NIKHEF/2007-020

IMAFF/FM-07/02
Introduction to String Theory

Based on lectures given at the Radboud Universiteit, Nijmegen.

1
These notes follow rather closely the course given in the fall semester of 1999. They
are based to a large extent on the books by Green, Schwarz and Witten, [1], the book
by Polchinski [2], the book of L ust and Theisen [3], and the review article [5]. I refer to
these sources for further references. A very useful introductory text that appeared after
the original version of these notes was written is the book by Zwiebach [4].
1 Current problems in particle physics
Before getting into string theory, it may be good to know to which questions we hope it
will provide answers.
At present, we know about four fundamental interactions. Three of them, electromag-
netism, the weak and the strong interaction are very successfully described by a theory
called modestly the Standard Model. The standard model seems to be in very good
shape experimentally. All quantities that we can compute reliably, and that we can mea-
sure, are in agreement with each other. There is only one thing that we are quite sure
is missing. The standard model predicts (at least) one additional particle needed for its
consistency, namely the Higgs boson. If it is found, which indeed should happen within
the next two decades, there is really no question left within the standard model that
requires an answer.
Certainly there are many questions which we may hope to see answered. We may
wish to compute the quark and lepton masses, explain hierarchies of scales, or understand
the general structure of the model, but without internal inconsistencies or conicts with
experiment we may have no other choice than to accept the model as it is. Many ideas
have been proposed to address some of these questions. Examples are Grand Unication
or Supersymmetry. Usually they improve the Standard Model in some ways, while
making other problems much worse. Supersymmetry, for example, improves the hierarchy
problem, but at the price of a drastic increase in the number of parameters. One can even
imagine extensions of the Standard Model that from our limited point of view are not
improvements at all. There may be additional particles with extremely weak interactions,
that do not seem to do anything useful. If we are lucky we may get experimental
information about what, if anything, lies beyond the Standard Model, but it seems at
least as plausible that we will never know.
The fourth interaction, gravity, is in a less good shape theoretically. Classically, it
is described by Einsteins theory of General Relativity. Experimentally this theory in
impressive agreement with experiment, but when one tries to quantize it one gets into
serious diculty. This is a fundamental problem, which does require a solution. Unlike
the standard model, where the remaining problems are mainly aesthetic, quantum gravity
has serious internal inconsistencies (more about this below). The strongest point in favor
of string theory is that it seems to solve these problems. At the moment it is the only
serious candidate for a consistent theory of quantum gravity. This does not mean that
there is a proof that all problems of quantum gravity are solved, nor that it is known that
no other possibilities exist.
2
This fact would already be enough to study String Theory seriously. But there is more.
String Theory does not just contain gravity, it comes inevitably with a large number of
other particles and interactions. These particles and interactions have the same features as
the Standard Model. Typically there are quark and lepton-like particles, coupling to gauge
bosons. Furthermore there are usually some scalars. The interactions are often chiral,
which means that left- and right-handed particles couple in dierent ways. This feature is
observed in the weak interactions. In fact any observed property of the Standard Model
can be identied within String Theory. Among the possibilities are also Grand Unication
or Supersymmetry.
Unfortunately, the particles and interactions String Theory predicts in addition to
gravity are very far from unique. There are so many possibilities that at the moment
nobody is able to nd out whether the Standard Model is among the possibilities. Nev-
ertheless, if String Theory is correct as a theory of gravity, there is really no other option
than that it is also a theory of all other interactions. According to our present under-
standing, String Theory is far more robust than the usual eld-theoretic description of
particles and interactions. We cannot take a String Theory description of gravity and
simple add the Standard Model to it. One cannot add particles or interactions to a given
String Theory, or remove particles one does not like.
This seems almost in contradiction with the fact that String Theory is far from unique.
The dierence between the String Theory and Field Theory description of nature is illus-
trated by the following picture
A B
A
B
The rst picture shows two possible string theories, the second two possible eld
theories. The idea is that in eld theory we can extend theory A to theory B without
throwing any part of theory A away. In String Theory we cannot do that. We can modify
theory A to theory B by throwing part of theory A away and adding something new in
its place. Suppose we know experimentally that we need everything in theory A. Then
String theory would actually predict that there are no other particles or interactions. If
an experiment does nd something not predicted by theory A, we would be forced to give
up String Theory altogether. On the other hand, if we are in the situation of the second
picture, we could simply add something to our theory.
For this reason one sometimes uses the phrase Theory of Everything when talking
about string theory. Not everybody thinks it is a good idea to use this phrase.
3
1.1 Problems of quantum gravity
One of the rst problems one encounters when one attempts to quantize a eld theory is
that certain calculations seem to lead to innities. For example, loop diagrams may lead
to integrals of the following type
_
d
4
p
1
(p
2
+ m
2
)
2
(1.1)
where p

is a four-momentum, and m some mass. It is easy to see that this integral

diverges for large p.
If one thinks for a moment about what one is doing here, it seems rather silly to inte-
grate over arbitrarily large momenta. Large momenta correspond to short distances, and
by integrating to innity, we are pretending that we understand physics to arbitrarily short
distances. This is a ridiculous assumption. Perhaps space-time is discrete at extremely
small distances, so that there is a maximum momentum, the inverse of the smallest dis-
tance. Perhaps something else changes at short distances. If anything changes at short
distances, it may happen that our integral is actually nite, proportional to log(), where
is some large, but nite, momentum scale. The trouble is that without detailed knowl-
edge about the short distance physics we cannot compute the integral, all we can do is
introduce a new parameter equal to the value of the integral. However, in quantum eld
theory there are integrals of this type for any loop diagram, and if we cannot compute
any of them we would be forced to introduces an innity of new parameters. To know the
theory completely we would have to measure all these parameters, and that is obviously
impossible.
However, some theories are nicely behaved. In such theories the unknown parameters
appear only in (innite) linear combinations such as
m + m
1
+ m
2
+ m
3
. . . (1.2)
where m is a parameter of the original theory (for example a mass), while m
1
. . . are the
unknown parameters from loop diagrams. If all physical quantities depend only on this
particular combination there is no need to determine them individually. We simply give
the whole sum a new name, m
R
, and if we know the value of m
R
we know enough. The
net result of this procedure is that we end up with an equal number of parameters as we
started with. If a theory has this property, we say that it is renormalizable. The Standard
Model has this property, and for this reason we regard it as a consistent quantum theory.
But gravity does not have this property. In the case of gravity the loop corrections
lead to an innity of new parameters that cannot be absorbed into the ones we started
with (there is just one, namely Newtons constant). Strictly speaking that means that
any physical quantity we compute in quantum gravity depends on an innite number of
parameters, and hence the predictive power of the theory is zero. This is of course an
exaggeration. If true, it would imply that we cannot even believe gravity in its classical
limit, where it is so successful.
4
It turns out that for any reasonable value of the unknown parameters, their eect is
completely negligible at observable energies.
The coupling constant of gravity is Newtons constant G
N
, dened through
F = G
N
m
1
m
1
r
2
(1.3)
Using and c we can construct from it a mass, the Planck mass
M
P
=
_
c
G
N
2 10
8
kg (1.4)
In high energy physics it is customary to multiply masses with c
2
and express them in
units of energy, for which the GeV is used. Then we get
M
P
=

c
5
G
N
= 1.22 10
19
GeV (1.5)
Using other combinations of and c we can convert this to an inverse length. This length,
the Planck length is 1.6 10
33
cm. Usually we use units where = c = 1, so that the
distinction between energy, mass and inverse length disappears.
The gravitational potential can be obtained from graviton exchange between the two
particles. The rst quantum correction is due to two-graviton exchange, and is pro-
portional to G
2
N
. The ratio of the two and the one-graviton term must therefore be
proportional to G
N
. Since the ratio must be dimensionless, it must be of the form G
N
E
2
,
where E is some energy of the process under consideration. Hence we see that the rst
correction must be proportional to (E/M
P
)
2
, which is extremely small as long as E is
much smaller than 10
19
GeV, and as long as the unknown coecient is of order 1. This
is far beyond anything we can ever hope to achieve with accelerators.
So the problem of nding a quantum theory of gravity is a purely theoretical one.
There does not appear to be any chance of observing its eects directly.
There are other problems associated with quantum mechanics and gravity. Some very
interesting problems arise in the discussion of black holes. One of the most intriguing
features of a black hole is the existence of a horizon. Outside observers cannot get infor-
mation about what happens inside that horizon. They can allow themselves to fall into the
black hole, inside the horizon, but then they cannot communicate with the outside world
anymore. Objects thrown into the black hole cannot be recovered. This seems to clash
5
with unitarity of the quantum mechanical S-matrix, which states that the total proba-
bility is conserved. Indeed, it has been proposed that the existence of black holes would
force us to give up unitarity of the S-matrix. This is called the information-loss prob-
lem. It is made more severe because of black hole evaporation. Given long enough time,
a black hole radiates away all its energy through Hawking radiation until it evaporates
completely. Then apparently all the information that went into it disappears forever.
Another related problem is that of the black hole entropy. Bekenstein and Hawking
have shown that a black hole has thermodynamic properties: it has black body radiation
with a certain temperature, and it behaves as if it has a certain entropy, proportional
to the area of the horizon. Normally one associates entropy with the logarithm of the
number of states of an object, but such a notion was not available here. Presumably
understanding the counting of states of a black hole in relation to its entropy will also tell
us how it can store information.
String theory has made interesting contributions to both problems. First of all string
theory does indeed change the short distance physics. It has been shown that in certain
string theories the unknown or innite corrections are in fact nite and calculable. Fur-
thermore it has been possible to obtain a description of the states of special black holes
(in special limits) so that the correct entropy could be obtained.
All of this is still very vague, so it is time to take a closer look.
1.2 String diagrams
In our present description of fundamental physics the basic objects are particles leptons,
quarks, gauge bosons etc. The starting point of String Theory is the assumption that the
basic objects have a one-dimensional extension, that they are like small pieces of rope.
This principle allows two kinds of fundamental objects: open and closed strings. When
such objects move through space they sweep out ribbons or cylinders.
Particle Open string
Closed string
t
t
t
There is still one more distinction to be made: strings may or may not have a denite
orientation. Depending on this we distinguish oriented and unoriented strings. All to-
gether there are at this point then four types of strings: open unoriented, open oriented,
closed unoriented and closed oriented. Oriented strings must keep their orientation when
they interact.
Strings interact by splitting and joining. We can represent this by plotting the strings
as a function of time. For example the following picture represents an open string splitting
6
t
It is also easy to draw the splitting of closed strings
t
A more interesting process is that of an open string decaying into an open and a closed
string.
t
This shows that open strings necessarily imply closed strings. The other way around
this is not true. It is impossible to draw a diagram where a closed string emits an open
one.
In general a string diagram can be any two-dimensional surface. Such a surface can
be interpreted in various ways by choosing a time coordinate on it. Then one can give a
string interpretation to the interactions by slicing the surface with planes orthogonal to
the time direction. For example, the previous diagram is obtained by time-slicing in a
suitable way a piece of a plane to which a cylinder is attached.
On a cylinder there are two obvious ways of choosing a time coordinate
7
t
Closed string
propagation
Open string loop
t
=
t
One corresponds to the propagation of a closed string, the other to an open string
making a closed loop.
One of the most important features of string theory is that to each such diagram there
corresponds a certain amplitude. We may try to interpret that amplitude anyway we like,
but there is only one amplitude. For example, to compute the scattering amplitude for
two closed strings to two closed strings one begins by drawing all valid diagrams with two
initial and two nal closed strings, propagating from or to + respectively.
There is a precise procedure that assigns to each such diagram a number, the transition
amplitude. This procedure is based on path integrals. These are innite dimensional
integrals over all surfaces with certain initial and nal states one is interested in. The
surfaces are weighted by an exponential; very roughly the time evolution from initial to
nal states is then given by
out[ U [in) =
_
Surfaces
e
iS/
, (1.6)
where S is the (classical) action. In the simplest case S is simply the area of the surface.
Of course such innite dimensional integrals require a much more detailed discussion.
Which diagrams are allowed depends on the kind of string theory one considers. In
the simplest case, oriented closed strings, the allowed diagrams would be the sphere,
the torus, the double torus, etc, each with four innite tubes attached. Two diagrams
are considered equivalent if we can deform them into each other in a continuous way,
without breaking anything. It is a basic result in mathematics that all such surfaces
are characterized topologically by the number of handles (the genus), i.e. a torus is a
sphere with one handle, etc. In the case of oriented open and closed strings one must
allow the surface to have holes, corresponding to the ends of open strings. In the case of
unoriented closed strings one cannot have holes, but one must allow closed surfaces that
do not have a denite orientation, such as the Klein bottle (a surface has a well-dened
orientation if one can consistently dene a normal vector in a continuous way, on all points
8
on the surface). Finally, in the case of unoriented open strings one must also allow open,
unoriented surfaces. The best known example is the Moebius strip.
External lines are either tubes glued to the surface itself, or strips glued to the bound-
aries. The latter correspond of course to external open strings.
All of this may be compared to the Feynman diagram expansion of quantum eld
theory. One obvious dierence is that in that case one obtains a strongly increasing
number of diagrams in each order. This is already clear if we look at a theory with
a three-point interaction. Also in quantum eld theory there is a precise calculational
procedure for computing an amplitude belonging to such a diagram.
At lowest order, there are three diagrams, whereas in oriented closed string theory there
is just one. At one loop order, there is still only one string diagram, but many more eld
theory diagrams.
But there is a more important dierence between the two situations. In eld theory
diagrams there are special points corresponding to the splitting of a particle into two
(or more) particles. This implies that we have to give additional information about the
strength of the interaction at these points. In a string theory there are no special points.
The surfaces are smooth, and every point is equivalent to any other point. If we time-slice
a surface, there may appear to be special points where the string splits. But this depends
on how we dene the equal-time slices. This is not a Lorentz-invariant concept, and hence
two dierent observers will not agree on a point on the two-dimensional surface which is
special. Hence there are no special points.
One consequence is that by specifying a free string theory, we have already specied
the interactions. As we will see, a string theory is specied by giving an action. The
action is always some two-dimensional Lagrangian density, integrated over the surface. In
the simplest cases the action is nothing but the area of the surface. The action does not
know about the topology of the surface. Computing dierent string diagrams amounts
then to computing the integrals of the same integrand over dierent surfaces.
The fact that string theory has no ultra-violet (short distance) divergencies is also
related to this fact. The divergencies in quantum eld theory can be traced back to the
existence of pointlike interactions, and can in fact be removed by smearing out the
interactions over a nite region. But in string theory there are no point-like interactions
to begin with.
String theory may nevertheless have divergencies, but they are of a dierent kind, and
can be avoided in certain cases.
9
2 Bosonic string action
The simplest string theory is the bosonic string. Although simple, it has serious defects:
its spectrum contains no space-time fermions, but it does contain a particle with imaginary
mass, a tachyon. It is therefore not a candidate for a theory of nature, but it is very useful
as a kind of toy-model.
The description of the bosonic string is modelled on that of a relativistic point particle.
We could describe the motion of such a particle by giving its position as a function of
time, x(t). However, this does not allow a Lorentz-invariant description. Instead one
could choose to specify X

(t) with the condition X

0
(t) = t, but this is still not Lorentz
invariant. The solution is to introduce a separate variable to parametrize the world line
of the particle. Each point on the world line is then specied by a set of D functions
X

(). We work here in D space-time dimensions. We will set the speed of light, c, equal
to 1.
For the action of such a relativistic point particle we may take the length of the world
line, i.e.
S = m
_
d
_

dX

d
dX

d
(2.1)
The minus sign in the square root reects our choice of metric, namely (, +, +, +). For
example a particle at rest at the origin can be described by X
0
= const , X
i
= 0,
which yields a positive square root argument. An important feature of this action is that
the answer does not depend on how we choose to parametrize the world line, i.e. if we
replace to any (monotonic) function f(). This is important, because it is the world
line itself that matters, not how we choose to parametrize it. When one considers the
equations of motion it turns out that m is the mass of the particle.
2.1 The Nambu-Goto action
As already mentioned above, we want to apply the same idea to String Theory. Now
we generalize from lines to surfaces, which are now called world sheets. To specify the
embedding of a world sheet in space-time we give D functions X

(, ). Now there are

two variables, one to describe the direction along the string (), and one to parametrize
a time-like direction. The action is proportional to the surface-area of the world sheet,
which can be written as
S[X] =
1
2

_
dd
_
det h

(2.2)
where h

is a two-by-two matrix dened as

h

, (2.3)
where

means

, and where instead of and we use variables

, = 0, 1, with

0
= and
1
= . The sign of the determinant is determined by the space-time metric. A
10
string at rest could be given by X
0
= , X
i
= X
i
(). Then h
00
= 1 and h
11
= (

X
i
)
2
,
h
01
= h
10
= 0, so that the determinant is negative. Then it will stay negative under
continuous deformations (since det h = 0 implies the manifold is singular at that point).
The action (2.2) is called the Nambu-Goto action.
2.2 The free boson action
It has the unpleasant property that it contains a square root. However, by introducing an
extra set of variables

(, ) we can get a dierent action which is classically equivalent

S[X, ] =
1
4

_
dd
_
det

(2.4)
Here

is a symmetric, non-singular two-by-two matrix that may depend on and .

By

we mean the inverse of

. The determinant of

is assumed to be negative
here; this will be justied in a moment.
To see the classical equivalence, note that

is not a dynamical degree of freedom:

it occurs without any derivatives. Hence we can eliminate it by using the equations of
motion. The latter are dened by requiring that the variation of the action with respect
to

vanishes:

S = 0 (2.5)
The action depends on

via the determinant and the inverse. The variation of the

inverse is computed as follows:
= 1 [ ] = 0 (2.6)
Therefore, using the chain rule
[ ] + = 0 (2.7)
Multiplying with we get then
[ ] = [] (2.8)
To vary

det we proceed as follows

[
_
det ] = [exp ( Tr (
1
2
log()))] =
1
2
(
_
det )[ Tr log()] (2.9)
The variation of the logarithm of a matrix can be computed in the same way as
x
log(x):
[log()] log(( + )) log()
= log(( + ) )
= log(1 + [] )
= []
(2.10)

As indicated,

(, ) may depend on and , but to simplify the notation we drop the arguments
as long as they are not essential.
11
The nal result is then
[
_
det ] =
1
2
_
det Tr (2.11)
The variation of the action is therefore

S =
1
4

_
dd
_
det [
1
2

]()

, (2.12)
with implicit summation over all repeated indices. This can only vanish for arbitrary
()

if the terms between square brackets add up to zero; this yields the matrix identity
h =
1
2
Tr ( h) (2.13)
Multiplying on both sides with gives
h =
1
2
Tr ( h) (, ) (2.14)
This allows us to express in terms of h, up to some function . When we substitute this
into the action S[X, ] the function drops out, and we do indeed obtain the Nambu-Goto
action S[X].
It is customary to introduce upper and lower indices (analogous to the space-time
Lorentz indices) also for the parameters

. Then one denes

, (2.15)
and we use this matrix to raise indices, and

to lower them. Since these matrices are

eachothers inverse, this makes sense. The action reads now
S[X, ] =
1
4

_
dd

, (2.16)
where denotes the determinant of

. From the relation between and h we see that

the natural sign of the determinant is negative. This action can be recognized as that of
D free scalar elds X

, = 1, . . . , D coupling to a two-dimensional metric

.
2.3 Symmetries
The bosonic string action S[X, ] has the following symmetries
Poincare invariance in D dimensions. This transformation acts only on X, not on

(, ) =

(, ) + a

(2.17)
12
Reparametrization invariance in 2 dimensions. This is taken over from the reparametriza-
tion invariance of the Nambu-Goto action, and must be present for the same reason.
It is also called dieomorphism invariance, and is in fact nothing but a general co-
ordinate transformation in two dimensions. Such a transformations is performed as
follows. We can introduce a new function X

) which describes the space-time

position of every point on the string in terms of a new parametrization

(, ) and

(, ). By denition this new function is therefore X

) = X

(, ). Now
we write the action entirely in terms of the new variables X

and

. For most
variables, this simply implies that we replace them in (2.4) by the primed variables.
In particular, dd is replaced by d

and

by

. Any action could be rewrit-

ten in this manner, but this is only a symmetry if we re-obtain the original action
when we express the primed variables back in terms of the original ones. In this
particular, changing back to the original variables introduces two potential prob-
lems: the change of integration measure when expressing

in terms of

, and the
change from

back to

. Even without prior knowledge of general relativity, one

realizes that in order for this to be a symmetry we have to transform as well. The
complete transformation is then:
X

) = X

(, )

) =

(, ) ,
(2.18)
where the new coordinates

and

are functions of the old ones and . Note

that we are free to change as we wish. If we manage to nd a transformation that
gives us back the original action, then we can call the combined transformation of
X and a symmetry.
Weyl invariance. This is a symmetry not seen in the Nambu-Goto action, and is a
consequence of the free function we found above. Indeed, it is easy to see that
the action is invariant under

(, ) = (, )

(, ) (2.19)
without changing X

.
The last two symmetries are redundancies of the two-dimensional theory on the world
sheet. his means that the two-dimensional action has fewer variables than it seems to have.
This is completely analogous to gauge symmetry in electrodynamics: the invariance under
the gauge transformation A

implies that one degree of freedom of A

can
be removed (in addition, A
0
is non-dynamical, so that the number of physical degrees of
freedom of the photon in D dimensions is reduced from D to D 2). As is also familiar
from electrodynamics, one may x the redundancy by choosing a gauge. This has to be
done in a clever way in order to make quantization of the theory manageable.
13
2.4 Conformal gauge
One way to x the gauge in our case is to x the two-dimensional metric

. It is in-
structive to consider the system in an arbitrary number N of world sheet dimensions.
The action and the rst two symmetries then remain exactly the same, but the Weyl
symmetry is broken:

transforms with
1
, and

transforms with a factor
N/2
.
They cancel only for N = 2. The number of components of the metric is
1
2
N(N + 1).
Using reparametrization invariance one can put N components of the metric to zero (at
least locally, in a nite neighborhood of a point), reducing the number of components
to
1
2
N(N 1). In two dimensions the extra Weyl invariance then allows us to remove
one more degree of freedom, so that nally none is left! This illustrates why two dimen-
sions is special, and why string theory is much easier than theories of membranes, whose
worldsheets are N-dimensional, with N > 2.
We can use this freedom to transform the metric to the at, Minkowski metric
ab
=
diag (1, 1) (with a 1 in the direction). This is called conformal gauge. Then the
action looks even simpler

S[X] =
1
4

_
dd
ab

a
X

b
X

(2.20)
This is the simplest action one can write down in two-dimensional eld theory. It is the
action of a free boson. More precisely, of D free bosons, because the index on the bosons
is from the two-dimensional point of view nothing other than a multiplicity. The action
S[X, ] we had earlier is in fact the action of a free boson coupled to two-dimensional
gravity.
2.5 world sheet versus space-time
The two-dimensional metric

(the world sheet metric) should not be confused with the

D-dimensional metric g

. We can make that metric manifest in the action by writing it

as
S[X] =
1
4

_
dd

(2.21)
Here g

may depend on the space-time position, but not on the word-sheet coordinates

a
, like and X. Hence from the two-dimensional point of view and X are elds,
but g

is not. If we are considering a string theory embedded in at, Minkowski space-

time, we must choose g

, = diag (1, 1, . . . , 1). One may also consider a string

propagating through curved space-time. This is obtained by simply substituting the
metric g

of that space-time. In these lectures, however, we restrict ourselves to at

space-times. We remarked above that is just a multiplicity. This is not quite true:
because we embed the string in minkowski space, one of the bosons will have an action
of the wrong sign. This has important consequences, as will be discussed later.

Our convention is to use indices , , . . . for world-sheet indices in arbitrary metrics, and a, b, . . . in
conformal gauge.
14
2.6 The equations of motion
Since the action S[X, ] depends on two elds X and there are two equations of motion.
If you are familiar with general relativity, you can immediately write down the one for the
two-dimensional metric: this of course leads to the Einstein equations, R

, where
the left hand side is the Ricci tensor. This term is derived from the gravitational action
_

R, which we do not have in our action.

Hence the equation of motion is simply

T

= 0. If you are not familiar with general relativity you can derive this by varying
S[X, ] with respect to . This is exactly the calculation we did earlier. The denition of
the Energy-Momentum tensor is
T

(, )
4

S[X, ] (2.22)
The factor 4

is for convenience. The computation yields

T

(, ) =

1
2

(2.23)
To compute this, compute rst

S[X, ] as was done before. This yields an integral of

the form

S =
_
d

Tr (F(

)) , (2.24)
for some matrix function F. Now we use the functional derivative

(, )
=

)(

) , (2.25)
which is functional analog of the discrete variable result
q
i
q
j
=
ij
(2.26)
The other equation of motion is the one for X

. Although it can be computed for

arbitrary , for simplicity we go to conformal gauge. Then the equation is simply
_

2

2


2

2
_
X

= 0 (2.27)
In two dimensions this equation can be solved very easily by writing it interms of variables

+
= + and

= . The general solution is

X

(, ) = X

L
( + ) + X

R
( ) , (2.28)
where X

L
and X

R
are arbitrary functions. A special point of X
L
, for example the max-
imum, stays at a xed value of + . Hence as increases, must decrease. Then the
maximum of X
L
moves to the left, and analogously the maximum of X
R
moves to the
right. Therefore we call these left- and right-moving components.

It can be added, and plays an interesting role. In two dimensions this term is however purely topological,
and does not contribute to the equations of motion.
15
2.7 Conformal invariance
It is known from classical electrodynamics that in certain cases some symmetries are left
even though the gauge was xed. For example, if we x the Lorentz gauge,

= 0,
then we still have not xed gauge transformations A

, with = 0.
The same occurs here. It turns out that there are special coordinate transformations
that change the metric by an overall factor:

) = (, )

(, ) (2.29)
Then one can remove using a Weyl transformation, and the net eect is that the metric
has not changed at all. Hence if we start with the conformal gauge metric
ab
we keep
the same metric, and therefore we have a residual symmetry.
This problem can be formulated in any dimension. One can completely classify all
transformations of the coordinates that satisfy (2.29). It should be clear that this in-
cludes in any case (Lorentz)-rotations,
a
=
a
b

b
, and translations,
a
+
a
, since they
do not change the metric at all. Another transformation with this property is a scale
transformation,
a
=
a
.
The name of this kind of a transformation derives from the fact that it preserves
angles, but not lengths. Consider two vectors v

and w

, which may be in any dimension.

The angle between these vectors (or its obvious generalization to Minkowski space) is
v

_
(v

)(w

)
(2.30)
A scale factor (, ) in the metric obviously drops out in the ratio, so angles are indeed
preserved. So in some sense these transformation preserve the form of an object, but
not its size. This is why they are called conformal transformations.
Just as symmetries in other systems, these transformations form a group, called the
conformal group. We have already seen that it contains the Lorentz group as a subgroup.
In p space and q time dimensions, the Lorentz group is SO(p, q). The conformal group
turns out to be SO(p + 1, q + 1), if p + q > 2.
If p + q = 2, however, the conformal group is much larger. This can easily be seen as
follows. Let us try the transformation

= f( + ) + g( )

= f( + ) g( )
(2.31)
where f and g are arbitrary continuous functions. The transformation of the Minkowski
metric
ab
under such a transformation leads to the matrix multiplication
_
f

+ g

+ g

__
1 0
0 1
__
f

+ g

+ g

_
= 4f

_
1 0
0 1
_
(2.32)
which proves that the metric transforms to a multiple of itself, and hence that this is
a conformal transformation. Thus we see that we may replace + and by
16
arbitrary functions, f( + ) and g( ). This is a symmetry with an innite number
of generators, certainly much larger than the conformal group SO(2, 2) we would expect
by extrapolating the higher dimensional result to two dimensions. This group is in fact
contained in these transformations: the SO(2, 2) transformations turn out to correspond
to functions f and g of the form f(x) = a + bx + cx
2
.
3 String spectra
Consider an observer looking at a string moving through space time. The string will
have a center-of-mass motion, around which it moves around. The motion around the
center-of-mass can be decomposed in normal modes of vibration. When the string in one
of these modes, the observer will view the energy of the mode as the mass of a particle.
There is an innite number of modes of vibration, and when we quantize the theory, each
of these modes will have its own harmonic oscillator wave function.
3.1 Mode expansion
3.1.1 Closed Strings
Now we have to make this intuitive notion precise. Let us rst consider closed strings.
The world sheet of a closed string is a cylinder. On this cylinder the action is, in conformal
gauge,
S =
1
4

_
d
2

a
X

a
X

, (3.1)
where d
2
dd. The fact that the elds X live on a cylinder implies that they satisfy
periodic boundary conditions:
X

(0, ) = X

(2, ) (3.2)
The point = 2 is the standard choice, but one may show that if we would take any
other choice, the spectrum does not depend on it. Any eld satisfying these boundary
condition can be Fourier expanded
X

(, ) =

n=
e
in
f

n
() (3.3)
The classical equation of motion for X is
[
2

]X

(, ) = 0 (3.4)
For the Fourier modes of a classical solution this implies

n
() = n
2
f

n
() . (3.5)
17
The solution is
f

n
() = a

n
e
in
+ b

n
e
in
, n ,= 0 (3.6)
and
f

0
() = p

+ q

. (3.7)
Putting all this together, and introducing a few convenient factors, we may write the
result as
X

(, ) = q

+ i
_

n=0
_
1
n
(

n
e
in(+)
+

n
e
in()
)
_
, (3.8)
Note that X

should be real. This immediately implies

(

n
)

n
(3.9)
3.1.2 Open String boundary conditions
For open strings we choose

the interval as 0 < . Clearly we need a boundary

condition at the two extremes. The possible boundary conditions can be determined by
considering the derivation of the equations of motion, with careful attention to boundary
terms. When varying the integrand of the action (the Lagrangian density),
L(X)
a
X

a
X

, (3.10)
the variation is dened as

X
L L(X + X) L(X)
=
a
(X

)
a
X

+
a
X

a
(X

= 2
a
(X

)
a
X

(3.11)
Hence

X
S =
1
2

_
d
_

0
d
a
(X

)
a
X

(3.12)
which after an integration by parts yields

X
S =
1
2

_
d
_

0
d(X

)
a

a
X

1
2

_
dX

=
=0
(3.13)
Requiring the rst term to vanish for arbitrary X

leads to the equations of motion,

a
X

. The second term must also vanish, and this happens most easily by choosing

= 0 at the boundaries = 0 and = . This is called a Neumann boundary. A

second possibility would be to assume that X

vanishes at the boundary, so that X

is
constant. Then X

( = 0, ) = c

and X

( = , ) = d

, where c and d are constant

vectors. Clearly this is not Poincare invariant, and in particular translation invariance is
always broken. This is called a Dirichlet boundary condition. We will come back to it
later, but for now we restrict ourselves to Neumann boundaries.

