersion relation, k, while the matrix model fermions have
k
2
. Looking at the Hamiltonian, we see that the cubic scalar interaction comes from
the term quadratic in momentum. We can think about this as follows. Relativistic fermions
have no dispersion, the velocity
k
being a constant. A collective pulse propagates without
changing its shape. The matrix model fermions have a k-dependent velocity, so that a pulse
deforms as it propagates. This translates into interactions among the bosons.
Second, the endpoint of the distribution, where p
+
p

0, causes diculties. At
this point the dynamics is determined not by the Hamiltonian but by the constraint that
the fermion density p
+
p

be positive. Further, perturbation theory in the collective

Hamiltonian suers various divergences at the endpoint, which need to be appropriately
regulated. It is sometimes useful to work with the Fourier transformed fermion

(p, t), in
terms of which the Hamiltonian looks almost the same [149],
H

N
F
=
_
dp
_

1
2

p
+
p
2
2

_
. (5.3.13)
The collective motion can now be described in terms of the surface x(p, t). In this variable
124
the sea has only an upper edge, and there is no endpoint to worry about.
5.4 General Issues
So what does one learn from all this? The object of study is the amplitude for m incoming
strings of energies
1
, . . . ,
m
to evolve into n outgoing strings of energies

1
, . . . ,

n
. The
matrix model allows the exact calculation of these amplitudes [150]. One starts with the
incoming strings, translates into an incoming fermion state using bosonization (asymptot-
ically we have noted that the bosons become free, so one can use the familiar relativistic
bosonization), evolve the free fermions in the inverted harmonic potential, and then trans-
late back using bosonization. The nal closed form from combining all these steps is a
little complicated,
5
but for given m and n one can follow through the steps and nd the
exact answer, without resorting to expansion in the closed string coupling
1
. The same
calculations remain exceedingly hard within the continuum string theory. One must in prin-
ciple carry out the path integral over X

, integrate over moduli space, and sum over genus.

Even the rst of these is nontrivial, because the path integral is not gaussian, due to the
tachyon background. The path integral on the sphere has been carried out using dicult
analytic arguments, and for some amplitudes by taking advantage of the large symmetry of
the problem, to be developed later.
Undoubtedly the main physical lesson learned thus far is the discovery of the large non-
perturbative eects, which as we have seen turned out to be a general feature of string
theory [115]. In the matrix model, there is an obvious nonperturbative process, the tunneling
of a single fermion through the potential barrier. Now we have to consider an issue that we
have ignored so far, namely what is going on on the other side of the barrier. There are
many possibilitiesthat the other side is empty (so the state we are considering is unstable);
that it is lled to the same Fermi level as the rst well; that the potential is modied, say
by an innite barrier at some x 0. None of these would change the perturbation theory,
so they represent a nonperturbative ambiguity. Later we will examine the problem of the
nonperturbative denition of the theory, but for now we will just use this ambiguity to
estimate the inaccuracy of perturbation theory. The fermion tunneling amplitude is e
B
with B given in the WKB approximation by
B =
_

2

2
dx
_
2 x
2
= . (5.4.1)
5
The expressions do simplify in the leading order, string tree level. We will work this out later.
125
So the tunneling amplitude is indeed e
O()
= e
O(g
1
s
)
. From the discussion in section 4,
one would expect this to be associated with a (2h)! growth of the perturbation theory. This
growth is indeed found when the amplitudes are expanded in powers of
2
[142].
In this case one can see what the nonperturbative eect associated with the large order
behavior is, namely single-fermion tunneling. The matrix model is an example of this picture
in which the strings are collective excitations of something simpler, the free fermions. The
bosonic collective Hamiltonian can presumably be used to generate all orders of perturbation
theory, but it is not at all clear that one can make sense of it non-perturbatively. The
description of the single-fermion tunneling in terms of the collective Hamiltonian is extremely
clumsy and ad hoc. More generally, the collective description seems to break down near the
end of the eigenvalue distribution, or for processes involving interference between two edges
of the Fermi surface. For example, the raising operators of the o eld create a fermion-hole
pair. An operator with momentum greater than p
+
p

tries to create a hole below the

lower edge p

of the Fermi sea, which is impossible. So the algebra of the bosonic raising and
lowering operatorsthe enumeration of statesbreaks down. A very similar thing happens
in the case of D = 2 Yang-Mills theory. The string representation can be obtained as a
bosonization of a non-relativistic free fermion representation just as here, and again the
enumeration of string states breaks down non-perturbatively, when the number of boxes in
a Young tableau column exceeds N.
One would of course like to use this to get some insight into the nature of the nonpertur-
bative eects and of the hypothetical fundamental degrees of freedom in higher dimensions.
One can try looking at limits in the higher dimensional theory as I have discussed, but an-
other avenue is to try to make as explicit as possible the connection between the free fermion
picture and the string picture here in the D = 2 theory. At present the connection is quite
roundabout, going through the matrix model and double scaling limit.
6
In fact, the connec-
tion is rather subtle. The string theory tachyon is related in an obvious way to the collective
motion of the fermions, but in fact there is more in the string theory. The graviton-dilaton
sector of the D = 2 string theory is the same as the dilaton gravity theories discussed by
Strominger and Verlinde. So although there is no propagating graviton or dilaton, there are
6
Something I have tried without success is to go directly from the string path integral to a spacetime
Hamiltonian by a good choice of gauge. Unfortunately, the obvious gauge choice,
0
= X
0
so as to identify
world-sheet and embedding times, does not seem to work even for the simpler problem of a relativistic
particle. The gauge choice assumes that X
0
is a monotonic function of
0
, so that if the bath backtracks one
must build it by combining monotonic paths. Unfortunately, the typical relativistic particle path backtracks
an innite number of times in the continuum limit. Light cone quantization avoids this backtracking problem
in at spacetime, but has been rather hard to implement with the linear dilaton background [151]. Kawai
and collaborators have made some progress along slightly dierent lines [152].
126
still physical eects from gravity, including a black hole solution. These are not evident in
the matrix model. It is important to clarify this, both to develop the matrix model string
connection, and because the gravitational eects are of interest in their own right.
We have been discussing the D = 1 (also called c = 1) matrix model. The D = 0 model,
based on a single matrix M without time, leads in a similar way to the D = 1 string (pure
world-sheet gravity = one Liouville dimension). Multi-matrix models M
i
lead to 0 < c < 1,
the minimal models plus a Liouville dimension. Also, ne-tuned multicritical points in the
D = 0 matrix model lead to non-unitary minimal models plus the Liouville dimension. The
extension to c > 1 is problematic. The matrix model can no longer be reduced to eigenvalues,
and no solution is known. Also, as we have noted the tachyon really is tachyonic beyond
D > 2, so the vacuum is unstable and presumably does not correspond to an attractive xed
point of the discrete theory.
Even at D 2 it would be useful to have a greater variety of models, so as to get
some perspective. The matrix model can be generalized to open plus closed strings [153] (a
dierent approach is given in ref. [154]). This reduces to a theory of interacting fermions. It is
similar to the U(N) Calogero-Sutherland model recently considered by Haldane and others,
but has not been solved. Even short of a full solution, it would be useful to understand in
the classical limit the connection between the collective excitations and the open and closed
strings. Another extension is spacetime supersymmetry. Supersymmetry breaking is after
all one of the most important non-perturbative phenomenon in string theory; spacetime
supersymmetry would also help with the tachyon instability of the c > 1 string. There is
nothing useful here yet. A straightforward supersymmetrization of the D = 1 matrix model
exists and has a double-scaling limit [155]. But Shyamoli Chaudhuri and I have recently
shown that the naive world-sheet theory is not conformally invariant, and the theory ows
to some exotic xed point that does not seem to be connected with the string physics that
is of interest [156].
To conclude these lectures, I will rst develop some details of the tree-level scattering
amplitudes. I will then discuss the gravitational physics of the D = 2 string, including the
black hole solution, as well as the generalization to higher string levels. Finally, I look for
the corresponding physics in the matrix model. I show how gravitational physics emerges
in the weak eld limit, then look at physics in strong eldsan (as yet unsuccessful) search
for the black hole solution, some exotic strong-eld physics that arises in the matrix model,
and the nonperturbative ambiguity.
127
5.5 Tree-Level Scattering
Consider a small incoming pulse on top of the static solution p
,static
, eq. (5.3.7). As we
have discussed, dispersion changes the shape of the pulse: higher points on the pulse travel
deeper into the potential and emerge later. Let us work this out at the classical level. The
equation of motion (5.3.4) implies that
v = (x p)e
t
, w = (x +p)e
t
(5.5.1)
are constant for a particle moving in the inverted harmonic potential. Then so also are the
integrals
v
mn
= e
(nm)t
_
FF
0
dp dx
2
(x p)
m
(x +p)
n
. (5.5.2)
The integral runs over the Fermi sea F, with the integral over the static Fermi sea F
0
subtracted, so this will converge for a pulse of nite width. Evaluating this in the limits
t and t relates the incoming and outgoing pulses. As t points approach
the line p = x, so
v e
qt

