String Theory in Two Dimensions
String Theory in Two Dimensions
String Theory in Two Dimensions
July 1991
⋆
STRING THEORY IN TWO DIMENSIONS
†
Igor R. Klebanov
arXiv:hep-th/9108019v2 25 May 2003
ABSTRACT
I review some of the recent progress in two-dimensional string
theory, which is formulated as a sum over surfaces embedded in one
dimension.
1. INTRODUCTION.
These notes are an expanded version of the lectures I gave at the 1991 ICTP
Spring School on String Theory and Quantum Gravity. Here I have attempted to
review, from my own personal viewpoint, some of the exciting developments in two-
dimensional string theory that have taken place over the last year and a half. Because
of the multitude of new results, and since the field is still developing rapidly, a com-
prehensive review must await a later date. These notes are mainly devoted to the
matrix model approach1 that has truly revolutionized the two-dimensional Euclidean
quantum gravity. Recently this approach has led to the exact solution of quantum
gravity coupled to conformal matter systems with c ≤ 1.2−6
These lectures are about the c = 1 model5,6 that is both the richest physically
and among the most easily soluble. This model is defined by the sum over geometries
⋆ Lectures delivered at the ICTP Spring School on String Theory and Quantum Gravity, Trieste,
April 1991.
† Supported in part by DOE grant DE-AC02-76WRO3072 and by an NSF Presidential Young
Investigator Award.
embedded in 1 dimension. The resulting string theory is, however, 2-dimensional
because the dynamical conformal factor of the world sheet geometry acts as an extra
hidden string coordinate.7 In fact, this is the maximal bosonic string theory that
is well-defined perturbatively. If the dimensionality is increased any further, then a
tachyon appears in the string spectrum and renders the theory unstable. The matrix
model maps the theory of surfaces embedded in 1 dimension into a non-relativistic
quantum mechanics of free fermions,8 from which virtually any quantity can be
calculated to all orders in the genus expansion. The second-quantized field theory
of non-relativistic fermions can be regarded here as the exact string field theory. A
transformation of this theory to a rather simple interacting boson representation9,10
will also be given. I will argue that the 2-dimensional string theory is the kind
of toy model which possesses a remarkably simple structure, and at the same time
incorporates some of the physics of string theories embedded in higher dimensions.
The simplicity of the theory is apparent in the matrix model approach, but much of
it remains obscure and mysterious from the point of view of the continuum Polyakov
path integral. Until we develop a better insight into the “miracles” of the matrix
models, we may be sure that our understanding of string physics is very incomplete.
These notes are mainly a review of published papers, but I also included a few
new observations and results. In section 2 I review the formulation of quantum grav-
ity coupled to a scalar field as a sum over discretized random surfaces embedded in 1
dimension. I will show that the matrix quantum mechanics generates the necessary
statistical sum. In section 3 I exhibit the reduction to free fermions and define the
double-scaling limit.5 The sum over continuous surfaces is calculated to all orders in
the genus expansion. In section 4 I show how to calculate the exact string amplitudes
using the free fermion formalism. Section 5 is devoted to the continuum path integral
formalism11 where the Liouville field acts as an extra dimension of string theory.
Some of the exact matrix model results will be reproduced, but this approach still
falls far short of the power of the matrix model. In section 6 I discuss the special
states which exist in the spectrum in addition to the massless “tachyon”. These
states, occurring only at integer momenta, are left-overs of the transverse excitations
2
of string theory. Remarkably, they generate a chiral W1+∞ algebra. In section 7 I
use the matrix model to formulate the exact string field theory both in the fermionic
and in bosonic terms. I present some manifestly finite bosonic calculations which are
in perfect agreement with the fermionic ones. In section 8 I discuss the new physical
effects that arise when the random surfaces are embedded in a circle of radius R: the
R → 1/R duality and its breaking due to Kosterlitz-Thouless vortices.6 This model
can be regarded either as a compactified Euclidean string theory or as a Minkowski
signature string theory at finite temperature. In section 9 I show that in the matrix
quantum the effects of vortices are implemented by the states in the non-trivial repre-
sentations of SU(N).6,12 Including only the SU(N) singlet sector gives the vortex-free
continuum limit where the sum over surfaces respects R → 1/R duality.6 In section
10 I use the thermal field theory of non-relativistic fermions to find exact amplitudes
of the compactified string theory.13 Finally, in section 11 I use the matrix chain model
to solve a string theory with a discretized embedding dimension.6 Remarkably, when
the lattice spacing is not too large, this theory is exactly equivalent to string theory
with a continuous embedding.
In this section I will introduce the discretized approach to summation over random
surfaces embedded in one dimension and its matrix model implementation. If we
parametrize continuous surfaces by coordinates (σ1 , σ2 ), then the Euclidean geometry
is described by the metric gµν (σ), and the embedding – by the scalar field X(σ).
Thus, the theory of random surfaces in one embedding dimension is equivalent to 2-d
quantum gravity coupled to a scalar field. In the Euclidean path integral approach to
such a theory, we have to sum over all the compact connected geometries and their
embeddings,
X Z
Z= [Dgµν ][DX]e−S . (2.1)
topologies
3
tions describing the same geometry-embedding in different coordinates should not be
counted separately. We assume the simplest generally covariant massless action,
1 √ 1
Z
S= d2 σ g( ′ g µν ∂µ X∂ν X + ΦR + 4λ) . (2.2)
4π α
In 2 dimensions the Einstein term is well-known to give the Euler characteristic, the
topological invariant which depends only on the genus of the surface h (the number
of handles),
1 √
Z
d2 σ gR = 2(1 − h) . (2.3)
4π
Thus, the weighting factor for a surface of genus h is (exp Φ)2h−2 . Since the sum over
genus h surfaces can be thought of as a diagram of string theory with h loops, we
identify the string coupling constant as g0 = eΦ .
The main problem faced in the calculation of the Euclidean path integral of eq.
(2.1) is the generally covariant definition of the measure for the sum over metrics
[Dgµν ] and of the cut-off. One may attempt to do this directly in the continuum.
Considerable success along this route has been achieved when metrics are described in
the light-cone gauge,14 or in the more traditional conformal gauge.15 We will carry
out some comparisons with the continuum approach in the course of the lectures.
The main subject of this section is a different approach to summing over geome-
tries, which has so far proven to be far more powerful than the continuum methods.
In this approach one sums over discrete approximations to smooth surfaces, and then
defines the continuum sum by taking the lattice spacing to zero.1 We may, for in-
stance, choose to approximate surfaces by collections of equilateral triangles of side a.
A small section of such a triangulated surface is shown in fig. 1. The dotted lines show
the lattice of coordination number 3, which is dual to the triangular lattice. Each face
of the triangular lattice is thought of as flat, and the curvature is entirely concentrated
at the vertices. Indeed, each vertex I of the triangular lattice has a conical singularity
with deficit angle π3 (6 − qI ), where qI is the number of triangles that meet at I. Thus,
4
I i
J
j
Fig. 1. A small section of triangulated surface. Solid lines denote the trian-
gular lattice Λ, and dotted lines – the dual lattice Λ̃.
the vertex I has a δ-function of curvature with positive, zero, or negative strength
if qI < 6, = 6, > 6 respectively. Such a distribution of curvature may seem like a
“poor man’s version” of geometry. However, in the continuum limit the size of each
face becomes infinitesimal, and we may define smoothed out curvature by averaging
over many triangles. In this way, the continuum field for the metric should appear
similarly to how the quantum field description emerges in the continuum limit of the
more familiar statistical systems, such as the Ising model, the XY model, etc. Later
on, we will show some strong indications that the discretized approach to summing
over geometries is indeed equivalent to the continuum field-theoretic approach.
The essential assumption of the discretized approach is that sum over genus h
R
geometries, [Dgµν ]h , may be defined as the sum over all distinct lattices Λ, with the
R
lattice spacing subsequently taken to zero. [DX] is then defined as the integral over
all possible embeddings of the lattice Λ in the real line. For simplicity, the lattices
Λ may be taken to be triangular, but admixture of higher polygons should not affect
the continuum limit. In order to complete the definition of the discretized approach,
5
we need to specify the weight attached to each configuration. This can be defined by
simply discretizing the action of eq. (2.2) and counting each distinct configuration
with weight e−S . We will find it convenient to specify the embedding coordinates
X at the centers of the triangles, i.e. at the vertices i of the dual lattice. Then the
1
R 2 √ µν P 2
discretized version of 4πα′ d σ gg ∂µ X∂ν X is ∼ <ij> (Xi − Xj ) , where the
sum runs over all the links < ij > of the dual lattice. Similarly,
√
√ 3 2
Z
d2 σ g → a V (2.4)
4
where V is the number of triangles, or, equivalently, the number of vertices of the
dual lattice Λ̃. Thus, the discretized version of the path integral (2.1) is
V Z
2(h−1)
X X Y Y
V
Z(g0 , κ) = g0 κ dXi G(Xi − Xj ) , (2.5)
h Λ i=1 hiji
√
where Λ are all triangular lattices of genus h, κ ∼ exp(− 3λa2 /4π), and (for some
choice of α′ ) G(X) = exp(− 21 X 2 ). If the continuum limit of eq. (2.5) indeed describes
quantum gravity coupled to a scalar field, then there should exist a whole universality
class of link factors G, of which the gaussian is only one representative, that result in
the same continuum theory.
A direct evaluation of the lattice sum (2.5) seems to be a daunting task, which
is still outside the numerical power of modern computers. Fortunately, there exists
a remarkable trick which allows us to exactly evaluate sums over surfaces of any
topology using only analytic tools. As was first noted by Kazakov and Migdal,16 a
statistical sum of the form (2.5) is generated in the Feynman graph expansion of
⋆
the quantum mechanics of a N × N hermitian matrix. Consider the Euclidean path
⋆ Similar tools work for other simple matter systems coupled to two-dimensional gravity. For
example, as discussed in other lectures in this volume, in the case of pure gravity the sum over
discretized surfaces is generated simply by an integral over an N × N hermitian matrix.
6
integral
ZT /2 2 !
∂Φ
Z
2
Z= D N Φ(x) exp −β dx Tr 1
2 + U(Φ) . (2.6)
∂x
−T /2
1
where x is the Euclidean time and U = 2α′ Φ
2 − 3!1 Φ3 . The parameter β enters as the
inverse Planck constant. By a rescaling of Φ eq. (2.6) can be brought to the form
ZT /2 2 !
∂Φ 1 2 κ 3
Z
2
Z∼ D N Φ(x) exp −N dx Tr 1
2 + Φ − Φ , (2.7)
∂x 2α′ 3!
−T /2
p
where κ = N/β is the cubic coupling constant. The connection with the statistical
sum (2.5) follows when we develop the graph expansion of Z in powers of κ. The
Feynman graphs all have coordination number 3, and are in one-to-one correspon-
dence with the dual lattices Λ̃ of the discretized random surfaces (fig. 1). The lattices
Λ dual to the Feynman graphs can be thought of as the basic triangulations. One
easily obtains the sum over all connected graphs ln Z,
V Z∞
X X Y Y ′
2−2h V
lim ln Z = N κ dxi e−|xi −xj |/α . (2.8)
T →∞
h Λ i=1−∞ hiji
This is precisely of the same form as eq. (2.5) which arises in two-dimensional quan-
tum gravity! The Euclidean time x assumes the role of the embedding coordinate
X in eq. (2.5). We note that the role of the link factor G is played by the one-
dimensional Euclidean propagator. Only for this exponential G can we establish the
exact equivalence with the matrix model. This does not pose a problem, however, as
we will find plenty of evidence that the continuum limit of the model (2.5) with the
exponential G indeed describes quantum gravity coupled to a scalar, eq. (2.1). It is
evident from (2.8) that the parameter α′ sets the scale of the embedding coordinate.
In fact, we have normalized α′ so that in the continuum limit it will precisely coincide
with the definition in eq. (2.2). Whenever α′ is not explicitly mentioned, its value
has been set to 1.
7
Further comparing eqs. (2.8) and (2.5), we note that the size of the matrix
N enters as 1/g0 . Let us show why. Each vertex contributes a factor ∼ N, each
edge (propagator) ∼ 1/N, and each face contributes N because there are as many
index loops as there are faces. Thus, each graph is weighted by N V −E+F which,
by Euler’s theorem, equals N 2−2h . The expansion of the free energy of the matrix
quantum mechanics in powers of 1/N 2 automatically classifies surfaces according to
topology. It would seem that, in order to define a theory with a finite string coupling,
it is necessary to consider finite N which, as we will see, is associated with severe
difficulties. Fortunately, this naive expectation is false: in the continuum limit the
“bare” string coupling 1/N becomes infinitely multiplicatively renormalized. Thus, if
N is taken to ∞ simultaneously with the world sheet continuum limit, then we may
obtain a theory with a finite string coupling. This remarkable phenomenon, known as
the “double-scaling limit”2,5 , will allow us to calculate sums over continuum surfaces
of any topology.
In order to reach the continuum limit, it is necessary to increase the cubic coupling
κ = N/β ∼ N 0 until it reaches the critical value κc where the average number of
p
triangles in a surface begins to diverge. In this limit we may think of each triangle as
√
being of infinitesimal area 3a2 /4, so that the whole surface has a finite area and is
continuous. In the continuum limit κ → κc we define the cosmological constant
∆ = π(κ2c − κ2 ) . (2.9)
√
Recalling that κ = κc exp(− 3λa2 /4π), we establish the connection between ∆ and
the physical cosmological constant λ: λ ∼ ∆/a2 .
Above we have sketched the connection between matrix quantum mechanics and
triangulated random surfaces. It is easy to generalize this to the case where, in
addition to triangles, the surfaces consist of other n-gons: we simply have to add to
the matrix potential U(Φ) monomials Φn , n > 3. For consistency, the continuum limit
should not be sensitive to the precise manner in which the surfaces are discretized.
In the next section we will give the exact solution of the matrix model and show that
the continuum limit is indeed universal.