The size of the interval is irrelevant because of scale invariance: any other choice can be rescaled to .
18
3.1.3 Open String mode expansion
The rest of the discussion is the same as in the closed case, except that the -modes are
slightly dierent now. We nd
X

(, ) = q

+ 2

+ i

n=0
1
n
(

n
e
in
cos n) , (3.14)
3.1.4 Open versus closed
This shows why we dened the open string with a range 0 < and we used twice this
range for the closed string. The advantage of doing it that way is that the modes come
out the same way, as a sum of terms of the form exp (in( )). This is convenient,
but not essential: with any other choice for the range our nal result for the spectrum
would be the same. This is a consequence of scale invariance of the action: any range can
always be scaled back to the one above.
The dependence on

is also dierent in the open and the closed case. The reason for
that will become clear in the next section: it ensures that the quantized modes will have
the same commutation relations in both cases.
Finally, note that the open string mode expansion can be written in the form (2.28),
but with X
L
= X
R
. Intuitively the reason for this is that in closed strings left and right-
moving modes are decoupled: a mode that moves to the left will continue to do so forever.
In the open string theory such a mode will bounce o the boundary and start moving to
the right.
3.2 Quantization
The system under consideration can be treated using the methods explained in Appendix
A. The Lagrangian is
L =
1
4

_

0
d
a
X

a
X

, (3.15)
so that S =
_
dL. Here = 1 for open strings and = 2 for closed strings.
Now it is easy to derive the canonical momentum of X

=
1
2

(3.16)
From the Poisson brackets of the classical theory we infer the correct form of the commu-
tator in the quantum theory:
[X

(, ),

, )] = i

)
[X

(, ), X

, )] = 0
[

(, ),

, )] = 0 .
(3.17)
19
These relations can be expressed in terms of modes. The result is

k
,

l
] = [

k
,

l
] = k

k+l,0
[

k
,

l
] = 0 (3.18)
and
[q

, p

] = i

(3.19)
We conclude that q

and p

can be interpreted as the center-of-mass coordinate and

momentum, as is already clear from the classical expression.
The -operators look a lot like harmonic oscillator operators. The quantum version
of (3.20) is
(

n
)

n
(3.20)
If we now dene
a

n
=
1

n
, (a

n
)

=
1

n
(3.21)
we obtain a harmonic oscillator algebra for each value of and n.
3.3 Negative norm states
Note however the following. A standard harmonic oscillator commutation relation has
the form
[a, a

] = 1 (3.22)
One denes the vacuum by the requirement a [0) = 0. If one computes the norm of the
state a

[0) one nds

[[a

[0) [[ = 0[ aa

[0) = 0[ [a, a

] [0) = 1 . (3.23)
However, because of the appearance of

one of the oscillators will have the commutation

relation [a, a

] = 1. This leads to states of negative norm. States with negative norm

often occur in quantum eld theory in intermediate stages of a calculation. They are
called ghosts. In sensible theories they should cancel. In particular, there should not exist
scattering processes of physical particles in which ghosts are produced. At this stage, the
string spectrum seems to contains such quantum states. If this were the end of the story,
the theory would be sick.

Note that the commutators (3.17) are at some world-sheet time =

0
. Hence also the modes
k
are
the coecients in the expansion of X at
0
. The classical time evolution of these modes as given (3.8) is
only used to distinguish left- and right-moving modes at
0
, but of course the time evolution is governed
by quantum mechanics, and not by the classical evolution (3.8).
20
3.4 Constraints
However, it is not the end of the story. So far we have completely ignored the second
equation of motion, T
ab
= 0. Written in terms of components this tensor becomes
T
00
= T
11
=
1
2
(

X
2
+ X
2
) (3.24)
T
10
= T
01
=

X X

(3.25)
where the dot and the prime indicate and derivatives, and Lorentz-indices are implic-
itly contracted.
We may compare this with the Hamiltonian of our system
H =
_

0
d(

X ) L =
1
4

_

0
(

X
2
+ X
2
) (3.26)
Hence we see that, up to normalization, H is just the integral over T
00
, which has to
vanish. The Hamiltonian is the generator of time translations. The integral over T
01
is
proportional to P, the generator of translations in the direction. Obviously, it also has
to vanish. But that is still not all, because T
ab
has to vanish without integration. Put
dierently, there is an innite number of modes of T
ab
that has to vanish.
3.5 Mode expansion of the constraints
To write down these modes it is convenient to go rst to dierent coordinates,

= (3.27)
Then the derivatives are

=
1
2
(

) . (3.28)
In these coordinates the components of the energy-momentum tensor are
T
++
=
1
2
(T
00
+ T
01
) =
+
X
+
X
T

=
1
2
(T
00
T
01
) =

X
(3.29)
Now we dene the modes as follows (for closed strings)
L
n
=
1
2

_
2
0
de
in
T
++

L
n
=
1
2

_
2
0
de
in
T

(3.30)
Here the bar on the second L should not be confused with complex conjugation; it only
serves to distinguish left- from right-moving modes. Under complex conjugation we have
in fact
L

n
= L
n
,

L

n
=

L
n
(3.31)
21
In the quantum theory the becomes a . These mode expansions are done at = 0.
With a little eort, we can substitute the mode expansions for X

and express the Ls

in terms of the modes of X. The result is
L
n
=
1
2

nm

m
,

L
n
=
1
2

m

nm

m
(3.32)
here we have dened

0
=

0
=
_
1
2

(3.33)
The advantage of the

coordinates is that the right- and leftmoving operators

n
and

n
are almost completely separated, with the exception of the zero mode p

, which has
no left or right orientation.
For open strings we nd in a similar way
L
n
=
1
2

_

0
d(e
in
T

+ e
in
T
++
)
=
1
2

nm

m
(3.34)
with

0
=

.
3.6 The Virasoro constraints
If we ignore the problem of ghost states for the moment, then the full Hilbert space of
states would be obtained as follows. First we choose a vacuum that is annihilated by all
the annihilation operators,

n
[0) =

n
[0) , n > 0 (3.35)
and that furthermore has zero momentum
p

[0) = 0 (3.36)
This zero-momentum state can be transformed into a ground state with momentum in
the usual way
[k

) = e
iq

k
[0) , p

[k

) = k

[k

) , (3.37)
where of course p

is an operator and k

a momentum vector. Then we obtain all ex-

cited states by action with the negative modes of the operators and . A typical,
unnormalized state the looks like
[n
1
, . . . , n

,
1
, . . . ,

, m
1
, . . . , m

,
1
, . . . ,

, k

) =

1
n
1
. . .

1
m
1
. . .

m
[k

) .
For open strings the space of states looks similar, except that one doesnt have the
operators.
22
Now we have to see what the eect is of the energy-momentum constraints. We have
just introduced the modes of the energy-momentum tensor. These modes, the operators
L
n
and

L
n
, are called the Virasoro operators. Since the energy-momentum tensor T
ab
classically has to vanish, one might come to the conclusion that in the quantum theory
we should impose the constraint
L
n
[phys) =

L
n
[phys) = 0 (3.38)
This would restrict the number of quantum states, and one might hope that the ghost
states disappear.
3.7 Operator ordering
However, there are two problems with this idea. First of all the Virasoro operators are not
yet well-dened. This is because in the quantum theory we replace the classical quantities
and by non-commuting operators. The ordering of the classical objects is irrelevant,
but the ordering of the quantum operators makes a dierence.
Since the only non-commuting objects are
k
and
k
, this ordering problem only
aects the operator L
0
.

The solution to this problem is to impose an ordering by hand,

and to dene L
0
as
L
0

1
2
:

m
: (3.39)
where the colons, :, indicate normal ordering. This means that all creation operators are
written to the left of all the annihilation operators. Consequently the vacuum expectation
value of any non-trivial normal-ordered operator O is zero, 0[ : O: [0)=0, because either
the state [0) is destroyed by the annihilation operators acting to the right, or 0[ is
destroyed by the creation operators action to the left. Only when an operator contains
no operators
k
, k ,= 0 at all it may have a non-trivial vacuum expectation value. The
normal-ordered Virasoro zero-mode operator may also be written as
L
0
=
1
2

2
0
+

m>0

m
. (3.40)
It is clear that changing the order of two oscillators changes L
0
by a constant. The
fact that we have imposed an ordering does not mean we can ignore that constant. It only
means that there is an unknown constant in the relation between L
0
and the Hamiltonian:
H = (L
0
a) + (

L
0
a) closed strings
H = (L
0
a) open strings .
(3.41)
This also aects the condition we must impose on the spectrum:
(L
0
a) [phys) = (

L
0
a) [phys) = 0 (3.42)

Here and in the following we will omit the analogous expressions for the operators that appear in
the closed string theory.
23
3.8 Commutators of constraints
But that is not the only problem. For example, consider the state
L
2
[0) =
1
2

2+p

p
[0) (3.43)
In the innite sum, the only term that contributes is the one with p = 1:
L
2
[0) =
1
2

1

1
[0) (3.44)
If we try to impose (3.38) we would like the norm of this state to vanish. But that is
clearly nonsense:
[[L
2
[0) [[ = 0[
1
2

1,
1
2

1,
[0) =
1
2

=
1
2
D ,= 0 (3.45)
This problem is a consequence of the fact that the L
n
s have non-trivial commutation
relations. It turns out to be an interesting, but rather cumbersome excercise to compute
[L
n
, L
m
]. One may rst do this classically, using Poisson brackets instead of commutators
for the modes
n
. Although the Poisson brackets have a non-trivial right hand side,

k
,

PB
=

k
,

PB
= ik

k+l,0

k
,

PB
= 0 , (3.46)
the modes commute with each other, and there is no ordering problem. It is rather easy
to compute
L
m
, L
n

PB
= i(mn)L
m+n
(3.47)
The calculation in the quantum case is essentially identical as long as we neglect ordering.
Hence we get
[L
m
, L
n
] = (mn)L
m+n
+ re-ordering contributions (3.48)
The reordering contributions can be computed by carefully commuting the oscillators that
are in the wrong place, but there is a more clever way.
3.9 Computation of the central charge
Since the reordering only aects L
0
, it follows that the correct form of the commutation
relation can only be
[L
m
, L
n
] = (mn)L
m+n
+ A(m)
m+n
, (3.49)
where A(m) is a c-number, i.e. it commutes with the L
n
s (such a contribution is called
a central charge of an algebra). Obviously A(m) = A(m). Furthermore we know
already that A(2) is non-zero:
0[ L
2
L
2
[0) =
1
2
D
0[ L
2
L
2
[0) = 0
_
0[ [L
2
, L
2
] [0) =
1
2
D , (3.50)

Poisson brackets are denoted here as . . .

PB
.
24
using 0[ L
0
[0) = 0. Hence A(2) =
1
2
D. In a similar way one nds that A(0) = A(1) = 0.
An elementary property of commutators is the Jacobi-identity
[A, [B, C]] + [B, [C, A]] + [C, [A, B]] = 0 (3.51)
which should hold independent of what A, B and C are. In the case it yields, for A, B, C
= L
1
, L
1n
and L
n
(2n + 1)A(1) (n + 2)A(n) (1 n)A(n + 1) = 0 (3.52)
This is a recursion relation for the coecients. Since A(1) = 0 the equation becomes
A(n + 1) =
_
n + 2
n 1
_
A(n) (3.53)
The solution is
A(n) =
1
12
cn(n + 1)(n 1) , (3.54)
and c can be xed by means of the special case n = 2: this gives c = D. The nal result
is a famous commutation relation,
[L
m
, L
n
] = (mn)L
m+n
+
1
12
c(m
3
m)
m+n
, (3.55)
which is called the Virasoro algebra.
3.10 The Virasoro algebra
The Virasoro algebra is an example of a Lie-algebra, just as the familiar angular momen-
tum algebra [J
i
, J
j
] = i
ijk
J
k
. Strictly speaking the Virasoro algebra given above is not
really a Lie-algebra. One of the properties of a Lie-algebra is a product that closes
and satises the Jacobi identies. For physicists, that product is usually a commutator.
However, if we regard the L
n
s as the elements, then the product does not close unless
c = 0. The solution is to add one extra element to the algebra, C, which commutes with
all the other elements and whose eigenvalues are c. An element that commutes with all
others but appears on the right hand side of commutators is called a central charge.
The Virasoro algebra has appeared here in a theory of D free bosons, with c = D. It
appears in fact in all two-dimensional eld theories that have conformal invariance, but
with dierent values of c. Field theories with conformal invariance are called conformal
eld theories.
There is another way to look at c. Remember that in the classical theory c = 0, but
that c ,= 0 is generated by quantum eects. The classical form of the algebra is changed by
quantum eects. This means that the classical symmetry is broken. When this happens,
we speak of an anomaly. The coecient c is often called the conformal anomaly.
25
3.11 Imposing the Virasoro constraints
In any case, the constant c ,= 0 in the Virasoro algebra tells us once again that it is
inconsistent to require L
n
[phys) = 0 for all n. Note however that for classical/quantum
correspondence we may accept a weaker condition, namely
phys

[ L
n
[phys) = 0 (n ,= 0) . (3.56)
For this to be satised it is sucient to demand that
L
n
[phys) = 0 for n > 0 , (3.57)
plus the reality condition L

n
= L
n
. In addition there is the zero-mode condition (3.42).
3.12 The mass shell condition
3.12.1 Closed strings
Let us rst consider (3.42). For closed strings, substituting

0
=
_
1
2

, this yields
(
1
4

p
2
+

m>0

m

m
a) [phys) = 0 (3.58)
plus the same condition with bars on all relevant quantities. These two conditions can be
written as
p
2
= M
2
L
= M
2
R
, (3.59)
with
M
2
L
=
4

m>0

m

m
a
_
(3.60)
and
M
2
R
=
4

m>0

m

m
a
_
(3.61)
The condition p
2
= M
2
is of course the mass-shell condition for a relativistic particle
with mass M, and D-momentum (
_

k
2
+ M
2
,

k). Depending on the excited state in

which we nd the string, it behaves as a particle with dierent mass. It is essential for
this interpretation that M
2
> 0. If M
2
< 0, the particle is called a tachyon, which
is a general name for particles exceeding the speed of light. To see this note that for
normal particles the relativistic energy is given by E =
_
p
2
+ M
2
= M, where p is the
3-momentum and = (1 v
2
)

1
2
(with c = 1). We can solve this relation for v, and
we get v
2
= p
2
/( p
2
+ M
2
). This is smaller than 1 for a massive particle, equal to 1 for
a massless particle and larger than 1 for a tachyon, which therefore moves faster than
the speed of light. In eld theory the presence of tachyons in the spectrum indicates an
26
instability of the vacuum. The standard eld theory description of a (free) massive scalar
particle involves a scalar eld (x) with a Hamiltonian
H =
_
d
3
x
_
(
d
dt
)
2
+ ()
2
+ M
2

2
_
(3.62)
If M
2
> 0 the conguration of minimal energy is = 0, but if M
2
< 0 the energy is not
bounded from below, and in any case the standard vacuum (corresponding to = 0) is
not a minimum. It is clear that the occurrence of tachyons in the spectrum is not a good
thing.
3.12.2 Open strings
For open strings a similar analysis gives p
2
= M
2
with
M
2
=
1

m>0

m

m
a
_
(3.63)
In principle a could be dierent than in the closed case, but we will see that it is in
fact the same. The factor of 4 in the overall factor in comparison to closed strings may
look like it was the result of some conventions we made in the past but it is not. For
example, we made a dierent choice for the range of in both cases, but because of
scale invariance that should not matter, and indeed it doesnt. We also normalized the
momenta and oscillators in a certain way, but that was simply done in order to get the
correct commutation relations (3.18).
3.13 Unphysical state decoupling
Now we should face the problem of the unphysical, negative norm states generated by

with = 0. There are three main approaches to doing this

Light-cone gauge
Covariant operator method
Covariant path integral method
In the rst method, which will be explained in a moment, one sacrices explicit
Lorentz-invariance (temporarily), but one gets only manifestly physical, positive norm
states. In the end Lorentz invariance is checked. One then discovers that this requires
a = 1 and D = 26.
27
In the second method one continues along the lines of the previous section, and now
imposes the remaining Virasoro constraints. Lorentz invariance is explicit at every stage,
but the absence of unphysical states leads to the requirement a = 1,D = 26.

Absence of such unphysical states in this formalism is guaranteed by the so-called

no-ghost theorem.
The last method requires knowledge of path integrals and gauge xing (the Fadeev-
Popov procedure), as well as BRST-quantization. Nowadays it is considered the most
attractive method, but too much machinery is required to explain it here.
3.14 Light cone gauge
The idea of light cone gauge is to use the residual invariance (2.31). It allows us to choose
a new world sheet time parameter

= f( + ) + g( ). (3.64)
We start by choosing new coordinates in space-time
X
i
, X

=
1

2
(X
0
X
D1
) (3.65)
As we have seen before, the equations of motion tell us that we can write
X

(, ) = X

L
( + ) + X

R
( ) (3.66)
This splitting in left- and right-moving modes is obvious for the oscillator modes, but the
center-of-mass coordinates require a bit more attention. Explicitly we have, for closed
strings
X

L
( + ) =
1
2
q

+
1
2

( + ) + i
_

n=0
_
1
n
(

n
e
in(+)
)
_
X

R
( ) =
1
2
q

+
1
2

( ) + i
_

n=0
_
1
n
(

n
e
in()
)
_
(3.67)
and for open strings
X

L
( + ) =
1
2
q

( + ) + i

n=0
_
1
n
(

n
e
in(+)
)
_
X

R
( ) =
1
2
q

( ) + i

n=0
_
1
n
(

n
e
in()
)
_ (3.68)

This is somewhat oversimplied. In fact, absence of unphysical states for the non-interacting string
leads to a 1, D 26. However, if a ,= 1 and D ,= 26 it is not known how to make the string theory
interacting. If D < 26 it turns out that classical conformal invariance is broken by quantum eects (the
conformal anomaly), and therefore it is inconsistent to go to light-cone gauge. Therefore there is no
contradiction with the results obtained in light-cone gauge.
28
Now choose in (3.64) f = b
L
X

L
+ a
L
, g = b
R
X

R
+ a
R
. Obviously we can do that
for at most one choice for , and we choose = +. Note that the constants b
L
, a
L
and
b
R
, a
R
are allowed as free parameters, because the requirement that f (or g) depend only
on + (or ) does not x them. Now we get

= b
L
X
+
L
+ b
R
X
+
R
+ a
L
+ a
R
(3.69)
In order to combine the rst two terms into X
+
we choose b
L
= b
R
. Finally we x the
values of the parameters in such a way that
X
+
= q
+
+

p
+
, (3.70)
dropping the accent on for convenience. We choose = 1 for closed strings and = 2
for open strings, so that the nal outcome is that X
+
looks exactly like before, except
that all the oscillators have been set to zero.
Note that this almost completely xes conformal invariance, since b
L
= b
R
= 1/

p
+
and a
L
+ a
R
= q
+
/

p
+
. Using (2.31) we nd for

=
1

p
+
_
X
+
L
X
+
R
_
+ a
L
a
R
= + a
L
a
R
+ oscillators ,
(3.71)
in other words

is determined up to a constant shift a

L
a
R
. Such a shift does not
respect the open string range 0 < , but is allowed for closed strings, where we only
have a periodicity condition, +2. Note that the oscillator terms do not contribute
to the boundary conditions.
Now let us look at the constraints

X
2
+ X
2
= 0
X

X = 0
_
(

X X

)
2
= 0 . (3.72)
In light cone coordinates, the inner product of two vectors V

and W

becomes
V

= V
i
W
i
V
+
W

W
+
(3.73)
Hence the constraints become
2(

X

)(

X
+
X
+
) = (

X
i
X
i
)
2
(3.74)
substituting X
+
we get
(

X

) =
1
2

p
+
(

X
i
X
i
)
2
(3.75)
The choice of upper or lower sign gives separate constraints for left- and right-moving
modes. These constraints allow us to express X

into X
i
(up to a constant), so that now
not only X
+
but also X

is eliminated as an independent dynamical degree of freedom.

29
In terms of modes, it allows us to express

n
(and

n
) in terms of a bilinear function
of
i
n
(resp
i
n
). The result is, classically,

n
=
_
2

1
p
+
_
1
2

i
nm

i
m
_
(3.76)
and similarly with bars, in the case of closed strings. This explicitly solves all modes
of the two constraints with n ,= 0. A similar formula is also valid for n = 0, with

0
=

0
=
_
1
2

. However, if n = 0 we have once again an ordering problem in

the quantized theory, which we solve by normal ordering. This introduces an unknown
constant a, the same constant we saw earlier when we only considered L
0
. Furthermore,
in the case of the closed string we do not get separate equations for

0
and

0
, since

0
only appears in

X

but not in X

. Let us rst consider the constraint on

X

, i.e. the
sum of the left- and right constraint. After a little bit of work we nd then
_
1
2

=
1
2
_
2

1
p
+
_
1
2

i
m

i
m
a +
1
2

m

i
m

i
m
a
_
(Closed strings) (3.77)
2
_
1
2

=
_
2

1
2p
+
_
1
2

i
m

i
m
a
_
(Open strings) (3.78)
Bringing all momenta to one side (note that the terms
i
m

i
m
contain the contribution
of the p
i
components of the momenta) we nd for the closed string
p
2
=
1
2
(M
2
L
+ M
2
R
)
M
2
L
=
4

m>0

i
m

i
m
a
_
; M
2
R
=
4

m>0

i
m

i
m
a
_
(3.79)
and for the open string
p
2
= M
2
M
2
=
1

m>0

i
m

i
m
a
_
(3.80)
Did we now solve all the constraints? Clearly we did for all the non-zero modes, but for
the zero-modes we have only considered so far the sum of the left and right constraint, and
we still need to consider the dierence, i.e. X

. For the open string the corresponding

condition is automatically satised, but for the closed string it is non-trivial. It simply
corresponds to (3.76) with n = 0, taking the dierence between the equations for and
, and using

0
=

0
. This directly leads to the requirement M
L
= M
R
.
30
3.15 The spectrum
Let us rst consider open strings. The physical space-time state corresponding to the
vacuum[0) in the two-dimensional theory has a mass M
2
= a(
1

). At the next excitation

level we nd a state

i
1
[0) (3.81)
It has mass M
2
= (1a)(
1

), and there are D2 of them. The index i suggests that this

states wants to be a space-time vector. Covariantly a space-time vector has D components.
Not all these components are physical. If a D-dimensional particle is massive, we can go
to its rest frame. We know then that it has D 1 physical components, corresponding
to its polarizations. If a particle is massless, we can not go to its rest frame, but we
can Lorentz-rotate its momentum to the form (E, k, 0, . . . , 0), with k = E. The physical
components correspond to the rotational degrees of freedom in the last D2 components.
There are thus D 2 physical degrees of freedom.
This situation is well-known in electrodynamics. A photon A

starts with D degrees

of freedom, but one is eliminated because of the equation of motion for A
0
, which contains
no time derivative and hence is just a constraint (
i
E
i
= 0, also known as Gauss law),
whereas another one is eliminated by gauge invariance. The equation of motion for a
vector boson of mass M is

= M
2
A

, (3.82)
which has gauge invariance A

only for M = 0 (whereas the non-dynamical

nature of A
0
holds independent of M).
Turning the argument around, the fact that we have here D2 components strongly
suggests that this state must be a massless vector boson. This means that a must be
equal to 1.
At the next excited level we nd the following states

i
1

j
1
[0) ,
i
2
[0) (3.83)
Note that in the rst set the state with indices (i, j) is the same as the one with indices
(j, i), and we should count it only once! In total there are then
1
2
(D2)(D1)+(D2) =
1
2
(D2)(D + 1) physical components. We expect to nd (at least) a massive symmetric
tensor T

.
The analysis of physical states for higher spin (massive or massless) tensor elds is somewhat more
complicated than for vector bosons, but the idea is the same. The result is that the physical states for
a massless particle in D dimensions are given by a representation of the transverse group SO(D 2),
whereas for a massive particle they are given by an SO(D 1) representation.
Some interesting representations of SO(N) are the singlet (dimension 1), and the vector (dimension N).
Furthermore we will need the rank-2 tensor representation. A general rank-2 tensor can be obtained
by combining two vectors V
i
and W
j
: T
ij
= V
i
W
j
. Clearly there are N
2
components. But these N
2
components can be split into three groups that transform into themselves:
T
ij
= [
1
2
(T
ij
+T
ji

2
N

ij
Tr T)] + [
1
2
(T
ij
T
ji
)] +
1
N
[
ij
Tr T] . (3.84)
31
The three components are called the rank-2 traceless symmetric tensor, the rank-2 anti-symmetric tensor
and the singlet. Since they cannot be split further, they are called irreducible representations, To verify
this splitting, note that a vector transforms as V
i
=

j
O
ij
V
j
, where O is an orthogonal N N matrix,
i.e. O
T
O = OO
T
= 1. The number of components of these tensors is respectively
1
2
N(N + 1) 1 =
1
2
(N + 2)(N 1),
1
2
N(N 1) and 1. Later we will also need another type of representation, the spinor
representation.
In D dimensions such a tensor has
1
2
(D+1)(D2) physical components (the dimension
of the traceless symmetric tensor of SO(D1)), and we see that this precisely uses up all
physical components. Hence there is no inconsistency with the assumption that this state
is massive. We could go on like this, but we would not nd any further inconsistencies.
For the closed string the reasoning is similar. The ground state [0) has a mass M
2
=

4a

. The rst excited state satisfying M

L
= M
R
is
i
1

j
1
[0). Note that there is no
symmetry in i and j in this case, hence there are (D 2)
2
states. One expects therefore
at least a covariant symmetric tensor eld g

and an anti-symmetric tensor eld B

.
One can decompose these into physical components, assuming that they are massive, or
assuming that they are massless. Just as for the vector eld, there are not enough degrees
of freedom to allow massive elds, and therefore they have to be massless. This again
gives a = 1. It turns out that there is just one degree of freedom left over, a scalar .
This follows from the decomposition of the (D 2)
2
states into irreducible SO(D 2)
representations.
What are these massless elds? The symmetric tensor can be identied with the
graviton. The scalar is called the dilaton, and in addition there is the antisymmetric
tensor B

. Experimentally only the rst has been observed (or more precisely, its eects
have been observed in the form of gravity). There is no evidence for the other two. In
four dimensions, the eld B

can be transformed into a pseudo-scalar (a scalar that is

odd under parity), which is often called an axion.
3.16 The critical dimension
The constant a, which was just determined to be 1 can also be determined by carefully re-
ordering the oscillators. For a single harmonic oscillator this gives the zero-point energy.
Usually one starts with a Hamiltonian (we choose the mass equal to 1)
H =
1
2
(p
2
+
2
q
2
) , (3.85)
with [q, p] = i. Then one introduces new operators a =
1

2
(p iq), a

=
1

2
(p + iq).
The Hamiltonian becomes then
H =
1
2
(aa

+ a

a) = [a

a +
1
2
] (3.86)
Hence the energy of the vacuum is
1
2
.
32
One may proceed in the same way here. The mass of a string excitation in light-come
gauge is an eigenvalue of the Hamiltonian
H =
1
4

_
d
D2

i=1
[(

X
i
)
2
+ (X
i
)
2
] (3.87)
Substituting the quantum mode expansion for X
i
without interchanging any operators
one gets (for open strings)
H =
1
2

D2

i=1

i
m

i
m
=
1

(
1
2
D2

i=1

m
:
i
m

i
m
: +
1
2
(D 2)

n=1
n) (3.88)
The sum diverges: it is a sum over the zero-mode energy of an innite number of oscilla-
tors! However, if we dene
(s) =

n=1
n
s
, (3.89)
then the sum converges for Re s > 1. This function is the Riemann zeta-function. It can
be uniquely continued to s = 1, and the result is
(1) =
1
12
(3.90)
Hence we conclude
a =
1
24
(D 2) , (3.91)
and since we have already concluded that a = 1, this can only work if D = 26. This is
called the critical dimension.
3.17 The Lorentz algebra

A better way to arrive at the conclusion that the space-time dimension must be 26 is to
consider closure of the Lorentz algebra. In general, one has a set of generators M

that
must satisfy the commutation relation
[M

, M

] = i

+ i

+ i

(3.92)
One can work out the precise form of these quantities in string theory. The result is
M

= q

n=1
_
1
n
(

n
)
_
(3.93)
(plus an contribution for closed strings). It is not too dicult to check that this does
indeed satisfy the commutation relation (3.92).
In light cone gauge, M

split into components M

ij
, M
i
, M
i+
and M
+
. The
oscillators

n
can be expressed in terms of a quadratic relation in
i
(see (3.76)). This
33
makes M
i
tri-linear in
i
. The computation of the commutation relations of these
operators M
i
with the others is technically of course totally dierent than before, but
the result should be the same.
The hardest one to check is [M
i
, M
j
], which should be zero. However, after some
tedious work one nds
[M
i
, M
j
] =
1
(p
+
)
2

m=1
_
m
_
26 D
12
_
+
1
m
_
D 26
12
+ 2(1 a)
__
(
i
m

j
m

j
m

i
m
)
(3.94)
This vanishes if and only if a = 1 and D = 26.
Note that this also implies that if we choose these values, then the Lorentz algebra
closes on the sets of states. In particular, all the states must then fall into Lorentz
multiplets. For example, in the open string theory we found at the rst excited level a
symmetric tensor plus a vector of SO(D 2). These nicely combined into a traceless
symmetric tensor of SO(D 1). Such miracles have to occur at any level, or else
Lorentz invariance is violated. The closure of the Lorentz algebra guarantees this.
3.18 Unoriented strings
We have up to now discussed open and closed strings, but only oriented ones. The
transformation that ips the orientation is the world-sheet parity transformation

= 2 (closed strings)

= (open strings)
(3.95)
These transformations map the world sheet into itself, but with a sign change of .
Oriented strings are not invariant under this transformation, whereas unoriented strings
by denition are invariant.
When one applies this to the mode expansions for open and closed strings one readily
discovers that their eect is equivalent to the following transformation of the modes

n
(1)
n

n
(open strings)

n
(closed strings) .
(3.96)
We can implement this on the quantum theory by introducing and orientation-reversal
operator with the property

1
= (1)
n

n
(open strings)

1
=

n
(closed strings) .
(3.97)
On the states we then impose the requirement that they be invariant under the action
of . Now everything is xed if we just give a rule for the action of on the vacuum.
The correct choice is [0) = [0) (note that choosing a sign here removes the vacuum
from the two-dimensional spectrum, and hence it removes the tachyon from the space-
time spectrum. Although this looks attractive at rst sight, it leads to inconsistencies).
34
Then at the rst excited level of open strings the states
i
1
[0) is transformed into minus
itself by , and therefore is not in the spectrum. At the rst excited level of the closed
string the states
i
1

j
1
[0) is transformed into the states with i and j interchanged, and
hence only the symmetric part of the tensor survives. Hence we only get a graviton and
a dilaton, but no anti-symmetric tensor.
This can all be summarized as follows
String Theory Massless Spectrum
Closed, Oriented g

, B

,
Closed, Unoriented g

,
Open, Oriented A

Open, Unoriented
One should remember that open strings cannot exist by themselves, but always require the
existence of closed strings ((un)oriented open strings require (un)oriented closed strings).
3.19 Chan-Paton labels

In the case of open strings there is one additional freedom. One may attach labels to
the endpoints of the string. In other words, there can be N distinct kind of endpoints,
labelled by an integer i = 1, . . . N. This goes back to work by Chan and Paton in 1969,
and is related to the fact that string theory was invented to describe hadronic interactions.
Open strings represented mesons, and the labels at the end of the string were associated
with quarks and anti-quarks.
To take this possibility into account when we quantize an open string, we may assign
a separate vacuum state to each open string with the boundary at = 0 labeled by a,
and the boundary at = labeled by b. Hence we add the two labels to the vacuum
state: [0; a, b).
If the open string is oriented the two labels are independent, but consistency requires
them to have the same multiplicity N. Then we nd N
2
distinct vacuum states, and at
the rst excited level we nd N
2
vector bosons. In the free string theory we can only
count them, but when we make the theory interaction we can nd out more. It turns out
that these N
2
vector bosons must be gauge bosons of a U(N) gauge group.
If the string theory is unoriented, we have to remember that the world sheet parity
transformation interchanges the two boundaries, and hence the labels a and b. Then the
invariant vacuum states are the symmetric linear combinations
1

2
([0; a, b) +[0; b, a)) (3.98)
This is a total of
1
2
N(N+1) states. At the rst excited level one has the following invariant
states
1

2
(
i
1
[0; a, b)
i
1
[0; b, a)) (3.99)
35
This is anti-symmetric, so there are
1
2
N(N 1) vector bosons. They turn out to form a
gauge group O(N).
Note that the unoriented open string has gauge bosons now. Previously we were
looking at the special case N = 1, and since O(1) is trivial but U(1) is not, we found no
vector bosons for the unoriented open string.
There is one additional possibility if N is even. Then, instead of dening [0; a, b) =
+[0; b, a) the denition [0; a, b) = [0; b, a). is also allowed. The vacuum state is then
1

2
([0; a, b) [0; b, a)) , (3.100)
which is just a single state for the minimal choice, N = 2 At the rst excited level, the
states are
1

2
(
i
1
[0; a, b) +
i
1
[0; b, a)) , (3.101)
which are
1
2
N(N+1) vector bosons. It turns out that they are gauge bosons of a symplectic
group, Sp(N).
3.20 Discussion
So far we have looked at free strings. That is to say, we have considered world sheets that
are either an innite cylinder or an innite strip. Interactions are taken into account by
allowing world sheets with holes, corresponding to strings splitting and joining, or with
semi-innite tubes attached to the surface (external closed string states) or semi-innite
strips attached to the boundaries (external open string states). Hence what we know so
far gives only limited information, but let us look at it anyway.
The main good feature that all string theories seen so far have in common is the
existence of a graviton g

in the spectrum. To be precise, we see that there is a massless

symmetric tensor, which is a candidate for a graviton. To see that it really is a graviton
we need to look at the interactions.
It is known, however, that there is not much choice in making a massless symmetric
tensor interact in a Lorentz-covariant way. The only possibility is Einsteins theory of
gravity (with possibly higher curvature corrections). Hence either string theory can be
made interacting in a consistent way, and then it must produce general relativity, or it is
inconsistent. The former possibility turns out to be the correct one.
For example, one may compute the three-graviton interaction in string theory and
compare with the analogous computation in general relativity. As expected, the form of
the interaction agrees, and one gets a relation between the coupling constant of gravity
Newtons constant and the only parameter in the string action,

.
This is somewhat misleading since we are in 26 dimensions, and not in four. Newtons
law of gravity in D dimensions is
F = G
N
m
1
m
2
r
D2
(3.102)
36
which gives the famous
1
r
2
force in 4 dimensions. It follows that the dimension of G
N
is
dimension dependent:
[G
N
] = [F][r]
D2
[m]
2
= [m]
1
[l]
D1
[t]
2
(3.103)
If we work with the convention c = = 1, so that [t] = [l] = [m]
1
, we get that
[G
N
] = [l]
D2
.
The parameter dimension of

is determined as follows. X clearly has the dimension

of length, and an action has the same dimension as (so that exp (iS/) makes sense),
which is 1 in our case. The dimension of the world sheet parameters and is irrelevant,
since it cancels. Hence

has the dimension of [l]

2
.
From this analysis we learn that it is inevitable that G
N
is given by
G
N
= (

)
1
2
(D2)
dimensionless constants (3.104)
If we manage to construct strings in four dimensions, we would expect G
N
(

).
Conversely, this implies the

is of order the Planck length, or the fundamental mass

scale 1/

is of order the Planck mass.

