Preprint typeset in JHEP style - PAPER VERSION

STRING THEORY

Paul K. Townsend
Department of Applied Mathematics and Theoretical Physics
Centre for Mathematical Sciences, University of Cambridge
Wilberforce Road, Cambridge, CB3 0WA, UK.
Email: [email protected]

Abstract: Part III Course Notes. 24 Lectures.

Contents

1. Why study String Theory? 2

2. The relativistic point particle 3

2.1 Gauge invariance 3
2.2 Hamiltonian formulation 4
2.2.1 Gauge invariance and first-class constraints 8
2.2.2 Gauge fixing 9
2.2.3 Continuous symmetries and Noether’s theorem 10
2.2.4 Quantization: canonical and Dirac’s method 12

3. The Nambu-Goto string 13

3.1 Hamiltonian formulation 15
3.1.1 Alternative form of phase-space action 16
3.1.2 Gauge invariances 16
3.1.3 Symmetries of NG action 17
3.2 Monge gauge 18
3.3 Polyakov action 19
3.3.1 Relation to phase-space action 20
3.4 Conformal gauge 21
3.4.1 “Conformal gauge” for the particle 23
3.4.2 Residual invariance of the string in conformal gauge 24
3.4.3 Conformal invariance 26
3.5 Solving the NG equations in conformal gauge 28
3.6 Open string boundary conditions 30
3.7 Fourier expansion: closed string 31
3.8 Open string 34
3.8.1 Fourier expansion: Free-ends 34
3.8.2 Mixed free-end/fixed-end boundary conditions 36
3.9 The NG string in light-cone gauge 38
3.9.1 Light-cone gauge for mixed Neumann/Dirichlet b.c.s 40
3.9.2 Closed string in light-cone gauge 40

4. Interlude: Light-cone gauge in field theory 41

4.0.3 Maxwell in light-cone gauge 41
4.0.4 Linearized Einstein in light-cone gauge 42

–1–
5. Quantum NG string 43
5.1 Light-cone gauge quantization: open string 43
5.1.1 Critical dimension 45
5.1.2 Quantum string with mixed N/D b.c.s 47
5.1.3 Quantum closed string 48
5.2 “Old covariant” quantization 49
5.2.1 The Virasoro constraints 53

6. Interlude: Path integrals and the point particle 56

6.1 Faddeev-Popov determinant 57
6.2 Fadeev-Popov ghosts 59
6.3 BRST invariance 62
6.4 BRST Quantization 63

7. BRST for the NG string 66

7.1 Quantum BRST 69
7.1.1 Critical dimension again 71

8. Interactions 73
8.1 Ghost zero modes and Sl(2; C) 74
8.2 Virasoro-Shapiro amplitude from the path integral 76
8.2.1 The Virasoro amplitude and its properties 80
8.3 Other amplitudes and vertex operators 82
8.3.1 The dilaton and the string-loop expansion 84
8.4 String theory at 1-loop: taming UV divergences 88
8.5 Beyond String Theory 90

1. Why study String Theory?

• Because it’s a consistent theory of perturbative quantum gravity, i.e. of gravi-
tons. But the perturbation theory has zero radius of convergence, and so
doesn’t actually define a “theory”. You need M-theory for that, but that’s not
yet a “theory” either.

• As a possible UV completion of the standard model. But the “landscape prob-

lem” makes it unclear whether String Theory will have any predictive power.

• Because it has applications to everything else (GR, QCD, fluid dynamics, con-
densed matter, . . . ) via AdS/CFT correspondence. But maybe you don’t really
need String Theory.

–2–
 • Because it lies on the borderline between the trivial and the insoluble. It’s the
“harmonic oscillator of the 21st century”.

• None of the above.

2. The relativistic point particle

What is an elementary particle?

• (Maths) “A unitary irrep of the Poincaré group”. These are classified by mass
and spin.

• (Physics) “A particle without structure”. The classical action for such a particle
should depend only on the geometry of its worldline (plus possible variables
describing its spin).

Let’s pursue the physicist’s answer, in the context of a D-dimensional Minkowski

space-time. For zero spin the simplest geometrical action for a particle of mass m is
Z B Z B√ Z tB √
2 dx
I = −mc dτ = −mc −ds2 = −mc −ẋ2 dt , ẋ = , (2.1)
A A tA dt

where t is an arbitrary worldline parameter. In words, the action is the elapsed proper
time between an initial point A and a final point B on the particle’s worldline.
We could include terms involving the extrinsic curvature K of the worldline,
which is essentially the D-acceleration, or yet higher derivative terms, i.e.
Z √ "  2 #
`K
I = −mc dt −ẋ2 1 + + ... , (2.2)
c2

where ` is a new length scale, which must be characteristic of some internal structure.
In the long-wavelength approximation c2 K −1  ` this structure is invisible and we
can neglect any extrinsic curvature corrections. Or perhaps the particle is truly
elementary, and ` = 0. In either case, quantization should yield a Hilbert space
carrying a unitary irrep of the Poincaré group. For zero spin this means that the
particle’s wavefunction Ψ should satisfy the Klein-Gordan equation ( − m2 ) Ψ = 0.
There are many ways to see that this is true.

2.1 Gauge invariance

Think of the particle action
Z √
I[x] = −mc dt −ẋ2 (2.3)

–3–
as a “1-dim. field theory” for D “scalar fields” xm (t) (m = 0, 1, · · · D − 1). For a
different parameterization, with parameter t0 , we will have “scalar fields” x0 (t0 ), s.t.
x0 (t0 ) = x(t). If t0 = t − ξ(t) for infinitesimal function ξ, then

x(t) = x0 (t − ξ) = x(t − ξ) + δξ x(t) = x(t) − ξ ẋ(t) + δξ x(t) , (2.4)

and hence
δξ x(t) = ξ(t)ẋ(t) . (2.5)
This is a gauge transformation with parameter ξ(t). Check:
√ 1 d(δξ x) 1 ˙ 2 
δξ −ẋ = − √
2 ẋ · = −√ ξ ẋ + ξ ẋ · ẍ
−ẋ2 dt −ẋ2
√
√ d −ẋ2 d  √ 2
= ξ˙ −ẋ + ξ
2 = ξ −ẋ , (2.6)
dt dt
so the action is invariant for any ξ(t) subject to the b.c.s ξ(tA ) = ξ(tB ) = 0. The
algebra of these gauge transformations is that of Diff 1 , i.e. 1-dim. diffeomorphisms
(maths) or 1-dim. general coordinate transformations (phys).
Gauge invariance is not a symmetry. Instead it implies a redundancy in the
description. We can remove the redundancy by imposing a gauge-fixing condition.
For example, in Minkowski space coords. (x0 , ~x) we may choose the “temporal gauge”

x0 (t) = c t . (2.7)

Since δξ x0 = c ξ when x0 = c t, insisting on this gauge choice implies ξ = 0; i.e. no

gauge transformation is compatible with the gauge choice, so the gauge is fixed. In
this gauge
Z p Z  
2 2 2 2 1 2 2 2


I = −mc dt 1 − v /c = dt −mc + mv 1 + O v /c , (2.8)
2

where v = |~x˙ |. The potential energy is therefore the rest mass energy mc2 , which we
can subtract because it is constant. We can then take the c → ∞ limit to get the
non-relativistic particle action
Z
1
IN R = m dt |d~x/dt|2 . (2.9)
2
From now on we set c = 1.

2.2 Hamiltonian formulation

√
If we start from the gauge-invariant action with L = −m −ẋ2 , then
∂L mẋ
p= =√ ⇒ p2 + m 2 ≡ 0 . (2.10)
∂ ẋ −ẋ2

–4–
So not all components of p are independent, which means that we cannot solve for
ẋ in terms of p. Another problem is that

mẋ2 √
H = ẋ · p − L = √ + m −ẋ2 ≡ 0 , (2.11)
−ẋ2
so the canonical Hamiltonian is zero.
What do we do? Around 1950 Dirac developed methods to deal with such cases.
We call the mass-shell condition p2 + m2 = 0 a “primary” constraint because it
is a direct consequence of the definition of conjugate momenta. Sometimes there
are “secondary” constraints but we will never encounter them. According to Dirac
we should take the Hamiltonian to be the mass-shell constraint times a Lagrange
multiplier, so that Z  
1 2 2


I = dt ẋ · p − e p + m , (2.12)
2
where e(t) is the Lagrange multiplier. We do not need to develop the ideas that lead
to this conclusion because we can easily check the result by eliminating the variables
p and e:

• Use the p equation of motion p = e−1 ẋ to get the new action

Z
1
dt e−1 ẋ2 − em2 .

I[x; e] = (2.13)
2
At this point it looks as though we have ‘1-dim. scalar fields” coupled to 1-
dim. “gravity”, with “cosmological constant” m2 ; in this interpretation e is
the square root of the 1-dim. metric, i.e. the “einbein”.
√
• Now eliminate e from (2.13) using the e equation of motion me = −ẋ2 , to
R √
get the standard point particle action I = −m dt −ẋ2 .

Elimination lemma. When is it legitimate to solve an equation of motion and sub-

stitute the result back into the action to get a new action? Let the action I[ψ, φ]
depend on two sets of variables ψ and φ, such that the equation δI/δφ = 0 can be
solved algebraically for the variables φ as functions of the variables ψ, i.e. φ = φ(ψ).
In this case 
δI 
≡ 0. (2.14)
δφ φ=φ(ψ)
The remaining equations of motion for ψ are then equivalent to those obtained by
ˆ = I[ψ, φ(ψ)], i.e. that obtained by back-substitution.
variation of the new action I[ψ]
This follows from the chain rule and (2.14):

δ Iˆ
  
δI  δφ(ψ) δI  δI 
= + = . (2.15)
δψ δψ φ=φ(ψ) δψ δφ φ=φ(ψ) δψ φ=φ(ψ)

–5–
Moral: If you use the field equations to eliminate a set of variables then you can
substitute the result into the action, to get a new action for the remaining variables,
only if the equations you used are those found by varying the original action with
respect to the set of variables you eliminate. You can’t back-substitute into the
action if you use the equations of motion of A to solve for B (although you can still
substitute into the remaining equations of motion).

The action (2.12) is still Diff 1 invariant. The gauge transformations are now
d
δξ x = ξ ẋ , δξ p = ξ ṗ , δξ e =
(eξ) . (2.16)
dt
However, the action is also invariant under the much simpler gauge transformations
δα x = α(t)p , δα p = 0 , δα e = α̇ . (2.17)
Let’s call this the “canonical” gauge transformation (for reasons that will become
clear). In fact,
1
t
δα I = α p2 − m2 tB (2.18)
2 A

which is zero if α(tA ) = α(tB ) = 0. The Diff 1 and canonical gauge transformations
are equivalent because they differ by a “trivial” gauge transformation.

Trivial gauge invariances. Consider I[ψ, φ] again and transformations

δI δI
δf ψ = f , δf φ = −f , (2.19)
δφ δψ
for arbitrary function f . This gives δf I = 0, so the action is gauge invariant. As
the gauge transformations are zero “on-shell” (i.e. using equations of motion) they
have no physical effect. Any two sets of gauge transformations that differ by a trivial
gauge transformation have equivalent physical implications.

If we fix the gauge invariance by choosing the temporal gauge x0 (t) = t we have
ẋm pm = ~x˙ · p~ − p0 , (2.20)
so in this gauge the canonical Hamiltonian is
p
H = p0 = ± |~p|2 + m2 , (2.21)
where we have used the constraint to solve for p0 . The sign ambiguity is typical for
a relativistic particle.
The canonical Hamiltonian depends on the choice of gauge. Another possible
gauge choice is light-cone gauge. Choose phase-space coordinates
1
x± = √ x1 ± x0 , x = x2 , . . . , xD−1



2
1
p± = √ (p1 ± p0 ) , p = (p2 , . . . , pD−1 ) . (2.22)
2

–6–
Then
ẋm pm = ẋ+ p+ + ẋ− p− + ẋ · p , p2 = 2p+ p− + |p|2 . (2.23)
Notice too that
p + = p− , p− = p+ . (2.24)
The light-cone gauge is
x+ (t) = t . (2.25)
Since δα x+ = αp+ = αp− the gauge is fixed provided that p− 6= 0. In this gauge
ẋm pm = ẋ · p + ẋ− p− + p+ , (2.26)
so the canonical Hamiltonian is now
|p|2 + m2
H = −p+ = , (2.27)
2p−
where we have used the mass-shell constraint to solve for p+ .

• Poisson brackets. For mechanical model with action

Z
I[q, p] = dt q̇ I pI − H(q, p)
 
(2.28)

the Poisson bracket of any two functions (f, g) on phase space is

∂f ∂g ∂f ∂g
{f, g}P B = I
− . (2.29)
∂q ∂pI ∂pI ∂q I
In particular,
 I
q , pJ P B = δJI .

(2.30)

• More generally, we start from a symplectic manifold, a phase-space with coor-

dinates z A and a symplectic (closed, invertible) 2-form Ω = 21 ΩAB dz A ∧ dz B .
Locally, since dΩ = 0,
Ω = dω , ω = dz A fA (z) , (2.31)
and the action in local coordinates is
Z
I = dt ż A fA (z) − H(z) .
 
(2.32)

The PB of functions (f, g) is defined as

∂f ∂g
{f, g}P B = ΩAB , (2.33)
∂z A ∂z B
where ΩAB is the inverse of ΩAB . The PB is an antisymmetric bilinear product,
from its definition. Also, for any three functions (f, g, h),
dΩ = 0 ⇔ {{f, g}P B , h}P B + cyclic permutations ≡ 0 . (2.34)
In other words, the PB satisfies the Jacobi identity, and is therefore a Lie
bracket, as a consequence of the closure of the symplectic 2-form.

–7–
 • Darboux theorem. This states that there exist local coordinates such that

Ω = dpI ∧ dq I ⇒ ω = pI dq I + d() . (2.35)

This leads to the definition (2.29) of the PB.

2.2.1 Gauge invariance and first-class constraints

Consider the action
Z
I = dt q̇ I pI − λi ϕi (q, p) ,

I = 1, . . . , N ; i = 1, . . . , n < N . (2.36)

The Lagrange multipliers λi impose the phase-space constraints ϕi = 0. Let us

suppose that
{ϕi , ϕj }P B = fij k ϕk (2.37)
for some phase-space structure functions fij k = −fji k . In this case we say that
the constraints are “first-class” (Dirac’s terminology. There may be “second-class”
constraints, but we don’t need to know about that now).

Lemma. For any phase space function, call it Q, we can define an infinitesimal
transformation of any phase-space function f , with infinitesimal parameter (t), by
δ f =  {f, Q}P B . In particular,
∂Q ∂Q
δ x I =  , δ pI = − . (2.38)
∂pI ∂xI
This transformation is such that
  
I

d ∂Q
δ ẋ pI = Q
˙ +  pI −Q . (2.39)
dt ∂pI

The special feature of first-class constraints is that they generate gauge invari-
ances. From the lemma we see that the linear combination i ϕi generates a transfor-
mation such that
d
δ ẋI pI = ˙i ϕi + ()


(2.40)
dt
and we also have

δ λi ϕi = δ λi ϕi + λi j {ϕi , ϕj }P B

= δ λk + λi j fij k ϕk


(2.41)

where we use (2.37) to get to the second line. Putting these result together, we have
Z  
k k i j k

d
δ I = dt ˙ − δ λ − λ  fij ϕk + () . (2.42)
dt

–8–
As the Lagrange multipliers are not functions of canonical variables, their transfor-
mations can be chosen independently. If we choose

δ λk = ˙k + i λj fij k (2.43)

then δ I is a surface term, which is zero if we impose the b.c.s i (tA ) = i (tB ) = 0.
The point particle provides a very simple (abelian) example. The one constraint
is
1 2
p + m2 ,


ϕ= (2.44)
2
and it is trivially first-class. It generates the canonical gauge transformations:
1  2
α x, p + m2 P B = αp ,

δα x =
2
1 
δα p = α p, p2 + m2 P B = 0 ,

(2.45)
2
and if we apply the formula (2.43) to get the gauge transformation of the einbein,
we find that δα e = α̇.
The general model (2.36) also includes the string, as we shall see later. This is
still a rather simple case because the structure functions are constants, which means
that the constraint functions ϕi span a (non-abelian) Lie algebra. In such cases the
transformation (2.43) is a Yang-Mills gauge transformation for a 1-dim. YM gauge
potential.

2.2.2 Gauge fixing

We can fix the gauge generated by a set of n first-class constraints by imposing n
gauge-fixing conditions
χi (q, p) = 0 i = 1, . . . , n . (2.46)
The gauge transformation of these constraints is

δ χi = χi , ϕj P B j ,

(2.47)

so if we want δ χi = 0 to imply j = 0 for all j (which is exactly what we do want in

order to fix the gauge completely) then we must choose the functions χi such that

det χi , ϕj P B 6= 0 .

(2.48)

This is a useful test for any proposed gauge fixing condition.

In addition to requiring that the gauge-fixing conditions χi = 0 actually do fix
the gauge, it should also be possible to make a gauge transformation to ensure that
χi = 0 if this is not already the case. In particular, if χi = f i for infinitesimal
functions f i , and χ̂i = χi + δ χi , then χ̂i = f i + δ χi and we should be able to find
parameters i such that χ̂i = 0. This requires us to solve the equation
 i
χ , ϕj P B j = −f i (2.49)

–9–
for i , but a solution exists for arbitrary f i iff the matrix {χi , ϕj }P B has non-zero
determinant.

Corollary. Whenever {χi , ϕj }P B has zero determinant, two problems arise. One
is that the gauge fixing conditions don’t completely fix the gauge, and the other
is that you can’t always arrange for the gauge fixing conditions to be satisfied by
making a gauge transformation. This is a very general point. Consider the Lorenz
gauge ∂ · A = 0 in electrodynamics (yes, that’s Ludwig Lorenz, not Henrik Lorentz
of the Lorentz transformation). A gauge transformation A → A + dα of the gauge
condition gives α = 0, which does not imply that α = 0; the gauge has not been
fixed completely. It is also true, and for the same reason, that you can’t always
make a gauge transformation to get to the Lorenz gauge if ∂ · A is not zero, even
if it is arbitrarily close to zero: the reason is that the operator  is not invertible
because there are non-zero solutions of the wave equation that cannot be eliminated
by imposing the b.c.s permissible for hyperbolic partial differential operators. The
Coulomb gauge ∇ · A = 0 does not have this problem because ∇2 is invertible for
appropriate b.c.s (but it breaks manifest Lorentz invariance).
The same problem will arise if we try to fix the gauge invariance of the action
(2.36) by imposing conditions on the Lagrange multipliers. More on this later.

2.2.3 Continuous symmetries and Noether’s theorem

In addition to its gauge invariance, the point particle action is invariant (in a
Minkowski background) under the Poincaré transformations

δΛ X m = Am + Λm n X n , δΛ Pm = Λm n Pn , (2.50)

where Am and Λmn = −Λnm are constant parameters of a spacetime translation

and Lorentz transformation, respectively. Noether’s theorem implies that there are
associated constants of the motion, i.e. conserved charges. These can be found easily
using the following version of Noether’s theorem:

• Noether’s Theorem. Let I[φ] be an action functional invariant under an

infinitesimal transformation δ φ for constant parameter . Then its variation
when  is an arbitrary function of t must be of the form
Z
δ I = dt ˙ Q . (2.51)

The quantity Q is a constant of motion. To see this, choose (t) to be zero at

the endpoints of integration. In this case, integration by parts gives us
Z
δ I = − dt  Q̇ . (2.52)

– 10 –
 But the left hand side is zero if we use the field equations because these ex-
tremize the action for any variation of φ, whereas the right-hand side is zero
for any (t) (with the specified endpoint conditions) only if Q̇(t) = 0 for any
time t (within the integration limits).
This proves Noether’s theorem: a continuous symmetry implies a conserved
charge (i.e. constant of the motion); it has to be continuous for us to be able
to consider its infinitesimal form. The proof is constructive in that it also gives
us the corresponding Noether charge: it is Q. Also, given Q we can recover
the symmetry transformation from the formula δ φ = {φ, Q}P B . There may
be conserved charges for which the RHS of this formula is zero. These are
“topological charges”, which do not generate symmetries; they are not Noether
charges.

To apply this proof of Noether’s theorem to Poincaré invariance of the point particle
action, we allow the parameters A and Λm n of (2.50) to be time-dependent. A
calculation then shows that
Z  
m 1 n m
δI = dt Ȧ Pm + Λ̇ m J n , (2.53)
2

where
Pm = Pm , J m n = X m Pn − X n P m , (2.54)
which are therefore the Poincaré charges. Notice that they are gauge-invariant; this
is obvious for Pm , and for J m n we have

δα J m n = α (P m Pn − Pn P m ) = 0 . (2.55)

Gauge-fixing and symmetries. If we have fixed a gauge invariance by imposing gauge-

fixing conditions χi = 0, then what happens if our gauge choice does not respect a
symmetry with Noether charge Q, i.e. what happens if {Q, χi }P B is non-zero.
The answer is that the symmetry is not broken. The reason is that there is an
intrinsic ambiguity in the symmetry transformation generated by Q whenever there
are gauge invariances. We may take the symmetry transformation to be

δ f = {f, Q}P B  + {f, ϕj }P B αj . (2.56)

That is, a symmetry transformation with parameter  combined with a gauge trans-
formtion for which the parameters αi are fixed, in a way to be determined, in terms
of . Because gauge transformations have no physical effect, such a transformation
is as good as the one generated by Q alone. The parameters αi () are determined by

– 11 –
requiring that the modified symmetry transformation respect the gauge conditions
χi = 0, i.e.
0 = χi , Q P B  + χi , ϕj P B αj () .
 
(2.57)
As long as {χi , ϕj }P B has non-zero determinant, we can solve this equation for all
αi in terms of .

Moral: gauge-fixing never breaks symmetries, because it just removes redundancies.

If a symmetry is broken by some gauge choice then there is something wrong with
the gauge choice!

2.2.4 Quantization: canonical and Dirac’s method

We will use the prescription
i I
q I , pJ
 
PB
→− q̂ , p̂J , (2.58)
~
which gives us the canonical commutation relations for the operators q̂ I and p̂I that
replace the classical phase-space coordinates:
 I
q̂ , p̂J = i~δJI .

(2.59)

Let’s apply this to the point particle in temporal gauge. In this case the canonical
commutation relations are precisely (2.59) where I = 1, . . . , D−1. We can realise this
om eigenfunctions of x̂I , with e-value xI , by setting p̂I = −i~∂I . The Schroedinger
equation is
∂Ψ √
HΨ = i~ , H = ± −~2 ∇2 + m2 (2.60)
∂t
Iterating we deduce that

−∇2 + ∂t2 + (m/~)2 Ψ(x) = 0

 
(2.61)

Since t = x0 , this is the Klein-Gordon equation for a scalar field Ψ and mass param-
eter m/~ (the mass parameter of the field equation is the particle mass divided by
~). The final result is Lorentz invariant even though this was not evident at each
step.
An alternative procedure is provided by Dirac’s method for quantization of sys-
tems with first-class constraints. We’ll use the point particle to illustrate the idea.

• Step 1. We start from the manifestly Lorentz invariant, but also gauge invari-
ant, action, and we quantise as if there were no constraint. This means that
we have the canonical commutation relations

[q̂ m , p̂n ] = i~δnm . (2.62)

We can realise this on eigenfunctions Ψ(x) of x̂m by setting p̂m = −i~∂m .