8
3. MATRIX QUANTUM MECHANICS AND FREE FERMIONS.
In the previous section we established the equivalence of the sum over connected
discretized surfaces to the logarithm of the path integral of the matrix quantum
mechanics. On the other hand, in the hamiltonian language,
Thus, as long as the initial and final states have some overlap with the ground state,
the ground state energy E0 will dominate the T → ∞ limit of eq. (3.1),
ln Z
lim = −βE0 . (3.2)
T →∞ T
The divergence of the sum over surfaces proportional to the length of the embedding
dimension arises due to the translation invariance. In order to calculate the sum
over surfaces embedded in the infinite real line R1 , all we need to find is the ground
state energy of the matrix quantum mechanics. To this end we carry out canonical
quantization of the SU(N) symmetric hermitian matrix quantum mechanics.8,17 The
Minkowski time lagrangian
9
matrix Ω̇Ω† can be decomposed in terms of the generators of SU(N),
N −1
i X X
Ω̇Ω† = √ α̇ij Tij + β̇ij T̃ij + α̇i Hi (3.6)
2 i<j i=1
where Hi are the diagonal generators of the Cartan subalgebra, and the other gener-
ators are defined as follows. Tij is the matrix M such that Mij = Mji = 1, and all
other entries are 0; T̃ij is the matrix M such that Mij = −Mji = −i, and all other
entries are 0. Substituting eq. (3.6) into eqs. (3.5) and (3.3), we find
X X
1 2
L= 2 λ̇i + U(λi ) + 12 (λi − λj )2 (α̇ij
2 2
+ β̇ij ) . (3.7)
i i<j
N
1 X 1 d 2 d
− 2 2
∆ (λ) . (3.8)
2β ∆ (λ) dλi dλi
i=1
It is left as an exercise for the reader to show that this can be rewritten as
1 X d2
− ∆(λ) . (3.9)
2β 2 ∆(λ)
i
dλi 2
1 X d2 X X Π2ij + Π̃2ij
H=− 2 ∆(λ) + U(λi ) + , (3.10)
2β ∆(λ) dλi 2 i
(λi − λj )2
i i<j
where Πij and Π̃ij are the momenta conjugate to αij and βij respectively, i.e., they are
the generators of left rotations on Ω, Ω → AΩ. Furthermore, since L is independent
of α̇i , the wave functions must obey the constraints Πi Ψ = 0. These constraints
arise because transformations Ω → AΩ, where A is a diagonal SU(N) matrix, do not
change Φ.
10
The constraints require that only those irreducible representations of SU(N)
that have a state with all weights equal to zero are allowed in the matrix quan-
tum mechanics.12 For instance, the fundamental representations are excluded, and
the simplest non-trivial representation is the adjoint. It is not hard to classify all the
irreducible representations allowed in the matrix quantum mechanics, i.e., the ones
that have states with zero weights. In the standard notation, the Young tableaux
of such representations consist of i + j columns, with the numbers of boxes in the
columns (from left to right)
(N − m1 , N − m2 , . . . , N − mj , ni , . . . , n2 , n1 ) (3.11)
Since the angular kinetic terms in (3.10) are positive definite, it is clear that we
have to look for the ground state among the wave functions that are annihilated
by Πij and Π̃ij , i.e. in the trivial (singlet) representation of SU(N). The singlet
wave functions are independent of the angles Ω, and are symmetric functions of the
eigenvalues, χsym (λ). Because of the special form of the hamiltonian (3.10), the
eigenvalue problem Hχsym = Eχsym assumes the remarkable form8
N
!
X
hi Ψ(λ) = EΨ(λ) ,
i=1 (3.12)
1 d2
hi = − 2 + U(λi ) ,
2β dλi 2
The N-fermion ground state is obtained by filling the lowest N energy levels of
11
h,
N
X
E0 = ǫi . (3.13)
i=1
The potential arising from triangulated random surfaces is U(λ) = 12 λ2 − 3!1 λ3 , which is
unbounded from below and does not seem support any stable states. Recall, however,
that in order to find the genus expansion of the sum over surfaces, we are only
interested in the expansion about the classical limit β → ∞ (h̄ → 0) in powers of
1/β 2 . Classically, there is a continuum of bounded orbits to the left of the potential
barrier. Semiclassically, these orbits are associated with energy levels (fig. 1) whose
spacing is O(1/β), and whose decay time is exponentially long as β → ∞. This
instability cannot be seen in the context of the semiclassical expansion in powers of
1/β 2 . Our goal, therefore, is to find the expansion of the ground state energy (3.13),
where ǫi are the semiclassical energy levels.
Zλ+p
N
2(µF − U(λ)) = π , (3.15)
β
λ−
where λ− and λ+ are the classical turning points. Eq. (3.15) shows that µF is an
increasing function of N/β. Clearly, both µF and E0 become singular at the point
where N/β → κ2c such that µF equals the height of the barrier µc . Recall that the
p
cubic coupling of the matrix quantum mechanics is κ = N/β, and that we expect
to find a singularity in the sum over graphs when κ becomes large enough that the
average number of vertices in a Feynman graph (or equivalently, of triangles in a
12
0
µF . βµ
..
..
.
random surface) begins to diverge. In the vicinity of this singularity the continuum
limit of quantum gravity can be defined. Now we have identified the physics of this
singularity in the equivalent free fermion system: it is associated with spilling of
fermions over the barrier.
1X
ρ(ǫ) = δ(ǫ − ǫn ), (3.16)
β n
in terms of which
13
ZµF
N
κ2 = = ρ(ǫ)dǫ ;
β
0
(3.17)
ZµF
− ln Z
lim = βE = β 2 ρ(ǫ)ǫdǫ .
T →∞ T
0
∂∆
= πρ(µF ) , (3.18)
∂µ
∂E 1
= βµ (3.19)
∂∆ π
These equations show that the continuum limit of the sum over surfaces is determined
by the singularity of the single particle density of states near the top of the barrier.
Using the WKB approximation, it was found in ref. 16 that
Zλ+
1 dλ 1
ρ(µF ) = p ∼ − ln µ + O(1/β 2 ) (3.20)
π 2(µF − U(λ)) π
λ−
1 1 β 2 ∆2
−βE = (βµ)2 ln µ + . . . ≈ + ... (3.21)
2π 2π ln ∆
The critical behavior of ρ comes from the part of the integral near the quadratic
maximum of the potential. In the continuum limit µ → 0, the classical trajectory at
µF spends logarithmically diverging amount of time near the maximum. Therefore,
the WKB wave functions are peaked there. In fact, we will show that, at any order
14
0
. βµ
.. ..
.. ..
. .
2
1 d 1 2
p 3
2 dy 2 + 2y + O 1/ β y ψ(y) = βeψ(y) (3.22)
For a finite βe, all the terms beyond the quadratic are irrelevant in the limit β → ∞
because they are suppressed by powers of β. This demonstrates the universality of
the continuum limit: any potential possessing a quadratic maximum will yield the
same sum over continuous surfaces. In effect, our rescaling has swept all the non-
universal details of the potential out to infinity. This implies, for instance, that,
instead of working with the potential of fig. 2, we can use the double well potential
√
U(λ) = − 12 λ2 + λ4 (fig. 3). After rescaling βλ = y, the Schrödinger equation again
reduces to the motion in the inverted harmonic oscillator. The only difference is that
now fermions fill both sides of the maximum. To all orders of the WKB expansion
this simply multiplies the density of states, and therefore the free energy, by a factor
of 2. The double well potential does not suffer from the non-perturbative instabilities:
the ground state energy can in principle be calculated even for finite N and β. We
will often adopt the symmetric case because it is more convenient for calculations,
remembering that we have to divide by 2 to find the sum over triangulated surfaces.
15
From eq. (3.21) we know the sum over all surfaces of spherical topology embedded
in R1 . In order to find the sums over surfaces of any topology, we need to develop a
systematic WKB expansion of the ground state energy. According to eqs. (3.18) and
(3.19), all we need to know is the complete expansion of the single particle density of
states ρ(µ). If we consider µ such that βµ is O(1), then this problem has a remarkably
simple solution,5 because eq. (3.22) shows that in this energy range the system is
purely quadratic. Let us write the density of states as
1 1
ρ(µ) = Tr δ(h0 + βµ) = Im Tr ,
π h0 + βµ − iǫ
(3.23)
1 d2 1
h0 = − 2
− y2 .
2 dy 2
We will evaluate the trace in position space in order to take advantage of the fact
that the resolvent of the hamiltonian of a simple harmonic oscillator is well known,
Z∞ Zyf
1 RT
dt 21 (ẏ 2 +ω 2 y 2 )
hyf | |yi i = dT e −βµT
Dy(t)e− 0
− 21 d2
dy 2
1 2 2
+ 2 ω y + βµ − iǫ
0 yi
(3.24)
Z∞ r
ω
(yf2 +yi2 ) cosh ωT −2yf yi /2 sinh ωT
= dT e−βµT e−ω .
2πsinhωT
0
In our case, the frequency ω is actually imaginary (we have an inverted harmonic
oscillator), so we must define the resolvent by analytic continuation. We rotate ω →
−i while simultaneously rotating T → iT , and the relevant resolvent is
Z∞ r
1 −i
−iβµT i (yf2 +yi2 ) cosh T −2yf yi /2 sinh T
hyf | |yii = i dT e e .
h0 + βµ − iǫ 2πsinhT
0
(3.25)
Now it is easy to derive the derivative of the density of states5
Z∞ Z∞
∂ρ 1 ∂ 1 1 T /2
= Im dyhy| |yi = Im dT e−iβµT .
∂(βµ) π ∂(βµ) h0 + βµ − iǫ π sinh(T /2)
−∞ 0
(3.26)
Note that only for the purposes of deriving the large βµ expansion can we reduce
16
the problem to pure inverted harmonic oscillator. The integral representation (3.26)
is not designed to correctly give the non-perturbative terms O(e−βµ ). In fact, such
terms depend on the details of the potential away from the quadratic maximum, and
are not universal. While each term in the large βµ expansion is related to the sum
over all geometries of a certain genus, we do not yet fully understand the geometrical
or physical meaning of the non-perturbative terms.
After expanding eq. (3.26) and integrating once, we find the complete series of
corrections to the leading order estimate in eq. (3.20),
∞
( )
1 X |B2m |
ρ(µ) = − ln µ + (22m−1 − 1) (2βµ)−2m . (3.27)
π m
m=1
∆ = −µ0 ln µ0 = µ0 | ln µ0 | . (3.28)
Then,
∞
( )
1 X 2m−1 |B2m | 1
µ = µ0 1− (2 − 1) (2βµ0 )−2m + O . (3.29)
ln µ0
m=1
m(2m − 1) ln2 µ0
Integrating (3.19) we finally arrive at the complete genus expansion of the sum over
surfaces
∞
( )
1 2 1 X (22m+1 − 1)|B2m+2 |
−βE = (2βµ0 ) ln µ0 − ln µ0 + (2βµ0)−2m .
8π 3 m(m + 1)(2m + 1)
m=1
(3.30)
Clearly, the sensible way to approach the continuum limit is through “double scaling”
β → ∞ and µ0 → 0 in such a way that βµ0 = 1/gst is kept fixed.5,6 Thus, quite
remarkably, we have found closed form expressions for sums over continuous surfaces
of any topology embedded in one dimension. The sum over genus h surfaces, given
2h−2
by the coefficient of gst , grows as (2h)!. This behavior is characteristic of closed
17
string theories,18 while in field theory we typically find that the sum over h-loop
diagrams exhibits a slower growth ∼ h!. The badly divergent perturbation series
(3.30) indicates that the theory is not well-defined non-perturbatively.
Eq. (3.30) shows explicitly how the bare string coupling 1/β gets multiplicatively
renormalized. Compared to the c < 1 matter coupled to gravity,2−4 the new feature
here is that the renormalization is not simply by a power of the cosmological constant
∆. It appears that ∆ itself is multiplicatively renormalized: ∆ → µ0 ≈ ∆/| ln ∆|.
Another peculiarity of the c = 1 matter is that the sums over spherical and toroidal
surfaces diverge in the continuum limit as | ln µ0 |. In section 5 we will argue that all
these features have a natural explanation in the context of Liouville theory. Before
we do that, however, we further exhibit the power of the matrix model by calculating
some correlation functions to all orders in the topological expansion.
4. CORRELATION FUNCTIONS.
The simplest kind of matrix model operator is Tr Φn (x) where n is any finite
integer. This operator inserts a vertex of order n in all possible places on the dual
graph Λ̃. In terms of the basic lattice Λ, this amounts to cutting an n-gon hole in
the surface, pinning it at the embedding coordinate x, and integrating the position
of the hole all over the surface. In the continuum limit, for any finite n the size of
the hole becomes infinitesimal, and it is commonly referred to as “a puncture”. The
continuum expression for the operator above should be roughly
√
Z
d2 σ gδ X(σ) − x) . (4.1)
√
Z Z
iqx n
dxe Tr Φ (x) → d2 σ geiqX , (4.2)
which resembles the basic “tachyon” operator of string theory. Below we outline the
procedure for calculating the correlation functions of such operators in the matrix
18
model. In section 5 we show that the results can indeed be reproduced in the string
theoretic formalism.
Since the matrix model has been reduced to a system of free fermions, it is con-
venient to perform calculations in the formalism of non-relativistic second quantized
field theory. The second quantized fermion field is defined as
Z∞
ψ̂(λ, x) = dνeiνx aǫ (ν)ψ ǫ (ν, λ) , (4.3)
−∞
where ψ ǫ (ν, λ) are the single fermion wave functions of energy ν, and the associated
† √
annihilation operators aǫ (ν) satisfy {aǫ (ν), aǫ (ν ′ )} = δǫ,ǫ′ δ(ν − ν ′ ). Setting λ = 2y
in eq. (3.22), we find that ψ ǫ (ν, λ) are solutions of the Whittaker equation
d2 λ2
( + )ψ = νψ. (4.4)
dλ2 4
and can therefore be expressed in terms of Whittaker functions W (ν, λ).19 Their
spectrum is continuous and doubly degenerate. Here ǫ denotes the parity of the
wavefunction ψ ǫ (ν, λ) and repeated ǫ are summed over.
where x is the Euclidean time. The ground state of the system (fig. 3b) satisfies
aǫ (ν)|βµi = 0, ν < βµ ,
(4.6)
a†ǫ (ν)|βµi = 0, ν > βµ .