We have seen that the masses of all the excited states are multiples of this parameter,
and hence inevitably all excited states have masses that are of order 10
19
GeV. These
particles are therefore unobservable.
The only observable string states are the massless ones (assuming we manage to get
rid of the tachyon). Of course very few massless particles have been seen: the photon, (in-
directly) the graviton, arguably the gluons and perhaps (but probably not) the neutrinos.
However, the standard model has a symmetry breaking mechanism, the Higgs mechanism,
and we only see the broken phase. In the unbroken theory all particles (except the scalars
that become the Higgs boson) are massless.
The idea is now that this unbroken theory corresponds to the massless particles in
string theory. The origin of the Higgs mechanism and the smallness of the scale at which
it occurs remains a mystery. The smallness of the ratio M
W
/M
Planck
10
17
is a mystery
independent of string theory, and at present string theory oers us no solution to this
problem.
If we forget this problem, let us compare the massless string spectrum to what we
observe:
The last column lists the solutions to the mismatch in the rst two columns. This
is also a preview of some of the subjects we still have to discuss. Here is very brief
explanation, to be made more detailed later.
Fermionic strings are string theories which have, in addition to the bosons X

,
also two-dimensional fermions. It can be shown that they always contain space-
time fermions in their spectrum. Hence they can give rise to quarks and leptons.
Fermionic strings have a critical dimension of 10, and they may or may not have
tachyons.
Superstrings are fermionic strings with space-time supersymmetry.
37
Bosonic string spectrum Observed Solution
Tachyon Fermionic string,
Superstrings
Graviton Yes
B

No Low-E mass, axion?

dilaton No Low-E mass
gauge symmetry SU(3) SU(2) U(1) String theory choice
(U(N), O(N), Sp(N))
D = 26 D = 4 Compactication
quarks, leptons Fermionic string
An axion is a pseudo-scalar that was proposed some time ago to solve certain prob-
lems with QCD. The B-particle is a pseudo-scalar, and hence could perhaps play
the role of the axion.
Supersymmetry is a symmetry between fermions and bosons. So far it has not
been observed in nature. Supersymmetry implies absence of tachyons (but is not
necessary for that).
Compactication means that some of the space-time dimensions are rolled up into
a small space, so small that it is unobservable to us.
Moduli are changes in the shape of the compactication space. They lead to massless
scalars in the spectrum.
Low-E mass means that by some mechanism these particles are assumed to get a
mass much below the Planck mass.
4 Compactication
The idea of space-time compactication precedes string theory by many decades. already
in the rst half of the century Kaluza, Klein and also Einstein considered the idea that
space-time is higher-dimensional.
4.1 Space-time compactication for particles
Suppose our world is actually ve-dimensional, with one dimension rolled up. This means
that x
4
is periodic, x
4
x
4
+2R. The other four dimensions remain unconstrained. We
denote the ve dimensions by m = (, 4) with = 0, 1, 2, 3.
Now take a particle in this ve-dimensional world. It satises the mass shell condition
p
2
= p

+ (p
4
)
2
= M
2
(4.1)
38
Now p
4
is not on equal footing with the space-time momenta p
i
. Wave functions must
respect the periodicity:
(x

, x
4
) = (x

, x
4
+ 2R) (4.2)
A free particle wave function,
= e
ip
m
xm
(4.3)
is only periodic if p
4
is quantized:
p
4
=
n
R
, n Z (4.4)
It makes then sense to put (p
4
)
2
on the other side of the equal sign in (4.12). Then
we get
p

= (p
4
)
2
M
2
(4.5)
Hence a four-dimensional observer sees a discrete spectrum of particles of mass-squared
M
2
+ n
2
/R
2
.
4.2 Space-time compactication for strings
We can do the same in string theory. If we do this for closed strings there is an additional
possibility. The string can wrap a couple of times around the compact dimension:
X

(, + 2) = X

(, ) + 2mR

(4.6)
where R

= R
,25
if we compactify the 25
th
dimension. Now = 0, . . . , 24, so that we
are assuming that we are in 25 space-time dimensions. This will turn out to be the critical
dimension. The rule is that the sum of compactied plus uncompactied dimensions must
be 26.
If we impose this periodicity we must rederive the string mode expansion. This is very
easy, and results in
X

(, ) = q

+ mR

+ i
_

n=0
_
1
n
(

n
e
in(+)
+

n
e
in()
)
_
, (4.7)
It is convenient to write this in terms of left and right-moving components
X

L
=
1
2
q

+
_

2
p

L
( + ) + i
_

n=0
_
1
n
(

n
e
in(+)
)
_
, (4.8)
and
X

R
=
1
2
q

+
_

2
p

R
( ) + i
_

n=0
_
1
n
(

n
e
in()
)
_
, (4.9)
with
p

L
=
1

2
_

+
1

mR

_
p

R
=
1

2
_

mR

_ (4.10)
39
4.3 The spectrum
Now we quantize this theory and solve the constraints. This is all nearly analogous to the
uncompactied case. The result is
M
2
= M
2
L
= M
2
R
(4.11)
with
M
2
L
=
4

_
1
2
(p
2
L
)
comp
+

m>0
23

i=1

i
m

i
m
+

m>0

25
m

25
m
a
_
M
2
R
=
4

_
1
2
(p
2
R
)
comp
+

m>0
23

i=1

i
m

i
m
+

m>0

25
m

25
m
a
_
(4.12)
Where, as above, only the compactied components of p
L
and p
R
are added to the mass.
When going to light-cone gauge one now uses the components 0 and 24 to dene X
+
and
X

, and one checks Lorentz invariance only in the uncompactied dimensions (since it
is manifestly violated in the compactied ones). The computation is essentially identi-
cal, and yields the result mentioned above, namely that the total dimension (compacti-
ed+uncompactied) must be 26. The constant a comes out as 1.
The quantization of the momenta in the compactied directions is exactly as for par-
ticles. Hence we get
p
L
=
_

2
_
n
R
+
mR

_
p
R
=
_

2
_
n
R

mR

_
(4.13)
Here p
L
and p
R
refer to the 25
th
components.
The spectrum again contains a tachyon. At the massless level we now have

i
1

i
1
[0, 0) g, B,

i
1

25
1
[0, 0) vector boson

25
1

i
1
[0, 0) vector boson

25
1

25
1
[0, 0) scalar
(4.14)
Here we use the notation [p
L
, p
R
). In addition it may happen that (p
L
)
2
and/or (p
R
)
2
is
precisely equal to 2. In that case there are extra massless particles:

i
1
[0, p
R
) vector boson

25
1
[0, p
R
) scalar

i
1
[p
L
, 0) vector boson

25
1
[p
L
, 0) scalar
[p
L
, p
R
) scalar
(4.15)
40
It turns out that the vector bosons are in fact gauge bosons of some new gauge in-
teraction. This is a second way to obtain gauge symmetry in addition to Chan-Paton
labels.
4.4 T-duality
The spectrum we just found has an interesting symmetry
R

R
n m
(4.16)
Under this transformation p
L
is unchanged and p
R
changes sign, but this has no inuence
on the spectrum.
It turns out that this remains true even for the interacting string theory. This is
remarkable, because it says that strings compactied on a very large circle are equivalent
to strings compactied on a very small circle. Or, in other words, one cannot do any
experiment to determine if the internal compactication manifold is large or small. This
is counter-intuitive, but there is a simple reason why this happens: the momentum and
winding states are interchanged. At small R, the momentum states are very widely
spaced, but it costs very little energy to wind the string around the circle. Therefore
winding states have very little spacing. If we replace R by

/R it is just the other

way around. Now momentum states take over the role of winding states and vice-versa.
Experimentally winding and momentum states cannot be distinguished.
momentum
winding
If you take this to the extreme limit R it leads to the conclusion that at space is
equivalent to a point! It is however not very clear what this would mean.
This phenomenon is know as T-duality. Although we have seen it here in a very special
system, it is seen in all compactications, and is believed to be a fundamental property of
string theory. It seems as if short distances (smaller than the Planck scale) are equivalent
to long distances. In other words, there seems to be a fundamental shortest length scale,
41
the Planck scale. If one probes smaller distances, nothing new is found, but one simply
sees the same phenomena one has seen at larger distances already.
This is also related to the absence of short-distance singularities (ultra-violet diver-
gences) in string loop calculations. What happens is that all integrals are cut o in a
natural way at the Planck scale. Integrating over momenta larger than the Planck mo-
mentum would be double counting: those contributions already are taken into account in
the integral over small momenta.
4.5 The self-dual point
We observed earlier that there is a possibility of having extra massless vector bosons. This
requires p
2
L
= 2, p
2
R
= 0 or vice versa. The condition p
R
= 0 relates m to n. Substituting
this in p
2
L
= 2 leads to the condition R = [n[

. Then we may compute m, and we nd

m =
1
n
. Hence the only possibility is n = m = 1, R =

. Note that the T-dual of

this theory is the same theory. This is called therefore called the self dual point on the
space of compactied string theories. In the selfdual point there are precisely two states of
the form
i
1
[0, p
R
) (namely (n, m) = (1, 1) or (1, 1)) and two of the form
i
1
[p
L
, 0)
(namely (n, m) = (1, 1) or (1, 1)).
One may expect these states to correspond to gauge bosons. Then what is the cor-
responding gauge group? If we are not in the selfdual point there are just two vector
bosons, and then there is not much choice; it can only be U(1) U(1). In the selfdual
point there are six vector bosons, and then the only possibilities are U(1)
6
, SU(2) U(1)
3
or SU(2)
2
. The latter turns out to be the correct one, but to decide that one would have
to study the interacting string.
4.6 Field theory interpretation

The appearance of massless gauge bosons and scalars in the spectrum can also be under-
stood as follows. We have seen that before compactication the spectrum contained a
graviton g

(i.e. a symmetric tensor), an antisymmetric tensor B

and a dilaton as
massless particles. Massive particles in 26 dimensions can be ignored: their momentum
and winding modes will be even more massive. The tachyon can contribute to the massless
spectrum, due to its momentum and winding modes. These are the extra, non-generic
massless states found above.
The presence of the massless particles in the string spectrum implies the existence
of a corresponding quantum eld theory in target space. This is a rather subtle step.
Up to now we have worked with a single string theory, and we have creation operators
that create modes on the string. This is not the same as particle creation in quantum
eld theory. Formally it is the same operation if we look at the string theory as a two-
dimensional quantum eld theory. But that is world sheet physics. In target space (i.e.
in 26 dimensions) a quantum eld theory has similar creation operators, but they create
particles. The operators
i
m
dont do that; they just change the vibrations of the string
moving through target space. The vacuum state [0) is a two-dimensional vacuum state.
42
It corresponds to a string without any vibrations, not to an empty universe. In other
words, what is missing is an operator that create strings from the target space vacuum.
Attempts to construct such operators have been partly successful, and have led to string
eld theory. This is a dicult subject we do not want to go into here.
Usually one bypasses this issue, and one simply postulates that for every particle in
the string spectrum there will be a multi-particle Hilbert space, and that the interactions
that create or destroy such particles can be described by means of a quantum eld theory.
Although it is hard to derive this quantum eld theory directly from string theory,
we can learn enough about it to determine it. First of all one can determine the classical
equation of motion from string theory. This already determines the classical action almost
completely. Secondly one can compute multi-particle scattering amplitudes (S-matrix
elements) directly from string theory, and compare with a putative quantum eld theory.
The presence of a massless spin-2 particle then implies that there must be a corre-
sponding eld theory. This is in fact general relativity. Likewise, a massless spin-1 particle
implies a gauge theory.

These eld theories are usually called low-energy eective eld

theories. The word low-energy refers to the fact that we are ignoring massive string
excitations. It might be possible to include them in the eld theory, but since there are
innitely many that is not practical. Since massive excitations are ignored, any eld the-
ory description must fail once we reach energies as big as the string scale (i.e. the Planck
scale. The word eective means that it is a good substitute for the real thing, which is
string theory. For example, one can have an eective theory of pion interactions, which
is not as fundamental as QCD but can give good results in certain approximations.
Once we have made this step, we can try to interpret what we have found in terms
of the eective low energy eld theory. The relevant eld are the graviton, the anti-
symmetric tensor and the dilaton. These massless elds can be decomposed in terms of
25-dimensional elds. This works as follows:
g

_
g
MN
A
M
A
M

_
(4.17)
B

_
B
MN
B
M
B
M
0
_
(4.18)
Here = 0 . . . 25 and M = 0 . . . 24. We see that the graviton decomposes into the
25-dimensional graviton plus a vector boson and a scalar, whereas B

yields B
MN
plus
another vector. Hence we get two massless vector bosons and a massless scalar , in agree-
ment with what we found. The dilaton in 26 dimensions just yields the 25-dimensional
dilaton. On top of this we get all the momentum and winding excitations of these parti-
cles.
In a Kaluza-Klein compactication one only has the momentum modes of string theory.
They can be seen in the eective eld theory by a mode-expansion in the compactied

The terminology spin-2 and spin-1 is strictly four-dimensional. In general it should be interpreted
as symmetric tensor and vector.
43
coordinates, e.g.
(x
M
, x
25
) =

n
(x
M
) cos(2nx
25
/R) +

n
(x
M
) sin(2nx
25
/R) (4.19)
The elds
n
(x
M
) and
n
(x
M
) (n ,= 0) are 25-dimensional massive scalar elds, and
0
is
a massless eld. The winding modes of string theory however can never be found in this
way. Here we see the limitations of eective eld theory.
4.7 Compactication of more dimensions
The foregoing can easily be generalized to more compactied dimensions. The general-
ization of a circle is a torus. However, rather than trying to imagine a higher dimensional
manifold, one usually prefers an algebraic description.
A circle was specied earlier by means of a single vector R

. The appropriate gener-

alization is to introduce N such vectors (linearly independent, of course) if you want to
compactify N dimensions. The meaning of this set of vectors is that points in space-time
that dier by any such vector are considered identical. If this identication is done in one
direction, space gets rolled up into a cylinder. If we add a second direction, the cylinder
gets rolled up into a torus. Beyond this we cannot even draw the resulting gure anymore,
but it is not hard to describe mathematically.
Obviously, if x

is the same point as (i.e. is identied with) x

+ 2R

1
, but also the
same point as x

+2R

2
, and if this is true for any x

, the it follows that x

is identied
with x

+2R

1
+2R

2
. We see then that the set of vectors R

i
must close under addition.
Such a set of vectors is called a lattice.
A lattice can always be described by choosing a set of basis vectors e
i
, i = 1 . . . N,
such that any vector

on the lattice can be written as

=
N

i=1
n
i
e
i
(4.20)
The allowed momenta must have integral inner product with all these lattice vectors in
order for wave functions to have to correct periodicity. A necessary and sucient condition
for this is
p e
i
Z for all i (4.21)
This is easily seen to imply that the momenta p must themselves lie on a lattice, the
so-called dual lattice

. (This situation also occurs in solid state physics, where such a

lattice is called reciprocal lattice; however string theorists like the word dual).
In this general situation the mass formula stays essentially the same, but p
L
and p
R
44
are higher-dimensional vectors of the form
p
L
=
_

2
( +
1

)
p
R
=
_

2
(
1

)
(4.22)
with

and

. The mass formula is usually written as

1
4

M
2
L
=
1
2
p
2
L
+^ 1
1
4

M
2
R
=
1
2
p
2
R
+

^ 1
M
2
L
= M
2
R
,
(4.23)
where p
L
and p
R
are always the compactied components (for uncompactied ones the left-
right splitting is not very useful anyway), and ^ is the sum over the number operators
of the uncompactied and compactied oscillators, i.e.
^ =
24

i=1

m>0

i
m

i
m
, (4.24)
and analogously for

^.
4.8 Narain lattices

However, there is a more general solution.

Dene a lattice consisting of vectors of the

form ( p
L
, p
R
). We observe that this lattice has the following properties
p
L
p

L
p
R
p

R
Z (4.25)
Lattices with this property are called Lorentzian integral. The adjective Lorentzian
refers to the fact that there is a minus sign: this can be reproduced by introducing a
Lorentzian metric diag (1, . . . , 1, 1, . . . , 1), with N positive and N negative entries.
We then treat ( p
L
, p
R
) as a 2N dimensional vector in a space with this metric. The
lattice vectors also satisfy
p
2
L
p
2
R
2Z (4.26)
Such a lattice is called Lorentzian even (actually this already implies Lorentzian integral,
but not the other way around). Finally we can dene a Lorentzian dual to a lattice: the

The compactication discussed here is the general solution to the condition of modular invariance,
which will be discussed in section 5.
45
Lorentzian dual

contains all vectors that have integral inner product with all vectors
in , where we use the Lorentzian inner product dened above.
One can easily show that the lattice constructed above has the property that =

.
Such a lattice is called a Lorentzian self-dual lattice. We see thus that is a Lorentzian
even self-dual lattice. A compactication on the corresponding torus is called a Narain
compactication. It can be shown (and we will return to this in chapter 5) that any such
compactication yields a consistent string theory.
The point is now that there exist lattices that are Lorentzian even self-dual, but are
not of the form (4.22). The generalization has in fact the following form (here we denote
the components of the vectors with an index I = 1, . . . N; the original 26-dimensional
index is split into (M, I), with the uncompactied space-time index M running from 0
to 26 N 1)
p
I
L
=
_

2
(
I
+
1

(
I
B
IJ

J
)
p
I
R
=
_

2
(
I

(
I
+ B
IJ

J
)) ,
(4.27)
where B
IJ
is an arbitrary anti-symmetric matrix. Upper and lower indices have no special
signicance here.
It is interesting to count parameters. The anti-symmetric matrix B
IJ
has
1
2
N(N 1)
real parameters. The lattice we started with is completely specied by the lengths and
relative angles of the basis vectors. This denes a matrix
g
ij
=

I
e
I
i
e
I
j
(4.28)
This is a real, symmetric matrix, which has
1
2
N(N + 1) parameters.
4.9 Background elds

4.9.1 The metric

The notation used in the previous section suggests already that g
IJ
and B
IJ
have some-
thing to do with the metric g

and the anti-symmetric tensor eld B

. This is indeed
the case. To see this, note that X
I
has periodicities X
I
X
I
+2
I
= X
I
+2e
i
n
i
. We
can take the information about the sizes and shape of the lattice out of X
I
by dening
X
I
=

i
e
I
i

X
i
(4.29)
Then

X
i
has standard, rectangular periodicities

X
i
X
i
+ 2n
i
. If we do this in the
action, the following happens
S[X] =
1
4

_
dd
ab

a
X
I

b
X
J

IJ
=
1
4

_
dd
ab

a

X
i

b

X
j
g
ij
(4.30)
46
Now all the structure is encoded in the matrix g
ij
, and if we vary the parameters the
periodicities of

X
i
do not change at all. We saw (see (2.21)) that we can write down the
string action in such a way that it is valid in an arbitrary curved space. Here we see the
simplest example. In general, the properties of such a space are encoded in the metric
g

(x). In the case under consideration here the metric does not depend on x, but in the
compactied part of space it is a non-trivial constant matrix.
4.9.2 the anti-symmetric tensor
What about B
IJ
? The string action written down so far does not have any room for these
parameters, but it is possible to write down a string action that is valid in the presence
of a non-trivial background B

. This action is
S[X] =
1
4

_
dd
_

ab

a

X
i

b

X
j
g
ij
+
ab

a

X
i

b

X
j
B
ij
_
, (4.31)
where
ab
is the two-dimensional Levi-Civita tensor:
10
=
01
= 1;
00
=
11
= 0.
Because of the anti-symmetry of B we clearly need an anti-symmetric tensor here. We
can write this action back into the old basis, with a diagonal space-time metric. The
advantage of that basis is that it is more convenient for computing the spectrum, and
indeed, that is the basis we used for previous calculations. If we dene
B
ij
=

IJ
e
I
i
e
J
j
B
IJ
(4.32)
we get
S[X] =
1
4

_
dd
_

ab

a
X
I

b
X
J

IJ
+
ab

a
X
I

b
X
J
B
IJ
_
, (4.33)
It may seem that this new term forces us to start from the beginning with the derivation
of the equations of motion and the whole quatization procedure. But (for constant B
IJ
at least) the eect is minimal. The extra term in the equations of motion is proportional
to
a

ab

b
X
I
. But this vanishes since the two derivatives commute.
The reason this term vanishes is that the extra term in the action is a total derivative:

ab

a
X
I

b
X
J
B
IJ
=
a
_

ab
X
I

b
X
J
B
IJ

(4.34)
When the the action is integrated over a surface without boundaries the contribution of
such a term vanishes. Therefore it also does not aect the equations of motion, which
describe variations of the action.
But then one may wonder how it can have any eect at all. The reason is that the
canonical momenta do change. We nd

I
=
1
2

(
IJ

0
+ B
IJ

1
) X
J
(4.35)
It is thus this quantity that satises the canonical quantization rule (3.17).
47
It is not dicult to see that the extra term does not aect the quantization of the
oscillator modes, but there is a change in the commutators of the zero modes q and p.
The eect of the extra term is, using (4.7), that p
I
is replaced by
I
= p
I
+
1

B
IJ

J
(

is the higher dimensional generalization of mR

in (4.7).
The expression for the mass spectrum is still (4.23), with p
I
L
expressed in terms p
I
and
I
exactly as before. But the canonical momentum is
I
, not p
I
, and hence the
quantization condition is that the eigenvalues of the operator
I
must lie on the dual
lattice. Then it is more useful to express p
I
L
and p
I
R
in terms of the canonical momenta.
This gives
p
I
L
=
1

2
_

p
I
+
1

I
_
=
1

2
_

I
+
1

(
I
B
IJ

J
)
_
p
I
R
=
1

2
_

p
I

I
_
=
1

2
_

(
I
+ B
IJ

J
)
_ (4.36)
If we now put the eigenvalues of on the dual lattice we do indeed recover (4.27).
4.10 Moduli

Combining what we saw in section 3.6 and in the previous one, we arrive now at the
following picture. In a compactication to D = 26 N dimensions, the 26-dimensional
metric decomposes as follows
g

_
g
MN
A
I
M
A
I
M

IJ
_
(4.37)
B

_
B
MN
B
I
M
B
I
M

IJ
_
(4.38)
Here A
I
M
and B
I
M
are two sets of N vector bosons each, and
IJ
and
IJ
are two sets
of
1
2
N(N + 1) and
1
2
N(N 1) scalars. All these quantities are elds in D = 26 N
dimensions, i.e. they are functions of the coordinate x
M
. They are not functions of the
compactied coordinates, however. The dependence on the compactied coordinate y
I
has been mode expanded, as in the case N = 1 discussed above.
We observe now that the parameters of the compactication can be viewed as vacuum
expectation values of the massless scalars in D = 26 N dimensions.
In general, a classical eld theory is quantized by expanding around a certain classical
background: =
cl
+
q
. The quantum eld is expanded in modes, and the vacuum
state is chosen in such as way that 0[
q
[0) = 0, so that 0[ [0) =
cl
. Then one calls

cl
the vacuum expectation value of . Also the term background eld is often used for

cl
.
If we want a theory to respect certains symmetries, then we must require that at
least the classical values of the elds,
cl
, respect it. For example translation invariance
requires that
cl
is constant. Lorentz-invariance requires that a vector cannot have a
vacuum expectation value, but a scalar can (whereas for a symmetric tensor the v.e.v.
must be proportional to

).
48
A scalar eld can get a vacuum expectation value without breaking Poincare invari-
ance. This is precisely what is occurring here. We may expand the scalar elds in D
dimensions as follows

IJ
= e
I
i
e
J
j
g
ij
+
IJ
qu

IJ
= B
IJ
+
IJ
qu
(4.39)
We see now that all parameters of the torus compactication can be interpreted as vacuum
expectation values of scalar elds.
This is not limited to torus compactication. It is often expressed by saying that
string theory does not have free parameters. All freedom is due to changing the vacuum.
There is no parameter one xes by hand, like the quark masses or the standard model
couplings.
The dierence is mainly philosophical. If the parameters are related to v.e.vs of elds,
there is a hope that they can be determined dynamically. Maybe one set of parameters
is energetically favorable. If a parameter is in the Lagrangian one starts with, there is no
hope to determine it from the theory itself.
Note that also the string coupling constant is not a parameter, but related to the
v.e.v. of the dilaton. One might object that

is a parameter, but that is not true. It is

a quantity with a dimension, and changing it just changes all dimensionful quantities in
the same way. If we make all the masses in our universe twice as heavy, and also change
Newtons constant accordingly, this can never be observed. Only dimensionless ratios are
true parameters.
The continuous parameters that relate dierent compactied strings are often called
moduli. The name comes from mathematics, where it is used for parameters that describe
deformations of a manifold. The space formed by the moduli is called moduli space. For
circle compactications in eld theory it is just a half-line R 0, correponding to the
value of the radius R. In string theory it gets a more interesting structure due to T-
duality, which identies two parts of the half-line. But even if one compacties just one
dimension in string theory there is already more structure.
4.11 Vacuum selection

What determines the vacuum? In scalar eld theories one often has a potential, i.e.
L =
1
2
_

(x, t)

(x, t) m
2
(x, t)
2

V () . (4.40)
For constant elds only the mass term and the potential are relevant, and the Hamil-
tonian density for constant elds is
H =
1
2
_
m
2
(x, t)
2

+ V () (4.41)
The Hamiltonian is the space-integral of this quantity. Since for constant the integrand
is independent of x, the integral is innite or at best zero, but if we cut o the innity
49
by putting the system in a box, we can compare dierent choice of the v.e.v. of by
comparing the resulting energy. Presumably the true vacuum is then the one of minimal
energy.
For a typical potential the equation of motion only allow a nite number of constant
solutions. Those solutions are simply the minimum of potential plus mass-term.
The situation we encounter in string theory diers from this picture in various ways.
First of all there is a continuum of solutions, not just a nite number. This means that
V is constant (furthermore m = 0, because we consider the massless scalars and ).
But then all solutions must be equivalent. [We didnt really show that the equations of
motion are satised for all radii R, but this can be seen in various ways. In the low-energy
eld theory limit, the equations of motion for are just Einsteins equations, and any at
space satises those equations. A torus is a at space.]
However, there is a problem with this point of view. We have a tachyon T, which
is a scalar with m
2
< 0. This means that the point < T >= 0 is an instable point in
the tachyon potential. If we knew the full potential V (T) we could try to determine the
correct minimum, but we do not know it. The tachyon couples to all other particles, and
in its presence it makes no sense to try to discuss the minima of the moduli potentials.
This will all be dierent in superstrings. There the moduli are in fact at directions in
the potential (and this can be shown also in the presence of interactions), and there is no
tachyon. Hence we get an enormous vacuum degeneracy: there is a huge set of possible
string compactications, and no principle that selects one over any other. The only hope
is then that some non-pertubative eect will lift the degeneracy.
4.12 Orbifolds

There are many other compactications one might consider. A very popular one is the
orbifold. We will just sketch the procedure for the one-dimensional case. We start with
the compactication on a circle with radius R, and the declare that all points related by
a reection are identied. This imposes rst of all a condition on the closed string states.
Only string states that are invariant under X X can live in such a world. It is easy
to see that this eliminates, for instance, all states created by an odd number of oscillators.
In addition one now has to allow closed strings that begin and end at identied points.
This gives new states in the Hilbert space. This part of the Hilbert space is called the
twisted sector.
50
Identification
plane
Twisted strings
Since this can be done for any circle radius, one obtains a second innite half-line
of solutions, with again a T-duality relating small R to large R. For one compactied
dimension (relevant for 25-dimensional strings) the full moduli space is known, and there
are two surprises. They can be seen in the following picture (here N = R
_
2

).
Circle
Orbifold
Isolated
theories
(self-dual
point)
The rst is that the orbifold line and the circle line meet each other. This means that
an orbifold theory (in fact the orbifold of the self-dual circle theory) is in fact equivalent to
51
a circle theory at a dierent radius. This is a non-trivial fact that can only be seen in string
theory (indeed, orbifolds are singular as manifolds, and a eld theory compactication on
such a manifold is not even well-dened). It is again a kind of duality: two seemingly
dierent theories are in fact equivalent. The second surprise is the existence of three
isolated theories that have no moduli at all. These were discovered by P. Ginsparg.
5 Fermions on the world sheet
5.1 Generalizations of the bosonic string
When we started with the bosonic string, we had a nice intuitive action, which was just
the surface area of the world sheet. This Nambu-Goto action was then shown to be
classically equivalent to a dierent action, which was the action of D free bosons. We saw
that this action had Poincare invariance in target space, and reparametrization invariance
and Weyl invariance on the world sheet.
The basic idea underlying all other string construction methods is to modify this action
to some other two-dimensional action, but in such a way that the essential symmetries are
preserved. Poincare invariance is clearly essential, but only in the D dimensions we are
interested in (which ultimately will be D = 4). Weyl invariance and reparametrization
invariance (and conformal invariance, a subset of these symmetries) clearly played an
important role in the discussion of the foregoing sections, although the lightcone gauge
derivation of the spectrum does not really show very clearly that these symmetries are
essential. But they are, in the sense that all important results and in particular the
appearance of gravity remain true provide we make sure that these symmetries are
present in the theory.
Often one does not write explicitly reparametrization invariant theories but instead one
works directly in conformal gauge, and demands that the theory has conformal invariance.
An rather trivial example is the compactied string. The world sheet action is

S[X, ] =
1
4

_
dd
ab
_

a
X

b
X

+
a
X
I

b
X
I

(5.1)
Since X
I
lives on a torus and X
M
does not, 26-dimensional Poincare invariance is broken
to D-dimensional Poincare invariance. But the theory still has conformal invariance (and
it can be written in a reparametrization invariant way be re-introducing a world-sheet
metric

).
In this example the compactied part of the theory (often called the internal theory)
is still written in terms of free bosons X
I
, but this is not necessary. Just as in four
dimensions, there are other eld theories than those made out of free bosons. In general
any two-dimensional eld theory with conformal invariance is acceptable. This kind of
theory is called a conformal eld theory, usually abbreviated as CFT. Such theories exist

From now on we use indices and for D-dimensional space-time; There will be no need anymore for
an index that covers compactied and uncompactied components simultaneously.
52
also in more than two dimensions, but in particular in two dimensions a lot is known, and
many such theories can be constructed. Hence in general one has
S[X, ] =
1
4

_
dd
_

ab

a
X

b
X

+ some conformal eld theory

(5.2)
5.2 The role of the conformal anomaly

There is one important restriction on this internal CFT, which is the generalization of
the restriction to 26 dimensions. To each conformal eld theory belongs a number, the
conformal anomaly. It can be computed by means of a two-dimensional one-loop quantum
correction, which will not be explained here.
Each conformal eld theory has an energy-momentum tensor that can be expanded in
modes L
n
and

L
n
, just as we did for the bosonic string. Classically these objects satisfy
L
m
, L
n

PB
= i(mn)L
m+n
. (5.3)
It can be shown that this relation is equivalent to having classical conformal invariance.
One can work out the commutator of two such operators in the quantum theory, and in
general the result is
[L
m
, L
n
] = (mn)L
m+n
+
c
12
(m
3
m)
m+n
. (5.4)
The rst term is what one expects on the basis of classical/quantum correspondence. We
have seen before that if one allows an extra term in the algebra, the central charge term,
it must have the form shown in (5.4).