+
(q t), w 2e
q+t
. (5.5.3)
I have used the expansion (5.3.10), with the free eld equation of motion
+
(q, t)
+
(q t)
for the incoming wave. Similarly, as t ,
v 2e
qt
, w e
q+t

(q +t) (5.5.4)
with

the outgoing wave. Thus,

v
mn
=
2
n
2(m + 1)
_

dt e
(nm)(tq)
_
(
+
(t q))
m+1

m+1
_
=
2
m
2(n + 1)
_

dt e
(nm)(t+q)
_
(

(t +q))
n+1

n+1
_
. (5.5.5)
Applying this for m = 0, n = i gives the Fourier transform of the incoming wave
as a nonlinear function of the outgoing wave [157], [158]. Expanding around the static
background,

(q

) = +

(q

), gives
_

dq

e
iq

+
(q

) = (2)
i
_

dq
+
e
iq+

k=1

1k
k!
(1 +i)
(2 k +i)

+
(q
+
)
k
. (5.5.6)
128
In terms of the Fourier modes,

(q

) = +
1

2
_

()e
iq
(5.5.7)
this becomes

+
() =

k=1
(

2)
1k
k!
(1 +i)
(2 k +i)
(5.5.8)
_
d
1
. . . d
k

(
1
) . . .

(
k
)(
1
+. . . +
k
) .
It is a useful fact that classical scattering is the same as tree level quantum scattering.
The quantum operators satisfy the same equation of motion as the classical eld, so the
solution (5.5.8) also holds, except that we would have to be careful about operator ordering.
But commutators are of order h and so do not matter to leading order. The modes have
been normalized to satisfy
[

(),

)] = 2(

). (5.5.9)
Dening the in-states
[
1
, . . . ,
n
; in) =
+
(
1
) . . .
+
(
n
)[0),
i
> 0, (5.5.10)
and similarly for the out-states, the classical S-matrix (5.5.8) plus the canonical commuta-
tor (5.5.9) give the tree level S-matrix. For example, the 1 n amplitude is [157]

1
, . . . ,
n
; out[; in) =
_
2

_
n1
(1 +i)
(2 n +i)
2(
1
+. . . +
n
). (5.5.11)
This same calculation can also be done by string methods, but only with great eort
and ingenuity. The main complication is the exponential term in the string action, from
the tachyon background. Far from the Liouville wall this is small, so for so-called bulk
processes which happen far away from the wall we can use the free action or at least expand
in powers of . (One can think about using wavepackets to separate out the bulk processes,
as I will describe later). We then have a free-eld string calculation of the type that gave the
Virasoro-Shapiro amplitude. In D > 2 this reduces to -functions only for four particles, but
in D = 2 the kinematic restrictions make it possible to evaluate for any number. Also, an
129
analytic continuation makes it possible to deduce the full amplitude from the bulk amplitude.
The result is almost the same as the matrix model result (5.5.11). That is, comparison of the
string and matrix model S-matrices shows that the latter is related to the former by [159]-
[161].

+
() =
(i)
(i)

+
()

() =
(i)
(i)

(). (5.5.12)
The ratio of gamma functions, known as the leg pole factor, is a pure phase for real frequen-
cies, so this is a unitary transformation on the states. Such a redenition is not surprising.
The matrix model gives a discrete approximation to the local vertex operator e
iqX+X
1

2
.
One expects to nd a renormalization for any such cuto construction. Following this line
of thought one can, purely within the matrix model, deduce the poles and zeroes of the leg
pole factor.
7
I will not pursue this, but simply take the leg pole factor from the comparison of ampli-
tudes. Although a pure phase, it will play an essential role soon.
5.6 Spacetime Gravity in the D = 2 String
Now let us take a closer look at the graviton, dilaton, and antisymmetric tensor states in
the D = 2 string,
[e, k) = e

1
[0, k). (5.6.1)
Recall from section 1.6 the Virasoro generators
L
0
=
1
2
k
2
+
a
2
2
+

n=1

n

n
L
m
=
1
2

n=

n

mn
+i(m + 1)a
m
, m ,= 0, (5.6.2)
7
To be precise, the correspondence between the string and matrix model S-matrices has only been veried
explicitly at tree level. However, there is strong reason to believe that the leg pole factor is the same to all
orders of perturbation theory. The renormalization comes from small distances on the world sheet, and the
string coupling is a relevant interaction. That is, short distance is the free asymptotic region of the critical
string picture.
130
where a

= (0,

2). Also recall

0
= k

ia

. Acting on the state (5.6.1), the physical state

conditions become (we use the convenient OCQ form)
L
0
,

L
0
: k
2
= a
2
L
1
,

L
1
: (k +ia)

= (k +ia)

= 0. (5.6.3)
Also, acting with L
1
on the state f

1
[0, k) and with

L
1
on

f

1
[0, k) we nd the
following spurious states
spurious: e

= (k ia)

+f

(k ia)

(5.6.4)
for any f,

f.
The L
1
and

L
1
conditions require e

to be orthogonal to k + ia on both indices, or

e

= n

where n

is in the one-dimensional space orthogonal to k+ia. The L

0
conditions
gives (k +ia) (k ia) = 0, so n k ia. But this means that e

is spurious, so there
are no observable states.
This is correct at generic momenta but breaks down at special points [160]. If k = ia,
the L
1
and

L
1
conditions are empty. States with polarizations proportional to a

are null in
this case, so we can remove the spacelike polarizations and be left with the observable state

0
1

0
1
[0, ia) . (5.6.5)
If k = ia the spurious state vanishes. The physical state condition requires e

to be
orthogonal to a

, so the observable state is

0
1

0
1
[0, +ia) . (5.6.6)
The corresponding vertex operators are respectively
: X
0

X
0
e
2

2X
1
: , : X
0

X
0
: . (5.6.7)
What does this mean physically? The absence of observable states at general momenta
means that there are no particle-like statesone cannot make a wavepacket. The vertex
operator : X
0