19
where f is any function. In the second quantized formalism, they translate into
Z Z
O(q) = dxeiqx
dλf (λ)ψ̂ † ψ̂(x, λ) . (4.8)
Thus, to find the connected correlation functions of any set of such operators, we
need to calculate
n Z
dxi eiqi xi hβµ|ψ̂ †ψ̂(x1 , λ1 ) . . . ψ̂ † ψ̂(xn , λn )|βµic . (4.9)
Y
G(q1 , λ1 ; . . . ; qn , λn ) =
i=1
Application of Wick’s theorem reduces this to the sum over one-loop diagrams with
fermion bilinear insertions in all possible orders around the loop. A convenient formula
for the Euclidean Green function is19
Z∞
E
S (x1 , λ1 ; x2 , λ2 ) = dνe−(ν−βµ)∆x {θ(∆x)θ(ν − βµ) − θ(−∆x)θ(βµ − ν)}×
−∞
Z∞ Z∞
ǫ ǫ dp −ip∆x i
ψ (ν, λ1 )ψ (ν, λ2 ) = e dν ψ ǫ (ν, λ1 )ψ ǫ (ν, λ2 )
2π p + i(ν − βµ)
−∞ −∞
Z∞ sgn(p)∞
dp −ip∆x
Z
=i e dse−sp+iβµs hλ1 |e2isH |λ2 i
2π
−∞ 0
(4.10)
where ∆x = x1 − x2 , and we have used19
Z∞
dνe−iνs ψ ǫ (ν, λ1 )ψ ǫ (ν, λ2 ) = hλ1 |e2isH |λ2 i =
−∞ (4.11)
i λ21 + λ22 2λ1 λ2
1
√ exp − −
−4πi sh s 4 th s sh s
with H = 1/2p2 − 1/8λ2 . With the Green function (4.10), after integration over the
20
loop momentum, Moore derived the representation
∞ ǫZ1 ∞ ǫn−1
Z ∞
∂ X X Z
G(q1 , λ1 ; . . . ; qn , λn ) = in+1 δ( qi ) dξeiµξ ds1 . . . dsn−1
∂µ
σ∈Σn −∞ 0 0
Pn−1
−s1 Qσ1 −...−sn−1 Qσn−1 2is1 H 2i(ξ− si )H
e hλσ(1) |e |λσ(2) i . . . hλσ(n) |e 1 |λσ(1) i
(4.12)
where Qσk = qσ(1) + . . . + qσ(k) , and ǫk = sgn[Qσk ]. In principle, any correlation
function of operators of type (4.7) can be found by integrating G. A convenient way
of calculating correlations of puncture operators is to first introduce the operator that
creates a finite boundary of length l, with momentum q injected into the boundary19
Z
O(l, q) = dxeiqx Tr e−lΦ(x) . (4.13)
The puncture operator should be the leading term in the small l expansion of O(l, q).
The correlations of puncture operators P (q) can be extracted according to
* + * +
Y Y |q | Y
O(li , qi ) ∼ li i P (qj ) . (4.14)
i i j
The details of the calculations are highly technical and can be found in ref. 19. Here
we will simply state the results for the two, three, and four point functions.
" #
1 1
hP (q)P (k)i = −δ(q + k)[Γ(1 − |q|)]2 µ|q| − (|q| − 1)(q 2 − |q| − 1) + . . . .
|q| 24(βµ)2
(4.15)
3
"
X Y |qi |
1
hP (q1 )P (q2 )P (q3 )i = δ( qi ) Γ(1 − |qi |)µ 2 1
βµ
i=1
1
− (|q3 | − 1)(|q3 | − 2)(q12 + q22 − |q3 | − 1)
24(βµ)2
Y4
4 4 2 2 2 2 2
+ (|q3 | − r) 3(q1 + q2 ) + 10q1 q2 − 10(q1 + q2 )|q3 | − 5(q1 − q2 ) + 12|q3 | + 7
r=1
#
1
× + ... ,
5760(βµ)2
(4.16)
21
where q1 , q2 > 0.
4
|qi |
X Y
hP (q1 )P (q2 )P (q3 )P (q4 )i = −δ( qi ) Γ(1 − |qi |)µ 2 ×
i=1 (4.17)
1 2
(|q1 + q2 | + |q1 + q3 | + |q1 + q4 | − 2) + O(1/(βµ) ) .
2(βµ)2
The new feature of the c = 1 correlators, compared to theories with c < 1, is that they
have genuine divergences occurring at quantized values of the external momenta.20 In
this sense, the c = 1 model is the most similar to critical string theory where the
amplitudes are well-known to have divergences associated with the production of on-
shell physical particles. The remarkable feature of the c = 1 amplitudes is that all
the divergences factorize into external leg factors.
Because of the divergences, eqs. (4.15)-(4.17) are strictly valid only if none of
qi are integer. If any of the momenta have integer values, then additional terms
need to be taken into account, which leads to cancellation of the infinity.20 The
physical meaning of this is simple: in the matrix model there is an explicit ultravi-
olet and infrared cut-off, and therefore there are no genuine zeroes or poles of the
correlation functions. Instead, they are regularized according to 1/0 → | ln µ0 |, and
0 → 1/| ln µ0 |. Note that in eqs. (4.15)-(4.17) µ can be replaced by µ0 . Thus, the
correlation functions scale as powers of µ0 , and not as powers of ∆ found in all the
c < 1 theories coupled to gravity.2−4 This provides further evidence that, for c = 1,
a renormalization of the cosmological constant takes place. In the next section we
sketch a Liouville theory explanation of some of the new effects peculiar to c = 1.
22
5. LIOUVILLE GRAVITY COUPLED TO c = 1 MATTER.
In this section we outline the continuum approach to evaluating the sum over
surfaces (2.1). In order to integrate over all metrics, in 2 dimensions we may pick the
conformal gauge gµν = ĝµν (τ )e−φ where ĝµν depends on a finite number of modular
parameters collectively denoted as τ . For spherical topology, there are no moduli,
and we may choose ĝµν = δµν . Eq. (2.2) shows that classically the Liouville field φ is
non-dynamical. It is well-known, however,11 that the kinetic term for φ is induced
through quantum effects arising from dependence on φ of the path integral measure.
Additional modification of the dynamics of φ has been proposed by David and by
Distler and Kawai,15 who argued that we can simplify the integration measure at the
expense of further renormalization of the action for φ such that the ghost, X and φ
fields combine into a conformally invariant theory with net conformal anomaly equal
to zero. Following their approach, we assume that the original path integral (2.1) is
transformed in the conformal gauge into
X Z
2h−2
Z= g0 [dτ ][DbDc][Dφ][DX]e−Sb,c −S ,
h (5.1)
1
Z p
2 µ µ −2φ
S= d σ ĝ ∂µ X∂ X + ∂µ φ∂ φ − 4R̂φ + 4∆e ,
4π
and Sb,c is the standard free action for the b and c ghosts. The exponential interaction
⋆
term originates from the cosmological term in the action (2.2).
With the action (5.1), the simplest set of conformally invariant operators are
1
Z
d2 σ ĝeiqX+ǫ(q)φ
p
T (q) = (5.2)
πβ
These operators inject embedding momentum q into the world sheet. They should be
thought of as the Liouville theory implementation of the operators (4.2). In a theory
⋆ Polchinski has argued21 that, for c = 1, the interaction term is not simply an exponential.
We will instead work with the exponential interaction, and will show that it is also possible
to explain correlation functions within this framework. Our argument does not substantially
differ from Polchinski’s, but we prefer working with the exponential because it was shown by
Curtwright and Thorn22 that with this form of interaction the conformal invariance can be
maintained exactly in the continuum limit.
23
with ∆ = 0, T (q) is conformally invariant for ǫ(q) = −2 ± |q|. For ∆ > 0, however,
it was shown in ref. 23, 24 that only the operator with ǫ(q) = −2 + |q| exists. This
conclusion agrees with the scaling of correlation functions in the matrix model. Let
us now use the Liouville formalism to calculate the spherical correlation functions.
*N + s
Y X ∆
T (qi ) = g02h−2 δ( qi ) 21 Γ(−s)
π
i=1 h
Z s Y
1
Z Z
−Sb,c −S0 2 −2φ̃
d2 σ
p p iq X+(−2+|q |)φ
[dτ ][D X̃][D φ̃][DbDc]e d σ ĝe ĝe i i
πβ
i
(5.3)
where
1
Z p
S0 = d2 σ ĝ ∂µ X̃∂ µ X̃ + ∂µ φ̃∂ µ φ̃ (5.4)
4π
1 PN
and s = 2(1 − h) − N + 2 i=1 |qi |. The factor Γ(−s) indicates that correlation
functions with non-negative integer s have a divergence from the volume of φ0 . This
explains the divergences in the 0-point functions on the sphere and torus found in
eq. (3.30). Now, the difficulty is to deal with the insertion of T (0)s . Goulian and
Lee26 proposed to calculate eq. (5.3) first for integer s, where the rules are perfectly
clear, and then to define it for non-integer s by analytic continuation. This gives
answers in agreement with the matrix model.26,27
24
for the coefficient of 12 Γ(0) can be performed explicitly27 and gives
M −1
β 2−M Y Γ(1 − |qi |)
, (5.5)
(M − 3)! Γ(|qi |)
i=1
The physical meaning of the zero of order m is that the “tachyon” decouples at zero
momentum. This is a sign of the fact that the “tachyon” is actually a massless particle
in two-dimensional string theory. We will show this more explicitly later on.
The continuation of eq. (5.6) to non-integer s gives the N-point tachyon ampli-
tudes with qN < 0, q1 , . . . , qN −1 > 0,27
*N + N N −3
Y
1 2−N
Y Γ(1 − |qi |) d
T (qi ) = 2 β µs+N
0
−3
. (5.7)
Γ(|qi |) dµ0
i=1 i=1
This agrees with the matrix model results (4.15)-(4.17) if the matrix model and
Liouville theory operators are related by momentum dependent normalization factor
1
T (q) = P (q) (5.8)
Γ(|q|)
With this correspondence, we can use the matrix model results to obtain tachyon
correlation functions to all orders in the genus expansion. This is remarkable because
25
direct Liouville theory calculations of correlation functions have not yet been per-
formed even for genus 1. For h > 3 we do not know the region of modular integration
and could not perform a direct Liouville calculation even in principle. From eq. (4.15)
we find, for instance,
2 " #
Γ(1 − |q|) |q| 1 1 2
hT (q)T (k)i = −δ(q + k) µ0 − (|q| −1)(q −|q| −1) + . . . .
Γ(|q|) |q| 24(βµ)2
(5.9)
A natural continuation of this formula to q = 0 is to replace limq→0 |q| by 1/| ln µ0 |.
Then we obtain,
∂2
hT (0)T (0)i ∼ E. (5.10)
∂∆2
This is a simple consistency check: differentiation of the path integral with respect
to ∆ inserts a zero-momentum tachyon.
To conclude this section, we will describe the only h > 0 calculation that has
been performed to date. This is the sum over toroidal surfaces with no insertions.28
The reason this calculation is relatively straightforward is that s = 0, and the com-
plicated insertion of T (0)s is absent. Sending s → 0 in eq. (5.3), and remembering
the renormalization ∆ → µ0 , we obtain the free field path integral
Z Z
2
Z1 = − 12 ln µ0 d τ [D X̃][D φ̃][DbDc]e−S0 (5.11)
F
where τ = τ1 + iτ2 is the modular parameter of the torus, and F denotes the funda-
mental region of the modular group: τ2 > 0, |τ | > 1, − 12 ≤ τ1 < 12 . This path integral
is standard, and we find
√ |η(q)|4 √
√
Z
2
Z1 (R/ α′ ) = − 12 ln µ0 d τ (2π τ2 |η(q)|2 )−1 Z(R/ α′ , τ, τ̄ ) , (5.12)
2τ2
F
where η is the Dedekind function and q = e2πiτ . In the above integrand, the first
term is from the ghost determinant, the second is from the Liouville determinant, and
26
√
Z(R/ α′ , τ, τ̄ ) is the partition function of the scalar field compactified on a circle,
√ ∞
πR2 |n − mτ |2
R 1 X
Z(R/ α′ , τ, τ̄ ) =√ √ exp − . (5.13)
α′ τ 2 |η(q)|2 n,m=−∞ α′ τ2
The double sum is over the soliton winding numbers about the two non-trivial cycles
of the torus. From eqs. (5.12) and (5.13), we obtain
√ R √
Z1 (R/ α′ ) = − 12 ln µ0 √ F (R/ α′ ) ,
4π α′
√ Z
d2 τ X
πR2 |n − mτ |2
(5.14)
′
F (R/ α ) = exp − .
τ22 n,m α′ τ2
F
The η-functions cancel out, and we end up only with the contributions of the zero
modes. This is due to the absence of particles corresponding to the transverse string
⋆
excitations, and reflects the two-dimensional nature of this string theory. Remark-
ably, the integral in eq. (5.14) can be easily performed using the trick of ref. 30.
The idea is to trade the m-sum over the winding sectors of the string for a sum over
many inequivalent fundamental regions which together cover the strip − 12 ≤ τ1 < 1
2
in the upper half-plane. In 26 dimensions this gives a representation of the string free
energy in terms of the free energies of all the modes of the string31 . Similarly, eq.