The extra term has no classical correspondence,

and can be interpreted as a violation of the classical symmetry by quantum eects. Such
an eect is generally called an anomaly. Since the symmetry aected by it is the con-
formal symmetry, c is often called the conformal anomaly. It may also be thought of
as counting the number of degrees of freedom of a theory (although it is not necessarily
integer). Indeed, for a free boson theory with N bosons the conformal anomaly is N. The
consistency requirement is now that the total conformal anomaly must be 26, i.e.
D + c
int
= 26 , (5.5)
where c
int
is the conformal anomaly of the internal CFT.
The best way to understand this, although this falls outside the scope of these lectures, is
as follows. In a path integral quantization of string theory one has to perform a gauge xing
procedure. This introduces extra elds, the so-called Fadeev-Popov ghosts. It turns out that
these elds also contribute to the conformal anomaly, and their contribution is 26. Hence if
we get +26 from the other elds in the theory, the space-time coordinate bosons X

and the
internal CFT, then the total conformal anomaly is zero. Then exact conformal invariance holds
for the quantum theory. This is the condition one needs.

More accurately: one requires the vacuum to satisfy L

n
[0) = 0 for n 1. Then in particular
L
1
[0) = 0 so that A(1) = 0.
53
Note that it is not very clear that theories constructed in this way can be thought of
as space-time compactications of a 26-dimensional theory. In general, we do not have
bosons X
I
as remnants of 26-dimensional bosons. In many cases

the resulting theories

can be viewed as compactications on non-at manifolds (a torus is at), but this is not
essential. The nice world-sheet surface interpretation of the string action will also be more
and more lost.
5.3 Internal free fermions
In the spirit of the foregoing discussion one may, as a rst attempt to build a new theory,
consider fermionic theories. We do this directly in conformal gauge (coupling free fermions
to non-trivial metric
ab
is possible, but it is a technical complication that is not going to
give any additional insight). The action is
S =
1
4

_
dd
_

a
X

a
X

(5.6)
with an implicit summation over A. Here
a
are the two-dimensional Dirac matrices, the
equivalent of

in four dimensions. These matrices satisfy

a
,
b
= 2
ab
1 (5.7)
In D = 2n dimensions, the Dirac matrices are 2
n
2
n
matrices. Hence in two dimensions
we get 2 2 matrices, and it is easy to show that the following set satises the anti-
commutation relation

0
=
_
0 i
i 0
_

1
=
_
0 i
i 0
_
(5.8)
Analogous to four dimensions,

0
. (5.9)
The representation for the Dirac matrices chosen here is the so-called Majorana rep-
resentation. In this representation the Dirac operator i
a

a
is real, and hence we may
consistently impose the condition that
A
is real. This is called a Majorana fermion.
Since it is real, we have in fact

=
T

0
. (5.10)
The Dirac equation is
i
a

A
= 0 (5.11)
Writing this in matrix notation we get
_
0
0

1
0
__

A
+
_
= 0 (5.12)

In fact this has not been studied extensively for bosonic strings, but the remark is based on experience
with fermionic strings.
54
This leads to two decoupled equations
(
0

1
)
A
+
= 0 , (
0
+
1
)
A

= 0 (5.13)
which are satised if
A
+
depends only on + and
A

depends only on . Hence

just like the bosons, the free fermions split into left and right-moving components.
5.4 Quantization
The quantization of the fermionic action works similar to that of bosons, the main dif-
ference being that instead of commutators we get anti-commutators. The canonical mo-
mentum of a fermion is
(
A
k
) =
i
2

A
k
(5.14)
The canonical commutation relation yields
_

A
k
(),
B
l
(

)
_
= 2

AB

kl
(

) . (5.15)
There is a perhaps unexpected factor of two in the canonical momentum. This is related to the fact that
and

are actually the same variable. The standard lagrangian density for complex (Dirac) fermions
is L =

,, and and

are treated as independent variables. This leads to a quantization condition

, = i(

). The standard lagrangian for Majorana fermions diers from that of Dirac fermions by
an additional factor 1/2, and and

are treated as related variables. Despite these dierences, one gets
same quantization condition (string actions dier from the standard eld theory actions by a factor
1/2

, which appears in the quantization condition in the obvious way). A proper treatment of the
quantization of free fermions requires the Dirac constraint formalism, which produces the correct factor
of two; see e.g. [3] for details. Furthermore there are several sign choices to be made in the quantization
procedure because fermionic variables must be treated as anti-commuting rather than commuting, even
classically. These are just subtleties in the quantization of free fermions in any dimension, and that have
nothing to do with string theory. Therefore this will not be discussed in more detail here.
Now we want to expand the fermionic elds in modes. To do this we need to know
their boundary conditions. For closed strings the possible boundary conditions follow
from the requirement that all physical quantities are periodic under +2. Physical
quantities for fermions are always bilinears, and hence the fermions themselves can be
either periodic or anti-periodic. This then leads to two distinct kinds of mode expansions
Periodic fermions

(, ) =

nZ

d
A
n
e
in()

A
+
(, ) =

nZ
d
A
n
e
in(+)
(5.16)
55
Anti-periodic fermions

(, ) =

rZ+
1
2

b
A
r
e
ir()

A
+
(, ) =

rZ+
1
2
b
A
r
e
ir(+)
(5.17)
For historical reasons such fermions are called Ramond fermions and Neveu-Schwarz
fermions respectively. The notation b and d for their modes is also traditional. As usual,
the bar distinguishes left- and rightmoving modes, and is not meant to indicate complex
conjugation.
It is now straightforward to substitute these mode expansions in the canonical com-
mutation relation(5.15) and obtain the commutation relations for the modes:
_
b
A
r
, b
B
s
_
=
AB

r+s
_
d
A
n
, d
B
m
_
=
AB

n+m
(5.18)
and similarly for the barred components. Furthermore barred and unbarred components
commute (or anti-commute, if one prefers).
For open strings the discussion is slightly dierent. As we have seen for the bosonic
string, the boundary condition arise from the surface term that one gets in a careful
derivation of the equations of motion:

AS() =
i
4

_
dd
_
(

A
)
a

A
+ (

A
)
a

=
i
4

_
dd
_
2(

A
)
a

A
+
a
_
(

A
)
a

A
_
(5.19)
If we require this variation to vanish for general variations
A
, then the rst term requires
the equations of motion to be satised, i.e.
a

A
= 0, while the second term is a
derivative, so it only contributes at the boundary. Since there are boundaries only in the
direction, we get the requirement

A
= (
A
)
T

A
= 0 (5.20)
at the boundary. This yields

A
+

A
+

= 0 (5.21)
Note that we may consider variations for any value of A independently, so that this is
a condition for every index A separately. It can be satised by relating
+
and

at the boundary (so that also their variations are related in the same way), i.e.
A
+
=

,
A
+
=
A

. This would appear to allow in total four possibilities, namely the

choices at = 0 and = . But actually there are just two, since the overall relative
sign between the + and components of is purely conventional. Hence we may x
56

A
+
(=0) =
A

(=0) and then we have just two choices left. This yields exactly the same
mode expansions is found above, except that one must replace

b by b and

d by d. Then
periodic refers to the boundary condition
A
+
(=) =
A

(=) and anti-periodic to

A
+
(=) =
A

(=).
The quantization condition then comes out exactly the same as in the closed case.
5.5 The energy-momentum tensor
The energy-momentum tensor was previously dened as
T
ab
(, )
4

ab
S[] (5.22)
where S is an action depending on the two-dimensional metric and other elds. This
denition is only usable if we have the action in a form where is explicit. In the
case of the bosonic string we started with the action in such a form, but then we went
to conformal gauge, and the explicit dependence on disappeared. In the case of free
fermions we started directly in conformal gauge. An action involving (i.e. an action
that is reparametrization or general coordinate invariant) can be written down but for
fermions this leads to some extra complications.
5.6 Noether currents
There is a second method for obtaining the energy momentum tensor, the Noether method.
In general this works as follows. Suppose we have an action that is invariant under some
(innitesimal) symmetry:
S[] = S[ +
S
] (5.23)
where stands for the set of elds, is a constant parameter, and
S
is a variation of
that corresponds to some symmetry. For example translation invariance corresponds to

S
= n
a

a
, where n
a
is the direction of the translation. Now suppose we make depend
on the space-time point x (we work here in arbitrary dimensions, but we will apply the
results to elds on a string world-sheet). Then the action need not be invariant, but since
it must be invariant when is constant, it must be true that
S =
_
d
D
xJ

(x) , (5.24)
for some J

(there may be higher derivatives as well, but by partial integration one can
always get this form). On the other hand, when is a classical solution, any variation of
the action around it should vanish, and this should in particular be true for the variation
(x). Hence
_
d
D
xJ

(x) = 0 (5.25)
57
or, after integrating by parts
_
d
D
x(x)

= 0 . (5.26)
This must be true for arbitrary functions (x), and that implies that the current J

is
conserved,

= 0 (5.27)
Note that this is true only under the assumption that the set of elds is a classical solution.
Therefore we expect that to verify

= 0 one has to use the equations of motion. In

this way one associates with every continuous symmetry a conserved current.
5.7 The Noether current of translation invariance
Let us now apply this principle to translation invariance. It turns out that the current
associated to translation invariance in the direction n

is J

. We will not prove

this in general here, but it can easily veried for the bosonic string action.
Suppose we consider the translation
a

a
+ n
a
. Then X X + n
a

a
X (for
simplicity we leave out the target space index on X). The action then transforms as
follows
S =
1
4

_
d
2

ab

a
X
b
X S
1
2

_
d
2

ab

a
[n
c
(
c
X)]
b
X +O(
2
) (5.28)
Hence
S =
1
2

_
d
2

ab

a
[()n
c
(
c
X)]
b
X
=
1
2

_
d
2

ab
[(
a
())n
c
(
c
X)
b
X + ()n
c
(
a

c
X)
b
X]
(5.29)
The last term can be written as
_
d
2

ab
()n
c
(
a

c
X)
b
X =
_
d
2

ab

c
()n
c
(
a
X)
b
X

_
d
2

ab
()n
c
(
a
X)
c

b
X
=
_
d
2

ab

c
()n
c
(
a
X)
b
X

_
d
2

ab
()n
c
(
b
X)
c

a
X
(5.30)
Hence _
d
2

ab
()n
c
(
a

c
X)
b
X =
1
2
_
d
2

ab

c
()n
c
(
a
X)
b
X (5.31)
58
The variation of S is therefore
S =
1
2

n
c
_
d
2

ab

a
()(
c
X)
b
X
1
2

ab

c
()
a
X
b
X
_
=
1
2

n
c
_
d
2

a
()(
c
X)
a
X
1
2

db

ac

a
()
d
X
b
X
_
=
1
2

n
c
_
d
2

a
()T
ac
,
(5.32)
where in the last step we substituted (2.23). The last line will now be taken as the
denition of the energy-momentum tensor in the general case.
Applying this to the free fermion term in the action we get (dropping the index A for
convenience)
= n
c

c
(5.33)
so that
2

S[] =
1
2
in
c
_
d
2

_
(
c

)
a

a
+

a
(
c
)

(5.34)
If we use the equation of motion,
a

a
= 0, this reduces to

1
2
in
c
_
d
2

a
(

c
) (5.35)
from which we read o
T
ac
() =
1
2
i

c
(5.36)
Note that this tensor is not symmetric, and hence not equal to the energy momentum
tensor one would get from a variation with respect to
ac
. This is due to the fact that in
general a current J

does not follow uniquely from the Noether procedure, but only up
to terms of the form

, where A

is an anti-symmetric tensor. One can indeed add

such a term to T
ac
to symmetrize it (this is called an improvement term; see exercise 5.6
in [6] for more details). In two dimensions the tensor turns out to be both symmetric and
traceless provided one uses the equations of motion. To see the former, note that (with

0
= 1)
T
01
=
1
2
i

=
1
2
i(

+
+

+
)
(5.37)
while
T
10
=
1
2
i

=
1
2
i(

+
)
(5.38)
59
Now use
0

=
1

and
0

+
=
1

+
to show that T
01
= T
10
For the ++ and
components of the tensor we nd
T
++
=
1
2
(T
00
+ T
01
) =
i
4

0
(
0
+
1
) =
i
4

T
(
0
+
1
) (5.39)
and analogously for T

. Obviously
+
appears only in T
++
and

only in T

.
5.8 The mass spectrum
It is now easy to compute the mass spectrum. As we did before we expand Virasoro
generators in modes, dened exactly as before.

Now we get
L
m
=
1
2

i
n

i
m+n
+
1
2

r
(r +
1
2
m)b
A
r
b
A
m+r
(5.40)
in the Neveu-Schwarz case, and
L
m
=
1
2

i
n

i
m+n
+
1
2

n
(n +
1
2
m)d
A
n
d
A
m+n
(5.41)
in the Ramond case. The space-time part has been written in light-cone gauge. As
before, there are ordering ambiguities. They only aect L
0
. In the free boson case we
had argued that for each boson separately they resulted in (L
0
)
cl
(X) = (L
0
)
qu
(X)
1
24
,
where (L
0
)
cl
is the expression derived directly from the action, and (L
0
)
qu
is the normal
ordered expression. For 24 transverse boson X
i
this produced exactly the required shift
a = 1.
5.9 The Neveu-Schwarz ground state
We can do the same for fermions. For a single Neveu-Schwarz fermion the naive classical
contribution to L
0
is
(L
0
)
cl
() =
1
2

r=
rb
r
b
r
(5.42)
Just as in the bosonic case, we dene the vacuum by
b
r
[0) = 0 for r > 0 (5.43)
so that the negative modes create states. Normal ordering is dened as putting all creation
operators on the left of all the annihilation operators. We see then that in the innite
sum all the terms with r < 0 are in the wrong order, and we use the anti-commutator to
re-order them. This yields
rb
r
b
r
= rb
r
b
r
+ r = sb
s
b
s
s , (5.44)

To get the result directly in this form one should use the symmetrized form of the energy-momentum
tensor. Note that the terms proportional to m are actually zero due to the anti-commutators.
60
with s = r. Hence for each positive, half-integer value s we get a contribution s. They
can be added as follows
1
2

sZ+
1
2
;s>0
(s) =
1
4

n=0
(2n + 1)
=
1
4
_

m=1
m

n=1
2n
_
=
1
4
_

m=1
m2

n=1
n
_
=
1
4

m=1
m =
1
48
(5.45)
Hence we nd that the contribution to a of a Neveu-Schwarz fermion is precisely half that
of a boson. As was the case for bosons, there are better ways to arrive at this result, but
this would require more discussion.
5.10 The Ramond ground state
For Ramond fermions the discussion is slightly more complicated. This is due to the
presence of a zero-mode fermion d
0
. Should we impose d
A
0
[0) = 0? If one attempts to do
that one would nd a contradiction with the anti-commutator
_
d
A
0
, d
B
0
_
=
AB
(5.46)
In fact there is another problem. One may easily show that
[L
0
, d
m
] = md
m
, (5.47)
so that acting with d
m
increase the L
0
eigenvalue of a state by m (this holds also for
the other operators
m
and b
r
). Hence acting with d
0
does not change the L
0
eigenvalue,
or in other words, the energy of a state. Hence the state d
A
0
[0) is degenerate with the
vacuum.
This leads one to consider as the ground state of the Ramond fermions a set of states
[a)
R
with the property
d
A
0
[a)
R
=

b
M
A
ab
[b)
R
(5.48)
Here M
A
ab
is a set of matrices (as many as there are fermions) which must satisfy
_
M
A
, M
B
_
=
AB
(5.49)
Such a relation is called a Cliord algebra. In general one prefers to write
M
A
=
1

A
, (5.50)
61
and then a solution is obtained by taking for
A
the (Euclidean) Dirac matrices, which
satisfy
_

A
,
B
_
= 2
AB
(5.51)
For example, for a single fermion the solution is
1
= 1, for two fermions one has

1
=
1
,
2
=
2
, (5.52)
for three fermions one gets
i
=
i
, for four fermions one gets the four-dimensional
Euclidean Dirac matrices (which are 4 4 matrices) etc. In general, in 2n + 1 and 2n
dimension one gets 2
n
2
n
matrices, and the ground states [a)
R
has 2
n
components.
The energy of the ground state (which yields the mass in target space) can be evaluated
as we did for the Neveu-Schwarz fermions, and one easily nds that the shift in L
0
is in
fact
1
2

n=1
n = +
1
24
per fermion, the opposite of free boson.
5.11 Excited states
The higher excited states in the spectrum are as follows. In a model with NS (Neveu-
Schwarz) fermions the target space state corresponding to the vacuum [0) has mass a.

The next excited state is b

A

1
2
[0), and it has mass
1
2
a. Then we get a state
i
1
[0) plus
the states [A, B) = b
A

1
2
b
B

1
2
[0) all with mass 1a. Note that [A, B) must be antisymmetric
in A and B. The value of a is
D2
24
+
N
48
, where D is the number of space-time dimensions
and N the number of fermions (A = 1, . . . , N).
In a Ramond model the ground state is [a)
R
with mass a
R
, with a
R
=
D2
24

N
24
.
The rst excited states are the states
i
1
[a)
R
and d
A
1
[a)
R
with mass 1 a
R
. Here we
have normal ordered the operators d and with respect to the states [a)
R
, using

m
[a)
R
= d
m
[a)
R
= 0 for m > 0 . (5.53)
Note that the naive Lorentz symmetry argument is not of much help if N > 1. The
argument was that the state
i
1
[0) had only D 2 components, and hence has to be
a massless vector. This then determines a, and then also D. In the present case in the
NS model there are additional states at the second excited level; this is also true in the
R model, and furthermore the ground states [a)
R
are themselves degenerate. As before,
the correct approach is to consider the Lorentz algebra. This yields D+
1
2
N = 26 in both
cases.

Hence a = 1 and a
R
= 1
N
16
. Note that if we could build a Ramond model with
sixteen or more fermions, the ground state would have non-negative mass-squared, i.e.
would not be tachyonic. However, we will see that this is not possible: Purely Ramond
or purely Neveu-Schwarz string theories are not consistent. Inconsistencies arise as soon
as one studies interactions.

Here we use units 4/

for closed strings and 1/

for open strings, and we consider only the left-moving

sector of the closed string. As before, left and right sectors must be combined, with the condition L
0
=

L
0
;
mass is an abbreviation for mass-squared.

This is a consequence of the fact that each fermion
contributes
1
2
to the conformal anomaly, as can easily be veried using the Virasoro generators.
62
Bosonization

It turns out that what we have constructed so far is nothing but a compactied bosonic string. If one
compacties a bosonic string on a suitably chosen lattice, the resulting spectrum is exactly the one we got
above, with a precise relation between R and NS sectors. The number of bosons must be twice the number
of fermions. The proof of this statement is related to bosonization: in two dimensions two fermions
1
and

2
are related to a boson . The relation is
1
= e
i
and
2
= e
i
. This is however beyond the scope of
these lectures.
It should be noted that even though we started with fermions in two dimensions, the space-time spectrum
contains only Lorentz tensors, and no spinors. Hence we only get bosons in the target space spectrum.
5.12 Fermionic strings
One can obtain a much more interesting result by making a small modication in the
foregoing discussion. Instead of giving the fermions an internal index A we give them
a Lorentz index . If we only have the bosons X

and the fermions

it is easy to derive
some interesting consequences. The action is of course
S =
1
4

_
dd
_

a
X

a
X

(5.54)
The canonical commutation relations now become
d

n
, d

m
=

n+m
, b

r
, b

s
=

r+s
(5.55)
Now we see immediately that there is a problem: the components d
0
n
and b
0
r
have anti-
commutators of the wrong sign, and will lead to ghost states in the spectrum. In the
case of the bosonic string this was solved by imposing the T
ab
= 0 constraints. Those
constraints were explicitly solve by going to light-cone gauge, using conformal invariance.
This wont help us with

. Since we have already used up conformal invariance to reduce

X

to X
i
, we cannot achieve the same reduction for

. The way out turns out to be

an additional symmetry called super-conformal invariance. We will come back to it later,
but let us assume that it does what we expect, and xes two components of

, reducing
it to
i
.
5.13 The critical dimension
The following holds for open strings, or the left- or rightmoving modes of the closed string
(up to a factor 4/

for closed strings or 1/

for open strings). Consider the rst excited

state in the NS sector, b

1
2
[0) is a vector boson. In light cone gauge it would become
b
i

1
2
[0), which has only D2 components, and is the only state at its mass level. By the
arguments used before, it must be massless. This requires a =
1
2
. On the other hand,
a =
D2
24
+
Ntr
48
. Here N
tr
is the number of transverse fermions, i.e. the fermions that
actually appear in the mass formula, i.e.
i
, i = 1, . . . , D 2. Hence N
tr
= D 2, and
63
from a =
1
2
one easily deduces that D = 10. If we consider the state
i
1
[0) that gave
rise to a massless vector boson in the bosonic string, then we nd that now there is no
problem making it massive. It is accompanied by the set of states b
i

1
2
b
j

1
2
[0), which is an
anti-symmetric tensor. The total number of components at this level is
1
2
(D 2)(D 3) + (D 2) =
1
2
(D 2)(D 1) , (5.56)
precisely the number of components of an asymmetric tensor of SO(D1). Hence at this
level we will have a massive anti-symmetric tensor.
We conclude therefore that the fermionic string considered here has a critical di-
mension D = 10. The number of world-sheet fermions and bosons is equal, namely
N = D = 10. Note that this does not satisfy the relation D +
1
2
N = 26 mentioned in the
previous section.
5.14 The Lorentz algebra

This can be understood as follows. We have argued that the relation D+

1
2
N = 26 follows
from closure of the Lorentz-algebra in lightcone gauge. Indeed, in the bosonic string that
will always lead to relation (5.5): D + c
int
= 26. However, the set of Lorentz generators
of the bosonic string, as shown in (3.93), acts only on the bosonic degrees of freedom X

.
This is also true if we introduce compactied coordinates, or internal fermions
A
. Since
the Lorentz-rotations do not act on these internal degrees of freedom, the form of (3.93)
stays exactly the same. If one goes to lightcone gauge the generators M
i
do depend on
the internal degrees of freedom because we eliminate

n
by solving the condition T
ab
= 0,
and T
ab
contains all degrees of freedom. However, the generators M

still act only on

X

.
Clearly we get new Lorentz generators if we give the fermions a space-time index
. These generators act on

as well as X

. The derivation is easy: the generators

are just the Noether currents associated with the global transformation X

, where

is a Lorentz transformation. The result written in terms of

modes consist of the terms we already had, (3.93), plus the following terms:
NS: i

r=
1
2
_
b

r
b

r
b

r
b

r
_
R:
i
2
[d

0
, d

0
] i

n=1
_
d

n
d

n
d

n
d

n
_
(5.57)
This of course changes the discussion of the Lorentz drastically, and it is no surprise that
the result is dierent. Indeed, in the NS-sector one nds a =
1
2
and D = 10, and in the
R-sector a
R
= 0, D = 10.
64
5.15 The Ramond ground state
We now arrive at a very interesting conclusion regarding the Ramond ground state. Con-
sider the rst term in the expression for the Lorentz-generator. We have seen before that
d
A
0
acts on the Ramond ground states as a Dirac matrix. This does not change if we
replace the index A by a space-time index :
d

0
[a)
R
=
1

ab
[b)
R
(5.58)
Consider now the action of the Lorentz transformation (5.57) on these states. Innitesimal
Lorentz transformations are generated by acting with the operator

on a state,
where is a set of small parameters. We see that under Lorentz transformations [a)
R
transforms as
[a)
R
[a)
R

i
4
[

]
ab

[b)
R
. (5.59)
The matrix
i
4
[

] (often called

) is the innitesimal Lorentz generator in the

spinor representation. Apparently the Ramond ground states rotate like spinors. The
corresponding particles must thus be space-time spinors and assuming the spin-statistics
relation holds they must be fermions. Then all their excited states are fermions as well,
since they are obtained from the ground states by action with space-time bosonic operators
d

and

.
5.16 World sheet supersymmetry

An important missing point in the foregoing is the condition that allows us to x the
light-cone gauge for

. This condition originates from an extra symmetry of the action.

One may verify that it is invariant under the transformation
X

= i
a

a
X

,
(5.60)
where is a spinor that anti-commutes with itself and with

.
A symmetry like this one, which mixes bosons and fermions, is called a supersymme-
try. It is not a symmetry of nature: for example, there are no bosonic partners of the
quarks and leptons. There is a lot of speculation that supersymmetry may be an approx-
imate symmetry of nature, and that the superpartners of the standard model particles
will soon be discovered, but that is irrelevant for our purposes. The supersymmetry we
consider here is a world-sheet symmetry, and its presence does not imply that the target
space spectrum we will get is supersymmetric.
By considering -dependent parameters we may work out the Noether current of
space-time supersymmetry. The result is
S =
1

_
d
2

a
J
a
(5.61)
65
with
J
a
=
1
2

b
X

(5.62)
It is easy to verify that this current is indeed conserved. To do this both the free fermion
and the free boson equations of motion are needed.
The current J
a
is called the supercurrent.
5.17 World sheet supergravity

Up to now the presence of world-sheet supersymmetry was a mere coincidence, but in

fact it turns out to be crucial. However, it is not completely trivial how to maintain it
when we try to write the action in a manifestly reparametrization invariant way. Then
we would introduce a metric

, which means introducing a new two-dimensional eld in

addition to X

and

. Supersymmetry acts on this eld. It cannot act trivially, because

it transforms bosons into fermions. Hence we are obliged to introduce an additional eld,
the superpartner of the word-sheet metric

.
Instead of

a more convenient set of elds is the so-called zwei-bein e

a
,

=
ab
e
a

e
b

(5.63)
The supersymmetry transformation of the zweibein is
e
a

= 2i
a

, (5.64)
where

is a new eld. It is a spinor with an extra two-dimensional vector index. If we

construct such a eld in four dimensions, we get a spin-3/2 elds (spin-1/2 times spin-1
yields spin-3/2), but in two dimensions the notion of spin is rather trivial (the little groups
are SO(1) for massive and SO(0) for massless particles). In any case, this eld is called a
gravitino. It turns out that if we want to write down an action that is reparametrization
invariant and has supersymmetry, we are forced to write down a supergravity theory. The
complete action is
S =
1
4

_
d
2

+ 2

+
1
2

_
.
(5.65)
For more details about this action and how it is obtained see [1]. The rst two terms are
just the generally covariant form of the conformal gauge action. The third term couples
the gravitino to the supercurrent. The last term is a four-fermi interaction required by
supersymmetry. Note that there are no kinetic terms for the gravitino. In more than two
dimensions there would be such a term, but it vanishes in two dimensions. This action is
not just supersymmetric, it is in fact invariant under local supersymmetry, which means
that the transformations can be -dependent.
66
5.18 The constraints

The correct quantization procedure for the fermionic string is to start with the supergrav-
ity action, derive the equations of motion, and only then go to conformal gauge, if one
wishes. This yields one additional equation of motion, namely the one for the gravitino.
In conformal gauge this equation is simply
J
a
= 0 . (5.66)
(conformal gauge means a little more than for the bosonic string. It turns out that
one can choose e
a

=
a

, and one can use the extra fermionic symmetries to set

= 0.)
The other three equations of motion are the usual ones for X

and

, plus the condition

T
ab
= 0.
An obvious guess is now that the new condition is the constraint we were looking
for earlier. This is indeed correct. In light-cone coordinates

splits into
i
,
+
and

. Using conformal symmetry and a supersymmetry transformation that survives the

conformal gauge choice, we may bring X
+
to the same form as before, and we can set
+
to zero. Then the conditions J
a
= T
ab
= 0 can be solved, yielding an expression X

and

in terms of the other components:

+
X

=
1
p
+
T

++

(
+
) =
2
p
+
J

+
(5.67)
where T

and J

are the the energy momentum tensor and the supercurrent, but with
the summation indices = 0, . . . , D1 replaced by i = 1, . . . , D2. In the case of closed
strings there are analogous relations for the right-moving components

and

),
whereas in the open string left- and rightmoving components of all elds are the same.
These relations express

n
and b

r
(or d

n
) in terms of all the other oscillators. To get
the full spectrum it therefore sucient to act with all transverse oscillators, as we already
did above, anticipating this result.
5.19 The conformal anomaly
In the case of the bosonic string the critical dimension could be obtained by cancelling the
conformal anomaly of the elds X

against the contribution of the ghost of superconformal

invariance. This yields D 26 = 0. How does this change if we consider fermionic strings? It
turns out that the extra local supersymmetry we got requires additional Fadeev-Popov ghosts.
These ghosts also contribute to the conformal anomaly, and their contribution is +11. For D
bosons and D fermions we get then the equation D +
1
2
D 26 + 11 = 0, which gives D = 10.
67
6 One loop diagrams
6.1 Neveu-Schwarz or Ramond?
The discussion so far give the impression that we can make a choice: that one can have
Neveu-Schwarz or Ramond strings. If this were true the result would be undesirable,
because we would get either a theory with only bosons, or a theory with only fermions.
However, just as one cannot make a theory with only open strings, one cannot make a
theory with only NS or R strings. As soon as one takes interactions into account, both are
needed. This is seem most elegantly if one considers the string one-loop diagram. This
is what we will compute now. To make life easy we consider a one loop diagram without
external lines. Furthermore we consider a closed string diagram, because that will in any
case occur in any string theory, open or closed.
6.2 The torus
A closed string loop diagram is of course a torus, as shown here
What we mean by computing the diagram is to compute a number, an amplitude,
that corresponds to the diagram. Since there are no external lines, this amplitude is not
the amplitude for some scattering process, but it turns out that it can be interpreted as
a contribution to the energy of the vacuum in target space. Experimentally, this number
would be measurable because it contributes to the cosmological constant, which aects
the evolution of the universe. Although this is a very interesting subject, it is not our
main interest now. We will not be interested in the result of the calculation, but in its
internal consistency. This will impose strong constraints on the theory, which determine
precisely the relation between R and NS.
6.3 The Polyakov path integral

The procedure one uses for doing these calculations is the path-integral. Conceptually this
is rather easy to understand, but the details are messy, and require advanced mathematical
68
methods. Below the main steps are sketched. Details can be found, for example, in [1],[3]
and [2]. Schematically, the calculation goes as follows.
A =
_
T[elds]T[]T[]e
S
E
[elds,,]
(6.1)
Here S
E
is the Euclidean action, obtained from the action in two-dimensional Minkowski
space by dening
0
= i
2
(see appendix A). Furthermore elds stands for all elds
in the theory, such as X

and

and any internal degrees of freedom. One integrates

over all values of these elds in all points on the torus, and also over all values of the
two-dimensional metric and the gravitino. Hence all X

(
a
) are treated as integration
variables, labelled by ,
0
and
1
. Obviously this is an innite-dimensional integral.