– 12 –
 • Step 2. Because of the gauge invariance there are unphysical states in the
Hilbert space. We need to remove these with a constraint. The mass-shell
constraint encodes the full dynamics of the particle, so we now impose this in
the quantum theory as the physical state condition

p̂2 + m2 |Ψi = 0 .


(2.63)

This is equivalent to the Klein-Gordon equation

D − (m/~)2 Ψ(x) = 0 ,
 
Ψ(x) = hx|Ψi . (2.64)

where D = η mn ∂m ∂n is the wave operator in D-dimensions.

More generally, for the general model with first-class constraints, we impose the
physical state conditions

ϕ̂i |Ψi = 0 , i = 1, . . . , n . (2.65)

The consistency of these conditions requires that

[ϕ̂i , ϕ̂j ] |Ψi = 0 ∀ i, j . (2.66)

This would be guaranteed if we could apply the PB-to-commutator prescription of

(2.58) to arbitrary phase-space functions, because this would give

[ϕ̂i , ϕ̂j ] = i~fij k ϕ̂k , (?) (2.67)

and the RHS annihilates physical states. However, because of operator ordering
ambiguities there is no guarantee that (2.67) will be true when the functions ϕi are
non-linear. We can use some of the ambiguity to redefine what we mean by ϕ̂i ,
but this may not be sufficient. There could be a quantum anomaly. The string will
provide an example of this.
From now on we set ~ = 1.

3. The Nambu-Goto string

The string analog of a particle’s worldline is its “worldsheet”: the 2-dimensional
surface in spacetime that the string sweeps out in the course of its time evolution.
Strings can be open, with two ends, or closed, with no ends. We shall start by
considering a closed string. This means that the parameter σ specifying position on
the string is subject to a periodic identification. The choice of period has no physical
significance; we will choose it to be 2π; i.e. (∼ means “is identified with”)

σ ∼ σ + 2π . (3.1)

– 13 –
The worldsheet of a closed string is topologically a cylinder, parametrised by σ and
some arbitrary time parameter t. We can consider these together as σ µ (µ = 0, 1),
i.e.
σ µ = (t, σ). (3.2)
The map from the worldsheet to Minkowski space-time is specified by worldsheet
fields X m (t, σ). Using this map we can pull back the Minkowski metric on space-
time to the worldsheet to get the induced worldsheet metric
gµν = ∂µ X m ∂ν X n ηmn . (3.3)
The natural string analog of the point particle action proportional to the proper
length of the worldline (i.e. the elapsed proper time) is the Nambu-Goto action,
which is proportional to the area of the worldsheet in the induced metric, i.e.
Z I p
IN G = −T dt dσ − det g , (3.4)

where the constant T is the string tension. Varying with respect to X we get the
NG equation of motion p 
∂µ − det g g µν ∂ν X = 0 . (3.5)
This is just the 2-dimensional massless wave equation for a set of scalar fields {X m }
(scalars with respect to the 2D local Lorentz group) propagating on a 2-dimensional
spacetime, but with a metric g that depends on the scalar fields.
Denoting derivatives with respect to t by an overdot and derivatives with respect
to σ by a prime, we have
Ẋ 2 Ẋ · X 0
 
gµν = , (3.6)
Ẋ · X 0 X 02
and hence the following alternative form of the NG action
Z I q
IN G = −T dt dσ (Ẋ · X 0 )2 − Ẋ 2 X 02 , (3.7)

where
X(t, σ + 2π) = X(t, σ) . (3.8)
This action is Diff 2 invariant; i.e. invariant under arbitrary local reparametrization
of the worldsheet coordinates. From an active point of view (transform fields rather
than the coordinates) a Diff 2 transformation of X is
δξ X m = ξ µ ∂ µ X m , (3.9)
where ξ(t, σ) is an infinitesimal worldsheet vector field. This implies that
p   p 
δξ − det g = ∂µ ξ µ − det g , (3.10)

and hence that the action is invariant if ξ is zero at the initial and final times.

– 14 –
3.1 Hamiltonian formulation
The worldsheet momentum density Pm (t, σ) canonically conjugate to the wordsheet
fields X m (t, σ) is
I
δL
q
Pm = , L = −T dσ (Ẋ · X 0 )2 − Ẋ 2 X 02 , (3.11)
δ Ẋ m
which gives
T h
02 0

0
i
Pm = √ Ẋm X − Xm Ẋ · X . (3.12)
− det g
This implies the following identities
2
P 2 + (T X 0 ) ≡ 0 , X 0m Pm ≡ 0 . (3.13)
In addition, the canonical Hamiltonian is
I
H = dσ Ẋ m Pm − L ≡ 0 . (3.14)

As for the particle, we should take the Hamiltonian to be a sum of Lagrange multi-
pliers times the constraints, so we should expect the phase-space form of the action
to be Z I  
m 1 h 2 0 2
i
0m
I = dt dσ Ẋ Pm − e P + (T X ) − u X Pm (3.15)
2
where e(t, σ) and u(t, σ) are Lagrange multipliers (analogous to the the “lapse” and
“shift” functions appearing in the Hamiltonian formulation of GR). To check this,
we may eliminate P by using its equation of motion:
P = e−1 Dt X , Dt X ≡ Ẋ − uX 0 . (3.16)
We are assuming here that e is nowhere zero (but we pass over this point). Back
substitution takes us to the action
Z I
1 n
2
o
I= dt dσ e−1 (Dt X)2 − e (T X 0 ) . (3.17)
2
Varying u in this new action we find that
Ẋ · X 0 det g
u= ⇒ Dt X 2 = . (3.18)
X 02 X 02
Here we assume that X 02 is non-zero (but we pass over this point too). Eliminating
u we arrive at the action
Z I  
1 −1 det g 0 2
I= dt dσ e − e (T X ) . (3.19)
2 X 02
Varying this action with respect to e we find that
p
T e = − det g/X 02 , (3.20)
and back-substitution returns us to the Nambu-Goto action in its original form.

– 15 –
3.1.1 Alternative form of phase-space action
Notice that the phase-space constraints are equivalent to
1 2
H± = 0 , H± ≡ (P ± T X 0 ) , (3.21)
4T
so we may rewrite the action as
Z I n o
I = dt dσ Ẋ m Pm − λ− H− − λ+ H+ , λ± = T e ± u . (3.22)

3.1.2 Gauge invariances

From the Hamiltonian form of the NG string action (3.15) we read off the canonical
Poisson bracket relations1

{X m (σ), Pn (σ 0 )}P B = δnm δ(σ − σ 0 ) . (3.23)

Using this one may now compute the PBs of the constraint functions. One finds that

{H+ (σ), H+ (σ 0 )}P B = [H+ (σ) + H+ (σ 0 )] δ 0 (σ − σ 0 ) ,

{H− (σ), H− (σ 0 )}P B = − [H− (σ) + H− (σ 0 )] δ 0 (σ − σ 0 ) ,
{H+ (σ), H− (σ 0 )}P B = 0 . (3.24)

This shows that

• The constraints are “first-class”, with constant structure functions, which are
therefore the structure constants of a Lie algebra

• This Lie algebra is a direct sum of two isomorphic algebras (−H− obeys the
same algebra as H+ ). In fact, it is the algebra

Diff 1 ⊕ Diff 1 . (3.25)

We will verify this later. Notice that this is a proper subalgebra of Diff 2 .
Only the Diff 1 ⊕ Diff 1 subalgebra has physical significance because all other
gauge transformations of Diff 2 are “trivial” in the sense explained earlier for
the particle.

The gauge transformation of any function F on phase space is

 I 
− +


δξ F = F, dσ ξ H− − ξ H+ . (3.26)
PB

1
We can put the action into the form (2.36) by expressing the worldsheet fields as Fourier series;
we will do this later. Then we can read off the PBs of the Fourier components, and use them to
get the PBs of the worldsheet fields. The result is as given.

– 16 –
where ξ ± are arbitrary parameters2 . This gives
1 − 1 +
δX = ξ (P − T X 0 ) + ξ (P + T X 0 ) ,
2T 2T
1h − i0 1  0
δP = − ξ (P − T X ) + ξ + (P + T X 0 ) .
0
(3.27)
2 2
Notice that
δξ− (P + T X 0 ) = 0 , δξ+ (P − T X 0 ) = 0 , (3.28)
and hence δξ∓ H± = 0, as expected from the fact that the algebra is a direct sum
(H+ has zero PB with H− ).
To get invariance of the action we have to transform the Lagrange multipliers
too. One finds that

0
0
δλ− = ξ˙− + λ− ξ − , δλ+ = ξ˙+ − λ+ ξ + . (3.29)

We see that λ± is a gauge potential for the ξ ± -transformation, with each being inert
under the gauge transformation associated with the other, as expected from the
direct sum structure of the gauge algebra.

3.1.3 Symmetries of NG action

The closed NG action has manifest Poincaré invariance, with Noether charges
I I
Pm = dσPm , Jmn = 2 dσX[m Pn] . (3.30)

These are constants of the motion. [Exercise: verify that the NG equations of motion
imply that Ṗm = 0 and J˙mn = 0.]

N.B. We use the following notation

1 1
T[mn] = (Tmn − Tnm ) , T(mn) = (Tmn + Tnm ) . (3.31)
2 2
In other words, we use square brackets for antisymmetrisation and round brackets for
symmetrisation, in both cases with “unit strength” (which means, for tensors of any
rank, that A[m1 ...mn ] = Am1 ...mn if A is totally antisymmetric, and S(m1 ...mn ) = Sm1 ...mn
if S is totally symmetric).

The closed NG string is also invariant under worldsheet parity: σ → −σ (mod

2π). The worldsheet fields (X, P ) are parity even, which means that X 0 is parity odd
and hence (P + T X 0 ) and (P − T X 0 ) are exchanged by parity. This implies that H±
are exchanged by parity.
2
The relative sign cancels effects due to the sign difference in the algebras of H+ and H− .

– 17 –
3.2 Monge gauge
A natural analogue of the temporal gauge for the particle is a gauge in which we
set not only X 0 = t, to fix the time-reparametrization invariance, but also (say)
X 1 = σ, to fix the reparametrization invariance of the string3 . This is often called
the “static gauge” but this is not a good name because there is no restriction to
static configurations. A better name is “Monge gauge”, after the 18th century French
geometer who used it in the study of surfaces. So, the Monge gauge for the NG string
is
X 0 (t, σ) = t X 1 (t, σ) = σ . (3.32)
In this gauge the action (3.15) becomes
Z I n
I = dt dσ Ẋ I PI + P0 − u P1 + X 0I PI


1  2 2 2 2 0 2


− e −P0 + P1 + |P| + T 1 + |X | , (3.33)
2
where I = 1, . . . , D − 2, and X is the (D − 2)-vector with components X I (and
similarly for P). We may solve the constraints for P1 , and P02 . Choosing the sign of
P0 corresponding to positive energy, we arrive at the action
Z I n p o
I 0 2 −2 2 0 2
I = dt dσ Ẋ PI − T 1 + |X | + T [|P| + (X · P) ] . (3.34)

The expression for the Hamiltonian in Monge gauge simplifies if P is momentarily

zero; we then have
I p
H = T dσ 1 + |X0 |2 (P = 0) . (3.35)

The integral equals the proper length L of the string. To see this, we observe that
the induced worldsheet metric in Monge gauge is
 2
0

2 2 2
ds ind = −dt + dσ + Ẋdt + X dσ 
 
   
2
= − 1 − |Ẋ|2 dt2 + 2Ẋ · X0 dσdt + 1 + |X0 | dσ 2 , (3.36)

and hence I I
p p
L= dσ ds2 |ind (t = const.) = dσ 1 + |X0 |2 . (3.37)

Also, when P = 0 the equations of motion in Monge gauge imply that Ẋ = 0,

so the string is momentarily at rest. The energy of such a string is H = T L, and
hence the (potential) energy per unit length, or energy density, of the string is

E =T. (3.38)
3
We could choose any linear combination of the space components of X to equal σ but locally
we can always orient the axes such that this combination equals X 1 .

– 18 –
This is the defining feature of an ultra-relativistic string. For small amplitude vibra-
tions of a uniform string of energy density E, the velocity of transverse waves on the
string is p
v = T /E c , (3.39)
so v ≤ c requires
T ≤E. (3.40)
For non-relativistic strings T  E (note that E includes any rest-mass energy, so
a string made of any available material will be non-relativistic). The Schwarzshild
solution of GR viewed as a string solution of 5D GR has T = E/2, so it is a relativistic
string, but not ultra-relativistic. Ultra-relativistic strings can occur as defects in
relativistic scalar fields; these are the relativistic analogs of the vortices that can
appear in superfluids. In the context of cosmology, such string are called “cosmic
strings”; they are well-described by the NG action at lengths scales large compared
to the size of the string core, but the NG action is then just an “effective action”,
which will have worldsheet curvature corrections, similar to those mentioned earlier
for the particle. In String Theory, we take the NG action at face value, as the action
for an “elementary” string.
Because the NG string is ultra-relativistic, it cannot support tangential momen-
tum. This is what the constraint X 0m Pm = 0 tells us; given that X 0 = t, we have
~ 0 · P~ = 0 ,
X (3.41)
which states that the (space) momentum is orthogonal to the tangent to the string.
This has various consequences, One is that there can be no longitudinal waves on
the string (i.e. sound waves). Only transverse fluctuations are physical.
It also means that a plane circular loop of NG string cannot be supported against
collapse by rotation in the plane (which can be done if T < E). This does not mean
that a plane circular loop of string cannot be supported against collapse by rotation
in other planes; we’ll see an example later.

3.3 Polyakov action

Consider the “Polyakov” action4
Z
T p
I[X, γ] = − d2 σ − det γ γ µν gµν , (3.42)
2
where γ is a new independent worldsheet metric. As before, gµν = ∂µ X · ∂ν X is
the induced metric. The action depends on the metric γ only through its conformal
class; in other words, given an everywhere non-zero function Ω, a rescaling
γµν → Ω2 γµν (3.43)
4
The Polyakov action was actually introduced by Brink, DiVecchia and Howe, and by Deser and
Zumino. Polyakov used it in the context of a path-integral quantization of the NG string, which
we will consider later.

– 19 –
has no effect; the action is “Weyl invariant”. This is easily seen since
p p
γ µν → Ω−2 γ µν , − det γ → Ω2 − det γ , (3.44)
√
so the factors of Ω cancel from − det γ γ µν , which is how γ appears in the Polyakov
action.
Varying the Polyakov action with respect to γ µν we get the equation
1
gµν − γµν (γ ρσ gρσ ) = 0 , (3.45)
2
which we can rewrite as
 
2 −2 1 µν
γµν = Ω gµν Ω = γ gµν . (3.46)
2
In other words, the equation of motion for γµν sets it equal to the induced metric
gµν up to an irrelevant conformal factor. Back substitution gives us
Z Z
T p p
I→− 2
d σ − det g g gµν = −T d2 σ − det g ,
µν
(3.47)
2
so the Polyakov action is equivalent to the NG action.

3.3.1 Relation to phase-space action

Recall that elimination of P , using P = e−1 Dt X, gives
Z
1 n
2
o  
I = d2 σ e−1 (Dt X)2 − e (T X 0 ) Dt = Ẋ − uX 0
2
Z
1 n
2
o
d2 σ e−1 Ẋ 2 − 2e−1 uẊ · X 0 + e−1 u2 − T 2 e2 (X 0 ) .


= (3.48)
2
This has the Polyakov form
Z
T p
I=− d2 σ − det γ γ µν ∂µ X · ∂ν X , (3.49)
2
with  
p µν 1 −1 u
− det γ γ = . (3.50)
Te u T 2 e2 − u2
This tells us that
u2 − T 2 e2 u
 
2
γµν = Ω , (3.51)
u 1
for some irrelevant conformal factor Ω. Equivalently
ds2 (γ) = Ω2 u2 − T 2 e2 dt2 + 2udtdσ + dσ 2



= Ω2 dσ + λ+ dt dσ − λ− dt ,



(3.52)
where
λ± = T e ± u . (3.53)
You should recognise these as the Lagrange multipliers of the phase-space action in
the form (3.22).

– 20 –
3.4 Conformal gauge
In 2D there exist local coordinates σ µ , and a conformal factor Ω, such that

γµν = Ω2 ηµν , (3.54)

i.e. such that the metric is conformally flat (in the same conformal class as the flat
Minkowski metric). This mathematical fact is plausible because the metric has three
independent functions, two of which can be eliminated by a coordinate transforma-
tion5 . In these coordinate the Polyakov action becomes
Z
T
IP oly → − d2 σ η µν ∂µ X · ∂ν X , (3.55)
2
i.e. a Minkowski space 2D field theory for D scalar fields X m (t, σ). By introducing
worldsheet light cone coordinates
1
σ ± = √ (σ ± t) (3.56)
2
we can rewrite the action (3.55) as
Z
I = −T d2 σ ∂+ X · ∂− X (3.57)

where ∂± = ∂/∂σ ± .
In the conformal gauge the NG equation of motion is the 2D wave equation

X m = 0 ,  ≡ −∂t2 + ∂σ2 ≡ η µν ∂µ ∂ν ≡ 2∂+ ∂− . (3.58)

This can be seen either by using the conformal gauge condition (3.54) in the NG
equation (3.5), or by varying X in the conformal gauge action (3.55). The general
solution is
X = XL (σ + ) + XR (σ − ) , (3.59)
where XL depends only on σ + and XR depends only on σ − . In other words, we have
some wave profile XL that moves to the left at the speed of light, superposed on
another wave profile XR that moves to the right at the speed of light. However, not
all components of XL and XR are independent because the equation of motion for
the independent metric, i.e. eq. (3.46), tells us that it is in the same conformal class
as the induced metric, and in the conformal gauge this condition is gµν = Ω2 ηµν , or

Ẋ 2 Ẋ · X 0 −Ω2 0
   
= , (3.60)
Ẋ · X 0 (X 0 )2 0 Ω2
5
The proof starts from the observation that the curvature of a 2D space is entirely determined by
its Ricci scalar R, which becomes a function of Ω and its derivatives for a conformally-flat metric;
so given R we get a differential equation for Ω with R as a “source” term. Locally, this can be
solved for Ω in terms of R, so for any R there exists an Ω for which the metric is locally conformally
flat.

– 21 –
This equation determines the irrelevant conformal factor Ω; more importantly, it also
imposes the constraints
 2
2 0 2 0 0
Ẋ + (X ) = 0 & Ẋ · X = 0 ⇔ Ẋ ± X = 0. (3.61)
These constraints are equivalent to the Hamiltonian constraints H± = 0. To see
this recall that P = T Ẋ in conformal gauge, so that
1 0 2 T  0
2
H± ≡ (P ± T X ) → Ẋ + X . (3.62)
4T 4
Notice that we can write the constraints using the worldsheet light cone coordinates
σ ± as
(∂− X)2 = 0 & (∂+ X)2 = 0 . (3.63)
A puzzle: We found the constraints, starting from the Polyakov action, from the
conformal gauge condition that we imposed to reduce this action to the simple form
(3.55) of a massless 2D scalar field theory. If we were given only the conformal gauge
action, and not told where it came from, how would we deduce the constraints? It
looks as though the action no longer “knows” about the constraints once we have gone
to conformal gauge. We shall address this puzzle in the context of the Hamiltonian
formulation.

From (3.52) we see that the conformal gauge condition (3.54) is equivalent to
the conditions
 
+ − 1
λ =λ =1 ⇔ e= & u=0 . (3.64)
T
In other words, it corresponds to imposing particular conditions on the Lagrange
multipliers that appear in the Hamiltonian form of the action.

Check. Setting e = 1/T and u = 0 we get the action

Z  
2 m 1 2 T 0 2
I = d σ Ẋ Pm − P − (X ) (3.65)
2T 2
Eliminating P by its equation of motion P = T Ẋ, we find the same action I[X]
for massless 2D fields as we found from the Polyakov action. Variation of I[X] with
respect to X yields the 2D wave equation X = 0, but that’s it. We can’t vary λ±
to deduce the constraints because we have set them both to unity. You might think
“yes, but when you set a field to a constant you should expect to lose the equation
that you found before by varying it, and that just means that you should add this
equation ‘by hand’ to those you get by varying the gauge-fixed action.”
However, it is not possible to “lose” equations by fixing a gauge, as long
as the gauge fixing is legitimate. This is because we remove redundancy, and
only redundancy, when we fix a gauge. It follows that the conformal gauge is not
legitimate, strictly speaking, because it must remove more than redundancy. To see
what is going on, it is helpful to return to the particle.

– 22 –
3.4.1 “Conformal gauge” for the particle
The particle analog of the conformal gauge is

e = 1/m . (3.66)

Substituting into the action we find that

p2
Z  
m m
I → dt ẋ pm − − . (3.67)
2m 2

Eliminating p by its equation of motion p = mẋ we arrive at the action

Z
m
dt ẋ2 − 1 .

I[x] = (3.68)
2
This is the particle analog of the conformal gauge action. The equation of motion is

ẍ = 0 , (3.69)

which is the “1D wave equation”. That’s all we get from I[x] but if we recall that the
√
particle equations of motion, prior to gauge fixing, imply that me = −ẋ2 , which
√
becomes 1 = −ẋ2 when e = 1/m, we see that the e = 1/m choice of gauge is
equivalent, using the equations of motion, to the constraint

ẋ2 = −1 . (3.70)

We appear to have lost this constraint from the gauge-fixed action. But this is
impossible, so there must be something wrong with the gauge choice.
To see why we cannot use the gauge invariance of the particle action to set
e = 1/m, recall that the gauge variation of e is δe = α̇, so
Z tB 
δα dt e = [α]ttBA = 0 by b.c.s on α(t). (3.71)
tA

The integral is the lapsed proper time divided by the mass. This quantity is gauge
invariant and cannot be changed by a gauge transformation, so setting e to any
particular constant fixes a gauge-invariant quantity; this is more than just fixing the
gauge. The best that we can do is to set

e = s, (3.72)

for variable constant s, on which the gauge-fixed action still depends. Varying the
gauge-fixed phase-space action with respect to s yields the integrated constraint
Z tB
dt p2 + m2 = 0 .


(3.73)
tA

– 23 –
However, since p2 +m2 is a constant of the motion, this is equivalent, when combined
with the equation of motion ṗ = 0 to the unintegrated constraint p2 + m2 = 0, which
we can think of as a initial condition. This interpretation would be obvious if,
instead of choosing e = s, we choose e = 1/m almost everywhere, leaving it free in a
neighbourhood of the initial time.
There is another, related, problem with the gauge-fixing condition e = 1/m,
which is that it doesn’t completely fix the gauge. Setting to zero the gauge variation
of the gauge-fixing condition we find that

0 = δα (e − 1/m) = α̇ (3.74)

which does not imply that α = 0. Instead it implies that α = ᾱ for some constant ᾱ.
So the gauge choice leaves a residual invariance, a rather simple one for the particle.
This residual invariance is a symmetry of the action (3.68) but it is really a gauge
invariance once we take into account that the Noether charge is precisely what the
apparently lost constraint sets to zero.

3.4.2 Residual invariance of the string in conformal gauge

The analog of the conformal gauge for the particle was both too strong and too weak

• Too strong because it sets to a gauge invariant variable, precisely the variable
needed to derive the constraint as an initial condition by varying the action.