X
0
: has an obvious interpretationit corresponds to an innitesimal rescal-
ing of the at metric G
00
, and can be absorbed into a rescaling of X
0
.
8
The other operator
8
We are considering X
0
to be noncompact, so this rescaling has no eect. If X
0
were periodic, it would
correspond to a change in the periodicity.
131
is more problematic because it corresponds to G
00
e
2

2X
1
, diverging at innity. In
fact, you have seen this before, in the talks by Strominger and Verlinde. The low energy
graviton-dilaton action for the D = 2 string,
S =
1
2
_
d
D
X

Ge
2
_
8 +R+ 4

_
(5.6.8)
is the same as they considered. As they discussed, this has a black hole solution [162], [163],
which in one set of coordinates is
G
11
= 1, G
00
= tanh
2
_
X
1

2 +
1
2
lnM
_
=
1
2
ln(M/2) + lncosh
_
X
1

2 +
1
2
lnM
_
. (5.6.9)
As M 0 this approaches the linear dilation vacuum (5.1.1). To rst order in M it is just
the perturbation
G
00
Me
2

2X
1
. (5.6.10)
We should note that, unlike the case of four dimensions where the size of the black
hole is proportional to M, here the size is independent of M. The mass appears in the
background (5.6.9) only as an additive term in and as an additive shift of X
1
. The size is
set by the only other scale in the problem, the string scale. The spacetime action (5.6.8) is
only valid at scales long compared to the string scale, so the derivation of the solution (5.6.9)
is not strictly valid.
9
However, an exact CFT corresponding to this solution, a variation of
the WZW theory, is known [162]. Also, because the scale of the geometry is of the same size
as the string itself, we cannot be sure that the string sees it as a black hole. For example,
we cannot make a clean causality argument in the information paradox, as we can do in
four dimensions. But the exact solution does appear to resemble a black holefor example
in the Bogoliubov transformation between the asymptotic tachyon modes and those on the
horizon [164].
Now let me state the result for the observable Hilbert space at higher levels (I am following
here refs. [165]; see these for earlier references). Extra physical states (the discrete states) are
found at momenta (k
0
, k
1
) = (in

2, is

2) for 2s and s n integers such that [n[ s 1,

as shown in g. 27. The corresponding operators are denoted W

s,n
(z)

W

s,n
( z). The lines
9
This comment does not apply to dilaton gravity as a eld theory, where the scales are set by hand.
132

ik
0
1
ik
2
2
W
+

W
+
W

W
Figure 27: Extra physical states in the (ik
0
, ik
1
) plane.
ik
0
= ik
1
correspond to (imaginary) tachyon momenta. The L
0
,

L
0
conditions imply that
the state (s, n) is at level s
2
n
2
.
Note also the conserved currents X
0
and

X
0
, closely related to the operator W
+
1,0

W
+
1,0
=
X
0

X
0
. Actually, although these currents are separately conserved, their dierence cou-
ples to the winding number in the X
0
direction. For the non-compact theory that we are
concerned with this vanishes, so only one symmetry is nontrivial. At higher levels there
are additional symmetries A
s,n
, in one-to-one correspondence with the W
+
s,n

W
+
s,n
(recall the
result (4.1.2) that a spacetime symmetry corresponds to a BRST invariant operator of ap-
propriate ghost number).
So the D = 2 string theory does have a few graviton-dilaton states, and some interesting
physics associated with them. In addition it has a large number of similar states at higher
levels. The state W

1,0

W

1,0
corresponds to a black hole background. It seems likely that
the other W

s,n

s,n
give rise to some innite-parameter generalization of the black hole, but
little is known about this; a formal argument for the existence of these solutions is given in
ref. [166].
The tachyon state was obvious in the matrix model. So are the symmetries A
s,n
they
133
are just the v
mn
introduced above for positive integer m, n [167]-[170],
10
A
s,n
v
s+n,sn
. (5.6.11)
This identication is consistent with the (k
0
, k
1
) values. Also the linear term in the variation
of the tachyon eld is known to be the same, as is the algebra
v
mn
, v
m

n
= 2(m

n n

m)v
m+m

1,n+n

1
. (5.6.12)
This is the same as the w

commutator (1.7.13), with 4V

i
m
v
i+m+1,im+1
. Only the range
of the indices diers.
However, the remainder of the physical states are not so evident. The W
+
s,n

W
+
s,n
do mix
with the tachyons at the appropriate momenta and can be seen in matrix model correla-
tors (reviewed in chapter 6 of ref. [134]). The W

s,n

s,n
are, like the black hole, singular
at large X
1
and do not appear in the matrix model in any simple way [136]. In the next
section we will see that much of the physics of the discrete states is not contained within
the matrix model by itself but also involves the mapping between the matrix model and the
D = 2 string theory.
5.7 Spacetime Gravity in the Matrix Model
The results of the previous section seem quite exciting. The D = 2 string theory has black
hole solutions, and we have in the matrix model an exact quantum solution to this theory.
But things are not yet so rosy. First, we do not know how to describe the black hole in
the matrix model languagewhat deformation of the matrix model corresponds to nonzero
Schwarzchild mass M? Second, as we will discuss in section 5.8, we do not yet have the
exact solution to the theory, only its perturbation expansion.
To begin to study this issue, we will rst ask a much more basic question [171]. How, in
the matrix model, do we see the gravitational interaction even between lumps of matter too
diuse to form a black hole? I will phrase this as an S-matrix question. Imagine sending two
successive tachyon pulses in from the asymptotic region. The rst pulse carries energy and
momentum, so the second pulse must feel its gravitational eld. There will then be some
amplitude for the process shown in g. 28a, in which part of the second pulse backscatters
from the gravitational eld of the rst pulse. But this does not happen in the matrix model!
10
Incidentally, the v
mn
are well-dened for general m, n only if the sea is entirely in the quadrant x < |p|,
but for integer m, n they are always well-dened.
134
2 1
t
q
2 1
I
II
III
a) b)
Figure 28: a) Successive pulses moving in the t q plane. Gravitational eld (dotted) of
pulse 1 will cause part of pulse 2 to backscatter, producing an outgoing wave (dashed) which
precedes the main reection from the wall. b) How the matrix model represents this process.
The initial wavefunction renormalization (I) produces a tail on pulse 2 which overlaps pulse 1;
the combined pulse reects from the wall (II); and the nal renormalization (III) produces
the outgoing wave.
The pulses consist of free fermions, which move independently in the inverted oscillator
potential.
This would seem to be a contradiction, because I have set this up as an S-matrix question,
and earlier I told you that the string and matrix model S-matrices agree. However, I also
told you that there is a renormalization of the states, the leg pole factor (5.5.12), and this
is in fact essential to seeing all of the physics. To see how the renormalization can have this
eect, let us write it in coordinate space, in terms of the asymptotic incoming and outgoing
tachyon elds:
(I) : o
+
(t q) =
_

d K()o
+
(t q )
(III) : o

(t +q) =
_

d K()o

(t +q ). (5.7.1)
135
Here, a bar is again used to distinguish the matrix model quantity from the corresponding
string quantity. The same kernel appears in both transformations,
K() =
_

d
2
e
i
_

2
_
i/4
(i)
(i)
=
z
2
J
1
(z), z = 2(2/)
1/8
e
/2
. (5.7.2)
This has asymptotic behaviors
K()
_

2
_
1/4
e

2
_
1/16
e
/4

cos(z +/4), . (5.7.3)