The analytic continuation to imaginary time has the advantage of making the path
integral better behaved, because in Minkowski space the weight factor is exp (iS/). In
the classical limit, 0, this integral is dominated by the extrema of the action, the
classical solutions. In the Euclidean formulation it is clear that uctuations around the
classical solutions, which increase the action, are suppressed by the negative exponential.
However, there may be uctuations that do not change the action. These are usually
related to symmetries. Let us rst consider the integration over the two-dimensional
metric . Then the symmetries we have to worry about are reparametrization and Weyl
invariance. If we take the denition we gave above, integrating these symmetries would
simply give rise to an innite factor. Therefore we change the denition:
A
x
=
A
Vol(Di)Vol(Weyl)Vol(Super)
(6.2)
where Vol(Di) denotes the total integration over the reparametrizations (dieomor-
phisms),
Vol(Di) the integration over the Weyl symmetry and Vol(Super) the integration over
the local supersymmetries. Note that we are not yet considering the space-time Poincare
symmetries of X

, because we are discussing the integral over the two-dimensional metric,

which is not aected by these symmetries.
Of course A
x
is a ratio of two innite quantities, which requires some care. This is
done by means of the Fadeev-Popov procedure, which will not be described in detail here.
Essentially one writes the integral over and as an integral over uctuations of these
elds around the conformal gauge, and the latter integrals then cancel against the Vol
factors in the denominator. In doing these manipulations one has to make a change of
integration variables for and , and this results in some Jacobian factors for the change
of measure. The result is thus
A
x
=
_
d
_
T[elds]Jac ()Jac()e
S
E
[elds,,0]
, (6.3)

An additional complication is that fact that some elds are fermionic. The the corresponding path
integral variables must be anti-commuting, which requires a special formalism which will not be discussed
here.
69
where Jac () and Jac () are the Jacobians. The integral over is over all two-
dimensional surfaces that are really inequivalent: surfaces that cannot be transformed
into each other by coordinate and Weyl-transformations. We will return to this later.
To take care of the Jacobians one introduces extra elds, the Fadeev-Popov ghost
b, c, and , such that

Jac ()Jac () =
_
TbTcTTe
S
E
(b,c,,)
, (6.4)
for some suitable ghost action. It turns out that this action is gaussian, so that the
integral can be done explicitly. A gaussian path integral is just a generalization of the
well-known result for a nite-dimensional integral
_
dx
1
. . . dx
n
e
x
i
A
ij
x
j
=
n/2
[det A]

1
2
. (6.5)
If we consider uncompactied strings, the action of X and is gaussian as well, and can
also be done.
In the special case of the torus it turns out that the ghost contribution precisely cancels
against two degrees of freedom of X

and

, so that we are nally left with

A
x
=
_
d
_
T[X]T[]e
S
E
[X
i
,
i
,,0]
, (6.6)
with the action in conformal gauge. This cancellation only works in 26 dimensions for the
bosonic string and 10 dimensions for the superstring.
6.4 Riemann surfaces

The foregoing discussion (with the exception of the last paragraph) was not limited to
the torus, but holds for any surface. Before continuing with the torus, let us make a few
general remarks about string perturbation theory.
The full perturbative expression for any amplitude has the form
out[ U [in) =

topologies
_
d
_
T[elds]
_
TbTcTTe
S
E
(elds,b,c,,)
(6.7)
Here the sum is over all surfaces that cannot be deformed continuously into each other,
and the integral is over all parameters of each surface that do not correspond to coordinate
changes and Weyl transformations.
Since string world sheets are two-dimensional we are in the lucky situation that we
can make use of powerful mathematical results regarding such surfaces, which are called
Riemann surfaces. The parameters of such a surface are called the moduli, and for each
topology the moduli are known.

In the literature is usually called , but unfortunately that character is already in use in this context.
70
The topology of Riemann surfaces is encoded in three integers: the number of handles,
also known as the genus g, the number of holes b and the number of crosscaps c (a
crosscap is a hole of which opposite sides are glued together with opposite orientation.).
For example, a sphere has g = b = c = 0; a torus is a sphere with one handle added to it,
so g = 1, b = c = 0; a disk is a sphere with one hole cut out, so g = c = 0 and b = 1.
Which surfaces are allowed depends on the kind of string theory one considers. For
example, the presence of a hole means that strings can have endpoints. Hence surfaces
with holes appear only for open string theories. A crosscap reverses orientation, and
appears only for unoriented strings. In general one has
Closed, Oriented b = c = 0
Closed, Unoriented b = 0
Open, Oriented c = 0
Open, Unoriented no restrictions
where open means, as usual, open and closed.
The amplitudes of the closed oriented string are sums over the genus g, and hence the
contributing surfaces are the sphere, the torus, the double torus, etc. This is reminiscent
of a loop expansion in eld theory, and hence one might expect that there is an expansion
parameter, a coupling constant, in terms of which higher genus surfaces are suppressed.
6.5 The role of the dilaton

In the discussion of bosonic string compactication we have seen that we can describe
string propagation in non-trivial gravitational backgrounds g

and anti-symmetric tensor

backgrounds B

. There is one other massless eld in the uncompactied closed string,

namely the dilaton. It seems natural to look for a modication of the action that involves
the dilaton . There are in fact several natural candidates, for example
S(X, , )
_
dd

(6.8)
or
S(X, , )
_
dd

(6.9)
to be added to the action we already have. The problem with both proposal is that they
both break Weyl invariance. Weyl-invariance is a transformation of the two-dimensional
elds X and . From the two-dimensional point of view the space-time background elds
g, B and are not elds, but more like coupling constants. They are not supposed to
transform under reparametrizations and Weyl transformations.
If Weyl invariance is broken, we loose the possibility to go to the conformal gauge,
and later to lightcone gauge. This turns out to have fatal consequences. However, there
71
is one possibility that turns out to be Weyl invariant:
S(X, , )
_
dd

R() , (6.10)
where R is the two-dimensional curvature scalar. In four dimensions, and without the
factor , this is the Einstein action for gravity. In two dimensions, however, gravity is
trivial: the equations of motion obtained by varying
_
dd

R() with respect to

vanish identically.
This means that the gravitational action (or (6.10) with constant ) does not change
if we change in a continuous way. Hence the action is the same for a sphere and an
egg-shaped surface, that can be transformed into each other by continuously varying the
metric. However, for two surfaces that cannot be continuously transformed into each
other, such as a sphere and a torus, the action may be dierent. This is indeed the case.
It turns out that the integral of

R() over a manifold is a topological invariant. The
quantity
=
1
4
_
dd

R() (6.11)
is called the Euler number of the surface. It is an integer that does not depend on the
metric, but only on the topology, and is given by
= 2 2g b c . (6.12)
For example a sphere has Euler number two, a disk (a sphere with one hole) has Euler
number 1, and a torus (a sphere with one handle) has Euler number 0.
Unlike g and B, a background value for has no immediate eect on the classical or
quantum treatment of the free boson X. However, the action (6.10) plays an important
role. Suppose we write the dilaton as = c+

. Here c is some constant, and

represents
the uctuations of the dilaton eld around that constant value. Then the action can be
written as S
E
=

S
E
+c (here we have chosen a convenient normalization for the dilaton
action, and we dene

S as the c-independent part of the action). The result is
out[ U [in) =

topologies
e
c
_
d
_
T[elds]
_
TbTcTTe

S
E
(elds,b,c,,)
(6.13)
We see that c appears as a topology-dependent weight factor for the terms in the sum. It
is customary to dene e
c
= , which is called the string coupling constant. The motivation
for this terminology is that each time we add a handle to the surface, we get an extra
factor
2
in the contribution to the amplitude. It is as if there is one factor for each
emission of a closed string.
The string coupling constant is thus related to the classical value (vacuum expectation
value) of a eld, the dilaton. In this sense it is not a parameter of the theory. We would
speak of a parameter if its value had to be dened a priori. But since is related to a
vacuum expectation value of a eld, it is possible that some dynamical mechanism (for
example the minimization of a potential) determines its value. However, as long as this
mechanism is unknown there is not much practical dierence between the two descriptions.
72
6.6 Moduli

The parameters of the surfaces integrated over by the d integral are called the moduli of
the surface. They should not be confused with the moduli of a compactication surface:
the latter is a space-time surface, whereas here we talk about world-sheets. Mathemati-
cally, however, they are the same.
The number of moduli of a surface of genus g (with b = c = 0) is 2g. This implies
that a sphere has no moduli. Conformally, all spheres are equivalent.
In general, any manifold can be covered with coordinate patches. On each patch on
can go to conformal gauge, so that the metric is
ab
. If the manifold has moduli, they are
related to the dierent ways of sewing the patches together.
On a sphere, one can choose a patch that covers the entire sphere except for one
point. This patch can be mapped to the two-dimensional plane. Innity in this plane
corresponds to the missing point on the sphere. In this way all spheres are mapped to
the same plane, and hence they are all equivalent.
A torus is made by rst making a cylinder, and then gluing the ends of the cylinder
together. A cylinder has just one modulus. One can map it to a rectangle (by cutting it
open), and by scale invariance one can make sure that one side of the rectangle has length
one. The length of the other side, L, can then not be changed anymore, and corresponds
to a modulus parameter.
1
L
To see that there really is a free parameter one may a walk on the surface starting in one
corner at a 45 degree angle between the two coordinate axes. Since we are in a metric
ab
this angle is unambiguously dened. If L = 1 we end in another corner, if L ,= 1 we end
on one of the sides. No coordinate or Weyl transformation can change that.
The gluing of the two ends of the cylinder introduces a second parameter. This
parameter can be understood as a rotation of the top of the cylinder before gluing it to
the bottom. Usually one represents a torus as a two-dimensional plane modulo a lattice
73
1
X
Y
One of the axes is chosen along the x-axis, and the length of a basic x-interval is chosen
equal to 1. Then the lattice is specied by a two-dimensional vector as indicated in the
picture. The projection of this vector along the y-axis corresponds to the length of the
cylinder, the projection along the x-axis to the rotation before gluing.
6.7 Complex coordinates
When we go to Euclidean space the factors e
in()
e
in(
0

1
)
appearing in mode
expansions become e
n(
2
i
1
)
. This suggests that it may be more natural to introduce
complex coordinates
z =
2
+ i
1
, z =
2
i
1
(6.14)
This is convenient for other reasons as well. For example, Riemann surfaces are most
easily described as complex surfaces (often called complex curves, since they have just
one complex coordinate). The origin of the simplications is Cauchys theorem, which is
a very powerful tool for evaluating contour integrals. Indeed, these techniques are used
abundantly in string theory. Of course Cauchys theorem is special to two real, or one
complex dimension. This is another reason why one can get a lot further with strings
than with higher dimensional objects.
In terms of complex coordinates the torus is described by the complex plane modulo a
lattice, and the lattice is described by a point in the complex plane. This point is usually
called .

The integral over the moduli, previously denote

_
d, becomes in the case of the torus
_
d
2

( Im )
2
, (6.15)
with d
2
= d Re d Im . The denominator is a measure factor, that gives the correct
weight to dierent values of . It emerges from a careful analysis of the path integral. We
will not derive it here, but we will see later some justication for its presence.

This should not be confused with the world-sheet time coordinate used before. Unfortunately is
used for two dierent quantities in the string literature. From now on we will use Euclidean coordinates

1
and
2
, so that no confusion is possible.
74
6.8 Modular transformations
Up to now we had dened the torus in terms of a lattice. This lattice was dened by two
basis vectors, corresponding to the points 1 and in the complex plane. However,
the same lattice and hence the same torus can be described just as well by choosing
dierent basis vectors. For example the choice 1, + 1 clearly describes the same
lattice
Im
Re
0
1
! !+1
This can be generalized further. One should keep in mind that the torus was dened
by rotating one basis vector along the real axis in the complex plane, and scaling it to 1.
The choice of this basis vector is free; we can also choose the direction . This has the
eect of interchanging the two basis vectors. This rotation, combined with a rescaling the
new basis vector along the real axis, has the eect of replacing by
1

. This is most
easily illustrated by taking purely imaginary:
1
!
rotated
rescaled
-
1
!
75
The set of such transformations of the torus forms a group, called the modular group.
We have identied two elements of that group, namely
T : + 1
S :
1

(6.16)
It turns out that these two transformations generate the entire group. The most general
modular transformation has the form

a + b
c + d
, a, b, c, d Z; ad bc = 1 . (6.17)
This group is isomorphic to SL
2
(Z)/Z
2
. The group SL
2
can be dened by the set of 22
matrices
_
a b
c d
_
(6.18)
with determinant 1. The group SL
2
contains the element 1. In the modular transforma-
tion this is indistinguishable from the identity, and for this reason the modular group is
actually isomorphic to SL
2
(Z)/Z
2
rather than SL
2
(Z). One may check that the modular
transformations satisfy
(ST)
3
= S
2
= 1 . (6.19)
We see thus that not all parameters give dierent tori.
6.9 Integration over moduli
This has implications for the integral over moduli. The integral over is not over the
full positive upper half plane, but should be restricted to a region that covers the set of
distinct tori just once. An example of such a region is shown below
1
2
1
2
- 1 -1
76
The entire upper half plane is covered with an innite number of such regions of dierent
shapes and sizes. For example, the lower part of the strip
1
2
Re <
1
2
contains an
innite number of identied regions. This can be seen as follows. The map 1/
takes the outside of the circle to the inside. Suppose we take the dashed region and shift
it by n units, using + 1 n times. This gives an innity of regions. Now apply
1/. This maps any such region to the inside of the circle, and in fact within the
range
1
2
Re <
1
2
. Hence the lower part of the strip
1
2
Re <
1
2
contains an
innite number of copies of the standard region, the dashed area in the picture.
When we integrate over each such region we get the same answer. At least we should
get the same answer, since otherwise the result depends on the choice we make among the
equivalent regions, and this would clearly be absurd. The requirement that we always get
the same answer is called modular invariance. Although it is formulated here as a property
of the integral, it turns out that the only way for it to hold in general (for example also
for diagrams with external lines) is that in fact the integrand is invariant.
It turns out that the measure (6.15) is already modular invariant by itself. One may
verify that
Im (
a + b
c + d
) = [c + d[
2
Im (6.20)
and that
d
2

= [c + d[
4
d
2
,

=
a + b
c + d
, (6.21)
so that the factors precisely cancel. In other words, the integration measure used in (6.15)
is modular invariant.
It follows that the integrand, obtained from the path integral, must also be modular
invariant.
6.10 Computing the path integral
Now we would like to compute the path integral for the physical degrees of freedom of
the bosonic string. For the bosons X
i
this can be done as follows. Since the integral is
gaussian, it gives simply a determinant, [det ( )]
1/2
. But this is not a useful expression.
Instead we will make use of a standard formula for path integrals in quantum mechanics
of one variable: _
PBC
Tqe
S
E
(q)
= Tr e
H
(6.22)
Here q(t) is a coordinate, and one considers a time interval 0 t . The abbreviation
PBC means periodic boundary conditions. Concretely it means that q(0) = q(). On
the left-hand side one is integrating over all paths that return to their starting point after
a Euclidean time interval .
This result can be straightforwardly generalized to theories with many dynamical
variables q
k
. Here we have innitely many such variables. The quantities X
i
(
1
), for each
i and for each value of
1
should all be viewed as dynamical variables q
k
. But this is not
a serious diculty.
77
In the lattice description of the torus we regard the real axis as the
1
direction, and
the imaginary axis as the Euclidean time direction. If Re = 0 we would nd for the
path integral, as a straightforward generalization of (6.22):

_
TXe
S
E
(X)
= Tr e
2 Im H
. (6.23)
The only small subtlety here is the factor 2. It appears because the torus as depicted
earlier has a periodicity 1 rather than 2 along the
1
direction. To make contact with
earlier conventions we had to scale up the entire lattice by a factor of 2, so that we get
2 instead of .
What happens if Re ,= 0? In that case we have to twist the torus before gluing it
together again, and the periodic boundary conditions in the Euclidean time direction are
dened including such a shift. The operator performing such a shift in the
1
direction is
the momentum operator P. A shift by an amount 2 Re is achieved by the operator
e
iP(2 Re )
(6.24)
The correct result is obtained by inserting this shift operator in the trace, so that we get
_
TXe
S
E
(X)
= Tr e
2 Im H
e
iP(2 Re )
. (6.25)
The operators H and P are the time and space translation operators on the cylinder,
and they can be derived from the energy-momentum tensor. We have already seen that
H = (L
0
1) + (

L
0
1). The expression for the momentum operator is
P =
1
2

_
2
0
d
1
T
01
= L
0

L
0
(6.26)
Putting everything together we nd then
_
TXe
S
E
(X)
= Tr e
2 Im [(L
0
1)(

L
0
1)]+2i Re (L
0

L
0
)
= Tr e
2i(L
0
1)
e
2i (

L
0
1)
(6.27)
6.11 The bosonic string partition function
How do we compute this trace? Consider rst of all only the left-moving part of the
string. A trace of the kind we are trying to compute can be expanded as
Tr e
2i(L
0
1)
=

n
d
n
q
n
(6.28)

To be precise: an innite factor has been dropped on both sides of the equation. On the left hand side
the integral over constant X gives innity; on the right hand the trace is proportional to the normalization
of the vacuum 0[0) = (0). This is a momentum conservation -function which is trivial because there are
no external momenta. This factor is due to the translation invariance of the action; the rest of Poincare
invariance gives a nite integral over the rotations of X
i
, which requires no special treatment.
78
where
q e
2i
(6.29)
and d
n
is the number of states with L
0
1 eigenvalue n. Such a function is usually called
a partition function. So now it is simply a matter of counting states as each level. This
is a simple combinatorial exercise. Remember that the Hilbert space of physical states is
obtained by taking the vacuum [0) and acting with the oscillators
i
n
, n > 0. Suppose
we had just one oscillator
n
. The set of states is then
State Mass
[0) 1

n
[0) n 1

2
n
[0) 2n 1

3
n
[0) 3n 1
. . . . . .
If these were all the states, the trace would yield
q
1
+ q
n1
+ q
2n1
+ q
3n1
+ . . . = q
1
1
1 q
n
(6.30)
Now we give the oscillator an index i, 1 i D 2 ( D will of course have to be
equal to 26, but we leave it as a parameter for the moment). All these oscillators act
independently on the states we already have, and this means that any additional oscillator
has a multiplicative eect on the partition function of the states that were already taken
into account. Hence the full contribution of a set of oscillators
i
n
is
q
1
(
1
1 q
n
)
D2
(6.31)
By the same logic, oscillators with a dierent mode index n also have a multiplicative
eect, so that we get
q
1

n=1
(
1
1 q
n
)
D2
(6.32)
This is the complete partition function of the left-moving sector. It is multiplied by a
similar factor from the right-moving sector, with q replaced by q. Often the result is
expressed in terms of the Dedekind -function
() = e
i/12

n=1
(1 e
2in
) . (6.33)
This function with precisely this factor e
(1/24)2i
turns out to have nice properties
under modular transformations. We will see them in a moment. Comparing this with
79
the string partition function we see that precisely for D = 26 the result can be expressed
completely in terms of -functions:
P
osc.
(, ) = [()( )]
24
. (6.34)
note that ( ) = (())

.
But we are not quite nished. We have forgotten that our Hilbert space also contains
momentum excitations of the vacuum, [p
i
). These excitations work completely indepen-
dently of the oscillators, and again give a multiplicative factor. Since the momenta are
continuous, the trace now becomes an integral
P
mom.
(, )
_
d
D2
p
(2)
D2
e
2 Im (
1
2

p
2
)
(6.35)
This is a gaussian integral, which is easy to do. It yields
P
mom.
(, ) = [2

Im ]
2D
(6.36)
Putting everything together we nd for the bosonic string one-loop path integral (up to
normalization)
_
TXe
S
E
(X)
=
_
1
2

_
24
(6.37)
_
TXe
S
E
(X)

_
d
2

( Im )
2
1
( Im )
12
()
24
( )
24
(6.38)
6.12 Modular invariance of the bosonic string
To check if indeed the bosonic string satises the requirement of modular invariance we
need the transformation of the function. This can be found in many books. The
+ 1 transformation is trivial:
( + 1) = e
i/12
() (6.39)
For the other basic transformation one nds
(1/) =

i() (6.40)
Hence the combination that appears in the string partition function transforms as
[( + 1)[
2
= [()[
2
; [(1/)[
2
= [[[()[
2
(6.41)
The factor [[ in the second transformation cancel against the transformation factor of the
momentum contribution, [

Im ]
1
for each dimension. Since these two basic transfor-
mations generate the entire modular group, it is clear that the bosonic string is modular
invariant.
Note that D = 26 is essential. If D ,= 26 we are left with factors q q to some power
if we try to express the partition function in terms of -functions. These factors are not
invariant under 1/. We discover here a second reason why the critical dimension
of the string must be 26.
80
6.13 Strings versus particles

We have seen that the string spectrum consists of an innity of modes, which an observer
interprets as particles. One may therefore think that a string loop diagram is in some
way related to the sum of particle loop diagrams. To see if that is true, we will now do
the same computation for a single particle, and at the end sum over all particles.
We follow here the discussion in [1], part II, page 55. The string loop diagrams are
supposed to contribute to the cosmological constant . For a massive real free boson
eld theory the relevant expression is
e

=
_
Texp (
_
d
D
x
1
2
_

+ m
2

) =
1
det

+ m
2
(6.42)
The reason we compare the string calculation with the logarithm of the eld theory path
integral is that the latter generates not only the one loop diagram, but also all disconnected
diagrams composed by putting loops together. The string diagram we have computed is
however strictly one-loop. Pictorially the eld theory path integral is
1 +
+ 1/2 + 1/6 + ....
The factors
1
N!
are combinatorial factors, and the whole diagrammatic expansion expo-
nentiates. Hence by taking the logarithm we pick out precisely the one-loop contribution.
The result is equal to
=
1
2
log det ( + m
2
)
=
1
2
Tr log
_
+ m
2

=
1
2
_
d
D
p
(2)
D
p[ log( + m
2
) [p)
=
1
2
_
d
D
p
(2)
D
log(p
2
+ m
2
)
(6.43)
There is another way of seeing that this is the right expression. Up to irrelevant
factors, (p
2
+m
2
)
1
is the propagator of a massive particle. When computing a Feynman
diagram one associates with every internal line such a propagator. If one dierentiates a
propagator with respect to m
2
it splits into two propagators:
d
dm
2
1
p
2
+ m
2
=
1
(p
2
+ m
2
)
2
(6.44)
81
The expression
_
d
D
p
(2)
D
1
(p
2
+ m
2
)
N
(6.45)
is the loop integral for a one loop diagram with N external lines, each with zero mo-
mentum and trivial coupling. Diagrammatically such a diagram is an N-sided polygon.
By dierentiating with respect to m
2
we go from an N-sided to an N + 1-sided polygon.
Our interest is in a one-loop diagram with zero external lines, which is what we com-
puted in string theory. By inverting the recursion, we should be able to compute that
by integrating the N = 1 diagram with respect to m
2
. This of course precisely gives the
logarithm.
To evaluate (6.43) we use the following integral representation
log(X) =
_

0
dt
t
e
tX
(6.46)
This can be derived by dierentiating w.r.t. X on both sides. The divergence at t = 0 is
independent of X and drops out when one dierentiates. So in fact the correct, regulated
expression has a constant term + on the left-hand side:
log(X) = lim
0
__

dt
t
e
tX

dt
t
e
t
_
(6.47)
We drop the m-independent, but innite constant. The rest of the calculation goes as
follows
=
1
2
_
d
D
p
(2)
D
log(p
2
+ m
2
)
=
1
2
_

0
dt
t
_
d
D
p
(2)
D
e
t(p
2
+m
2
)
=
1
2
_

0
dt
t
e
tm
2
_
d
D
p
(2)
D
e
tp
2
=
1
2
_

0
dt
t
e
tm
2 1
(4t)
D/2
(6.48)
As remarked above, in this calculation an m-independent innite term was discarded,
proportional to the volume of momentum space. Apart from that, the integral has now
an even stronger divergence at t = 0. The latter divergence can be recognized as the
usual UV divergence in quantum eld theory. This can be seen as follows. As seen
above, dierentiating (6.43) N times with respect to m
2
gives (6.45), which is the basic
integral one encounters when one computes a one-loop diagram with N external lines.
For N = D/2 this diagram diverges logarithmically for large p, and for N < D/2 there is
a power divergence. These are called an ultra-violet divergences. On the other hand, if
we dierentiate the result of the integration N-times w.r.t. m
2
we get

_
dt
t
t
ND/2
e
tm
2
(6.49)
82
which is also logarithmically divergent for N = D/2 and power divergent for N < D/2.
We conclude that the divergence at t = 0 must be interpreted as the eld theory ultraviolet
divergence. The rst, m-independent divergence has a dierent interpretation. It amounts
to an innite shift in the vacuum energy, and can be attributed to the contribution of the
ground state energies of the innitely many oscillators. This is the target space analog of
the innite sum over n we encountered in string theory. Dropping this term amounts
to setting the energy of the vacuum to zero.
If we treat string theory as an innite set of eld theories, we expect to get

string
=

m
log
_
det (p
2
+ m
2
) (6.50)
where the sum is over all physical states in the theory. Using the integral representation
of a single particle we get (for the uncompactied bosonic string)

string
=
1
2
_

0
dt
t
_
1
4t
_
13

m
e
tm
2
(6.51)
The sum can be evaluated in terms of the partition function (remember that q = e
2i
)

24
()
24
( ) = (q q)
1

n
(1 q
n
)
24

n
(1 q
n
)
24
(6.52)
If we replace q by e
m
2
this is almost the sum we are looking for, except that the partition
function contains more than just physical states. It also includes states with m
L
,= m
R
,
and we have seen before that such states are not physical. We can project out these
unmatched states by integrating over Re . For the imaginary part of the partition
function we get
e
4 Im
[1 + . . .] (6.53)
The tachyon mass is m
2
= 4/

. From the leading term we read o

(4/

)t = 4 Im (6.54)
Hence

m
e
tm
2
=
_ 1
2

1
2
d(Re)
24
()
24
( ) (6.55)
with Im = t/(

). This results in the following expression for

string
=
1
2
_

0
dt
t
_
1
4t
_
13
_ 1
2

1
2
d(Re)
24
()
24
( ) (6.56)

We may ignore the spins of the particles here. We simple sum over all dierent spin states, and treat
the particles as scalars.
83
Changing integration variables to Im we nd

string
=
1
2
_
1
4
2

_
13
_
d
2

( Im )
2
( Im )
12

24
()
24
( ) (6.57)
Apart from overall factors, which are not of interest here, we see that we have obtained
the string loop diagram, with one important dierence: instead of integrating over a
fundamental domain, as shown in the gure in section (5.9), we are integrating over the
entire strip
1
2
< Re
1
2
. But this strip includes the true string contribution an
innite number of times. Hence if the string contribution is nite, the eld theory sum is
necessarily innite.
Furthermore we see that if we take the standard integration region in string theory, as
shown in the gure, the dangerous Im 0 region is cut out. This is how string theory
avoids the eld theory divergence. It includes a natural, built in cut-o that is required
by its internal consistency.
Nevertheless the bosonic string loop integral is not nite. This is entirely due to the
tachyon. Indeed, for a negative mass squared m
2
= A we get an integral of the form
_

0
dt
t
_
1
4t
_
13
e
At
, (6.58)
which diverges badly for large t. Massless particles (A = 0) do not give a divergence for
large t and massive particles are also completely safe. So if we get rid of the tachyon, we
may expect a nite result.
Note that the tachyon divergence comes from the opposite side of the integration
region as the UV-divergence. This is usually called an infrared divergence. The origin of
such divergences is rather dierent, and are usually associated with the denition of the
vacuum.
6.14 Partition function for compactied strings
It is instructive to consider the compactied string partition function. It can be obtained
rather easily from the uncompactied string partition function. The only dierence is
that we do not have continuous components p
I
for the compactied momenta. Since
their contribution is absent, we have to drop a factor (

Im )
N
from the bosonic string
partition function. However, there are also additional contributions to the mass due to
the momenta p
L
and p
R
, see (4.12). Since these momenta can be assigned to any state,
completely independently from the oscillators, this leads to a multiplicative factor in the
partition function. This factor is
P

() =

( p
L
, p
R
)
e
i p
2
L
e
i p
2
R
, (6.59)
where we are assuming that the momenta ( p
L
, p
R
) lie on a lattice . The compactied
string partition function is thus given by
P
comp.
= P
uncomp.
(

Im )
N
P

(6.60)
84
Since P
uncomp
is modular invariant, we have to check modular invariance of the remaining
factors.
To check + 1 is, as always, easy. We see that this requires that p
2
L
p
2
R
is an
even integer. We recognize this is the requirement that should be a Lorentzian even
lattice.
To derive the other transformation rule requires a trick, the Poisson resummation
formula. This works as follows. Suppose we have a lattice , and a continuous function
f( w). Consider the quantity

w
f( w) (6.61)
The Poisson formula re-writes this sum as a sum over the dual lattice

. First dene
F(z) =

w
f( w +z) (6.62)
Obviously F(z) is periodic in z with the periodicity of . Hence we may Fourier expand
it. The Fourier components of functions with periodicity lie on the dual lattice

F(z) =

e
2izv
F

(v) (6.63)
The coecient F

are determined using an inverse Fourier transform:

F

(v) =
1
vol()
_
unit cell of
d
N
ye
2iyv
F(y) , (6.64)
where vol(V ) denotes the volume of the unit cell, i.e. the result of the integral of 1 over
the unit cell. Now we substitute F(y):
F

(v) =
1
vol()
_
unit cell of
d
N
ye
2iyv

w
f( w +y) , (6.65)
and we observe that the sum over all cells combines with the integral over the unit cell to
an integral over all of R
N
, where N is the dimension of . Hence we may write
F

(v) =
1
vol()
_
unit cell of
d
N
ye
2i( w+y)v

w
f( w +y)
=
1
vol()
_
R
N
d
N
xe
2ixv
f(x)
(6.66)
Applying this to F(0) yields the formula

w
f( w) =
1
vol()