• Too weak because it doesn’t completely fix the gauge; it leaves a residual
invariance, which is a symmetry of the gauge fixed action, but it’s really a
gauge invariance because the Noether charge is precisely what is set to zero by
the constraint.

As mentioned earlier, the same happens for Lorenz gauge in electrodynamics, and
the same is true of the conformal gauge of the string. The conformal gauge is too
strong; we cannot choose the conformal gauge everywhere on the string worldsheet.
A careful analysis would lead to the conclusion that the conformal gauge action
should depend on additional variables, and variation with respect to them leads to
the constraints (∂± X)2 = 0 being imposed, e.g. as initial conditions.
The conformal gauge is also too weak. It leaves a residual invariance, which turns
out to play a major role in String Theory. We can see what this invariance is by re-
turning to the gauge transformations of the Lagrange multipliers λ± , given in (3.29).
Setting λ± = 1, and requiring that this be maintained by a gauge transformation,
leads to
√
0 = δ λ− − 1 |λ− =1 = ξ˙− + (ξ − )0 = 2∂+ ξ − ,


√
0 = δ λ+ − 1 |λ+ =1 = ξ˙− − (ξ + )0 = 2∂− ξ + ,


(3.75)

– 24 –
so we have a residual invariance with parameters ξ ± satisfying

∂− ξ + = ∂+ ξ − = 0 . (3.76)

In other words,
ξ − = ξ − (σ − ) , ξ + = ξ + σ+ .


(3.77)
Let’s check this. Recalling the gauge transformation of X given in (3.27), and
using the conformal gauge relation P = T Ẋ, we see that
1 
δξ X = √ ξ − ∂− X + ξ + ∂+ X .

(3.78)
2
R
Using this in the conformal gauge action I[X] = T d2 σ ∂+ X · ∂− X, we find that the
ξ + variation of I[X] is
Z
T
d2 σ ∂+ ξ + ∂+ X · ∂− X + ∂+ X · ∂− ξ + ∂+ X .



δξ + I = √ (3.79)
2

Now we integrate by parts in the first term (with respect to ∂+ ); in doing so we pick
up a total time derivative, which we can omit. We are left with a term that almost
cancels the second term, leaving
Z
T
d2 σ ∂− ξ + (∂+ X)2 .

δξ + I = √ (3.80)
2
A similar calculation yields
Z
T
d2 σ ∂+ ξ − (∂− X)2 .

δξ − I = √ (3.81)
2

We thus confirm that the action is invariant if the parameters ξ ± are restricted by
(3.76).
The residual invariance is a symmetry of the conformal gauge action. Using the
above calculation, we can compute the Noether charges by allowing the parameters
to have an additional t-dependence; i.e.

ξ − = ξ − (σ − , t) , ξ + = ξ + (σ + , t) , (3.82)

then we find that

Z
T
d2 σ ∂t ξ + (∂+ X)2 + ∂t ξ − (∂− X)2 ,

δξ I = (3.83)
2

so the Noether charges are

Θ++ = T (∂+ X)2 & Θ−− = T (∂− X)2 . (3.84)

– 25 –
These are the non-zero components of the stress tensor in light-cone coordinates.
Viewing the conformal gauge action as a massless 2D scalar field theory, its stress
tensor is  
1 ρσ
Θµν = T ∂µ X · ∂ν X − ηµν η (∂ρ X · ∂σ X) . (3.85)
2
Because this tensor is traceless its Θ+− component is zero (Exercise: check this).
The only non-zero components are Θ++ and Θ−− .
Puzzle: How can the residual invariance be a symmetry when it is a residual gauge
invariance? The answer is that it is a symmetry of the conformal gauge action, rather
than a gauge invariance, precisely because the Hamiltonian constraints are “missing”
from this action. Notice that the Noether charges of the residual symmetry
are precisely what the missing constraints set to zero. So if we properly take
into account the origin of the conformal gauge action, and include the constraints
as initial conditions, then we have a residual gauge invariance. But if we view the
conformal gauge action as just a 2D scalar field theory, it is a symmetry.

3.4.3 Conformal invariance

What is the residual symmetry of the string in conformal gauge?
The conformal group of any space-time, with metric g, is the group of conformal
isometries, which is the subgroup of the group of diffeomorphisms, generated by
vector fields, that leave invariant the conformal class of g. This subgroup is generated
by the conformal Killing vector fields, for which6

(Lξ g)µν = χgµν (3.86)

for any function χ. The symbol Lξ means “Lie derivative with respect to vector field
ξ”, and its definition is such that

(Lξ g)µν = ξ ρ ∂ρ gµν + 2∂(µ ξ ρ gν)ρ . (3.87)

The infinitesimal action of this group on a scalar X is X → X + δξ X, where

δξ X = Lξ X = ξ µ ∂µ X . (3.88)

For Minkowski space-time with metric η, which is constant in cartesian coordinates,

a vector field ξ is a conformal Killing vector field if 2∂(µ ξν) = χηµν . The number
of linearly independent conformal Killing vector fields is finite for n > 2. Let’s
6
If χ = 0 the vector field ξ generates an isometry, which is a special case of a conformal isometry.
The isometry group of Minkowski space-time is the Poincaré group, which is a proper subgroup of
the conformal group.

– 26 –
now consider the n = 2 case; in light-cone coordinates the condition (3.87) can be
expressed as
2∂+ ξ − ∂+ ξ + + ∂− ξ −
   
0χ
= . (3.89)
∂+ ξ + + ∂− ξ − 2∂− ξ + χ0
This equation determines the irrelevant factor χ but it also restricts ξ ± to satisfy
∂∓ ξ ± = 0, which is precisely (3.76). We conclude that the residual symmetry of the
NG action in conformal gauge is 2D conformal invariance.

What is the algebra of the residual symmetry transformations?

A symmetry of the action is also a symmetry of the equations of motion7 . For
any solution of the equations of motion we have

X = XL (σ + ) + XR (σ − ) (3.90)

subject to the constraints

(∂− XR )2 = 0 & (∂+ XL )2 = 0 , (3.91)

which follow from (3.63). The residual symmetry transformation (3.78) can now be
written as
   
+ 1 + + − 1 − −
δξ+ XL (σ ) − √ ξ (σ+ )∂+ XL (σ ) + δξ− XR (σ ) − √ ξ (σ− )∂− XR (σ ) = 0 .
2 2
(3.92)
+ −
As the terms in the first (second) bracket are all functions of σ (σ ) only, both sets
of bracketed terms are separately zero8 , so
1 1
δξ− XR (σ − ) = √ ξ − (σ− )∂− XR (σ − ) .
δξ+ XL (σ + ) = √ ξ + (σ+ )∂+ XL (σ + ) ,
2 2
(3.93)
These are just the transformations due to separate 1D coordinate transformations;
the algebra of these transformations is

Diff 1 ⊕ Diff 1 . (3.94)

Whereas this was originally an algebra of gauge transformations of the phase-space

form of the NG action, with unrestricted parameters, it is now the algebra of a
symmetry of the conformal gauge action.

N.B. A basis for Diff 1 is

{z n ∂z ; n = 0, 1, 2, . . .} (3.95)
7
But the converse is not true; there can be symmetries of the equations of motion that are not
symmetries of the action.
8
Actually, they are both equal to constants, which have to sum to zero, but we shall ignore shifts
of XL and XR by constants.

– 27 –
where z can be either σ + or σ − . An arbitrary Diff 1 symmetry transformation, with
constant parameters {ξn ; n = 0, 1, 2 . . .} has a parameter

1
ξ(z) = ξ0 + ξ1 z + ξ2 z 2 + . . . (3.96)
2
In other words, the linear combination of parameters ξ depend only on z, not on
t. This accords with the fact that the parameters ξ ± of the residual invariance
in conformal gauge are functions only of σ ± . If we allow them to also have an
independent dependence on t, then both ξ ± become arbitrary functions of both σ
and t, which is what they were before we chose the conformal gauge.

Moral: The NG action is a gauge theory of the 2D conformal group, for which the
algebra is Diff 1 ⊕ Diff 1 . In conformal gauge there is a residual conformal symmetry,
with the same algebra but the parameters are now constants.

The algebra Diff 1 has a finite-dimensional subalgebra, for which a basis of vector
fields is  
1 2
J− = ∂z , J3 = z∂z , J+ = z ∂z . (3.97)
2
The commutation relations of these vector fields are [Exercise: check this]

[J3 , J± ] = ±J± , [J− , J+ ] = J3 . (3.98)

This is the algebra of Sl(2; R), so the finite dimensional conformal algebra is Sl(2; R)⊕
Sl(2; R) ∼
= SO(2, 2). In n dimensions the conformal algebra is SO(2, n); for example,
for D = 4 it is SO(2, 4).

3.5 Solving the NG equations in conformal gauge

Locally, the NG solutions reduce to the 2D wave equation in conformal gauge, which
is easily solved. In particular, we can solve the 2D wave-equation for X 0 by setting
X 0 (t, σ) = t. In fact, we can use the residual conformal symmetry in conformal
gauge to set X 0 = t without loss of generality. Let’s check whether this does fix the
residual conformal symmetry. Recall that the residual conformal transformation of
X is δξ X = ξ + (σ + )∂+ X + ξ − (σ − )∂− X, so
√

0 = 2 δξ X 0 − t  = ξ + (σ + ) − ξ − (σ − ) , (3.99)
X =t0

¯ a constant. The
which implies that ξ + (σ + ) = ξ − (σ − ) and hence that ξ + = ξ − = ξ,
only surviving part of the conformal transformation is therefore δX ∝ ξX ¯ 0 , which
corresponds to a constant shift of σ (a change of where we choose σ = 0 on the
string).

– 28 –
 We can easily write down the general solution for X~ that satisfies the 2D wave
equation but to have a global solution we must also solve the conformal gauge con-
straints, which are now
√ 2

 2∂ ~
±  = 1.
X

(3.100)

Given a solution of the 2D wave equation for X ~ we may check directly to see whether
the constraints are satisfied. Alternative, we may compute the induced metric to see
if it is conformally flat; if it is then the conformal gauge constraints will be satisfied
because they are precisely the conditions for conformal flatness of the induced metric.
Let’s apply these ideas to the closed string configuration in a 5-dimensional
space-time with X 0 = t and

1 in(σ−t) 1 im(σ+t)
Z ≡ X 1 + iX 2 = e , W ≡ X 3 + iX 4 = e . (3.101)
2n 2m
This configuration clearly solves the 2D wave equation. If the induced metric is
conformally flat then it will also solve the full NG equations, including the constraints.
A calculation gives

2 1
ds2 ind = − dX 0 + |dZ|2 + |dW |2 = −dt2 +

dt2 + dσ 2


2
1
−dt2 + dσ 2 .


= (3.102)
2
In other words,
1
gµν = ηµν . (3.103)
2
This is flat, and hence conformally flat, so the given configuration is a solution of
the NG equations.
This solution has the special property of being stationary; the string is motionless
in a particular rotating frame. To see this, we first compute the proper length L of
the string. Setting t = t0 in the induced worldsheet metric (for some constant t0 ) we
see that d`2 = 21 dσ 2 , so
1
I √
L= dσ = 2 π . (3.104)
2
It is rather surprising that this should be constant, i.e. independent of t0 ; it means
that the motion of the string is supporting it against collapse due to its tension. To
check this, we may compute the total energy, which is
I I
H = dσP = T dσ Ẋ 0 = 2πT .
0
(3.105)

We see that
√ √ 
H= 2TL = TL + 2 − 1 TL (3.106)

– 29 –
The first term is the potential energy of the string. The second term is therefore
kinetic energy. The string is supported against collapse by rotation in the Z and
W planes. The string is circular for n = m, and planar, so a circular planar loop
of string can be supported against collapse by rotation in two orthogonal planes
provided that neither of them coincides with the plane of the string loop.

3.6 Open string boundary conditions

An open string has two ends. We shall choose the parameter length to be π, so the
action in Hamiltonian form is
Z Z π  
m 1  2 0 2
 0
I = dt dσ Ẋ Pm − e P + (T X ) − uX · P . (3.107)
0 2

What are the possible b.c.s at the ends of the string?

Principle: the action should be stationary when the equations of motion are sat-
isfied. In other words, when we vary the action to get the equations of motion,
the boundary terms arising from integration by parts must be zero; otherwise the
functional derivative of the action is not defined.

Applying this principle to the above action, we see that boundary terms can arise
only when we vary X 0 and integrate by parts to get the derivative with respect to σ
off the δX variation (we can ignore any boundary terms in time). These boundary
terms are Z
σ=π
δI|on−shell = − dt T 2 eX 0 + uP · δX σ=0 .


(3.108)

Here, “on-shell” is shorthand for “using the equations of motion”. [Exercise: check
this].
It would make no physical sense to fix X 0 at the endpoints, and if X 0 is free
then so is Ẋ 0 and hence P 0 when we use the equations of motion, so the boundary
term with the factor of δX 0 will be zero only if we impose the conditions


0 0
u|ends = 0 , X  = 0. (3.109)
ends

Given that e 6= 0 (we’ll pass over this point) we conclude that


~ 0 ~
X · δX 

= 0. (3.110)
ends

~ we
What this means is that at each end and for each component, call it X∗ , of X,
have

either X∗0 = 0 (Neumann b.c.s) ,

or δX∗ = 0 ⇒ X∗ = X̄∗ (constant) (Dirichlet b.c.s) (3.111)

– 30 –
 ~ 0 = 0 at
There are many possibilities. The simplest is free-end b.c.s for which X
both ends. In this case
X 0 |ends = 0 . (3.112)
This implies that (X 0 )2 is zero at the ends of the string. The open string mass-shell
constraint then implies that P 2 is zero at the endpoints, and since
  
P |ends = e−1 Ẋ − uX 0  = e−1 Ẋ  = 0, (3.113)
 
ends ends

we deduce that Ẋ 2 is zero at the ends of the string; i.e. the string endpoints move
at the speed of light.

3.7 Fourier expansion: closed string

The worldsheet fields of the closed string are periodic functions of σ with (by con-
vention) period 2π, so we can express them as Fourier series. It is convenient to
express (X, P ) as Fourier series by starting with the combinations P ± T X 0 (because
the gauge transformations act separately on P + T X 0 and P − T X 0 ), so we write
r
0 T X ikσ
P − TX = e αk (α−k (t) = αk∗ )
π
k∈Z
r
T X
P + T X0 = e−ikσ α̃k (α̃−k (t) = α̃k∗ ) (3.114)
π
k∈Z

Recall that worldsheet parity σ → −σ exchanges P + T X 0 with P − T X 0 . Because

of the relative minus sign in the exponent of the Fourier series, this means that
worldsheet parity exchanges αk with α̃k :

αk ↔ α̃k . (3.115)

We can integrate either of the above equations to determine the total D-momentum
in terms of Fourier modes, since X 0 integrates to zero for a closed string; this gives
us I √
4πT α0 p
p = P dσ = √ ⇒ α0 = α̃0 = √ . (3.116)
4πT α̃0 4πT
By adding the Fourier series expressions for P ± T X 0 we now get
r
p(t) T X ikσ
P (t, σ) = + e [αk (t) + α̃−k (t)] , (3.117)
2π 4π k6=0

By subtracting we get
1 X ikσ
X0 = −√ e (αk − α̃−k ) , (3.118)
4πT k6=0

– 31 –
which we may integrate to get the Fourier series expansion for X:
1 X i ikσ
X(t, σ) = x(t) + √ e [αk (t) − α̃−k (t)] . (3.119)
4πT k6=0 k

The integration constant (actually a function of t) can be interpreted as the position

of the centre of mass of the string; we should expect it to behave like a free particle.
Using the Fourier series expansions of (X, P ) we now find that
I Xi
d
dσ Ẋ m Pm = ẋm pm + α̇k · α−k + α̃˙ k · α̃−k + () (3.120)
k=1
k dt

Exercise: check this [Hint. Cross terms that mix α with α̃ are all in the total time
derivative term, and the k < 0 terms in the resulting sum double the k > 0 terms].
Next we Fourier expand the constraint functions H± :
1 X inσ 1 X −inσ
H− = e Ln , H+ = e L̃n . (3.121)
2π 2π
n∈Z n∈Z

Inverting to get the Fourier coefficients in terms of the functions H± , we get

I I
−inσ 1 2
Ln = dσ e H− = dσ e−inσ (P − T X 0 )
4T
I I
1 2
inσ
L̃n = dσ e H+ = dσ einσ (P + T X 0 ) . (3.122)
4T
Inserting the Fourier expansions (3.114) we find that (Exercise: check this)
1X 1X
Ln = αk · αn−k , L̃n = α̃k · α̃n−k . (3.123)
2 2
k∈Z k∈Z

We may similarly expand the Lagrange multipliers as Fourier series but it should
be clear in advance that there will be one Fourier mode of λ− for each Ln (let’s call
this λ−n ) and one Fourier mode of λ+ for each L̃n (let’s call this λ̃−n ). We may now
write down the closed string action in terms of Fourier modes. It is
 
Z  Xi X   
I = dt ẋm pm + α̇k · α−k + α̃˙ k · α̃−k −


λ−n Ln + λ̃−n L̃n . (3.124)

k=1
k 
n∈Z

This action is manifestly Poincaré invariant. The Noether charges are

I
Pm = dσPm = pm ,
I
J mn = 2 dσX [m P n] = 2x[m pn] + S mn (3.125)

where the spin part of the Lorentz charge is (Exercise: check this)
∞
mn
X i  [m n] [m n]

S = −2 α−k αk + α̃−k α̃k . (3.126)
k=1
k

– 32 –
 • Lemma. For a Lagrangian of the form
i
L= α̇α∗ − H(α, α∗ ) (3.127)
c
for constant c, the PB of the canonical variables takes the form

{α, α∗ }P B = −ic . (3.128)

p
To see this set α = c/2[q + i sign(c)p] to get L = q̇p − H, for which we know
that {q, p}P B = 1. This implies the above PB for α and α∗ .

Using this lemma we may read off from the action that the non-zero Poisson brackets
of canonical variables are {xm , pn }P B = δnm and
 m n
αk , α−k P B = −ikη mn ,
 m n
α̃k , α̃−k P B = −ikη mn . (3.129)

Using these PBs, and the Fourier series expressions for (X, P ), we may compute the
PB of X(σ) with P (σ 0 ). [Exercise: check that the result agrees with (3.23).]
We may also use the PBs (3.129) to compute the PBs of the constraint functions
(Ln , L̃n ). The non-zero PBs are (Exercise: check this)
n o
{Lk , Lj }P B = −i (k − j) Lk+j , L̃k , L̃j = −i (k − j) L̃k+j . (3.130)
PB

We may draw a number of conclusions from this result:

• The constraints are first class, so the Ln and L̃n generate gauge transformations,
for each n ∈ Z.

• The structure functions of the algebra of first-class constraints are constants.

This means that the (Ln , L̃n ) span an infinite dimensional Lie algebra.

• The Lie algebra of the gauge group is a direct sum of two copies of the same
algebra, sometimes called the Witt algebra.

The Witt algebra is also the algebra of diffeomorphisms of the circle. Suppose we
have a circle parameterized by θ ∼ θ + 2π (we could take θ to be σ + or σ − ). The
algebra Diff 1 of diffeomorphisms is spanned by the vector fields on the circle, and
since these are periodic we may take as a basis set the vector fields {Vn ; n ∈ Z},
where
d
Vn = einθ . (3.131)
dθ
The commutator of two basis vector fields is

[Vk , Vj ] = −i (k − j) Vk+j . (3.132)

Corollary: the algebra of the gauge group is Diff 1 ⊕ Diff 1 , as claimed previously.

– 33 –
3.8 Open string
The open string has two ends. We will choose the ends to be at σ = 0 and σ = π,
so the parameter length of the string is π (this is just a convention). We shall first
consider the case of free-end (Neumann) boundary conditions. Then we shall go on
to see how the results change when the string ends are not free to move in certain
directions (mixed Neumann/Dirichlet b.c.s).

3.8.1 Fourier expansion: Free-ends

If the ends of the string are free, we must require X 0 to be zero at the ends, i.e. at
σ = 0 and σ = π. We shall proceed in a way that will allow us to take over results
from the closed string; we shall use a “doubling trick”.
• First we extend the definition of (X, P ) from the interval [0, π] to the interval
[0, 2π] in such a way that (X, P ) are periodic on this doubled interval. This
will allow us to use the closed string Fourier series expressions.

• Next, we impose a condition that relates (X, P ) in the interval [π, 2π] to (X, P )
in the interval [0, π]; this will ensure that any additional degrees of freedom
that we have introduced by doubling the interval are removed; everything will
depend only on what the piece of string in the interval [0, π] is doing. The con-
dition we impose should be consistent with periodicity in the doubled interval,
but it should also imply the free-end b.c.s at σ = 0, π.

• Since we have periodicity in σ with period 2π, if σ ∈ [0, π] then −σ ∼ −σ+2π ∈

[π, 2π], so we need to relate the worldsheet fields at σ to their values at −σ.
The solution to these requirements is to impose the condition

(P + T X 0 ) (σ) = (P − T X 0 ) (−σ) . (3.133)

This is consistent with periodicity, and setting σ = 0 it implies that X 0 (0) = 0. It

also implies that X 0 (π) = 0 because −π ∼ π by periodicity. In terms of Fourier
modes, the condition (3.133) becomes

α̃k = αk (k ∈ Z). (3.134)

Using this in (3.114) we have

r
T X ∓ikσ
P ± T X0 = e αk (α−k = αk∗ ) . (3.135)
π
k∈Z

Equivalently,
r
T X i X
P = cos(kσ) αk , X0 = −√ sin(kσ) αk (3.136)
π πT k∈Z
k∈Z

– 34 –
Integrating to get X, and defining p(t) by
p
α0 = √ , (3.137)
πT
we have
1 Xi
X(t, σ) = x(t) + √ cos(kσ) αk ,
πT k6=0 k
r
p(t) T X
P (t, σ) = + cos(kσ) αk . (3.138)
π π k6=0

Notice that p is again the total momentum since

Z π
p= P (t, σ)dσ , (3.139)
0

but its relation to α0 differs from that of the closed string.