To describe the scattering of an incoming (+) string tachyon pulse, one must (I) transform
to the matrix model tachyon eld via (5.7.1), (II) evolve the pulse as described in section 5.5,
and (III) transform back. The transformation (5.7.1) is nonlocal, and it is the early-time
behavior that is important. Although this falls exponentially, so does the gravitational eect
(5.6.10) that we seek. As long as we use narrow enough pulses, gaussians, the exponential
tail can be discerned [172], [171].
The string amplitude I-II-III does indeed display the gravitational scattering. This works
as shown in g. 28b. The transformation I on the incoming pulse 2 produces an early tail
that overlaps pulse 1.
11
These propagate together through the turning point and interact
locally, and the transformation III produces the outgoing gravitationally scattered wave.
It will be useful for the later discussion to provide some details, but it is simpler for this
to look at the slightly simpler process of 2 1 bulk tachyon scattering. That is, as shown
in g. 29a, a pair of tachyons in one pulse can interact and scatter into one outgoing tachyon
before reaching the wall. This occurs in the D = 2 string theory, from direct calculation of
the three string interaction, but it again does not occur in the matrix model because the
fermions are freethey travel on hyperbolic phase space orbits and so reach the turning
point of the potential. The leg pole transformation III is again responsible for the dierence,
as shown in g. 29b; the tail on transformation I is not important in this case. Explicitly,
lim
t+q
o

(t +q) = 2
1/4

1/4
e
t
_

t
dt

e
t

(t

+q)
11
The amplitude of K grows at late times, but it oscillates so rapidly that it falls eectively to zero, at
least in all cases I have yet encountered.
136
t
q
a) b)
III
A
B
Figure 29: a) Two tachyons in an incoming pulse can backscatter into one (dashed line) at a
given point A. In the matrix model all tachyons pass through the turning region B. b) The
tail from transformation III produces the outgoing wave.
= 2
1/4

1/4
e
t
_

dt

e
t

(t

+q)
= 2
3/4

1/4
e
t+q
v
10
. (5.7.4)
In the second line we have used the narrowness of the wavepacket to extend the range of
integration, and in the third we have noted that the result is simply proportional to the
conserved charge v
10
. Now expressing this in terms of the incoming eld gives
lim
t+q
o

(t +q) = 2
3/4

1/4
e
t+q
_

dt

e
t

+q
_
(
t
o
+
(t q))
2
+
t
o
+
(t q)/

_
= 2
1/2
e
t+q
_

dt

e
t

+q
(o
+
(t

q))
2
+ O() . (5.7.5)
In the second line we have carried out the renormalization (5.7.1), leading to a simple re-
sult after integration by parts. The O() term is from 1 1 scattering on the tachyon
background.
The result (5.7.5) is the correct bulk scattering, as found from a string theory or eective
137
eld theory calculation [171]. Note in particular the exponential factor e
tt

+2q
. At t = t

,
which is the point A in gure 29 where the incoming and outgoing rays meet, this is the
trilinear string coupling e
2q
.
In a sense, nothing here is new. In momentum space, we have simply found the graviton
pole. It is well-known that this pole is not present in the matrix model amplitude but comes
from the pole factors for the external legs [159]-[161], [173]-[175]. However, this coordinate
space analysis makes a number of things much clearer, and dispels some confusions in the
literature. One confusion is the frequent assertion that the leg pole factor is not relevant
for physics in Minkowski space (real ), because it is a pure phase. We have seen that it is
indeed essential for important Minkowski space physicsthis is possible because the phase
is a function of the momenta of the particles and so can be seen in interference. A second
confusion is an identication of the string tachyon eld with the macroscopic loop amplitude.
This is not the same as the identication (5.7.1), and does not have the correct spacetime
physics.
The coordinate space calculation also provides a physical interpretation for the discrete
states in the cohomology. That is, they correspond to long-range forces not associated with
propagating degrees of freedom.
12
The imaginary values of the momenta for the discrete
states correspond to exponential fallo of the forces in spacetime.
Let me emphasize that it is only the full string S-matrix I-II-III that is physical. That this
factorizes into three separate parts, the rst and third being linear in terms of the bosons,
and the second being linear in terms of the fermions, is an accident of the D = 2 kinematics
and one should not give particular signicance to the individual factors. Note also that the
matrix model, in spite of its qualitative resemblance to the string spacetime, misrepresents
the causal structure of the physics. The 2 1 bulk process actually occurs at the point A
in g. 29a, before the pulse reaches the Liouville wall, but in the matrix model calculation
the information in the incoming pulse always travels through the turning region B.
The scattering calculations here become more tedious at higher orders in the gravitational
eld and at higher levels. One would like to carry out the redenition (5.7.1) directly in the
Hamiltonian, converting the known matrix model Hamiltonian into a string Hamiltonian,
with propagating tachyons plus long-ranged elds. The latter would presumably correspond
to some gauge-xing of string eld theory. But one must add to the linear transforma-
tion (5.7.1) appropriate nonlinear terms to bring the Hamiltonian to this form, and this is
12
Although they fall exponentially, I use the term long-range because they extend far enough to allow
them to be distinguished from the essential gaussian nonlocality of the string.
138
does not seem to be simple.
5.8 Strong Nonlinearities
Now that we have detected weak eld gravity, we would like to nd the black hole solution.
The rst thing that comes to mind is to keep the eld while omitting the source pulse 1,
producing a source-free solution. But it is not evident how to do this, since the construction in
g. 28a depends in an essential way on the direct interaction between the two pulses. Various
groups have proposed matrix model representations of the black hole background [176]. They
take points of view dierent from my own, so it is hard to make a direct comparison. Several
do emphasize the important point that the string theory is dened not by the matrix model
alone (step II) but also by the mapping between the matrix model and string theory (steps I
and III). But it is not clear that any of the proposal pass the test that they give the correct
scattering in the weak eld, weak coupling region studied in the previous section.
Since we have a complete description of the scattering of incoming pulses, we might try to
make a black hole, by sending in a large enough pulse of matter. The perturbation (5.6.10)
of the metric is M/x
4
in matrix model variables, so in order to form a black hole we need
at least for this to be of order 1. Consider now a pulse on the Fermi sea, of width x and
height p. The total energy M in the pulse is of order the number of fermions times the
height above the Fermi level, or xp (p
2
). Now, x
<