(v) (6.67)
85
with
f

(v) =
_
R
N
d
N
xe
2ixv
f(x) (6.68)
For our purpose we apply this to the transformed term in the lattice partition function
f( p
L
, p
R
) = e
i p
2
L
/
e
i p
2
R
/
. (6.69)
It turns out that we should regard ( p
L
, p
R
) as a vector on a Lorentzian lattice; only then
do we get a simple answer. The foregoing discussion holds also for such lattices. The
volume of the unit cell is dened as
vol() =
_
[det g[ , g
ij
= e
i
e
j
, (6.70)
where e
i
is a set of basis vectors. This denition generalizes to the Lorentzian case by
simply replacing the inner product by a Lorentzian one. The same is done with all other
inner products in the foregoing paragraph. The Poisson resummation formula allows us
to write the lattice sum of (6.69) as a sum over the Lorentzian dual lattice. The summand
is
f

(v
L
, v
R
) =
_
R
2N
d
2N
xe
2i(x
L
v
L
x
R
v
R
)
f(x
L
, x
R
) . (6.71)
As long as Im > 0 this is a well-dened gaussian integral, which is easy to do, and
yields
f

(v
L
, v
R
) = [[
N
e
iv
2
L
iv
2
R
(6.72)
Hence
P

(
1

) =
1
vol()
[[
N

(v
L
,v
R
)

e
iv
2
L
e
i v
2
R
, (6.73)
where

is the Lorentzian dual of We observe that the factor [[

N
cancels against the
contribution from ( Im )
N/2
. The only way to get modular invariance is clearly =

and vol() = 1. The latter condition follows in fact from the self-duality, since vol() is
always the inverse of vol(

).
The conclusion is thus that the torus compactied string is modular invariant if and
only the momenta ( p
L
, p
R
) lie on a Lorentzian even, self-dual lattice.
7 Superstrings
Now we will consider the fermionic string theory discussed in chapter 4. Remember that
its critical dimension was 10, and that in lightcone gauge the spectrum was built out of
8 transverse bosons and 8 transverse fermions. The contribution of the bosons to the
partition function does not dier from what we have seen in the previous section. It is
( Im )
4
(()( ))
8
(7.1)
Here we have split the coecient a in L
0
a into two parts, one associated with the bosons
and another associated with the fermions. For the bosonic contribution we make a choice
86
per boson as in the case of the bosonic string, and the rest will have to be supplied by the
fermions. Only in this way can we express the bosonic partition in terms of -functions.
Now we compute the fermionic contributions.
7.1 Free fermion partition functions
7.2 Neveu-Schwarz states
Consider rst the Neveu-Schwarz sector. We have seen that the modes satisfy the following
anti-commutator
b
r
, b
s
=
r+s,0
, (7.2)
and that the vacuum state is annihilated by the negative modes of b
r
.
Let us rst see what a single left-moving Neveu-Schwarz eld contributes. At the rst
few levels, we nd the following states, denoting the L
0
eigenvalue by h:
h = 0 [0)
h =
1
2
b

1
2
[0)
h = 1 none
h =
3
2
b

3
2
[0)
h = 2 b

3
2
b

1
2
[0)
h =
5
2
b

5
2
[0)
h = 3 b

5
2
b

1
2
[0)
h =
7
2
b

7
2
[0)
h = 4 b

7
2
b

1
2
[0) , b

5
2
b

3
2
[0)
(7.3)
Note that fermionic oscillators must satisfy the Pauli exclusion principle, so that for
example b
1/2
b
1/2
is zero. For this reason there is no state at level h = 1, and we have
to go to h = 4 to nd more than one state.
It is straightforward to compute the partition function. Each oscillator b
r
can act
once or zero times on the ground state. If there were just one oscillator b
r
there would
just be two states, [0) and b
r
[0) with h = 0 and h = r. The single oscillator contribution
is thus Tr q
L
0
= 1 + q
r
, where q = exp 2i. All oscillators with dierent modes act
independently, and it is easy to see that each contributes via additional factors of this
form. Furthermore we have to take into account the shift in the ground state energy.
For a free boson it is natural to assign 1/24 to each boson, so that for 24 transverse
bosons we get precisely the required 1. This gives the factor q
1/24
in the -function.
For fermionic strings we have seen that the critical dimension is 10, so that there are
8 transverse dimensions. To shift in the ground state energy should be
1
2
, to which
87
the bosons contribute 1/3. Hence the fermions must contribute 1/6, or 1/48 per
transverse fermion.
The result is thus, for each fermion,
P
NS
(q) = Tr q
L
0

1
48
= q

1
48

r=1/2
(1 + q
r
) . (7.4)
7.3 Ramond states
In the Ramond sector the fermionic oscillators are integer moded. The ground state is
in this case not a single state, but a set of dimension 2
N/2
, where N is the number of
fermions. The ground state forms in fact a spinor representation of SO(N). For the shift
in the vacuum energy we have seen before that it is the same as for a free boson. Hence
we get for the contribution to the partition function for a single fermion
P
R
(q) = Tr q
L
0

1
24
=

2q
1
24

n=1
(1 +q
n
) , (7.5)
by exactly the same arguments as used above. The normalization looks a bit strange,
since one would expect coecients in partition functions to count states. However, for
N fermions we get P
R
(q)
N
, and there is then no problem for even N. If we include left-
and rightmovers, we get P
R
(q)
N
P
R
( q)
N
, and then all coecients are integers for any N.
7.4 Boundary conditions
We now have two kinds of fermionic partition functions, NS and R. We have seen before
that Neveu-Schwarz fermions appear when one choose anti-periodic boundary conditions
around the cylinder, while Ramond fermions appear for periodic boundary conditions.
On the torus this corresponds to boundary conditions along the real axis. One might
expect then that it should also be possible to change the boundary conditions along the
imaginary axis, since the choice of the axes is merely a matter of convention.
Indeed, this is possible. But rst there is something else to be said about the path
integral for fermions. A bosonic path integral is an in integral over real (or complex)
numbers. For example, in the case of one coordinate q(t) we integrate over commuting
numbers q, not over operators. In the case of fermions, that would not be correct, because
even classically fermionic degrees of freedom must anti-commute. This can be taken into
account by dening integrals over anti-commuting c-numbers, also known as Grassmann
variables. Using this formalism one can dene the path integral, and derive the relation
with a trace over a Hilbert space, as we did for bosonic variables. The result is almost
the same, except for one small detail:
_
ABC
Te
S
E
()
= Tr e
H
(7.6)
88
Here ABC stands for anti-periodic boundary conditions. These are the partition func-
tions we have just computed.
We may also consider periodic boundary conditions for fermions, but then we get a
dierent answer. The answer can be obtained by ipping the sign of each fermion before
gluing the top of the cylinder to the bottom, thus changing the boundary condition from
anti-periodic to periodic. This can be achieved by inserting an operator into the trace.
This operator must have eigenvalue 1 for states made with an odd number of fermions,
and +1 for states made with an even number. To construct it we introduce the fermion
number operator F.
F =
1
2

_
2
0
(7.7)
F =

r
: b
r
b
r
; F =

r
: d
r
d
r
(7.8)
Now we insert into the trace the operator (1)
F
. This operator has the property
(1)
F
b = b(1)
F
; (1)
F
d = d(1)
F
, (7.9)
for any fermionic oscillator b
n
or d
n
. Then
_
PBC
Te
S
E
()
= Tr (1)
F
e
H
(7.10)
To compute such a trace we must know how (1)
F
acts on the ground states. In the NS
sector it is natural to require that the vacuum has zero fermion number.
F [0) = 0 ; (1)
F
[0) = [0) (7.11)
In the Ramond sector it is a bit more subtle. We have seen that d
0
transforms the set of
ground states into themselves. On the other hand d
0
changes the (1)
F
eigenvalue when
it acts. The only way out is then that the ground states [a) must split into two subsets,
one with eigenvalue + and the other with eigenvalue 1.
It is quite easy to see how this insertion changes our expressions for the partition
functions. Each fermionic oscillator now contributes a factor 1 q
r
instead of 1 + q
r
.
Hence
P
NS
(q)
PBC
= Tr (1)
F
q
L
0

1
48
= q

1
48

r=
1
2
(1 q
r
) . (7.12)
In the Ramond sector we get
P
R
(q)
PBC
= 0 (7.13)
since the trace over (1)
F
on the ground state manifestly vanishes. It is easy to see that
the same is true for all excited states as well.
Now we can compute four kinds of fermionic partition functions, labelled by the peri-
odicities along each of the two cycles of the torus. For N fermions the result is
89
AA Tr
NS
q
L
0

c
24
(

)
N/2
AP Tr
NS
(1)
F
q
L
0

c
24
(

)
N/2
PA Tr
R
q
L
0

c
24
(

)
N/2
PP Tr
R
(1)
F
q
L
0

c
24
(

)
N/2
Here the letters AP indicate anti-periodicity along the space direction and periodicity
along the time direction on the torus, etc. It turns out that these four partition functions
can be expressed in terms of standard mathematical functions, namely the Jacobi -
functions and the Dedekind -function, which we have already seen. The -functions are
dened as follows

_
a
b
_
(z[) =

n
e
i[(n+a)
2
+2(n+a)(z+b)]
(7.14)
with the additional denitions

1
=
_
1/2
1/2
_
;
2
=
_
1/2
0
_
;
3
=
_
0
0
_
;
4
=
_
0
1/2
_
(7.15)
The last column above indicates the identication of each partition function with ratios
of - and -functions. It is far from obvious that these ratios are equal to the formulas
we derived above. But one may check by expanding in q
n
that indeed equality holds for
any nite number of levels. Of course the equality can be proved to all orders, but we
will not do that here.
Note that the Jacobi -functions have two arguments, but we are only using them at
z = 0 here. The function
1
(z[) vanishes for z = 0, as does the partition function in the
PP sector, but it can be made plausible that the identication given here is the correct
one. To do this one may examine the dependence on the rst argument, which can be
done by considering tori with external legs. Then the contribution from the PP-sector
does not vanish, and the identication with
1
makes sense.
7.5 Modular transformations
Now we discuss modular invariance. Clearly modular transformations change the fermion
boundary conditions. For example, the transformation
1

interchanges the two

cycles (space and time) on the torus, and hence it interchanges AP and PA. The
transformation + 1 maps XY to X(XY) as shown in the gure, where X and Y
stand for A or P, and the multiplication rule is AA=P, AP=A and PP=P. In other words,
it interchanges AA and AP. Since these two transformations generate the modular group,
we generate all permutations of AA, AP and PA. On the other hand PP transforms into
itself.
90
!
!+1
X
Y
XY
X
The explicit form of the modular transformations can be worked out by means of the
Poisson resummation formula, applied to the -function form of the partition function.
Note that this form is rather reminiscent of a summation over a lattice, namely the lattice
of integers. Using this Poisson resummation for this lattice it is rather easy to derive

1
(
1

) = i

i
1
();
2
(
1

) =

i
4
();

3
(
1

) =

i
3
();
4
(
1

) =

i
2
()
(7.16)

1
( + 1) = e
i/4

1
();
2
( + 1) = e
i/4

2
();

3
( + 1) =
4
();
4
( + 1) =
3
()
(7.17)
We have already seen that
(
1

) =

i(); ( + 1) = e
i/12
() (7.18)
7.6 The modular invariant partition function
To construct something modular invariant we clearly need some sort of sum over the
various boundary conditions. Let us consider the correct number of elds for the fermionic
string, namely eight transverse bosons and fermions. The four partition functions are then
of the form
P
i
() =
(
i
())
4

12
()
(7.19)
Under modular transformations
P
1
(
1

) =
4
P
1
() ; P
2
(
1

) =
4
P
4
()
P
4
(
1

) =
4
P
2
() ; P
3
(
1

) =
4
P
3
()
(7.20)
P
1
( + 1) = P
1
() ; P
2
( + 1) = P
2
()
P
3
( + 1) = P
4
() ; P
3
( + 1) = P
3
()
(7.21)
91
The full partition function contains a factor [ Im ()]
(2D)
2
just as for the bosonic string.
This cancels factors
4
in the 1/ transformation. Apart from that we see a sign
in the transformation between P
3
and P
4
. This is cancelled by taking as the partition
function
P
GSO
() =
1
2
(P
3
() P
4
() P
2
() + xP
1
()) (7.22)
Note that the coecient of P
1
is not determined by modular invariance, because it only
transforms into itself. By considering modular invariance at two loops one does get a
constraint on x, and it turns out that x = 1.
The full partition function is then
P
II
(, ) = [ Im ()]
(2D)
2
P
GSO
()P
GSO
( ) (7.23)
Several points require some more discussion. In the foregoing paragraph GSO stands
for Gliozzi, Scherk and Olive, who invented this construction in a somewhat dierent way.
The subscript II stands for type-II, the name given for historical reasons to this kind
of string theory. Let us see what the linear combination does to the spectrum. In the
Neveu-Schwarz sector the combination
1
2
(P
3
P
4
) = Tr
NS
1
2
(1 (1)
F
)q
L
0

1
2
(7.24)
is a partition function with a projection operator inserted. Indeed, the combination
1
2
(1 (1)
F
) is a projection operator, that removes all states from the spectrum that are
created by and even number of fermion operators (the GSO-projection). This is good
news, because the tachyon corresponds to the state [0) with F = 0. Hence in this string
theory the tachyon is gone! At the rst excited level we nd the states
b
i
1/2
[0) (7.25)
which stay in the spectrum (note that the normalization factor
1
2
ensures that there is
precisely one such multiplet). As we have already seen, these states are massless. Together
with the ones from the rightmoving sector they produce a graviton, a dilaton and a B

.
In the Ramond sector a similar projection occurs. Here we have
1
2
(P
2
P
1
) = Tr
R
1
2
(1 (1)
F
)q
L
0
1
. (7.26)
To interpret this we have to think about the action of (1)
F
on the ground state. We
have seen that (1)
F
anti-commutes with all the d
i
0
operators, and we have also seen that
d
i
0
can be represented up to a factor

2 by the Dirac matrices. Then there is a

natural candidate for (1)
F
, namely

= i
N/2

i
(7.27)
92
(the factor i
N/2
is inserted to make sure that

is Hermitean). This matrix is the general-

ization to N Euclidean

dimensions of the familiar operator

5
in four dimension. In four
dimensions the projection operators
1
2
(1
5
) are also familiar. They project on fermions
with chirality + or 1. The former have their spin aligned with the direction of motion,
and the latter have it anti-aligned. States with denite chirality must be massless, since
otherwise we can ip the velocity relative to the spin by overtaking the particle.
In other dimensions the notion of spin changes, but chirality can still be dened,
namely as the eigenvalue of

. Hence the operators

1
2
(1

) project on the two chiralities.

In case of the 10-dimensional fermionic string the Ramond ground state has multiplicity
2
(2D)/2
= 16. It consists out of 8 states with chirality + and eight with chirality . We
see thus that the GSO-projection selects half the states. We can choose the chirality by
selecting the + sign or the sign in front of P
1
. Since we can do that for left- and right-
moving modes separately, there are in fact four type-II strings, which we could denote as
II
++
, II
+
, II
+
and II

. But by making a reection in 10 dimensions we can ip all

chiralities simultaneously, so that II
++
is equivalent to II

and II
+
to II
+
. Hence there
are really only two string theories, called type-IIA (opposite chiralities) and type-IIB (the
same chiralities).
Note that the Ramond partition function in P
GSO
is multiplied with 1. This is
good news, because we have seen that Ramond states are space-time fermions. A well-
known feature of fermions is that they contribute to loop diagrams with a 1 sign. This
is closely related to the spin-statistics relation. Since the torus diagram is a one-loop
diagram, bosons and fermions should contribute to it with opposite sign. Remarkably,
this is automatically guaranteed by modular invariance.
7.7 Multi-loop modular invariance
In principle one should also worry about higher loop modular invariance. It turns out
that this gives very few constraints. From modular invariance for two-loop surfaces we
do learn one interesting fact, the normalization of the
1
contribution.
The essence of the argument goes as follows. Consider a two-loop diagram in the limit
where it almost separates into two one-loop tori. In that limit the two-loop partition
function factorizes into two one-loop partition functions, which are described by Jacobi
-functions. An n-loop surface has 2n independent loops that cannot be contracted to a
point. This is shown below for the three-torus.

Remember that we are only considering the transverse fermions here.

93
a
1
a
2
a
3
b
1
b
2
b
3
On each loop the fermions can have boundary condition + or . This gives a total of 2
2n
boundary conditions on an n-loop surface: Four for the torus, as we have seen, 16 for the
two-torus, etc. On the two-torus there are then 16 distinct terms in the partition function.
In the limit two-torus torus torus those 16 partition functions approach the product
P
i
P
j
. It turns out that modular transformations of the two-torus mix up the boundary
conditions. The 10 combinations (i, j) = (2 . . . 4, 2 . . . 4) and (1,1) are transformed into
each other, and the six combinations (1, 2 . . . 4) and (2 . . . 4, 1) are also mixed with each
other. Invariance under modular transformations at two loops thus requires the coecient
of P
1
P
1
, x
2
, to be equal to the coecient of P
3
P
3
, namely 1 (there is no sign change in
this transformation).
7.8 Superstrings
The complete massless spectrum consists out of states coming from four sectors: NS-NS,
NS-R, R-NS and R-R. The NS-NS states produce the graviton, the dilaton and the B

,
as mentioned above. The NS-R and R-NS produce, in lightcone coordinates, a chiral
spinor times a vector. In the type-IIA string NS-R and R-NS sectors have states with
opposite space-time chirality, whereas in the type-IIB theory they have the same chirality.
A theory which is invariant under chirality-ip of all particles is called non-chiral. We see
thus that type-IIA is non-chiral, whereas type-IIB is chiral. Note that the standard model
is chiral: for example, left-handed fermions couple to the W boson whereas right-handed
fermions do not.
Covariantly, the product spinor times vector produces a gravitino and an ordinary
spinor with opposite chirality. In four dimensions a gravitino is a particle with spin 3/2.
In any dimension, the presence of a massless gravitino in the spectrum implies that the
spectrum is supersymmetric.
7.9 Supersymmetry and Supergravity

We have already discussed supersymmetry in two dimensions. The basic idea is the same
in any dimension. One introduces an innitesimal transformation which takes bosons into
fermions, and vice-versa. In general there can be several supersymmetries.
94
In gravitational theories the supersymmetry acts in particular on the graviton. Each
supersymmetry produces one gravitino. In the present case we get in fact two gravi-
tini: one from R-NS and one from NS-R. One says therefore that the type-IIA and -IIB
theories have N=2 supersymmetry. Theories with gravity that are supersymmetric are
in fact invariant under local supersymmetry. This means that they are invariant under
innitesimal transformations with a space-time dependent parameter (x). Such theories
are called supergravity theories.
Given a number of supersymmetries one can work out the possible super-multiplets,
combinations of bosons and fermions that are closed under the action of the supersym-
metry. For example in two dimensions a boson and a fermion form a supermultiplet.
As one may expect, the sizes of the multiplets increase with the number of supersym-
metries. In four dimensions, the maximum number is N=4 if one wants to restrict to
global supersymmetry (i.e. no graviton and gravitino), and N=8 if one does not want to
have particles of spin larger than 2. It turns out to be impossible to make interacting
theories with particles of spin larger than 2, so that the absolute maximum number of
supersymmetries in four dimensions is 8. Similar maxima exist in other dimensions; for
D = 5, 6 the maximum is N=4; for D = 7, 8, 9 and 10 the maximum is N=2, and for
D = 11 the maximum is 1. Beyond D = 11 there are no supergravity theories at all!
7.10 Supersymmetry for the massive states
We see thus that the type-II theories produce as their massless spectrum N=2 super-
gravity theories. Indeed they produce precisely the only two known N=2 supergravity
theories in ten dimensions. These theories were constructed independently of string the-
ory. Supersymmetry implies that the number of physical bosonic states must be equal
to the number of physical fermionic states. In light-cone gauge, the bosonic states from
one side of the theory are a transverse vector b
i
1/2
[0) with D 2 components, and the
fermionic states are a chiral spinor [a) with
1
2
2
(D2)/2
components (there is a factor
1
2
because of the projection on one chirality). For D = 10 these numbers are both equal to
8, for other dimensions they dier. String theory has in addition to these massless states
innitely many massive states, on which supersymmetry has to act as well. The action
of supersymmetry has a spectacular eect on the partition functions. The -functions
satisfy the so-called abstruse identity of Jacobi:
(
3
)
4
(
4
)
4
(
2
)
4
= 0 (7.28)
The word abstruse means dicult to comprehend, but to quote Gliozzi, Scherk and
Olive: This identity is not abstruse to us because it is a consequence of supersymmetry
at the massive string levels. Indeed, the identity is a consequence of the fact that at
every level the number of fermions and bosons is equal, and that bosons and fermions
95
contribute to the partition function with opposite sign!

7.11 Ramond-Ramond particles

The sector we have not considered yet is the RR-sector. The particles coming from
this sector are products of fermions and fermions, and hence have integer spin (more
precisely, they are in tensor representations of the Lorentz-group SO(D 1, 1) rather
than in spinor representations) and bosonic statistics. To work out their representations
requires some knowledge about computing tensor products of two spinor representations
in SO(N). The answer is that in the type-IIA theory the RR-sector contains a vector
boson and a rank-3 anti-symmetric tensor, while in the type-IIB theory one gets a scalar,
an anti-symmetric rank-2 tensor and a self-dual anti-symmetric rank-4 tensor.

7.12 Other modular invariant string theories

In the foregoing discussion we have found a solution to the condition of modular invariance
by imposing the condition on left- and rightmovers separately. But this is not necessary.
Instead we might consider
[ Im ()]
(2D)
2

ij
a
ij
P
i
()P
j
( ) (7.29)
and require invariance under modular transformations. Assuming that the P
1
contri-
butions behave as before we nd that there are two independent solutions, namely the
type-II strings seen above and furthermore (with a P
1
-term added)
P
0
(, ) =
1
2
[ Im ()]
(2D)
2
[P
3
()P
3
( ) + P
4
()P
4
( ) + P
2
()P
2
( ) P
1
()P
1
( )]
(7.30)
This is called the type-0 string theory. It also comes in two kinds, namely type-0A and
type-0B, which have opposite signs in the last term. These theories have a tachyon,
because the NS ground states is not projected out, and they have no space-time fermions
since there are no NS-R and R-NS sectors. They are not supersymmetric. In fact it can
be shown that supersymmetry forbids a tachyon. The converse is not true: there do exist
non-supersymmetric strings without tachyons.

Technical remark: Only physical boson and fermion degrees of freedom should be counted. These are
also called the on-shell degrees of freedom. For example a massless vector boson has D2 physical degrees
of freedom. In the light-cone gauge one only sees physical degrees of freedom, so that one automatically
obtains the correct count.

perhaps surprisingly, rank-4 tensors in 10 dimension can have a chirality
+ or , just like fermions, and one nds a rank-4 tensor of chirality + or (depending in whether one
considers the II
++
or the II

version of the IIB string). Such tensors are called self-dual or anti-self-dual.
96
7.13 Internal fermions
When we discussed fermions on the world-sheet we started with fermions
A
that have
an internal index A rather than a space-time index . If one builds theories with such
fermions, the issue of modular invariance also arises.
We are now back at the bosonic string. We have seen that if we have 2N internal
fermions, the resulting string theory lives in 26 N dimensions. The partition function
can be built out of functions
F
i
() =

i
()
()
, (7.31)
and similarly for the right-moving modes. A function F
i
represents a pair of fermions. A
partition function of a fermionic string contains some combination of N such functions
times a product of 24N functions
1
originating from the transverse bosons X
i
. Hence
the full partition function is a linear combination of terms of the form
1
()
24N
N

m=1
F
im
() =
1
()
24
N

m=1

im
() (7.32)
The -function to the power 24 is completely modular invariant, and hence the -functions
must be modular invariant as well (note that we may ignore the factors

, since they will

cancel no matter what we do). For simplicity we will try to make the other factors cancel
separately for left- and right-movers. From the transformation of
2
under + 1 we
see then that N must be a multiple of 8: otherwise there can be no
2
term, and then the
others must vanish as well. If N = 8 there is indeed a simple solution, which we write in
terms of the ratios F as
F
8
=
1
2
_
(F
3
)
8
+ (F
4
)
8
+ (F
2
)
8
(F
1
)
8

(7.33)
If N = 16 there are two solutions, namely (F
8
)
2
and F
16
, the latter dened in the obvious
way.
The free fermion Lagrangian for 2N fermions is symmetric under SO(2N) rotations

A
=

B
O
B
A

B
(7.34)
where O is an orthogonal matrix. (Note that since
A
is real SU(2N) is not a symmetry).
This symmetry can be seen in the spectrum provided that the boundary conditions respect
it. In other words, only if all 2N fermions always have the same boundary conditions.
If we consider for example the F
16
theory we expect the states in the spectrum to form
SO(32) multiplets. Indeed, b
A
1/2
transforms as an SO(32) vector. For the (F
8
)
2
theory we
only expect an SO(16) SO(16) symmetry.
Consider now a theory in 18 dimensions plus 16 internal fermions. The only way to
construct a partition function that is separately modular invariant on the right and the
left is
( Im )
8
()
16
( )
16
F
8
()F
8
( ) (7.35)
97
Let us compute the spectrum of this theory, starting with the left-moving NS sector. Since
the relative sign between F
3
and F
4
is positive, we keep all the states created by an even
number of bs. Hence the lowest states are the tachyon and the states b
A
1/2
b
B
1/2
[0). The
latter are massless and form an anti-symmetric tensor of the symmetry group, SO(16),
with dimension 120. In the Ramond sector we get as the lowest state a chiral SO(16)
spinor. This has
1
2
2
8
= 128 components. Furthermore there is a space-time vector created
by

1
. When we combine left and right sector we get the following massless states:
248+248 vector bosons, 248 248 scalars, and of course the graviton, the dilaton and a
B

.
This is probably reminiscent of something we have seen before (see appendix C): these
are precisely the states that one obtains if one compacties the bosonic string to 18
dimensions on the (E
8
)
L
(E
8
)
R
Lorentzian ESDL. It is now not a very bold conjecture
that perhaps these theories are the same. This is indeed true, and we can get additional
evidence for this conjecture by showing that the state multiplicities are the same at any
level, as a consequence of the following identity
1
2
_
(

)
8
+ (

)
8
+ (

)
8
_
=
1

vE
8
e
iv
2
(7.36)
This identity is easy to prove if we write
i
as a sum, as in 7.14). This equivalence is an
example of bosonization. On the left hand side we have a theory constructed out of 16
free fermions, on the right hand side a theory constructed out of 8 free bosons. Indeed,
in general a pair of fermions with appropriate boundary conditions is equivalent to a free
boson on an appropriate torus.
Similarly the two fermionic theories in 26 16 = 10 dimensions can be identied
with the E
8
E
8
and D
16
Euclidean ESDLs in 16 dimensions. If we combine left and
rightmovers we can in fact built three theories: (E
8
E
8
)
L
(E
8
E
8
)
R
, (D
16
)
L
(D
16
)
R
and also the hybrid combination (E
8
E
8
)
L
(D
16
)
R
. The partition functions are of
course
( Im )
4
()
8
( )
8
F
16
()F
16
( ) (7.37)
for the (D
16
)
L
(D
16
)
R
theory, and obvious variations on this theme for the other two.
It should be emphasized that there are far more bosonic string compactications; we
have only considered here the ones that have separate modular invariance in the left and
right.
8 Heterotic strings
The foregoing discussion leads us naturally to the idea of combining the right sector of
the type-II string with the left sector of one of the 10-dimensional compactied bosonic
strings. Since both are separately modular invariant in the left and right sectors, cutting
each theory in half and gluing the parts back together should give a modular invariant
theory. This thought is indeed correct, and the result was so important that a new name
was given to it: the heterotic string.
98
Since there are two distinct leftmoving partition functions in the 10-dimensional bosonic
string, we can build two heterotic strings. The partition functions are given by
1
( Im )
4
P
GSO
( )()
8
F
8
()F
8
() (8.1)
and
1
( Im )
4
P
GSO
( )()
8
F
16
() (8.2)
These two are called respectively the E
8
E
8
heterotic string and the SO(32) heterotic
string. In principle we also have the option of choosing the sign in front of P
1
in P
GSO
.
But this only ips the chirality of all the fermions, and has no observable eect, because
there is no other chirality to compare with.
8.1 The massless spectrum
Note rst of all that the spectrum does not contain a tachyon. The leftmoving bosonic
sector contains one, but the rightmoving superstring does not. Hence there is no physical
tachyon state, because the condition M
L
= M
R
cannot be satised.
Analyzing the spectrum of the heterotic strings is easy. The right sector gives us a
vector representation and a spinor representation of the transverse SO(8) group in 10
dimensions. The left sector gives as a vector of SO(8) plus a number of scalars from the
(F
8
)
2
or F
16
part of the partition function. These scalars come from the NS parts in both
cases, and from the R part in the case of F
8
. The total number is
(F
8
)
2
: 2 (
1
2
(16 15)
NS
+
1
2
(2
8
)
R
) = 496
(D
16
) :
1
2
(32 31)
NS
= 496
(8.3)
Note that in the rst case the subscript R refers only to the rst (or second) group of
16 fermions. The other group of 16 is always in the NS ground state. The total ground
state mass contribution for the 8 bosons, 16 NS fermions and 16 R fermions is thus
(8(
1
24
)+16(
1
48
)+16(+
1
24
)) = 0. For the second case the analogous computation
yields (8 (
1
24
) + 32 (+
1
24
) = 1, and hence the Ramond ground state is massive.
The complete massless spectrum is thus (V =vector and S =spinor)
(V + S) (V + 496) (8.4)
which yields the usual graviton, B

and dilaton (from V V ), 496 massless vector bosons,

a gravitino and an ordinary spinor with opposite chirality (from SV ) and 496 additional
spinors.
99
8.2 Supersymmetry

The presence of the gravitino suggests that there must be supersymmetry. Indeed, this
theory has N = 1 supersymmetry (only N = 1 because in this case there is only one grav-
itino). As for type-II it is in fact a local supersymmetry, i.e. supergravity. Any symmetry
must act on all elds in a theory. Since supersymmetry switches bosons an fermions, it
cannot act trivially. Hence it must act non-trivially on the 496 vector bosons. Obviously,
they are transformed into the 496 fermions and vice-versa. In any supersymmetric theory
vector bosons are accompanied by fermions in the same representation (i.e. the adjoint
representation). These fermions are called gauginos (in the supersymmetric extensions
of the standard model one speaks of photinos, gluinos, Winos and Zinos; experiments
are looking for such particles, but so far they havent been found).
8.3 Gauge symmetry

Just as the presence of a vector-spinor or gravitino implies supergravity, in string theory

the presence of a vector boson always implies the existence of a local internal symmetry or,
in other words, a gauge symmetry. Gauge interactions can be seen as a generalization of
electrodynamics (which is a U(1) gauge theory) to non-abelian symmetries (i.e. symmetry
algebras whose generators do not commute with each other). In this context, the non-
abelian symmetries always form a (semi-)simple Lie-algebra (the most familiar example
is SU(2)). The number of gauge bosons is always equal to the number of generators of
the Lie-algebra. One can write down a gauge-invariant action for the vector bosons, and
one can couple them to other particles. For example, the coupling to fermions
I
is as

I,J

igA
a

T
a
IJ
)
J
(8.5)
The sum is over all components of the fermion eld. The matrices T
a
form a representation
of the Lie-algebra, which means that they satisfy the commutation relations. The number
of components of
I
is called the dimension of the representation.
In the present case, the gauge symmetry includes the global SO(2N) symmetry that
rotates the world-sheet fermions into each other. In the case of the SO(32) heterotic string
this is indeed precisely the full gauge symmetry. However, in the case of the E
8
E
8
heterotic string one would expect only an SO(16) SO(16) symmetry on the basis of
this argument. Here something special happens. From the Ramond sector we get (two
times) 128 extra massless vector bosons. As a result, the gauge symmetry gets extended
as well. The extension is described by a Lie algebra called E
8
, which has 248 generators
and contains SO(16) as a sub-algebra. The gauginos are in the same representation as
the gauge bosons, the adjoint representation.
The spinor S of SO(8) is chiral because of the GSO-projection in the right sector. It
follows that the gauginos are chiral as well (as is in fact the gravitino).
100
8.4 Compactication
When it was discovered in 1984, the heterotic string was almost immediately seen as an
excellent theory for particle physics. The main reasons are that it has a suciently large
gauge symmetry to contain the standard model, and that it has massless chiral fermions.
For some people even the presence of supersymmetry in the spectrum is an advantage,
although it has not (yet) been seen in nature. The reason is that already at that time
supersymmetry was considered as an attractive mechanism to solve the so called hierarchy
problem. Simply stated, this is the question how to explain the large discrepancy between
the weak interaction scale ( 100 GeV) and the natural scale of quantum gravity, the
Planck scale ( 10
19
GeV) (in fact, supersymmetry does not explain that discrepancy, but
explains why it is stable against quantum uctuations). Obviously, since supersymmetry
is not observed it can at best be an approximate or broken symmetry, but from the
Planck scale point of view it is to rst approximation exact, and the breaking is only a
higher order correction, which is usually ignored in string phenomenology.
The most striking disagreement of the heterotic string description with our world is
the number of space-time dimensions. But we have already seen that the number can
be reduced to 4 by compactifying 6 dimensions. Just as the bosonic string, the heterotic
string can be compactied on a torus, but it turns out that this is not a good idea. The
reason is that in torus compactications all higher-dimensional elds are decomposed into
lower-dimensional ones according to simple group-theoretical rules:
SO(D 1, 1) SO(D N 1, 1) SO(N) (8.6)
(in our case D = 10 and N = 6) In such a decomposition, a vector V decomposes as
(V, 1) +(1, V ). In matrix notation this is simply a matter of decomposing the matrix into
blocks
_
SO(3, 1) matrix 0
0 SO(6) matrix
_
(8.7)
An observer in four dimensions only sees the rst four components, and views the internal
components as a multiplicity. Hence he/she sees a vector and N scalars.
Now we have to determine how this works for spinors. To simplify the notation we do
this in Euclidean space, for a decomposition
SO(D) SO(D N) SO(N) . (8.8)
The only dierence with the Minkowski case are a few irrelevant factors of i in some
expressions. We assume that both D and N are even. In that case there are two kinds
of fundamental spinor representations. They are distinguished by the eigenvalue of the
matrix

, the product of all Dirac matrices (usually called

5
in four dimensions).
Innitesimal rotations of these spinors are generated by the matrices

=
1
4
i [

] , (8.9)
101
where

are the Dirac matrices. In N even dimensions, these are 2

N/2
2
N/2
matrices.
Since we have three dierent spaces, we have to introduce three dierent set of Dirac
matrices:
SO(D) :
M
SO(D N) :

SO(N) :
I
(8.10)
All these matrices can be written down explicitly in many dierent ways, but it can be
shown that all choices are equivalent.
Now we compute

, and we get

= (1)
N/2

DN

N
(8.11)
The two chirality eigenspaces of

are projected out by the operators

P

D
=
1
2
(1

) (8.12)
Writing this in terms of the subgroup we nd
P
+
D
= P
+
DN
P
+
N
+ P

DN
P

N
for N = 0 mod 4 (8.13)
and
P
+
D
= P
+
DN
P

N
+ P

DN
P
+
N
for N = 2 mod 4 (8.14)
Denoting the spinor representations as S (chirality +) and C (chirality ) we may write
this in the case of interest as
S
10
(S
4
, C
6
) + (C
4
, S
6
) (8.15)
An observer in four dimensions sees a 10-dimensional chiral spinor as 4 chiral spinors
S
4
and 4 anti-chiral ones C
4
. The 4 components are the dimensions of S
6
and C
6
; an
observer in the four uncompactied dimensions cannot rotate the internal components of
the spinor, but observes the number of components.
However, we are only interested in massless spinors, so we have to work out the
mass that is observed. This is determined by solving the Dirac equation. Its internal
space eigenvalues are observed as a mass (of course we start with a massless spinor in 10
dimensions). A four-dimensional massless fermion is a ten-dimensional massless spinor
that does not get a mass contribution from the internal dimensions:
_
(

1)

+ (1
I
)
I

= 0 (8.16)
If the second term vanishes, this is observed as zero mass in four dimensions.