(5.14) becomes
∞∞
√ Z 2 2 k2
d τ dτ πR
Z
2
X
F (R/ α′ ) = +2 exp − ′ . (5.15)
τ22 τ22 α τ2
F 0 k=1
The second term is simply the temperature-dependent one-loop free energy for a
single massless boson in 2 dimensions, expressed in the proper time representation.
The first term is the one-loop cosmological constant of the massless boson, which has
been automatically supplied with a ‘stringy’ ultraviolet cut-off: here the τ integral
⋆ The isolated transverse states at discrete momenta, to be discussed in section 6, do not affect
the torus partition function.29
27
is over the fundamental region, not over the strip. Performing the integrals in eq.
√
(5.15), we find F (R/ α′ ) = π3 (1 + α′ /R2 ) and
√
√ 1 R α′
Z1 (R/ α′ ) = − ( √ + ) ln µ0 . (5.16)
24 α′ R
As we will show in section 9, this answer agrees with the matrix model result. This
calculation gives us hope that the Liouville path integral exactly describes sums over
geometries beyond the tree level.
The most striking feature of the tachyon correlation functions is the occurrence
Γ(1−|qi |)
of the external leg factors Γ(|qi |) , which contain factorized poles whenever any of
the momenta approaches an integer value. Although this exact factorization is not
yet fully understood, it was noted in ref. 20 that the poles are related to the presence
of other physical states in the theory besides the tachyons (5.2). These states are
remnants of the transverse excitations of the string and occur only at integer q.
20,32,33
The origin of these states can be traced to the pure c = 1 conformal field theory
where there are special primary fields of the form34
with conformal weight J 2 . These states form SU(2) multiplets with spin J and
magnetic number m. The connection with SU(2) becomes apparent once we note
that the states (6.1) are the full set of primary fields in the compact c = 1 theory
with the self-dual radius R = 1, where there is a well-known level 1 SU(2) current
algebra. For each spin J, VJ,±J = 1, which gives a tachyon operator. Other mem-
bers of the multiplet can be constructed by applying raising and lowering operators
28
1
H
2πi dz exp(±2iX(z)). After coupling to gravity, we can “dress” the special c = 1
operators to obtain new (1, 0) operators of the form
†
Recently, it was established that these operators are related to a W1+∞ algebra.35
Undoubtedly, this algebra plays an important role in determining the properties of
the theory.
Z∞
dxe2miX (λ − λc )2J (6.4)
−∞
where λc is the coordinate of the quadratic maximum. Using this identification, the
correlation functions of all operators with m = 0 were calculated in ref. 10. Recently,
it was shown37,38 that the operators (6.4) generate a W1+∞ algebra. The connection
of this algebra with the algebra of special operators in Liouville theory has been
established by E. Witten.35
¯ .
V1,0 V̄1,0 = ∂X ∂X (6.5)
Thus, of the full dilaton field of the critical string theory only its zero-momentum
part remains here. Similarly, the higher J excitations correspond to remnants of the
higher mass particles. The connection of the special operators to the divergences in
29
the correlation functions is provided by the fusion rules. There are terms ∼ 1/|z −w|2
in the fusion rules of the physical operators (6.3). Integration over one operator near
another can thus produce logarithmic divergences.
1
einX+(−2+|n|)φ (z, z̄)e−inX+(−2+|n|)φ (w, w̄) ∼ . . . + V V̄ + ... .
|z − w|2 |n|−1,0 |n|−1,0
(6.6)
Integrating over z near w, we obtain a logarithmic divergence for any integer |q| > 0
due to the appearance of a physical operator in the fusion rule. This explains the
infinite sequence of poles in the two-point function.
The success of the Liouville approach shows that the theory of surfaces embed-
ded in one dimension can also be viewed as critical string theory embedded in two
dimensions.7,21 The field φ, which starts out as the conformal factor of the world sheet
metric, assumes the role of another embedding dimension similar to X. Indeed, the
path integral (5.1) can be thought of as the sigma model for the bosonic string prop-
agating in two-dimensional Euclidean target space with the metric Gµν = δµν , the
dilaton background D(φ) = −4φ, and the tachyon background T (φ) = ∆e−2φ . The
two-dimensionality of the string theory has many important physical consequences.
First of all, the mass of the particle corresponding to motion of the string in its
ground state, “the tachyon” is
2−D
m2T = . (7.1)
6
Thus, D = c+1 = 2 is the “critical dimension of non-critical string theory” where “the
tachyon” is exactly massless. Indeed, in the calculation of the Liouville path integral
30
on the torus, we found a massless two-dimensional boson propagating around the
loop. From the decoupling at zero momentum, observed in the study of the tachyon
correlation functions, we deduce that the theory is invariant under constant shifts of
the tachyon field.
In this section we show that a string field theory with the features anticipated
above can be derived directly from the matrix quantum mechanics. This further
strengthens the connection of c = 1 quantum gravity with string theory in D = 2.
The matrix quantum mechanics directly reduces to the exact string field theory, which
is the second-quantized field theory of non-relativistic fermions discussed in section 4.
In this section we will simply recast this formalism so that some of its physical features
become more transparent. First, we will show that the non-relativistic field theory
can be expressed in the double-scaling limit as a theory of free quasi-relativistic chiral
fermions.10,39 These chiral fermions have the kinetic term that is relativistic to order
0 , but receives translationally non-invariant corrections of order g . Further, we will
gst st
carry out the conventional bosonization40 of this fermionic hamiltonian, and obtain
an interacting theory of massless bosons in D = 2 first derived by Das and Jevicki
with somewhat different methods.9 In conclusion, we will perform a few manifestly
finite calculations in the bosonic formalism to show that it is fully equivalent to the
original representation in terms of non-relativistic fermions.
31
7.1. Chiral Fermions.
1 ∂Ψ† ∂Ψ
Z
ĥ = dλ 2
+ U(λ)Ψ† Ψ − µF (Ψ† Ψ − N) , (7.2)
2β ∂λ ∂λ
where µF is the Lagrange multiplier necessary to fix the total number of fermions to
equal N. As usual, it will be adjusted so as to equal the Fermi level of the N fermion
system, µF = ǫN . The fermion field has the expansion
X
Ψ(λ, t) = αi ψi (λ)e−iβǫi t , (7.3)
i
where ψi (λ) are the single particle wave functions and αi are the respective annihila-
tion operators. The time t is now taken to be Minkowskian.
All the known exact results of the c = 1 matrix model have been derived from
this nonrelativistic field theory. An important feature of this theory is that it is two-
dimensional: the field Ψ(λ, t) depends on t, and also on the eigenvalue coordinate
λ. This is the simplest way to see how the hidden dimension, originating from the
Liouville mode of quantum gravity, emerges in the matrix model.
This suggests that the behavior of excitations near the Fermi surface is quasi-relativistic.
The spectrum of relativistic fermions in a box is, of course, exactly linear. We will
therefore find non-relativistic corrections arising from the violations of the linearity of
ǫn due to higher orders of the semiclassical expansion. As the Fermi level approaches
32
the maximum, µ = µc − µF → 0, the spacing of the spectrum vanishes, ω ≈ π/| ln µ|.
Thus, we expect that the size of the box in which the quasi-relativistic fermions are
enclosed diverges ∼ | ln µ|.
where v(λ) is the velocity of the classical trajectory of a particle at the Fermi level in
the potential U(λ),
dλ p
v(λ) = = 2(µF − U(λ)) . (7.6)
dτ
ZT /2
1
dτ iΨ†R ∂τ ΨR − iΨ†L ∂τ ΨL + † †
H = 2β ĥ = ∂ τ Ψ L ∂τ Ψ L + ∂τ Ψ R ∂τ Ψ R
2βv 2
0 (7.7)
′′ ′ 2
1 v 5(v )
+ Ψ†L ΨL + Ψ†R ΨR − ,
4β v3 2v 4
where v ′ ≡ dv/dτ . Here we see that the natural spatial coordinate, in terms of which
the fermion has a standard Dirac action to leading order in β, is τ – the classical time
33
of motion at the Fermi level – rather than λ. An important feature of the hamiltonian
is that it does not mix ΨL and ΨR ,
H = HL + HR . (7.8)
The only mixing between the different chiralities is through the boundary conditions.
In order to determine the boundary conditions, consider the leading semiclassical
expression for Ψ(λ, 0),
X X
Ψ(λ, 0) = ψn an + ψm b†m ,
n>0 m≤0
Zλ (7.9)
2 π
ψn = √ cos β dλ′ vn (λ′ ) −
Tv 4
p n
vn = 2(ǫn − U(λ)) ,
T T
ΨR (τ = 0) = ΨL (τ = 0), ΨR (τ = ) = ΨL (τ = ) . (7.11)
2 2
These ensure that the fermion number current not flow out of the finite interval, i.e.,
that Ψ̄(τ )γ1 Ψ(τ ) = Ψ†R ΨR −Ψ†L ΨL vanish at the boundary. They also guarantee that
34
ΨR and ΨL are not independent fields and that we are including the correct number of
degrees of freedom. The issue of dynamics at the boundary is quite subtle, however,
since the corrections to the relativistic hamiltonian in eq. (7.7) blow up precisely
at the boundary. The problem is that the semiclassical approximation breaks down
at the points where v(τ ) = 0 even as β → 0. One possibility of dealing with this
breakdown is to carefully regularize the physics at the boundary points. This has
lead to some success.41,42 We will suggest another method, which is problem-free
from the beginning: to approach double-scaling limit with µ < 0. As we will show,
an equivalent procedure is to work with µ > 0 but to construct the theory in terms
of Ψ(p, t), where p is the momentum conjugate to λ.
Z∞
1 dτ 2 1 2 5 2
Hgst = |∂τ Ψi | + 2 |Ψi | 1 − coth (τ ) , (7.12)
4βµ sinh2 (τ ) 2
0
where i runs over L and R. This correction does not change the non-interacting
nature of the fermions, but it does render the fermion propagator non-standard.
35
we find
Z∞
1
dτ iΨ†R ∂τ ΨR − iΨ†L ∂τ ΨL + ∂τ Ψ†L ∂τ ΨL + ∂τ Ψ†R ∂τ ΨR
H== 2
4β|µ| ch τ
−∞
1 5 2 † †
+ (1 − th τ ) ΨL ΨL + ΨR ΨR .
8β|µ| ch 2 τ 2
(7.13)
By choosing µ < 0 we have eliminated two problems at once. First, now the hamil-
tonian has no divergence at τ = 0. Second, there is now no relevant boundary at
all, ΨL and ΨR do not mix, and there is no need to impose boundary conditions.
The fact that the two chiralities do not mix to all orders of the string perturbation
theory is remarkable. Its impact on the scattering processes will be discussed after
we introduce the bosonized formalism.
In this section we will explicitly confirm the expectation that the space-time
picture of the c = 1 string theory involves the dynamics of one massless scalar field
in two dimensions. To this end we will bosonize the fields ΨL and ΨR following the
standard bosonization rules for Dirac fermions. 40 A two dimensional free massless
Dirac fermion is equivalent to a single free massless scalar boson. In our case, although
the fermions are free, they are not truly relativistic beyond the semiclassical limit.
This will give rise to interaction terms in the equivalent bosonic field theory.
Zτ
1 √
(P − X ′ )dτ ′ : ,
ΨL = √ : exp i π
2π
(7.14)
Zτ
1 √
(P + X ′ )dτ ′ : ,
ΨR = √ : exp i π
2π
36
explicit calculations of section 6.3. To convert eq. (7.7), we make use of the following
easily derived expressions
i
: Ψ†L ∂τ ΨL − Ψ†R ∂τ ΨR : = : P 2 + (X ′ )2 :
2
X′
: Ψ†L ΨL + Ψ†R ΨR : = − √ (7.15)
π
√ 1 1 ′′′
: ∂τ Ψ†L ∂τ ΨL + ∂τ Ψ†R ∂τ ΨR : = − π : P X ′ P + (X ′ )3 + X : .
3 6π
ZT /2 √
1 π 1 1 ′′′
: H := 2 ′ 2
dτ : P + (X ) − 2 P X ′ P + (X ′ )3 + X −
2 βv 3 6π
0 (7.16)
′′
5(v ′ )2
1 ′ v
√ X − : .
2β π v3 2v 4
If we integrate by parts and discard the boundary terms, this reduces to [10]
ZT /2 √ ′′
(v ′ )2
1 2 ′ 2 π ′ 1 ′ 3 1 ′ v
: H := dτ : P + (X ) − 2 (P X P + (X ) ) − √ X − : .
2 βv 3 2β π 3v 3 2v 4
0
(7.17)
The boundary conditions obeyed by the field X are determined by those of Ψ, eq.
(7.11), which ensure that fermion number not flow out of the τ box. As we have shown,
the current density Ψ̄(τ )γ1 Ψ(τ ) = Ψ†R ΨR − Ψ†L ΨL vanishes at the boundary. Since
this density is proportional to ∂t X(t, τ ), we deduce that X(t, 0) and X(t, T2 ) must
be constant, i.e., X satisfies Dirichlet boundary conditions. The constraint on the
total fermion number requires that X(t, 0) − X(t, T2 ) = 0. We are free to choose
X(t, 0) = X(t, T2 ) = 0. These boundary conditions eliminate all the winding and
momentum modes of X. As a result, there is no need to worry about the periodic
nature of X: it acts like an ordinary scalar field in a box with Dirichlet boundary
conditions.
37
The hamiltonian (7.17) and the boundary condition agree with the collective
field approach.9,41 Physically, the massless field X describes small fluctuations of the
Fermi surface. In the double-scaling limit the hamiltonian reduces to (for µ > 0)
Z∞ √
1 − 23 cth 2 τ
1 2 ′ 2π ′ 1 ′ 3 ′
: H := dτ : P + (X ) − P X P + (X ) − √ X : .
2 2βµ sh 2 τ 3 12βµ π sh 2 τ
0
(7.18)
This hamiltonian, as its fermionic counterpart, suffers from a divergence at τ = 0.