Using the Fourier series expansions for (X, P ) we find that
Z π ∞
m m
X i d
Ẋ Pm dσ = ẋ pm + α̇k · α−k + () . (3.140)
0 k=1
k dt

Because of (3.133) we also have H+ (σ) = H− (−σ), so we should impose a similar

relation on the Lagrange multipliers

λ+ (σ) = λ− (−σ) . (3.141)

Then
Z π Z π Z π
− −
+
dσ λ− (−σ)H− (−σ)


dσ λ H− + λ H+ = dσ λ (σ)H− (σ) +
0
Z0 π Z0 2π
= dσ λ− (σ)H− (σ) + dσ λ− (σ)H− (σ)
I0 π

= dσ λ− H− , (3.142)

where we can now use the Fourier series expansions of the closed string.
The final result for the open string action in Fourier modes is
 
Z  Xi X 
I = dt ẋm pm + α̇k · α−k − λ−n Ln . (3.143)

k=1
k 
n∈Z

The difference with the closed string is that we have one set of oscillator variables
instead of two. We can now read off the non-zero PB relations

{xm , pn }P B = δnm ,
 m n
αk , α−k P B = −ikη mn , (3.144)

– 35 –
and we can use this to show that

{Ln , αkm }P B = ikαn+k . (3.145)

This means that the gauge variation of αk is

X X
δξ αkm = ξ−n {αkm , Ln }P B = −ik m
ξ−n αn+k , (3.146)
n∈Z n∈Z

where ξn are parameters. To compute the gauge transformation of (x, p) we use the
fact that
1
L0 = α02 + . . . , Ln = α0 · αn + . . . (3.147)
2
√
where the dots indicate terms that do not involve α0 , and the relation p = πT α0
to compute
1 X
δξ xm = √ ξ−n αn , δpm = 0 . (3.148)
πT n∈Z
Finally, one may verify that the action is invariant if
X
δξ λn = ξ˙n + i (2k − n) ξk λn−k . (3.149)
k∈Z

3.8.2 Mixed free-end/fixed-end boundary conditions

Let’s now suppose that we have mixed Neumann and Dirichlet boundary conditions.
We will divide the cartesian coordinates X m into two sets
 m̂  m̌
X ; m̂ = 0, 1, . . . , p , X ; m̌ = p + 1, . . . , D − 1 . (3.150)

We will suppose that X m̂ (t, σ) are subject to Neumann b.c.s. and that X m̌ (t, σ) are
subject to Dirichlet b.c.s, so the string is stretched between a p-plane at the origin
and a parallel p-plane situated at X m̌ = Lm̌ . The boundary conditions corresponding
to this situation are


m̂ 0 
 
X  =0 X m̌ σ=0 = 0 & X m̌ σ=π = Lm̌ . (3.151)
ends

Notice that these boundary conditions break invariance under the SO(1, D − 1)
Lorentz group to invariance under the subgroup SO(1, p) × SO(D − p − 1). In
particular the D-dimensional Lorentz invariance is broken to a (p + 1)-dimensional
Lorentz invariance.
To get the Fourier series expansions of P ± T X 0 for these b.c.s. we may again
use the doubling trick, but the constraint relating the components of (P ± T X 0 ) at
σ to the components at −σ now depends on whether it is a m̂ component or a m̌
component. For the m̂ components we choose the (3.133) condition, which implies
that (X m̂ )0 is zero at the endpoints. For the m̌ components we impose the condition
m̌ m̌
(P + T X 0 ) (σ) = − (P + T X 0 ) (−σ) , (3.152)

– 36 –
which implies that Pm̌ is zero at the endpoints, and hence (since the equations of
motion imply that P = e−1 Ẋ at the endpoints) that

m̌ 
Ẋ  = 0. (3.153)
ends

In terms of the oscillator variables, the condition (3.152) becomes

α̃km̌ = −αkm̌ , (3.154)
which gives r
0 m̌ T X ∓ikσ m̌
(P ± T X ) = ∓ e αk . (3.155)
π
k∈Z
Taking sums and differences, we find that
r
T X
P m̌ = i sin(kσ)αkm̌ , (3.156)
π
k∈Z

which is indeed zero at σ = 0, π, and

0 αm̌ 1 X
X m̌ = √ 0 + √ cos(kσ)αkm̌ , (3.157)
πT πT k6=0

which we may integrate to get X m̌ . For the b.c.s of (3.150), we get

r
m̌ Lm̌ σ 1 X1 m̌ m̌ π m̌
X = −√ sin(kσ)αk , L = α . (3.158)
π πT k6=0 k T 0

Using the Fourier series expressions for X m̌ and Pm̌ , we find that
Z π ∞
m̌
X i X m̌ m̌ d
Ẋ Pm̌ = αk α−k + () (3.159)
0 k=1
k m̌
dt
The only effect of the Dirichlet b.c.s (relative to Neumann b.c.s) is the absence of
any zero mode term.
Irrespective of whether the b.c.s are Neumann or Dirichlet, we still have H+ (σ) =
H− (−σ), so the Fourier series for the constraint functions are formally independent
of the b.c.s. So, adding the contributions from the Neumann and Dirichlet directions,
we arrive at the action
 
Z ∞
 X i X 
I = dt ẋm̂ pm̂ + α̇k · α−k − λ−n Ln , (3.160)
 k
k=1

n∈Z

1
where, as before, Ln = 2 k αk · αn−k . Notice the absence of a ẋm̌ pm̌ term. Apart
P

from this, the only difference to the√free-end case is in the zero-mode contribution
to the Ln . Whereas we had α0 = p/ πT , we now have
r
m̂ pm̂ m̌ T m̌
α0 = √ , α0 = L . (3.161)
πT π

– 37 –
3.9 The NG string in light-cone gauge
We shall start with the open string (with free-end b.c.s). We shall impose the gauge
conditions
X + (t, σ) = x+ (t) , P− (t, σ) = p− (t) . (3.162)
It customary to also set x+ (t) = t, as for the particle, but it is simpler not to do this.
This means that we will not be fixing the gauge completely since we will still be free
to make σ-independent reparametrizations of the worldsheet time t.
The above gauge-fixing conditions are equivalent to
+
(P ± T X 0 ) = p− (t) ⇔ αk+ = 0 ∀k 6= 0 . (3.163)

In other words, we impose a light-cone gauge condition only on the oscillator variables
of the string, not on the zero modes (centre of mass variables). Let’s check that
the gauge has been otherwise fixed. We can investigate this using the criterion
summarised by the formula (2.48); we compute
+ +

Ln , α−k = −ikαn−k
= −ikα0+ (using gauge condition)
kp−
= −i √ δnk (3.164)
πT
This is invertible if we exclude n = 0 and k = 0, so we have fixed all but the gauge
transformation generated by L0 . Now we have, since αk+ = 0,
∞ ∞
X i X i
α̇k · α−k = α̇k · α−k , (3.165)
k=1
k k=1
k

where the (D − 2)-vectors αk are the transverse oscillator variables.

We also have,
∞
1 2 X 1
p2 + 2πT N ,


L0 = α0 + α−k · αk = (3.166)
2 k=1
2πT

where N is the level number:

∞
X
N= α−k · αk . (3.167)
k=1

For n 6= 0,
1X + −
1X
Ln = αk αn−k + αk− αn−k
+
+ αk · αn−k
2 2
k∈Z k∈Z
1 X
= α0+ αn− + αk · αn−k (using gauge condition). (3.168)
2
k∈Z

– 38 –
 √
We can solve this for αn− ; using p = πT α0 , we get
√
− πT X
αn = − αk · αn−k (n 6= 0). (3.169)
2p−
k∈Z

As we have solved the constraints Ln = 0 for n 6= 0, only the L0 = 0 constraint

will be imposed by a Lagrange multiplier in the gauge-fixed action, which is
Z ( ∞
)
X i 1
I = dt ẋm pm + α̇k · α−k − e0 p2 + M 2


(3.170)
k=1
k 2

where e0 = λ0 /(πT ) and

 
0 1
2
M = 2πT N = N/α α0 ≡ . (3.171)
2πT

Notice that the action does not involve αn− (for n 6= 0) but the Lorentz charges do.
Recall that the spin part of the Lorentz charge J mn is S mn = −2 ∞ i [m n]
P
k=1 k α−k αk . Its
non-zero components of S mn in light-cone gauge are (I, J, = 1, . . . D − 2)
∞
X i [I J]
S IJ = −2 α−k αk ,
k=1
k
∞
X i − I
−I I
αk− .


S =− α−k αk − α−k (3.172)
k=1
k

The canonical PB relations that we read off from the action (3.170) are

{xm , pn }P B = δnm , αkI , α−k

J
= −ik δ IJ .

PB
(3.173)

These may be used to compute the PBs of the Lorentz generators; since J = L + S
where {L, S}P B = 0, the PB relations among the components of S alone must be the
same as those of J . The PBs of S IJ are those of the Lie algebra of the transverse
rotation group SO(D − 2), and their PBs with S −K are those expected from the fact
that S −K is a (D − 2) vector. Finally, Lorentz invariance requires that

S −I , S −J

PB
= 0. (3.174)

This can be confirmed by making use of the PB relations

√ √
 − − πT −
 − I πT I
αk , α` P B = i (k − `) αk+j , αk , α` P B = −i ` αk+` . (3.175)
p− p−

This has to work because gauge fixing cannot break symmetries; it can only obscure
them.

– 39 –
3.9.1 Light-cone gauge for mixed Neumann/Dirichlet b.c.s
What changes if we change the boundary conditions to the mixed Neumann/Dirichlet
case? Assuming that Dirichlet b.c.s apply only in some, or all, of the (D − 2)
transverse directions9 we can still impose the gauge-fixing condition10

αk+ = 0 ∀k , (3.176)

and then proceed as before. We can again solve the constraints Ln = 0 (n 6= 0) for
αk− (k 6= 0). The L0 = 0 constraint, which still has to be imposed via a Lagrange
multiplier, is
1 1
0 = α02 + N = p̂2 + (T L)2 + 2πT N


(3.177)
2 2πT
where
 
p̂2 = pm̂ pm̂ , L2 = Lm̌  . (3.178)
In other words, the boundary conditions affect only the zero modes. The action is
Z ( ∞
)
X i 1
I = dt ẋm̂ pm̂ + α̇k · α−k − e0 p̂2 + M 2


, (3.179)
k=1
k 2

where
M 2 = (T L)2 + N/α0 . (3.180)
Classically, N ≥ 0 and the minimum energy configuration has N = 0. In this case,
M = T L, which can be interpreted as the statement that the minimal energy string
is a straight string stretched orthogonally between the two p-planes; since they are
separated by a distance L the potential energy in the string is T L.

3.9.2 Closed string in light-cone gauge

Now we fix the gauge invariances associated with Ln and L̃n for n 6= 0 by setting

αk+ = 0 & α̃k+ = 0 ∀k 6= 0 . (3.181)

This leaves unfixed the gauge invariances generated by L0 and L̃0 , which are now
∞
1 2 X p2
L0 = α0 + α−k · αk = +N,
2 k=1
8πT
∞
1 2 X p2
L̃0 = α̃0 + α̃−k · α̃k = + Ñ . (3.182)
2 k=1
8πT

9
This excludes only the case in which one end is free and other end is completely fixed, i.e.
tethered to a fixed point.
10
This is no longer looks so simple in terms of X and P .

– 40 –
Here we have used the closed string relation (3.116) between p and α0 = α̃0 . By
adding and subtracting the two constraints L0 = 0 and L̃0 = 0 we get the two
equivalent constraints
 
2
p + 4πT N + Ñ = 0 & Ñ − N = 0 , (3.183)

which will be imposed by Lagrange multipliers e0 and u0 in the gauge-fixed action.

The remaining constraints Ln = 0 and L̃n = 0 for n 6= 0 we solve for αk− and α̃k− , as
for the open string. The closed string action in light-cone gauge is therefore
Z ( ∞
)
X i   1 
˙ k · α̃−k − e0 p2 + M 2 − u0 N − Ñ

I = dt ẋm pm +


α̇k · α−k + α̃ ,
k=1
k 2
(3.184)
where
 
M 2 = 4πT N + Ñ
 
= 8πT N using Ñ = N constraint . (3.185)

4. Interlude: Light-cone gauge in field theory

We shall consider Maxwell’s equation, for the vector potential Am , and the linearised
Einstein equations for a symmetric tensor potential hmn , which may be interpreted
as the perturbation of the space-time metric about a Minkowski vacuum metric.
We choose light-cone coordinates (x+ , x− , xI ) (I = 1, . . . , D − 2). Recall that
for the particle we assumed that p− 6= 0, which is equivalent to the assumption that
the differential operator ∂− is invertible. We shall make the same assumption in the
application to field theory.

4.0.3 Maxwell in light-cone gauge

Maxwell’s equations are
D Am − ∂m (∂ · A) = 0 , m = 0, 1, . . . , D − 1 . (4.1)
They are invariant under the gauge transformation Am → Am + ∂m α. The light-cone
gauge is
A− = 0 . (4.2)
To see that the gauge is fixed we set to zero a gauge variation of the gauge-fixing
condition: 0 = δα A− = ∂− α. This implies that α = 0 if the differential operator ∂−
is invertible.
In this gauge the m = − Maxwell equation is ∂− (∂ · A) = 0, which implies that
∂ · A = 0 since we assume invertibility of ∂− . And since 0 = ∂ · A = ∂− A+ + ∂I AI ,
we can solve for A+ :
A+ = −∂−−1 (∂I AI ) . (4.3)

– 41 –
This leaves AI as the only independent variables. The m = + equation is D A+ = 0,
but this is a consequence of the m = I equation, which is

D AI = 0 , I = 1, . . . , D − 2 (4.4)

So this is what Maxwell’s equations look like in light-cone gauge: wave equations for
D − 2 independent polarisations.

4.0.4 Linearized Einstein in light-cone gauge

The linearised Einstein equations are

D hmn − 2∂(m hn) + ∂m ∂n h = 0 , hm ≡ ∂ n hnm , h = η mn hmn . (4.5)

They are invariant under the gauge transformation (Exercise: verify this)

hmn → hmn + 2∂(m ξn) . (4.6)

The light-cone gauge choice is

h−n = 0 (n = −, +, I) ⇒ h− = 0 & h = hJJ . (4.7)

In the light-cone gauge the “m = −” equation is −∂− hn + ∂− ∂n hJJ = 0, which

can be solved for hn :
hn = ∂n hJJ . (4.8)
But since we already know that h− = 0, this tells us that hJJ = 0, and hence that
hn = 0 and h = 0. At this point we see that the equations reduce to D hmn = 0,
but

h+ = 0 ⇒ h++ = −∂−−1 (∂I hI+ ) ,

hI = 0 ⇒ h+I = −∂−−1 (∂J hJI ) , (4.9)

so the only independent components of hmn are hIJ , and this has zero trace. We
conclude that the linearised Einstein equations in light-cone gauge are

D hIJ = 0 & hII = 0 . (4.10)

The number of polarisation states of the graviton in D dimensions is therefore

1 1
(D − 2)(D − 1) − 1 = D(D − 3) . (4.11)
2 2
For example, for D = 4 there are two polarisation states, and the graviton is a
massless particle of spin-2.

– 42 –
5. Quantum NG string
Now we pass to the quantum theory. We start by fixing the gauge invariances by
a variant of the light-cone gauge, which has the advantage of eliminating, prior to
quantization, all unphysical components of the oscillator variables. Then we consider
how the same results could be found by Lorentz-covariant quantization; this is the
“old covariant” approach.

5.1 Light-cone gauge quantization: open string

The canonical PB relations of the open string in Fourier modes are (3.173). These
become the canonical commutation relations

[x̂m , p̂n ] = iδnm ,

 I J 
α̂k , α̂−k = kδ IJ , (5.1)

where the hats now indicate operateors, and the hermiticity of the operators (X̂, P̂ )
requires that
α̂−k = α̂†k . (5.2)
A state of the string of definite momentum is the tensor product of a momentum
eigenstate |pi with a state in the oscillator Fock space, built upon the Fock vacuum
state |0i annihilated by all annihilation operators:

α̂k |0i = 0 ∀ k ∈ Z+ . (5.3)

We get other states in the Fock space by acting on the oscillator vacuum with the
creation operators α̂−k any number of times, and for any k > 0. This gives us a
basis for the entire infinite-dimensional space.
Next, we need to replace the level number N by a level number operator N̂ ,
but there is an operator ordering ambiguity; different orderings lead to operators N̂
that differ by a constant. We shall choose to call N̂ the particular operator that
annihilates the oscillator vacuum; i.e.
∞
X
N̂ = α̂−k · α̂k ⇒ N̂ |0i = 0 . (5.4)
k=1

So the oscillator vacuum has level number zero. Notice now that
h i
N̂ , α̂−k = k α̂−k . (5.5)

This tell us that acting on a state with any component of α̂−k raises the level number
by k, and this tells that N̂ is diagonal in the Fock state basis constructed in the
way described above, and that the possible level numbers (eigenvalues of N̂ ) are
N = 0, 1, 2, . . . , ∞. We can therefore organise the states according to their level
number. There is only one state in the Fock space with N = 0, the oscillator

– 43 –
 I
vacuum. At N = 1 we have the (D − 2) states α̂−1 |0i. At N = 2 we have the states
I I J
α̂−2 |0i and α̂−1 α̂−1 |0i. At N = 3 we have the states

I I J I J K
α̂−3 |0i , α̂−2 α̂−1 |0i , α̂−1 α̂−1 α̂−1 |0i , (5.6)

and so on.
A generic state of the string at level N in a momentum eigenstate takes the form

|pi ⊗ |ΨN i , (5.7)

where p is the D-momentum and ΨN some state in the oscillator Fock space with
level number N . The mass-shell constraint for such a state implies that p2 = −M 2 ,
where
M 2 = 2πT (N − a) . (5.8)
The constant a is introduced to take care of the operator ordering ambiguity in
passing from the classical to the quantum theory. We chose to define N̂ in a particular
way, such that its eigenvalues are non-negative integers, but there is nothing to tell
us that this is what we must use in the quantum mass-shell constraint (imposed as
a physical state condition, as for the particle). As mentioned, different definitions of
N would differ by a constant, so we introduce the constant a to allow for this.
In fact, if we had defined N̂ using the conventional Weyl ordering that leads to the
usual zero-point energy for a harmonic oscillator, we would find that its eigenvalues
are
∞
(D − 2) X
N+ k, (5.9)
2 k=1

This is because the (D − 2) oscillators associated to the pair (αk , α†k ) have angular
frequency |k|, and we have to sum over all oscillators. This would lead us to make
the identification
∞
(D − 2) X
−a= k (5.10)
2 k=1

The sum on the RHS is infinite, it would seem. In fact, it is ill-defined. One way to
define it is as the s → −1 limit of
∞
X
ζ(s) = k −s . (5.11)
k=1

When this sum converges, it defines an analytic function in the complex s-plane: the
Riemann zeta function. Remarkably, ζ(s) can be analytically continued to s = −1,
where it is finite; in fact
1
ζ(−1) = − . (5.12)
12

– 44 –
Using this in (5.10) we find that
(D − 2)
a= . (5.13)
24
This looks rather dubious, so let’s leave it aside for the moment and proceed
to analyse the string spectrum for arbitrary a, and level by level. We shall use the
standard notation
2πT = 1/α0 . (5.14)

• N = 0. There is one state, and hence a scalar, with M 2 = −a/α0 . For a > 0
(as (5.13) suggests) this scalar is a tachyon.

• N = 1. There are now (D − 2) states, α̂−1 |0i with M 2 = (1 − a)/α0 . The

only way that these states could be part of a Lorentz-invariant theory is if
they describe the polarization states of a massless vector (a massive vector has
(D − 1) polarisation states), so Lorentz invariance requires

a = 1. (5.15)

• N = 2. Since a = 1 the N = 2 states are massive, with M 2 = 1/α0 . The states

are those of (5.6), which are in the symmetric 2nd-rank tensor plus vector
representation of SO(D − 2). These states form a symmetric traceless tensor
of SO(D − 1) and hence describe a massive spin-2 particle11 .

We now know that the string ground state is a scalar tachyon, and its first excited
state is a massless vector, a “photon”. All higher level states are massive, and so
should be in SO(D − 2) representations that can be combined to form SO(D − 1)
representations (i.e. representations of the rotation group). We have seen that this
is true for N = 2 and it can be shown to be true for all N ≥ 2. The N = 1 states
are exceptional in this respect.
Notice that if a = 1 is used in (5.13) we find that D = 26. Remarkably, it is
indeed true that Lorentz invariance requires D = 26, as we shall now see.

5.1.1 Critical dimension

Because Lorentz invariance is not manifest in light-cone gauge it might be broken
when we pass to the quantum theory. To check Lorentz invariance we have to com-
pute the commutators of the quantum Lorentz charges Jˆmn . In fact, Jˆ = L̂ + Ŝ
and it is easy to see that [L̂, Ŝ] = 0 so we can focus on the spin Ŝ mn ; its components
should obey the same algebra as those of Jˆ, and this requires that
h i
Ŝ −I , Ŝ −J = 0 . (5.16)
11
A symmetric traceless tensor field of rank n is usually said to describe a particle of “spin n”
even though “spin” is not sufficient to label states in space times of dimension D > 4.

– 45 –
If the {, }P B → −i[, ] rule were to apply to these charges then Lorentz invariance
of the quantum string would be guaranteed because the classical theory is Lorentz
invariant, even in the light-cone gauge. But it does not apply because the S −I are
cubic in the transverse oscillators; a product of two of them is therefore 6th-order in
transverse oscillators, but the commutator reduces this to 4th order. The classical
PB computation gives zero for this 4th order term, but to achieve this in the quantum
theory we might have to change the order of operators, which would produce a term
quadratic in oscillators. So, potentially, the RHS of (5.16) might end up being an
expression quadratic in transverse oscillators.
Because of this possibility, we need to check (5.16); there is no guarantee that it
will be true. We can do this calculation once we have the quantum analogs of the PB
relations (3.175). The commutator [α̂k− , α̂`− ] is the one we have to examine carefully.
There is no ordering ambiguity in the quantum version of the expression (3.169) for
αn− as long as n 6= 0, so we are taking the commutator of well-defined operators as
long as k` 6= 0. Looking first at the k + ` = 0 case, we find that12
πT |p|2
   
 − −  2πT D − 2
k3 − k .


α̂k , α̂−k = 2k 2 + N̂ + 2 (5.17)
p− 2πT p− 24
Using the mass-shell condition in the operator form
 2  h
|p| p+ p− i
+ N̂ = a − 2 , (5.18)
2πT 2πT
which is valid
√ when the operators act on any physical state13 , and using the the fact
that p+ = πT α0− , we can rewrite (5.17) as
√    
 − −  πT − 2πT (D − 2) (D − 2)) 3
α̂k , α̂−k = −2k α + 2 k a− + k (5.19)
p− 0 p− 24 24
More generally, one finds that
√    
 − − πT − 2πT (D − 2) (D − 2) 3
α̂k , α̂` = − (k − `) α̂k+` + 2 k a− + k δk+`
p− p− 24 24
(5.20)
where 
1 n=0
δn = (5.21)
0 n 6= 0 .
Compare this result with the analogous PB relation of (3.175); he second term in
(5.20) has no classical counterpart. Using this result leads to the further result that
∞    
h
−I −J
i 4πT X (D − 2) 1 (D − 2) [I J]
Ŝ , Ŝ = 2 −2 k+ 2a − α̂−k α̂k , (5.22)
p− k=1 12 k 12
12
A very similar calculation will be explained in more detail later.
13
In the light-cone gauge used here the mass-shell constraint is the only physical-state condition.

– 46 –
which is zero for D > 3 iff
a = 1 & D = 26 . (5.23)
We therefore confirm that Lorentz invariance requires a = 1, but we now see that it
also requires D = 26; this is the critical dimension of the NG string.

From now on we drop the hats on operators.