The Dirac
equation in internal space is just
I

I
= 0, and it has 2
N/2
solutions (of both chiralities):

Note that in general it is a bit misleading to speak of eigenvalues, because the Dirac matrix acts
o-diagonally; however, if the eigenvalue is zero, this does not matter.
102

= constant
a
, a = 1 . . . 2
N/2
, where is a spinor index and a labels the dierent
solutions. These are all the independent constant vectors in spinor space. The chiral and
anti-chiral spinors are in the same representation of any gauge group, and hence we end
up with a non-chiral theory: we cannot get the standard model.
The number of solutions in internal space gives the number of observed massless
particles in four dimensions. These particles are distinguished from each other by having
a dierent wave function in the compactied dimensions.
The foregoing problem is fortunately special to torus compactications. If one com-
pacties on other manifolds, a chiral spinor decomposes in a certain number N
L
of chiral
spinors and another number N
R
of anti-chiral ones.
On other spaces one has to generalize 8.16) to take into account curvature. This implies
that the ordinary derivatives
I
are replaced by covariant derivatives which contain the
spin connection of the manifold (this was trivial for the torus, since a torus is at). An
additional complication is that sometimes one allows some of the elds in the space-time
theory to have classical values (an example of that was considered in the discussion of
torus compactication where we allowed a non-vanishing but constant background
eld B
IJ
). Often one chooses a non-vanishing classical gauge eld, and in that case the
covariant derivative must also contain a gauge potential term. All of these issues are
beyond the scope of these lectures. However, it is clear that in general a dierent, more
complicated equation must be solved in the internal dimensions. In general N
L
,= N
R
so
that a chiral theory may be obtained after compactication.
8.5 Calabi-Yau compactication

One cannot simply compactify space-time on any manifold. String theory imposes con-
straints on the space-times it can propagate in. These constraints follow from conformal
invariance. A very popular class of solutions is a class of six-dimensional manifolds called
Calabi-Yau spaces. They satisfy the additional requirement that in compactication one
supersymmetry is preserved. Note that the heterotic string in 10 dimensions has N = 1
supersymmetry: there is one gravitino. The rules for decomposing a gravitino in torus
compactications are the same as for ordinary fermions, as discussed above. Hence in
four dimensions the torus-compactied heterotic string has four chiral and four anti-chiral
gravitinos. In four dimensions the anti-chiral ones are just the anti-particles of the chiral
ones, so eectively there are four gravitinos and one calls this N = 4 supersymmetry. It
turns out that chiral theories in four dimensions are possible only for N = 0 and N = 1
supersymmetry.
The numbers N
L
and N
R
in Calabi-Yau compactication can be computed. They are
topological, which means that they do not change under continuous deformations of the
manifold. The dierence is equal to half the Euler number of the manifold.
103
8.6 The standard model in string theory

We must rst review a few basic properties of the standard model. It contains particles of
spin 1/2, the quarks and leptons, particles of spin 1, the carriers of the electromagnetic,
strong and weak forces, and most likely a particle of spin 0, the Higgs boson. The spin
1 bosons belong to a gauge group SU(3) SU(2) U(1), which below about 100 GeV
is broken down to SU(3) U(1), QCD and electrodynamics. If we ignore the weak
interactions the standard model is non-chiral: QED and QCD couple in the same way to
both chiralities. However, the weak interactions do distinguish chirality, and couple only
to the left-handed and not to the right-handed electron.
A remarkable feature of the standard model is that the particles are grouped in families:
it is natural to associate (e,
e
, u, d), (,

, c, s) and (,

, t, b) with each other. Although

this grouping is rather arbitrary (it is determined only by the mass hierarchy, which is not
understood, and is furthermore disturbed by mixing), it is certainly true that the basic
SU(3) SU(2) U(1) quantum number repeat three times.
When writing one family of the standard model, it is sometimes useful to replace
certain particles by their anti-particles. In general, anti-particles have opposite charge and
complex conjugate non-abelian group representations. Furthermore in four dimensions
(and in all 4m dimensions) particles and anti-particles have opposite chirality (in 4m+2
dimensions they have the same chirality). For massless particles chirality is equivalent to
helicity: a particle with chirality + is left-handed. Hence, for example, the anti-particle of
e

R
is e
+
L
. This can be used to make all particles in a standard model family left-handed.
If one does that, the SU(3) SU(2) U(1) representations are
Representation Family 1 Family 2 Family 3
(3, 2,
1
6
) u, d c, s t, b
(

3, 1,
2
3
) u c

t
(

3, 1,
1
3
)

d s

d
(1, 2,
1
2
) e

,
e

(1, 2, 1) e
+

+
Here the bar denotes the conjugate of a representation, and representations are
denoted as (dimension).
The standard model gauge group can be embedded into larger groups. The most
familiar embedding chain is
104
The idea is that just as the low energy gauge group SU(3) U(1) is extended at 100
GeV to the standard model gauge group SU(3) SU(2) U(1), perhaps at still higher
energies the standard model gauge group is enlarged further. This can even be done in
such a way that there is just one gauge group and one representation per family. From
this point of view SO(10) looks perhaps most attractive, but the next step, E
6
, has also
been studied in the literature.
The point is now that there is a class of compactications of the E
8
E
8
heterotic
string in which the gauge group is broken to E
6
E
8
. Furthermore in these compacti-
cations the massless fermions are automatically in the representation (27). One can have
N
L
lefthanded fermions and N
R
righthanded ones, with N
L
,= N
R
. If we represent the
righthanded ones by anti-particles, we may also write that as N
L
(27)
L
+ N
R
(27)
L
, i.e.
with only left-handed particles.
In switching from particles to anti-particles one has to be careful with the notion of
chirality of a theory. The rule is simple: if after ipping all chiralities one obtains a
dierent theory, the theory is chiral. However, if we can write one theory in the form of
another one by trading some particles for anti-particles, then those two theories are not
dierent. For example
(27)
L

chirality ip
(27)
R
(27)
L
chiral
(27)
L
+ (27)
L

chirality ip
(27)
R
+ (27)
R
(27)
L
+ (27)
L
non-chiral
(10)
L

chirality ip
(10)
R
(10)
L
= (10)
L
non-chiral
Here the double arrow indicates a particle - anti-particle interchange. In the last line
the representation (10) of SO(10) appears. This is the vector representation, and it is
real: (10) = (10). Hence a four-dimensional fermion in such a representation can never
be chiral.
105
Chiral particles cannot have a mass. The reason is that in a mass term of the form
m

the chiral projection operator is ipped:

1
2
(1 +
5
)

4
1
2
(1 +
5
) =

1
2
(1
5
)
4
= (
1
2
(1
5
))

4
(8.17)
Hence if the second has chirality +, the rst must have chirality . This implies that
one cannot write down a mass term for (27)
L
unless there is also a (27)
R
. On the other
hand, if both (27)
L
and (27)
R
are present a mass term is allowed.
An often used principle in quantum eld theory is anything that is allowed is obliga-
tory. In other words, if there is no reason why a certain parameter should be zero, then
it is reasonable to assume that it is not zero. On the basis of this principle one usually as-
sumes that if for example N
L
> N
R
, then N
R
(27)
R
s combine with N
L
(27)
L
s to acquire
a large mass (in this context large means too large to be observed), so that there are
N
L
N
R
unpaired (27)s left over. They cannot get a mass, and hence they must appear
as particles in the low-energy spectrum. By the same logic, when E
6
break to SO(10) the
leftover (10) + (1) are assumed to acquire a mass of order the symmetry breaking scale
(which must be much larger than 100 GeV). The leftover (1) in the breaking to SU(5) can
also become massive, so that nally only the standard model particles remain massless.
If we accept all this, then we must be looking for a Calabi-Yau manifold with [N
L

N
R
[ = 3. One of the best studied examples has N
L
= 1 and N
R
= 101, but thousands of
Calabi-Yau manifolds have been found, including some yielding three families. In addition
there are many variations on the basic idea. These allow breaking E
6
to something closer
to the standard model gauge group, while at the same time lowering the number of
families. Although there are many examples that get rather close to the standard model,
there is none that really ts like a glove. On the other hand, nothing like a systematic
search has been possible so far due to the overwhelming number of possibilities.
Furthermore it is dicult to do detailed calculations. For example, one may wish
to compute the quark and lepton masses. They get their masses from a three-point
coupling to the Higgs scalar,
x

x
, where x denote a particle, and
x
is a coupling
constant. This is called a Yukawa coupling. The mass of a particle is equal to the vacuum
expectation value of the Higgs a universal parameter with the dimension of mass
times
x
. The values of
x
can be computed in string theory, but the problem is that the
massless spectrum typically contains a few hundred scalars, so that it is not clear which
one is going to play the role of the Higgs scalar . In addition the Higgs potential and
the whole symmetry breaking mechanism is poorly understood within string theory. One
has tried to remedy this by adding some eld-theoretic assumptions, but pretty soon any
predictive power is completely lost. It seems that in making contact with reality one still
has a long way to go.
However, it is encouraging that the gross features of the standard model are repro-
duced, and that they come out essentially automatically from a theory that in principle
could only hope to be a theory of gravity. These features are
Chiral fermions
106
A suciently large gauge group
The correct family structure
Replication of families
8.7 Other ten-dimensional Heterotic strings

It turns out that one can nd other solutions to the conditions of modular invariance,
by dropping the requirement of separate modular invariance in the left- and right-moving
sectors. An example is
1
4
[()( )]
8
( Im )
4
_
(F
4
3
F
4
4
)
R
_
(F
8
3
+ F
8
4
)
2
+ (F
8
2
F
8
1
)
2

L
+(F
4
3
+ F
4
4
)
R
_
(F
8
3
F
8
4
)(F
8
1
+ F
8
2
) + (F
8
1
+ F
8
2
)(F
8
3
F
8
4
)

L
+(F
4
2
+ F
4
1
)
R
_
(F
8
3
F
8
4
)
2
+ (F
8
2
+ F
8
1
)
2

L
+(F
4
2
F
4
1
)
R
_
(F
8
3
+ F
8
4
)(F
8
1
F
8
2
) + (F
8
1
F
8
2
)(F
8
3
+ F
8
4
)

L
_
(8.18)
Here the right-moving factors (with subscript R) have an implicit argument whereas the
left-moving ones (L) have argument . In the left-moving sector the fermions are grouped
into two groups of 16, and e.g. in (F
8
3
F
8
4
)
2
the rst factor refers to the rst group of 16
and the second factor to the second.
From the partition function one reads o that the gauge group is SO(16) SO(16),
and is not extended, and that there are no gravitinos, and hence no supersymmetry. Fur-
thermore there is no tachyon. This shows that it is not necessary to have supersymmetry
to get rid of the tachyon.
There are six other non-supersymmetric heterotic string theories in ten dimensions,
but the other six all have a tachyon.
8.8 Four-dimensional strings

The main idea behind four-dimensional strings is to drop the requirement that all right-
moving fermions carry a space-time index . For conformal invariance we still need to
have a total of 10 fermions, but we can divide them up into four space-time and six
internal fermions. Of course we do the same with the 10 right-moving bosons. In the
left sector we have 26 bosons, which were previously divided into 10 space-time and 16
internal. Instead we choose now 4 space-time and 22 internal.
It should be clear from the foregoing discussion that a group of 2N internal fermions
gives rise to an SO(2N) symmetry if and only if all 2N fermions always have the same
boundary conditions. The same is true for fermions with a space-time index. The reason
the theories constructed so far are ten-dimensional strings is the fact that in light-cone
gauge there is an SO(8) symmetry: all partition functions depend on (F
i
( ))
4
. If instead
we allow combinations like F
i
1
F
i
2
F
i
3
F
i
4
with dierent labels i
m
, then clearly the SO(8)
107
symmetry is broken, in general to SO(2)
4
(one SO(2) for each pair of fermions). If we
associate the rst factor F
i
1
with the light-cone components of the fermions

, then the
resulting string theory is four-dimensional.
The diculty in constructing such a theory is to construct a modular invariant par-
tition function out of the internal degrees of freedom. Concretely, suppose one the right
the 10-dimensional space-time index is split into (, M), M = 1 . . . , 6 and on the left the
26-dimensional space-time index is split into (, I), I = 1 . . . , 22, then the internal degrees
of freedom are
M
, X
M
and X
I
. Now the internal degrees of freedom are partly bosons
and partly fermions, and it is dicult to build modular invariant partition functions in
such a hybrid system. The solution is bosonization: in 2 dimension one (compactied)
boson can be replaced by a pair of fermions or vice-versa. Hence we can either replace
M
by 3 bosons and construct a modular invariant partition function out of free bosons (this
leads to a description in terms of Lorentzian even self-dual lattices) or replace X
M
and
X
I
by 12 and 44 fermions. Then one proceeds as above, by trying to construct modular
invariant combinations of fermion partition functions. A technical complication in both
cases is how to preserve the two-dimensional supersymmetry that relates
M
and X
M
.
We will not give further details here. Both methods work, and yield huge numbers of
solutions. So many solutions that a complete listing is impossible. In any case one can
get chiral and non-chiral theories, a large number of possible gauge groups, theories with
and without tachyons and with any allowed number of supersymmetries.
It may seem that there are two dierent ways to get to four dimensions: compactify a
ten-dimensional string, or construct a string theory directly in 4 dimensions. It turns out
that these two methods are closely related. In many examples a given four-dimensional
string theory yields exactly the same spectrum as a denite Calabi-Yau compactication.
It is believed that then not only the spectrum but also all interactions are the same: the
two theories are identical.
There are other methods for constructing strings in four dimensions, but they are all
(believed to be) connected to each other. The full set of four-dimensional theories is a
complicated parameter space that is still to a large extent unexplored.
9 Dualities
Also in ten dimensions there are connections among seemingly dierent string theories.
The theories involved in these dualities are the ten-dimensional supersymmetric string
theories: they include the type-IIA, type-IIB, heterotic SO(32) and heterotic E
8
E
8
the-
ories introduced in the previous chapter. Before discussing dualities, we have to introduce
one more family member, namely
9.1 Open strings
All theories seen so far are oriented closed strings. There exists also a theory of unoriented
open and closed strings in ten dimensions that satises all known consistency conditions.
108
Just as one of the heterotic strings, this theory has N = 1 supersymmetry and an SO(32)
gauge group. It is called the type-I string. The reason why SO(32) appears is rather
dierent than in the case of heterotic strings. In that case the reason was modular
invariance.
9.2 Finiteness
As we have seen, modular invariance for closed strings allowed us to restrict the one-loop
integration to one modular domain, depicted in section (5.9). Furthermore we have
seen that this domain excludes the dangerous Im = 0 limit where the eld theory
divergencies are. Then, if there is no tachyon, the one loop integral is nite. Indeed,
the two supersymmetric heterotic strings have nite one-loop integral, and so does the
non-supersymmetric SO(16) SO(16) string.
Furthermore the one-loop integral (without external legs) is not just nite, but zero for
superstrings, because the partition function (the integrand) vanishes due to the abstruse
identity. This is important for higher loop divergencies.
Consider for example the two-loop diagram, the double torus. The modular integration
for the double torus contains a limit where the double torus looks like this
One may show that in the limit where the size of the tube connecting the tori goes to zero
and the length of that tube goes to innity, only one particle contributes: the dilaton. In
that limit the amplitude factorizes into two so-called tadpole diagrams.
A tadpole is a diagram with a single external line. Four-momentum conservation requires
that the four-momentum into that line vanishes: p

= 0. Then p
2
= M
2
= 0, so this
must be a massless particle. Lorentz-invariance requires that it must be a scalar. Then
it can only be the dilaton.
It can be shown that if the one-loop diagram without external lines give a contribution
, then the dilaton tadpole is proportional to . The the limit of the double torus diagram,
treated as a eld theory Feynman diagram with one propagator ending on both sides on a
one-point vertex, yields something like [1/p
2
]. However, momentum conservation tells
109
us that p
2
= 0. This implies that the double torus contribution becomes innite, unless
= 0. In that case this simple argument is inconclusive.
Although the discussion here was not very precise, the argument is indeed correct.
The SO(16) SO(16) heterotic string has a nite, non-zero . Therefore it is nite at
one loop, but divergent at two loops. The supersymmetric heterotic strings have = 0
at one loop. It turns out that the full two-loop integral again yields zero, which does not
follow from the foregoing discussion, but is consistent with it. In fact the corresponding
integral vanishes at any loop order. Diagrams with external lines do not vanish, but give
a nite answer at any loop order (or at least that is what is generally believed, but only
partly proved).
In the case of open and unoriented strings the same arguments apply to the closed
sub-sector, i.e. the torus and higher order generalizations, but the open and non-oriented
diagrams are a separate story. At the lowest non-trivial order, Euler number 1, there two
surfaces, characterized by b = 1 and another by c = 1 (see eq. (6.12)). To both surfaces
one can attach an external dilaton, so that one obtains two dilaton tadpole diagrams (disk
and crosscap)
The point is now that the contributions of these two dilaton tadpoles must cancel. If not,
then they lead to a divergence at Euler number zero due to a dilaton propagator in an
intermediate line. At Euler number zero the relevant diagrams are annulus, Klein bottle
and Mobius strip. It turns out that the integrations corresponding to these diagrams
have limits that look like a dilaton propagation between two disks (annulus), between
two crosscaps (Klein bottle) and between a disk and a crosscap (Mobius strip). Summing
these diagrams we get, schematically
D[
1
p
2
]D + C[
1
p
2
]C + C[
1
p
2
]D + D[
1
p
2
]C = (D + C)[
1
p
2
](D + C) . (9.1)
The dangerous 1/p
2
divergence cancels if D + C = 0.
Direct computation shows that in suitable units C = 32. It turns out that each
boundary can appear with a certain multiplicity factor N, the Chan-Paton multiplicity
that was introduced before. Then, in the same units, the disk contribution D is equal to
N, and the tadpoles cancel if N = 32. In that case one nds precisely the gauge bosons
of SO(32) in the spectrum.
110
It should be emphasized that this argument is not as nice as the one used for heterotic
strings. There niteness was a consequence of a sound principle, modular invariance,
whereas here it has to be imposed by hand.
We will not discuss the SO(32) open string theory in more detail here, except to
remark that the massless spectrum is identical to that of the heterotic SO(32) string. In
particular both have N = 1 supersymmetry. However, the massive spectra are dierent.
For example, the heterotic string has massive states in the SO(32) spinor representation,
whereas the open string theory only has symmetric and anti-symmetric tensors in its
spectrum.
9.3 Dual models
The term duality is very frequently used in string theory to describe a variety of phe-
nomena. The rst use of the term goes back to 1968, and is related to the historical origin
of string theory.
This notion of duality occurs in four point functions at tree level. For example, consider
two particles that scatter by the exchange of some other particle, as in the following
diagram:
Two-to-two particle scattering is kinematically described in terms of the Mandelstam
variables s, t and u:
s = (p
1
+ p
2
)
2
, t = (p
2
+ p
3
)
2
, u = (p
1
+ p
3
)
2
, (9.2)
where all momenta are assumed to be incoming (in other words, the nal state particles
have outgoing momenta p
3
and p
4
). If the exchanged particle has spin 0, this particular
diagram has an amplitude of the form
A = g
2
1
t M
2
(9.3)
where g is a coupling constant and M the mass of the exchanged particle.
111
String theory was born by considering the scattering of hadrons by exchanging mesons.
If one considers the exchange of several mesons of dierent spins J, the amplitude gener-
alizes to
A
t
(s, t) =

J
g
2
J
s
J
t M
2
J
(9.4)
One may also consider annihilation diagrams of the form
This yields an amplitude of the form
A
s
(s, t) =

J
g
2
J
t
J
s M
2
J
(9.5)
It was observed empirically that in hadronic scattering A
s
(s, t) A
t
(s, t). This led
Veneziano in to a formula that explicitly had this s-t duality built in:
A(s, t) =
((s))((t))
((s) (t))
(9.6)
with
(s) = (0) +

s (9.7)
The function satises
(u + 1) = u(u) , (1) = 1 (9.8)
so that (u) = (u + 1)! for positive integer arguments. For negative arguments there are
singularities at u = 0, 1, 2, . . ., and near u = n the behavior is
(u)
(1)
n
n!(u + n)
(9.9)
We see then that the Veneziano amplitude has poles if (s) = n or (t) = n, or
s =
1

(n (0)) , t =
1

(n (0)) (9.10)
If one interprets these poles as particles exchanged in the t-channel (t = M
2
) or the
s-channel (s = M
2
), then we recognize here precisely the open string spectrum (3.63)
(apart from multiplicities and spin-dependence, which with a more careful analysis also
112
come out correctly). The parameter (0) is called the intercept and plays the role of the
vacuum energy subtraction a. The parameter

is called the Regge slope (this explains

nally why the fundamental length scale of string theory was called

, and in particular
why there is a prime).
The foregoing remarks suggest that the Veneziano amplitude has a string theory in-
terpretation, and indeed it does. Pictorially the observation in the previous paragraph
is
Although it is beyond the scope of these lectures, the four-point string diagram is not
hard to compute, and yields indeed the Veneziano amplitude. In special limits (namely
near the poles in s or near the poles in t) this single amplitude behaves at the same time
as sum of an innite number of f-channel exchanges, or as the sum of an innite number
of s-channel annihilations. It is historically rather strange that rst the amplitude was
written down, and only afterwards its origin was understood, but that is indeed what
happened.
Note that by deforming the open string tree diagram one can get the s and t diagrams
but not the u diagram
113
since we cannot interchange external lines. However, if we work with closed instead of
open strings we can interchange all lines, since they are simple four tubes attached to
a sphere. The tubes can be moved around each other on the sphere. Consequently the
four-point amplitude for closed strings has a symmetry in s, t and u, and has the form
A(s, t, u) =
(
1
4
(s))(
1
4
(t))(
1
4
(u))
(
1
4
(s)
1
4
(t)
1
4
(u))
(9.11)
The factors
1
4
reect the factor of 4 dierence in spacing between the open and closed
string spectra.
9.4 Historical remarks
After the discovery of the Veneziano amplitude and the realization that it had a string
interpretation there was a short but intense period of trying to develop it as a strong in-
teraction model. These models were called dual models, because of the duality property
described above. These models ran into serious diculties when the tachyon was found
and the critical dimension turned out to be 26.
Around 1972 QCD was formulated as a theory of the strong interactions, and it quickly
gained in popularity over dual models. Meanwhile fermionic strings were constructed, and
in the early 70s, completely independently of string theory, supersymmetry and super-
gravity were discovered. In 1976 Gliozzi, Scherk and Olive formulated their projection
on the fermionic string spectrum, which connected these developments for the rst time:
superstrings were born, and the supergravity theories appeared as their low-energy limits.
An important shift of perspective took place when it was realized that the innite
tower of string states contained a massless particle of spin 2 (a symmetric tensor in 10 or
26 dimensions). It was in particular Scherk who pushed the idea the string theory should
be considered as a theory of gravity rather than the strong interactions. But after 1975 the
interest in string theory diminished rapidly, and all attention turned to the development
of the standard model, which was going through an exciting period.
In 1984 the second string revolution began, after an almost ten-year period of near
inactivity. The Heterotic strings were constructed, and the E
8
E
8
theory looked espe-
cially promising for phenomenology. In the years that followed there was a huge outburst
114
of activity. Calabi-Yau compactications were studied, and many other compactications
and four-dimensional string constructions were investigated.
After another period of relatively little activity (although much more than between
1975-1984) in 1994 the third string revolution occurred. This was sparked by the
discovery of a set of string dualities, although a rather dierent kind of dualities than
those discussed above.
9.5 T-duality

In the section on bosonic string torus compactication we saw that the compactied
string has a remarkable duality under R

/R. Similar dualities can be studied for the

compactied type-II and heterotic strings. If we compactify on a circle to 9 dimensions,
it turns out that the two compactied type-II strings are mapped into each other, and
that the two heterotic strings are also mapped to each other. We will not study this in
detail, but only give a rough idea how this works.
The two type-II strings only dier in the relative chirality of the left and right GSO
partition functions. However, chirality is lost if one make a torus compactication, as we
have seen in the previous chapter. Therefore it is not a big surprise that the dierence
between IIA and IIB is lost upon compactication.
For the Heterotic compactication we can be more explicit. A general compactication
of one dimension can be described by a
1,1
Lorentzian even self-dual lattice. If we combine
this with the 16-dimensional lattice in the left sector (using the bosonic description of
that sector) we get a Lorentzian even self-dual lattice
17,1
. A general theorem on such
lattices says that they are all related by Lorentz transformations. It can be shown that
in string theory these Lorentz-transformations are parametrized by the moduli of the
compactication torus, which are massless scalar elds. Giving these scalars a vacuum
expectation value we move from one
17,1
lattice to another, and in particular we can
move from E
8
E
8

1,1
to D
16

1,1
. This shows that these two theories are on the
same moduli space, and hence they are clearly related. That the relation is actually a
T-duality requires a bit more discussion.
9.6 Perturbative and non-perturbative physics

String theory denes a perturbation series. Any amplitude comes out as an innite sum
of the form

n
a
n
g
n
, where g is the string coupling, related to the exponential of the
vacuum expectation value of the dilaton.
It is not obvious that this is all there is. Suppose the exact answer is of the form
F(g) =

n
a
n
g
n
+ Ae

1
g
2
(9.12)
The second term has the property that all its derivatives at g = 0 vanish. Hence its Taylor
expansion around g = 0 is trivial. One can never see such a term in perturbation theory.
115
Therefore, no matter how far we expand, we will never be able to compute the coef-
cient A from a loop diagram. Eects that behave like this are called non-perturbative
eects. It is known that such contributions occur, for example, in QCD.
In QCD the appearance of such eects is related to the existence of classical solutions
to the equations of motion (without sources). The rst terms in (9.12) are coming from
an expansion of the action around the trivial classical solution (with all elds equal to
zero), whereas the second term comes from an expansion around the non-trivial solution
(there are additional perturbative corrections to the last term). In string theory it is a
priori much more dicult to nd such classical solutions. This is because in string theory
we only have a perturbative expansion.
To understand this note that in QCD the Feynman diagram are derived from an action,
and that action can also be used to derive classical equations of motion that can be solve.
In string theory we only have the Feynman diagrams, but no action to derive them
from (we would need an action in target space, not on the world-sheet).
One way to get around this is to derive from string theory a low-energy eective action.
This is an action for the massless elds only, such that the interactions derived from it
are the same as those derived directly from string theory. This action includes of course
the Einstein action for gravity, as well as the usual actions of the gauge bosons. Then one
can solve the equations of motion of this low-energy eective action, and postulate that
it is the rst order approximation to a genuine string theory classical solution.
Classical solutions to eld theory equations of motion can be classied according to
their extension in space. A solution can be point-like, meaning that its energy density is
concentrated around a point in space; analogously it can be string-like or membrane-like.
In general we call a solution a p-brane if it has p spatial dimensions. Hence a particle-like
solution is a 0-brane, a string-like solution a 1-brane, etc.
9.7 D-branes

There exists an exception to the statement that we cannot describe classical solutions
exactly in string theory. These are the so-called D-branes (where D stands for Dirichlet).
As we have seen, open strings can have two kinds of boundary condition. A Neumann
boundary condition has the property that