This divergence was regularized in ref. 41 with zeta-function techniques. Alterna-
tively, we may take the double-scaling limit with µ < 0. Then
Z∞ √
1 − 23 th 2 τ
1 2 ′ 2 π ′ 1 ′ 3 ′
: H := dτ : P + (X ) − P X P + (X ) − √ X :,
2 2β|µ| ch 2 τ 3 12β|µ| π ch 2 τ
0
(7.19)
and the sickness at τ = 0 has disappeared. The bosonized theory of non-relativistic
fermions has a remarkable structure. In the double-scaling limit there is a single cubic
interaction term of order gst , and a tadpole of order gst . Therefore, we can develop
an expansion of correlation functions in powers of gst using conventional perturbation
theory. As we will show, this expansion reproduces the genus expansion of the string
amplitudes.
Z∞
1
Z
dxeiqx
dτ e−lλ(τ ) : Ψ†L ΨL + Ψ†R ΨR (τ, t) : , (7.20)
2
0
38
where λ(τ ) is the classical trajectory at the Fermi level. Upon bosonization, we find
Z Z Z∞
O(l, q) ∼ dxeiqx dτ e−lλ(τ ) ∂τ X ∼ i dkF (k, l)k X̃(q, k) (7.21)
−∞
where
Z∞
F (k, l) = dτ e−lλ(τ ) cos(kτ ) ,
0 (7.22)
Z Z
X(x, τ ) = dxe−iqx dk sin(kτ )X(q, k) .
√
Evaluating F (k, l) with λ(τ ) = 2µ ch τ , we find38
p π p p
F (k, l) = Kik (l 2µ) = (I−ik (l 2µ) − Iik (l 2µ)) . (7.23)
2 sin(ikπ)
For small z, we will replace
1
Iν (z) → (z/2)ν . (7.24)
Γ(ν + 1)
39
same factor for each external leg appears in the direct fermionic calculations, (4.15)-
(4.17). The lesson from this rather technical exercise is that the Das-Jevicki field
theory assembles the correlation functions in a rather remarkable way. Each scattering
amplitude of the X-quanta has to be multiplied by the external leg factors which arise
because of the unusual form (7.21) of the external operators.43 On the other hand,
these factors containing the poles have an important string theoretic meaning and
certainly cannot be discarded. Thus, we are still missing the precise interpretation of
the Das-Jevicki field X in the string theory with Euclidean signature.
E Γ(iE)
− exp −i ln µ (7.27)
2 Γ(−iE)
which does not affect any scattering cross-sections. Since the external leg phases are
unobservable, they can be absorbed in the definition of the vertex operators. Thus,
we find that the Minkowskian scattering amplitudes of tachyons are given by the
corresponding amplitudes of the X-quanta.
In fact, we may consider two kinds of scattering experiments. The usual scattering
involves colliding right-moving and left-moving wave packets. In this theory such an
experiment yields trivial results since H = HL + HR 43 , and the wave packets pass
through each other with no influence or time delay.44 Thus, there is no conventional
“bulk” scattering, whose rate is finite per unit spatial volume when plane waves are
being scattered.
We can, however, consider another kind of scattering.44 For µ > 0 we may prepare
a left-moving wave packet incident on the boundary and wait for it to reflect. This
can be interpreted as the scattering of n left-moving particles into m right-moving
particles. (Recall that the number of massless particles in one spatial dimension is
40
a subtle concept that needs to be carefully defined.) For µ < 0, a left- (right- )
moving wave packet stays left- (right- ) moving to all orders of perturbation theory in
gst . However, now there are two asymptotic regions, τ → ±∞, and the wave packet
undergoes some deformation as it passes through the interaction region near τ = 0.
This deformation can be interpreted as a change of state (and number) of particles.
It is this “non-bulk” scattering,44,41,42 whose rate is not proportional to the spatial
volume, that gives, upon Euclidean continuation, the matrix model correlators. To
demonstrate this explicitly, we will now calculate tree-level amplitudes for 2 → 1 and
2 → 2 particle scattering.
We will work with the µ < 0 hamiltonian of eq. (7.19) and show that the calcula-
tions are manifestly finite. Since the chiralities do not mix, we will consider scattering
of right-moving particles described by the hamiltonian
Z∞ " √ #
1 π 1 − 23 th 2 τ
HR = : (P − X ′ )2 + ′ 3
(P − X ) + √ ′
(P − X ) : .
4 6β|µ| ch 2 τ 12β|µ| π ch 2 τ
−∞
(7.28)
Following ref. 41, we will perform our calculation in the hamiltonian formalism. This
is advantageous here because bosonization maps the fermionic hamiltonian into the
normal-ordered bosonic one. Since fermionic calculations are finite, we expect that
the normal ordering will remove all the divergences in the bosonic approach. In other
words, the hamiltonian approach provides a definite set of rules on how to handle the
bosonized theory so that it is perfectly equivalent to the original fermionic theory.
Z∞
dk
a(k)ei(kτ −|k|t) + a† (k)e−i(kτ −|k|t) ,
X(t, τ ) = p
4π|k|
−∞
(7.29)
Z∞
dk
|k| a(k)ei(kτ −|k|t) − a† (k)e−i(kτ −|k|t) ,
P (t, τ ) = −i p
4π|k|
−∞
41
H3 + H1 .
Z∞
H2 = dkka† (k)a(k) ,
0
Z∞
i p
H3 = dk1 dk2 dk3 k1 k2 k3 f (k1 + k2 + k3 )a(k1 )a(k2 )a(k3 )−
24πβ|µ| (7.30)
0
3f (k1 + k2 − k3 ) : a(k1 )a(k2 )a† (k3 ) : + h.c. ,
Z∞ √
i
dk kg(k) a(k) − a† (k) ,
H1 = −
48πβ|µ|
0
where
Z∞
1 πk
f (k) = dτ 2
eikτ = ,
ch τ sh (πk/2)
−∞
(7.31)
Z∞ 3 2τ
1 − th π(k 3 + 2k)
g(k) = 2
eikτ = .
ch 2 τ 4 sh (πk/2)
−∞
S = 1 − 2πiδ(Ei − Ef )T . (7.32)
where |k1 k2 >= a† (k1 )a† (k2 )|0 >. Substituting the cubic interaction term, we find
The fact that this amplitude describes “non-bulk” scattering is evident from the
absence of a separate delta-function for momentum conservation. In order to relate
42
this to the matrix model three-puncture correlation function, we perform Euclidean
continuation Ej → i|qj |. Including the external leg factors according to eq. (7.26),
the result is in agreement with the tree-level contribution to eq. (4.16).
where the sum runs over all the intermediate states i. The s-channel contribution is
easily seen to be41
4 ∞
2 Y
gst
Z 2
kf (k1 + k2 − k) kf 2 (k1 + k2 + k)
(s)
S (k1 , k2 ; k3 , k4 ) = −i Ej dk − ,
8π k1 + k2 − k + iǫ k1 + k2 + k − iǫ
j=1 0
(7.36)
where the first term arises from the one-particle intermediate state, and the second
one – from the five-particle intermediate state. Eq. (7.36) can be expressed as an
integral over k from −∞ to ∞. The t- and u-channel contributions assume a similar
form. Summing over the three channels, we obtain41
4
πg 2 Y
S(k1 , k2 ; k3 , k4 ) = −i st δ(E1 +E2 −E3 −E4 )
Ej F (ps )+F (pt )+F (pu ) , (7.37)
8
j=1
Z∞
(p − k)2
k
F (p) = dk . (7.38)
sh 2 (π(p − k)/2) p − k + iǫ sgn(k)
−∞
Using
1 1
= P + πiδ(x) (7.39)
x − iǫ x
43
x
2
Fig. 4a) The one-loop O gst Fig. 4b) The tadpole graphs
contributions to the 1 → 1 2.
of order gst
amplitude.
we find F (p) = −i 4p
π −
8
3π . The total amplitude thus becomes
2 4
gst Y
S(E1,2 ; E3,4 ) = − δ(E1 +E2 −E3 −E4 ) Ej E1 +E2 +|E1 −E3 |+|E1 −E4 |−2i .
2
j=1
(7.40)
Upon the Euclidean continuation Ej → i|qj |, and inclusion of the external leg factors,
this precisely agrees with the non-relativistic fermion calculation of the 4-puncture
correlator (4.17).
2 k2 Z∞
f 2 (k1 + k2 − k) f 2 (k1 + k2 + k)
18gst
dk1 dk2 k1 k2 − (7.41)
(24π)2 k − k1 − k2 + iǫ k + k1 + k2 − iǫ
0
where the first term is from the 2-particle intermediate state, and the second – from
the 4-particle intermediate state (fig. 4a). After changing variables to s = k1 + k2
and k2 , and integrating over k2 , this reduces to
2 k2 Z∞
f 2 (k − s) 2
3gst 3 gst 2 3 2 8
dss =− k ik + 2k + (7.42)
(24π)2 k − s + iǫ sgn(s) 48π 15
−∞
44
where we evaluated the integral similarly to that of eq. (7.38). This is not the
complete answer because there are also the diagrams of fig. 4b) arising from the
tadpole term H1 . They contribute equally, giving
2 k2 Z∞ Z∞
12gst f (k1)g(k1 ) g2 k2 k14 + 2k12 2
gst 7k 2
dk1 k1 = − st dk1 = − × .
(24π)(48π) −k1 + iǫ 384 sh 2 (πk1 /2) 48π 15
0 0
(7.43)
Adding this to the one-loop contribution (7.42), we find
2
gst
E 2 iE 3 + 2E 2 + 1 .
T (E, E) = − (7.44)
48π
Upon the Euclidean continuation and inclusion of the external leg factors, this agrees
with eq. (4.15).
The agreement with the string one-loop calculation gives us further confidence
that the bosonic string field theory is finite and exact. This implies that the perturba-
tion series exhibits the (2h)! behavior, which is unlike the h! found in the conventional
field theory. This appears to be connected with the lack of translational invariance of
the hamiltonian and of the scattering amplitudes.41 In general, owing to the expo-
nential fall-off of the form-factors f (k) and g(k), the hamiltonian (7.30) is completely
free of ultraviolet divergences. Thus, our innovation of taking the double-scaling limit
with µ < 0 has rendered the entire perturbative expansion manifestly finite. On the
other hand, the end results of our calculations do not change under the continuation
µ → −µ. This suggests that the divergences for µ > 0, connected with the fact that
there f (k) and g(k) did not fall off as k → ∞,41 are simply a result of using an
inconvenient formalism. This is also indicated by the fact that all divergences are
correctly removed by zeta-function regularization.41 In fact, positive and negative µ
are related by the transformation that interchanges the classical coordinate λ with
its conjugate momentum p. Thus, the equation
∂2
(− 21 − 1 λ2 )ψ = −βµψ (7.45)
∂λ2 2
45
can be written in terms of p as
∂2
(− 21 − 1 p2 )ψ = βµψ . (7.46)
∂p2 2
The lesson is that, while for µ < 0 it is convenient to regard λ as the coordinate, for
µ > 0 the natural coordinate is p. If we define the original non-relativistic fermion
theory in terms of Ψ(p, t), then the formalism of this section goes through nicely
for µ > 0: there are no boundaries, and the chiralities do not mix. Thus, all the
calculations above can also be regarded as µ > 0 calculations which are now manifestly
finite. Using this approach, the tree-level 4-point amplitude was calculated in the
lagrangian formalism in ref. 42. Another technique, which works efficiently for all the
tree amplitudes, but is not easily generalized to the loop corrections, was introduced
in ref. 44.
Despite the impressive progress in the formulation of the bosonic string field
theory of the c = 1 quantum gravity, there are many puzzles remaining. In particular,
how should we interpret the coordinate τ and the field X(τ )? Initially, it was argued
that τ is essentially the zero mode of the Liouville field φ. If so, then it is natural
to think of X as the tachyon field. However, it was shown recently that τ is not
locally related to the scale factor, but is instead a conjugate variable in a complicated
integral transform.38 This transform is necessary to turn a string field theory, which is
highly non-local in φ-space, into a simple local hamiltonian in τ -space. Miraculously,
the matrix model has automatically provided the coordinates in which the theory
looks the simplest. However, it does not appear possible to interpret X simply as the
tachyon. Recall that, in addition to the tachyon, the theory has a discrete infinity
of other non-field degrees of freedom. It appears that the theory has soaked up the
discrete degrees of freedom together with the tachyon into a single massless field X.
The form of the matrix model operators for the discrete observables, eq. (6.4), indeed
shows that they are mixed in the X-field together with the tachyon. Clearly, we need
to attain a much better understanding of the precise string theoretic meaning of the
bosonized field theory of the matrix model.
46
8. COMPACT TARGET SPACE AND DUALITY.
In this section we consider the discretized formulation of the sum over surfaces
embedded in a circle of radius R.6 We will adopt the basic definition (2.5), with the
new condition
that ensures the periodicity around the circle. In this model we encounter a new set
of physical issues related to the R-dependence of the sum. The c = 1 conformal field
theory is symmetric under the transformation45
√
α′ α′
R→ ; gst → gst . (8.2)
R R
The change of gst , equivalent to a constant shift of the dilaton background, is necessary
to preserve the invariance to all orders of the genus expansion. A symmetry of the
matter system, such as (8.2), is expected to survive the coupling to gravity. In
this section we will find, however, that in the discretized formulation the duality is
generally broken, and can only be reinstated by a careful fine-tuning. This breaking
of duality is due not to the coupling to gravity, but to the introduction of the explicit
lattice cut-off. Indeed, a generic lattice formulation of the compact c = 1 model on a
fixed geometry has no dual symmetry. The phase transition that separates the small
R (high temperature) phase from the large R (low temperature) phase is well-known
in statistical mechanics as the Kosterlitz-Thouless transition.46 Physically, it is due
to condensation of vortices, configurations that are ignored in the “naive” continuum
limit, but are included in the continuum limit of a generic lattice theory. The vortices
are irrelevant for R > RKT , but they condense and change the behavior of the theory
for R < RKT .