x and (p
2
) xp, so we need
(p)
2
>

x
2
(5.8.1)
to produce a gravitational eect of order 1. That is, the height of the pulse is comparable
to the height of the whole sea. The tachyon self-interaction is then also of order 1, since the
dispersion of the fermions in the pulse is large. In contrast to four dimensions, we cannot
get into a situation where the gravitational interaction dominates.
13
It has a long enough
range to be detected, but not long enough to dominate. So we cannot be sure a priori that
it is possible to make a black hole. We will have to be experimentalists, throwing in pulses
with various properties and seeing if the result has any of the characteristic signatures of a
black hole.
In fact, two interesting qualitative behaviors are found with large pulses [147]. The
rst is depicted in g. 30a. For a suciently tall or steep pulse dispersion will cause it to
13
The D = 2 string also diers from dilaton gravity, where the matter is introduced by hand and one may
take it to be free and conformally invariant.
139
p
x
a)
p
x
b)
Figure 30: a) Outgoing pulse in which dispersion has produced a fold. At the indicated x,
the Fermi sea has two lled bands separated by an empty band. b) Incoming pulse extending
above the line p = x. The fermions above the line pass over the potential barrier to x = .
broaden past the vertical, producing an empty band in the Fermi sea and a double-valued
Fermi surface. What does this correspond to in bosonic language? There is a one-to-one
correspondence between single-valued Fermi surfaces and classical states of the bosonic eld.
By a classical state I mean one for which A
2
) = A)
2
for any observable A, in the classical
limit where the fermions become a continuous uid. The double-valued surface does not, in
this classical limit, correspond to a classical bosonic state. One can check this explicitly by
calculating in such a sea the one-point and two-point functions of the boson modes (),
which are just the fermion currents. Even without calculation, a little thought will show that
by moving fermions down into the empty band one can reduce the energy
1
2
_
(
t
o)
2
+(
q
o)
2
_
while keeping the expectation values of
t
o and
q
o xed.
The double-valued surface is thus a bosonic state with a large amount of energy in
addition to that in the classical eldthat is, a large amount of radiation. There is a
supercial resemblance here to formation of a black hole. In both cases, for incoming matter
below a threshold (to form a horizon or a fold respectively), the incoming energy comes out
140
again in the classical eld. Above the threshold a non-negligible fraction of the incoming
classical energy returns as radiation, Hawking radiation in the black hole and fold radiation
in the matrix model. But here the resemblance ends. The fold radiation comes out too
promptly, and it is too hottypical energies of the quanta scale as h
1/2
rather than h
1
.
It is not obvious how we would see the fold radiation if we had only the stringy, bosonic,
description and not the fermionic one. In bosonic language the fold corresponds to an
intersection of characteristics (null lines), which occurs because the velocity is a function of
the tachyon eld. In most situations where such a thing occurs, there are higher-derivative
terms in the equation of motion which prevent double-valuedness and lead to formation of a
shock wave. But that is not the case hereit is evident in the free fermionic language that
pulse turns into a fold. So instead of a shock wave we get radiation, but I dont know how
to see this purely from the bosonic Hamiltonian. One rather dierent possibility is that the
fold is a coordinate singularity, and that we need to unfold it by an appropriate coordinate
transformation in phase space, but I do not know how to make sense of this.
The second interesting behavior is shown in g. 30b. This is a pulse which extends above
the line p = x, so that some of the fermions pass over the barrier to the asymptotic region
on the right. This certainly sounds like a black holewe throw matter in and it doesnt all
come back out! But now we have to face a problem that we have deferred, the interpretation
the other side of the potential barrier. In fact there is an innite number of ways to dene
the matrix model, all of which give a unitary S-matrix and all of which have the same
perturbation expansion [150] (and presumably the same as that of the string). One class
of theories (type I in the terminology of ref. [150]) eliminates the second asymptotic region
by modifying the potential. For example, a sharp innite barrier, V (x) = for x > A,
A 0, leaves the perturbation theory with xed numbers of incoming strings unaected
for any A 0. So does any other modication such that V (x) is
1
2
x
2
for x < 0 and rises
to innity as x . All incoming fermions eventually return ot x = , so these are
unitary quantum theories within the Hilbert space of incoming and outgoing fermions in the
left asymptotic region (or the bosonized equivalent). The type II theory, on the other hand,
leaves the potential unmodied and and lls both sides of the barrier to the same level. It is
a unitary quantum theory but with two asymptotic Hilbert spaces. So there appears to be
an innite number of consistent nonperturbative denitions of the matrix model. Note that
the ambiguity appears not only in the nonperturbative single-fermion tunneling amplitudes,
but also in the large-eld classical behavior. This is because a large incoming pulse can
propagate into the strongly coupled region.
141
Our study of gravitational and bulk scattering now pays an unexpected bonus, for we
have a new consistency condition which must be satised. Suppose we modify the potential
by an innite barrier at x = A 0. A fermion reaching this barrier will jump suddenly
from phase space point (x, p) = (A, p) to (A, p). The quantities v = (x p)e
t
and
w = (x +p)e
t
are no longer conserved, and neither will be the v
mn
. Eq. (5.7.5) no longer
holds and we get the wrong bulk amplitude. This violates causalitythe bulk scattering
occurs at the point A of g. 29a, where string perturbation theory is valid and the incoming
wave weak; the nonlinearity becomes large only later, at point B. A further acausality will be
a nonconservation of gravitational mass in the process of g. 28. This same argument applies
to all modications of type I, even if the potential rises more smoothly. Since (xp)e