5.1.2 Quantum string with mixed N/D b.c.s

We shall consider only the case for which both ends of an open string are constrained
to move in the same p-plane. In this case the light-cone gauge action is
Z ( ∞
)
m̂
X i 1 2


I = dt ẋ pm̂ + α̇k · α−k − e0 p̂ + 2πT N , (5.24)
k=1
k 2

where the level number N is exactly the same as it was for the open string with
free ends. The hats here are not “operator hats”; they indicate directions in the
(p + 1)-dimensional subspace of D-dimensional Minkowski space-time in which the
string ends (and centre-or-mass) move: m̂ = 0, 1, . . . , p.
Quantization proceeds exactly as for the string with free ends, except that the
mass-shell condition at given level N is now a wave-equation in the (p+1)-dimensional
Minkowski space-time. In particular, the mass-squared at level N is again 2πT (N −
a), and the N = 1 excited states are |p̂i ⊗ α−1 |0i, where α−1 is a transverse (D − 2)-
vector. However, the boundary conditions preserve only an SO(D − p − 1) sub-
group of the transverse rotation group SO(D − 2) (we’ll assume that p ≥ 2, so that
D − p − 1 < D − 2). This means that the generic level-1 state of (p + 1)-momentum
p̂ is h i
Iˆ Iˇ
|p̂i ⊗ AIˆ(p̂)α−1 + AIˇ(p̂)α−1 |0i . (5.25)
We can identify AIˆ as the (p−2) physical components of a (p+1)-vector potential, and
AIˇ as (D −p−1) scalars, all propagating in the Minkp+1 subspace of MinkD . Because
a massive photon would have (p − 1) physical components, it must be massless14 .,
and this again tells us that a = 1.
To check that the (p + 1)-dimensional Lorentz invariance is preserved in the
quantum theory we need to check that
h i
ˆ ˆ
S −I , S −J = 0 , (5.26)

and this again turns out to be true only if a = 1 and D = 26.

Because the b.c.s break translation invariance in the directions orthogonal to
the fixed p-plane on which the string ends, the string can lose or gain momentum in
14
We can’t use one of the (D − p − 1) scalars as the extra component of the massive photon field
because this would break the transverse rotation invariance

– 47 –
these directions. This was considered unphysical for many years, but there is now a
physical interpretation. The fixed p-plane is now a p-brane, of a special kind called
a D-brane (or Dp-brane) where the D here is for “Dirichlet”. What makes this
interpretation possible is that small fluctuations of a p-brane in D dimensions are
described by D − p − 1 massless scalar fileds “on the brane”, and this is precisely the
number of such fields that we have found at level-one; all higher levels give massive
fields. The ground state is still a tachyon, which indicates that NG D-branes are
unstable.

5.1.3 Quantum closed string

Finally, we consider light-cone gauge quantization of the closed string. There are
now two sets of oscillator operators, with commutation relations
 I J 
αk , α−k = kδ IJ = α̃kI , α̃−k
J
 
(5.27)

The oscillator vacuum is now

|0i = |0iR ⊗ |0iL , αk |0iR = 0 & α̃k |0iL = 0 ∀k > 0 . (5.28)

ˆ such that they annihilate the oscillator

We define the level number operators N̂ and Ñ
vacuum. Again, we can choose a basis for the Fock space built on the oscillator
vacuum for which these operators are diagonal, with eigenvalues N and Ñ . In the
space of states of definite p and definite (N, Ñ ) the mass-shell and level matching
constraints are
h  i  
p2 + 4πT (N − aR ) + Ñ − aL = 0 & (N − aR ) = Ñ − aL . (5.29)

Since we want |0i to be a physical state we must choose aL = aR = a, and then we

have
p2 + 8πT (N − a) = 0 & Ñ = N . (5.30)
This means that we can organise the states according to the level N , with M 2 =
8πT (N − a). We must do this respecting the level-matching condition Ñ = N . Let’s
look at the first few levels

• N = 0. There is one state, and hence a scalar, with M 2 = −4a/α0 .

• N = 1. There are now (D − 2) × (D − 2) states

I J
α−1 |0iR ⊗ α−1 |0iL (5.31)

We can split these into irreducible representations by taking the combinations

I J
[hIJ (p) + δIJ φ(p) + bIJ (p)] α−1 |0iR ⊗ α−1 |0iL , (5.32)

– 48 –
 where hIJ is symmetric traceless tensor, bIJ an antisymmetric tensor and φ a
scalar. The only way that these could be part of a Lorentz-invariant theory i s if
hIJ are the physical components of a massless spin-2 field because massive spin-
2 would require a symmetric traceless tensor of the full rotation group SO(D −
1). Then bIJ must be the physical components of a massless antisymmetric
tensor field, and φ a massless scalar (the dilaton).
Since we require M 2 = 0 we must choose a = 1 again15 . This means that the
ground state is a tachyon, indicating an instability of the Minkowski vacuum.

• N = 2. Since a = 1 the N = 2 states are massive, with M 2 = 8πT . Recalling

that the open string states at level-2 combined into a symmetric traceless tensor
of SO(D − 1), we see that the level-2 states of the closed string will combine
into those SO(D − 1) representations found in the product of two symmetric
traceless SO(D −1) tensors. This includes a 4-order symmetric traceless tensor
describing a particle of spin-4; there will be several lower spins too.

The most remarkable fact about these results is that the closed string spectrum
contains a massless spin-2 particle, suggesting that a closed string theory will be a
theory of quantum gravity. As for the open string, one finds that Lorentz invariance
is preserved only if D = 26 (the calculation needed to prove this is a repeat of the
open string calculation because the spin operator is a sum of a contribution from
“left” oscillators and a contribution from “right” oscillators). The ground state is
a tachyon, but the tachyon is absent in superstring theory, for which the critical
dimension is D = 10, and there are various ways to compactify dimensions so as to
arrive at more-or-less realistic models of gravity coupled to matter.

5.2 “Old covariant” quantization

Dirac’s method of dealing with first-class constraints would appear to allow us to
quantise the string in a way that preserves manifest Lorentz invariance. Let’s consider
the open string with free-end b.c.s. Recall that the action in terms of Fourier modes
is
 
Z  ∞
X X  1X
I = dt ẋm pm + α̇k · α−k − λ−n Ln , Ln = αk · αn−k , (5.33)

k=1
 2
n∈Z k∈Z

and that {Lk , L` }P B = −i (k − `) Lk+` . Applying the {, } → −i[, ] rule to the PBs of
the canonical variables, we get the canonical commutation relations

[xm , pn ] = iδnm ,
 m n 
αk , α−k = kη mn . (5.34)
15
This is usually given as a = 2 but that’s due to a different definition of a for the closed string.

– 49 –
Now we define the oscillator vacuum |0i by

αkm |0i = 0 ∀k > 0 (m = 0, 1, . . . , D − 1). (5.35)

m
The Fock space is built on |0i by the action of the creation operators α−k , but this
gives a space with many unphysical states since we now have D creation operators
for each k, whereas we know (from light-cone gauge quantization) that D − 2 suffice
to construct the physical states.
Can we remove unphysical states by imposing the physical state conditions

Ln |physi = 0 ∀n ? (5.36)

Notice that we do not encounter an operator ordering ambiguity when passing from
the classical phase-space function Ln to the corresponding operator Ln except when
n = 0, so the operator Ln is unambiguous for n 6= 0 and it is easy to see that

Ln |0i = 0 , n > 0. (5.37)

However, it is also easy to see [exercise: check these statements] that

1X
L−1 |0i ≡ αk · α−1−k |0i = α0 · α−1 |0i
2 k
 
1X 1 2
L−2 |0i ≡ αk · α−2−k |0i = α0 · α−2 + α−1 |0i , (5.38)
2 k 2

so it looks as though not even |0i is physical. In fact, there are no states in the
Fock space satisfying (5.36) because the algebra of the operators Ln has a quantum
anomaly, which is such that the set of operators {Ln ; n ∈ Z} is not “first-class”. That
is what we shall now prove.
Since the Ln are quadratic in oscillator variables, the product of two of them
is quartic but the commutator [Lm , Ln ] is again quadratic. That is what we expect
from the PB, which is proportional to Lm+n , but to get the operator Lm+n from the
expression that results from computing the commutator [Lm , Ln ], we may need to
re-order operators, and that would produce a constant term. So, we must find that

[Lm , Ln ] = (m − n) Lm+n + Amn (5.39)

for some constants Amn . We can compute the commutator using the fact that

[Lm , αk ] = −k αk+m . (5.40)

This can be verified directly but it also follows from the corresponding PB result
because no ordering ambiguity is possible either on the LHS or the RHS. Using this,

– 50 –
we find that
X
[Lm , Ln ] = ([Lm , αk ] · αn−k + αk · [Lm , αn−k ])
k
1X 1X
=− k αk+m · αn−k − (n − k) αk · αn+m−k . (5.41)
2 k 2 k

As long as n + m 6= 0 this expression is not affected by any change in the order of

operators, so it must equal what one gets from an application of the {, }P B → −i[, ]
rule. In other words, Amn = 0 unless m + n = 0. We can check this by using the
fact that α−k = αk† , so that

αk |0i = 0 ⇔ h0|α−k = 0 . (5.42)

From this we see that for m + n 6= 0

h0|αk+m · αn−k |0i = 0 = h0|αk · αn+m−k |0i (m + n 6= 0) (5.43)

and hence that h0|[Lm , Ln ]|0i = 0 unless m + n = 0. This tells us that Amn =
A(m)δm+n .
We now focus on the m + n = 0 case, for which

[Lm , L−m ] = 2mL0 + A(m) ⇒ A(−m) = −A(m) . (5.44)

Because of an operator ordering ambiguity, the operator L0 is only defined up to the

addition of a constant, so what we find for A(m) will obviously depend on how we
define the operator L0 . We shall define it as
∞
1 X
L0 = α02 + α−k · αk , (5.45)
2 k=1

but we should keep in mind that it is possible to redefine L0 by adding a constant to

it. We could now return to (5.41), set n = −m, and then complete the computation
to find A(m). This can be done, but it has to be done with great care to avoid
illegitimate manipulations of infinite sums. For that reason we here take an indirect
route.
First we use the Jacobi identity

[Lk , [Lm , Ln ]] + cyclic permutations ≡ 0 , (5.46)

to deduce that [Exercise]

[(m − n)A(k) + (n − k)A(m) + (k − m)A(n)] δm+n+k = 0 . (5.47)

Now set k = 1 and m = −n − 1 (so that m + n + k = 0) to deduce that

(n + 2)A(n) − (2n + 1)A(1)
A(n + 1) = n ≥ 2. (5.48)
n−1

– 51 –
This is a recursion relation that determines A(n) for n ≥ 3 in terms of A(1) and A(2),
so there are two independent solutions of the recursion relation. You may verify that
A(m) = m and A(m) = m3 are solutions, so now we have
[Lm , L−m ] = 2mL0 + c1 m + c2 m3 , (5.49)
for some constants c1 and c2 . Observing that
h0| [Lm , L−m ] |0i = h0|Lm L−m |0i = ||L−m |0i||2 , (5.50)
and that
1 p2
h0|L0 |0i = α02 = , (5.51)
2 2πT
we deduce that  2
2 p
||L−m |0i|| − m = c1 m + c2 m3 . (5.52)
πT
We can now get two equations for the two unknown constants (c1 , c2 ) by evalu-
ating ||L−m |0i||2 for m = 1 and m = 2. Using (5.38) we find that
1 pm pn p2
||L−1 |0i||2 = h0|p · α−1 p · α−1 |0i = h0|α1m α−1
n
|0i = , (5.53)
πT πT πT
and that
  
2 1 2 1 2
||L−2 |0i|| = h0| α0 · α2 + α1 α0 · α−2 + α−1 |0i
2 2
1 1
= h0|p · α2 p · α−2 |0i + h0|α12 α−1 2
|0i
πT 4
2p2 D
= + , (5.54)
πT 2
from which we see that
p2 2p2 D
||L−1 |0i||2 − = 0, ||L−2 |0i||2 − = , (5.55)
πT πT 2
and hence that
D D
c1 + c2 = 0 , c1 + 4c2 = ⇒ c2 = −c1 = . (5.56)
4 12
Inserting this result into (5.49), we have
D
m3 − m ,


[Lm , L−m ] = 2mL0 + (5.57)
12
and hence that
D
m3 − m δm+n .


[Lm , Ln ] = (m − n) Lm+n + (5.58)
12
This is an example of the Virasoro algebra. In general, the Virasoro algebra
takes the form
c
m3 − m δm+n ,


[Lm , Ln ] = (m − n) Lm+n + (5.59)
12
where c is the central charge. In the current context we get this algebra with c = D.

– 52 –
5.2.1 The Virasoro constraints
We have just seen that without breaking manifest Lorentz covariance, it is not possi-
ble to impose the physical state conditions that we would need to impose to eliminate
all unphysical degrees of freedom. In view of our light-cone gauge results, this should
not be too much of a surprise. We saw that the level-one states are the polarisation
states of a massless spin-1 particle. The covariant description of a massless spin-1
particle in terms of a vector field (which is what is needed to allow interactions)
necessarily involves unphysical degrees of freedom.
This argument suggests that we should aim to impose weaker conditions that
leave unphysical states associated to gauge invariances, but that remove all other
unphysical states. For correspondence with the classical NG string in a semi-classical
limit, these weaker conditions should have the property that

hΨ0 | (Ln − aδn ) |Ψi = 0 ∀n , (5.60)

for any two allowed states |Ψi and |Ψ0 i. Because of the operator ordering ambiguity
in L0 , we should allow for the possibility that the L0 operator of relevance here is
shifted by some constant relative to how we defined it in (5.45), hence the constant a
(which will turn out to be the same constant a that we introduced in the light-cone
gauge quantization). We can achieve (5.60) without encountering inconsistencies by
imposing the Virasoro constraints

Ln |Ψi = 0 ∀n > 0 & (L0 − a) |Ψi = 0 . (5.61)

Let’s call states |Ψi satisfying the Virasoro constraints “Virasoro-allowed”.

As for the light-cone gauge quantization, we can define the level number operator
∞
X
N= α−k · αk . (5.62)
k=1

This differs from the level-number operator of the light-cone gauge in that it includes
all D components of the oscillator annihilation and creation operators, not just the
D − 2 transverse components. Otherwise it plays a similar role. We can choose a
basis such that the operators p and N are diagonal, in which case p and N will mean
their respective eigenvalues. For given p and N , the Virasoro mass-shell condition is

p2 + M 2 = 0 , M 2 = 2πT (N − a) . (5.63)

The mass-squared is a linearly increasing function of the level number, as in the

light-cone case. Let’s consider the first few levels. The only N = 0 state in the
oscillator Fock space is |0i (which gives the string states |pi ⊗ |0i).
The general N = 1 state is
m
Am (p)α−1 |0i . (5.64)

– 53 –
The norm-squared of this state is
||A · α−1 |0i||2 = Am An h0|α1m α−1
n
|0i = η mn Am An . (5.65)
This could be negative but we still have to impose the other Virasoro conditions
Lk (A · α−1 )|0i = 0 for k > 0. For k ≥ 2 these conditions are trivially satisfied, but
for k = 1 we find that
0 = L1 (A · α−1 )|0i = A · α0 |0i ⇒ p · A = 0. (5.66)
Let’s now consider the implications of this for the constant a:
• a > 1. Then M 2 < 0 so p is spacelike. In a frame where p = (0; p, 0) the
constraint p · A = 0 is equivalent to A1 = 0 and then
||A · α−1 |0i||2 = −A20 + |A|2 (5.67)
which allows “ghosts”, i.e. states of negative norm (e.g. those with A = 0 but
A0 6= 0). This implies a violation of unitarity (non-conservation of probability)
so we should not allow a > 1.

• a = 1. Then M 2 = 0, which agrees with the light-cone gauge result. In this

case p is null. In a frame where p = (1; 1, 0) the constraint p · A = 0 implies
that A0 = A1 , and hence that
||A · α−1 |0i||2 = |A|2 ≥ 0 . (5.68)
0 1
There are no ghosts but the state |χi = (α−1 + α−1 )|0i is one with p · A = 0
but has zero-norm. This state is also orthogonal to all other N = 1 states
satisfying P · A = 0, so given one such state |ψi we have
|||ψi + c|χi||2 = |||ψi||2 (5.69)
for any complex constant c. This means that we may identify the physical
polarisation with equivalence classes of states, where any two states that differ
by a null state are considered equivalent. The space of such equivalence classes
has dimension D − 2.
What we are finding here is essentially the Gupta-Bleuler quantization of elec-
trodynamics (in D space-time dimensions).

• a < 1. Then M 2 > 0 , so p is timelike. In a frame where p = (p, ~0), the

constraint p · A = 0 implies that A0 = 0, so
||A · α−1 |0i||2 = |A|
~ 2 ≥ 0, (5.70)
and now there are no non-zero zero-norm states satisfying p · A = 0.
There is nothing obviously unphysical about a < 1 but it involves the inclusion
of states that are absent in the light-cone gauge. The quantization procedure
has introduced new degrees of freedom that were absent classically.

– 54 –
 Let’s now look at the level-2 states. The general level-2 state is
m n m


|Ψ2 i = Amn α−1 α−1 + Bm α−2 |0i (5.71)
This is trivially annihilated by Lk for k > 2. However, L1 |Ψ2 i = 0 imposes the
condition
Bn = −α0m Amn , (5.72)
and L2 |Ψ2 i = 0 imposes the condition
η mn Amn = −2α0 · B (5.73)
This means that only the traceless part of Amn is algebraically independent, so the
dimension of the Virasoro-allowed level-2 space is
 
1 1
D(D + 1) − 1 = D(D − 1) − 1 + D (5.74)
2 2
The dimension is D larger than the physical level-2 space that we found from light-
cone gauge quantization (that was spanned by the polarisation states of a massive
spin-2 particle, so a symmetric traceless tensor of the SO(D − 1) rotation group).
However, equivalence with the light-cone gauge results is still possible if there are
sufficient null states, and no ghosts.
To analyse this we need to consider the norm-squared of |Ψ2 i, which is
|||Ψ2 i||2 = 2Amn Amn + 2B 2 . (5.75)
Then we need to consider the implications for this norm of (5.72) and (5.73). We
will not pursue this further, but the final result is that there are no ghosts only if
D ≤ 26 and then there are sufficient null states for equivalence with the light-cone
gauge results iff D = 26.
It is simple to see that there are ghosts if D > 26. Consider the special case of
(5.71) for which
Amn = η mn + k1 α0m α0n , B m = k2 α0m . (5.76)
The conditions (5.72) and (5.73) determine the constants (k1 , k2 ) to be
D+4 D−1
k1 = , k2 = , (5.77)
10 5
and then one finds that the norm-squared is
2
− (D − 1) (D − 26) . (5.78)
25
This is negative for D > 26, so the state being considered is a ghost. It is null for
D = 26. For D < 26 the state has positive norm but it now counts as a physical
state that increases the dimension of the space of level-2 states. For D < 26 there are
more states than those that can be found from quantization of the physical degrees
of freedom of the classical NG string, so it is unclear to what extent any “sub-critical
string theory” would be related to strings.

– 55 –
6. Interlude: Path integrals and the point particle
Let A(X) be the quantum-mechanical amplitude for a particle to go from the origin
of Minkowski coordinates to some other point in Minkowski space-time with cartesian
coordinates X. As shown by Feynman, A(X) has a path-integral representation. In
the case of a relativistic particle of mass m, with phase-space action I[x, p; e] we have
Z Z
A(x) = [de] [dxdp] eiI , x(0) = 0 , x(1) = X . (6.1)

Here we are parametrising the path such that it takes unit parameter time to get
from the space-time origin to the space-time point with coordinates X. The integrals
have still to be defined, but we proceed formally for the moment.
We now allow t to be complex and we “Wick rotate”: first set t = −it̃ to get
Z i  
m i 2 2



I= dt̃ ẋ pm + e p + m ẋ = dx/dt̃ . (6.2)
0 2

As it stands, t̃ is pure imaginary, but we can rotate the contour in the complex t̃-
plane back to the real axis; if we choose to call this real integration variable t then
this procedure takes
Z 1  
m 1 2 2


−iI → dt −iẋ pm + e p + m = IE (6.3)
0 2

where IE is the “Euclidian action”. Notice that on successive elimination of p and e,

Z 1 √
1 1  −1 2
Z
2

IE [x, p; e] → dt e ẋ + m e → m dt ẋ2 , (6.4)
2 0 0

which is positive. The amplitude A is now given by the “Euclidean path integral”
Z Z
A = [de] [dxdp]e−IE . (6.5)

We will fix the gauge invariance by setting e = s for constant s. As discussed

previously, the variable s is gauge-invariant, so it is not possible to use gauge invari-
ance to bring it to a particular value, so we have to integrate over s, which (being
proportional to the elapsed proper time) could be any number from zero to infinity.
In other words, we can use gauge invariance to reduce the functional integral over
e(t) to an ordinary integral over s. We now have
Z ∞ Z R1
ds [dxdp]e 0 {iẋ pm − 2 (p +m )} .
m s 2 2
A(x) = (6.6)
0

This is not quite right, for reasons to be explained soon, but it will suffice for the
moment.

– 56 –
 R
To define the [dx dp] integral we first approximate the path in some way. We
could do this by n straight-line segments. We would then have n D-momentum in-
tegral to do (one for each segment) and (n − 1) integrals over the D-vector positions
of the joins. This illustrates the general point that in any multiple-integral approx-
imation to the phase-space path integral there will be some number of phase-space
pairs of integrals plus one extra D-momentum integral, which is the average of p(t).
Consider the n = 1 case, for which
s
−IE = iX m Pm − P 2 + m2 .


x(t) = Xt & p(t) = P ⇒ (6.7)
2
The only free variable on which the Euclidean action depends is the particle’s D-
R R
momentum P , so [dxdp] is approximated by the momentum-space integral dD P ,
and we find that
Z ∞
eiX·P
Z Z
D iX·P − 2s (P 2 +m2 )
A1 (X) = d P e ds e ∝ dD P 2 , (6.8)
0 P + m2

which is the Fourier transform of the standard momentum-space Feynman propagator

for a particle of mass m and zero spin.
That was just the n = 1 approximation! A simpler alternative to approximation
by segments (which differs from it only for n > 1) is approximation by polynomials,
nth order for x(t) and (n − 1)th order for p(t). Consider the n = 2 case:
 
2 1
x(t) = (X − x1 ) t + x1 t & p(t) = P + 2q t − . (6.9)
2

Notice that x(t) satisfies the b.c.s and P is the integral of p(t) (i.e. the average
D-momentum). We have to integrate over the pair (x1 , q) and P . Using these
expressions, we find that
s
s 1
−IE = iX · P − P 2 + m2 − p21 − x21 , (p1 = q − ix1 /s). (6.10)
2 6 6s
This gives
Z Z ∞ Z
x2
Z
sp2

D iX·P − 2s (P 2 +m2 ) D − 6s1 D − 1
A2 (X) = d Pe ds e d x1 e d p1 e 6 . (6.11)
0

The bracketed pair of Gaussian integrals is an s-independent constant, so A2 (X) ∝

A1 (X). One finds, similarly, that An (X) = A1 (X). Taking the n → ∞ limit we then
have A(X) ∝ A1 (X), so A(X) is the Feynman propagator in configuration space.

6.1 Faddeev-Popov determinant

Now we return to the problem of gauge-fixing. The problem with the formula (6.1)
is that, because of gauge-invariance, we are integrating over too many functions.