In this section we will demonstrate the lack of duality of the discretized sum over
surfaces. We will also show that, if the lattice sum is modified to exclude the vortices,
then the partition function is explicitly dual. These general results will be of interest
47
when, in the next section, we consider the specific discretized sum generated by the
matrix model. We will be able to isolate the effects of vortices in the matrix model,
and will calculate the dual partition function in the vortex-free continuum limit.
In the conformal field theory the target space duality (8.2) is generated by the
dual transformation on the world sheet47
∂a X → ǫab ∂b X . (8.3)
Its lattice analogue is the transformation from the lattice Λ to the dual lattice Λ̃.
Recall that, when we sum over lattices Λ, we define the target space variables xi at the
centers of the faces of Λ, i.e. at the vertices of λ̃. We will carry out a transformation,
after which the new variables, pI , are defined at the vertices of Λ.
First, to each directed link < ij > of Λ̃ associate the link < IJ > of Λ which
~ points out of the surface (fig.
~ × IJ
intersects it, directed so that the cross product ij
1). Now we define Dij = DIJ = xi − xj . Let us change variables in the integral (2.5)
so that the integral is over DIJ instead of xi . There are E links but, apart from the
⋆
zero mode, there are only V − 1 independent x’s. Therefore, the DIJ ’s are not all
independent but must satisfy F − 1 + 2h constraints. The constraint associated with
P
each face is that hiji Dij = 0, where the sum runs over the directed boundary of
a face of Λ̃. This is equivalent to the condition that the sum of the DIJ ’s emerging
P
from each vertex of Λ must vanish, J DIJ = 0. Similarly, for each independent
P
non-contractable loop on Λ̃, of which there are 2h, we find loop Dij = 0. In terms
of λ this means that hIJi ǫaIJ DIJ = 0, where the symbol ǫaIJ is non-zero only if the
P
link IJ intersects the specially chosen non-contractable loop a on Λ̃. We direct the
~ × ~a points into the surface, and −1 if it points out.
loop a, and define ǫaIJ = 1 if IJ
⋆ Here V , E and F are, respectively, the numbers of vertices, edges and faces of Λ̃, and h is the
genus.
48
each face of Λ̃ and la for each non-contractable loop, the integral over the variables
xi , for each discretization of the surface, can be replaced with
2h
Z Y Z Y " #Z
X Y X
dDIJ G(DIJ ) dpI exp ipI DIJ dla exp ila ǫaIJ DIJ =
hIJi I J a=1 hIJi
Z Y 2h
Y Y
dpI dla G̃(pI − pJ + la ǫaIJ )
I a=1 hIJi
(8.4)
where G̃ is the Fourier transform of G. We see that, after the transformation (8.4),
the Lagrange multipliers pI assume the role of new integration variables residing at
the vertices of Λ. On surfaces of genus > 0 there are 2h additional l-integrations.
Geometrically speaking, one introduces a cut on the surface along each canonically
chosen non-contractable loop, so that the values of p undergo a discontinuity of la
across the a-th cut, and subsequently integrates over la . On a sphere these additional
integrations do not arise, and one simply includes a factor G̃(pI − pJ ) for each link
of the lattice Λ.
The discussion above literally applies to the string on a real line. When G is
defined on a circle of radius R, then the dual variables la and pI assume discrete
n
values R. The integrals over pI and la in eq. (8.4) are replaced by discrete sums,
and the partition function, when expressed in terms of the dual variables, bears little
resemblance to the original expression. In fact, the new variables pI describe the
embedding of the world sheet in the discretized real line with a lattice spacing 1/R.
On a sphere, where la do not arise, the resulting lattice sum is precisely what is
needed to describe string theory with such an embedding. For higher genus, the few
extra variables la spoil the precise equivalence, but this does not alter the essential
“bulk” physics.
We have found that the transformation to the dual lattice, unlike the transforma-
tion (8.3) in the “naive” continuum limit, does not prove R → α′ /R duality. Instead,
it establishes that string theory embedded in a compact dimension of radius R is
essentially equivalent to string theory embedded in a discretized dimension of lattice
49
spacing 1/R. To conclude that R is equivalent to 1/R is just as counterintuitive as
to argue that theories with very small lattice spacing in target space have the same
physics as theories with enormous lattice spacing. In section 11 we will directly study
the theory with the discrete target space. We will show that, if the lattice spacing
is less than a critical value, then the target space lattice is “smoothed out”, but if
it exceeds the critical value, then the theory drastically changes its properties. The
two phases are separated by a Kosterlitz-Thouless phase transition. Equivalently, in
the circle embedding the massless c = 1 phase for R > RKT is separated by the K-T
transition from the disordered (massive) c = 0 phase for R < RKT . The mechanism
of this phase transition involves deconfinement of vortices.46
Consider a fixed geometry ĝµν , and assume that in the continuum limit the action
reduces to eq. (2.2). A vortex of winding number n located at the origin is described
by the configuration
where θ is the azimuthal angle. This configuration is singular at the origin: the values
of X have a branch cut in the continuum limit. On a lattice, however, there is no
singularity, and the vortex configurations are typically included in the statistical sum.
√
If we introduce lattice spacing µ, then the action of a vortex is Sn = n2 R2 | ln µ|/4α′ ,
and it seems that the vortices are suppressed in the continuum limit. We will now
show that this expectation is only true for large enough R. Let us consider the
dynamics of elementary vortices with n = ±1. Although each one is suppressed by
the action, it has a large entropy: there are ∼ 1/µ places on the surface where it
can be found. Thus, the contribution of each vortex or antivortex to the partition
function is of the order
1 −S1 2 ′
e = µ(R /4α )−1 . (8.6)
µ
√
It follows that, for R > 2 α′ , the vortices are irrelevant in the continuum limit. On
√
the other hand, for R < 2 α′ they dominate the partition function, invalidating the
“naive” continuum limit assumed in the conformal field theory.46 In fact, in this
50
phase the proliferation of vortices disorders X and makes its correlations short range.
The critical properties of this phase are those of pure gravity. Is there any way to
salvage the “naive” continuum limit? The answer is yes, but only at the expense of
fine tuning the model, so that the vortices do not appear even on the lattice. Below
we find an explicit example of a model without vortices, and show that it does exhibit
the R → 1/R duality.
∞
X 1 2
G(x) = e− 2 (x+2πmR) (8.7)
m=−∞
where the sum over m renders G periodic under x → x+2πR. Now, eq. (2.5) becomes
V 2πR ∞
dxi Y X − 1 (xi −xj +2πmij R)2
Z
2(h−1)
X X Y
Area
Z(g0 , κ) = g0 κ √ e 2 . (8.8)
h Λ i=1
2π m
hiji ij =−∞
0
where ∂I is the directed boundary of I. After the dual transformation, one obtains a
discrete sum with
1 a 2
G̃(pI − pJ + la ǫaIJ ) = e− 2 (pI −pJ +la ǫIJ )
nI na . (8.10)
pI = , la =
R R
This sum is closely connected to the sum over all possible vortex numbers inside each
face of Λ̃.48
51
We shall now change the definition of the partition function (8.8) to exclude the
vortex configurations. It is helpful to think of mij as link gauge fields. The vortex
numbers vI defined in eq. (8.9) are then analogous to field strengths. If there are no
vortices, the field strength is zero everywhere. However, we still need to sum over all
P
possible windings around non-contractable loops on the lattice Λ̃, lA = loop a mij .
Thus, the space of mij we need to sum over is
mij = ǫA
ij lA + mi − mj , (8.11)
where mi range over all integers and play the role of gauge transformations. ǫA
ij is
defined above and is non-zero only for the links hiji intersecting the specially chosen
non-trivial loop A on the lattice Λ. Thus, lA is the winding number for the non-trivial
cycle a on Λ̃ which intersects the loop A. Summing over mij from eq. (8.11) only,
the partition function becomes
−1 Z∞
VY ∞
X 2(h−1)
X
V dxi X Y − 1 (xi −xj +2πlA ǫAij R)2
Z(g0 , κ) = 2πR g0 κ √ e 2 ,
h Λ i=1 −∞
2π l =−∞ hiji
A
(8.12)
where the factor of 2πR arises from integration over the zero mode of x. After
transforming to the dual lattice we obtain
2h−2 X −1 Z∞
FY ∞
1X g dpI X Y − 1 (pI −pJ +la ǫaIJ /R)2
Z(g0 , κ) = √ 0 κV
√ e 2 .
R 2πR 2π l =−∞ hIJi
h Λ I=1 −∞ a
(8.13)
Thus, after elimination of the vortices, the transformation to the dual lattice clearly
exhibits duality under
√ 1 g0
2πR → √ , g0 → √ . (8.14)
2πR 2πR
Remarkably, the duality is manifest even before the continuum limit is taken. In
the next section we consider the matrix model for random surfaces embedded in a
52
circle. We will find a clear separation between the vortex contributions, and the
“naive” vortex-free continuum sum. This will allow us to calculate explicitly the dual
partition function of the vortex-free model.
The Euclidean matrix quantum mechanics is easily modified to simulate the sum
over discretized random surfaces embedded in a circle of radius R: we simply define
the matrix variable Φ(x) on a circle of radius R, i.e. Φ(x + 2πR) = Φ(x). The path
integral becomes6
2πR 2 !
∂Φ
Z Z
2
Z= D N Φ(x) exp −β dx Tr 1
2 + U(Φ) . (9.1)
∂x
0
∞
X
G(xi − xj ) = e−|xi −xj +2πmR| (9.2)
m=−∞
gives the weight for each link in the random surface interpretation. In terms of the
hamiltonian (3.10), eq. (9.1) is simply a path integral representation for the partition
function
ZR = Tr e−2πRβH , (9.3)
so that 2πR plays the role of the inverse temperature. The problem of calculating
the finite temperature partition function seems to be drastically more complicated
than the zero temperature problem, where only the ground state energy was relevant.
Now we have to know all the energies and degeneracies of states in arbitrarily high
representations of SU(N). We also have to explain the sudden enormous jump in
the number of degrees of freedom as we slightly increase the temperature from zero.
53
In view of the discussion in the last section, the reader should not be surprised if we
claim that the new degrees of freedom, incorporated in the non-trivial representations
of SU(N), are due to the K-T vortices. Since the vortices are dynamically suppressed
for R > RKT , the problem of calculating ZR is not as complicated as it may seem.
Instead of working with a fixed number of fermions N, we will take the well-known
route of introducing a chemical potential µF adjusted so that
Z∞
N 1
κ2 = = ρ(ǫ) dǫ . (9.4)
β 1 + e2πRβ(ǫ−µF )
0
∂F
= µF . (9.5)
∂N
Z∞
∂∆ ∂ 1
= π ρ̃(µ) = π deρ(e)
∂µ ∂µ 1 + e2πRβ(µ−e)
−∞ (9.6)
∂F 1
= βµ ,
∂∆ π
to emphasize resemblance with eqs. (3.18) and (3.19) in the R = ∞ case. Indeed,
we will calculate F (∆) using the same method. The change introduced by a finite R
only affects the first of the equations, determining µ(∆). We can think of ρ̃(µ) as the
54
temperature-modified density of states. Differentiating it, we find
Z∞
1 ∂ ρ̃ 1 ∂ρ πRβ
= de 2
. (9.7)
β ∂µ β ∂e 2 cosh [πRβ(µ − e)]
−∞
Now, substituting the integral representation (3.26) and performing the integral over
e, we find
Z∞
1 ∂ ρ̃ 1 T /2 T /2R
= Im dT e−iβµT , (9.8)
β ∂µ π sinh(T /2) sinh(T /2R)
0
up to terms O(e−βµ ) and O(e−βµ2πR ) which are invisible in the large-β asymptotic
expansion. Thus, this integral representation should only be regarded as the generator
of the correct expansion in powers of gst . Integrating this equation and fixing the
integration constant to agree with the WKB expansion, we find
Z∞
∂∆ dt −it t/2βµ t/2βµR
= Re e . (9.9)
∂µ t sinh t/2βµ sinh(t/2βµR)
µ
To demonstrate the duality in the genus expansion, consider the asymptotic ex-
pansion of eq. (9.9)
∞
" #
∂∆ X √ −2m
= − ln µ + 2βµ R fm (R) ,
∂µ
m=1
m
(9.11)
X
2k 2(m−k)
|B2k | |B2(m−k) | m−2k
fm (R) = (2m − 1)! |2 − 2| |2 − 2| R .
(2k)![2(m − k)]!
k=0
55
10 + 7R−2 ), and so forth. Solving for µ(∆) and integrating eq. (9.6) we find the sum
over connected surfaces Z = −2πRβF :
∞
( )
1 √ 2 X fm+1 (R) √ −2m
Z= (2βµ0 R) ln µ0 − 2f1 (R) ln µ0 + (2βµ0 R) .
4 m(2m + 1)
m=1
(9.12)
Comparing this with eq. (3.30), valid in the case of infinite radius, we find that the
coupling constant 1/(βµ0) has been replaced by
1
gef f (R) = √ , (9.13)
βµ0 R
and that the contribution of each genus h contains the function fh (R). Since both
gef f (R) and fh (R) are invariant under the dual transformation (9.10), we have con-
firmed that the genus expansion of the sum over surfaces is manifestly dual. As in the
critical string theory, the effective coupling constant depends on the radius. Indeed,
the transformation R → 1/R keeping βµ0 fixed is not a symmetry, as it interchanges
a weakly and a strongly coupled theory.
1
From eq. (9.12) we find that the sum over toroidal surfaces is − 12 (R + 1/R) ln µ0 .
This agrees with the direct calculation in the continuum, eq. (5.16). The extra
factor of 2 in the matrix model is due to the doubling of free energy for models with
symmetric potentials. To get the basic sum over triangulated surfaces, we have to
divide eq. (9.12) by 2, which then gives perfect agreement with Liouville theory.