X vanishes at the boundary. This amounts to

conservation of momentum at the boundary. For a Dirichlet boundary condition

X = 0
at the boundary. This means that X(0) and X() are xed as a function of time.
One may consider situations where for the bosonic string D space-like coordinates
have Dirichlet boundary conditions, and the remaining 26 D have Neumann boundary
conditions. Then we have
X
i
(0) = v
i
, X
i
() = w
i
, i = 1, . . . , D (9.13)
where v
i
and w
i
are some xed vectors. The remaining space-like directions are not
xed. Below we illustrate this for a total of 3 dimensions instead of 26. We plot the two
space-like dimensions x
1
and x
2
116
D1-branes D0-branes
In the rst picture we have Dirichlet boundary conditions in the x
1
direction. Hence the
open string has its endpoints xed to two lines at xed x
1
. The endpoints are allowed to
move in the x
2
direction. In the second picture we have Dirichlet boundary conditions in
both directions. Then the endpoints are completely xed to points in space. Finally we
may take Neumann conditions in all directions. Then the endpoints are free to move in
all space directions. This is what we normally mean by open strings.
In the rst two cases translation invariance is broken in one or two directions. This
can be understood in terms of objects that are present in space. A particle xed at
some position obviously breaks translation invariance in all directions. A wall breaks
translation invariance in the direction orthogonal to it. Therefore we interpret the rst
picture in terms of two lines, or 1-branes, whose presence breaks translation invariance,
and the second picture is interpreted in terms of two points or 0-branes. The space-time
momentum of modes on the string is not conserved because it can be transferred to these
objects. Because their presence is due to Dirichlet boundary conditions these objects are
called D-branes or D
p
-branes: they are p-branes dened by the property that open strings
can end on them.
This only makes sense if we view the D-branes as dynamical objects, that can them-
selves move around and vibrate. This point of view is indeed correct. It turns out that
these D-branes are the exact string analog of certain approximate eld theory solutions.
There may be additional solutions, but those are harder to describe.
Note that the D-branes are objects that are not obtained in string perturbation theory.
There are precise rules for which kind of D-branes can occur in which string theories.
There are no D-branes in heterotic strings, but they do exist in type-II strings. This may
look surprising at rst, since type-II strings were constructed as closed string theories.
However, that was a perturbative description. It turns out that these theories nevertheless
contain non-perturbative states that are described by open strings.
The rules are that type-IIA theories can have D
even
-branes whereas type-IIB theories
can have D
odd
-branes. Of special interest are the D
0
-branes in type-IIA and the D
9
branes in type-IIB. Let us start with the latter. A D
9
brane lls all of space, and the
open strings ending on it have endpoints that can move freely through space. Such strings
are traditional open strings with only Neumann conditions, and no Dirichlet conditions.
The type-IIB string either has no D
9
branes at all, or it must have exactly 32. In the
latter case one is forced to allow also dierent world-sheet orientations, and one is led to
117
the type-I string. One can assign a label to each of the 32 branes. This corresponds to
the Chan-Paton labels. From this point of view the type-I string lives in a space built out
of 32 D
9
branes stacked on top of each other.
9.8 S-duality

The D
0
-branes of type-IIA correspond to extra, non-perturbative particle-like states in
the spectrum. These states turn out to carry a charge with respect to the massless vector
boson A

in the perturbative spectrum. They are the only states that are charged, because
it is easy to show that all perturbative states are neutral.
One may also compute the masses of the D
0
particles, and then one nds that they
are
|q|
g
, where q is the charge and g the string coupling constant.
Witten made the observation that these particles could be interpreted as Kaluza-Klein
modes of a circle compactication of 11-dimensional supergravity.
The latter theory has two bosonic elds: a graviton, and a rank-3 anti-symmetric
tensor. Upon compactication (as we have seen in section (3.6)) the graviton yields a
10-dimensional graviton, a photon and a scalar. The rank-3 tensor yields a rank-3 tensor
and a rank-2 anti-symmetric tensor. These are precisely the bosonic elds of the ten-
dimensional type-II string at the massless level.
But Kaluza-Klein compactication yields more than that. If we expand any eld
in modes, we get also massive excitations. These excitations turn out to carry charge
with respect to the photon, and turn out to match precisely the D
0
-particle spectrum.
Furthermore the masses of these excitations are inversely proportional to the radius of
the circle: the smaller the circle, the more energy it costs to excite a mode in the compact
direction.
This suggests to identify the radius of the circle with the strength of the type-IIA
coupling constant. At small coupling constant the masses of the Kaluza-Klein modes or
D
0
-particles become large and can be neglected; this correspond to the D = 11 theory
compactied on a very small circle. In the limit of large coupling the D
0
particles become
massless, and the radius goes to innity. In this limit we approach 11 at dimensions.
The conclusion which is supported by other arguments is that the strong coupling
limit of the type-IIA theory is an 11-dimensional theory. This is denitely not a string
theory. It is also denitely not 11-dimensional supergravity: that theory has innities in
its perturbation theory, which should be absent if it is the strong coupling limit of a nite
string theory. The speculation is now that it is an unknown theory, called M-theory
whose low-energy limit is D = 11 supergravity.
One can also consider the strong coupling limit of all other string theories and one
nds
Heterotic SO(32) Type-ISO(32)
Type-IIB Type-IIB
Heterotic E
8
E
8
11-dimensional supergravity on S
1
/Z
2
(9.14)
The rst duality requires the existence of SO(32) spinors in the spectrum of the type-I
118
string. They can only because of non-perturbative eects. The second case is, for obvious
reasons, called self-duality. The strong coupling limit of the E
8
E
8
heterotic string is
D = 11 supergravity compactied on a circle with mirror image points identied to each
other, as shown here:
The resulting space is a line segment, so that the 11-dimensional space becomes two planes
separated by the length of the segment. The two E
8
factors can be associated with the
two planes.
There is a lot more to say about dualities in 10 and 11 dimensions, and even more in
less than 10 dimensions. But we will stop here and conclude with the duality picture that
summarizes the present understanding.
S
T
T
S
S
119
A Functional methods
In classical eld theory we deal with dynamical systems with an innite number of coor-
dinates q
i
. For example, suppose we want to describe the classical motion of gauge elds
A

(x, t) in electrodynamics. Then the coordinates that enter into the familiar Lagrangian
or Hamiltonian description are A

(x).
An additional complication is that these coordinates are continuously innite: the
label i of q
i
gets replaced by the set (, x). This implies that summations become
integrations, and dierentiations are replaced by functional derivatives.
Let us rst review the kind of formulas we want to generalize to classical eld theory.
A system with a nite number of degrees of freedom is usually described by a Lagrangian
L(q
i
, q
i
). To every path with initial time t
0
and nal time t
1
one associates an action
S =
_
t
1
t
0
dtL(q(t), q(t)) (A.1)
Hamiltons principle states that the classical trajectory followed by the system is given by
an extremum of the action: S[q
i
(t)] = 0. From this one derives the equations of motion,
the Euler-Lagrange equations,
d
dt
_
L
q
i
_

_
L
q
i
_
= 0 . (A.2)
To quantize such a theory one denes canonical momenta
p
i
=
L
q
i
, (A.3)
and a Hamiltonian
H =

i
p
i
q
i
L . (A.4)
For two classical quantities a and B one may dene a Poisson bracket
A, B
PB
=

_
A
q

B
p

A
p

B
q

_
(A.5)
To go to the quantum theory one replace A and B by by Hilbert space operators, and
furthermore one replaces the Poisson brackets A, B
PB
by the commutator
1
i
[A, B]. For
A = q
i
and B = p
j
this procedure leads to the well-known Heisenberg relation
[q
i
, p
j
] = i
ij
. (A.6)
A classical eld theory is given by a set of coordinates
a
(x), where x is a space-time
point and a some additional discrete label, for example a space-time vector index or an
index of some internal symmetry of the theory. The dynamics is governed by a Lagrangian
L(
a
(x, t),

a
(x, t) (A.7)
120
from which one can compute the action as S =
_
dtL. To write down the Euler-Lagrange
equations one needs a functional derivative. A functional is an object that assigns a
number to a function. To only example we will ever encounter is an integral. A functional
derivative can be viewed as the derivative of the functional with respect to the function.
For example, consider a function (x) and the functional
F[] =
_
d
3
xf((x)) (A.8)
This should be though of as a continuous version of
F(q
i
) =

i
f(q
i
) (A.9)
To dierentiate this, all we need is the rule
q
i
q
j
=
ij
(A.10)
In the present case this generalizes to
(x)
(y)
= (x y) , (A.11)
where we use a instead of to distinguish a functional derivative from an ordinary one.
Apart from the notation, the rules for functional dierentiation are quite similar to those
of ordinary dierentiation. For example, one may use the chain rule, and consequently
the functional derivative of F is
F[]
(y)
=
_
d
3
xf

((x))
(x)
(y)
=
_
d
3
xf

((x))(x y)
= f

((y)) .
(A.12)
The Euler-Lagrange equations are in this case
d
dt
_
L

a
(y)
_

_
L

a
(y)
_
= 0 . (A.13)
This is not the most convenient form, since in Lorentz covariant eld theories time deriva-
tives appear in combination with space derivatives, as

. The symmetry between space

and time is restored by dening a Lagrangian density, which is simply the integrand of
the Lagrangian
L =
_
d
3
xL (A.14)
121
In terms of L the Euler lagrange equations are

_
L
(

)
_

= 0 , (A.15)
These are ordinary derivatives, because we see L as a function (not a functional) of and

. This version of the Euler-Lagrangian can either be derived from (A.13), or directly
from the variation of S =
_
d
4
xL.
Canonical momenta are dened as

a
(x) =
L

a
(x)
(A.16)
A bit more care is needed when there is a distinction between upper and lower indices. If
g
ab
is a symmetric matrix and the action depends on combinations
a
g
ab

b
it is customary
to dene
a
=

b
g
ab

b
. Indices can be raised by using g
ab
, which is the inverse of g
ab
:

b
g
ab
g
bc
=
c
a
. Suppose we have a functional F =
_
d
3
xV
c

c
(x), with V
c
a constant
vector. Then
F

c
(x)
= V
c
, (A.17)
using

a
(x)

b
(x)
=
a
b
(x y) (A.18)
One see this more explicitly by leaving all indices upper, and writing the matrix g
ab
explicitly. Hence the derivative with respect to a quantity with an upper (lower) index
yields a quantity with a lower (upper) index.
This applies also to canonical momenta and Poisson brackets. The canonical momen-
tum of
a
is then (if there is a distinction between upper and lower indices)

a
=
L

a
(x)
(A.19)
The Hamiltonian is
H =
_
d
3
x

a
(x)

a
(x) L (A.20)
and canonical quantization yields
_

a
(x),
b
(y)

= i
b
a
(x y) (A.21)
Often one converts the rst index to an upper index, and then the result is of course
_

a
(x),
b
(y)

= ig
ab
(x y) (A.22)
122
A.1 The Euclidean action
The Euclidean action is obtained by a continuation from Minkowski space, by making
time imaginary. This could mean either t = it
E
or or t = it
E
. The choice between
these two options is determined by the requirement that the analytic continuation should
make the path integral a damped exponential. Consider the weight factor in the path
integral, exp (iS). In this expression S =
_
dtL, and we consider a simple Lagrangian
of the form
1
2
q
2
V (q), so that the Hamiltonian is H =
1
2
p
2
+ V (q). Once we know the
correct continuation for this action, we know it for all action.
If the potential V is bounded from below, as it should in a physical system, then
clearly the correct denition for the potential term alone is t = it
E
. If the Minkowski
time integral is from 0 to T we get now
iS = i
_
T
0
dt(
1
2
(
dq
dt
)
2
V (q))) =
_
T
E
0
dt
E
(
1
2
(
dq
dt
E
)
2
V (q)) (A.23)
with T = iT
E
. We see that both the kinetic and the potential term give a negative
exponential. It is conventional to choose the Euclidean action in such a way that the
path-integral is dominated by its minima (and not its maxima). Hence one denes
exp (iS(t = it
E
)) exp (S
E
) , (A.24)
with S
E
a positive denite expression. In our case
S
E
=
_
T
E
0
dt
E
(
1
2
(
dq
dt
E
)
2
+ V (q)) (A.25)
Note that this denition contains an extra minus sign compared to the naively expected
relation, and that in particular the V (q) contribution appears with opposite sign in S and
S
E
:
S
E
=
_
T
E
0
dt
E
L
M
(t = it
E
) , (A.26)
where L
M
is the Lagrangian in Minkowski space.
B Path integrals
The derivation of (6.22) goes as follows (this discussion is based on appendix A of [2])).
We consider a quantum mechanical system with one coordinate q and a momentum p.
The corresponding quantum operators q and p satisfy [ q, p] = i. Furthermore we dene
eigenstates [q) (with q [q) = q [q)) and [p) (with p [p) = p [p)), and we have
q[p) = e
ipq
q[q

) = (q q

)
p[p

) = 2(p p

)
(B.1)
123
The integration measures for position and momentum integral are respectively
_
dq [q) q[ = 1;
_
dp
2
[p) p[ = 1. (B.2)
We divide the time-interval T into N equal steps , T = N, and dene t
m
= m. At
t = 0 a complete set of states is [q). We dene a set of states [q, t) as
[q, t) = e
iHt
[q) (B.3)
This introduces a complete set of states at every time t
m
. In general we have the com-
pleteness relation
A[B) =
_
dqA[q, t)q, t[B) . (B.4)
Applying this to each intermediate time t
m
yields for a transition amplitude
q
f
, T[q
i
, 0) =
_
dq
1
. . . dq
N1
N1

m=0
q
m+1
, t
m+1
[q
m
, t
m
) (B.5)
Now we express all [q, t) in terms of [q) [q, 0) using (B.3):
q
m+1
, t
m+1
[q
m
, t
m
) = q
m+1
[ e
iHt
m+1
e
iHtm
[q
m
)
= q
m+1
[ e
iH
[q
m
)
(B.6)
Now we insert a complete set of momentum states
q
m+1
, t
m+1
[q
m
, t
m
) =
_
dp
m
2
q
m+1
[p
m
) p
m
[ e
iH
[q
m
) (B.7)
The classical Hamiltonian is a function of p and q, H(p, q). The quantum Hamiltonian is
ambiguous because of the ordering of p and q. We will dene it by requiring that all p
are always to the left of q. Then
p
m
[ H( p, q) [q
m
) = H(p
m
, q
m
)q[p) (B.8)
Even if p and q are ordered correctly in H, the exponential exp (H) still contains
improperly ordered terms. However, they are always of order
2
and will be dropped.
Then we get
q
m+1
, t
m+1
[q
m
, t
m
) =
_
dp
m
2
e
iH(pm,qm)
q
m+1
[p
m
)p
m
[q
m
)
=
_
dp
m
2
e
i[H(pm,qm)pm(q
m+1
qm)]
,
(B.9)
where in the last step (B.1) was used.
124
We will only be interested in Hamiltonians of the form H =
1
2
ap
2
+ V (q), where a is
some constant. In that case the p
m
integrations in (B.9) can be done explicitly. At this
point it is better to go to Euclidean space, by dening = i. The sign choice is dictated
by the requirement that we get a damped gaussian integral. The relevant integral is
_
dp
2
exp (
1
2
ap
2
+ ipd)) =
_
dp
2
exp (
1
2
[a(p id/a)
2
+ d
2
/a])
=
1

2a
exp (
1
2
d
2
/a)) ,
(B.10)
with d = q
m+1
q
m
. Now we get
q
f
, T[q
i
, 0) =
_
N1

i=1
dq
i
_
1

2a
_
N
exp (
N1

m=0
[
1
2
(q
m+1
q
m
)
2
/a V (q
m
)])
=
_
N1

i=1
dq
i
_
1

2a
_
N
exp (
N1

m=0
[
1
2
(
q
m+1
q
m

)
2
/a V (q
m
)])
(B.11)
The limit 0 denes the path-integral as (writing T = i, so that = N)
q
f
, [q
i
, 0) =
_
Tqexp
_

0
dt[
1
2a
q
2
V (q)])
=
_
Tqexp (S
E
( q, q))
(B.12)
Here we have absorbed the prefactor into the denition of Tq. Note that the normalization
is anyhow xed by the limit 0, which should yield the normalization of the state.
The result can also be written as
q
f
, 0[e
H
[q
i
, 0) =
_
Tqexp (S
E
) (B.13)
The path integral is over all paths q(t) with q(0) = q
i
and q() = q
f
. Now we take
q
i
= q
f
q and integrate over all q
i
. Then we get
Tr e
H

_
dqq, 0[e
H
[q, 0) =
_
PBC
Tqexp (S
E
) (B.14)
The extra integral has extended the path-integral with an integral over the initial position.
We are now integrating over all paths satisfying q(0) = q() = q, with an integration over
q. Hence this is a path integral over all paths satisfying a periodic boundary condition.
C Self-dual lattices
Here we summarize a few results on even self-dual lattices.
125
Lorentzian even self-dual lattices
p,q
are dened by the requirement that all vectors
have even length, and the
p,q
is self-dual. Both length are duality are dened with an
inner product of the form
(v, w)
2
=
p

i=1
v
2
i

q

j=1
w
2
j
, (C.1)
where (v, w) is a vector on the lattice.
Lorentzian even self-dual lattices (ESDLs) are classied modulo Lorentz transforma-
tions. Up to such transformations there is just one lattice, i.e. all ESDLs are related to
each other by Lorentz transformations. Physically this information is useless. Particle
spectra depend on v
2
and w
2
separately, not just on their dierence. Only rotations in
the rst p or the last q components make no dierence. In other words, the condition for
modular invariance is Lorentz invariant, but the spectrum is not.
Since all Lorentzian ESDLs are Lorentz rotations of each other, the parameters of the
Lorentz transformation are the moduli of the compactication. The full parameter space
(moduli-space) is in fact the coset space
SO(p, q)
SO(p) SO(q)
(C.2)
This means that we consider the Lorentz-transformations in SO(p, q), but we identify
those Lorentz-transformations that are related to each other by rotations in the rst p
or the last q components. The precise moduli space is still a bit smaller: because of T-
duality and its generalizations on higher-dimensional tori some parts of the moduli space
are identied with each other.
The foregoing holds if p ,= 0, q ,= 0. If q = 0, ESDLs exist only if p is a multiple of 8.
The rst case is p = 8, and it turns out that the lattice is then unique and given by
(0, . . . , 0, 1, 0, . . . , 0, 1, 0, . . . , 0)
(
1
2
, . . . ,
1
2
)
(C.3)
This is an overcomplete basis, but in any case the idea is that one builds a lattice by
adding all these vectors in any integer linear combination. The vectors shown here have
length 2, and in total there are 240 of them: 112 of the rst type, and 128 of the second
type (note that by adding the rst set one can get all even sign changes of the second
vector). This lattice is called E
8
(it is in fact the root lattice of the Lie-algebra E
8
).
By taking this lattice for left- and rightmovers we can build a lorentzian ESDL
8,8
=
(E
8
)
L
(E
8
)
R
. The vectors of this lattice are of the form (v
L
, wv
R
) with v
L
and v
R
vectors
of the above form. On this lattice we may compactify the bosonic string to 16 dimensions.
In the resulting spectrum the states of the form
i
1
[v
L
, 0) and
I
1

i
1
[0, 0) (with v
2
L
= 2
and I = 1, . . . 8) are massless vector bosons. There are 240 + 8 = 248 such bosons from
the left-moving and the same number from the right-moving modes. It turns out that
massless vector bosons in string theory are always gauge bosons. In this case they are
gauge bosons with a Lie-algebra (E
8
)
L
(E
8
)
R
. Here E
8
is a Lie-algebra specied, just
126
as SU(2), by a set of commutation relations. We will not explain its construction here.
It has of course precisely 248 generators.
If p = 16 there are two solutions: (E
8
)
2
and a lattice one might call E
16
. It is dened as
E
8
above, but with 16-component vectors. Because of this the second vector, (
1
2
, . . . ,
1
2
)
now has length 4 and would not produce a massless gauge boson when we compactify
on
16,16
= (E
16
)
L
(E
16
)
R
. Indeed, now there are 496 massless gauge bosons from
the left and the same number from the right. They all originate from the rst type of
lattice vector in (C.3), plus the 16 oscillators
I
1
. The number of length-2 vectors is
1
2
(16 15) 4 = 480. The gauge group is in fact SO(32)
L
SO(32)
R
. The name E
16
is not standard. Usually one calls the lattice without the half-integer vectors D
16
: it is
the root lattice of SO(32). The half-integer spin vector is an SO(32) spinor weight. So
the correct description of the E
16
lattice would be the D
16
root lattice with a spinor
conjugacy class added.
For profound reasons 496 = 2248, so that both p = 16 lattices give the same number
of gauge bosons, but dierent gauge groups.
If p = 24 there are 24 ESDLs, and for p = 32 the number explodes to an unknown
number larger than 10
7
(and undoubtedly much larger)!
D Lie algebra dictionary
Here we collect some formulas and conventions for Lie-algebras. This is not a review of
group theory, but rather an encyclopedic dictionary of some relevant facts with few
explanations.
D.1 The algebra
We will mainly use compact groups (see below). Their Lie algebras can be characterized
by a set of dim (A) hermitean generators T
a
, a = 1, dim (A), where A stands for adjoint.
Provided a suitable basis choice is made, the generators satisfy the following algebra
[T
a
, T
b
] = if
abc
T
c
, (D.1)
with structure constants f
abc
that are real and completely anti-symmetric.
D.2 Exponentiation
Locally, near the identity, the corresponding Lie group can be obtained by exponentiation
g(
a
) = e
i
a
T
a
. (D.2)
The global properties of the group, involving element not close to 1, are not fully
described by the Lie-algebra alone, but will not be discussed here. The space formed by
all the group elements is called the group manifold.
127
D.3 Real forms
A Lie-algebra is a vector space of dimension dim (A) with an additional operation, the
commutator. An arbitrary element of the vector space has the form

a
T
a
. In ap-
plications to physics
a
is either a real or a complex number. If the coecients
a
are
all real and the generators Hermitean, the group manifold is a compact space. For a
given compact group there is a unique complex Lie-algebra, which is obtained simply by
allowing all coecients
a
to be complex. Within the complex algebra there are several
real sub-algebras, called real forms. The generators of such a sub-algebra can be chosen
so that (D.1) is satised with all structure constants real, but with generators that are
not necessarily Hermitean. One can always obtain the real forms from the compact real
form (which has hermitean generators) by choosing a basis so that the generators split
into two sets, H and /, so that [H, H] H and [/, /] H. Then one may consistently
replace all generators K / by iK without aecting the reality of the coecients f
abc
.
The most common case in physics are the real forms SO(n, m) of the compact real form
SO(n + m). Most of the following results hold for the compact real form of the algebra,
unless an explicit statement about non-compact forms is made.
D.4 The classical Lie groups
The group SU(N) is the group of unitary N N matrices with determinant 1; SO(N) is
the group of real orthogonal matrices with determinant 1, and Sp(2r) the group of real
2r2r matrices S that satises S
T
MS = M, where M is a matrix which is block-diagonal
in term of 2 2 blocks of the form
_
0 1
1 0
_
(D.3)
Mathematicians (and some physicist) write Sp(r) instead of Sp(2r).
D.5 Simple Lie-algebras
Lie-algebras are in general a product (or, more accurately, a direct sum) of a semi-
simple Lie-algebra and some U(1)s. The latter require no further discussion. Semi-simple
algebras are a product of various simple ones; the simple Lie-algebras have been classied
completely, see below.
D.6 The Cartan subalgebra
This is the maximal set of commuting generators of the simple algebra. All such sets can
be shown to be equivalent. The dimension of this space is called the rank (denoted r) of
the algebra.
128
D.7 Roots
If we denote the Cartan subalgebra generators as H
i
, i = 1, . . . , r, then the remaining
generators can be chosen so that
[H
i
, E

] =
i
E

. (D.4)
The eigenvalues with respect to the Cartan subalgebra are vectors in a space of dimension
r. We label the remaining generators by their eigenvalues . These eigenvalues are called
the roots of the algebra.
D.8 Positive roots
A positive root is a root whose rst component
1
is positive in some xed basis. This
basis must be chosen so that, for all roots,
1
,= 0.
D.9 Simple roots
Simple roots are positive roots that cannot be written as positive linear combinations
of other positive roots. There are precisely r of them. They form a basis of the vector
space of all the roots. The set of simple roots of a given algebra is unique up to O(r)
rotations. In particular it does not depend on the choice of the Cartan subalgebra or
the basis choice in root space. This set is thus completely specied by their relative
lengths and mutual inner products. The inner product used here, denoted

, is the
straightforward Euclidean one.
D.10 The Cartan matrix
The Cartan matrix is dened as
A
ij
= 2

i

j

j

j
, (D.5)
where
i
is a simple root. This matrix is unique for a given algebra, up to permutations
of the simple roots. One of the non-trivial results of Cartans classication of the simple
Lie algebras is that all elements of A are integers. The diagonal elements are all equal to
2 by construction; the o-diagonal ones are equal to 0, 1, 2 or 3.
D.11 Dynkin diagrams
are a graphical representation of the Cartan matrix. Each root is represented by a dot.
The dots are connected by n lines, where n is the maximum of [A
ij
[ and [A
ji
[. If [A
ij
[ >
[A
ji
[ an arrow from root i to root j is added to the line. The simple algebras are divided
into 7 classes, labelled AG , with Dynkin diagrams as shown below.
129
A
r
B
r
C
r
D
r
G
2
F
4
E
6
E
7
E
8
1 2 3 4 r-1 r
1 2 3 4 r-1
r
1 2 3 4 r-1 r
1 2 3 4 r-2
r-1
r
1 2
1 2 3 4
1 2 3 4 5
6
1 2 3 4 5 6
7
1 2 3 4 5 6 7
8
D.12 Long and short roots
If a line from i to j has an arrow, A
ij
,= A
ji
and hence the lengths of roots i and j are not
the same. An arrow points always from a root to another root with smaller length. Lines
without arrows connect roots of equal length. There is at most one line with an arrow
per diagram, and therefore there are at most two dierent lengths. This is not only true
for the simple roots, but for all roots. One frequently used convention is to give all the
long roots length-squared equal to two. Then the short roots have length 1 if they are
connected to the long ones by a double line, and length-squared
2
3
if they are connected
by a triple line. Often the short roots are labelled by closed dots, and the long ones by
open dots, although this is strictly speaking superuous. If all roots have the same length
the algebra is called simply laced. This is true for types A,D and E.
130
D.13 Realizations
The compact Lie-algebras corresponding to types A D are realized by the algebras
SU(n), SO(n) and Sp(n). The correspondence is as follows
A
r
:SU(r + 1)
B
r
:SO(2r + 1)
C
r
:Sp(2r)
D
r
:SO(2r)
(D.6)
There is no such simple characterization for the algebras of types E, F and G, the excep-
tional algebras.
D.14 Representations
A set of unitary N N matrices satisfying the algebra (D.1) is said to form a (unitary
matrix) representation of dimension N.
D.15 Equivalence
If a set of hermitean generators T
a
satisfy the algebra, then so do

T
a
= U

T
a
U, if U is
unitary. Then T
a
and

T
a
are called equivalent.
D.16 Real, complex and pseudo-real representations
The complex conjugate representation is the set of generators (T
a
)

, which obviously
satisfy the algebra if T
a
does. A representation is real if a basis exists so that for all
a (T
a
)

= T
a
(in other words, if a

T
a
= U

T
a
U exists so that all generators are
purely imaginary). An example of a real representation is the adjoint representation. A
representation is pseudo-real if it is not real, but only real up to equivalence, i.e. (T
a
)

=
C

T
a
C for some unitary matrix C. Otherwise a representation is called complex.
A frequently occurring example of a pseudo-real representation is the two-dimensional
one of SU(2). The generators are the Pauli matrices, and only
2
is purely imaginary.
However, if one conjugates with U = i
2
the other two matrices change sign, so that
indeed

i
= U

i
U.
D.17 Irreducible representations
If a non-trivial subspace of the vector space on which a representation acts is mapped
onto itself (an invariant subspace) the representation is called reducible. Then all Ts
can be simultaneously block-diagonalized, and each block is by itself a representation. If
there are no invariant subspaces the representation is called irreducible.
131
D.18 Weights
In any representation the matrices representing the Cartan sub-algebra generators H
i
can
be diagonalized simultaneously. The space on which the representation acts decomposes
in this way into eigenspaces with a set of eigenvalues

, i.e.H
i
v

=
i
v

. The

s, which
are vectors in the vector space spanned by the roots, are called weights. The vector space
is usually called weight space.
D.19 Weight space versus representation space
We are now working in two quite dierent vector spaces: the r-dimensional weight space,
and the N dimensional space on which the representation matrices act. The former is a
real space, the latter in general a complex space. Often the vectors in the latter space are
referred to as states, a terminology borrowed from quantum mechanics. Although this
may be somewhat misleading in applications to classical physics, it has the advantage of
avoiding confusion between the two spaces.
D.20 Weight multiplicities
In the basis in which all Cartan sub-algebra generators are simultaneously diagonal each
state in a representation are characterized by some weight vector . However, this does
not characterize states completely, since several states can have the same weight. The
number of states in a representation R that have weight is called the multiplicity of
in R.
D.21 Coroots
Coroots are dened as =
2

.
D.22 Dynkin labels
For any vector in weight space we can dene Dynkin labels l
i
as l
i
=
i
= 2

i

i
.
Since the simple (co)roots form a complete basis, these Dynkin labels are nothing but
the components of a weight written with respect to a dierent basis. The advantage of
this basis is that it can be shown that for any unitary representation of the algebra the
Dynkin labels are integers.
D.23 Highest weights
Every irreducible representation of a simple Lie-algebra has a unique weight so that
on the corresponding weight vector E

= 0 for all positive roots . Then is called

the highest weight of the representation. Its Dynkin labels are non-negative integers.
Furthermore for every set of non-negative Dynkin labels there is precisely one irreducible
representation whose highest weight has these Dynkin labels.
132
D.24 The irreducible representations of a simple Lie-algebra
can thus be enumerated by writing a non-negative integer next to each node of the Dynkin
diagram. The states in a representation can all be constructed by acting with the gener-
ators E

on the highest weight state. This state always has multiplicity 1.

D.25 Special representations
Fundamental representations
The representations with Dynkin labels (0, 0, , . . . , 0, 1, 0, . . . , 0) are called the fun-
damental representations.
The adjoint representation
The adjoint representation is the set of generators (T
a
)
bc
= if
abc
; it has dimension
dim (A).
Vector representations
The N N matrices that were used above to dene the classical Lie groups form
the vector representation of those groups; the expansion of these matrices around
the identity yields the vector representation of the corresponding Lie-algebra.
Fundamental spinor representations
They are dened only for SO(N). If N is odd, they have Dynkin label (0, 0, . . . , 0, 1).
If N is even there are two fundamental spinor representations with Dynkin labels
(0, 0, 0, . . . , 1, 0) and (0, 0, 0, . . . , 0, 1).
D.26 Tensor products
If V
i
1
transforms according to some representation R
1
and W
i
2
according to some repre-
sentation R
2
, then obviously the set of products V
i
1
W
i
2
forms a representation as well.
This is called the tensor product representation R
1
R
2
; it has dimension dim R
1
dim R
2
.
This representation is usually not irreducible. It can thus be decomposed into irreducible
representations:
R
1
R
2
=

j
N
12j
R
j
, (D.7)
where N
12j
is the number of times R
j
appears in the tensor product.
References
[1] M. Green, J. Schwarz and E. Witten, Superstring Theory, parts 1 and 2, Cambridge
Univ. Press, (1986).
[2] J. Polchinski String Theory, parts 1 and 2, Cambridge Univ. Press, (1998).
133
[3] D. L ust and S. Theisen, Lectures on String Theory. Springer (1989).
[4] B. Zwiebach, A rst course in String Theory. Cambridge University Press, Cam-
bridge. 2005.
[5] W. Lerche, A. Schellekens and N. Warner, Lattices and Strings, Physics Reports 177
(1989) 1.
[6] B. de Wit and J. Smith, Field Theory and Particle Physics,(North Holland).
134