– 57 –
Implicitly, we are integrating over functions α(t) that are maps from the (one-
dimensional ) gauge group to the worldline. If this integral were explicit we could
just omit the integral, but it is only implicit, so it is not immediately obvious how
we should proceed.
Since we can choose a gauge for which e(t) = s, for variable constant s, it must
be possible to write an arbitrary function e(t) as a gauge transform of e = s:

e(t) = s + α̇(t) = es [α(t)] . (6.12)

We have now expressed the general e(t) in terms of the gauge group parameter α(t)
and the constant s. As a corollary, we have
Z Z ∞ Z
[de] = ds [dα] ∆F P (6.13)
0

where ∆F P is the Jacobian for the change of variables from e(t) to {s, α(t)}:
 
δes [α(t)]
∆F P = det . (6.14)
δα(t0 )

This is the Faddeev-Popov determinant. Using this result in the formula

Z
1 = [de]δ[e(t) − s] , (6.15)

which definees what we mean by the delta functional, we deduce that

Z ∞ Z
1= ds [dα] ∆F P δ[e(t) − s] (6.16)
0

We now return to the initial path-integral expression for A(X) and “insert 1”
into the integrand; i.e. we insert the RHS of (6.16) to get
Z Z ∞ Z Z
A(X) = [de] ds [dα] ∆F P δ[e(t) − s] [dxdp]eiI . (6.17)
0

Re-ordering the integrals and using the delta functional to do the [de] integral (this
sets e = s elsewhere in the integrand) we get
Z Z ∞ Z
A(X) = [dα] ds ∆F P [dxdp]eiIe=s . (6.18)
0

By these manipulations we have made explicit the integral over maps from the gauge
group to the worldline, so we can remedy the problem of too many integrals by
simply omitting the [dα] integral. This gives us
Z ∞ Z
A(X) = ds ∆F P [dxdp]eiIe=s . (6.19)
0

– 58 –
The ∆F P factor was missing from (6.6), which is why that formula was “not quite
right”, but the ∆F P factor only effects the normalisation of A(X), which anyway
depends on the detailed definitions of the path integral measures.
Although the FP determinant is not very relevant to the computation of A(X)
it is relevant to other computations, and it is very important to the path-integral
quantization of the NG string, which we will get to soon.

6.2 Fadeev-Popov ghosts

Let (bi , ci ) (i = 1, . . . , n) be n pairs of anticommuting variables. This means that

bi , cj = 0 ,
  i j
{bi , bj } = 0 , c , c = 0 ∀ i, j = 1, . . . , n , (6.20)

where {, } means anticommutator: {A, B} = AB + BA. Any function of anticom-

muting variables has a terminating Taylor expansion because no one anticommuting
variable can appear twice. Consider the n = 1 case

f (b, c) = f0 + bf1 + cf−1 + bcf˜0 , (6.21)

where (f0 , f±1 , f˜0 ) are independent of both b and c. Then

∂ ∂ ∂
f = f1 + cf˜0 ⇒ f = f˜0 (6.22)
∂b ∂c ∂b
Essentially, a derivative with respect to b removes the part of f that is independent of
b and then strips b off what is left. However, we should move b to the left of anything
else before stripping it off; this is equivalent to the definition of the derivative as a
“left derivative”. Using this definition we have
∂ ∂ ∂
f = f−1 − bf˜0 ⇒ f = −f˜0 . (6.23)
∂c ∂b ∂c
There is minus sign relative to (6.22) because we had to move c to the left of b. This
result shows that  
∂ ∂
, = 0. (6.24)
∂b ∂c
That is, partial derivatives with respect to anticommuting variables anti-commute.
We can also integrate over anticommuting variables. The (Berezin) integral
over an anticommuting variable is defined to the same as the partial derivative with
respect to it. Consider, for example, the Gaussian integral
Z   
−bi M i j cj ∂ ∂ ∂ ∂ i
n n
d cd b e = ··· n
· · · 1 e−bi M j cj . (6.25)
∂bn ∂b1 ∂c ∂c

If we expand the integrand in powers of bM c the expansion terminates at the nth

term, which is also the only term that contributes to the integral because it is the

– 59 –
only one to contain all bi and all ci . Because of the anti-commutativity of the partial
derivatives, we then find that
Z
i 1
dn c dn b e−bi M j cj ∝ εi ···i M i1 j1 · · · M in jn εj1 ···jn = det M (6.26)
n! 1 n
We can use a functional variant of this result to write the FP determinant as a
Gaussian integral over anticommuting “worldline fields” b(t) and c(t):
Z  Z 1 Z 1 
0 0 0 0
det [δ(t − t )∂t ] = [dbdc] exp − dt dt b(t ) [δ(t − t )∂t ] c(t)
0 0
Z  Z 1 
= [dbdc] exp − dt bċ . (6.27)
0

The anticommuting worldline fields are known collectively as the FP ghosts, although
it is useful to distinguish between them by calling c the ghost and b the anti-ghost.
N.B. There is no relation between the FP ghosts and the ghosts that
appear in the NG string spectrum for D > 26. The same word is being
used for two entirely different things!
Using (6.27) in the expression (6.19) we arrive at the result
Z ∞ Z Z
A(X) = ds [dxdp] [dcdb] eiIqu , (6.28)
0

where the “quantum” action is16

Z
s 2
Iqu = dt {ẋm pm + ibċ − Hqu } , p + m2 .


Hqu = (6.29)
2
We now have a mechanical system with an extended phase-space with additional,
anticommuting, coordinates (b, c). The factor of i multiplying the bċ term stems
from the convention (and it is only a convention) that a product of two “real” anti-
commuting variables is “imaginary”, so an i is needed for reality17 .
There is a simple extension of these ideas to the general mechanical system with
n first class constraints ϕi imposed by Lagrange multipliers λi . The FP operator is
found by varying the gauge-fixing condition with respect to the gauge parameter, so
in the gauge λi = λ̄i , for constants λ̄i , we find that
δ λi (t0 ) 
  
∆F P = det (6.30)
δj (t) λi =λ̄i
16
We can see now why the FP determinant does not affect our previous computation of A(X); it
changes only the overall normalisation factor, which we did not compute.
17
Of course, an anticommuting number cannot really be real; it is “real” if we declare it to be
unchanged by complex conjugation. Given two such anti-commuting numbers µ and ν we may
construct the complex anti-commuting number µ + iν, which will then have complex conjugate
µ − iν. According to the convention, the product iµν is real.

– 60 –
where δ λi is the gauge variation of λi with parameters i , given in (2.43). This gives
Z
0 i k i
[dbdc] eiIF P [b,c] ,


∆F P = det δ(t − t ) δj ∂t + λ̄ fjk ∝ (6.31)

where the FP action is

Z
dt ibi ċ + λ̄k cj fjk i .
  
IF P [b, c] = (6.32)

This must be added to the original action to get the “quantum action”
Z
Iqu = dt q̇ I pi + ibi ċi − Hqu , Hqu = λ̄k ϕk + icj fjk i bi .


(6.33)

We now have an action for a mechanical system with an extended phase space
(actually a superspace) for which some coordinates are anticommuting. On this space
we have the following closed non-degenerate (i.e. invertible) 2-form

Ω = dpm ∧ dxm + dbi ∧ dci . (6.34)

This is “orthosymplectic” rather than “symplectic” because the anticommutativity

of b and c means that18
dbi ∧ dci = dci ∧ dbi . (6.35)
This leads to a canonical Poisson bracket for b and c that is symmetric rather than
antisymmetric
bi , cj P B = cj , bi P B = −iδij .
 
(6.36)
The PB defined by the inverse of an orthosymplectic 2-form is no longer a Lie
bracket because it is not always antisymmetric, but it is a super-Lie bracket satisfying
a super-Jacobi identity. Let C, C 0 , C 00 be commuting functions on phase space, and
let A, A0 , A00 be anticommuting functions on phase space. Then the super-Jacobi
identity states that

{C, {C 0 , C 00 }P B }P B + cyclic perms. ≡ 0

{A, {C, C 0 }P B }P B + cyclic perms. ≡ 0
{A, {A0 , A00 }P B }P B + cyclic perms. ≡ 0 , (6.37)

but

{C, {A, A0 }P B }P B − {A0 , {C, A}P B }P B + {A, {A0 , C}P B }P B ≡ 0 . (6.38)

We get a minus sign in the second term because we had to cycle A00 past A0 , and this
sign is changed back to plus in the next term for the same reason.
18
The usual minus sign coming from the antisymmetry of the wedge product of 1-forms is cancelled
by the minus sign coming from changing the order of b and c.

– 61 –
 In the quantum theory, these identities become
h h ii
Ĉ, Ĉ 0 , Ĉ 00 + cyclic perms. ≡ 0
h h ii
0
Â, Ĉ, Ĉ + cyclic perms. ≡ 0 (6.39)

and
[A, {A0 , A00 }] + cyclic perms. ≡ 0 (6.40)
but
[C {A, A0 }] − {A0 [C, A]} + {A [A0 , C]} ≡ 0 . (6.41)
All of these identities hold for arbitrary operators Â, Â0 , Â00 and Ĉ, Ĉ 0 , Ĉ 00 ; e.g. for
arbitrary square matrices [Exercise: check this].

6.3 BRST invariance

The “quantum” point particle action (6.29) is invariant under the following trans-
formations with constant anticommuting parameter Λ:
1 2
p + m2 Λ ,


δΛ x = iΛcp , δΛ p = 0 , δΛ b = − δΛ c = 0 . (6.42)
2
Notice that the transformations of (x, p) are gauge transformations with parameter
α(t) = −iΛc(t); the factor of i is included because Λ is assumed to anticommute
with c and we use the convention that complex conjugation changes the order of
anticommuting numbers, so we need an i for “reality”. Allowing Λ to be t-dependent
we find [Exercise: check this]
Z
i
dt Λ̇ p2 + m2 c .


δΛ Iqu = (6.43)
2
This confirms the invariance for constant Λ and tells us that the (anti-commuting)
Noether charge is
1
QBRST = c p2 + m2 .


(6.44)
2
This is the BRST charge19 . It generates the BRST transformations (6.42) via
Poisson brackets defined for functions on the extended phase space. Let’s check this:
 
1 2
δΛ x = {x, iΛQBRST }P B = iΛc x, p = iΛcp
2 PB
i 1
δΛ b = {b, iΛQBRST }P B = − Λ {b, c}P B p2 + m2 = − p2 + m2 Λ . (6.45)



2 2
BRST symmetry seems rather mysterious, but it’s not just a special feature of
the point particle. Consider the quantum action (6.33) of the general mechanical
19
BRST stands for Becchi, Rouet, Stora and Tyutin, who discovered BRST symmetry in the
context of YM theory.

– 62 –
model with first class constraints (which includes, as we have seen, the NG string).
If we assume that the constraint functions ϕi span a Lie algebra then the structure
functions fijk will be constants satisfying (as a consequence of the Jacobi identity)

f[ij ` fk]` m ≡ 0 . (6.46)

In this case20 the action is invariant under the transformations generated by

i
QBRST = ci ϕi + ci ck fki j bj , (6.47)
2
which satisfies
{QBRST , QBRST }P B = 0 . (6.48)
This is not a trivial property of QBRST because the PB is symmetric under exchange
of its two arguments if these are both anticommuting.
We still need to check the invariance. Because the variation of (q̇ I pI + ibċ) is
guaranteed to be a total time derivative by an extension to the extended phase-space
of the lemma summarised by (2.39), we need only check that

Hqu = λ̄k ϕk + icj fjk i bi .

0 = {QBRST , Hqu }P B , (6.49)

This can be checked directly, but an alternative is to first check that

Hqu = i λ̄i bi , QBRST P B ,


(6.50)

and then use the super-Jacobi identity to show that

1
QBRST , λ̄i bi , QBRST P B P B = {QBRST , QBRST }P B , λ̄i bi P B ,
 
(6.51)
2
which is zero by (6.48).

6.4 BRST Quantization

Consider first general model with “quantum” action (6.33). The canonical commu-
tation relations for the extended phase-space coordinates are

[x̂m , p̂n ] = iδnm , {b̂i , ĉj } = δij . (6.52)

In addition,
ĉi , ĉj = 0 .

{b̂i , b̂j } = 0 (6.53)
The Hamiltonian now becomes the operator
 
Ĥ = λ̄k ϕ̂i + iĉj fjk i b̂i . (6.54)

20
There is still a BRST charge if the structure functions are not constants, but it is more com-
plicated.

– 63 –
Notice that Ĥ commutes with the ghost number operator

ngh = ĉi b̂i . (6.55)

This operator also has the property that

h i
ngh , ĉi = ci
 
ngh , b̂i = −b̂i , (6.56)

so the c-ghosts have ghost number 1 and the b-ghosts have ghost number −1.
The BRST charge becomes the following operator of unit ghost number:
i
Q̂BRST = ĉi ϕ̂i + ĉi ĉk fki j b̂j . (6.57)
2
Assuming that the PB to (anti)commutator rule applies, we learn from (6.48) that

Q̂2BRST = 0 . (6.58)

This is the fundamental property of the BRST charge. As a consequence of this

property, it is consistent to impose the physical state condition

Q̂BRST |physi = 0 . (6.59)

This condition has the following motivation. The Hamiltonian operator Ĥ is gauge-
dependent21 so its matrix elements cannot be physical. However, the quantum version
of (6.50) is h i
Ĥ = λ̄i b̂i , Q̂BRST , (6.60)

so it follows from (6.59) that, for any two physical states |physi and |phys0 i,

hphys0 |Ĥ|physi = 0 . (6.61)

In other words, the BRST physical state condition ensure that all physical matrix
elements are gauge-independent.
The “physical state” condition (6.59) does not actually remove all unphysical
states because for any state |χi the state Q̂BRST |χi will be “physical”, by the defi-
nition (6.59), as a consequence of the nilpotency of QBRST , but it will also be null if
we assume an inner product for which Q̂BRST is hermitian:
 2
†
Q̂BRST |χi = hχ|Q̂BRST Q̂BRST |χi
 
 
= hχ|Q̂2BRST |χi if Q̂†BRST = Q̂BRST
=0 (6.62)
21
Although the constants λ̄i may be gauge invariant if the gauge condition λi = λ̄i is assumed
to hold for all t, these constants could still be changed locally. The main point is that different
gauge-fixing conditions lead to a different Ĥ.

– 64 –
So we should really define physical states as equivalence classes (cohomology classes
of Q̂BRST ), where the equivalence relation is

|Ψi ∼ |Ψi + Q̂BRST |χi (6.63)

for any state |χi. This is consistent because Q̂BRST |χi is orthogonal to all states
that are physical by the definition (6.59).
Let’s now see how these ideas apply to the point particle. In a basis for which
(x̂, ĉ) are diagonal, with eigenvalues (x, c), the the canonical (anti)commutation re-
lations are realised by the operators
∂
p̂m = −i∂m , b̂ = , (6.64)
∂c
acting on wavefunctions Ψ(x, c), which we can expand as

Ψ(x, c) = ψ0 (x) + cψ1 (x) . (6.65)

The BRST charge is now

1
Q̂BRST = − c  − m2 ,


(6.66)
2
so the physical state condition is c( − m2 )ψ0 (x) = 0, which implies that ψ0 is
a solution to the Klein-Gordon equation. We learn nothing about ψ1 (x), but the
equivalence relation (6.63) tells us that

ψ1 ∼ ψ1 +  − m2 χ0 ,


(6.67)

for any function χ0 (x), which implies that ψ1 is equivalent to zero unless it too is
a solution of the KG equation22 . So we actually get a doubling of the expected
physical states (solutions of the KG equation). For this reason, we have to impose
the additional condition

b |physi = 0 ⇒ ψ1 = 0 . (6.68)

All this depends on a choice of inner product for which Q̂BRST is hermitian,
despite being nilpotent. This is achieved in the point particle case by the choice
Z Z Z
0 ∗ 0 ∂
hΨ|Ψ i = d x dc Ψ Ψ = dD x
D
[Ψ∗ Ψ0 ] . (6.69)
∂c

With respect to this inner product, the operators b̂ and ĉ are hermitian, and hence
Q̂BRST is hermitian. Using this inner product we can construct a field theory action
22
Expand both sides in terms of eigenfunctions of the KG operator, and compare coefficients; all
coefficients in the expansion of ( − m2 )χ0 are arbitrary except the coefficients of zero modes of
the KG operator, which are zero.

– 65 –
from which the BRST physical state condition emerges as a field squation. This
action is

S[Ψ(x, c)] = hΨ|Q̂BRST |Ψi

Z Z
1 D ∂ 
ψ0 c  − m2 ψ0


= d x
2 ∂c
Z
1
dD x ψ0  − m2 ψ0 ,


= (6.70)
2
which is the Klein-Gordon action.

7. BRST for the NG string

We now aim to use the conformal gauge in a path-integral approach to quantisation
of the closed NG string. Recall that the conformal gauge for the NG string is λ± = 1,
and that in this gauge the gauge variation of λ± is
√
δλ± = ∓ 2 ∂∓ ξ ± . (7.1)

The FP contribution to the action is therefore

√ Z I n o
IF P = 2 dt dσ ib∂+ c + b̃∂− c̃ . (7.2)

Notice that it is precisely because the conformal gauge fails to completely fix the
gauge invariance that we get a non-trivial FP action. Whenever the gauge is fixed
completely, the FP action is one for which the FP ghosts can be trivially eliminated;
that’s not the case here because we can’t invert ∂± , and that is also the reason that
there is a residual conformal symmetry. From this fact, one can see that the FP
ghosts are, in a sense made precise by the BRST formalism, subtracting out the
residual unphysical degrees of freedom that survive the gauge-fixing condition.
Adding the FP action we get the “quantum” action for the closed NG string in
conformal gauge
Z I n o
Iqu = dt dσ ẊPm + ibċ + ib̃c̃˙ − Hqu ,
P2  
Hqu = + T (X 0 )2 − i bc0 − b̃c̃0 . (7.3)
2T
From this action we can read off the PB relations; in particular

{b(σ), c(σ 0 )}P B = −iδ(σ − σ 0 ) = {b̃(σ), c̃(σ 0 )}P B . (7.4)

The BRST charge can be written as

QBRST = Q− + Q+ , (7.5)

– 66 –
and it follows from the general formula, using the fact that the algebra of constraint
functions is Diff 1 ⊕ Diff 1 , that
I I n o
0 0
Q− = dσ {cH− + icc b} , Q+ = dσ c̃H+ − ic̃c̃ b̃ . (7.6)

Using the PB relations obeyed by H± , given in (3.24), it is not difficult to verify that

{Q± , Q± }P B = 0
⇒ {QBRST , QBRST }P B = 0 . (7.7)
{Q+ , Q− }P B = 0

We now pass to the Fourier mode formulation. In addition to the Fourier series
for P ± T X 0 , we will need the Fourier series expansions
X 1 X ikσ
c= eikσ ck , b= e bk ,
2π
k∈Z k∈Z
X 1 X −ikσ
c̃ = e−ikσ c̃k , b̃ = e b̃k . (7.8)
2π
k∈Z k∈Z

The “quantum” action in terms of Fourier modes is

 
Z ∞
 X i X   
Iqu = dt ẋm pm + α̇k · α−k + α̃˙ k · α̃−k + i b−n ċn + b̃−n c̃˙ n − Hqu ,



k=1
k 
n∈Z
  h i
Hqu = L0 + N(gh) + L̃0 + Ñ(gh) , (7.9)

where the ghost level numbers are

∞
X ∞
X
N(gh) = k (b−k ck + c−k bk ) , Ñ(gh) = k(b̃−k c̃k + c̃−k b̃k ) . (7.10)
k=1 k=1

It will be convenient to define

L0 = L0 + N(gh) = 21 α02 + N + N(gh)


⇒ Hqu = L0 + L̃0 . (7.11)
L̃0 = L̃0 + Ñ(gh) = 21 α02 + Ñ + Ñ(gh)

We can now read off the PBs of the Fourier modes. For the new, anticommuting,
variables we have
n o
{cn , b−n }P B = −i , c̃n , b̃−n = −i , (n ∈ Z). (7.12)
PB

Notice that n = 0 is included, although the (anti)ghost zero modes (b0 , c0 ) and (b̃0 , c̃0 )
do not appear in the Hamiltonian. These anticommutation relations are equivalent
to (7.4).

– 67 –
 A Fourier decomposition of Q± yields the result
X 1 XX
Q− = c−n Ln − (p − q) c−p c−q bp+q ,
2
n∈Z p∈Z q∈Z
X 1 XX
Q+ = c̃−n L̃n − (p − q) c̃−p c̃−q b̃p+q . (7.13)
2
n∈Z p∈Z q∈Z

Let us rewrite these expressions as

X 1 (gh)
 X 1 (gh)

Q− = Lm + Lm c−m , Q+ = L̃m + L̃m c−m , (7.14)
2 2
m∈Z m∈Z

where
X X
L(gh)
m = (m − n) bm+n c−n , L̃(gh)
m = (m − n) b̃m+n c̃−n . (7.15)
n n

Notice that
(gh) (gh)
L0 ≡ N(gh) , L̃0 ≡ Ñ(gh) . (7.16)
A check on the expressions for Q± is to verify that

L0 = i {b0 , Q− }P B n  o
⇒ Hqu = i b0 + b̃0 , QBRST , (7.17)
L̃0 = i{b̃0 , Q+ }P B PB

which is the formula (6.50) for our case. In fact, if we define

Lm = Lm + L(gh)
m , L̃m = L̃m + L̃(gh)
m , (7.18)

then we have the generalisation

Lm = i {bm , Q− }P B , L̃m = i{b̃m , Q+ }P B . (7.19)

A calculation using the PBs (7.12) shows that

 (gh) (gh) (gh)
Lm , Ln PB
= −i (m − n) Lm+n , (7.20)
(gh)
and similarly for L̃m , from which it follows that

{Lm , Ln }P B = −i (m − n) Lm+n
{Ln , L̃m }P B = 0
{L̃m , L̃n }P B = −i (m − n) L̃m+n . (7.21)

This was to be expected; the conformal gauge preserves a residual conformal invari-
ance, which is preserved by the FP ghost action. The FP ghosts therefore contribute
to the associated Noether charges, which obey the same conformal algebra as before.

– 68 –
 In fact, the Ln are the Fourier modes of the non-zero components of the energy-
momentum stress tensor of the quantum action (7.3). The PBs of the Lm with
the Fourier modes of the various fields determine the transformations of these fields
under the residual conformal invariance of the conformal gauge. For example

{Lm , αn }P B = inαn+m ,
{Lm , cn }P B = i (2m + n) cn+m ,
{Lm , bn }P B = −i (m − n) bn+m (7.22)

In general, for any phase-space function O(J) of conformal dimension J,

(J)
Lm , On(J)

PB
= −i [m (J − 1) − n] On+m , (7.23)

so we can see that P ± T X 0 have conformal dimension 1 while b has conformal

dimension 2 and c has conformal dimension −1. This is all classical, but it carries
over to the quantum theory; in particular the conformal dimensions of P ± T X 0 and
0
(b, c) are as above. However, whereas the product O(J) O(J ) of phase-space functions
has conformal dimension J + J 0 , this will not generally be true in the quantum
theory; products of operators can have anomalous conformal dimensions.
We should also consider the transformations generated by L̃m ; these are entirely
analogous, so the worldsheet fields and (b, c) actually have two equal conformal di-
mensions: P ± T X 0 has conformal dimensions (1, 1), the ghost c has conformal
dimensions (−1, −1) and the antighost b has conformal dimensions (2, 2). For any
closed string theory preserving worldsheet parity, the left/right conformal dimensions
will be equal.