This provides additional evidence that the singlet free energy in the matrix model
evaluates the sum over surfaces in the vortex-free continuum limit. To see this more
explicitly, we have to consider the non-singlet corrections to the free energy and show
that, like the vortices, they are irrelevant for R > RKT .
This problem was discussed in ref. 12, where it was argued that the total partition
function can be factorized as
!
X
Tr e−2πRβH = Trsinglet e−2πRβH Dn e−2πRβδEn (9.14)
n
where δEn is the energy gap between the ground state and the lowest state in the nth
56
representation, and Dn is the degeneracy factor. Let us discuss the leading correction
to the free energy coming from the adjoint representation. Although the degeneracy
factor Dadj = N 2 − 1 diverges as N → ∞, an estimate of ref. 12 shows that the
energy gap also diverges in the continuum limit,
βδEadj = c| ln µ| . (9.15)
1
is negligible for R > πc in the continuum limit where Nµ is kept constant. Taking
the logarithm of eq. (9.14), we find that the total free energy is
F = Fsinglet + O N 2 µ2πRc + . . . (9.17)
Thus, on the one hand, the higher representations are enhanced by enormous degen-
eracy factors; on the other hand, they are suppressed by energy gaps which diverge
logarithmically in the continuum limit. This struggle of entropy with energy is pre-
cisely of the same type as occurs in the physics of the Kosterlitz-Thouless vortices, as
discussed above. In fact, we will now argue that the leading non-singlet correction in
eq. (9.17) is associated with a single vortex-antivortex pair. We perturb the action
of the continuum theory with the operator which creates elementary vortices and
anti-vortices,
R
Z
d2 σ
p
Ov = ĝ cos (X L − X R ) (9.18)
α′
where XL,R are the chiral components of the scalar field X. In string theory language,
this is the sum of vertex operators which create states of winding number 1 and −1.
The insertion of such an operator on a surface creates an endpoint of a cut in the
2
R
values of X. The conformal weight of Ov is h = h̄ = 4α ′ ; therefore, Ov becomes
√ √
relevant for R < 2 α′ , causing a phase transition at RKT = 2 α′ 49 . The phase
transition is connected with the instability towards creation of cuts in the values of
X, i.e., with vortex condensation.
57
This picture of the Kosterlitz-Thouless phase transition can be easily adapted
for coupling to two-dimensional quantum gravity. The sum over genus zero surfaces
coupled to the periodic scalar reduces to the path integral
R
Z Z
2 2
p βφ
Z =N [DX][Dφ] exp −SL − S0 − k d σ ĝe cos ′ (XL − XR ) (9.19)
α
where SL is the Liouville gravity action for c = 1, and β = −2 + √Rα′ so as to make the
conformal weight of the perturbing operator (1, 1). The gravitational dimension of
√
the perturbation is then √R ′ −1. For small values of the cut-off µ, and for R > 2 α′ ,
2 α
Z √ √
2 2 R/ α′ 4 (2R/ α′ )−2
= −µ R ln µ + O k µ + O k µ + ... , (9.20)
N2
√
which demonstrates that the Kosterlitz-Thouless transition occurs at Rc = 2 α′ , the
same value as in a fixed gravitational background. The term O(k 2 ) in Z originates
from the 2-point function of dressed winding mode vertex operators, i.e. from surfaces
with one cut. This leading correction has the same form as eq. (9.17) found in
the context of matrix quantum mechanics. The further corrections, coming from
additional vortex-antivortex pairs, are even more suppressed for R > RKT in the
continuum limit.
For R < RKT , however, the expansion in k diverges badly, indicating an instabil-
ity with respect to condensation of vortices. In this phase we expect the field X to
become massive, and the continuum limit of the theory should behave as pure gravity.
In the matrix model this is easy to show for very small R. The argument is standard:
at very high temperature a d + 1-dimensional theory reduces to d-dimensional theory
because the compact Euclidean dimension of length 2πR = 1/T becomes so small that
variation of fields along it can be ignored. Thus, for small R, the matrix quantum
mechanics (9.1) reduces to the integral over a matrix (zero-dimensional theory)50
Z
2
D N Φ exp[−2πRβ Tr U(Φ)] , (9.21)
58
We have accumulated a considerable amount of evidence that the non-singlet
wave functions of the matrix quantum mechanics implement the physical effects of
the Kosterlitz-Thouless vortices. For R > RKT their corrections to the partition
function are negligible, and the singlet free energy (9.12) gives the sum over surfaces.
For R < RKT the non-singlets dominate the partition function, and the matrix model
no longer describes c = 1 conformal field theory coupled to gravity. In this phase the
singlet free energy gives the sum over surfaces in the vortex-free “naive” continuum
limit. As we showed in section 8, this continuum limit can be achieved at the expense
of a careful fine tuning of the lattice model. In the matrix model the equivalent
operation turns out to be simple: just discard the non-singlet states. We conclude,
therefore, that for any R the singlet sector of the matrix quantum mechanics describes
c = 1 conformal field theory coupled to gravity. This remarkable fact means that, like
the R = ∞ theory, the compact theory is exactly soluble in terms of free fermions. In
the next section, we will use this to calculate compact correlation functions exactly
to all orders in the genus expansion.
The free fermions are now described by the thermal Euclidean second-quantized
action
Z∞ 2πR
d d2 λ2
Z
S= dλ dxψ̂ † (− + 2+ + βµ)ψ̂ , (10.1)
dx dλ 4
−∞ 0
59
and the fermion field satisfies the antiperiodic boundary conditions ψ̂(λ, x + 2πR) =
−ψ̂(λ, x). Compactifying the Euclidean time is a standard device for describing the
theory at a finite temperature T = 1/(2πR). The thermal vacuum satisfies
† 1
hβµ|aǫ (ν)aǫ′ (ν)|βµi = δǫǫ′ f (ν) ≡ δǫǫ′ (10.2)
e(βµ−ν)/T +1
where f (ν) is the Fermi function and µ is the Fermi level. The thermal Euclidean
Green function may be obtained by replacing θ(ν − βµ) by f (ν) in the T = 0 Green
function of eq. (4.10),
Z∞
S E (x1 , λ1 ; x2 , λ2 ) = dνe−(ν−βµ)∆x {θ(∆x)f (ν) − θ(−∆x)(1 − f (ν))}×
−∞
sgn(ω
(10.3)
Z n)∞
i X
ψ ǫ (ν, λ1 )ψ ǫ (ν, λ2 ) = e−iωn ∆x dse−sωn +iβµs hλ1 |e2isH |λ2 i .
2πR ωn 0
The only change compared to the R = ∞ formula (4.10) is that the allowed fre-
quencies become discrete, ωn = (n + 12 )/R, where n are integers. This ensures the
antiperiodicity of SE around the compact direction.
60
as q1 , q2 , . . . , qn . Factoring out the delta-function, the contribution of this diagram is
sgn(ω
Z n)∞ sgn(ωZn +q1 )∞
in X
dα1 e−α1 ωn +iβµα1 hλn |e2iα1 H |λ1 i dα2 e−α2 (ωn +q1 )+iβµα2 ×
R ω
n
0 0
Pn−1
sgn(ωn + 1
qi )∞
Z Pn−1
2iα2 H
hλ1 |e |λ2 i . . . dαn e−αn (ωn + 1
qi )+iβµαn
hλn−1 |e2iαn H |λn i
0
.
(10.4)
It is necessary to break the sum over ωn into parts where the sgn functions in the
exponents are constant. Thus, we form pk = − k1 qk and put the p’s in the increasing
P
order. If p and p′ are two consecutive p’s, then the summation that needs to be
performed is, defining ξ = n1 αi ,
P
p′ − 2R
1
′
1 X
−ωn ξ e−pξ − e−p ξ ξ/2R
e = . (10.5)
R 1
ξ sh (ξ/2R)
ωn =p+ 2R
This factor multiplies the integrand of an n-fold αi -integral. Comparing this with
the R = ∞ case, we find that there the loop momentum integration between p and
′
p′ gives the factor (e−pξ − e−p ξ )/ξ multiplying the same n-fold integrand. It follows
that, in every single contribution to the final answer, the compactification of the
ξ/2R
target space simply inserts the extra factor sh (ξ/2R)
into the integrand. Thus, from
eq. (4.12) we arrive at
∞ ǫ1 ∞
1 ∂ X Z ξ/2R
Z
GR (q1 , λ1 ; . . . ; qn , λn ) = in+1 RδP qi dξeiβµξ ds1 . . .
β ∂µ sh (ξ/2R)
σ∈Σn −∞ 0
ǫn−1
Z ∞
σ σ
Pn−1
dsn−1 e−s1 Q1 −...−sn−1 Qn−1 hλσ(1) |e2is1 H |λσ(2) i . . . hλσ(n) |e2i(ξ− 1
si )H
|λσ(1) i
0
(10.6)
This formula allows us to calculate correlation functions of any set of operators of type
(4.7) in the compact c = 1 model. The integral representation of any such correlator
61
is obtained from that in the non-compact case by insertion of the factor shξ/2R
(ξ/2R) into
P
the integrand, and by replacement δ( qi ) → RδP qi . A special case of this rule is
evident in passing from the R = ∞ equation (3.16) to the finite R equation (9.8).
X
hO1 (q1 ) . . . On (qn )iR=∞ = δ( qi )F (q1 , . . . , qn ; µ, β)
then
hO1 (q1 ) . . . On (qn )iR = RδP qi FR (q1 , . . . , qn ; µ, β) ,
1 ∂
2Rβ ∂µ (10.7)
FR (q1 , . . . , qn ; µ, β) = 1 ∂
F (q1 , . . . , qn ; µ, β) .
sin( 2Rβ ∂µ )
This remarkable relation exists because the entire dependence on µ is through the
factor eiβµξ in the integrands of eqs. (4.12) and (10.6). Therefore, a function f (ξ)
can be introduced into the integrand simply by acting on the integral with operator
f (− βi ∂µ
∂
).
1 ∂ ∞ 2k
2Rβ ∂µ
X (1 − 2−2k+1 )|B2k | ∂
1 ∂
=1+ . (10.8)
sin( 2Rβ ∂µ )
(βR)2k (2k)! ∂µ
k=1
Substituting this expansion into eq. (10.7), we determine the finite R genus g corre-
62
lators in terms of the R = ∞ genus ≤ g correlators,
FRg=0(qi ; µ) = F g=0(qi ; µ) ,
1 ∂ 2 g=0
FRg=1(qi ; µ) = F g=1(qi ; µ) + F (qi ; µ) ,
24R2 ∂µ2
g=2 g=2 1 ∂ 2 g=1 7 ∂ 4 g=0
FR (qi ; µ) = F (qi ; µ) + F (qi ; µ) + F (qi ; µ) , etc.
24R2 ∂µ2 5760R4 ∂µ4
(10.9)
In particular, apart from the discretization of momenta and the change in the overall
δ-function, all spherical correlation functions are not affected by the compactification.
This was, of course, expected in the Liouville approach. The advantage of the matrix
model formalism is that the entire genus expansion of any correlator can be calculated.
Surprisingly, as we have demonstrated, the compact case is hardly any more difficult
than the non-compact case. From the point of view of the Liouville approach, this
conclusion is quite remarkable. There the finite R calculations have the additional
complication of having to perform 2h sums over the winding numbers around the non-
trivial cycles of a genus h surface. Thus, eqs. (10.7) and (10.9), which are completely
general, pose a new challenge for the Liouville approach.
Below we give genus expansions for some correlation functions of puncture oper-
ators P (q). For the two-point function up to genus three we get
" √
2 |q| 1 ( Rβµ)−2
hP (q)P (−q)i = −R[Γ(1 − |q|)] µ − (|q| − 1)×
|q| 24
√ 3
( Rβµ)−4 Y
2 1
R(q − |q| − 1) − + (|q| − r)×
R 5760
r=1
2 4 3 2 2 7
R (3q − 10|q| − 5q + 12|q| + 7) − 10(q − |q| − 1) + 2
R
√ 5
( Rβµ)−6 Y
− (|q| − r) R3 (9q 6 − 63|q|5 + 42q 4 + 217|q|3 − 205|q| − 93)
2903040
r=1
#
4 3 2 147 2 93
− 21R(3q − 10|q| − 5q + 12|q| + 7) + (q − |q| − 1) − 3 + . . .
R R
(10.10)
We see that the external leg factors are the same as in the non-compact case. The
63
relation between the tachyon operators T (q) and P (q) is also the same, eq. (5.8). For
the same reason as in the non-compact case, eq. (10.10) is, strictly speaking, valid
only for non-integer q, where we may replace µ by µ0 . As in the sum over surfaces, we
find that the effective string coupling constant is gef f ∼ √ 1 . The R-dependence
Rβµ0
of the correlator at each genus is not dual: the duality is broken by insertions of
momentum into the world sheet. In the limit qi → 0 the duality is restored, and the
tachyon correlators reduce to derivatives of the sum over surfaces Z(µ0 , R),
∂2
hT (0)T (0)i ∼ Z(µ0 , R) . (10.11)
∂∆2
The three-point function for the case q1 , q2 > 0 is, up to genus two,
3 "
Y Γ(1 − |qi |) |qi | 1
hT (q1 )T (q2 )T (q3 )i = RδP qi µ 2 1
Γ(|qi |) βµ
i=1
√ 4
( Rβµ)−2
Y
2 2 1
− (|q3 | − 1)(|q3 | − 2) R(q1 + q2 − |q3 | − 1) − + (|q3 | − r)×
24 R
r=1
2 4 4 2 2 2 2 2
R 3(q1 + q2 ) + 10q1 q2 − 10(q1 + q2 )|q3 | − 5(q1 − q2 ) + 12|q3 | + 7
√ −4
#
7 ( Rβµ)
− 10(q12 + q22 − |q3 | − 1) + 2 + ...