t
is
negative for incoming fermions with p > x but positive on all outgoing trajectories on the
left side, it can never be conserved. So all theories of type I are inconsistent. One can also
see this another waythe v
mn
correspond to unbroken spacetime gauge symmetries and one
would expect these to be conserved exactly.
The type II theory with unmodied potential conserves the v
mn
, but it still gives rise to
an inconsistent string theory, at least with the natural interpretation that the two asymptotic
regions of the matrix model correspond to two asymptotic regions of spacetime. The point
is that part of the conserved v
mn
passes over the barrier, so the rst equality in eq. (5.7.5),
relating v
10
to the outgoing eld on the left only, no longer holds.
So I am saying that in spite of the substantial eort that went into matrix models no
consistent nonperturbative construction of the D = 2 string has yet been given. I think
that this is much better than the previous situation of an innite number of theories. The
study of the bulk and gravitational scattering in coordinate space has not only dispels some
confusions and claries the nature of the discrete states, but also provides a criterion for
selecting among the possible nonperturbative denitions of the theory. The next step is
to search for a solution to the consistency conditions, considering modication both of the
matrix model itself and of the mapping between the matrix model and string theory. This
is an important problem, bearing directly on the e
C/gc
behavior, and I am optimistic that
progress can be made.
5.9 Conclusion
In conclusion, let me say again that I think it likely that string theory will involve rich
new dynamics that play an essential role in determining the ground state. Sometimes I
142
feel that the tools I have described, such as the matrix model and the various high energy
limits, are crude and unsophisticated. This is why I found the parallels with Wilsons lecture
so encouragingboth the feeling of clutching at straws, and the resemblance between his
models and the matrix model. So perhaps the same method of attack that was so successful
in quantum eld theory will be useful again.
Acknowledgments
I would like to thank Mark Bowick, Shyamoli Chaudhuri, Jacques Distler, Je Harvey, Igor
Klebanov, David Lowe, Makoto Natsuume, Phil Nelson, Andy Strominger, Lenny Susskind,
and Lars Thorlacius for discussions of various points. I would also like to thank the stu-
dents, organizers, and fellow lecturers of the school for their questions and comments, for
an enjoyable experience, and for foosball lessons. This work was supported by NSF grants
PHY-89-04035, PHY-91-16964, and PHY-94-07194.
143
References
[1] K. Wilson, Rev. Mod. Phys. 55, 583 (1983).
[2] A. Giveon, M. Porrati, and E. Rabinovici, Phys. Rept. 244, 77 (1994).
[3] P. Ginsparg, in Fields, Strings and Critical Phenomena, Les Houches Session XLIX,
edited by E. Brezin and J. Zinn-Justin (Elsevier, 1989) 1.
[4] J. L. Cardy, in Fields, Strings and Critical Phenomena, Les Houches Session XLIX,
edited by E. Brezin and J. Zinn-Justin (Elsevier, 1989) 169.
[5] D. Friedan, in Recent Advances in Field Theory and Statistical Mechanics, Les Houches
Session XXXIX, edited by J.-B. Zuber and R. Stora (Elsevier, 1984) 839.
[6] M. Peskin, in From the Planck Scale to the Weak Scale, TASI 1986, edited by H. Haber
(World Scientic, 1987) 277.
[7] T. Banks, in The Santa Fe TASI-87, edited by R. Slansky and G. West (World Scientic,
1988) 572.
[8] S. B. Giddings, Phys.Rept. 170, 167 (1988).
[9] E. DHoker, in Recent Directions in Particle Theory, TASI 1992, edited by J. Harvey
and J. Polchinski (World Scientic, 1993) 1.
[10] M. Green, J. Schwarz, and E. Witten, Superstring Theory, Vols. 1 and 2 (Cambridge,
1987).
[11] A. M. Polyakov, Gauge Fields and Strings (Harwood, 1987).
[12] D. Lust and S. Theisen, Lectures on String Theory (Springer-Verlag, 1989).
[13] M. Kaku, Introduction to Superstrings (Springer-Verlag, 1988); Strings, Conformal
Fields, and Topology (Springer-Verlag, 1991).
[14] B. Hateld, Quantum Field Theory of Point Particles and Strings (Addison-Wesley,
1992).
[15] J. M. Droue and C. Itzykson, Statistical Field Theory, Vols. 1 and 2 (Cambridge,
1989).
144
[16] C. Itzykson, H. Saleur, and J.-B. Zuber, eds., Conformal Invariance and Applications
to Statistical Mechanics (World Scientic, 1988).
[17] J. Polchinski, Joes Big Book of String (Cambridge, 1995?).
[18] A. A. Belavin, A. M. Polyakov, and A. B. Zamolodchikov, Nucl. Phys. B241, 333
(1984).
[19] D. Friedan, E. Martinec, and S. Shenker, Nucl. Phys. B271, 93 (1986).
[20] S. Fubini, A. J. Hansen, and R. Jackiw, Phys. Rev. D7, 1732 (1973).
[21] A. Chodos and C. B. Thorn, Nucl. Phys. B72, 509 (1974).
[22] D. Friedan, Ann. Phys. 163, 318 (1985).
[23] E. S. Fradkin and A. A. Tseytlin, Phys. Lett. 158B, 316 (1985); 160B, 69 (1985).
[24] C.G. Callan, E.J. Martinec, M.J. Perry, and D. Friedan, Nucl. Phys. B262, 593 (1985).
[25] E. Witten, Comm. Math. Phys. 92, 451 (1984).
[26] V. Knizhnik and A. B. Zamolodchikov, Nucl. Phys. B247, 83 (1984).
[27] D. Gepner and E. Witten, Nucl. Phys. B278, 493 (1986).
[28] D. Friedan, Z. Qiu, and S. Shenker, Phys. Rev. Lett. 52, 1575 (1984).
[29] V. S. Dotsenko and V. A. Fateev, Nucl. Phys. B240, 312 (1984); B251 691 (1985).
[30] P. Goddard and D. Olive, Int. J. Mod. Phys. A1, 303 (1986).
[31] P. Ramond, Phys. Rev. D3, 2415 (1971).
[32] A. Neveu and J. H. Schwarz, Nucl. Phys. B31, 86 (1971).
[33] M. Ademollo, et. al., Nucl. Phys. B111, 77 (1976); B114, 297 (1976); Phys. Lett. 62B,
105 (1976).
[34] T. Eguchi and A. Taormina, Phys. Lett. B196, 75 (1987).
145
[35] C. N. Pope. L. J. Romans, and X. Shen, in Proceedings of Strings 90, edited by R.
Arnowitt, et. al. (World Scientic, 1991) 287; X. Shen, Int. J. Mod. Phys. A7, 6953
(1992) hep-th/9202072.
[36] P. Bouwknegt and K. Schoutens, Phys. Rept. B223, 183 (1993) hep-th/9210010.
[37] C. Vafa, Phys. Lett. 199B, 195 (1987).
[38] A. LeClair, M. E. Peskin, and C. R. Preitschopf, Nucl. Phys. B317, 411, 464 (1989).
[39] J. Polchinski, Nucl. Phys. B307, 61 (1988).
[40] H. Sonoda, Nucl. Phys. B311, 401,417 (1988).
[41] J. L. Cardy, Nucl. Phys. B270, 186 (1986).
[42] O. Alvarez, Nucl. Phys. B216, 125 (1983).
[43] E. DHoker and D. H. Phong, Rev. Mod. Phys. 60, 917 (1988).
[44] S. Weinberg, in General Relativity, An Einstein Centenary Survey, edited by S. W.