7.1 Quantum BRST

Passing to the quantum theory, the non-zero anticommutators of the (anti)ghost
modes are
{cn , b−n } = 1 , {c̃n , b̃−n } = 1 . (7.24)
We define the (anti)ghost oscillator vacuum as the state

|0igh = |0igh gh
R ⊗ |0iL , (7.25)

such that
cn |0igh
R = 0, n>0 & bn |0igh
R = 0, n ≥ 0. (7.26)
and similarly for the tilde operators acting on |0igh L . We can act on this with
the (anti)ghost creation operators c0 and (c−n , b−n ) for n > 0 to get states in an
(anti)ghost Fock space. The full oscillator vacuum is now the tensor product state

|0) = |0i ⊗ |0igh . (7.27)

– 69 –
 We should first deal with some operator ordering ambiguities in expressions
(gh)
involving (anti)ghost operators. There is no ambiguity in the expression for Lm as
long as m 6= 0; for m = 0 we choose the ordering given in (7.10), which ensures that
∞
X
(gh)
L0 ≡ N(gh) = k (b−k ck + c−k bk ) ⇒ N(gh) |0i = 0 . (7.28)
k=1

We then have
L(gh)
m |0igh = 0 , m ≥ 0. (7.29)
Notice that
   
N(gh) , c−k = kc−k , N(gh) , b−k = kb−k , (7.30)
so that acting with either the ghost creation operator c−k or the anti-ghost creation
operator b−k increases the ghost level number by k. It follows that the eigenvalues
of N(gh) and Ñ(gh) are the non-negative integers.
There is also an ordering ambiguity in Q− that allows us to add to it any multiple
of ĉ0 ; we choose the order such that
  ∞    
1 X 1 (gh) 1 (gh)
Q− = L0 + N(gh) − a ĉ0 + L−m + L−m cm + c−m Lm + Lm ,
2 m=1
2 2
(7.31)
for some constant a. This definition is such that

Q− |0i = 0 ⇒ (L0 − a) |0i = 0 . (7.32)

It is also such that [Exercise: check this]

{bm , Q− } = Lm − aδm , (7.33)

which is the quantum version of (7.19). Let’s also record here that

[Lm , bn ] = (m − n) bn+m , (7.34)

which is the statement that the operator b has conformal dimension 2.

Now we show how the Virasoro anomaly in the algebra of the Ln is related to a
BRST anomaly. Using (7.33) we find that

[Lm , Ln ] = [{bm , Q− } , Ln ]
= − {[Ln , bm ] , Q− } + {[Q− , Ln ] , bm }
= (m − n) {bm+n , Q− } + {[Q− , Ln ] , bm } , (7.35)

where the second line follows from the super-Jacobi identity, and the last line uses
(7.34). Now we use (7.33) again, and again the super-Jacobi identity, to show that

[Q− , Ln ] = [Q− , {bn , Q− }] = − Q2− , bn .

 
(7.36)

– 70 –
Using this in (7.35) we deduce that
 2 
[Lm , Ln ] = (m − n) (Lm+n − aδm+n ) + Q− , bn , bm . (7.37)

This shows that Q2− = 0 implies no Virasoro anomaly (i.e. zero central charge c).
If Q2− is non-zero it will be some expression quadratic in oscillator operators (the
classical result ensures that the quartic term cancels) and it must have ghost number
2, so
1X
Q2− = ck c−k A(k) (7.38)
2
k∈Z

for some function A(k). We then find that

[Lm , Ln ] = (m − n) (Lm+n − aδm+n ) + A(m)δm+n (7.39)

This shows that no Virasoro anomaly implies Q2− = 0. The same argument applies
to Q+ , so we now see that

Q2BRST = 0 ⇔ c = 0. (7.40)

For there to be no BRST anomaly we require that the Ln − aδn satisfy the Witt
algebra for some constant a; i.e. we require that

[Lm , Ln ] = (m − n) (Lm+n − aδm+n ) . (7.41)

7.1.1 Critical dimension again

We already know that the algebra satisfied by the Lm is the Virasoro algebra with
central charge D, i.e.
D
m3 − m δm+n .


[Lm , Ln ] = (m − n) Lm+n + (7.42)
12
(gh)
Now we need to determine the algebra satisfied by the Lm , which must take the
form
  c
 (gh) (gh)  (gh) gh
m3 − m δm+n .


Lm , Ln = (m − n) Lm+n − aδm+n + (7.43)
12
for some constants a and cgh . Setting n = −m and taking the oscillator vacuum
(gh)
expectation value of both sides, we deduce (using the fact that Lm |0i = 0 for
m ≥ 0)
 (gh) 2
  cgh
m3 − m .


L−m |0i = −2ma + (7.44)
12
In particular

 (gh) 2
 
 (gh) 2
  cgh
L
 −1 |0i = −2a , L
 −2 |0i  = −4a + . (7.45)
2


– 71 –
But we may also compute the LHSs directly using
(gh)
L−1 |0i = − (b−1 c0 + 2b0 c−1 ) |0i
(gh)
L−2 |0i = − (b−2 c0 + 3b−1 c−1 + 4b0 c−2 ) |0i (7.46)

For example23

 (gh) 2
 
L
 −1 |0i = h0| (c0 b1 + 2c1 b0 ) (b−1 c0 + 2b0 c−1 ) |0i
= −2h0| (c0 b0 b1 c−1 + b0 c0 c1 b−1 ) |0i (using b20 = c20 = 0)
= −2h0| (c0 b0 {b1 , c−1 } + {c1 , b−1 } b0 c0 ) |0i (using b1 |0i = c1 |0i = 0)
= −2h0| {c0 , b0 } |0i = −2 , (7.47)

from which we conclude that a = 1. Similarly,

 (gh) 2
 
L
 −2 |0i = h0| (2c0 b2 + 3c1 b1 + 4c2 b0 ) (2b−2 c0 + 3b−1 c−1 + 4b0 c−2 ) |0i
= −8h0| (c0 b0 b2 c−2 + b0 c0 c2 b−2 ) |0i − 9h0|c1 b−1 b1 c−1 |0i
= −8h0| (c0 b0 {b2 , c−2 } + b0 c0 {c2 , b−2 }) |0i − 9h0| {c1 , b−1 } {b1 , c−1 } |0i
= −8h0|({c0 , b0 } |0i − 9 = −17 , (7.48)

from which we conclude that

cgh
−4 + = −17 ⇒ cgh = −26 . (7.49)
2
Using these values for a and cgh in (7.43) we find that
  26
(gh)
L(gh) (gh)
m3 − m δm+n ,
 

m , Ln = (m − n) Lm+n − δ m+n − (7.50)
12
and combining this with (7.42), we deduce that

D − 26
m3 − m .


[Lm , Ln ] = (m − n) (Lm+n − δm+n ) + (7.51)
12
This is a Virasoro algebra with central charge D − 26, which is zero (as required for
Q2− = 0) iff D = 26.
In conclusion, we find agreement with the light-cone gauge result that a = 1
and D = 26. In this case it is consistent to impose the BRST physical-state condi-
tion QBRST |physi = 0. The physical states are then cohomology classes of QBRST .
Consider the state
|ψ)R = |ΨiR × |0igh
R (7.52)
23
Subtleties of the inner product for the (anti)ghost zero modes can be passed over here because
all that is needed is their anti-commutation relation.

– 72 –
where |ψiR is a state in the α-oscillator Fock space. Then
∞
X
Q− |ψ)R = (L0 − 1) |ψiR ⊗ |0igh
R + Lm |ψiR ⊗ c−m |0igh
R , (7.53)
m=1

which is zero only if ψ satisfies

(L0 − 1) |ψiR = 0 & Lm |ψiR = 0 ∀m > 0 . (7.54)

These are the Virasoro conditions of the “old covariant” method of quantization.
There is an entirely analogous result for the states built on the |0)L vacuum, and
it can be shown that the physical states of the light-cone gauge are in one-to-one
correspondence with the cohomology classes of QBRST at ghost number −1/2. This
is known as the “no-ghost” theorem.

8. Interactions
So far we have seen that each excited state of a string can be viewed as a particle
with a particular mass and spin. Now we are going to see that the stringy origin of
these particles leads naturally to interactions between them. We shall explore this
in the context of a path integral quantization of a closed string, where we sum over
world sheets weighted by the NG string action. We shall assume that the conformal
gauge has been chosen, and that we have done the Gaussian integral over P , which
effectively sets P = T Ẋ, so the conformal gauge action is
Z I
∂
Ic.g. = −T dt dσ ∂+ X · ∂− X , ∂± = . (8.1)
∂σ ±
We now “Wick rotate”: set t = −iτ and then take τ to be real rather than imaginary:
1
σ − → √ (σ + iτ ) = z , σ + → z̄ . (8.2)
2
Then Z
¯ , ∂ ∂
−iIc.g → IE = T d2 z ∂X · ∂X ∂= , ∂¯ = . (8.3)
∂z ∂ z̄
The functional IE [X] is the Euclidean NG action in conformal gauge; the integral is
over the cylindrical worldsheet. The Euclidean FP ghost action is
Z n o
(F P ) 2 ¯
IE = d z ib∂c + ib̃∂c̃ . (8.4)

For path integral purposes, the conformal gauge action is the “quantum” Euclidean
action
(qu) (F P )
IE = IE [X] + IE [b, c; b̃, c̃] . (8.5)

– 73 –
 Now suppose that the cylindrical worldsheet is infinite in both directions, corre-
sponding to a string in the far past propagating into the far future. It can be formally
mapped to the complex plane with the complex coordinate w = e−iz . Notice that

σ → σ + 2π ⇔ w → e2πi w , (8.6)

so the constant τ slices of the cylinder are mapped to circles centred on the origin in
the complex w-plane. Also

τ → −∞ ⇔ w → 0, τ →∞ ⇔ w→∞ (8.7)

Here we view the complex w-plane as the Riemann sphere, with two punctures: one
at the south pole (w = 0) associated to the incoming string, and one at the north
pole (w = ∞) associated to the outgoing string. Conformal invariance is essential
to the consistency of this idea, as is the cancellation of the conformal anomaly; for
example
c
L0 |cylinder = L0 |R.sphere + . (8.8)
24
The central charge represents a Casimir energy due to the periodic identification of
the string coordinate σ on the cylinder. As we have seen c = 0 when D = 26 due to
the FP ghost contribution to it.

8.1 Ghost zero modes and Sl(2; C)

When we go from the cylinder to the R. sphere we must take account of how the
(F P )
various fields behave under conformal transformations. Integrals such as IE will
take the the same form in terms of the transformed fields, now functions of w, only
if the integrand has conformal dimensions (1, 1) (i.e. 1 with respect to both “left”
(F P )
and “right” factors of the conformal group). This is the case for IE ; for example,
¯ has conformal dimension (1, 1) because b has conformal dimension (2, 0) and c
b∂c
has conformal dimension (−1, 0), while ∂¯ has conformal dimension (0, 1).
We shall now suppose that all fields are defined on the R. sphere, and we relabel
the coordinate w → z, so the Euclidean FP action is again given precisely by (8.4).
Path integrals representing amplitudes should include a functional integral over the
FP ghosts. For the amplitudes to be discussed here, this is just a product of gaussian
integrals that only contribute to an overall normalisation, which we are ignoring,
except that it would not be a gaussian for modes of b or c on the R. sphere for which

either ¯ =0
∂c or ¯ = 0.
∂b (8.9)

In either case the FP action would be zero so Berezin integration over these modes
would give us a zero amplitude. We should make the replacement
Z Z
[dbdc] → [dbdc]0 (8.10)

– 74 –
where the prime indicates an absence of integration over (anti)ghost “zero modes”.
What are these (anti)ghost zero modes. From (8.9) we see that they are analytic,
but they are not analytic functions because both b and c have a non-zero conformal
dimension. The fact that b has conformal dimension 2 means that it is a quadratic
differential. By a variant of the argument to follow for c, it can be shown that
there are no analytic quadratic differentials on the R. sphere. The fact that c has
conformal dimension −1 means that it is a vector (a one-form, or co-vector, has
conformal dimension 1). This is as expected because c must be of the same tensorial
type as the gauge parameter ξ − , which is the one component of the vector ξ − ∂−
which becomes ξ(z)∂ in Euclidean space. To see how this behaves at infinity, set
z = −1/ζ to get
ξ(z)∂z = ξ (z(ζ)) ζ 2 ∂ζ (ζ = −1/z). (8.11)

This is finite at ζ = 0 as long as ξζ 2 is finite as ζ → 0, which is equivalent to the

requirement that ξ(z) grow no faster that z 2 as z → ∞. So the analytic vector fields
that are well defined on the R. sphere have the form

ξ(z) = α1 + α2 z + α3 z 2 , αk ∈ C k = 1, 2, 3. (8.12)

These vector fields are the conformal Killing vector fields of the R. sphere. They
span the Lie algebra of Sl(2; C). Recall that the infinite-dimensional 2D conformal
algebra Diff 1 ⊕Diff 1 has sl(2; R)⊕sl(2; R) as a finite-dimensional sub algebra. This 6-
dimensional algebra becomes the 3-complex dimensional algebra sl(2; C) in Euclidean
signature.
The defining representation of the group Sl(2; C) is in terms of the following
matrices  
ab
, a, b, c, d ∈ C , ad − bc = 1 . (8.13)
cd

The group Sl(2; C) acts on the coordinate z of the R. sphere via the fractional linear
transformation
az + b
z → z0 = . (8.14)
cz + d
Near the identity we can write
     
ab 10 1 α1 2α2
= + + ... (8.15)
cd 01 2 2α3 −α1

where (α1 , α2 , α3 ) are three complex parameters and the omitted terms are at least
quadratic in these parameters; to first order in them, one finds that

z 0 = z + ξ(z) , ξ(z) = α1 + α2 z + α3 z 2 . (8.16)

– 75 –
8.2 Virasoro-Shapiro amplitude from the path integral
If an incoming or outgoing string can be mapped to the poles of the R. sphere,
what is to stop us considering more strings, incoming or outgoing, mapped to other
points? For example, a four-punctured sphere could represent the scattering of two
strings into two other strings. Actually, each string can be in any of its excited
states, so we should have an option at each puncture of an insertion into the path-
integral corresponding to each excited state of the string. We shall explore this
idea by answering the following question. What is the amplitude A(X1 , . . . , XN )
for the string to pass through N points in Minkowski space-time with coordinates
(X1 , . . . , XN )? We could represent this formally as the path integral
Z N Z
Y
−IE
A(X1 , . . . , XN ) = [dX] e dzi2 δ D (X(zi ) − Xi ) , (8.17)
i=1

where the functional integral is over maps X(z, z̄) from the Riemann sphere to
Minkowski space-time (we are supposing here that we have done the gaussian in-
R
tegrals over P and over the (anti)ghosts), and the dzi integrals are over the R.
sphere. In the path integral computation of the propagator for the point particle,
we had to sum over the gauge-invariant proper time of the particle worldline (pro-
portional to the “modular” parameter s). If there were some analogous parameters
of the R. sphere then we would have to integrate over them, and we would have to
consider how the P and the (anti)ghosts path integrals dependence on them, but
the R. sphere has no modular parameters24 ; as Riemann proved, the conformally flat
metric on the R. sphere is unique.
It is simpler to consider the Fourier transform25
N Z
Y
A(p1 , . . . , pN ) = dD Xi eipi ·Xi A(X1 , . . . , XN )
i=1
Z N Z
Y
−IE
= [dX] e d2 zi eipi ·Xi
i=1
Z N Z N
!
Y X
= [dX] d2 zi exp −IE + i p j · Xj (8.18)
i=1 j=1

Now we observe that

N Z ( " N # )
X
¯ − i X
−IE + i pj · Xj = −T d2 z ∂X · ∂X δ 2 (z − zj )pj · X . (8.19)
j=1
T j=1
24 ¯ = 0 on the R.
This is actually equivalent to the earlier claim that there are no solutions of ∂b
sphere since any analytic quadratic differential on a Riemann surface corresponds to the possibility
of deforming a conformally flat metric metric to a conformally inequivalent one.
25
Although we should set D = 26 to avoid the Virasoro anomaly, this really becomes an unavoid-
able issue only when we consider string loops, so we keep the space-time dimension D arbitrary
here.

– 76 –
 ¯ = −X · ∇2 X + ∂(), where ∇2 = ∂ ∂¯ is the
Integrating by parts to write ∂X · ∂X
Laplacian on the R. sphere, we have26
N Z ( " N
#)
X
2 2 i X 2
−IE + i pj · Xj = T d z X · ∇ X + δ (z − zj )pj . (8.20)
j=1
T j=1

The idea now is to complete the square in X(z) but to do this we need to invert ∇2
and there is a problem with this because ∇2 has a zero eigenvalue on the sphere.
The eigenfunction is the constant function, i.e. X(z) = X0 , so we should write
Z Z Z
[dX] = d X0 [dX]0 ,
D
(8.21)

where [dX]0 is an integral over all functions except the constant function. Isolating
the X0 -dependence we now have
Z 
i( j pj )·X0
P
D
A(p1 , . . . , pN ) = d X0 e Â(p1 , . . . , pN )
N
!
X
∝ δ pj Â(p1 , . . . , pN ) , (8.22)
j=1

where the path integral for  excludes the integration over the constant function.
The delta-function prefactor imposes conservation of the total D-momentum.
We can now invert ∇2 ; the inverse is the Green function G(z, zi ) on the R. sphere:
1
∇2 G(z, zi ) = δ 2 (z − zi ) ⇒ G(z, zi ) = ln |z − zi |2 . (8.23)
2π
Setting
N
i X
X(z) = Y (z) − G(z, zi )pi , (8.24)
2T i=1
we have27 [dX]0 = [dY ]0 , and [Exercise]
N Z
X 1 XX
−IE + pj · Xj = T d2 z Y · ∇ 2 Y + pi · pj ln |zi − zj |2 . (8.25)
j=1
8πT i j

The terms in the double sum are infinite when i = j, so we omit these terms (this
amounts to a renormalisation to remove “self-energies”). We can now do the gaussian
[dY ]0 path integral. This contributes only to the overall normalisation, and we are
left with
N Z
Y Y pi · p j
Â(p1 , . . . , pN ) ∝ d2 zi |zj − zk |αjk , αij = . (8.26)
i=1 j<k
2πT
26
There is no surface term because the sphere has no boundary.
27
A shift in the integration variable has no effect because we integrate over all values of the
(non-constant) functions X.

– 77 –
 As the derivation of this formula assumed conformal invariance, this result should
be invariant under the Sl(2; C) conformal isometry group of the R. sphere. Let’s check
this; using (8.14) one finds that

zi − zj dz 2
zi0 − zj0 = , d2 z 0 = , (8.27)
(czi + d) (czj + d) (cz + d)4

and hence that

N
" N
#" N #
−4− 0j αij
Y Y Y Y Y P
d2 zi0 |zj0 − zk0 |αjk = d2 zi |zj − zk |αjk |czi + d| , (8.28)
i=1 j<k i=1 j<k i=1

where
X N
X
0
αij = αij − αii (i = 1, . . . , N )
j j=1
N
!
1 X p2i
= pi · pj −
2πT j=1
2πT
p2i
=− (by momentum conservation). (8.29)
2πT
We see from this that the amplitude is Sl(2; C) invariant only if

p2i
−4 + =0 ⇔ p2i = 8πT . (8.30)
2πT
This is the mass-shell condition for the tachyonic ground state of the string! We
learn a few things from this computation

• We might have supposed that the N momenta p1 , . . . , pN could be chosen arbi-

trarily, consistent with overall momentum conservation, but we have now seen
that these must be the momenta of strings in their ground state, i.e. we have
a tachyon scattering amplitude.

• The tachyon scattering amplitude that we have constructed is “on-shell”, i.e.

the external particles have momenta satisfying their mass-shell condition. This
can be contrasted with the amplitude that we found earlier for propagation
of a particle from one point to another in Minkowski space time; this was the
Fourier transform of 1/(p2 + m2 ), which is infinite on the mass shell. This
should lead us to expect that the string scattering amplitude of (8.26) makes
sense only for N > 2.

Recall that the FP ghosts in the “quantum” action used in the path integral
are the result of a gauge-fixing that is needed to eliminate unnecessary integration

– 78 –
over redundant variables; it amounts to a trick that allows us to divide by the gauge
group volume. If this volume were finite it would only affect the normalisation but
it is usually infinite. Recall also, that we had to restrict the functional integral over
the (anti)ghosts; we excluded integration over the ghost “zero modes”, which are
conformal Killing vectors of the R. sphere. As a result, the amplitude (8.26) is still
infinite because of its Sl(2; C) invariance. We can deal with this by gauge fixing.
We choose the gauge fixing conditions
fi ≡ zi − ui = 0 , i = 1, 2, 3. (8.31)
That is, we chose three of the N points on the sphere to be at zi = ui for arbitrarily
chosen complex numbers ui . This will involve an insertion of delta functions into
the integrals defining the amplitude, but this must be accompanied by an FP deter-
minant. Since the infinitesimal Sl(2; C) variation of z is δz = α1 + α2 z + α3 z 2 , we
have
  1 z1 z12 
 

 ∂fi   
2 
 ∂αj  =  1 z2 z2  = (z1 − z2 )(z2 − z3 )(z3 − z1 ) . (8.32)
  
 1 z z 2 
3 3

The FP determinant is the modulus squared of this, so

∆F P = |z1 − z2 |2 |z2 − z3 |2 |z3 − z1 |2 . (8.33)
Following the earlier argument for gauge fixing the particle action, the insertion of
the delta functions with the FP determinant allows us to factor out the (infinite)
volume Ω of Sl(2; C); dividing by this volume we then get
N Z
Y Y
−1
Ω Â(p1 , . . . , pN ) ∝ d2 zi δ 2 (f1 )δ 2 (f2 )δ 2 (f3 )∆F P |zj − zk |αjk . (8.34)
i=1 j<k

This can be checked as follows. Multiply both sides by |(u1 − u2 )(u2 − u3 )(u3 − u1 )|−2
and integrate over (u1 , u2 , u3 ). On the RHS the u integrals can be done using the
delta functions, the ∆F P is then cancelled and we recover the expression (8.26). On
the LHS the integral cancels the factor of Ω−1 because, formally,
d2 u1 d2 u2 d2 u3
Z
Ω= . (8.35)
|(u1 − u2 )(u2 − u3 )2 (u3 − u1 )|2
This integral is infinite but the integrand is the Sl(2; C) invariant measure on the
Sl(2; C) group manifold, parametrised by three complex coordinates on which Sl(2; C)
acts by the fractional linear transformation (8.14).
We may now do the (z1 , z2 , z3 ) integrals of (8.34) to get
Ω−1 Â(p1 , . . . , pN ) ∝ |u1 − u2 |2+α12 |u2 − u3 |2+α23 |u3 − u1 |2+α13 × (8.36)
YN Z YN Y
× d2 zi |u1 − zi |α1i |u2 − zi |α2i |u3 − zi |α3i |zi − zj |αjk .
i=4 i=4 4≤j<k

– 79 –
This can be simplified enormously by the choice

u3 = 1 , u2 = 0 , u1 → ∞ . (8.37)

In this limit we get a factor of

2
P
|u1 |4−α11 + i α1i
= 1, (8.38)

where the equality follows upon using both the mass-shell condition and momentum
conservation. The remaining terms give the Virasoro-Shapiro amplitude
N Z
Y N
Y Y
2
ÂV S (p1 , . . . , pN ) = d zi |zi |α2i |zi − 1|α3i |zi − zj |αjk . (8.39)
i=4 i=4 4≤j<k

The result for N = 3 is a constant, which can be interpreted as a coupling constant.