R 5760
(10.12)
In the case of the four-point function we must consider two independent kinematic
configurations which will give amplitudes that are not related by analytic continuation.19 The
simplest case is when q1 , q2 , q3 > 0,
64
4 "
Y Γ(1 − |qi |) |qi |
hT (q1 )T (q2 )T (q3 )T (q4 )i = −RδP qi µ 2 (βµ)−2 (|q4 | − 1)
Γ(|qi |)
i=1
√ 3
Rβµ)−2 Y
( 2 2 2 1
− (|q4 | − r) R(q1 + q2 + q3 − |q4 | − 1) −
24 R
r=1
√ 5
( Rβµ)−4 Y
+ (|q4 | − r) R2 3(q14 + q24 + q34 ) + 10(q12 q22 + q12 q32 + q22 q32 )
5760
r=1
− 10|q4 |(q12 + q22 + q32 ) − 5(q12 + q22 + q32 ) + 10(q1 q2 + q1 q3 + q2 q3 ) + 12|q4 | + 7
#
7
− 10(q12 + q22 + q32 − |q4 | − 1) + 2 + ...
R
(10.13)
The other kinematic configuration is q1 , q2 > 0, q3 , q4 < 0 with min(|qi |) = q1 ,
max(|qi |) = q2 . This leads to a much more complicated formula, so we just give the
answer out to genus one
4
" √
( Rβµ)−2
Y Γ(1 − |qi |) |qi | q2 − 1
hT (q1 )T (q2 )T (q3 )T (q4 )i = −RδP qi µ 2 1−
Γ(|qi |) (βµ)2 24
i=1
× R S 3 (S − 2q2 − 6) + q22 (S 2 + q32 + q42 ) + 13S 2 q2 − 8Sq22 − 5q2 (q32 + q42 )
#
1
+ 6S(S − 2q2 ) + 6(q32 + q42 ) + 10q22 + S − 2q2 − 6 − (S − 2)(S − 3) + . . .
R
(10.14)
where we have defined S = 21 |qi |.
P
Using the matrix model, we have directly calculated the correlation functions of
the tachyon operators for the momentum modes,
1
Z
d2 σ ĝeiq(X+X̄) e(−2+|q|)φ ,
p
T (q) = (10.15)
β
where X and X̄ are the holomorphic and anti-holomorphic parts of the X-field. In
string theory we also have the winding operators
1
Z
d2 σ ĝeiq̃(X−X̄) e(−2+|q̃|)φ ,
p
T̃ (q̃) = (10.16)
β
65
where q̃ is quantized as nR. Although these operators are hard to introduce in the
matrix model directly, their correlators are obtained from those of the momentum
operators through the dual transformation
1
X + X̄ → X − X̄ , R→ , β → βR. (10.17)
R
2 "
D E Γ(1 − |q̃|) 1
T̃ (q̃)T̃ (−q̃) =R µ|q̃|
Γ(|q̃|) |q̃|
√ # (10.18)
( Rβµ)−2
1 2
− (|q̃| − 1) (q̃ − |q̃| − 1) − R + . . . ,
24 R
etc. In general, if
X
hT (q1 ) . . . T (qn )iR=∞ = δ( qi )F (q1 , . . . , qn ; µ, β) , then
D E 1 ∂ (10.19)
2β ∂µ
T̃ (q̃1 ) . . . T̃ (q̃n ) = Rn−1 δP q̃i 1 ∂
F (q̃1 , . . . , q̃n ; µ, βR) .
R sin( 2β ∂µ )
This formula was obtained by carrying out the dual transformation on the similar
formula (10.7) for the momentum states. Since we do not have a direct matrix model
calculation of the winding operator correlations, we do not yet know the expressions
for correlators where both the momentum and the winding operators are present.
Is there any real time interpretation of the exact correlation functions we have
calculated? In the R = ∞ case we found that, after the continuation to Minkowski
signature, the external leg factors turn into pure phases, and the correlation functions
become scattering amplitudes of the X-quanta of the Das-Jevicki field theory. We can
perform the same continuation, |qj | → −iEj , for the finite R correlation functions,
expecting to obtain the same field theory at a finite temperature T = 1/(2πR). The
external leg factors again turn into phases and can be absorbed in the definition of
66
the vertex operators, and we find the scattering amplitudes at temperature T . For
instance,
"
S(E1 , E2 ; E3 ) = −gstr δ(E1 + E2 − E3 ) E1 E2 E3 1+
# (10.20)
2
gstr
(1 + iE3 )(2 + iE3 ) 1 − iE3 + E12 + E22 + (2πT )2 + . . .
24
67
11. DISCRETIZED TARGET SPACE.
In this section we will study another interesting model: the theory of surfaces
embedded in a discretized real line with lattice spacing ǫ. As shown in section 8,
the lattice duality transformation relates it to the model of surfaces embedded in a
circle of radius 1/ǫ. Although for genus > 0 this transformation generates 2h extra
variables which violate the precise equivalence of the two theories, the basic physical
effect – the Kosterlitz-Thouless phase transition – is of the same nature.
In fact, the model with discretized target space is very interesting in its own right.
Here we can test how string theory responds to introduction of a small lattice spacing
in space-time. It has been argued52 that in this respect string theory is very different
from any field theory: if the lattice spacing is smaller than some critical value, then
the theory on a lattice is precisely the same as the theory in the continuum. This is to
be contrasted with any known field theory, where the continuum behavior may only
be recovered as the lattice spacing is sent to zero. The exactly soluble c = 1 string
theory is an ideal ground for testing this remarkable stringy effect. Below we give the
exact solution of the model with the discretized target space and confirm that, for
ǫ < ǫc , the lattice is smoothed out so that the theory is identical to the embedding
in R1 .
It is easy to check that the perturbative expansion of this integral gives the statistical
sum of the form
V −1
ln Z(ǫ) X 2(h−1) X V Y X Y −ǫ|ni −nj |
lim = g0 κ e , (11.2)
M →∞ M n
h Λ i=1 i hiji
which is precisely what is needed to describe the embedding in a discretized real line
with lattice spacing ǫ. As in the matrix quantum mechanics (2.6), the integral (11.1)
68
can be expressed in terms of the eigenvalues of the matrices Φi . The only modification
here is that, instead of quantum mechanics of N identical non-interacting fermions,
we now find quantum mechanics with a discrete time step ǫ. Thus,
N
ln Z(ǫ) X
lim = ln µi , (11.3)
M →∞ M
i=1
Z∞
µi fi (x) = dyK(x, y)fi(y),
−∞ (11.4)
1/2
− y)2
β β (x
K(x, y) = exp − + ǫ W (x) + W (y) .
2πǫ 2 ǫ
Although it is hard to solve this problem exactly, as usual things simplify in the con-
tinuum limit where we expect to find universal behavior controlled by the quadratic
maximum of W (x). Indeed, if we take
√
and rescale x β = z, we find
(z − w)2 1
1 2 2 1 3 3
K(z, w) = √ exp − + ǫ z +w +O √ (z + w ) . (11.6)
2πǫ 2ǫ 4 β
Thus, the terms beyond the quadratic order are suppressed by powers of β and are,
therefore, irrelevant. The quadratic problem can be solved exactly through finding the
equivalent quantum mechanical problem with Planck constant 1/β ′ and hamiltonian
H(ǫ) such that
′
K(x, y) = hx| e−ǫβ H(ǫ) |yi . (11.7)
Then, µi = exp(−ǫβ ′ ei ) where ei are the N lowest eigenvalues of H(ǫ). For the
69
quadratic transfer matrix,
1/2
(x − y)2 1 2
β β
K(x, y) = exp − − 2 ǫ(x + y 2 ) , (11.8)
2πǫ 2 ǫ
p2 ω 2 x2
the hamiltonian is also quadratic H = 2 − 2 . Comparing
1/2
mωβ ′ mωβ ′
−ǫβ ′ H 2 2 −1
hx| e |yi = exp − (x + y ) cot ωǫ − 2xy sin ωǫ
2π sin ωǫ 2
(11.9)
with eq. (11.8), we find
sin(ωǫ)
cos ǫω = 1 − 12 ǫ2 , β′ = β . (11.10)
ωǫ
Thus, we have reduced the problem of random surfaces embedded in discretized target
space to quantum mechanics of N fermions moving in an inverted harmonic oscillator
potential, i.e. to random surfaces embedded in R1 ! Eq. (11.10) dictates how the cur-
1
vature of the potential and β ′ depend on ǫ. Recalling that α′ = 1/ω 2 and gst = β ′ µ0 ,
we find that the target space scale and the coupling constant of the equivalent con-
tinuum string theory depend on ǫ. Thus, changing the lattice spacing in target space
simply changes the parameters of the equivalent string theory embedded in continu-
ous space. This effect was originally demonstrated in higher-dimensional string theory
using completely different techniques.52 It appears to be a general property of string
theory.
Intuitively, we do not expect the equivalence with the continuum string theory
to be there for arbitrarily large ǫ. Indeed, eq. (11.10) shows that, if ǫ exceeds ǫc = 2
then the solution for ω becomes complex, which is a sign of instability. Since large
lattice spacing corresponds, in the dual language, to small radius, it is clear that the
instability is with respect to condensation of vortices. Thus, for ǫ > ǫc , we expect
the vortices to eliminate the massless field on the world sheet corresponding to the
70
embedding dimension, so that the pure gravity results. This is easy to check for very
large ǫ where the sites of the matrix chain become decoupled,
ln Z(ǫ)
Z
2
lim → D N Φe−βǫ Tr W (Φ) , (11.11)
M →∞ M
and we obtain the one-matrix model well-known to describe the pure gravity.
√
ǫc = π α′ (11.12)
We can compare this result with the position of the K-T transition in the matrix
√
chain model, where we found ǫc = 2 and α′ (ǫc ) = 2/π. The agreement of these
values with the continuum relation (11.12) strengthens our explanation of the phase
transition in the matrix model. It is satisfying that we have found the matrix model
confirmation of a general effect, smoothing out of a target space lattice, which applies
to string theories in any dimension.
71
12. CONCLUSION.
I hope that I have convinced the reader that the 2-dimensional string theory is
a highly non-trivial toy model for string theory in higher dimensions. Its matrix
model formulation as a sum over surfaces embedded in 1 dimension is an example of
a perfectly regularized generally covariant definition of the Polyakov path integral.
It also turns out to be remarkably powerful, giving us the exact solution of a non-
trivial string theory. Many of its physical features, such as the R → 1/R duality
(and its breaking due to the vortices), smoothing out of the target space discreteness,
presence of poles in the correlation functions, etc., carry over to string theories in
higher dimensions. Also, even though this string theory is two-dimensional, it has
some interesting remnants of transverse excitations, manifested in the isolated states
at integer momenta which generate the W1+∞ algebra. Thanks to the matrix models,
we have acquired a wealth of exact information on the perturbative properties of
this string theory. Unfortunately, we still do not have a detailed understanding or
interpretation of most of it in the conventional continuum formalism. Hopefully, this
gap will be closed in the not too distant future.
72
REFERENCES
2. D. J. Gross and A. A. Migdal, Phys. Rev. Lett. 64 (1990) 717; M. Douglas and
S. Shenker, Nucl. Phys. B335 (1990) 635; E. Brezin and V. Kazakov, Phys.
Lett. 236B (1990) 144.
3. D. J. Gross and A. Migdal, Phys. Rev. Lett. 64 (1990) 717; E. Brezin, M. Dou-
glas, V. Kazakov and S. Shenker, Phys. Lett. 237B (1990) 43; C. Crnkovic,
P. Ginsparg and G. Moore, Phys. Lett. 237B (1990) 196.
7. S. Das, S. Naik and S. Wadia, Mod. Phys. Lett. A4 (1989) 1033; J. Polchinski,
Nucl. Phys. B324 (1989) 123; S. Das, A. Dhar and S. Wadia, Mod. Phys.
Lett. A5 (1990) 799; T. Banks and J. Lykken, Nucl. Phys. B331 (1990) 173;
A. Tseytlin, Int. Jour. Mod. Phys. A5 (1990) 1833.
73
13. I. R. Klebanov and D. Lowe, Princeton preprint PUPT-1256, to appear in Nucl.
Phys. B.
15. F. David, Mod. Phys. Lett. A3 (1988) 1651; J. Distler and H. Kawai, Nucl.
Phys. B321 (1989) 509.
20. D. J. Gross, I. R. Klebanov and M. J. Newman, Nucl. Phys. B350 (1991), 621.
25. S. Gupta, A. Trivedi and M. Wise, Nucl. Phys. B340 (1990) 475.
28. M. Bershadsky and I. R. Klebanov, Phys. Rev. Lett. 65 (1990), 3088; N. Sakai
and Y. Tanii, Tokyo Inst. of Tech. preprint TIT/HEP-160 (1990).
30. B. McClain and B. D. B. Roth, Comm. Math. Phys. 111 (1987) 539; K. H.
O’Brien and C.-I. Tan, Phys. Rev. D36 (1987) 1184.
74
33. B. Lian and G. Zuckerman, Yale preprint YCTP-P18-91.
34. R. Dijkgraaf, E. Verlinde and H. Verlinde, Comm. Math. Phys. 115 (1988) 649.
38. G. Moore and N. Seiberg, Rutgers and Yale preprint RU-91-29, YCTP-P19-91.
39. A. Sengupta and S. Wadia, Int. Jour. Mod. Phys. A6 (1991) 1961.
40. S. Coleman, Phys. Rev. D11 (1975) 2088; S. Mandelstam, Phys. Rev. D11
(1975) 3026.
45. P. Ginsparg and C. Vafa, Nucl. Phys. B289 (1987) 414; E. Alvarez and M.
Osorio, Phys. Rev. D40 (1989), 1150.
47. J. Molera and B. Ovrut, Phys. Rev. D40 (1989), 1146; M. J. Duff, preprint
CTP-TAMU-53/89.
49. Ya. I. Kogan, JETP Lett. 45 (1987), 709; B. Sathiapalan, Phys. Rev. D35
(1987), 3277.
75
50. Z. Yang, Phys. Lett. 243B (1990) 365.
76