Hawking and W. Israel (Cambridge, 1979) 790.
[45] T. Yoneya, Nouvo Cim. Lett. 8, 951 (1973).
[46] J. Scherk and J. H. Schwarz, Nucl. Phys. B81, 118 (1974).
[47] P. Candelas, G. Horowitz, A. Strominger, and E. Witten, Nucl. B258, 46 (1985).
[48] U. Amaldi, W. de Boer, and H. Furstenau, Phys. Lett. B260, 447 (1991).
[49] A. M. Polyakov, Phys. Lett. B103, 207, 211 (1981).
[50] J. Polchinski and A. Strominger, Phys. Rev. Lett. 67, 1681 (1991).
[51] E. DHoker and S. B. Giddings, Nucl. Phys. B291, 90 (1987).
[52] R. C. Brower, Phys. Rev. D6, 1655 (1972); P. Goddard and C. B. Thorn, Phys. Lett.
40B, 235 (1972).
[53] D. M. Capper and M. J. Du, Nuovo Cim. 23A, 173 (1974); Phys. Lett. 53A, 361
(1975).
146
[54] L. Alvarez-Gaume and E. Witten, Nucl. Phys. B234, 269 (1984).
[55] T. Kugo and I. Ojima, Phys. Lett. 73B, 459 (1978); Suppl. Prog. Theor. Phys. 66, 1
(1979).
[56] M. Henneaux and C. Teitelboim, Quantization of Gauge Systems (Princeton, 1992).
[57] C. Becchi, A. Rouet, and R. Stora, Phys. Lett. 52B, 344 (1974); Ann. Phys. 98, 287
(1976).
[58] M. Kato and K. Ogawa, Nucl. Phys. B212, 443 (1983).
[59] M.D. Freeman and D.I. Olive, Phys. Lett. 175B, 151 (1986); C. B. Thorn, Nucl. Phys.
B286, 61 (1987); I.B. Frenkel, H. Garland, and G.J. Zuckerman, Proc. Natl. Acad. Sci.
83, 8442 (1986).
[60] D. J. Gross, J. A. Harvey, E. Martinec, and R. Rohm, Phys. Rev. Lett. 54, 502 (1985);
Nucl. Phys. B256, 53 (1985); B267, 75 (1986).
[61] M. B. Green and J. H. Schwarz, Nucl. Phys. B255, 93 (1985).
[62] H. Ooguri and C. Vafa, Nucl. Phys. B361, 469 (1991).
[63] S.-H. H. Tye, Status of Fractional Superstrings, presented at Strings 93, Berke-
ley (1993) hep-th/9311021; K. R. Dienes, Recent Developments in Fractional Super-
strings, presented at SUSY-93, Northeastern (1993) hep-th/9306156.
[64] C. N. Pope, W-Strings 93, presented at Strings 93, Berkeley (1993) hep-th/9309125.
[65] J. L. Cardy and D. C. Lewellen, Phys. Lett. B259, 274 (1991); D. C. Lewellen, Nucl.
Phys. B372, 654 (1992).
[66] E. Martinec, Nucl. Phys. B281, 157 (1987).
[67] B. Zwiebach, Nucl. Phys. B390, 33 (1993).
[68] W. Fischler and L. Susskind, Phys. Lett. B171, 383 (1986); 173, 262 (1986).
[69] P. Manseld, Nucl. Phys. B283, 551 (1987).
[70] J. H. Schwarz, Int. J. Mod. Phys. A4, 2653 (1989).
147
[71] V. Kaplunovsky, Phys. Rev. Lett. 55, 299 (1985).
[72] S. Jain, R. Shankar, and S. R. Wadia, Phys. Rev. D32, 2713 (1985).
[73] A. B. Zamolodchikov, Sov. J. Nucl. Phys. 44; J. L. Cardy, in Phase Transitions and
Critical Phenomena, Vol. 11, edited by C. Domb and J. L. Lebowitz (Academic, 1986).
[74] T. Banks, L. Dixon, D. Friedan, and E. Martinec, Nucl. Phys. B299, 613 (1988).
[75] C. L. Kane, M. P. A. Fisher, and J. Polchinski, Phys. Rev. Lett. 72, 4129 (1994) cond-
mat/9402108.
[76] K. Kikkawa and M. Yamanaka, Phys. Lett. B149, 357 (1984); N. Sakai and I. Senda,
Prog. Theor. Phys. 75, 692 (1986).
[77] A. A. Tseytlin, presented at the 2nd Journee Cosmologie, Observatoire de Paris (1994)
hep-th/9407099.
[78] T. Buscher, Phys. Lett. B159, 127 (1985); B194, 59 (1987); B201, 466 (1988).
[79] M. Dine, P. Huet, and N. Seiberg, Nucl. Phys. B322, 301 (1989).
[80] S. Giddings, J. Polchinski, and A. Strominger, Phys. Rev. D48, 5784 (1993).
[81] R. Dijkgraaf, E. Verlinde, and H. Verlinde, in Perspectives in String Theory, edited by
P. Di Vecchia and J. L. Petersen (World Scientic, 1988); A. Shapere and F. Wilczek,
Nucl. Phys. B320, 669 (1989).
[82] S.-T. Yau, ed., Essays on Mirror Manifolds, (International Press, 1992).
[83] A. Giveon and E. Witten, Phys. Lett. B332, 44 (1994) hep-th/9404184.
[84] J. Dai, R. Leigh, and J. Polchinski, Mod. Phys. Lett. A4, 2073 (1989).
[85] R. Leigh, Mod. Phys. Lett. A4, 2767 (1989).
[86] N. Berkovits and C. Vafa, Mod. Phys. Lett. A9, 653 (1994) hep-th/9310170.
[87] F. Bastianelli, N. Ohta, and J. L. Petersen, Phys. Rev. Lett. 73, 1199 (1994) hep-
th/9403150.
[88] N. Berkovits, M. Freeman, and P. West, Phys. Lett. B325, 63 (1994) hep-th/9312013.
148
[89] H. Kunitomo, M. Sakaguchi, and A. Tokura, A Universal w String Theory, Osaka
preprint OU-HET-187 (1994) hep-th/9403086; A Heirachy of Super w Strings, Osaka
preprint OU-HET-199 (1994) hep-th/9408007.
[90] H. Kunitomo, On the Nonlinear Realization of the Superconformal Symmetry, Osaka
preprint OU-HET-196 (1994) hep-th/9407052.
[91] A. Font, L. Ibanez, D. L ust and F. Quevedo, Phys. Lett. B249, 35 (1990).
[92] A. Sen, Int. J. Mod. Phys. A9, 3707 (1994) hep-th/9402002.
[93] P. A. M. Dirac, Proc. R. Soc. A133, 60 (1931).
[94] E. Witten, Phys. Lett. B86, 286 (1979).
[95] C. Montonen and D. Olive, Phys. Lett. B72, 117 (1977).
[96] P. Goddard, J. Nuyts, and D. Olive, Nucl. Phys. B125, 1 (1977).
[97] E. Witten and D. Olive, Phys. Lett. B78, 97 (1978).
[98] H. Osborn, Phys. Lett. B83, 321 (1979).
[99] A. Sen, Phys. Lett. B329, 217 (1994) hep-th/9402032.
[100] N. Seiberg and E. Witten, Nucl. Phys. B426, 19 (1994) hep-th/9407087.
[101] Monopoles, Duality, and Chiral Symmetry Breaking in N = 2 Supersymmetric QCD,
N. Seiberg and E. Witten, preprint RU-94-60 (1994) hep-th/9408099.
[102] A Strong Coupling Test of S-Duality, C. Vafa and E. Witten, preprint HUTP-94-
A017 (1994) hep-th/9408074.
[103] M. J. Du, Class. Quantum. Grav. 5, 189 (1988); A. Strominger, Nucl. Phys. B343,
167 (1990).
[104] W. Siegel, Phys. Lett. 151B, 391, 396 (1984).
[105] E. Witten, Nucl. Phys. B268, 253 (1986).
[106] C. Thorn, Phys. Rept. 175, 1 (1989).
149
[107] W. Siegel, Introduction to String Field Theory (World Scientic, 1988).
[108] R. Brustein and S. P. de Alwis, Nucl. Phys. B352, 451 (1991); R. Brustein and K.
Roland, Nucl. Phys. B372, 201 (1992).
[109] S. B. Giddings, E. Martinec, and E. Witten, Phys. Lett. B176, 362 (1986).
[110] W. Siegel, Phys. Rev. D49, 4144 (1994) hep-th/9312117.
[111] A. Strominger, Nucl. Phys. B294, 93 (1987).
[112] J. Polchinski, Comm. Math. Phys. 104, 37 (1986).
[113] J.-C. Le Guillou and J. Zinn-Justin, eds., Large-Order Behavior of Perturbation The-
ory, Current Physics Sources and Comments, Vol. 7 (North-Holland, 1990).
[114] D. Gross and V. Periwal, Phys. Rev. Lett. 60, 2105 (1988).
[115] S. Shenker, in Random Surfaces and Quantum Gravity, edited by O. Alvarez, E. Mari-
nari, and P. Windey (Plenum, 1991).
[116] J. Polchinski, Combinatorics of Boundaries in String Theory, preprint NSF-ITP-94-
52 (1994) hep-th/9407031.
[117] A. DAdda, M. L uscher and P. Di Vecchia, Nucl. Phys. B146, 63 (1978); B152, 125
(1979).