The N = 4 case was found by Virasoro, and generalised to arbitrary N by Shapiro.

8.2.1 The Virasoro amplitude and its properties

The N = 4 amplitude is
Z
Â(p1 , p2 , p3 , p4 ) = d2 z |z|α24 |z − 1|α34 . (8.40)

If we view this as the amplitude for two identical particles of momenta (p1 , p2 ) to
scatter into two other identical particles of the same kind with momenta (−p3 , −p4 )
then we can define the associated (rescaled) Mandelstam invariants as
1 1
s=−(p1 + p2 )2 = −2 − α12
8πT 2
1 1
t==− (p1 + p3 )2 = −2 − α13
8πT 2
1 1
u=− (p1 + p4 )2 = −2 − α14 . (8.41)
8πT 2
Recall that, for the amplitude of interest here,

p21 = p22 = p23 = p24 = 8πT , (8.42)

and that momentum conservation requires28

p1 + p2 = −p3 − p4 . (8.43)

In the centre of mass frame we have

p1 = (E, p~ ) , p2 = (E, −~p ) & p3 = (−E, p~ 0 ) , p4 = (−E, −~p 0 ) , (8.44)

28
The Virasoro amplitude was derived assuming, for simplicity, that all particles were either
incoming. In this convention, outgoing particles will have negative energy.

– 80 –
where
|~p|2 = |~p 0 |2 = E 2 + 8πT . (8.45)
From this we see that
E2  s
s= , t = −2 1 + (1 − cos θs ) , (8.46)
2πT 4
where θs is the scattering angle. Of course, it doesn’t make much sense to talk about
scattering of tachyons, but the amplitude provides an illustration of some properties
that are typical of scattering amplitudes in string theory, and superstring theory
(which has no tachyons).
From (8.41) and momentum conservation it follows that (Exercise)

α34 = α12 = −4 − 2s ,
α24 = α13 = −4 − 2t ,
α23 = α14 = −4 − 2u , (8.47)

and
s + t + u = −4 . (8.48)
We will now use the following identity

Γ(α + 1)Γ(β + 1)Γ(−α − β − 1)

Z
d2 z |z|2α |z − 1|2β ≡ π . (8.49)
Γ(α + β + 2)Γ(−α)Γ(−β)

We can use this to do the integrals of (8.40) by setting

α = −2 − t , β = −2 − s . (8.50)

This gives the Virasoro amplitude

Γ(−1 − t)Γ(−1 − s)Γ(−1 − u)

A(s, t) ∝ (u = −4 − s − t). (8.51)
Γ(u + 2)Γ(s + 2)Γ(t + 2)

For fixed t the Virasoro amplitude A becomes a function of s, which has poles
when29
s = −1, 0, 1, 2, . . . (8.52)
These poles correspond to resonances, i.e. to other particles in the spectrum (stable
particles, in fact, because the poles are on the real axis in the complex s-plane). The
position of the pole on the real axis gives the mass-squared of the particle in units of
−m2 = 8πT . The pole at s = −1 is therefore the tachyon. The pole at s = 0 implies
29
The poles come from the Γ(−1 − s) factor. Recall that Γ(z) has no zeros in the complex z-plane
but has simple poles on the real axis at non-negative integers.

– 81 –
the existence of massless particles in the spectrum. The residue of the pole at s = 0
is
Γ(−1 − t)Γ(3 + t)
− = t2 − 4 . (8.53)
Γ(−2 + t)Γ(t + 2)
This is a quadratic function of t and hence of cos θs , which implies that there must
be a massless particle of spin 2 (but none of higher spin). The residue of the pole
at s = n is a polynomial in t of order 2(n + 1), so that 2(n + 1) is the maximum
spin of particles in the spectrum with mass-squared n × (8πT ). In a plot of Jmax
against s, such particles appear at integer values of Jmax on a straight line with slope
α0 /2 and intercept 2 (value of Jmax at s = 0). This is the leading Regge trajectory.
All other particles in the spectrum appear on parallel “daughter” trajectories in the
J − s plane (e.g. the massless spin-zero particle in the spectrum is the first one on
the trajectory with zero intercept.
Another feature of the Virasoro amplitude is its s ↔ t symmetry. Poles in A as
a function of s at fixed t therefore reappear as poles in A as a function of t at fixed
s. These correspond to the exchange of a particle. In particular, a massless spin-2
particle is exchanged, and general arguments imply that such a particle must be
the quantum associated to the gravitational force, so a theory of interacting closed
strings is a theory of quantum gravity.
There is an analog of the Virasoro amplitude for open strings, which was proposed
earlier; this is the Veneziano amplitude30
Γ(−1 − s)Γ(−1 − t)
A= , (8.54)
Γ(−2 − s − t)
where now
1 1
s=− (p1 + p2 )2 , t=− (p1 + p3 )2 . (8.55)
2πT 2πT
Thie Veneziano amplitude also has poles at s = −1, 0, 1, 2, . . ., but the maximum
spin for s = n is now Jmax = n + 1, and the leading Regge trajectory has slope α0 and
intercept 1 (this is the constant a that equals the zero point energy in the light-cone
gauge quantization of the open string).

8.3 Other amplitudes and vertex operators

Consider the following modification of the Euclidean conformal gauge action to in-
clude a potential function U (X):
Z
IE → I = IE + d2 z U (X) . (8.56)

The term involving U is not conformal invariant. However, if we write the action as
Z
1 2

¯ + 2πα0 U (X)

I= d z ∂X · ∂X (8.57)
2πα0
30
Veneziano did not do a computation to get this amplitude; he just proposed it on the basis of its
properties, but it turns out to be the amplitude for elastic scattering of two open-string tachyons.

– 82 –
then we see that α0 plays the role ~ so that a term in the integrand with a factor
of α0 , such as the U (X) term, must be considered along with first-order quantum
corrections to the 2D classical field theory defined by the action (8.56). Under
renormalisation, all operators of a given classical conformal dimension mix and those
of definite quantum dimension are eigenfunctions of an anomalous dimension matrix.
In this case, the matrix is −D /(8πT ) and the functions U (X) satisfying
D
− U = λU (8.58)
8πT
have quantum conformal dimension (λ, λ). Its integral over the worldsheet will be
conformal invariant if λ = 1, and this is the case if p2 = 8πT . This agrees with the
calculation that we have already done; if we treat U (X) as a perturbation and write
∞ Y
X N Z
−I −IE
e =e d2 zi U (X(zi )) , (8.59)
N =0 i=1

and take U (X) to be some linear combination of the functions Up (X) = eip·X , then
the corresponding expansion of the vacuum to vacuum amplitude includes scattering
amplitudes of the type already considered, which we found to be conformal invariant
iff p2 = 8πT .
In an operator version of this calculation we would have to compute vacuum
expectation values of products of integrals of “vertex operators” of the form

V0 =: eip·X : (8.60)

where the colons indicate “normal ordering” (all creation operators appear to the
left of all annihilation operators, so that V0 annihilates the oscillator vacuum). Each
particle in the string spectrum has its own vertex operator of conformal dimension
(1, 1). For example, at level 1 we have the vertex “operator”
¯ n eip·X : ,
V1 = mn : ∂X m ∂X (8.61)

where mn is a constant polarisation tensor. Since ∂X m ∂X ¯ n already has dimension

(1, 1) we need the eip·X factor to have conformal dimension zero, which it does if
p2 = 0. It turns out that one also needs to impose the conditions pm mn = pn mn = 0.
So V1 is a vertex operator for massless particles in the string spectrum. Notice that
if mn = ξm pn then
Z Z
d z V1 = −iξm d2 z∂X m ∂¯ eip·X ,
2


(8.62)

so integrating by parts we have

Z Z
d zV1 = iξm d2 z∇2 X m eip·X ,
2
(8.63)

– 83 –
which is zero if we use the X field equation31 . This shows that there is a gauge
invariance for zero mass. Notice that the polarisation tensor has the decomposition

mn = (mn) + [mn] . (8.64)

The symmetric part is the polarisation tensor for the massless graviton and the
antisymmetric part is the polarisation tensor for the massless antisymmetric tensor.
Now consider the level-2 vertex operator
(4) ¯ p ∂X
¯ q eip·X :
V2 = mnpq : ∂X m ∂X n ∂X (8.65)

¯ p ∂X
Since ∂X m ∂X n ∂X ¯ q has dimension (2, 2) we need : eip·X : to have conformal
dimension (−1, −1), which will be the case if p2 = −8πT and the polarisation tensor
mnpq is orthogonal to p on all four indices. Another level-2 vertex operator is
(3) ¯ n ∂X
¯ p eip·X :
V2 = mnp : ∂ 2 X m ∂X (8.66)

Notice that level-matching now becomes the requirement that both “left” and “right”
conformal dimensions equal 1.
As for the tachyon field, we can also introduce the massless level-1 fields as
background fields32 . For example, we may modify the Euclidean conformal gauge
action to Z
1 ¯ n [ηmn + hmn (X) + bmn (X)] ,
d2 z ∂X m ∂X

I= 0
(8.67)
2πα
where hmn (X) is a symmetric tensor field and bmn (X) is an antisymmetric tensor
field. Notice that we can define ηmn + hmn (X) to be the space-time metric gmn (X).
The conditions for conformal invariance to first-order in h and b are, respectively,
the linear Einstein equations and Maxwell-type equations.
However, we are missing one of the level-1 states: the scalar “dilaton”. The dila-
ton field Φ(X) couples to the string, through its derivative ∂m Φ, via the interaction
Z
1 2
h i
¯ m + c̃b̃∂X m ∂m Φ(X) .
Idil ∝ d z cb ∂X (8.68)
2πα0

However, this is equivalent, because of a ghost number anomaly to a much simpler

term that we will now discuss.

8.3.1 The dilaton and the string-loop expansion

The way that the dilaton field Φ(X) couples to the string in its Polyakov formulation
is through the scalar curvature of the independent worldsheet metric γµν . In two
31
Such terms can be removed by a field redefinition.
32
This is also true, but less useful, for higher-level fields.

– 84 –
dimensions the Riemann curvature tensor is entirely determined by its double trace,
the Ricci scalar R(γ), but this allows us to add to the Euclidean NG action the term

√
Z
1
IΦ = d2 z Φ(X) γ R(γ) . (8.69)
4π
Here are some features of this term:

• If Φ = φ0 , a constant then
√
Z
1
IΦ = φ0 χ , χ= d2 z γ R(γ) . (8.70)
4π
The integral χ is a topological invariant of the worldsheet, called the Euler
number. For a compact Riemann surface with boundary (which we’ll abbrevi-
ate to “Riemann surface” in what follows) the Euler number is related to the
genus g (the number of doughnut-type “holes”) by the formula

χ = 2 (1 − g) . (8.71)

• In conformal gauge, we can write the line element for the (Euclidean signature)
metric γ as ds2 (γ) = 2eσ dzdz̄, i.e. a conformal factor eσ (an arbitrary function
of z and z̄) times the Euclidean metric. We then find that
√
γ R(γ) = 2∇2 σ (8.72)

and hence, after integrating by parts,

Z
1 ¯ n ∂n ∂m Φ .
IΦ = d2 z σ ∂X m ∂X (8.73)
2π
This shows that IΦ is not conformal invariant because it depends on the con-
formal factor (through its σ-dependence). This is allowed because (like the
tachyon potential term) it comes with an additional factor of α0 relative to the
other terms; we have to consider the lack of conformal invariance of IΦ at the
same time that we consider possible conformal anomalies of action (8.67).

• When we consider conformal invariance to second-order in α0 we find the con-

dition D Φ = 0 to leading order in Φ; this confirms that Φ(X) is a massless
field.

So far we have considered amplitudes for which the Euclidean worldsheet is

a Riemann sphere with some number of punctures (corresponding to some vertex
“operator” insertion into the path integral). However, in QM we have to sum over
all possibilities, and hence over Euclidean worldsheets of any genus g. So far we have
considered only the g = 0 case, i.e. the Riemann sphere. How should we weight this

– 85 –
sum over different topologies? To answer this question, at least partially, suppose
that
Φ = φ0 + φ(X) , (8.74)
where φ(X) is zero in the vacuum; i.e. the constant φ0 is the “vacuum expectation
value” of Φ(X). Then there will appear a factor in the path integral of the form

g−1
e−φ0 χ = gs2 , gs ≡ eφ0 . (8.75)

For g = 0 this tells us that the R. sphere contribution to scattering amplitudes is

weighted by a factor of 1/gs2 . If we use these amplitudes to construct an effective field
theory action S from which we could read off the amplitudes directly (by looking at
the various interaction terms) then this action will come with a factor of gs (we can
then absorb all other dimensionless factors into a redefinition of gs , i.e. of φ0 ). If we
focus on the amplitudes for scattering of level-1 particles then we find that
Z  
1 1 D
p −2φ 1 2 2 0
S[h, g, φ] = 2 (D−2) d x − det g e 2Λ + R(g) − H + 4 (∂φ) + O(α ) ,
gs `s 3
(8.76)
where the cosmological constant is
(D − 26)
Λ= . (8.77)
3α0
and `s is a length scale, required by dimensional analysis. This scale must be set by
the string tension, for the following reason. The NG string action in a background
metric gmn is unchanged if

α0 → λα0 & gµν → λgµν , (8.78)

√
and this is also true of the spacetime effective action if `s ∝ α0 . The constant
of proportionality can be absorbed into the dimensionless constant gs , so we may
choose, without loss of generality, to set
√
`s = α0 . (8.79)

Some other features of the effective space-time action are

• The exact result for S will involve a series of all order in α0 since the coupling
of the background fields to the string introduces interactions into the 2D QFT
defined by the string worldsheet action.

• The leading term is the cosmological constant Λ. Unless we know the entire
infinite series in α0 , we must set Λ = 0; i.e. we must choose D = 26. It is then
consistent to consider the string as perturbation about the Minkowski vacuum,
which is what we implicitly assumed when we earlier derived the condition
D = 26.

– 86 –
 • It is consistent to exclude the coupling to the string of the fields associated to
massive modes of the string because without them the worldsheet action defines
a renormalizable 2D QFT. Coupling to the fields associated to the massive
particles in the string spectrum leads to a non-renormalizable 2D QFT for
which it is necessary to consider all possible terms of all dimensions. But if all
fields of level N > 1 are all zero initially then they stay zero.

• The integrand involves a factor of e−2φ . This is because the action must be
such that φ0 ≡ ln gs and φ(X) must appear only through the combination
Φ = φ0 + φ(X).

In the effective spacetime action, gs2 plays the role of ~. This suggests that we have
been considering so far only the leading term in a semi-classical expansion. This is
true because we have still to consider R. surfaces with genus g > 0, and a string
amplitude at genus g is weighted, according to to (8.75), by a factor of (gs2 )g−1 , i.e. a
factor of (gs2 )g relative to the zero-loop amplitude. This confirms that the string-loop
expansion is a semi-classical expansion in powers of gs2 . Taking into account all string
loops gives us a double expansion of the effective field theory33
Z ∞ ∞
1 D
p X
2g 2(g−1)φ
X
S= (D−2)
d x − det g gs e Lg , Lg = `2l (l)
s Lg . (8.80)
2
gs `s g=0 l=0

In effect, the expansion in powers of `s comes from first-quantisation of the string, and
the expansion in powers of gs comes from second-quantisation. How can we quantise
twice? Is there not a single ~? The situation is actually not so different from that of
the point particle. When we first-quantise we get a Klein-Gordon equaltion but with
a mass m/~; we then relabel this as m so that it becomes the mass parameter of the
classical field equation, and then we quantise again. For the string, first quantization
would have led to α0 ~ as the expansion parameter if we had not set ~ = 1; if we
relabel this as α0 then ~ appears only in the combination gs2 ~.
To lowest order in α0 we have what looks like GR coupled to an antisymmetric
tensor and a scalar. The D-dimensional Newton gravitational constant GD is

GD ∝ gs2 `(D−2)
s . (8.81)

Consistency of the string-loop expansion (in powers of gs2 ) relies on this formula.
Particles in the string spectrum have masses proportional to 1/`s , independent of gs ,
33 (l)
There is a lot of freedom in the form of the Lagrangians Lg beyond leading order. Recall that
the construction of S involves a prior determination of scattering amplitudes of the level-1 fields,
which we then arrange to replicate from a local spacetime Lagrangian. Since the amplitudes are
all “on-shell” they actually determine only the field equations for the background fields, and then
only up to field redefinitions. Even with all this freedom it is not obvious why it should be possible
to replicate the string theory scattering amplitudes in this way, although this has been checked to
low orders in the expansion and there are general arguments that purport to prove it.

– 87 –
so their contribution to the gravitational potential in D dimensions is proportional
to gs2 , and hence zero at zero string coupling. This means that the strings of free
(gs = 0) string theory do not back-react on the space-time metric; the metric is
changed by the presence of strings only within perturbation theory. If this had not
been the case it would not have been consistent to start (as we did) by considering
a string in Minkowski spacetime.
Why is gs called the string coupling constant? Consider a g string-loop vacuum
to vacuum diagram with the appearance of a chain of g tori connected by long
“throats”, and think of it as “fattened” Feymann diagram in which a chain of g
loops connected by lines; where each line meets a loop we have a 3-point vertex. As
there are (g − 1) lines that link the loops, and each of the two ends of each line ends
at a vertex, we have a total of 2(g − 1) vertices. If we associate a coupling constant
to each vertex, call it gs , we see that this particular diagram comes with a factor of
(gs2 )g−1 , which agrees with our earlier result.
Is there a g-loop R. surface with the appearance just postulated. Yes, there is.
For g > 0 there is no longer a unique flat metric; for example, there is a one complex
parameter f smily of conformally inequivalent flat metrics. For g ≥ 2 there is a
3(g − 1)-parameter family of conformally inequivalent flat metrics; these parameters
are called “moduli”. This number can be understood intuitively from the “chain of
tori” diagram if we associate one parameter with each propagator. For g loops we
have, in addition to the (g − 1) links, (g − 2) “interior” loops with 2-propagators
each, and two “end of chain” loops with one propagator each. The total number of
propagators is therefore

(g − 1) + 2(g − 2) + 2 = 3(g − 1) . (8.82)

This is also, and not coincidentally, the dimension of the space of quadratic differen-
tials on a Riemann surface of genus g ≥ 2.

8.4 String theory at 1-loop: taming UV divergences

We will now take a brief look at what happens at one string-loop. In this case
amplitudes are found from the path integral by considering vertex operator insertions
at points on a conformally flat complex torus. In contrast to the R. sphere, for
which all conformally flat metrics are conformally equivalent, there is a one complex
parameter family of conformally flat tori. We can define a flat torus by a doubly
periodic identification in the complex z-plane. Without loss of generality we may
choose one identification to be under a unit shift along the real axis. The other
identification must then be by some complex number τ (as it is traditionally called)
that has a non-zero imaginary part, so

z ∼z+1 & z ∼z+τ, =τ > 0 . (8.83)

– 88 –
We may restrict τ to take values in the upper-half τ -plane without loss of general-
ity since otherwise we could take the complex conjugate and consider the z̄-plane.
Different values of the “modulus” τ correspond to “shapes” that are conformally
inequivalent, which we should sum over in the path integral. However not all points
on the upper-half τ -plane correspond to different shapes. It is not difficult to see
that the unit translation
τ →τ +1 (8.84)
gives the same torus; the z = τ + 1 in the z-plane is equivalent to the point z = τ
because z ∼ z + 1. It is also true that the inversion
1
τ →− (8.85)
τ
gives the same torus (after a rescaling). Consider the special case for which τ is pure
imaginary, say τ = 2i; we then have a rectangular torus with sides in the ratio 2 : 1.
After taking τ → −1/τ , which takes 2i to i/2 we get a rectangular torus with sides
in the ratio 1 : 2, but this is the same as 2 : 1 if we don’t care which side we consider
first. More generally, the lengths of the two sides are in equivalent ratios and the
angle between them is preserved by any analytic map, so we get an equivalent torus.
The operations of unit translation and inversion do not commute; they close to the
discrete fractional linear transformations
aτ + b
τ→ , a, b, c, d ∈ Z , ad − bc = 1. (8.86)
cτ + d
This is a realisation of the group

P Sl(2; Z) ∼
= Sl(2; Z)/ {±1} . (8.87)

We have to identify Sl(2; Z) matrices that differ by a sign since this sign cancels from
(8.86). This is called the “modular group” of the torus. Because tori with modular
parameters related by the action of the modular group have the same shape, we
should integrate τ in the path integral only over one fundamental domain of the
modular group.
It is convenient to choose this fundamental domain to be the one, let’s call it F ,
in which we may take =τ → ∞. This is because in this limit the torus becomes long
and thin and it starts to look like a one-loop Feynmann diagram (with vertices at
various points if we had vertex operators at points on the torus). Specifically, we can
think of =τ as the modular parameter s of a particle worldline in the s → ∞ limit,
which is an IR limit. We should expect string theory to reduce to a field theory in
this limit because the massive particles in the string spectrum should be unimportant
and we should be able to replace the string theory by its effective field theory.
In the particle case we have to integrate s from s = 0 to s = ∞, and s = 0 is
the UV limit in which we find the UV divergences of QFT. The situation is quite

– 89 –
different for string theory because the fundamental domain F is the region of the
upper-half τ -plane defined by
1 1
|τ | ≥ 1 & − ≤ <τ < . (8.88)
2 2
This does not include the imaginary axis below τ = i, so there is no UV limit! A
better way to say this is that an approach to theUV limit is equivalent, by a modular
transfermation, to an approach to the IR limit. So UV divergences will be absent as
long as the IR divergences cancel.

8.5 Beyond String Theory

So far we have considered only the NG string, or “bosonic string” as it is often
called to distinguish it from string theories that incorporate fermions in some way.
There are not many consistent ways to do this. One, called the “spinning string”
involves a worldline supersymmetric extension of the NG string. The “spinning
string” has critical dimension D = 10 but still has a tachyon. However, it turns
out that it is consistent to simply “throw away” the tachyon, along with a lot of
other states, leaving a spectrum that exhibits spacetime supersymmetry. This is
called the “superstring”, and another formulation of it was later found for which
the spacetime supersymmetry is manifest (as is then the absence of a tachyon since
tachyonic unitary irreps of the super-Poincaré group do not exist). Actually, there are
several superstring theories, which are roughly in correspondence with the possible
low-energy D = 10 supergravity theories (after excluding the anomalous ones).
Superstring theory provides a way of consistently computing perturbative cor-
rections to a D = 10 supergravity theory. However the string-loop expansions an
asymptotic one with zero radius of divergence; it does not define a theory and is
useful only if gs  1. Because of (8.81) this means that
 
1/(D−2)
ls  `P lanck `P lanck = GD , (8.89)

so we are not going to learn much from perturbative string theory about the real
problems of quantum gravity, which occur at the Planck length. For that a non-
perturbative extension of string theory is required. Fortunately, supersymmetry in
D = 10 is a strong constraint from which a lot has been learned about this non-
perturbative theory. What we have have learned is

• The non-perturbative theory has D = 11, not D = 10. From this perspective
all the different D = 10 perturbative superstring theories are unified into one
theory, called “M-theory” (best not to ask why).

• Strings are just one of many kinds of “branes” and M-theory is an Owellian
democracy in which all branes are equal but some are more equal than others.

– 90 –