[118] E. Fradkin, Field Theories of Condensed Matter Systems (Addison-Wesley, 1991).
[119] S. Weinberg and E. Witten, Phys. Lett. 96B, 59 (1980).
[120] J. J. Atick and E. Witten, Nucl. Phys. B310, 291 (1988).
[121] C. B. Thorn, in Sakharov Memorial Lectures in Physics, proceedings of the First In-
ternational Sakharov Conference on Physics, Moscow, edited by L.V. Keldysh and V.
Ya. Fainberg (Nova Science, 1991) 447; hep-th/9405069.
[122] L. Susskind, The World as a Hologram, preprint SU-ITP-94-33 (1994) hep-
th/9409089.
[123] D. J. Gross and P. F. Mende, Phys. Lett. 197B, 129 (1987); Nucl. Phys. B303, 407
(1988).
150
[124] P. F. Mende and H. Ooguri, Nucl. Phys. B339, 641 (1990).
[125] D. J. Gross, Phys. Rev. Lett. 60, 1229 (1988).
[126] D. Amati, M. Ciafaloni, and G. Veneziano, Phys. Lett. B216, 41 (1989).
[127] L. Susskind, Phys. Rev. Lett. 71, 2367 (1993) hep-th/9307168.
[128] A. Mezhlumian, A. Peet, and L. Thorlacius, Phys. Rev. D50, 2725 (1994) hep-
th/9402125.
[129] P. Salomonson and B.-S. Skagerstam, Nucl. Phys. B268, 349 (1986).
[130] D. Mitchell and N. Turok, Phys. Rev. Lett. 58, 1577 (1987); Nucl. Phys. B294, 1138
(1987).
[131] R. Brandenberger and C. Vafa, Nucl. Phys. B316, 391 (1989).
[132] D. A. Lowe and L. Thorlacius, Hot String Soup, preprint NSF-ITP-94-83, UCSBTH-
94-30 (1994) hep-th/9408134.
[133] P. Ginsparg and G. Moore, in Recent Directions in Particle Theory, Proceedings of
the 1992 TASI, edited by J. Harvey and J. Polchinski (World Scientic 1993) hep-
th/9304011.
[134] I. Klebanov, in String Theory and Quantum Gravity, Proceedings of the Trieste School
1991, edited by J. Harvey et. al. (World Scientic, 1992) hep-th/9108019.
[135] R. C. Myers, Phys. Lett. B199, 371 (1988); I. Antoniades, C. Bachas, J. Ellis, and D.
Nanopoulos, Phys. Lett. B211, 393 (1988).
[136] N. Seiberg, Prog. Theor. Phys. Suppl. 102, 319 (1990).
[137] J. Polchinski, in Proceedings of Strings 90, edited by R. Arnowitt, et. al. (World
Scientic, 1991) 62.
[138] J. Polchinski, Nucl. Phys. B346, 253 (1990).
[139] E. Brezin, C. Itzykson, G. Parisi, and J.-B. Zuber, Comm. Math. Phys. 59, 35 (1978).
[140] E. Brezin, C. Itzykson, and J.-B. Zuber, Adv. App. Math. 1, 109 (1980).
151
[141] V. A. Kazakov and A. A. Migdal, Nucl. Phys. B311, 171 (1988).
[142] E. Brezin, V. A. Kazakov, and A. B. Zamolodchikov, Nucl. Phys. B333 (1990) 673;
P. Ginsparg and J. Zinn-Justin, Phys. Lett. B240 (1990) 333; G. Parisi, Phys. Lett.
B238, 209 (1990); D. J. Gross and N. Miljkovic, Phys. Lett. B238, 217 (1990).
[143] D. Gross and A. Migdal, Phys. Rev. Lett. 64, 127 (1990); M. Douglas and S. Shenker,
Nucl. Phys. B335, 635 (1990); E. Brezin and V. Kazakov, Phys. Lett. 236, 144 (1990).
[144] T. L. Curtright and C. B. Thorn, Phys. Rev. Lett. 48, 1309 (1982); V. Knizhnik, A.
Polyakov, and A. Zamolodchikov, Mod. Phys. Lett. A3, 819 (1988); F. David, Mod.
Phys. Lett. A3, 1651 (1988); J. Distler and H. Kawai, Nucl. Phys. B321 (1989) 509.
[145] E. DHoker, in Particle Physics, VI J. A. Swieca Summer School, edited by O. Eboli,
M. Gomes, and A. Santoro, (World Scientic, 1992) 282.
[146] S. R. Das and A. Jevicki, Mod. Phys. Lett. A5, 1639 (1990).
[147] J. Polchinski, Nucl. Phys. B362, 125 (1991).
[148] A. M. Sengupta and S. R. Wadia, Int. J. Mod. Phys. A6, 1961 (1991).
[149] I. Klebanov, private communication.
[150] G. Moore, M. R. Plesser, and S. Ramgoolam, Nucl. Phys. B377, 143 (1992) hep-
th/9111035.
[151] E. Smith, Nucl. Phys. B382, 229 (1992) hep-th/9201049; Phys. Lett. B305, 344 (1993)
hep-th/9208039; R. E. Rudd, Ph. D. Thesis, Princeton (1992) unpublished.
[152] N. Ishibashi and H. Kawai, Phys. Lett. B314, 190 (1993) hep-th/9307045; M. Ikehara,
N. Ishibashi, H. Kawai, T. Mogami, R. Nakayama, and N. Sasakura, String Field
Theory in the Temporal Gauge, preprint KEK-TH-402 (1994) hep-th/9406207.
[153] Z. Yang (Texas U.), Phys. Lett. B257, 40 (1991); J. A. Minahan, Int. J. Mod. Phys.
A8, 3599 (1993) hep-th/9204013.
[154] V. A. Kazakov and I. K. Kostov, Nucl. Phys. B386, 520 (1992) hep-th/9205059; I.
K. Kostov (Saclay), Field Theory of Open and Closed Strings with Discrete Target
Space, preprint SACLAY-SPHT-94-097 (1994) hep-th/9410164.
152
[155] E. Marinari and G. Parisi, Phys. Lett. B240 (1990) 375.
[156] S. Chaudhuri and J. Polchinski, Critical Behavior of the Marinari-Parisi Model,
preprint NSF-ITP-94-53 (1994) hep-th/9408005.
[157] G. Moore and R. Plesser, Phys. Rev. D46, 1730 (1992) hep-th/9203060.
[158] J. Polchinski, On the Nonperturbative Consistency of d = 2 String Theory, preprint
NSF-ITP-94-96 (1994) hep-th/9409168.
[159] P. Di Francesco and D. Kutasov, Phys. Lett. B261, 385 (1991); Nucl. Phys. B375,
119 (1992).
[160] A. M. Polyakov, Mod. Phys. Lett. A6, 635 (1991); in Two Dimensional Quantum
Gravity and Random Surfaces, 8th Jerusalem Winter School for Theoretical Physics,
Jerusalem, edited by D.J. Gross, T. Piran and S. Weinberg (1990-91) 175.
[161] D. J. Gross and I. R. Klebanov, Nucl. Phys. B359, 3 (1991).
[162] E. Witten, Phys. Rev. D44 (1991) 314.
[163] G. Mandal, A Sengupta, and S. Wadia, Mod. Phys. Lett. A6 (1991) 1685.
[164] R. Dijkgraaf, E. Verlinde, and H. Verlinde, Nucl. Phys. B371, 269 (1992).
[165] E. Witten, Nucl. Phys. B373, 187 (1992) hep-th/9108004; E. Witten and B. Zwiebach,
Nucl. Phys. B377, 55 (1992) hep-th/9201056.
[166] S. Mukherji, S. Mukhi, and A. Sen, Phys. Lett. B275, 39 (1992).
[167] J. Avan and A. Jevicki, Phys. Lett. B266, 35 (1991).
[168] G. Moore and N. Seiberg, Int. J. Mod. Phys. A7, 2601 (1992).
[169] S. R. Das, A. Dhar, G. Mandal, and S. R. Wadia, Int. J. Mod. Phys. A7, 5165 (1992)
hep-th/9110021.
[170] D. Minic, J. Polchinski, and Z. Yang, Nucl. Phys. B369, 324 (1992).
[171] M. Natsuume and J. Polchinski, Nucl. Phys. B424, 137 (1994) hep-th/9402156.
153
[172] M. Natsuume, Zero Mode Divergence Problem in String Theory, preprint NSF-ITP-
93-157 (1993) hep-th/9401013.
[173] N. Sakai and Y. Tanii, Prog. Theor. Phys. Suppl. 110, 117 (1992) hep-th/9111049;
Phys. Lett. B276, 41 (1992) hep-th/9108027.
[174] D. Minic and Z. Yang, Phys. Lett. B274, 27 (1992) hep-th/9109012.
[175] D. Lowe, Mod. Phys. Lett. A7, 2647 (1992) hep-th/9204084.
[176] A. Dhar, G. Mandal, and S. R. Wadia, Mod. Phys. Lett. A7, 3703 (1992) hep-
th/9210120; ; A8 1701, (1993) hep-th/9304072; S. R. Das, Mod. Phys. Lett. A8, 69
(1993) hep-th/9210107; A8, 1331 (1993) hep-th/9303116; A. Jevicki and T. Yoneya,
Nucl. Phys. B411, 64 (1994) hep-th/9305109.
154